Initial import

58 files changed, 27459 insertions, 0 deletions

diff --git a/arch/tile/kernel/Makefile b/arch/tile/kernel/Makefile
new file mode 100644
index 000000000..21f77bf68
--- /dev/null
+++ b/arch/tile/kernel/Makefile
@@ -0,0 +1,36 @@
+#
+# Makefile for the Linux/TILE kernel.
+#
+
+extra-y := vmlinux.lds head_$(BITS).o
+obj-y := backtrace.o entry.o hvglue.o irq.o messaging.o \
+	pci-dma.o proc.o process.o ptrace.o reboot.o \
+	setup.o signal.o single_step.o stack.o sys.o \
+	sysfs.o time.o traps.o unaligned.o vdso.o \
+	intvec_$(BITS).o regs_$(BITS).o tile-desc_$(BITS).o
+
+ifdef CONFIG_FUNCTION_TRACER
+CFLAGS_REMOVE_ftrace.o = -pg
+CFLAGS_REMOVE_early_printk.o = -pg
+endif
+
+obj-$(CONFIG_HARDWALL)		+= hardwall.o
+obj-$(CONFIG_COMPAT)		+= compat.o compat_signal.o
+obj-$(CONFIG_SMP)		+= smpboot.o smp.o tlb.o
+obj-$(CONFIG_MODULES)		+= module.o
+obj-$(CONFIG_EARLY_PRINTK)	+= early_printk.o
+obj-$(CONFIG_KEXEC)		+= machine_kexec.o relocate_kernel_$(BITS).o
+ifdef CONFIG_TILEGX
+obj-$(CONFIG_PCI)		+= pci_gx.o
+else
+obj-$(CONFIG_PCI)		+= pci.o
+endif
+obj-$(CONFIG_PERF_EVENTS)	+= perf_event.o
+obj-$(CONFIG_USE_PMC)		+= pmc.o
+obj-$(CONFIG_TILE_USB)		+= usb.o
+obj-$(CONFIG_TILE_HVGLUE_TRACE)	+= hvglue_trace.o
+obj-$(CONFIG_FUNCTION_TRACER)	+= ftrace.o mcount_64.o
+obj-$(CONFIG_KPROBES)		+= kprobes.o
+obj-$(CONFIG_KGDB)		+= kgdb.o
+
+obj-y				+= vdso/
diff --git a/arch/tile/kernel/asm-offsets.c b/arch/tile/kernel/asm-offsets.c
new file mode 100644
index 000000000..375e7c321
--- /dev/null
+++ b/arch/tile/kernel/asm-offsets.c
@@ -0,0 +1,84 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Generates definitions from c-type structures used by assembly sources.
+ */
+
+/* Check for compatible compiler early in the build. */
+#ifdef CONFIG_TILEGX
+# ifndef __tilegx__
+#  error Can only build TILE-Gx configurations with tilegx compiler
+# endif
+# ifndef __LP64__
+#  error Must not specify -m32 when building the TILE-Gx kernel
+# endif
+#else
+# ifdef __tilegx__
+#  error Can not build TILEPro configurations with tilegx compiler
+# endif
+#endif
+
+#include <linux/kbuild.h>
+#include <linux/thread_info.h>
+#include <linux/sched.h>
+#include <linux/hardirq.h>
+#include <linux/ptrace.h>
+#include <hv/hypervisor.h>
+
+void foo(void)
+{
+	DEFINE(SINGLESTEP_STATE_BUFFER_OFFSET,
+	       offsetof(struct single_step_state, buffer));
+	DEFINE(SINGLESTEP_STATE_FLAGS_OFFSET,
+	       offsetof(struct single_step_state, flags));
+	DEFINE(SINGLESTEP_STATE_ORIG_PC_OFFSET,
+	       offsetof(struct single_step_state, orig_pc));
+	DEFINE(SINGLESTEP_STATE_NEXT_PC_OFFSET,
+	       offsetof(struct single_step_state, next_pc));
+	DEFINE(SINGLESTEP_STATE_BRANCH_NEXT_PC_OFFSET,
+	       offsetof(struct single_step_state, branch_next_pc));
+	DEFINE(SINGLESTEP_STATE_UPDATE_VALUE_OFFSET,
+	       offsetof(struct single_step_state, update_value));
+
+	DEFINE(THREAD_INFO_TASK_OFFSET,
+	       offsetof(struct thread_info, task));
+	DEFINE(THREAD_INFO_FLAGS_OFFSET,
+	       offsetof(struct thread_info, flags));
+	DEFINE(THREAD_INFO_STATUS_OFFSET,
+	       offsetof(struct thread_info, status));
+	DEFINE(THREAD_INFO_HOMECACHE_CPU_OFFSET,
+	       offsetof(struct thread_info, homecache_cpu));
+	DEFINE(THREAD_INFO_PREEMPT_COUNT_OFFSET,
+	       offsetof(struct thread_info, preempt_count));
+	DEFINE(THREAD_INFO_STEP_STATE_OFFSET,
+	       offsetof(struct thread_info, step_state));
+#ifdef __tilegx__
+	DEFINE(THREAD_INFO_UNALIGN_JIT_BASE_OFFSET,
+	       offsetof(struct thread_info, unalign_jit_base));
+	DEFINE(THREAD_INFO_UNALIGN_JIT_TMP_OFFSET,
+	       offsetof(struct thread_info, unalign_jit_tmp));
+#endif
+
+	DEFINE(TASK_STRUCT_THREAD_KSP_OFFSET,
+	       offsetof(struct task_struct, thread.ksp));
+	DEFINE(TASK_STRUCT_THREAD_PC_OFFSET,
+	       offsetof(struct task_struct, thread.pc));
+
+	DEFINE(HV_TOPOLOGY_WIDTH_OFFSET,
+	       offsetof(HV_Topology, width));
+	DEFINE(HV_TOPOLOGY_HEIGHT_OFFSET,
+	       offsetof(HV_Topology, height));
+
+	DEFINE(IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET,
+	       offsetof(irq_cpustat_t, irq_syscall_count));
+}
diff --git a/arch/tile/kernel/backtrace.c b/arch/tile/kernel/backtrace.c
new file mode 100644
index 000000000..f8b74ca83
--- /dev/null
+++ b/arch/tile/kernel/backtrace.c
@@ -0,0 +1,683 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/kernel.h>
+#include <linux/string.h>
+#include <asm/byteorder.h>
+#include <asm/backtrace.h>
+#include <asm/tile-desc.h>
+#include <arch/abi.h>
+
+#ifdef __tilegx__
+#define TILE_MAX_INSTRUCTIONS_PER_BUNDLE TILEGX_MAX_INSTRUCTIONS_PER_BUNDLE
+#define tile_decoded_instruction tilegx_decoded_instruction
+#define tile_mnemonic tilegx_mnemonic
+#define parse_insn_tile parse_insn_tilegx
+#define TILE_OPC_IRET TILEGX_OPC_IRET
+#define TILE_OPC_ADDI TILEGX_OPC_ADDI
+#define TILE_OPC_ADDLI TILEGX_OPC_ADDLI
+#define TILE_OPC_INFO TILEGX_OPC_INFO
+#define TILE_OPC_INFOL TILEGX_OPC_INFOL
+#define TILE_OPC_JRP TILEGX_OPC_JRP
+#define TILE_OPC_MOVE TILEGX_OPC_MOVE
+#define OPCODE_STORE TILEGX_OPC_ST
+typedef long long bt_int_reg_t;
+#else
+#define TILE_MAX_INSTRUCTIONS_PER_BUNDLE TILEPRO_MAX_INSTRUCTIONS_PER_BUNDLE
+#define tile_decoded_instruction tilepro_decoded_instruction
+#define tile_mnemonic tilepro_mnemonic
+#define parse_insn_tile parse_insn_tilepro
+#define TILE_OPC_IRET TILEPRO_OPC_IRET
+#define TILE_OPC_ADDI TILEPRO_OPC_ADDI
+#define TILE_OPC_ADDLI TILEPRO_OPC_ADDLI
+#define TILE_OPC_INFO TILEPRO_OPC_INFO
+#define TILE_OPC_INFOL TILEPRO_OPC_INFOL
+#define TILE_OPC_JRP TILEPRO_OPC_JRP
+#define TILE_OPC_MOVE TILEPRO_OPC_MOVE
+#define OPCODE_STORE TILEPRO_OPC_SW
+typedef int bt_int_reg_t;
+#endif
+
+/* A decoded bundle used for backtracer analysis. */
+struct BacktraceBundle {
+	tile_bundle_bits bits;
+	int num_insns;
+	struct tile_decoded_instruction
+	insns[TILE_MAX_INSTRUCTIONS_PER_BUNDLE];
+};
+
+
+/* Locates an instruction inside the given bundle that
+ * has the specified mnemonic, and whose first 'num_operands_to_match'
+ * operands exactly match those in 'operand_values'.
+ */
+static const struct tile_decoded_instruction *find_matching_insn(
+	const struct BacktraceBundle *bundle,
+	tile_mnemonic mnemonic,
+	const int *operand_values,
+	int num_operands_to_match)
+{
+	int i, j;
+	bool match;
+
+	for (i = 0; i < bundle->num_insns; i++) {
+		const struct tile_decoded_instruction *insn =
+			&bundle->insns[i];
+
+		if (insn->opcode->mnemonic != mnemonic)
+			continue;
+
+		match = true;
+		for (j = 0; j < num_operands_to_match; j++) {
+			if (operand_values[j] != insn->operand_values[j]) {
+				match = false;
+				break;
+			}
+		}
+
+		if (match)
+			return insn;
+	}
+
+	return NULL;
+}
+
+/* Does this bundle contain an 'iret' instruction? */
+static inline bool bt_has_iret(const struct BacktraceBundle *bundle)
+{
+	return find_matching_insn(bundle, TILE_OPC_IRET, NULL, 0) != NULL;
+}
+
+/* Does this bundle contain an 'addi sp, sp, OFFSET' or
+ * 'addli sp, sp, OFFSET' instruction, and if so, what is OFFSET?
+ */
+static bool bt_has_addi_sp(const struct BacktraceBundle *bundle, int *adjust)
+{
+	static const int vals[2] = { TREG_SP, TREG_SP };
+
+	const struct tile_decoded_instruction *insn =
+		find_matching_insn(bundle, TILE_OPC_ADDI, vals, 2);
+	if (insn == NULL)
+		insn = find_matching_insn(bundle, TILE_OPC_ADDLI, vals, 2);
+#ifdef __tilegx__
+	if (insn == NULL)
+		insn = find_matching_insn(bundle, TILEGX_OPC_ADDXLI, vals, 2);
+	if (insn == NULL)
+		insn = find_matching_insn(bundle, TILEGX_OPC_ADDXI, vals, 2);
+#endif
+	if (insn == NULL)
+		return false;
+
+	*adjust = insn->operand_values[2];
+	return true;
+}
+
+/* Does this bundle contain any 'info OP' or 'infol OP'
+ * instruction, and if so, what are their OP?  Note that OP is interpreted
+ * as an unsigned value by this code since that's what the caller wants.
+ * Returns the number of info ops found.
+ */
+static int bt_get_info_ops(const struct BacktraceBundle *bundle,
+		int operands[MAX_INFO_OPS_PER_BUNDLE])
+{
+	int num_ops = 0;
+	int i;
+
+	for (i = 0; i < bundle->num_insns; i++) {
+		const struct tile_decoded_instruction *insn =
+			&bundle->insns[i];
+
+		if (insn->opcode->mnemonic == TILE_OPC_INFO ||
+		    insn->opcode->mnemonic == TILE_OPC_INFOL) {
+			operands[num_ops++] = insn->operand_values[0];
+		}
+	}
+
+	return num_ops;
+}
+
+/* Does this bundle contain a jrp instruction, and if so, to which
+ * register is it jumping?
+ */
+static bool bt_has_jrp(const struct BacktraceBundle *bundle, int *target_reg)
+{
+	const struct tile_decoded_instruction *insn =
+		find_matching_insn(bundle, TILE_OPC_JRP, NULL, 0);
+	if (insn == NULL)
+		return false;
+
+	*target_reg = insn->operand_values[0];
+	return true;
+}
+
+/* Does this bundle modify the specified register in any way? */
+static bool bt_modifies_reg(const struct BacktraceBundle *bundle, int reg)
+{
+	int i, j;
+	for (i = 0; i < bundle->num_insns; i++) {
+		const struct tile_decoded_instruction *insn =
+			&bundle->insns[i];
+
+		if (insn->opcode->implicitly_written_register == reg)
+			return true;
+
+		for (j = 0; j < insn->opcode->num_operands; j++)
+			if (insn->operands[j]->is_dest_reg &&
+			    insn->operand_values[j] == reg)
+				return true;
+	}
+
+	return false;
+}
+
+/* Does this bundle modify sp? */
+static inline bool bt_modifies_sp(const struct BacktraceBundle *bundle)
+{
+	return bt_modifies_reg(bundle, TREG_SP);
+}
+
+/* Does this bundle modify lr? */
+static inline bool bt_modifies_lr(const struct BacktraceBundle *bundle)
+{
+	return bt_modifies_reg(bundle, TREG_LR);
+}
+
+/* Does this bundle contain the instruction 'move fp, sp'? */
+static inline bool bt_has_move_r52_sp(const struct BacktraceBundle *bundle)
+{
+	static const int vals[2] = { 52, TREG_SP };
+	return find_matching_insn(bundle, TILE_OPC_MOVE, vals, 2) != NULL;
+}
+
+/* Does this bundle contain a store of lr to sp? */
+static inline bool bt_has_sw_sp_lr(const struct BacktraceBundle *bundle)
+{
+	static const int vals[2] = { TREG_SP, TREG_LR };
+	return find_matching_insn(bundle, OPCODE_STORE, vals, 2) != NULL;
+}
+
+#ifdef __tilegx__
+/* Track moveli values placed into registers. */
+static inline void bt_update_moveli(const struct BacktraceBundle *bundle,
+				    int moveli_args[])
+{
+	int i;
+	for (i = 0; i < bundle->num_insns; i++) {
+		const struct tile_decoded_instruction *insn =
+			&bundle->insns[i];
+
+		if (insn->opcode->mnemonic == TILEGX_OPC_MOVELI) {
+			int reg = insn->operand_values[0];
+			moveli_args[reg] = insn->operand_values[1];
+		}
+	}
+}
+
+/* Does this bundle contain an 'add sp, sp, reg' instruction
+ * from a register that we saw a moveli into, and if so, what
+ * is the value in the register?
+ */
+static bool bt_has_add_sp(const struct BacktraceBundle *bundle, int *adjust,
+			  int moveli_args[])
+{
+	static const int vals[2] = { TREG_SP, TREG_SP };
+
+	const struct tile_decoded_instruction *insn =
+		find_matching_insn(bundle, TILEGX_OPC_ADDX, vals, 2);
+	if (insn) {
+		int reg = insn->operand_values[2];
+		if (moveli_args[reg]) {
+			*adjust = moveli_args[reg];
+			return true;
+		}
+	}
+	return false;
+}
+#endif
+
+/* Locates the caller's PC and SP for a program starting at the
+ * given address.
+ */
+static void find_caller_pc_and_caller_sp(CallerLocation *location,
+					 const unsigned long start_pc,
+					 BacktraceMemoryReader read_memory_func,
+					 void *read_memory_func_extra)
+{
+	/* Have we explicitly decided what the sp is,
+	 * rather than just the default?
+	 */
+	bool sp_determined = false;
+
+	/* Has any bundle seen so far modified lr? */
+	bool lr_modified = false;
+
+	/* Have we seen a move from sp to fp? */
+	bool sp_moved_to_r52 = false;
+
+	/* Have we seen a terminating bundle? */
+	bool seen_terminating_bundle = false;
+
+	/* Cut down on round-trip reading overhead by reading several
+	 * bundles at a time.
+	 */
+	tile_bundle_bits prefetched_bundles[32];
+	int num_bundles_prefetched = 0;
+	int next_bundle = 0;
+	unsigned long pc;
+
+#ifdef __tilegx__
+	/* Naively try to track moveli values to support addx for -m32. */
+	int moveli_args[TILEGX_NUM_REGISTERS] = { 0 };
+#endif
+
+	/* Default to assuming that the caller's sp is the current sp.
+	 * This is necessary to handle the case where we start backtracing
+	 * right at the end of the epilog.
+	 */
+	location->sp_location = SP_LOC_OFFSET;
+	location->sp_offset = 0;
+
+	/* Default to having no idea where the caller PC is. */
+	location->pc_location = PC_LOC_UNKNOWN;
+
+	/* Don't even try if the PC is not aligned. */
+	if (start_pc % TILE_BUNDLE_ALIGNMENT_IN_BYTES != 0)
+		return;
+
+	for (pc = start_pc;; pc += sizeof(tile_bundle_bits)) {
+
+		struct BacktraceBundle bundle;
+		int num_info_ops, info_operands[MAX_INFO_OPS_PER_BUNDLE];
+		int one_ago, jrp_reg;
+		bool has_jrp;
+
+		if (next_bundle >= num_bundles_prefetched) {
+			/* Prefetch some bytes, but don't cross a page
+			 * boundary since that might cause a read failure we
+			 * don't care about if we only need the first few
+			 * bytes. Note: we don't care what the actual page
+			 * size is; using the minimum possible page size will
+			 * prevent any problems.
+			 */
+			unsigned int bytes_to_prefetch = 4096 - (pc & 4095);
+			if (bytes_to_prefetch > sizeof prefetched_bundles)
+				bytes_to_prefetch = sizeof prefetched_bundles;
+
+			if (!read_memory_func(prefetched_bundles, pc,
+					      bytes_to_prefetch,
+					      read_memory_func_extra)) {
+				if (pc == start_pc) {
+					/* The program probably called a bad
+					 * address, such as a NULL pointer.
+					 * So treat this as if we are at the
+					 * start of the function prolog so the
+					 * backtrace will show how we got here.
+					 */
+					location->pc_location = PC_LOC_IN_LR;
+					return;
+				}
+
+				/* Unreadable address. Give up. */
+				break;
+			}
+
+			next_bundle = 0;
+			num_bundles_prefetched =
+				bytes_to_prefetch / sizeof(tile_bundle_bits);
+		}
+
+		/*
+		 * Decode the next bundle.
+		 * TILE always stores instruction bundles in little-endian
+		 * mode, even when the chip is running in big-endian mode.
+		 */
+		bundle.bits = le64_to_cpu(prefetched_bundles[next_bundle++]);
+		bundle.num_insns =
+			parse_insn_tile(bundle.bits, pc, bundle.insns);
+		num_info_ops = bt_get_info_ops(&bundle, info_operands);
+
+		/* First look at any one_ago info ops if they are interesting,
+		 * since they should shadow any non-one-ago info ops.
+		 */
+		for (one_ago = (pc != start_pc) ? 1 : 0;
+		     one_ago >= 0; one_ago--) {
+			int i;
+			for (i = 0; i < num_info_ops; i++) {
+				int info_operand = info_operands[i];
+				if (info_operand < CALLER_UNKNOWN_BASE)	{
+					/* Weird; reserved value, ignore it. */
+					continue;
+				}
+
+				/* Skip info ops which are not in the
+				 * "one_ago" mode we want right now.
+				 */
+				if (((info_operand & ONE_BUNDLE_AGO_FLAG) != 0)
+				    != (one_ago != 0))
+					continue;
+
+				/* Clear the flag to make later checking
+				 * easier. */
+				info_operand &= ~ONE_BUNDLE_AGO_FLAG;
+
+				/* Default to looking at PC_IN_LR_FLAG. */
+				if (info_operand & PC_IN_LR_FLAG)
+					location->pc_location =
+						PC_LOC_IN_LR;
+				else
+					location->pc_location =
+						PC_LOC_ON_STACK;
+
+				switch (info_operand) {
+				case CALLER_UNKNOWN_BASE:
+					location->pc_location = PC_LOC_UNKNOWN;
+					location->sp_location = SP_LOC_UNKNOWN;
+					return;
+
+				case CALLER_SP_IN_R52_BASE:
+				case CALLER_SP_IN_R52_BASE | PC_IN_LR_FLAG:
+					location->sp_location = SP_LOC_IN_R52;
+					return;
+
+				default:
+				{
+					const unsigned int val = info_operand
+						- CALLER_SP_OFFSET_BASE;
+					const unsigned int sp_offset =
+						(val >> NUM_INFO_OP_FLAGS) * 8;
+					if (sp_offset < 32768) {
+						/* This is a properly encoded
+						 * SP offset. */
+						location->sp_location =
+							SP_LOC_OFFSET;
+						location->sp_offset =
+							sp_offset;
+						return;
+					} else {
+						/* This looked like an SP
+						 * offset, but it's outside
+						 * the legal range, so this
+						 * must be an unrecognized
+						 * info operand.  Ignore it.
+						 */
+					}
+				}
+				break;
+				}
+			}
+		}
+
+		if (seen_terminating_bundle) {
+			/* We saw a terminating bundle during the previous
+			 * iteration, so we were only looking for an info op.
+			 */
+			break;
+		}
+
+		if (bundle.bits == 0) {
+			/* Wacky terminating bundle. Stop looping, and hope
+			 * we've already seen enough to find the caller.
+			 */
+			break;
+		}
+
+		/*
+		 * Try to determine caller's SP.
+		 */
+
+		if (!sp_determined) {
+			int adjust;
+			if (bt_has_addi_sp(&bundle, &adjust)
+#ifdef __tilegx__
+			    || bt_has_add_sp(&bundle, &adjust, moveli_args)
+#endif
+				) {
+				location->sp_location = SP_LOC_OFFSET;
+
+				if (adjust <= 0) {
+					/* We are in prolog about to adjust
+					 * SP. */
+					location->sp_offset = 0;
+				} else {
+					/* We are in epilog restoring SP. */
+					location->sp_offset = adjust;
+				}
+
+				sp_determined = true;
+			} else {
+				if (bt_has_move_r52_sp(&bundle)) {
+					/* Maybe in prolog, creating an
+					 * alloca-style frame.  But maybe in
+					 * the middle of a fixed-size frame
+					 * clobbering r52 with SP.
+					 */
+					sp_moved_to_r52 = true;
+				}
+
+				if (bt_modifies_sp(&bundle)) {
+					if (sp_moved_to_r52) {
+						/* We saw SP get saved into
+						 * r52 earlier (or now), which
+						 * must have been in the
+						 * prolog, so we now know that
+						 * SP is still holding the
+						 * caller's sp value.
+						 */
+						location->sp_location =
+							SP_LOC_OFFSET;
+						location->sp_offset = 0;
+					} else {
+						/* Someone must have saved
+						 * aside the caller's SP value
+						 * into r52, so r52 holds the
+						 * current value.
+						 */
+						location->sp_location =
+							SP_LOC_IN_R52;
+					}
+					sp_determined = true;
+				}
+			}
+
+#ifdef __tilegx__
+			/* Track moveli arguments for -m32 mode. */
+			bt_update_moveli(&bundle, moveli_args);
+#endif
+		}
+
+		if (bt_has_iret(&bundle)) {
+			/* This is a terminating bundle. */
+			seen_terminating_bundle = true;
+			continue;
+		}
+
+		/*
+		 * Try to determine caller's PC.
+		 */
+
+		jrp_reg = -1;
+		has_jrp = bt_has_jrp(&bundle, &jrp_reg);
+		if (has_jrp)
+			seen_terminating_bundle = true;
+
+		if (location->pc_location == PC_LOC_UNKNOWN) {
+			if (has_jrp) {
+				if (jrp_reg == TREG_LR && !lr_modified) {
+					/* Looks like a leaf function, or else
+					 * lr is already restored. */
+					location->pc_location =
+						PC_LOC_IN_LR;
+				} else {
+					location->pc_location =
+						PC_LOC_ON_STACK;
+				}
+			} else if (bt_has_sw_sp_lr(&bundle)) {
+				/* In prolog, spilling initial lr to stack. */
+				location->pc_location = PC_LOC_IN_LR;
+			} else if (bt_modifies_lr(&bundle)) {
+				lr_modified = true;
+			}
+		}
+	}
+}
+
+/* Initializes a backtracer to start from the given location.
+ *
+ * If the frame pointer cannot be determined it is set to -1.
+ *
+ * state: The state to be filled in.
+ * read_memory_func: A callback that reads memory.
+ * read_memory_func_extra: An arbitrary argument to read_memory_func.
+ * pc: The current PC.
+ * lr: The current value of the 'lr' register.
+ * sp: The current value of the 'sp' register.
+ * r52: The current value of the 'r52' register.
+ */
+void backtrace_init(BacktraceIterator *state,
+		    BacktraceMemoryReader read_memory_func,
+		    void *read_memory_func_extra,
+		    unsigned long pc, unsigned long lr,
+		    unsigned long sp, unsigned long r52)
+{
+	CallerLocation location;
+	unsigned long fp, initial_frame_caller_pc;
+
+	/* Find out where we are in the initial frame. */
+	find_caller_pc_and_caller_sp(&location, pc,
+				     read_memory_func, read_memory_func_extra);
+
+	switch (location.sp_location) {
+	case SP_LOC_UNKNOWN:
+		/* Give up. */
+		fp = -1;
+		break;
+
+	case SP_LOC_IN_R52:
+		fp = r52;
+		break;
+
+	case SP_LOC_OFFSET:
+		fp = sp + location.sp_offset;
+		break;
+
+	default:
+		/* Give up. */
+		fp = -1;
+		break;
+	}
+
+	/* If the frame pointer is not aligned to the basic word size
+	 * something terrible happened and we should mark it as invalid.
+	 */
+	if (fp % sizeof(bt_int_reg_t) != 0)
+		fp = -1;
+
+	/* -1 means "don't know initial_frame_caller_pc". */
+	initial_frame_caller_pc = -1;
+
+	switch (location.pc_location) {
+	case PC_LOC_UNKNOWN:
+		/* Give up. */
+		fp = -1;
+		break;
+
+	case PC_LOC_IN_LR:
+		if (lr == 0 || lr % TILE_BUNDLE_ALIGNMENT_IN_BYTES != 0) {
+			/* Give up. */
+			fp = -1;
+		} else {
+			initial_frame_caller_pc = lr;
+		}
+		break;
+
+	case PC_LOC_ON_STACK:
+		/* Leave initial_frame_caller_pc as -1,
+		 * meaning check the stack.
+		 */
+		break;
+
+	default:
+		/* Give up. */
+		fp = -1;
+		break;
+	}
+
+	state->pc = pc;
+	state->sp = sp;
+	state->fp = fp;
+	state->initial_frame_caller_pc = initial_frame_caller_pc;
+	state->read_memory_func = read_memory_func;
+	state->read_memory_func_extra = read_memory_func_extra;
+}
+
+/* Handle the case where the register holds more bits than the VA. */
+static bool valid_addr_reg(bt_int_reg_t reg)
+{
+	return ((unsigned long)reg == reg);
+}
+
+/* Advances the backtracing state to the calling frame, returning
+ * true iff successful.
+ */
+bool backtrace_next(BacktraceIterator *state)
+{
+	unsigned long next_fp, next_pc;
+	bt_int_reg_t next_frame[2];
+
+	if (state->fp == -1) {
+		/* No parent frame. */
+		return false;
+	}
+
+	/* Try to read the frame linkage data chaining to the next function. */
+	if (!state->read_memory_func(&next_frame, state->fp, sizeof next_frame,
+				     state->read_memory_func_extra)) {
+		return false;
+	}
+
+	next_fp = next_frame[1];
+	if (!valid_addr_reg(next_frame[1]) ||
+	    next_fp % sizeof(bt_int_reg_t) != 0) {
+		/* Caller's frame pointer is suspect, so give up. */
+		return false;
+	}
+
+	if (state->initial_frame_caller_pc != -1) {
+		/* We must be in the initial stack frame and already know the
+		 * caller PC.
+		 */
+		next_pc = state->initial_frame_caller_pc;
+
+		/* Force reading stack next time, in case we were in the
+		 * initial frame.  We don't do this above just to paranoidly
+		 * avoid changing the struct at all when we return false.
+		 */
+		state->initial_frame_caller_pc = -1;
+	} else {
+		/* Get the caller PC from the frame linkage area. */
+		next_pc = next_frame[0];
+		if (!valid_addr_reg(next_frame[0]) || next_pc == 0 ||
+		    next_pc % TILE_BUNDLE_ALIGNMENT_IN_BYTES != 0) {
+			/* The PC is suspect, so give up. */
+			return false;
+		}
+	}
+
+	/* Update state to become the caller's stack frame. */
+	state->pc = next_pc;
+	state->sp = state->fp;
+	state->fp = next_fp;
+
+	return true;
+}
diff --git a/arch/tile/kernel/compat.c b/arch/tile/kernel/compat.c
new file mode 100644
index 000000000..49120843f
--- /dev/null
+++ b/arch/tile/kernel/compat.c
@@ -0,0 +1,108 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/* Adjust unistd.h to provide 32-bit numbers and functions. */
+#define __SYSCALL_COMPAT
+
+#include <linux/compat.h>
+#include <linux/syscalls.h>
+#include <linux/kdev_t.h>
+#include <linux/fs.h>
+#include <linux/fcntl.h>
+#include <linux/uaccess.h>
+#include <linux/signal.h>
+#include <asm/syscalls.h>
+
+/*
+ * Syscalls that take 64-bit numbers traditionally take them in 32-bit
+ * "high" and "low" value parts on 32-bit architectures.
+ * In principle, one could imagine passing some register arguments as
+ * fully 64-bit on TILE-Gx in 32-bit mode, but it seems easier to
+ * adapt the usual convention.
+ */
+
+COMPAT_SYSCALL_DEFINE4(truncate64, char __user *, filename, u32, dummy,
+                       u32, low, u32, high)
+{
+	return sys_truncate(filename, ((loff_t)high << 32) | low);
+}
+
+COMPAT_SYSCALL_DEFINE4(ftruncate64, unsigned int, fd, u32, dummy,
+                       u32, low, u32, high)
+{
+	return sys_ftruncate(fd, ((loff_t)high << 32) | low);
+}
+
+COMPAT_SYSCALL_DEFINE6(pread64, unsigned int, fd, char __user *, ubuf,
+                       size_t, count, u32, dummy, u32, low, u32, high)
+{
+	return sys_pread64(fd, ubuf, count, ((loff_t)high << 32) | low);
+}
+
+COMPAT_SYSCALL_DEFINE6(pwrite64, unsigned int, fd, char __user *, ubuf,
+                       size_t, count, u32, dummy, u32, low, u32, high)
+{
+	return sys_pwrite64(fd, ubuf, count, ((loff_t)high << 32) | low);
+}
+
+COMPAT_SYSCALL_DEFINE6(sync_file_range2, int, fd, unsigned int, flags,
+                       u32, offset_lo, u32, offset_hi,
+                       u32, nbytes_lo, u32, nbytes_hi)
+{
+	return sys_sync_file_range(fd, ((loff_t)offset_hi << 32) | offset_lo,
+				   ((loff_t)nbytes_hi << 32) | nbytes_lo,
+				   flags);
+}
+
+COMPAT_SYSCALL_DEFINE6(fallocate, int, fd, int, mode,
+                       u32, offset_lo, u32, offset_hi,
+                       u32, len_lo, u32, len_hi)
+{
+	return sys_fallocate(fd, mode, ((loff_t)offset_hi << 32) | offset_lo,
+			     ((loff_t)len_hi << 32) | len_lo);
+}
+
+/*
+ * Avoid bug in generic sys_llseek() that specifies offset_high and
+ * offset_low as "unsigned long", thus making it possible to pass
+ * a sign-extended high 32 bits in offset_low.
+ */
+COMPAT_SYSCALL_DEFINE5(llseek, unsigned int, fd, unsigned int, offset_high,
+		       unsigned int, offset_low, loff_t __user *, result,
+		       unsigned int, origin)
+{
+	return sys_llseek(fd, offset_high, offset_low, result, origin);
+}
+
+/* Provide the compat syscall number to call mapping. */
+#undef __SYSCALL
+#define __SYSCALL(nr, call) [nr] = (call),
+
+/* See comments in sys.c */
+#define compat_sys_fadvise64_64 sys32_fadvise64_64
+#define compat_sys_readahead sys32_readahead
+#define sys_llseek compat_sys_llseek
+
+/* Call the assembly trampolines where necessary. */
+#define compat_sys_rt_sigreturn _compat_sys_rt_sigreturn
+#define sys_clone _sys_clone
+
+/*
+ * Note that we can't include <linux/unistd.h> here since the header
+ * guard will defeat us; <asm/unistd.h> checks for __SYSCALL as well.
+ */
+void *compat_sys_call_table[__NR_syscalls] = {
+	[0 ... __NR_syscalls-1] = sys_ni_syscall,
+#include <asm/unistd.h>
+};
diff --git a/arch/tile/kernel/compat_signal.c b/arch/tile/kernel/compat_signal.c
new file mode 100644
index 000000000..e8c2c0414
--- /dev/null
+++ b/arch/tile/kernel/compat_signal.c
@@ -0,0 +1,248 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/sched.h>
+#include <linux/mm.h>
+#include <linux/smp.h>
+#include <linux/kernel.h>
+#include <linux/signal.h>
+#include <linux/errno.h>
+#include <linux/wait.h>
+#include <linux/unistd.h>
+#include <linux/stddef.h>
+#include <linux/personality.h>
+#include <linux/suspend.h>
+#include <linux/ptrace.h>
+#include <linux/elf.h>
+#include <linux/compat.h>
+#include <linux/syscalls.h>
+#include <linux/uaccess.h>
+#include <asm/processor.h>
+#include <asm/ucontext.h>
+#include <asm/sigframe.h>
+#include <asm/syscalls.h>
+#include <asm/vdso.h>
+#include <arch/interrupts.h>
+
+struct compat_ucontext {
+	compat_ulong_t	  uc_flags;
+	compat_uptr_t     uc_link;
+	struct compat_sigaltstack	  uc_stack;
+	struct sigcontext uc_mcontext;
+	sigset_t	  uc_sigmask;	/* mask last for extensibility */
+};
+
+struct compat_rt_sigframe {
+	unsigned char save_area[C_ABI_SAVE_AREA_SIZE]; /* caller save area */
+	struct compat_siginfo info;
+	struct compat_ucontext uc;
+};
+
+int copy_siginfo_to_user32(struct compat_siginfo __user *to, const siginfo_t *from)
+{
+	int err;
+
+	if (!access_ok(VERIFY_WRITE, to, sizeof(struct compat_siginfo)))
+		return -EFAULT;
+
+	/* If you change siginfo_t structure, please make sure that
+	   this code is fixed accordingly.
+	   It should never copy any pad contained in the structure
+	   to avoid security leaks, but must copy the generic
+	   3 ints plus the relevant union member.  */
+	err = __put_user(from->si_signo, &to->si_signo);
+	err |= __put_user(from->si_errno, &to->si_errno);
+	err |= __put_user((short)from->si_code, &to->si_code);
+
+	if (from->si_code < 0) {
+		err |= __put_user(from->si_pid, &to->si_pid);
+		err |= __put_user(from->si_uid, &to->si_uid);
+		err |= __put_user(from->si_int, &to->si_int);
+	} else {
+		/*
+		 * First 32bits of unions are always present:
+		 * si_pid === si_band === si_tid === si_addr(LS half)
+		 */
+		err |= __put_user(from->_sifields._pad[0],
+				  &to->_sifields._pad[0]);
+		switch (from->si_code >> 16) {
+		case __SI_FAULT >> 16:
+			break;
+		case __SI_CHLD >> 16:
+			err |= __put_user(from->si_utime, &to->si_utime);
+			err |= __put_user(from->si_stime, &to->si_stime);
+			err |= __put_user(from->si_status, &to->si_status);
+			/* FALL THROUGH */
+		default:
+		case __SI_KILL >> 16:
+			err |= __put_user(from->si_uid, &to->si_uid);
+			break;
+		case __SI_POLL >> 16:
+			err |= __put_user(from->si_fd, &to->si_fd);
+			break;
+		case __SI_TIMER >> 16:
+			err |= __put_user(from->si_overrun, &to->si_overrun);
+			err |= __put_user(from->si_int, &to->si_int);
+			break;
+			 /* This is not generated by the kernel as of now.  */
+		case __SI_RT >> 16:
+		case __SI_MESGQ >> 16:
+			err |= __put_user(from->si_uid, &to->si_uid);
+			err |= __put_user(from->si_int, &to->si_int);
+			break;
+		}
+	}
+	return err;
+}
+
+int copy_siginfo_from_user32(siginfo_t *to, struct compat_siginfo __user *from)
+{
+	int err;
+
+	if (!access_ok(VERIFY_READ, from, sizeof(struct compat_siginfo)))
+		return -EFAULT;
+
+	memset(to, 0, sizeof(*to));
+
+	err = __get_user(to->si_signo, &from->si_signo);
+	err |= __get_user(to->si_errno, &from->si_errno);
+	err |= __get_user(to->si_code, &from->si_code);
+
+	err |= __get_user(to->si_pid, &from->si_pid);
+	err |= __get_user(to->si_uid, &from->si_uid);
+	err |= __get_user(to->si_int, &from->si_int);
+
+	return err;
+}
+
+/* The assembly shim for this function arranges to ignore the return value. */
+long compat_sys_rt_sigreturn(void)
+{
+	struct pt_regs *regs = current_pt_regs();
+	struct compat_rt_sigframe __user *frame =
+		(struct compat_rt_sigframe __user *) compat_ptr(regs->sp);
+	sigset_t set;
+
+	if (!access_ok(VERIFY_READ, frame, sizeof(*frame)))
+		goto badframe;
+	if (__copy_from_user(&set, &frame->uc.uc_sigmask, sizeof(set)))
+		goto badframe;
+
+	set_current_blocked(&set);
+
+	if (restore_sigcontext(regs, &frame->uc.uc_mcontext))
+		goto badframe;
+
+	if (compat_restore_altstack(&frame->uc.uc_stack))
+		goto badframe;
+
+	return 0;
+
+badframe:
+	signal_fault("bad sigreturn frame", regs, frame, 0);
+	return 0;
+}
+
+/*
+ * Determine which stack to use..
+ */
+static inline void __user *compat_get_sigframe(struct k_sigaction *ka,
+					       struct pt_regs *regs,
+					       size_t frame_size)
+{
+	unsigned long sp;
+
+	/* Default to using normal stack */
+	sp = (unsigned long)compat_ptr(regs->sp);
+
+	/*
+	 * If we are on the alternate signal stack and would overflow
+	 * it, don't.  Return an always-bogus address instead so we
+	 * will die with SIGSEGV.
+	 */
+	if (on_sig_stack(sp) && !likely(on_sig_stack(sp - frame_size)))
+		return (void __user __force *)-1UL;
+
+	/* This is the X/Open sanctioned signal stack switching.  */
+	if (ka->sa.sa_flags & SA_ONSTACK) {
+		if (sas_ss_flags(sp) == 0)
+			sp = current->sas_ss_sp + current->sas_ss_size;
+	}
+
+	sp -= frame_size;
+	/*
+	 * Align the stack pointer according to the TILE ABI,
+	 * i.e. so that on function entry (sp & 15) == 0.
+	 */
+	sp &= -16UL;
+	return (void __user *) sp;
+}
+
+int compat_setup_rt_frame(struct ksignal *ksig, sigset_t *set,
+			  struct pt_regs *regs)
+{
+	unsigned long restorer;
+	struct compat_rt_sigframe __user *frame;
+	int err = 0, sig = ksig->sig;
+
+	frame = compat_get_sigframe(&ksig->ka, regs, sizeof(*frame));
+
+	if (!access_ok(VERIFY_WRITE, frame, sizeof(*frame)))
+		goto err;
+
+	/* Always write at least the signal number for the stack backtracer. */
+	if (ksig->ka.sa.sa_flags & SA_SIGINFO) {
+		/* At sigreturn time, restore the callee-save registers too. */
+		err |= copy_siginfo_to_user32(&frame->info, &ksig->info);
+		regs->flags |= PT_FLAGS_RESTORE_REGS;
+	} else {
+		err |= __put_user(ksig->info.si_signo, &frame->info.si_signo);
+	}
+
+	/* Create the ucontext.  */
+	err |= __clear_user(&frame->save_area, sizeof(frame->save_area));
+	err |= __put_user(0, &frame->uc.uc_flags);
+	err |= __put_user(0, &frame->uc.uc_link);
+	err |= __compat_save_altstack(&frame->uc.uc_stack, regs->sp);
+	err |= setup_sigcontext(&frame->uc.uc_mcontext, regs);
+	err |= __copy_to_user(&frame->uc.uc_sigmask, set, sizeof(*set));
+	if (err)
+		goto err;
+
+	restorer = VDSO_SYM(&__vdso_rt_sigreturn);
+	if (ksig->ka.sa.sa_flags & SA_RESTORER)
+		restorer = ptr_to_compat_reg(ksig->ka.sa.sa_restorer);
+
+	/*
+	 * Set up registers for signal handler.
+	 * Registers that we don't modify keep the value they had from
+	 * user-space at the time we took the signal.
+	 * We always pass siginfo and mcontext, regardless of SA_SIGINFO,
+	 * since some things rely on this (e.g. glibc's debug/segfault.c).
+	 */
+	regs->pc = ptr_to_compat_reg(ksig->ka.sa.sa_handler);
+	regs->ex1 = PL_ICS_EX1(USER_PL, 1); /* set crit sec in handler */
+	regs->sp = ptr_to_compat_reg(frame);
+	regs->lr = restorer;
+	regs->regs[0] = (unsigned long) sig;
+	regs->regs[1] = ptr_to_compat_reg(&frame->info);
+	regs->regs[2] = ptr_to_compat_reg(&frame->uc);
+	regs->flags |= PT_FLAGS_CALLER_SAVES;
+	return 0;
+
+err:
+	trace_unhandled_signal("bad sigreturn frame", regs,
+			      (unsigned long)frame, SIGSEGV);
+	return -EFAULT;
+}
diff --git a/arch/tile/kernel/early_printk.c b/arch/tile/kernel/early_printk.c
new file mode 100644
index 000000000..aefb2c086
--- /dev/null
+++ b/arch/tile/kernel/early_printk.c
@@ -0,0 +1,75 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/console.h>
+#include <linux/kernel.h>
+#include <linux/init.h>
+#include <linux/string.h>
+#include <linux/irqflags.h>
+#include <linux/printk.h>
+#include <asm/setup.h>
+#include <hv/hypervisor.h>
+
+static void early_hv_write(struct console *con, const char *s, unsigned n)
+{
+	tile_console_write(s, n);
+
+	/*
+	 * Convert NL to NLCR (close enough to CRNL) during early boot.
+	 * We assume newlines are at the ends of strings, which turns out
+	 * to be good enough for early boot console output.
+	 */
+	if (n && s[n-1] == '\n')
+		tile_console_write("\r", 1);
+}
+
+static struct console early_hv_console = {
+	.name =		"earlyhv",
+	.write =	early_hv_write,
+	.flags =	CON_PRINTBUFFER | CON_BOOT,
+	.index =	-1,
+};
+
+void early_panic(const char *fmt, ...)
+{
+	struct va_format vaf;
+	va_list args;
+
+	arch_local_irq_disable_all();
+
+	va_start(args, fmt);
+
+	vaf.fmt = fmt;
+	vaf.va = &args;
+
+	early_printk("Kernel panic - not syncing: %pV", &vaf);
+
+	va_end(args);
+
+	dump_stack();
+	hv_halt();
+}
+
+static int __init setup_early_printk(char *str)
+{
+	if (early_console)
+		return 1;
+
+	early_console = &early_hv_console;
+	register_console(early_console);
+
+	return 0;
+}
+
+early_param("earlyprintk", setup_early_printk);
diff --git a/arch/tile/kernel/entry.S b/arch/tile/kernel/entry.S
new file mode 100644
index 000000000..3d9175992
--- /dev/null
+++ b/arch/tile/kernel/entry.S
@@ -0,0 +1,71 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/linkage.h>
+#include <linux/unistd.h>
+#include <asm/irqflags.h>
+#include <asm/processor.h>
+#include <arch/abi.h>
+#include <arch/spr_def.h>
+
+#ifdef __tilegx__
+#define bnzt bnezt
+#endif
+
+STD_ENTRY(current_text_addr)
+	{ move r0, lr; jrp lr }
+	STD_ENDPROC(current_text_addr)
+
+STD_ENTRY(dump_stack)
+	{ move r2, lr; lnk r1 }
+	{ move r4, r52; addli r1, r1, dump_stack - . }
+	{ move r3, sp; j _dump_stack }
+	jrp lr   /* keep backtracer happy */
+	STD_ENDPROC(dump_stack)
+
+STD_ENTRY(KBacktraceIterator_init_current)
+	{ move r2, lr; lnk r1 }
+	{ move r4, r52; addli r1, r1, KBacktraceIterator_init_current - . }
+	{ move r3, sp; j _KBacktraceIterator_init_current }
+	jrp lr   /* keep backtracer happy */
+	STD_ENDPROC(KBacktraceIterator_init_current)
+
+/* Loop forever on a nap during SMP boot. */
+STD_ENTRY(smp_nap)
+	nap
+	nop       /* avoid provoking the icache prefetch with a jump */
+	j smp_nap /* we are not architecturally guaranteed not to exit nap */
+	jrp lr    /* clue in the backtracer */
+	STD_ENDPROC(smp_nap)
+
+/*
+ * Enable interrupts racelessly and then nap until interrupted.
+ * Architecturally, we are guaranteed that enabling interrupts via
+ * mtspr to INTERRUPT_CRITICAL_SECTION only interrupts at the next PC.
+ * This function's _cpu_idle_nap address is special; see intvec.S.
+ * When interrupted at _cpu_idle_nap, we bump the PC forward 8, and
+ * as a result return to the function that called _cpu_idle().
+ */
+STD_ENTRY(_cpu_idle)
+	movei r1, 1
+	IRQ_ENABLE_LOAD(r2, r3)
+	mtspr INTERRUPT_CRITICAL_SECTION, r1
+	IRQ_ENABLE_APPLY(r2, r3)       /* unmask, but still with ICS set */
+	mtspr INTERRUPT_CRITICAL_SECTION, zero
+	.global _cpu_idle_nap
+_cpu_idle_nap:
+	nap
+	nop       /* avoid provoking the icache prefetch with a jump */
+	jrp lr
+	STD_ENDPROC(_cpu_idle)
diff --git a/arch/tile/kernel/ftrace.c b/arch/tile/kernel/ftrace.c
new file mode 100644
index 000000000..0c0996175
--- /dev/null
+++ b/arch/tile/kernel/ftrace.c
@@ -0,0 +1,248 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * TILE-Gx specific ftrace support
+ */
+
+#include <linux/ftrace.h>
+#include <linux/uaccess.h>
+
+#include <asm/cacheflush.h>
+#include <asm/ftrace.h>
+#include <asm/sections.h>
+
+#include <arch/opcode.h>
+
+#ifdef CONFIG_DYNAMIC_FTRACE
+
+static inline tilegx_bundle_bits NOP(void)
+{
+	return create_UnaryOpcodeExtension_X0(FNOP_UNARY_OPCODE_X0) |
+		create_RRROpcodeExtension_X0(UNARY_RRR_0_OPCODE_X0) |
+		create_Opcode_X0(RRR_0_OPCODE_X0) |
+		create_UnaryOpcodeExtension_X1(NOP_UNARY_OPCODE_X1) |
+		create_RRROpcodeExtension_X1(UNARY_RRR_0_OPCODE_X1) |
+		create_Opcode_X1(RRR_0_OPCODE_X1);
+}
+
+static int machine_stopped __read_mostly;
+
+int ftrace_arch_code_modify_prepare(void)
+{
+	machine_stopped = 1;
+	return 0;
+}
+
+int ftrace_arch_code_modify_post_process(void)
+{
+	flush_icache_range(0, CHIP_L1I_CACHE_SIZE());
+	machine_stopped = 0;
+	return 0;
+}
+
+/*
+ * Put { move r10, lr; jal ftrace_caller } in a bundle, this lets dynamic
+ * tracer just add one cycle overhead to every kernel function when disabled.
+ */
+static unsigned long ftrace_gen_branch(unsigned long pc, unsigned long addr,
+				       bool link)
+{
+	tilegx_bundle_bits opcode_x0, opcode_x1;
+	long pcrel_by_instr = (addr - pc) >> TILEGX_LOG2_BUNDLE_SIZE_IN_BYTES;
+
+	if (link) {
+		/* opcode: jal addr */
+		opcode_x1 =
+			create_Opcode_X1(JUMP_OPCODE_X1) |
+			create_JumpOpcodeExtension_X1(JAL_JUMP_OPCODE_X1) |
+			create_JumpOff_X1(pcrel_by_instr);
+	} else {
+		/* opcode: j addr */
+		opcode_x1 =
+			create_Opcode_X1(JUMP_OPCODE_X1) |
+			create_JumpOpcodeExtension_X1(J_JUMP_OPCODE_X1) |
+			create_JumpOff_X1(pcrel_by_instr);
+	}
+
+	/*
+	 * Also put { move r10, lr; jal ftrace_stub } in a bundle, which
+	 * is used to replace the instruction in address ftrace_call.
+	 */
+	if (addr == FTRACE_ADDR || addr == (unsigned long)ftrace_stub) {
+		/* opcode: or r10, lr, zero */
+		opcode_x0 =
+			create_Dest_X0(10) |
+			create_SrcA_X0(TREG_LR) |
+			create_SrcB_X0(TREG_ZERO) |
+			create_RRROpcodeExtension_X0(OR_RRR_0_OPCODE_X0) |
+			create_Opcode_X0(RRR_0_OPCODE_X0);
+	} else {
+		/* opcode: fnop */
+		opcode_x0 =
+			create_UnaryOpcodeExtension_X0(FNOP_UNARY_OPCODE_X0) |
+			create_RRROpcodeExtension_X0(UNARY_RRR_0_OPCODE_X0) |
+			create_Opcode_X0(RRR_0_OPCODE_X0);
+	}
+
+	return opcode_x1 | opcode_x0;
+}
+
+static unsigned long ftrace_nop_replace(struct dyn_ftrace *rec)
+{
+	return NOP();
+}
+
+static unsigned long ftrace_call_replace(unsigned long pc, unsigned long addr)
+{
+	return ftrace_gen_branch(pc, addr, true);
+}
+
+static int ftrace_modify_code(unsigned long pc, unsigned long old,
+			      unsigned long new)
+{
+	unsigned long pc_wr;
+
+	/* Check if the address is in kernel text space and module space. */
+	if (!kernel_text_address(pc))
+		return -EINVAL;
+
+	/* Operate on writable kernel text mapping. */
+	pc_wr = pc - MEM_SV_START + PAGE_OFFSET;
+
+	if (probe_kernel_write((void *)pc_wr, &new, MCOUNT_INSN_SIZE))
+		return -EPERM;
+
+	smp_wmb();
+
+	if (!machine_stopped && num_online_cpus() > 1)
+		flush_icache_range(pc, pc + MCOUNT_INSN_SIZE);
+
+	return 0;
+}
+
+int ftrace_update_ftrace_func(ftrace_func_t func)
+{
+	unsigned long pc, old;
+	unsigned long new;
+	int ret;
+
+	pc = (unsigned long)&ftrace_call;
+	memcpy(&old, &ftrace_call, MCOUNT_INSN_SIZE);
+	new = ftrace_call_replace(pc, (unsigned long)func);
+
+	ret = ftrace_modify_code(pc, old, new);
+
+	return ret;
+}
+
+int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
+{
+	unsigned long new, old;
+	unsigned long ip = rec->ip;
+
+	old = ftrace_nop_replace(rec);
+	new = ftrace_call_replace(ip, addr);
+
+	return ftrace_modify_code(rec->ip, old, new);
+}
+
+int ftrace_make_nop(struct module *mod,
+		    struct dyn_ftrace *rec, unsigned long addr)
+{
+	unsigned long ip = rec->ip;
+	unsigned long old;
+	unsigned long new;
+	int ret;
+
+	old = ftrace_call_replace(ip, addr);
+	new = ftrace_nop_replace(rec);
+	ret = ftrace_modify_code(ip, old, new);
+
+	return ret;
+}
+
+int __init ftrace_dyn_arch_init(void)
+{
+	return 0;
+}
+#endif /* CONFIG_DYNAMIC_FTRACE */
+
+#ifdef CONFIG_FUNCTION_GRAPH_TRACER
+void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr,
+			   unsigned long frame_pointer)
+{
+	unsigned long return_hooker = (unsigned long) &return_to_handler;
+	struct ftrace_graph_ent trace;
+	unsigned long old;
+	int err;
+
+	if (unlikely(atomic_read(&current->tracing_graph_pause)))
+		return;
+
+	old = *parent;
+	*parent = return_hooker;
+
+	err = ftrace_push_return_trace(old, self_addr, &trace.depth,
+				       frame_pointer);
+	if (err == -EBUSY) {
+		*parent = old;
+		return;
+	}
+
+	trace.func = self_addr;
+
+	/* Only trace if the calling function expects to */
+	if (!ftrace_graph_entry(&trace)) {
+		current->curr_ret_stack--;
+		*parent = old;
+	}
+}
+
+#ifdef CONFIG_DYNAMIC_FTRACE
+extern unsigned long ftrace_graph_call;
+
+static int __ftrace_modify_caller(unsigned long *callsite,
+				  void (*func) (void), bool enable)
+{
+	unsigned long caller_fn = (unsigned long) func;
+	unsigned long pc = (unsigned long) callsite;
+	unsigned long branch = ftrace_gen_branch(pc, caller_fn, false);
+	unsigned long nop = NOP();
+	unsigned long old = enable ? nop : branch;
+	unsigned long new = enable ? branch : nop;
+
+	return ftrace_modify_code(pc, old, new);
+}
+
+static int ftrace_modify_graph_caller(bool enable)
+{
+	int ret;
+
+	ret = __ftrace_modify_caller(&ftrace_graph_call,
+				     ftrace_graph_caller,
+				     enable);
+
+	return ret;
+}
+
+int ftrace_enable_ftrace_graph_caller(void)
+{
+	return ftrace_modify_graph_caller(true);
+}
+
+int ftrace_disable_ftrace_graph_caller(void)
+{
+	return ftrace_modify_graph_caller(false);
+}
+#endif /* CONFIG_DYNAMIC_FTRACE */
+#endif /* CONFIG_FUNCTION_GRAPH_TRACER */
diff --git a/arch/tile/kernel/hardwall.c b/arch/tile/kernel/hardwall.c
new file mode 100644
index 000000000..2fd1694ac
--- /dev/null
+++ b/arch/tile/kernel/hardwall.c
@@ -0,0 +1,1096 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/fs.h>
+#include <linux/proc_fs.h>
+#include <linux/seq_file.h>
+#include <linux/rwsem.h>
+#include <linux/kprobes.h>
+#include <linux/sched.h>
+#include <linux/hardirq.h>
+#include <linux/uaccess.h>
+#include <linux/smp.h>
+#include <linux/cdev.h>
+#include <linux/compat.h>
+#include <asm/hardwall.h>
+#include <asm/traps.h>
+#include <asm/siginfo.h>
+#include <asm/irq_regs.h>
+
+#include <arch/interrupts.h>
+#include <arch/spr_def.h>
+
+
+/*
+ * Implement a per-cpu "hardwall" resource class such as UDN or IPI.
+ * We use "hardwall" nomenclature throughout for historical reasons.
+ * The lock here controls access to the list data structure as well as
+ * to the items on the list.
+ */
+struct hardwall_type {
+	int index;
+	int is_xdn;
+	int is_idn;
+	int disabled;
+	const char *name;
+	struct list_head list;
+	spinlock_t lock;
+	struct proc_dir_entry *proc_dir;
+};
+
+enum hardwall_index {
+	HARDWALL_UDN = 0,
+#ifndef __tilepro__
+	HARDWALL_IDN = 1,
+	HARDWALL_IPI = 2,
+#endif
+	_HARDWALL_TYPES
+};
+
+static struct hardwall_type hardwall_types[] = {
+	{  /* user-space access to UDN */
+		0,
+		1,
+		0,
+		0,
+		"udn",
+		LIST_HEAD_INIT(hardwall_types[HARDWALL_UDN].list),
+		__SPIN_LOCK_UNLOCKED(hardwall_types[HARDWALL_UDN].lock),
+		NULL
+	},
+#ifndef __tilepro__
+	{  /* user-space access to IDN */
+		1,
+		1,
+		1,
+		1,  /* disabled pending hypervisor support */
+		"idn",
+		LIST_HEAD_INIT(hardwall_types[HARDWALL_IDN].list),
+		__SPIN_LOCK_UNLOCKED(hardwall_types[HARDWALL_IDN].lock),
+		NULL
+	},
+	{  /* access to user-space IPI */
+		2,
+		0,
+		0,
+		0,
+		"ipi",
+		LIST_HEAD_INIT(hardwall_types[HARDWALL_IPI].list),
+		__SPIN_LOCK_UNLOCKED(hardwall_types[HARDWALL_IPI].lock),
+		NULL
+	},
+#endif
+};
+
+/*
+ * This data structure tracks the cpu data, etc., associated
+ * one-to-one with a "struct file *" from opening a hardwall device file.
+ * Note that the file's private data points back to this structure.
+ */
+struct hardwall_info {
+	struct list_head list;             /* for hardwall_types.list */
+	struct list_head task_head;        /* head of tasks in this hardwall */
+	struct hardwall_type *type;        /* type of this resource */
+	struct cpumask cpumask;            /* cpus reserved */
+	int id;                            /* integer id for this hardwall */
+	int teardown_in_progress;          /* are we tearing this one down? */
+
+	/* Remaining fields only valid for user-network resources. */
+	int ulhc_x;                        /* upper left hand corner x coord */
+	int ulhc_y;                        /* upper left hand corner y coord */
+	int width;                         /* rectangle width */
+	int height;                        /* rectangle height */
+#if CHIP_HAS_REV1_XDN()
+	atomic_t xdn_pending_count;        /* cores in phase 1 of drain */
+#endif
+};
+
+
+/* /proc/tile/hardwall */
+static struct proc_dir_entry *hardwall_proc_dir;
+
+/* Functions to manage files in /proc/tile/hardwall. */
+static void hardwall_add_proc(struct hardwall_info *);
+static void hardwall_remove_proc(struct hardwall_info *);
+
+/* Allow disabling UDN access. */
+static int __init noudn(char *str)
+{
+	pr_info("User-space UDN access is disabled\n");
+	hardwall_types[HARDWALL_UDN].disabled = 1;
+	return 0;
+}
+early_param("noudn", noudn);
+
+#ifndef __tilepro__
+/* Allow disabling IDN access. */
+static int __init noidn(char *str)
+{
+	pr_info("User-space IDN access is disabled\n");
+	hardwall_types[HARDWALL_IDN].disabled = 1;
+	return 0;
+}
+early_param("noidn", noidn);
+
+/* Allow disabling IPI access. */
+static int __init noipi(char *str)
+{
+	pr_info("User-space IPI access is disabled\n");
+	hardwall_types[HARDWALL_IPI].disabled = 1;
+	return 0;
+}
+early_param("noipi", noipi);
+#endif
+
+
+/*
+ * Low-level primitives for UDN/IDN
+ */
+
+#ifdef __tilepro__
+#define mtspr_XDN(hwt, name, val) \
+	do { (void)(hwt); __insn_mtspr(SPR_UDN_##name, (val)); } while (0)
+#define mtspr_MPL_XDN(hwt, name, val) \
+	do { (void)(hwt); __insn_mtspr(SPR_MPL_UDN_##name, (val)); } while (0)
+#define mfspr_XDN(hwt, name) \
+	((void)(hwt), __insn_mfspr(SPR_UDN_##name))
+#else
+#define mtspr_XDN(hwt, name, val)					\
+	do {								\
+		if ((hwt)->is_idn)					\
+			__insn_mtspr(SPR_IDN_##name, (val));		\
+		else							\
+			__insn_mtspr(SPR_UDN_##name, (val));		\
+	} while (0)
+#define mtspr_MPL_XDN(hwt, name, val)					\
+	do {								\
+		if ((hwt)->is_idn)					\
+			__insn_mtspr(SPR_MPL_IDN_##name, (val));	\
+		else							\
+			__insn_mtspr(SPR_MPL_UDN_##name, (val));	\
+	} while (0)
+#define mfspr_XDN(hwt, name) \
+  ((hwt)->is_idn ? __insn_mfspr(SPR_IDN_##name) : __insn_mfspr(SPR_UDN_##name))
+#endif
+
+/* Set a CPU bit if the CPU is online. */
+#define cpu_online_set(cpu, dst) do { \
+	if (cpu_online(cpu))          \
+		cpumask_set_cpu(cpu, dst);    \
+} while (0)
+
+
+/* Does the given rectangle contain the given x,y coordinate? */
+static int contains(struct hardwall_info *r, int x, int y)
+{
+	return (x >= r->ulhc_x && x < r->ulhc_x + r->width) &&
+		(y >= r->ulhc_y && y < r->ulhc_y + r->height);
+}
+
+/* Compute the rectangle parameters and validate the cpumask. */
+static int check_rectangle(struct hardwall_info *r, struct cpumask *mask)
+{
+	int x, y, cpu, ulhc, lrhc;
+
+	/* The first cpu is the ULHC, the last the LRHC. */
+	ulhc = find_first_bit(cpumask_bits(mask), nr_cpumask_bits);
+	lrhc = find_last_bit(cpumask_bits(mask), nr_cpumask_bits);
+
+	/* Compute the rectangle attributes from the cpus. */
+	r->ulhc_x = cpu_x(ulhc);
+	r->ulhc_y = cpu_y(ulhc);
+	r->width = cpu_x(lrhc) - r->ulhc_x + 1;
+	r->height = cpu_y(lrhc) - r->ulhc_y + 1;
+
+	/* Width and height must be positive */
+	if (r->width <= 0 || r->height <= 0)
+		return -EINVAL;
+
+	/* Confirm that the cpumask is exactly the rectangle. */
+	for (y = 0, cpu = 0; y < smp_height; ++y)
+		for (x = 0; x < smp_width; ++x, ++cpu)
+			if (cpumask_test_cpu(cpu, mask) != contains(r, x, y))
+				return -EINVAL;
+
+	/*
+	 * Note that offline cpus can't be drained when this user network
+	 * rectangle eventually closes.  We used to detect this
+	 * situation and print a warning, but it annoyed users and
+	 * they ignored it anyway, so now we just return without a
+	 * warning.
+	 */
+	return 0;
+}
+
+/*
+ * Hardware management of hardwall setup, teardown, trapping,
+ * and enabling/disabling PL0 access to the networks.
+ */
+
+/* Bit field values to mask together for writes to SPR_XDN_DIRECTION_PROTECT */
+enum direction_protect {
+	N_PROTECT = (1 << 0),
+	E_PROTECT = (1 << 1),
+	S_PROTECT = (1 << 2),
+	W_PROTECT = (1 << 3),
+	C_PROTECT = (1 << 4),
+};
+
+static inline int xdn_which_interrupt(struct hardwall_type *hwt)
+{
+#ifndef __tilepro__
+	if (hwt->is_idn)
+		return INT_IDN_FIREWALL;
+#endif
+	return INT_UDN_FIREWALL;
+}
+
+static void enable_firewall_interrupts(struct hardwall_type *hwt)
+{
+	arch_local_irq_unmask_now(xdn_which_interrupt(hwt));
+}
+
+static void disable_firewall_interrupts(struct hardwall_type *hwt)
+{
+	arch_local_irq_mask_now(xdn_which_interrupt(hwt));
+}
+
+/* Set up hardwall on this cpu based on the passed hardwall_info. */
+static void hardwall_setup_func(void *info)
+{
+	struct hardwall_info *r = info;
+	struct hardwall_type *hwt = r->type;
+
+	int cpu = smp_processor_id();  /* on_each_cpu disables preemption */
+	int x = cpu_x(cpu);
+	int y = cpu_y(cpu);
+	int bits = 0;
+	if (x == r->ulhc_x)
+		bits |= W_PROTECT;
+	if (x == r->ulhc_x + r->width - 1)
+		bits |= E_PROTECT;
+	if (y == r->ulhc_y)
+		bits |= N_PROTECT;
+	if (y == r->ulhc_y + r->height - 1)
+		bits |= S_PROTECT;
+	BUG_ON(bits == 0);
+	mtspr_XDN(hwt, DIRECTION_PROTECT, bits);
+	enable_firewall_interrupts(hwt);
+}
+
+/* Set up all cpus on edge of rectangle to enable/disable hardwall SPRs. */
+static void hardwall_protect_rectangle(struct hardwall_info *r)
+{
+	int x, y, cpu, delta;
+	struct cpumask rect_cpus;
+
+	cpumask_clear(&rect_cpus);
+
+	/* First include the top and bottom edges */
+	cpu = r->ulhc_y * smp_width + r->ulhc_x;
+	delta = (r->height - 1) * smp_width;
+	for (x = 0; x < r->width; ++x, ++cpu) {
+		cpu_online_set(cpu, &rect_cpus);
+		cpu_online_set(cpu + delta, &rect_cpus);
+	}
+
+	/* Then the left and right edges */
+	cpu -= r->width;
+	delta = r->width - 1;
+	for (y = 0; y < r->height; ++y, cpu += smp_width) {
+		cpu_online_set(cpu, &rect_cpus);
+		cpu_online_set(cpu + delta, &rect_cpus);
+	}
+
+	/* Then tell all the cpus to set up their protection SPR */
+	on_each_cpu_mask(&rect_cpus, hardwall_setup_func, r, 1);
+}
+
+/* Entered from INT_xDN_FIREWALL interrupt vector with irqs disabled. */
+void __kprobes do_hardwall_trap(struct pt_regs* regs, int fault_num)
+{
+	struct hardwall_info *rect;
+	struct hardwall_type *hwt;
+	struct task_struct *p;
+	struct siginfo info;
+	int cpu = smp_processor_id();
+	int found_processes;
+	struct pt_regs *old_regs = set_irq_regs(regs);
+
+	irq_enter();
+
+	/* Figure out which network trapped. */
+	switch (fault_num) {
+#ifndef __tilepro__
+	case INT_IDN_FIREWALL:
+		hwt = &hardwall_types[HARDWALL_IDN];
+		break;
+#endif
+	case INT_UDN_FIREWALL:
+		hwt = &hardwall_types[HARDWALL_UDN];
+		break;
+	default:
+		BUG();
+	}
+	BUG_ON(hwt->disabled);
+
+	/* This tile trapped a network access; find the rectangle. */
+	spin_lock(&hwt->lock);
+	list_for_each_entry(rect, &hwt->list, list) {
+		if (cpumask_test_cpu(cpu, &rect->cpumask))
+			break;
+	}
+
+	/*
+	 * It shouldn't be possible not to find this cpu on the
+	 * rectangle list, since only cpus in rectangles get hardwalled.
+	 * The hardwall is only removed after the user network is drained.
+	 */
+	BUG_ON(&rect->list == &hwt->list);
+
+	/*
+	 * If we already started teardown on this hardwall, don't worry;
+	 * the abort signal has been sent and we are just waiting for things
+	 * to quiesce.
+	 */
+	if (rect->teardown_in_progress) {
+		pr_notice("cpu %d: detected %s hardwall violation %#lx while teardown already in progress\n",
+			  cpu, hwt->name,
+			  (long)mfspr_XDN(hwt, DIRECTION_PROTECT));
+		goto done;
+	}
+
+	/*
+	 * Kill off any process that is activated in this rectangle.
+	 * We bypass security to deliver the signal, since it must be
+	 * one of the activated processes that generated the user network
+	 * message that caused this trap, and all the activated
+	 * processes shared a single open file so are pretty tightly
+	 * bound together from a security point of view to begin with.
+	 */
+	rect->teardown_in_progress = 1;
+	wmb(); /* Ensure visibility of rectangle before notifying processes. */
+	pr_notice("cpu %d: detected %s hardwall violation %#lx...\n",
+		  cpu, hwt->name, (long)mfspr_XDN(hwt, DIRECTION_PROTECT));
+	info.si_signo = SIGILL;
+	info.si_errno = 0;
+	info.si_code = ILL_HARDWALL;
+	found_processes = 0;
+	list_for_each_entry(p, &rect->task_head,
+			    thread.hardwall[hwt->index].list) {
+		BUG_ON(p->thread.hardwall[hwt->index].info != rect);
+		if (!(p->flags & PF_EXITING)) {
+			found_processes = 1;
+			pr_notice("hardwall: killing %d\n", p->pid);
+			do_send_sig_info(info.si_signo, &info, p, false);
+		}
+	}
+	if (!found_processes)
+		pr_notice("hardwall: no associated processes!\n");
+
+ done:
+	spin_unlock(&hwt->lock);
+
+	/*
+	 * We have to disable firewall interrupts now, or else when we
+	 * return from this handler, we will simply re-interrupt back to
+	 * it.  However, we can't clear the protection bits, since we
+	 * haven't yet drained the network, and that would allow packets
+	 * to cross out of the hardwall region.
+	 */
+	disable_firewall_interrupts(hwt);
+
+	irq_exit();
+	set_irq_regs(old_regs);
+}
+
+/* Allow access from user space to the user network. */
+void grant_hardwall_mpls(struct hardwall_type *hwt)
+{
+#ifndef __tilepro__
+	if (!hwt->is_xdn) {
+		__insn_mtspr(SPR_MPL_IPI_0_SET_0, 1);
+		return;
+	}
+#endif
+	mtspr_MPL_XDN(hwt, ACCESS_SET_0, 1);
+	mtspr_MPL_XDN(hwt, AVAIL_SET_0, 1);
+	mtspr_MPL_XDN(hwt, COMPLETE_SET_0, 1);
+	mtspr_MPL_XDN(hwt, TIMER_SET_0, 1);
+#if !CHIP_HAS_REV1_XDN()
+	mtspr_MPL_XDN(hwt, REFILL_SET_0, 1);
+	mtspr_MPL_XDN(hwt, CA_SET_0, 1);
+#endif
+}
+
+/* Deny access from user space to the user network. */
+void restrict_hardwall_mpls(struct hardwall_type *hwt)
+{
+#ifndef __tilepro__
+	if (!hwt->is_xdn) {
+		__insn_mtspr(SPR_MPL_IPI_0_SET_1, 1);
+		return;
+	}
+#endif
+	mtspr_MPL_XDN(hwt, ACCESS_SET_1, 1);
+	mtspr_MPL_XDN(hwt, AVAIL_SET_1, 1);
+	mtspr_MPL_XDN(hwt, COMPLETE_SET_1, 1);
+	mtspr_MPL_XDN(hwt, TIMER_SET_1, 1);
+#if !CHIP_HAS_REV1_XDN()
+	mtspr_MPL_XDN(hwt, REFILL_SET_1, 1);
+	mtspr_MPL_XDN(hwt, CA_SET_1, 1);
+#endif
+}
+
+/* Restrict or deny as necessary for the task we're switching to. */
+void hardwall_switch_tasks(struct task_struct *prev,
+			   struct task_struct *next)
+{
+	int i;
+	for (i = 0; i < HARDWALL_TYPES; ++i) {
+		if (prev->thread.hardwall[i].info != NULL) {
+			if (next->thread.hardwall[i].info == NULL)
+				restrict_hardwall_mpls(&hardwall_types[i]);
+		} else if (next->thread.hardwall[i].info != NULL) {
+			grant_hardwall_mpls(&hardwall_types[i]);
+		}
+	}
+}
+
+/* Does this task have the right to IPI the given cpu? */
+int hardwall_ipi_valid(int cpu)
+{
+#ifdef __tilegx__
+	struct hardwall_info *info =
+		current->thread.hardwall[HARDWALL_IPI].info;
+	return info && cpumask_test_cpu(cpu, &info->cpumask);
+#else
+	return 0;
+#endif
+}
+
+/*
+ * Code to create, activate, deactivate, and destroy hardwall resources.
+ */
+
+/* Create a hardwall for the given resource */
+static struct hardwall_info *hardwall_create(struct hardwall_type *hwt,
+					     size_t size,
+					     const unsigned char __user *bits)
+{
+	struct hardwall_info *iter, *info;
+	struct cpumask mask;
+	unsigned long flags;
+	int rc;
+
+	/* Reject crazy sizes out of hand, a la sys_mbind(). */
+	if (size > PAGE_SIZE)
+		return ERR_PTR(-EINVAL);
+
+	/* Copy whatever fits into a cpumask. */
+	if (copy_from_user(&mask, bits, min(sizeof(struct cpumask), size)))
+		return ERR_PTR(-EFAULT);
+
+	/*
+	 * If the size was short, clear the rest of the mask;
+	 * otherwise validate that the rest of the user mask was zero
+	 * (we don't try hard to be efficient when validating huge masks).
+	 */
+	if (size < sizeof(struct cpumask)) {
+		memset((char *)&mask + size, 0, sizeof(struct cpumask) - size);
+	} else if (size > sizeof(struct cpumask)) {
+		size_t i;
+		for (i = sizeof(struct cpumask); i < size; ++i) {
+			char c;
+			if (get_user(c, &bits[i]))
+				return ERR_PTR(-EFAULT);
+			if (c)
+				return ERR_PTR(-EINVAL);
+		}
+	}
+
+	/* Allocate a new hardwall_info optimistically. */
+	info = kmalloc(sizeof(struct hardwall_info),
+			GFP_KERNEL | __GFP_ZERO);
+	if (info == NULL)
+		return ERR_PTR(-ENOMEM);
+	INIT_LIST_HEAD(&info->task_head);
+	info->type = hwt;
+
+	/* Compute the rectangle size and validate that it's plausible. */
+	cpumask_copy(&info->cpumask, &mask);
+	info->id = find_first_bit(cpumask_bits(&mask), nr_cpumask_bits);
+	if (hwt->is_xdn) {
+		rc = check_rectangle(info, &mask);
+		if (rc != 0) {
+			kfree(info);
+			return ERR_PTR(rc);
+		}
+	}
+
+	/*
+	 * Eliminate cpus that are not part of this Linux client.
+	 * Note that this allows for configurations that we might not want to
+	 * support, such as one client on every even cpu, another client on
+	 * every odd cpu.
+	 */
+	cpumask_and(&info->cpumask, &info->cpumask, cpu_online_mask);
+
+	/* Confirm it doesn't overlap and add it to the list. */
+	spin_lock_irqsave(&hwt->lock, flags);
+	list_for_each_entry(iter, &hwt->list, list) {
+		if (cpumask_intersects(&iter->cpumask, &info->cpumask)) {
+			spin_unlock_irqrestore(&hwt->lock, flags);
+			kfree(info);
+			return ERR_PTR(-EBUSY);
+		}
+	}
+	list_add_tail(&info->list, &hwt->list);
+	spin_unlock_irqrestore(&hwt->lock, flags);
+
+	/* Set up appropriate hardwalling on all affected cpus. */
+	if (hwt->is_xdn)
+		hardwall_protect_rectangle(info);
+
+	/* Create a /proc/tile/hardwall entry. */
+	hardwall_add_proc(info);
+
+	return info;
+}
+
+/* Activate a given hardwall on this cpu for this process. */
+static int hardwall_activate(struct hardwall_info *info)
+{
+	int cpu;
+	unsigned long flags;
+	struct task_struct *p = current;
+	struct thread_struct *ts = &p->thread;
+	struct hardwall_type *hwt;
+
+	/* Require a hardwall. */
+	if (info == NULL)
+		return -ENODATA;
+
+	/* Not allowed to activate a hardwall that is being torn down. */
+	if (info->teardown_in_progress)
+		return -EINVAL;
+
+	/*
+	 * Get our affinity; if we're not bound to this tile uniquely,
+	 * we can't access the network registers.
+	 */
+	if (cpumask_weight(&p->cpus_allowed) != 1)
+		return -EPERM;
+
+	/* Make sure we are bound to a cpu assigned to this resource. */
+	cpu = smp_processor_id();
+	BUG_ON(cpumask_first(&p->cpus_allowed) != cpu);
+	if (!cpumask_test_cpu(cpu, &info->cpumask))
+		return -EINVAL;
+
+	/* If we are already bound to this hardwall, it's a no-op. */
+	hwt = info->type;
+	if (ts->hardwall[hwt->index].info) {
+		BUG_ON(ts->hardwall[hwt->index].info != info);
+		return 0;
+	}
+
+	/* Success!  This process gets to use the resource on this cpu. */
+	ts->hardwall[hwt->index].info = info;
+	spin_lock_irqsave(&hwt->lock, flags);
+	list_add(&ts->hardwall[hwt->index].list, &info->task_head);
+	spin_unlock_irqrestore(&hwt->lock, flags);
+	grant_hardwall_mpls(hwt);
+	printk(KERN_DEBUG "Pid %d (%s) activated for %s hardwall: cpu %d\n",
+	       p->pid, p->comm, hwt->name, cpu);
+	return 0;
+}
+
+/*
+ * Deactivate a task's hardwall.  Must hold lock for hardwall_type.
+ * This method may be called from exit_thread(), so we don't want to
+ * rely on too many fields of struct task_struct still being valid.
+ * We assume the cpus_allowed, pid, and comm fields are still valid.
+ */
+static void _hardwall_deactivate(struct hardwall_type *hwt,
+				 struct task_struct *task)
+{
+	struct thread_struct *ts = &task->thread;
+
+	if (cpumask_weight(&task->cpus_allowed) != 1) {
+		pr_err("pid %d (%s) releasing %s hardwall with an affinity mask containing %d cpus!\n",
+		       task->pid, task->comm, hwt->name,
+		       cpumask_weight(&task->cpus_allowed));
+		BUG();
+	}
+
+	BUG_ON(ts->hardwall[hwt->index].info == NULL);
+	ts->hardwall[hwt->index].info = NULL;
+	list_del(&ts->hardwall[hwt->index].list);
+	if (task == current)
+		restrict_hardwall_mpls(hwt);
+}
+
+/* Deactivate a task's hardwall. */
+static int hardwall_deactivate(struct hardwall_type *hwt,
+			       struct task_struct *task)
+{
+	unsigned long flags;
+	int activated;
+
+	spin_lock_irqsave(&hwt->lock, flags);
+	activated = (task->thread.hardwall[hwt->index].info != NULL);
+	if (activated)
+		_hardwall_deactivate(hwt, task);
+	spin_unlock_irqrestore(&hwt->lock, flags);
+
+	if (!activated)
+		return -EINVAL;
+
+	printk(KERN_DEBUG "Pid %d (%s) deactivated for %s hardwall: cpu %d\n",
+	       task->pid, task->comm, hwt->name, raw_smp_processor_id());
+	return 0;
+}
+
+void hardwall_deactivate_all(struct task_struct *task)
+{
+	int i;
+	for (i = 0; i < HARDWALL_TYPES; ++i)
+		if (task->thread.hardwall[i].info)
+			hardwall_deactivate(&hardwall_types[i], task);
+}
+
+/* Stop the switch before draining the network. */
+static void stop_xdn_switch(void *arg)
+{
+#if !CHIP_HAS_REV1_XDN()
+	/* Freeze the switch and the demux. */
+	__insn_mtspr(SPR_UDN_SP_FREEZE,
+		     SPR_UDN_SP_FREEZE__SP_FRZ_MASK |
+		     SPR_UDN_SP_FREEZE__DEMUX_FRZ_MASK |
+		     SPR_UDN_SP_FREEZE__NON_DEST_EXT_MASK);
+#else
+	/*
+	 * Drop all packets bound for the core or off the edge.
+	 * We rely on the normal hardwall protection setup code
+	 * to have set the low four bits to trigger firewall interrupts,
+	 * and shift those bits up to trigger "drop on send" semantics,
+	 * plus adding "drop on send to core" for all switches.
+	 * In practice it seems the switches latch the DIRECTION_PROTECT
+	 * SPR so they won't start dropping if they're already
+	 * delivering the last message to the core, but it doesn't
+	 * hurt to enable it here.
+	 */
+	struct hardwall_type *hwt = arg;
+	unsigned long protect = mfspr_XDN(hwt, DIRECTION_PROTECT);
+	mtspr_XDN(hwt, DIRECTION_PROTECT, (protect | C_PROTECT) << 5);
+#endif
+}
+
+static void empty_xdn_demuxes(struct hardwall_type *hwt)
+{
+#ifndef __tilepro__
+	if (hwt->is_idn) {
+		while (__insn_mfspr(SPR_IDN_DATA_AVAIL) & (1 << 0))
+			(void) __tile_idn0_receive();
+		while (__insn_mfspr(SPR_IDN_DATA_AVAIL) & (1 << 1))
+			(void) __tile_idn1_receive();
+		return;
+	}
+#endif
+	while (__insn_mfspr(SPR_UDN_DATA_AVAIL) & (1 << 0))
+		(void) __tile_udn0_receive();
+	while (__insn_mfspr(SPR_UDN_DATA_AVAIL) & (1 << 1))
+		(void) __tile_udn1_receive();
+	while (__insn_mfspr(SPR_UDN_DATA_AVAIL) & (1 << 2))
+		(void) __tile_udn2_receive();
+	while (__insn_mfspr(SPR_UDN_DATA_AVAIL) & (1 << 3))
+		(void) __tile_udn3_receive();
+}
+
+/* Drain all the state from a stopped switch. */
+static void drain_xdn_switch(void *arg)
+{
+	struct hardwall_info *info = arg;
+	struct hardwall_type *hwt = info->type;
+
+#if CHIP_HAS_REV1_XDN()
+	/*
+	 * The switches have been configured to drop any messages
+	 * destined for cores (or off the edge of the rectangle).
+	 * But the current message may continue to be delivered,
+	 * so we wait until all the cores have finished any pending
+	 * messages before we stop draining.
+	 */
+	int pending = mfspr_XDN(hwt, PENDING);
+	while (pending--) {
+		empty_xdn_demuxes(hwt);
+		if (hwt->is_idn)
+			__tile_idn_send(0);
+		else
+			__tile_udn_send(0);
+	}
+	atomic_dec(&info->xdn_pending_count);
+	while (atomic_read(&info->xdn_pending_count))
+		empty_xdn_demuxes(hwt);
+#else
+	int i;
+	int from_tile_words, ca_count;
+
+	/* Empty out the 5 switch point fifos. */
+	for (i = 0; i < 5; i++) {
+		int words, j;
+		__insn_mtspr(SPR_UDN_SP_FIFO_SEL, i);
+		words = __insn_mfspr(SPR_UDN_SP_STATE) & 0xF;
+		for (j = 0; j < words; j++)
+			(void) __insn_mfspr(SPR_UDN_SP_FIFO_DATA);
+		BUG_ON((__insn_mfspr(SPR_UDN_SP_STATE) & 0xF) != 0);
+	}
+
+	/* Dump out the 3 word fifo at top. */
+	from_tile_words = (__insn_mfspr(SPR_UDN_DEMUX_STATUS) >> 10) & 0x3;
+	for (i = 0; i < from_tile_words; i++)
+		(void) __insn_mfspr(SPR_UDN_DEMUX_WRITE_FIFO);
+
+	/* Empty out demuxes. */
+	empty_xdn_demuxes(hwt);
+
+	/* Empty out catch all. */
+	ca_count = __insn_mfspr(SPR_UDN_DEMUX_CA_COUNT);
+	for (i = 0; i < ca_count; i++)
+		(void) __insn_mfspr(SPR_UDN_CA_DATA);
+	BUG_ON(__insn_mfspr(SPR_UDN_DEMUX_CA_COUNT) != 0);
+
+	/* Clear demux logic. */
+	__insn_mtspr(SPR_UDN_DEMUX_CTL, 1);
+
+	/*
+	 * Write switch state; experimentation indicates that 0xc3000
+	 * is an idle switch point.
+	 */
+	for (i = 0; i < 5; i++) {
+		__insn_mtspr(SPR_UDN_SP_FIFO_SEL, i);
+		__insn_mtspr(SPR_UDN_SP_STATE, 0xc3000);
+	}
+#endif
+}
+
+/* Reset random XDN state registers at boot up and during hardwall teardown. */
+static void reset_xdn_network_state(struct hardwall_type *hwt)
+{
+	if (hwt->disabled)
+		return;
+
+	/* Clear out other random registers so we have a clean slate. */
+	mtspr_XDN(hwt, DIRECTION_PROTECT, 0);
+	mtspr_XDN(hwt, AVAIL_EN, 0);
+	mtspr_XDN(hwt, DEADLOCK_TIMEOUT, 0);
+
+#if !CHIP_HAS_REV1_XDN()
+	/* Reset UDN coordinates to their standard value */
+	{
+		unsigned int cpu = smp_processor_id();
+		unsigned int x = cpu_x(cpu);
+		unsigned int y = cpu_y(cpu);
+		__insn_mtspr(SPR_UDN_TILE_COORD, (x << 18) | (y << 7));
+	}
+
+	/* Set demux tags to predefined values and enable them. */
+	__insn_mtspr(SPR_UDN_TAG_VALID, 0xf);
+	__insn_mtspr(SPR_UDN_TAG_0, (1 << 0));
+	__insn_mtspr(SPR_UDN_TAG_1, (1 << 1));
+	__insn_mtspr(SPR_UDN_TAG_2, (1 << 2));
+	__insn_mtspr(SPR_UDN_TAG_3, (1 << 3));
+
+	/* Set other rev0 random registers to a clean state. */
+	__insn_mtspr(SPR_UDN_REFILL_EN, 0);
+	__insn_mtspr(SPR_UDN_DEMUX_QUEUE_SEL, 0);
+	__insn_mtspr(SPR_UDN_SP_FIFO_SEL, 0);
+
+	/* Start the switch and demux. */
+	__insn_mtspr(SPR_UDN_SP_FREEZE, 0);
+#endif
+}
+
+void reset_network_state(void)
+{
+	reset_xdn_network_state(&hardwall_types[HARDWALL_UDN]);
+#ifndef __tilepro__
+	reset_xdn_network_state(&hardwall_types[HARDWALL_IDN]);
+#endif
+}
+
+/* Restart an XDN switch after draining. */
+static void restart_xdn_switch(void *arg)
+{
+	struct hardwall_type *hwt = arg;
+
+#if CHIP_HAS_REV1_XDN()
+	/* One last drain step to avoid races with injection and draining. */
+	empty_xdn_demuxes(hwt);
+#endif
+
+	reset_xdn_network_state(hwt);
+
+	/* Disable firewall interrupts. */
+	disable_firewall_interrupts(hwt);
+}
+
+/* Last reference to a hardwall is gone, so clear the network. */
+static void hardwall_destroy(struct hardwall_info *info)
+{
+	struct task_struct *task;
+	struct hardwall_type *hwt;
+	unsigned long flags;
+
+	/* Make sure this file actually represents a hardwall. */
+	if (info == NULL)
+		return;
+
+	/*
+	 * Deactivate any remaining tasks.  It's possible to race with
+	 * some other thread that is exiting and hasn't yet called
+	 * deactivate (when freeing its thread_info), so we carefully
+	 * deactivate any remaining tasks before freeing the
+	 * hardwall_info object itself.
+	 */
+	hwt = info->type;
+	info->teardown_in_progress = 1;
+	spin_lock_irqsave(&hwt->lock, flags);
+	list_for_each_entry(task, &info->task_head,
+			    thread.hardwall[hwt->index].list)
+		_hardwall_deactivate(hwt, task);
+	spin_unlock_irqrestore(&hwt->lock, flags);
+
+	if (hwt->is_xdn) {
+		/* Configure the switches for draining the user network. */
+		printk(KERN_DEBUG
+		       "Clearing %s hardwall rectangle %dx%d %d,%d\n",
+		       hwt->name, info->width, info->height,
+		       info->ulhc_x, info->ulhc_y);
+		on_each_cpu_mask(&info->cpumask, stop_xdn_switch, hwt, 1);
+
+		/* Drain the network. */
+#if CHIP_HAS_REV1_XDN()
+		atomic_set(&info->xdn_pending_count,
+			   cpumask_weight(&info->cpumask));
+		on_each_cpu_mask(&info->cpumask, drain_xdn_switch, info, 0);
+#else
+		on_each_cpu_mask(&info->cpumask, drain_xdn_switch, info, 1);
+#endif
+
+		/* Restart switch and disable firewall. */
+		on_each_cpu_mask(&info->cpumask, restart_xdn_switch, hwt, 1);
+	}
+
+	/* Remove the /proc/tile/hardwall entry. */
+	hardwall_remove_proc(info);
+
+	/* Now free the hardwall from the list. */
+	spin_lock_irqsave(&hwt->lock, flags);
+	BUG_ON(!list_empty(&info->task_head));
+	list_del(&info->list);
+	spin_unlock_irqrestore(&hwt->lock, flags);
+	kfree(info);
+}
+
+
+static int hardwall_proc_show(struct seq_file *sf, void *v)
+{
+	struct hardwall_info *info = sf->private;
+
+	seq_printf(sf, "%*pbl\n", cpumask_pr_args(&info->cpumask));
+	return 0;
+}
+
+static int hardwall_proc_open(struct inode *inode,
+			      struct file *file)
+{
+	return single_open(file, hardwall_proc_show, PDE_DATA(inode));
+}
+
+static const struct file_operations hardwall_proc_fops = {
+	.open		= hardwall_proc_open,
+	.read		= seq_read,
+	.llseek		= seq_lseek,
+	.release	= single_release,
+};
+
+static void hardwall_add_proc(struct hardwall_info *info)
+{
+	char buf[64];
+	snprintf(buf, sizeof(buf), "%d", info->id);
+	proc_create_data(buf, 0444, info->type->proc_dir,
+			 &hardwall_proc_fops, info);
+}
+
+static void hardwall_remove_proc(struct hardwall_info *info)
+{
+	char buf[64];
+	snprintf(buf, sizeof(buf), "%d", info->id);
+	remove_proc_entry(buf, info->type->proc_dir);
+}
+
+int proc_pid_hardwall(struct seq_file *m, struct pid_namespace *ns,
+		      struct pid *pid, struct task_struct *task)
+{
+	int i;
+	int n = 0;
+	for (i = 0; i < HARDWALL_TYPES; ++i) {
+		struct hardwall_info *info = task->thread.hardwall[i].info;
+		if (info)
+			seq_printf(m, "%s: %d\n", info->type->name, info->id);
+	}
+	return n;
+}
+
+void proc_tile_hardwall_init(struct proc_dir_entry *root)
+{
+	int i;
+	for (i = 0; i < HARDWALL_TYPES; ++i) {
+		struct hardwall_type *hwt = &hardwall_types[i];
+		if (hwt->disabled)
+			continue;
+		if (hardwall_proc_dir == NULL)
+			hardwall_proc_dir = proc_mkdir("hardwall", root);
+		hwt->proc_dir = proc_mkdir(hwt->name, hardwall_proc_dir);
+	}
+}
+
+
+/*
+ * Character device support via ioctl/close.
+ */
+
+static long hardwall_ioctl(struct file *file, unsigned int a, unsigned long b)
+{
+	struct hardwall_info *info = file->private_data;
+	int minor = iminor(file->f_mapping->host);
+	struct hardwall_type* hwt;
+
+	if (_IOC_TYPE(a) != HARDWALL_IOCTL_BASE)
+		return -EINVAL;
+
+	BUILD_BUG_ON(HARDWALL_TYPES != _HARDWALL_TYPES);
+	BUILD_BUG_ON(HARDWALL_TYPES !=
+		     sizeof(hardwall_types)/sizeof(hardwall_types[0]));
+
+	if (minor < 0 || minor >= HARDWALL_TYPES)
+		return -EINVAL;
+	hwt = &hardwall_types[minor];
+	WARN_ON(info && hwt != info->type);
+
+	switch (_IOC_NR(a)) {
+	case _HARDWALL_CREATE:
+		if (hwt->disabled)
+			return -ENOSYS;
+		if (info != NULL)
+			return -EALREADY;
+		info = hardwall_create(hwt, _IOC_SIZE(a),
+				       (const unsigned char __user *)b);
+		if (IS_ERR(info))
+			return PTR_ERR(info);
+		file->private_data = info;
+		return 0;
+
+	case _HARDWALL_ACTIVATE:
+		return hardwall_activate(info);
+
+	case _HARDWALL_DEACTIVATE:
+		if (current->thread.hardwall[hwt->index].info != info)
+			return -EINVAL;
+		return hardwall_deactivate(hwt, current);
+
+	case _HARDWALL_GET_ID:
+		return info ? info->id : -EINVAL;
+
+	default:
+		return -EINVAL;
+	}
+}
+
+#ifdef CONFIG_COMPAT
+static long hardwall_compat_ioctl(struct file *file,
+				  unsigned int a, unsigned long b)
+{
+	/* Sign-extend the argument so it can be used as a pointer. */
+	return hardwall_ioctl(file, a, (unsigned long)compat_ptr(b));
+}
+#endif
+
+/* The user process closed the file; revoke access to user networks. */
+static int hardwall_flush(struct file *file, fl_owner_t owner)
+{
+	struct hardwall_info *info = file->private_data;
+	struct task_struct *task, *tmp;
+	unsigned long flags;
+
+	if (info) {
+		/*
+		 * NOTE: if multiple threads are activated on this hardwall
+		 * file, the other threads will continue having access to the
+		 * user network until they are context-switched out and back
+		 * in again.
+		 *
+		 * NOTE: A NULL files pointer means the task is being torn
+		 * down, so in that case we also deactivate it.
+		 */
+		struct hardwall_type *hwt = info->type;
+		spin_lock_irqsave(&hwt->lock, flags);
+		list_for_each_entry_safe(task, tmp, &info->task_head,
+					 thread.hardwall[hwt->index].list) {
+			if (task->files == owner || task->files == NULL)
+				_hardwall_deactivate(hwt, task);
+		}
+		spin_unlock_irqrestore(&hwt->lock, flags);
+	}
+
+	return 0;
+}
+
+/* This hardwall is gone, so destroy it. */
+static int hardwall_release(struct inode *inode, struct file *file)
+{
+	hardwall_destroy(file->private_data);
+	return 0;
+}
+
+static const struct file_operations dev_hardwall_fops = {
+	.open           = nonseekable_open,
+	.unlocked_ioctl = hardwall_ioctl,
+#ifdef CONFIG_COMPAT
+	.compat_ioctl   = hardwall_compat_ioctl,
+#endif
+	.flush          = hardwall_flush,
+	.release        = hardwall_release,
+};
+
+static struct cdev hardwall_dev;
+
+static int __init dev_hardwall_init(void)
+{
+	int rc;
+	dev_t dev;
+
+	rc = alloc_chrdev_region(&dev, 0, HARDWALL_TYPES, "hardwall");
+	if (rc < 0)
+		return rc;
+	cdev_init(&hardwall_dev, &dev_hardwall_fops);
+	rc = cdev_add(&hardwall_dev, dev, HARDWALL_TYPES);
+	if (rc < 0)
+		return rc;
+
+	return 0;
+}
+late_initcall(dev_hardwall_init);
diff --git a/arch/tile/kernel/head_32.S b/arch/tile/kernel/head_32.S
new file mode 100644
index 000000000..8d5b40ff2
--- /dev/null
+++ b/arch/tile/kernel/head_32.S
@@ -0,0 +1,183 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * TILE startup code.
+ */
+
+#include <linux/linkage.h>
+#include <linux/init.h>
+#include <asm/page.h>
+#include <asm/pgtable.h>
+#include <asm/thread_info.h>
+#include <asm/processor.h>
+#include <asm/asm-offsets.h>
+#include <hv/hypervisor.h>
+#include <arch/chip.h>
+#include <arch/spr_def.h>
+
+/*
+ * This module contains the entry code for kernel images. It performs the
+ * minimal setup needed to call the generic C routines.
+ */
+
+	__HEAD
+ENTRY(_start)
+	/* Notify the hypervisor of what version of the API we want */
+	{
+	  movei r1, TILE_CHIP
+	  movei r2, TILE_CHIP_REV
+	}
+	{
+	  moveli r0, _HV_VERSION_OLD_HV_INIT
+	  jal _hv_init
+	}
+	/* Get a reasonable default ASID in r0 */
+	{
+	  move r0, zero
+	  jal _hv_inquire_asid
+	}
+	/* Install the default page table */
+	{
+	  moveli r6, lo16(swapper_pgprot - PAGE_OFFSET)
+	  move r4, r0     /* use starting ASID of range for this page table */
+	}
+	{
+	  moveli r0, lo16(swapper_pg_dir - PAGE_OFFSET)
+	  auli r6, r6, ha16(swapper_pgprot - PAGE_OFFSET)
+	}
+	{
+	  lw r2, r6
+	  addi r6, r6, 4
+	}
+	{
+	  lw r3, r6
+	  auli r0, r0, ha16(swapper_pg_dir - PAGE_OFFSET)
+	}
+	{
+	  finv r6
+	  move r1, zero   /* high 32 bits of CPA is zero */
+	}
+	{
+	  moveli lr, lo16(1f)
+	  moveli r5, CTX_PAGE_FLAG
+	}
+	{
+	  auli lr, lr, ha16(1f)
+	  j _hv_install_context
+	}
+1:
+
+	/* Get our processor number and save it away in SAVE_K_0. */
+	jal _hv_inquire_topology
+	mulll_uu r4, r1, r2        /* r1 == y, r2 == width */
+	add r4, r4, r0             /* r0 == x, so r4 == cpu == y*width + x */
+
+#ifdef CONFIG_SMP
+	/*
+	 * Load up our per-cpu offset.  When the first (master) tile
+	 * boots, this value is still zero, so we will load boot_pc
+	 * with start_kernel, and boot_sp at the top of init_stack.
+	 * The master tile initializes the per-cpu offset array, so that
+	 * when subsequent (secondary) tiles boot, they will instead load
+	 * from their per-cpu versions of boot_sp and boot_pc.
+	 */
+	moveli r5, lo16(__per_cpu_offset)
+	auli r5, r5, ha16(__per_cpu_offset)
+	s2a r5, r4, r5
+	lw r5, r5
+	bnz r5, 1f
+
+	/*
+	 * Save the width and height to the smp_topology variable
+	 * for later use.
+	 */
+	moveli r0, lo16(smp_topology + HV_TOPOLOGY_WIDTH_OFFSET)
+	auli r0, r0, ha16(smp_topology + HV_TOPOLOGY_WIDTH_OFFSET)
+	{
+	  sw r0, r2
+	  addi r0, r0, (HV_TOPOLOGY_HEIGHT_OFFSET - HV_TOPOLOGY_WIDTH_OFFSET)
+	}
+	sw r0, r3
+1:
+#else
+	move r5, zero
+#endif
+
+	/* Load and go with the correct pc and sp. */
+	{
+	  addli r1, r5, lo16(boot_sp)
+	  addli r0, r5, lo16(boot_pc)
+	}
+	{
+	  auli r1, r1, ha16(boot_sp)
+	  auli r0, r0, ha16(boot_pc)
+	}
+	lw r0, r0
+	lw sp, r1
+	or r4, sp, r4
+	mtspr SPR_SYSTEM_SAVE_K_0, r4  /* save ksp0 + cpu */
+	{
+	  move lr, zero   /* stop backtraces in the called function */
+	  jr r0
+	}
+	ENDPROC(_start)
+
+__PAGE_ALIGNED_BSS
+	.align PAGE_SIZE
+ENTRY(empty_zero_page)
+	.fill PAGE_SIZE,1,0
+	END(empty_zero_page)
+
+	.macro PTE va, cpa, bits1, no_org=0
+	.ifeq \no_org
+	.org swapper_pg_dir + PGD_INDEX(\va) * HV_PTE_SIZE
+	.endif
+	.word HV_PTE_PAGE | HV_PTE_DIRTY | HV_PTE_PRESENT | HV_PTE_ACCESSED | \
+	      (HV_PTE_MODE_CACHE_NO_L3 << HV_PTE_INDEX_MODE)
+	.word (\bits1) | (HV_CPA_TO_PTFN(\cpa) << (HV_PTE_INDEX_PTFN - 32))
+	.endm
+
+__PAGE_ALIGNED_DATA
+	.align PAGE_SIZE
+ENTRY(swapper_pg_dir)
+	/*
+	 * All data pages from PAGE_OFFSET to MEM_USER_INTRPT are mapped as
+	 * VA = PA + PAGE_OFFSET.  We remap things with more precise access
+	 * permissions and more respect for size of RAM later.
+	 */
+	.set addr, 0
+	.rept (MEM_USER_INTRPT - PAGE_OFFSET) >> PGDIR_SHIFT
+	PTE addr + PAGE_OFFSET, addr, (1 << (HV_PTE_INDEX_READABLE - 32)) | \
+				      (1 << (HV_PTE_INDEX_WRITABLE - 32))
+	.set addr, addr + PGDIR_SIZE
+	.endr
+
+	/* The true text VAs are mapped as VA = PA + MEM_SV_START */
+	PTE MEM_SV_START, 0, (1 << (HV_PTE_INDEX_READABLE - 32)) | \
+			      (1 << (HV_PTE_INDEX_EXECUTABLE - 32))
+	.org swapper_pg_dir + PGDIR_SIZE
+	END(swapper_pg_dir)
+
+	/*
+	 * Isolate swapper_pgprot to its own cache line, since each cpu
+	 * starting up will read it using VA-is-PA and local homing.
+	 * This would otherwise likely conflict with other data on the cache
+	 * line, once we have set its permanent home in the page tables.
+	 */
+	__INITDATA
+	.align CHIP_L2_LINE_SIZE()
+ENTRY(swapper_pgprot)
+	PTE	0, 0, (1 << (HV_PTE_INDEX_READABLE - 32)) | \
+		      (1 << (HV_PTE_INDEX_WRITABLE - 32)), 1
+	.align CHIP_L2_LINE_SIZE()
+	END(swapper_pgprot)
diff --git a/arch/tile/kernel/head_64.S b/arch/tile/kernel/head_64.S
new file mode 100644
index 000000000..bd0e12f28
--- /dev/null
+++ b/arch/tile/kernel/head_64.S
@@ -0,0 +1,279 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * TILE startup code.
+ */
+
+#include <linux/linkage.h>
+#include <linux/init.h>
+#include <asm/page.h>
+#include <asm/pgtable.h>
+#include <asm/thread_info.h>
+#include <asm/processor.h>
+#include <asm/asm-offsets.h>
+#include <hv/hypervisor.h>
+#include <arch/chip.h>
+#include <arch/spr_def.h>
+
+/* Extract two 32-bit bit values that were read into one register. */
+#ifdef __BIG_ENDIAN__
+#define GET_FIRST_INT(rd, rs) shrsi rd, rs, 32
+#define GET_SECOND_INT(rd, rs) addxi rd, rs, 0
+#else
+#define GET_FIRST_INT(rd, rs) addxi rd, rs, 0
+#define GET_SECOND_INT(rd, rs) shrsi rd, rs, 32
+#endif
+
+/*
+ * This module contains the entry code for kernel images. It performs the
+ * minimal setup needed to call the generic C routines.
+ */
+
+	__HEAD
+ENTRY(_start)
+	/* Notify the hypervisor of what version of the API we want */
+	{
+#if KERNEL_PL == 1 && _HV_VERSION == 13
+	  /* Support older hypervisors by asking for API version 12. */
+	  movei r0, _HV_VERSION_OLD_HV_INIT
+#else
+	  movei r0, _HV_VERSION
+#endif
+	  movei r1, TILE_CHIP
+	}
+	{
+	  movei r2, TILE_CHIP_REV
+	  movei r3, KERNEL_PL
+	}
+	jal _hv_init
+	/* Get a reasonable default ASID in r0 */
+	{
+	  move r0, zero
+	  jal _hv_inquire_asid
+	}
+
+	/*
+	 * Install the default page table.  The relocation required to
+	 * statically define the table is a bit too complex, so we have
+	 * to plug in the pointer from the L0 to the L1 table by hand.
+	 * We only do this on the first cpu to boot, though, since the
+	 * other CPUs should see a properly-constructed page table.
+	 */
+	{
+	  GET_FIRST_INT(r2, r0)    /* ASID for hv_install_context */
+	  moveli r4, hw1_last(swapper_pgprot - PAGE_OFFSET)
+	}
+	{
+	  shl16insli r4, r4, hw0(swapper_pgprot - PAGE_OFFSET)
+	}
+	{
+	  ld r1, r4               /* access_pte for hv_install_context */
+	}
+	{
+	  moveli r0, hw1_last(.Lsv_data_pmd - PAGE_OFFSET)
+	  moveli r6, hw1_last(temp_data_pmd - PAGE_OFFSET)
+	}
+	{
+	  /* After initializing swapper_pgprot, HV_PTE_GLOBAL is set. */
+	  bfextu r7, r1, HV_PTE_INDEX_GLOBAL, HV_PTE_INDEX_GLOBAL
+	  finv r4
+	}
+	bnez r7, .Lno_write
+	{
+	  shl16insli r0, r0, hw0(.Lsv_data_pmd - PAGE_OFFSET)
+	  shl16insli r6, r6, hw0(temp_data_pmd - PAGE_OFFSET)
+	}
+	{
+	  /* Cut off the low bits of the PT address. */
+	  shrui r6, r6, HV_LOG2_PAGE_TABLE_ALIGN
+	  /* Start with our access pte. */
+	  move r5, r1
+	}
+	{
+	  /* Stuff the address into the page table pointer slot of the PTE. */
+	  bfins r5, r6, HV_PTE_INDEX_PTFN, \
+			HV_PTE_INDEX_PTFN + HV_PTE_PTFN_BITS - 1
+	}
+	{
+	  /* Store the L0 data PTE. */
+	  st r0, r5
+	  addli r6, r6, (temp_code_pmd - temp_data_pmd) >> \
+			HV_LOG2_PAGE_TABLE_ALIGN
+	}
+	{
+	  addli r0, r0, .Lsv_code_pmd - .Lsv_data_pmd
+	  bfins r5, r6, HV_PTE_INDEX_PTFN, \
+			HV_PTE_INDEX_PTFN + HV_PTE_PTFN_BITS - 1
+	}
+	/* Store the L0 code PTE. */
+	st r0, r5
+
+.Lno_write:
+	moveli lr, hw2_last(1f)
+	{
+	  shl16insli lr, lr, hw1(1f)
+	  moveli r0, hw1_last(swapper_pg_dir - PAGE_OFFSET)
+	}
+	{
+	  shl16insli lr, lr, hw0(1f)
+	  shl16insli r0, r0, hw0(swapper_pg_dir - PAGE_OFFSET)
+	}
+	{
+	  moveli r3, CTX_PAGE_FLAG
+	  j _hv_install_context
+	}
+1:
+
+	/* Install the interrupt base. */
+	moveli r0, hw2_last(intrpt_start)
+	shl16insli r0, r0, hw1(intrpt_start)
+	shl16insli r0, r0, hw0(intrpt_start)
+	mtspr SPR_INTERRUPT_VECTOR_BASE_K, r0
+
+	/* Get our processor number and save it away in SAVE_K_0. */
+	jal _hv_inquire_topology
+	{
+	  GET_FIRST_INT(r5, r1)   /* r5 = width */
+	  GET_SECOND_INT(r4, r0)  /* r4 = y */
+	}
+	{
+	  GET_FIRST_INT(r6, r0)   /* r6 = x */
+	  mul_lu_lu r4, r4, r5
+	}
+	{
+	  add r4, r4, r6          /* r4 == cpu == y*width + x */
+	}
+
+#ifdef CONFIG_SMP
+	/*
+	 * Load up our per-cpu offset.  When the first (master) tile
+	 * boots, this value is still zero, so we will load boot_pc
+	 * with start_kernel, and boot_sp with at the top of init_stack.
+	 * The master tile initializes the per-cpu offset array, so that
+	 * when subsequent (secondary) tiles boot, they will instead load
+	 * from their per-cpu versions of boot_sp and boot_pc.
+	 */
+	moveli r5, hw2_last(__per_cpu_offset)
+	shl16insli r5, r5, hw1(__per_cpu_offset)
+	shl16insli r5, r5, hw0(__per_cpu_offset)
+	shl3add r5, r4, r5
+	ld r5, r5
+	bnez r5, 1f
+
+	/*
+	 * Save the width and height to the smp_topology variable
+	 * for later use.
+	 */
+	moveli r0, hw2_last(smp_topology + HV_TOPOLOGY_WIDTH_OFFSET)
+	shl16insli r0, r0, hw1(smp_topology + HV_TOPOLOGY_WIDTH_OFFSET)
+	shl16insli r0, r0, hw0(smp_topology + HV_TOPOLOGY_WIDTH_OFFSET)
+	st r0, r1
+1:
+#else
+	move r5, zero
+#endif
+
+	/* Load and go with the correct pc and sp. */
+	{
+	  moveli r1, hw2_last(boot_sp)
+	  moveli r0, hw2_last(boot_pc)
+	}
+	{
+	  shl16insli r1, r1, hw1(boot_sp)
+	  shl16insli r0, r0, hw1(boot_pc)
+	}
+	{
+	  shl16insli r1, r1, hw0(boot_sp)
+	  shl16insli r0, r0, hw0(boot_pc)
+	}
+	{
+	  add r1, r1, r5
+	  add r0, r0, r5
+	}
+	ld r0, r0
+	ld sp, r1
+	shli r4, r4, CPU_SHIFT
+	bfins r4, sp, 0, CPU_SHIFT-1
+	mtspr SPR_SYSTEM_SAVE_K_0, r4  /* save ksp0 + cpu */
+	{
+	  move lr, zero   /* stop backtraces in the called function */
+	  jr r0
+	}
+	ENDPROC(_start)
+
+__PAGE_ALIGNED_BSS
+	.align PAGE_SIZE
+ENTRY(empty_zero_page)
+	.fill PAGE_SIZE,1,0
+	END(empty_zero_page)
+
+	.macro PTE cpa, bits1
+	.quad HV_PTE_PAGE | HV_PTE_DIRTY | HV_PTE_PRESENT | HV_PTE_ACCESSED |\
+	      HV_PTE_GLOBAL | (HV_PTE_MODE_CACHE_NO_L3 << HV_PTE_INDEX_MODE) |\
+	      (\bits1) | (HV_CPA_TO_PTFN(\cpa) << HV_PTE_INDEX_PTFN)
+	.endm
+
+__PAGE_ALIGNED_DATA
+	.align PAGE_SIZE
+ENTRY(swapper_pg_dir)
+	.org swapper_pg_dir + PGD_INDEX(PAGE_OFFSET) * HV_PTE_SIZE
+.Lsv_data_pmd:
+	.quad 0  /* PTE temp_data_pmd - PAGE_OFFSET, 0 */
+	.org swapper_pg_dir + PGD_INDEX(MEM_SV_START) * HV_PTE_SIZE
+.Lsv_code_pmd:
+	.quad 0  /* PTE temp_code_pmd - PAGE_OFFSET, 0 */
+	.org swapper_pg_dir + SIZEOF_PGD
+	END(swapper_pg_dir)
+
+	.align HV_PAGE_TABLE_ALIGN
+ENTRY(temp_data_pmd)
+	/*
+	 * We fill the PAGE_OFFSET pmd with huge pages with
+	 * VA = PA + PAGE_OFFSET.  We remap things with more precise access
+	 * permissions later.
+	 */
+	.set addr, 0
+	.rept PTRS_PER_PMD
+	PTE addr, HV_PTE_READABLE | HV_PTE_WRITABLE
+	.set addr, addr + HPAGE_SIZE
+	.endr
+	.org temp_data_pmd + SIZEOF_PMD
+	END(temp_data_pmd)
+
+	.align HV_PAGE_TABLE_ALIGN
+ENTRY(temp_code_pmd)
+	/*
+	 * We fill the MEM_SV_START pmd with huge pages with
+	 * VA = PA + PAGE_OFFSET.  We remap things with more precise access
+	 * permissions later.
+	 */
+	.set addr, 0
+	.rept PTRS_PER_PMD
+	PTE addr, HV_PTE_READABLE | HV_PTE_EXECUTABLE
+	.set addr, addr + HPAGE_SIZE
+	.endr
+	.org temp_code_pmd + SIZEOF_PMD
+	END(temp_code_pmd)
+
+	/*
+	 * Isolate swapper_pgprot to its own cache line, since each cpu
+	 * starting up will read it using VA-is-PA and local homing.
+	 * This would otherwise likely conflict with other data on the cache
+	 * line, once we have set its permanent home in the page tables.
+	 */
+	__INITDATA
+	.align CHIP_L2_LINE_SIZE()
+ENTRY(swapper_pgprot)
+	.quad HV_PTE_PRESENT | (HV_PTE_MODE_CACHE_NO_L3 << HV_PTE_INDEX_MODE)
+	.align CHIP_L2_LINE_SIZE()
+	END(swapper_pgprot)
diff --git a/arch/tile/kernel/hvglue.S b/arch/tile/kernel/hvglue.S
new file mode 100644
index 000000000..2ab456622
--- /dev/null
+++ b/arch/tile/kernel/hvglue.S
@@ -0,0 +1,74 @@
+/* Hypervisor call vector addresses; see <hv/hypervisor.h> */
+.macro gensym sym, val, size
+.org \val
+.global _\sym
+.type _\sym,function
+_\sym:
+.size _\sym,\size
+#ifndef CONFIG_TILE_HVGLUE_TRACE
+.globl \sym
+.set \sym,_\sym
+#endif
+.endm
+
+.section .hvglue,"x",@nobits
+.align 8
+gensym hv_init, 0x20, 32
+gensym hv_install_context, 0x40, 32
+gensym hv_sysconf, 0x60, 32
+gensym hv_get_rtc, 0x80, 32
+gensym hv_set_rtc, 0xa0, 32
+gensym hv_flush_asid, 0xc0, 32
+gensym hv_flush_page, 0xe0, 32
+gensym hv_flush_pages, 0x100, 32
+gensym hv_restart, 0x120, 32
+gensym hv_halt, 0x140, 32
+gensym hv_power_off, 0x160, 32
+gensym hv_inquire_physical, 0x180, 32
+gensym hv_inquire_memory_controller, 0x1a0, 32
+gensym hv_inquire_virtual, 0x1c0, 32
+gensym hv_inquire_asid, 0x1e0, 32
+gensym hv_nanosleep, 0x200, 32
+gensym hv_console_read_if_ready, 0x220, 32
+gensym hv_console_write, 0x240, 32
+gensym hv_downcall_dispatch, 0x260, 32
+gensym hv_inquire_topology, 0x280, 32
+gensym hv_fs_findfile, 0x2a0, 32
+gensym hv_fs_fstat, 0x2c0, 32
+gensym hv_fs_pread, 0x2e0, 32
+gensym hv_physaddr_read64, 0x300, 32
+gensym hv_physaddr_write64, 0x320, 32
+gensym hv_get_command_line, 0x340, 32
+gensym hv_set_caching, 0x360, 32
+gensym hv_bzero_page, 0x380, 32
+gensym hv_register_message_state, 0x3a0, 32
+gensym hv_send_message, 0x3c0, 32
+gensym hv_receive_message, 0x3e0, 32
+gensym hv_inquire_context, 0x400, 32
+gensym hv_start_all_tiles, 0x420, 32
+gensym hv_dev_open, 0x440, 32
+gensym hv_dev_close, 0x460, 32
+gensym hv_dev_pread, 0x480, 32
+gensym hv_dev_pwrite, 0x4a0, 32
+gensym hv_dev_poll, 0x4c0, 32
+gensym hv_dev_poll_cancel, 0x4e0, 32
+gensym hv_dev_preada, 0x500, 32
+gensym hv_dev_pwritea, 0x520, 32
+gensym hv_flush_remote, 0x540, 32
+gensym hv_console_putc, 0x560, 32
+gensym hv_inquire_tiles, 0x580, 32
+gensym hv_confstr, 0x5a0, 32
+gensym hv_reexec, 0x5c0, 32
+gensym hv_set_command_line, 0x5e0, 32
+gensym hv_clear_intr, 0x600, 32
+gensym hv_enable_intr, 0x620, 32
+gensym hv_disable_intr, 0x640, 32
+gensym hv_raise_intr, 0x660, 32
+gensym hv_trigger_ipi, 0x680, 32
+gensym hv_store_mapping, 0x6a0, 32
+gensym hv_inquire_realpa, 0x6c0, 32
+gensym hv_flush_all, 0x6e0, 32
+gensym hv_get_ipi_pte, 0x700, 32
+gensym hv_set_pte_super_shift, 0x720, 32
+gensym hv_console_set_ipi, 0x7e0, 32
+gensym hv_glue_internals, 0x800, 30720
diff --git a/arch/tile/kernel/hvglue_trace.c b/arch/tile/kernel/hvglue_trace.c
new file mode 100644
index 000000000..85c74ad29
--- /dev/null
+++ b/arch/tile/kernel/hvglue_trace.c
@@ -0,0 +1,266 @@
+/*
+ * Copyright 2013 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/*
+ * Pull in the hypervisor header so we declare all the ABI functions
+ * with the underscore versions, then undef the names so that we can
+ * provide our own wrapper versions.
+ */
+#define hv_init _hv_init
+#define hv_install_context _hv_install_context
+#define hv_sysconf _hv_sysconf
+#define hv_get_rtc _hv_get_rtc
+#define hv_set_rtc _hv_set_rtc
+#define hv_flush_asid _hv_flush_asid
+#define hv_flush_page _hv_flush_page
+#define hv_flush_pages _hv_flush_pages
+#define hv_restart _hv_restart
+#define hv_halt _hv_halt
+#define hv_power_off _hv_power_off
+#define hv_inquire_physical _hv_inquire_physical
+#define hv_inquire_memory_controller _hv_inquire_memory_controller
+#define hv_inquire_virtual _hv_inquire_virtual
+#define hv_inquire_asid _hv_inquire_asid
+#define hv_nanosleep _hv_nanosleep
+#define hv_console_read_if_ready _hv_console_read_if_ready
+#define hv_console_write _hv_console_write
+#define hv_downcall_dispatch _hv_downcall_dispatch
+#define hv_inquire_topology _hv_inquire_topology
+#define hv_fs_findfile _hv_fs_findfile
+#define hv_fs_fstat _hv_fs_fstat
+#define hv_fs_pread _hv_fs_pread
+#define hv_physaddr_read64 _hv_physaddr_read64
+#define hv_physaddr_write64 _hv_physaddr_write64
+#define hv_get_command_line _hv_get_command_line
+#define hv_set_caching _hv_set_caching
+#define hv_bzero_page _hv_bzero_page
+#define hv_register_message_state _hv_register_message_state
+#define hv_send_message _hv_send_message
+#define hv_receive_message _hv_receive_message
+#define hv_inquire_context _hv_inquire_context
+#define hv_start_all_tiles _hv_start_all_tiles
+#define hv_dev_open _hv_dev_open
+#define hv_dev_close _hv_dev_close
+#define hv_dev_pread _hv_dev_pread
+#define hv_dev_pwrite _hv_dev_pwrite
+#define hv_dev_poll _hv_dev_poll
+#define hv_dev_poll_cancel _hv_dev_poll_cancel
+#define hv_dev_preada _hv_dev_preada
+#define hv_dev_pwritea _hv_dev_pwritea
+#define hv_flush_remote _hv_flush_remote
+#define hv_console_putc _hv_console_putc
+#define hv_inquire_tiles _hv_inquire_tiles
+#define hv_confstr _hv_confstr
+#define hv_reexec _hv_reexec
+#define hv_set_command_line _hv_set_command_line
+#define hv_clear_intr _hv_clear_intr
+#define hv_enable_intr _hv_enable_intr
+#define hv_disable_intr _hv_disable_intr
+#define hv_raise_intr _hv_raise_intr
+#define hv_trigger_ipi _hv_trigger_ipi
+#define hv_store_mapping _hv_store_mapping
+#define hv_inquire_realpa _hv_inquire_realpa
+#define hv_flush_all _hv_flush_all
+#define hv_get_ipi_pte _hv_get_ipi_pte
+#define hv_set_pte_super_shift _hv_set_pte_super_shift
+#define hv_console_set_ipi _hv_console_set_ipi
+#include <hv/hypervisor.h>
+#undef hv_init
+#undef hv_install_context
+#undef hv_sysconf
+#undef hv_get_rtc
+#undef hv_set_rtc
+#undef hv_flush_asid
+#undef hv_flush_page
+#undef hv_flush_pages
+#undef hv_restart
+#undef hv_halt
+#undef hv_power_off
+#undef hv_inquire_physical
+#undef hv_inquire_memory_controller
+#undef hv_inquire_virtual
+#undef hv_inquire_asid
+#undef hv_nanosleep
+#undef hv_console_read_if_ready
+#undef hv_console_write
+#undef hv_downcall_dispatch
+#undef hv_inquire_topology
+#undef hv_fs_findfile
+#undef hv_fs_fstat
+#undef hv_fs_pread
+#undef hv_physaddr_read64
+#undef hv_physaddr_write64
+#undef hv_get_command_line
+#undef hv_set_caching
+#undef hv_bzero_page
+#undef hv_register_message_state
+#undef hv_send_message
+#undef hv_receive_message
+#undef hv_inquire_context
+#undef hv_start_all_tiles
+#undef hv_dev_open
+#undef hv_dev_close
+#undef hv_dev_pread
+#undef hv_dev_pwrite
+#undef hv_dev_poll
+#undef hv_dev_poll_cancel
+#undef hv_dev_preada
+#undef hv_dev_pwritea
+#undef hv_flush_remote
+#undef hv_console_putc
+#undef hv_inquire_tiles
+#undef hv_confstr
+#undef hv_reexec
+#undef hv_set_command_line
+#undef hv_clear_intr
+#undef hv_enable_intr
+#undef hv_disable_intr
+#undef hv_raise_intr
+#undef hv_trigger_ipi
+#undef hv_store_mapping
+#undef hv_inquire_realpa
+#undef hv_flush_all
+#undef hv_get_ipi_pte
+#undef hv_set_pte_super_shift
+#undef hv_console_set_ipi
+
+/*
+ * Provide macros based on <linux/syscalls.h> to provide a wrapper
+ * function that invokes the same function with an underscore prefix.
+ * We can't use the existing __SC_xxx macros because we need to
+ * support up to nine arguments rather than up to six, and also this
+ * way the file stands alone from possible changes in the
+ * implementation of <linux/syscalls.h>.
+ */
+#define HV_WRAP0(type, name)					\
+	type name(void);					\
+	type name(void)						\
+	{							\
+		return _##name();				\
+	}
+#define __HV_DECL1(t1, a1)	t1 a1
+#define __HV_DECL2(t2, a2, ...) t2 a2, __HV_DECL1(__VA_ARGS__)
+#define __HV_DECL3(t3, a3, ...) t3 a3, __HV_DECL2(__VA_ARGS__)
+#define __HV_DECL4(t4, a4, ...) t4 a4, __HV_DECL3(__VA_ARGS__)
+#define __HV_DECL5(t5, a5, ...) t5 a5, __HV_DECL4(__VA_ARGS__)
+#define __HV_DECL6(t6, a6, ...) t6 a6, __HV_DECL5(__VA_ARGS__)
+#define __HV_DECL7(t7, a7, ...) t7 a7, __HV_DECL6(__VA_ARGS__)
+#define __HV_DECL8(t8, a8, ...) t8 a8, __HV_DECL7(__VA_ARGS__)
+#define __HV_DECL9(t9, a9, ...) t9 a9, __HV_DECL8(__VA_ARGS__)
+#define __HV_PASS1(t1, a1)	a1
+#define __HV_PASS2(t2, a2, ...) a2, __HV_PASS1(__VA_ARGS__)
+#define __HV_PASS3(t3, a3, ...) a3, __HV_PASS2(__VA_ARGS__)
+#define __HV_PASS4(t4, a4, ...) a4, __HV_PASS3(__VA_ARGS__)
+#define __HV_PASS5(t5, a5, ...) a5, __HV_PASS4(__VA_ARGS__)
+#define __HV_PASS6(t6, a6, ...) a6, __HV_PASS5(__VA_ARGS__)
+#define __HV_PASS7(t7, a7, ...) a7, __HV_PASS6(__VA_ARGS__)
+#define __HV_PASS8(t8, a8, ...) a8, __HV_PASS7(__VA_ARGS__)
+#define __HV_PASS9(t9, a9, ...) a9, __HV_PASS8(__VA_ARGS__)
+#define HV_WRAPx(x, type, name, ...)				\
+	type name(__HV_DECL##x(__VA_ARGS__));			\
+	type name(__HV_DECL##x(__VA_ARGS__))			\
+	{							\
+		return _##name(__HV_PASS##x(__VA_ARGS__));	\
+	}
+#define HV_WRAP1(type, name, ...) HV_WRAPx(1, type, name, __VA_ARGS__)
+#define HV_WRAP2(type, name, ...) HV_WRAPx(2, type, name, __VA_ARGS__)
+#define HV_WRAP3(type, name, ...) HV_WRAPx(3, type, name, __VA_ARGS__)
+#define HV_WRAP4(type, name, ...) HV_WRAPx(4, type, name, __VA_ARGS__)
+#define HV_WRAP5(type, name, ...) HV_WRAPx(5, type, name, __VA_ARGS__)
+#define HV_WRAP6(type, name, ...) HV_WRAPx(6, type, name, __VA_ARGS__)
+#define HV_WRAP7(type, name, ...) HV_WRAPx(7, type, name, __VA_ARGS__)
+#define HV_WRAP8(type, name, ...) HV_WRAPx(8, type, name, __VA_ARGS__)
+#define HV_WRAP9(type, name, ...) HV_WRAPx(9, type, name, __VA_ARGS__)
+
+/* List all the hypervisor API functions. */
+HV_WRAP4(void, hv_init, HV_VersionNumber, interface_version_number,
+	 int, chip_num, int, chip_rev_num, int, client_pl)
+HV_WRAP1(long, hv_sysconf, HV_SysconfQuery, query)
+HV_WRAP3(int, hv_confstr, HV_ConfstrQuery, query, HV_VirtAddr, buf, int, len)
+#if CHIP_HAS_IPI()
+HV_WRAP3(int, hv_get_ipi_pte, HV_Coord, tile, int, pl, HV_PTE*, pte)
+HV_WRAP3(int, hv_console_set_ipi, int, ipi, int, event, HV_Coord, coord);
+#else
+HV_WRAP1(void, hv_enable_intr, HV_IntrMask, enab_mask)
+HV_WRAP1(void, hv_disable_intr, HV_IntrMask, disab_mask)
+HV_WRAP1(void, hv_clear_intr, HV_IntrMask, clear_mask)
+HV_WRAP1(void, hv_raise_intr, HV_IntrMask, raise_mask)
+HV_WRAP2(HV_Errno, hv_trigger_ipi, HV_Coord, tile, int, interrupt)
+#endif /* !CHIP_HAS_IPI() */
+HV_WRAP3(int, hv_store_mapping, HV_VirtAddr, va, unsigned int, len,
+	 HV_PhysAddr, pa)
+HV_WRAP2(HV_PhysAddr, hv_inquire_realpa, HV_PhysAddr, cpa, unsigned int, len)
+HV_WRAP0(HV_RTCTime, hv_get_rtc)
+HV_WRAP1(void, hv_set_rtc, HV_RTCTime, time)
+HV_WRAP4(int, hv_install_context, HV_PhysAddr, page_table, HV_PTE, access,
+	 HV_ASID, asid, __hv32, flags)
+HV_WRAP2(int, hv_set_pte_super_shift, int, level, int, log2_count)
+HV_WRAP0(HV_Context, hv_inquire_context)
+HV_WRAP1(int, hv_flush_asid, HV_ASID, asid)
+HV_WRAP2(int, hv_flush_page, HV_VirtAddr, address, HV_PageSize, page_size)
+HV_WRAP3(int, hv_flush_pages, HV_VirtAddr, start, HV_PageSize, page_size,
+	 unsigned long, size)
+HV_WRAP1(int, hv_flush_all, int, preserve_global)
+HV_WRAP2(void, hv_restart, HV_VirtAddr, cmd, HV_VirtAddr, args)
+HV_WRAP0(void, hv_halt)
+HV_WRAP0(void, hv_power_off)
+HV_WRAP1(int, hv_reexec, HV_PhysAddr, entry)
+HV_WRAP0(HV_Topology, hv_inquire_topology)
+HV_WRAP3(HV_Errno, hv_inquire_tiles, HV_InqTileSet, set, HV_VirtAddr, cpumask,
+	 int, length)
+HV_WRAP1(HV_PhysAddrRange, hv_inquire_physical, int, idx)
+HV_WRAP2(HV_MemoryControllerInfo, hv_inquire_memory_controller, HV_Coord, coord,
+	 int, controller)
+HV_WRAP1(HV_VirtAddrRange, hv_inquire_virtual, int, idx)
+HV_WRAP1(HV_ASIDRange, hv_inquire_asid, int, idx)
+HV_WRAP1(void, hv_nanosleep, int, nanosecs)
+HV_WRAP0(int, hv_console_read_if_ready)
+HV_WRAP1(void, hv_console_putc, int, byte)
+HV_WRAP2(int, hv_console_write, HV_VirtAddr, bytes, int, len)
+HV_WRAP0(void, hv_downcall_dispatch)
+HV_WRAP1(int, hv_fs_findfile, HV_VirtAddr, filename)
+HV_WRAP1(HV_FS_StatInfo, hv_fs_fstat, int, inode)
+HV_WRAP4(int, hv_fs_pread, int, inode, HV_VirtAddr, buf,
+	 int, length, int, offset)
+HV_WRAP2(unsigned long long, hv_physaddr_read64, HV_PhysAddr, addr,
+	 HV_PTE, access)
+HV_WRAP3(void, hv_physaddr_write64, HV_PhysAddr, addr, HV_PTE, access,
+	 unsigned long long, val)
+HV_WRAP2(int, hv_get_command_line, HV_VirtAddr, buf, int, length)
+HV_WRAP2(HV_Errno, hv_set_command_line, HV_VirtAddr, buf, int, length)
+HV_WRAP1(void, hv_set_caching, unsigned long, bitmask)
+HV_WRAP2(void, hv_bzero_page, HV_VirtAddr, va, unsigned int, size)
+HV_WRAP1(HV_Errno, hv_register_message_state, HV_MsgState*, msgstate)
+HV_WRAP4(int, hv_send_message, HV_Recipient *, recips, int, nrecip,
+	 HV_VirtAddr, buf, int, buflen)
+HV_WRAP3(HV_RcvMsgInfo, hv_receive_message, HV_MsgState, msgstate,
+	 HV_VirtAddr, buf, int, buflen)
+HV_WRAP0(void, hv_start_all_tiles)
+HV_WRAP2(int, hv_dev_open, HV_VirtAddr, name, __hv32, flags)
+HV_WRAP1(int, hv_dev_close, int, devhdl)
+HV_WRAP5(int, hv_dev_pread, int, devhdl, __hv32, flags, HV_VirtAddr, va,
+	 __hv32, len, __hv64, offset)
+HV_WRAP5(int, hv_dev_pwrite, int, devhdl, __hv32, flags, HV_VirtAddr, va,
+	 __hv32, len, __hv64, offset)
+HV_WRAP3(int, hv_dev_poll, int, devhdl, __hv32, events, HV_IntArg, intarg)
+HV_WRAP1(int, hv_dev_poll_cancel, int, devhdl)
+HV_WRAP6(int, hv_dev_preada, int, devhdl, __hv32, flags, __hv32, sgl_len,
+	 HV_SGL *, sglp, __hv64, offset, HV_IntArg, intarg)
+HV_WRAP6(int, hv_dev_pwritea, int, devhdl, __hv32, flags, __hv32, sgl_len,
+	 HV_SGL *, sglp, __hv64, offset, HV_IntArg, intarg)
+HV_WRAP9(int, hv_flush_remote, HV_PhysAddr, cache_pa,
+	 unsigned long, cache_control, unsigned long*, cache_cpumask,
+	 HV_VirtAddr, tlb_va, unsigned long, tlb_length,
+	 unsigned long, tlb_pgsize, unsigned long*, tlb_cpumask,
+	 HV_Remote_ASID*, asids, int, asidcount)
diff --git a/arch/tile/kernel/intvec_32.S b/arch/tile/kernel/intvec_32.S
new file mode 100644
index 000000000..cdbda45a4
--- /dev/null
+++ b/arch/tile/kernel/intvec_32.S
@@ -0,0 +1,1920 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Linux interrupt vectors.
+ */
+
+#include <linux/linkage.h>
+#include <linux/errno.h>
+#include <linux/init.h>
+#include <linux/unistd.h>
+#include <asm/ptrace.h>
+#include <asm/thread_info.h>
+#include <asm/irqflags.h>
+#include <asm/atomic_32.h>
+#include <asm/asm-offsets.h>
+#include <hv/hypervisor.h>
+#include <arch/abi.h>
+#include <arch/interrupts.h>
+#include <arch/spr_def.h>
+
+#define PTREGS_PTR(reg, ptreg) addli reg, sp, C_ABI_SAVE_AREA_SIZE + (ptreg)
+
+#define PTREGS_OFFSET_SYSCALL PTREGS_OFFSET_REG(TREG_SYSCALL_NR)
+
+	.macro  push_reg reg, ptr=sp, delta=-4
+	{
+	 sw     \ptr, \reg
+	 addli  \ptr, \ptr, \delta
+	}
+	.endm
+
+	.macro  pop_reg reg, ptr=sp, delta=4
+	{
+	 lw     \reg, \ptr
+	 addli  \ptr, \ptr, \delta
+	}
+	.endm
+
+	.macro  pop_reg_zero reg, zreg, ptr=sp, delta=4
+	{
+	 move   \zreg, zero
+	 lw     \reg, \ptr
+	 addi   \ptr, \ptr, \delta
+	}
+	.endm
+
+	.macro  push_extra_callee_saves reg
+	PTREGS_PTR(\reg, PTREGS_OFFSET_REG(51))
+	push_reg r51, \reg
+	push_reg r50, \reg
+	push_reg r49, \reg
+	push_reg r48, \reg
+	push_reg r47, \reg
+	push_reg r46, \reg
+	push_reg r45, \reg
+	push_reg r44, \reg
+	push_reg r43, \reg
+	push_reg r42, \reg
+	push_reg r41, \reg
+	push_reg r40, \reg
+	push_reg r39, \reg
+	push_reg r38, \reg
+	push_reg r37, \reg
+	push_reg r36, \reg
+	push_reg r35, \reg
+	push_reg r34, \reg, PTREGS_OFFSET_BASE - PTREGS_OFFSET_REG(34)
+	.endm
+
+	.macro  panic str
+	.pushsection .rodata, "a"
+1:
+	.asciz  "\str"
+	.popsection
+	{
+	 moveli r0, lo16(1b)
+	}
+	{
+	 auli   r0, r0, ha16(1b)
+	 jal    panic
+	}
+	.endm
+
+#ifdef __COLLECT_LINKER_FEEDBACK__
+	.pushsection .text.intvec_feedback,"ax"
+intvec_feedback:
+	.popsection
+#endif
+
+	/*
+	 * Default interrupt handler.
+	 *
+	 * vecnum is where we'll put this code.
+	 * c_routine is the C routine we'll call.
+	 *
+	 * The C routine is passed two arguments:
+	 * - A pointer to the pt_regs state.
+	 * - The interrupt vector number.
+	 *
+	 * The "processing" argument specifies the code for processing
+	 * the interrupt. Defaults to "handle_interrupt".
+	 */
+	.macro  int_hand vecnum, vecname, c_routine, processing=handle_interrupt
+	.org    (\vecnum << 8)
+intvec_\vecname:
+	.ifc    \vecnum, INT_SWINT_1
+	blz     TREG_SYSCALL_NR_NAME, sys_cmpxchg
+	.endif
+
+	/* Temporarily save a register so we have somewhere to work. */
+
+	mtspr   SPR_SYSTEM_SAVE_K_1, r0
+	mfspr   r0, SPR_EX_CONTEXT_K_1
+
+	/* The cmpxchg code clears sp to force us to reset it here on fault. */
+	{
+	 bz     sp, 2f
+	 andi   r0, r0, SPR_EX_CONTEXT_1_1__PL_MASK  /* mask off ICS */
+	}
+
+	.ifc    \vecnum, INT_DOUBLE_FAULT
+	/*
+	 * For double-faults from user-space, fall through to the normal
+	 * register save and stack setup path.  Otherwise, it's the
+	 * hypervisor giving us one last chance to dump diagnostics, and we
+	 * branch to the kernel_double_fault routine to do so.
+	 */
+	bz      r0, 1f
+	j       _kernel_double_fault
+1:
+	.else
+	/*
+	 * If we're coming from user-space, then set sp to the top of
+	 * the kernel stack.  Otherwise, assume sp is already valid.
+	 */
+	{
+	 bnz    r0, 0f
+	 move   r0, sp
+	}
+	.endif
+
+	.ifc    \c_routine, do_page_fault
+	/*
+	 * The page_fault handler may be downcalled directly by the
+	 * hypervisor even when Linux is running and has ICS set.
+	 *
+	 * In this case the contents of EX_CONTEXT_K_1 reflect the
+	 * previous fault and can't be relied on to choose whether or
+	 * not to reinitialize the stack pointer.  So we add a test
+	 * to see whether SYSTEM_SAVE_K_2 has the high bit set,
+	 * and if so we don't reinitialize sp, since we must be coming
+	 * from Linux.  (In fact the precise case is !(val & ~1),
+	 * but any Linux PC has to have the high bit set.)
+	 *
+	 * Note that the hypervisor *always* sets SYSTEM_SAVE_K_2 for
+	 * any path that turns into a downcall to one of our TLB handlers.
+	 */
+	mfspr   r0, SPR_SYSTEM_SAVE_K_2
+	{
+	 blz    r0, 0f    /* high bit in S_S_1_2 is for a PC to use */
+	 move   r0, sp
+	}
+	.endif
+
+2:
+	/*
+	 * SYSTEM_SAVE_K_0 holds the cpu number in the low bits, and
+	 * the current stack top in the higher bits.  So we recover
+	 * our stack top by just masking off the low bits, then
+	 * point sp at the top aligned address on the actual stack page.
+	 */
+	mfspr   r0, SPR_SYSTEM_SAVE_K_0
+	mm      r0, r0, zero, LOG2_NR_CPU_IDS, 31
+
+0:
+	/*
+	 * Align the stack mod 64 so we can properly predict what
+	 * cache lines we need to write-hint to reduce memory fetch
+	 * latency as we enter the kernel.  The layout of memory is
+	 * as follows, with cache line 0 at the lowest VA, and cache
+	 * line 4 just below the r0 value this "andi" computes.
+	 * Note that we never write to cache line 4, and we skip
+	 * cache line 1 for syscalls.
+	 *
+	 *    cache line 4: ptregs padding (two words)
+	 *    cache line 3: r46...lr, pc, ex1, faultnum, orig_r0, flags, pad
+	 *    cache line 2: r30...r45
+	 *    cache line 1: r14...r29
+	 *    cache line 0: 2 x frame, r0..r13
+	 */
+#if STACK_TOP_DELTA != 64
+#error STACK_TOP_DELTA must be 64 for assumptions here and in task_pt_regs()
+#endif
+	andi    r0, r0, -64
+
+	/*
+	 * Push the first four registers on the stack, so that we can set
+	 * them to vector-unique values before we jump to the common code.
+	 *
+	 * Registers are pushed on the stack as a struct pt_regs,
+	 * with the sp initially just above the struct, and when we're
+	 * done, sp points to the base of the struct, minus
+	 * C_ABI_SAVE_AREA_SIZE, so we can directly jal to C code.
+	 *
+	 * This routine saves just the first four registers, plus the
+	 * stack context so we can do proper backtracing right away,
+	 * and defers to handle_interrupt to save the rest.
+	 * The backtracer needs pc, ex1, lr, sp, r52, and faultnum.
+	 */
+	addli   r0, r0, PTREGS_OFFSET_LR - (PTREGS_SIZE + KSTK_PTREGS_GAP)
+	wh64    r0    /* cache line 3 */
+	{
+	 sw     r0, lr
+	 addli  r0, r0, PTREGS_OFFSET_SP - PTREGS_OFFSET_LR
+	}
+	{
+	 sw     r0, sp
+	 addli  sp, r0, PTREGS_OFFSET_REG(52) - PTREGS_OFFSET_SP
+	}
+	{
+	 sw     sp, r52
+	 addli  sp, sp, PTREGS_OFFSET_REG(1) - PTREGS_OFFSET_REG(52)
+	}
+	wh64    sp    /* cache line 0 */
+	{
+	 sw     sp, r1
+	 addli  sp, sp, PTREGS_OFFSET_REG(2) - PTREGS_OFFSET_REG(1)
+	}
+	{
+	 sw     sp, r2
+	 addli  sp, sp, PTREGS_OFFSET_REG(3) - PTREGS_OFFSET_REG(2)
+	}
+	{
+	 sw     sp, r3
+	 addli  sp, sp, PTREGS_OFFSET_PC - PTREGS_OFFSET_REG(3)
+	}
+	mfspr   r0, SPR_EX_CONTEXT_K_0
+	.ifc \processing,handle_syscall
+	/*
+	 * Bump the saved PC by one bundle so that when we return, we won't
+	 * execute the same swint instruction again.  We need to do this while
+	 * we're in the critical section.
+	 */
+	addi    r0, r0, 8
+	.endif
+	{
+	 sw     sp, r0
+	 addli  sp, sp, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_PC
+	}
+	mfspr   r0, SPR_EX_CONTEXT_K_1
+	{
+	 sw     sp, r0
+	 addi   sp, sp, PTREGS_OFFSET_FAULTNUM - PTREGS_OFFSET_EX1
+	/*
+	 * Use r0 for syscalls so it's a temporary; use r1 for interrupts
+	 * so that it gets passed through unchanged to the handler routine.
+	 * Note that the .if conditional confusingly spans bundles.
+	 */
+	 .ifc \processing,handle_syscall
+	 movei  r0, \vecnum
+	}
+	{
+	 sw     sp, r0
+	 .else
+	 movei  r1, \vecnum
+	}
+	{
+	 sw     sp, r1
+	 .endif
+	 addli  sp, sp, PTREGS_OFFSET_REG(0) - PTREGS_OFFSET_FAULTNUM
+	}
+	mfspr   r0, SPR_SYSTEM_SAVE_K_1    /* Original r0 */
+	{
+	 sw     sp, r0
+	 addi   sp, sp, -PTREGS_OFFSET_REG(0) - 4
+	}
+	{
+	 sw     sp, zero        /* write zero into "Next SP" frame pointer */
+	 addi   sp, sp, -4      /* leave SP pointing at bottom of frame */
+	}
+	.ifc \processing,handle_syscall
+	j       handle_syscall
+	.else
+	/*
+	 * Capture per-interrupt SPR context to registers.
+	 * We overload the meaning of r3 on this path such that if its bit 31
+	 * is set, we have to mask all interrupts including NMIs before
+	 * clearing the interrupt critical section bit.
+	 * See discussion below at "finish_interrupt_save".
+	 */
+	.ifc \c_routine, do_page_fault
+	mfspr   r2, SPR_SYSTEM_SAVE_K_3   /* address of page fault */
+	mfspr   r3, SPR_SYSTEM_SAVE_K_2   /* info about page fault */
+	.else
+	.ifc \vecnum, INT_DOUBLE_FAULT
+	{
+	 mfspr  r2, SPR_SYSTEM_SAVE_K_2   /* double fault info from HV */
+	 movei  r3, 0
+	}
+	.else
+	.ifc \c_routine, do_trap
+	{
+	 mfspr  r2, GPV_REASON
+	 movei  r3, 0
+	}
+	.else
+	.ifc \c_routine, handle_perf_interrupt
+	{
+	 mfspr  r2, PERF_COUNT_STS
+	 movei  r3, -1   /* not used, but set for consistency */
+	}
+	.else
+	.ifc \c_routine, handle_perf_interrupt
+	{
+	 mfspr  r2, AUX_PERF_COUNT_STS
+	 movei  r3, -1   /* not used, but set for consistency */
+	}
+	.else
+	movei   r3, 0
+	.endif
+	.endif
+	.endif
+	.endif
+	.endif
+	/* Put function pointer in r0 */
+	moveli  r0, lo16(\c_routine)
+	{
+	 auli   r0, r0, ha16(\c_routine)
+	 j       \processing
+	}
+	.endif
+	ENDPROC(intvec_\vecname)
+
+#ifdef __COLLECT_LINKER_FEEDBACK__
+	.pushsection .text.intvec_feedback,"ax"
+	.org    (\vecnum << 5)
+	FEEDBACK_ENTER_EXPLICIT(intvec_\vecname, .intrpt, 1 << 8)
+	jrp     lr
+	.popsection
+#endif
+
+	.endm
+
+
+	/*
+	 * Save the rest of the registers that we didn't save in the actual
+	 * vector itself.  We can't use r0-r10 inclusive here.
+	 */
+	.macro  finish_interrupt_save, function
+
+	/* If it's a syscall, save a proper orig_r0, otherwise just zero. */
+	PTREGS_PTR(r52, PTREGS_OFFSET_ORIG_R0)
+	{
+	 .ifc \function,handle_syscall
+	 sw     r52, r0
+	 .else
+	 sw     r52, zero
+	 .endif
+	 PTREGS_PTR(r52, PTREGS_OFFSET_TP)
+	}
+
+	/*
+	 * For ordinary syscalls, we save neither caller- nor callee-
+	 * save registers, since the syscall invoker doesn't expect the
+	 * caller-saves to be saved, and the called kernel functions will
+	 * take care of saving the callee-saves for us.
+	 *
+	 * For interrupts we save just the caller-save registers.  Saving
+	 * them is required (since the "caller" can't save them).  Again,
+	 * the called kernel functions will restore the callee-save
+	 * registers for us appropriately.
+	 *
+	 * On return, we normally restore nothing special for syscalls,
+	 * and just the caller-save registers for interrupts.
+	 *
+	 * However, there are some important caveats to all this:
+	 *
+	 * - We always save a few callee-save registers to give us
+	 *   some scratchpad registers to carry across function calls.
+	 *
+	 * - fork/vfork/etc require us to save all the callee-save
+	 *   registers, which we do in PTREGS_SYSCALL_ALL_REGS, below.
+	 *
+	 * - We always save r0..r5 and r10 for syscalls, since we need
+	 *   to reload them a bit later for the actual kernel call, and
+	 *   since we might need them for -ERESTARTNOINTR, etc.
+	 *
+	 * - Before invoking a signal handler, we save the unsaved
+	 *   callee-save registers so they are visible to the
+	 *   signal handler or any ptracer.
+	 *
+	 * - If the unsaved callee-save registers are modified, we set
+	 *   a bit in pt_regs so we know to reload them from pt_regs
+	 *   and not just rely on the kernel function unwinding.
+	 *   (Done for ptrace register writes and SA_SIGINFO handler.)
+	 */
+	{
+	 sw     r52, tp
+	 PTREGS_PTR(r52, PTREGS_OFFSET_REG(33))
+	}
+	wh64    r52    /* cache line 2 */
+	push_reg r33, r52
+	push_reg r32, r52
+	push_reg r31, r52
+	.ifc \function,handle_syscall
+	push_reg r30, r52, PTREGS_OFFSET_SYSCALL - PTREGS_OFFSET_REG(30)
+	push_reg TREG_SYSCALL_NR_NAME, r52, \
+	  PTREGS_OFFSET_REG(5) - PTREGS_OFFSET_SYSCALL
+	.else
+
+	push_reg r30, r52, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(30)
+	wh64    r52    /* cache line 1 */
+	push_reg r29, r52
+	push_reg r28, r52
+	push_reg r27, r52
+	push_reg r26, r52
+	push_reg r25, r52
+	push_reg r24, r52
+	push_reg r23, r52
+	push_reg r22, r52
+	push_reg r21, r52
+	push_reg r20, r52
+	push_reg r19, r52
+	push_reg r18, r52
+	push_reg r17, r52
+	push_reg r16, r52
+	push_reg r15, r52
+	push_reg r14, r52
+	push_reg r13, r52
+	push_reg r12, r52
+	push_reg r11, r52
+	push_reg r10, r52
+	push_reg r9, r52
+	push_reg r8, r52
+	push_reg r7, r52
+	push_reg r6, r52
+
+	.endif
+
+	push_reg r5, r52
+	sw      r52, r4
+
+	/* Load tp with our per-cpu offset. */
+#ifdef CONFIG_SMP
+	{
+	 mfspr  r20, SPR_SYSTEM_SAVE_K_0
+	 moveli r21, lo16(__per_cpu_offset)
+	}
+	{
+	 auli   r21, r21, ha16(__per_cpu_offset)
+	 mm     r20, r20, zero, 0, LOG2_NR_CPU_IDS-1
+	}
+	s2a     r20, r20, r21
+	lw      tp, r20
+#else
+	move    tp, zero
+#endif
+
+	/*
+	 * If we will be returning to the kernel, we will need to
+	 * reset the interrupt masks to the state they had before.
+	 * Set DISABLE_IRQ in flags iff we came from PL1 with irqs disabled.
+	 * We load flags in r32 here so we can jump to .Lrestore_regs
+	 * directly after do_page_fault_ics() if necessary.
+	 */
+	mfspr   r32, SPR_EX_CONTEXT_K_1
+	{
+	 andi   r32, r32, SPR_EX_CONTEXT_1_1__PL_MASK  /* mask off ICS */
+	 PTREGS_PTR(r21, PTREGS_OFFSET_FLAGS)
+	}
+	bzt     r32, 1f       /* zero if from user space */
+	IRQS_DISABLED(r32)    /* zero if irqs enabled */
+#if PT_FLAGS_DISABLE_IRQ != 1
+# error Value of IRQS_DISABLED used to set PT_FLAGS_DISABLE_IRQ; fix
+#endif
+1:
+	.ifnc \function,handle_syscall
+	/* Record the fact that we saved the caller-save registers above. */
+	ori     r32, r32, PT_FLAGS_CALLER_SAVES
+	.endif
+	sw      r21, r32
+
+#ifdef __COLLECT_LINKER_FEEDBACK__
+	/*
+	 * Notify the feedback routines that we were in the
+	 * appropriate fixed interrupt vector area.  Note that we
+	 * still have ICS set at this point, so we can't invoke any
+	 * atomic operations or we will panic.  The feedback
+	 * routines internally preserve r0..r10 and r30 up.
+	 */
+	.ifnc \function,handle_syscall
+	shli    r20, r1, 5
+	.else
+	moveli  r20, INT_SWINT_1 << 5
+	.endif
+	addli   r20, r20, lo16(intvec_feedback)
+	auli    r20, r20, ha16(intvec_feedback)
+	jalr    r20
+
+	/* And now notify the feedback routines that we are here. */
+	FEEDBACK_ENTER(\function)
+#endif
+
+	/*
+	 * we've captured enough state to the stack (including in
+	 * particular our EX_CONTEXT state) that we can now release
+	 * the interrupt critical section and replace it with our
+	 * standard "interrupts disabled" mask value.  This allows
+	 * synchronous interrupts (and profile interrupts) to punch
+	 * through from this point onwards.
+	 *
+	 * If bit 31 of r3 is set during a non-NMI interrupt, we know we
+	 * are on the path where the hypervisor has punched through our
+	 * ICS with a page fault, so we call out to do_page_fault_ics()
+	 * to figure out what to do with it.  If the fault was in
+	 * an atomic op, we unlock the atomic lock, adjust the
+	 * saved register state a little, and return "zero" in r4,
+	 * falling through into the normal page-fault interrupt code.
+	 * If the fault was in a kernel-space atomic operation, then
+	 * do_page_fault_ics() resolves it itself, returns "one" in r4,
+	 * and as a result goes directly to restoring registers and iret,
+	 * without trying to adjust the interrupt masks at all.
+	 * The do_page_fault_ics() API involves passing and returning
+	 * a five-word struct (in registers) to avoid writing the
+	 * save and restore code here.
+	 */
+	.ifc \function,handle_nmi
+	IRQ_DISABLE_ALL(r20)
+	.else
+	.ifnc \function,handle_syscall
+	bgezt   r3, 1f
+	{
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	 jal    do_page_fault_ics
+	}
+	FEEDBACK_REENTER(\function)
+	bzt     r4, 1f
+	j       .Lrestore_regs
+1:
+	.endif
+	IRQ_DISABLE(r20, r21)
+	.endif
+	mtspr   INTERRUPT_CRITICAL_SECTION, zero
+
+	/*
+	 * Prepare the first 256 stack bytes to be rapidly accessible
+	 * without having to fetch the background data.  We don't really
+	 * know how far to write-hint, but kernel stacks generally
+	 * aren't that big, and write-hinting here does take some time.
+	 */
+	addi    r52, sp, -64
+	{
+	 wh64   r52
+	 addi   r52, r52, -64
+	}
+	{
+	 wh64   r52
+	 addi   r52, r52, -64
+	}
+	{
+	 wh64   r52
+	 addi   r52, r52, -64
+	}
+	wh64    r52
+
+#ifdef CONFIG_TRACE_IRQFLAGS
+	.ifnc \function,handle_nmi
+	/*
+	 * We finally have enough state set up to notify the irq
+	 * tracing code that irqs were disabled on entry to the handler.
+	 * The TRACE_IRQS_OFF call clobbers registers r0-r29.
+	 * For syscalls, we already have the register state saved away
+	 * on the stack, so we don't bother to do any register saves here,
+	 * and later we pop the registers back off the kernel stack.
+	 * For interrupt handlers, save r0-r3 in callee-saved registers.
+	 */
+	.ifnc \function,handle_syscall
+	{ move r30, r0; move r31, r1 }
+	{ move r32, r2; move r33, r3 }
+	.endif
+	TRACE_IRQS_OFF
+	.ifnc \function,handle_syscall
+	{ move r0, r30; move r1, r31 }
+	{ move r2, r32; move r3, r33 }
+	.endif
+	.endif
+#endif
+
+	.endm
+
+	.macro  check_single_stepping, kind, not_single_stepping
+	/*
+	 * Check for single stepping in user-level priv
+	 *   kind can be "normal", "ill", or "syscall"
+	 * At end, if fall-thru
+	 *   r29: thread_info->step_state
+	 *   r28: &pt_regs->pc
+	 *   r27: pt_regs->pc
+	 *   r26: thread_info->step_state->buffer
+	 */
+
+	/* Check for single stepping */
+	GET_THREAD_INFO(r29)
+	{
+	 /* Get pointer to field holding step state */
+	 addi   r29, r29, THREAD_INFO_STEP_STATE_OFFSET
+
+	 /* Get pointer to EX1 in register state */
+	 PTREGS_PTR(r27, PTREGS_OFFSET_EX1)
+	}
+	{
+	 /* Get pointer to field holding PC */
+	 PTREGS_PTR(r28, PTREGS_OFFSET_PC)
+
+	 /* Load the pointer to the step state */
+	 lw     r29, r29
+	}
+	/* Load EX1 */
+	lw      r27, r27
+	{
+	 /* Points to flags */
+	 addi   r23, r29, SINGLESTEP_STATE_FLAGS_OFFSET
+
+	 /* No single stepping if there is no step state structure */
+	 bzt    r29, \not_single_stepping
+	}
+	{
+	 /* mask off ICS and any other high bits */
+	 andi   r27, r27, SPR_EX_CONTEXT_1_1__PL_MASK
+
+	 /* Load pointer to single step instruction buffer */
+	 lw     r26, r29
+	}
+	/* Check priv state */
+	bnz     r27, \not_single_stepping
+
+	/* Get flags */
+	lw      r22, r23
+	{
+	 /* Branch if single-step mode not enabled */
+	 bbnst  r22, \not_single_stepping
+
+	 /* Clear enabled flag */
+	 andi   r22, r22, ~SINGLESTEP_STATE_MASK_IS_ENABLED
+	}
+	.ifc \kind,normal
+	{
+	 /* Load PC */
+	 lw     r27, r28
+
+	 /* Point to the entry containing the original PC */
+	 addi   r24, r29, SINGLESTEP_STATE_ORIG_PC_OFFSET
+	}
+	{
+	 /* Disable single stepping flag */
+	 sw     r23, r22
+	}
+	{
+	 /* Get the original pc */
+	 lw     r24, r24
+
+	 /* See if the PC is at the start of the single step buffer */
+	 seq    r25, r26, r27
+	}
+	/*
+	 * NOTE: it is really expected that the PC be in the single step buffer
+	 *       at this point
+	 */
+	bzt     r25, \not_single_stepping
+
+	/* Restore the original PC */
+	sw      r28, r24
+	.else
+	.ifc \kind,syscall
+	{
+	 /* Load PC */
+	 lw     r27, r28
+
+	 /* Point to the entry containing the next PC */
+	 addi   r24, r29, SINGLESTEP_STATE_NEXT_PC_OFFSET
+	}
+	{
+	 /* Increment the stopped PC by the bundle size */
+	 addi   r26, r26, 8
+
+	 /* Disable single stepping flag */
+	 sw     r23, r22
+	}
+	{
+	 /* Get the next pc */
+	 lw     r24, r24
+
+	 /*
+	  * See if the PC is one bundle past the start of the
+	  * single step buffer
+	  */
+	 seq    r25, r26, r27
+	}
+	{
+	 /*
+	  * NOTE: it is really expected that the PC be in the
+	  * single step buffer at this point
+	  */
+	 bzt    r25, \not_single_stepping
+	}
+	/* Set to the next PC */
+	sw      r28, r24
+	.else
+	{
+	 /* Point to 3rd bundle in buffer */
+	 addi   r25, r26, 16
+
+	 /* Load PC */
+	 lw      r27, r28
+	}
+	{
+	 /* Disable single stepping flag */
+	 sw      r23, r22
+
+	 /* See if the PC is in the single step buffer */
+	 slte_u  r24, r26, r27
+	}
+	{
+	 slte_u r25, r27, r25
+
+	 /*
+	  * NOTE: it is really expected that the PC be in the
+	  * single step buffer at this point
+	  */
+	 bzt    r24, \not_single_stepping
+	}
+	bzt     r25, \not_single_stepping
+	.endif
+	.endif
+	.endm
+
+	/*
+	 * Redispatch a downcall.
+	 */
+	.macro  dc_dispatch vecnum, vecname
+	.org    (\vecnum << 8)
+intvec_\vecname:
+	j       _hv_downcall_dispatch
+	ENDPROC(intvec_\vecname)
+	.endm
+
+	/*
+	 * Common code for most interrupts.  The C function we're eventually
+	 * going to is in r0, and the faultnum is in r1; the original
+	 * values for those registers are on the stack.
+	 */
+	.pushsection .text.handle_interrupt,"ax"
+handle_interrupt:
+	finish_interrupt_save handle_interrupt
+
+	/*
+	 * Check for if we are single stepping in user level. If so, then
+	 * we need to restore the PC.
+	 */
+
+	check_single_stepping normal, .Ldispatch_interrupt
+.Ldispatch_interrupt:
+
+	/* Jump to the C routine; it should enable irqs as soon as possible. */
+	{
+	 jalr   r0
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	}
+	FEEDBACK_REENTER(handle_interrupt)
+	{
+	 movei  r30, 0   /* not an NMI */
+	 j      interrupt_return
+	}
+	STD_ENDPROC(handle_interrupt)
+
+/*
+ * This routine takes a boolean in r30 indicating if this is an NMI.
+ * If so, we also expect a boolean in r31 indicating whether to
+ * re-enable the oprofile interrupts.
+ *
+ * Note that .Lresume_userspace is jumped to directly in several
+ * places, and we need to make sure r30 is set correctly in those
+ * callers as well.
+ */
+STD_ENTRY(interrupt_return)
+	/* If we're resuming to kernel space, don't check thread flags. */
+	{
+	 bnz    r30, .Lrestore_all  /* NMIs don't special-case user-space */
+	 PTREGS_PTR(r29, PTREGS_OFFSET_EX1)
+	}
+	lw      r29, r29
+	andi    r29, r29, SPR_EX_CONTEXT_1_1__PL_MASK  /* mask off ICS */
+	bzt     r29, .Lresume_userspace
+
+#ifdef CONFIG_PREEMPT
+	/* Returning to kernel space. Check if we need preemption. */
+	GET_THREAD_INFO(r29)
+	addli   r28, r29, THREAD_INFO_FLAGS_OFFSET
+	{
+	 lw     r28, r28
+	 addli  r29, r29, THREAD_INFO_PREEMPT_COUNT_OFFSET
+	}
+	{
+	 andi   r28, r28, _TIF_NEED_RESCHED
+	 lw     r29, r29
+	}
+	bzt     r28, 1f
+	bnz     r29, 1f
+	/* Disable interrupts explicitly for preemption. */
+	IRQ_DISABLE(r20,r21)
+	TRACE_IRQS_OFF
+	jal     preempt_schedule_irq
+	FEEDBACK_REENTER(interrupt_return)
+1:
+#endif
+
+	/* If we're resuming to _cpu_idle_nap, bump PC forward by 8. */
+	{
+	 PTREGS_PTR(r29, PTREGS_OFFSET_PC)
+	 moveli r27, lo16(_cpu_idle_nap)
+	}
+	{
+	 lw     r28, r29
+	 auli   r27, r27, ha16(_cpu_idle_nap)
+	}
+	{
+	 seq    r27, r27, r28
+	}
+	{
+	 bbns   r27, .Lrestore_all
+	 addi   r28, r28, 8
+	}
+	sw      r29, r28
+	j       .Lrestore_all
+
+.Lresume_userspace:
+	FEEDBACK_REENTER(interrupt_return)
+
+	/*
+	 * Use r33 to hold whether we have already loaded the callee-saves
+	 * into ptregs.  We don't want to do it twice in this loop, since
+	 * then we'd clobber whatever changes are made by ptrace, etc.
+	 * Get base of stack in r32.
+	 */
+	{
+	 GET_THREAD_INFO(r32)
+	 movei  r33, 0
+	}
+
+.Lretry_work_pending:
+	/*
+	 * Disable interrupts so as to make sure we don't
+	 * miss an interrupt that sets any of the thread flags (like
+	 * need_resched or sigpending) between sampling and the iret.
+	 * Routines like schedule() or do_signal() may re-enable
+	 * interrupts before returning.
+	 */
+	IRQ_DISABLE(r20, r21)
+	TRACE_IRQS_OFF  /* Note: clobbers registers r0-r29 */
+
+
+	/* Check to see if there is any work to do before returning to user. */
+	{
+	 addi   r29, r32, THREAD_INFO_FLAGS_OFFSET
+	 moveli r1, lo16(_TIF_ALLWORK_MASK)
+	}
+	{
+	 lw     r29, r29
+	 auli   r1, r1, ha16(_TIF_ALLWORK_MASK)
+	}
+	and     r1, r29, r1
+	bzt     r1, .Lrestore_all
+
+	/*
+	 * Make sure we have all the registers saved for signal
+	 * handling, notify-resume, or single-step.  Call out to C
+	 * code to figure out exactly what we need to do for each flag bit,
+	 * then if necessary, reload the flags and recheck.
+	 */
+	{
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	 bnz    r33, 1f
+	}
+	push_extra_callee_saves r0
+	movei   r33, 1
+1:	jal     do_work_pending
+	bnz     r0, .Lretry_work_pending
+
+	/*
+	 * In the NMI case we
+	 * omit the call to single_process_check_nohz, which normally checks
+	 * to see if we should start or stop the scheduler tick, because
+	 * we can't call arbitrary Linux code from an NMI context.
+	 * We always call the homecache TLB deferral code to re-trigger
+	 * the deferral mechanism.
+	 *
+	 * The other chunk of responsibility this code has is to reset the
+	 * interrupt masks appropriately to reset irqs and NMIs.  We have
+	 * to call TRACE_IRQS_OFF and TRACE_IRQS_ON to support all the
+	 * lockdep-type stuff, but we can't set ICS until afterwards, since
+	 * ICS can only be used in very tight chunks of code to avoid
+	 * tripping over various assertions that it is off.
+	 *
+	 * (There is what looks like a window of vulnerability here since
+	 * we might take a profile interrupt between the two SPR writes
+	 * that set the mask, but since we write the low SPR word first,
+	 * and our interrupt entry code checks the low SPR word, any
+	 * profile interrupt will actually disable interrupts in both SPRs
+	 * before returning, which is OK.)
+	 */
+.Lrestore_all:
+	PTREGS_PTR(r0, PTREGS_OFFSET_EX1)
+	{
+	 lw     r0, r0
+	 PTREGS_PTR(r32, PTREGS_OFFSET_FLAGS)
+	}
+	{
+	 andi   r0, r0, SPR_EX_CONTEXT_1_1__PL_MASK
+	 lw     r32, r32
+	}
+	bnz    r0, 1f
+	j       2f
+#if PT_FLAGS_DISABLE_IRQ != 1
+# error Assuming PT_FLAGS_DISABLE_IRQ == 1 so we can use bbnst below
+#endif
+1:	bbnst   r32, 2f
+	IRQ_DISABLE(r20,r21)
+	TRACE_IRQS_OFF
+	movei   r0, 1
+	mtspr   INTERRUPT_CRITICAL_SECTION, r0
+	bzt     r30, .Lrestore_regs
+	j       3f
+2:	TRACE_IRQS_ON
+	movei   r0, 1
+	mtspr   INTERRUPT_CRITICAL_SECTION, r0
+	IRQ_ENABLE(r20, r21)
+	bzt     r30, .Lrestore_regs
+3:
+
+	/* We are relying on INT_PERF_COUNT at 33, and AUX_PERF_COUNT at 48 */
+	{
+	 moveli r0, lo16(1 << (INT_PERF_COUNT - 32))
+	 bz     r31, .Lrestore_regs
+	}
+	auli    r0, r0, ha16(1 << (INT_AUX_PERF_COUNT - 32))
+	mtspr   SPR_INTERRUPT_MASK_RESET_K_1, r0
+
+	/*
+	 * We now commit to returning from this interrupt, since we will be
+	 * doing things like setting EX_CONTEXT SPRs and unwinding the stack
+	 * frame.  No calls should be made to any other code after this point.
+	 * This code should only be entered with ICS set.
+	 * r32 must still be set to ptregs.flags.
+	 * We launch loads to each cache line separately first, so we can
+	 * get some parallelism out of the memory subsystem.
+	 * We start zeroing caller-saved registers throughout, since
+	 * that will save some cycles if this turns out to be a syscall.
+	 */
+.Lrestore_regs:
+	FEEDBACK_REENTER(interrupt_return)   /* called from elsewhere */
+
+	/*
+	 * Rotate so we have one high bit and one low bit to test.
+	 * - low bit says whether to restore all the callee-saved registers,
+	 *   or just r30-r33, and r52 up.
+	 * - high bit (i.e. sign bit) says whether to restore all the
+	 *   caller-saved registers, or just r0.
+	 */
+#if PT_FLAGS_CALLER_SAVES != 2 || PT_FLAGS_RESTORE_REGS != 4
+# error Rotate trick does not work :-)
+#endif
+	{
+	 rli    r20, r32, 30
+	 PTREGS_PTR(sp, PTREGS_OFFSET_REG(0))
+	}
+
+	/*
+	 * Load cache lines 0, 2, and 3 in that order, then use
+	 * the last loaded value, which makes it likely that the other
+	 * cache lines have also loaded, at which point we should be
+	 * able to safely read all the remaining words on those cache
+	 * lines without waiting for the memory subsystem.
+	 */
+	pop_reg_zero r0, r28, sp, PTREGS_OFFSET_REG(30) - PTREGS_OFFSET_REG(0)
+	pop_reg_zero r30, r2, sp, PTREGS_OFFSET_PC - PTREGS_OFFSET_REG(30)
+	pop_reg_zero r21, r3, sp, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_PC
+	pop_reg_zero lr, r4, sp, PTREGS_OFFSET_REG(52) - PTREGS_OFFSET_EX1
+	{
+	 mtspr  SPR_EX_CONTEXT_K_0, r21
+	 move   r5, zero
+	}
+	{
+	 mtspr  SPR_EX_CONTEXT_K_1, lr
+	 andi   lr, lr, SPR_EX_CONTEXT_1_1__PL_MASK  /* mask off ICS */
+	}
+
+	/* Restore callee-saveds that we actually use. */
+	pop_reg_zero r52, r6, sp, PTREGS_OFFSET_REG(31) - PTREGS_OFFSET_REG(52)
+	pop_reg_zero r31, r7
+	pop_reg_zero r32, r8
+	pop_reg_zero r33, r9, sp, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(33)
+
+	/*
+	 * If we modified other callee-saveds, restore them now.
+	 * This is rare, but could be via ptrace or signal handler.
+	 */
+	{
+	 move   r10, zero
+	 bbs    r20, .Lrestore_callees
+	}
+.Lcontinue_restore_regs:
+
+	/* Check if we're returning from a syscall. */
+	{
+	 move   r11, zero
+	 blzt   r20, 1f  /* no, so go restore callee-save registers */
+	}
+
+	/*
+	 * Check if we're returning to userspace.
+	 * Note that if we're not, we don't worry about zeroing everything.
+	 */
+	{
+	 addli  sp, sp, PTREGS_OFFSET_LR - PTREGS_OFFSET_REG(29)
+	 bnz    lr, .Lkernel_return
+	}
+
+	/*
+	 * On return from syscall, we've restored r0 from pt_regs, but we
+	 * clear the remainder of the caller-saved registers.  We could
+	 * restore the syscall arguments, but there's not much point,
+	 * and it ensures user programs aren't trying to use the
+	 * caller-saves if we clear them, as well as avoiding leaking
+	 * kernel pointers into userspace.
+	 */
+	pop_reg_zero lr, r12, sp, PTREGS_OFFSET_TP - PTREGS_OFFSET_LR
+	pop_reg_zero tp, r13, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_TP
+	{
+	 lw     sp, sp
+	 move   r14, zero
+	 move   r15, zero
+	}
+	{ move r16, zero; move r17, zero }
+	{ move r18, zero; move r19, zero }
+	{ move r20, zero; move r21, zero }
+	{ move r22, zero; move r23, zero }
+	{ move r24, zero; move r25, zero }
+	{ move r26, zero; move r27, zero }
+
+	/* Set r1 to errno if we are returning an error, otherwise zero. */
+	{
+	 moveli r29, 4096
+	 sub    r1, zero, r0
+	}
+	slt_u   r29, r1, r29
+	{
+	 mnz    r1, r29, r1
+	 move   r29, zero
+	}
+	iret
+
+	/*
+	 * Not a syscall, so restore caller-saved registers.
+	 * First kick off a load for cache line 1, which we're touching
+	 * for the first time here.
+	 */
+	.align 64
+1:	pop_reg r29, sp, PTREGS_OFFSET_REG(1) - PTREGS_OFFSET_REG(29)
+	pop_reg r1
+	pop_reg r2
+	pop_reg r3
+	pop_reg r4
+	pop_reg r5
+	pop_reg r6
+	pop_reg r7
+	pop_reg r8
+	pop_reg r9
+	pop_reg r10
+	pop_reg r11
+	pop_reg r12
+	pop_reg r13
+	pop_reg r14
+	pop_reg r15
+	pop_reg r16
+	pop_reg r17
+	pop_reg r18
+	pop_reg r19
+	pop_reg r20
+	pop_reg r21
+	pop_reg r22
+	pop_reg r23
+	pop_reg r24
+	pop_reg r25
+	pop_reg r26
+	pop_reg r27
+	pop_reg r28, sp, PTREGS_OFFSET_LR - PTREGS_OFFSET_REG(28)
+	/* r29 already restored above */
+	bnz     lr, .Lkernel_return
+	pop_reg lr, sp, PTREGS_OFFSET_TP - PTREGS_OFFSET_LR
+	pop_reg tp, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_TP
+	lw      sp, sp
+	iret
+
+	/*
+	 * We can't restore tp when in kernel mode, since a thread might
+	 * have migrated from another cpu and brought a stale tp value.
+	 */
+.Lkernel_return:
+	pop_reg lr, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_LR
+	lw      sp, sp
+	iret
+
+	/* Restore callee-saved registers from r34 to r51. */
+.Lrestore_callees:
+	addli  sp, sp, PTREGS_OFFSET_REG(34) - PTREGS_OFFSET_REG(29)
+	pop_reg r34
+	pop_reg r35
+	pop_reg r36
+	pop_reg r37
+	pop_reg r38
+	pop_reg r39
+	pop_reg r40
+	pop_reg r41
+	pop_reg r42
+	pop_reg r43
+	pop_reg r44
+	pop_reg r45
+	pop_reg r46
+	pop_reg r47
+	pop_reg r48
+	pop_reg r49
+	pop_reg r50
+	pop_reg r51, sp, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(51)
+	j .Lcontinue_restore_regs
+	STD_ENDPROC(interrupt_return)
+
+	/*
+	 * Some interrupts don't check for single stepping
+	 */
+	.pushsection .text.handle_interrupt_no_single_step,"ax"
+handle_interrupt_no_single_step:
+	finish_interrupt_save handle_interrupt_no_single_step
+	{
+	 jalr   r0
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	}
+	FEEDBACK_REENTER(handle_interrupt_no_single_step)
+	{
+	 movei  r30, 0   /* not an NMI */
+	 j      interrupt_return
+	}
+	STD_ENDPROC(handle_interrupt_no_single_step)
+
+	/*
+	 * "NMI" interrupts mask ALL interrupts before calling the
+	 * handler, and don't check thread flags, etc., on the way
+	 * back out.  In general, the only things we do here for NMIs
+	 * are the register save/restore, fixing the PC if we were
+	 * doing single step, and the dataplane kernel-TLB management.
+	 * We don't (for example) deal with start/stop of the sched tick.
+	 */
+	.pushsection .text.handle_nmi,"ax"
+handle_nmi:
+	finish_interrupt_save handle_nmi
+	check_single_stepping normal, .Ldispatch_nmi
+.Ldispatch_nmi:
+	{
+	 jalr   r0
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	}
+	FEEDBACK_REENTER(handle_nmi)
+	{
+	 movei  r30, 1
+	 seq    r31, r0, zero
+	}
+	j       interrupt_return
+	STD_ENDPROC(handle_nmi)
+
+	/*
+	 * Parallel code for syscalls to handle_interrupt.
+	 */
+	.pushsection .text.handle_syscall,"ax"
+handle_syscall:
+	finish_interrupt_save handle_syscall
+
+	/*
+	 * Check for if we are single stepping in user level. If so, then
+	 * we need to restore the PC.
+	 */
+	check_single_stepping syscall, .Ldispatch_syscall
+.Ldispatch_syscall:
+
+	/* Enable irqs. */
+	TRACE_IRQS_ON
+	IRQ_ENABLE(r20, r21)
+
+	/* Bump the counter for syscalls made on this tile. */
+	moveli  r20, lo16(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
+	auli    r20, r20, ha16(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
+	add     r20, r20, tp
+	lw      r21, r20
+	addi    r21, r21, 1
+	{
+	 sw     r20, r21
+	 GET_THREAD_INFO(r31)
+	}
+
+	/* Trace syscalls, if requested. */
+	addi	r31, r31, THREAD_INFO_FLAGS_OFFSET
+	lw	r30, r31
+	andi    r30, r30, _TIF_SYSCALL_TRACE
+	bzt	r30, .Lrestore_syscall_regs
+	{
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	 jal    do_syscall_trace_enter
+	}
+	FEEDBACK_REENTER(handle_syscall)
+
+	/*
+	 * We always reload our registers from the stack at this
+	 * point.  They might be valid, if we didn't build with
+	 * TRACE_IRQFLAGS, and this isn't a dataplane tile, and we're not
+	 * doing syscall tracing, but there are enough cases now that it
+	 * seems simplest just to do the reload unconditionally.
+	 */
+.Lrestore_syscall_regs:
+	PTREGS_PTR(r11, PTREGS_OFFSET_REG(0))
+	pop_reg r0, r11
+	pop_reg r1, r11
+	pop_reg r2, r11
+	pop_reg r3, r11
+	pop_reg r4, r11
+	pop_reg r5, r11, PTREGS_OFFSET_SYSCALL - PTREGS_OFFSET_REG(5)
+	pop_reg TREG_SYSCALL_NR_NAME, r11
+
+	/* Ensure that the syscall number is within the legal range. */
+	moveli  r21, __NR_syscalls
+	{
+	 slt_u  r21, TREG_SYSCALL_NR_NAME, r21
+	 moveli r20, lo16(sys_call_table)
+	}
+	{
+	 bbns   r21, .Linvalid_syscall
+	 auli   r20, r20, ha16(sys_call_table)
+	}
+	s2a     r20, TREG_SYSCALL_NR_NAME, r20
+	lw      r20, r20
+
+	/* Jump to syscall handler. */
+	jalr    r20
+.Lhandle_syscall_link: /* value of "lr" after "jalr r20" above */
+
+	/*
+	 * Write our r0 onto the stack so it gets restored instead
+	 * of whatever the user had there before.
+	 */
+	PTREGS_PTR(r29, PTREGS_OFFSET_REG(0))
+	sw      r29, r0
+
+.Lsyscall_sigreturn_skip:
+	FEEDBACK_REENTER(handle_syscall)
+
+	/* Do syscall trace again, if requested. */
+	lw	r30, r31
+	andi    r30, r30, _TIF_SYSCALL_TRACE
+	bzt     r30, 1f
+	{
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	 jal    do_syscall_trace_exit
+	}
+	FEEDBACK_REENTER(handle_syscall)
+1:	{
+	 movei  r30, 0               /* not an NMI */
+	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
+	}
+
+.Linvalid_syscall:
+	/* Report an invalid syscall back to the user program */
+	{
+	 PTREGS_PTR(r29, PTREGS_OFFSET_REG(0))
+	 movei  r28, -ENOSYS
+	}
+	sw      r29, r28
+	{
+	 movei  r30, 0               /* not an NMI */
+	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
+	}
+	STD_ENDPROC(handle_syscall)
+
+	/* Return the address for oprofile to suppress in backtraces. */
+STD_ENTRY_SECTION(handle_syscall_link_address, .text.handle_syscall)
+	lnk     r0
+	{
+	 addli  r0, r0, .Lhandle_syscall_link - .
+	 jrp    lr
+	}
+	STD_ENDPROC(handle_syscall_link_address)
+
+STD_ENTRY(ret_from_fork)
+	jal     sim_notify_fork
+	jal     schedule_tail
+	FEEDBACK_REENTER(ret_from_fork)
+	{
+	 movei  r30, 0               /* not an NMI */
+	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
+	}
+	STD_ENDPROC(ret_from_fork)
+
+STD_ENTRY(ret_from_kernel_thread)
+	jal     sim_notify_fork
+	jal     schedule_tail
+	FEEDBACK_REENTER(ret_from_fork)
+	{
+	 move   r0, r31
+	 jalr   r30
+	}
+	FEEDBACK_REENTER(ret_from_kernel_thread)
+	{
+	 movei  r30, 0               /* not an NMI */
+	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
+	}
+	STD_ENDPROC(ret_from_kernel_thread)
+
+	/*
+	 * Code for ill interrupt.
+	 */
+	.pushsection .text.handle_ill,"ax"
+handle_ill:
+	finish_interrupt_save handle_ill
+
+	/*
+	 * Check for if we are single stepping in user level. If so, then
+	 * we need to restore the PC.
+	 */
+	check_single_stepping ill, .Ldispatch_normal_ill
+
+	{
+	 /* See if the PC is the 1st bundle in the buffer */
+	 seq    r25, r27, r26
+
+	 /* Point to the 2nd bundle in the buffer */
+	 addi   r26, r26, 8
+	}
+	{
+	 /* Point to the original pc */
+	 addi   r24, r29, SINGLESTEP_STATE_ORIG_PC_OFFSET
+
+	 /* Branch if the PC is the 1st bundle in the buffer */
+	 bnz    r25, 3f
+	}
+	{
+	 /* See if the PC is the 2nd bundle of the buffer */
+	 seq    r25, r27, r26
+
+	 /* Set PC to next instruction */
+	 addi   r24, r29, SINGLESTEP_STATE_NEXT_PC_OFFSET
+	}
+	{
+	 /* Point to flags */
+	 addi   r25, r29, SINGLESTEP_STATE_FLAGS_OFFSET
+
+	 /* Branch if PC is in the second bundle */
+	 bz     r25, 2f
+	}
+	/* Load flags */
+	lw      r25, r25
+	{
+	 /*
+	  * Get the offset for the register to restore
+	  * Note: the lower bound is 2, so we have implicit scaling by 4.
+	  *  No multiplication of the register number by the size of a register
+	  *  is needed.
+	  */
+	 mm     r27, r25, zero, SINGLESTEP_STATE_TARGET_LB, \
+		SINGLESTEP_STATE_TARGET_UB
+
+	 /* Mask Rewrite_LR */
+	 andi   r25, r25, SINGLESTEP_STATE_MASK_UPDATE
+	}
+	{
+	 addi   r29, r29, SINGLESTEP_STATE_UPDATE_VALUE_OFFSET
+
+	 /* Don't rewrite temp register */
+	 bz     r25, 3f
+	}
+	{
+	 /* Get the temp value */
+	 lw     r29, r29
+
+	 /* Point to where the register is stored */
+	 add    r27, r27, sp
+	}
+
+	/* Add in the C ABI save area size to the register offset */
+	addi    r27, r27, C_ABI_SAVE_AREA_SIZE
+
+	/* Restore the user's register with the temp value */
+	sw      r27, r29
+	j       3f
+
+2:
+	/* Must be in the third bundle */
+	addi    r24, r29, SINGLESTEP_STATE_BRANCH_NEXT_PC_OFFSET
+
+3:
+	/* set PC and continue */
+	lw      r26, r24
+	{
+	 sw     r28, r26
+	 GET_THREAD_INFO(r0)
+	}
+
+	/*
+	 * Clear TIF_SINGLESTEP to prevent recursion if we execute an ill.
+	 * The normal non-arch flow redundantly clears TIF_SINGLESTEP, but we
+	 * need to clear it here and can't really impose on all other arches.
+	 * So what's another write between friends?
+	 */
+
+	addi    r1, r0, THREAD_INFO_FLAGS_OFFSET
+	{
+	 lw     r2, r1
+	 addi   r0, r0, THREAD_INFO_TASK_OFFSET  /* currently a no-op */
+	}
+	andi    r2, r2, ~_TIF_SINGLESTEP
+	sw      r1, r2
+
+	/* Issue a sigtrap */
+	{
+	 lw     r0, r0          /* indirect thru thread_info to get task_info*/
+	 addi   r1, sp, C_ABI_SAVE_AREA_SIZE  /* put ptregs pointer into r1 */
+	}
+
+	jal     send_sigtrap    /* issue a SIGTRAP */
+	FEEDBACK_REENTER(handle_ill)
+	{
+	 movei  r30, 0               /* not an NMI */
+	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
+	}
+
+.Ldispatch_normal_ill:
+	{
+	 jalr   r0
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	}
+	FEEDBACK_REENTER(handle_ill)
+	{
+	 movei  r30, 0   /* not an NMI */
+	 j      interrupt_return
+	}
+	STD_ENDPROC(handle_ill)
+
+/* Various stub interrupt handlers and syscall handlers */
+
+STD_ENTRY_LOCAL(_kernel_double_fault)
+	mfspr   r1, SPR_EX_CONTEXT_K_0
+	move    r2, lr
+	move    r3, sp
+	move    r4, r52
+	addi    sp, sp, -C_ABI_SAVE_AREA_SIZE
+	j       kernel_double_fault
+	STD_ENDPROC(_kernel_double_fault)
+
+STD_ENTRY_LOCAL(bad_intr)
+	mfspr   r2, SPR_EX_CONTEXT_K_0
+	panic   "Unhandled interrupt %#x: PC %#lx"
+	STD_ENDPROC(bad_intr)
+
+/*
+ * Special-case sigreturn to not write r0 to the stack on return.
+ * This is technically more efficient, but it also avoids difficulties
+ * in the 64-bit OS when handling 32-bit compat code, since we must not
+ * sign-extend r0 for the sigreturn return-value case.
+ */
+#define PTREGS_SYSCALL_SIGRETURN(x, reg)                \
+	STD_ENTRY(_##x);                                \
+	addli   lr, lr, .Lsyscall_sigreturn_skip - .Lhandle_syscall_link; \
+	{                                               \
+	 PTREGS_PTR(reg, PTREGS_OFFSET_BASE);           \
+	 j      x                                       \
+	};                                              \
+	STD_ENDPROC(_##x)
+
+PTREGS_SYSCALL_SIGRETURN(sys_rt_sigreturn, r0)
+
+/* Save additional callee-saves to pt_regs and jump to standard function. */
+STD_ENTRY(_sys_clone)
+	push_extra_callee_saves r4
+	j       sys_clone
+	STD_ENDPROC(_sys_clone)
+
+/*
+ * This entrypoint is taken for the cmpxchg and atomic_update fast
+ * swints.  We may wish to generalize it to other fast swints at some
+ * point, but for now there are just two very similar ones, which
+ * makes it faster.
+ *
+ * The fast swint code is designed to have a small footprint.  It does
+ * not save or restore any GPRs, counting on the caller-save registers
+ * to be available to it on entry.  It does not modify any callee-save
+ * registers (including "lr").  It does not check what PL it is being
+ * called at, so you'd better not call it other than at PL0.
+ * The <atomic.h> wrapper assumes it only clobbers r20-r29, so if
+ * it ever is necessary to use more registers, be aware.
+ *
+ * It does not use the stack, but since it might be re-interrupted by
+ * a page fault which would assume the stack was valid, it does
+ * save/restore the stack pointer and zero it out to make sure it gets reset.
+ * Since we always keep interrupts disabled, the hypervisor won't
+ * clobber our EX_CONTEXT_K_x registers, so we don't save/restore them
+ * (other than to advance the PC on return).
+ *
+ * We have to manually validate the user vs kernel address range
+ * (since at PL1 we can read/write both), and for performance reasons
+ * we don't allow cmpxchg on the fc000000 memory region, since we only
+ * validate that the user address is below PAGE_OFFSET.
+ *
+ * We place it in the __HEAD section to ensure it is relatively
+ * near to the intvec_SWINT_1 code (reachable by a conditional branch).
+ *
+ * Our use of ATOMIC_LOCK_REG here must match do_page_fault_ics().
+ *
+ * As we do in lib/atomic_asm_32.S, we bypass a store if the value we
+ * would store is the same as the value we just loaded.
+ */
+	__HEAD
+	.align 64
+	/* Align much later jump on the start of a cache line. */
+	nop
+#if PAGE_SIZE >= 0x10000
+	nop
+#endif
+ENTRY(sys_cmpxchg)
+
+	/*
+	 * Save "sp" and set it zero for any possible page fault.
+	 *
+	 * HACK: We want to both zero sp and check r0's alignment,
+	 * so we do both at once. If "sp" becomes nonzero we
+	 * know r0 is unaligned and branch to the error handler that
+	 * restores sp, so this is OK.
+	 *
+	 * ICS is disabled right now so having a garbage but nonzero
+	 * sp is OK, since we won't execute any faulting instructions
+	 * when it is nonzero.
+	 */
+	{
+	 move   r27, sp
+	 andi	sp, r0, 3
+	}
+
+	/*
+	 * Get the lock address in ATOMIC_LOCK_REG, and also validate that the
+	 * address is less than PAGE_OFFSET, since that won't trap at PL1.
+	 * We only use bits less than PAGE_SHIFT to avoid having to worry
+	 * about aliasing among multiple mappings of the same physical page,
+	 * and we ignore the low 3 bits so we have one lock that covers
+	 * both a cmpxchg64() and a cmpxchg() on either its low or high word.
+	 * NOTE: this must match __atomic_hashed_lock() in lib/atomic_32.c.
+	 */
+
+#if (PAGE_OFFSET & 0xffff) != 0
+# error Code here assumes PAGE_OFFSET can be loaded with just hi16()
+#endif
+
+	{
+	 /* Check for unaligned input. */
+	 bnz    sp, .Lcmpxchg_badaddr
+	 auli   r23, zero, hi16(PAGE_OFFSET)  /* hugepage-aligned */
+	}
+	{
+	 /*
+	  * Slide bits into position for 'mm'. We want to ignore
+	  * the low 3 bits of r0, and consider only the next
+	  * ATOMIC_HASH_SHIFT bits.
+	  * Because of C pointer arithmetic, we want to compute this:
+	  *
+	  * ((char*)atomic_locks +
+	  *  (((r0 >> 3) & ((1 << ATOMIC_HASH_SHIFT) - 1)) << 2))
+	  *
+	  * Instead of two shifts we just ">> 1", and use 'mm'
+	  * to ignore the low and high bits we don't want.
+	  */
+	 shri	r25, r0, 1
+
+	 slt_u  r23, r0, r23
+
+	 /*
+	  * Ensure that the TLB is loaded before we take out the lock.
+	  * This will start fetching the value all the way into our L1
+	  * as well (and if it gets modified before we grab the lock,
+	  * it will be invalidated from our cache before we reload it).
+	  */
+	 lw	r26, r0
+	}
+	{
+	 auli	r21, zero, ha16(atomic_locks)
+
+	 bbns   r23, .Lcmpxchg_badaddr
+	}
+#if PAGE_SIZE < 0x10000
+	/* atomic_locks is page-aligned so for big pages we don't need this. */
+	addli   r21, r21, lo16(atomic_locks)
+#endif
+	{
+	 /*
+	  * Insert the hash bits into the page-aligned pointer.
+	  * ATOMIC_HASH_SHIFT is so big that we don't actually hash
+	  * the unmasked address bits, as that may cause unnecessary
+	  * collisions.
+	  */
+	 mm	ATOMIC_LOCK_REG_NAME, r25, r21, 2, (ATOMIC_HASH_SHIFT + 2) - 1
+
+	 seqi	r23, TREG_SYSCALL_NR_NAME, __NR_FAST_cmpxchg64
+	}
+	{
+	 /* Branch away at this point if we're doing a 64-bit cmpxchg. */
+	 bbs    r23, .Lcmpxchg64
+	 andi   r23, r0, 7       /* Precompute alignment for cmpxchg64. */
+	}
+	{
+	 /*
+	  * We very carefully align the code that actually runs with
+	  * the lock held (twelve bundles) so that we know it is all in
+	  * the icache when we start.  This instruction (the jump) is
+	  * at the start of the first cache line, address zero mod 64;
+	  * we jump to the very end of the second cache line to get that
+	  * line loaded in the icache, then fall through to issue the tns
+	  * in the third cache line, at which point it's all cached.
+	  * Note that is for performance, not correctness.
+	  */
+	 j      .Lcmpxchg32_tns
+	}
+
+/* Symbol for do_page_fault_ics() to use to compare against the PC. */
+.global __sys_cmpxchg_grab_lock
+__sys_cmpxchg_grab_lock:
+
+	/*
+	 * Perform the actual cmpxchg or atomic_update.
+	 */
+.Ldo_cmpxchg32:
+	{
+	 lw     r21, r0
+	 seqi	r23, TREG_SYSCALL_NR_NAME, __NR_FAST_atomic_update
+	 move	r24, r2
+	}
+	{
+	 seq    r22, r21, r1     /* See if cmpxchg matches. */
+	 and	r25, r21, r1     /* If atomic_update, compute (*mem & mask) */
+	}
+	{
+	 or	r22, r22, r23    /* Skip compare branch for atomic_update. */
+	 add	r25, r25, r2     /* Compute (*mem & mask) + addend. */
+	}
+	{
+	 mvnz	r24, r23, r25    /* Use atomic_update value if appropriate. */
+	 bbns   r22, .Lcmpxchg32_nostore
+	}
+	seq     r22, r24, r21    /* Are we storing the value we loaded? */
+	bbs     r22, .Lcmpxchg32_nostore
+	sw      r0, r24
+
+	/* The following instruction is the start of the second cache line. */
+	/* Do slow mtspr here so the following "mf" waits less. */
+	{
+	 move   sp, r27
+	 mtspr  SPR_EX_CONTEXT_K_0, r28
+	}
+	mf
+
+	{
+	 move   r0, r21
+	 sw     ATOMIC_LOCK_REG_NAME, zero
+	}
+	iret
+
+	/* Duplicated code here in the case where we don't overlap "mf" */
+.Lcmpxchg32_nostore:
+	{
+	 move   r0, r21
+	 sw     ATOMIC_LOCK_REG_NAME, zero
+	}
+	{
+	 move   sp, r27
+	 mtspr  SPR_EX_CONTEXT_K_0, r28
+	}
+	iret
+
+	/*
+	 * The locking code is the same for 32-bit cmpxchg/atomic_update,
+	 * and for 64-bit cmpxchg.  We provide it as a macro and put
+	 * it into both versions.  We can't share the code literally
+	 * since it depends on having the right branch-back address.
+	 */
+	.macro  cmpxchg_lock, bitwidth
+
+	/* Lock; if we succeed, jump back up to the read-modify-write. */
+#ifdef CONFIG_SMP
+	tns     r21, ATOMIC_LOCK_REG_NAME
+#else
+	/*
+	 * Non-SMP preserves all the lock infrastructure, to keep the
+	 * code simpler for the interesting (SMP) case.  However, we do
+	 * one small optimization here and in atomic_asm.S, which is
+	 * to fake out acquiring the actual lock in the atomic_lock table.
+	 */
+	movei	r21, 0
+#endif
+
+	/* Issue the slow SPR here while the tns result is in flight. */
+	mfspr   r28, SPR_EX_CONTEXT_K_0
+
+	{
+	 addi   r28, r28, 8    /* return to the instruction after the swint1 */
+	 bzt    r21, .Ldo_cmpxchg\bitwidth
+	}
+	/*
+	 * The preceding instruction is the last thing that must be
+	 * hot in the icache before we do the "tns" above.
+	 */
+
+#ifdef CONFIG_SMP
+	/*
+	 * We failed to acquire the tns lock on our first try.  Now use
+	 * bounded exponential backoff to retry, like __atomic_spinlock().
+	 */
+	{
+	 moveli r23, 2048       /* maximum backoff time in cycles */
+	 moveli r25, 32         /* starting backoff time in cycles */
+	}
+1:	mfspr   r26, CYCLE_LOW  /* get start point for this backoff */
+2:	mfspr   r22, CYCLE_LOW  /* test to see if we've backed off enough */
+	sub     r22, r22, r26
+	slt     r22, r22, r25
+	bbst    r22, 2b
+	{
+	 shli   r25, r25, 1     /* double the backoff; retry the tns */
+	 tns    r21, ATOMIC_LOCK_REG_NAME
+	}
+	slt     r26, r23, r25   /* is the proposed backoff too big? */
+	{
+	 mvnz   r25, r26, r23
+	 bzt    r21, .Ldo_cmpxchg\bitwidth
+	}
+	j       1b
+#endif /* CONFIG_SMP */
+	.endm
+
+.Lcmpxchg32_tns:
+	/*
+	 * This is the last instruction on the second cache line.
+	 * The nop here loads the second line, then we fall through
+	 * to the tns to load the third line before we take the lock.
+	 */
+	nop
+	cmpxchg_lock 32
+
+	/*
+	 * This code is invoked from sys_cmpxchg after most of the
+	 * preconditions have been checked.  We still need to check
+	 * that r0 is 8-byte aligned, since if it's not we won't
+	 * actually be atomic.  However, ATOMIC_LOCK_REG has the atomic
+	 * lock pointer and r27/r28 have the saved SP/PC.
+	 * r23 is holding "r0 & 7" so we can test for alignment.
+	 * The compare value is in r2/r3; the new value is in r4/r5.
+	 * On return, we must put the old value in r0/r1.
+	 */
+	.align 64
+.Lcmpxchg64:
+	{
+	 bzt     r23, .Lcmpxchg64_tns
+	}
+	j       .Lcmpxchg_badaddr
+
+.Ldo_cmpxchg64:
+	{
+	 lw     r21, r0
+	 addi   r25, r0, 4
+	}
+	{
+	 lw     r1, r25
+	}
+	seq     r26, r21, r2
+	{
+	 bz     r26, .Lcmpxchg64_mismatch
+	 seq    r26, r1, r3
+	}
+	{
+	 bz     r26, .Lcmpxchg64_mismatch
+	}
+	sw      r0, r4
+	sw      r25, r5
+
+	/*
+	 * The 32-bit path provides optimized "match" and "mismatch"
+	 * iret paths, but we don't have enough bundles in this cache line
+	 * to do that, so we just make even the "mismatch" path do an "mf".
+	 */
+.Lcmpxchg64_mismatch:
+	{
+	 move   sp, r27
+	 mtspr  SPR_EX_CONTEXT_K_0, r28
+	}
+	mf
+	{
+	 move   r0, r21
+	 sw     ATOMIC_LOCK_REG_NAME, zero
+	}
+	iret
+
+.Lcmpxchg64_tns:
+	cmpxchg_lock 64
+
+
+	/*
+	 * Reset sp and revector to sys_cmpxchg_badaddr(), which will
+	 * just raise the appropriate signal and exit.  Doing it this
+	 * way means we don't have to duplicate the code in intvec.S's
+	 * int_hand macro that locates the top of the stack.
+	 */
+.Lcmpxchg_badaddr:
+	{
+	 moveli TREG_SYSCALL_NR_NAME, __NR_cmpxchg_badaddr
+	 move   sp, r27
+	}
+	j       intvec_SWINT_1
+	ENDPROC(sys_cmpxchg)
+	ENTRY(__sys_cmpxchg_end)
+
+
+/* The single-step support may need to read all the registers. */
+int_unalign:
+	push_extra_callee_saves r0
+	j       do_trap
+
+/* Include .intrpt array of interrupt vectors */
+	.section ".intrpt", "ax"
+
+#ifndef CONFIG_USE_PMC
+#define handle_perf_interrupt bad_intr
+#endif
+
+#ifndef CONFIG_HARDWALL
+#define do_hardwall_trap bad_intr
+#endif
+
+	int_hand     INT_ITLB_MISS, ITLB_MISS, \
+		     do_page_fault, handle_interrupt_no_single_step
+	int_hand     INT_MEM_ERROR, MEM_ERROR, bad_intr
+	int_hand     INT_ILL, ILL, do_trap, handle_ill
+	int_hand     INT_GPV, GPV, do_trap
+	int_hand     INT_SN_ACCESS, SN_ACCESS, do_trap
+	int_hand     INT_IDN_ACCESS, IDN_ACCESS, do_trap
+	int_hand     INT_UDN_ACCESS, UDN_ACCESS, do_trap
+	int_hand     INT_IDN_REFILL, IDN_REFILL, bad_intr
+	int_hand     INT_UDN_REFILL, UDN_REFILL, bad_intr
+	int_hand     INT_IDN_COMPLETE, IDN_COMPLETE, bad_intr
+	int_hand     INT_UDN_COMPLETE, UDN_COMPLETE, bad_intr
+	int_hand     INT_SWINT_3, SWINT_3, do_trap
+	int_hand     INT_SWINT_2, SWINT_2, do_trap
+	int_hand     INT_SWINT_1, SWINT_1, SYSCALL, handle_syscall
+	int_hand     INT_SWINT_0, SWINT_0, do_trap
+	int_hand     INT_UNALIGN_DATA, UNALIGN_DATA, int_unalign
+	int_hand     INT_DTLB_MISS, DTLB_MISS, do_page_fault
+	int_hand     INT_DTLB_ACCESS, DTLB_ACCESS, do_page_fault
+	int_hand     INT_DMATLB_MISS, DMATLB_MISS, do_page_fault
+	int_hand     INT_DMATLB_ACCESS, DMATLB_ACCESS, do_page_fault
+	int_hand     INT_SNITLB_MISS, SNITLB_MISS, do_page_fault
+	int_hand     INT_SN_NOTIFY, SN_NOTIFY, bad_intr
+	int_hand     INT_SN_FIREWALL, SN_FIREWALL, do_hardwall_trap
+	int_hand     INT_IDN_FIREWALL, IDN_FIREWALL, bad_intr
+	int_hand     INT_UDN_FIREWALL, UDN_FIREWALL, do_hardwall_trap
+	int_hand     INT_TILE_TIMER, TILE_TIMER, do_timer_interrupt
+	int_hand     INT_IDN_TIMER, IDN_TIMER, bad_intr
+	int_hand     INT_UDN_TIMER, UDN_TIMER, bad_intr
+	int_hand     INT_DMA_NOTIFY, DMA_NOTIFY, bad_intr
+	int_hand     INT_IDN_CA, IDN_CA, bad_intr
+	int_hand     INT_UDN_CA, UDN_CA, bad_intr
+	int_hand     INT_IDN_AVAIL, IDN_AVAIL, bad_intr
+	int_hand     INT_UDN_AVAIL, UDN_AVAIL, bad_intr
+	int_hand     INT_PERF_COUNT, PERF_COUNT, \
+		     handle_perf_interrupt, handle_nmi
+	int_hand     INT_INTCTRL_3, INTCTRL_3, bad_intr
+#if CONFIG_KERNEL_PL == 2
+	dc_dispatch  INT_INTCTRL_2, INTCTRL_2
+	int_hand     INT_INTCTRL_1, INTCTRL_1, bad_intr
+#else
+	int_hand     INT_INTCTRL_2, INTCTRL_2, bad_intr
+	dc_dispatch  INT_INTCTRL_1, INTCTRL_1
+#endif
+	int_hand     INT_INTCTRL_0, INTCTRL_0, bad_intr
+	int_hand     INT_MESSAGE_RCV_DWNCL, MESSAGE_RCV_DWNCL, \
+		     hv_message_intr
+	int_hand     INT_DEV_INTR_DWNCL, DEV_INTR_DWNCL, \
+		     tile_dev_intr
+	int_hand     INT_I_ASID, I_ASID, bad_intr
+	int_hand     INT_D_ASID, D_ASID, bad_intr
+	int_hand     INT_DMATLB_MISS_DWNCL, DMATLB_MISS_DWNCL, \
+		     do_page_fault
+	int_hand     INT_SNITLB_MISS_DWNCL, SNITLB_MISS_DWNCL, \
+		     do_page_fault
+	int_hand     INT_DMATLB_ACCESS_DWNCL, DMATLB_ACCESS_DWNCL, \
+		     do_page_fault
+	int_hand     INT_SN_CPL, SN_CPL, bad_intr
+	int_hand     INT_DOUBLE_FAULT, DOUBLE_FAULT, do_trap
+	int_hand     INT_AUX_PERF_COUNT, AUX_PERF_COUNT, \
+		     handle_perf_interrupt, handle_nmi
+
+	/* Synthetic interrupt delivered only by the simulator */
+	int_hand     INT_BREAKPOINT, BREAKPOINT, do_breakpoint
diff --git a/arch/tile/kernel/intvec_64.S b/arch/tile/kernel/intvec_64.S
new file mode 100644
index 000000000..5b67efcec
--- /dev/null
+++ b/arch/tile/kernel/intvec_64.S
@@ -0,0 +1,1573 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Linux interrupt vectors.
+ */
+
+#include <linux/linkage.h>
+#include <linux/errno.h>
+#include <linux/unistd.h>
+#include <linux/init.h>
+#include <asm/ptrace.h>
+#include <asm/thread_info.h>
+#include <asm/irqflags.h>
+#include <asm/asm-offsets.h>
+#include <asm/types.h>
+#include <asm/traps.h>
+#include <asm/signal.h>
+#include <hv/hypervisor.h>
+#include <arch/abi.h>
+#include <arch/interrupts.h>
+#include <arch/spr_def.h>
+
+#define PTREGS_PTR(reg, ptreg) addli reg, sp, C_ABI_SAVE_AREA_SIZE + (ptreg)
+
+#define PTREGS_OFFSET_SYSCALL PTREGS_OFFSET_REG(TREG_SYSCALL_NR)
+
+#if CONFIG_KERNEL_PL == 1 || CONFIG_KERNEL_PL == 2
+/*
+ * Set "result" non-zero if ex1 holds the PL of the kernel
+ * (with or without ICS being set).  Note this works only
+ * because we never find the PL at level 3.
+ */
+# define IS_KERNEL_EX1(result, ex1) andi result, ex1, CONFIG_KERNEL_PL
+#else
+# error Recode IS_KERNEL_EX1 for CONFIG_KERNEL_PL
+#endif
+
+	.macro  push_reg reg, ptr=sp, delta=-8
+	{
+	 st     \ptr, \reg
+	 addli  \ptr, \ptr, \delta
+	}
+	.endm
+
+	.macro  pop_reg reg, ptr=sp, delta=8
+	{
+	 ld     \reg, \ptr
+	 addli  \ptr, \ptr, \delta
+	}
+	.endm
+
+	.macro  pop_reg_zero reg, zreg, ptr=sp, delta=8
+	{
+	 move   \zreg, zero
+	 ld     \reg, \ptr
+	 addi   \ptr, \ptr, \delta
+	}
+	.endm
+
+	.macro  push_extra_callee_saves reg
+	PTREGS_PTR(\reg, PTREGS_OFFSET_REG(51))
+	push_reg r51, \reg
+	push_reg r50, \reg
+	push_reg r49, \reg
+	push_reg r48, \reg
+	push_reg r47, \reg
+	push_reg r46, \reg
+	push_reg r45, \reg
+	push_reg r44, \reg
+	push_reg r43, \reg
+	push_reg r42, \reg
+	push_reg r41, \reg
+	push_reg r40, \reg
+	push_reg r39, \reg
+	push_reg r38, \reg
+	push_reg r37, \reg
+	push_reg r36, \reg
+	push_reg r35, \reg
+	push_reg r34, \reg, PTREGS_OFFSET_BASE - PTREGS_OFFSET_REG(34)
+	.endm
+
+	.macro  panic str
+	.pushsection .rodata, "a"
+1:
+	.asciz  "\str"
+	.popsection
+	{
+	 moveli r0, hw2_last(1b)
+	}
+	{
+	 shl16insli r0, r0, hw1(1b)
+	}
+	{
+	 shl16insli r0, r0, hw0(1b)
+	 jal    panic
+	}
+	.endm
+
+	/*
+	 * Unalign data exception fast handling: In order to handle
+	 * unaligned data access, a fast JIT version is generated and stored
+	 * in a specific area in user space. We first need to do a quick poke
+	 * to see if the JIT is available. We use certain bits in the fault
+	 * PC (3 to 9 is used for 16KB page size) as index to address the JIT
+	 * code area. The first 64bit word is the fault PC, and the 2nd one is
+	 * the fault bundle itself. If these 2 words both match, then we
+	 * directly "iret" to JIT code. If not, a slow path is invoked to
+	 * generate new JIT code. Note: the current JIT code WILL be
+	 * overwritten if it existed. So, ideally we can handle 128 unalign
+	 * fixups via JIT. For lookup efficiency and to effectively support
+	 * tight loops with multiple unaligned reference, a simple
+	 * direct-mapped cache is used.
+	 *
+	 * SPR_EX_CONTEXT_K_0 is modified to return to JIT code.
+	 * SPR_EX_CONTEXT_K_1 has ICS set.
+	 * SPR_EX_CONTEXT_0_0 is setup to user program's next PC.
+	 * SPR_EX_CONTEXT_0_1 = 0.
+	 */
+	.macro int_hand_unalign_fast  vecnum, vecname
+	.org  (\vecnum << 8)
+intvec_\vecname:
+	/* Put r3 in SPR_SYSTEM_SAVE_K_1.  */
+	mtspr   SPR_SYSTEM_SAVE_K_1, r3
+
+	mfspr   r3, SPR_EX_CONTEXT_K_1
+	/*
+	 * Examine if exception comes from user without ICS set.
+	 * If not, just go directly to the slow path.
+	 */
+	bnez    r3, hand_unalign_slow_nonuser
+
+	mfspr   r3, SPR_SYSTEM_SAVE_K_0
+
+	/* Get &thread_info->unalign_jit_tmp[0] in r3. */
+	bfexts  r3, r3, 0, CPU_SHIFT-1
+	mm      r3, zero, LOG2_THREAD_SIZE, 63
+	addli   r3, r3, THREAD_INFO_UNALIGN_JIT_TMP_OFFSET
+
+	/*
+	 * Save r0, r1, r2 into thread_info array r3 points to
+	 * from low to high memory in order.
+	 */
+	st_add  r3, r0, 8
+	st_add  r3, r1, 8
+	{
+	 st_add r3, r2, 8
+	 andi   r2, sp, 7
+	}
+
+	/* Save stored r3 value so we can revert it on a page fault. */
+	mfspr   r1, SPR_SYSTEM_SAVE_K_1
+	st      r3, r1
+
+	{
+	 /* Generate a SIGBUS if sp is not 8-byte aligned. */
+	 bnez   r2, hand_unalign_slow_badsp
+	}
+
+	/*
+	 * Get the thread_info in r0; load r1 with pc. Set the low bit of sp
+	 * as an indicator to the page fault code in case we fault.
+	 */
+	{
+	 ori    sp, sp, 1
+	 mfspr  r1, SPR_EX_CONTEXT_K_0
+	}
+
+	/* Add the jit_info offset in thread_info; extract r1 [3:9] into r2. */
+	{
+	 addli  r0, r3, THREAD_INFO_UNALIGN_JIT_BASE_OFFSET - \
+	  (THREAD_INFO_UNALIGN_JIT_TMP_OFFSET + (3 * 8))
+	 bfextu r2, r1, 3, (2 + PAGE_SHIFT - UNALIGN_JIT_SHIFT)
+	}
+
+	/* Load the jit_info; multiply r2 by 128. */
+	{
+	 ld     r0, r0
+	 shli   r2, r2, UNALIGN_JIT_SHIFT
+	}
+
+	/*
+	 * If r0 is NULL, the JIT page is not mapped, so go to slow path;
+	 * add offset r2 to r0 at the same time.
+	 */
+	{
+	 beqz   r0, hand_unalign_slow
+	 add    r2, r0, r2
+	}
+
+        /*
+	 * We are loading from userspace (both the JIT info PC and
+	 * instruction word, and the instruction word we executed)
+	 * and since either could fault while holding the interrupt
+	 * critical section, we must tag this region and check it in
+	 * do_page_fault() to handle it properly.
+	 */
+ENTRY(__start_unalign_asm_code)
+
+	/* Load first word of JIT in r0 and increment r2 by 8. */
+	ld_add  r0, r2, 8
+
+	/*
+	 * Compare the PC with the 1st word in JIT; load the fault bundle
+	 * into r1.
+	 */
+	{
+	 cmpeq  r0, r0, r1
+	 ld     r1, r1
+	}
+
+	/* Go to slow path if PC doesn't match. */
+	beqz    r0, hand_unalign_slow
+
+	/*
+	 * Load the 2nd word of JIT, which is supposed to be the fault
+	 * bundle for a cache hit. Increment r2; after this bundle r2 will
+	 * point to the potential start of the JIT code we want to run.
+	 */
+	ld_add  r0, r2, 8
+
+	/* No further accesses to userspace are done after this point. */
+ENTRY(__end_unalign_asm_code)
+
+	/* Compare the real bundle with what is saved in the JIT area. */
+	{
+	 cmpeq  r0, r1, r0
+	 mtspr  SPR_EX_CONTEXT_0_1, zero
+	}
+
+	/* Go to slow path if the fault bundle does not match. */
+	beqz    r0, hand_unalign_slow
+
+	/*
+	 * A cache hit is found.
+	 * r2 points to start of JIT code (3rd word).
+	 * r0 is the fault pc.
+	 * r1 is the fault bundle.
+	 * Reset the low bit of sp.
+	 */
+	{
+	 mfspr  r0, SPR_EX_CONTEXT_K_0
+	 andi   sp, sp, ~1
+	}
+
+	/* Write r2 into EX_CONTEXT_K_0 and increment PC. */
+	{
+	 mtspr  SPR_EX_CONTEXT_K_0, r2
+	 addi   r0, r0, 8
+	}
+
+	/*
+	 * Set ICS on kernel EX_CONTEXT_K_1 in order to "iret" to
+	 * user with ICS set. This way, if the JIT fixup causes another
+	 * unalign exception (which shouldn't be possible) the user
+	 * process will be terminated with SIGBUS. Also, our fixup will
+	 * run without interleaving with external interrupts.
+	 * Each fixup is at most 14 bundles, so it won't hold ICS for long.
+	 */
+	{
+	 movei  r1, PL_ICS_EX1(USER_PL, 1)
+	 mtspr  SPR_EX_CONTEXT_0_0, r0
+	}
+
+	{
+	 mtspr  SPR_EX_CONTEXT_K_1, r1
+	 addi   r3, r3, -(3 * 8)
+	}
+
+	/* Restore r0..r3. */
+	ld_add  r0, r3, 8
+	ld_add  r1, r3, 8
+	ld_add  r2, r3, 8
+	ld      r3, r3
+
+	iret
+	ENDPROC(intvec_\vecname)
+	.endm
+
+#ifdef __COLLECT_LINKER_FEEDBACK__
+	.pushsection .text.intvec_feedback,"ax"
+intvec_feedback:
+	.popsection
+#endif
+
+	/*
+	 * Default interrupt handler.
+	 *
+	 * vecnum is where we'll put this code.
+	 * c_routine is the C routine we'll call.
+	 *
+	 * The C routine is passed two arguments:
+	 * - A pointer to the pt_regs state.
+	 * - The interrupt vector number.
+	 *
+	 * The "processing" argument specifies the code for processing
+	 * the interrupt. Defaults to "handle_interrupt".
+	 */
+	.macro __int_hand vecnum, vecname, c_routine,processing=handle_interrupt
+intvec_\vecname:
+	/* Temporarily save a register so we have somewhere to work. */
+
+	mtspr   SPR_SYSTEM_SAVE_K_1, r0
+	mfspr   r0, SPR_EX_CONTEXT_K_1
+
+	/*
+	 * The unalign data fastpath code sets the low bit in sp to
+	 * force us to reset it here on fault.
+	 */
+	{
+	 blbs   sp, 2f
+	 IS_KERNEL_EX1(r0, r0)
+	}
+
+	.ifc    \vecnum, INT_DOUBLE_FAULT
+	/*
+	 * For double-faults from user-space, fall through to the normal
+	 * register save and stack setup path.  Otherwise, it's the
+	 * hypervisor giving us one last chance to dump diagnostics, and we
+	 * branch to the kernel_double_fault routine to do so.
+	 */
+	beqz    r0, 1f
+	j       _kernel_double_fault
+1:
+	.else
+	/*
+	 * If we're coming from user-space, then set sp to the top of
+	 * the kernel stack.  Otherwise, assume sp is already valid.
+	 */
+	{
+	 bnez   r0, 0f
+	 move   r0, sp
+	}
+	.endif
+
+	.ifc    \c_routine, do_page_fault
+	/*
+	 * The page_fault handler may be downcalled directly by the
+	 * hypervisor even when Linux is running and has ICS set.
+	 *
+	 * In this case the contents of EX_CONTEXT_K_1 reflect the
+	 * previous fault and can't be relied on to choose whether or
+	 * not to reinitialize the stack pointer.  So we add a test
+	 * to see whether SYSTEM_SAVE_K_2 has the high bit set,
+	 * and if so we don't reinitialize sp, since we must be coming
+	 * from Linux.  (In fact the precise case is !(val & ~1),
+	 * but any Linux PC has to have the high bit set.)
+	 *
+	 * Note that the hypervisor *always* sets SYSTEM_SAVE_K_2 for
+	 * any path that turns into a downcall to one of our TLB handlers.
+	 *
+	 * FIXME: if we end up never using this path, perhaps we should
+	 * prevent the hypervisor from generating downcalls in this case.
+	 * The advantage of getting a downcall is we can panic in Linux.
+	 */
+	mfspr   r0, SPR_SYSTEM_SAVE_K_2
+	{
+	 bltz   r0, 0f    /* high bit in S_S_1_2 is for a PC to use */
+	 move   r0, sp
+	}
+	.endif
+
+2:
+	/*
+	 * SYSTEM_SAVE_K_0 holds the cpu number in the high bits, and
+	 * the current stack top in the lower bits.  So we recover
+	 * our starting stack value by sign-extending the low bits, then
+	 * point sp at the top aligned address on the actual stack page.
+	 */
+	mfspr   r0, SPR_SYSTEM_SAVE_K_0
+	bfexts  r0, r0, 0, CPU_SHIFT-1
+
+0:
+	/*
+	 * Align the stack mod 64 so we can properly predict what
+	 * cache lines we need to write-hint to reduce memory fetch
+	 * latency as we enter the kernel.  The layout of memory is
+	 * as follows, with cache line 0 at the lowest VA, and cache
+	 * line 8 just below the r0 value this "andi" computes.
+	 * Note that we never write to cache line 8, and we skip
+	 * cache lines 1-3 for syscalls.
+	 *
+	 *    cache line 8: ptregs padding (two words)
+	 *    cache line 7: sp, lr, pc, ex1, faultnum, orig_r0, flags, cmpexch
+	 *    cache line 6: r46...r53 (tp)
+	 *    cache line 5: r38...r45
+	 *    cache line 4: r30...r37
+	 *    cache line 3: r22...r29
+	 *    cache line 2: r14...r21
+	 *    cache line 1: r6...r13
+	 *    cache line 0: 2 x frame, r0..r5
+	 */
+#if STACK_TOP_DELTA != 64
+#error STACK_TOP_DELTA must be 64 for assumptions here and in task_pt_regs()
+#endif
+	andi    r0, r0, -64
+
+	/*
+	 * Push the first four registers on the stack, so that we can set
+	 * them to vector-unique values before we jump to the common code.
+	 *
+	 * Registers are pushed on the stack as a struct pt_regs,
+	 * with the sp initially just above the struct, and when we're
+	 * done, sp points to the base of the struct, minus
+	 * C_ABI_SAVE_AREA_SIZE, so we can directly jal to C code.
+	 *
+	 * This routine saves just the first four registers, plus the
+	 * stack context so we can do proper backtracing right away,
+	 * and defers to handle_interrupt to save the rest.
+	 * The backtracer needs pc, ex1, lr, sp, r52, and faultnum,
+	 * and needs sp set to its final location at the bottom of
+	 * the stack frame.
+	 */
+	addli   r0, r0, PTREGS_OFFSET_LR - (PTREGS_SIZE + KSTK_PTREGS_GAP)
+	wh64    r0   /* cache line 7 */
+	{
+	 st     r0, lr
+	 addli  r0, r0, PTREGS_OFFSET_SP - PTREGS_OFFSET_LR
+	}
+	{
+	 st     r0, sp
+	 addli  sp, r0, PTREGS_OFFSET_REG(52) - PTREGS_OFFSET_SP
+	}
+	wh64    sp   /* cache line 6 */
+	{
+	 st     sp, r52
+	 addli  sp, sp, PTREGS_OFFSET_REG(1) - PTREGS_OFFSET_REG(52)
+	}
+	wh64    sp   /* cache line 0 */
+	{
+	 st     sp, r1
+	 addli  sp, sp, PTREGS_OFFSET_REG(2) - PTREGS_OFFSET_REG(1)
+	}
+	{
+	 st     sp, r2
+	 addli  sp, sp, PTREGS_OFFSET_REG(3) - PTREGS_OFFSET_REG(2)
+	}
+	{
+	 st     sp, r3
+	 addli  sp, sp, PTREGS_OFFSET_PC - PTREGS_OFFSET_REG(3)
+	}
+	mfspr   r0, SPR_EX_CONTEXT_K_0
+	.ifc \processing,handle_syscall
+	/*
+	 * Bump the saved PC by one bundle so that when we return, we won't
+	 * execute the same swint instruction again.  We need to do this while
+	 * we're in the critical section.
+	 */
+	addi    r0, r0, 8
+	.endif
+	{
+	 st     sp, r0
+	 addli  sp, sp, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_PC
+	}
+	mfspr   r0, SPR_EX_CONTEXT_K_1
+	{
+	 st     sp, r0
+	 addi   sp, sp, PTREGS_OFFSET_FAULTNUM - PTREGS_OFFSET_EX1
+	/*
+	 * Use r0 for syscalls so it's a temporary; use r1 for interrupts
+	 * so that it gets passed through unchanged to the handler routine.
+	 * Note that the .if conditional confusingly spans bundles.
+	 */
+	 .ifc \processing,handle_syscall
+	 movei  r0, \vecnum
+	}
+	{
+	 st     sp, r0
+	 .else
+	 movei  r1, \vecnum
+	}
+	{
+	 st     sp, r1
+	 .endif
+	 addli  sp, sp, PTREGS_OFFSET_REG(0) - PTREGS_OFFSET_FAULTNUM
+	}
+	mfspr   r0, SPR_SYSTEM_SAVE_K_1    /* Original r0 */
+	{
+	 st     sp, r0
+	 addi   sp, sp, -PTREGS_OFFSET_REG(0) - 8
+	}
+	{
+	 st     sp, zero        /* write zero into "Next SP" frame pointer */
+	 addi   sp, sp, -8      /* leave SP pointing at bottom of frame */
+	}
+	.ifc \processing,handle_syscall
+	j       handle_syscall
+	.else
+	/* Capture per-interrupt SPR context to registers. */
+	.ifc \c_routine, do_page_fault
+	mfspr   r2, SPR_SYSTEM_SAVE_K_3   /* address of page fault */
+	mfspr   r3, SPR_SYSTEM_SAVE_K_2   /* info about page fault */
+	.else
+	.ifc \vecnum, INT_ILL_TRANS
+	mfspr   r2, ILL_VA_PC
+	.else
+	.ifc \vecnum, INT_DOUBLE_FAULT
+	mfspr   r2, SPR_SYSTEM_SAVE_K_2   /* double fault info from HV */
+	.else
+	.ifc \c_routine, do_trap
+	mfspr   r2, GPV_REASON
+	.else
+	.ifc \c_routine, handle_perf_interrupt
+	mfspr   r2, PERF_COUNT_STS
+	.else
+	.ifc \c_routine, handle_perf_interrupt
+	mfspr   r2, AUX_PERF_COUNT_STS
+	.endif
+	.endif
+	.endif
+	.endif
+	.endif
+	.endif
+	/* Put function pointer in r0 */
+	moveli  r0, hw2_last(\c_routine)
+	shl16insli r0, r0, hw1(\c_routine)
+	{
+	 shl16insli r0, r0, hw0(\c_routine)
+	 j       \processing
+	}
+	.endif
+	ENDPROC(intvec_\vecname)
+
+#ifdef __COLLECT_LINKER_FEEDBACK__
+	.pushsection .text.intvec_feedback,"ax"
+	.org    (\vecnum << 5)
+	FEEDBACK_ENTER_EXPLICIT(intvec_\vecname, .intrpt, 1 << 8)
+	jrp     lr
+	.popsection
+#endif
+
+	.endm
+
+
+	/*
+	 * Save the rest of the registers that we didn't save in the actual
+	 * vector itself.  We can't use r0-r10 inclusive here.
+	 */
+	.macro  finish_interrupt_save, function
+
+	/* If it's a syscall, save a proper orig_r0, otherwise just zero. */
+	PTREGS_PTR(r52, PTREGS_OFFSET_ORIG_R0)
+	{
+	 .ifc \function,handle_syscall
+	 st     r52, r0
+	 .else
+	 st     r52, zero
+	 .endif
+	 PTREGS_PTR(r52, PTREGS_OFFSET_TP)
+	}
+	st      r52, tp
+	{
+	 mfspr  tp, CMPEXCH_VALUE
+	 PTREGS_PTR(r52, PTREGS_OFFSET_CMPEXCH)
+	}
+
+	/*
+	 * For ordinary syscalls, we save neither caller- nor callee-
+	 * save registers, since the syscall invoker doesn't expect the
+	 * caller-saves to be saved, and the called kernel functions will
+	 * take care of saving the callee-saves for us.
+	 *
+	 * For interrupts we save just the caller-save registers.  Saving
+	 * them is required (since the "caller" can't save them).  Again,
+	 * the called kernel functions will restore the callee-save
+	 * registers for us appropriately.
+	 *
+	 * On return, we normally restore nothing special for syscalls,
+	 * and just the caller-save registers for interrupts.
+	 *
+	 * However, there are some important caveats to all this:
+	 *
+	 * - We always save a few callee-save registers to give us
+	 *   some scratchpad registers to carry across function calls.
+	 *
+	 * - fork/vfork/etc require us to save all the callee-save
+	 *   registers, which we do in PTREGS_SYSCALL_ALL_REGS, below.
+	 *
+	 * - We always save r0..r5 and r10 for syscalls, since we need
+	 *   to reload them a bit later for the actual kernel call, and
+	 *   since we might need them for -ERESTARTNOINTR, etc.
+	 *
+	 * - Before invoking a signal handler, we save the unsaved
+	 *   callee-save registers so they are visible to the
+	 *   signal handler or any ptracer.
+	 *
+	 * - If the unsaved callee-save registers are modified, we set
+	 *   a bit in pt_regs so we know to reload them from pt_regs
+	 *   and not just rely on the kernel function unwinding.
+	 *   (Done for ptrace register writes and SA_SIGINFO handler.)
+	 */
+	{
+	 st     r52, tp
+	 PTREGS_PTR(r52, PTREGS_OFFSET_REG(33))
+	}
+	wh64    r52    /* cache line 4 */
+	push_reg r33, r52
+	push_reg r32, r52
+	push_reg r31, r52
+	.ifc \function,handle_syscall
+	push_reg r30, r52, PTREGS_OFFSET_SYSCALL - PTREGS_OFFSET_REG(30)
+	push_reg TREG_SYSCALL_NR_NAME, r52, \
+	  PTREGS_OFFSET_REG(5) - PTREGS_OFFSET_SYSCALL
+	.else
+
+	push_reg r30, r52, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(30)
+	wh64    r52   /* cache line 3 */
+	push_reg r29, r52
+	push_reg r28, r52
+	push_reg r27, r52
+	push_reg r26, r52
+	push_reg r25, r52
+	push_reg r24, r52
+	push_reg r23, r52
+	push_reg r22, r52
+	wh64    r52   /* cache line 2 */
+	push_reg r21, r52
+	push_reg r20, r52
+	push_reg r19, r52
+	push_reg r18, r52
+	push_reg r17, r52
+	push_reg r16, r52
+	push_reg r15, r52
+	push_reg r14, r52
+	wh64    r52   /* cache line 1 */
+	push_reg r13, r52
+	push_reg r12, r52
+	push_reg r11, r52
+	push_reg r10, r52
+	push_reg r9, r52
+	push_reg r8, r52
+	push_reg r7, r52
+	push_reg r6, r52
+
+	.endif
+
+	push_reg r5, r52
+	st      r52, r4
+
+	/*
+	 * If we will be returning to the kernel, we will need to
+	 * reset the interrupt masks to the state they had before.
+	 * Set DISABLE_IRQ in flags iff we came from kernel pl with
+	 * irqs disabled.
+	 */
+	mfspr   r32, SPR_EX_CONTEXT_K_1
+	{
+	 IS_KERNEL_EX1(r22, r22)
+	 PTREGS_PTR(r21, PTREGS_OFFSET_FLAGS)
+	}
+	beqzt   r32, 1f       /* zero if from user space */
+	IRQS_DISABLED(r32)    /* zero if irqs enabled */
+#if PT_FLAGS_DISABLE_IRQ != 1
+# error Value of IRQS_DISABLED used to set PT_FLAGS_DISABLE_IRQ; fix
+#endif
+1:
+	.ifnc \function,handle_syscall
+	/* Record the fact that we saved the caller-save registers above. */
+	ori     r32, r32, PT_FLAGS_CALLER_SAVES
+	.endif
+	st      r21, r32
+
+	/*
+	 * we've captured enough state to the stack (including in
+	 * particular our EX_CONTEXT state) that we can now release
+	 * the interrupt critical section and replace it with our
+	 * standard "interrupts disabled" mask value.  This allows
+	 * synchronous interrupts (and profile interrupts) to punch
+	 * through from this point onwards.
+	 *
+	 * It's important that no code before this point touch memory
+	 * other than our own stack (to keep the invariant that this
+	 * is all that gets touched under ICS), and that no code after
+	 * this point reference any interrupt-specific SPR, in particular
+	 * the EX_CONTEXT_K_ values.
+	 */
+	.ifc \function,handle_nmi
+	IRQ_DISABLE_ALL(r20)
+	.else
+	IRQ_DISABLE(r20, r21)
+	.endif
+	mtspr   INTERRUPT_CRITICAL_SECTION, zero
+
+	/* Load tp with our per-cpu offset. */
+#ifdef CONFIG_SMP
+	{
+	 mfspr  r20, SPR_SYSTEM_SAVE_K_0
+	 moveli r21, hw2_last(__per_cpu_offset)
+	}
+	{
+	 shl16insli r21, r21, hw1(__per_cpu_offset)
+	 bfextu r20, r20, CPU_SHIFT, 63
+	}
+	shl16insli r21, r21, hw0(__per_cpu_offset)
+	shl3add r20, r20, r21
+	ld      tp, r20
+#else
+	move    tp, zero
+#endif
+
+#ifdef __COLLECT_LINKER_FEEDBACK__
+	/*
+	 * Notify the feedback routines that we were in the
+	 * appropriate fixed interrupt vector area.  Note that we
+	 * still have ICS set at this point, so we can't invoke any
+	 * atomic operations or we will panic.  The feedback
+	 * routines internally preserve r0..r10 and r30 up.
+	 */
+	.ifnc \function,handle_syscall
+	shli    r20, r1, 5
+	.else
+	moveli  r20, INT_SWINT_1 << 5
+	.endif
+	moveli  r21, hw2_last(intvec_feedback)
+	shl16insli r21, r21, hw1(intvec_feedback)
+	shl16insli r21, r21, hw0(intvec_feedback)
+	add     r20, r20, r21
+	jalr    r20
+
+	/* And now notify the feedback routines that we are here. */
+	FEEDBACK_ENTER(\function)
+#endif
+
+	/*
+	 * Prepare the first 256 stack bytes to be rapidly accessible
+	 * without having to fetch the background data.
+	 */
+	addi    r52, sp, -64
+	{
+	 wh64   r52
+	 addi   r52, r52, -64
+	}
+	{
+	 wh64   r52
+	 addi   r52, r52, -64
+	}
+	{
+	 wh64   r52
+	 addi   r52, r52, -64
+	}
+	wh64    r52
+
+#ifdef CONFIG_TRACE_IRQFLAGS
+	.ifnc \function,handle_nmi
+	/*
+	 * We finally have enough state set up to notify the irq
+	 * tracing code that irqs were disabled on entry to the handler.
+	 * The TRACE_IRQS_OFF call clobbers registers r0-r29.
+	 * For syscalls, we already have the register state saved away
+	 * on the stack, so we don't bother to do any register saves here,
+	 * and later we pop the registers back off the kernel stack.
+	 * For interrupt handlers, save r0-r3 in callee-saved registers.
+	 */
+	.ifnc \function,handle_syscall
+	{ move r30, r0; move r31, r1 }
+	{ move r32, r2; move r33, r3 }
+	.endif
+	TRACE_IRQS_OFF
+	.ifnc \function,handle_syscall
+	{ move r0, r30; move r1, r31 }
+	{ move r2, r32; move r3, r33 }
+	.endif
+	.endif
+#endif
+
+	.endm
+
+	/*
+	 * Redispatch a downcall.
+	 */
+	.macro  dc_dispatch vecnum, vecname
+	.org    (\vecnum << 8)
+intvec_\vecname:
+	j       _hv_downcall_dispatch
+	ENDPROC(intvec_\vecname)
+	.endm
+
+	/*
+	 * Common code for most interrupts.  The C function we're eventually
+	 * going to is in r0, and the faultnum is in r1; the original
+	 * values for those registers are on the stack.
+	 */
+	.pushsection .text.handle_interrupt,"ax"
+handle_interrupt:
+	finish_interrupt_save handle_interrupt
+
+	/* Jump to the C routine; it should enable irqs as soon as possible. */
+	{
+	 jalr   r0
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	}
+	FEEDBACK_REENTER(handle_interrupt)
+	{
+	 movei  r30, 0   /* not an NMI */
+	 j      interrupt_return
+	}
+	STD_ENDPROC(handle_interrupt)
+
+/*
+ * This routine takes a boolean in r30 indicating if this is an NMI.
+ * If so, we also expect a boolean in r31 indicating whether to
+ * re-enable the oprofile interrupts.
+ *
+ * Note that .Lresume_userspace is jumped to directly in several
+ * places, and we need to make sure r30 is set correctly in those
+ * callers as well.
+ */
+STD_ENTRY(interrupt_return)
+	/* If we're resuming to kernel space, don't check thread flags. */
+	{
+	 bnez   r30, .Lrestore_all  /* NMIs don't special-case user-space */
+	 PTREGS_PTR(r29, PTREGS_OFFSET_EX1)
+	}
+	ld      r29, r29
+	IS_KERNEL_EX1(r29, r29)
+	{
+	 beqzt  r29, .Lresume_userspace
+	 move   r29, sp
+	}
+
+#ifdef CONFIG_PREEMPT
+	/* Returning to kernel space. Check if we need preemption. */
+	EXTRACT_THREAD_INFO(r29)
+	addli   r28, r29, THREAD_INFO_FLAGS_OFFSET
+	{
+	 ld     r28, r28
+	 addli  r29, r29, THREAD_INFO_PREEMPT_COUNT_OFFSET
+	}
+	{
+	 andi   r28, r28, _TIF_NEED_RESCHED
+	 ld4s   r29, r29
+	}
+	beqzt   r28, 1f
+	bnez    r29, 1f
+	/* Disable interrupts explicitly for preemption. */
+	IRQ_DISABLE(r20,r21)
+	TRACE_IRQS_OFF
+	jal     preempt_schedule_irq
+	FEEDBACK_REENTER(interrupt_return)
+1:
+#endif
+
+	/* If we're resuming to _cpu_idle_nap, bump PC forward by 8. */
+	{
+	 moveli r27, hw2_last(_cpu_idle_nap)
+	 PTREGS_PTR(r29, PTREGS_OFFSET_PC)
+	}
+	{
+	 ld     r28, r29
+	 shl16insli r27, r27, hw1(_cpu_idle_nap)
+	}
+	{
+	 shl16insli r27, r27, hw0(_cpu_idle_nap)
+	}
+	{
+	 cmpeq  r27, r27, r28
+	}
+	{
+	 blbc   r27, .Lrestore_all
+	 addi   r28, r28, 8
+	}
+	st      r29, r28
+	j       .Lrestore_all
+
+.Lresume_userspace:
+	FEEDBACK_REENTER(interrupt_return)
+
+	/*
+	 * Use r33 to hold whether we have already loaded the callee-saves
+	 * into ptregs.  We don't want to do it twice in this loop, since
+	 * then we'd clobber whatever changes are made by ptrace, etc.
+	 */
+	{
+	 movei  r33, 0
+	 move   r32, sp
+	}
+
+	/* Get base of stack in r32. */
+	EXTRACT_THREAD_INFO(r32)
+
+.Lretry_work_pending:
+	/*
+	 * Disable interrupts so as to make sure we don't
+	 * miss an interrupt that sets any of the thread flags (like
+	 * need_resched or sigpending) between sampling and the iret.
+	 * Routines like schedule() or do_signal() may re-enable
+	 * interrupts before returning.
+	 */
+	IRQ_DISABLE(r20, r21)
+	TRACE_IRQS_OFF  /* Note: clobbers registers r0-r29 */
+
+
+	/* Check to see if there is any work to do before returning to user. */
+	{
+	 addi   r29, r32, THREAD_INFO_FLAGS_OFFSET
+	 moveli r1, hw1_last(_TIF_ALLWORK_MASK)
+	}
+	{
+	 ld     r29, r29
+	 shl16insli r1, r1, hw0(_TIF_ALLWORK_MASK)
+	}
+	and     r1, r29, r1
+	beqzt   r1, .Lrestore_all
+
+	/*
+	 * Make sure we have all the registers saved for signal
+	 * handling or notify-resume.  Call out to C code to figure out
+	 * exactly what we need to do for each flag bit, then if
+	 * necessary, reload the flags and recheck.
+	 */
+	{
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	 bnez   r33, 1f
+	}
+	push_extra_callee_saves r0
+	movei   r33, 1
+1:	jal     do_work_pending
+	bnez    r0, .Lretry_work_pending
+
+	/*
+	 * In the NMI case we
+	 * omit the call to single_process_check_nohz, which normally checks
+	 * to see if we should start or stop the scheduler tick, because
+	 * we can't call arbitrary Linux code from an NMI context.
+	 * We always call the homecache TLB deferral code to re-trigger
+	 * the deferral mechanism.
+	 *
+	 * The other chunk of responsibility this code has is to reset the
+	 * interrupt masks appropriately to reset irqs and NMIs.  We have
+	 * to call TRACE_IRQS_OFF and TRACE_IRQS_ON to support all the
+	 * lockdep-type stuff, but we can't set ICS until afterwards, since
+	 * ICS can only be used in very tight chunks of code to avoid
+	 * tripping over various assertions that it is off.
+	 */
+.Lrestore_all:
+	PTREGS_PTR(r0, PTREGS_OFFSET_EX1)
+	{
+	 ld      r0, r0
+	 PTREGS_PTR(r32, PTREGS_OFFSET_FLAGS)
+	}
+	{
+	 IS_KERNEL_EX1(r0, r0)
+	 ld     r32, r32
+	}
+	bnez    r0, 1f
+	j       2f
+#if PT_FLAGS_DISABLE_IRQ != 1
+# error Assuming PT_FLAGS_DISABLE_IRQ == 1 so we can use blbct below
+#endif
+1:	blbct   r32, 2f
+	IRQ_DISABLE(r20,r21)
+	TRACE_IRQS_OFF
+	movei   r0, 1
+	mtspr   INTERRUPT_CRITICAL_SECTION, r0
+	beqzt   r30, .Lrestore_regs
+	j       3f
+2:	TRACE_IRQS_ON
+	IRQ_ENABLE_LOAD(r20, r21)
+	movei   r0, 1
+	mtspr   INTERRUPT_CRITICAL_SECTION, r0
+	IRQ_ENABLE_APPLY(r20, r21)
+	beqzt   r30, .Lrestore_regs
+3:
+
+#if INT_PERF_COUNT + 1 != INT_AUX_PERF_COUNT
+# error Bad interrupt assumption
+#endif
+	{
+	 movei  r0, 3   /* two adjacent bits for the PERF_COUNT mask */
+	 beqz   r31, .Lrestore_regs
+	}
+	shli    r0, r0, INT_PERF_COUNT
+	mtspr   SPR_INTERRUPT_MASK_RESET_K, r0
+
+	/*
+	 * We now commit to returning from this interrupt, since we will be
+	 * doing things like setting EX_CONTEXT SPRs and unwinding the stack
+	 * frame.  No calls should be made to any other code after this point.
+	 * This code should only be entered with ICS set.
+	 * r32 must still be set to ptregs.flags.
+	 * We launch loads to each cache line separately first, so we can
+	 * get some parallelism out of the memory subsystem.
+	 * We start zeroing caller-saved registers throughout, since
+	 * that will save some cycles if this turns out to be a syscall.
+	 */
+.Lrestore_regs:
+
+	/*
+	 * Rotate so we have one high bit and one low bit to test.
+	 * - low bit says whether to restore all the callee-saved registers,
+	 *   or just r30-r33, and r52 up.
+	 * - high bit (i.e. sign bit) says whether to restore all the
+	 *   caller-saved registers, or just r0.
+	 */
+#if PT_FLAGS_CALLER_SAVES != 2 || PT_FLAGS_RESTORE_REGS != 4
+# error Rotate trick does not work :-)
+#endif
+	{
+	 rotli  r20, r32, 62
+	 PTREGS_PTR(sp, PTREGS_OFFSET_REG(0))
+	}
+
+	/*
+	 * Load cache lines 0, 4, 6 and 7, in that order, then use
+	 * the last loaded value, which makes it likely that the other
+	 * cache lines have also loaded, at which point we should be
+	 * able to safely read all the remaining words on those cache
+	 * lines without waiting for the memory subsystem.
+	 */
+	pop_reg r0, sp, PTREGS_OFFSET_REG(30) - PTREGS_OFFSET_REG(0)
+	pop_reg r30, sp, PTREGS_OFFSET_REG(52) - PTREGS_OFFSET_REG(30)
+	pop_reg_zero r52, r3, sp, PTREGS_OFFSET_CMPEXCH - PTREGS_OFFSET_REG(52)
+	pop_reg_zero r21, r27, sp, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_CMPEXCH
+	pop_reg_zero lr, r2, sp, PTREGS_OFFSET_PC - PTREGS_OFFSET_EX1
+	{
+	 mtspr  CMPEXCH_VALUE, r21
+	 move   r4, zero
+	}
+	pop_reg r21, sp, PTREGS_OFFSET_REG(31) - PTREGS_OFFSET_PC
+	{
+	 mtspr  SPR_EX_CONTEXT_K_1, lr
+	 IS_KERNEL_EX1(lr, lr)
+	}
+	{
+	 mtspr  SPR_EX_CONTEXT_K_0, r21
+	 move   r5, zero
+	}
+
+	/* Restore callee-saveds that we actually use. */
+	pop_reg_zero r31, r6
+	pop_reg_zero r32, r7
+	pop_reg_zero r33, r8, sp, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(33)
+
+	/*
+	 * If we modified other callee-saveds, restore them now.
+	 * This is rare, but could be via ptrace or signal handler.
+	 */
+	{
+	 move   r9, zero
+	 blbs   r20, .Lrestore_callees
+	}
+.Lcontinue_restore_regs:
+
+	/* Check if we're returning from a syscall. */
+	{
+	 move   r10, zero
+	 bltzt  r20, 1f  /* no, so go restore callee-save registers */
+	}
+
+	/*
+	 * Check if we're returning to userspace.
+	 * Note that if we're not, we don't worry about zeroing everything.
+	 */
+	{
+	 addli  sp, sp, PTREGS_OFFSET_LR - PTREGS_OFFSET_REG(29)
+	 bnez   lr, .Lkernel_return
+	}
+
+	/*
+	 * On return from syscall, we've restored r0 from pt_regs, but we
+	 * clear the remainder of the caller-saved registers.  We could
+	 * restore the syscall arguments, but there's not much point,
+	 * and it ensures user programs aren't trying to use the
+	 * caller-saves if we clear them, as well as avoiding leaking
+	 * kernel pointers into userspace.
+	 */
+	pop_reg_zero lr, r11, sp, PTREGS_OFFSET_TP - PTREGS_OFFSET_LR
+	pop_reg_zero tp, r12, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_TP
+	{
+	 ld     sp, sp
+	 move   r13, zero
+	 move   r14, zero
+	}
+	{ move r15, zero; move r16, zero }
+	{ move r17, zero; move r18, zero }
+	{ move r19, zero; move r20, zero }
+	{ move r21, zero; move r22, zero }
+	{ move r23, zero; move r24, zero }
+	{ move r25, zero; move r26, zero }
+
+	/* Set r1 to errno if we are returning an error, otherwise zero. */
+	{
+	 moveli r29, 4096
+	 sub    r1, zero, r0
+	}
+	{
+	 move   r28, zero
+	 cmpltu r29, r1, r29
+	}
+	{
+	 mnz    r1, r29, r1
+	 move   r29, zero
+	}
+	iret
+
+	/*
+	 * Not a syscall, so restore caller-saved registers.
+	 * First kick off loads for cache lines 1-3, which we're touching
+	 * for the first time here.
+	 */
+	.align 64
+1:	pop_reg r29, sp, PTREGS_OFFSET_REG(21) - PTREGS_OFFSET_REG(29)
+	pop_reg r21, sp, PTREGS_OFFSET_REG(13) - PTREGS_OFFSET_REG(21)
+	pop_reg r13, sp, PTREGS_OFFSET_REG(1) - PTREGS_OFFSET_REG(13)
+	pop_reg r1
+	pop_reg r2
+	pop_reg r3
+	pop_reg r4
+	pop_reg r5
+	pop_reg r6
+	pop_reg r7
+	pop_reg r8
+	pop_reg r9
+	pop_reg r10
+	pop_reg r11
+	pop_reg r12, sp, 16
+	/* r13 already restored above */
+	pop_reg r14
+	pop_reg r15
+	pop_reg r16
+	pop_reg r17
+	pop_reg r18
+	pop_reg r19
+	pop_reg r20, sp, 16
+	/* r21 already restored above */
+	pop_reg r22
+	pop_reg r23
+	pop_reg r24
+	pop_reg r25
+	pop_reg r26
+	pop_reg r27
+	pop_reg r28, sp, PTREGS_OFFSET_LR - PTREGS_OFFSET_REG(28)
+	/* r29 already restored above */
+	bnez    lr, .Lkernel_return
+	pop_reg lr, sp, PTREGS_OFFSET_TP - PTREGS_OFFSET_LR
+	pop_reg tp, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_TP
+	ld      sp, sp
+	iret
+
+	/*
+	 * We can't restore tp when in kernel mode, since a thread might
+	 * have migrated from another cpu and brought a stale tp value.
+	 */
+.Lkernel_return:
+	pop_reg lr, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_LR
+	ld      sp, sp
+	iret
+
+	/* Restore callee-saved registers from r34 to r51. */
+.Lrestore_callees:
+	addli  sp, sp, PTREGS_OFFSET_REG(34) - PTREGS_OFFSET_REG(29)
+	pop_reg r34
+	pop_reg r35
+	pop_reg r36
+	pop_reg r37
+	pop_reg r38
+	pop_reg r39
+	pop_reg r40
+	pop_reg r41
+	pop_reg r42
+	pop_reg r43
+	pop_reg r44
+	pop_reg r45
+	pop_reg r46
+	pop_reg r47
+	pop_reg r48
+	pop_reg r49
+	pop_reg r50
+	pop_reg r51, sp, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(51)
+	j .Lcontinue_restore_regs
+	STD_ENDPROC(interrupt_return)
+
+	/*
+	 * "NMI" interrupts mask ALL interrupts before calling the
+	 * handler, and don't check thread flags, etc., on the way
+	 * back out.  In general, the only things we do here for NMIs
+	 * are register save/restore and dataplane kernel-TLB management.
+	 * We don't (for example) deal with start/stop of the sched tick.
+	 */
+	.pushsection .text.handle_nmi,"ax"
+handle_nmi:
+	finish_interrupt_save handle_nmi
+	{
+	 jalr   r0
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	}
+	FEEDBACK_REENTER(handle_nmi)
+	{
+	 movei  r30, 1
+	 cmpeq  r31, r0, zero
+	}
+	j       interrupt_return
+	STD_ENDPROC(handle_nmi)
+
+	/*
+	 * Parallel code for syscalls to handle_interrupt.
+	 */
+	.pushsection .text.handle_syscall,"ax"
+handle_syscall:
+	finish_interrupt_save handle_syscall
+
+	/* Enable irqs. */
+	TRACE_IRQS_ON
+	IRQ_ENABLE(r20, r21)
+
+	/* Bump the counter for syscalls made on this tile. */
+	moveli r20, hw2_last(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
+	shl16insli r20, r20, hw1(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
+	shl16insli r20, r20, hw0(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
+	add     r20, r20, tp
+	ld4s    r21, r20
+	{
+	 addi   r21, r21, 1
+	 move   r31, sp
+	}
+	{
+	 st4    r20, r21
+	 EXTRACT_THREAD_INFO(r31)
+	}
+
+	/* Trace syscalls, if requested. */
+	addi	r31, r31, THREAD_INFO_FLAGS_OFFSET
+	{
+	 ld     r30, r31
+	 moveli r32, _TIF_SYSCALL_ENTRY_WORK
+	}
+	and     r30, r30, r32
+	{
+	 addi   r30, r31, THREAD_INFO_STATUS_OFFSET - THREAD_INFO_FLAGS_OFFSET
+	 beqzt	r30, .Lrestore_syscall_regs
+	}
+	{
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	 jal    do_syscall_trace_enter
+	}
+	FEEDBACK_REENTER(handle_syscall)
+
+	/*
+	 * We always reload our registers from the stack at this
+	 * point.  They might be valid, if we didn't build with
+	 * TRACE_IRQFLAGS, and this isn't a dataplane tile, and we're not
+	 * doing syscall tracing, but there are enough cases now that it
+	 * seems simplest just to do the reload unconditionally.
+	 */
+.Lrestore_syscall_regs:
+	{
+	 ld     r30, r30
+	 PTREGS_PTR(r11, PTREGS_OFFSET_REG(0))
+	}
+	pop_reg r0,  r11
+	pop_reg r1,  r11
+	pop_reg r2,  r11
+	pop_reg r3,  r11
+	pop_reg r4,  r11
+	pop_reg r5,  r11, PTREGS_OFFSET_SYSCALL - PTREGS_OFFSET_REG(5)
+	{
+	 ld     TREG_SYSCALL_NR_NAME, r11
+	 moveli r21, __NR_syscalls
+	}
+
+	/* Ensure that the syscall number is within the legal range. */
+	{
+	 moveli r20, hw2(sys_call_table)
+#ifdef CONFIG_COMPAT
+	 blbs   r30, .Lcompat_syscall
+#endif
+	}
+	{
+	 cmpltu r21, TREG_SYSCALL_NR_NAME, r21
+	 shl16insli r20, r20, hw1(sys_call_table)
+	}
+	{
+	 blbc   r21, .Linvalid_syscall
+	 shl16insli r20, r20, hw0(sys_call_table)
+	}
+.Lload_syscall_pointer:
+	shl3add r20, TREG_SYSCALL_NR_NAME, r20
+	ld      r20, r20
+
+	/* Jump to syscall handler. */
+	jalr    r20
+.Lhandle_syscall_link: /* value of "lr" after "jalr r20" above */
+
+	/*
+	 * Write our r0 onto the stack so it gets restored instead
+	 * of whatever the user had there before.
+	 * In compat mode, sign-extend r0 before storing it.
+	 */
+	{
+	 PTREGS_PTR(r29, PTREGS_OFFSET_REG(0))
+	 blbct  r30, 1f
+	}
+	addxi   r0, r0, 0
+1:	st      r29, r0
+
+.Lsyscall_sigreturn_skip:
+	FEEDBACK_REENTER(handle_syscall)
+
+	/* Do syscall trace again, if requested. */
+	{
+	 ld      r30, r31
+	 moveli  r32, _TIF_SYSCALL_EXIT_WORK
+	}
+	and      r0, r30, r32
+	{
+	 andi    r0, r30, _TIF_SINGLESTEP
+	 beqzt   r0, 1f
+	}
+	{
+	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
+	 jal    do_syscall_trace_exit
+	}
+	FEEDBACK_REENTER(handle_syscall)
+	andi    r0, r30, _TIF_SINGLESTEP
+
+1:	beqzt	r0, 2f
+
+	/* Single stepping -- notify ptrace. */
+	{
+	 movei   r0, SIGTRAP
+	 jal     ptrace_notify
+	}
+	FEEDBACK_REENTER(handle_syscall)
+
+2:	{
+	 movei  r30, 0               /* not an NMI */
+	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
+	}
+
+#ifdef CONFIG_COMPAT
+.Lcompat_syscall:
+	/*
+	 * Load the base of the compat syscall table in r20, and
+	 * range-check the syscall number (duplicated from 64-bit path).
+	 * Sign-extend all the user's passed arguments to make them consistent.
+	 * Also save the original "r(n)" values away in "r(11+n)" in
+	 * case the syscall table entry wants to validate them.
+	 */
+	moveli  r20, hw2(compat_sys_call_table)
+	{
+	 cmpltu r21, TREG_SYSCALL_NR_NAME, r21
+	 shl16insli r20, r20, hw1(compat_sys_call_table)
+	}
+	{
+	 blbc   r21, .Linvalid_syscall
+	 shl16insli r20, r20, hw0(compat_sys_call_table)
+	}
+	{ move r11, r0; addxi r0, r0, 0 }
+	{ move r12, r1; addxi r1, r1, 0 }
+	{ move r13, r2; addxi r2, r2, 0 }
+	{ move r14, r3; addxi r3, r3, 0 }
+	{ move r15, r4; addxi r4, r4, 0 }
+	{ move r16, r5; addxi r5, r5, 0 }
+	j .Lload_syscall_pointer
+#endif
+
+.Linvalid_syscall:
+	/* Report an invalid syscall back to the user program */
+	{
+	 PTREGS_PTR(r29, PTREGS_OFFSET_REG(0))
+	 movei  r28, -ENOSYS
+	}
+	st      r29, r28
+	{
+	 movei  r30, 0               /* not an NMI */
+	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
+	}
+	STD_ENDPROC(handle_syscall)
+
+	/* Return the address for oprofile to suppress in backtraces. */
+STD_ENTRY_SECTION(handle_syscall_link_address, .text.handle_syscall)
+	lnk     r0
+	{
+	 addli  r0, r0, .Lhandle_syscall_link - .
+	 jrp    lr
+	}
+	STD_ENDPROC(handle_syscall_link_address)
+
+STD_ENTRY(ret_from_fork)
+	jal     sim_notify_fork
+	jal     schedule_tail
+	FEEDBACK_REENTER(ret_from_fork)
+	{
+	 movei  r30, 0               /* not an NMI */
+	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
+	}
+	STD_ENDPROC(ret_from_fork)
+
+STD_ENTRY(ret_from_kernel_thread)
+	jal     sim_notify_fork
+	jal     schedule_tail
+	FEEDBACK_REENTER(ret_from_fork)
+	{
+	 move   r0, r31
+	 jalr   r30
+	}
+	FEEDBACK_REENTER(ret_from_kernel_thread)
+	{
+	 movei  r30, 0               /* not an NMI */
+	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
+	}
+	STD_ENDPROC(ret_from_kernel_thread)
+
+/* Various stub interrupt handlers and syscall handlers */
+
+STD_ENTRY_LOCAL(_kernel_double_fault)
+	mfspr   r1, SPR_EX_CONTEXT_K_0
+	move    r2, lr
+	move    r3, sp
+	move    r4, r52
+	addi    sp, sp, -C_ABI_SAVE_AREA_SIZE
+	j       kernel_double_fault
+	STD_ENDPROC(_kernel_double_fault)
+
+STD_ENTRY_LOCAL(bad_intr)
+	mfspr   r2, SPR_EX_CONTEXT_K_0
+	panic   "Unhandled interrupt %#x: PC %#lx"
+	STD_ENDPROC(bad_intr)
+
+/*
+ * Special-case sigreturn to not write r0 to the stack on return.
+ * This is technically more efficient, but it also avoids difficulties
+ * in the 64-bit OS when handling 32-bit compat code, since we must not
+ * sign-extend r0 for the sigreturn return-value case.
+ */
+#define PTREGS_SYSCALL_SIGRETURN(x, reg)                \
+	STD_ENTRY(_##x);                                \
+	addli   lr, lr, .Lsyscall_sigreturn_skip - .Lhandle_syscall_link; \
+	{                                               \
+	 PTREGS_PTR(reg, PTREGS_OFFSET_BASE);           \
+	 j      x                                       \
+	};                                              \
+	STD_ENDPROC(_##x)
+
+PTREGS_SYSCALL_SIGRETURN(sys_rt_sigreturn, r0)
+#ifdef CONFIG_COMPAT
+PTREGS_SYSCALL_SIGRETURN(compat_sys_rt_sigreturn, r0)
+#endif
+
+/* Save additional callee-saves to pt_regs and jump to standard function. */
+STD_ENTRY(_sys_clone)
+	push_extra_callee_saves r4
+	j       sys_clone
+	STD_ENDPROC(_sys_clone)
+
+	/*
+	 * Recover r3, r2, r1 and r0 here saved by unalign fast vector.
+	 * The vector area limit is 32 bundles, so we handle the reload here.
+	 * r0, r1, r2 are in thread_info from low to high memory in order.
+	 * r3 points to location the original r3 was saved.
+	 * We put this code in the __HEAD section so it can be reached
+	 * via a conditional branch from the fast path.
+	 */
+	__HEAD
+hand_unalign_slow:
+	andi    sp, sp, ~1
+hand_unalign_slow_badsp:
+	addi    r3, r3, -(3 * 8)
+	ld_add  r0, r3, 8
+	ld_add  r1, r3, 8
+	ld      r2, r3
+hand_unalign_slow_nonuser:
+	mfspr   r3, SPR_SYSTEM_SAVE_K_1
+	__int_hand     INT_UNALIGN_DATA, UNALIGN_DATA_SLOW, int_unalign
+
+/* The unaligned data support needs to read all the registers. */
+int_unalign:
+	push_extra_callee_saves r0
+	j       do_unaligned
+ENDPROC(hand_unalign_slow)
+
+/* Fill the return address stack with nonzero entries. */
+STD_ENTRY(fill_ra_stack)
+	{
+	 move	r0, lr
+	 jal	1f
+	}
+1:	jal	2f
+2:	jal	3f
+3:	jal	4f
+4:	jrp	r0
+	STD_ENDPROC(fill_ra_stack)
+
+	.macro int_hand  vecnum, vecname, c_routine, processing=handle_interrupt
+	.org   (\vecnum << 8)
+		__int_hand   \vecnum, \vecname, \c_routine, \processing
+	.endm
+
+/* Include .intrpt array of interrupt vectors */
+	.section ".intrpt", "ax"
+	.global intrpt_start
+intrpt_start:
+
+#ifndef CONFIG_USE_PMC
+#define handle_perf_interrupt bad_intr
+#endif
+
+#ifndef CONFIG_HARDWALL
+#define do_hardwall_trap bad_intr
+#endif
+
+	int_hand     INT_MEM_ERROR, MEM_ERROR, do_trap
+	int_hand     INT_SINGLE_STEP_3, SINGLE_STEP_3, bad_intr
+#if CONFIG_KERNEL_PL == 2
+	int_hand     INT_SINGLE_STEP_2, SINGLE_STEP_2, gx_singlestep_handle
+	int_hand     INT_SINGLE_STEP_1, SINGLE_STEP_1, bad_intr
+#else
+	int_hand     INT_SINGLE_STEP_2, SINGLE_STEP_2, bad_intr
+	int_hand     INT_SINGLE_STEP_1, SINGLE_STEP_1, gx_singlestep_handle
+#endif
+	int_hand     INT_SINGLE_STEP_0, SINGLE_STEP_0, bad_intr
+	int_hand     INT_IDN_COMPLETE, IDN_COMPLETE, bad_intr
+	int_hand     INT_UDN_COMPLETE, UDN_COMPLETE, bad_intr
+	int_hand     INT_ITLB_MISS, ITLB_MISS, do_page_fault
+	int_hand     INT_ILL, ILL, do_trap
+	int_hand     INT_GPV, GPV, do_trap
+	int_hand     INT_IDN_ACCESS, IDN_ACCESS, do_trap
+	int_hand     INT_UDN_ACCESS, UDN_ACCESS, do_trap
+	int_hand     INT_SWINT_3, SWINT_3, do_trap
+	int_hand     INT_SWINT_2, SWINT_2, do_trap
+	int_hand     INT_SWINT_1, SWINT_1, SYSCALL, handle_syscall
+	int_hand     INT_SWINT_0, SWINT_0, do_trap
+	int_hand     INT_ILL_TRANS, ILL_TRANS, do_trap
+	int_hand_unalign_fast INT_UNALIGN_DATA, UNALIGN_DATA
+	int_hand     INT_DTLB_MISS, DTLB_MISS, do_page_fault
+	int_hand     INT_DTLB_ACCESS, DTLB_ACCESS, do_page_fault
+	int_hand     INT_IDN_FIREWALL, IDN_FIREWALL, do_hardwall_trap
+	int_hand     INT_UDN_FIREWALL, UDN_FIREWALL, do_hardwall_trap
+	int_hand     INT_TILE_TIMER, TILE_TIMER, do_timer_interrupt
+	int_hand     INT_IDN_TIMER, IDN_TIMER, bad_intr
+	int_hand     INT_UDN_TIMER, UDN_TIMER, bad_intr
+	int_hand     INT_IDN_AVAIL, IDN_AVAIL, bad_intr
+	int_hand     INT_UDN_AVAIL, UDN_AVAIL, bad_intr
+	int_hand     INT_IPI_3, IPI_3, bad_intr
+#if CONFIG_KERNEL_PL == 2
+	int_hand     INT_IPI_2, IPI_2, tile_dev_intr
+	int_hand     INT_IPI_1, IPI_1, bad_intr
+#else
+	int_hand     INT_IPI_2, IPI_2, bad_intr
+	int_hand     INT_IPI_1, IPI_1, tile_dev_intr
+#endif
+	int_hand     INT_IPI_0, IPI_0, bad_intr
+	int_hand     INT_PERF_COUNT, PERF_COUNT, \
+		     handle_perf_interrupt, handle_nmi
+	int_hand     INT_AUX_PERF_COUNT, AUX_PERF_COUNT, \
+		     handle_perf_interrupt, handle_nmi
+	int_hand     INT_INTCTRL_3, INTCTRL_3, bad_intr
+#if CONFIG_KERNEL_PL == 2
+	dc_dispatch  INT_INTCTRL_2, INTCTRL_2
+	int_hand     INT_INTCTRL_1, INTCTRL_1, bad_intr
+#else
+	int_hand     INT_INTCTRL_2, INTCTRL_2, bad_intr
+	dc_dispatch  INT_INTCTRL_1, INTCTRL_1
+#endif
+	int_hand     INT_INTCTRL_0, INTCTRL_0, bad_intr
+	int_hand     INT_MESSAGE_RCV_DWNCL, MESSAGE_RCV_DWNCL, \
+		     hv_message_intr
+	int_hand     INT_DEV_INTR_DWNCL, DEV_INTR_DWNCL, bad_intr
+	int_hand     INT_I_ASID, I_ASID, bad_intr
+	int_hand     INT_D_ASID, D_ASID, bad_intr
+	int_hand     INT_DOUBLE_FAULT, DOUBLE_FAULT, do_trap
+
+	/* Synthetic interrupt delivered only by the simulator */
+	int_hand     INT_BREAKPOINT, BREAKPOINT, do_breakpoint
diff --git a/arch/tile/kernel/irq.c b/arch/tile/kernel/irq.c
new file mode 100644
index 000000000..22044fc69
--- /dev/null
+++ b/arch/tile/kernel/irq.c
@@ -0,0 +1,280 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/module.h>
+#include <linux/seq_file.h>
+#include <linux/interrupt.h>
+#include <linux/irq.h>
+#include <linux/kernel_stat.h>
+#include <linux/uaccess.h>
+#include <hv/drv_pcie_rc_intf.h>
+#include <arch/spr_def.h>
+#include <asm/traps.h>
+#include <linux/perf_event.h>
+
+/* Bit-flag stored in irq_desc->chip_data to indicate HW-cleared irqs. */
+#define IS_HW_CLEARED 1
+
+/*
+ * The set of interrupts we enable for arch_local_irq_enable().
+ * This is initialized to have just a single interrupt that the kernel
+ * doesn't actually use as a sentinel.  During kernel init,
+ * interrupts are added as the kernel gets prepared to support them.
+ * NOTE: we could probably initialize them all statically up front.
+ */
+DEFINE_PER_CPU(unsigned long long, interrupts_enabled_mask) =
+  INITIAL_INTERRUPTS_ENABLED;
+EXPORT_PER_CPU_SYMBOL(interrupts_enabled_mask);
+
+/* Define per-tile device interrupt statistics state. */
+DEFINE_PER_CPU(irq_cpustat_t, irq_stat) ____cacheline_internodealigned_in_smp;
+EXPORT_PER_CPU_SYMBOL(irq_stat);
+
+/*
+ * Define per-tile irq disable mask; the hardware/HV only has a single
+ * mask that we use to implement both masking and disabling.
+ */
+static DEFINE_PER_CPU(unsigned long, irq_disable_mask)
+	____cacheline_internodealigned_in_smp;
+
+/*
+ * Per-tile IRQ nesting depth.  Used to make sure we enable newly
+ * enabled IRQs before exiting the outermost interrupt.
+ */
+static DEFINE_PER_CPU(int, irq_depth);
+
+#if CHIP_HAS_IPI()
+/* Use SPRs to manipulate device interrupts. */
+#define mask_irqs(irq_mask) __insn_mtspr(SPR_IPI_MASK_SET_K, irq_mask)
+#define unmask_irqs(irq_mask) __insn_mtspr(SPR_IPI_MASK_RESET_K, irq_mask)
+#define clear_irqs(irq_mask) __insn_mtspr(SPR_IPI_EVENT_RESET_K, irq_mask)
+#else
+/* Use HV to manipulate device interrupts. */
+#define mask_irqs(irq_mask) hv_disable_intr(irq_mask)
+#define unmask_irqs(irq_mask) hv_enable_intr(irq_mask)
+#define clear_irqs(irq_mask) hv_clear_intr(irq_mask)
+#endif
+
+/*
+ * The interrupt handling path, implemented in terms of HV interrupt
+ * emulation on TILEPro, and IPI hardware on TILE-Gx.
+ * Entered with interrupts disabled.
+ */
+void tile_dev_intr(struct pt_regs *regs, int intnum)
+{
+	int depth = __this_cpu_inc_return(irq_depth);
+	unsigned long original_irqs;
+	unsigned long remaining_irqs;
+	struct pt_regs *old_regs;
+
+#if CHIP_HAS_IPI()
+	/*
+	 * Pending interrupts are listed in an SPR.  We might be
+	 * nested, so be sure to only handle irqs that weren't already
+	 * masked by a previous interrupt.  Then, mask out the ones
+	 * we're going to handle.
+	 */
+	unsigned long masked = __insn_mfspr(SPR_IPI_MASK_K);
+	original_irqs = __insn_mfspr(SPR_IPI_EVENT_K) & ~masked;
+	__insn_mtspr(SPR_IPI_MASK_SET_K, original_irqs);
+#else
+	/*
+	 * Hypervisor performs the equivalent of the Gx code above and
+	 * then puts the pending interrupt mask into a system save reg
+	 * for us to find.
+	 */
+	original_irqs = __insn_mfspr(SPR_SYSTEM_SAVE_K_3);
+#endif
+	remaining_irqs = original_irqs;
+
+	/* Track time spent here in an interrupt context. */
+	old_regs = set_irq_regs(regs);
+	irq_enter();
+
+#ifdef CONFIG_DEBUG_STACKOVERFLOW
+	/* Debugging check for stack overflow: less than 1/8th stack free? */
+	{
+		long sp = stack_pointer - (long) current_thread_info();
+		if (unlikely(sp < (sizeof(struct thread_info) + STACK_WARN))) {
+			pr_emerg("%s: stack overflow: %ld\n",
+				 __func__, sp - sizeof(struct thread_info));
+			dump_stack();
+		}
+	}
+#endif
+	while (remaining_irqs) {
+		unsigned long irq = __ffs(remaining_irqs);
+		remaining_irqs &= ~(1UL << irq);
+
+		/* Count device irqs; Linux IPIs are counted elsewhere. */
+		if (irq != IRQ_RESCHEDULE)
+			__this_cpu_inc(irq_stat.irq_dev_intr_count);
+
+		generic_handle_irq(irq);
+	}
+
+	/*
+	 * If we weren't nested, turn on all enabled interrupts,
+	 * including any that were reenabled during interrupt
+	 * handling.
+	 */
+	if (depth == 1)
+		unmask_irqs(~__this_cpu_read(irq_disable_mask));
+
+	__this_cpu_dec(irq_depth);
+
+	/*
+	 * Track time spent against the current process again and
+	 * process any softirqs if they are waiting.
+	 */
+	irq_exit();
+	set_irq_regs(old_regs);
+}
+
+
+/*
+ * Remove an irq from the disabled mask.  If we're in an interrupt
+ * context, defer enabling the HW interrupt until we leave.
+ */
+static void tile_irq_chip_enable(struct irq_data *d)
+{
+	get_cpu_var(irq_disable_mask) &= ~(1UL << d->irq);
+	if (__this_cpu_read(irq_depth) == 0)
+		unmask_irqs(1UL << d->irq);
+	put_cpu_var(irq_disable_mask);
+}
+
+/*
+ * Add an irq to the disabled mask.  We disable the HW interrupt
+ * immediately so that there's no possibility of it firing.  If we're
+ * in an interrupt context, the return path is careful to avoid
+ * unmasking a newly disabled interrupt.
+ */
+static void tile_irq_chip_disable(struct irq_data *d)
+{
+	get_cpu_var(irq_disable_mask) |= (1UL << d->irq);
+	mask_irqs(1UL << d->irq);
+	put_cpu_var(irq_disable_mask);
+}
+
+/* Mask an interrupt. */
+static void tile_irq_chip_mask(struct irq_data *d)
+{
+	mask_irqs(1UL << d->irq);
+}
+
+/* Unmask an interrupt. */
+static void tile_irq_chip_unmask(struct irq_data *d)
+{
+	unmask_irqs(1UL << d->irq);
+}
+
+/*
+ * Clear an interrupt before processing it so that any new assertions
+ * will trigger another irq.
+ */
+static void tile_irq_chip_ack(struct irq_data *d)
+{
+	if ((unsigned long)irq_data_get_irq_chip_data(d) != IS_HW_CLEARED)
+		clear_irqs(1UL << d->irq);
+}
+
+/*
+ * For per-cpu interrupts, we need to avoid unmasking any interrupts
+ * that we disabled via disable_percpu_irq().
+ */
+static void tile_irq_chip_eoi(struct irq_data *d)
+{
+	if (!(__this_cpu_read(irq_disable_mask) & (1UL << d->irq)))
+		unmask_irqs(1UL << d->irq);
+}
+
+static struct irq_chip tile_irq_chip = {
+	.name = "tile_irq_chip",
+	.irq_enable = tile_irq_chip_enable,
+	.irq_disable = tile_irq_chip_disable,
+	.irq_ack = tile_irq_chip_ack,
+	.irq_eoi = tile_irq_chip_eoi,
+	.irq_mask = tile_irq_chip_mask,
+	.irq_unmask = tile_irq_chip_unmask,
+};
+
+void __init init_IRQ(void)
+{
+	ipi_init();
+}
+
+void setup_irq_regs(void)
+{
+	/* Enable interrupt delivery. */
+	unmask_irqs(~0UL);
+#if CHIP_HAS_IPI()
+	arch_local_irq_unmask(INT_IPI_K);
+#endif
+}
+
+void tile_irq_activate(unsigned int irq, int tile_irq_type)
+{
+	/*
+	 * We use handle_level_irq() by default because the pending
+	 * interrupt vector (whether modeled by the HV on
+	 * TILEPro or implemented in hardware on TILE-Gx) has
+	 * level-style semantics for each bit.  An interrupt fires
+	 * whenever a bit is high, not just at edges.
+	 */
+	irq_flow_handler_t handle = handle_level_irq;
+	if (tile_irq_type == TILE_IRQ_PERCPU)
+		handle = handle_percpu_irq;
+	irq_set_chip_and_handler(irq, &tile_irq_chip, handle);
+
+	/*
+	 * Flag interrupts that are hardware-cleared so that ack()
+	 * won't clear them.
+	 */
+	if (tile_irq_type == TILE_IRQ_HW_CLEAR)
+		irq_set_chip_data(irq, (void *)IS_HW_CLEARED);
+}
+EXPORT_SYMBOL(tile_irq_activate);
+
+
+void ack_bad_irq(unsigned int irq)
+{
+	pr_err("unexpected IRQ trap at vector %02x\n", irq);
+}
+
+/*
+ * /proc/interrupts printing:
+ */
+int arch_show_interrupts(struct seq_file *p, int prec)
+{
+#ifdef CONFIG_PERF_EVENTS
+	int i;
+
+	seq_printf(p, "%*s: ", prec, "PMI");
+
+	for_each_online_cpu(i)
+		seq_printf(p, "%10llu ", per_cpu(perf_irqs, i));
+	seq_puts(p, "  perf_events\n");
+#endif
+	return 0;
+}
+
+#if CHIP_HAS_IPI()
+int arch_setup_hwirq(unsigned int irq, int node)
+{
+	return irq >= NR_IRQS ? -EINVAL : 0;
+}
+
+void arch_teardown_hwirq(unsigned int irq) { }
+#endif
diff --git a/arch/tile/kernel/kgdb.c b/arch/tile/kernel/kgdb.c
new file mode 100644
index 000000000..ff5335ae0
--- /dev/null
+++ b/arch/tile/kernel/kgdb.c
@@ -0,0 +1,495 @@
+/*
+ * Copyright 2013 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * TILE-Gx KGDB support.
+ */
+
+#include <linux/ptrace.h>
+#include <linux/kgdb.h>
+#include <linux/kdebug.h>
+#include <linux/uaccess.h>
+#include <linux/module.h>
+#include <asm/cacheflush.h>
+
+static tile_bundle_bits singlestep_insn = TILEGX_BPT_BUNDLE | DIE_SSTEPBP;
+static unsigned long stepped_addr;
+static tile_bundle_bits stepped_instr;
+
+struct dbg_reg_def_t dbg_reg_def[DBG_MAX_REG_NUM] = {
+	{ "r0", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[0])},
+	{ "r1", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[1])},
+	{ "r2", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[2])},
+	{ "r3", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[3])},
+	{ "r4", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[4])},
+	{ "r5", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[5])},
+	{ "r6", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[6])},
+	{ "r7", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[7])},
+	{ "r8", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[8])},
+	{ "r9", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[9])},
+	{ "r10", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[10])},
+	{ "r11", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[11])},
+	{ "r12", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[12])},
+	{ "r13", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[13])},
+	{ "r14", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[14])},
+	{ "r15", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[15])},
+	{ "r16", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[16])},
+	{ "r17", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[17])},
+	{ "r18", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[18])},
+	{ "r19", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[19])},
+	{ "r20", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[20])},
+	{ "r21", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[21])},
+	{ "r22", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[22])},
+	{ "r23", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[23])},
+	{ "r24", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[24])},
+	{ "r25", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[25])},
+	{ "r26", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[26])},
+	{ "r27", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[27])},
+	{ "r28", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[28])},
+	{ "r29", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[29])},
+	{ "r30", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[30])},
+	{ "r31", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[31])},
+	{ "r32", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[32])},
+	{ "r33", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[33])},
+	{ "r34", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[34])},
+	{ "r35", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[35])},
+	{ "r36", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[36])},
+	{ "r37", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[37])},
+	{ "r38", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[38])},
+	{ "r39", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[39])},
+	{ "r40", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[40])},
+	{ "r41", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[41])},
+	{ "r42", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[42])},
+	{ "r43", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[43])},
+	{ "r44", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[44])},
+	{ "r45", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[45])},
+	{ "r46", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[46])},
+	{ "r47", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[47])},
+	{ "r48", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[48])},
+	{ "r49", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[49])},
+	{ "r50", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[50])},
+	{ "r51", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[51])},
+	{ "r52", GDB_SIZEOF_REG, offsetof(struct pt_regs, regs[52])},
+	{ "tp", GDB_SIZEOF_REG, offsetof(struct pt_regs, tp)},
+	{ "sp", GDB_SIZEOF_REG, offsetof(struct pt_regs, sp)},
+	{ "lr", GDB_SIZEOF_REG, offsetof(struct pt_regs, lr)},
+	{ "sn", GDB_SIZEOF_REG, -1},
+	{ "idn0", GDB_SIZEOF_REG, -1},
+	{ "idn1", GDB_SIZEOF_REG, -1},
+	{ "udn0", GDB_SIZEOF_REG, -1},
+	{ "udn1", GDB_SIZEOF_REG, -1},
+	{ "udn2", GDB_SIZEOF_REG, -1},
+	{ "udn3", GDB_SIZEOF_REG, -1},
+	{ "zero", GDB_SIZEOF_REG, -1},
+	{ "pc", GDB_SIZEOF_REG, offsetof(struct pt_regs, pc)},
+	{ "faultnum", GDB_SIZEOF_REG, offsetof(struct pt_regs, faultnum)},
+};
+
+char *dbg_get_reg(int regno, void *mem, struct pt_regs *regs)
+{
+	if (regno >= DBG_MAX_REG_NUM || regno < 0)
+		return NULL;
+
+	if (dbg_reg_def[regno].offset != -1)
+		memcpy(mem, (void *)regs + dbg_reg_def[regno].offset,
+		       dbg_reg_def[regno].size);
+	else
+		memset(mem, 0, dbg_reg_def[regno].size);
+	return dbg_reg_def[regno].name;
+}
+
+int dbg_set_reg(int regno, void *mem, struct pt_regs *regs)
+{
+	if (regno >= DBG_MAX_REG_NUM || regno < 0)
+		return -EINVAL;
+
+	if (dbg_reg_def[regno].offset != -1)
+		memcpy((void *)regs + dbg_reg_def[regno].offset, mem,
+		       dbg_reg_def[regno].size);
+	return 0;
+}
+
+/*
+ * Similar to pt_regs_to_gdb_regs() except that process is sleeping and so
+ * we may not be able to get all the info.
+ */
+void
+sleeping_thread_to_gdb_regs(unsigned long *gdb_regs, struct task_struct *task)
+{
+	struct pt_regs *thread_regs;
+
+	if (task == NULL)
+		return;
+
+	/* Initialize to zero. */
+	memset(gdb_regs, 0, NUMREGBYTES);
+
+	thread_regs = task_pt_regs(task);
+	memcpy(gdb_regs, thread_regs, TREG_LAST_GPR * sizeof(unsigned long));
+	gdb_regs[TILEGX_PC_REGNUM] = thread_regs->pc;
+	gdb_regs[TILEGX_FAULTNUM_REGNUM] = thread_regs->faultnum;
+}
+
+void kgdb_arch_set_pc(struct pt_regs *regs, unsigned long pc)
+{
+	regs->pc = pc;
+}
+
+static void kgdb_call_nmi_hook(void *ignored)
+{
+	kgdb_nmicallback(raw_smp_processor_id(), NULL);
+}
+
+void kgdb_roundup_cpus(unsigned long flags)
+{
+	local_irq_enable();
+	smp_call_function(kgdb_call_nmi_hook, NULL, 0);
+	local_irq_disable();
+}
+
+/*
+ * Convert a kernel address to the writable kernel text mapping.
+ */
+static unsigned long writable_address(unsigned long addr)
+{
+	unsigned long ret = 0;
+
+	if (core_kernel_text(addr))
+		ret = addr - MEM_SV_START + PAGE_OFFSET;
+	else if (is_module_text_address(addr))
+		ret = addr;
+	else
+		pr_err("Unknown virtual address 0x%lx\n", addr);
+
+	return ret;
+}
+
+/*
+ * Calculate the new address for after a step.
+ */
+static unsigned long get_step_address(struct pt_regs *regs)
+{
+	int src_reg;
+	int jump_off;
+	int br_off;
+	unsigned long addr;
+	unsigned int opcode;
+	tile_bundle_bits bundle;
+
+	/* Move to the next instruction by default. */
+	addr = regs->pc + TILEGX_BUNDLE_SIZE_IN_BYTES;
+	bundle = *(unsigned long *)instruction_pointer(regs);
+
+	/* 0: X mode, Otherwise: Y mode. */
+	if (bundle & TILEGX_BUNDLE_MODE_MASK) {
+		if (get_Opcode_Y1(bundle) == RRR_1_OPCODE_Y1 &&
+		    get_RRROpcodeExtension_Y1(bundle) ==
+		    UNARY_RRR_1_OPCODE_Y1) {
+			opcode = get_UnaryOpcodeExtension_Y1(bundle);
+
+			switch (opcode) {
+			case JALR_UNARY_OPCODE_Y1:
+			case JALRP_UNARY_OPCODE_Y1:
+			case JR_UNARY_OPCODE_Y1:
+			case JRP_UNARY_OPCODE_Y1:
+				src_reg = get_SrcA_Y1(bundle);
+				dbg_get_reg(src_reg, &addr, regs);
+				break;
+			}
+		}
+	} else if (get_Opcode_X1(bundle) == RRR_0_OPCODE_X1) {
+		if (get_RRROpcodeExtension_X1(bundle) ==
+		    UNARY_RRR_0_OPCODE_X1) {
+			opcode = get_UnaryOpcodeExtension_X1(bundle);
+
+			switch (opcode) {
+			case JALR_UNARY_OPCODE_X1:
+			case JALRP_UNARY_OPCODE_X1:
+			case JR_UNARY_OPCODE_X1:
+			case JRP_UNARY_OPCODE_X1:
+				src_reg = get_SrcA_X1(bundle);
+				dbg_get_reg(src_reg, &addr, regs);
+				break;
+			}
+		}
+	} else if (get_Opcode_X1(bundle) == JUMP_OPCODE_X1) {
+		opcode = get_JumpOpcodeExtension_X1(bundle);
+
+		switch (opcode) {
+		case JAL_JUMP_OPCODE_X1:
+		case J_JUMP_OPCODE_X1:
+			jump_off = sign_extend(get_JumpOff_X1(bundle), 27);
+			addr = regs->pc +
+				(jump_off << TILEGX_LOG2_BUNDLE_SIZE_IN_BYTES);
+			break;
+		}
+	} else if (get_Opcode_X1(bundle) == BRANCH_OPCODE_X1) {
+		br_off = 0;
+		opcode = get_BrType_X1(bundle);
+
+		switch (opcode) {
+		case BEQZT_BRANCH_OPCODE_X1:
+		case BEQZ_BRANCH_OPCODE_X1:
+			if (get_SrcA_X1(bundle) == 0)
+				br_off = get_BrOff_X1(bundle);
+			break;
+		case BGEZT_BRANCH_OPCODE_X1:
+		case BGEZ_BRANCH_OPCODE_X1:
+			if (get_SrcA_X1(bundle) >= 0)
+				br_off = get_BrOff_X1(bundle);
+			break;
+		case BGTZT_BRANCH_OPCODE_X1:
+		case BGTZ_BRANCH_OPCODE_X1:
+			if (get_SrcA_X1(bundle) > 0)
+				br_off = get_BrOff_X1(bundle);
+			break;
+		case BLBCT_BRANCH_OPCODE_X1:
+		case BLBC_BRANCH_OPCODE_X1:
+			if (!(get_SrcA_X1(bundle) & 1))
+				br_off = get_BrOff_X1(bundle);
+			break;
+		case BLBST_BRANCH_OPCODE_X1:
+		case BLBS_BRANCH_OPCODE_X1:
+			if (get_SrcA_X1(bundle) & 1)
+				br_off = get_BrOff_X1(bundle);
+			break;
+		case BLEZT_BRANCH_OPCODE_X1:
+		case BLEZ_BRANCH_OPCODE_X1:
+			if (get_SrcA_X1(bundle) <= 0)
+				br_off = get_BrOff_X1(bundle);
+			break;
+		case BLTZT_BRANCH_OPCODE_X1:
+		case BLTZ_BRANCH_OPCODE_X1:
+			if (get_SrcA_X1(bundle) < 0)
+				br_off = get_BrOff_X1(bundle);
+			break;
+		case BNEZT_BRANCH_OPCODE_X1:
+		case BNEZ_BRANCH_OPCODE_X1:
+			if (get_SrcA_X1(bundle) != 0)
+				br_off = get_BrOff_X1(bundle);
+			break;
+		}
+
+		if (br_off != 0) {
+			br_off = sign_extend(br_off, 17);
+			addr = regs->pc +
+				(br_off << TILEGX_LOG2_BUNDLE_SIZE_IN_BYTES);
+		}
+	}
+
+	return addr;
+}
+
+/*
+ * Replace the next instruction after the current instruction with a
+ * breakpoint instruction.
+ */
+static void do_single_step(struct pt_regs *regs)
+{
+	unsigned long addr_wr;
+
+	/* Determine where the target instruction will send us to. */
+	stepped_addr = get_step_address(regs);
+	probe_kernel_read((char *)&stepped_instr, (char *)stepped_addr,
+			  BREAK_INSTR_SIZE);
+
+	addr_wr = writable_address(stepped_addr);
+	probe_kernel_write((char *)addr_wr, (char *)&singlestep_insn,
+			   BREAK_INSTR_SIZE);
+	smp_wmb();
+	flush_icache_range(stepped_addr, stepped_addr + BREAK_INSTR_SIZE);
+}
+
+static void undo_single_step(struct pt_regs *regs)
+{
+	unsigned long addr_wr;
+
+	if (stepped_instr == 0)
+		return;
+
+	addr_wr = writable_address(stepped_addr);
+	probe_kernel_write((char *)addr_wr, (char *)&stepped_instr,
+			   BREAK_INSTR_SIZE);
+	stepped_instr = 0;
+	smp_wmb();
+	flush_icache_range(stepped_addr, stepped_addr + BREAK_INSTR_SIZE);
+}
+
+/*
+ * Calls linux_debug_hook before the kernel dies. If KGDB is enabled,
+ * then try to fall into the debugger.
+ */
+static int
+kgdb_notify(struct notifier_block *self, unsigned long cmd, void *ptr)
+{
+	int ret;
+	unsigned long flags;
+	struct die_args *args = (struct die_args *)ptr;
+	struct pt_regs *regs = args->regs;
+
+#ifdef CONFIG_KPROBES
+	/*
+	 * Return immediately if the kprobes fault notifier has set
+	 * DIE_PAGE_FAULT.
+	 */
+	if (cmd == DIE_PAGE_FAULT)
+		return NOTIFY_DONE;
+#endif /* CONFIG_KPROBES */
+
+	switch (cmd) {
+	case DIE_BREAK:
+	case DIE_COMPILED_BPT:
+		break;
+	case DIE_SSTEPBP:
+		local_irq_save(flags);
+		kgdb_handle_exception(0, SIGTRAP, 0, regs);
+		local_irq_restore(flags);
+		return NOTIFY_STOP;
+	default:
+		/* Userspace events, ignore. */
+		if (user_mode(regs))
+			return NOTIFY_DONE;
+	}
+
+	local_irq_save(flags);
+	ret = kgdb_handle_exception(args->trapnr, args->signr, args->err, regs);
+	local_irq_restore(flags);
+	if (ret)
+		return NOTIFY_DONE;
+
+	return NOTIFY_STOP;
+}
+
+static struct notifier_block kgdb_notifier = {
+	.notifier_call = kgdb_notify,
+};
+
+/*
+ * kgdb_arch_handle_exception - Handle architecture specific GDB packets.
+ * @vector: The error vector of the exception that happened.
+ * @signo: The signal number of the exception that happened.
+ * @err_code: The error code of the exception that happened.
+ * @remcom_in_buffer: The buffer of the packet we have read.
+ * @remcom_out_buffer: The buffer of %BUFMAX bytes to write a packet into.
+ * @regs: The &struct pt_regs of the current process.
+ *
+ * This function MUST handle the 'c' and 's' command packets,
+ * as well packets to set / remove a hardware breakpoint, if used.
+ * If there are additional packets which the hardware needs to handle,
+ * they are handled here. The code should return -1 if it wants to
+ * process more packets, and a %0 or %1 if it wants to exit from the
+ * kgdb callback.
+ */
+int kgdb_arch_handle_exception(int vector, int signo, int err_code,
+			       char *remcom_in_buffer, char *remcom_out_buffer,
+			       struct pt_regs *regs)
+{
+	char *ptr;
+	unsigned long address;
+
+	/* Undo any stepping we may have done. */
+	undo_single_step(regs);
+
+	switch (remcom_in_buffer[0]) {
+	case 'c':
+	case 's':
+	case 'D':
+	case 'k':
+		/*
+		 * Try to read optional parameter, pc unchanged if no parm.
+		 * If this was a compiled-in breakpoint, we need to move
+		 * to the next instruction or we will just breakpoint
+		 * over and over again.
+		 */
+		ptr = &remcom_in_buffer[1];
+		if (kgdb_hex2long(&ptr, &address))
+			regs->pc = address;
+		else if (*(unsigned long *)regs->pc == compiled_bpt)
+			regs->pc += BREAK_INSTR_SIZE;
+
+		if (remcom_in_buffer[0] == 's') {
+			do_single_step(regs);
+			kgdb_single_step = 1;
+			atomic_set(&kgdb_cpu_doing_single_step,
+				   raw_smp_processor_id());
+		} else
+			atomic_set(&kgdb_cpu_doing_single_step, -1);
+
+		return 0;
+	}
+
+	return -1; /* this means that we do not want to exit from the handler */
+}
+
+struct kgdb_arch arch_kgdb_ops;
+
+/*
+ * kgdb_arch_init - Perform any architecture specific initalization.
+ *
+ * This function will handle the initalization of any architecture
+ * specific callbacks.
+ */
+int kgdb_arch_init(void)
+{
+	tile_bundle_bits bundle = TILEGX_BPT_BUNDLE;
+
+	memcpy(arch_kgdb_ops.gdb_bpt_instr, &bundle, BREAK_INSTR_SIZE);
+	return register_die_notifier(&kgdb_notifier);
+}
+
+/*
+ * kgdb_arch_exit - Perform any architecture specific uninitalization.
+ *
+ * This function will handle the uninitalization of any architecture
+ * specific callbacks, for dynamic registration and unregistration.
+ */
+void kgdb_arch_exit(void)
+{
+	unregister_die_notifier(&kgdb_notifier);
+}
+
+int kgdb_arch_set_breakpoint(struct kgdb_bkpt *bpt)
+{
+	int err;
+	unsigned long addr_wr = writable_address(bpt->bpt_addr);
+
+	if (addr_wr == 0)
+		return -1;
+
+	err = probe_kernel_read(bpt->saved_instr, (char *)bpt->bpt_addr,
+				BREAK_INSTR_SIZE);
+	if (err)
+		return err;
+
+	err = probe_kernel_write((char *)addr_wr, arch_kgdb_ops.gdb_bpt_instr,
+				 BREAK_INSTR_SIZE);
+	smp_wmb();
+	flush_icache_range((unsigned long)bpt->bpt_addr,
+			   (unsigned long)bpt->bpt_addr + BREAK_INSTR_SIZE);
+	return err;
+}
+
+int kgdb_arch_remove_breakpoint(struct kgdb_bkpt *bpt)
+{
+	int err;
+	unsigned long addr_wr = writable_address(bpt->bpt_addr);
+
+	if (addr_wr == 0)
+		return -1;
+
+	err = probe_kernel_write((char *)addr_wr, (char *)bpt->saved_instr,
+				 BREAK_INSTR_SIZE);
+	smp_wmb();
+	flush_icache_range((unsigned long)bpt->bpt_addr,
+			   (unsigned long)bpt->bpt_addr + BREAK_INSTR_SIZE);
+	return err;
+}
diff --git a/arch/tile/kernel/kprobes.c b/arch/tile/kernel/kprobes.c
new file mode 100644
index 000000000..f8a45c51e
--- /dev/null
+++ b/arch/tile/kernel/kprobes.c
@@ -0,0 +1,527 @@
+/*
+ * arch/tile/kernel/kprobes.c
+ * Kprobes on TILE-Gx
+ *
+ * Some portions copied from the MIPS version.
+ *
+ * Copyright (C) IBM Corporation, 2002, 2004
+ * Copyright 2006 Sony Corp.
+ * Copyright 2010 Cavium Networks
+ *
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/kprobes.h>
+#include <linux/kdebug.h>
+#include <linux/module.h>
+#include <linux/slab.h>
+#include <linux/uaccess.h>
+#include <asm/cacheflush.h>
+
+#include <arch/opcode.h>
+
+DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
+DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
+
+tile_bundle_bits breakpoint_insn = TILEGX_BPT_BUNDLE;
+tile_bundle_bits breakpoint2_insn = TILEGX_BPT_BUNDLE | DIE_SSTEPBP;
+
+/*
+ * Check whether instruction is branch or jump, or if executing it
+ * has different results depending on where it is executed (e.g. lnk).
+ */
+static int __kprobes insn_has_control(kprobe_opcode_t insn)
+{
+	if (get_Mode(insn) != 0) {   /* Y-format bundle */
+		if (get_Opcode_Y1(insn) != RRR_1_OPCODE_Y1 ||
+		    get_RRROpcodeExtension_Y1(insn) != UNARY_RRR_1_OPCODE_Y1)
+			return 0;
+
+		switch (get_UnaryOpcodeExtension_Y1(insn)) {
+		case JALRP_UNARY_OPCODE_Y1:
+		case JALR_UNARY_OPCODE_Y1:
+		case JRP_UNARY_OPCODE_Y1:
+		case JR_UNARY_OPCODE_Y1:
+		case LNK_UNARY_OPCODE_Y1:
+			return 1;
+		default:
+			return 0;
+		}
+	}
+
+	switch (get_Opcode_X1(insn)) {
+	case BRANCH_OPCODE_X1:	/* branch instructions */
+	case JUMP_OPCODE_X1:	/* jump instructions: j and jal */
+		return 1;
+
+	case RRR_0_OPCODE_X1:   /* other jump instructions */
+		if (get_RRROpcodeExtension_X1(insn) != UNARY_RRR_0_OPCODE_X1)
+			return 0;
+		switch (get_UnaryOpcodeExtension_X1(insn)) {
+		case JALRP_UNARY_OPCODE_X1:
+		case JALR_UNARY_OPCODE_X1:
+		case JRP_UNARY_OPCODE_X1:
+		case JR_UNARY_OPCODE_X1:
+		case LNK_UNARY_OPCODE_X1:
+			return 1;
+		default:
+			return 0;
+		}
+	default:
+		return 0;
+	}
+}
+
+int __kprobes arch_prepare_kprobe(struct kprobe *p)
+{
+	unsigned long addr = (unsigned long)p->addr;
+
+	if (addr & (sizeof(kprobe_opcode_t) - 1))
+		return -EINVAL;
+
+	if (insn_has_control(*p->addr)) {
+		pr_notice("Kprobes for control instructions are not supported\n");
+		return -EINVAL;
+	}
+
+	/* insn: must be on special executable page on tile. */
+	p->ainsn.insn = get_insn_slot();
+	if (!p->ainsn.insn)
+		return -ENOMEM;
+
+	/*
+	 * In the kprobe->ainsn.insn[] array we store the original
+	 * instruction at index zero and a break trap instruction at
+	 * index one.
+	 */
+	memcpy(&p->ainsn.insn[0], p->addr, sizeof(kprobe_opcode_t));
+	p->ainsn.insn[1] = breakpoint2_insn;
+	p->opcode = *p->addr;
+
+	return 0;
+}
+
+void __kprobes arch_arm_kprobe(struct kprobe *p)
+{
+	unsigned long addr_wr;
+
+	/* Operate on writable kernel text mapping. */
+	addr_wr = (unsigned long)p->addr - MEM_SV_START + PAGE_OFFSET;
+
+	if (probe_kernel_write((void *)addr_wr, &breakpoint_insn,
+		sizeof(breakpoint_insn)))
+		pr_err("%s: failed to enable kprobe\n", __func__);
+
+	smp_wmb();
+	flush_insn_slot(p);
+}
+
+void __kprobes arch_disarm_kprobe(struct kprobe *kp)
+{
+	unsigned long addr_wr;
+
+	/* Operate on writable kernel text mapping. */
+	addr_wr = (unsigned long)kp->addr - MEM_SV_START + PAGE_OFFSET;
+
+	if (probe_kernel_write((void *)addr_wr, &kp->opcode,
+		sizeof(kp->opcode)))
+		pr_err("%s: failed to enable kprobe\n", __func__);
+
+	smp_wmb();
+	flush_insn_slot(kp);
+}
+
+void __kprobes arch_remove_kprobe(struct kprobe *p)
+{
+	if (p->ainsn.insn) {
+		free_insn_slot(p->ainsn.insn, 0);
+		p->ainsn.insn = NULL;
+	}
+}
+
+static void __kprobes save_previous_kprobe(struct kprobe_ctlblk *kcb)
+{
+	kcb->prev_kprobe.kp = kprobe_running();
+	kcb->prev_kprobe.status = kcb->kprobe_status;
+	kcb->prev_kprobe.saved_pc = kcb->kprobe_saved_pc;
+}
+
+static void __kprobes restore_previous_kprobe(struct kprobe_ctlblk *kcb)
+{
+	__this_cpu_write(current_kprobe, kcb->prev_kprobe.kp);
+	kcb->kprobe_status = kcb->prev_kprobe.status;
+	kcb->kprobe_saved_pc = kcb->prev_kprobe.saved_pc;
+}
+
+static void __kprobes set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
+			struct kprobe_ctlblk *kcb)
+{
+	__this_cpu_write(current_kprobe, p);
+	kcb->kprobe_saved_pc = regs->pc;
+}
+
+static void __kprobes prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
+{
+	/* Single step inline if the instruction is a break. */
+	if (p->opcode == breakpoint_insn ||
+	    p->opcode == breakpoint2_insn)
+		regs->pc = (unsigned long)p->addr;
+	else
+		regs->pc = (unsigned long)&p->ainsn.insn[0];
+}
+
+static int __kprobes kprobe_handler(struct pt_regs *regs)
+{
+	struct kprobe *p;
+	int ret = 0;
+	kprobe_opcode_t *addr;
+	struct kprobe_ctlblk *kcb;
+
+	addr = (kprobe_opcode_t *)regs->pc;
+
+	/*
+	 * We don't want to be preempted for the entire
+	 * duration of kprobe processing.
+	 */
+	preempt_disable();
+	kcb = get_kprobe_ctlblk();
+
+	/* Check we're not actually recursing. */
+	if (kprobe_running()) {
+		p = get_kprobe(addr);
+		if (p) {
+			if (kcb->kprobe_status == KPROBE_HIT_SS &&
+			    p->ainsn.insn[0] == breakpoint_insn) {
+				goto no_kprobe;
+			}
+			/*
+			 * We have reentered the kprobe_handler(), since
+			 * another probe was hit while within the handler.
+			 * We here save the original kprobes variables and
+			 * just single step on the instruction of the new probe
+			 * without calling any user handlers.
+			 */
+			save_previous_kprobe(kcb);
+			set_current_kprobe(p, regs, kcb);
+			kprobes_inc_nmissed_count(p);
+			prepare_singlestep(p, regs);
+			kcb->kprobe_status = KPROBE_REENTER;
+			return 1;
+		} else {
+			if (*addr != breakpoint_insn) {
+				/*
+				 * The breakpoint instruction was removed by
+				 * another cpu right after we hit, no further
+				 * handling of this interrupt is appropriate.
+				 */
+				ret = 1;
+				goto no_kprobe;
+			}
+			p = __this_cpu_read(current_kprobe);
+			if (p->break_handler && p->break_handler(p, regs))
+				goto ss_probe;
+		}
+		goto no_kprobe;
+	}
+
+	p = get_kprobe(addr);
+	if (!p) {
+		if (*addr != breakpoint_insn) {
+			/*
+			 * The breakpoint instruction was removed right
+			 * after we hit it.  Another cpu has removed
+			 * either a probepoint or a debugger breakpoint
+			 * at this address.  In either case, no further
+			 * handling of this interrupt is appropriate.
+			 */
+			ret = 1;
+		}
+		/* Not one of ours: let kernel handle it. */
+		goto no_kprobe;
+	}
+
+	set_current_kprobe(p, regs, kcb);
+	kcb->kprobe_status = KPROBE_HIT_ACTIVE;
+
+	if (p->pre_handler && p->pre_handler(p, regs)) {
+		/* Handler has already set things up, so skip ss setup. */
+		return 1;
+	}
+
+ss_probe:
+	prepare_singlestep(p, regs);
+	kcb->kprobe_status = KPROBE_HIT_SS;
+	return 1;
+
+no_kprobe:
+	preempt_enable_no_resched();
+	return ret;
+}
+
+/*
+ * Called after single-stepping.  p->addr is the address of the
+ * instruction that has been replaced by the breakpoint. To avoid the
+ * SMP problems that can occur when we temporarily put back the
+ * original opcode to single-step, we single-stepped a copy of the
+ * instruction.  The address of this copy is p->ainsn.insn.
+ *
+ * This function prepares to return from the post-single-step
+ * breakpoint trap.
+ */
+static void __kprobes resume_execution(struct kprobe *p,
+				       struct pt_regs *regs,
+				       struct kprobe_ctlblk *kcb)
+{
+	unsigned long orig_pc = kcb->kprobe_saved_pc;
+	regs->pc = orig_pc + 8;
+}
+
+static inline int post_kprobe_handler(struct pt_regs *regs)
+{
+	struct kprobe *cur = kprobe_running();
+	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
+
+	if (!cur)
+		return 0;
+
+	if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
+		kcb->kprobe_status = KPROBE_HIT_SSDONE;
+		cur->post_handler(cur, regs, 0);
+	}
+
+	resume_execution(cur, regs, kcb);
+
+	/* Restore back the original saved kprobes variables and continue. */
+	if (kcb->kprobe_status == KPROBE_REENTER) {
+		restore_previous_kprobe(kcb);
+		goto out;
+	}
+	reset_current_kprobe();
+out:
+	preempt_enable_no_resched();
+
+	return 1;
+}
+
+static inline int kprobe_fault_handler(struct pt_regs *regs, int trapnr)
+{
+	struct kprobe *cur = kprobe_running();
+	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
+
+	if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
+		return 1;
+
+	if (kcb->kprobe_status & KPROBE_HIT_SS) {
+		/*
+		 * We are here because the instruction being single
+		 * stepped caused a page fault. We reset the current
+		 * kprobe and the ip points back to the probe address
+		 * and allow the page fault handler to continue as a
+		 * normal page fault.
+		 */
+		resume_execution(cur, regs, kcb);
+		reset_current_kprobe();
+		preempt_enable_no_resched();
+	}
+	return 0;
+}
+
+/*
+ * Wrapper routine for handling exceptions.
+ */
+int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
+				       unsigned long val, void *data)
+{
+	struct die_args *args = (struct die_args *)data;
+	int ret = NOTIFY_DONE;
+
+	switch (val) {
+	case DIE_BREAK:
+		if (kprobe_handler(args->regs))
+			ret = NOTIFY_STOP;
+		break;
+	case DIE_SSTEPBP:
+		if (post_kprobe_handler(args->regs))
+			ret = NOTIFY_STOP;
+		break;
+	case DIE_PAGE_FAULT:
+		/* kprobe_running() needs smp_processor_id(). */
+		preempt_disable();
+
+		if (kprobe_running()
+		    && kprobe_fault_handler(args->regs, args->trapnr))
+			ret = NOTIFY_STOP;
+		preempt_enable();
+		break;
+	default:
+		break;
+	}
+	return ret;
+}
+
+int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
+{
+	struct jprobe *jp = container_of(p, struct jprobe, kp);
+	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
+
+	kcb->jprobe_saved_regs = *regs;
+	kcb->jprobe_saved_sp = regs->sp;
+
+	memcpy(kcb->jprobes_stack, (void *)kcb->jprobe_saved_sp,
+	       MIN_JPROBES_STACK_SIZE(kcb->jprobe_saved_sp));
+
+	regs->pc = (unsigned long)(jp->entry);
+
+	return 1;
+}
+
+/* Defined in the inline asm below. */
+void jprobe_return_end(void);
+
+void __kprobes jprobe_return(void)
+{
+	asm volatile(
+		"bpt\n\t"
+		".globl jprobe_return_end\n"
+		"jprobe_return_end:\n");
+}
+
+int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
+{
+	struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
+
+	if (regs->pc >= (unsigned long)jprobe_return &&
+	    regs->pc <= (unsigned long)jprobe_return_end) {
+		*regs = kcb->jprobe_saved_regs;
+		memcpy((void *)kcb->jprobe_saved_sp, kcb->jprobes_stack,
+		       MIN_JPROBES_STACK_SIZE(kcb->jprobe_saved_sp));
+		preempt_enable_no_resched();
+
+		return 1;
+	}
+	return 0;
+}
+
+/*
+ * Function return probe trampoline:
+ * - init_kprobes() establishes a probepoint here
+ * - When the probed function returns, this probe causes the
+ *   handlers to fire
+ */
+static void __used kretprobe_trampoline_holder(void)
+{
+	asm volatile(
+		"nop\n\t"
+		".global kretprobe_trampoline\n"
+		"kretprobe_trampoline:\n\t"
+		"nop\n\t"
+		: : : "memory");
+}
+
+void kretprobe_trampoline(void);
+
+void __kprobes arch_prepare_kretprobe(struct kretprobe_instance *ri,
+				      struct pt_regs *regs)
+{
+	ri->ret_addr = (kprobe_opcode_t *) regs->lr;
+
+	/* Replace the return addr with trampoline addr */
+	regs->lr = (unsigned long)kretprobe_trampoline;
+}
+
+/*
+ * Called when the probe at kretprobe trampoline is hit.
+ */
+static int __kprobes trampoline_probe_handler(struct kprobe *p,
+						struct pt_regs *regs)
+{
+	struct kretprobe_instance *ri = NULL;
+	struct hlist_head *head, empty_rp;
+	struct hlist_node *tmp;
+	unsigned long flags, orig_ret_address = 0;
+	unsigned long trampoline_address = (unsigned long)kretprobe_trampoline;
+
+	INIT_HLIST_HEAD(&empty_rp);
+	kretprobe_hash_lock(current, &head, &flags);
+
+	/*
+	 * It is possible to have multiple instances associated with a given
+	 * task either because multiple functions in the call path have
+	 * a return probe installed on them, and/or more than one return
+	 * return probe was registered for a target function.
+	 *
+	 * We can handle this because:
+	 *     - instances are always inserted at the head of the list
+	 *     - when multiple return probes are registered for the same
+	 *       function, the first instance's ret_addr will point to the
+	 *       real return address, and all the rest will point to
+	 *       kretprobe_trampoline
+	 */
+	hlist_for_each_entry_safe(ri, tmp, head, hlist) {
+		if (ri->task != current)
+			/* another task is sharing our hash bucket */
+			continue;
+
+		if (ri->rp && ri->rp->handler)
+			ri->rp->handler(ri, regs);
+
+		orig_ret_address = (unsigned long)ri->ret_addr;
+		recycle_rp_inst(ri, &empty_rp);
+
+		if (orig_ret_address != trampoline_address) {
+			/*
+			 * This is the real return address. Any other
+			 * instances associated with this task are for
+			 * other calls deeper on the call stack
+			 */
+			break;
+		}
+	}
+
+	kretprobe_assert(ri, orig_ret_address, trampoline_address);
+	instruction_pointer(regs) = orig_ret_address;
+
+	reset_current_kprobe();
+	kretprobe_hash_unlock(current, &flags);
+	preempt_enable_no_resched();
+
+	hlist_for_each_entry_safe(ri, tmp, &empty_rp, hlist) {
+		hlist_del(&ri->hlist);
+		kfree(ri);
+	}
+	/*
+	 * By returning a non-zero value, we are telling
+	 * kprobe_handler() that we don't want the post_handler
+	 * to run (and have re-enabled preemption)
+	 */
+	return 1;
+}
+
+int __kprobes arch_trampoline_kprobe(struct kprobe *p)
+{
+	if (p->addr == (kprobe_opcode_t *)kretprobe_trampoline)
+		return 1;
+
+	return 0;
+}
+
+static struct kprobe trampoline_p = {
+	.addr = (kprobe_opcode_t *)kretprobe_trampoline,
+	.pre_handler = trampoline_probe_handler
+};
+
+int __init arch_init_kprobes(void)
+{
+	register_kprobe(&trampoline_p);
+	return 0;
+}
diff --git a/arch/tile/kernel/machine_kexec.c b/arch/tile/kernel/machine_kexec.c
new file mode 100644
index 000000000..008aa2fae
--- /dev/null
+++ b/arch/tile/kernel/machine_kexec.c
@@ -0,0 +1,298 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * based on machine_kexec.c from other architectures in linux-2.6.18
+ */
+
+#include <linux/mm.h>
+#include <linux/kexec.h>
+#include <linux/delay.h>
+#include <linux/reboot.h>
+#include <linux/errno.h>
+#include <linux/vmalloc.h>
+#include <linux/cpumask.h>
+#include <linux/kernel.h>
+#include <linux/elf.h>
+#include <linux/highmem.h>
+#include <linux/mmu_context.h>
+#include <linux/io.h>
+#include <linux/timex.h>
+#include <asm/pgtable.h>
+#include <asm/pgalloc.h>
+#include <asm/cacheflush.h>
+#include <asm/checksum.h>
+#include <asm/tlbflush.h>
+#include <asm/homecache.h>
+#include <hv/hypervisor.h>
+
+
+/*
+ * This stuff is not in elf.h and is not in any other kernel include.
+ * This stuff is needed below in the little boot notes parser to
+ * extract the command line so we can pass it to the hypervisor.
+ */
+struct Elf32_Bhdr {
+	Elf32_Word b_signature;
+	Elf32_Word b_size;
+	Elf32_Half b_checksum;
+	Elf32_Half b_records;
+};
+#define ELF_BOOT_MAGIC		0x0E1FB007
+#define EBN_COMMAND_LINE	0x00000004
+#define roundupsz(X) (((X) + 3) & ~3)
+
+/* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
+
+
+void machine_shutdown(void)
+{
+	/*
+	 * Normally we would stop all the other processors here, but
+	 * the check in machine_kexec_prepare below ensures we'll only
+	 * get this far if we've been booted with "nosmp" on the
+	 * command line or without CONFIG_SMP so there's nothing to do
+	 * here (for now).
+	 */
+}
+
+void machine_crash_shutdown(struct pt_regs *regs)
+{
+	/*
+	 * Cannot happen.  This type of kexec is disabled on this
+	 * architecture (and enforced in machine_kexec_prepare below).
+	 */
+}
+
+
+int machine_kexec_prepare(struct kimage *image)
+{
+	if (num_online_cpus() > 1) {
+		pr_warn("%s: detected attempt to kexec with num_online_cpus() > 1\n",
+			__func__);
+		return -ENOSYS;
+	}
+	if (image->type != KEXEC_TYPE_DEFAULT) {
+		pr_warn("%s: detected attempt to kexec with unsupported type: %d\n",
+			__func__, image->type);
+		return -ENOSYS;
+	}
+	return 0;
+}
+
+void machine_kexec_cleanup(struct kimage *image)
+{
+	/*
+	 * We did nothing in machine_kexec_prepare,
+	 * so we have nothing to do here.
+	 */
+}
+
+/*
+ * If we can find elf boot notes on this page, return the command
+ * line.  Otherwise, silently return null.  Somewhat kludgy, but no
+ * good way to do this without significantly rearchitecting the
+ * architecture-independent kexec code.
+ */
+
+static unsigned char *kexec_bn2cl(void *pg)
+{
+	struct Elf32_Bhdr *bhdrp;
+	Elf32_Nhdr *nhdrp;
+	unsigned char *desc;
+	unsigned char *command_line;
+	__sum16 csum;
+
+	bhdrp = (struct Elf32_Bhdr *) pg;
+
+	/*
+	 * This routine is invoked for every source page, so make
+	 * sure to quietly ignore every impossible page.
+	 */
+	if (bhdrp->b_signature != ELF_BOOT_MAGIC ||
+	    bhdrp->b_size > PAGE_SIZE)
+		return 0;
+
+	/*
+	 * If we get a checksum mismatch, warn with the checksum
+	 * so we can diagnose better.
+	 */
+	csum = ip_compute_csum(pg, bhdrp->b_size);
+	if (csum != 0) {
+		pr_warn("%s: bad checksum %#x (size %d)\n",
+			__func__, csum, bhdrp->b_size);
+		return 0;
+	}
+
+	nhdrp = (Elf32_Nhdr *) (bhdrp + 1);
+
+	while (nhdrp->n_type != EBN_COMMAND_LINE) {
+
+		desc = (unsigned char *) (nhdrp + 1);
+		desc += roundupsz(nhdrp->n_descsz);
+
+		nhdrp = (Elf32_Nhdr *) desc;
+
+		/* still in bounds? */
+		if ((unsigned char *) (nhdrp + 1) >
+		    ((unsigned char *) pg) + bhdrp->b_size) {
+
+			pr_info("%s: out of bounds\n", __func__);
+			return 0;
+		}
+	}
+
+	command_line = (unsigned char *) (nhdrp + 1);
+	desc = command_line;
+
+	while (*desc != '\0') {
+		desc++;
+		if (((unsigned long)desc & PAGE_MASK) != (unsigned long)pg) {
+			pr_info("%s: ran off end of page\n", __func__);
+			return 0;
+		}
+	}
+
+	return command_line;
+}
+
+static void kexec_find_and_set_command_line(struct kimage *image)
+{
+	kimage_entry_t *ptr, entry;
+
+	unsigned char *command_line = 0;
+	unsigned char *r;
+	HV_Errno hverr;
+
+	for (ptr = &image->head;
+	     (entry = *ptr) && !(entry & IND_DONE);
+	     ptr = (entry & IND_INDIRECTION) ?
+		     phys_to_virt((entry & PAGE_MASK)) : ptr + 1) {
+
+		if ((entry & IND_SOURCE)) {
+			void *va =
+				kmap_atomic_pfn(entry >> PAGE_SHIFT);
+			r = kexec_bn2cl(va);
+			if (r) {
+				command_line = r;
+				break;
+			}
+			kunmap_atomic(va);
+		}
+	}
+
+	if (command_line != 0) {
+		pr_info("setting new command line to \"%s\"\n", command_line);
+
+		hverr = hv_set_command_line(
+			(HV_VirtAddr) command_line, strlen(command_line));
+		kunmap_atomic(command_line);
+	} else {
+		pr_info("%s: no command line found; making empty\n", __func__);
+		hverr = hv_set_command_line((HV_VirtAddr) command_line, 0);
+	}
+	if (hverr)
+		pr_warn("%s: hv_set_command_line returned error: %d\n",
+			__func__, hverr);
+}
+
+/*
+ * The kexec code range-checks all its PAs, so to avoid having it run
+ * amok and allocate memory and then sequester it from every other
+ * controller, we force it to come from controller zero.  We also
+ * disable the oom-killer since if we do end up running out of memory,
+ * that almost certainly won't help.
+ */
+struct page *kimage_alloc_pages_arch(gfp_t gfp_mask, unsigned int order)
+{
+	gfp_mask |= __GFP_THISNODE | __GFP_NORETRY;
+	return alloc_pages_node(0, gfp_mask, order);
+}
+
+/*
+ * Address range in which pa=va mapping is set in setup_quasi_va_is_pa().
+ * For tilepro, PAGE_OFFSET is used since this is the largest possbile value
+ * for tilepro, while for tilegx, we limit it to entire middle level page
+ * table which we assume has been allocated and is undoubtedly large enough.
+ */
+#ifndef __tilegx__
+#define	QUASI_VA_IS_PA_ADDR_RANGE PAGE_OFFSET
+#else
+#define	QUASI_VA_IS_PA_ADDR_RANGE PGDIR_SIZE
+#endif
+
+static void setup_quasi_va_is_pa(void)
+{
+	HV_PTE pte;
+	unsigned long i;
+
+	/*
+	 * Flush our TLB to prevent conflicts between the previous contents
+	 * and the new stuff we're about to add.
+	 */
+	local_flush_tlb_all();
+
+	/*
+	 * setup VA is PA, at least up to QUASI_VA_IS_PA_ADDR_RANGE.
+	 * Note here we assume that level-1 page table is defined by
+	 * HPAGE_SIZE.
+	 */
+	pte = hv_pte(_PAGE_KERNEL | _PAGE_HUGE_PAGE);
+	pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_NO_L3);
+	for (i = 0; i < (QUASI_VA_IS_PA_ADDR_RANGE >> HPAGE_SHIFT); i++) {
+		unsigned long vaddr = i << HPAGE_SHIFT;
+		pgd_t *pgd = pgd_offset(current->mm, vaddr);
+		pud_t *pud = pud_offset(pgd, vaddr);
+		pte_t *ptep = (pte_t *) pmd_offset(pud, vaddr);
+		unsigned long pfn = i << (HPAGE_SHIFT - PAGE_SHIFT);
+
+		if (pfn_valid(pfn))
+			__set_pte(ptep, pfn_pte(pfn, pte));
+	}
+}
+
+
+void machine_kexec(struct kimage *image)
+{
+	void *reboot_code_buffer;
+	pte_t *ptep;
+	void (*rnk)(unsigned long, void *, unsigned long)
+		__noreturn;
+
+	/* Mask all interrupts before starting to reboot. */
+	interrupt_mask_set_mask(~0ULL);
+
+	kexec_find_and_set_command_line(image);
+
+	/*
+	 * Adjust the home caching of the control page to be cached on
+	 * this cpu, and copy the assembly helper into the control
+	 * code page, which we map in the vmalloc area.
+	 */
+	homecache_change_page_home(image->control_code_page, 0,
+				   smp_processor_id());
+	reboot_code_buffer = page_address(image->control_code_page);
+	BUG_ON(reboot_code_buffer == NULL);
+	ptep = virt_to_pte(NULL, (unsigned long)reboot_code_buffer);
+	__set_pte(ptep, pte_mkexec(*ptep));
+	memcpy(reboot_code_buffer, relocate_new_kernel,
+	       relocate_new_kernel_size);
+	__flush_icache_range(
+		(unsigned long) reboot_code_buffer,
+		(unsigned long) reboot_code_buffer + relocate_new_kernel_size);
+
+	setup_quasi_va_is_pa();
+
+	/* now call it */
+	rnk = reboot_code_buffer;
+	(*rnk)(image->head, reboot_code_buffer, image->start);
+}
diff --git a/arch/tile/kernel/mcount_64.S b/arch/tile/kernel/mcount_64.S
new file mode 100644
index 000000000..6c6702451
--- /dev/null
+++ b/arch/tile/kernel/mcount_64.S
@@ -0,0 +1,211 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * TILE-Gx specific __mcount support
+ */
+
+#include <linux/linkage.h>
+#include <asm/ftrace.h>
+
+#define REGSIZE 8
+
+	.text
+	.global __mcount
+
+	.macro	MCOUNT_SAVE_REGS
+	addli	sp, sp, -REGSIZE
+	{
+	 st     sp, lr
+	 addli	r29, sp, - (12 * REGSIZE)
+	}
+	{
+	 addli	sp, sp, - (13 * REGSIZE)
+	 st     r29, sp
+	}
+	addli	r29, r29, REGSIZE
+	{ st	r29, r0; addli	r29, r29, REGSIZE }
+	{ st	r29, r1; addli	r29, r29, REGSIZE }
+	{ st	r29, r2; addli	r29, r29, REGSIZE }
+	{ st	r29, r3; addli	r29, r29, REGSIZE }
+	{ st	r29, r4; addli	r29, r29, REGSIZE }
+	{ st	r29, r5; addli	r29, r29, REGSIZE }
+	{ st	r29, r6; addli	r29, r29, REGSIZE }
+	{ st	r29, r7; addli	r29, r29, REGSIZE }
+	{ st	r29, r8; addli	r29, r29, REGSIZE }
+	{ st	r29, r9; addli	r29, r29, REGSIZE }
+	{ st	r29, r10; addli	r29, r29, REGSIZE }
+	.endm
+
+	.macro	MCOUNT_RESTORE_REGS
+	addli	r29, sp, (2 * REGSIZE)
+	{ ld	r0, r29; addli	r29, r29, REGSIZE }
+	{ ld	r1, r29; addli	r29, r29, REGSIZE }
+	{ ld	r2, r29; addli	r29, r29, REGSIZE }
+	{ ld	r3, r29; addli	r29, r29, REGSIZE }
+	{ ld	r4, r29; addli	r29, r29, REGSIZE }
+	{ ld	r5, r29; addli	r29, r29, REGSIZE }
+	{ ld	r6, r29; addli	r29, r29, REGSIZE }
+	{ ld	r7, r29; addli	r29, r29, REGSIZE }
+	{ ld	r8, r29; addli	r29, r29, REGSIZE }
+	{ ld	r9, r29; addli	r29, r29, REGSIZE }
+	{ ld	r10, r29; addli	lr, sp, (13 * REGSIZE) }
+	{ ld	lr, lr;  addli	sp, sp, (14 * REGSIZE) }
+	.endm
+
+	.macro  RETURN_BACK
+	{ move	r12, lr; move	lr, r10 }
+	jrp	r12
+	.endm
+
+#ifdef CONFIG_DYNAMIC_FTRACE
+
+	.align	64
+STD_ENTRY(__mcount)
+__mcount:
+	j	ftrace_stub
+STD_ENDPROC(__mcount)
+
+	.align	64
+STD_ENTRY(ftrace_caller)
+	MCOUNT_SAVE_REGS
+
+	/* arg1: self return address */
+	/* arg2: parent's return address */
+	/* arg3: ftrace_ops */
+	/* arg4: regs (but make it NULL) */
+	{ move	r0, lr;  moveli        r2, hw2_last(function_trace_op) }
+	{ move	r1, r10; shl16insli    r2, r2, hw1(function_trace_op) }
+	{ movei	r3, 0;   shl16insli    r2, r2, hw0(function_trace_op) }
+	ld	r2,r2
+
+	.global	ftrace_call
+ftrace_call:
+	/*
+	 * a placeholder for the call to a real tracing function, i.e.
+	 * ftrace_trace_function()
+	 */
+	nop
+
+#ifdef CONFIG_FUNCTION_GRAPH_TRACER
+	.global	ftrace_graph_call
+ftrace_graph_call:
+	/*
+	 * a placeholder for the call to a real tracing function, i.e.
+	 * ftrace_graph_caller()
+	 */
+	nop
+#endif
+	MCOUNT_RESTORE_REGS
+	.global	ftrace_stub
+ftrace_stub:
+	RETURN_BACK
+STD_ENDPROC(ftrace_caller)
+
+#else /* ! CONFIG_DYNAMIC_FTRACE */
+
+	.align	64
+STD_ENTRY(__mcount)
+	{
+	 moveli	r11, hw2_last(ftrace_trace_function)
+	 moveli	r13, hw2_last(ftrace_stub)
+	}
+	{
+	 shl16insli	r11, r11, hw1(ftrace_trace_function)
+	 shl16insli	r13, r13, hw1(ftrace_stub)
+	}
+	{
+	 shl16insli	r11, r11, hw0(ftrace_trace_function)
+	 shl16insli	r13, r13, hw0(ftrace_stub)
+	}
+
+	ld	r11, r11
+	sub	r14, r13, r11
+	bnez	r14, static_trace
+
+#ifdef	CONFIG_FUNCTION_GRAPH_TRACER
+	moveli	r15, hw2_last(ftrace_graph_return)
+	shl16insli	r15, r15, hw1(ftrace_graph_return)
+	shl16insli	r15, r15, hw0(ftrace_graph_return)
+	ld	r15, r15
+	sub	r15, r15, r13
+	bnez	r15, ftrace_graph_caller
+
+	{
+	 moveli	r16, hw2_last(ftrace_graph_entry)
+	 moveli	r17, hw2_last(ftrace_graph_entry_stub)
+	}
+	{
+	 shl16insli	r16, r16, hw1(ftrace_graph_entry)
+	 shl16insli	r17, r17, hw1(ftrace_graph_entry_stub)
+	}
+	{
+	 shl16insli	r16, r16, hw0(ftrace_graph_entry)
+	 shl16insli	r17, r17, hw0(ftrace_graph_entry_stub)
+	}
+	ld	r16, r16
+	sub	r17, r16, r17
+	bnez	r17, ftrace_graph_caller
+
+#endif
+	RETURN_BACK
+
+static_trace:
+	MCOUNT_SAVE_REGS
+
+	/* arg1: self return address */
+	/* arg2: parent's return address */
+	{ move	r0, lr; move	r1, r10 }
+
+	/* call ftrace_trace_function() */
+	jalr	r11
+
+	MCOUNT_RESTORE_REGS
+
+	.global ftrace_stub
+ftrace_stub:
+	RETURN_BACK
+STD_ENDPROC(__mcount)
+
+#endif	/* ! CONFIG_DYNAMIC_FTRACE */
+
+#ifdef CONFIG_FUNCTION_GRAPH_TRACER
+
+STD_ENTRY(ftrace_graph_caller)
+ftrace_graph_caller:
+#ifndef CONFIG_DYNAMIC_FTRACE
+	MCOUNT_SAVE_REGS
+#endif
+
+	/* arg1: Get the location of the parent's return address */
+	addi	r0, sp, 12 * REGSIZE
+	/* arg2: Get self return address */
+	move	r1, lr
+
+	jal prepare_ftrace_return
+
+	MCOUNT_RESTORE_REGS
+	RETURN_BACK
+STD_ENDPROC(ftrace_graph_caller)
+
+	.global return_to_handler
+return_to_handler:
+	MCOUNT_SAVE_REGS
+
+	jal	ftrace_return_to_handler
+	/* restore the real parent address */
+	move	r11, r0
+
+	MCOUNT_RESTORE_REGS
+	jr	r11
+
+#endif /* CONFIG_FUNCTION_GRAPH_TRACER */
diff --git a/arch/tile/kernel/messaging.c b/arch/tile/kernel/messaging.c
new file mode 100644
index 000000000..7475af3aa
--- /dev/null
+++ b/arch/tile/kernel/messaging.c
@@ -0,0 +1,115 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/percpu.h>
+#include <linux/smp.h>
+#include <linux/hardirq.h>
+#include <linux/ptrace.h>
+#include <asm/hv_driver.h>
+#include <asm/irq_regs.h>
+#include <asm/traps.h>
+#include <hv/hypervisor.h>
+#include <arch/interrupts.h>
+
+/* All messages are stored here */
+static DEFINE_PER_CPU(HV_MsgState, msg_state);
+
+void init_messaging(void)
+{
+	/* Allocate storage for messages in kernel space */
+	HV_MsgState *state = this_cpu_ptr(&msg_state);
+	int rc = hv_register_message_state(state);
+	if (rc != HV_OK)
+		panic("hv_register_message_state: error %d", rc);
+
+	/* Make sure downcall interrupts will be enabled. */
+	arch_local_irq_unmask(INT_INTCTRL_K);
+}
+
+void hv_message_intr(struct pt_regs *regs, int intnum)
+{
+	/*
+	 * We enter with interrupts disabled and leave them disabled,
+	 * to match expectations of called functions (e.g.
+	 * do_ccupdate_local() in mm/slab.c).  This is also consistent
+	 * with normal call entry for device interrupts.
+	 */
+
+	int message[HV_MAX_MESSAGE_SIZE/sizeof(int)];
+	HV_RcvMsgInfo rmi;
+	int nmsgs = 0;
+
+	/* Track time spent here in an interrupt context */
+	struct pt_regs *old_regs = set_irq_regs(regs);
+	irq_enter();
+
+#ifdef CONFIG_DEBUG_STACKOVERFLOW
+	/* Debugging check for stack overflow: less than 1/8th stack free? */
+	{
+		long sp = stack_pointer - (long) current_thread_info();
+		if (unlikely(sp < (sizeof(struct thread_info) + STACK_WARN))) {
+			pr_emerg("%s: stack overflow: %ld\n",
+				 __func__, sp - sizeof(struct thread_info));
+			dump_stack();
+		}
+	}
+#endif
+
+	while (1) {
+		HV_MsgState *state = this_cpu_ptr(&msg_state);
+		rmi = hv_receive_message(*state, (HV_VirtAddr) message,
+					 sizeof(message));
+		if (rmi.msglen == 0)
+			break;
+
+		if (rmi.msglen < 0)
+			panic("hv_receive_message failed: %d", rmi.msglen);
+
+		++nmsgs;
+
+		if (rmi.source == HV_MSG_TILE) {
+			int tag;
+
+			/* we just send tags for now */
+			BUG_ON(rmi.msglen != sizeof(int));
+
+			tag = message[0];
+#ifdef CONFIG_SMP
+			evaluate_message(message[0]);
+#else
+			panic("Received IPI message %d in UP mode", tag);
+#endif
+		} else if (rmi.source == HV_MSG_INTR) {
+			HV_IntrMsg *him = (HV_IntrMsg *)message;
+			struct hv_driver_cb *cb =
+				(struct hv_driver_cb *)him->intarg;
+			cb->callback(cb, him->intdata);
+			__this_cpu_inc(irq_stat.irq_hv_msg_count);
+		}
+	}
+
+	/*
+	 * We shouldn't have gotten a message downcall with no
+	 * messages available.
+	 */
+	if (nmsgs == 0)
+		panic("Message downcall invoked with no messages!");
+
+	/*
+	 * Track time spent against the current process again and
+	 * process any softirqs if they are waiting.
+	 */
+	irq_exit();
+	set_irq_regs(old_regs);
+}
diff --git a/arch/tile/kernel/module.c b/arch/tile/kernel/module.c
new file mode 100644
index 000000000..2305084c9
--- /dev/null
+++ b/arch/tile/kernel/module.c
@@ -0,0 +1,232 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Based on i386 version, copyright (C) 2001 Rusty Russell.
+ */
+
+#include <linux/moduleloader.h>
+#include <linux/elf.h>
+#include <linux/vmalloc.h>
+#include <linux/fs.h>
+#include <linux/string.h>
+#include <linux/kernel.h>
+#include <asm/pgtable.h>
+#include <asm/homecache.h>
+#include <arch/opcode.h>
+
+#ifdef MODULE_DEBUG
+#define DEBUGP printk
+#else
+#define DEBUGP(fmt...)
+#endif
+
+/*
+ * Allocate some address space in the range MEM_MODULE_START to
+ * MEM_MODULE_END and populate it with memory.
+ */
+void *module_alloc(unsigned long size)
+{
+	struct page **pages;
+	pgprot_t prot_rwx = __pgprot(_PAGE_KERNEL | _PAGE_KERNEL_EXEC);
+	struct vm_struct *area;
+	int i = 0;
+	int npages;
+
+	npages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
+	pages = kmalloc(npages * sizeof(struct page *), GFP_KERNEL);
+	if (pages == NULL)
+		return NULL;
+	for (; i < npages; ++i) {
+		pages[i] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
+		if (!pages[i])
+			goto error;
+	}
+
+	area = __get_vm_area(size, VM_ALLOC, MEM_MODULE_START, MEM_MODULE_END);
+	if (!area)
+		goto error;
+	area->nr_pages = npages;
+	area->pages = pages;
+
+	if (map_vm_area(area, prot_rwx, pages)) {
+		vunmap(area->addr);
+		goto error;
+	}
+
+	return area->addr;
+
+error:
+	while (--i >= 0)
+		__free_page(pages[i]);
+	kfree(pages);
+	return NULL;
+}
+
+
+/* Free memory returned from module_alloc */
+void module_memfree(void *module_region)
+{
+	vfree(module_region);
+
+	/* Globally flush the L1 icache. */
+	flush_remote(0, HV_FLUSH_EVICT_L1I, cpu_online_mask,
+		     0, 0, 0, NULL, NULL, 0);
+
+	/*
+	 * FIXME: Add module_arch_freeing_init to trim exception
+	 * table entries.
+	 */
+}
+
+#ifdef __tilegx__
+/*
+ * Validate that the high 16 bits of "value" is just the sign-extension of
+ * the low 48 bits.
+ */
+static int validate_hw2_last(long value, struct module *me)
+{
+	if (((value << 16) >> 16) != value) {
+		pr_warn("module %s: Out of range HW2_LAST value %#lx\n",
+			me->name, value);
+		return 0;
+	}
+	return 1;
+}
+
+/*
+ * Validate that "value" isn't too big to hold in a JumpOff relocation.
+ */
+static int validate_jumpoff(long value)
+{
+	/* Determine size of jump offset. */
+	int shift = __builtin_clzl(get_JumpOff_X1(create_JumpOff_X1(-1)));
+
+	/* Check to see if it fits into the relocation slot. */
+	long f = get_JumpOff_X1(create_JumpOff_X1(value));
+	f = (f << shift) >> shift;
+
+	return f == value;
+}
+#endif
+
+int apply_relocate_add(Elf_Shdr *sechdrs,
+		       const char *strtab,
+		       unsigned int symindex,
+		       unsigned int relsec,
+		       struct module *me)
+{
+	unsigned int i;
+	Elf_Rela *rel = (void *)sechdrs[relsec].sh_addr;
+	Elf_Sym *sym;
+	u64 *location;
+	unsigned long value;
+
+	DEBUGP("Applying relocate section %u to %u\n", relsec,
+	       sechdrs[relsec].sh_info);
+	for (i = 0; i < sechdrs[relsec].sh_size / sizeof(*rel); i++) {
+		/* This is where to make the change */
+		location = (void *)sechdrs[sechdrs[relsec].sh_info].sh_addr
+			+ rel[i].r_offset;
+		/*
+		 * This is the symbol it is referring to.
+		 * Note that all undefined symbols have been resolved.
+		 */
+		sym = (Elf_Sym *)sechdrs[symindex].sh_addr
+			+ ELF_R_SYM(rel[i].r_info);
+		value = sym->st_value + rel[i].r_addend;
+
+		switch (ELF_R_TYPE(rel[i].r_info)) {
+
+#ifdef __LITTLE_ENDIAN
+# define MUNGE(func) \
+	(*location = ((*location & ~func(-1)) | func(value)))
+#else
+/*
+ * Instructions are always little-endian, so when we read them as data,
+ * we have to swap them around before and after modifying them.
+ */
+# define MUNGE(func) \
+	(*location = swab64((swab64(*location) & ~func(-1)) | func(value)))
+#endif
+
+#ifndef __tilegx__
+		case R_TILE_32:
+			*(uint32_t *)location = value;
+			break;
+		case R_TILE_IMM16_X0_HA:
+			value = (value + 0x8000) >> 16;
+			/*FALLTHROUGH*/
+		case R_TILE_IMM16_X0_LO:
+			MUNGE(create_Imm16_X0);
+			break;
+		case R_TILE_IMM16_X1_HA:
+			value = (value + 0x8000) >> 16;
+			/*FALLTHROUGH*/
+		case R_TILE_IMM16_X1_LO:
+			MUNGE(create_Imm16_X1);
+			break;
+		case R_TILE_JOFFLONG_X1:
+			value -= (unsigned long) location;  /* pc-relative */
+			value = (long) value >> 3;     /* count by instrs */
+			MUNGE(create_JOffLong_X1);
+			break;
+#else
+		case R_TILEGX_64:
+			*location = value;
+			break;
+		case R_TILEGX_IMM16_X0_HW2_LAST:
+			if (!validate_hw2_last(value, me))
+				return -ENOEXEC;
+			value >>= 16;
+			/*FALLTHROUGH*/
+		case R_TILEGX_IMM16_X0_HW1:
+			value >>= 16;
+			/*FALLTHROUGH*/
+		case R_TILEGX_IMM16_X0_HW0:
+			MUNGE(create_Imm16_X0);
+			break;
+		case R_TILEGX_IMM16_X1_HW2_LAST:
+			if (!validate_hw2_last(value, me))
+				return -ENOEXEC;
+			value >>= 16;
+			/*FALLTHROUGH*/
+		case R_TILEGX_IMM16_X1_HW1:
+			value >>= 16;
+			/*FALLTHROUGH*/
+		case R_TILEGX_IMM16_X1_HW0:
+			MUNGE(create_Imm16_X1);
+			break;
+		case R_TILEGX_JUMPOFF_X1:
+			value -= (unsigned long) location;  /* pc-relative */
+			value = (long) value >> 3;     /* count by instrs */
+			if (!validate_jumpoff(value)) {
+				pr_warn("module %s: Out of range jump to %#llx at %#llx (%p)\n",
+					me->name,
+					sym->st_value + rel[i].r_addend,
+					rel[i].r_offset, location);
+				return -ENOEXEC;
+			}
+			MUNGE(create_JumpOff_X1);
+			break;
+#endif
+
+#undef MUNGE
+
+		default:
+			pr_err("module %s: Unknown relocation: %d\n",
+			       me->name, (int) ELF_R_TYPE(rel[i].r_info));
+			return -ENOEXEC;
+		}
+	}
+	return 0;
+}
diff --git a/arch/tile/kernel/pci-dma.c b/arch/tile/kernel/pci-dma.c
new file mode 100644
index 000000000..09b58703a
--- /dev/null
+++ b/arch/tile/kernel/pci-dma.c
@@ -0,0 +1,630 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/mm.h>
+#include <linux/dma-mapping.h>
+#include <linux/swiotlb.h>
+#include <linux/vmalloc.h>
+#include <linux/export.h>
+#include <asm/tlbflush.h>
+#include <asm/homecache.h>
+
+/* Generic DMA mapping functions: */
+
+/*
+ * Allocate what Linux calls "coherent" memory.  On TILEPro this is
+ * uncached memory; on TILE-Gx it is hash-for-home memory.
+ */
+#ifdef __tilepro__
+#define PAGE_HOME_DMA PAGE_HOME_UNCACHED
+#else
+#define PAGE_HOME_DMA PAGE_HOME_HASH
+#endif
+
+static void *tile_dma_alloc_coherent(struct device *dev, size_t size,
+				     dma_addr_t *dma_handle, gfp_t gfp,
+				     struct dma_attrs *attrs)
+{
+	u64 dma_mask = (dev && dev->coherent_dma_mask) ?
+		dev->coherent_dma_mask : DMA_BIT_MASK(32);
+	int node = dev ? dev_to_node(dev) : 0;
+	int order = get_order(size);
+	struct page *pg;
+	dma_addr_t addr;
+
+	gfp |= __GFP_ZERO;
+
+	/*
+	 * If the mask specifies that the memory be in the first 4 GB, then
+	 * we force the allocation to come from the DMA zone.  We also
+	 * force the node to 0 since that's the only node where the DMA
+	 * zone isn't empty.  If the mask size is smaller than 32 bits, we
+	 * may still not be able to guarantee a suitable memory address, in
+	 * which case we will return NULL.  But such devices are uncommon.
+	 */
+	if (dma_mask <= DMA_BIT_MASK(32)) {
+		gfp |= GFP_DMA;
+		node = 0;
+	}
+
+	pg = homecache_alloc_pages_node(node, gfp, order, PAGE_HOME_DMA);
+	if (pg == NULL)
+		return NULL;
+
+	addr = page_to_phys(pg);
+	if (addr + size > dma_mask) {
+		__homecache_free_pages(pg, order);
+		return NULL;
+	}
+
+	*dma_handle = addr;
+
+	return page_address(pg);
+}
+
+/*
+ * Free memory that was allocated with tile_dma_alloc_coherent.
+ */
+static void tile_dma_free_coherent(struct device *dev, size_t size,
+				   void *vaddr, dma_addr_t dma_handle,
+				   struct dma_attrs *attrs)
+{
+	homecache_free_pages((unsigned long)vaddr, get_order(size));
+}
+
+/*
+ * The map routines "map" the specified address range for DMA
+ * accesses.  The memory belongs to the device after this call is
+ * issued, until it is unmapped with dma_unmap_single.
+ *
+ * We don't need to do any mapping, we just flush the address range
+ * out of the cache and return a DMA address.
+ *
+ * The unmap routines do whatever is necessary before the processor
+ * accesses the memory again, and must be called before the driver
+ * touches the memory.  We can get away with a cache invalidate if we
+ * can count on nothing having been touched.
+ */
+
+/* Set up a single page for DMA access. */
+static void __dma_prep_page(struct page *page, unsigned long offset,
+			    size_t size, enum dma_data_direction direction)
+{
+	/*
+	 * Flush the page from cache if necessary.
+	 * On tilegx, data is delivered to hash-for-home L3; on tilepro,
+	 * data is delivered direct to memory.
+	 *
+	 * NOTE: If we were just doing DMA_TO_DEVICE we could optimize
+	 * this to be a "flush" not a "finv" and keep some of the
+	 * state in cache across the DMA operation, but it doesn't seem
+	 * worth creating the necessary flush_buffer_xxx() infrastructure.
+	 */
+	int home = page_home(page);
+	switch (home) {
+	case PAGE_HOME_HASH:
+#ifdef __tilegx__
+		return;
+#endif
+		break;
+	case PAGE_HOME_UNCACHED:
+#ifdef __tilepro__
+		return;
+#endif
+		break;
+	case PAGE_HOME_IMMUTABLE:
+		/* Should be going to the device only. */
+		BUG_ON(direction == DMA_FROM_DEVICE ||
+		       direction == DMA_BIDIRECTIONAL);
+		return;
+	case PAGE_HOME_INCOHERENT:
+		/* Incoherent anyway, so no need to work hard here. */
+		return;
+	default:
+		BUG_ON(home < 0 || home >= NR_CPUS);
+		break;
+	}
+	homecache_finv_page(page);
+
+#ifdef DEBUG_ALIGNMENT
+	/* Warn if the region isn't cacheline aligned. */
+	if (offset & (L2_CACHE_BYTES - 1) || (size & (L2_CACHE_BYTES - 1)))
+		pr_warn("Unaligned DMA to non-hfh memory: PA %#llx/%#lx\n",
+			PFN_PHYS(page_to_pfn(page)) + offset, size);
+#endif
+}
+
+/* Make the page ready to be read by the core. */
+static void __dma_complete_page(struct page *page, unsigned long offset,
+				size_t size, enum dma_data_direction direction)
+{
+#ifdef __tilegx__
+	switch (page_home(page)) {
+	case PAGE_HOME_HASH:
+		/* I/O device delivered data the way the cpu wanted it. */
+		break;
+	case PAGE_HOME_INCOHERENT:
+		/* Incoherent anyway, so no need to work hard here. */
+		break;
+	case PAGE_HOME_IMMUTABLE:
+		/* Extra read-only copies are not a problem. */
+		break;
+	default:
+		/* Flush the bogus hash-for-home I/O entries to memory. */
+		homecache_finv_map_page(page, PAGE_HOME_HASH);
+		break;
+	}
+#endif
+}
+
+static void __dma_prep_pa_range(dma_addr_t dma_addr, size_t size,
+				enum dma_data_direction direction)
+{
+	struct page *page = pfn_to_page(PFN_DOWN(dma_addr));
+	unsigned long offset = dma_addr & (PAGE_SIZE - 1);
+	size_t bytes = min(size, (size_t)(PAGE_SIZE - offset));
+
+	while (size != 0) {
+		__dma_prep_page(page, offset, bytes, direction);
+		size -= bytes;
+		++page;
+		offset = 0;
+		bytes = min((size_t)PAGE_SIZE, size);
+	}
+}
+
+static void __dma_complete_pa_range(dma_addr_t dma_addr, size_t size,
+				    enum dma_data_direction direction)
+{
+	struct page *page = pfn_to_page(PFN_DOWN(dma_addr));
+	unsigned long offset = dma_addr & (PAGE_SIZE - 1);
+	size_t bytes = min(size, (size_t)(PAGE_SIZE - offset));
+
+	while (size != 0) {
+		__dma_complete_page(page, offset, bytes, direction);
+		size -= bytes;
+		++page;
+		offset = 0;
+		bytes = min((size_t)PAGE_SIZE, size);
+	}
+}
+
+static int tile_dma_map_sg(struct device *dev, struct scatterlist *sglist,
+			   int nents, enum dma_data_direction direction,
+			   struct dma_attrs *attrs)
+{
+	struct scatterlist *sg;
+	int i;
+
+	BUG_ON(!valid_dma_direction(direction));
+
+	WARN_ON(nents == 0 || sglist->length == 0);
+
+	for_each_sg(sglist, sg, nents, i) {
+		sg->dma_address = sg_phys(sg);
+		__dma_prep_pa_range(sg->dma_address, sg->length, direction);
+#ifdef CONFIG_NEED_SG_DMA_LENGTH
+		sg->dma_length = sg->length;
+#endif
+	}
+
+	return nents;
+}
+
+static void tile_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
+			      int nents, enum dma_data_direction direction,
+			      struct dma_attrs *attrs)
+{
+	struct scatterlist *sg;
+	int i;
+
+	BUG_ON(!valid_dma_direction(direction));
+	for_each_sg(sglist, sg, nents, i) {
+		sg->dma_address = sg_phys(sg);
+		__dma_complete_pa_range(sg->dma_address, sg->length,
+					direction);
+	}
+}
+
+static dma_addr_t tile_dma_map_page(struct device *dev, struct page *page,
+				    unsigned long offset, size_t size,
+				    enum dma_data_direction direction,
+				    struct dma_attrs *attrs)
+{
+	BUG_ON(!valid_dma_direction(direction));
+
+	BUG_ON(offset + size > PAGE_SIZE);
+	__dma_prep_page(page, offset, size, direction);
+
+	return page_to_pa(page) + offset;
+}
+
+static void tile_dma_unmap_page(struct device *dev, dma_addr_t dma_address,
+				size_t size, enum dma_data_direction direction,
+				struct dma_attrs *attrs)
+{
+	BUG_ON(!valid_dma_direction(direction));
+
+	__dma_complete_page(pfn_to_page(PFN_DOWN(dma_address)),
+			    dma_address & (PAGE_SIZE - 1), size, direction);
+}
+
+static void tile_dma_sync_single_for_cpu(struct device *dev,
+					 dma_addr_t dma_handle,
+					 size_t size,
+					 enum dma_data_direction direction)
+{
+	BUG_ON(!valid_dma_direction(direction));
+
+	__dma_complete_pa_range(dma_handle, size, direction);
+}
+
+static void tile_dma_sync_single_for_device(struct device *dev,
+					    dma_addr_t dma_handle, size_t size,
+					    enum dma_data_direction direction)
+{
+	__dma_prep_pa_range(dma_handle, size, direction);
+}
+
+static void tile_dma_sync_sg_for_cpu(struct device *dev,
+				     struct scatterlist *sglist, int nelems,
+				     enum dma_data_direction direction)
+{
+	struct scatterlist *sg;
+	int i;
+
+	BUG_ON(!valid_dma_direction(direction));
+	WARN_ON(nelems == 0 || sglist->length == 0);
+
+	for_each_sg(sglist, sg, nelems, i) {
+		dma_sync_single_for_cpu(dev, sg->dma_address,
+					sg_dma_len(sg), direction);
+	}
+}
+
+static void tile_dma_sync_sg_for_device(struct device *dev,
+					struct scatterlist *sglist, int nelems,
+					enum dma_data_direction direction)
+{
+	struct scatterlist *sg;
+	int i;
+
+	BUG_ON(!valid_dma_direction(direction));
+	WARN_ON(nelems == 0 || sglist->length == 0);
+
+	for_each_sg(sglist, sg, nelems, i) {
+		dma_sync_single_for_device(dev, sg->dma_address,
+					   sg_dma_len(sg), direction);
+	}
+}
+
+static inline int
+tile_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
+{
+	return 0;
+}
+
+static inline int
+tile_dma_supported(struct device *dev, u64 mask)
+{
+	return 1;
+}
+
+static struct dma_map_ops tile_default_dma_map_ops = {
+	.alloc = tile_dma_alloc_coherent,
+	.free = tile_dma_free_coherent,
+	.map_page = tile_dma_map_page,
+	.unmap_page = tile_dma_unmap_page,
+	.map_sg = tile_dma_map_sg,
+	.unmap_sg = tile_dma_unmap_sg,
+	.sync_single_for_cpu = tile_dma_sync_single_for_cpu,
+	.sync_single_for_device = tile_dma_sync_single_for_device,
+	.sync_sg_for_cpu = tile_dma_sync_sg_for_cpu,
+	.sync_sg_for_device = tile_dma_sync_sg_for_device,
+	.mapping_error = tile_dma_mapping_error,
+	.dma_supported = tile_dma_supported
+};
+
+struct dma_map_ops *tile_dma_map_ops = &tile_default_dma_map_ops;
+EXPORT_SYMBOL(tile_dma_map_ops);
+
+/* Generic PCI DMA mapping functions */
+
+static void *tile_pci_dma_alloc_coherent(struct device *dev, size_t size,
+					 dma_addr_t *dma_handle, gfp_t gfp,
+					 struct dma_attrs *attrs)
+{
+	int node = dev_to_node(dev);
+	int order = get_order(size);
+	struct page *pg;
+	dma_addr_t addr;
+
+	gfp |= __GFP_ZERO;
+
+	pg = homecache_alloc_pages_node(node, gfp, order, PAGE_HOME_DMA);
+	if (pg == NULL)
+		return NULL;
+
+	addr = page_to_phys(pg);
+
+	*dma_handle = addr + get_dma_offset(dev);
+
+	return page_address(pg);
+}
+
+/*
+ * Free memory that was allocated with tile_pci_dma_alloc_coherent.
+ */
+static void tile_pci_dma_free_coherent(struct device *dev, size_t size,
+				       void *vaddr, dma_addr_t dma_handle,
+				       struct dma_attrs *attrs)
+{
+	homecache_free_pages((unsigned long)vaddr, get_order(size));
+}
+
+static int tile_pci_dma_map_sg(struct device *dev, struct scatterlist *sglist,
+			       int nents, enum dma_data_direction direction,
+			       struct dma_attrs *attrs)
+{
+	struct scatterlist *sg;
+	int i;
+
+	BUG_ON(!valid_dma_direction(direction));
+
+	WARN_ON(nents == 0 || sglist->length == 0);
+
+	for_each_sg(sglist, sg, nents, i) {
+		sg->dma_address = sg_phys(sg);
+		__dma_prep_pa_range(sg->dma_address, sg->length, direction);
+
+		sg->dma_address = sg->dma_address + get_dma_offset(dev);
+#ifdef CONFIG_NEED_SG_DMA_LENGTH
+		sg->dma_length = sg->length;
+#endif
+	}
+
+	return nents;
+}
+
+static void tile_pci_dma_unmap_sg(struct device *dev,
+				  struct scatterlist *sglist, int nents,
+				  enum dma_data_direction direction,
+				  struct dma_attrs *attrs)
+{
+	struct scatterlist *sg;
+	int i;
+
+	BUG_ON(!valid_dma_direction(direction));
+	for_each_sg(sglist, sg, nents, i) {
+		sg->dma_address = sg_phys(sg);
+		__dma_complete_pa_range(sg->dma_address, sg->length,
+					direction);
+	}
+}
+
+static dma_addr_t tile_pci_dma_map_page(struct device *dev, struct page *page,
+					unsigned long offset, size_t size,
+					enum dma_data_direction direction,
+					struct dma_attrs *attrs)
+{
+	BUG_ON(!valid_dma_direction(direction));
+
+	BUG_ON(offset + size > PAGE_SIZE);
+	__dma_prep_page(page, offset, size, direction);
+
+	return page_to_pa(page) + offset + get_dma_offset(dev);
+}
+
+static void tile_pci_dma_unmap_page(struct device *dev, dma_addr_t dma_address,
+				    size_t size,
+				    enum dma_data_direction direction,
+				    struct dma_attrs *attrs)
+{
+	BUG_ON(!valid_dma_direction(direction));
+
+	dma_address -= get_dma_offset(dev);
+
+	__dma_complete_page(pfn_to_page(PFN_DOWN(dma_address)),
+			    dma_address & (PAGE_SIZE - 1), size, direction);
+}
+
+static void tile_pci_dma_sync_single_for_cpu(struct device *dev,
+					     dma_addr_t dma_handle,
+					     size_t size,
+					     enum dma_data_direction direction)
+{
+	BUG_ON(!valid_dma_direction(direction));
+
+	dma_handle -= get_dma_offset(dev);
+
+	__dma_complete_pa_range(dma_handle, size, direction);
+}
+
+static void tile_pci_dma_sync_single_for_device(struct device *dev,
+						dma_addr_t dma_handle,
+						size_t size,
+						enum dma_data_direction
+						direction)
+{
+	dma_handle -= get_dma_offset(dev);
+
+	__dma_prep_pa_range(dma_handle, size, direction);
+}
+
+static void tile_pci_dma_sync_sg_for_cpu(struct device *dev,
+					 struct scatterlist *sglist,
+					 int nelems,
+					 enum dma_data_direction direction)
+{
+	struct scatterlist *sg;
+	int i;
+
+	BUG_ON(!valid_dma_direction(direction));
+	WARN_ON(nelems == 0 || sglist->length == 0);
+
+	for_each_sg(sglist, sg, nelems, i) {
+		dma_sync_single_for_cpu(dev, sg->dma_address,
+					sg_dma_len(sg), direction);
+	}
+}
+
+static void tile_pci_dma_sync_sg_for_device(struct device *dev,
+					    struct scatterlist *sglist,
+					    int nelems,
+					    enum dma_data_direction direction)
+{
+	struct scatterlist *sg;
+	int i;
+
+	BUG_ON(!valid_dma_direction(direction));
+	WARN_ON(nelems == 0 || sglist->length == 0);
+
+	for_each_sg(sglist, sg, nelems, i) {
+		dma_sync_single_for_device(dev, sg->dma_address,
+					   sg_dma_len(sg), direction);
+	}
+}
+
+static inline int
+tile_pci_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
+{
+	return 0;
+}
+
+static inline int
+tile_pci_dma_supported(struct device *dev, u64 mask)
+{
+	return 1;
+}
+
+static struct dma_map_ops tile_pci_default_dma_map_ops = {
+	.alloc = tile_pci_dma_alloc_coherent,
+	.free = tile_pci_dma_free_coherent,
+	.map_page = tile_pci_dma_map_page,
+	.unmap_page = tile_pci_dma_unmap_page,
+	.map_sg = tile_pci_dma_map_sg,
+	.unmap_sg = tile_pci_dma_unmap_sg,
+	.sync_single_for_cpu = tile_pci_dma_sync_single_for_cpu,
+	.sync_single_for_device = tile_pci_dma_sync_single_for_device,
+	.sync_sg_for_cpu = tile_pci_dma_sync_sg_for_cpu,
+	.sync_sg_for_device = tile_pci_dma_sync_sg_for_device,
+	.mapping_error = tile_pci_dma_mapping_error,
+	.dma_supported = tile_pci_dma_supported
+};
+
+struct dma_map_ops *gx_pci_dma_map_ops = &tile_pci_default_dma_map_ops;
+EXPORT_SYMBOL(gx_pci_dma_map_ops);
+
+/* PCI DMA mapping functions for legacy PCI devices */
+
+#ifdef CONFIG_SWIOTLB
+static void *tile_swiotlb_alloc_coherent(struct device *dev, size_t size,
+					 dma_addr_t *dma_handle, gfp_t gfp,
+					 struct dma_attrs *attrs)
+{
+	gfp |= GFP_DMA;
+	return swiotlb_alloc_coherent(dev, size, dma_handle, gfp);
+}
+
+static void tile_swiotlb_free_coherent(struct device *dev, size_t size,
+				       void *vaddr, dma_addr_t dma_addr,
+				       struct dma_attrs *attrs)
+{
+	swiotlb_free_coherent(dev, size, vaddr, dma_addr);
+}
+
+static struct dma_map_ops pci_swiotlb_dma_ops = {
+	.alloc = tile_swiotlb_alloc_coherent,
+	.free = tile_swiotlb_free_coherent,
+	.map_page = swiotlb_map_page,
+	.unmap_page = swiotlb_unmap_page,
+	.map_sg = swiotlb_map_sg_attrs,
+	.unmap_sg = swiotlb_unmap_sg_attrs,
+	.sync_single_for_cpu = swiotlb_sync_single_for_cpu,
+	.sync_single_for_device = swiotlb_sync_single_for_device,
+	.sync_sg_for_cpu = swiotlb_sync_sg_for_cpu,
+	.sync_sg_for_device = swiotlb_sync_sg_for_device,
+	.dma_supported = swiotlb_dma_supported,
+	.mapping_error = swiotlb_dma_mapping_error,
+};
+
+static struct dma_map_ops pci_hybrid_dma_ops = {
+	.alloc = tile_swiotlb_alloc_coherent,
+	.free = tile_swiotlb_free_coherent,
+	.map_page = tile_pci_dma_map_page,
+	.unmap_page = tile_pci_dma_unmap_page,
+	.map_sg = tile_pci_dma_map_sg,
+	.unmap_sg = tile_pci_dma_unmap_sg,
+	.sync_single_for_cpu = tile_pci_dma_sync_single_for_cpu,
+	.sync_single_for_device = tile_pci_dma_sync_single_for_device,
+	.sync_sg_for_cpu = tile_pci_dma_sync_sg_for_cpu,
+	.sync_sg_for_device = tile_pci_dma_sync_sg_for_device,
+	.mapping_error = tile_pci_dma_mapping_error,
+	.dma_supported = tile_pci_dma_supported
+};
+
+struct dma_map_ops *gx_legacy_pci_dma_map_ops = &pci_swiotlb_dma_ops;
+struct dma_map_ops *gx_hybrid_pci_dma_map_ops = &pci_hybrid_dma_ops;
+#else
+struct dma_map_ops *gx_legacy_pci_dma_map_ops;
+struct dma_map_ops *gx_hybrid_pci_dma_map_ops;
+#endif
+EXPORT_SYMBOL(gx_legacy_pci_dma_map_ops);
+EXPORT_SYMBOL(gx_hybrid_pci_dma_map_ops);
+
+#ifdef CONFIG_ARCH_HAS_DMA_SET_COHERENT_MASK
+int dma_set_coherent_mask(struct device *dev, u64 mask)
+{
+	struct dma_map_ops *dma_ops = get_dma_ops(dev);
+
+	/*
+	 * For PCI devices with 64-bit DMA addressing capability, promote
+	 * the dma_ops to full capability for both streams and consistent
+	 * memory access. For 32-bit capable devices, limit the consistent 
+	 * memory DMA range to max_direct_dma_addr.
+	 */
+	if (dma_ops == gx_pci_dma_map_ops ||
+	    dma_ops == gx_hybrid_pci_dma_map_ops ||
+	    dma_ops == gx_legacy_pci_dma_map_ops) {
+		if (mask == DMA_BIT_MASK(64))
+			set_dma_ops(dev, gx_pci_dma_map_ops);
+		else if (mask > dev->archdata.max_direct_dma_addr)
+			mask = dev->archdata.max_direct_dma_addr;
+	}
+
+	if (!dma_supported(dev, mask))
+		return -EIO;
+	dev->coherent_dma_mask = mask;
+	return 0;
+}
+EXPORT_SYMBOL(dma_set_coherent_mask);
+#endif
+
+#ifdef ARCH_HAS_DMA_GET_REQUIRED_MASK
+/*
+ * The generic dma_get_required_mask() uses the highest physical address
+ * (max_pfn) to provide the hint to the PCI drivers regarding 32-bit or
+ * 64-bit DMA configuration. Since TILEGx has I/O TLB/MMU, allowing the
+ * DMAs to use the full 64-bit PCI address space and not limited by
+ * the physical memory space, we always let the PCI devices use
+ * 64-bit DMA if they have that capability, by returning the 64-bit
+ * DMA mask here. The device driver has the option to use 32-bit DMA if
+ * the device is not capable of 64-bit DMA.
+ */
+u64 dma_get_required_mask(struct device *dev)
+{
+	return DMA_BIT_MASK(64);
+}
+EXPORT_SYMBOL_GPL(dma_get_required_mask);
+#endif
diff --git a/arch/tile/kernel/pci.c b/arch/tile/kernel/pci.c
new file mode 100644
index 000000000..9475a74cd
--- /dev/null
+++ b/arch/tile/kernel/pci.c
@@ -0,0 +1,599 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/kernel.h>
+#include <linux/pci.h>
+#include <linux/delay.h>
+#include <linux/string.h>
+#include <linux/init.h>
+#include <linux/capability.h>
+#include <linux/sched.h>
+#include <linux/errno.h>
+#include <linux/irq.h>
+#include <linux/io.h>
+#include <linux/uaccess.h>
+#include <linux/export.h>
+
+#include <asm/processor.h>
+#include <asm/sections.h>
+#include <asm/byteorder.h>
+#include <asm/hv_driver.h>
+#include <hv/drv_pcie_rc_intf.h>
+
+
+/*
+ * Initialization flow and process
+ * -------------------------------
+ *
+ * This files contains the routines to search for PCI buses,
+ * enumerate the buses, and configure any attached devices.
+ *
+ * There are two entry points here:
+ * 1) tile_pci_init
+ *    This sets up the pci_controller structs, and opens the
+ *    FDs to the hypervisor.  This is called from setup_arch() early
+ *    in the boot process.
+ * 2) pcibios_init
+ *    This probes the PCI bus(es) for any attached hardware.  It's
+ *    called by subsys_initcall.  All of the real work is done by the
+ *    generic Linux PCI layer.
+ *
+ */
+
+static int pci_probe = 1;
+
+/*
+ * This flag tells if the platform is TILEmpower that needs
+ * special configuration for the PLX switch chip.
+ */
+int __write_once tile_plx_gen1;
+
+static struct pci_controller controllers[TILE_NUM_PCIE];
+static int num_controllers;
+static int pci_scan_flags[TILE_NUM_PCIE];
+
+static struct pci_ops tile_cfg_ops;
+
+
+/*
+ * We don't need to worry about the alignment of resources.
+ */
+resource_size_t pcibios_align_resource(void *data, const struct resource *res,
+			    resource_size_t size, resource_size_t align)
+{
+	return res->start;
+}
+EXPORT_SYMBOL(pcibios_align_resource);
+
+/*
+ * Open a FD to the hypervisor PCI device.
+ *
+ * controller_id is the controller number, config type is 0 or 1 for
+ * config0 or config1 operations.
+ */
+static int tile_pcie_open(int controller_id, int config_type)
+{
+	char filename[32];
+	int fd;
+
+	sprintf(filename, "pcie/%d/config%d", controller_id, config_type);
+
+	fd = hv_dev_open((HV_VirtAddr)filename, 0);
+
+	return fd;
+}
+
+
+/*
+ * Get the IRQ numbers from the HV and set up the handlers for them.
+ */
+static int tile_init_irqs(int controller_id, struct pci_controller *controller)
+{
+	char filename[32];
+	int fd;
+	int ret;
+	int x;
+	struct pcie_rc_config rc_config;
+
+	sprintf(filename, "pcie/%d/ctl", controller_id);
+	fd = hv_dev_open((HV_VirtAddr)filename, 0);
+	if (fd < 0) {
+		pr_err("PCI: hv_dev_open(%s) failed\n", filename);
+		return -1;
+	}
+	ret = hv_dev_pread(fd, 0, (HV_VirtAddr)(&rc_config),
+			   sizeof(rc_config), PCIE_RC_CONFIG_MASK_OFF);
+	hv_dev_close(fd);
+	if (ret != sizeof(rc_config)) {
+		pr_err("PCI: wanted %zd bytes, got %d\n",
+		       sizeof(rc_config), ret);
+		return -1;
+	}
+	/* Record irq_base so that we can map INTx to IRQ # later. */
+	controller->irq_base = rc_config.intr;
+
+	for (x = 0; x < 4; x++)
+		tile_irq_activate(rc_config.intr + x,
+				  TILE_IRQ_HW_CLEAR);
+
+	if (rc_config.plx_gen1)
+		controller->plx_gen1 = 1;
+
+	return 0;
+}
+
+/*
+ * First initialization entry point, called from setup_arch().
+ *
+ * Find valid controllers and fill in pci_controller structs for each
+ * of them.
+ *
+ * Returns the number of controllers discovered.
+ */
+int __init tile_pci_init(void)
+{
+	int i;
+
+	if (!pci_probe) {
+		pr_info("PCI: disabled by boot argument\n");
+		return 0;
+	}
+
+	pr_info("PCI: Searching for controllers...\n");
+
+	/* Re-init number of PCIe controllers to support hot-plug feature. */
+	num_controllers = 0;
+
+	/* Do any configuration we need before using the PCIe */
+
+	for (i = 0; i < TILE_NUM_PCIE; i++) {
+		/*
+		 * To see whether we need a real config op based on
+		 * the results of pcibios_init(), to support PCIe hot-plug.
+		 */
+		if (pci_scan_flags[i] == 0) {
+			int hv_cfg_fd0 = -1;
+			int hv_cfg_fd1 = -1;
+			int hv_mem_fd = -1;
+			char name[32];
+			struct pci_controller *controller;
+
+			/*
+			 * Open the fd to the HV.  If it fails then this
+			 * device doesn't exist.
+			 */
+			hv_cfg_fd0 = tile_pcie_open(i, 0);
+			if (hv_cfg_fd0 < 0)
+				continue;
+			hv_cfg_fd1 = tile_pcie_open(i, 1);
+			if (hv_cfg_fd1 < 0) {
+				pr_err("PCI: Couldn't open config fd to HV for controller %d\n",
+				       i);
+				goto err_cont;
+			}
+
+			sprintf(name, "pcie/%d/mem", i);
+			hv_mem_fd = hv_dev_open((HV_VirtAddr)name, 0);
+			if (hv_mem_fd < 0) {
+				pr_err("PCI: Could not open mem fd to HV!\n");
+				goto err_cont;
+			}
+
+			pr_info("PCI: Found PCI controller #%d\n", i);
+
+			controller = &controllers[i];
+
+			controller->index = i;
+			controller->hv_cfg_fd[0] = hv_cfg_fd0;
+			controller->hv_cfg_fd[1] = hv_cfg_fd1;
+			controller->hv_mem_fd = hv_mem_fd;
+			controller->last_busno = 0xff;
+			controller->ops = &tile_cfg_ops;
+
+			num_controllers++;
+			continue;
+
+err_cont:
+			if (hv_cfg_fd0 >= 0)
+				hv_dev_close(hv_cfg_fd0);
+			if (hv_cfg_fd1 >= 0)
+				hv_dev_close(hv_cfg_fd1);
+			if (hv_mem_fd >= 0)
+				hv_dev_close(hv_mem_fd);
+			continue;
+		}
+	}
+
+	/*
+	 * Before using the PCIe, see if we need to do any platform-specific
+	 * configuration, such as the PLX switch Gen 1 issue on TILEmpower.
+	 */
+	for (i = 0; i < num_controllers; i++) {
+		struct pci_controller *controller = &controllers[i];
+
+		if (controller->plx_gen1)
+			tile_plx_gen1 = 1;
+	}
+
+	return num_controllers;
+}
+
+/*
+ * (pin - 1) converts from the PCI standard's [1:4] convention to
+ * a normal [0:3] range.
+ */
+static int tile_map_irq(const struct pci_dev *dev, u8 slot, u8 pin)
+{
+	struct pci_controller *controller =
+		(struct pci_controller *)dev->sysdata;
+	return (pin - 1) + controller->irq_base;
+}
+
+
+static void fixup_read_and_payload_sizes(void)
+{
+	struct pci_dev *dev = NULL;
+	int smallest_max_payload = 0x1; /* Tile maxes out at 256 bytes. */
+	int max_read_size = PCI_EXP_DEVCTL_READRQ_512B;
+	u16 new_values;
+
+	/* Scan for the smallest maximum payload size. */
+	for_each_pci_dev(dev) {
+		if (!pci_is_pcie(dev))
+			continue;
+
+		if (dev->pcie_mpss < smallest_max_payload)
+			smallest_max_payload = dev->pcie_mpss;
+	}
+
+	/* Now, set the max_payload_size for all devices to that value. */
+	new_values = max_read_size | (smallest_max_payload << 5);
+	for_each_pci_dev(dev)
+		pcie_capability_clear_and_set_word(dev, PCI_EXP_DEVCTL,
+				PCI_EXP_DEVCTL_PAYLOAD | PCI_EXP_DEVCTL_READRQ,
+				new_values);
+}
+
+
+/*
+ * Second PCI initialization entry point, called by subsys_initcall.
+ *
+ * The controllers have been set up by the time we get here, by a call to
+ * tile_pci_init.
+ */
+int __init pcibios_init(void)
+{
+	int i;
+
+	pr_info("PCI: Probing PCI hardware\n");
+
+	/*
+	 * Delay a bit in case devices aren't ready.  Some devices are
+	 * known to require at least 20ms here, but we use a more
+	 * conservative value.
+	 */
+	msleep(250);
+
+	/* Scan all of the recorded PCI controllers.  */
+	for (i = 0; i < TILE_NUM_PCIE; i++) {
+		/*
+		 * Do real pcibios init ops if the controller is initialized
+		 * by tile_pci_init() successfully and not initialized by
+		 * pcibios_init() yet to support PCIe hot-plug.
+		 */
+		if (pci_scan_flags[i] == 0 && controllers[i].ops != NULL) {
+			struct pci_controller *controller = &controllers[i];
+			struct pci_bus *bus;
+			LIST_HEAD(resources);
+
+			if (tile_init_irqs(i, controller)) {
+				pr_err("PCI: Could not initialize IRQs\n");
+				continue;
+			}
+
+			pr_info("PCI: initializing controller #%d\n", i);
+
+			pci_add_resource(&resources, &ioport_resource);
+			pci_add_resource(&resources, &iomem_resource);
+			bus = pci_scan_root_bus(NULL, 0, controller->ops,
+						controller, &resources);
+			controller->root_bus = bus;
+			controller->last_busno = bus->busn_res.end;
+		}
+	}
+
+	/* Do machine dependent PCI interrupt routing */
+	pci_fixup_irqs(pci_common_swizzle, tile_map_irq);
+
+	/*
+	 * This comes from the generic Linux PCI driver.
+	 *
+	 * It allocates all of the resources (I/O memory, etc)
+	 * associated with the devices read in above.
+	 */
+	pci_assign_unassigned_resources();
+
+	/* Configure the max_read_size and max_payload_size values. */
+	fixup_read_and_payload_sizes();
+
+	/* Record the I/O resources in the PCI controller structure. */
+	for (i = 0; i < TILE_NUM_PCIE; i++) {
+		/*
+		 * Do real pcibios init ops if the controller is initialized
+		 * by tile_pci_init() successfully and not initialized by
+		 * pcibios_init() yet to support PCIe hot-plug.
+		 */
+		if (pci_scan_flags[i] == 0 && controllers[i].ops != NULL) {
+			struct pci_bus *root_bus = controllers[i].root_bus;
+			struct pci_bus *next_bus;
+			struct pci_dev *dev;
+
+			pci_bus_add_devices(root_bus);
+
+			list_for_each_entry(dev, &root_bus->devices, bus_list) {
+				/*
+				 * Find the PCI host controller, ie. the 1st
+				 * bridge.
+				 */
+				if ((dev->class >> 8) == PCI_CLASS_BRIDGE_PCI &&
+					(PCI_SLOT(dev->devfn) == 0)) {
+					next_bus = dev->subordinate;
+					controllers[i].mem_resources[0] =
+						*next_bus->resource[0];
+					controllers[i].mem_resources[1] =
+						 *next_bus->resource[1];
+					controllers[i].mem_resources[2] =
+						 *next_bus->resource[2];
+
+					/* Setup flags. */
+					pci_scan_flags[i] = 1;
+
+					break;
+				}
+			}
+		}
+	}
+
+	return 0;
+}
+subsys_initcall(pcibios_init);
+
+/*
+ * No bus fixups needed.
+ */
+void pcibios_fixup_bus(struct pci_bus *bus)
+{
+	/* Nothing needs to be done. */
+}
+
+void pcibios_set_master(struct pci_dev *dev)
+{
+	/* No special bus mastering setup handling. */
+}
+
+/* Process any "pci=" kernel boot arguments. */
+char *__init pcibios_setup(char *str)
+{
+	if (!strcmp(str, "off")) {
+		pci_probe = 0;
+		return NULL;
+	}
+	return str;
+}
+
+/*
+ * Enable memory and/or address decoding, as appropriate, for the
+ * device described by the 'dev' struct.
+ *
+ * This is called from the generic PCI layer, and can be called
+ * for bridges or endpoints.
+ */
+int pcibios_enable_device(struct pci_dev *dev, int mask)
+{
+	u16 cmd, old_cmd;
+	u8 header_type;
+	int i;
+	struct resource *r;
+
+	pci_read_config_byte(dev, PCI_HEADER_TYPE, &header_type);
+
+	pci_read_config_word(dev, PCI_COMMAND, &cmd);
+	old_cmd = cmd;
+	if ((header_type & 0x7F) == PCI_HEADER_TYPE_BRIDGE) {
+		/*
+		 * For bridges, we enable both memory and I/O decoding
+		 * in call cases.
+		 */
+		cmd |= PCI_COMMAND_IO;
+		cmd |= PCI_COMMAND_MEMORY;
+	} else {
+		/*
+		 * For endpoints, we enable memory and/or I/O decoding
+		 * only if they have a memory resource of that type.
+		 */
+		for (i = 0; i < 6; i++) {
+			r = &dev->resource[i];
+			if (r->flags & IORESOURCE_UNSET) {
+				pr_err("PCI: Device %s not available because of resource collisions\n",
+				       pci_name(dev));
+				return -EINVAL;
+			}
+			if (r->flags & IORESOURCE_IO)
+				cmd |= PCI_COMMAND_IO;
+			if (r->flags & IORESOURCE_MEM)
+				cmd |= PCI_COMMAND_MEMORY;
+		}
+	}
+
+	/*
+	 * We only write the command if it changed.
+	 */
+	if (cmd != old_cmd)
+		pci_write_config_word(dev, PCI_COMMAND, cmd);
+	return 0;
+}
+
+/****************************************************************
+ *
+ * Tile PCI config space read/write routines
+ *
+ ****************************************************************/
+
+/*
+ * These are the normal read and write ops
+ * These are expanded with macros from  pci_bus_read_config_byte() etc.
+ *
+ * devfn is the combined PCI slot & function.
+ *
+ * offset is in bytes, from the start of config space for the
+ * specified bus & slot.
+ */
+
+static int tile_cfg_read(struct pci_bus *bus, unsigned int devfn, int offset,
+			 int size, u32 *val)
+{
+	struct pci_controller *controller = bus->sysdata;
+	int busnum = bus->number & 0xff;
+	int slot = (devfn >> 3) & 0x1f;
+	int function = devfn & 0x7;
+	u32 addr;
+	int config_mode = 1;
+
+	/*
+	 * There is no bridge between the Tile and bus 0, so we
+	 * use config0 to talk to bus 0.
+	 *
+	 * If we're talking to a bus other than zero then we
+	 * must have found a bridge.
+	 */
+	if (busnum == 0) {
+		/*
+		 * We fake an empty slot for (busnum == 0) && (slot > 0),
+		 * since there is only one slot on bus 0.
+		 */
+		if (slot) {
+			*val = 0xFFFFFFFF;
+			return 0;
+		}
+		config_mode = 0;
+	}
+
+	addr = busnum << 20;		/* Bus in 27:20 */
+	addr |= slot << 15;		/* Slot (device) in 19:15 */
+	addr |= function << 12;		/* Function is in 14:12 */
+	addr |= (offset & 0xFFF);	/* byte address in 0:11 */
+
+	return hv_dev_pread(controller->hv_cfg_fd[config_mode], 0,
+			    (HV_VirtAddr)(val), size, addr);
+}
+
+
+/*
+ * See tile_cfg_read() for relevant comments.
+ * Note that "val" is the value to write, not a pointer to that value.
+ */
+static int tile_cfg_write(struct pci_bus *bus, unsigned int devfn, int offset,
+			  int size, u32 val)
+{
+	struct pci_controller *controller = bus->sysdata;
+	int busnum = bus->number & 0xff;
+	int slot = (devfn >> 3) & 0x1f;
+	int function = devfn & 0x7;
+	u32 addr;
+	int config_mode = 1;
+	HV_VirtAddr valp = (HV_VirtAddr)&val;
+
+	/*
+	 * For bus 0 slot 0 we use config 0 accesses.
+	 */
+	if (busnum == 0) {
+		/*
+		 * We fake an empty slot for (busnum == 0) && (slot > 0),
+		 * since there is only one slot on bus 0.
+		 */
+		if (slot)
+			return 0;
+		config_mode = 0;
+	}
+
+	addr = busnum << 20;		/* Bus in 27:20 */
+	addr |= slot << 15;		/* Slot (device) in 19:15 */
+	addr |= function << 12;		/* Function is in 14:12 */
+	addr |= (offset & 0xFFF);	/* byte address in 0:11 */
+
+#ifdef __BIG_ENDIAN
+	/* Point to the correct part of the 32-bit "val". */
+	valp += 4 - size;
+#endif
+
+	return hv_dev_pwrite(controller->hv_cfg_fd[config_mode], 0,
+			     valp, size, addr);
+}
+
+
+static struct pci_ops tile_cfg_ops = {
+	.read =         tile_cfg_read,
+	.write =        tile_cfg_write,
+};
+
+
+/*
+ * In the following, each PCI controller's mem_resources[1]
+ * represents its (non-prefetchable) PCI memory resource.
+ * mem_resources[0] and mem_resources[2] refer to its PCI I/O and
+ * prefetchable PCI memory resources, respectively.
+ * For more details, see pci_setup_bridge() in setup-bus.c.
+ * By comparing the target PCI memory address against the
+ * end address of controller 0, we can determine the controller
+ * that should accept the PCI memory access.
+ */
+#define TILE_READ(size, type)						\
+type _tile_read##size(unsigned long addr)				\
+{									\
+	type val;							\
+	int idx = 0;							\
+	if (addr > controllers[0].mem_resources[1].end &&		\
+	    addr > controllers[0].mem_resources[2].end)			\
+		idx = 1;                                                \
+	if (hv_dev_pread(controllers[idx].hv_mem_fd, 0,			\
+			 (HV_VirtAddr)(&val), sizeof(type), addr))	\
+		pr_err("PCI: read %zd bytes at 0x%lX failed\n",		\
+		       sizeof(type), addr);				\
+	return val;							\
+}									\
+EXPORT_SYMBOL(_tile_read##size)
+
+TILE_READ(b, u8);
+TILE_READ(w, u16);
+TILE_READ(l, u32);
+TILE_READ(q, u64);
+
+#define TILE_WRITE(size, type)						\
+void _tile_write##size(type val, unsigned long addr)			\
+{									\
+	int idx = 0;							\
+	if (addr > controllers[0].mem_resources[1].end &&		\
+	    addr > controllers[0].mem_resources[2].end)			\
+		idx = 1;                                                \
+	if (hv_dev_pwrite(controllers[idx].hv_mem_fd, 0,		\
+			  (HV_VirtAddr)(&val), sizeof(type), addr))	\
+		pr_err("PCI: write %zd bytes at 0x%lX failed\n",	\
+		       sizeof(type), addr);				\
+}									\
+EXPORT_SYMBOL(_tile_write##size)
+
+TILE_WRITE(b, u8);
+TILE_WRITE(w, u16);
+TILE_WRITE(l, u32);
+TILE_WRITE(q, u64);
diff --git a/arch/tile/kernel/pci_gx.c b/arch/tile/kernel/pci_gx.c
new file mode 100644
index 000000000..b1df847d0
--- /dev/null
+++ b/arch/tile/kernel/pci_gx.c
@@ -0,0 +1,1595 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/kernel.h>
+#include <linux/mmzone.h>
+#include <linux/pci.h>
+#include <linux/delay.h>
+#include <linux/string.h>
+#include <linux/init.h>
+#include <linux/capability.h>
+#include <linux/sched.h>
+#include <linux/errno.h>
+#include <linux/irq.h>
+#include <linux/msi.h>
+#include <linux/io.h>
+#include <linux/uaccess.h>
+#include <linux/ctype.h>
+
+#include <asm/processor.h>
+#include <asm/sections.h>
+#include <asm/byteorder.h>
+
+#include <gxio/iorpc_globals.h>
+#include <gxio/kiorpc.h>
+#include <gxio/trio.h>
+#include <gxio/iorpc_trio.h>
+#include <hv/drv_trio_intf.h>
+
+#include <arch/sim.h>
+
+/*
+ * This file containes the routines to search for PCI buses,
+ * enumerate the buses, and configure any attached devices.
+ */
+
+#define DEBUG_PCI_CFG	0
+
+#if DEBUG_PCI_CFG
+#define TRACE_CFG_WR(size, val, bus, dev, func, offset) \
+	pr_info("CFG WR %d-byte VAL %#x to bus %d dev %d func %d addr %u\n", \
+		size, val, bus, dev, func, offset & 0xFFF);
+#define TRACE_CFG_RD(size, val, bus, dev, func, offset) \
+	pr_info("CFG RD %d-byte VAL %#x from bus %d dev %d func %d addr %u\n", \
+		size, val, bus, dev, func, offset & 0xFFF);
+#else
+#define TRACE_CFG_WR(...)
+#define TRACE_CFG_RD(...)
+#endif
+
+static int pci_probe = 1;
+
+/* Information on the PCIe RC ports configuration. */
+static int pcie_rc[TILEGX_NUM_TRIO][TILEGX_TRIO_PCIES];
+
+/*
+ * On some platforms with one or more Gx endpoint ports, we need to
+ * delay the PCIe RC port probe for a few seconds to work around
+ * a HW PCIe link-training bug. The exact delay is specified with
+ * a kernel boot argument in the form of "pcie_rc_delay=T,P,S",
+ * where T is the TRIO instance number, P is the port number and S is
+ * the delay in seconds. If the argument is specified, but the delay is
+ * not provided, the value will be DEFAULT_RC_DELAY.
+ */
+static int rc_delay[TILEGX_NUM_TRIO][TILEGX_TRIO_PCIES];
+
+/* Default number of seconds that the PCIe RC port probe can be delayed. */
+#define DEFAULT_RC_DELAY	10
+
+/* The PCI I/O space size in each PCI domain. */
+#define IO_SPACE_SIZE		0x10000
+
+/* Provide shorter versions of some very long constant names. */
+#define AUTO_CONFIG_RC	\
+	TRIO_PCIE_INTFC_PORT_CONFIG__STRAP_STATE_VAL_AUTO_CONFIG_RC
+#define AUTO_CONFIG_RC_G1	\
+	TRIO_PCIE_INTFC_PORT_CONFIG__STRAP_STATE_VAL_AUTO_CONFIG_RC_G1
+#define AUTO_CONFIG_EP	\
+	TRIO_PCIE_INTFC_PORT_CONFIG__STRAP_STATE_VAL_AUTO_CONFIG_ENDPOINT
+#define AUTO_CONFIG_EP_G1	\
+	TRIO_PCIE_INTFC_PORT_CONFIG__STRAP_STATE_VAL_AUTO_CONFIG_ENDPOINT_G1
+
+/* Array of the PCIe ports configuration info obtained from the BIB. */
+struct pcie_trio_ports_property pcie_ports[TILEGX_NUM_TRIO];
+
+/* Number of configured TRIO instances. */
+int num_trio_shims;
+
+/* All drivers share the TRIO contexts defined here. */
+gxio_trio_context_t trio_contexts[TILEGX_NUM_TRIO];
+
+/* Pointer to an array of PCIe RC controllers. */
+struct pci_controller pci_controllers[TILEGX_NUM_TRIO * TILEGX_TRIO_PCIES];
+int num_rc_controllers;
+
+static struct pci_ops tile_cfg_ops;
+
+/* Mask of CPUs that should receive PCIe interrupts. */
+static struct cpumask intr_cpus_map;
+
+/* We don't need to worry about the alignment of resources. */
+resource_size_t pcibios_align_resource(void *data, const struct resource *res,
+				       resource_size_t size,
+				       resource_size_t align)
+{
+	return res->start;
+}
+EXPORT_SYMBOL(pcibios_align_resource);
+
+/*
+ * Pick a CPU to receive and handle the PCIe interrupts, based on the IRQ #.
+ * For now, we simply send interrupts to non-dataplane CPUs.
+ * We may implement methods to allow user to specify the target CPUs,
+ * e.g. via boot arguments.
+ */
+static int tile_irq_cpu(int irq)
+{
+	unsigned int count;
+	int i = 0;
+	int cpu;
+
+	count = cpumask_weight(&intr_cpus_map);
+	if (unlikely(count == 0)) {
+		pr_warn("intr_cpus_map empty, interrupts will be delievered to dataplane tiles\n");
+		return irq % (smp_height * smp_width);
+	}
+
+	count = irq % count;
+	for_each_cpu(cpu, &intr_cpus_map) {
+		if (i++ == count)
+			break;
+	}
+	return cpu;
+}
+
+/* Open a file descriptor to the TRIO shim. */
+static int tile_pcie_open(int trio_index)
+{
+	gxio_trio_context_t *context = &trio_contexts[trio_index];
+	int ret;
+	int mac;
+
+	/* This opens a file descriptor to the TRIO shim. */
+	ret = gxio_trio_init(context, trio_index);
+	if (ret < 0)
+		goto gxio_trio_init_failure;
+
+	/* Allocate an ASID for the kernel. */
+	ret = gxio_trio_alloc_asids(context, 1, 0, 0);
+	if (ret < 0) {
+		pr_err("PCI: ASID alloc failure on TRIO %d, give up\n",
+			trio_index);
+		goto asid_alloc_failure;
+	}
+
+	context->asid = ret;
+
+#ifdef USE_SHARED_PCIE_CONFIG_REGION
+	/*
+	 * Alloc a PIO region for config access, shared by all MACs per TRIO.
+	 * This shouldn't fail since the kernel is supposed to the first
+	 * client of the TRIO's PIO regions.
+	 */
+	ret = gxio_trio_alloc_pio_regions(context, 1, 0, 0);
+	if (ret < 0) {
+		pr_err("PCI: CFG PIO alloc failure on TRIO %d, give up\n",
+			trio_index);
+		goto pio_alloc_failure;
+	}
+
+	context->pio_cfg_index = ret;
+
+	/*
+	 * For PIO CFG, the bus_address_hi parameter is 0. The mac parameter
+	 * is also 0 because it is specified in PIO_REGION_SETUP_CFG_ADDR.
+	 */
+	ret = gxio_trio_init_pio_region_aux(context, context->pio_cfg_index,
+		0, 0, HV_TRIO_PIO_FLAG_CONFIG_SPACE);
+	if (ret < 0) {
+		pr_err("PCI: CFG PIO init failure on TRIO %d, give up\n",
+			trio_index);
+		goto pio_alloc_failure;
+	}
+#endif
+
+	/* Get the properties of the PCIe ports on this TRIO instance. */
+	ret = gxio_trio_get_port_property(context, &pcie_ports[trio_index]);
+	if (ret < 0) {
+		pr_err("PCI: PCIE_GET_PORT_PROPERTY failure, error %d, on TRIO %d\n",
+		       ret, trio_index);
+		goto get_port_property_failure;
+	}
+
+	context->mmio_base_mac =
+		iorpc_ioremap(context->fd, 0, HV_TRIO_CONFIG_IOREMAP_SIZE);
+	if (context->mmio_base_mac == NULL) {
+		pr_err("PCI: TRIO config space mapping failure, error %d, on TRIO %d\n",
+		       ret, trio_index);
+		ret = -ENOMEM;
+
+		goto trio_mmio_mapping_failure;
+	}
+
+	/* Check the port strap state which will override the BIB setting. */
+	for (mac = 0; mac < TILEGX_TRIO_PCIES; mac++) {
+		TRIO_PCIE_INTFC_PORT_CONFIG_t port_config;
+		unsigned int reg_offset;
+
+		/* Ignore ports that are not specified in the BIB. */
+		if (!pcie_ports[trio_index].ports[mac].allow_rc &&
+		    !pcie_ports[trio_index].ports[mac].allow_ep)
+			continue;
+
+		reg_offset =
+			(TRIO_PCIE_INTFC_PORT_CONFIG <<
+				TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+			(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_INTERFACE <<
+				TRIO_CFG_REGION_ADDR__INTFC_SHIFT) |
+			(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+		port_config.word =
+			__gxio_mmio_read(context->mmio_base_mac + reg_offset);
+
+		if (port_config.strap_state != AUTO_CONFIG_RC &&
+		    port_config.strap_state != AUTO_CONFIG_RC_G1) {
+			/*
+			 * If this is really intended to be an EP port, record
+			 * it so that the endpoint driver will know about it.
+			 */
+			if (port_config.strap_state == AUTO_CONFIG_EP ||
+			    port_config.strap_state == AUTO_CONFIG_EP_G1)
+				pcie_ports[trio_index].ports[mac].allow_ep = 1;
+		}
+	}
+
+	return ret;
+
+trio_mmio_mapping_failure:
+get_port_property_failure:
+asid_alloc_failure:
+#ifdef USE_SHARED_PCIE_CONFIG_REGION
+pio_alloc_failure:
+#endif
+	hv_dev_close(context->fd);
+gxio_trio_init_failure:
+	context->fd = -1;
+
+	return ret;
+}
+
+static int __init tile_trio_init(void)
+{
+	int i;
+
+	/* We loop over all the TRIO shims. */
+	for (i = 0; i < TILEGX_NUM_TRIO; i++) {
+		if (tile_pcie_open(i) < 0)
+			continue;
+		num_trio_shims++;
+	}
+
+	return 0;
+}
+postcore_initcall(tile_trio_init);
+
+static void tilegx_legacy_irq_ack(struct irq_data *d)
+{
+	__insn_mtspr(SPR_IPI_EVENT_RESET_K, 1UL << d->irq);
+}
+
+static void tilegx_legacy_irq_mask(struct irq_data *d)
+{
+	__insn_mtspr(SPR_IPI_MASK_SET_K, 1UL << d->irq);
+}
+
+static void tilegx_legacy_irq_unmask(struct irq_data *d)
+{
+	__insn_mtspr(SPR_IPI_MASK_RESET_K, 1UL << d->irq);
+}
+
+static struct irq_chip tilegx_legacy_irq_chip = {
+	.name			= "tilegx_legacy_irq",
+	.irq_ack		= tilegx_legacy_irq_ack,
+	.irq_mask		= tilegx_legacy_irq_mask,
+	.irq_unmask		= tilegx_legacy_irq_unmask,
+
+	/* TBD: support set_affinity. */
+};
+
+/*
+ * This is a wrapper function of the kernel level-trigger interrupt
+ * handler handle_level_irq() for PCI legacy interrupts. The TRIO
+ * is configured such that only INTx Assert interrupts are proxied
+ * to Linux which just calls handle_level_irq() after clearing the
+ * MAC INTx Assert status bit associated with this interrupt.
+ */
+static void trio_handle_level_irq(unsigned int irq, struct irq_desc *desc)
+{
+	struct pci_controller *controller = irq_desc_get_handler_data(desc);
+	gxio_trio_context_t *trio_context = controller->trio;
+	uint64_t intx = (uint64_t)irq_desc_get_chip_data(desc);
+	int mac = controller->mac;
+	unsigned int reg_offset;
+	uint64_t level_mask;
+
+	handle_level_irq(irq, desc);
+
+	/*
+	 * Clear the INTx Level status, otherwise future interrupts are
+	 * not sent.
+	 */
+	reg_offset = (TRIO_PCIE_INTFC_MAC_INT_STS <<
+		TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+		(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_INTERFACE <<
+		TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+		(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+	level_mask = TRIO_PCIE_INTFC_MAC_INT_STS__INT_LEVEL_MASK << intx;
+
+	__gxio_mmio_write(trio_context->mmio_base_mac + reg_offset, level_mask);
+}
+
+/*
+ * Create kernel irqs and set up the handlers for the legacy interrupts.
+ * Also some minimum initialization for the MSI support.
+ */
+static int tile_init_irqs(struct pci_controller *controller)
+{
+	int i;
+	int j;
+	int irq;
+	int result;
+
+	cpumask_copy(&intr_cpus_map, cpu_online_mask);
+
+
+	for (i = 0; i < 4; i++) {
+		gxio_trio_context_t *context = controller->trio;
+		int cpu;
+
+		/* Ask the kernel to allocate an IRQ. */
+		irq = irq_alloc_hwirq(-1);
+		if (!irq) {
+			pr_err("PCI: no free irq vectors, failed for %d\n", i);
+			goto free_irqs;
+		}
+		controller->irq_intx_table[i] = irq;
+
+		/* Distribute the 4 IRQs to different tiles. */
+		cpu = tile_irq_cpu(irq);
+
+		/* Configure the TRIO intr binding for this IRQ. */
+		result = gxio_trio_config_legacy_intr(context, cpu_x(cpu),
+						      cpu_y(cpu), KERNEL_PL,
+						      irq, controller->mac, i);
+		if (result < 0) {
+			pr_err("PCI: MAC intx config failed for %d\n", i);
+
+			goto free_irqs;
+		}
+
+		/* Register the IRQ handler with the kernel. */
+		irq_set_chip_and_handler(irq, &tilegx_legacy_irq_chip,
+					trio_handle_level_irq);
+		irq_set_chip_data(irq, (void *)(uint64_t)i);
+		irq_set_handler_data(irq, controller);
+	}
+
+	return 0;
+
+free_irqs:
+	for (j = 0; j < i; j++)
+		irq_free_hwirq(controller->irq_intx_table[j]);
+
+	return -1;
+}
+
+/*
+ * Return 1 if the port is strapped to operate in RC mode.
+ */
+static int
+strapped_for_rc(gxio_trio_context_t *trio_context, int mac)
+{
+	TRIO_PCIE_INTFC_PORT_CONFIG_t port_config;
+	unsigned int reg_offset;
+
+	/* Check the port configuration. */
+	reg_offset =
+		(TRIO_PCIE_INTFC_PORT_CONFIG <<
+			TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+		(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_INTERFACE <<
+			TRIO_CFG_REGION_ADDR__INTFC_SHIFT) |
+		(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+	port_config.word =
+		__gxio_mmio_read(trio_context->mmio_base_mac + reg_offset);
+
+	if (port_config.strap_state == AUTO_CONFIG_RC ||
+	    port_config.strap_state == AUTO_CONFIG_RC_G1)
+		return 1;
+	else
+		return 0;
+}
+
+/*
+ * Find valid controllers and fill in pci_controller structs for each
+ * of them.
+ *
+ * Return the number of controllers discovered.
+ */
+int __init tile_pci_init(void)
+{
+	int ctl_index = 0;
+	int i, j;
+
+	if (!pci_probe) {
+		pr_info("PCI: disabled by boot argument\n");
+		return 0;
+	}
+
+	pr_info("PCI: Searching for controllers...\n");
+
+	if (num_trio_shims == 0 || sim_is_simulator())
+		return 0;
+
+	/*
+	 * Now determine which PCIe ports are configured to operate in RC
+	 * mode. There is a differece in the port configuration capability
+	 * between the Gx36 and Gx72 devices.
+	 *
+	 * The Gx36 has configuration capability for each of the 3 PCIe
+	 * interfaces (disable, auto endpoint, auto RC, etc.).
+	 * On the Gx72, you can only select one of the 3 PCIe interfaces per
+	 * TRIO to train automatically. Further, the allowable training modes
+	 * are reduced to four options (auto endpoint, auto RC, stream x1,
+	 * stream x4).
+	 *
+	 * For Gx36 ports, it must be allowed to be in RC mode by the
+	 * Board Information Block, and the hardware strapping pins must be
+	 * set to RC mode.
+	 *
+	 * For Gx72 ports, the port will operate in RC mode if either of the
+	 * following is true:
+	 * 1. It is allowed to be in RC mode by the Board Information Block,
+	 *    and the BIB doesn't allow the EP mode.
+	 * 2. It is allowed to be in either the RC or the EP mode by the BIB,
+	 *    and the hardware strapping pin is set to RC mode.
+	 */
+	for (i = 0; i < TILEGX_NUM_TRIO; i++) {
+		gxio_trio_context_t *context = &trio_contexts[i];
+
+		if (context->fd < 0)
+			continue;
+
+		for (j = 0; j < TILEGX_TRIO_PCIES; j++) {
+			int is_rc = 0;
+
+			if (pcie_ports[i].is_gx72 &&
+			    pcie_ports[i].ports[j].allow_rc) {
+				if (!pcie_ports[i].ports[j].allow_ep ||
+				    strapped_for_rc(context, j))
+					is_rc = 1;
+			} else if (pcie_ports[i].ports[j].allow_rc &&
+				   strapped_for_rc(context, j)) {
+				is_rc = 1;
+			}
+			if (is_rc) {
+				pcie_rc[i][j] = 1;
+				num_rc_controllers++;
+			}
+		}
+	}
+
+	/* Return if no PCIe ports are configured to operate in RC mode. */
+	if (num_rc_controllers == 0)
+		return 0;
+
+	/* Set the TRIO pointer and MAC index for each PCIe RC port. */
+	for (i = 0; i < TILEGX_NUM_TRIO; i++) {
+		for (j = 0; j < TILEGX_TRIO_PCIES; j++) {
+			if (pcie_rc[i][j]) {
+				pci_controllers[ctl_index].trio =
+					&trio_contexts[i];
+				pci_controllers[ctl_index].mac = j;
+				pci_controllers[ctl_index].trio_index = i;
+				ctl_index++;
+				if (ctl_index == num_rc_controllers)
+					goto out;
+			}
+		}
+	}
+
+out:
+	/* Configure each PCIe RC port. */
+	for (i = 0; i < num_rc_controllers; i++) {
+
+		/* Configure the PCIe MAC to run in RC mode. */
+		struct pci_controller *controller = &pci_controllers[i];
+
+		controller->index = i;
+		controller->ops = &tile_cfg_ops;
+
+		controller->io_space.start = PCIBIOS_MIN_IO +
+			(i * IO_SPACE_SIZE);
+		controller->io_space.end = controller->io_space.start +
+			IO_SPACE_SIZE - 1;
+		BUG_ON(controller->io_space.end > IO_SPACE_LIMIT);
+		controller->io_space.flags = IORESOURCE_IO;
+		snprintf(controller->io_space_name,
+			 sizeof(controller->io_space_name),
+			 "PCI I/O domain %d", i);
+		controller->io_space.name = controller->io_space_name;
+
+		/*
+		 * The PCI memory resource is located above the PA space.
+		 * For every host bridge, the BAR window or the MMIO aperture
+		 * is in range [3GB, 4GB - 1] of a 4GB space beyond the
+		 * PA space.
+		 */
+		controller->mem_offset = TILE_PCI_MEM_START +
+			(i * TILE_PCI_BAR_WINDOW_TOP);
+		controller->mem_space.start = controller->mem_offset +
+			TILE_PCI_BAR_WINDOW_TOP - TILE_PCI_BAR_WINDOW_SIZE;
+		controller->mem_space.end = controller->mem_offset +
+			TILE_PCI_BAR_WINDOW_TOP - 1;
+		controller->mem_space.flags = IORESOURCE_MEM;
+		snprintf(controller->mem_space_name,
+			 sizeof(controller->mem_space_name),
+			 "PCI mem domain %d", i);
+		controller->mem_space.name = controller->mem_space_name;
+	}
+
+	return num_rc_controllers;
+}
+
+/*
+ * (pin - 1) converts from the PCI standard's [1:4] convention to
+ * a normal [0:3] range.
+ */
+static int tile_map_irq(const struct pci_dev *dev, u8 device, u8 pin)
+{
+	struct pci_controller *controller =
+		(struct pci_controller *)dev->sysdata;
+	return controller->irq_intx_table[pin - 1];
+}
+
+static void fixup_read_and_payload_sizes(struct pci_controller *controller)
+{
+	gxio_trio_context_t *trio_context = controller->trio;
+	struct pci_bus *root_bus = controller->root_bus;
+	TRIO_PCIE_RC_DEVICE_CONTROL_t dev_control;
+	TRIO_PCIE_RC_DEVICE_CAP_t rc_dev_cap;
+	unsigned int reg_offset;
+	struct pci_bus *child;
+	int mac;
+	int err;
+
+	mac = controller->mac;
+
+	/* Set our max read request size to be 4KB. */
+	reg_offset =
+		(TRIO_PCIE_RC_DEVICE_CONTROL <<
+			TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+		(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_STANDARD <<
+			TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+		(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+	dev_control.word = __gxio_mmio_read32(trio_context->mmio_base_mac +
+					      reg_offset);
+	dev_control.max_read_req_sz = 5;
+	__gxio_mmio_write32(trio_context->mmio_base_mac + reg_offset,
+			    dev_control.word);
+
+	/*
+	 * Set the max payload size supported by this Gx PCIe MAC.
+	 * Though Gx PCIe supports Max Payload Size of up to 1024 bytes,
+	 * experiments have shown that setting MPS to 256 yields the
+	 * best performance.
+	 */
+	reg_offset =
+		(TRIO_PCIE_RC_DEVICE_CAP <<
+			TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+		(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_STANDARD <<
+			TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+		(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+	rc_dev_cap.word = __gxio_mmio_read32(trio_context->mmio_base_mac +
+					     reg_offset);
+	rc_dev_cap.mps_sup = 1;
+	__gxio_mmio_write32(trio_context->mmio_base_mac + reg_offset,
+			    rc_dev_cap.word);
+
+	/* Configure PCI Express MPS setting. */
+	list_for_each_entry(child, &root_bus->children, node)
+		pcie_bus_configure_settings(child);
+
+	/*
+	 * Set the mac_config register in trio based on the MPS/MRS of the link.
+	 */
+	reg_offset =
+		(TRIO_PCIE_RC_DEVICE_CONTROL <<
+			TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+		(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_STANDARD <<
+			TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+		(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+	dev_control.word = __gxio_mmio_read32(trio_context->mmio_base_mac +
+						reg_offset);
+
+	err = gxio_trio_set_mps_mrs(trio_context,
+				    dev_control.max_payload_size,
+				    dev_control.max_read_req_sz,
+				    mac);
+	if (err < 0) {
+		pr_err("PCI: PCIE_CONFIGURE_MAC_MPS_MRS failure, MAC %d on TRIO %d\n",
+		       mac, controller->trio_index);
+	}
+}
+
+static int setup_pcie_rc_delay(char *str)
+{
+	unsigned long delay = 0;
+	unsigned long trio_index;
+	unsigned long mac;
+
+	if (str == NULL || !isdigit(*str))
+		return -EINVAL;
+	trio_index = simple_strtoul(str, (char **)&str, 10);
+	if (trio_index >= TILEGX_NUM_TRIO)
+		return -EINVAL;
+
+	if (*str != ',')
+		return -EINVAL;
+
+	str++;
+	if (!isdigit(*str))
+		return -EINVAL;
+	mac = simple_strtoul(str, (char **)&str, 10);
+	if (mac >= TILEGX_TRIO_PCIES)
+		return -EINVAL;
+
+	if (*str != '\0') {
+		if (*str != ',')
+			return -EINVAL;
+
+		str++;
+		if (!isdigit(*str))
+			return -EINVAL;
+		delay = simple_strtoul(str, (char **)&str, 10);
+	}
+
+	rc_delay[trio_index][mac] = delay ? : DEFAULT_RC_DELAY;
+	return 0;
+}
+early_param("pcie_rc_delay", setup_pcie_rc_delay);
+
+/* PCI initialization entry point, called by subsys_initcall. */
+int __init pcibios_init(void)
+{
+	resource_size_t offset;
+	LIST_HEAD(resources);
+	int next_busno;
+	int i;
+
+	tile_pci_init();
+
+	if (num_rc_controllers == 0)
+		return 0;
+
+	/*
+	 * Delay a bit in case devices aren't ready.  Some devices are
+	 * known to require at least 20ms here, but we use a more
+	 * conservative value.
+	 */
+	msleep(250);
+
+	/* Scan all of the recorded PCI controllers.  */
+	for (next_busno = 0, i = 0; i < num_rc_controllers; i++) {
+		struct pci_controller *controller = &pci_controllers[i];
+		gxio_trio_context_t *trio_context = controller->trio;
+		TRIO_PCIE_INTFC_PORT_STATUS_t port_status;
+		TRIO_PCIE_INTFC_TX_FIFO_CTL_t tx_fifo_ctl;
+		struct pci_bus *bus;
+		unsigned int reg_offset;
+		unsigned int class_code_revision;
+		int trio_index;
+		int mac;
+		int ret;
+
+		if (trio_context->fd < 0)
+			continue;
+
+		trio_index = controller->trio_index;
+		mac = controller->mac;
+
+		/*
+		 * Check for PCIe link-up status to decide if we need
+		 * to force the link to come up.
+		 */
+		reg_offset =
+			(TRIO_PCIE_INTFC_PORT_STATUS <<
+				TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+			(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_INTERFACE <<
+				TRIO_CFG_REGION_ADDR__INTFC_SHIFT) |
+			(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+		port_status.word =
+			__gxio_mmio_read(trio_context->mmio_base_mac +
+					 reg_offset);
+		if (!port_status.dl_up) {
+			if (rc_delay[trio_index][mac]) {
+				pr_info("Delaying PCIe RC TRIO init %d sec on MAC %d on TRIO %d\n",
+					rc_delay[trio_index][mac], mac,
+					trio_index);
+				msleep(rc_delay[trio_index][mac] * 1000);
+			}
+			ret = gxio_trio_force_rc_link_up(trio_context, mac);
+			if (ret < 0)
+				pr_err("PCI: PCIE_FORCE_LINK_UP failure, MAC %d on TRIO %d\n",
+				       mac, trio_index);
+		}
+
+		pr_info("PCI: Found PCI controller #%d on TRIO %d MAC %d\n",
+			i, trio_index, controller->mac);
+
+		/* Delay the bus probe if needed. */
+		if (rc_delay[trio_index][mac]) {
+			pr_info("Delaying PCIe RC bus enumerating %d sec on MAC %d on TRIO %d\n",
+				rc_delay[trio_index][mac], mac, trio_index);
+			msleep(rc_delay[trio_index][mac] * 1000);
+		} else {
+			/*
+			 * Wait a bit here because some EP devices
+			 * take longer to come up.
+			 */
+			msleep(1000);
+		}
+
+		/* Check for PCIe link-up status again. */
+		port_status.word =
+			__gxio_mmio_read(trio_context->mmio_base_mac +
+					 reg_offset);
+		if (!port_status.dl_up) {
+			if (pcie_ports[trio_index].ports[mac].removable) {
+				pr_info("PCI: link is down, MAC %d on TRIO %d\n",
+					mac, trio_index);
+				pr_info("This is expected if no PCIe card is connected to this link\n");
+			} else
+				pr_err("PCI: link is down, MAC %d on TRIO %d\n",
+				       mac, trio_index);
+			continue;
+		}
+
+		/*
+		 * Ensure that the link can come out of L1 power down state.
+		 * Strictly speaking, this is needed only in the case of
+		 * heavy RC-initiated DMAs.
+		 */
+		reg_offset =
+			(TRIO_PCIE_INTFC_TX_FIFO_CTL <<
+				TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+			(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_INTERFACE <<
+				TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+			(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+		tx_fifo_ctl.word =
+			__gxio_mmio_read(trio_context->mmio_base_mac +
+					 reg_offset);
+		tx_fifo_ctl.min_p_credits = 0;
+		__gxio_mmio_write(trio_context->mmio_base_mac + reg_offset,
+				  tx_fifo_ctl.word);
+
+		/*
+		 * Change the device ID so that Linux bus crawl doesn't confuse
+		 * the internal bridge with any Tilera endpoints.
+		 */
+		reg_offset =
+			(TRIO_PCIE_RC_DEVICE_ID_VEN_ID <<
+				TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+			(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_STANDARD <<
+				TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+			(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+		__gxio_mmio_write32(trio_context->mmio_base_mac + reg_offset,
+				    (TILERA_GX36_RC_DEV_ID <<
+				    TRIO_PCIE_RC_DEVICE_ID_VEN_ID__DEV_ID_SHIFT) |
+				    TILERA_VENDOR_ID);
+
+		/* Set the internal P2P bridge class code. */
+		reg_offset =
+			(TRIO_PCIE_RC_REVISION_ID <<
+				TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+			(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_STANDARD <<
+				TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+			(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+		class_code_revision =
+			__gxio_mmio_read32(trio_context->mmio_base_mac +
+					   reg_offset);
+		class_code_revision = (class_code_revision & 0xff) |
+			(PCI_CLASS_BRIDGE_PCI << 16);
+
+		__gxio_mmio_write32(trio_context->mmio_base_mac +
+				    reg_offset, class_code_revision);
+
+#ifdef USE_SHARED_PCIE_CONFIG_REGION
+
+		/* Map in the MMIO space for the PIO region. */
+		offset = HV_TRIO_PIO_OFFSET(trio_context->pio_cfg_index) |
+			(((unsigned long long)mac) <<
+			TRIO_TILE_PIO_REGION_SETUP_CFG_ADDR__MAC_SHIFT);
+
+#else
+
+		/* Alloc a PIO region for PCI config access per MAC. */
+		ret = gxio_trio_alloc_pio_regions(trio_context, 1, 0, 0);
+		if (ret < 0) {
+			pr_err("PCI: PCI CFG PIO alloc failure for mac %d on TRIO %d, give up\n",
+			       mac, trio_index);
+
+			continue;
+		}
+
+		trio_context->pio_cfg_index[mac] = ret;
+
+		/* For PIO CFG, the bus_address_hi parameter is 0. */
+		ret = gxio_trio_init_pio_region_aux(trio_context,
+			trio_context->pio_cfg_index[mac],
+			mac, 0, HV_TRIO_PIO_FLAG_CONFIG_SPACE);
+		if (ret < 0) {
+			pr_err("PCI: PCI CFG PIO init failure for mac %d on TRIO %d, give up\n",
+			       mac, trio_index);
+
+			continue;
+		}
+
+		offset = HV_TRIO_PIO_OFFSET(trio_context->pio_cfg_index[mac]) |
+			(((unsigned long long)mac) <<
+			TRIO_TILE_PIO_REGION_SETUP_CFG_ADDR__MAC_SHIFT);
+
+#endif
+
+		/*
+		 * To save VMALLOC space, we take advantage of the fact that
+		 * bit 29 in the PIO CFG address format is reserved 0. With
+		 * TRIO_TILE_PIO_REGION_SETUP_CFG_ADDR__MAC_SHIFT being 30,
+		 * this cuts VMALLOC space usage from 1GB to 512MB per mac.
+		 */
+		trio_context->mmio_base_pio_cfg[mac] =
+			iorpc_ioremap(trio_context->fd, offset, (1UL <<
+			(TRIO_TILE_PIO_REGION_SETUP_CFG_ADDR__MAC_SHIFT - 1)));
+		if (trio_context->mmio_base_pio_cfg[mac] == NULL) {
+			pr_err("PCI: PIO map failure for mac %d on TRIO %d\n",
+			       mac, trio_index);
+
+			continue;
+		}
+
+		/* Initialize the PCIe interrupts. */
+		if (tile_init_irqs(controller)) {
+			pr_err("PCI: IRQs init failure for mac %d on TRIO %d\n",
+				mac, trio_index);
+
+			continue;
+		}
+
+		/*
+		 * The PCI memory resource is located above the PA space.
+		 * The memory range for the PCI root bus should not overlap
+		 * with the physical RAM.
+		 */
+		pci_add_resource_offset(&resources, &controller->mem_space,
+					controller->mem_offset);
+		pci_add_resource(&resources, &controller->io_space);
+		controller->first_busno = next_busno;
+		bus = pci_scan_root_bus(NULL, next_busno, controller->ops,
+					controller, &resources);
+		controller->root_bus = bus;
+		next_busno = bus->busn_res.end + 1;
+	}
+
+	/* Do machine dependent PCI interrupt routing */
+	pci_fixup_irqs(pci_common_swizzle, tile_map_irq);
+
+	/*
+	 * This comes from the generic Linux PCI driver.
+	 *
+	 * It allocates all of the resources (I/O memory, etc)
+	 * associated with the devices read in above.
+	 */
+	pci_assign_unassigned_resources();
+
+	/* Record the I/O resources in the PCI controller structure. */
+	for (i = 0; i < num_rc_controllers; i++) {
+		struct pci_controller *controller = &pci_controllers[i];
+		gxio_trio_context_t *trio_context = controller->trio;
+		struct pci_bus *root_bus = pci_controllers[i].root_bus;
+		int ret;
+		int j;
+
+		/*
+		 * Skip controllers that are not properly initialized or
+		 * have down links.
+		 */
+		if (root_bus == NULL)
+			continue;
+
+		/* Configure the max_payload_size values for this domain. */
+		fixup_read_and_payload_sizes(controller);
+
+		/* Alloc a PIO region for PCI memory access for each RC port. */
+		ret = gxio_trio_alloc_pio_regions(trio_context, 1, 0, 0);
+		if (ret < 0) {
+			pr_err("PCI: MEM PIO alloc failure on TRIO %d mac %d, give up\n",
+			       controller->trio_index, controller->mac);
+
+			continue;
+		}
+
+		controller->pio_mem_index = ret;
+
+		/*
+		 * For PIO MEM, the bus_address_hi parameter is hard-coded 0
+		 * because we always assign 32-bit PCI bus BAR ranges.
+		 */
+		ret = gxio_trio_init_pio_region_aux(trio_context,
+						    controller->pio_mem_index,
+						    controller->mac,
+						    0,
+						    0);
+		if (ret < 0) {
+			pr_err("PCI: MEM PIO init failure on TRIO %d mac %d, give up\n",
+			       controller->trio_index, controller->mac);
+
+			continue;
+		}
+
+#ifdef CONFIG_TILE_PCI_IO
+		/*
+		 * Alloc a PIO region for PCI I/O space access for each RC port.
+		 */
+		ret = gxio_trio_alloc_pio_regions(trio_context, 1, 0, 0);
+		if (ret < 0) {
+			pr_err("PCI: I/O PIO alloc failure on TRIO %d mac %d, give up\n",
+			       controller->trio_index, controller->mac);
+
+			continue;
+		}
+
+		controller->pio_io_index = ret;
+
+		/*
+		 * For PIO IO, the bus_address_hi parameter is hard-coded 0
+		 * because PCI I/O address space is 32-bit.
+		 */
+		ret = gxio_trio_init_pio_region_aux(trio_context,
+						    controller->pio_io_index,
+						    controller->mac,
+						    0,
+						    HV_TRIO_PIO_FLAG_IO_SPACE);
+		if (ret < 0) {
+			pr_err("PCI: I/O PIO init failure on TRIO %d mac %d, give up\n",
+			       controller->trio_index, controller->mac);
+
+			continue;
+		}
+#endif
+
+		/*
+		 * Configure a Mem-Map region for each memory controller so
+		 * that Linux can map all of its PA space to the PCI bus.
+		 * Use the IOMMU to handle hash-for-home memory.
+		 */
+		for_each_online_node(j) {
+			unsigned long start_pfn = node_start_pfn[j];
+			unsigned long end_pfn = node_end_pfn[j];
+			unsigned long nr_pages = end_pfn - start_pfn;
+
+			ret = gxio_trio_alloc_memory_maps(trio_context, 1, 0,
+							  0);
+			if (ret < 0) {
+				pr_err("PCI: Mem-Map alloc failure on TRIO %d mac %d for MC %d, give up\n",
+				       controller->trio_index, controller->mac,
+				       j);
+
+				goto alloc_mem_map_failed;
+			}
+
+			controller->mem_maps[j] = ret;
+
+			/*
+			 * Initialize the Mem-Map and the I/O MMU so that all
+			 * the physical memory can be accessed by the endpoint
+			 * devices. The base bus address is set to the base CPA
+			 * of this memory controller plus an offset (see pci.h).
+			 * The region's base VA is set to the base CPA. The
+			 * I/O MMU table essentially translates the CPA to
+			 * the real PA. Implicitly, for node 0, we create
+			 * a separate Mem-Map region that serves as the inbound
+			 * window for legacy 32-bit devices. This is a direct
+			 * map of the low 4GB CPA space.
+			 */
+			ret = gxio_trio_init_memory_map_mmu_aux(trio_context,
+				controller->mem_maps[j],
+				start_pfn << PAGE_SHIFT,
+				nr_pages << PAGE_SHIFT,
+				trio_context->asid,
+				controller->mac,
+				(start_pfn << PAGE_SHIFT) +
+				TILE_PCI_MEM_MAP_BASE_OFFSET,
+				j,
+				GXIO_TRIO_ORDER_MODE_UNORDERED);
+			if (ret < 0) {
+				pr_err("PCI: Mem-Map init failure on TRIO %d mac %d for MC %d, give up\n",
+				       controller->trio_index, controller->mac,
+				       j);
+
+				goto alloc_mem_map_failed;
+			}
+			continue;
+
+alloc_mem_map_failed:
+			break;
+		}
+
+		pci_bus_add_devices(root_bus);
+	}
+
+	return 0;
+}
+subsys_initcall(pcibios_init);
+
+/* No bus fixups needed. */
+void pcibios_fixup_bus(struct pci_bus *bus)
+{
+}
+
+/* Process any "pci=" kernel boot arguments. */
+char *__init pcibios_setup(char *str)
+{
+	if (!strcmp(str, "off")) {
+		pci_probe = 0;
+		return NULL;
+	}
+	return str;
+}
+
+/*
+ * Called for each device after PCI setup is done.
+ * We initialize the PCI device capabilities conservatively, assuming that
+ * all devices can only address the 32-bit DMA space. The exception here is
+ * that the device dma_offset is set to the value that matches the 64-bit
+ * capable devices. This is OK because dma_offset is not used by legacy
+ * dma_ops, nor by the hybrid dma_ops's streaming DMAs, which are 64-bit ops.
+ * This implementation matches the kernel design of setting PCI devices'
+ * coherent_dma_mask to 0xffffffffull by default, allowing the device drivers
+ * to skip calling pci_set_consistent_dma_mask(DMA_BIT_MASK(32)).
+ */
+static void pcibios_fixup_final(struct pci_dev *pdev)
+{
+	set_dma_ops(&pdev->dev, gx_legacy_pci_dma_map_ops);
+	set_dma_offset(&pdev->dev, TILE_PCI_MEM_MAP_BASE_OFFSET);
+	pdev->dev.archdata.max_direct_dma_addr =
+		TILE_PCI_MAX_DIRECT_DMA_ADDRESS;
+	pdev->dev.coherent_dma_mask = TILE_PCI_MAX_DIRECT_DMA_ADDRESS;
+}
+DECLARE_PCI_FIXUP_FINAL(PCI_ANY_ID, PCI_ANY_ID, pcibios_fixup_final);
+
+/* Map a PCI MMIO bus address into VA space. */
+void __iomem *ioremap(resource_size_t phys_addr, unsigned long size)
+{
+	struct pci_controller *controller = NULL;
+	resource_size_t bar_start;
+	resource_size_t bar_end;
+	resource_size_t offset;
+	resource_size_t start;
+	resource_size_t end;
+	int trio_fd;
+	int i;
+
+	start = phys_addr;
+	end = phys_addr + size - 1;
+
+	/*
+	 * By searching phys_addr in each controller's mem_space, we can
+	 * determine the controller that should accept the PCI memory access.
+	 */
+	for (i = 0; i < num_rc_controllers; i++) {
+		/*
+		 * Skip controllers that are not properly initialized or
+		 * have down links.
+		 */
+		if (pci_controllers[i].root_bus == NULL)
+			continue;
+
+		bar_start = pci_controllers[i].mem_space.start;
+		bar_end = pci_controllers[i].mem_space.end;
+
+		if ((start >= bar_start) && (end <= bar_end)) {
+			controller = &pci_controllers[i];
+			break;
+		}
+	}
+
+	if (controller == NULL)
+		return NULL;
+
+	trio_fd = controller->trio->fd;
+
+	/* Convert the resource start to the bus address offset. */
+	start = phys_addr - controller->mem_offset;
+
+	offset = HV_TRIO_PIO_OFFSET(controller->pio_mem_index) + start;
+
+	/* We need to keep the PCI bus address's in-page offset in the VA. */
+	return iorpc_ioremap(trio_fd, offset, size) +
+		(start & (PAGE_SIZE - 1));
+}
+EXPORT_SYMBOL(ioremap);
+
+#ifdef CONFIG_TILE_PCI_IO
+/* Map a PCI I/O address into VA space. */
+void __iomem *ioport_map(unsigned long port, unsigned int size)
+{
+	struct pci_controller *controller = NULL;
+	resource_size_t bar_start;
+	resource_size_t bar_end;
+	resource_size_t offset;
+	resource_size_t start;
+	resource_size_t end;
+	int trio_fd;
+	int i;
+
+	start = port;
+	end = port + size - 1;
+
+	/*
+	 * By searching the port in each controller's io_space, we can
+	 * determine the controller that should accept the PCI I/O access.
+	 */
+	for (i = 0; i < num_rc_controllers; i++) {
+		/*
+		 * Skip controllers that are not properly initialized or
+		 * have down links.
+		 */
+		if (pci_controllers[i].root_bus == NULL)
+			continue;
+
+		bar_start = pci_controllers[i].io_space.start;
+		bar_end = pci_controllers[i].io_space.end;
+
+		if ((start >= bar_start) && (end <= bar_end)) {
+			controller = &pci_controllers[i];
+			break;
+		}
+	}
+
+	if (controller == NULL)
+		return NULL;
+
+	trio_fd = controller->trio->fd;
+
+	/* Convert the resource start to the bus address offset. */
+	port -= controller->io_space.start;
+
+	offset = HV_TRIO_PIO_OFFSET(controller->pio_io_index) + port;
+
+	/* We need to keep the PCI bus address's in-page offset in the VA. */
+	return iorpc_ioremap(trio_fd, offset, size) + (port & (PAGE_SIZE - 1));
+}
+EXPORT_SYMBOL(ioport_map);
+
+void ioport_unmap(void __iomem *addr)
+{
+	iounmap(addr);
+}
+EXPORT_SYMBOL(ioport_unmap);
+#endif
+
+void pci_iounmap(struct pci_dev *dev, void __iomem *addr)
+{
+	iounmap(addr);
+}
+EXPORT_SYMBOL(pci_iounmap);
+
+/****************************************************************
+ *
+ * Tile PCI config space read/write routines
+ *
+ ****************************************************************/
+
+/*
+ * These are the normal read and write ops
+ * These are expanded with macros from  pci_bus_read_config_byte() etc.
+ *
+ * devfn is the combined PCI device & function.
+ *
+ * offset is in bytes, from the start of config space for the
+ * specified bus & device.
+ */
+static int tile_cfg_read(struct pci_bus *bus, unsigned int devfn, int offset,
+			 int size, u32 *val)
+{
+	struct pci_controller *controller = bus->sysdata;
+	gxio_trio_context_t *trio_context = controller->trio;
+	int busnum = bus->number & 0xff;
+	int device = PCI_SLOT(devfn);
+	int function = PCI_FUNC(devfn);
+	int config_type = 1;
+	TRIO_TILE_PIO_REGION_SETUP_CFG_ADDR_t cfg_addr;
+	void *mmio_addr;
+
+	/*
+	 * Map all accesses to the local device on root bus into the
+	 * MMIO space of the MAC. Accesses to the downstream devices
+	 * go to the PIO space.
+	 */
+	if (pci_is_root_bus(bus)) {
+		if (device == 0) {
+			/*
+			 * This is the internal downstream P2P bridge,
+			 * access directly.
+			 */
+			unsigned int reg_offset;
+
+			reg_offset = ((offset & 0xFFF) <<
+				TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+				(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_PROTECTED
+				<< TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+				(controller->mac <<
+					TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+			mmio_addr = trio_context->mmio_base_mac + reg_offset;
+
+			goto valid_device;
+
+		} else {
+			/*
+			 * We fake an empty device for (device > 0),
+			 * since there is only one device on bus 0.
+			 */
+			goto invalid_device;
+		}
+	}
+
+	/*
+	 * Accesses to the directly attached device have to be
+	 * sent as type-0 configs.
+	 */
+	if (busnum == (controller->first_busno + 1)) {
+		/*
+		 * There is only one device off of our built-in P2P bridge.
+		 */
+		if (device != 0)
+			goto invalid_device;
+
+		config_type = 0;
+	}
+
+	cfg_addr.word = 0;
+	cfg_addr.reg_addr = (offset & 0xFFF);
+	cfg_addr.fn = function;
+	cfg_addr.dev = device;
+	cfg_addr.bus = busnum;
+	cfg_addr.type = config_type;
+
+	/*
+	 * Note that we don't set the mac field in cfg_addr because the
+	 * mapping is per port.
+	 */
+	mmio_addr = trio_context->mmio_base_pio_cfg[controller->mac] +
+		cfg_addr.word;
+
+valid_device:
+
+	switch (size) {
+	case 4:
+		*val = __gxio_mmio_read32(mmio_addr);
+		break;
+
+	case 2:
+		*val = __gxio_mmio_read16(mmio_addr);
+		break;
+
+	case 1:
+		*val = __gxio_mmio_read8(mmio_addr);
+		break;
+
+	default:
+		return PCIBIOS_FUNC_NOT_SUPPORTED;
+	}
+
+	TRACE_CFG_RD(size, *val, busnum, device, function, offset);
+
+	return 0;
+
+invalid_device:
+
+	switch (size) {
+	case 4:
+		*val = 0xFFFFFFFF;
+		break;
+
+	case 2:
+		*val = 0xFFFF;
+		break;
+
+	case 1:
+		*val = 0xFF;
+		break;
+
+	default:
+		return PCIBIOS_FUNC_NOT_SUPPORTED;
+	}
+
+	return 0;
+}
+
+
+/*
+ * See tile_cfg_read() for relevent comments.
+ * Note that "val" is the value to write, not a pointer to that value.
+ */
+static int tile_cfg_write(struct pci_bus *bus, unsigned int devfn, int offset,
+			  int size, u32 val)
+{
+	struct pci_controller *controller = bus->sysdata;
+	gxio_trio_context_t *trio_context = controller->trio;
+	int busnum = bus->number & 0xff;
+	int device = PCI_SLOT(devfn);
+	int function = PCI_FUNC(devfn);
+	int config_type = 1;
+	TRIO_TILE_PIO_REGION_SETUP_CFG_ADDR_t cfg_addr;
+	void *mmio_addr;
+	u32 val_32 = (u32)val;
+	u16 val_16 = (u16)val;
+	u8 val_8 = (u8)val;
+
+	/*
+	 * Map all accesses to the local device on root bus into the
+	 * MMIO space of the MAC. Accesses to the downstream devices
+	 * go to the PIO space.
+	 */
+	if (pci_is_root_bus(bus)) {
+		if (device == 0) {
+			/*
+			 * This is the internal downstream P2P bridge,
+			 * access directly.
+			 */
+			unsigned int reg_offset;
+
+			reg_offset = ((offset & 0xFFF) <<
+				TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+				(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_PROTECTED
+				<< TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+				(controller->mac <<
+					TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+			mmio_addr = trio_context->mmio_base_mac + reg_offset;
+
+			goto valid_device;
+
+		} else {
+			/*
+			 * We fake an empty device for (device > 0),
+			 * since there is only one device on bus 0.
+			 */
+			goto invalid_device;
+		}
+	}
+
+	/*
+	 * Accesses to the directly attached device have to be
+	 * sent as type-0 configs.
+	 */
+	if (busnum == (controller->first_busno + 1)) {
+		/*
+		 * There is only one device off of our built-in P2P bridge.
+		 */
+		if (device != 0)
+			goto invalid_device;
+
+		config_type = 0;
+	}
+
+	cfg_addr.word = 0;
+	cfg_addr.reg_addr = (offset & 0xFFF);
+	cfg_addr.fn = function;
+	cfg_addr.dev = device;
+	cfg_addr.bus = busnum;
+	cfg_addr.type = config_type;
+
+	/*
+	 * Note that we don't set the mac field in cfg_addr because the
+	 * mapping is per port.
+	 */
+	mmio_addr = trio_context->mmio_base_pio_cfg[controller->mac] +
+			cfg_addr.word;
+
+valid_device:
+
+	switch (size) {
+	case 4:
+		__gxio_mmio_write32(mmio_addr, val_32);
+		TRACE_CFG_WR(size, val_32, busnum, device, function, offset);
+		break;
+
+	case 2:
+		__gxio_mmio_write16(mmio_addr, val_16);
+		TRACE_CFG_WR(size, val_16, busnum, device, function, offset);
+		break;
+
+	case 1:
+		__gxio_mmio_write8(mmio_addr, val_8);
+		TRACE_CFG_WR(size, val_8, busnum, device, function, offset);
+		break;
+
+	default:
+		return PCIBIOS_FUNC_NOT_SUPPORTED;
+	}
+
+invalid_device:
+
+	return 0;
+}
+
+
+static struct pci_ops tile_cfg_ops = {
+	.read =         tile_cfg_read,
+	.write =        tile_cfg_write,
+};
+
+
+/* MSI support starts here. */
+static unsigned int tilegx_msi_startup(struct irq_data *d)
+{
+	if (d->msi_desc)
+		pci_msi_unmask_irq(d);
+
+	return 0;
+}
+
+static void tilegx_msi_ack(struct irq_data *d)
+{
+	__insn_mtspr(SPR_IPI_EVENT_RESET_K, 1UL << d->irq);
+}
+
+static void tilegx_msi_mask(struct irq_data *d)
+{
+	pci_msi_mask_irq(d);
+	__insn_mtspr(SPR_IPI_MASK_SET_K, 1UL << d->irq);
+}
+
+static void tilegx_msi_unmask(struct irq_data *d)
+{
+	__insn_mtspr(SPR_IPI_MASK_RESET_K, 1UL << d->irq);
+	pci_msi_unmask_irq(d);
+}
+
+static struct irq_chip tilegx_msi_chip = {
+	.name			= "tilegx_msi",
+	.irq_startup		= tilegx_msi_startup,
+	.irq_ack		= tilegx_msi_ack,
+	.irq_mask		= tilegx_msi_mask,
+	.irq_unmask		= tilegx_msi_unmask,
+
+	/* TBD: support set_affinity. */
+};
+
+int arch_setup_msi_irq(struct pci_dev *pdev, struct msi_desc *desc)
+{
+	struct pci_controller *controller;
+	gxio_trio_context_t *trio_context;
+	struct msi_msg msg;
+	int default_irq;
+	uint64_t mem_map_base;
+	uint64_t mem_map_limit;
+	u64 msi_addr;
+	int mem_map;
+	int cpu;
+	int irq;
+	int ret;
+
+	irq = irq_alloc_hwirq(-1);
+	if (!irq)
+		return -ENOSPC;
+
+	/*
+	 * Since we use a 64-bit Mem-Map to accept the MSI write, we fail
+	 * devices that are not capable of generating a 64-bit message address.
+	 * These devices will fall back to using the legacy interrupts.
+	 * Most PCIe endpoint devices do support 64-bit message addressing.
+	 */
+	if (desc->msi_attrib.is_64 == 0) {
+		dev_info(&pdev->dev, "64-bit MSI message address not supported, falling back to legacy interrupts\n");
+
+		ret = -ENOMEM;
+		goto is_64_failure;
+	}
+
+	default_irq = desc->msi_attrib.default_irq;
+	controller = irq_get_handler_data(default_irq);
+
+	BUG_ON(!controller);
+
+	trio_context = controller->trio;
+
+	/*
+	 * Allocate a scatter-queue that will accept the MSI write and
+	 * trigger the TILE-side interrupts. We use the scatter-queue regions
+	 * before the mem map regions, because the latter are needed by more
+	 * applications.
+	 */
+	mem_map = gxio_trio_alloc_scatter_queues(trio_context, 1, 0, 0);
+	if (mem_map >= 0) {
+		TRIO_MAP_SQ_DOORBELL_FMT_t doorbell_template = {{
+			.pop = 0,
+			.doorbell = 1,
+		}};
+
+		mem_map += TRIO_NUM_MAP_MEM_REGIONS;
+		mem_map_base = MEM_MAP_INTR_REGIONS_BASE +
+			mem_map * MEM_MAP_INTR_REGION_SIZE;
+		mem_map_limit = mem_map_base + MEM_MAP_INTR_REGION_SIZE - 1;
+
+		msi_addr = mem_map_base + MEM_MAP_INTR_REGION_SIZE - 8;
+		msg.data = (unsigned int)doorbell_template.word;
+	} else {
+		/* SQ regions are out, allocate from map mem regions. */
+		mem_map = gxio_trio_alloc_memory_maps(trio_context, 1, 0, 0);
+		if (mem_map < 0) {
+			dev_info(&pdev->dev, "%s Mem-Map alloc failure - failed to initialize MSI interrupts - falling back to legacy interrupts\n",
+				 desc->msi_attrib.is_msix ? "MSI-X" : "MSI");
+			ret = -ENOMEM;
+			goto msi_mem_map_alloc_failure;
+		}
+
+		mem_map_base = MEM_MAP_INTR_REGIONS_BASE +
+			mem_map * MEM_MAP_INTR_REGION_SIZE;
+		mem_map_limit = mem_map_base + MEM_MAP_INTR_REGION_SIZE - 1;
+
+		msi_addr = mem_map_base + TRIO_MAP_MEM_REG_INT3 -
+			TRIO_MAP_MEM_REG_INT0;
+
+		msg.data = mem_map;
+	}
+
+	/* We try to distribute different IRQs to different tiles. */
+	cpu = tile_irq_cpu(irq);
+
+	/*
+	 * Now call up to the HV to configure the MSI interrupt and
+	 * set up the IPI binding.
+	 */
+	ret = gxio_trio_config_msi_intr(trio_context, cpu_x(cpu), cpu_y(cpu),
+					KERNEL_PL, irq, controller->mac,
+					mem_map, mem_map_base, mem_map_limit,
+					trio_context->asid);
+	if (ret < 0) {
+		dev_info(&pdev->dev, "HV MSI config failed\n");
+
+		goto hv_msi_config_failure;
+	}
+
+	irq_set_msi_desc(irq, desc);
+
+	msg.address_hi = msi_addr >> 32;
+	msg.address_lo = msi_addr & 0xffffffff;
+
+	pci_write_msi_msg(irq, &msg);
+	irq_set_chip_and_handler(irq, &tilegx_msi_chip, handle_level_irq);
+	irq_set_handler_data(irq, controller);
+
+	return 0;
+
+hv_msi_config_failure:
+	/* Free mem-map */
+msi_mem_map_alloc_failure:
+is_64_failure:
+	irq_free_hwirq(irq);
+	return ret;
+}
+
+void arch_teardown_msi_irq(unsigned int irq)
+{
+	irq_free_hwirq(irq);
+}
diff --git a/arch/tile/kernel/perf_event.c b/arch/tile/kernel/perf_event.c
new file mode 100644
index 000000000..bb509cee3
--- /dev/null
+++ b/arch/tile/kernel/perf_event.c
@@ -0,0 +1,1005 @@
+/*
+ * Copyright 2014 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ *
+ * Perf_events support for Tile processor.
+ *
+ * This code is based upon the x86 perf event
+ * code, which is:
+ *
+ *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
+ *  Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
+ *  Copyright (C) 2009 Jaswinder Singh Rajput
+ *  Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter
+ *  Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
+ *  Copyright (C) 2009 Intel Corporation, <markus.t.metzger@intel.com>
+ *  Copyright (C) 2009 Google, Inc., Stephane Eranian
+ */
+
+#include <linux/kprobes.h>
+#include <linux/kernel.h>
+#include <linux/kdebug.h>
+#include <linux/mutex.h>
+#include <linux/bitmap.h>
+#include <linux/irq.h>
+#include <linux/interrupt.h>
+#include <linux/perf_event.h>
+#include <linux/atomic.h>
+#include <asm/traps.h>
+#include <asm/stack.h>
+#include <asm/pmc.h>
+#include <hv/hypervisor.h>
+
+#define TILE_MAX_COUNTERS	4
+
+#define PERF_COUNT_0_IDX	0
+#define PERF_COUNT_1_IDX	1
+#define AUX_PERF_COUNT_0_IDX	2
+#define AUX_PERF_COUNT_1_IDX	3
+
+struct cpu_hw_events {
+	int			n_events;
+	struct perf_event	*events[TILE_MAX_COUNTERS]; /* counter order */
+	struct perf_event	*event_list[TILE_MAX_COUNTERS]; /* enabled
+								order */
+	int			assign[TILE_MAX_COUNTERS];
+	unsigned long		active_mask[BITS_TO_LONGS(TILE_MAX_COUNTERS)];
+	unsigned long		used_mask;
+};
+
+/* TILE arch specific performance monitor unit */
+struct tile_pmu {
+	const char	*name;
+	int		version;
+	const int	*hw_events;	/* generic hw events table */
+	/* generic hw cache events table */
+	const int	(*cache_events)[PERF_COUNT_HW_CACHE_MAX]
+				       [PERF_COUNT_HW_CACHE_OP_MAX]
+				       [PERF_COUNT_HW_CACHE_RESULT_MAX];
+	int		(*map_hw_event)(u64);	 /*method used to map
+						  hw events */
+	int		(*map_cache_event)(u64); /*method used to map
+						  cache events */
+
+	u64		max_period;		/* max sampling period */
+	u64		cntval_mask;		/* counter width mask */
+	int		cntval_bits;		/* counter width */
+	int		max_events;		/* max generic hw events
+						in map */
+	int		num_counters;		/* number base + aux counters */
+	int		num_base_counters;	/* number base counters */
+};
+
+DEFINE_PER_CPU(u64, perf_irqs);
+static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
+
+#define TILE_OP_UNSUPP		(-1)
+
+#ifndef __tilegx__
+/* TILEPro hardware events map */
+static const int tile_hw_event_map[] = {
+	[PERF_COUNT_HW_CPU_CYCLES]		= 0x01, /* ONE */
+	[PERF_COUNT_HW_INSTRUCTIONS]		= 0x06, /* MP_BUNDLE_RETIRED */
+	[PERF_COUNT_HW_CACHE_REFERENCES]	= TILE_OP_UNSUPP,
+	[PERF_COUNT_HW_CACHE_MISSES]		= TILE_OP_UNSUPP,
+	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= 0x16, /*
+					  MP_CONDITIONAL_BRANCH_ISSUED */
+	[PERF_COUNT_HW_BRANCH_MISSES]		= 0x14, /*
+					  MP_CONDITIONAL_BRANCH_MISSPREDICT */
+	[PERF_COUNT_HW_BUS_CYCLES]		= TILE_OP_UNSUPP,
+};
+#else
+/* TILEGx hardware events map */
+static const int tile_hw_event_map[] = {
+	[PERF_COUNT_HW_CPU_CYCLES]		= 0x181, /* ONE */
+	[PERF_COUNT_HW_INSTRUCTIONS]		= 0xdb, /* INSTRUCTION_BUNDLE */
+	[PERF_COUNT_HW_CACHE_REFERENCES]	= TILE_OP_UNSUPP,
+	[PERF_COUNT_HW_CACHE_MISSES]		= TILE_OP_UNSUPP,
+	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= 0xd9, /*
+						COND_BRANCH_PRED_CORRECT */
+	[PERF_COUNT_HW_BRANCH_MISSES]		= 0xda, /*
+						COND_BRANCH_PRED_INCORRECT */
+	[PERF_COUNT_HW_BUS_CYCLES]		= TILE_OP_UNSUPP,
+};
+#endif
+
+#define C(x) PERF_COUNT_HW_CACHE_##x
+
+/*
+ * Generalized hw caching related hw_event table, filled
+ * in on a per model basis. A value of -1 means
+ * 'not supported', any other value means the
+ * raw hw_event ID.
+ */
+#ifndef __tilegx__
+/* TILEPro hardware cache event map */
+static const int tile_cache_event_map[PERF_COUNT_HW_CACHE_MAX]
+				     [PERF_COUNT_HW_CACHE_OP_MAX]
+				     [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
+[C(L1D)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = 0x21, /* RD_MISS */
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = 0x22, /* WR_MISS */
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(L1I)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = 0x12, /* MP_ICACHE_HIT_ISSUED */
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(LL)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(DTLB)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = 0x1d, /* TLB_CNT */
+		[C(RESULT_MISS)] = 0x20, /* TLB_EXCEPTION */
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(ITLB)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = 0x13, /* MP_ITLB_HIT_ISSUED */
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(BPU)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+};
+#else
+/* TILEGx hardware events map */
+static const int tile_cache_event_map[PERF_COUNT_HW_CACHE_MAX]
+				     [PERF_COUNT_HW_CACHE_OP_MAX]
+				     [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
+[C(L1D)] = {
+	/*
+	 * Like some other architectures (e.g. ARM), the performance
+	 * counters don't differentiate between read and write
+	 * accesses/misses, so this isn't strictly correct, but it's the
+	 * best we can do. Writes and reads get combined.
+	 */
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = 0x44, /* RD_MISS */
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = 0x45, /* WR_MISS */
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(L1I)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(LL)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(DTLB)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = 0x40, /* TLB_CNT */
+		[C(RESULT_MISS)] = 0x43, /* TLB_EXCEPTION */
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = 0x40, /* TLB_CNT */
+		[C(RESULT_MISS)] = 0x43, /* TLB_EXCEPTION */
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(ITLB)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = 0xd4, /* ITLB_MISS_INT */
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = 0xd4, /* ITLB_MISS_INT */
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(BPU)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+};
+#endif
+
+static atomic_t tile_active_events;
+static DEFINE_MUTEX(perf_intr_reserve_mutex);
+
+static int tile_map_hw_event(u64 config);
+static int tile_map_cache_event(u64 config);
+
+static int tile_pmu_handle_irq(struct pt_regs *regs, int fault);
+
+/*
+ * To avoid new_raw_count getting larger then pre_raw_count
+ * in tile_perf_event_update(), we limit the value of max_period to 2^31 - 1.
+ */
+static const struct tile_pmu tilepmu = {
+#ifndef __tilegx__
+	.name = "tilepro",
+#else
+	.name = "tilegx",
+#endif
+	.max_events = ARRAY_SIZE(tile_hw_event_map),
+	.map_hw_event = tile_map_hw_event,
+	.hw_events = tile_hw_event_map,
+	.map_cache_event = tile_map_cache_event,
+	.cache_events = &tile_cache_event_map,
+	.cntval_bits = 32,
+	.cntval_mask = (1ULL << 32) - 1,
+	.max_period = (1ULL << 31) - 1,
+	.num_counters = TILE_MAX_COUNTERS,
+	.num_base_counters = TILE_BASE_COUNTERS,
+};
+
+static const struct tile_pmu *tile_pmu __read_mostly;
+
+/*
+ * Check whether perf event is enabled.
+ */
+int tile_perf_enabled(void)
+{
+	return atomic_read(&tile_active_events) != 0;
+}
+
+/*
+ * Read Performance Counters.
+ */
+static inline u64 read_counter(int idx)
+{
+	u64 val = 0;
+
+	/* __insn_mfspr() only takes an immediate argument */
+	switch (idx) {
+	case PERF_COUNT_0_IDX:
+		val = __insn_mfspr(SPR_PERF_COUNT_0);
+		break;
+	case PERF_COUNT_1_IDX:
+		val = __insn_mfspr(SPR_PERF_COUNT_1);
+		break;
+	case AUX_PERF_COUNT_0_IDX:
+		val = __insn_mfspr(SPR_AUX_PERF_COUNT_0);
+		break;
+	case AUX_PERF_COUNT_1_IDX:
+		val = __insn_mfspr(SPR_AUX_PERF_COUNT_1);
+		break;
+	default:
+		WARN_ON_ONCE(idx > AUX_PERF_COUNT_1_IDX ||
+				idx < PERF_COUNT_0_IDX);
+	}
+
+	return val;
+}
+
+/*
+ * Write Performance Counters.
+ */
+static inline void write_counter(int idx, u64 value)
+{
+	/* __insn_mtspr() only takes an immediate argument */
+	switch (idx) {
+	case PERF_COUNT_0_IDX:
+		__insn_mtspr(SPR_PERF_COUNT_0, value);
+		break;
+	case PERF_COUNT_1_IDX:
+		__insn_mtspr(SPR_PERF_COUNT_1, value);
+		break;
+	case AUX_PERF_COUNT_0_IDX:
+		__insn_mtspr(SPR_AUX_PERF_COUNT_0, value);
+		break;
+	case AUX_PERF_COUNT_1_IDX:
+		__insn_mtspr(SPR_AUX_PERF_COUNT_1, value);
+		break;
+	default:
+		WARN_ON_ONCE(idx > AUX_PERF_COUNT_1_IDX ||
+				idx < PERF_COUNT_0_IDX);
+	}
+}
+
+/*
+ * Enable performance event by setting
+ * Performance Counter Control registers.
+ */
+static inline void tile_pmu_enable_event(struct perf_event *event)
+{
+	struct hw_perf_event *hwc = &event->hw;
+	unsigned long cfg, mask;
+	int shift, idx = hwc->idx;
+
+	/*
+	 * prevent early activation from tile_pmu_start() in hw_perf_enable
+	 */
+
+	if (WARN_ON_ONCE(idx == -1))
+		return;
+
+	if (idx < tile_pmu->num_base_counters)
+		cfg = __insn_mfspr(SPR_PERF_COUNT_CTL);
+	else
+		cfg = __insn_mfspr(SPR_AUX_PERF_COUNT_CTL);
+
+	switch (idx) {
+	case PERF_COUNT_0_IDX:
+	case AUX_PERF_COUNT_0_IDX:
+		mask = TILE_EVENT_MASK;
+		shift = 0;
+		break;
+	case PERF_COUNT_1_IDX:
+	case AUX_PERF_COUNT_1_IDX:
+		mask = TILE_EVENT_MASK << 16;
+		shift = 16;
+		break;
+	default:
+		WARN_ON_ONCE(idx < PERF_COUNT_0_IDX ||
+			idx > AUX_PERF_COUNT_1_IDX);
+		return;
+	}
+
+	/* Clear mask bits to enable the event. */
+	cfg &= ~mask;
+	cfg |= hwc->config << shift;
+
+	if (idx < tile_pmu->num_base_counters)
+		__insn_mtspr(SPR_PERF_COUNT_CTL, cfg);
+	else
+		__insn_mtspr(SPR_AUX_PERF_COUNT_CTL, cfg);
+}
+
+/*
+ * Disable performance event by clearing
+ * Performance Counter Control registers.
+ */
+static inline void tile_pmu_disable_event(struct perf_event *event)
+{
+	struct hw_perf_event *hwc = &event->hw;
+	unsigned long cfg, mask;
+	int idx = hwc->idx;
+
+	if (idx == -1)
+		return;
+
+	if (idx < tile_pmu->num_base_counters)
+		cfg = __insn_mfspr(SPR_PERF_COUNT_CTL);
+	else
+		cfg = __insn_mfspr(SPR_AUX_PERF_COUNT_CTL);
+
+	switch (idx) {
+	case PERF_COUNT_0_IDX:
+	case AUX_PERF_COUNT_0_IDX:
+		mask = TILE_PLM_MASK;
+		break;
+	case PERF_COUNT_1_IDX:
+	case AUX_PERF_COUNT_1_IDX:
+		mask = TILE_PLM_MASK << 16;
+		break;
+	default:
+		WARN_ON_ONCE(idx < PERF_COUNT_0_IDX ||
+			idx > AUX_PERF_COUNT_1_IDX);
+		return;
+	}
+
+	/* Set mask bits to disable the event. */
+	cfg |= mask;
+
+	if (idx < tile_pmu->num_base_counters)
+		__insn_mtspr(SPR_PERF_COUNT_CTL, cfg);
+	else
+		__insn_mtspr(SPR_AUX_PERF_COUNT_CTL, cfg);
+}
+
+/*
+ * Propagate event elapsed time into the generic event.
+ * Can only be executed on the CPU where the event is active.
+ * Returns the delta events processed.
+ */
+static u64 tile_perf_event_update(struct perf_event *event)
+{
+	struct hw_perf_event *hwc = &event->hw;
+	int shift = 64 - tile_pmu->cntval_bits;
+	u64 prev_raw_count, new_raw_count;
+	u64 oldval;
+	int idx = hwc->idx;
+	u64 delta;
+
+	/*
+	 * Careful: an NMI might modify the previous event value.
+	 *
+	 * Our tactic to handle this is to first atomically read and
+	 * exchange a new raw count - then add that new-prev delta
+	 * count to the generic event atomically:
+	 */
+again:
+	prev_raw_count = local64_read(&hwc->prev_count);
+	new_raw_count = read_counter(idx);
+
+	oldval = local64_cmpxchg(&hwc->prev_count, prev_raw_count,
+				 new_raw_count);
+	if (oldval != prev_raw_count)
+		goto again;
+
+	/*
+	 * Now we have the new raw value and have updated the prev
+	 * timestamp already. We can now calculate the elapsed delta
+	 * (event-)time and add that to the generic event.
+	 *
+	 * Careful, not all hw sign-extends above the physical width
+	 * of the count.
+	 */
+	delta = (new_raw_count << shift) - (prev_raw_count << shift);
+	delta >>= shift;
+
+	local64_add(delta, &event->count);
+	local64_sub(delta, &hwc->period_left);
+
+	return new_raw_count;
+}
+
+/*
+ * Set the next IRQ period, based on the hwc->period_left value.
+ * To be called with the event disabled in hw:
+ */
+static int tile_event_set_period(struct perf_event *event)
+{
+	struct hw_perf_event *hwc = &event->hw;
+	int idx = hwc->idx;
+	s64 left = local64_read(&hwc->period_left);
+	s64 period = hwc->sample_period;
+	int ret = 0;
+
+	/*
+	 * If we are way outside a reasonable range then just skip forward:
+	 */
+	if (unlikely(left <= -period)) {
+		left = period;
+		local64_set(&hwc->period_left, left);
+		hwc->last_period = period;
+		ret = 1;
+	}
+
+	if (unlikely(left <= 0)) {
+		left += period;
+		local64_set(&hwc->period_left, left);
+		hwc->last_period = period;
+		ret = 1;
+	}
+	if (left > tile_pmu->max_period)
+		left = tile_pmu->max_period;
+
+	/*
+	 * The hw event starts counting from this event offset,
+	 * mark it to be able to extra future deltas:
+	 */
+	local64_set(&hwc->prev_count, (u64)-left);
+
+	write_counter(idx, (u64)(-left) & tile_pmu->cntval_mask);
+
+	perf_event_update_userpage(event);
+
+	return ret;
+}
+
+/*
+ * Stop the event but do not release the PMU counter
+ */
+static void tile_pmu_stop(struct perf_event *event, int flags)
+{
+	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
+	struct hw_perf_event *hwc = &event->hw;
+	int idx = hwc->idx;
+
+	if (__test_and_clear_bit(idx, cpuc->active_mask)) {
+		tile_pmu_disable_event(event);
+		cpuc->events[hwc->idx] = NULL;
+		WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
+		hwc->state |= PERF_HES_STOPPED;
+	}
+
+	if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
+		/*
+		 * Drain the remaining delta count out of a event
+		 * that we are disabling:
+		 */
+		tile_perf_event_update(event);
+		hwc->state |= PERF_HES_UPTODATE;
+	}
+}
+
+/*
+ * Start an event (without re-assigning counter)
+ */
+static void tile_pmu_start(struct perf_event *event, int flags)
+{
+	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
+	int idx = event->hw.idx;
+
+	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
+		return;
+
+	if (WARN_ON_ONCE(idx == -1))
+		return;
+
+	if (flags & PERF_EF_RELOAD) {
+		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
+		tile_event_set_period(event);
+	}
+
+	event->hw.state = 0;
+
+	cpuc->events[idx] = event;
+	__set_bit(idx, cpuc->active_mask);
+
+	unmask_pmc_interrupts();
+
+	tile_pmu_enable_event(event);
+
+	perf_event_update_userpage(event);
+}
+
+/*
+ * Add a single event to the PMU.
+ *
+ * The event is added to the group of enabled events
+ * but only if it can be scehduled with existing events.
+ */
+static int tile_pmu_add(struct perf_event *event, int flags)
+{
+	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
+	struct hw_perf_event *hwc;
+	unsigned long mask;
+	int b, max_cnt;
+
+	hwc = &event->hw;
+
+	/*
+	 * We are full.
+	 */
+	if (cpuc->n_events == tile_pmu->num_counters)
+		return -ENOSPC;
+
+	cpuc->event_list[cpuc->n_events] = event;
+	cpuc->n_events++;
+
+	hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
+	if (!(flags & PERF_EF_START))
+		hwc->state |= PERF_HES_ARCH;
+
+	/*
+	 * Find first empty counter.
+	 */
+	max_cnt = tile_pmu->num_counters;
+	mask = ~cpuc->used_mask;
+
+	/* Find next free counter. */
+	b = find_next_bit(&mask, max_cnt, 0);
+
+	/* Should not happen. */
+	if (WARN_ON_ONCE(b == max_cnt))
+		return -ENOSPC;
+
+	/*
+	 * Assign counter to event.
+	 */
+	event->hw.idx = b;
+	__set_bit(b, &cpuc->used_mask);
+
+	/*
+	 * Start if requested.
+	 */
+	if (flags & PERF_EF_START)
+		tile_pmu_start(event, PERF_EF_RELOAD);
+
+	return 0;
+}
+
+/*
+ * Delete a single event from the PMU.
+ *
+ * The event is deleted from the group of enabled events.
+ * If it is the last event, disable PMU interrupt.
+ */
+static void tile_pmu_del(struct perf_event *event, int flags)
+{
+	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
+	int i;
+
+	/*
+	 * Remove event from list, compact list if necessary.
+	 */
+	for (i = 0; i < cpuc->n_events; i++) {
+		if (cpuc->event_list[i] == event) {
+			while (++i < cpuc->n_events)
+				cpuc->event_list[i-1] = cpuc->event_list[i];
+			--cpuc->n_events;
+			cpuc->events[event->hw.idx] = NULL;
+			__clear_bit(event->hw.idx, &cpuc->used_mask);
+			tile_pmu_stop(event, PERF_EF_UPDATE);
+			break;
+		}
+	}
+	/*
+	 * If there are no events left, then mask PMU interrupt.
+	 */
+	if (cpuc->n_events == 0)
+		mask_pmc_interrupts();
+	perf_event_update_userpage(event);
+}
+
+/*
+ * Propagate event elapsed time into the event.
+ */
+static inline void tile_pmu_read(struct perf_event *event)
+{
+	tile_perf_event_update(event);
+}
+
+/*
+ * Map generic events to Tile PMU.
+ */
+static int tile_map_hw_event(u64 config)
+{
+	if (config >= tile_pmu->max_events)
+		return -EINVAL;
+	return tile_pmu->hw_events[config];
+}
+
+/*
+ * Map generic hardware cache events to Tile PMU.
+ */
+static int tile_map_cache_event(u64 config)
+{
+	unsigned int cache_type, cache_op, cache_result;
+	int code;
+
+	if (!tile_pmu->cache_events)
+		return -ENOENT;
+
+	cache_type = (config >>  0) & 0xff;
+	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
+		return -EINVAL;
+
+	cache_op = (config >>  8) & 0xff;
+	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
+		return -EINVAL;
+
+	cache_result = (config >> 16) & 0xff;
+	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
+		return -EINVAL;
+
+	code = (*tile_pmu->cache_events)[cache_type][cache_op][cache_result];
+	if (code == TILE_OP_UNSUPP)
+		return -EINVAL;
+
+	return code;
+}
+
+static void tile_event_destroy(struct perf_event *event)
+{
+	if (atomic_dec_return(&tile_active_events) == 0)
+		release_pmc_hardware();
+}
+
+static int __tile_event_init(struct perf_event *event)
+{
+	struct perf_event_attr *attr = &event->attr;
+	struct hw_perf_event *hwc = &event->hw;
+	int code;
+
+	switch (attr->type) {
+	case PERF_TYPE_HARDWARE:
+		code = tile_pmu->map_hw_event(attr->config);
+		break;
+	case PERF_TYPE_HW_CACHE:
+		code = tile_pmu->map_cache_event(attr->config);
+		break;
+	case PERF_TYPE_RAW:
+		code = attr->config & TILE_EVENT_MASK;
+		break;
+	default:
+		/* Should not happen. */
+		return -EOPNOTSUPP;
+	}
+
+	if (code < 0)
+		return code;
+
+	hwc->config = code;
+	hwc->idx = -1;
+
+	if (attr->exclude_user)
+		hwc->config |= TILE_CTL_EXCL_USER;
+
+	if (attr->exclude_kernel)
+		hwc->config |= TILE_CTL_EXCL_KERNEL;
+
+	if (attr->exclude_hv)
+		hwc->config |= TILE_CTL_EXCL_HV;
+
+	if (!hwc->sample_period) {
+		hwc->sample_period = tile_pmu->max_period;
+		hwc->last_period = hwc->sample_period;
+		local64_set(&hwc->period_left, hwc->sample_period);
+	}
+	event->destroy = tile_event_destroy;
+	return 0;
+}
+
+static int tile_event_init(struct perf_event *event)
+{
+	int err = 0;
+	perf_irq_t old_irq_handler = NULL;
+
+	if (atomic_inc_return(&tile_active_events) == 1)
+		old_irq_handler = reserve_pmc_hardware(tile_pmu_handle_irq);
+
+	if (old_irq_handler) {
+		pr_warn("PMC hardware busy (reserved by oprofile)\n");
+
+		atomic_dec(&tile_active_events);
+		return -EBUSY;
+	}
+
+	switch (event->attr.type) {
+	case PERF_TYPE_RAW:
+	case PERF_TYPE_HARDWARE:
+	case PERF_TYPE_HW_CACHE:
+		break;
+
+	default:
+		return -ENOENT;
+	}
+
+	err = __tile_event_init(event);
+	if (err) {
+		if (event->destroy)
+			event->destroy(event);
+	}
+	return err;
+}
+
+static struct pmu tilera_pmu = {
+	.event_init	= tile_event_init,
+	.add		= tile_pmu_add,
+	.del		= tile_pmu_del,
+
+	.start		= tile_pmu_start,
+	.stop		= tile_pmu_stop,
+
+	.read		= tile_pmu_read,
+};
+
+/*
+ * PMU's IRQ handler, PMU has 2 interrupts, they share the same handler.
+ */
+int tile_pmu_handle_irq(struct pt_regs *regs, int fault)
+{
+	struct perf_sample_data data;
+	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
+	struct perf_event *event;
+	struct hw_perf_event *hwc;
+	u64 val;
+	unsigned long status;
+	int bit;
+
+	__this_cpu_inc(perf_irqs);
+
+	if (!atomic_read(&tile_active_events))
+		return 0;
+
+	status = pmc_get_overflow();
+	pmc_ack_overflow(status);
+
+	for_each_set_bit(bit, &status, tile_pmu->num_counters) {
+
+		event = cpuc->events[bit];
+
+		if (!event)
+			continue;
+
+		if (!test_bit(bit, cpuc->active_mask))
+			continue;
+
+		hwc = &event->hw;
+
+		val = tile_perf_event_update(event);
+		if (val & (1ULL << (tile_pmu->cntval_bits - 1)))
+			continue;
+
+		perf_sample_data_init(&data, 0, event->hw.last_period);
+		if (!tile_event_set_period(event))
+			continue;
+
+		if (perf_event_overflow(event, &data, regs))
+			tile_pmu_stop(event, 0);
+	}
+
+	return 0;
+}
+
+static bool __init supported_pmu(void)
+{
+	tile_pmu = &tilepmu;
+	return true;
+}
+
+int __init init_hw_perf_events(void)
+{
+	supported_pmu();
+	perf_pmu_register(&tilera_pmu, "cpu", PERF_TYPE_RAW);
+	return 0;
+}
+arch_initcall(init_hw_perf_events);
+
+/* Callchain handling code. */
+
+/*
+ * Tile specific backtracing code for perf_events.
+ */
+static inline void perf_callchain(struct perf_callchain_entry *entry,
+		    struct pt_regs *regs)
+{
+	struct KBacktraceIterator kbt;
+	unsigned int i;
+
+	/*
+	 * Get the address just after the "jalr" instruction that
+	 * jumps to the handler for a syscall.  When we find this
+	 * address in a backtrace, we silently ignore it, which gives
+	 * us a one-step backtrace connection from the sys_xxx()
+	 * function in the kernel to the xxx() function in libc.
+	 * Otherwise, we lose the ability to properly attribute time
+	 * from the libc calls to the kernel implementations, since
+	 * oprofile only considers PCs from backtraces a pair at a time.
+	 */
+	unsigned long handle_syscall_pc = handle_syscall_link_address();
+
+	KBacktraceIterator_init(&kbt, NULL, regs);
+	kbt.profile = 1;
+
+	/*
+	 * The sample for the pc is already recorded.  Now we are adding the
+	 * address of the callsites on the stack.  Our iterator starts
+	 * with the frame of the (already sampled) call site.  If our
+	 * iterator contained a "return address" field, we could have just
+	 * used it and wouldn't have needed to skip the first
+	 * frame.  That's in effect what the arm and x86 versions do.
+	 * Instead we peel off the first iteration to get the equivalent
+	 * behavior.
+	 */
+
+	if (KBacktraceIterator_end(&kbt))
+		return;
+	KBacktraceIterator_next(&kbt);
+
+	/*
+	 * Set stack depth to 16 for user and kernel space respectively, that
+	 * is, total 32 stack frames.
+	 */
+	for (i = 0; i < 16; ++i) {
+		unsigned long pc;
+		if (KBacktraceIterator_end(&kbt))
+			break;
+		pc = kbt.it.pc;
+		if (pc != handle_syscall_pc)
+			perf_callchain_store(entry, pc);
+		KBacktraceIterator_next(&kbt);
+	}
+}
+
+void perf_callchain_user(struct perf_callchain_entry *entry,
+		    struct pt_regs *regs)
+{
+	perf_callchain(entry, regs);
+}
+
+void perf_callchain_kernel(struct perf_callchain_entry *entry,
+		      struct pt_regs *regs)
+{
+	perf_callchain(entry, regs);
+}
diff --git a/arch/tile/kernel/pmc.c b/arch/tile/kernel/pmc.c
new file mode 100644
index 000000000..db62cc34b
--- /dev/null
+++ b/arch/tile/kernel/pmc.c
@@ -0,0 +1,121 @@
+/*
+ * Copyright 2014 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/errno.h>
+#include <linux/spinlock.h>
+#include <linux/module.h>
+#include <linux/atomic.h>
+#include <linux/interrupt.h>
+
+#include <asm/processor.h>
+#include <asm/pmc.h>
+
+perf_irq_t perf_irq = NULL;
+int handle_perf_interrupt(struct pt_regs *regs, int fault)
+{
+	int retval;
+
+	if (!perf_irq)
+		panic("Unexpected PERF_COUNT interrupt %d\n", fault);
+
+	nmi_enter();
+	retval = perf_irq(regs, fault);
+	nmi_exit();
+	return retval;
+}
+
+/* Reserve PMC hardware if it is available. */
+perf_irq_t reserve_pmc_hardware(perf_irq_t new_perf_irq)
+{
+	return cmpxchg(&perf_irq, NULL, new_perf_irq);
+}
+EXPORT_SYMBOL(reserve_pmc_hardware);
+
+/* Release PMC hardware. */
+void release_pmc_hardware(void)
+{
+	perf_irq = NULL;
+}
+EXPORT_SYMBOL(release_pmc_hardware);
+
+
+/*
+ * Get current overflow status of each performance counter,
+ * and auxiliary performance counter.
+ */
+unsigned long
+pmc_get_overflow(void)
+{
+	unsigned long status;
+
+	/*
+	 * merge base+aux into a single vector
+	 */
+	status = __insn_mfspr(SPR_PERF_COUNT_STS);
+	status |= __insn_mfspr(SPR_AUX_PERF_COUNT_STS) << TILE_BASE_COUNTERS;
+	return status;
+}
+
+/*
+ * Clear the status bit for the corresponding counter, if written
+ * with a one.
+ */
+void
+pmc_ack_overflow(unsigned long status)
+{
+	/*
+	 * clear overflow status by writing ones
+	 */
+	__insn_mtspr(SPR_PERF_COUNT_STS, status);
+	__insn_mtspr(SPR_AUX_PERF_COUNT_STS, status >> TILE_BASE_COUNTERS);
+}
+
+/*
+ * The perf count interrupts are masked and unmasked explicitly,
+ * and only here.  The normal irq_enable() does not enable them,
+ * and irq_disable() does not disable them.  That lets these
+ * routines drive the perf count interrupts orthogonally.
+ *
+ * We also mask the perf count interrupts on entry to the perf count
+ * interrupt handler in assembly code, and by default unmask them
+ * again (with interrupt critical section protection) just before
+ * returning from the interrupt.  If the perf count handler returns
+ * a non-zero error code, then we don't re-enable them before returning.
+ *
+ * For Pro, we rely on both interrupts being in the same word to update
+ * them atomically so we never have one enabled and one disabled.
+ */
+
+#if CHIP_HAS_SPLIT_INTR_MASK()
+# if INT_PERF_COUNT < 32 || INT_AUX_PERF_COUNT < 32
+#  error Fix assumptions about which word PERF_COUNT interrupts are in
+# endif
+#endif
+
+static inline unsigned long long pmc_mask(void)
+{
+	unsigned long long mask = 1ULL << INT_PERF_COUNT;
+	mask |= 1ULL << INT_AUX_PERF_COUNT;
+	return mask;
+}
+
+void unmask_pmc_interrupts(void)
+{
+	interrupt_mask_reset_mask(pmc_mask());
+}
+
+void mask_pmc_interrupts(void)
+{
+	interrupt_mask_set_mask(pmc_mask());
+}
diff --git a/arch/tile/kernel/proc.c b/arch/tile/kernel/proc.c
new file mode 100644
index 000000000..7983e9868
--- /dev/null
+++ b/arch/tile/kernel/proc.c
@@ -0,0 +1,160 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/smp.h>
+#include <linux/seq_file.h>
+#include <linux/threads.h>
+#include <linux/cpumask.h>
+#include <linux/timex.h>
+#include <linux/delay.h>
+#include <linux/fs.h>
+#include <linux/proc_fs.h>
+#include <linux/sysctl.h>
+#include <linux/hardirq.h>
+#include <linux/hugetlb.h>
+#include <linux/mman.h>
+#include <asm/unaligned.h>
+#include <asm/pgtable.h>
+#include <asm/processor.h>
+#include <asm/sections.h>
+#include <asm/homecache.h>
+#include <asm/hardwall.h>
+#include <arch/chip.h>
+
+
+/*
+ * Support /proc/cpuinfo
+ */
+
+#define cpu_to_ptr(n) ((void *)((long)(n)+1))
+#define ptr_to_cpu(p) ((long)(p) - 1)
+
+static int show_cpuinfo(struct seq_file *m, void *v)
+{
+	int n = ptr_to_cpu(v);
+
+	if (n == 0) {
+		seq_printf(m, "cpu count\t: %d\n", num_online_cpus());
+		seq_printf(m, "cpu list\t: %*pbl\n",
+			   cpumask_pr_args(cpu_online_mask));
+		seq_printf(m, "model name\t: %s\n", chip_model);
+		seq_printf(m, "flags\t\t:\n");  /* nothing for now */
+		seq_printf(m, "cpu MHz\t\t: %llu.%06llu\n",
+			   get_clock_rate() / 1000000,
+			   (get_clock_rate() % 1000000));
+		seq_printf(m, "bogomips\t: %lu.%02lu\n\n",
+			   loops_per_jiffy/(500000/HZ),
+			   (loops_per_jiffy/(5000/HZ)) % 100);
+	}
+
+#ifdef CONFIG_SMP
+	if (!cpu_online(n))
+		return 0;
+#endif
+
+	seq_printf(m, "processor\t: %d\n", n);
+
+	/* Print only num_online_cpus() blank lines total. */
+	if (cpumask_next(n, cpu_online_mask) < nr_cpu_ids)
+		seq_printf(m, "\n");
+
+	return 0;
+}
+
+static void *c_start(struct seq_file *m, loff_t *pos)
+{
+	return *pos < nr_cpu_ids ? cpu_to_ptr(*pos) : NULL;
+}
+static void *c_next(struct seq_file *m, void *v, loff_t *pos)
+{
+	++*pos;
+	return c_start(m, pos);
+}
+static void c_stop(struct seq_file *m, void *v)
+{
+}
+const struct seq_operations cpuinfo_op = {
+	.start	= c_start,
+	.next	= c_next,
+	.stop	= c_stop,
+	.show	= show_cpuinfo,
+};
+
+/*
+ * Support /proc/tile directory
+ */
+
+static int __init proc_tile_init(void)
+{
+	struct proc_dir_entry *root = proc_mkdir("tile", NULL);
+	if (root == NULL)
+		return 0;
+
+	proc_tile_hardwall_init(root);
+
+	return 0;
+}
+
+arch_initcall(proc_tile_init);
+
+/*
+ * Support /proc/sys/tile directory
+ */
+
+static struct ctl_table unaligned_subtable[] = {
+	{
+		.procname	= "enabled",
+		.data		= &unaligned_fixup,
+		.maxlen		= sizeof(int),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec
+	},
+	{
+		.procname	= "printk",
+		.data		= &unaligned_printk,
+		.maxlen		= sizeof(int),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec
+	},
+	{
+		.procname	= "count",
+		.data		= &unaligned_fixup_count,
+		.maxlen		= sizeof(int),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec
+	},
+	{}
+};
+
+static struct ctl_table unaligned_table[] = {
+	{
+		.procname	= "unaligned_fixup",
+		.mode		= 0555,
+		.child		= unaligned_subtable
+	},
+	{}
+};
+
+static struct ctl_path tile_path[] = {
+	{ .procname = "tile" },
+	{ }
+};
+
+static int __init proc_sys_tile_init(void)
+{
+	register_sysctl_paths(tile_path, unaligned_table);
+	return 0;
+}
+
+arch_initcall(proc_sys_tile_init);
diff --git a/arch/tile/kernel/process.c b/arch/tile/kernel/process.c
new file mode 100644
index 000000000..b403c2e3e
--- /dev/null
+++ b/arch/tile/kernel/process.c
@@ -0,0 +1,576 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/sched.h>
+#include <linux/preempt.h>
+#include <linux/module.h>
+#include <linux/fs.h>
+#include <linux/kprobes.h>
+#include <linux/elfcore.h>
+#include <linux/tick.h>
+#include <linux/init.h>
+#include <linux/mm.h>
+#include <linux/compat.h>
+#include <linux/hardirq.h>
+#include <linux/syscalls.h>
+#include <linux/kernel.h>
+#include <linux/tracehook.h>
+#include <linux/signal.h>
+#include <linux/context_tracking.h>
+#include <asm/stack.h>
+#include <asm/switch_to.h>
+#include <asm/homecache.h>
+#include <asm/syscalls.h>
+#include <asm/traps.h>
+#include <asm/setup.h>
+#include <asm/uaccess.h>
+#ifdef CONFIG_HARDWALL
+#include <asm/hardwall.h>
+#endif
+#include <arch/chip.h>
+#include <arch/abi.h>
+#include <arch/sim_def.h>
+
+/*
+ * Use the (x86) "idle=poll" option to prefer low latency when leaving the
+ * idle loop over low power while in the idle loop, e.g. if we have
+ * one thread per core and we want to get threads out of futex waits fast.
+ */
+static int __init idle_setup(char *str)
+{
+	if (!str)
+		return -EINVAL;
+
+	if (!strcmp(str, "poll")) {
+		pr_info("using polling idle threads\n");
+		cpu_idle_poll_ctrl(true);
+		return 0;
+	} else if (!strcmp(str, "halt")) {
+		return 0;
+	}
+	return -1;
+}
+early_param("idle", idle_setup);
+
+void arch_cpu_idle(void)
+{
+	__this_cpu_write(irq_stat.idle_timestamp, jiffies);
+	_cpu_idle();
+}
+
+/*
+ * Release a thread_info structure
+ */
+void arch_release_thread_info(struct thread_info *info)
+{
+	struct single_step_state *step_state = info->step_state;
+
+	if (step_state) {
+
+		/*
+		 * FIXME: we don't munmap step_state->buffer
+		 * because the mm_struct for this process (info->task->mm)
+		 * has already been zeroed in exit_mm().  Keeping a
+		 * reference to it here seems like a bad move, so this
+		 * means we can't munmap() the buffer, and therefore if we
+		 * ptrace multiple threads in a process, we will slowly
+		 * leak user memory.  (Note that as soon as the last
+		 * thread in a process dies, we will reclaim all user
+		 * memory including single-step buffers in the usual way.)
+		 * We should either assign a kernel VA to this buffer
+		 * somehow, or we should associate the buffer(s) with the
+		 * mm itself so we can clean them up that way.
+		 */
+		kfree(step_state);
+	}
+}
+
+static void save_arch_state(struct thread_struct *t);
+
+int copy_thread(unsigned long clone_flags, unsigned long sp,
+		unsigned long arg, struct task_struct *p)
+{
+	struct pt_regs *childregs = task_pt_regs(p);
+	unsigned long ksp;
+	unsigned long *callee_regs;
+
+	/*
+	 * Set up the stack and stack pointer appropriately for the
+	 * new child to find itself woken up in __switch_to().
+	 * The callee-saved registers must be on the stack to be read;
+	 * the new task will then jump to assembly support to handle
+	 * calling schedule_tail(), etc., and (for userspace tasks)
+	 * returning to the context set up in the pt_regs.
+	 */
+	ksp = (unsigned long) childregs;
+	ksp -= C_ABI_SAVE_AREA_SIZE;   /* interrupt-entry save area */
+	((long *)ksp)[0] = ((long *)ksp)[1] = 0;
+	ksp -= CALLEE_SAVED_REGS_COUNT * sizeof(unsigned long);
+	callee_regs = (unsigned long *)ksp;
+	ksp -= C_ABI_SAVE_AREA_SIZE;   /* __switch_to() save area */
+	((long *)ksp)[0] = ((long *)ksp)[1] = 0;
+	p->thread.ksp = ksp;
+
+	/* Record the pid of the task that created this one. */
+	p->thread.creator_pid = current->pid;
+
+	if (unlikely(p->flags & PF_KTHREAD)) {
+		/* kernel thread */
+		memset(childregs, 0, sizeof(struct pt_regs));
+		memset(&callee_regs[2], 0,
+		       (CALLEE_SAVED_REGS_COUNT - 2) * sizeof(unsigned long));
+		callee_regs[0] = sp;   /* r30 = function */
+		callee_regs[1] = arg;  /* r31 = arg */
+		childregs->ex1 = PL_ICS_EX1(KERNEL_PL, 0);
+		p->thread.pc = (unsigned long) ret_from_kernel_thread;
+		return 0;
+	}
+
+	/*
+	 * Start new thread in ret_from_fork so it schedules properly
+	 * and then return from interrupt like the parent.
+	 */
+	p->thread.pc = (unsigned long) ret_from_fork;
+
+	/*
+	 * Do not clone step state from the parent; each thread
+	 * must make its own lazily.
+	 */
+	task_thread_info(p)->step_state = NULL;
+
+#ifdef __tilegx__
+	/*
+	 * Do not clone unalign jit fixup from the parent; each thread
+	 * must allocate its own on demand.
+	 */
+	task_thread_info(p)->unalign_jit_base = NULL;
+#endif
+
+	/*
+	 * Copy the registers onto the kernel stack so the
+	 * return-from-interrupt code will reload it into registers.
+	 */
+	*childregs = *current_pt_regs();
+	childregs->regs[0] = 0;         /* return value is zero */
+	if (sp)
+		childregs->sp = sp;  /* override with new user stack pointer */
+	memcpy(callee_regs, &childregs->regs[CALLEE_SAVED_FIRST_REG],
+	       CALLEE_SAVED_REGS_COUNT * sizeof(unsigned long));
+
+	/* Save user stack top pointer so we can ID the stack vm area later. */
+	p->thread.usp0 = childregs->sp;
+
+	/*
+	 * If CLONE_SETTLS is set, set "tp" in the new task to "r4",
+	 * which is passed in as arg #5 to sys_clone().
+	 */
+	if (clone_flags & CLONE_SETTLS)
+		childregs->tp = childregs->regs[4];
+
+
+#if CHIP_HAS_TILE_DMA()
+	/*
+	 * No DMA in the new thread.  We model this on the fact that
+	 * fork() clears the pending signals, alarms, and aio for the child.
+	 */
+	memset(&p->thread.tile_dma_state, 0, sizeof(struct tile_dma_state));
+	memset(&p->thread.dma_async_tlb, 0, sizeof(struct async_tlb));
+#endif
+
+	/* New thread has its miscellaneous processor state bits clear. */
+	p->thread.proc_status = 0;
+
+#ifdef CONFIG_HARDWALL
+	/* New thread does not own any networks. */
+	memset(&p->thread.hardwall[0], 0,
+	       sizeof(struct hardwall_task) * HARDWALL_TYPES);
+#endif
+
+
+	/*
+	 * Start the new thread with the current architecture state
+	 * (user interrupt masks, etc.).
+	 */
+	save_arch_state(&p->thread);
+
+	return 0;
+}
+
+int set_unalign_ctl(struct task_struct *tsk, unsigned int val)
+{
+	task_thread_info(tsk)->align_ctl = val;
+	return 0;
+}
+
+int get_unalign_ctl(struct task_struct *tsk, unsigned long adr)
+{
+	return put_user(task_thread_info(tsk)->align_ctl,
+			(unsigned int __user *)adr);
+}
+
+static struct task_struct corrupt_current = { .comm = "<corrupt>" };
+
+/*
+ * Return "current" if it looks plausible, or else a pointer to a dummy.
+ * This can be helpful if we are just trying to emit a clean panic.
+ */
+struct task_struct *validate_current(void)
+{
+	struct task_struct *tsk = current;
+	if (unlikely((unsigned long)tsk < PAGE_OFFSET ||
+		     (high_memory && (void *)tsk > high_memory) ||
+		     ((unsigned long)tsk & (__alignof__(*tsk) - 1)) != 0)) {
+		pr_err("Corrupt 'current' %p (sp %#lx)\n", tsk, stack_pointer);
+		tsk = &corrupt_current;
+	}
+	return tsk;
+}
+
+/* Take and return the pointer to the previous task, for schedule_tail(). */
+struct task_struct *sim_notify_fork(struct task_struct *prev)
+{
+	struct task_struct *tsk = current;
+	__insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_OS_FORK_PARENT |
+		     (tsk->thread.creator_pid << _SIM_CONTROL_OPERATOR_BITS));
+	__insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_OS_FORK |
+		     (tsk->pid << _SIM_CONTROL_OPERATOR_BITS));
+	return prev;
+}
+
+int dump_task_regs(struct task_struct *tsk, elf_gregset_t *regs)
+{
+	struct pt_regs *ptregs = task_pt_regs(tsk);
+	elf_core_copy_regs(regs, ptregs);
+	return 1;
+}
+
+#if CHIP_HAS_TILE_DMA()
+
+/* Allow user processes to access the DMA SPRs */
+void grant_dma_mpls(void)
+{
+#if CONFIG_KERNEL_PL == 2
+	__insn_mtspr(SPR_MPL_DMA_CPL_SET_1, 1);
+	__insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_1, 1);
+#else
+	__insn_mtspr(SPR_MPL_DMA_CPL_SET_0, 1);
+	__insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_0, 1);
+#endif
+}
+
+/* Forbid user processes from accessing the DMA SPRs */
+void restrict_dma_mpls(void)
+{
+#if CONFIG_KERNEL_PL == 2
+	__insn_mtspr(SPR_MPL_DMA_CPL_SET_2, 1);
+	__insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_2, 1);
+#else
+	__insn_mtspr(SPR_MPL_DMA_CPL_SET_1, 1);
+	__insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_1, 1);
+#endif
+}
+
+/* Pause the DMA engine, then save off its state registers. */
+static void save_tile_dma_state(struct tile_dma_state *dma)
+{
+	unsigned long state = __insn_mfspr(SPR_DMA_USER_STATUS);
+	unsigned long post_suspend_state;
+
+	/* If we're running, suspend the engine. */
+	if ((state & DMA_STATUS_MASK) == SPR_DMA_STATUS__RUNNING_MASK)
+		__insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__SUSPEND_MASK);
+
+	/*
+	 * Wait for the engine to idle, then save regs.  Note that we
+	 * want to record the "running" bit from before suspension,
+	 * and the "done" bit from after, so that we can properly
+	 * distinguish a case where the user suspended the engine from
+	 * the case where the kernel suspended as part of the context
+	 * swap.
+	 */
+	do {
+		post_suspend_state = __insn_mfspr(SPR_DMA_USER_STATUS);
+	} while (post_suspend_state & SPR_DMA_STATUS__BUSY_MASK);
+
+	dma->src = __insn_mfspr(SPR_DMA_SRC_ADDR);
+	dma->src_chunk = __insn_mfspr(SPR_DMA_SRC_CHUNK_ADDR);
+	dma->dest = __insn_mfspr(SPR_DMA_DST_ADDR);
+	dma->dest_chunk = __insn_mfspr(SPR_DMA_DST_CHUNK_ADDR);
+	dma->strides = __insn_mfspr(SPR_DMA_STRIDE);
+	dma->chunk_size = __insn_mfspr(SPR_DMA_CHUNK_SIZE);
+	dma->byte = __insn_mfspr(SPR_DMA_BYTE);
+	dma->status = (state & SPR_DMA_STATUS__RUNNING_MASK) |
+		(post_suspend_state & SPR_DMA_STATUS__DONE_MASK);
+}
+
+/* Restart a DMA that was running before we were context-switched out. */
+static void restore_tile_dma_state(struct thread_struct *t)
+{
+	const struct tile_dma_state *dma = &t->tile_dma_state;
+
+	/*
+	 * The only way to restore the done bit is to run a zero
+	 * length transaction.
+	 */
+	if ((dma->status & SPR_DMA_STATUS__DONE_MASK) &&
+	    !(__insn_mfspr(SPR_DMA_USER_STATUS) & SPR_DMA_STATUS__DONE_MASK)) {
+		__insn_mtspr(SPR_DMA_BYTE, 0);
+		__insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__REQUEST_MASK);
+		while (__insn_mfspr(SPR_DMA_USER_STATUS) &
+		       SPR_DMA_STATUS__BUSY_MASK)
+			;
+	}
+
+	__insn_mtspr(SPR_DMA_SRC_ADDR, dma->src);
+	__insn_mtspr(SPR_DMA_SRC_CHUNK_ADDR, dma->src_chunk);
+	__insn_mtspr(SPR_DMA_DST_ADDR, dma->dest);
+	__insn_mtspr(SPR_DMA_DST_CHUNK_ADDR, dma->dest_chunk);
+	__insn_mtspr(SPR_DMA_STRIDE, dma->strides);
+	__insn_mtspr(SPR_DMA_CHUNK_SIZE, dma->chunk_size);
+	__insn_mtspr(SPR_DMA_BYTE, dma->byte);
+
+	/*
+	 * Restart the engine if we were running and not done.
+	 * Clear a pending async DMA fault that we were waiting on return
+	 * to user space to execute, since we expect the DMA engine
+	 * to regenerate those faults for us now.  Note that we don't
+	 * try to clear the TIF_ASYNC_TLB flag, since it's relatively
+	 * harmless if set, and it covers both DMA and the SN processor.
+	 */
+	if ((dma->status & DMA_STATUS_MASK) == SPR_DMA_STATUS__RUNNING_MASK) {
+		t->dma_async_tlb.fault_num = 0;
+		__insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__REQUEST_MASK);
+	}
+}
+
+#endif
+
+static void save_arch_state(struct thread_struct *t)
+{
+#if CHIP_HAS_SPLIT_INTR_MASK()
+	t->interrupt_mask = __insn_mfspr(SPR_INTERRUPT_MASK_0_0) |
+		((u64)__insn_mfspr(SPR_INTERRUPT_MASK_0_1) << 32);
+#else
+	t->interrupt_mask = __insn_mfspr(SPR_INTERRUPT_MASK_0);
+#endif
+	t->ex_context[0] = __insn_mfspr(SPR_EX_CONTEXT_0_0);
+	t->ex_context[1] = __insn_mfspr(SPR_EX_CONTEXT_0_1);
+	t->system_save[0] = __insn_mfspr(SPR_SYSTEM_SAVE_0_0);
+	t->system_save[1] = __insn_mfspr(SPR_SYSTEM_SAVE_0_1);
+	t->system_save[2] = __insn_mfspr(SPR_SYSTEM_SAVE_0_2);
+	t->system_save[3] = __insn_mfspr(SPR_SYSTEM_SAVE_0_3);
+	t->intctrl_0 = __insn_mfspr(SPR_INTCTRL_0_STATUS);
+	t->proc_status = __insn_mfspr(SPR_PROC_STATUS);
+#if !CHIP_HAS_FIXED_INTVEC_BASE()
+	t->interrupt_vector_base = __insn_mfspr(SPR_INTERRUPT_VECTOR_BASE_0);
+#endif
+	t->tile_rtf_hwm = __insn_mfspr(SPR_TILE_RTF_HWM);
+#if CHIP_HAS_DSTREAM_PF()
+	t->dstream_pf = __insn_mfspr(SPR_DSTREAM_PF);
+#endif
+}
+
+static void restore_arch_state(const struct thread_struct *t)
+{
+#if CHIP_HAS_SPLIT_INTR_MASK()
+	__insn_mtspr(SPR_INTERRUPT_MASK_0_0, (u32) t->interrupt_mask);
+	__insn_mtspr(SPR_INTERRUPT_MASK_0_1, t->interrupt_mask >> 32);
+#else
+	__insn_mtspr(SPR_INTERRUPT_MASK_0, t->interrupt_mask);
+#endif
+	__insn_mtspr(SPR_EX_CONTEXT_0_0, t->ex_context[0]);
+	__insn_mtspr(SPR_EX_CONTEXT_0_1, t->ex_context[1]);
+	__insn_mtspr(SPR_SYSTEM_SAVE_0_0, t->system_save[0]);
+	__insn_mtspr(SPR_SYSTEM_SAVE_0_1, t->system_save[1]);
+	__insn_mtspr(SPR_SYSTEM_SAVE_0_2, t->system_save[2]);
+	__insn_mtspr(SPR_SYSTEM_SAVE_0_3, t->system_save[3]);
+	__insn_mtspr(SPR_INTCTRL_0_STATUS, t->intctrl_0);
+	__insn_mtspr(SPR_PROC_STATUS, t->proc_status);
+#if !CHIP_HAS_FIXED_INTVEC_BASE()
+	__insn_mtspr(SPR_INTERRUPT_VECTOR_BASE_0, t->interrupt_vector_base);
+#endif
+	__insn_mtspr(SPR_TILE_RTF_HWM, t->tile_rtf_hwm);
+#if CHIP_HAS_DSTREAM_PF()
+	__insn_mtspr(SPR_DSTREAM_PF, t->dstream_pf);
+#endif
+}
+
+
+void _prepare_arch_switch(struct task_struct *next)
+{
+#if CHIP_HAS_TILE_DMA()
+	struct tile_dma_state *dma = &current->thread.tile_dma_state;
+	if (dma->enabled)
+		save_tile_dma_state(dma);
+#endif
+}
+
+
+struct task_struct *__sched _switch_to(struct task_struct *prev,
+				       struct task_struct *next)
+{
+	/* DMA state is already saved; save off other arch state. */
+	save_arch_state(&prev->thread);
+
+#if CHIP_HAS_TILE_DMA()
+	/*
+	 * Restore DMA in new task if desired.
+	 * Note that it is only safe to restart here since interrupts
+	 * are disabled, so we can't take any DMATLB miss or access
+	 * interrupts before we have finished switching stacks.
+	 */
+	if (next->thread.tile_dma_state.enabled) {
+		restore_tile_dma_state(&next->thread);
+		grant_dma_mpls();
+	} else {
+		restrict_dma_mpls();
+	}
+#endif
+
+	/* Restore other arch state. */
+	restore_arch_state(&next->thread);
+
+#ifdef CONFIG_HARDWALL
+	/* Enable or disable access to the network registers appropriately. */
+	hardwall_switch_tasks(prev, next);
+#endif
+
+	/*
+	 * Switch kernel SP, PC, and callee-saved registers.
+	 * In the context of the new task, return the old task pointer
+	 * (i.e. the task that actually called __switch_to).
+	 * Pass the value to use for SYSTEM_SAVE_K_0 when we reset our sp.
+	 */
+	return __switch_to(prev, next, next_current_ksp0(next));
+}
+
+/*
+ * This routine is called on return from interrupt if any of the
+ * TIF_WORK_MASK flags are set in thread_info->flags.  It is
+ * entered with interrupts disabled so we don't miss an event
+ * that modified the thread_info flags.  If any flag is set, we
+ * handle it and return, and the calling assembly code will
+ * re-disable interrupts, reload the thread flags, and call back
+ * if more flags need to be handled.
+ *
+ * We return whether we need to check the thread_info flags again
+ * or not.  Note that we don't clear TIF_SINGLESTEP here, so it's
+ * important that it be tested last, and then claim that we don't
+ * need to recheck the flags.
+ */
+int do_work_pending(struct pt_regs *regs, u32 thread_info_flags)
+{
+	/* If we enter in kernel mode, do nothing and exit the caller loop. */
+	if (!user_mode(regs))
+		return 0;
+
+	user_exit();
+
+	/* Enable interrupts; they are disabled again on return to caller. */
+	local_irq_enable();
+
+	if (thread_info_flags & _TIF_NEED_RESCHED) {
+		schedule();
+		return 1;
+	}
+#if CHIP_HAS_TILE_DMA()
+	if (thread_info_flags & _TIF_ASYNC_TLB) {
+		do_async_page_fault(regs);
+		return 1;
+	}
+#endif
+	if (thread_info_flags & _TIF_SIGPENDING) {
+		do_signal(regs);
+		return 1;
+	}
+	if (thread_info_flags & _TIF_NOTIFY_RESUME) {
+		clear_thread_flag(TIF_NOTIFY_RESUME);
+		tracehook_notify_resume(regs);
+		return 1;
+	}
+	if (thread_info_flags & _TIF_SINGLESTEP)
+		single_step_once(regs);
+
+	user_enter();
+
+	return 0;
+}
+
+unsigned long get_wchan(struct task_struct *p)
+{
+	struct KBacktraceIterator kbt;
+
+	if (!p || p == current || p->state == TASK_RUNNING)
+		return 0;
+
+	for (KBacktraceIterator_init(&kbt, p, NULL);
+	     !KBacktraceIterator_end(&kbt);
+	     KBacktraceIterator_next(&kbt)) {
+		if (!in_sched_functions(kbt.it.pc))
+			return kbt.it.pc;
+	}
+
+	return 0;
+}
+
+/* Flush thread state. */
+void flush_thread(void)
+{
+	/* Nothing */
+}
+
+/*
+ * Free current thread data structures etc..
+ */
+void exit_thread(void)
+{
+#ifdef CONFIG_HARDWALL
+	/*
+	 * Remove the task from the list of tasks that are associated
+	 * with any live hardwalls.  (If the task that is exiting held
+	 * the last reference to a hardwall fd, it would already have
+	 * been released and deactivated at this point.)
+	 */
+	hardwall_deactivate_all(current);
+#endif
+}
+
+void show_regs(struct pt_regs *regs)
+{
+	struct task_struct *tsk = validate_current();
+	int i;
+
+	if (tsk != &corrupt_current)
+		show_regs_print_info(KERN_ERR);
+#ifdef __tilegx__
+	for (i = 0; i < 17; i++)
+		pr_err(" r%-2d: " REGFMT " r%-2d: " REGFMT " r%-2d: " REGFMT "\n",
+		       i, regs->regs[i], i+18, regs->regs[i+18],
+		       i+36, regs->regs[i+36]);
+	pr_err(" r17: " REGFMT " r35: " REGFMT " tp : " REGFMT "\n",
+	       regs->regs[17], regs->regs[35], regs->tp);
+	pr_err(" sp : " REGFMT " lr : " REGFMT "\n", regs->sp, regs->lr);
+#else
+	for (i = 0; i < 13; i++)
+		pr_err(" r%-2d: " REGFMT " r%-2d: " REGFMT " r%-2d: " REGFMT " r%-2d: " REGFMT "\n",
+		       i, regs->regs[i], i+14, regs->regs[i+14],
+		       i+27, regs->regs[i+27], i+40, regs->regs[i+40]);
+	pr_err(" r13: " REGFMT " tp : " REGFMT " sp : " REGFMT " lr : " REGFMT "\n",
+	       regs->regs[13], regs->tp, regs->sp, regs->lr);
+#endif
+	pr_err(" pc : " REGFMT " ex1: %ld     faultnum: %ld\n",
+	       regs->pc, regs->ex1, regs->faultnum);
+
+	dump_stack_regs(regs);
+}
diff --git a/arch/tile/kernel/ptrace.c b/arch/tile/kernel/ptrace.c
new file mode 100644
index 000000000..f84eed824
--- /dev/null
+++ b/arch/tile/kernel/ptrace.c
@@ -0,0 +1,325 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Copied from i386: Ross Biro 1/23/92
+ */
+
+#include <linux/kernel.h>
+#include <linux/ptrace.h>
+#include <linux/kprobes.h>
+#include <linux/compat.h>
+#include <linux/uaccess.h>
+#include <linux/regset.h>
+#include <linux/elf.h>
+#include <linux/tracehook.h>
+#include <linux/context_tracking.h>
+#include <asm/traps.h>
+#include <arch/chip.h>
+
+#define CREATE_TRACE_POINTS
+#include <trace/events/syscalls.h>
+
+void user_enable_single_step(struct task_struct *child)
+{
+	set_tsk_thread_flag(child, TIF_SINGLESTEP);
+}
+
+void user_disable_single_step(struct task_struct *child)
+{
+	clear_tsk_thread_flag(child, TIF_SINGLESTEP);
+}
+
+/*
+ * Called by kernel/ptrace.c when detaching..
+ */
+void ptrace_disable(struct task_struct *child)
+{
+	clear_tsk_thread_flag(child, TIF_SINGLESTEP);
+
+	/*
+	 * These two are currently unused, but will be set by arch_ptrace()
+	 * and used in the syscall assembly when we do support them.
+	 */
+	clear_tsk_thread_flag(child, TIF_SYSCALL_TRACE);
+}
+
+/*
+ * Get registers from task and ready the result for userspace.
+ * Note that we localize the API issues to getregs() and putregs() at
+ * some cost in performance, e.g. we need a full pt_regs copy for
+ * PEEKUSR, and two copies for POKEUSR.  But in general we expect
+ * GETREGS/PUTREGS to be the API of choice anyway.
+ */
+static char *getregs(struct task_struct *child, struct pt_regs *uregs)
+{
+	*uregs = *task_pt_regs(child);
+
+	/* Set up flags ABI bits. */
+	uregs->flags = 0;
+#ifdef CONFIG_COMPAT
+	if (task_thread_info(child)->status & TS_COMPAT)
+		uregs->flags |= PT_FLAGS_COMPAT;
+#endif
+
+	return (char *)uregs;
+}
+
+/* Put registers back to task. */
+static void putregs(struct task_struct *child, struct pt_regs *uregs)
+{
+	struct pt_regs *regs = task_pt_regs(child);
+
+	/* Don't allow overwriting the kernel-internal flags word. */
+	uregs->flags = regs->flags;
+
+	/* Only allow setting the ICS bit in the ex1 word. */
+	uregs->ex1 = PL_ICS_EX1(USER_PL, EX1_ICS(uregs->ex1));
+
+	*regs = *uregs;
+}
+
+enum tile_regset {
+	REGSET_GPR,
+};
+
+static int tile_gpr_get(struct task_struct *target,
+			  const struct user_regset *regset,
+			  unsigned int pos, unsigned int count,
+			  void *kbuf, void __user *ubuf)
+{
+	struct pt_regs regs;
+
+	getregs(target, &regs);
+
+	return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &regs, 0,
+				   sizeof(regs));
+}
+
+static int tile_gpr_set(struct task_struct *target,
+			  const struct user_regset *regset,
+			  unsigned int pos, unsigned int count,
+			  const void *kbuf, const void __user *ubuf)
+{
+	int ret;
+	struct pt_regs regs;
+
+	ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &regs, 0,
+				 sizeof(regs));
+	if (ret)
+		return ret;
+
+	putregs(target, &regs);
+
+	return 0;
+}
+
+static const struct user_regset tile_user_regset[] = {
+	[REGSET_GPR] = {
+		.core_note_type = NT_PRSTATUS,
+		.n = ELF_NGREG,
+		.size = sizeof(elf_greg_t),
+		.align = sizeof(elf_greg_t),
+		.get = tile_gpr_get,
+		.set = tile_gpr_set,
+	},
+};
+
+static const struct user_regset_view tile_user_regset_view = {
+	.name = CHIP_ARCH_NAME,
+	.e_machine = ELF_ARCH,
+	.ei_osabi = ELF_OSABI,
+	.regsets = tile_user_regset,
+	.n = ARRAY_SIZE(tile_user_regset),
+};
+
+const struct user_regset_view *task_user_regset_view(struct task_struct *task)
+{
+	return &tile_user_regset_view;
+}
+
+long arch_ptrace(struct task_struct *child, long request,
+		 unsigned long addr, unsigned long data)
+{
+	unsigned long __user *datap = (long __user __force *)data;
+	unsigned long tmp;
+	long ret = -EIO;
+	char *childreg;
+	struct pt_regs copyregs;
+
+	switch (request) {
+
+	case PTRACE_PEEKUSR:  /* Read register from pt_regs. */
+		if (addr >= PTREGS_SIZE)
+			break;
+		childreg = getregs(child, &copyregs) + addr;
+#ifdef CONFIG_COMPAT
+		if (is_compat_task()) {
+			if (addr & (sizeof(compat_long_t)-1))
+				break;
+			ret = put_user(*(compat_long_t *)childreg,
+				       (compat_long_t __user *)datap);
+		} else
+#endif
+		{
+			if (addr & (sizeof(long)-1))
+				break;
+			ret = put_user(*(long *)childreg, datap);
+		}
+		break;
+
+	case PTRACE_POKEUSR:  /* Write register in pt_regs. */
+		if (addr >= PTREGS_SIZE)
+			break;
+		childreg = getregs(child, &copyregs) + addr;
+#ifdef CONFIG_COMPAT
+		if (is_compat_task()) {
+			if (addr & (sizeof(compat_long_t)-1))
+				break;
+			*(compat_long_t *)childreg = data;
+		} else
+#endif
+		{
+			if (addr & (sizeof(long)-1))
+				break;
+			*(long *)childreg = data;
+		}
+		putregs(child, &copyregs);
+		ret = 0;
+		break;
+
+	case PTRACE_GETREGS:  /* Get all registers from the child. */
+		ret = copy_regset_to_user(child, &tile_user_regset_view,
+					  REGSET_GPR, 0,
+					  sizeof(struct pt_regs), datap);
+		break;
+
+	case PTRACE_SETREGS:  /* Set all registers in the child. */
+		ret = copy_regset_from_user(child, &tile_user_regset_view,
+					    REGSET_GPR, 0,
+					    sizeof(struct pt_regs), datap);
+		break;
+
+	case PTRACE_GETFPREGS:  /* Get the child FPU state. */
+	case PTRACE_SETFPREGS:  /* Set the child FPU state. */
+		break;
+
+	case PTRACE_SETOPTIONS:
+		/* Support TILE-specific ptrace options. */
+		BUILD_BUG_ON(PTRACE_O_MASK_TILE & PTRACE_O_MASK);
+		tmp = data & PTRACE_O_MASK_TILE;
+		data &= ~PTRACE_O_MASK_TILE;
+		ret = ptrace_request(child, request, addr, data);
+		if (ret == 0) {
+			unsigned int flags = child->ptrace;
+			flags &= ~(PTRACE_O_MASK_TILE << PT_OPT_FLAG_SHIFT);
+			flags |= (tmp << PT_OPT_FLAG_SHIFT);
+			child->ptrace = flags;
+		}
+		break;
+
+	default:
+#ifdef CONFIG_COMPAT
+		if (task_thread_info(current)->status & TS_COMPAT) {
+			ret = compat_ptrace_request(child, request,
+						    addr, data);
+			break;
+		}
+#endif
+		ret = ptrace_request(child, request, addr, data);
+		break;
+	}
+
+	return ret;
+}
+
+#ifdef CONFIG_COMPAT
+/* Not used; we handle compat issues in arch_ptrace() directly. */
+long compat_arch_ptrace(struct task_struct *child, compat_long_t request,
+			       compat_ulong_t addr, compat_ulong_t data)
+{
+	BUG();
+}
+#endif
+
+int do_syscall_trace_enter(struct pt_regs *regs)
+{
+	u32 work = ACCESS_ONCE(current_thread_info()->flags);
+
+	/*
+	 * If TIF_NOHZ is set, we are required to call user_exit() before
+	 * doing anything that could touch RCU.
+	 */
+	if (work & _TIF_NOHZ)
+		user_exit();
+
+	if (work & _TIF_SYSCALL_TRACE) {
+		if (tracehook_report_syscall_entry(regs))
+			regs->regs[TREG_SYSCALL_NR] = -1;
+	}
+
+	if (work & _TIF_SYSCALL_TRACEPOINT)
+		trace_sys_enter(regs, regs->regs[TREG_SYSCALL_NR]);
+
+	return regs->regs[TREG_SYSCALL_NR];
+}
+
+void do_syscall_trace_exit(struct pt_regs *regs)
+{
+	long errno;
+
+	/*
+	 * We may come here right after calling schedule_user()
+	 * in which case we can be in RCU user mode.
+	 */
+	user_exit();
+
+	/*
+	 * The standard tile calling convention returns the value (or negative
+	 * errno) in r0, and zero (or positive errno) in r1.
+	 * It saves a couple of cycles on the hot path to do this work in
+	 * registers only as we return, rather than updating the in-memory
+	 * struct ptregs.
+	 */
+	errno = (long) regs->regs[0];
+	if (errno < 0 && errno > -4096)
+		regs->regs[1] = -errno;
+	else
+		regs->regs[1] = 0;
+
+	if (test_thread_flag(TIF_SYSCALL_TRACE))
+		tracehook_report_syscall_exit(regs, 0);
+
+	if (test_thread_flag(TIF_SYSCALL_TRACEPOINT))
+		trace_sys_exit(regs, regs->regs[0]);
+}
+
+void send_sigtrap(struct task_struct *tsk, struct pt_regs *regs)
+{
+	struct siginfo info;
+
+	memset(&info, 0, sizeof(info));
+	info.si_signo = SIGTRAP;
+	info.si_code  = TRAP_BRKPT;
+	info.si_addr  = (void __user *) regs->pc;
+
+	/* Send us the fakey SIGTRAP */
+	force_sig_info(SIGTRAP, &info, tsk);
+}
+
+/* Handle synthetic interrupt delivered only by the simulator. */
+void __kprobes do_breakpoint(struct pt_regs* regs, int fault_num)
+{
+	enum ctx_state prev_state = exception_enter();
+	send_sigtrap(current, regs);
+	exception_exit(prev_state);
+}
diff --git a/arch/tile/kernel/reboot.c b/arch/tile/kernel/reboot.c
new file mode 100644
index 000000000..6c5d2c070
--- /dev/null
+++ b/arch/tile/kernel/reboot.c
@@ -0,0 +1,51 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/stddef.h>
+#include <linux/reboot.h>
+#include <linux/smp.h>
+#include <linux/pm.h>
+#include <linux/export.h>
+#include <asm/page.h>
+#include <asm/setup.h>
+#include <hv/hypervisor.h>
+
+#ifndef CONFIG_SMP
+#define smp_send_stop()
+#endif
+
+void machine_halt(void)
+{
+	arch_local_irq_disable_all();
+	smp_send_stop();
+	hv_halt();
+}
+
+void machine_power_off(void)
+{
+	arch_local_irq_disable_all();
+	smp_send_stop();
+	hv_power_off();
+}
+
+void machine_restart(char *cmd)
+{
+	arch_local_irq_disable_all();
+	smp_send_stop();
+	hv_restart((HV_VirtAddr) "vmlinux", (HV_VirtAddr) cmd);
+}
+
+/* No interesting distinction to be made here. */
+void (*pm_power_off)(void) = NULL;
+EXPORT_SYMBOL(pm_power_off);
diff --git a/arch/tile/kernel/regs_32.S b/arch/tile/kernel/regs_32.S
new file mode 100644
index 000000000..542cae17a
--- /dev/null
+++ b/arch/tile/kernel/regs_32.S
@@ -0,0 +1,145 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/linkage.h>
+#include <asm/ptrace.h>
+#include <asm/asm-offsets.h>
+#include <arch/spr_def.h>
+#include <asm/processor.h>
+#include <asm/switch_to.h>
+
+/*
+ * See <asm/switch_to.h>; called with prev and next task_struct pointers.
+ * "prev" is returned in r0 for _switch_to and also for ret_from_fork.
+ *
+ * We want to save pc/sp in "prev", and get the new pc/sp from "next".
+ * We also need to save all the callee-saved registers on the stack.
+ *
+ * Intel enables/disables access to the hardware cycle counter in
+ * seccomp (secure computing) environments if necessary, based on
+ * has_secure_computing().  We might want to do this at some point,
+ * though it would require virtualizing the other SPRs under WORLD_ACCESS.
+ *
+ * Since we're saving to the stack, we omit sp from this list.
+ * And for parallels with other architectures, we save lr separately,
+ * in the thread_struct itself (as the "pc" field).
+ *
+ * This code also needs to be aligned with process.c copy_thread()
+ */
+
+#if CALLEE_SAVED_REGS_COUNT != 24
+# error Mismatch between <asm/switch_to.h> and kernel/entry.S
+#endif
+#define FRAME_SIZE ((2 + CALLEE_SAVED_REGS_COUNT) * 4)
+
+#define SAVE_REG(r) { sw r12, r; addi r12, r12, 4 }
+#define LOAD_REG(r) { lw r, r12; addi r12, r12, 4 }
+#define FOR_EACH_CALLEE_SAVED_REG(f)					\
+							f(r30); f(r31); \
+	f(r32); f(r33); f(r34); f(r35);	f(r36); f(r37); f(r38); f(r39); \
+	f(r40); f(r41); f(r42); f(r43); f(r44); f(r45); f(r46); f(r47); \
+	f(r48); f(r49); f(r50); f(r51); f(r52);
+
+STD_ENTRY_SECTION(__switch_to, .sched.text)
+	{
+	  move r10, sp
+	  sw sp, lr
+	  addi sp, sp, -FRAME_SIZE
+	}
+	{
+	  addi r11, sp, 4
+	  addi r12, sp, 8
+	}
+	{
+	  sw r11, r10
+	  addli r4, r1, TASK_STRUCT_THREAD_KSP_OFFSET
+	}
+	{
+	  lw r13, r4   /* Load new sp to a temp register early. */
+	  addli r3, r0, TASK_STRUCT_THREAD_KSP_OFFSET
+	}
+	FOR_EACH_CALLEE_SAVED_REG(SAVE_REG)
+	{
+	  sw r3, sp
+	  addli r3, r0, TASK_STRUCT_THREAD_PC_OFFSET
+	}
+	{
+	  sw r3, lr
+	  addli r4, r1, TASK_STRUCT_THREAD_PC_OFFSET
+	}
+	{
+	  lw lr, r4
+	  addi r12, r13, 8
+	}
+	{
+	  /* Update sp and ksp0 simultaneously to avoid backtracer warnings. */
+	  move sp, r13
+	  mtspr SPR_SYSTEM_SAVE_K_0, r2
+	}
+	FOR_EACH_CALLEE_SAVED_REG(LOAD_REG)
+.L__switch_to_pc:
+	{
+	  addi sp, sp, FRAME_SIZE
+	  jrp lr   /* r0 is still valid here, so return it */
+	}
+	STD_ENDPROC(__switch_to)
+
+/* Return a suitable address for the backtracer for suspended threads */
+STD_ENTRY_SECTION(get_switch_to_pc, .sched.text)
+	lnk r0
+	{
+	  addli r0, r0, .L__switch_to_pc - .
+	  jrp lr
+	}
+	STD_ENDPROC(get_switch_to_pc)
+
+STD_ENTRY(get_pt_regs)
+	.irp reg, r0, r1, r2, r3, r4, r5, r6, r7, \
+		 r8, r9, r10, r11, r12, r13, r14, r15, \
+		 r16, r17, r18, r19, r20, r21, r22, r23, \
+		 r24, r25, r26, r27, r28, r29, r30, r31, \
+		 r32, r33, r34, r35, r36, r37, r38, r39, \
+		 r40, r41, r42, r43, r44, r45, r46, r47, \
+		 r48, r49, r50, r51, r52, tp, sp
+	{
+	 sw r0, \reg
+	 addi r0, r0, 4
+	}
+	.endr
+	{
+	 sw r0, lr
+	 addi r0, r0, PTREGS_OFFSET_PC - PTREGS_OFFSET_LR
+	}
+	lnk r1
+	{
+	 sw r0, r1
+	 addi r0, r0, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_PC
+	}
+	mfspr r1, INTERRUPT_CRITICAL_SECTION
+	shli r1, r1, SPR_EX_CONTEXT_1_1__ICS_SHIFT
+	ori r1, r1, KERNEL_PL
+	{
+	 sw r0, r1
+	 addi r0, r0, PTREGS_OFFSET_FAULTNUM - PTREGS_OFFSET_EX1
+	}
+	{
+	 sw r0, zero       /* clear faultnum */
+	 addi r0, r0, PTREGS_OFFSET_ORIG_R0 - PTREGS_OFFSET_FAULTNUM
+	}
+	{
+	 sw r0, zero       /* clear orig_r0 */
+	 addli r0, r0, -PTREGS_OFFSET_ORIG_R0    /* restore r0 to base */
+	}
+	jrp lr
+	STD_ENDPROC(get_pt_regs)
diff --git a/arch/tile/kernel/regs_64.S b/arch/tile/kernel/regs_64.S
new file mode 100644
index 000000000..bbffcc6f3
--- /dev/null
+++ b/arch/tile/kernel/regs_64.S
@@ -0,0 +1,145 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/linkage.h>
+#include <asm/ptrace.h>
+#include <asm/asm-offsets.h>
+#include <arch/spr_def.h>
+#include <asm/processor.h>
+#include <asm/switch_to.h>
+
+/*
+ * See <asm/switch_to.h>; called with prev and next task_struct pointers.
+ * "prev" is returned in r0 for _switch_to and also for ret_from_fork.
+ *
+ * We want to save pc/sp in "prev", and get the new pc/sp from "next".
+ * We also need to save all the callee-saved registers on the stack.
+ *
+ * Intel enables/disables access to the hardware cycle counter in
+ * seccomp (secure computing) environments if necessary, based on
+ * has_secure_computing().  We might want to do this at some point,
+ * though it would require virtualizing the other SPRs under WORLD_ACCESS.
+ *
+ * Since we're saving to the stack, we omit sp from this list.
+ * And for parallels with other architectures, we save lr separately,
+ * in the thread_struct itself (as the "pc" field).
+ *
+ * This code also needs to be aligned with process.c copy_thread()
+ */
+
+#if CALLEE_SAVED_REGS_COUNT != 24
+# error Mismatch between <asm/switch_to.h> and kernel/entry.S
+#endif
+#define FRAME_SIZE ((2 + CALLEE_SAVED_REGS_COUNT) * 8)
+
+#define SAVE_REG(r) { st r12, r; addi r12, r12, 8 }
+#define LOAD_REG(r) { ld r, r12; addi r12, r12, 8 }
+#define FOR_EACH_CALLEE_SAVED_REG(f)					\
+							f(r30); f(r31); \
+	f(r32); f(r33); f(r34); f(r35);	f(r36); f(r37); f(r38); f(r39); \
+	f(r40); f(r41); f(r42); f(r43); f(r44); f(r45); f(r46); f(r47); \
+	f(r48); f(r49); f(r50); f(r51); f(r52);
+
+STD_ENTRY_SECTION(__switch_to, .sched.text)
+	{
+	  move r10, sp
+	  st sp, lr
+	}
+	{
+	  addli r11, sp, -FRAME_SIZE + 8
+	  addli sp, sp, -FRAME_SIZE
+	}
+	{
+	  st r11, r10
+	  addli r4, r1, TASK_STRUCT_THREAD_KSP_OFFSET
+	}
+	{
+	  ld r13, r4   /* Load new sp to a temp register early. */
+	  addi r12, sp, 16
+	}
+	FOR_EACH_CALLEE_SAVED_REG(SAVE_REG)
+	addli r3, r0, TASK_STRUCT_THREAD_KSP_OFFSET
+	{
+	  st r3, sp
+	  addli r3, r0, TASK_STRUCT_THREAD_PC_OFFSET
+	}
+	{
+	  st r3, lr
+	  addli r4, r1, TASK_STRUCT_THREAD_PC_OFFSET
+	}
+	{
+	  ld lr, r4
+	  addi r12, r13, 16
+	}
+	{
+	  /* Update sp and ksp0 simultaneously to avoid backtracer warnings. */
+	  move sp, r13
+	  mtspr SPR_SYSTEM_SAVE_K_0, r2
+	}
+	FOR_EACH_CALLEE_SAVED_REG(LOAD_REG)
+.L__switch_to_pc:
+	{
+	  addli sp, sp, FRAME_SIZE
+	  jrp lr   /* r0 is still valid here, so return it */
+	}
+	STD_ENDPROC(__switch_to)
+
+/* Return a suitable address for the backtracer for suspended threads */
+STD_ENTRY_SECTION(get_switch_to_pc, .sched.text)
+	lnk r0
+	{
+	  addli r0, r0, .L__switch_to_pc - .
+	  jrp lr
+	}
+	STD_ENDPROC(get_switch_to_pc)
+
+STD_ENTRY(get_pt_regs)
+	.irp reg, r0, r1, r2, r3, r4, r5, r6, r7, \
+		 r8, r9, r10, r11, r12, r13, r14, r15, \
+		 r16, r17, r18, r19, r20, r21, r22, r23, \
+		 r24, r25, r26, r27, r28, r29, r30, r31, \
+		 r32, r33, r34, r35, r36, r37, r38, r39, \
+		 r40, r41, r42, r43, r44, r45, r46, r47, \
+		 r48, r49, r50, r51, r52, tp, sp
+	{
+	 st r0, \reg
+	 addi r0, r0, 8
+	}
+	.endr
+	{
+	 st r0, lr
+	 addi r0, r0, PTREGS_OFFSET_PC - PTREGS_OFFSET_LR
+	}
+	lnk r1
+	{
+	 st r0, r1
+	 addi r0, r0, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_PC
+	}
+	mfspr r1, INTERRUPT_CRITICAL_SECTION
+	shli r1, r1, SPR_EX_CONTEXT_1_1__ICS_SHIFT
+	ori r1, r1, KERNEL_PL
+	{
+	 st r0, r1
+	 addi r0, r0, PTREGS_OFFSET_FAULTNUM - PTREGS_OFFSET_EX1
+	}
+	{
+	 st r0, zero       /* clear faultnum */
+	 addi r0, r0, PTREGS_OFFSET_ORIG_R0 - PTREGS_OFFSET_FAULTNUM
+	}
+	{
+	 st r0, zero       /* clear orig_r0 */
+	 addli r0, r0, -PTREGS_OFFSET_ORIG_R0    /* restore r0 to base */
+	}
+	jrp lr
+	STD_ENDPROC(get_pt_regs)
diff --git a/arch/tile/kernel/relocate_kernel_32.S b/arch/tile/kernel/relocate_kernel_32.S
new file mode 100644
index 000000000..e44fbcf8c
--- /dev/null
+++ b/arch/tile/kernel/relocate_kernel_32.S
@@ -0,0 +1,269 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * copy new kernel into place and then call hv_reexec
+ *
+ */
+
+#include <linux/linkage.h>
+#include <arch/chip.h>
+#include <asm/page.h>
+#include <hv/hypervisor.h>
+
+#undef RELOCATE_NEW_KERNEL_VERBOSE
+
+STD_ENTRY(relocate_new_kernel)
+
+	move	r30, r0		/* page list */
+	move	r31, r1		/* address of page we are on */
+	move	r32, r2		/* start address of new kernel */
+
+	shri	r1, r1, PAGE_SHIFT
+	addi	r1, r1, 1
+	shli	sp, r1, PAGE_SHIFT
+	addi	sp, sp, -8
+	/* we now have a stack (whether we need one or not) */
+
+	moveli	r40, lo16(hv_console_putc)
+	auli	r40, r40, ha16(hv_console_putc)
+
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r0, 'r'
+	jalr	r40
+
+	moveli	r0, '_'
+	jalr	r40
+
+	moveli	r0, 'n'
+	jalr	r40
+
+	moveli	r0, '_'
+	jalr	r40
+
+	moveli	r0, 'k'
+	jalr	r40
+
+	moveli	r0, '\n'
+	jalr	r40
+#endif
+
+	/*
+	 * Throughout this code r30 is pointer to the element of page
+	 * list we are working on.
+	 *
+	 * Normally we get to the next element of the page list by
+	 * incrementing r30 by four.  The exception is if the element
+	 * on the page list is an IND_INDIRECTION in which case we use
+	 * the element with the low bits masked off as the new value
+	 * of r30.
+	 *
+	 * To get this started, we need the value passed to us (which
+	 * will always be an IND_INDIRECTION) in memory somewhere with
+	 * r30 pointing at it.  To do that, we push the value passed
+	 * to us on the stack and make r30 point to it.
+	 */
+
+	sw	sp, r30
+	move	r30, sp
+	addi	sp, sp, -8
+
+	/*
+	 * On TILEPro, we need to flush all tiles' caches, since we may
+	 * have been doing hash-for-home caching there.  Note that we
+	 * must do this _after_ we're completely done modifying any memory
+	 * other than our output buffer (which we know is locally cached).
+	 * We want the caches to be fully clean when we do the reexec,
+	 * because the hypervisor is going to do this flush again at that
+	 * point, and we don't want that second flush to overwrite any memory.
+	 */
+	{
+	 move	r0, zero	 /* cache_pa */
+	 move	r1, zero
+	}
+	{
+	 auli	r2, zero, ha16(HV_FLUSH_EVICT_L2) /* cache_control */
+	 movei	r3, -1		 /* cache_cpumask; -1 means all client tiles */
+	}
+	{
+	 move	r4, zero	 /* tlb_va */
+	 move	r5, zero	 /* tlb_length */
+	}
+	{
+	 move	r6, zero	 /* tlb_pgsize */
+	 move	r7, zero	 /* tlb_cpumask */
+	}
+	{
+	 move	r8, zero	 /* asids */
+	 moveli	r20, lo16(hv_flush_remote)
+	}
+	{
+	 move	r9, zero	 /* asidcount */
+	 auli	r20, r20, ha16(hv_flush_remote)
+	}
+
+	jalr	r20
+
+	/* r33 is destination pointer, default to zero */
+
+	moveli	r33, 0
+
+.Lloop:	lw	r10, r30
+
+	andi	r9, r10, 0xf	/* low 4 bits tell us what type it is */
+	xor	r10, r10, r9	/* r10 is now value with low 4 bits stripped */
+
+	seqi	r0, r9, 0x1	/* IND_DESTINATION */
+	bzt	r0, .Ltry2
+
+	move	r33, r10
+
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r0, 'd'
+	jalr	r40
+#endif
+
+	addi	r30, r30, 4
+	j	.Lloop
+
+.Ltry2:
+	seqi	r0, r9, 0x2	/* IND_INDIRECTION */
+	bzt	r0, .Ltry4
+
+	move	r30, r10
+
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r0, 'i'
+	jalr	r40
+#endif
+
+	j	.Lloop
+
+.Ltry4:
+	seqi	r0, r9, 0x4	/* IND_DONE */
+	bzt	r0, .Ltry8
+
+	mf
+
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r0, 'D'
+	jalr	r40
+	moveli	r0, '\n'
+	jalr	r40
+#endif
+
+	move	r0, r32
+	moveli	r1, 0		/* arg to hv_reexec is 64 bits */
+
+	moveli	r41, lo16(hv_reexec)
+	auli	r41, r41, ha16(hv_reexec)
+
+	jalr	r41
+
+	/* we should not get here */
+
+	moveli	r0, '?'
+	jalr	r40
+	moveli	r0, '\n'
+	jalr	r40
+
+	j	.Lhalt
+
+.Ltry8:	seqi	r0, r9, 0x8	/* IND_SOURCE */
+	bz	r0, .Lerr	/* unknown type */
+
+	/* copy page at r10 to page at r33 */
+
+	move	r11, r33
+
+	moveli	r0, lo16(PAGE_SIZE)
+	auli	r0, r0, ha16(PAGE_SIZE)
+	add	r33, r33, r0
+
+	/* copy word at r10 to word at r11 until r11 equals r33 */
+
+	/* We know page size must be multiple of 16, so we can unroll
+	 * 16 times safely without any edge case checking.
+	 *
+	 * Issue a flush of the destination every 16 words to avoid
+	 * incoherence when starting the new kernel.  (Now this is
+	 * just good paranoia because the hv_reexec call will also
+	 * take care of this.)
+	 */
+
+1:
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0; addi	r11, r11, 4 }
+	{ lw	r0, r10; addi	r10, r10, 4 }
+	{ sw	r11, r0 }
+	{ flush r11    ; addi	r11, r11, 4 }
+
+	seq	r0, r33, r11
+	bzt	r0, 1b
+
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r0, 's'
+	jalr	r40
+#endif
+
+	addi	r30, r30, 4
+	j	.Lloop
+
+
+.Lerr:	moveli	r0, 'e'
+	jalr	r40
+	moveli	r0, 'r'
+	jalr	r40
+	moveli	r0, 'r'
+	jalr	r40
+	moveli	r0, '\n'
+	jalr	r40
+.Lhalt:
+	moveli	r41, lo16(hv_halt)
+	auli	r41, r41, ha16(hv_halt)
+
+	jalr	r41
+	STD_ENDPROC(relocate_new_kernel)
+
+	.section .rodata,"a"
+
+	.globl relocate_new_kernel_size
+relocate_new_kernel_size:
+	.long .Lend_relocate_new_kernel - relocate_new_kernel
diff --git a/arch/tile/kernel/relocate_kernel_64.S b/arch/tile/kernel/relocate_kernel_64.S
new file mode 100644
index 000000000..d9d8cf617
--- /dev/null
+++ b/arch/tile/kernel/relocate_kernel_64.S
@@ -0,0 +1,263 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * copy new kernel into place and then call hv_reexec
+ *
+ */
+
+#include <linux/linkage.h>
+#include <arch/chip.h>
+#include <asm/page.h>
+#include <hv/hypervisor.h>
+
+#undef RELOCATE_NEW_KERNEL_VERBOSE
+
+STD_ENTRY(relocate_new_kernel)
+
+	move	r30, r0		/* page list */
+	move	r31, r1		/* address of page we are on */
+	move	r32, r2		/* start address of new kernel */
+
+	shrui	r1, r1, PAGE_SHIFT
+	addi	r1, r1, 1
+	shli	sp, r1, PAGE_SHIFT
+	addi	sp, sp, -8
+	/* we now have a stack (whether we need one or not) */
+
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r40, hw2_last(hv_console_putc)
+	shl16insli r40, r40, hw1(hv_console_putc)
+	shl16insli r40, r40, hw0(hv_console_putc)
+
+	moveli	r0, 'r'
+	jalr	r40
+
+	moveli	r0, '_'
+	jalr	r40
+
+	moveli	r0, 'n'
+	jalr	r40
+
+	moveli	r0, '_'
+	jalr	r40
+
+	moveli	r0, 'k'
+	jalr	r40
+
+	moveli	r0, '\n'
+	jalr	r40
+#endif
+
+	/*
+	 * Throughout this code r30 is pointer to the element of page
+	 * list we are working on.
+	 *
+	 * Normally we get to the next element of the page list by
+	 * incrementing r30 by eight.  The exception is if the element
+	 * on the page list is an IND_INDIRECTION in which case we use
+	 * the element with the low bits masked off as the new value
+	 * of r30.
+	 *
+	 * To get this started, we need the value passed to us (which
+	 * will always be an IND_INDIRECTION) in memory somewhere with
+	 * r30 pointing at it.  To do that, we push the value passed
+	 * to us on the stack and make r30 point to it.
+	 */
+
+	st	sp, r30
+	move	r30, sp
+	addi	sp, sp, -16
+
+	/*
+	 * On TILE-GX, we need to flush all tiles' caches, since we may
+	 * have been doing hash-for-home caching there.  Note that we
+	 * must do this _after_ we're completely done modifying any memory
+	 * other than our output buffer (which we know is locally cached).
+	 * We want the caches to be fully clean when we do the reexec,
+	 * because the hypervisor is going to do this flush again at that
+	 * point, and we don't want that second flush to overwrite any memory.
+	 */
+	{
+	 move	r0, zero	 /* cache_pa */
+	 moveli	r1, hw2_last(HV_FLUSH_EVICT_L2)
+	}
+	{
+	 shl16insli	r1, r1, hw1(HV_FLUSH_EVICT_L2)
+	 movei	r2, -1		 /* cache_cpumask; -1 means all client tiles */
+	}
+	{
+	 shl16insli	r1, r1, hw0(HV_FLUSH_EVICT_L2)  /* cache_control */
+	 move	r3, zero	 /* tlb_va */
+	}
+	{
+	 move	r4, zero	 /* tlb_length */
+	 move	r5, zero	 /* tlb_pgsize */
+	}
+	{
+	 move	r6, zero	 /* tlb_cpumask */
+	 move	r7, zero	 /* asids */
+	}
+	{
+	 moveli	r20, hw2_last(hv_flush_remote)
+	 move	r8, zero	 /* asidcount */
+	}
+	shl16insli	r20, r20, hw1(hv_flush_remote)
+	shl16insli	r20, r20, hw0(hv_flush_remote)
+
+	jalr	r20
+
+	/* r33 is destination pointer, default to zero */
+
+	moveli	r33, 0
+
+.Lloop:	ld	r10, r30
+
+	andi	r9, r10, 0xf	/* low 4 bits tell us what type it is */
+	xor	r10, r10, r9	/* r10 is now value with low 4 bits stripped */
+
+	cmpeqi	r0, r9, 0x1	/* IND_DESTINATION */
+	beqzt	r0, .Ltry2
+
+	move	r33, r10
+
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r0, 'd'
+	jalr	r40
+#endif
+
+	addi	r30, r30, 8
+	j	.Lloop
+
+.Ltry2:
+	cmpeqi	r0, r9, 0x2	/* IND_INDIRECTION */
+	beqzt	r0, .Ltry4
+
+	move	r30, r10
+
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r0, 'i'
+	jalr	r40
+#endif
+
+	j	.Lloop
+
+.Ltry4:
+	cmpeqi	r0, r9, 0x4	/* IND_DONE */
+	beqzt	r0, .Ltry8
+
+	mf
+
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r0, 'D'
+	jalr	r40
+	moveli	r0, '\n'
+	jalr	r40
+#endif
+
+	move	r0, r32
+
+	moveli	r41, hw2_last(hv_reexec)
+	shl16insli	r41, r41, hw1(hv_reexec)
+	shl16insli	r41, r41, hw0(hv_reexec)
+
+	jalr	r41
+
+	/* we should not get here */
+
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r0, '?'
+	jalr	r40
+	moveli	r0, '\n'
+	jalr	r40
+#endif
+
+	j	.Lhalt
+
+.Ltry8:	cmpeqi	r0, r9, 0x8	/* IND_SOURCE */
+	beqz	r0, .Lerr	/* unknown type */
+
+	/* copy page at r10 to page at r33 */
+
+	move	r11, r33
+
+	moveli	r0, hw2_last(PAGE_SIZE)
+	shl16insli	r0, r0, hw1(PAGE_SIZE)
+	shl16insli	r0, r0, hw0(PAGE_SIZE)
+	add	r33, r33, r0
+
+	/* copy word at r10 to word at r11 until r11 equals r33 */
+
+	/* We know page size must be multiple of 8, so we can unroll
+	 * 8 times safely without any edge case checking.
+	 *
+	 * Issue a flush of the destination every 8 words to avoid
+	 * incoherence when starting the new kernel.  (Now this is
+	 * just good paranoia because the hv_reexec call will also
+	 * take care of this.)
+	 */
+
+1:
+	{ ld	r0, r10; addi	r10, r10, 8 }
+	{ st	r11, r0; addi	r11, r11, 8 }
+	{ ld	r0, r10; addi	r10, r10, 8 }
+	{ st	r11, r0; addi	r11, r11, 8 }
+	{ ld	r0, r10; addi	r10, r10, 8 }
+	{ st	r11, r0; addi	r11, r11, 8 }
+	{ ld	r0, r10; addi	r10, r10, 8 }
+	{ st	r11, r0; addi	r11, r11, 8 }
+	{ ld	r0, r10; addi	r10, r10, 8 }
+	{ st	r11, r0; addi	r11, r11, 8 }
+	{ ld	r0, r10; addi	r10, r10, 8 }
+	{ st	r11, r0; addi	r11, r11, 8 }
+	{ ld	r0, r10; addi	r10, r10, 8 }
+	{ st	r11, r0; addi	r11, r11, 8 }
+	{ ld	r0, r10; addi	r10, r10, 8 }
+	{ st	r11, r0 }
+	{ flush r11    ; addi	r11, r11, 8 }
+
+	cmpeq	r0, r33, r11
+	beqzt	r0, 1b
+
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r0, 's'
+	jalr	r40
+#endif
+
+	addi	r30, r30, 8
+	j	.Lloop
+
+
+.Lerr:
+#ifdef RELOCATE_NEW_KERNEL_VERBOSE
+	moveli	r0, 'e'
+	jalr	r40
+	moveli	r0, 'r'
+	jalr	r40
+	moveli	r0, 'r'
+	jalr	r40
+	moveli	r0, '\n'
+	jalr	r40
+#endif
+.Lhalt:
+	moveli r41, hw2_last(hv_halt)
+	shl16insli r41, r41, hw1(hv_halt)
+	shl16insli r41, r41, hw0(hv_halt)
+
+	jalr	r41
+	STD_ENDPROC(relocate_new_kernel)
+
+	.section .rodata,"a"
+
+	.globl relocate_new_kernel_size
+relocate_new_kernel_size:
+	.long .Lend_relocate_new_kernel - relocate_new_kernel
diff --git a/arch/tile/kernel/setup.c b/arch/tile/kernel/setup.c
new file mode 100644
index 000000000..d366675e4
--- /dev/null
+++ b/arch/tile/kernel/setup.c
@@ -0,0 +1,1740 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/sched.h>
+#include <linux/kernel.h>
+#include <linux/mmzone.h>
+#include <linux/bootmem.h>
+#include <linux/module.h>
+#include <linux/node.h>
+#include <linux/cpu.h>
+#include <linux/ioport.h>
+#include <linux/irq.h>
+#include <linux/kexec.h>
+#include <linux/pci.h>
+#include <linux/swiotlb.h>
+#include <linux/initrd.h>
+#include <linux/io.h>
+#include <linux/highmem.h>
+#include <linux/smp.h>
+#include <linux/timex.h>
+#include <linux/hugetlb.h>
+#include <linux/start_kernel.h>
+#include <linux/screen_info.h>
+#include <linux/tick.h>
+#include <asm/setup.h>
+#include <asm/sections.h>
+#include <asm/cacheflush.h>
+#include <asm/pgalloc.h>
+#include <asm/mmu_context.h>
+#include <hv/hypervisor.h>
+#include <arch/interrupts.h>
+
+/* <linux/smp.h> doesn't provide this definition. */
+#ifndef CONFIG_SMP
+#define setup_max_cpus 1
+#endif
+
+static inline int ABS(int x) { return x >= 0 ? x : -x; }
+
+/* Chip information */
+char chip_model[64] __write_once;
+
+#ifdef CONFIG_VT
+struct screen_info screen_info;
+#endif
+
+struct pglist_data node_data[MAX_NUMNODES] __read_mostly;
+EXPORT_SYMBOL(node_data);
+
+/* Information on the NUMA nodes that we compute early */
+unsigned long node_start_pfn[MAX_NUMNODES];
+unsigned long node_end_pfn[MAX_NUMNODES];
+unsigned long __initdata node_memmap_pfn[MAX_NUMNODES];
+unsigned long __initdata node_percpu_pfn[MAX_NUMNODES];
+unsigned long __initdata node_free_pfn[MAX_NUMNODES];
+
+static unsigned long __initdata node_percpu[MAX_NUMNODES];
+
+/*
+ * per-CPU stack and boot info.
+ */
+DEFINE_PER_CPU(unsigned long, boot_sp) =
+	(unsigned long)init_stack + THREAD_SIZE;
+
+#ifdef CONFIG_SMP
+DEFINE_PER_CPU(unsigned long, boot_pc) = (unsigned long)start_kernel;
+#else
+/*
+ * The variable must be __initdata since it references __init code.
+ * With CONFIG_SMP it is per-cpu data, which is exempt from validation.
+ */
+unsigned long __initdata boot_pc = (unsigned long)start_kernel;
+#endif
+
+#ifdef CONFIG_HIGHMEM
+/* Page frame index of end of lowmem on each controller. */
+unsigned long node_lowmem_end_pfn[MAX_NUMNODES];
+
+/* Number of pages that can be mapped into lowmem. */
+static unsigned long __initdata mappable_physpages;
+#endif
+
+/* Data on which physical memory controller corresponds to which NUMA node */
+int node_controller[MAX_NUMNODES] = { [0 ... MAX_NUMNODES-1] = -1 };
+
+#ifdef CONFIG_HIGHMEM
+/* Map information from VAs to PAs */
+unsigned long pbase_map[1 << (32 - HPAGE_SHIFT)]
+  __write_once __attribute__((aligned(L2_CACHE_BYTES)));
+EXPORT_SYMBOL(pbase_map);
+
+/* Map information from PAs to VAs */
+void *vbase_map[NR_PA_HIGHBIT_VALUES]
+  __write_once __attribute__((aligned(L2_CACHE_BYTES)));
+EXPORT_SYMBOL(vbase_map);
+#endif
+
+/* Node number as a function of the high PA bits */
+int highbits_to_node[NR_PA_HIGHBIT_VALUES] __write_once;
+EXPORT_SYMBOL(highbits_to_node);
+
+static unsigned int __initdata maxmem_pfn = -1U;
+static unsigned int __initdata maxnodemem_pfn[MAX_NUMNODES] = {
+	[0 ... MAX_NUMNODES-1] = -1U
+};
+static nodemask_t __initdata isolnodes;
+
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
+enum { DEFAULT_PCI_RESERVE_MB = 64 };
+static unsigned int __initdata pci_reserve_mb = DEFAULT_PCI_RESERVE_MB;
+unsigned long __initdata pci_reserve_start_pfn = -1U;
+unsigned long __initdata pci_reserve_end_pfn = -1U;
+#endif
+
+static int __init setup_maxmem(char *str)
+{
+	unsigned long long maxmem;
+	if (str == NULL || (maxmem = memparse(str, NULL)) == 0)
+		return -EINVAL;
+
+	maxmem_pfn = (maxmem >> HPAGE_SHIFT) << (HPAGE_SHIFT - PAGE_SHIFT);
+	pr_info("Forcing RAM used to no more than %dMB\n",
+		maxmem_pfn >> (20 - PAGE_SHIFT));
+	return 0;
+}
+early_param("maxmem", setup_maxmem);
+
+static int __init setup_maxnodemem(char *str)
+{
+	char *endp;
+	unsigned long long maxnodemem;
+	long node;
+
+	node = str ? simple_strtoul(str, &endp, 0) : INT_MAX;
+	if (node >= MAX_NUMNODES || *endp != ':')
+		return -EINVAL;
+
+	maxnodemem = memparse(endp+1, NULL);
+	maxnodemem_pfn[node] = (maxnodemem >> HPAGE_SHIFT) <<
+		(HPAGE_SHIFT - PAGE_SHIFT);
+	pr_info("Forcing RAM used on node %ld to no more than %dMB\n",
+		node, maxnodemem_pfn[node] >> (20 - PAGE_SHIFT));
+	return 0;
+}
+early_param("maxnodemem", setup_maxnodemem);
+
+struct memmap_entry {
+	u64 addr;	/* start of memory segment */
+	u64 size;	/* size of memory segment */
+};
+static struct memmap_entry memmap_map[64];
+static int memmap_nr;
+
+static void add_memmap_region(u64 addr, u64 size)
+{
+	if (memmap_nr >= ARRAY_SIZE(memmap_map)) {
+		pr_err("Ooops! Too many entries in the memory map!\n");
+		return;
+	}
+	memmap_map[memmap_nr].addr = addr;
+	memmap_map[memmap_nr].size = size;
+	memmap_nr++;
+}
+
+static int __init setup_memmap(char *p)
+{
+	char *oldp;
+	u64 start_at, mem_size;
+
+	if (!p)
+		return -EINVAL;
+
+	if (!strncmp(p, "exactmap", 8)) {
+		pr_err("\"memmap=exactmap\" not valid on tile\n");
+		return 0;
+	}
+
+	oldp = p;
+	mem_size = memparse(p, &p);
+	if (p == oldp)
+		return -EINVAL;
+
+	if (*p == '@') {
+		pr_err("\"memmap=nn@ss\" (force RAM) invalid on tile\n");
+	} else if (*p == '#') {
+		pr_err("\"memmap=nn#ss\" (force ACPI data) invalid on tile\n");
+	} else if (*p == '$') {
+		start_at = memparse(p+1, &p);
+		add_memmap_region(start_at, mem_size);
+	} else {
+		if (mem_size == 0)
+			return -EINVAL;
+		maxmem_pfn = (mem_size >> HPAGE_SHIFT) <<
+			(HPAGE_SHIFT - PAGE_SHIFT);
+	}
+	return *p == '\0' ? 0 : -EINVAL;
+}
+early_param("memmap", setup_memmap);
+
+static int __init setup_mem(char *str)
+{
+	return setup_maxmem(str);
+}
+early_param("mem", setup_mem);  /* compatibility with x86 */
+
+static int __init setup_isolnodes(char *str)
+{
+	if (str == NULL || nodelist_parse(str, isolnodes) != 0)
+		return -EINVAL;
+
+	pr_info("Set isolnodes value to '%*pbl'\n",
+		nodemask_pr_args(&isolnodes));
+	return 0;
+}
+early_param("isolnodes", setup_isolnodes);
+
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
+static int __init setup_pci_reserve(char* str)
+{
+	if (str == NULL || kstrtouint(str, 0, &pci_reserve_mb) != 0 ||
+	    pci_reserve_mb > 3 * 1024)
+		return -EINVAL;
+
+	pr_info("Reserving %dMB for PCIE root complex mappings\n",
+		pci_reserve_mb);
+	return 0;
+}
+early_param("pci_reserve", setup_pci_reserve);
+#endif
+
+#ifndef __tilegx__
+/*
+ * vmalloc=size forces the vmalloc area to be exactly 'size' bytes.
+ * This can be used to increase (or decrease) the vmalloc area.
+ */
+static int __init parse_vmalloc(char *arg)
+{
+	if (!arg)
+		return -EINVAL;
+
+	VMALLOC_RESERVE = (memparse(arg, &arg) + PGDIR_SIZE - 1) & PGDIR_MASK;
+
+	/* See validate_va() for more on this test. */
+	if ((long)_VMALLOC_START >= 0)
+		early_panic("\"vmalloc=%#lx\" value too large: maximum %#lx\n",
+			    VMALLOC_RESERVE, _VMALLOC_END - 0x80000000UL);
+
+	return 0;
+}
+early_param("vmalloc", parse_vmalloc);
+#endif
+
+#ifdef CONFIG_HIGHMEM
+/*
+ * Determine for each controller where its lowmem is mapped and how much of
+ * it is mapped there.  On controller zero, the first few megabytes are
+ * already mapped in as code at MEM_SV_START, so in principle we could
+ * start our data mappings higher up, but for now we don't bother, to avoid
+ * additional confusion.
+ *
+ * One question is whether, on systems with more than 768 Mb and
+ * controllers of different sizes, to map in a proportionate amount of
+ * each one, or to try to map the same amount from each controller.
+ * (E.g. if we have three controllers with 256MB, 1GB, and 256MB
+ * respectively, do we map 256MB from each, or do we map 128 MB, 512
+ * MB, and 128 MB respectively?)  For now we use a proportionate
+ * solution like the latter.
+ *
+ * The VA/PA mapping demands that we align our decisions at 16 MB
+ * boundaries so that we can rapidly convert VA to PA.
+ */
+static void *__init setup_pa_va_mapping(void)
+{
+	unsigned long curr_pages = 0;
+	unsigned long vaddr = PAGE_OFFSET;
+	nodemask_t highonlynodes = isolnodes;
+	int i, j;
+
+	memset(pbase_map, -1, sizeof(pbase_map));
+	memset(vbase_map, -1, sizeof(vbase_map));
+
+	/* Node zero cannot be isolated for LOWMEM purposes. */
+	node_clear(0, highonlynodes);
+
+	/* Count up the number of pages on non-highonlynodes controllers. */
+	mappable_physpages = 0;
+	for_each_online_node(i) {
+		if (!node_isset(i, highonlynodes))
+			mappable_physpages +=
+				node_end_pfn[i] - node_start_pfn[i];
+	}
+
+	for_each_online_node(i) {
+		unsigned long start = node_start_pfn[i];
+		unsigned long end = node_end_pfn[i];
+		unsigned long size = end - start;
+		unsigned long vaddr_end;
+
+		if (node_isset(i, highonlynodes)) {
+			/* Mark this controller as having no lowmem. */
+			node_lowmem_end_pfn[i] = start;
+			continue;
+		}
+
+		curr_pages += size;
+		if (mappable_physpages > MAXMEM_PFN) {
+			vaddr_end = PAGE_OFFSET +
+				(((u64)curr_pages * MAXMEM_PFN /
+				  mappable_physpages)
+				 << PAGE_SHIFT);
+		} else {
+			vaddr_end = PAGE_OFFSET + (curr_pages << PAGE_SHIFT);
+		}
+		for (j = 0; vaddr < vaddr_end; vaddr += HPAGE_SIZE, ++j) {
+			unsigned long this_pfn =
+				start + (j << HUGETLB_PAGE_ORDER);
+			pbase_map[vaddr >> HPAGE_SHIFT] = this_pfn;
+			if (vbase_map[__pfn_to_highbits(this_pfn)] ==
+			    (void *)-1)
+				vbase_map[__pfn_to_highbits(this_pfn)] =
+					(void *)(vaddr & HPAGE_MASK);
+		}
+		node_lowmem_end_pfn[i] = start + (j << HUGETLB_PAGE_ORDER);
+		BUG_ON(node_lowmem_end_pfn[i] > end);
+	}
+
+	/* Return highest address of any mapped memory. */
+	return (void *)vaddr;
+}
+#endif /* CONFIG_HIGHMEM */
+
+/*
+ * Register our most important memory mappings with the debug stub.
+ *
+ * This is up to 4 mappings for lowmem, one mapping per memory
+ * controller, plus one for our text segment.
+ */
+static void store_permanent_mappings(void)
+{
+	int i;
+
+	for_each_online_node(i) {
+		HV_PhysAddr pa = ((HV_PhysAddr)node_start_pfn[i]) << PAGE_SHIFT;
+#ifdef CONFIG_HIGHMEM
+		HV_PhysAddr high_mapped_pa = node_lowmem_end_pfn[i];
+#else
+		HV_PhysAddr high_mapped_pa = node_end_pfn[i];
+#endif
+
+		unsigned long pages = high_mapped_pa - node_start_pfn[i];
+		HV_VirtAddr addr = (HV_VirtAddr) __va(pa);
+		hv_store_mapping(addr, pages << PAGE_SHIFT, pa);
+	}
+
+	hv_store_mapping((HV_VirtAddr)_text,
+			 (uint32_t)(_einittext - _text), 0);
+}
+
+/*
+ * Use hv_inquire_physical() to populate node_{start,end}_pfn[]
+ * and node_online_map, doing suitable sanity-checking.
+ * Also set min_low_pfn, max_low_pfn, and max_pfn.
+ */
+static void __init setup_memory(void)
+{
+	int i, j;
+	int highbits_seen[NR_PA_HIGHBIT_VALUES] = { 0 };
+#ifdef CONFIG_HIGHMEM
+	long highmem_pages;
+#endif
+#ifndef __tilegx__
+	int cap;
+#endif
+#if defined(CONFIG_HIGHMEM) || defined(__tilegx__)
+	long lowmem_pages;
+#endif
+	unsigned long physpages = 0;
+
+	/* We are using a char to hold the cpu_2_node[] mapping */
+	BUILD_BUG_ON(MAX_NUMNODES > 127);
+
+	/* Discover the ranges of memory available to us */
+	for (i = 0; ; ++i) {
+		unsigned long start, size, end, highbits;
+		HV_PhysAddrRange range = hv_inquire_physical(i);
+		if (range.size == 0)
+			break;
+#ifdef CONFIG_FLATMEM
+		if (i > 0) {
+			pr_err("Can't use discontiguous PAs: %#llx..%#llx\n",
+			       range.size, range.start + range.size);
+			continue;
+		}
+#endif
+#ifndef __tilegx__
+		if ((unsigned long)range.start) {
+			pr_err("Range not at 4GB multiple: %#llx..%#llx\n",
+			       range.start, range.start + range.size);
+			continue;
+		}
+#endif
+		if ((range.start & (HPAGE_SIZE-1)) != 0 ||
+		    (range.size & (HPAGE_SIZE-1)) != 0) {
+			unsigned long long start_pa = range.start;
+			unsigned long long orig_size = range.size;
+			range.start = (start_pa + HPAGE_SIZE - 1) & HPAGE_MASK;
+			range.size -= (range.start - start_pa);
+			range.size &= HPAGE_MASK;
+			pr_err("Range not hugepage-aligned: %#llx..%#llx: now %#llx-%#llx\n",
+			       start_pa, start_pa + orig_size,
+			       range.start, range.start + range.size);
+		}
+		highbits = __pa_to_highbits(range.start);
+		if (highbits >= NR_PA_HIGHBIT_VALUES) {
+			pr_err("PA high bits too high: %#llx..%#llx\n",
+			       range.start, range.start + range.size);
+			continue;
+		}
+		if (highbits_seen[highbits]) {
+			pr_err("Range overlaps in high bits: %#llx..%#llx\n",
+			       range.start, range.start + range.size);
+			continue;
+		}
+		highbits_seen[highbits] = 1;
+		if (PFN_DOWN(range.size) > maxnodemem_pfn[i]) {
+			int max_size = maxnodemem_pfn[i];
+			if (max_size > 0) {
+				pr_err("Maxnodemem reduced node %d to %d pages\n",
+				       i, max_size);
+				range.size = PFN_PHYS(max_size);
+			} else {
+				pr_err("Maxnodemem disabled node %d\n", i);
+				continue;
+			}
+		}
+		if (physpages + PFN_DOWN(range.size) > maxmem_pfn) {
+			int max_size = maxmem_pfn - physpages;
+			if (max_size > 0) {
+				pr_err("Maxmem reduced node %d to %d pages\n",
+				       i, max_size);
+				range.size = PFN_PHYS(max_size);
+			} else {
+				pr_err("Maxmem disabled node %d\n", i);
+				continue;
+			}
+		}
+		if (i >= MAX_NUMNODES) {
+			pr_err("Too many PA nodes (#%d): %#llx...%#llx\n",
+			       i, range.size, range.size + range.start);
+			continue;
+		}
+
+		start = range.start >> PAGE_SHIFT;
+		size = range.size >> PAGE_SHIFT;
+		end = start + size;
+
+#ifndef __tilegx__
+		if (((HV_PhysAddr)end << PAGE_SHIFT) !=
+		    (range.start + range.size)) {
+			pr_err("PAs too high to represent: %#llx..%#llx\n",
+			       range.start, range.start + range.size);
+			continue;
+		}
+#endif
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
+		/*
+		 * Blocks that overlap the pci reserved region must
+		 * have enough space to hold the maximum percpu data
+		 * region at the top of the range.  If there isn't
+		 * enough space above the reserved region, just
+		 * truncate the node.
+		 */
+		if (start <= pci_reserve_start_pfn &&
+		    end > pci_reserve_start_pfn) {
+			unsigned int per_cpu_size =
+				__per_cpu_end - __per_cpu_start;
+			unsigned int percpu_pages =
+				NR_CPUS * (PFN_UP(per_cpu_size) >> PAGE_SHIFT);
+			if (end < pci_reserve_end_pfn + percpu_pages) {
+				end = pci_reserve_start_pfn;
+				pr_err("PCI mapping region reduced node %d to %ld pages\n",
+				       i, end - start);
+			}
+		}
+#endif
+
+		for (j = __pfn_to_highbits(start);
+		     j <= __pfn_to_highbits(end - 1); j++)
+			highbits_to_node[j] = i;
+
+		node_start_pfn[i] = start;
+		node_end_pfn[i] = end;
+		node_controller[i] = range.controller;
+		physpages += size;
+		max_pfn = end;
+
+		/* Mark node as online */
+		node_set(i, node_online_map);
+		node_set(i, node_possible_map);
+	}
+
+#ifndef __tilegx__
+	/*
+	 * For 4KB pages, mem_map "struct page" data is 1% of the size
+	 * of the physical memory, so can be quite big (640 MB for
+	 * four 16G zones).  These structures must be mapped in
+	 * lowmem, and since we currently cap out at about 768 MB,
+	 * it's impractical to try to use this much address space.
+	 * For now, arbitrarily cap the amount of physical memory
+	 * we're willing to use at 8 million pages (32GB of 4KB pages).
+	 */
+	cap = 8 * 1024 * 1024;  /* 8 million pages */
+	if (physpages > cap) {
+		int num_nodes = num_online_nodes();
+		int cap_each = cap / num_nodes;
+		unsigned long dropped_pages = 0;
+		for (i = 0; i < num_nodes; ++i) {
+			int size = node_end_pfn[i] - node_start_pfn[i];
+			if (size > cap_each) {
+				dropped_pages += (size - cap_each);
+				node_end_pfn[i] = node_start_pfn[i] + cap_each;
+			}
+		}
+		physpages -= dropped_pages;
+		pr_warn("Only using %ldMB memory - ignoring %ldMB\n",
+			physpages >> (20 - PAGE_SHIFT),
+			dropped_pages >> (20 - PAGE_SHIFT));
+		pr_warn("Consider using a larger page size\n");
+	}
+#endif
+
+	/* Heap starts just above the last loaded address. */
+	min_low_pfn = PFN_UP((unsigned long)_end - PAGE_OFFSET);
+
+#ifdef CONFIG_HIGHMEM
+	/* Find where we map lowmem from each controller. */
+	high_memory = setup_pa_va_mapping();
+
+	/* Set max_low_pfn based on what node 0 can directly address. */
+	max_low_pfn = node_lowmem_end_pfn[0];
+
+	lowmem_pages = (mappable_physpages > MAXMEM_PFN) ?
+		MAXMEM_PFN : mappable_physpages;
+	highmem_pages = (long) (physpages - lowmem_pages);
+
+	pr_notice("%ldMB HIGHMEM available\n",
+		  pages_to_mb(highmem_pages > 0 ? highmem_pages : 0));
+	pr_notice("%ldMB LOWMEM available\n", pages_to_mb(lowmem_pages));
+#else
+	/* Set max_low_pfn based on what node 0 can directly address. */
+	max_low_pfn = node_end_pfn[0];
+
+#ifndef __tilegx__
+	if (node_end_pfn[0] > MAXMEM_PFN) {
+		pr_warn("Only using %ldMB LOWMEM\n", MAXMEM >> 20);
+		pr_warn("Use a HIGHMEM enabled kernel\n");
+		max_low_pfn = MAXMEM_PFN;
+		max_pfn = MAXMEM_PFN;
+		node_end_pfn[0] = MAXMEM_PFN;
+	} else {
+		pr_notice("%ldMB memory available\n",
+			  pages_to_mb(node_end_pfn[0]));
+	}
+	for (i = 1; i < MAX_NUMNODES; ++i) {
+		node_start_pfn[i] = 0;
+		node_end_pfn[i] = 0;
+	}
+	high_memory = __va(node_end_pfn[0]);
+#else
+	lowmem_pages = 0;
+	for (i = 0; i < MAX_NUMNODES; ++i) {
+		int pages = node_end_pfn[i] - node_start_pfn[i];
+		lowmem_pages += pages;
+		if (pages)
+			high_memory = pfn_to_kaddr(node_end_pfn[i]);
+	}
+	pr_notice("%ldMB memory available\n", pages_to_mb(lowmem_pages));
+#endif
+#endif
+}
+
+/*
+ * On 32-bit machines, we only put bootmem on the low controller,
+ * since PAs > 4GB can't be used in bootmem.  In principle one could
+ * imagine, e.g., multiple 1 GB controllers all of which could support
+ * bootmem, but in practice using controllers this small isn't a
+ * particularly interesting scenario, so we just keep it simple and
+ * use only the first controller for bootmem on 32-bit machines.
+ */
+static inline int node_has_bootmem(int nid)
+{
+#ifdef CONFIG_64BIT
+	return 1;
+#else
+	return nid == 0;
+#endif
+}
+
+static inline unsigned long alloc_bootmem_pfn(int nid,
+					      unsigned long size,
+					      unsigned long goal)
+{
+	void *kva = __alloc_bootmem_node(NODE_DATA(nid), size,
+					 PAGE_SIZE, goal);
+	unsigned long pfn = kaddr_to_pfn(kva);
+	BUG_ON(goal && PFN_PHYS(pfn) != goal);
+	return pfn;
+}
+
+static void __init setup_bootmem_allocator_node(int i)
+{
+	unsigned long start, end, mapsize, mapstart;
+
+	if (node_has_bootmem(i)) {
+		NODE_DATA(i)->bdata = &bootmem_node_data[i];
+	} else {
+		/* Share controller zero's bdata for now. */
+		NODE_DATA(i)->bdata = &bootmem_node_data[0];
+		return;
+	}
+
+	/* Skip up to after the bss in node 0. */
+	start = (i == 0) ? min_low_pfn : node_start_pfn[i];
+
+	/* Only lowmem, if we're a HIGHMEM build. */
+#ifdef CONFIG_HIGHMEM
+	end = node_lowmem_end_pfn[i];
+#else
+	end = node_end_pfn[i];
+#endif
+
+	/* No memory here. */
+	if (end == start)
+		return;
+
+	/* Figure out where the bootmem bitmap is located. */
+	mapsize = bootmem_bootmap_pages(end - start);
+	if (i == 0) {
+		/* Use some space right before the heap on node 0. */
+		mapstart = start;
+		start += mapsize;
+	} else {
+		/* Allocate bitmap on node 0 to avoid page table issues. */
+		mapstart = alloc_bootmem_pfn(0, PFN_PHYS(mapsize), 0);
+	}
+
+	/* Initialize a node. */
+	init_bootmem_node(NODE_DATA(i), mapstart, start, end);
+
+	/* Free all the space back into the allocator. */
+	free_bootmem(PFN_PHYS(start), PFN_PHYS(end - start));
+
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
+	/*
+	 * Throw away any memory aliased by the PCI region.
+	 */
+	if (pci_reserve_start_pfn < end && pci_reserve_end_pfn > start) {
+		start = max(pci_reserve_start_pfn, start);
+		end = min(pci_reserve_end_pfn, end);
+		reserve_bootmem(PFN_PHYS(start), PFN_PHYS(end - start),
+				BOOTMEM_EXCLUSIVE);
+	}
+#endif
+}
+
+static void __init setup_bootmem_allocator(void)
+{
+	int i;
+	for (i = 0; i < MAX_NUMNODES; ++i)
+		setup_bootmem_allocator_node(i);
+
+	/* Reserve any memory excluded by "memmap" arguments. */
+	for (i = 0; i < memmap_nr; ++i) {
+		struct memmap_entry *m = &memmap_map[i];
+		reserve_bootmem(m->addr, m->size, BOOTMEM_DEFAULT);
+	}
+
+#ifdef CONFIG_BLK_DEV_INITRD
+	if (initrd_start) {
+		/* Make sure the initrd memory region is not modified. */
+		if (reserve_bootmem(initrd_start, initrd_end - initrd_start,
+				    BOOTMEM_EXCLUSIVE)) {
+			pr_crit("The initrd memory region has been polluted. Disabling it.\n");
+			initrd_start = 0;
+			initrd_end = 0;
+		} else {
+			/*
+			 * Translate initrd_start & initrd_end from PA to VA for
+			 * future access.
+			 */
+			initrd_start += PAGE_OFFSET;
+			initrd_end += PAGE_OFFSET;
+		}
+	}
+#endif
+
+#ifdef CONFIG_KEXEC
+	if (crashk_res.start != crashk_res.end)
+		reserve_bootmem(crashk_res.start, resource_size(&crashk_res),
+				BOOTMEM_DEFAULT);
+#endif
+}
+
+void *__init alloc_remap(int nid, unsigned long size)
+{
+	int pages = node_end_pfn[nid] - node_start_pfn[nid];
+	void *map = pfn_to_kaddr(node_memmap_pfn[nid]);
+	BUG_ON(size != pages * sizeof(struct page));
+	memset(map, 0, size);
+	return map;
+}
+
+static int __init percpu_size(void)
+{
+	int size = __per_cpu_end - __per_cpu_start;
+	size += PERCPU_MODULE_RESERVE;
+	size += PERCPU_DYNAMIC_EARLY_SIZE;
+	if (size < PCPU_MIN_UNIT_SIZE)
+		size = PCPU_MIN_UNIT_SIZE;
+	size = roundup(size, PAGE_SIZE);
+
+	/* In several places we assume the per-cpu data fits on a huge page. */
+	BUG_ON(kdata_huge && size > HPAGE_SIZE);
+	return size;
+}
+
+static void __init zone_sizes_init(void)
+{
+	unsigned long zones_size[MAX_NR_ZONES] = { 0 };
+	int size = percpu_size();
+	int num_cpus = smp_height * smp_width;
+	const unsigned long dma_end = (1UL << (32 - PAGE_SHIFT));
+
+	int i;
+
+	for (i = 0; i < num_cpus; ++i)
+		node_percpu[cpu_to_node(i)] += size;
+
+	for_each_online_node(i) {
+		unsigned long start = node_start_pfn[i];
+		unsigned long end = node_end_pfn[i];
+#ifdef CONFIG_HIGHMEM
+		unsigned long lowmem_end = node_lowmem_end_pfn[i];
+#else
+		unsigned long lowmem_end = end;
+#endif
+		int memmap_size = (end - start) * sizeof(struct page);
+		node_free_pfn[i] = start;
+
+		/*
+		 * Set aside pages for per-cpu data and the mem_map array.
+		 *
+		 * Since the per-cpu data requires special homecaching,
+		 * if we are in kdata_huge mode, we put it at the end of
+		 * the lowmem region.  If we're not in kdata_huge mode,
+		 * we take the per-cpu pages from the bottom of the
+		 * controller, since that avoids fragmenting a huge page
+		 * that users might want.  We always take the memmap
+		 * from the bottom of the controller, since with
+		 * kdata_huge that lets it be under a huge TLB entry.
+		 *
+		 * If the user has requested isolnodes for a controller,
+		 * though, there'll be no lowmem, so we just alloc_bootmem
+		 * the memmap.  There will be no percpu memory either.
+		 */
+		if (i != 0 && node_isset(i, isolnodes)) {
+			node_memmap_pfn[i] =
+				alloc_bootmem_pfn(0, memmap_size, 0);
+			BUG_ON(node_percpu[i] != 0);
+		} else if (node_has_bootmem(start)) {
+			unsigned long goal = 0;
+			node_memmap_pfn[i] =
+				alloc_bootmem_pfn(i, memmap_size, 0);
+			if (kdata_huge)
+				goal = PFN_PHYS(lowmem_end) - node_percpu[i];
+			if (node_percpu[i])
+				node_percpu_pfn[i] =
+					alloc_bootmem_pfn(i, node_percpu[i],
+							  goal);
+		} else {
+			/* In non-bootmem zones, just reserve some pages. */
+			node_memmap_pfn[i] = node_free_pfn[i];
+			node_free_pfn[i] += PFN_UP(memmap_size);
+			if (!kdata_huge) {
+				node_percpu_pfn[i] = node_free_pfn[i];
+				node_free_pfn[i] += PFN_UP(node_percpu[i]);
+			} else {
+				node_percpu_pfn[i] =
+					lowmem_end - PFN_UP(node_percpu[i]);
+			}
+		}
+
+#ifdef CONFIG_HIGHMEM
+		if (start > lowmem_end) {
+			zones_size[ZONE_NORMAL] = 0;
+			zones_size[ZONE_HIGHMEM] = end - start;
+		} else {
+			zones_size[ZONE_NORMAL] = lowmem_end - start;
+			zones_size[ZONE_HIGHMEM] = end - lowmem_end;
+		}
+#else
+		zones_size[ZONE_NORMAL] = end - start;
+#endif
+
+		if (start < dma_end) {
+			zones_size[ZONE_DMA] = min(zones_size[ZONE_NORMAL],
+						   dma_end - start);
+			zones_size[ZONE_NORMAL] -= zones_size[ZONE_DMA];
+		} else {
+			zones_size[ZONE_DMA] = 0;
+		}
+
+		/* Take zone metadata from controller 0 if we're isolnode. */
+		if (node_isset(i, isolnodes))
+			NODE_DATA(i)->bdata = &bootmem_node_data[0];
+
+		free_area_init_node(i, zones_size, start, NULL);
+		printk(KERN_DEBUG "  Normal zone: %ld per-cpu pages\n",
+		       PFN_UP(node_percpu[i]));
+
+		/* Track the type of memory on each node */
+		if (zones_size[ZONE_NORMAL] || zones_size[ZONE_DMA])
+			node_set_state(i, N_NORMAL_MEMORY);
+#ifdef CONFIG_HIGHMEM
+		if (end != start)
+			node_set_state(i, N_HIGH_MEMORY);
+#endif
+
+		node_set_online(i);
+	}
+}
+
+#ifdef CONFIG_NUMA
+
+/* which logical CPUs are on which nodes */
+struct cpumask node_2_cpu_mask[MAX_NUMNODES] __write_once;
+EXPORT_SYMBOL(node_2_cpu_mask);
+
+/* which node each logical CPU is on */
+char cpu_2_node[NR_CPUS] __write_once __attribute__((aligned(L2_CACHE_BYTES)));
+EXPORT_SYMBOL(cpu_2_node);
+
+/* Return cpu_to_node() except for cpus not yet assigned, which return -1 */
+static int __init cpu_to_bound_node(int cpu, struct cpumask* unbound_cpus)
+{
+	if (!cpu_possible(cpu) || cpumask_test_cpu(cpu, unbound_cpus))
+		return -1;
+	else
+		return cpu_to_node(cpu);
+}
+
+/* Return number of immediately-adjacent tiles sharing the same NUMA node. */
+static int __init node_neighbors(int node, int cpu,
+				 struct cpumask *unbound_cpus)
+{
+	int neighbors = 0;
+	int w = smp_width;
+	int h = smp_height;
+	int x = cpu % w;
+	int y = cpu / w;
+	if (x > 0 && cpu_to_bound_node(cpu-1, unbound_cpus) == node)
+		++neighbors;
+	if (x < w-1 && cpu_to_bound_node(cpu+1, unbound_cpus) == node)
+		++neighbors;
+	if (y > 0 && cpu_to_bound_node(cpu-w, unbound_cpus) == node)
+		++neighbors;
+	if (y < h-1 && cpu_to_bound_node(cpu+w, unbound_cpus) == node)
+		++neighbors;
+	return neighbors;
+}
+
+static void __init setup_numa_mapping(void)
+{
+	int distance[MAX_NUMNODES][NR_CPUS];
+	HV_Coord coord;
+	int cpu, node, cpus, i, x, y;
+	int num_nodes = num_online_nodes();
+	struct cpumask unbound_cpus;
+	nodemask_t default_nodes;
+
+	cpumask_clear(&unbound_cpus);
+
+	/* Get set of nodes we will use for defaults */
+	nodes_andnot(default_nodes, node_online_map, isolnodes);
+	if (nodes_empty(default_nodes)) {
+		BUG_ON(!node_isset(0, node_online_map));
+		pr_err("Forcing NUMA node zero available as a default node\n");
+		node_set(0, default_nodes);
+	}
+
+	/* Populate the distance[] array */
+	memset(distance, -1, sizeof(distance));
+	cpu = 0;
+	for (coord.y = 0; coord.y < smp_height; ++coord.y) {
+		for (coord.x = 0; coord.x < smp_width;
+		     ++coord.x, ++cpu) {
+			BUG_ON(cpu >= nr_cpu_ids);
+			if (!cpu_possible(cpu)) {
+				cpu_2_node[cpu] = -1;
+				continue;
+			}
+			for_each_node_mask(node, default_nodes) {
+				HV_MemoryControllerInfo info =
+					hv_inquire_memory_controller(
+						coord, node_controller[node]);
+				distance[node][cpu] =
+					ABS(info.coord.x) + ABS(info.coord.y);
+			}
+			cpumask_set_cpu(cpu, &unbound_cpus);
+		}
+	}
+	cpus = cpu;
+
+	/*
+	 * Round-robin through the NUMA nodes until all the cpus are
+	 * assigned.  We could be more clever here (e.g. create four
+	 * sorted linked lists on the same set of cpu nodes, and pull
+	 * off them in round-robin sequence, removing from all four
+	 * lists each time) but given the relatively small numbers
+	 * involved, O(n^2) seem OK for a one-time cost.
+	 */
+	node = first_node(default_nodes);
+	while (!cpumask_empty(&unbound_cpus)) {
+		int best_cpu = -1;
+		int best_distance = INT_MAX;
+		for (cpu = 0; cpu < cpus; ++cpu) {
+			if (cpumask_test_cpu(cpu, &unbound_cpus)) {
+				/*
+				 * Compute metric, which is how much
+				 * closer the cpu is to this memory
+				 * controller than the others, shifted
+				 * up, and then the number of
+				 * neighbors already in the node as an
+				 * epsilon adjustment to try to keep
+				 * the nodes compact.
+				 */
+				int d = distance[node][cpu] * num_nodes;
+				for_each_node_mask(i, default_nodes) {
+					if (i != node)
+						d -= distance[i][cpu];
+				}
+				d *= 8;  /* allow space for epsilon */
+				d -= node_neighbors(node, cpu, &unbound_cpus);
+				if (d < best_distance) {
+					best_cpu = cpu;
+					best_distance = d;
+				}
+			}
+		}
+		BUG_ON(best_cpu < 0);
+		cpumask_set_cpu(best_cpu, &node_2_cpu_mask[node]);
+		cpu_2_node[best_cpu] = node;
+		cpumask_clear_cpu(best_cpu, &unbound_cpus);
+		node = next_node(node, default_nodes);
+		if (node == MAX_NUMNODES)
+			node = first_node(default_nodes);
+	}
+
+	/* Print out node assignments and set defaults for disabled cpus */
+	cpu = 0;
+	for (y = 0; y < smp_height; ++y) {
+		printk(KERN_DEBUG "NUMA cpu-to-node row %d:", y);
+		for (x = 0; x < smp_width; ++x, ++cpu) {
+			if (cpu_to_node(cpu) < 0) {
+				pr_cont(" -");
+				cpu_2_node[cpu] = first_node(default_nodes);
+			} else {
+				pr_cont(" %d", cpu_to_node(cpu));
+			}
+		}
+		pr_cont("\n");
+	}
+}
+
+static struct cpu cpu_devices[NR_CPUS];
+
+static int __init topology_init(void)
+{
+	int i;
+
+	for_each_online_node(i)
+		register_one_node(i);
+
+	for (i = 0; i < smp_height * smp_width; ++i)
+		register_cpu(&cpu_devices[i], i);
+
+	return 0;
+}
+
+subsys_initcall(topology_init);
+
+#else /* !CONFIG_NUMA */
+
+#define setup_numa_mapping() do { } while (0)
+
+#endif /* CONFIG_NUMA */
+
+/*
+ * Initialize hugepage support on this cpu.  We do this on all cores
+ * early in boot: before argument parsing for the boot cpu, and after
+ * argument parsing but before the init functions run on the secondaries.
+ * So the values we set up here in the hypervisor may be overridden on
+ * the boot cpu as arguments are parsed.
+ */
+static void init_super_pages(void)
+{
+#ifdef CONFIG_HUGETLB_SUPER_PAGES
+	int i;
+	for (i = 0; i < HUGE_SHIFT_ENTRIES; ++i)
+		hv_set_pte_super_shift(i, huge_shift[i]);
+#endif
+}
+
+/**
+ * setup_cpu() - Do all necessary per-cpu, tile-specific initialization.
+ * @boot: Is this the boot cpu?
+ *
+ * Called from setup_arch() on the boot cpu, or online_secondary().
+ */
+void setup_cpu(int boot)
+{
+	/* The boot cpu sets up its permanent mappings much earlier. */
+	if (!boot)
+		store_permanent_mappings();
+
+	/* Allow asynchronous TLB interrupts. */
+#if CHIP_HAS_TILE_DMA()
+	arch_local_irq_unmask(INT_DMATLB_MISS);
+	arch_local_irq_unmask(INT_DMATLB_ACCESS);
+#endif
+#ifdef __tilegx__
+	arch_local_irq_unmask(INT_SINGLE_STEP_K);
+#endif
+
+	/*
+	 * Allow user access to many generic SPRs, like the cycle
+	 * counter, PASS/FAIL/DONE, INTERRUPT_CRITICAL_SECTION, etc.
+	 */
+	__insn_mtspr(SPR_MPL_WORLD_ACCESS_SET_0, 1);
+
+#if CHIP_HAS_SN()
+	/* Static network is not restricted. */
+	__insn_mtspr(SPR_MPL_SN_ACCESS_SET_0, 1);
+#endif
+
+	/*
+	 * Set the MPL for interrupt control 0 & 1 to the corresponding
+	 * values.  This includes access to the SYSTEM_SAVE and EX_CONTEXT
+	 * SPRs, as well as the interrupt mask.
+	 */
+	__insn_mtspr(SPR_MPL_INTCTRL_0_SET_0, 1);
+	__insn_mtspr(SPR_MPL_INTCTRL_1_SET_1, 1);
+
+	/* Initialize IRQ support for this cpu. */
+	setup_irq_regs();
+
+#ifdef CONFIG_HARDWALL
+	/* Reset the network state on this cpu. */
+	reset_network_state();
+#endif
+
+	init_super_pages();
+}
+
+#ifdef CONFIG_BLK_DEV_INITRD
+
+static int __initdata set_initramfs_file;
+static char __initdata initramfs_file[128] = "initramfs";
+
+static int __init setup_initramfs_file(char *str)
+{
+	if (str == NULL)
+		return -EINVAL;
+	strncpy(initramfs_file, str, sizeof(initramfs_file) - 1);
+	set_initramfs_file = 1;
+
+	return 0;
+}
+early_param("initramfs_file", setup_initramfs_file);
+
+/*
+ * We look for a file called "initramfs" in the hvfs.  If there is one, we
+ * allocate some memory for it and it will be unpacked to the initramfs.
+ * If it's compressed, the initd code will uncompress it first.
+ */
+static void __init load_hv_initrd(void)
+{
+	HV_FS_StatInfo stat;
+	int fd, rc;
+	void *initrd;
+
+	/* If initrd has already been set, skip initramfs file in hvfs. */
+	if (initrd_start)
+		return;
+
+	fd = hv_fs_findfile((HV_VirtAddr) initramfs_file);
+	if (fd == HV_ENOENT) {
+		if (set_initramfs_file) {
+			pr_warn("No such hvfs initramfs file '%s'\n",
+				initramfs_file);
+			return;
+		} else {
+			/* Try old backwards-compatible name. */
+			fd = hv_fs_findfile((HV_VirtAddr)"initramfs.cpio.gz");
+			if (fd == HV_ENOENT)
+				return;
+		}
+	}
+	BUG_ON(fd < 0);
+	stat = hv_fs_fstat(fd);
+	BUG_ON(stat.size < 0);
+	if (stat.flags & HV_FS_ISDIR) {
+		pr_warn("Ignoring hvfs file '%s': it's a directory\n",
+			initramfs_file);
+		return;
+	}
+	initrd = alloc_bootmem_pages(stat.size);
+	rc = hv_fs_pread(fd, (HV_VirtAddr) initrd, stat.size, 0);
+	if (rc != stat.size) {
+		pr_err("Error reading %d bytes from hvfs file '%s': %d\n",
+		       stat.size, initramfs_file, rc);
+		free_initrd_mem((unsigned long) initrd, stat.size);
+		return;
+	}
+	initrd_start = (unsigned long) initrd;
+	initrd_end = initrd_start + stat.size;
+}
+
+void __init free_initrd_mem(unsigned long begin, unsigned long end)
+{
+	free_bootmem(__pa(begin), end - begin);
+}
+
+static int __init setup_initrd(char *str)
+{
+	char *endp;
+	unsigned long initrd_size;
+
+	initrd_size = str ? simple_strtoul(str, &endp, 0) : 0;
+	if (initrd_size == 0 || *endp != '@')
+		return -EINVAL;
+
+	initrd_start = simple_strtoul(endp+1, &endp, 0);
+	if (initrd_start == 0)
+		return -EINVAL;
+
+	initrd_end = initrd_start + initrd_size;
+
+	return 0;
+}
+early_param("initrd", setup_initrd);
+
+#else
+static inline void load_hv_initrd(void) {}
+#endif /* CONFIG_BLK_DEV_INITRD */
+
+static void __init validate_hv(void)
+{
+	/*
+	 * It may already be too late, but let's check our built-in
+	 * configuration against what the hypervisor is providing.
+	 */
+	unsigned long glue_size = hv_sysconf(HV_SYSCONF_GLUE_SIZE);
+	int hv_page_size = hv_sysconf(HV_SYSCONF_PAGE_SIZE_SMALL);
+	int hv_hpage_size = hv_sysconf(HV_SYSCONF_PAGE_SIZE_LARGE);
+	HV_ASIDRange asid_range;
+
+#ifndef CONFIG_SMP
+	HV_Topology topology = hv_inquire_topology();
+	BUG_ON(topology.coord.x != 0 || topology.coord.y != 0);
+	if (topology.width != 1 || topology.height != 1) {
+		pr_warn("Warning: booting UP kernel on %dx%d grid; will ignore all but first tile\n",
+			topology.width, topology.height);
+	}
+#endif
+
+	if (PAGE_OFFSET + HV_GLUE_START_CPA + glue_size > (unsigned long)_text)
+		early_panic("Hypervisor glue size %ld is too big!\n",
+			    glue_size);
+	if (hv_page_size != PAGE_SIZE)
+		early_panic("Hypervisor page size %#x != our %#lx\n",
+			    hv_page_size, PAGE_SIZE);
+	if (hv_hpage_size != HPAGE_SIZE)
+		early_panic("Hypervisor huge page size %#x != our %#lx\n",
+			    hv_hpage_size, HPAGE_SIZE);
+
+#ifdef CONFIG_SMP
+	/*
+	 * Some hypervisor APIs take a pointer to a bitmap array
+	 * whose size is at least the number of cpus on the chip.
+	 * We use a struct cpumask for this, so it must be big enough.
+	 */
+	if ((smp_height * smp_width) > nr_cpu_ids)
+		early_panic("Hypervisor %d x %d grid too big for Linux NR_CPUS %d\n",
+			    smp_height, smp_width, nr_cpu_ids);
+#endif
+
+	/*
+	 * Check that we're using allowed ASIDs, and initialize the
+	 * various asid variables to their appropriate initial states.
+	 */
+	asid_range = hv_inquire_asid(0);
+	min_asid = asid_range.start;
+	__this_cpu_write(current_asid, min_asid);
+	max_asid = asid_range.start + asid_range.size - 1;
+
+	if (hv_confstr(HV_CONFSTR_CHIP_MODEL, (HV_VirtAddr)chip_model,
+		       sizeof(chip_model)) < 0) {
+		pr_err("Warning: HV_CONFSTR_CHIP_MODEL not available\n");
+		strlcpy(chip_model, "unknown", sizeof(chip_model));
+	}
+}
+
+static void __init validate_va(void)
+{
+#ifndef __tilegx__   /* FIXME: GX: probably some validation relevant here */
+	/*
+	 * Similarly, make sure we're only using allowed VAs.
+	 * We assume we can contiguously use MEM_USER_INTRPT .. MEM_HV_START,
+	 * and 0 .. KERNEL_HIGH_VADDR.
+	 * In addition, make sure we CAN'T use the end of memory, since
+	 * we use the last chunk of each pgd for the pgd_list.
+	 */
+	int i, user_kernel_ok = 0;
+	unsigned long max_va = 0;
+	unsigned long list_va =
+		((PGD_LIST_OFFSET / sizeof(pgd_t)) << PGDIR_SHIFT);
+
+	for (i = 0; ; ++i) {
+		HV_VirtAddrRange range = hv_inquire_virtual(i);
+		if (range.size == 0)
+			break;
+		if (range.start <= MEM_USER_INTRPT &&
+		    range.start + range.size >= MEM_HV_START)
+			user_kernel_ok = 1;
+		if (range.start == 0)
+			max_va = range.size;
+		BUG_ON(range.start + range.size > list_va);
+	}
+	if (!user_kernel_ok)
+		early_panic("Hypervisor not configured for user/kernel VAs\n");
+	if (max_va == 0)
+		early_panic("Hypervisor not configured for low VAs\n");
+	if (max_va < KERNEL_HIGH_VADDR)
+		early_panic("Hypervisor max VA %#lx smaller than %#lx\n",
+			    max_va, KERNEL_HIGH_VADDR);
+
+	/* Kernel PCs must have their high bit set; see intvec.S. */
+	if ((long)VMALLOC_START >= 0)
+		early_panic("Linux VMALLOC region below the 2GB line (%#lx)!\n"
+			    "Reconfigure the kernel with smaller VMALLOC_RESERVE\n",
+			    VMALLOC_START);
+#endif
+}
+
+/*
+ * cpu_lotar_map lists all the cpus that are valid for the supervisor
+ * to cache data on at a page level, i.e. what cpus can be placed in
+ * the LOTAR field of a PTE.  It is equivalent to the set of possible
+ * cpus plus any other cpus that are willing to share their cache.
+ * It is set by hv_inquire_tiles(HV_INQ_TILES_LOTAR).
+ */
+struct cpumask __write_once cpu_lotar_map;
+EXPORT_SYMBOL(cpu_lotar_map);
+
+/*
+ * hash_for_home_map lists all the tiles that hash-for-home data
+ * will be cached on.  Note that this may includes tiles that are not
+ * valid for this supervisor to use otherwise (e.g. if a hypervisor
+ * device is being shared between multiple supervisors).
+ * It is set by hv_inquire_tiles(HV_INQ_TILES_HFH_CACHE).
+ */
+struct cpumask hash_for_home_map;
+EXPORT_SYMBOL(hash_for_home_map);
+
+/*
+ * cpu_cacheable_map lists all the cpus whose caches the hypervisor can
+ * flush on our behalf.  It is set to cpu_possible_mask OR'ed with
+ * hash_for_home_map, and it is what should be passed to
+ * hv_flush_remote() to flush all caches.  Note that if there are
+ * dedicated hypervisor driver tiles that have authorized use of their
+ * cache, those tiles will only appear in cpu_lotar_map, NOT in
+ * cpu_cacheable_map, as they are a special case.
+ */
+struct cpumask __write_once cpu_cacheable_map;
+EXPORT_SYMBOL(cpu_cacheable_map);
+
+static __initdata struct cpumask disabled_map;
+
+static int __init disabled_cpus(char *str)
+{
+	int boot_cpu = smp_processor_id();
+
+	if (str == NULL || cpulist_parse_crop(str, &disabled_map) != 0)
+		return -EINVAL;
+	if (cpumask_test_cpu(boot_cpu, &disabled_map)) {
+		pr_err("disabled_cpus: can't disable boot cpu %d\n", boot_cpu);
+		cpumask_clear_cpu(boot_cpu, &disabled_map);
+	}
+	return 0;
+}
+
+early_param("disabled_cpus", disabled_cpus);
+
+void __init print_disabled_cpus(void)
+{
+	if (!cpumask_empty(&disabled_map))
+		pr_info("CPUs not available for Linux: %*pbl\n",
+			cpumask_pr_args(&disabled_map));
+}
+
+static void __init setup_cpu_maps(void)
+{
+	struct cpumask hv_disabled_map, cpu_possible_init;
+	int boot_cpu = smp_processor_id();
+	int cpus, i, rc;
+
+	/* Learn which cpus are allowed by the hypervisor. */
+	rc = hv_inquire_tiles(HV_INQ_TILES_AVAIL,
+			      (HV_VirtAddr) cpumask_bits(&cpu_possible_init),
+			      sizeof(cpu_cacheable_map));
+	if (rc < 0)
+		early_panic("hv_inquire_tiles(AVAIL) failed: rc %d\n", rc);
+	if (!cpumask_test_cpu(boot_cpu, &cpu_possible_init))
+		early_panic("Boot CPU %d disabled by hypervisor!\n", boot_cpu);
+
+	/* Compute the cpus disabled by the hvconfig file. */
+	cpumask_complement(&hv_disabled_map, &cpu_possible_init);
+
+	/* Include them with the cpus disabled by "disabled_cpus". */
+	cpumask_or(&disabled_map, &disabled_map, &hv_disabled_map);
+
+	/*
+	 * Disable every cpu after "setup_max_cpus".  But don't mark
+	 * as disabled the cpus that are outside of our initial rectangle,
+	 * since that turns out to be confusing.
+	 */
+	cpus = 1;                          /* this cpu */
+	cpumask_set_cpu(boot_cpu, &disabled_map);   /* ignore this cpu */
+	for (i = 0; cpus < setup_max_cpus; ++i)
+		if (!cpumask_test_cpu(i, &disabled_map))
+			++cpus;
+	for (; i < smp_height * smp_width; ++i)
+		cpumask_set_cpu(i, &disabled_map);
+	cpumask_clear_cpu(boot_cpu, &disabled_map); /* reset this cpu */
+	for (i = smp_height * smp_width; i < NR_CPUS; ++i)
+		cpumask_clear_cpu(i, &disabled_map);
+
+	/*
+	 * Setup cpu_possible map as every cpu allocated to us, minus
+	 * the results of any "disabled_cpus" settings.
+	 */
+	cpumask_andnot(&cpu_possible_init, &cpu_possible_init, &disabled_map);
+	init_cpu_possible(&cpu_possible_init);
+
+	/* Learn which cpus are valid for LOTAR caching. */
+	rc = hv_inquire_tiles(HV_INQ_TILES_LOTAR,
+			      (HV_VirtAddr) cpumask_bits(&cpu_lotar_map),
+			      sizeof(cpu_lotar_map));
+	if (rc < 0) {
+		pr_err("warning: no HV_INQ_TILES_LOTAR; using AVAIL\n");
+		cpu_lotar_map = *cpu_possible_mask;
+	}
+
+	/* Retrieve set of CPUs used for hash-for-home caching */
+	rc = hv_inquire_tiles(HV_INQ_TILES_HFH_CACHE,
+			      (HV_VirtAddr) hash_for_home_map.bits,
+			      sizeof(hash_for_home_map));
+	if (rc < 0)
+		early_panic("hv_inquire_tiles(HFH_CACHE) failed: rc %d\n", rc);
+	cpumask_or(&cpu_cacheable_map, cpu_possible_mask, &hash_for_home_map);
+}
+
+
+static int __init dataplane(char *str)
+{
+	pr_warn("WARNING: dataplane support disabled in this kernel\n");
+	return 0;
+}
+
+early_param("dataplane", dataplane);
+
+#ifdef CONFIG_NO_HZ_FULL
+/* Warn if hypervisor shared cpus are marked as nohz_full. */
+static int __init check_nohz_full_cpus(void)
+{
+	struct cpumask shared;
+	int cpu;
+
+	if (hv_inquire_tiles(HV_INQ_TILES_SHARED,
+			     (HV_VirtAddr) shared.bits, sizeof(shared)) < 0) {
+		pr_warn("WARNING: No support for inquiring hv shared tiles\n");
+		return 0;
+	}
+	for_each_cpu(cpu, &shared) {
+		if (tick_nohz_full_cpu(cpu))
+			pr_warn("WARNING: nohz_full cpu %d receives hypervisor interrupts!\n",
+			       cpu);
+	}
+	return 0;
+}
+arch_initcall(check_nohz_full_cpus);
+#endif
+
+#ifdef CONFIG_CMDLINE_BOOL
+static char __initdata builtin_cmdline[COMMAND_LINE_SIZE] = CONFIG_CMDLINE;
+#endif
+
+void __init setup_arch(char **cmdline_p)
+{
+	int len;
+
+#if defined(CONFIG_CMDLINE_BOOL) && defined(CONFIG_CMDLINE_OVERRIDE)
+	len = hv_get_command_line((HV_VirtAddr) boot_command_line,
+				  COMMAND_LINE_SIZE);
+	if (boot_command_line[0])
+		pr_warn("WARNING: ignoring dynamic command line \"%s\"\n",
+			boot_command_line);
+	strlcpy(boot_command_line, builtin_cmdline, COMMAND_LINE_SIZE);
+#else
+	char *hv_cmdline;
+#if defined(CONFIG_CMDLINE_BOOL)
+	if (builtin_cmdline[0]) {
+		int builtin_len = strlcpy(boot_command_line, builtin_cmdline,
+					  COMMAND_LINE_SIZE);
+		if (builtin_len < COMMAND_LINE_SIZE-1)
+			boot_command_line[builtin_len++] = ' ';
+		hv_cmdline = &boot_command_line[builtin_len];
+		len = COMMAND_LINE_SIZE - builtin_len;
+	} else
+#endif
+	{
+		hv_cmdline = boot_command_line;
+		len = COMMAND_LINE_SIZE;
+	}
+	len = hv_get_command_line((HV_VirtAddr) hv_cmdline, len);
+	if (len < 0 || len > COMMAND_LINE_SIZE)
+		early_panic("hv_get_command_line failed: %d\n", len);
+#endif
+
+	*cmdline_p = boot_command_line;
+
+	/* Set disabled_map and setup_max_cpus very early */
+	parse_early_param();
+
+	/* Make sure the kernel is compatible with the hypervisor. */
+	validate_hv();
+	validate_va();
+
+	setup_cpu_maps();
+
+
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
+	/*
+	 * Initialize the PCI structures.  This is done before memory
+	 * setup so that we know whether or not a pci_reserve region
+	 * is necessary.
+	 */
+	if (tile_pci_init() == 0)
+		pci_reserve_mb = 0;
+
+	/* PCI systems reserve a region just below 4GB for mapping iomem. */
+	pci_reserve_end_pfn  = (1 << (32 - PAGE_SHIFT));
+	pci_reserve_start_pfn = pci_reserve_end_pfn -
+		(pci_reserve_mb << (20 - PAGE_SHIFT));
+#endif
+
+	init_mm.start_code = (unsigned long) _text;
+	init_mm.end_code = (unsigned long) _etext;
+	init_mm.end_data = (unsigned long) _edata;
+	init_mm.brk = (unsigned long) _end;
+
+	setup_memory();
+	store_permanent_mappings();
+	setup_bootmem_allocator();
+
+	/*
+	 * NOTE: before this point _nobody_ is allowed to allocate
+	 * any memory using the bootmem allocator.
+	 */
+
+#ifdef CONFIG_SWIOTLB
+	swiotlb_init(0);
+#endif
+
+	paging_init();
+	setup_numa_mapping();
+	zone_sizes_init();
+	set_page_homes();
+	setup_cpu(1);
+	setup_clock();
+	load_hv_initrd();
+}
+
+
+/*
+ * Set up per-cpu memory.
+ */
+
+unsigned long __per_cpu_offset[NR_CPUS] __write_once;
+EXPORT_SYMBOL(__per_cpu_offset);
+
+static size_t __initdata pfn_offset[MAX_NUMNODES] = { 0 };
+static unsigned long __initdata percpu_pfn[NR_CPUS] = { 0 };
+
+/*
+ * As the percpu code allocates pages, we return the pages from the
+ * end of the node for the specified cpu.
+ */
+static void *__init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align)
+{
+	int nid = cpu_to_node(cpu);
+	unsigned long pfn = node_percpu_pfn[nid] + pfn_offset[nid];
+
+	BUG_ON(size % PAGE_SIZE != 0);
+	pfn_offset[nid] += size / PAGE_SIZE;
+	BUG_ON(node_percpu[nid] < size);
+	node_percpu[nid] -= size;
+	if (percpu_pfn[cpu] == 0)
+		percpu_pfn[cpu] = pfn;
+	return pfn_to_kaddr(pfn);
+}
+
+/*
+ * Pages reserved for percpu memory are not freeable, and in any case we are
+ * on a short path to panic() in setup_per_cpu_area() at this point anyway.
+ */
+static void __init pcpu_fc_free(void *ptr, size_t size)
+{
+}
+
+/*
+ * Set up vmalloc page tables using bootmem for the percpu code.
+ */
+static void __init pcpu_fc_populate_pte(unsigned long addr)
+{
+	pgd_t *pgd;
+	pud_t *pud;
+	pmd_t *pmd;
+	pte_t *pte;
+
+	BUG_ON(pgd_addr_invalid(addr));
+	if (addr < VMALLOC_START || addr >= VMALLOC_END)
+		panic("PCPU addr %#lx outside vmalloc range %#lx..%#lx; try increasing CONFIG_VMALLOC_RESERVE\n",
+		      addr, VMALLOC_START, VMALLOC_END);
+
+	pgd = swapper_pg_dir + pgd_index(addr);
+	pud = pud_offset(pgd, addr);
+	BUG_ON(!pud_present(*pud));
+	pmd = pmd_offset(pud, addr);
+	if (pmd_present(*pmd)) {
+		BUG_ON(pmd_huge_page(*pmd));
+	} else {
+		pte = __alloc_bootmem(L2_KERNEL_PGTABLE_SIZE,
+				      HV_PAGE_TABLE_ALIGN, 0);
+		pmd_populate_kernel(&init_mm, pmd, pte);
+	}
+}
+
+void __init setup_per_cpu_areas(void)
+{
+	struct page *pg;
+	unsigned long delta, pfn, lowmem_va;
+	unsigned long size = percpu_size();
+	char *ptr;
+	int rc, cpu, i;
+
+	rc = pcpu_page_first_chunk(PERCPU_MODULE_RESERVE, pcpu_fc_alloc,
+				   pcpu_fc_free, pcpu_fc_populate_pte);
+	if (rc < 0)
+		panic("Cannot initialize percpu area (err=%d)", rc);
+
+	delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
+	for_each_possible_cpu(cpu) {
+		__per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
+
+		/* finv the copy out of cache so we can change homecache */
+		ptr = pcpu_base_addr + pcpu_unit_offsets[cpu];
+		__finv_buffer(ptr, size);
+		pfn = percpu_pfn[cpu];
+
+		/* Rewrite the page tables to cache on that cpu */
+		pg = pfn_to_page(pfn);
+		for (i = 0; i < size; i += PAGE_SIZE, ++pfn, ++pg) {
+
+			/* Update the vmalloc mapping and page home. */
+			unsigned long addr = (unsigned long)ptr + i;
+			pte_t *ptep = virt_to_kpte(addr);
+			pte_t pte = *ptep;
+			BUG_ON(pfn != pte_pfn(pte));
+			pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_TILE_L3);
+			pte = set_remote_cache_cpu(pte, cpu);
+			set_pte_at(&init_mm, addr, ptep, pte);
+
+			/* Update the lowmem mapping for consistency. */
+			lowmem_va = (unsigned long)pfn_to_kaddr(pfn);
+			ptep = virt_to_kpte(lowmem_va);
+			if (pte_huge(*ptep)) {
+				printk(KERN_DEBUG "early shatter of huge page at %#lx\n",
+				       lowmem_va);
+				shatter_pmd((pmd_t *)ptep);
+				ptep = virt_to_kpte(lowmem_va);
+				BUG_ON(pte_huge(*ptep));
+			}
+			BUG_ON(pfn != pte_pfn(*ptep));
+			set_pte_at(&init_mm, lowmem_va, ptep, pte);
+		}
+	}
+
+	/* Set our thread pointer appropriately. */
+	set_my_cpu_offset(__per_cpu_offset[smp_processor_id()]);
+
+	/* Make sure the finv's have completed. */
+	mb_incoherent();
+
+	/* Flush the TLB so we reference it properly from here on out. */
+	local_flush_tlb_all();
+}
+
+static struct resource data_resource = {
+	.name	= "Kernel data",
+	.start	= 0,
+	.end	= 0,
+	.flags	= IORESOURCE_BUSY | IORESOURCE_MEM
+};
+
+static struct resource code_resource = {
+	.name	= "Kernel code",
+	.start	= 0,
+	.end	= 0,
+	.flags	= IORESOURCE_BUSY | IORESOURCE_MEM
+};
+
+/*
+ * On Pro, we reserve all resources above 4GB so that PCI won't try to put
+ * mappings above 4GB.
+ */
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
+static struct resource* __init
+insert_non_bus_resource(void)
+{
+	struct resource *res =
+		kzalloc(sizeof(struct resource), GFP_ATOMIC);
+	if (!res)
+		return NULL;
+	res->name = "Non-Bus Physical Address Space";
+	res->start = (1ULL << 32);
+	res->end = -1LL;
+	res->flags = IORESOURCE_BUSY | IORESOURCE_MEM;
+	if (insert_resource(&iomem_resource, res)) {
+		kfree(res);
+		return NULL;
+	}
+	return res;
+}
+#endif
+
+static struct resource* __init
+insert_ram_resource(u64 start_pfn, u64 end_pfn, bool reserved)
+{
+	struct resource *res =
+		kzalloc(sizeof(struct resource), GFP_ATOMIC);
+	if (!res)
+		return NULL;
+	res->name = reserved ? "Reserved" : "System RAM";
+	res->start = start_pfn << PAGE_SHIFT;
+	res->end = (end_pfn << PAGE_SHIFT) - 1;
+	res->flags = IORESOURCE_BUSY | IORESOURCE_MEM;
+	if (insert_resource(&iomem_resource, res)) {
+		kfree(res);
+		return NULL;
+	}
+	return res;
+}
+
+/*
+ * Request address space for all standard resources
+ *
+ * If the system includes PCI root complex drivers, we need to create
+ * a window just below 4GB where PCI BARs can be mapped.
+ */
+static int __init request_standard_resources(void)
+{
+	int i;
+	enum { CODE_DELTA = MEM_SV_START - PAGE_OFFSET };
+
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
+	insert_non_bus_resource();
+#endif
+
+	for_each_online_node(i) {
+		u64 start_pfn = node_start_pfn[i];
+		u64 end_pfn = node_end_pfn[i];
+
+#if defined(CONFIG_PCI) && !defined(__tilegx__)
+		if (start_pfn <= pci_reserve_start_pfn &&
+		    end_pfn > pci_reserve_start_pfn) {
+			if (end_pfn > pci_reserve_end_pfn)
+				insert_ram_resource(pci_reserve_end_pfn,
+						    end_pfn, 0);
+			end_pfn = pci_reserve_start_pfn;
+		}
+#endif
+		insert_ram_resource(start_pfn, end_pfn, 0);
+	}
+
+	code_resource.start = __pa(_text - CODE_DELTA);
+	code_resource.end = __pa(_etext - CODE_DELTA)-1;
+	data_resource.start = __pa(_sdata);
+	data_resource.end = __pa(_end)-1;
+
+	insert_resource(&iomem_resource, &code_resource);
+	insert_resource(&iomem_resource, &data_resource);
+
+	/* Mark any "memmap" regions busy for the resource manager. */
+	for (i = 0; i < memmap_nr; ++i) {
+		struct memmap_entry *m = &memmap_map[i];
+		insert_ram_resource(PFN_DOWN(m->addr),
+				    PFN_UP(m->addr + m->size - 1), 1);
+	}
+
+#ifdef CONFIG_KEXEC
+	insert_resource(&iomem_resource, &crashk_res);
+#endif
+
+	return 0;
+}
+
+subsys_initcall(request_standard_resources);
diff --git a/arch/tile/kernel/signal.c b/arch/tile/kernel/signal.c
new file mode 100644
index 000000000..87299a6cf
--- /dev/null
+++ b/arch/tile/kernel/signal.c
@@ -0,0 +1,409 @@
+/*
+ * Copyright (C) 1991, 1992  Linus Torvalds
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/sched.h>
+#include <linux/mm.h>
+#include <linux/smp.h>
+#include <linux/kernel.h>
+#include <linux/signal.h>
+#include <linux/errno.h>
+#include <linux/wait.h>
+#include <linux/unistd.h>
+#include <linux/stddef.h>
+#include <linux/personality.h>
+#include <linux/suspend.h>
+#include <linux/ptrace.h>
+#include <linux/elf.h>
+#include <linux/compat.h>
+#include <linux/syscalls.h>
+#include <linux/uaccess.h>
+#include <asm/processor.h>
+#include <asm/ucontext.h>
+#include <asm/sigframe.h>
+#include <asm/syscalls.h>
+#include <asm/vdso.h>
+#include <arch/interrupts.h>
+
+#define DEBUG_SIG 0
+
+/*
+ * Do a signal return; undo the signal stack.
+ */
+
+int restore_sigcontext(struct pt_regs *regs,
+		       struct sigcontext __user *sc)
+{
+	int err;
+
+	/* Always make any pending restarted system calls return -EINTR */
+	current->restart_block.fn = do_no_restart_syscall;
+
+	/*
+	 * Enforce that sigcontext is like pt_regs, and doesn't mess
+	 * up our stack alignment rules.
+	 */
+	BUILD_BUG_ON(sizeof(struct sigcontext) != sizeof(struct pt_regs));
+	BUILD_BUG_ON(sizeof(struct sigcontext) % 8 != 0);
+	err = __copy_from_user(regs, sc, sizeof(*regs));
+
+	/* Ensure that the PL is always set to USER_PL. */
+	regs->ex1 = PL_ICS_EX1(USER_PL, EX1_ICS(regs->ex1));
+
+	regs->faultnum = INT_SWINT_1_SIGRETURN;
+
+	return err;
+}
+
+void signal_fault(const char *type, struct pt_regs *regs,
+		  void __user *frame, int sig)
+{
+	trace_unhandled_signal(type, regs, (unsigned long)frame, SIGSEGV);
+	force_sigsegv(sig, current);
+}
+
+/* The assembly shim for this function arranges to ignore the return value. */
+SYSCALL_DEFINE0(rt_sigreturn)
+{
+	struct pt_regs *regs = current_pt_regs();
+	struct rt_sigframe __user *frame =
+		(struct rt_sigframe __user *)(regs->sp);
+	sigset_t set;
+
+	if (!access_ok(VERIFY_READ, frame, sizeof(*frame)))
+		goto badframe;
+	if (__copy_from_user(&set, &frame->uc.uc_sigmask, sizeof(set)))
+		goto badframe;
+
+	set_current_blocked(&set);
+
+	if (restore_sigcontext(regs, &frame->uc.uc_mcontext))
+		goto badframe;
+
+	if (restore_altstack(&frame->uc.uc_stack))
+		goto badframe;
+
+	return 0;
+
+badframe:
+	signal_fault("bad sigreturn frame", regs, frame, 0);
+	return 0;
+}
+
+/*
+ * Set up a signal frame.
+ */
+
+int setup_sigcontext(struct sigcontext __user *sc, struct pt_regs *regs)
+{
+	return  __copy_to_user(sc, regs, sizeof(*regs));
+}
+
+/*
+ * Determine which stack to use..
+ */
+static inline void __user *get_sigframe(struct k_sigaction *ka,
+					struct pt_regs *regs,
+					size_t frame_size)
+{
+	unsigned long sp;
+
+	/* Default to using normal stack */
+	sp = regs->sp;
+
+	/*
+	 * If we are on the alternate signal stack and would overflow
+	 * it, don't.  Return an always-bogus address instead so we
+	 * will die with SIGSEGV.
+	 */
+	if (on_sig_stack(sp) && !likely(on_sig_stack(sp - frame_size)))
+		return (void __user __force *)-1UL;
+
+	/* This is the X/Open sanctioned signal stack switching.  */
+	if (ka->sa.sa_flags & SA_ONSTACK) {
+		if (sas_ss_flags(sp) == 0)
+			sp = current->sas_ss_sp + current->sas_ss_size;
+	}
+
+	sp -= frame_size;
+	/*
+	 * Align the stack pointer according to the TILE ABI,
+	 * i.e. so that on function entry (sp & 15) == 0.
+	 */
+	sp &= -16UL;
+	return (void __user *) sp;
+}
+
+static int setup_rt_frame(struct ksignal *ksig, sigset_t *set,
+			  struct pt_regs *regs)
+{
+	unsigned long restorer;
+	struct rt_sigframe __user *frame;
+	int err = 0, sig = ksig->sig;
+
+	frame = get_sigframe(&ksig->ka, regs, sizeof(*frame));
+
+	if (!access_ok(VERIFY_WRITE, frame, sizeof(*frame)))
+		goto err;
+
+	/* Always write at least the signal number for the stack backtracer. */
+	if (ksig->ka.sa.sa_flags & SA_SIGINFO) {
+		/* At sigreturn time, restore the callee-save registers too. */
+		err |= copy_siginfo_to_user(&frame->info, &ksig->info);
+		regs->flags |= PT_FLAGS_RESTORE_REGS;
+	} else {
+		err |= __put_user(ksig->info.si_signo, &frame->info.si_signo);
+	}
+
+	/* Create the ucontext.  */
+	err |= __clear_user(&frame->save_area, sizeof(frame->save_area));
+	err |= __put_user(0, &frame->uc.uc_flags);
+	err |= __put_user(NULL, &frame->uc.uc_link);
+	err |= __save_altstack(&frame->uc.uc_stack, regs->sp);
+	err |= setup_sigcontext(&frame->uc.uc_mcontext, regs);
+	err |= __copy_to_user(&frame->uc.uc_sigmask, set, sizeof(*set));
+	if (err)
+		goto err;
+
+	restorer = VDSO_SYM(&__vdso_rt_sigreturn);
+	if (ksig->ka.sa.sa_flags & SA_RESTORER)
+		restorer = (unsigned long) ksig->ka.sa.sa_restorer;
+
+	/*
+	 * Set up registers for signal handler.
+	 * Registers that we don't modify keep the value they had from
+	 * user-space at the time we took the signal.
+	 * We always pass siginfo and mcontext, regardless of SA_SIGINFO,
+	 * since some things rely on this (e.g. glibc's debug/segfault.c).
+	 */
+	regs->pc = (unsigned long) ksig->ka.sa.sa_handler;
+	regs->ex1 = PL_ICS_EX1(USER_PL, 1); /* set crit sec in handler */
+	regs->sp = (unsigned long) frame;
+	regs->lr = restorer;
+	regs->regs[0] = (unsigned long) sig;
+	regs->regs[1] = (unsigned long) &frame->info;
+	regs->regs[2] = (unsigned long) &frame->uc;
+	regs->flags |= PT_FLAGS_CALLER_SAVES;
+	return 0;
+
+err:
+	trace_unhandled_signal("bad sigreturn frame", regs,
+			      (unsigned long)frame, SIGSEGV);
+	return -EFAULT;
+}
+
+/*
+ * OK, we're invoking a handler
+ */
+
+static void handle_signal(struct ksignal *ksig, struct pt_regs *regs)
+{
+	sigset_t *oldset = sigmask_to_save();
+	int ret;
+
+	/* Are we from a system call? */
+	if (regs->faultnum == INT_SWINT_1) {
+		/* If so, check system call restarting.. */
+		switch (regs->regs[0]) {
+		case -ERESTART_RESTARTBLOCK:
+		case -ERESTARTNOHAND:
+			regs->regs[0] = -EINTR;
+			break;
+
+		case -ERESTARTSYS:
+			if (!(ksig->ka.sa.sa_flags & SA_RESTART)) {
+				regs->regs[0] = -EINTR;
+				break;
+			}
+			/* fallthrough */
+		case -ERESTARTNOINTR:
+			/* Reload caller-saves to restore r0..r5 and r10. */
+			regs->flags |= PT_FLAGS_CALLER_SAVES;
+			regs->regs[0] = regs->orig_r0;
+			regs->pc -= 8;
+		}
+	}
+
+	/* Set up the stack frame */
+#ifdef CONFIG_COMPAT
+	if (is_compat_task())
+		ret = compat_setup_rt_frame(ksig, oldset, regs);
+	else
+#endif
+		ret = setup_rt_frame(ksig, oldset, regs);
+
+	signal_setup_done(ret, ksig, test_thread_flag(TIF_SINGLESTEP));
+}
+
+/*
+ * Note that 'init' is a special process: it doesn't get signals it doesn't
+ * want to handle. Thus you cannot kill init even with a SIGKILL even by
+ * mistake.
+ */
+void do_signal(struct pt_regs *regs)
+{
+	struct ksignal ksig;
+
+	/*
+	 * i386 will check if we're coming from kernel mode and bail out
+	 * here.  In my experience this just turns weird crashes into
+	 * weird spin-hangs.  But if we find a case where this seems
+	 * helpful, we can reinstate the check on "!user_mode(regs)".
+	 */
+
+	if (get_signal(&ksig)) {
+		/* Whee! Actually deliver the signal.  */
+		handle_signal(&ksig, regs);
+		goto done;
+	}
+
+	/* Did we come from a system call? */
+	if (regs->faultnum == INT_SWINT_1) {
+		/* Restart the system call - no handlers present */
+		switch (regs->regs[0]) {
+		case -ERESTARTNOHAND:
+		case -ERESTARTSYS:
+		case -ERESTARTNOINTR:
+			regs->flags |= PT_FLAGS_CALLER_SAVES;
+			regs->regs[0] = regs->orig_r0;
+			regs->pc -= 8;
+			break;
+
+		case -ERESTART_RESTARTBLOCK:
+			regs->flags |= PT_FLAGS_CALLER_SAVES;
+			regs->regs[TREG_SYSCALL_NR] = __NR_restart_syscall;
+			regs->pc -= 8;
+			break;
+		}
+	}
+
+	/* If there's no signal to deliver, just put the saved sigmask back. */
+	restore_saved_sigmask();
+
+done:
+	/* Avoid double syscall restart if there are nested signals. */
+	regs->faultnum = INT_SWINT_1_SIGRETURN;
+}
+
+int show_unhandled_signals = 1;
+
+static int __init crashinfo(char *str)
+{
+	const char *word;
+
+	if (*str == '\0')
+		show_unhandled_signals = 2;
+	else if (*str != '=' || kstrtoint(++str, 0, &show_unhandled_signals) != 0)
+		return 0;
+
+	switch (show_unhandled_signals) {
+	case 0:
+		word = "No";
+		break;
+	case 1:
+		word = "One-line";
+		break;
+	default:
+		word = "Detailed";
+		break;
+	}
+	pr_info("%s crash reports will be generated on the console\n", word);
+	return 1;
+}
+__setup("crashinfo", crashinfo);
+
+static void dump_mem(void __user *address)
+{
+	void __user *addr;
+	enum { region_size = 256, bytes_per_line = 16 };
+	int i, j, k;
+	int found_readable_mem = 0;
+
+	if (!access_ok(VERIFY_READ, address, 1)) {
+		pr_err("Not dumping at address 0x%lx (kernel address)\n",
+		       (unsigned long)address);
+		return;
+	}
+
+	addr = (void __user *)
+		(((unsigned long)address & -bytes_per_line) - region_size/2);
+	if (addr > address)
+		addr = NULL;
+	for (i = 0; i < region_size;
+	     addr += bytes_per_line, i += bytes_per_line) {
+		unsigned char buf[bytes_per_line];
+		char line[100];
+		if (copy_from_user(buf, addr, bytes_per_line))
+			continue;
+		if (!found_readable_mem) {
+			pr_err("Dumping memory around address 0x%lx:\n",
+			       (unsigned long)address);
+			found_readable_mem = 1;
+		}
+		j = sprintf(line, REGFMT ":", (unsigned long)addr);
+		for (k = 0; k < bytes_per_line; ++k)
+			j += sprintf(&line[j], " %02x", buf[k]);
+		pr_err("%s\n", line);
+	}
+	if (!found_readable_mem)
+		pr_err("No readable memory around address 0x%lx\n",
+		       (unsigned long)address);
+}
+
+void trace_unhandled_signal(const char *type, struct pt_regs *regs,
+			    unsigned long address, int sig)
+{
+	struct task_struct *tsk = current;
+
+	if (show_unhandled_signals == 0)
+		return;
+
+	/* If the signal is handled, don't show it here. */
+	if (!is_global_init(tsk)) {
+		void __user *handler =
+			tsk->sighand->action[sig-1].sa.sa_handler;
+		if (handler != SIG_IGN && handler != SIG_DFL)
+			return;
+	}
+
+	/* Rate-limit the one-line output, not the detailed output. */
+	if (show_unhandled_signals <= 1 && !printk_ratelimit())
+		return;
+
+	printk("%s%s[%d]: %s at %lx pc "REGFMT" signal %d",
+	       task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
+	       tsk->comm, task_pid_nr(tsk), type, address, regs->pc, sig);
+
+	print_vma_addr(KERN_CONT " in ", regs->pc);
+
+	printk(KERN_CONT "\n");
+
+	if (show_unhandled_signals > 1) {
+		switch (sig) {
+		case SIGILL:
+		case SIGFPE:
+		case SIGSEGV:
+		case SIGBUS:
+			pr_err("User crash: signal %d, trap %ld, address 0x%lx\n",
+			       sig, regs->faultnum, address);
+			show_regs(regs);
+			dump_mem((void __user *)address);
+			break;
+		default:
+			pr_err("User crash: signal %d, trap %ld\n",
+			       sig, regs->faultnum);
+			break;
+		}
+	}
+}
diff --git a/arch/tile/kernel/single_step.c b/arch/tile/kernel/single_step.c
new file mode 100644
index 000000000..53f7b9def
--- /dev/null
+++ b/arch/tile/kernel/single_step.c
@@ -0,0 +1,785 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * A code-rewriter that enables instruction single-stepping.
+ */
+
+#include <linux/smp.h>
+#include <linux/ptrace.h>
+#include <linux/slab.h>
+#include <linux/thread_info.h>
+#include <linux/uaccess.h>
+#include <linux/mman.h>
+#include <linux/types.h>
+#include <linux/err.h>
+#include <linux/prctl.h>
+#include <linux/context_tracking.h>
+#include <asm/cacheflush.h>
+#include <asm/traps.h>
+#include <asm/uaccess.h>
+#include <asm/unaligned.h>
+#include <arch/abi.h>
+#include <arch/spr_def.h>
+#include <arch/opcode.h>
+
+
+#ifndef __tilegx__   /* Hardware support for single step unavailable. */
+
+#define signExtend17(val) sign_extend((val), 17)
+#define TILE_X1_MASK (0xffffffffULL << 31)
+
+enum mem_op {
+	MEMOP_NONE,
+	MEMOP_LOAD,
+	MEMOP_STORE,
+	MEMOP_LOAD_POSTINCR,
+	MEMOP_STORE_POSTINCR
+};
+
+static inline tilepro_bundle_bits set_BrOff_X1(tilepro_bundle_bits n,
+	s32 offset)
+{
+	tilepro_bundle_bits result;
+
+	/* mask out the old offset */
+	tilepro_bundle_bits mask = create_BrOff_X1(-1);
+	result = n & (~mask);
+
+	/* or in the new offset */
+	result |= create_BrOff_X1(offset);
+
+	return result;
+}
+
+static inline tilepro_bundle_bits move_X1(tilepro_bundle_bits n, int dest,
+	int src)
+{
+	tilepro_bundle_bits result;
+	tilepro_bundle_bits op;
+
+	result = n & (~TILE_X1_MASK);
+
+	op = create_Opcode_X1(SPECIAL_0_OPCODE_X1) |
+		create_RRROpcodeExtension_X1(OR_SPECIAL_0_OPCODE_X1) |
+		create_Dest_X1(dest) |
+		create_SrcB_X1(TREG_ZERO) |
+		create_SrcA_X1(src) ;
+
+	result |= op;
+	return result;
+}
+
+static inline tilepro_bundle_bits nop_X1(tilepro_bundle_bits n)
+{
+	return move_X1(n, TREG_ZERO, TREG_ZERO);
+}
+
+static inline tilepro_bundle_bits addi_X1(
+	tilepro_bundle_bits n, int dest, int src, int imm)
+{
+	n &= ~TILE_X1_MASK;
+
+	n |=  (create_SrcA_X1(src) |
+	       create_Dest_X1(dest) |
+	       create_Imm8_X1(imm) |
+	       create_S_X1(0) |
+	       create_Opcode_X1(IMM_0_OPCODE_X1) |
+	       create_ImmOpcodeExtension_X1(ADDI_IMM_0_OPCODE_X1));
+
+	return n;
+}
+
+static tilepro_bundle_bits rewrite_load_store_unaligned(
+	struct single_step_state *state,
+	tilepro_bundle_bits bundle,
+	struct pt_regs *regs,
+	enum mem_op mem_op,
+	int size, int sign_ext)
+{
+	unsigned char __user *addr;
+	int val_reg, addr_reg, err, val;
+	int align_ctl;
+
+	align_ctl = unaligned_fixup;
+	switch (task_thread_info(current)->align_ctl) {
+	case PR_UNALIGN_NOPRINT:
+		align_ctl = 1;
+		break;
+	case PR_UNALIGN_SIGBUS:
+		align_ctl = 0;
+		break;
+	}
+
+	/* Get address and value registers */
+	if (bundle & TILEPRO_BUNDLE_Y_ENCODING_MASK) {
+		addr_reg = get_SrcA_Y2(bundle);
+		val_reg = get_SrcBDest_Y2(bundle);
+	} else if (mem_op == MEMOP_LOAD || mem_op == MEMOP_LOAD_POSTINCR) {
+		addr_reg = get_SrcA_X1(bundle);
+		val_reg  = get_Dest_X1(bundle);
+	} else {
+		addr_reg = get_SrcA_X1(bundle);
+		val_reg  = get_SrcB_X1(bundle);
+	}
+
+	/*
+	 * If registers are not GPRs, don't try to handle it.
+	 *
+	 * FIXME: we could handle non-GPR loads by getting the real value
+	 * from memory, writing it to the single step buffer, using a
+	 * temp_reg to hold a pointer to that memory, then executing that
+	 * instruction and resetting temp_reg.  For non-GPR stores, it's a
+	 * little trickier; we could use the single step buffer for that
+	 * too, but we'd have to add some more state bits so that we could
+	 * call back in here to copy that value to the real target.  For
+	 * now, we just handle the simple case.
+	 */
+	if ((val_reg >= PTREGS_NR_GPRS &&
+	     (val_reg != TREG_ZERO ||
+	      mem_op == MEMOP_LOAD ||
+	      mem_op == MEMOP_LOAD_POSTINCR)) ||
+	    addr_reg >= PTREGS_NR_GPRS)
+		return bundle;
+
+	/* If it's aligned, don't handle it specially */
+	addr = (void __user *)regs->regs[addr_reg];
+	if (((unsigned long)addr % size) == 0)
+		return bundle;
+
+	/*
+	 * Return SIGBUS with the unaligned address, if requested.
+	 * Note that we return SIGBUS even for completely invalid addresses
+	 * as long as they are in fact unaligned; this matches what the
+	 * tilepro hardware would be doing, if it could provide us with the
+	 * actual bad address in an SPR, which it doesn't.
+	 */
+	if (align_ctl == 0) {
+		siginfo_t info = {
+			.si_signo = SIGBUS,
+			.si_code = BUS_ADRALN,
+			.si_addr = addr
+		};
+		trace_unhandled_signal("unaligned trap", regs,
+				       (unsigned long)addr, SIGBUS);
+		force_sig_info(info.si_signo, &info, current);
+		return (tilepro_bundle_bits) 0;
+	}
+
+	/* Handle unaligned load/store */
+	if (mem_op == MEMOP_LOAD || mem_op == MEMOP_LOAD_POSTINCR) {
+		unsigned short val_16;
+		switch (size) {
+		case 2:
+			err = copy_from_user(&val_16, addr, sizeof(val_16));
+			val = sign_ext ? ((short)val_16) : val_16;
+			break;
+		case 4:
+			err = copy_from_user(&val, addr, sizeof(val));
+			break;
+		default:
+			BUG();
+		}
+		if (err == 0) {
+			state->update_reg = val_reg;
+			state->update_value = val;
+			state->update = 1;
+		}
+	} else {
+		unsigned short val_16;
+		val = (val_reg == TREG_ZERO) ? 0 : regs->regs[val_reg];
+		switch (size) {
+		case 2:
+			val_16 = val;
+			err = copy_to_user(addr, &val_16, sizeof(val_16));
+			break;
+		case 4:
+			err = copy_to_user(addr, &val, sizeof(val));
+			break;
+		default:
+			BUG();
+		}
+	}
+
+	if (err) {
+		siginfo_t info = {
+			.si_signo = SIGBUS,
+			.si_code = BUS_ADRALN,
+			.si_addr = addr
+		};
+		trace_unhandled_signal("bad address for unaligned fixup", regs,
+				       (unsigned long)addr, SIGBUS);
+		force_sig_info(info.si_signo, &info, current);
+		return (tilepro_bundle_bits) 0;
+	}
+
+	if (unaligned_printk || unaligned_fixup_count == 0) {
+		pr_info("Process %d/%s: PC %#lx: Fixup of unaligned %s at %#lx\n",
+			current->pid, current->comm, regs->pc,
+			mem_op == MEMOP_LOAD || mem_op == MEMOP_LOAD_POSTINCR ?
+			"load" : "store",
+			(unsigned long)addr);
+		if (!unaligned_printk) {
+#define P pr_info
+P("\n");
+P("Unaligned fixups in the kernel will slow your application considerably.\n");
+P("To find them, write a \"1\" to /proc/sys/tile/unaligned_fixup/printk,\n");
+P("which requests the kernel show all unaligned fixups, or write a \"0\"\n");
+P("to /proc/sys/tile/unaligned_fixup/enabled, in which case each unaligned\n");
+P("access will become a SIGBUS you can debug. No further warnings will be\n");
+P("shown so as to avoid additional slowdown, but you can track the number\n");
+P("of fixups performed via /proc/sys/tile/unaligned_fixup/count.\n");
+P("Use the tile-addr2line command (see \"info addr2line\") to decode PCs.\n");
+P("\n");
+#undef P
+		}
+	}
+	++unaligned_fixup_count;
+
+	if (bundle & TILEPRO_BUNDLE_Y_ENCODING_MASK) {
+		/* Convert the Y2 instruction to a prefetch. */
+		bundle &= ~(create_SrcBDest_Y2(-1) |
+			    create_Opcode_Y2(-1));
+		bundle |= (create_SrcBDest_Y2(TREG_ZERO) |
+			   create_Opcode_Y2(LW_OPCODE_Y2));
+	/* Replace the load postincr with an addi */
+	} else if (mem_op == MEMOP_LOAD_POSTINCR) {
+		bundle = addi_X1(bundle, addr_reg, addr_reg,
+				 get_Imm8_X1(bundle));
+	/* Replace the store postincr with an addi */
+	} else if (mem_op == MEMOP_STORE_POSTINCR) {
+		bundle = addi_X1(bundle, addr_reg, addr_reg,
+				 get_Dest_Imm8_X1(bundle));
+	} else {
+		/* Convert the X1 instruction to a nop. */
+		bundle &= ~(create_Opcode_X1(-1) |
+			    create_UnShOpcodeExtension_X1(-1) |
+			    create_UnOpcodeExtension_X1(-1));
+		bundle |= (create_Opcode_X1(SHUN_0_OPCODE_X1) |
+			   create_UnShOpcodeExtension_X1(
+				   UN_0_SHUN_0_OPCODE_X1) |
+			   create_UnOpcodeExtension_X1(
+				   NOP_UN_0_SHUN_0_OPCODE_X1));
+	}
+
+	return bundle;
+}
+
+/*
+ * Called after execve() has started the new image.  This allows us
+ * to reset the info state.  Note that the the mmap'ed memory, if there
+ * was any, has already been unmapped by the exec.
+ */
+void single_step_execve(void)
+{
+	struct thread_info *ti = current_thread_info();
+	kfree(ti->step_state);
+	ti->step_state = NULL;
+}
+
+/*
+ * single_step_once() - entry point when single stepping has been triggered.
+ * @regs: The machine register state
+ *
+ *  When we arrive at this routine via a trampoline, the single step
+ *  engine copies the executing bundle to the single step buffer.
+ *  If the instruction is a condition branch, then the target is
+ *  reset to one past the next instruction. If the instruction
+ *  sets the lr, then that is noted. If the instruction is a jump
+ *  or call, then the new target pc is preserved and the current
+ *  bundle instruction set to null.
+ *
+ *  The necessary post-single-step rewriting information is stored in
+ *  single_step_state->  We use data segment values because the
+ *  stack will be rewound when we run the rewritten single-stepped
+ *  instruction.
+ */
+void single_step_once(struct pt_regs *regs)
+{
+	extern tilepro_bundle_bits __single_step_ill_insn;
+	extern tilepro_bundle_bits __single_step_j_insn;
+	extern tilepro_bundle_bits __single_step_addli_insn;
+	extern tilepro_bundle_bits __single_step_auli_insn;
+	struct thread_info *info = (void *)current_thread_info();
+	struct single_step_state *state = info->step_state;
+	int is_single_step = test_ti_thread_flag(info, TIF_SINGLESTEP);
+	tilepro_bundle_bits __user *buffer, *pc;
+	tilepro_bundle_bits bundle;
+	int temp_reg;
+	int target_reg = TREG_LR;
+	int err;
+	enum mem_op mem_op = MEMOP_NONE;
+	int size = 0, sign_ext = 0;  /* happy compiler */
+	int align_ctl;
+
+	align_ctl = unaligned_fixup;
+	switch (task_thread_info(current)->align_ctl) {
+	case PR_UNALIGN_NOPRINT:
+		align_ctl = 1;
+		break;
+	case PR_UNALIGN_SIGBUS:
+		align_ctl = 0;
+		break;
+	}
+
+	asm(
+"    .pushsection .rodata.single_step\n"
+"    .align 8\n"
+"    .globl    __single_step_ill_insn\n"
+"__single_step_ill_insn:\n"
+"    ill\n"
+"    .globl    __single_step_addli_insn\n"
+"__single_step_addli_insn:\n"
+"    { nop; addli r0, zero, 0 }\n"
+"    .globl    __single_step_auli_insn\n"
+"__single_step_auli_insn:\n"
+"    { nop; auli r0, r0, 0 }\n"
+"    .globl    __single_step_j_insn\n"
+"__single_step_j_insn:\n"
+"    j .\n"
+"    .popsection\n"
+	);
+
+	/*
+	 * Enable interrupts here to allow touching userspace and the like.
+	 * The callers expect this: do_trap() already has interrupts
+	 * enabled, and do_work_pending() handles functions that enable
+	 * interrupts internally.
+	 */
+	local_irq_enable();
+
+	if (state == NULL) {
+		/* allocate a page of writable, executable memory */
+		state = kmalloc(sizeof(struct single_step_state), GFP_KERNEL);
+		if (state == NULL) {
+			pr_err("Out of kernel memory trying to single-step\n");
+			return;
+		}
+
+		/* allocate a cache line of writable, executable memory */
+		buffer = (void __user *) vm_mmap(NULL, 0, 64,
+					  PROT_EXEC | PROT_READ | PROT_WRITE,
+					  MAP_PRIVATE | MAP_ANONYMOUS,
+					  0);
+
+		if (IS_ERR((void __force *)buffer)) {
+			kfree(state);
+			pr_err("Out of kernel pages trying to single-step\n");
+			return;
+		}
+
+		state->buffer = buffer;
+		state->is_enabled = 0;
+
+		info->step_state = state;
+
+		/* Validate our stored instruction patterns */
+		BUG_ON(get_Opcode_X1(__single_step_addli_insn) !=
+		       ADDLI_OPCODE_X1);
+		BUG_ON(get_Opcode_X1(__single_step_auli_insn) !=
+		       AULI_OPCODE_X1);
+		BUG_ON(get_SrcA_X1(__single_step_addli_insn) != TREG_ZERO);
+		BUG_ON(get_Dest_X1(__single_step_addli_insn) != 0);
+		BUG_ON(get_JOffLong_X1(__single_step_j_insn) != 0);
+	}
+
+	/*
+	 * If we are returning from a syscall, we still haven't hit the
+	 * "ill" for the swint1 instruction.  So back the PC up to be
+	 * pointing at the swint1, but we'll actually return directly
+	 * back to the "ill" so we come back in via SIGILL as if we
+	 * had "executed" the swint1 without ever being in kernel space.
+	 */
+	if (regs->faultnum == INT_SWINT_1)
+		regs->pc -= 8;
+
+	pc = (tilepro_bundle_bits __user *)(regs->pc);
+	if (get_user(bundle, pc) != 0) {
+		pr_err("Couldn't read instruction at %p trying to step\n", pc);
+		return;
+	}
+
+	/* We'll follow the instruction with 2 ill op bundles */
+	state->orig_pc = (unsigned long)pc;
+	state->next_pc = (unsigned long)(pc + 1);
+	state->branch_next_pc = 0;
+	state->update = 0;
+
+	if (!(bundle & TILEPRO_BUNDLE_Y_ENCODING_MASK)) {
+		/* two wide, check for control flow */
+		int opcode = get_Opcode_X1(bundle);
+
+		switch (opcode) {
+		/* branches */
+		case BRANCH_OPCODE_X1:
+		{
+			s32 offset = signExtend17(get_BrOff_X1(bundle));
+
+			/*
+			 * For branches, we use a rewriting trick to let the
+			 * hardware evaluate whether the branch is taken or
+			 * untaken.  We record the target offset and then
+			 * rewrite the branch instruction to target 1 insn
+			 * ahead if the branch is taken.  We then follow the
+			 * rewritten branch with two bundles, each containing
+			 * an "ill" instruction. The supervisor examines the
+			 * pc after the single step code is executed, and if
+			 * the pc is the first ill instruction, then the
+			 * branch (if any) was not taken.  If the pc is the
+			 * second ill instruction, then the branch was
+			 * taken. The new pc is computed for these cases, and
+			 * inserted into the registers for the thread.  If
+			 * the pc is the start of the single step code, then
+			 * an exception or interrupt was taken before the
+			 * code started processing, and the same "original"
+			 * pc is restored.  This change, different from the
+			 * original implementation, has the advantage of
+			 * executing a single user instruction.
+			 */
+			state->branch_next_pc = (unsigned long)(pc + offset);
+
+			/* rewrite branch offset to go forward one bundle */
+			bundle = set_BrOff_X1(bundle, 2);
+		}
+		break;
+
+		/* jumps */
+		case JALB_OPCODE_X1:
+		case JALF_OPCODE_X1:
+			state->update = 1;
+			state->next_pc =
+				(unsigned long) (pc + get_JOffLong_X1(bundle));
+			break;
+
+		case JB_OPCODE_X1:
+		case JF_OPCODE_X1:
+			state->next_pc =
+				(unsigned long) (pc + get_JOffLong_X1(bundle));
+			bundle = nop_X1(bundle);
+			break;
+
+		case SPECIAL_0_OPCODE_X1:
+			switch (get_RRROpcodeExtension_X1(bundle)) {
+			/* jump-register */
+			case JALRP_SPECIAL_0_OPCODE_X1:
+			case JALR_SPECIAL_0_OPCODE_X1:
+				state->update = 1;
+				state->next_pc =
+					regs->regs[get_SrcA_X1(bundle)];
+				break;
+
+			case JRP_SPECIAL_0_OPCODE_X1:
+			case JR_SPECIAL_0_OPCODE_X1:
+				state->next_pc =
+					regs->regs[get_SrcA_X1(bundle)];
+				bundle = nop_X1(bundle);
+				break;
+
+			case LNK_SPECIAL_0_OPCODE_X1:
+				state->update = 1;
+				target_reg = get_Dest_X1(bundle);
+				break;
+
+			/* stores */
+			case SH_SPECIAL_0_OPCODE_X1:
+				mem_op = MEMOP_STORE;
+				size = 2;
+				break;
+
+			case SW_SPECIAL_0_OPCODE_X1:
+				mem_op = MEMOP_STORE;
+				size = 4;
+				break;
+			}
+			break;
+
+		/* loads and iret */
+		case SHUN_0_OPCODE_X1:
+			if (get_UnShOpcodeExtension_X1(bundle) ==
+			    UN_0_SHUN_0_OPCODE_X1) {
+				switch (get_UnOpcodeExtension_X1(bundle)) {
+				case LH_UN_0_SHUN_0_OPCODE_X1:
+					mem_op = MEMOP_LOAD;
+					size = 2;
+					sign_ext = 1;
+					break;
+
+				case LH_U_UN_0_SHUN_0_OPCODE_X1:
+					mem_op = MEMOP_LOAD;
+					size = 2;
+					sign_ext = 0;
+					break;
+
+				case LW_UN_0_SHUN_0_OPCODE_X1:
+					mem_op = MEMOP_LOAD;
+					size = 4;
+					break;
+
+				case IRET_UN_0_SHUN_0_OPCODE_X1:
+				{
+					unsigned long ex0_0 = __insn_mfspr(
+						SPR_EX_CONTEXT_0_0);
+					unsigned long ex0_1 = __insn_mfspr(
+						SPR_EX_CONTEXT_0_1);
+					/*
+					 * Special-case it if we're iret'ing
+					 * to PL0 again.  Otherwise just let
+					 * it run and it will generate SIGILL.
+					 */
+					if (EX1_PL(ex0_1) == USER_PL) {
+						state->next_pc = ex0_0;
+						regs->ex1 = ex0_1;
+						bundle = nop_X1(bundle);
+					}
+				}
+				}
+			}
+			break;
+
+		/* postincrement operations */
+		case IMM_0_OPCODE_X1:
+			switch (get_ImmOpcodeExtension_X1(bundle)) {
+			case LWADD_IMM_0_OPCODE_X1:
+				mem_op = MEMOP_LOAD_POSTINCR;
+				size = 4;
+				break;
+
+			case LHADD_IMM_0_OPCODE_X1:
+				mem_op = MEMOP_LOAD_POSTINCR;
+				size = 2;
+				sign_ext = 1;
+				break;
+
+			case LHADD_U_IMM_0_OPCODE_X1:
+				mem_op = MEMOP_LOAD_POSTINCR;
+				size = 2;
+				sign_ext = 0;
+				break;
+
+			case SWADD_IMM_0_OPCODE_X1:
+				mem_op = MEMOP_STORE_POSTINCR;
+				size = 4;
+				break;
+
+			case SHADD_IMM_0_OPCODE_X1:
+				mem_op = MEMOP_STORE_POSTINCR;
+				size = 2;
+				break;
+
+			default:
+				break;
+			}
+			break;
+		}
+
+		if (state->update) {
+			/*
+			 * Get an available register.  We start with a
+			 * bitmask with 1's for available registers.
+			 * We truncate to the low 32 registers since
+			 * we are guaranteed to have set bits in the
+			 * low 32 bits, then use ctz to pick the first.
+			 */
+			u32 mask = (u32) ~((1ULL << get_Dest_X0(bundle)) |
+					   (1ULL << get_SrcA_X0(bundle)) |
+					   (1ULL << get_SrcB_X0(bundle)) |
+					   (1ULL << target_reg));
+			temp_reg = __builtin_ctz(mask);
+			state->update_reg = temp_reg;
+			state->update_value = regs->regs[temp_reg];
+			regs->regs[temp_reg] = (unsigned long) (pc+1);
+			regs->flags |= PT_FLAGS_RESTORE_REGS;
+			bundle = move_X1(bundle, target_reg, temp_reg);
+		}
+	} else {
+		int opcode = get_Opcode_Y2(bundle);
+
+		switch (opcode) {
+		/* loads */
+		case LH_OPCODE_Y2:
+			mem_op = MEMOP_LOAD;
+			size = 2;
+			sign_ext = 1;
+			break;
+
+		case LH_U_OPCODE_Y2:
+			mem_op = MEMOP_LOAD;
+			size = 2;
+			sign_ext = 0;
+			break;
+
+		case LW_OPCODE_Y2:
+			mem_op = MEMOP_LOAD;
+			size = 4;
+			break;
+
+		/* stores */
+		case SH_OPCODE_Y2:
+			mem_op = MEMOP_STORE;
+			size = 2;
+			break;
+
+		case SW_OPCODE_Y2:
+			mem_op = MEMOP_STORE;
+			size = 4;
+			break;
+		}
+	}
+
+	/*
+	 * Check if we need to rewrite an unaligned load/store.
+	 * Returning zero is a special value meaning we generated a signal.
+	 */
+	if (mem_op != MEMOP_NONE && align_ctl >= 0) {
+		bundle = rewrite_load_store_unaligned(state, bundle, regs,
+						      mem_op, size, sign_ext);
+		if (bundle == 0)
+			return;
+	}
+
+	/* write the bundle to our execution area */
+	buffer = state->buffer;
+	err = __put_user(bundle, buffer++);
+
+	/*
+	 * If we're really single-stepping, we take an INT_ILL after.
+	 * If we're just handling an unaligned access, we can just
+	 * jump directly back to where we were in user code.
+	 */
+	if (is_single_step) {
+		err |= __put_user(__single_step_ill_insn, buffer++);
+		err |= __put_user(__single_step_ill_insn, buffer++);
+	} else {
+		long delta;
+
+		if (state->update) {
+			/* We have some state to update; do it inline */
+			int ha16;
+			bundle = __single_step_addli_insn;
+			bundle |= create_Dest_X1(state->update_reg);
+			bundle |= create_Imm16_X1(state->update_value);
+			err |= __put_user(bundle, buffer++);
+			bundle = __single_step_auli_insn;
+			bundle |= create_Dest_X1(state->update_reg);
+			bundle |= create_SrcA_X1(state->update_reg);
+			ha16 = (state->update_value + 0x8000) >> 16;
+			bundle |= create_Imm16_X1(ha16);
+			err |= __put_user(bundle, buffer++);
+			state->update = 0;
+		}
+
+		/* End with a jump back to the next instruction */
+		delta = ((regs->pc + TILEPRO_BUNDLE_SIZE_IN_BYTES) -
+			(unsigned long)buffer) >>
+			TILEPRO_LOG2_BUNDLE_ALIGNMENT_IN_BYTES;
+		bundle = __single_step_j_insn;
+		bundle |= create_JOffLong_X1(delta);
+		err |= __put_user(bundle, buffer++);
+	}
+
+	if (err) {
+		pr_err("Fault when writing to single-step buffer\n");
+		return;
+	}
+
+	/*
+	 * Flush the buffer.
+	 * We do a local flush only, since this is a thread-specific buffer.
+	 */
+	__flush_icache_range((unsigned long)state->buffer,
+			     (unsigned long)buffer);
+
+	/* Indicate enabled */
+	state->is_enabled = is_single_step;
+	regs->pc = (unsigned long)state->buffer;
+
+	/* Fault immediately if we are coming back from a syscall. */
+	if (regs->faultnum == INT_SWINT_1)
+		regs->pc += 8;
+}
+
+#else
+
+static DEFINE_PER_CPU(unsigned long, ss_saved_pc);
+
+
+/*
+ * Called directly on the occasion of an interrupt.
+ *
+ * If the process doesn't have single step set, then we use this as an
+ * opportunity to turn single step off.
+ *
+ * It has been mentioned that we could conditionally turn off single stepping
+ * on each entry into the kernel and rely on single_step_once to turn it
+ * on for the processes that matter (as we already do), but this
+ * implementation is somewhat more efficient in that we muck with registers
+ * once on a bum interrupt rather than on every entry into the kernel.
+ *
+ * If SINGLE_STEP_CONTROL_K has CANCELED set, then an interrupt occurred,
+ * so we have to run through this process again before we can say that an
+ * instruction has executed.
+ *
+ * swint will set CANCELED, but it's a legitimate instruction.  Fortunately
+ * it changes the PC.  If it hasn't changed, then we know that the interrupt
+ * wasn't generated by swint and we'll need to run this process again before
+ * we can say an instruction has executed.
+ *
+ * If either CANCELED == 0 or the PC's changed, we send out SIGTRAPs and get
+ * on with our lives.
+ */
+
+void gx_singlestep_handle(struct pt_regs *regs, int fault_num)
+{
+	enum ctx_state prev_state = exception_enter();
+	unsigned long *ss_pc = this_cpu_ptr(&ss_saved_pc);
+	struct thread_info *info = (void *)current_thread_info();
+	int is_single_step = test_ti_thread_flag(info, TIF_SINGLESTEP);
+	unsigned long control = __insn_mfspr(SPR_SINGLE_STEP_CONTROL_K);
+
+	if (is_single_step == 0) {
+		__insn_mtspr(SPR_SINGLE_STEP_EN_K_K, 0);
+
+	} else if ((*ss_pc != regs->pc) ||
+		   (!(control & SPR_SINGLE_STEP_CONTROL_1__CANCELED_MASK))) {
+
+		control |= SPR_SINGLE_STEP_CONTROL_1__CANCELED_MASK;
+		control |= SPR_SINGLE_STEP_CONTROL_1__INHIBIT_MASK;
+		__insn_mtspr(SPR_SINGLE_STEP_CONTROL_K, control);
+		send_sigtrap(current, regs);
+	}
+	exception_exit(prev_state);
+}
+
+
+/*
+ * Called from need_singlestep.  Set up the control registers and the enable
+ * register, then return back.
+ */
+
+void single_step_once(struct pt_regs *regs)
+{
+	unsigned long *ss_pc = this_cpu_ptr(&ss_saved_pc);
+	unsigned long control = __insn_mfspr(SPR_SINGLE_STEP_CONTROL_K);
+
+	*ss_pc = regs->pc;
+	control |= SPR_SINGLE_STEP_CONTROL_1__CANCELED_MASK;
+	control |= SPR_SINGLE_STEP_CONTROL_1__INHIBIT_MASK;
+	__insn_mtspr(SPR_SINGLE_STEP_CONTROL_K, control);
+	__insn_mtspr(SPR_SINGLE_STEP_EN_K_K, 1 << USER_PL);
+}
+
+void single_step_execve(void)
+{
+	/* Nothing */
+}
+
+#endif /* !__tilegx__ */
diff --git a/arch/tile/kernel/smp.c b/arch/tile/kernel/smp.c
new file mode 100644
index 000000000..07e3ff5cc
--- /dev/null
+++ b/arch/tile/kernel/smp.c
@@ -0,0 +1,287 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * TILE SMP support routines.
+ */
+
+#include <linux/smp.h>
+#include <linux/interrupt.h>
+#include <linux/io.h>
+#include <linux/irq.h>
+#include <linux/irq_work.h>
+#include <linux/module.h>
+#include <asm/cacheflush.h>
+#include <asm/homecache.h>
+
+/*
+ * We write to width and height with a single store in head_NN.S,
+ * so make the variable aligned to "long".
+ */
+HV_Topology smp_topology __write_once __aligned(sizeof(long));
+EXPORT_SYMBOL(smp_topology);
+
+#if CHIP_HAS_IPI()
+static unsigned long __iomem *ipi_mappings[NR_CPUS];
+#endif
+
+/* Does messaging work correctly to the local cpu? */
+bool self_interrupt_ok;
+
+/*
+ * Top-level send_IPI*() functions to send messages to other cpus.
+ */
+
+/* Set by smp_send_stop() to avoid recursive panics. */
+static int stopping_cpus;
+
+static void __send_IPI_many(HV_Recipient *recip, int nrecip, int tag)
+{
+	int sent = 0;
+	while (sent < nrecip) {
+		int rc = hv_send_message(recip, nrecip,
+					 (HV_VirtAddr)&tag, sizeof(tag));
+		if (rc < 0) {
+			if (!stopping_cpus)  /* avoid recursive panic */
+				panic("hv_send_message returned %d", rc);
+			break;
+		}
+		WARN_ONCE(rc == 0, "hv_send_message() returned zero\n");
+		sent += rc;
+	}
+}
+
+void send_IPI_single(int cpu, int tag)
+{
+	HV_Recipient recip = {
+		.y = cpu / smp_width,
+		.x = cpu % smp_width,
+		.state = HV_TO_BE_SENT
+	};
+	__send_IPI_many(&recip, 1, tag);
+}
+
+void send_IPI_many(const struct cpumask *mask, int tag)
+{
+	HV_Recipient recip[NR_CPUS];
+	int cpu;
+	int nrecip = 0;
+	int my_cpu = smp_processor_id();
+	for_each_cpu(cpu, mask) {
+		HV_Recipient *r;
+		BUG_ON(cpu == my_cpu);
+		r = &recip[nrecip++];
+		r->y = cpu / smp_width;
+		r->x = cpu % smp_width;
+		r->state = HV_TO_BE_SENT;
+	}
+	__send_IPI_many(recip, nrecip, tag);
+}
+
+void send_IPI_allbutself(int tag)
+{
+	struct cpumask mask;
+	cpumask_copy(&mask, cpu_online_mask);
+	cpumask_clear_cpu(smp_processor_id(), &mask);
+	send_IPI_many(&mask, tag);
+}
+
+/*
+ * Functions related to starting/stopping cpus.
+ */
+
+/* Handler to start the current cpu. */
+static void smp_start_cpu_interrupt(void)
+{
+	get_irq_regs()->pc = start_cpu_function_addr;
+}
+
+/* Handler to stop the current cpu. */
+static void smp_stop_cpu_interrupt(void)
+{
+	arch_local_irq_disable_all();
+	set_cpu_online(smp_processor_id(), 0);
+	for (;;)
+		asm("nap; nop");
+}
+
+/* This function calls the 'stop' function on all other CPUs in the system. */
+void smp_send_stop(void)
+{
+	stopping_cpus = 1;
+	send_IPI_allbutself(MSG_TAG_STOP_CPU);
+}
+
+/* On panic, just wait; we may get an smp_send_stop() later on. */
+void panic_smp_self_stop(void)
+{
+	while (1)
+		asm("nap; nop");
+}
+
+/*
+ * Dispatch code called from hv_message_intr() for HV_MSG_TILE hv messages.
+ */
+void evaluate_message(int tag)
+{
+	switch (tag) {
+	case MSG_TAG_START_CPU: /* Start up a cpu */
+		smp_start_cpu_interrupt();
+		break;
+
+	case MSG_TAG_STOP_CPU: /* Sent to shut down slave CPU's */
+		smp_stop_cpu_interrupt();
+		break;
+
+	case MSG_TAG_CALL_FUNCTION_MANY: /* Call function on cpumask */
+		generic_smp_call_function_interrupt();
+		break;
+
+	case MSG_TAG_CALL_FUNCTION_SINGLE: /* Call function on one other CPU */
+		generic_smp_call_function_single_interrupt();
+		break;
+
+	case MSG_TAG_IRQ_WORK: /* Invoke IRQ work */
+		irq_work_run();
+		break;
+
+	default:
+		panic("Unknown IPI message tag %d", tag);
+		break;
+	}
+}
+
+
+/*
+ * flush_icache_range() code uses smp_call_function().
+ */
+
+struct ipi_flush {
+	unsigned long start;
+	unsigned long end;
+};
+
+static void ipi_flush_icache_range(void *info)
+{
+	struct ipi_flush *flush = (struct ipi_flush *) info;
+	__flush_icache_range(flush->start, flush->end);
+}
+
+void flush_icache_range(unsigned long start, unsigned long end)
+{
+	struct ipi_flush flush = { start, end };
+
+	/* If invoked with irqs disabled, we can not issue IPIs. */
+	if (irqs_disabled())
+		flush_remote(0, HV_FLUSH_EVICT_L1I, NULL, 0, 0, 0,
+			NULL, NULL, 0);
+	else {
+		preempt_disable();
+		on_each_cpu(ipi_flush_icache_range, &flush, 1);
+		preempt_enable();
+	}
+}
+EXPORT_SYMBOL(flush_icache_range);
+
+
+#ifdef CONFIG_IRQ_WORK
+void arch_irq_work_raise(void)
+{
+	if (arch_irq_work_has_interrupt())
+		send_IPI_single(smp_processor_id(), MSG_TAG_IRQ_WORK);
+}
+#endif
+
+
+/* Called when smp_send_reschedule() triggers IRQ_RESCHEDULE. */
+static irqreturn_t handle_reschedule_ipi(int irq, void *token)
+{
+	__this_cpu_inc(irq_stat.irq_resched_count);
+	scheduler_ipi();
+
+	return IRQ_HANDLED;
+}
+
+static struct irqaction resched_action = {
+	.handler = handle_reschedule_ipi,
+	.name = "resched",
+	.dev_id = handle_reschedule_ipi /* unique token */,
+};
+
+void __init ipi_init(void)
+{
+	int cpu = smp_processor_id();
+	HV_Recipient recip = { .y = cpu_y(cpu), .x = cpu_x(cpu),
+			       .state = HV_TO_BE_SENT };
+	int tag = MSG_TAG_CALL_FUNCTION_SINGLE;
+
+	/*
+	 * Test if we can message ourselves for arch_irq_work_raise.
+	 * This functionality is only available in the Tilera hypervisor
+	 * in versions 4.3.4 and following.
+	 */
+	if (hv_send_message(&recip, 1, (HV_VirtAddr)&tag, sizeof(tag)) == 1)
+		self_interrupt_ok = true;
+	else
+		pr_warn("Older hypervisor: disabling fast irq_work_raise\n");
+
+#if CHIP_HAS_IPI()
+	/* Map IPI trigger MMIO addresses. */
+	for_each_possible_cpu(cpu) {
+		HV_Coord tile;
+		HV_PTE pte;
+		unsigned long offset;
+
+		tile.x = cpu_x(cpu);
+		tile.y = cpu_y(cpu);
+		if (hv_get_ipi_pte(tile, KERNEL_PL, &pte) != 0)
+			panic("Failed to initialize IPI for cpu %d\n", cpu);
+
+		offset = PFN_PHYS(pte_pfn(pte));
+		ipi_mappings[cpu] = ioremap_prot(offset, PAGE_SIZE, pte);
+	}
+#endif
+
+	/* Bind handle_reschedule_ipi() to IRQ_RESCHEDULE. */
+	tile_irq_activate(IRQ_RESCHEDULE, TILE_IRQ_PERCPU);
+	BUG_ON(setup_irq(IRQ_RESCHEDULE, &resched_action));
+}
+
+#if CHIP_HAS_IPI()
+
+void smp_send_reschedule(int cpu)
+{
+	WARN_ON(cpu_is_offline(cpu));
+
+	/*
+	 * We just want to do an MMIO store.  The traditional writeq()
+	 * functions aren't really correct here, since they're always
+	 * directed at the PCI shim.  For now, just do a raw store,
+	 * casting away the __iomem attribute.
+	 */
+	((unsigned long __force *)ipi_mappings[cpu])[IRQ_RESCHEDULE] = 0;
+}
+
+#else
+
+void smp_send_reschedule(int cpu)
+{
+	HV_Coord coord;
+
+	WARN_ON(cpu_is_offline(cpu));
+
+	coord.y = cpu_y(cpu);
+	coord.x = cpu_x(cpu);
+	hv_trigger_ipi(coord, IRQ_RESCHEDULE);
+}
+
+#endif /* CHIP_HAS_IPI() */
diff --git a/arch/tile/kernel/smpboot.c b/arch/tile/kernel/smpboot.c
new file mode 100644
index 000000000..20d52a98e
--- /dev/null
+++ b/arch/tile/kernel/smpboot.c
@@ -0,0 +1,268 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/module.h>
+#include <linux/init.h>
+#include <linux/kernel.h>
+#include <linux/mm.h>
+#include <linux/sched.h>
+#include <linux/kernel_stat.h>
+#include <linux/bootmem.h>
+#include <linux/notifier.h>
+#include <linux/cpu.h>
+#include <linux/percpu.h>
+#include <linux/delay.h>
+#include <linux/err.h>
+#include <linux/irq.h>
+#include <asm/mmu_context.h>
+#include <asm/tlbflush.h>
+#include <asm/sections.h>
+
+/* State of each CPU. */
+static DEFINE_PER_CPU(int, cpu_state) = { 0 };
+
+/* The messaging code jumps to this pointer during boot-up */
+unsigned long start_cpu_function_addr;
+
+/* Called very early during startup to mark boot cpu as online */
+void __init smp_prepare_boot_cpu(void)
+{
+	int cpu = smp_processor_id();
+	set_cpu_online(cpu, 1);
+	set_cpu_present(cpu, 1);
+	__this_cpu_write(cpu_state, CPU_ONLINE);
+
+	init_messaging();
+}
+
+static void start_secondary(void);
+
+/*
+ * Called at the top of init() to launch all the other CPUs.
+ * They run free to complete their initialization and then wait
+ * until they get an IPI from the boot cpu to come online.
+ */
+void __init smp_prepare_cpus(unsigned int max_cpus)
+{
+	long rc;
+	int cpu, cpu_count;
+	int boot_cpu = smp_processor_id();
+
+	current_thread_info()->cpu = boot_cpu;
+
+	/*
+	 * Pin this task to the boot CPU while we bring up the others,
+	 * just to make sure we don't uselessly migrate as they come up.
+	 */
+	rc = sched_setaffinity(current->pid, cpumask_of(boot_cpu));
+	if (rc != 0)
+		pr_err("Couldn't set init affinity to boot cpu (%ld)\n", rc);
+
+	/* Print information about disabled and dataplane cpus. */
+	print_disabled_cpus();
+
+	/*
+	 * Tell the messaging subsystem how to respond to the
+	 * startup message.  We use a level of indirection to avoid
+	 * confusing the linker with the fact that the messaging
+	 * subsystem is calling __init code.
+	 */
+	start_cpu_function_addr = (unsigned long) &online_secondary;
+
+	/* Set up thread context for all new processors. */
+	cpu_count = 1;
+	for (cpu = 0; cpu < NR_CPUS; ++cpu)	{
+		struct task_struct *idle;
+
+		if (cpu == boot_cpu)
+			continue;
+
+		if (!cpu_possible(cpu)) {
+			/*
+			 * Make this processor do nothing on boot.
+			 * Note that we don't give the boot_pc function
+			 * a stack, so it has to be assembly code.
+			 */
+			per_cpu(boot_sp, cpu) = 0;
+			per_cpu(boot_pc, cpu) = (unsigned long) smp_nap;
+			continue;
+		}
+
+		/* Create a new idle thread to run start_secondary() */
+		idle = fork_idle(cpu);
+		if (IS_ERR(idle))
+			panic("failed fork for CPU %d", cpu);
+		idle->thread.pc = (unsigned long) start_secondary;
+
+		/* Make this thread the boot thread for this processor */
+		per_cpu(boot_sp, cpu) = task_ksp0(idle);
+		per_cpu(boot_pc, cpu) = idle->thread.pc;
+
+		++cpu_count;
+	}
+	BUG_ON(cpu_count > (max_cpus ? max_cpus : 1));
+
+	/* Fire up the other tiles, if any */
+	init_cpu_present(cpu_possible_mask);
+	if (cpumask_weight(cpu_present_mask) > 1) {
+		mb();  /* make sure all data is visible to new processors */
+		hv_start_all_tiles();
+	}
+}
+
+static __initdata struct cpumask init_affinity;
+
+static __init int reset_init_affinity(void)
+{
+	long rc = sched_setaffinity(current->pid, &init_affinity);
+	if (rc != 0)
+		pr_warn("couldn't reset init affinity (%ld)\n", rc);
+	return 0;
+}
+late_initcall(reset_init_affinity);
+
+static struct cpumask cpu_started;
+
+/*
+ * Activate a secondary processor.  Very minimal; don't add anything
+ * to this path without knowing what you're doing, since SMP booting
+ * is pretty fragile.
+ */
+static void start_secondary(void)
+{
+	int cpuid;
+
+	preempt_disable();
+
+	cpuid = smp_processor_id();
+
+	/* Set our thread pointer appropriately. */
+	set_my_cpu_offset(__per_cpu_offset[cpuid]);
+
+	/*
+	 * In large machines even this will slow us down, since we
+	 * will be contending for for the printk spinlock.
+	 */
+	/* printk(KERN_DEBUG "Initializing CPU#%d\n", cpuid); */
+
+	/* Initialize the current asid for our first page table. */
+	__this_cpu_write(current_asid, min_asid);
+
+	/* Set up this thread as another owner of the init_mm */
+	atomic_inc(&init_mm.mm_count);
+	current->active_mm = &init_mm;
+	if (current->mm)
+		BUG();
+	enter_lazy_tlb(&init_mm, current);
+
+	/* Allow hypervisor messages to be received */
+	init_messaging();
+	local_irq_enable();
+
+	/* Indicate that we're ready to come up. */
+	/* Must not do this before we're ready to receive messages */
+	if (cpumask_test_and_set_cpu(cpuid, &cpu_started)) {
+		pr_warn("CPU#%d already started!\n", cpuid);
+		for (;;)
+			local_irq_enable();
+	}
+
+	smp_nap();
+}
+
+/*
+ * Bring a secondary processor online.
+ */
+void online_secondary(void)
+{
+	/*
+	 * low-memory mappings have been cleared, flush them from
+	 * the local TLBs too.
+	 */
+	local_flush_tlb();
+
+	BUG_ON(in_interrupt());
+
+	/* This must be done before setting cpu_online_mask */
+	wmb();
+
+	notify_cpu_starting(smp_processor_id());
+
+	set_cpu_online(smp_processor_id(), 1);
+	__this_cpu_write(cpu_state, CPU_ONLINE);
+
+	/* Set up tile-specific state for this cpu. */
+	setup_cpu(0);
+
+	/* Set up tile-timer clock-event device on this cpu */
+	setup_tile_timer();
+
+	cpu_startup_entry(CPUHP_ONLINE);
+}
+
+int __cpu_up(unsigned int cpu, struct task_struct *tidle)
+{
+	/* Wait 5s total for all CPUs for them to come online */
+	static int timeout;
+	for (; !cpumask_test_cpu(cpu, &cpu_started); timeout++) {
+		if (timeout >= 50000) {
+			pr_info("skipping unresponsive cpu%d\n", cpu);
+			local_irq_enable();
+			return -EIO;
+		}
+		udelay(100);
+	}
+
+	local_irq_enable();
+	per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
+
+	/* Unleash the CPU! */
+	send_IPI_single(cpu, MSG_TAG_START_CPU);
+	while (!cpumask_test_cpu(cpu, cpu_online_mask))
+		cpu_relax();
+	return 0;
+}
+
+static void panic_start_cpu(void)
+{
+	panic("Received a MSG_START_CPU IPI after boot finished.");
+}
+
+void __init smp_cpus_done(unsigned int max_cpus)
+{
+	int cpu, next, rc;
+
+	/* Reset the response to a (now illegal) MSG_START_CPU IPI. */
+	start_cpu_function_addr = (unsigned long) &panic_start_cpu;
+
+	cpumask_copy(&init_affinity, cpu_online_mask);
+
+	/*
+	 * Pin ourselves to a single cpu in the initial affinity set
+	 * so that kernel mappings for the rootfs are not in the dataplane,
+	 * if set, and to avoid unnecessary migrating during bringup.
+	 * Use the last cpu just in case the whole chip has been
+	 * isolated from the scheduler, to keep init away from likely
+	 * more useful user code.  This also ensures that work scheduled
+	 * via schedule_delayed_work() in the init routines will land
+	 * on this cpu.
+	 */
+	for (cpu = cpumask_first(&init_affinity);
+	     (next = cpumask_next(cpu, &init_affinity)) < nr_cpu_ids;
+	     cpu = next)
+		;
+	rc = sched_setaffinity(current->pid, cpumask_of(cpu));
+	if (rc != 0)
+		pr_err("Couldn't set init affinity to cpu %d (%d)\n", cpu, rc);
+}
diff --git a/arch/tile/kernel/stack.c b/arch/tile/kernel/stack.c
new file mode 100644
index 000000000..c42dce50a
--- /dev/null
+++ b/arch/tile/kernel/stack.c
@@ -0,0 +1,515 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/sched.h>
+#include <linux/kernel.h>
+#include <linux/kprobes.h>
+#include <linux/module.h>
+#include <linux/pfn.h>
+#include <linux/kallsyms.h>
+#include <linux/stacktrace.h>
+#include <linux/uaccess.h>
+#include <linux/mmzone.h>
+#include <linux/dcache.h>
+#include <linux/fs.h>
+#include <linux/string.h>
+#include <asm/backtrace.h>
+#include <asm/page.h>
+#include <asm/ucontext.h>
+#include <asm/switch_to.h>
+#include <asm/sigframe.h>
+#include <asm/stack.h>
+#include <asm/vdso.h>
+#include <arch/abi.h>
+#include <arch/interrupts.h>
+
+#define KBT_ONGOING	0  /* Backtrace still ongoing */
+#define KBT_DONE	1  /* Backtrace cleanly completed */
+#define KBT_RUNNING	2  /* Can't run backtrace on a running task */
+#define KBT_LOOP	3  /* Backtrace entered a loop */
+
+/* Is address on the specified kernel stack? */
+static int in_kernel_stack(struct KBacktraceIterator *kbt, unsigned long sp)
+{
+	ulong kstack_base = (ulong) kbt->task->stack;
+	if (kstack_base == 0)  /* corrupt task pointer; just follow stack... */
+		return sp >= PAGE_OFFSET && sp < (unsigned long)high_memory;
+	return sp >= kstack_base && sp < kstack_base + THREAD_SIZE;
+}
+
+/* Callback for backtracer; basically a glorified memcpy */
+static bool read_memory_func(void *result, unsigned long address,
+			     unsigned int size, void *vkbt)
+{
+	int retval;
+	struct KBacktraceIterator *kbt = (struct KBacktraceIterator *)vkbt;
+
+	if (address == 0)
+		return 0;
+	if (__kernel_text_address(address)) {
+		/* OK to read kernel code. */
+	} else if (address >= PAGE_OFFSET) {
+		/* We only tolerate kernel-space reads of this task's stack */
+		if (!in_kernel_stack(kbt, address))
+			return 0;
+	} else if (!kbt->is_current) {
+		return 0;	/* can't read from other user address spaces */
+	}
+	pagefault_disable();
+	retval = __copy_from_user_inatomic(result,
+					   (void __user __force *)address,
+					   size);
+	pagefault_enable();
+	return (retval == 0);
+}
+
+/* Return a pt_regs pointer for a valid fault handler frame */
+static struct pt_regs *valid_fault_handler(struct KBacktraceIterator* kbt)
+{
+	const char *fault = NULL;  /* happy compiler */
+	char fault_buf[64];
+	unsigned long sp = kbt->it.sp;
+	struct pt_regs *p;
+
+	if (sp % sizeof(long) != 0)
+		return NULL;
+	if (!in_kernel_stack(kbt, sp))
+		return NULL;
+	if (!in_kernel_stack(kbt, sp + C_ABI_SAVE_AREA_SIZE + PTREGS_SIZE-1))
+		return NULL;
+	p = (struct pt_regs *)(sp + C_ABI_SAVE_AREA_SIZE);
+	if (p->faultnum == INT_SWINT_1 || p->faultnum == INT_SWINT_1_SIGRETURN)
+		fault = "syscall";
+	else {
+		if (kbt->verbose) {     /* else we aren't going to use it */
+			snprintf(fault_buf, sizeof(fault_buf),
+				 "interrupt %ld", p->faultnum);
+			fault = fault_buf;
+		}
+	}
+	if (EX1_PL(p->ex1) == KERNEL_PL &&
+	    __kernel_text_address(p->pc) &&
+	    in_kernel_stack(kbt, p->sp) &&
+	    p->sp >= sp) {
+		if (kbt->verbose)
+			pr_err("  <%s while in kernel mode>\n", fault);
+	} else if (user_mode(p) &&
+		   p->sp < PAGE_OFFSET && p->sp != 0) {
+		if (kbt->verbose)
+			pr_err("  <%s while in user mode>\n", fault);
+	} else {
+		if (kbt->verbose)
+			pr_err("  (odd fault: pc %#lx, sp %#lx, ex1 %#lx?)\n",
+			       p->pc, p->sp, p->ex1);
+		return NULL;
+	}
+	if (kbt->profile && ((1ULL << p->faultnum) & QUEUED_INTERRUPTS) != 0)
+		return NULL;
+	return p;
+}
+
+/* Is the pc pointing to a sigreturn trampoline? */
+static int is_sigreturn(unsigned long pc)
+{
+	return current->mm && (pc == VDSO_SYM(&__vdso_rt_sigreturn));
+}
+
+/* Return a pt_regs pointer for a valid signal handler frame */
+static struct pt_regs *valid_sigframe(struct KBacktraceIterator* kbt,
+				      struct rt_sigframe* kframe)
+{
+	BacktraceIterator *b = &kbt->it;
+
+	if (is_sigreturn(b->pc) && b->sp < PAGE_OFFSET &&
+	    b->sp % sizeof(long) == 0) {
+		int retval;
+		pagefault_disable();
+		retval = __copy_from_user_inatomic(
+			kframe, (void __user __force *)b->sp,
+			sizeof(*kframe));
+		pagefault_enable();
+		if (retval != 0 ||
+		    (unsigned int)(kframe->info.si_signo) >= _NSIG)
+			return NULL;
+		if (kbt->verbose) {
+			pr_err("  <received signal %d>\n",
+			       kframe->info.si_signo);
+		}
+		return (struct pt_regs *)&kframe->uc.uc_mcontext;
+	}
+	return NULL;
+}
+
+static int KBacktraceIterator_is_sigreturn(struct KBacktraceIterator *kbt)
+{
+	return is_sigreturn(kbt->it.pc);
+}
+
+static int KBacktraceIterator_restart(struct KBacktraceIterator *kbt)
+{
+	struct pt_regs *p;
+	struct rt_sigframe kframe;
+
+	p = valid_fault_handler(kbt);
+	if (p == NULL)
+		p = valid_sigframe(kbt, &kframe);
+	if (p == NULL)
+		return 0;
+	backtrace_init(&kbt->it, read_memory_func, kbt,
+		       p->pc, p->lr, p->sp, p->regs[52]);
+	kbt->new_context = 1;
+	return 1;
+}
+
+/* Find a frame that isn't a sigreturn, if there is one. */
+static int KBacktraceIterator_next_item_inclusive(
+	struct KBacktraceIterator *kbt)
+{
+	for (;;) {
+		do {
+			if (!KBacktraceIterator_is_sigreturn(kbt))
+				return KBT_ONGOING;
+		} while (backtrace_next(&kbt->it));
+
+		if (!KBacktraceIterator_restart(kbt))
+			return KBT_DONE;
+	}
+}
+
+/*
+ * If the current sp is on a page different than what we recorded
+ * as the top-of-kernel-stack last time we context switched, we have
+ * probably blown the stack, and nothing is going to work out well.
+ * If we can at least get out a warning, that may help the debug,
+ * though we probably won't be able to backtrace into the code that
+ * actually did the recursive damage.
+ */
+static void validate_stack(struct pt_regs *regs)
+{
+	int cpu = raw_smp_processor_id();
+	unsigned long ksp0 = get_current_ksp0();
+	unsigned long ksp0_base = ksp0 & -THREAD_SIZE;
+	unsigned long sp = stack_pointer;
+
+	if (EX1_PL(regs->ex1) == KERNEL_PL && regs->sp >= ksp0) {
+		pr_err("WARNING: cpu %d: kernel stack %#lx..%#lx underrun!\n"
+		       "  sp %#lx (%#lx in caller), caller pc %#lx, lr %#lx\n",
+		       cpu, ksp0_base, ksp0, sp, regs->sp, regs->pc, regs->lr);
+	}
+
+	else if (sp < ksp0_base + sizeof(struct thread_info)) {
+		pr_err("WARNING: cpu %d: kernel stack %#lx..%#lx overrun!\n"
+		       "  sp %#lx (%#lx in caller), caller pc %#lx, lr %#lx\n",
+		       cpu, ksp0_base, ksp0, sp, regs->sp, regs->pc, regs->lr);
+	}
+}
+
+void KBacktraceIterator_init(struct KBacktraceIterator *kbt,
+			     struct task_struct *t, struct pt_regs *regs)
+{
+	unsigned long pc, lr, sp, r52;
+	int is_current;
+
+	/*
+	 * Set up callback information.  We grab the kernel stack base
+	 * so we will allow reads of that address range.
+	 */
+	is_current = (t == NULL || t == current);
+	kbt->is_current = is_current;
+	if (is_current)
+		t = validate_current();
+	kbt->task = t;
+	kbt->verbose = 0;   /* override in caller if desired */
+	kbt->profile = 0;   /* override in caller if desired */
+	kbt->end = KBT_ONGOING;
+	kbt->new_context = 1;
+	if (is_current)
+		validate_stack(regs);
+
+	if (regs == NULL) {
+		if (is_current || t->state == TASK_RUNNING) {
+			/* Can't do this; we need registers */
+			kbt->end = KBT_RUNNING;
+			return;
+		}
+		pc = get_switch_to_pc();
+		lr = t->thread.pc;
+		sp = t->thread.ksp;
+		r52 = 0;
+	} else {
+		pc = regs->pc;
+		lr = regs->lr;
+		sp = regs->sp;
+		r52 = regs->regs[52];
+	}
+
+	backtrace_init(&kbt->it, read_memory_func, kbt, pc, lr, sp, r52);
+	kbt->end = KBacktraceIterator_next_item_inclusive(kbt);
+}
+EXPORT_SYMBOL(KBacktraceIterator_init);
+
+int KBacktraceIterator_end(struct KBacktraceIterator *kbt)
+{
+	return kbt->end != KBT_ONGOING;
+}
+EXPORT_SYMBOL(KBacktraceIterator_end);
+
+void KBacktraceIterator_next(struct KBacktraceIterator *kbt)
+{
+	unsigned long old_pc = kbt->it.pc, old_sp = kbt->it.sp;
+	kbt->new_context = 0;
+	if (!backtrace_next(&kbt->it) && !KBacktraceIterator_restart(kbt)) {
+		kbt->end = KBT_DONE;
+		return;
+	}
+	kbt->end = KBacktraceIterator_next_item_inclusive(kbt);
+	if (old_pc == kbt->it.pc && old_sp == kbt->it.sp) {
+		/* Trapped in a loop; give up. */
+		kbt->end = KBT_LOOP;
+	}
+}
+EXPORT_SYMBOL(KBacktraceIterator_next);
+
+static void describe_addr(struct KBacktraceIterator *kbt,
+			  unsigned long address,
+			  int have_mmap_sem, char *buf, size_t bufsize)
+{
+	struct vm_area_struct *vma;
+	size_t namelen, remaining;
+	unsigned long size, offset, adjust;
+	char *p, *modname;
+	const char *name;
+	int rc;
+
+	/*
+	 * Look one byte back for every caller frame (i.e. those that
+	 * aren't a new context) so we look up symbol data for the
+	 * call itself, not the following instruction, which may be on
+	 * a different line (or in a different function).
+	 */
+	adjust = !kbt->new_context;
+	address -= adjust;
+
+	if (address >= PAGE_OFFSET) {
+		/* Handle kernel symbols. */
+		BUG_ON(bufsize < KSYM_NAME_LEN);
+		name = kallsyms_lookup(address, &size, &offset,
+				       &modname, buf);
+		if (name == NULL) {
+			buf[0] = '\0';
+			return;
+		}
+		namelen = strlen(buf);
+		remaining = (bufsize - 1) - namelen;
+		p = buf + namelen;
+		rc = snprintf(p, remaining, "+%#lx/%#lx ",
+			      offset + adjust, size);
+		if (modname && rc < remaining)
+			snprintf(p + rc, remaining - rc, "[%s] ", modname);
+		buf[bufsize-1] = '\0';
+		return;
+	}
+
+	/* If we don't have the mmap_sem, we can't show any more info. */
+	buf[0] = '\0';
+	if (!have_mmap_sem)
+		return;
+
+	/* Find vma info. */
+	vma = find_vma(kbt->task->mm, address);
+	if (vma == NULL || address < vma->vm_start) {
+		snprintf(buf, bufsize, "[unmapped address] ");
+		return;
+	}
+
+	if (vma->vm_file) {
+		p = d_path(&vma->vm_file->f_path, buf, bufsize);
+		if (IS_ERR(p))
+			p = "?";
+		name = kbasename(p);
+	} else {
+		name = "anon";
+	}
+
+	/* Generate a string description of the vma info. */
+	namelen = strlen(name);
+	remaining = (bufsize - 1) - namelen;
+	memmove(buf, name, namelen);
+	snprintf(buf + namelen, remaining, "[%lx+%lx] ",
+		 vma->vm_start, vma->vm_end - vma->vm_start);
+}
+
+/*
+ * Avoid possible crash recursion during backtrace.  If it happens, it
+ * makes it easy to lose the actual root cause of the failure, so we
+ * put a simple guard on all the backtrace loops.
+ */
+static bool start_backtrace(void)
+{
+	if (current->thread.in_backtrace) {
+		pr_err("Backtrace requested while in backtrace!\n");
+		return false;
+	}
+	current->thread.in_backtrace = true;
+	return true;
+}
+
+static void end_backtrace(void)
+{
+	current->thread.in_backtrace = false;
+}
+
+/*
+ * This method wraps the backtracer's more generic support.
+ * It is only invoked from the architecture-specific code; show_stack()
+ * and dump_stack() (in entry.S) are architecture-independent entry points.
+ */
+void tile_show_stack(struct KBacktraceIterator *kbt, int headers)
+{
+	int i;
+	int have_mmap_sem = 0;
+
+	if (!start_backtrace())
+		return;
+	if (headers) {
+		/*
+		 * Add a blank line since if we are called from panic(),
+		 * then bust_spinlocks() spit out a space in front of us
+		 * and it will mess up our KERN_ERR.
+		 */
+		pr_err("Starting stack dump of tid %d, pid %d (%s) on cpu %d at cycle %lld\n",
+		       kbt->task->pid, kbt->task->tgid, kbt->task->comm,
+		       raw_smp_processor_id(), get_cycles());
+	}
+	kbt->verbose = 1;
+	i = 0;
+	for (; !KBacktraceIterator_end(kbt); KBacktraceIterator_next(kbt)) {
+		char namebuf[KSYM_NAME_LEN+100];
+		unsigned long address = kbt->it.pc;
+
+		/* Try to acquire the mmap_sem as we pass into userspace. */
+		if (address < PAGE_OFFSET && !have_mmap_sem && kbt->task->mm)
+			have_mmap_sem =
+				down_read_trylock(&kbt->task->mm->mmap_sem);
+
+		describe_addr(kbt, address, have_mmap_sem,
+			      namebuf, sizeof(namebuf));
+
+		pr_err("  frame %d: 0x%lx %s(sp 0x%lx)\n",
+		       i++, address, namebuf, (unsigned long)(kbt->it.sp));
+
+		if (i >= 100) {
+			pr_err("Stack dump truncated (%d frames)\n", i);
+			break;
+		}
+	}
+	if (kbt->end == KBT_LOOP)
+		pr_err("Stack dump stopped; next frame identical to this one\n");
+	if (headers)
+		pr_err("Stack dump complete\n");
+	if (have_mmap_sem)
+		up_read(&kbt->task->mm->mmap_sem);
+	end_backtrace();
+}
+EXPORT_SYMBOL(tile_show_stack);
+
+
+/* This is called from show_regs() and _dump_stack() */
+void dump_stack_regs(struct pt_regs *regs)
+{
+	struct KBacktraceIterator kbt;
+	KBacktraceIterator_init(&kbt, NULL, regs);
+	tile_show_stack(&kbt, 1);
+}
+EXPORT_SYMBOL(dump_stack_regs);
+
+static struct pt_regs *regs_to_pt_regs(struct pt_regs *regs,
+				       ulong pc, ulong lr, ulong sp, ulong r52)
+{
+	memset(regs, 0, sizeof(struct pt_regs));
+	regs->pc = pc;
+	regs->lr = lr;
+	regs->sp = sp;
+	regs->regs[52] = r52;
+	return regs;
+}
+
+/* This is called from dump_stack() and just converts to pt_regs */
+void _dump_stack(int dummy, ulong pc, ulong lr, ulong sp, ulong r52)
+{
+	struct pt_regs regs;
+	dump_stack_regs(regs_to_pt_regs(&regs, pc, lr, sp, r52));
+}
+
+/* This is called from KBacktraceIterator_init_current() */
+void _KBacktraceIterator_init_current(struct KBacktraceIterator *kbt, ulong pc,
+				      ulong lr, ulong sp, ulong r52)
+{
+	struct pt_regs regs;
+	KBacktraceIterator_init(kbt, NULL,
+				regs_to_pt_regs(&regs, pc, lr, sp, r52));
+}
+
+/* This is called only from kernel/sched/core.c, with esp == NULL */
+void show_stack(struct task_struct *task, unsigned long *esp)
+{
+	struct KBacktraceIterator kbt;
+	if (task == NULL || task == current)
+		KBacktraceIterator_init_current(&kbt);
+	else
+		KBacktraceIterator_init(&kbt, task, NULL);
+	tile_show_stack(&kbt, 0);
+}
+
+#ifdef CONFIG_STACKTRACE
+
+/* Support generic Linux stack API too */
+
+void save_stack_trace_tsk(struct task_struct *task, struct stack_trace *trace)
+{
+	struct KBacktraceIterator kbt;
+	int skip = trace->skip;
+	int i = 0;
+
+	if (!start_backtrace())
+		goto done;
+	if (task == NULL || task == current)
+		KBacktraceIterator_init_current(&kbt);
+	else
+		KBacktraceIterator_init(&kbt, task, NULL);
+	for (; !KBacktraceIterator_end(&kbt); KBacktraceIterator_next(&kbt)) {
+		if (skip) {
+			--skip;
+			continue;
+		}
+		if (i >= trace->max_entries || kbt.it.pc < PAGE_OFFSET)
+			break;
+		trace->entries[i++] = kbt.it.pc;
+	}
+	end_backtrace();
+done:
+	trace->nr_entries = i;
+}
+EXPORT_SYMBOL(save_stack_trace_tsk);
+
+void save_stack_trace(struct stack_trace *trace)
+{
+	save_stack_trace_tsk(NULL, trace);
+}
+EXPORT_SYMBOL_GPL(save_stack_trace);
+
+#endif
+
+/* In entry.S */
+EXPORT_SYMBOL(KBacktraceIterator_init_current);
diff --git a/arch/tile/kernel/sys.c b/arch/tile/kernel/sys.c
new file mode 100644
index 000000000..38debe706
--- /dev/null
+++ b/arch/tile/kernel/sys.c
@@ -0,0 +1,123 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * This file contains various random system calls that
+ * have a non-standard calling sequence on the Linux/TILE
+ * platform.
+ */
+
+#include <linux/errno.h>
+#include <linux/sched.h>
+#include <linux/mm.h>
+#include <linux/smp.h>
+#include <linux/syscalls.h>
+#include <linux/mman.h>
+#include <linux/file.h>
+#include <linux/mempolicy.h>
+#include <linux/binfmts.h>
+#include <linux/fs.h>
+#include <linux/compat.h>
+#include <linux/uaccess.h>
+#include <linux/signal.h>
+#include <asm/syscalls.h>
+#include <asm/pgtable.h>
+#include <asm/homecache.h>
+#include <asm/cachectl.h>
+#include <arch/chip.h>
+
+SYSCALL_DEFINE3(cacheflush, unsigned long, addr, unsigned long, len,
+		unsigned long, flags)
+{
+	/* DCACHE is not particularly effective if not bound to one cpu. */
+	if (flags & DCACHE)
+		homecache_evict(cpumask_of(raw_smp_processor_id()));
+
+	if (flags & ICACHE)
+		flush_remote(0, HV_FLUSH_EVICT_L1I, mm_cpumask(current->mm),
+			     0, 0, 0, NULL, NULL, 0);
+	return 0;
+}
+
+/*
+ * Syscalls that pass 64-bit values on 32-bit systems normally
+ * pass them as (low,high) word packed into the immediately adjacent
+ * registers.  If the low word naturally falls on an even register,
+ * our ABI makes it work correctly; if not, we adjust it here.
+ * Handling it here means we don't have to fix uclibc AND glibc AND
+ * any other standard libcs we want to support.
+ */
+
+#if !defined(__tilegx__) || defined(CONFIG_COMPAT)
+
+ssize_t sys32_readahead(int fd, u32 offset_lo, u32 offset_hi, u32 count)
+{
+	return sys_readahead(fd, ((loff_t)offset_hi << 32) | offset_lo, count);
+}
+
+int sys32_fadvise64_64(int fd, u32 offset_lo, u32 offset_hi,
+		       u32 len_lo, u32 len_hi, int advice)
+{
+	return sys_fadvise64_64(fd, ((loff_t)offset_hi << 32) | offset_lo,
+				((loff_t)len_hi << 32) | len_lo, advice);
+}
+
+#endif /* 32-bit syscall wrappers */
+
+/* Note: used by the compat code even in 64-bit Linux. */
+SYSCALL_DEFINE6(mmap2, unsigned long, addr, unsigned long, len,
+		unsigned long, prot, unsigned long, flags,
+		unsigned long, fd, unsigned long, off_4k)
+{
+#define PAGE_ADJUST (PAGE_SHIFT - 12)
+	if (off_4k & ((1 << PAGE_ADJUST) - 1))
+		return -EINVAL;
+	return sys_mmap_pgoff(addr, len, prot, flags, fd,
+			      off_4k >> PAGE_ADJUST);
+}
+
+#ifdef __tilegx__
+SYSCALL_DEFINE6(mmap, unsigned long, addr, unsigned long, len,
+		unsigned long, prot, unsigned long, flags,
+		unsigned long, fd, off_t, offset)
+{
+	if (offset & ((1 << PAGE_SHIFT) - 1))
+		return -EINVAL;
+	return sys_mmap_pgoff(addr, len, prot, flags, fd,
+			      offset >> PAGE_SHIFT);
+}
+#endif
+
+
+/* Provide the actual syscall number to call mapping. */
+#undef __SYSCALL
+#define __SYSCALL(nr, call) [nr] = (call),
+
+#ifndef __tilegx__
+/* See comments at the top of the file. */
+#define sys_fadvise64_64 sys32_fadvise64_64
+#define sys_readahead sys32_readahead
+#endif
+
+/* Call the assembly trampolines where necessary. */
+#undef sys_rt_sigreturn
+#define sys_rt_sigreturn _sys_rt_sigreturn
+#define sys_clone _sys_clone
+
+/*
+ * Note that we can't include <linux/unistd.h> here since the header
+ * guard will defeat us; <asm/unistd.h> checks for __SYSCALL as well.
+ */
+void *sys_call_table[__NR_syscalls] = {
+	[0 ... __NR_syscalls-1] = sys_ni_syscall,
+#include <asm/unistd.h>
+};
diff --git a/arch/tile/kernel/sysfs.c b/arch/tile/kernel/sysfs.c
new file mode 100644
index 000000000..a3ed12f8f
--- /dev/null
+++ b/arch/tile/kernel/sysfs.c
@@ -0,0 +1,269 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * /sys entry support.
+ */
+
+#include <linux/device.h>
+#include <linux/cpu.h>
+#include <linux/slab.h>
+#include <linux/smp.h>
+#include <linux/stat.h>
+#include <hv/hypervisor.h>
+
+/* Return a string queried from the hypervisor, truncated to page size. */
+static ssize_t get_hv_confstr(char *page, int query)
+{
+	ssize_t n = hv_confstr(query, (unsigned long)page, PAGE_SIZE - 1);
+	n = n < 0 ? 0 : min(n, (ssize_t)PAGE_SIZE - 1) - 1;
+	if (n)
+		page[n++] = '\n';
+	page[n] = '\0';
+	return n;
+}
+
+static ssize_t chip_width_show(struct device *dev,
+			       struct device_attribute *attr,
+			       char *page)
+{
+	return sprintf(page, "%u\n", smp_width);
+}
+static DEVICE_ATTR(chip_width, 0444, chip_width_show, NULL);
+
+static ssize_t chip_height_show(struct device *dev,
+				struct device_attribute *attr,
+				char *page)
+{
+	return sprintf(page, "%u\n", smp_height);
+}
+static DEVICE_ATTR(chip_height, 0444, chip_height_show, NULL);
+
+static ssize_t chip_serial_show(struct device *dev,
+				struct device_attribute *attr,
+				char *page)
+{
+	return get_hv_confstr(page, HV_CONFSTR_CHIP_SERIAL_NUM);
+}
+static DEVICE_ATTR(chip_serial, 0444, chip_serial_show, NULL);
+
+static ssize_t chip_revision_show(struct device *dev,
+				  struct device_attribute *attr,
+				  char *page)
+{
+	return get_hv_confstr(page, HV_CONFSTR_CHIP_REV);
+}
+static DEVICE_ATTR(chip_revision, 0444, chip_revision_show, NULL);
+
+
+static ssize_t type_show(struct device *dev,
+			    struct device_attribute *attr,
+			    char *page)
+{
+	return sprintf(page, "tilera\n");
+}
+static DEVICE_ATTR(type, 0444, type_show, NULL);
+
+#define HV_CONF_ATTR(name, conf)					\
+	static ssize_t name ## _show(struct device *dev,		\
+				     struct device_attribute *attr, \
+				     char *page)			\
+	{								\
+		return get_hv_confstr(page, conf);			\
+	}								\
+	static DEVICE_ATTR(name, 0444, name ## _show, NULL);
+
+HV_CONF_ATTR(version,		HV_CONFSTR_HV_SW_VER)
+HV_CONF_ATTR(config_version,	HV_CONFSTR_HV_CONFIG_VER)
+
+HV_CONF_ATTR(board_part,	HV_CONFSTR_BOARD_PART_NUM)
+HV_CONF_ATTR(board_serial,	HV_CONFSTR_BOARD_SERIAL_NUM)
+HV_CONF_ATTR(board_revision,	HV_CONFSTR_BOARD_REV)
+HV_CONF_ATTR(board_description,	HV_CONFSTR_BOARD_DESC)
+HV_CONF_ATTR(mezz_part,		HV_CONFSTR_MEZZ_PART_NUM)
+HV_CONF_ATTR(mezz_serial,	HV_CONFSTR_MEZZ_SERIAL_NUM)
+HV_CONF_ATTR(mezz_revision,	HV_CONFSTR_MEZZ_REV)
+HV_CONF_ATTR(mezz_description,	HV_CONFSTR_MEZZ_DESC)
+HV_CONF_ATTR(cpumod_part,	HV_CONFSTR_CPUMOD_PART_NUM)
+HV_CONF_ATTR(cpumod_serial,	HV_CONFSTR_CPUMOD_SERIAL_NUM)
+HV_CONF_ATTR(cpumod_revision,	HV_CONFSTR_CPUMOD_REV)
+HV_CONF_ATTR(cpumod_description,HV_CONFSTR_CPUMOD_DESC)
+HV_CONF_ATTR(switch_control,	HV_CONFSTR_SWITCH_CONTROL)
+
+static struct attribute *board_attrs[] = {
+	&dev_attr_board_part.attr,
+	&dev_attr_board_serial.attr,
+	&dev_attr_board_revision.attr,
+	&dev_attr_board_description.attr,
+	&dev_attr_mezz_part.attr,
+	&dev_attr_mezz_serial.attr,
+	&dev_attr_mezz_revision.attr,
+	&dev_attr_mezz_description.attr,
+	&dev_attr_cpumod_part.attr,
+	&dev_attr_cpumod_serial.attr,
+	&dev_attr_cpumod_revision.attr,
+	&dev_attr_cpumod_description.attr,
+	&dev_attr_switch_control.attr,
+	NULL
+};
+
+static struct attribute_group board_attr_group = {
+	.name   = "board",
+	.attrs  = board_attrs,
+};
+
+
+static struct bin_attribute hvconfig_bin;
+
+static ssize_t
+hvconfig_bin_read(struct file *filp, struct kobject *kobj,
+		  struct bin_attribute *bin_attr,
+		  char *buf, loff_t off, size_t count)
+{
+	static size_t size;
+
+	/* Lazily learn the true size (minus the trailing NUL). */
+	if (size == 0)
+		size = hv_confstr(HV_CONFSTR_HV_CONFIG, 0, 0) - 1;
+
+	/* Check and adjust input parameters. */
+	if (off > size)
+		return -EINVAL;
+	if (count > size - off)
+		count = size - off;
+
+	if (count) {
+		/* Get a copy of the hvc and copy out the relevant portion. */
+		char *hvc;
+
+		size = off + count;
+		hvc = kmalloc(size, GFP_KERNEL);
+		if (hvc == NULL)
+			return -ENOMEM;
+		hv_confstr(HV_CONFSTR_HV_CONFIG, (unsigned long)hvc, size);
+		memcpy(buf, hvc + off, count);
+		kfree(hvc);
+	}
+
+	return count;
+}
+
+static ssize_t hv_stats_show(struct device *dev,
+			     struct device_attribute *attr,
+			     char *page)
+{
+	int cpu = dev->id;
+	long lotar = HV_XY_TO_LOTAR(cpu_x(cpu), cpu_y(cpu));
+
+	ssize_t n = hv_confstr(HV_CONFSTR_HV_STATS,
+			       (unsigned long)page, PAGE_SIZE - 1,
+			       lotar, 0);
+	n = n < 0 ? 0 : min(n, (ssize_t)PAGE_SIZE - 1);
+	page[n] = '\0';
+	return n;
+}
+
+static ssize_t hv_stats_store(struct device *dev,
+			      struct device_attribute *attr,
+			      const char *page,
+			      size_t count)
+{
+	int cpu = dev->id;
+	long lotar = HV_XY_TO_LOTAR(cpu_x(cpu), cpu_y(cpu));
+
+	ssize_t n = hv_confstr(HV_CONFSTR_HV_STATS, 0, 0, lotar, 1);
+	return n < 0 ? n : count;
+}
+
+static DEVICE_ATTR(hv_stats, 0644, hv_stats_show, hv_stats_store);
+
+static int hv_stats_device_add(struct device *dev, struct subsys_interface *sif)
+{
+	int err, cpu = dev->id;
+
+	if (!cpu_online(cpu))
+		return 0;
+
+	err = sysfs_create_file(&dev->kobj, &dev_attr_hv_stats.attr);
+
+	return err;
+}
+
+static int hv_stats_device_remove(struct device *dev,
+				  struct subsys_interface *sif)
+{
+	int cpu = dev->id;
+
+	if (!cpu_online(cpu))
+		return 0;
+
+	sysfs_remove_file(&dev->kobj, &dev_attr_hv_stats.attr);
+	return 0;
+}
+
+
+static struct subsys_interface hv_stats_interface = {
+	.name			= "hv_stats",
+	.subsys			= &cpu_subsys,
+	.add_dev		= hv_stats_device_add,
+	.remove_dev		= hv_stats_device_remove,
+};
+
+static int __init create_sysfs_entries(void)
+{
+	int err = 0;
+
+#define create_cpu_attr(name)						\
+	if (!err)							\
+		err = device_create_file(cpu_subsys.dev_root, &dev_attr_##name);
+	create_cpu_attr(chip_width);
+	create_cpu_attr(chip_height);
+	create_cpu_attr(chip_serial);
+	create_cpu_attr(chip_revision);
+
+#define create_hv_attr(name)						\
+	if (!err)							\
+		err = sysfs_create_file(hypervisor_kobj, &dev_attr_##name.attr);
+	create_hv_attr(type);
+	create_hv_attr(version);
+	create_hv_attr(config_version);
+
+	if (!err)
+		err = sysfs_create_group(hypervisor_kobj, &board_attr_group);
+
+	if (!err) {
+		sysfs_bin_attr_init(&hvconfig_bin);
+		hvconfig_bin.attr.name = "hvconfig";
+		hvconfig_bin.attr.mode = S_IRUGO;
+		hvconfig_bin.read = hvconfig_bin_read;
+		hvconfig_bin.size = PAGE_SIZE;
+		err = sysfs_create_bin_file(hypervisor_kobj, &hvconfig_bin);
+	}
+
+	if (!err) {
+		/*
+		 * Don't bother adding the hv_stats files on each CPU if
+		 * our hypervisor doesn't supply statistics.
+		 */
+		int cpu = raw_smp_processor_id();
+		long lotar = HV_XY_TO_LOTAR(cpu_x(cpu), cpu_y(cpu));
+		char dummy;
+		ssize_t n = hv_confstr(HV_CONFSTR_HV_STATS,
+				       (unsigned long) &dummy, 1,
+				       lotar, 0);
+		if (n >= 0)
+			err = subsys_interface_register(&hv_stats_interface);
+	}
+
+	return err;
+}
+subsys_initcall(create_sysfs_entries);
diff --git a/arch/tile/kernel/tile-desc_32.c b/arch/tile/kernel/tile-desc_32.c
new file mode 100644
index 000000000..dd7bd1d85
--- /dev/null
+++ b/arch/tile/kernel/tile-desc_32.c
@@ -0,0 +1,2605 @@
+/* TILEPro opcode information.
+ *
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ *
+ *
+ *
+ *
+ */
+
+/* This define is BFD_RELOC_##x for real bfd, or -1 for everyone else. */
+#define BFD_RELOC(x) -1
+
+/* Special registers. */
+#define TREG_LR 55
+#define TREG_SN 56
+#define TREG_ZERO 63
+
+#include <linux/stddef.h>
+#include <asm/tile-desc.h>
+
+const struct tilepro_opcode tilepro_opcodes[395] =
+{
+ { "bpt", TILEPRO_OPC_BPT, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "info", TILEPRO_OPC_INFO, 0xf, 1, TREG_ZERO, 1,
+    { { 0 }, { 1 }, { 2 }, { 3 }, { 0, } },
+  },
+  { "infol", TILEPRO_OPC_INFOL, 0x3, 1, TREG_ZERO, 1,
+    { { 4 }, { 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "j", TILEPRO_OPC_J, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 6 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "jal", TILEPRO_OPC_JAL, 0x2, 1, TREG_LR, 1,
+    { { 0, }, { 6 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "move", TILEPRO_OPC_MOVE, 0xf, 2, TREG_ZERO, 1,
+    { { 7, 8 }, { 9, 10 }, { 11, 12 }, { 13, 14 }, { 0, } },
+  },
+  { "move.sn", TILEPRO_OPC_MOVE_SN, 0x3, 2, TREG_SN, 1,
+    { { 7, 8 }, { 9, 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "movei", TILEPRO_OPC_MOVEI, 0xf, 2, TREG_ZERO, 1,
+    { { 7, 0 }, { 9, 1 }, { 11, 2 }, { 13, 3 }, { 0, } },
+  },
+  { "movei.sn", TILEPRO_OPC_MOVEI_SN, 0x3, 2, TREG_SN, 1,
+    { { 7, 0 }, { 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "moveli", TILEPRO_OPC_MOVELI, 0x3, 2, TREG_ZERO, 1,
+    { { 7, 4 }, { 9, 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "moveli.sn", TILEPRO_OPC_MOVELI_SN, 0x3, 2, TREG_SN, 1,
+    { { 7, 4 }, { 9, 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "movelis", TILEPRO_OPC_MOVELIS, 0x3, 2, TREG_SN, 1,
+    { { 7, 4 }, { 9, 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "prefetch", TILEPRO_OPC_PREFETCH, 0x12, 1, TREG_ZERO, 1,
+    { { 0, }, { 10 }, { 0, }, { 0, }, { 15 } },
+  },
+  { "raise", TILEPRO_OPC_RAISE, 0x2, 0, TREG_ZERO, 1,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "add", TILEPRO_OPC_ADD, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "add.sn", TILEPRO_OPC_ADD_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addb", TILEPRO_OPC_ADDB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addb.sn", TILEPRO_OPC_ADDB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addbs_u", TILEPRO_OPC_ADDBS_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addbs_u.sn", TILEPRO_OPC_ADDBS_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addh", TILEPRO_OPC_ADDH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addh.sn", TILEPRO_OPC_ADDH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addhs", TILEPRO_OPC_ADDHS, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addhs.sn", TILEPRO_OPC_ADDHS_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addi", TILEPRO_OPC_ADDI, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 11, 12, 2 }, { 13, 14, 3 }, { 0, } },
+  },
+  { "addi.sn", TILEPRO_OPC_ADDI_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addib", TILEPRO_OPC_ADDIB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addib.sn", TILEPRO_OPC_ADDIB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addih", TILEPRO_OPC_ADDIH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addih.sn", TILEPRO_OPC_ADDIH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addli", TILEPRO_OPC_ADDLI, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 4 }, { 9, 10, 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addli.sn", TILEPRO_OPC_ADDLI_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 4 }, { 9, 10, 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addlis", TILEPRO_OPC_ADDLIS, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 4 }, { 9, 10, 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "adds", TILEPRO_OPC_ADDS, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "adds.sn", TILEPRO_OPC_ADDS_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "adiffb_u", TILEPRO_OPC_ADIFFB_U, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "adiffb_u.sn", TILEPRO_OPC_ADIFFB_U_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "adiffh", TILEPRO_OPC_ADIFFH, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "adiffh.sn", TILEPRO_OPC_ADIFFH_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "and", TILEPRO_OPC_AND, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "and.sn", TILEPRO_OPC_AND_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "andi", TILEPRO_OPC_ANDI, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 11, 12, 2 }, { 13, 14, 3 }, { 0, } },
+  },
+  { "andi.sn", TILEPRO_OPC_ANDI_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "auli", TILEPRO_OPC_AULI, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 4 }, { 9, 10, 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "avgb_u", TILEPRO_OPC_AVGB_U, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "avgb_u.sn", TILEPRO_OPC_AVGB_U_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "avgh", TILEPRO_OPC_AVGH, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "avgh.sn", TILEPRO_OPC_AVGH_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bbns", TILEPRO_OPC_BBNS, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bbns.sn", TILEPRO_OPC_BBNS_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bbnst", TILEPRO_OPC_BBNST, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bbnst.sn", TILEPRO_OPC_BBNST_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bbs", TILEPRO_OPC_BBS, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bbs.sn", TILEPRO_OPC_BBS_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bbst", TILEPRO_OPC_BBST, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bbst.sn", TILEPRO_OPC_BBST_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgez", TILEPRO_OPC_BGEZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgez.sn", TILEPRO_OPC_BGEZ_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgezt", TILEPRO_OPC_BGEZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgezt.sn", TILEPRO_OPC_BGEZT_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgz", TILEPRO_OPC_BGZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgz.sn", TILEPRO_OPC_BGZ_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgzt", TILEPRO_OPC_BGZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgzt.sn", TILEPRO_OPC_BGZT_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bitx", TILEPRO_OPC_BITX, 0x5, 2, TREG_ZERO, 1,
+    { { 7, 8 }, { 0, }, { 11, 12 }, { 0, }, { 0, } },
+  },
+  { "bitx.sn", TILEPRO_OPC_BITX_SN, 0x1, 2, TREG_SN, 1,
+    { { 7, 8 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blez", TILEPRO_OPC_BLEZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blez.sn", TILEPRO_OPC_BLEZ_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blezt", TILEPRO_OPC_BLEZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blezt.sn", TILEPRO_OPC_BLEZT_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blz", TILEPRO_OPC_BLZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blz.sn", TILEPRO_OPC_BLZ_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blzt", TILEPRO_OPC_BLZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blzt.sn", TILEPRO_OPC_BLZT_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bnz", TILEPRO_OPC_BNZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bnz.sn", TILEPRO_OPC_BNZ_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bnzt", TILEPRO_OPC_BNZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bnzt.sn", TILEPRO_OPC_BNZT_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bytex", TILEPRO_OPC_BYTEX, 0x5, 2, TREG_ZERO, 1,
+    { { 7, 8 }, { 0, }, { 11, 12 }, { 0, }, { 0, } },
+  },
+  { "bytex.sn", TILEPRO_OPC_BYTEX_SN, 0x1, 2, TREG_SN, 1,
+    { { 7, 8 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bz", TILEPRO_OPC_BZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bz.sn", TILEPRO_OPC_BZ_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bzt", TILEPRO_OPC_BZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bzt.sn", TILEPRO_OPC_BZT_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 10, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "clz", TILEPRO_OPC_CLZ, 0x5, 2, TREG_ZERO, 1,
+    { { 7, 8 }, { 0, }, { 11, 12 }, { 0, }, { 0, } },
+  },
+  { "clz.sn", TILEPRO_OPC_CLZ_SN, 0x1, 2, TREG_SN, 1,
+    { { 7, 8 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "crc32_32", TILEPRO_OPC_CRC32_32, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "crc32_32.sn", TILEPRO_OPC_CRC32_32_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "crc32_8", TILEPRO_OPC_CRC32_8, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "crc32_8.sn", TILEPRO_OPC_CRC32_8_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ctz", TILEPRO_OPC_CTZ, 0x5, 2, TREG_ZERO, 1,
+    { { 7, 8 }, { 0, }, { 11, 12 }, { 0, }, { 0, } },
+  },
+  { "ctz.sn", TILEPRO_OPC_CTZ_SN, 0x1, 2, TREG_SN, 1,
+    { { 7, 8 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "drain", TILEPRO_OPC_DRAIN, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "dtlbpr", TILEPRO_OPC_DTLBPR, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "dword_align", TILEPRO_OPC_DWORD_ALIGN, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "dword_align.sn", TILEPRO_OPC_DWORD_ALIGN_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "finv", TILEPRO_OPC_FINV, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "flush", TILEPRO_OPC_FLUSH, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fnop", TILEPRO_OPC_FNOP, 0xf, 0, TREG_ZERO, 1,
+    { {  }, {  }, {  }, {  }, { 0, } },
+  },
+  { "icoh", TILEPRO_OPC_ICOH, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ill", TILEPRO_OPC_ILL, 0xa, 0, TREG_ZERO, 1,
+    { { 0, }, {  }, { 0, }, {  }, { 0, } },
+  },
+  { "inthb", TILEPRO_OPC_INTHB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "inthb.sn", TILEPRO_OPC_INTHB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "inthh", TILEPRO_OPC_INTHH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "inthh.sn", TILEPRO_OPC_INTHH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "intlb", TILEPRO_OPC_INTLB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "intlb.sn", TILEPRO_OPC_INTLB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "intlh", TILEPRO_OPC_INTLH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "intlh.sn", TILEPRO_OPC_INTLH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "inv", TILEPRO_OPC_INV, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "iret", TILEPRO_OPC_IRET, 0x2, 0, TREG_ZERO, 1,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "jalb", TILEPRO_OPC_JALB, 0x2, 1, TREG_LR, 1,
+    { { 0, }, { 22 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "jalf", TILEPRO_OPC_JALF, 0x2, 1, TREG_LR, 1,
+    { { 0, }, { 22 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "jalr", TILEPRO_OPC_JALR, 0x2, 1, TREG_LR, 1,
+    { { 0, }, { 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "jalrp", TILEPRO_OPC_JALRP, 0x2, 1, TREG_LR, 1,
+    { { 0, }, { 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "jb", TILEPRO_OPC_JB, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 22 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "jf", TILEPRO_OPC_JF, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 22 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "jr", TILEPRO_OPC_JR, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "jrp", TILEPRO_OPC_JRP, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lb", TILEPRO_OPC_LB, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 23, 15 } },
+  },
+  { "lb.sn", TILEPRO_OPC_LB_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lb_u", TILEPRO_OPC_LB_U, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 23, 15 } },
+  },
+  { "lb_u.sn", TILEPRO_OPC_LB_U_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lbadd", TILEPRO_OPC_LBADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lbadd.sn", TILEPRO_OPC_LBADD_SN, 0x2, 3, TREG_SN, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lbadd_u", TILEPRO_OPC_LBADD_U, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lbadd_u.sn", TILEPRO_OPC_LBADD_U_SN, 0x2, 3, TREG_SN, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lh", TILEPRO_OPC_LH, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 23, 15 } },
+  },
+  { "lh.sn", TILEPRO_OPC_LH_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lh_u", TILEPRO_OPC_LH_U, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 23, 15 } },
+  },
+  { "lh_u.sn", TILEPRO_OPC_LH_U_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lhadd", TILEPRO_OPC_LHADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lhadd.sn", TILEPRO_OPC_LHADD_SN, 0x2, 3, TREG_SN, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lhadd_u", TILEPRO_OPC_LHADD_U, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lhadd_u.sn", TILEPRO_OPC_LHADD_U_SN, 0x2, 3, TREG_SN, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lnk", TILEPRO_OPC_LNK, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lnk.sn", TILEPRO_OPC_LNK_SN, 0x2, 1, TREG_SN, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lw", TILEPRO_OPC_LW, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 23, 15 } },
+  },
+  { "lw.sn", TILEPRO_OPC_LW_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lw_na", TILEPRO_OPC_LW_NA, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lw_na.sn", TILEPRO_OPC_LW_NA_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lwadd", TILEPRO_OPC_LWADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lwadd.sn", TILEPRO_OPC_LWADD_SN, 0x2, 3, TREG_SN, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lwadd_na", TILEPRO_OPC_LWADD_NA, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lwadd_na.sn", TILEPRO_OPC_LWADD_NA_SN, 0x2, 3, TREG_SN, 1,
+    { { 0, }, { 9, 24, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "maxb_u", TILEPRO_OPC_MAXB_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "maxb_u.sn", TILEPRO_OPC_MAXB_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "maxh", TILEPRO_OPC_MAXH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "maxh.sn", TILEPRO_OPC_MAXH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "maxib_u", TILEPRO_OPC_MAXIB_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "maxib_u.sn", TILEPRO_OPC_MAXIB_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "maxih", TILEPRO_OPC_MAXIH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "maxih.sn", TILEPRO_OPC_MAXIH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mf", TILEPRO_OPC_MF, 0x2, 0, TREG_ZERO, 1,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mfspr", TILEPRO_OPC_MFSPR, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 25 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "minb_u", TILEPRO_OPC_MINB_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "minb_u.sn", TILEPRO_OPC_MINB_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "minh", TILEPRO_OPC_MINH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "minh.sn", TILEPRO_OPC_MINH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "minib_u", TILEPRO_OPC_MINIB_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "minib_u.sn", TILEPRO_OPC_MINIB_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "minih", TILEPRO_OPC_MINIH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "minih.sn", TILEPRO_OPC_MINIH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mm", TILEPRO_OPC_MM, 0x3, 5, TREG_ZERO, 1,
+    { { 7, 8, 16, 26, 27 }, { 9, 10, 17, 28, 29 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mnz", TILEPRO_OPC_MNZ, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "mnz.sn", TILEPRO_OPC_MNZ_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mnzb", TILEPRO_OPC_MNZB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mnzb.sn", TILEPRO_OPC_MNZB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mnzh", TILEPRO_OPC_MNZH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mnzh.sn", TILEPRO_OPC_MNZH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mtspr", TILEPRO_OPC_MTSPR, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 30, 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhh_ss", TILEPRO_OPC_MULHH_SS, 0x5, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 11, 12, 18 }, { 0, }, { 0, } },
+  },
+  { "mulhh_ss.sn", TILEPRO_OPC_MULHH_SS_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhh_su", TILEPRO_OPC_MULHH_SU, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhh_su.sn", TILEPRO_OPC_MULHH_SU_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhh_uu", TILEPRO_OPC_MULHH_UU, 0x5, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 11, 12, 18 }, { 0, }, { 0, } },
+  },
+  { "mulhh_uu.sn", TILEPRO_OPC_MULHH_UU_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhha_ss", TILEPRO_OPC_MULHHA_SS, 0x5, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 31, 12, 18 }, { 0, }, { 0, } },
+  },
+  { "mulhha_ss.sn", TILEPRO_OPC_MULHHA_SS_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhha_su", TILEPRO_OPC_MULHHA_SU, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhha_su.sn", TILEPRO_OPC_MULHHA_SU_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhha_uu", TILEPRO_OPC_MULHHA_UU, 0x5, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 31, 12, 18 }, { 0, }, { 0, } },
+  },
+  { "mulhha_uu.sn", TILEPRO_OPC_MULHHA_UU_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhhsa_uu", TILEPRO_OPC_MULHHSA_UU, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhhsa_uu.sn", TILEPRO_OPC_MULHHSA_UU_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhl_ss", TILEPRO_OPC_MULHL_SS, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhl_ss.sn", TILEPRO_OPC_MULHL_SS_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhl_su", TILEPRO_OPC_MULHL_SU, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhl_su.sn", TILEPRO_OPC_MULHL_SU_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhl_us", TILEPRO_OPC_MULHL_US, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhl_us.sn", TILEPRO_OPC_MULHL_US_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhl_uu", TILEPRO_OPC_MULHL_UU, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhl_uu.sn", TILEPRO_OPC_MULHL_UU_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhla_ss", TILEPRO_OPC_MULHLA_SS, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhla_ss.sn", TILEPRO_OPC_MULHLA_SS_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhla_su", TILEPRO_OPC_MULHLA_SU, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhla_su.sn", TILEPRO_OPC_MULHLA_SU_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhla_us", TILEPRO_OPC_MULHLA_US, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhla_us.sn", TILEPRO_OPC_MULHLA_US_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhla_uu", TILEPRO_OPC_MULHLA_UU, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhla_uu.sn", TILEPRO_OPC_MULHLA_UU_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulhlsa_uu", TILEPRO_OPC_MULHLSA_UU, 0x5, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 31, 12, 18 }, { 0, }, { 0, } },
+  },
+  { "mulhlsa_uu.sn", TILEPRO_OPC_MULHLSA_UU_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulll_ss", TILEPRO_OPC_MULLL_SS, 0x5, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 11, 12, 18 }, { 0, }, { 0, } },
+  },
+  { "mulll_ss.sn", TILEPRO_OPC_MULLL_SS_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulll_su", TILEPRO_OPC_MULLL_SU, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulll_su.sn", TILEPRO_OPC_MULLL_SU_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulll_uu", TILEPRO_OPC_MULLL_UU, 0x5, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 11, 12, 18 }, { 0, }, { 0, } },
+  },
+  { "mulll_uu.sn", TILEPRO_OPC_MULLL_UU_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mullla_ss", TILEPRO_OPC_MULLLA_SS, 0x5, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 31, 12, 18 }, { 0, }, { 0, } },
+  },
+  { "mullla_ss.sn", TILEPRO_OPC_MULLLA_SS_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mullla_su", TILEPRO_OPC_MULLLA_SU, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mullla_su.sn", TILEPRO_OPC_MULLLA_SU_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mullla_uu", TILEPRO_OPC_MULLLA_UU, 0x5, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 31, 12, 18 }, { 0, }, { 0, } },
+  },
+  { "mullla_uu.sn", TILEPRO_OPC_MULLLA_UU_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulllsa_uu", TILEPRO_OPC_MULLLSA_UU, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mulllsa_uu.sn", TILEPRO_OPC_MULLLSA_UU_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mvnz", TILEPRO_OPC_MVNZ, 0x5, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 31, 12, 18 }, { 0, }, { 0, } },
+  },
+  { "mvnz.sn", TILEPRO_OPC_MVNZ_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mvz", TILEPRO_OPC_MVZ, 0x5, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 31, 12, 18 }, { 0, }, { 0, } },
+  },
+  { "mvz.sn", TILEPRO_OPC_MVZ_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mz", TILEPRO_OPC_MZ, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "mz.sn", TILEPRO_OPC_MZ_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mzb", TILEPRO_OPC_MZB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mzb.sn", TILEPRO_OPC_MZB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mzh", TILEPRO_OPC_MZH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mzh.sn", TILEPRO_OPC_MZH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "nap", TILEPRO_OPC_NAP, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "nop", TILEPRO_OPC_NOP, 0xf, 0, TREG_ZERO, 1,
+    { {  }, {  }, {  }, {  }, { 0, } },
+  },
+  { "nor", TILEPRO_OPC_NOR, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "nor.sn", TILEPRO_OPC_NOR_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "or", TILEPRO_OPC_OR, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "or.sn", TILEPRO_OPC_OR_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ori", TILEPRO_OPC_ORI, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 11, 12, 2 }, { 13, 14, 3 }, { 0, } },
+  },
+  { "ori.sn", TILEPRO_OPC_ORI_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "packbs_u", TILEPRO_OPC_PACKBS_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "packbs_u.sn", TILEPRO_OPC_PACKBS_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "packhb", TILEPRO_OPC_PACKHB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "packhb.sn", TILEPRO_OPC_PACKHB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "packhs", TILEPRO_OPC_PACKHS, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "packhs.sn", TILEPRO_OPC_PACKHS_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "packlb", TILEPRO_OPC_PACKLB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "packlb.sn", TILEPRO_OPC_PACKLB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "pcnt", TILEPRO_OPC_PCNT, 0x5, 2, TREG_ZERO, 1,
+    { { 7, 8 }, { 0, }, { 11, 12 }, { 0, }, { 0, } },
+  },
+  { "pcnt.sn", TILEPRO_OPC_PCNT_SN, 0x1, 2, TREG_SN, 1,
+    { { 7, 8 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "rl", TILEPRO_OPC_RL, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "rl.sn", TILEPRO_OPC_RL_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "rli", TILEPRO_OPC_RLI, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 11, 12, 34 }, { 13, 14, 35 }, { 0, } },
+  },
+  { "rli.sn", TILEPRO_OPC_RLI_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "s1a", TILEPRO_OPC_S1A, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "s1a.sn", TILEPRO_OPC_S1A_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "s2a", TILEPRO_OPC_S2A, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "s2a.sn", TILEPRO_OPC_S2A_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "s3a", TILEPRO_OPC_S3A, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "s3a.sn", TILEPRO_OPC_S3A_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadab_u", TILEPRO_OPC_SADAB_U, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadab_u.sn", TILEPRO_OPC_SADAB_U_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadah", TILEPRO_OPC_SADAH, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadah.sn", TILEPRO_OPC_SADAH_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadah_u", TILEPRO_OPC_SADAH_U, 0x1, 3, TREG_ZERO, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadah_u.sn", TILEPRO_OPC_SADAH_U_SN, 0x1, 3, TREG_SN, 1,
+    { { 21, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadb_u", TILEPRO_OPC_SADB_U, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadb_u.sn", TILEPRO_OPC_SADB_U_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadh", TILEPRO_OPC_SADH, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadh.sn", TILEPRO_OPC_SADH_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadh_u", TILEPRO_OPC_SADH_U, 0x1, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sadh_u.sn", TILEPRO_OPC_SADH_U_SN, 0x1, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sb", TILEPRO_OPC_SB, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 17 }, { 0, }, { 0, }, { 15, 36 } },
+  },
+  { "sbadd", TILEPRO_OPC_SBADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 24, 17, 37 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "seq", TILEPRO_OPC_SEQ, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "seq.sn", TILEPRO_OPC_SEQ_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "seqb", TILEPRO_OPC_SEQB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "seqb.sn", TILEPRO_OPC_SEQB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "seqh", TILEPRO_OPC_SEQH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "seqh.sn", TILEPRO_OPC_SEQH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "seqi", TILEPRO_OPC_SEQI, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 11, 12, 2 }, { 13, 14, 3 }, { 0, } },
+  },
+  { "seqi.sn", TILEPRO_OPC_SEQI_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "seqib", TILEPRO_OPC_SEQIB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "seqib.sn", TILEPRO_OPC_SEQIB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "seqih", TILEPRO_OPC_SEQIH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "seqih.sn", TILEPRO_OPC_SEQIH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sh", TILEPRO_OPC_SH, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 17 }, { 0, }, { 0, }, { 15, 36 } },
+  },
+  { "shadd", TILEPRO_OPC_SHADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 24, 17, 37 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shl", TILEPRO_OPC_SHL, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "shl.sn", TILEPRO_OPC_SHL_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shlb", TILEPRO_OPC_SHLB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shlb.sn", TILEPRO_OPC_SHLB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shlh", TILEPRO_OPC_SHLH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shlh.sn", TILEPRO_OPC_SHLH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shli", TILEPRO_OPC_SHLI, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 11, 12, 34 }, { 13, 14, 35 }, { 0, } },
+  },
+  { "shli.sn", TILEPRO_OPC_SHLI_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shlib", TILEPRO_OPC_SHLIB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shlib.sn", TILEPRO_OPC_SHLIB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shlih", TILEPRO_OPC_SHLIH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shlih.sn", TILEPRO_OPC_SHLIH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shr", TILEPRO_OPC_SHR, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "shr.sn", TILEPRO_OPC_SHR_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shrb", TILEPRO_OPC_SHRB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shrb.sn", TILEPRO_OPC_SHRB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shrh", TILEPRO_OPC_SHRH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shrh.sn", TILEPRO_OPC_SHRH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shri", TILEPRO_OPC_SHRI, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 11, 12, 34 }, { 13, 14, 35 }, { 0, } },
+  },
+  { "shri.sn", TILEPRO_OPC_SHRI_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shrib", TILEPRO_OPC_SHRIB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shrib.sn", TILEPRO_OPC_SHRIB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shrih", TILEPRO_OPC_SHRIH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shrih.sn", TILEPRO_OPC_SHRIH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slt", TILEPRO_OPC_SLT, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "slt.sn", TILEPRO_OPC_SLT_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slt_u", TILEPRO_OPC_SLT_U, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "slt_u.sn", TILEPRO_OPC_SLT_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltb", TILEPRO_OPC_SLTB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltb.sn", TILEPRO_OPC_SLTB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltb_u", TILEPRO_OPC_SLTB_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltb_u.sn", TILEPRO_OPC_SLTB_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slte", TILEPRO_OPC_SLTE, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "slte.sn", TILEPRO_OPC_SLTE_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slte_u", TILEPRO_OPC_SLTE_U, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "slte_u.sn", TILEPRO_OPC_SLTE_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slteb", TILEPRO_OPC_SLTEB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slteb.sn", TILEPRO_OPC_SLTEB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slteb_u", TILEPRO_OPC_SLTEB_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slteb_u.sn", TILEPRO_OPC_SLTEB_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slteh", TILEPRO_OPC_SLTEH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slteh.sn", TILEPRO_OPC_SLTEH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slteh_u", TILEPRO_OPC_SLTEH_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slteh_u.sn", TILEPRO_OPC_SLTEH_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slth", TILEPRO_OPC_SLTH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slth.sn", TILEPRO_OPC_SLTH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slth_u", TILEPRO_OPC_SLTH_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slth_u.sn", TILEPRO_OPC_SLTH_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slti", TILEPRO_OPC_SLTI, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 11, 12, 2 }, { 13, 14, 3 }, { 0, } },
+  },
+  { "slti.sn", TILEPRO_OPC_SLTI_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "slti_u", TILEPRO_OPC_SLTI_U, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 11, 12, 2 }, { 13, 14, 3 }, { 0, } },
+  },
+  { "slti_u.sn", TILEPRO_OPC_SLTI_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltib", TILEPRO_OPC_SLTIB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltib.sn", TILEPRO_OPC_SLTIB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltib_u", TILEPRO_OPC_SLTIB_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltib_u.sn", TILEPRO_OPC_SLTIB_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltih", TILEPRO_OPC_SLTIH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltih.sn", TILEPRO_OPC_SLTIH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltih_u", TILEPRO_OPC_SLTIH_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sltih_u.sn", TILEPRO_OPC_SLTIH_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sne", TILEPRO_OPC_SNE, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "sne.sn", TILEPRO_OPC_SNE_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sneb", TILEPRO_OPC_SNEB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sneb.sn", TILEPRO_OPC_SNEB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sneh", TILEPRO_OPC_SNEH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sneh.sn", TILEPRO_OPC_SNEH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sra", TILEPRO_OPC_SRA, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "sra.sn", TILEPRO_OPC_SRA_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "srab", TILEPRO_OPC_SRAB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "srab.sn", TILEPRO_OPC_SRAB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "srah", TILEPRO_OPC_SRAH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "srah.sn", TILEPRO_OPC_SRAH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "srai", TILEPRO_OPC_SRAI, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 11, 12, 34 }, { 13, 14, 35 }, { 0, } },
+  },
+  { "srai.sn", TILEPRO_OPC_SRAI_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sraib", TILEPRO_OPC_SRAIB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sraib.sn", TILEPRO_OPC_SRAIB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sraih", TILEPRO_OPC_SRAIH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sraih.sn", TILEPRO_OPC_SRAIH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 32 }, { 9, 10, 33 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sub", TILEPRO_OPC_SUB, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "sub.sn", TILEPRO_OPC_SUB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "subb", TILEPRO_OPC_SUBB, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "subb.sn", TILEPRO_OPC_SUBB_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "subbs_u", TILEPRO_OPC_SUBBS_U, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "subbs_u.sn", TILEPRO_OPC_SUBBS_U_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "subh", TILEPRO_OPC_SUBH, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "subh.sn", TILEPRO_OPC_SUBH_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "subhs", TILEPRO_OPC_SUBHS, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "subhs.sn", TILEPRO_OPC_SUBHS_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "subs", TILEPRO_OPC_SUBS, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "subs.sn", TILEPRO_OPC_SUBS_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sw", TILEPRO_OPC_SW, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 10, 17 }, { 0, }, { 0, }, { 15, 36 } },
+  },
+  { "swadd", TILEPRO_OPC_SWADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 24, 17, 37 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "swint0", TILEPRO_OPC_SWINT0, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "swint1", TILEPRO_OPC_SWINT1, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "swint2", TILEPRO_OPC_SWINT2, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "swint3", TILEPRO_OPC_SWINT3, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "tblidxb0", TILEPRO_OPC_TBLIDXB0, 0x5, 2, TREG_ZERO, 1,
+    { { 21, 8 }, { 0, }, { 31, 12 }, { 0, }, { 0, } },
+  },
+  { "tblidxb0.sn", TILEPRO_OPC_TBLIDXB0_SN, 0x1, 2, TREG_SN, 1,
+    { { 21, 8 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "tblidxb1", TILEPRO_OPC_TBLIDXB1, 0x5, 2, TREG_ZERO, 1,
+    { { 21, 8 }, { 0, }, { 31, 12 }, { 0, }, { 0, } },
+  },
+  { "tblidxb1.sn", TILEPRO_OPC_TBLIDXB1_SN, 0x1, 2, TREG_SN, 1,
+    { { 21, 8 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "tblidxb2", TILEPRO_OPC_TBLIDXB2, 0x5, 2, TREG_ZERO, 1,
+    { { 21, 8 }, { 0, }, { 31, 12 }, { 0, }, { 0, } },
+  },
+  { "tblidxb2.sn", TILEPRO_OPC_TBLIDXB2_SN, 0x1, 2, TREG_SN, 1,
+    { { 21, 8 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "tblidxb3", TILEPRO_OPC_TBLIDXB3, 0x5, 2, TREG_ZERO, 1,
+    { { 21, 8 }, { 0, }, { 31, 12 }, { 0, }, { 0, } },
+  },
+  { "tblidxb3.sn", TILEPRO_OPC_TBLIDXB3_SN, 0x1, 2, TREG_SN, 1,
+    { { 21, 8 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "tns", TILEPRO_OPC_TNS, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "tns.sn", TILEPRO_OPC_TNS_SN, 0x2, 2, TREG_SN, 1,
+    { { 0, }, { 9, 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "wh64", TILEPRO_OPC_WH64, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 10 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "xor", TILEPRO_OPC_XOR, 0xf, 3, TREG_ZERO, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 11, 12, 18 }, { 13, 14, 19 }, { 0, } },
+  },
+  { "xor.sn", TILEPRO_OPC_XOR_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 16 }, { 9, 10, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "xori", TILEPRO_OPC_XORI, 0x3, 3, TREG_ZERO, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "xori.sn", TILEPRO_OPC_XORI_SN, 0x3, 3, TREG_SN, 1,
+    { { 7, 8, 0 }, { 9, 10, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { NULL, TILEPRO_OPC_NONE, 0, 0, TREG_ZERO, 0, { { 0, } },
+  }
+};
+#define BITFIELD(start, size) ((start) | (((1 << (size)) - 1) << 6))
+#define CHILD(array_index) (TILEPRO_OPC_NONE + (array_index))
+
+static const unsigned short decode_X0_fsm[1153] =
+{
+  BITFIELD(22, 9) /* index 0 */,
+  CHILD(513), CHILD(530), CHILD(547), CHILD(564), CHILD(596), CHILD(613),
+  CHILD(630), TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, CHILD(663), CHILD(680), CHILD(697),
+  CHILD(714), CHILD(746), CHILD(763), CHILD(780), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813),
+  CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813),
+  CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813),
+  CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813),
+  CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813),
+  CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813),
+  CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813),
+  CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813),
+  CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813),
+  CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(813),
+  CHILD(813), CHILD(813), CHILD(813), CHILD(813), CHILD(828), CHILD(828),
+  CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828),
+  CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828),
+  CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828),
+  CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828),
+  CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828),
+  CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828),
+  CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828),
+  CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828),
+  CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828),
+  CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828), CHILD(828),
+  CHILD(828), CHILD(828), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(873), CHILD(878), CHILD(883), CHILD(903), CHILD(908),
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, CHILD(913),
+  CHILD(918), CHILD(923), CHILD(943), CHILD(948), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, CHILD(953), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, CHILD(988), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, CHILD(993), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, CHILD(1076), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(18, 4) /* index 513 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_ADDB, TILEPRO_OPC_ADDH, TILEPRO_OPC_ADD,
+  TILEPRO_OPC_ADIFFB_U, TILEPRO_OPC_ADIFFH, TILEPRO_OPC_AND,
+  TILEPRO_OPC_AVGB_U, TILEPRO_OPC_AVGH, TILEPRO_OPC_CRC32_32,
+  TILEPRO_OPC_CRC32_8, TILEPRO_OPC_INTHB, TILEPRO_OPC_INTHH,
+  TILEPRO_OPC_INTLB, TILEPRO_OPC_INTLH, TILEPRO_OPC_MAXB_U,
+  BITFIELD(18, 4) /* index 530 */,
+  TILEPRO_OPC_MAXH, TILEPRO_OPC_MINB_U, TILEPRO_OPC_MINH, TILEPRO_OPC_MNZB,
+  TILEPRO_OPC_MNZH, TILEPRO_OPC_MNZ, TILEPRO_OPC_MULHHA_SS,
+  TILEPRO_OPC_MULHHA_SU, TILEPRO_OPC_MULHHA_UU, TILEPRO_OPC_MULHHSA_UU,
+  TILEPRO_OPC_MULHH_SS, TILEPRO_OPC_MULHH_SU, TILEPRO_OPC_MULHH_UU,
+  TILEPRO_OPC_MULHLA_SS, TILEPRO_OPC_MULHLA_SU, TILEPRO_OPC_MULHLA_US,
+  BITFIELD(18, 4) /* index 547 */,
+  TILEPRO_OPC_MULHLA_UU, TILEPRO_OPC_MULHLSA_UU, TILEPRO_OPC_MULHL_SS,
+  TILEPRO_OPC_MULHL_SU, TILEPRO_OPC_MULHL_US, TILEPRO_OPC_MULHL_UU,
+  TILEPRO_OPC_MULLLA_SS, TILEPRO_OPC_MULLLA_SU, TILEPRO_OPC_MULLLA_UU,
+  TILEPRO_OPC_MULLLSA_UU, TILEPRO_OPC_MULLL_SS, TILEPRO_OPC_MULLL_SU,
+  TILEPRO_OPC_MULLL_UU, TILEPRO_OPC_MVNZ, TILEPRO_OPC_MVZ, TILEPRO_OPC_MZB,
+  BITFIELD(18, 4) /* index 564 */,
+  TILEPRO_OPC_MZH, TILEPRO_OPC_MZ, TILEPRO_OPC_NOR, CHILD(581),
+  TILEPRO_OPC_PACKHB, TILEPRO_OPC_PACKLB, TILEPRO_OPC_RL, TILEPRO_OPC_S1A,
+  TILEPRO_OPC_S2A, TILEPRO_OPC_S3A, TILEPRO_OPC_SADAB_U, TILEPRO_OPC_SADAH,
+  TILEPRO_OPC_SADAH_U, TILEPRO_OPC_SADB_U, TILEPRO_OPC_SADH,
+  TILEPRO_OPC_SADH_U,
+  BITFIELD(12, 2) /* index 581 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, CHILD(586),
+  BITFIELD(14, 2) /* index 586 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, CHILD(591),
+  BITFIELD(16, 2) /* index 591 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_MOVE,
+  BITFIELD(18, 4) /* index 596 */,
+  TILEPRO_OPC_SEQB, TILEPRO_OPC_SEQH, TILEPRO_OPC_SEQ, TILEPRO_OPC_SHLB,
+  TILEPRO_OPC_SHLH, TILEPRO_OPC_SHL, TILEPRO_OPC_SHRB, TILEPRO_OPC_SHRH,
+  TILEPRO_OPC_SHR, TILEPRO_OPC_SLTB, TILEPRO_OPC_SLTB_U, TILEPRO_OPC_SLTEB,
+  TILEPRO_OPC_SLTEB_U, TILEPRO_OPC_SLTEH, TILEPRO_OPC_SLTEH_U,
+  TILEPRO_OPC_SLTE,
+  BITFIELD(18, 4) /* index 613 */,
+  TILEPRO_OPC_SLTE_U, TILEPRO_OPC_SLTH, TILEPRO_OPC_SLTH_U, TILEPRO_OPC_SLT,
+  TILEPRO_OPC_SLT_U, TILEPRO_OPC_SNEB, TILEPRO_OPC_SNEH, TILEPRO_OPC_SNE,
+  TILEPRO_OPC_SRAB, TILEPRO_OPC_SRAH, TILEPRO_OPC_SRA, TILEPRO_OPC_SUBB,
+  TILEPRO_OPC_SUBH, TILEPRO_OPC_SUB, TILEPRO_OPC_XOR, TILEPRO_OPC_DWORD_ALIGN,
+  BITFIELD(18, 3) /* index 630 */,
+  CHILD(639), CHILD(642), CHILD(645), CHILD(648), CHILD(651), CHILD(654),
+  CHILD(657), CHILD(660),
+  BITFIELD(21, 1) /* index 639 */,
+  TILEPRO_OPC_ADDS, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 642 */,
+  TILEPRO_OPC_SUBS, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 645 */,
+  TILEPRO_OPC_ADDBS_U, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 648 */,
+  TILEPRO_OPC_ADDHS, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 651 */,
+  TILEPRO_OPC_SUBBS_U, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 654 */,
+  TILEPRO_OPC_SUBHS, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 657 */,
+  TILEPRO_OPC_PACKHS, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 660 */,
+  TILEPRO_OPC_PACKBS_U, TILEPRO_OPC_NONE,
+  BITFIELD(18, 4) /* index 663 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_ADDB_SN, TILEPRO_OPC_ADDH_SN,
+  TILEPRO_OPC_ADD_SN, TILEPRO_OPC_ADIFFB_U_SN, TILEPRO_OPC_ADIFFH_SN,
+  TILEPRO_OPC_AND_SN, TILEPRO_OPC_AVGB_U_SN, TILEPRO_OPC_AVGH_SN,
+  TILEPRO_OPC_CRC32_32_SN, TILEPRO_OPC_CRC32_8_SN, TILEPRO_OPC_INTHB_SN,
+  TILEPRO_OPC_INTHH_SN, TILEPRO_OPC_INTLB_SN, TILEPRO_OPC_INTLH_SN,
+  TILEPRO_OPC_MAXB_U_SN,
+  BITFIELD(18, 4) /* index 680 */,
+  TILEPRO_OPC_MAXH_SN, TILEPRO_OPC_MINB_U_SN, TILEPRO_OPC_MINH_SN,
+  TILEPRO_OPC_MNZB_SN, TILEPRO_OPC_MNZH_SN, TILEPRO_OPC_MNZ_SN,
+  TILEPRO_OPC_MULHHA_SS_SN, TILEPRO_OPC_MULHHA_SU_SN,
+  TILEPRO_OPC_MULHHA_UU_SN, TILEPRO_OPC_MULHHSA_UU_SN,
+  TILEPRO_OPC_MULHH_SS_SN, TILEPRO_OPC_MULHH_SU_SN, TILEPRO_OPC_MULHH_UU_SN,
+  TILEPRO_OPC_MULHLA_SS_SN, TILEPRO_OPC_MULHLA_SU_SN,
+  TILEPRO_OPC_MULHLA_US_SN,
+  BITFIELD(18, 4) /* index 697 */,
+  TILEPRO_OPC_MULHLA_UU_SN, TILEPRO_OPC_MULHLSA_UU_SN,
+  TILEPRO_OPC_MULHL_SS_SN, TILEPRO_OPC_MULHL_SU_SN, TILEPRO_OPC_MULHL_US_SN,
+  TILEPRO_OPC_MULHL_UU_SN, TILEPRO_OPC_MULLLA_SS_SN, TILEPRO_OPC_MULLLA_SU_SN,
+  TILEPRO_OPC_MULLLA_UU_SN, TILEPRO_OPC_MULLLSA_UU_SN,
+  TILEPRO_OPC_MULLL_SS_SN, TILEPRO_OPC_MULLL_SU_SN, TILEPRO_OPC_MULLL_UU_SN,
+  TILEPRO_OPC_MVNZ_SN, TILEPRO_OPC_MVZ_SN, TILEPRO_OPC_MZB_SN,
+  BITFIELD(18, 4) /* index 714 */,
+  TILEPRO_OPC_MZH_SN, TILEPRO_OPC_MZ_SN, TILEPRO_OPC_NOR_SN, CHILD(731),
+  TILEPRO_OPC_PACKHB_SN, TILEPRO_OPC_PACKLB_SN, TILEPRO_OPC_RL_SN,
+  TILEPRO_OPC_S1A_SN, TILEPRO_OPC_S2A_SN, TILEPRO_OPC_S3A_SN,
+  TILEPRO_OPC_SADAB_U_SN, TILEPRO_OPC_SADAH_SN, TILEPRO_OPC_SADAH_U_SN,
+  TILEPRO_OPC_SADB_U_SN, TILEPRO_OPC_SADH_SN, TILEPRO_OPC_SADH_U_SN,
+  BITFIELD(12, 2) /* index 731 */,
+  TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN, CHILD(736),
+  BITFIELD(14, 2) /* index 736 */,
+  TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN, CHILD(741),
+  BITFIELD(16, 2) /* index 741 */,
+  TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN,
+  TILEPRO_OPC_MOVE_SN,
+  BITFIELD(18, 4) /* index 746 */,
+  TILEPRO_OPC_SEQB_SN, TILEPRO_OPC_SEQH_SN, TILEPRO_OPC_SEQ_SN,
+  TILEPRO_OPC_SHLB_SN, TILEPRO_OPC_SHLH_SN, TILEPRO_OPC_SHL_SN,
+  TILEPRO_OPC_SHRB_SN, TILEPRO_OPC_SHRH_SN, TILEPRO_OPC_SHR_SN,
+  TILEPRO_OPC_SLTB_SN, TILEPRO_OPC_SLTB_U_SN, TILEPRO_OPC_SLTEB_SN,
+  TILEPRO_OPC_SLTEB_U_SN, TILEPRO_OPC_SLTEH_SN, TILEPRO_OPC_SLTEH_U_SN,
+  TILEPRO_OPC_SLTE_SN,
+  BITFIELD(18, 4) /* index 763 */,
+  TILEPRO_OPC_SLTE_U_SN, TILEPRO_OPC_SLTH_SN, TILEPRO_OPC_SLTH_U_SN,
+  TILEPRO_OPC_SLT_SN, TILEPRO_OPC_SLT_U_SN, TILEPRO_OPC_SNEB_SN,
+  TILEPRO_OPC_SNEH_SN, TILEPRO_OPC_SNE_SN, TILEPRO_OPC_SRAB_SN,
+  TILEPRO_OPC_SRAH_SN, TILEPRO_OPC_SRA_SN, TILEPRO_OPC_SUBB_SN,
+  TILEPRO_OPC_SUBH_SN, TILEPRO_OPC_SUB_SN, TILEPRO_OPC_XOR_SN,
+  TILEPRO_OPC_DWORD_ALIGN_SN,
+  BITFIELD(18, 3) /* index 780 */,
+  CHILD(789), CHILD(792), CHILD(795), CHILD(798), CHILD(801), CHILD(804),
+  CHILD(807), CHILD(810),
+  BITFIELD(21, 1) /* index 789 */,
+  TILEPRO_OPC_ADDS_SN, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 792 */,
+  TILEPRO_OPC_SUBS_SN, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 795 */,
+  TILEPRO_OPC_ADDBS_U_SN, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 798 */,
+  TILEPRO_OPC_ADDHS_SN, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 801 */,
+  TILEPRO_OPC_SUBBS_U_SN, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 804 */,
+  TILEPRO_OPC_SUBHS_SN, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 807 */,
+  TILEPRO_OPC_PACKHS_SN, TILEPRO_OPC_NONE,
+  BITFIELD(21, 1) /* index 810 */,
+  TILEPRO_OPC_PACKBS_U_SN, TILEPRO_OPC_NONE,
+  BITFIELD(6, 2) /* index 813 */,
+  TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN,
+  CHILD(818),
+  BITFIELD(8, 2) /* index 818 */,
+  TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN,
+  CHILD(823),
+  BITFIELD(10, 2) /* index 823 */,
+  TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN,
+  TILEPRO_OPC_MOVELI_SN,
+  BITFIELD(6, 2) /* index 828 */,
+  TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, CHILD(833),
+  BITFIELD(8, 2) /* index 833 */,
+  TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, CHILD(838),
+  BITFIELD(10, 2) /* index 838 */,
+  TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, TILEPRO_OPC_MOVELI,
+  BITFIELD(0, 2) /* index 843 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, CHILD(848),
+  BITFIELD(2, 2) /* index 848 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, CHILD(853),
+  BITFIELD(4, 2) /* index 853 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, CHILD(858),
+  BITFIELD(6, 2) /* index 858 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, CHILD(863),
+  BITFIELD(8, 2) /* index 863 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, CHILD(868),
+  BITFIELD(10, 2) /* index 868 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_INFOL,
+  BITFIELD(20, 2) /* index 873 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_ADDIB, TILEPRO_OPC_ADDIH, TILEPRO_OPC_ADDI,
+  BITFIELD(20, 2) /* index 878 */,
+  TILEPRO_OPC_MAXIB_U, TILEPRO_OPC_MAXIH, TILEPRO_OPC_MINIB_U,
+  TILEPRO_OPC_MINIH,
+  BITFIELD(20, 2) /* index 883 */,
+  CHILD(888), TILEPRO_OPC_SEQIB, TILEPRO_OPC_SEQIH, TILEPRO_OPC_SEQI,
+  BITFIELD(6, 2) /* index 888 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, CHILD(893),
+  BITFIELD(8, 2) /* index 893 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, CHILD(898),
+  BITFIELD(10, 2) /* index 898 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_MOVEI,
+  BITFIELD(20, 2) /* index 903 */,
+  TILEPRO_OPC_SLTIB, TILEPRO_OPC_SLTIB_U, TILEPRO_OPC_SLTIH,
+  TILEPRO_OPC_SLTIH_U,
+  BITFIELD(20, 2) /* index 908 */,
+  TILEPRO_OPC_SLTI, TILEPRO_OPC_SLTI_U, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(20, 2) /* index 913 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_ADDIB_SN, TILEPRO_OPC_ADDIH_SN,
+  TILEPRO_OPC_ADDI_SN,
+  BITFIELD(20, 2) /* index 918 */,
+  TILEPRO_OPC_MAXIB_U_SN, TILEPRO_OPC_MAXIH_SN, TILEPRO_OPC_MINIB_U_SN,
+  TILEPRO_OPC_MINIH_SN,
+  BITFIELD(20, 2) /* index 923 */,
+  CHILD(928), TILEPRO_OPC_SEQIB_SN, TILEPRO_OPC_SEQIH_SN, TILEPRO_OPC_SEQI_SN,
+  BITFIELD(6, 2) /* index 928 */,
+  TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN, CHILD(933),
+  BITFIELD(8, 2) /* index 933 */,
+  TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN, CHILD(938),
+  BITFIELD(10, 2) /* index 938 */,
+  TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN,
+  TILEPRO_OPC_MOVEI_SN,
+  BITFIELD(20, 2) /* index 943 */,
+  TILEPRO_OPC_SLTIB_SN, TILEPRO_OPC_SLTIB_U_SN, TILEPRO_OPC_SLTIH_SN,
+  TILEPRO_OPC_SLTIH_U_SN,
+  BITFIELD(20, 2) /* index 948 */,
+  TILEPRO_OPC_SLTI_SN, TILEPRO_OPC_SLTI_U_SN, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE,
+  BITFIELD(20, 2) /* index 953 */,
+  TILEPRO_OPC_NONE, CHILD(958), TILEPRO_OPC_XORI, TILEPRO_OPC_NONE,
+  BITFIELD(0, 2) /* index 958 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(963),
+  BITFIELD(2, 2) /* index 963 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(968),
+  BITFIELD(4, 2) /* index 968 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(973),
+  BITFIELD(6, 2) /* index 973 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(978),
+  BITFIELD(8, 2) /* index 978 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(983),
+  BITFIELD(10, 2) /* index 983 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_INFO,
+  BITFIELD(20, 2) /* index 988 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_ANDI_SN, TILEPRO_OPC_XORI_SN,
+  TILEPRO_OPC_NONE,
+  BITFIELD(17, 5) /* index 993 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_RLI, TILEPRO_OPC_SHLIB, TILEPRO_OPC_SHLIH,
+  TILEPRO_OPC_SHLI, TILEPRO_OPC_SHRIB, TILEPRO_OPC_SHRIH, TILEPRO_OPC_SHRI,
+  TILEPRO_OPC_SRAIB, TILEPRO_OPC_SRAIH, TILEPRO_OPC_SRAI, CHILD(1026),
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(12, 4) /* index 1026 */,
+  TILEPRO_OPC_NONE, CHILD(1043), CHILD(1046), CHILD(1049), CHILD(1052),
+  CHILD(1055), CHILD(1058), CHILD(1061), CHILD(1064), CHILD(1067),
+  CHILD(1070), CHILD(1073), TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1043 */,
+  TILEPRO_OPC_BITX, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1046 */,
+  TILEPRO_OPC_BYTEX, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1049 */,
+  TILEPRO_OPC_CLZ, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1052 */,
+  TILEPRO_OPC_CTZ, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1055 */,
+  TILEPRO_OPC_FNOP, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1058 */,
+  TILEPRO_OPC_NOP, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1061 */,
+  TILEPRO_OPC_PCNT, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1064 */,
+  TILEPRO_OPC_TBLIDXB0, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1067 */,
+  TILEPRO_OPC_TBLIDXB1, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1070 */,
+  TILEPRO_OPC_TBLIDXB2, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1073 */,
+  TILEPRO_OPC_TBLIDXB3, TILEPRO_OPC_NONE,
+  BITFIELD(17, 5) /* index 1076 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_RLI_SN, TILEPRO_OPC_SHLIB_SN,
+  TILEPRO_OPC_SHLIH_SN, TILEPRO_OPC_SHLI_SN, TILEPRO_OPC_SHRIB_SN,
+  TILEPRO_OPC_SHRIH_SN, TILEPRO_OPC_SHRI_SN, TILEPRO_OPC_SRAIB_SN,
+  TILEPRO_OPC_SRAIH_SN, TILEPRO_OPC_SRAI_SN, CHILD(1109), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(12, 4) /* index 1109 */,
+  TILEPRO_OPC_NONE, CHILD(1126), CHILD(1129), CHILD(1132), CHILD(1135),
+  CHILD(1055), CHILD(1058), CHILD(1138), CHILD(1141), CHILD(1144),
+  CHILD(1147), CHILD(1150), TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1126 */,
+  TILEPRO_OPC_BITX_SN, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1129 */,
+  TILEPRO_OPC_BYTEX_SN, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1132 */,
+  TILEPRO_OPC_CLZ_SN, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1135 */,
+  TILEPRO_OPC_CTZ_SN, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1138 */,
+  TILEPRO_OPC_PCNT_SN, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1141 */,
+  TILEPRO_OPC_TBLIDXB0_SN, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1144 */,
+  TILEPRO_OPC_TBLIDXB1_SN, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1147 */,
+  TILEPRO_OPC_TBLIDXB2_SN, TILEPRO_OPC_NONE,
+  BITFIELD(16, 1) /* index 1150 */,
+  TILEPRO_OPC_TBLIDXB3_SN, TILEPRO_OPC_NONE,
+};
+
+static const unsigned short decode_X1_fsm[1540] =
+{
+  BITFIELD(54, 9) /* index 0 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  CHILD(513), CHILD(561), CHILD(594), TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, CHILD(641),
+  CHILD(689), CHILD(722), TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, CHILD(766),
+  CHILD(766), CHILD(766), CHILD(766), CHILD(766), CHILD(766), CHILD(766),
+  CHILD(766), CHILD(766), CHILD(766), CHILD(766), CHILD(766), CHILD(766),
+  CHILD(766), CHILD(766), CHILD(766), CHILD(766), CHILD(766), CHILD(766),
+  CHILD(766), CHILD(766), CHILD(766), CHILD(766), CHILD(766), CHILD(766),
+  CHILD(766), CHILD(766), CHILD(766), CHILD(766), CHILD(766), CHILD(766),
+  CHILD(766), CHILD(781), CHILD(781), CHILD(781), CHILD(781), CHILD(781),
+  CHILD(781), CHILD(781), CHILD(781), CHILD(781), CHILD(781), CHILD(781),
+  CHILD(781), CHILD(781), CHILD(781), CHILD(781), CHILD(781), CHILD(781),
+  CHILD(781), CHILD(781), CHILD(781), CHILD(781), CHILD(781), CHILD(781),
+  CHILD(781), CHILD(781), CHILD(781), CHILD(781), CHILD(781), CHILD(781),
+  CHILD(781), CHILD(781), CHILD(781), CHILD(796), CHILD(796), CHILD(796),
+  CHILD(796), CHILD(796), CHILD(796), CHILD(796), CHILD(796), CHILD(796),
+  CHILD(796), CHILD(796), CHILD(796), CHILD(796), CHILD(796), CHILD(796),
+  CHILD(796), CHILD(796), CHILD(796), CHILD(796), CHILD(796), CHILD(796),
+  CHILD(796), CHILD(796), CHILD(796), CHILD(796), CHILD(796), CHILD(796),
+  CHILD(796), CHILD(796), CHILD(796), CHILD(796), CHILD(796), CHILD(826),
+  CHILD(826), CHILD(826), CHILD(826), CHILD(826), CHILD(826), CHILD(826),
+  CHILD(826), CHILD(826), CHILD(826), CHILD(826), CHILD(826), CHILD(826),
+  CHILD(826), CHILD(826), CHILD(826), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843), CHILD(843),
+  CHILD(843), CHILD(860), CHILD(899), CHILD(923), CHILD(932),
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  CHILD(941), CHILD(950), CHILD(974), CHILD(983), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM,
+  TILEPRO_OPC_MM, TILEPRO_OPC_MM, TILEPRO_OPC_MM, CHILD(992),
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, CHILD(1334),
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J,
+  TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_J, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL,
+  TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_JAL, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(49, 5) /* index 513 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_ADDB, TILEPRO_OPC_ADDH, TILEPRO_OPC_ADD,
+  TILEPRO_OPC_AND, TILEPRO_OPC_INTHB, TILEPRO_OPC_INTHH, TILEPRO_OPC_INTLB,
+  TILEPRO_OPC_INTLH, TILEPRO_OPC_JALRP, TILEPRO_OPC_JALR, TILEPRO_OPC_JRP,
+  TILEPRO_OPC_JR, TILEPRO_OPC_LNK, TILEPRO_OPC_MAXB_U, TILEPRO_OPC_MAXH,
+  TILEPRO_OPC_MINB_U, TILEPRO_OPC_MINH, TILEPRO_OPC_MNZB, TILEPRO_OPC_MNZH,
+  TILEPRO_OPC_MNZ, TILEPRO_OPC_MZB, TILEPRO_OPC_MZH, TILEPRO_OPC_MZ,
+  TILEPRO_OPC_NOR, CHILD(546), TILEPRO_OPC_PACKHB, TILEPRO_OPC_PACKLB,
+  TILEPRO_OPC_RL, TILEPRO_OPC_S1A, TILEPRO_OPC_S2A, TILEPRO_OPC_S3A,
+  BITFIELD(43, 2) /* index 546 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, CHILD(551),
+  BITFIELD(45, 2) /* index 551 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, CHILD(556),
+  BITFIELD(47, 2) /* index 556 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_MOVE,
+  BITFIELD(49, 5) /* index 561 */,
+  TILEPRO_OPC_SB, TILEPRO_OPC_SEQB, TILEPRO_OPC_SEQH, TILEPRO_OPC_SEQ,
+  TILEPRO_OPC_SHLB, TILEPRO_OPC_SHLH, TILEPRO_OPC_SHL, TILEPRO_OPC_SHRB,
+  TILEPRO_OPC_SHRH, TILEPRO_OPC_SHR, TILEPRO_OPC_SH, TILEPRO_OPC_SLTB,
+  TILEPRO_OPC_SLTB_U, TILEPRO_OPC_SLTEB, TILEPRO_OPC_SLTEB_U,
+  TILEPRO_OPC_SLTEH, TILEPRO_OPC_SLTEH_U, TILEPRO_OPC_SLTE,
+  TILEPRO_OPC_SLTE_U, TILEPRO_OPC_SLTH, TILEPRO_OPC_SLTH_U, TILEPRO_OPC_SLT,
+  TILEPRO_OPC_SLT_U, TILEPRO_OPC_SNEB, TILEPRO_OPC_SNEH, TILEPRO_OPC_SNE,
+  TILEPRO_OPC_SRAB, TILEPRO_OPC_SRAH, TILEPRO_OPC_SRA, TILEPRO_OPC_SUBB,
+  TILEPRO_OPC_SUBH, TILEPRO_OPC_SUB,
+  BITFIELD(49, 4) /* index 594 */,
+  CHILD(611), CHILD(614), CHILD(617), CHILD(620), CHILD(623), CHILD(626),
+  CHILD(629), CHILD(632), CHILD(635), CHILD(638), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 611 */,
+  TILEPRO_OPC_SW, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 614 */,
+  TILEPRO_OPC_XOR, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 617 */,
+  TILEPRO_OPC_ADDS, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 620 */,
+  TILEPRO_OPC_SUBS, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 623 */,
+  TILEPRO_OPC_ADDBS_U, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 626 */,
+  TILEPRO_OPC_ADDHS, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 629 */,
+  TILEPRO_OPC_SUBBS_U, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 632 */,
+  TILEPRO_OPC_SUBHS, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 635 */,
+  TILEPRO_OPC_PACKHS, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 638 */,
+  TILEPRO_OPC_PACKBS_U, TILEPRO_OPC_NONE,
+  BITFIELD(49, 5) /* index 641 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_ADDB_SN, TILEPRO_OPC_ADDH_SN,
+  TILEPRO_OPC_ADD_SN, TILEPRO_OPC_AND_SN, TILEPRO_OPC_INTHB_SN,
+  TILEPRO_OPC_INTHH_SN, TILEPRO_OPC_INTLB_SN, TILEPRO_OPC_INTLH_SN,
+  TILEPRO_OPC_JALRP, TILEPRO_OPC_JALR, TILEPRO_OPC_JRP, TILEPRO_OPC_JR,
+  TILEPRO_OPC_LNK_SN, TILEPRO_OPC_MAXB_U_SN, TILEPRO_OPC_MAXH_SN,
+  TILEPRO_OPC_MINB_U_SN, TILEPRO_OPC_MINH_SN, TILEPRO_OPC_MNZB_SN,
+  TILEPRO_OPC_MNZH_SN, TILEPRO_OPC_MNZ_SN, TILEPRO_OPC_MZB_SN,
+  TILEPRO_OPC_MZH_SN, TILEPRO_OPC_MZ_SN, TILEPRO_OPC_NOR_SN, CHILD(674),
+  TILEPRO_OPC_PACKHB_SN, TILEPRO_OPC_PACKLB_SN, TILEPRO_OPC_RL_SN,
+  TILEPRO_OPC_S1A_SN, TILEPRO_OPC_S2A_SN, TILEPRO_OPC_S3A_SN,
+  BITFIELD(43, 2) /* index 674 */,
+  TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN, CHILD(679),
+  BITFIELD(45, 2) /* index 679 */,
+  TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN, CHILD(684),
+  BITFIELD(47, 2) /* index 684 */,
+  TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN, TILEPRO_OPC_OR_SN,
+  TILEPRO_OPC_MOVE_SN,
+  BITFIELD(49, 5) /* index 689 */,
+  TILEPRO_OPC_SB, TILEPRO_OPC_SEQB_SN, TILEPRO_OPC_SEQH_SN,
+  TILEPRO_OPC_SEQ_SN, TILEPRO_OPC_SHLB_SN, TILEPRO_OPC_SHLH_SN,
+  TILEPRO_OPC_SHL_SN, TILEPRO_OPC_SHRB_SN, TILEPRO_OPC_SHRH_SN,
+  TILEPRO_OPC_SHR_SN, TILEPRO_OPC_SH, TILEPRO_OPC_SLTB_SN,
+  TILEPRO_OPC_SLTB_U_SN, TILEPRO_OPC_SLTEB_SN, TILEPRO_OPC_SLTEB_U_SN,
+  TILEPRO_OPC_SLTEH_SN, TILEPRO_OPC_SLTEH_U_SN, TILEPRO_OPC_SLTE_SN,
+  TILEPRO_OPC_SLTE_U_SN, TILEPRO_OPC_SLTH_SN, TILEPRO_OPC_SLTH_U_SN,
+  TILEPRO_OPC_SLT_SN, TILEPRO_OPC_SLT_U_SN, TILEPRO_OPC_SNEB_SN,
+  TILEPRO_OPC_SNEH_SN, TILEPRO_OPC_SNE_SN, TILEPRO_OPC_SRAB_SN,
+  TILEPRO_OPC_SRAH_SN, TILEPRO_OPC_SRA_SN, TILEPRO_OPC_SUBB_SN,
+  TILEPRO_OPC_SUBH_SN, TILEPRO_OPC_SUB_SN,
+  BITFIELD(49, 4) /* index 722 */,
+  CHILD(611), CHILD(739), CHILD(742), CHILD(745), CHILD(748), CHILD(751),
+  CHILD(754), CHILD(757), CHILD(760), CHILD(763), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 739 */,
+  TILEPRO_OPC_XOR_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 742 */,
+  TILEPRO_OPC_ADDS_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 745 */,
+  TILEPRO_OPC_SUBS_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 748 */,
+  TILEPRO_OPC_ADDBS_U_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 751 */,
+  TILEPRO_OPC_ADDHS_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 754 */,
+  TILEPRO_OPC_SUBBS_U_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 757 */,
+  TILEPRO_OPC_SUBHS_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 760 */,
+  TILEPRO_OPC_PACKHS_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 763 */,
+  TILEPRO_OPC_PACKBS_U_SN, TILEPRO_OPC_NONE,
+  BITFIELD(37, 2) /* index 766 */,
+  TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN,
+  CHILD(771),
+  BITFIELD(39, 2) /* index 771 */,
+  TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN,
+  CHILD(776),
+  BITFIELD(41, 2) /* index 776 */,
+  TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN, TILEPRO_OPC_ADDLI_SN,
+  TILEPRO_OPC_MOVELI_SN,
+  BITFIELD(37, 2) /* index 781 */,
+  TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, CHILD(786),
+  BITFIELD(39, 2) /* index 786 */,
+  TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, CHILD(791),
+  BITFIELD(41, 2) /* index 791 */,
+  TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, TILEPRO_OPC_ADDLI, TILEPRO_OPC_MOVELI,
+  BITFIELD(31, 2) /* index 796 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, CHILD(801),
+  BITFIELD(33, 2) /* index 801 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, CHILD(806),
+  BITFIELD(35, 2) /* index 806 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, CHILD(811),
+  BITFIELD(37, 2) /* index 811 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, CHILD(816),
+  BITFIELD(39, 2) /* index 816 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, CHILD(821),
+  BITFIELD(41, 2) /* index 821 */,
+  TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_AULI, TILEPRO_OPC_INFOL,
+  BITFIELD(31, 4) /* index 826 */,
+  TILEPRO_OPC_BZ, TILEPRO_OPC_BZT, TILEPRO_OPC_BNZ, TILEPRO_OPC_BNZT,
+  TILEPRO_OPC_BGZ, TILEPRO_OPC_BGZT, TILEPRO_OPC_BGEZ, TILEPRO_OPC_BGEZT,
+  TILEPRO_OPC_BLZ, TILEPRO_OPC_BLZT, TILEPRO_OPC_BLEZ, TILEPRO_OPC_BLEZT,
+  TILEPRO_OPC_BBS, TILEPRO_OPC_BBST, TILEPRO_OPC_BBNS, TILEPRO_OPC_BBNST,
+  BITFIELD(31, 4) /* index 843 */,
+  TILEPRO_OPC_BZ_SN, TILEPRO_OPC_BZT_SN, TILEPRO_OPC_BNZ_SN,
+  TILEPRO_OPC_BNZT_SN, TILEPRO_OPC_BGZ_SN, TILEPRO_OPC_BGZT_SN,
+  TILEPRO_OPC_BGEZ_SN, TILEPRO_OPC_BGEZT_SN, TILEPRO_OPC_BLZ_SN,
+  TILEPRO_OPC_BLZT_SN, TILEPRO_OPC_BLEZ_SN, TILEPRO_OPC_BLEZT_SN,
+  TILEPRO_OPC_BBS_SN, TILEPRO_OPC_BBST_SN, TILEPRO_OPC_BBNS_SN,
+  TILEPRO_OPC_BBNST_SN,
+  BITFIELD(51, 3) /* index 860 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_ADDIB, TILEPRO_OPC_ADDIH, TILEPRO_OPC_ADDI,
+  CHILD(869), TILEPRO_OPC_MAXIB_U, TILEPRO_OPC_MAXIH, TILEPRO_OPC_MFSPR,
+  BITFIELD(31, 2) /* index 869 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(874),
+  BITFIELD(33, 2) /* index 874 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(879),
+  BITFIELD(35, 2) /* index 879 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(884),
+  BITFIELD(37, 2) /* index 884 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(889),
+  BITFIELD(39, 2) /* index 889 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(894),
+  BITFIELD(41, 2) /* index 894 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_INFO,
+  BITFIELD(51, 3) /* index 899 */,
+  TILEPRO_OPC_MINIB_U, TILEPRO_OPC_MINIH, TILEPRO_OPC_MTSPR, CHILD(908),
+  TILEPRO_OPC_SEQIB, TILEPRO_OPC_SEQIH, TILEPRO_OPC_SEQI, TILEPRO_OPC_SLTIB,
+  BITFIELD(37, 2) /* index 908 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, CHILD(913),
+  BITFIELD(39, 2) /* index 913 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, CHILD(918),
+  BITFIELD(41, 2) /* index 918 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_MOVEI,
+  BITFIELD(51, 3) /* index 923 */,
+  TILEPRO_OPC_SLTIB_U, TILEPRO_OPC_SLTIH, TILEPRO_OPC_SLTIH_U,
+  TILEPRO_OPC_SLTI, TILEPRO_OPC_SLTI_U, TILEPRO_OPC_XORI, TILEPRO_OPC_LBADD,
+  TILEPRO_OPC_LBADD_U,
+  BITFIELD(51, 3) /* index 932 */,
+  TILEPRO_OPC_LHADD, TILEPRO_OPC_LHADD_U, TILEPRO_OPC_LWADD,
+  TILEPRO_OPC_LWADD_NA, TILEPRO_OPC_SBADD, TILEPRO_OPC_SHADD,
+  TILEPRO_OPC_SWADD, TILEPRO_OPC_NONE,
+  BITFIELD(51, 3) /* index 941 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_ADDIB_SN, TILEPRO_OPC_ADDIH_SN,
+  TILEPRO_OPC_ADDI_SN, TILEPRO_OPC_ANDI_SN, TILEPRO_OPC_MAXIB_U_SN,
+  TILEPRO_OPC_MAXIH_SN, TILEPRO_OPC_MFSPR,
+  BITFIELD(51, 3) /* index 950 */,
+  TILEPRO_OPC_MINIB_U_SN, TILEPRO_OPC_MINIH_SN, TILEPRO_OPC_MTSPR, CHILD(959),
+  TILEPRO_OPC_SEQIB_SN, TILEPRO_OPC_SEQIH_SN, TILEPRO_OPC_SEQI_SN,
+  TILEPRO_OPC_SLTIB_SN,
+  BITFIELD(37, 2) /* index 959 */,
+  TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN, CHILD(964),
+  BITFIELD(39, 2) /* index 964 */,
+  TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN, CHILD(969),
+  BITFIELD(41, 2) /* index 969 */,
+  TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN, TILEPRO_OPC_ORI_SN,
+  TILEPRO_OPC_MOVEI_SN,
+  BITFIELD(51, 3) /* index 974 */,
+  TILEPRO_OPC_SLTIB_U_SN, TILEPRO_OPC_SLTIH_SN, TILEPRO_OPC_SLTIH_U_SN,
+  TILEPRO_OPC_SLTI_SN, TILEPRO_OPC_SLTI_U_SN, TILEPRO_OPC_XORI_SN,
+  TILEPRO_OPC_LBADD_SN, TILEPRO_OPC_LBADD_U_SN,
+  BITFIELD(51, 3) /* index 983 */,
+  TILEPRO_OPC_LHADD_SN, TILEPRO_OPC_LHADD_U_SN, TILEPRO_OPC_LWADD_SN,
+  TILEPRO_OPC_LWADD_NA_SN, TILEPRO_OPC_SBADD, TILEPRO_OPC_SHADD,
+  TILEPRO_OPC_SWADD, TILEPRO_OPC_NONE,
+  BITFIELD(46, 7) /* index 992 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  CHILD(1121), CHILD(1121), CHILD(1121), CHILD(1121), CHILD(1124),
+  CHILD(1124), CHILD(1124), CHILD(1124), CHILD(1127), CHILD(1127),
+  CHILD(1127), CHILD(1127), CHILD(1130), CHILD(1130), CHILD(1130),
+  CHILD(1130), CHILD(1133), CHILD(1133), CHILD(1133), CHILD(1133),
+  CHILD(1136), CHILD(1136), CHILD(1136), CHILD(1136), CHILD(1139),
+  CHILD(1139), CHILD(1139), CHILD(1139), CHILD(1142), CHILD(1142),
+  CHILD(1142), CHILD(1142), CHILD(1145), CHILD(1145), CHILD(1145),
+  CHILD(1145), CHILD(1148), CHILD(1148), CHILD(1148), CHILD(1148),
+  CHILD(1151), CHILD(1242), CHILD(1290), CHILD(1323), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1121 */,
+  TILEPRO_OPC_RLI, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1124 */,
+  TILEPRO_OPC_SHLIB, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1127 */,
+  TILEPRO_OPC_SHLIH, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1130 */,
+  TILEPRO_OPC_SHLI, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1133 */,
+  TILEPRO_OPC_SHRIB, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1136 */,
+  TILEPRO_OPC_SHRIH, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1139 */,
+  TILEPRO_OPC_SHRI, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1142 */,
+  TILEPRO_OPC_SRAIB, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1145 */,
+  TILEPRO_OPC_SRAIH, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1148 */,
+  TILEPRO_OPC_SRAI, TILEPRO_OPC_NONE,
+  BITFIELD(43, 3) /* index 1151 */,
+  TILEPRO_OPC_NONE, CHILD(1160), CHILD(1163), CHILD(1166), CHILD(1169),
+  CHILD(1172), CHILD(1175), CHILD(1178),
+  BITFIELD(53, 1) /* index 1160 */,
+  TILEPRO_OPC_DRAIN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1163 */,
+  TILEPRO_OPC_DTLBPR, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1166 */,
+  TILEPRO_OPC_FINV, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1169 */,
+  TILEPRO_OPC_FLUSH, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1172 */,
+  TILEPRO_OPC_FNOP, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1175 */,
+  TILEPRO_OPC_ICOH, TILEPRO_OPC_NONE,
+  BITFIELD(31, 2) /* index 1178 */,
+  CHILD(1183), CHILD(1211), CHILD(1239), CHILD(1239),
+  BITFIELD(53, 1) /* index 1183 */,
+  CHILD(1186), TILEPRO_OPC_NONE,
+  BITFIELD(33, 2) /* index 1186 */,
+  TILEPRO_OPC_ILL, TILEPRO_OPC_ILL, TILEPRO_OPC_ILL, CHILD(1191),
+  BITFIELD(35, 2) /* index 1191 */,
+  TILEPRO_OPC_ILL, CHILD(1196), TILEPRO_OPC_ILL, TILEPRO_OPC_ILL,
+  BITFIELD(37, 2) /* index 1196 */,
+  TILEPRO_OPC_ILL, CHILD(1201), TILEPRO_OPC_ILL, TILEPRO_OPC_ILL,
+  BITFIELD(39, 2) /* index 1201 */,
+  TILEPRO_OPC_ILL, CHILD(1206), TILEPRO_OPC_ILL, TILEPRO_OPC_ILL,
+  BITFIELD(41, 2) /* index 1206 */,
+  TILEPRO_OPC_ILL, TILEPRO_OPC_ILL, TILEPRO_OPC_BPT, TILEPRO_OPC_ILL,
+  BITFIELD(53, 1) /* index 1211 */,
+  CHILD(1214), TILEPRO_OPC_NONE,
+  BITFIELD(33, 2) /* index 1214 */,
+  TILEPRO_OPC_ILL, TILEPRO_OPC_ILL, TILEPRO_OPC_ILL, CHILD(1219),
+  BITFIELD(35, 2) /* index 1219 */,
+  TILEPRO_OPC_ILL, CHILD(1224), TILEPRO_OPC_ILL, TILEPRO_OPC_ILL,
+  BITFIELD(37, 2) /* index 1224 */,
+  TILEPRO_OPC_ILL, CHILD(1229), TILEPRO_OPC_ILL, TILEPRO_OPC_ILL,
+  BITFIELD(39, 2) /* index 1229 */,
+  TILEPRO_OPC_ILL, CHILD(1234), TILEPRO_OPC_ILL, TILEPRO_OPC_ILL,
+  BITFIELD(41, 2) /* index 1234 */,
+  TILEPRO_OPC_ILL, TILEPRO_OPC_ILL, TILEPRO_OPC_RAISE, TILEPRO_OPC_ILL,
+  BITFIELD(53, 1) /* index 1239 */,
+  TILEPRO_OPC_ILL, TILEPRO_OPC_NONE,
+  BITFIELD(43, 3) /* index 1242 */,
+  CHILD(1251), CHILD(1254), CHILD(1257), CHILD(1275), CHILD(1278),
+  CHILD(1281), CHILD(1284), CHILD(1287),
+  BITFIELD(53, 1) /* index 1251 */,
+  TILEPRO_OPC_INV, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1254 */,
+  TILEPRO_OPC_IRET, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1257 */,
+  CHILD(1260), TILEPRO_OPC_NONE,
+  BITFIELD(31, 2) /* index 1260 */,
+  TILEPRO_OPC_LB, TILEPRO_OPC_LB, TILEPRO_OPC_LB, CHILD(1265),
+  BITFIELD(33, 2) /* index 1265 */,
+  TILEPRO_OPC_LB, TILEPRO_OPC_LB, TILEPRO_OPC_LB, CHILD(1270),
+  BITFIELD(35, 2) /* index 1270 */,
+  TILEPRO_OPC_LB, TILEPRO_OPC_LB, TILEPRO_OPC_LB, TILEPRO_OPC_PREFETCH,
+  BITFIELD(53, 1) /* index 1275 */,
+  TILEPRO_OPC_LB_U, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1278 */,
+  TILEPRO_OPC_LH, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1281 */,
+  TILEPRO_OPC_LH_U, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1284 */,
+  TILEPRO_OPC_LW, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1287 */,
+  TILEPRO_OPC_MF, TILEPRO_OPC_NONE,
+  BITFIELD(43, 3) /* index 1290 */,
+  CHILD(1299), CHILD(1302), CHILD(1305), CHILD(1308), CHILD(1311),
+  CHILD(1314), CHILD(1317), CHILD(1320),
+  BITFIELD(53, 1) /* index 1299 */,
+  TILEPRO_OPC_NAP, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1302 */,
+  TILEPRO_OPC_NOP, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1305 */,
+  TILEPRO_OPC_SWINT0, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1308 */,
+  TILEPRO_OPC_SWINT1, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1311 */,
+  TILEPRO_OPC_SWINT2, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1314 */,
+  TILEPRO_OPC_SWINT3, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1317 */,
+  TILEPRO_OPC_TNS, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1320 */,
+  TILEPRO_OPC_WH64, TILEPRO_OPC_NONE,
+  BITFIELD(43, 2) /* index 1323 */,
+  CHILD(1328), TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(45, 1) /* index 1328 */,
+  CHILD(1331), TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1331 */,
+  TILEPRO_OPC_LW_NA, TILEPRO_OPC_NONE,
+  BITFIELD(46, 7) /* index 1334 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  CHILD(1463), CHILD(1463), CHILD(1463), CHILD(1463), CHILD(1466),
+  CHILD(1466), CHILD(1466), CHILD(1466), CHILD(1469), CHILD(1469),
+  CHILD(1469), CHILD(1469), CHILD(1472), CHILD(1472), CHILD(1472),
+  CHILD(1472), CHILD(1475), CHILD(1475), CHILD(1475), CHILD(1475),
+  CHILD(1478), CHILD(1478), CHILD(1478), CHILD(1478), CHILD(1481),
+  CHILD(1481), CHILD(1481), CHILD(1481), CHILD(1484), CHILD(1484),
+  CHILD(1484), CHILD(1484), CHILD(1487), CHILD(1487), CHILD(1487),
+  CHILD(1487), CHILD(1490), CHILD(1490), CHILD(1490), CHILD(1490),
+  CHILD(1151), CHILD(1493), CHILD(1517), CHILD(1529), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1463 */,
+  TILEPRO_OPC_RLI_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1466 */,
+  TILEPRO_OPC_SHLIB_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1469 */,
+  TILEPRO_OPC_SHLIH_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1472 */,
+  TILEPRO_OPC_SHLI_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1475 */,
+  TILEPRO_OPC_SHRIB_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1478 */,
+  TILEPRO_OPC_SHRIH_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1481 */,
+  TILEPRO_OPC_SHRI_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1484 */,
+  TILEPRO_OPC_SRAIB_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1487 */,
+  TILEPRO_OPC_SRAIH_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1490 */,
+  TILEPRO_OPC_SRAI_SN, TILEPRO_OPC_NONE,
+  BITFIELD(43, 3) /* index 1493 */,
+  CHILD(1251), CHILD(1254), CHILD(1502), CHILD(1505), CHILD(1508),
+  CHILD(1511), CHILD(1514), CHILD(1287),
+  BITFIELD(53, 1) /* index 1502 */,
+  TILEPRO_OPC_LB_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1505 */,
+  TILEPRO_OPC_LB_U_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1508 */,
+  TILEPRO_OPC_LH_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1511 */,
+  TILEPRO_OPC_LH_U_SN, TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1514 */,
+  TILEPRO_OPC_LW_SN, TILEPRO_OPC_NONE,
+  BITFIELD(43, 3) /* index 1517 */,
+  CHILD(1299), CHILD(1302), CHILD(1305), CHILD(1308), CHILD(1311),
+  CHILD(1314), CHILD(1526), CHILD(1320),
+  BITFIELD(53, 1) /* index 1526 */,
+  TILEPRO_OPC_TNS_SN, TILEPRO_OPC_NONE,
+  BITFIELD(43, 2) /* index 1529 */,
+  CHILD(1534), TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(45, 1) /* index 1534 */,
+  CHILD(1537), TILEPRO_OPC_NONE,
+  BITFIELD(53, 1) /* index 1537 */,
+  TILEPRO_OPC_LW_NA_SN, TILEPRO_OPC_NONE,
+};
+
+static const unsigned short decode_Y0_fsm[168] =
+{
+  BITFIELD(27, 4) /* index 0 */,
+  TILEPRO_OPC_NONE, CHILD(17), CHILD(22), CHILD(27), CHILD(47), CHILD(52),
+  CHILD(57), CHILD(62), CHILD(67), TILEPRO_OPC_ADDI, CHILD(72), CHILD(102),
+  TILEPRO_OPC_SEQI, CHILD(117), TILEPRO_OPC_SLTI, TILEPRO_OPC_SLTI_U,
+  BITFIELD(18, 2) /* index 17 */,
+  TILEPRO_OPC_ADD, TILEPRO_OPC_S1A, TILEPRO_OPC_S2A, TILEPRO_OPC_SUB,
+  BITFIELD(18, 2) /* index 22 */,
+  TILEPRO_OPC_MNZ, TILEPRO_OPC_MVNZ, TILEPRO_OPC_MVZ, TILEPRO_OPC_MZ,
+  BITFIELD(18, 2) /* index 27 */,
+  TILEPRO_OPC_AND, TILEPRO_OPC_NOR, CHILD(32), TILEPRO_OPC_XOR,
+  BITFIELD(12, 2) /* index 32 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, CHILD(37),
+  BITFIELD(14, 2) /* index 37 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, CHILD(42),
+  BITFIELD(16, 2) /* index 42 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_MOVE,
+  BITFIELD(18, 2) /* index 47 */,
+  TILEPRO_OPC_RL, TILEPRO_OPC_SHL, TILEPRO_OPC_SHR, TILEPRO_OPC_SRA,
+  BITFIELD(18, 2) /* index 52 */,
+  TILEPRO_OPC_SLTE, TILEPRO_OPC_SLTE_U, TILEPRO_OPC_SLT, TILEPRO_OPC_SLT_U,
+  BITFIELD(18, 2) /* index 57 */,
+  TILEPRO_OPC_MULHLSA_UU, TILEPRO_OPC_S3A, TILEPRO_OPC_SEQ, TILEPRO_OPC_SNE,
+  BITFIELD(18, 2) /* index 62 */,
+  TILEPRO_OPC_MULHH_SS, TILEPRO_OPC_MULHH_UU, TILEPRO_OPC_MULLL_SS,
+  TILEPRO_OPC_MULLL_UU,
+  BITFIELD(18, 2) /* index 67 */,
+  TILEPRO_OPC_MULHHA_SS, TILEPRO_OPC_MULHHA_UU, TILEPRO_OPC_MULLLA_SS,
+  TILEPRO_OPC_MULLLA_UU,
+  BITFIELD(0, 2) /* index 72 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(77),
+  BITFIELD(2, 2) /* index 77 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(82),
+  BITFIELD(4, 2) /* index 82 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(87),
+  BITFIELD(6, 2) /* index 87 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(92),
+  BITFIELD(8, 2) /* index 92 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(97),
+  BITFIELD(10, 2) /* index 97 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_INFO,
+  BITFIELD(6, 2) /* index 102 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, CHILD(107),
+  BITFIELD(8, 2) /* index 107 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, CHILD(112),
+  BITFIELD(10, 2) /* index 112 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_MOVEI,
+  BITFIELD(15, 5) /* index 117 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_RLI, TILEPRO_OPC_RLI, TILEPRO_OPC_RLI, TILEPRO_OPC_RLI,
+  TILEPRO_OPC_SHLI, TILEPRO_OPC_SHLI, TILEPRO_OPC_SHLI, TILEPRO_OPC_SHLI,
+  TILEPRO_OPC_SHRI, TILEPRO_OPC_SHRI, TILEPRO_OPC_SHRI, TILEPRO_OPC_SHRI,
+  TILEPRO_OPC_SRAI, TILEPRO_OPC_SRAI, TILEPRO_OPC_SRAI, TILEPRO_OPC_SRAI,
+  CHILD(150), CHILD(159), TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(12, 3) /* index 150 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_BITX, TILEPRO_OPC_BYTEX, TILEPRO_OPC_CLZ,
+  TILEPRO_OPC_CTZ, TILEPRO_OPC_FNOP, TILEPRO_OPC_NOP, TILEPRO_OPC_PCNT,
+  BITFIELD(12, 3) /* index 159 */,
+  TILEPRO_OPC_TBLIDXB0, TILEPRO_OPC_TBLIDXB1, TILEPRO_OPC_TBLIDXB2,
+  TILEPRO_OPC_TBLIDXB3, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE,
+};
+
+static const unsigned short decode_Y1_fsm[140] =
+{
+  BITFIELD(59, 4) /* index 0 */,
+  TILEPRO_OPC_NONE, CHILD(17), CHILD(22), CHILD(27), CHILD(47), CHILD(52),
+  CHILD(57), TILEPRO_OPC_ADDI, CHILD(62), CHILD(92), TILEPRO_OPC_SEQI,
+  CHILD(107), TILEPRO_OPC_SLTI, TILEPRO_OPC_SLTI_U, TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE,
+  BITFIELD(49, 2) /* index 17 */,
+  TILEPRO_OPC_ADD, TILEPRO_OPC_S1A, TILEPRO_OPC_S2A, TILEPRO_OPC_SUB,
+  BITFIELD(49, 2) /* index 22 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_MNZ, TILEPRO_OPC_MZ, TILEPRO_OPC_NONE,
+  BITFIELD(49, 2) /* index 27 */,
+  TILEPRO_OPC_AND, TILEPRO_OPC_NOR, CHILD(32), TILEPRO_OPC_XOR,
+  BITFIELD(43, 2) /* index 32 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, CHILD(37),
+  BITFIELD(45, 2) /* index 37 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, CHILD(42),
+  BITFIELD(47, 2) /* index 42 */,
+  TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_OR, TILEPRO_OPC_MOVE,
+  BITFIELD(49, 2) /* index 47 */,
+  TILEPRO_OPC_RL, TILEPRO_OPC_SHL, TILEPRO_OPC_SHR, TILEPRO_OPC_SRA,
+  BITFIELD(49, 2) /* index 52 */,
+  TILEPRO_OPC_SLTE, TILEPRO_OPC_SLTE_U, TILEPRO_OPC_SLT, TILEPRO_OPC_SLT_U,
+  BITFIELD(49, 2) /* index 57 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_S3A, TILEPRO_OPC_SEQ, TILEPRO_OPC_SNE,
+  BITFIELD(31, 2) /* index 62 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(67),
+  BITFIELD(33, 2) /* index 67 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(72),
+  BITFIELD(35, 2) /* index 72 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(77),
+  BITFIELD(37, 2) /* index 77 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(82),
+  BITFIELD(39, 2) /* index 82 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, CHILD(87),
+  BITFIELD(41, 2) /* index 87 */,
+  TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_ANDI, TILEPRO_OPC_INFO,
+  BITFIELD(37, 2) /* index 92 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, CHILD(97),
+  BITFIELD(39, 2) /* index 97 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, CHILD(102),
+  BITFIELD(41, 2) /* index 102 */,
+  TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_ORI, TILEPRO_OPC_MOVEI,
+  BITFIELD(48, 3) /* index 107 */,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_RLI, TILEPRO_OPC_SHLI, TILEPRO_OPC_SHRI,
+  TILEPRO_OPC_SRAI, CHILD(116), TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(43, 3) /* index 116 */,
+  TILEPRO_OPC_NONE, CHILD(125), CHILD(130), CHILD(135), TILEPRO_OPC_NONE,
+  TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(46, 2) /* index 125 */,
+  TILEPRO_OPC_FNOP, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(46, 2) /* index 130 */,
+  TILEPRO_OPC_ILL, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+  BITFIELD(46, 2) /* index 135 */,
+  TILEPRO_OPC_NOP, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE, TILEPRO_OPC_NONE,
+};
+
+static const unsigned short decode_Y2_fsm[24] =
+{
+  BITFIELD(56, 3) /* index 0 */,
+  CHILD(9), TILEPRO_OPC_LB_U, TILEPRO_OPC_LH, TILEPRO_OPC_LH_U,
+  TILEPRO_OPC_LW, TILEPRO_OPC_SB, TILEPRO_OPC_SH, TILEPRO_OPC_SW,
+  BITFIELD(20, 2) /* index 9 */,
+  TILEPRO_OPC_LB, TILEPRO_OPC_LB, TILEPRO_OPC_LB, CHILD(14),
+  BITFIELD(22, 2) /* index 14 */,
+  TILEPRO_OPC_LB, TILEPRO_OPC_LB, TILEPRO_OPC_LB, CHILD(19),
+  BITFIELD(24, 2) /* index 19 */,
+  TILEPRO_OPC_LB, TILEPRO_OPC_LB, TILEPRO_OPC_LB, TILEPRO_OPC_PREFETCH,
+};
+
+#undef BITFIELD
+#undef CHILD
+const unsigned short * const
+tilepro_bundle_decoder_fsms[TILEPRO_NUM_PIPELINE_ENCODINGS] =
+{
+  decode_X0_fsm,
+  decode_X1_fsm,
+  decode_Y0_fsm,
+  decode_Y1_fsm,
+  decode_Y2_fsm
+};
+const struct tilepro_operand tilepro_operands[43] =
+{
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_IMM8_X0),
+    8, 1, 0, 0, 0, 0,
+    create_Imm8_X0, get_Imm8_X0
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_IMM8_X1),
+    8, 1, 0, 0, 0, 0,
+    create_Imm8_X1, get_Imm8_X1
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_IMM8_Y0),
+    8, 1, 0, 0, 0, 0,
+    create_Imm8_Y0, get_Imm8_Y0
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_IMM8_Y1),
+    8, 1, 0, 0, 0, 0,
+    create_Imm8_Y1, get_Imm8_Y1
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_IMM16_X0),
+    16, 1, 0, 0, 0, 0,
+    create_Imm16_X0, get_Imm16_X0
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_IMM16_X1),
+    16, 1, 0, 0, 0, 0,
+    create_Imm16_X1, get_Imm16_X1
+  },
+  {
+    TILEPRO_OP_TYPE_ADDRESS, BFD_RELOC(TILEPRO_JOFFLONG_X1),
+    29, 1, 0, 0, 1, TILEPRO_LOG2_BUNDLE_ALIGNMENT_IN_BYTES,
+    create_JOffLong_X1, get_JOffLong_X1
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 0, 1, 0, 0,
+    create_Dest_X0, get_Dest_X0
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcA_X0, get_SrcA_X0
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 0, 1, 0, 0,
+    create_Dest_X1, get_Dest_X1
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcA_X1, get_SrcA_X1
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 0, 1, 0, 0,
+    create_Dest_Y0, get_Dest_Y0
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcA_Y0, get_SrcA_Y0
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 0, 1, 0, 0,
+    create_Dest_Y1, get_Dest_Y1
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcA_Y1, get_SrcA_Y1
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcA_Y2, get_SrcA_Y2
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcB_X0, get_SrcB_X0
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcB_X1, get_SrcB_X1
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcB_Y0, get_SrcB_Y0
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcB_Y1, get_SrcB_Y1
+  },
+  {
+    TILEPRO_OP_TYPE_ADDRESS, BFD_RELOC(TILEPRO_BROFF_X1),
+    17, 1, 0, 0, 1, TILEPRO_LOG2_BUNDLE_ALIGNMENT_IN_BYTES,
+    create_BrOff_X1, get_BrOff_X1
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 1, 0, 0,
+    create_Dest_X0, get_Dest_X0
+  },
+  {
+    TILEPRO_OP_TYPE_ADDRESS, BFD_RELOC(NONE),
+    28, 1, 0, 0, 1, TILEPRO_LOG2_BUNDLE_ALIGNMENT_IN_BYTES,
+    create_JOff_X1, get_JOff_X1
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 0, 1, 0, 0,
+    create_SrcBDest_Y2, get_SrcBDest_Y2
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 1, 0, 0,
+    create_SrcA_X1, get_SrcA_X1
+  },
+  {
+    TILEPRO_OP_TYPE_SPR, BFD_RELOC(TILEPRO_MF_IMM15_X1),
+    15, 0, 0, 0, 0, 0,
+    create_MF_Imm15_X1, get_MF_Imm15_X1
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_MMSTART_X0),
+    5, 0, 0, 0, 0, 0,
+    create_MMStart_X0, get_MMStart_X0
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_MMEND_X0),
+    5, 0, 0, 0, 0, 0,
+    create_MMEnd_X0, get_MMEnd_X0
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_MMSTART_X1),
+    5, 0, 0, 0, 0, 0,
+    create_MMStart_X1, get_MMStart_X1
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_MMEND_X1),
+    5, 0, 0, 0, 0, 0,
+    create_MMEnd_X1, get_MMEnd_X1
+  },
+  {
+    TILEPRO_OP_TYPE_SPR, BFD_RELOC(TILEPRO_MT_IMM15_X1),
+    15, 0, 0, 0, 0, 0,
+    create_MT_Imm15_X1, get_MT_Imm15_X1
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 1, 0, 0,
+    create_Dest_Y0, get_Dest_Y0
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_SHAMT_X0),
+    5, 0, 0, 0, 0, 0,
+    create_ShAmt_X0, get_ShAmt_X0
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_SHAMT_X1),
+    5, 0, 0, 0, 0, 0,
+    create_ShAmt_X1, get_ShAmt_X1
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_SHAMT_Y0),
+    5, 0, 0, 0, 0, 0,
+    create_ShAmt_Y0, get_ShAmt_Y0
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_SHAMT_Y1),
+    5, 0, 0, 0, 0, 0,
+    create_ShAmt_Y1, get_ShAmt_Y1
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcBDest_Y2, get_SrcBDest_Y2
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEPRO_DEST_IMM8_X1),
+    8, 1, 0, 0, 0, 0,
+    create_Dest_Imm8_X1, get_Dest_Imm8_X1
+  },
+  {
+    TILEPRO_OP_TYPE_ADDRESS, BFD_RELOC(NONE),
+    10, 1, 0, 0, 1, TILEPRO_LOG2_SN_INSTRUCTION_SIZE_IN_BYTES,
+    create_BrOff_SN, get_BrOff_SN
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(NONE),
+    8, 0, 0, 0, 0, 0,
+    create_Imm8_SN, get_Imm8_SN
+  },
+  {
+    TILEPRO_OP_TYPE_IMMEDIATE, BFD_RELOC(NONE),
+    8, 1, 0, 0, 0, 0,
+    create_Imm8_SN, get_Imm8_SN
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    2, 0, 0, 1, 0, 0,
+    create_Dest_SN, get_Dest_SN
+  },
+  {
+    TILEPRO_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    2, 0, 1, 0, 0, 0,
+    create_Src_SN, get_Src_SN
+  }
+};
+
+
+
+
+/* Given a set of bundle bits and a specific pipe, returns which
+ * instruction the bundle contains in that pipe.
+ */
+const struct tilepro_opcode *
+find_opcode(tilepro_bundle_bits bits, tilepro_pipeline pipe)
+{
+  const unsigned short *table = tilepro_bundle_decoder_fsms[pipe];
+  int index = 0;
+
+  while (1)
+  {
+    unsigned short bitspec = table[index];
+    unsigned int bitfield =
+      ((unsigned int)(bits >> (bitspec & 63))) & (bitspec >> 6);
+
+    unsigned short next = table[index + 1 + bitfield];
+    if (next <= TILEPRO_OPC_NONE)
+      return &tilepro_opcodes[next];
+
+    index = next - TILEPRO_OPC_NONE;
+  }
+}
+
+
+int
+parse_insn_tilepro(tilepro_bundle_bits bits,
+                   unsigned int pc,
+                   struct tilepro_decoded_instruction
+                   decoded[TILEPRO_MAX_INSTRUCTIONS_PER_BUNDLE])
+{
+  int num_instructions = 0;
+  int pipe;
+
+  int min_pipe, max_pipe;
+  if ((bits & TILEPRO_BUNDLE_Y_ENCODING_MASK) == 0)
+  {
+    min_pipe = TILEPRO_PIPELINE_X0;
+    max_pipe = TILEPRO_PIPELINE_X1;
+  }
+  else
+  {
+    min_pipe = TILEPRO_PIPELINE_Y0;
+    max_pipe = TILEPRO_PIPELINE_Y2;
+  }
+
+  /* For each pipe, find an instruction that fits. */
+  for (pipe = min_pipe; pipe <= max_pipe; pipe++)
+  {
+    const struct tilepro_opcode *opc;
+    struct tilepro_decoded_instruction *d;
+    int i;
+
+    d = &decoded[num_instructions++];
+    opc = find_opcode (bits, (tilepro_pipeline)pipe);
+    d->opcode = opc;
+
+    /* Decode each operand, sign extending, etc. as appropriate. */
+    for (i = 0; i < opc->num_operands; i++)
+    {
+      const struct tilepro_operand *op =
+        &tilepro_operands[opc->operands[pipe][i]];
+      int opval = op->extract (bits);
+      if (op->is_signed)
+      {
+        /* Sign-extend the operand. */
+        int shift = (int)((sizeof(int) * 8) - op->num_bits);
+        opval = (opval << shift) >> shift;
+      }
+
+      /* Adjust PC-relative scaled branch offsets. */
+      if (op->type == TILEPRO_OP_TYPE_ADDRESS)
+      {
+        opval *= TILEPRO_BUNDLE_SIZE_IN_BYTES;
+        opval += (int)pc;
+      }
+
+      /* Record the final value. */
+      d->operands[i] = op;
+      d->operand_values[i] = opval;
+    }
+  }
+
+  return num_instructions;
+}
diff --git a/arch/tile/kernel/tile-desc_64.c b/arch/tile/kernel/tile-desc_64.c
new file mode 100644
index 000000000..65b5f8aca
--- /dev/null
+++ b/arch/tile/kernel/tile-desc_64.c
@@ -0,0 +1,2218 @@
+/* TILE-Gx opcode information.
+ *
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ *
+ *
+ *
+ *
+ */
+
+/* This define is BFD_RELOC_##x for real bfd, or -1 for everyone else. */
+#define BFD_RELOC(x) -1
+
+/* Special registers. */
+#define TREG_LR 55
+#define TREG_SN 56
+#define TREG_ZERO 63
+
+#include <linux/stddef.h>
+#include <asm/tile-desc.h>
+
+const struct tilegx_opcode tilegx_opcodes[334] =
+{
+ { "bpt", TILEGX_OPC_BPT, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "info", TILEGX_OPC_INFO, 0xf, 1, TREG_ZERO, 1,
+    { { 0 }, { 1 }, { 2 }, { 3 }, { 0, } },
+  },
+  { "infol", TILEGX_OPC_INFOL, 0x3, 1, TREG_ZERO, 1,
+    { { 4 }, { 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "move", TILEGX_OPC_MOVE, 0xf, 2, TREG_ZERO, 1,
+    { { 6, 7 }, { 8, 9 }, { 10, 11 }, { 12, 13 }, { 0, } },
+  },
+  { "movei", TILEGX_OPC_MOVEI, 0xf, 2, TREG_ZERO, 1,
+    { { 6, 0 }, { 8, 1 }, { 10, 2 }, { 12, 3 }, { 0, } },
+  },
+  { "moveli", TILEGX_OPC_MOVELI, 0x3, 2, TREG_ZERO, 1,
+    { { 6, 4 }, { 8, 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "prefetch", TILEGX_OPC_PREFETCH, 0x12, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 14 } },
+  },
+  { "prefetch_add_l1", TILEGX_OPC_PREFETCH_ADD_L1, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "prefetch_add_l1_fault", TILEGX_OPC_PREFETCH_ADD_L1_FAULT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "prefetch_add_l2", TILEGX_OPC_PREFETCH_ADD_L2, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "prefetch_add_l2_fault", TILEGX_OPC_PREFETCH_ADD_L2_FAULT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "prefetch_add_l3", TILEGX_OPC_PREFETCH_ADD_L3, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "prefetch_add_l3_fault", TILEGX_OPC_PREFETCH_ADD_L3_FAULT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "prefetch_l1", TILEGX_OPC_PREFETCH_L1, 0x12, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 14 } },
+  },
+  { "prefetch_l1_fault", TILEGX_OPC_PREFETCH_L1_FAULT, 0x12, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 14 } },
+  },
+  { "prefetch_l2", TILEGX_OPC_PREFETCH_L2, 0x12, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 14 } },
+  },
+  { "prefetch_l2_fault", TILEGX_OPC_PREFETCH_L2_FAULT, 0x12, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 14 } },
+  },
+  { "prefetch_l3", TILEGX_OPC_PREFETCH_L3, 0x12, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 14 } },
+  },
+  { "prefetch_l3_fault", TILEGX_OPC_PREFETCH_L3_FAULT, 0x12, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 14 } },
+  },
+  { "raise", TILEGX_OPC_RAISE, 0x2, 0, TREG_ZERO, 1,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "add", TILEGX_OPC_ADD, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "addi", TILEGX_OPC_ADDI, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 10, 11, 2 }, { 12, 13, 3 }, { 0, } },
+  },
+  { "addli", TILEGX_OPC_ADDLI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 4 }, { 8, 9, 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addx", TILEGX_OPC_ADDX, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "addxi", TILEGX_OPC_ADDXI, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 10, 11, 2 }, { 12, 13, 3 }, { 0, } },
+  },
+  { "addxli", TILEGX_OPC_ADDXLI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 4 }, { 8, 9, 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "addxsc", TILEGX_OPC_ADDXSC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "and", TILEGX_OPC_AND, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "andi", TILEGX_OPC_ANDI, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 10, 11, 2 }, { 12, 13, 3 }, { 0, } },
+  },
+  { "beqz", TILEGX_OPC_BEQZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "beqzt", TILEGX_OPC_BEQZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bfexts", TILEGX_OPC_BFEXTS, 0x1, 4, TREG_ZERO, 1,
+    { { 6, 7, 21, 22 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bfextu", TILEGX_OPC_BFEXTU, 0x1, 4, TREG_ZERO, 1,
+    { { 6, 7, 21, 22 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bfins", TILEGX_OPC_BFINS, 0x1, 4, TREG_ZERO, 1,
+    { { 23, 7, 21, 22 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgez", TILEGX_OPC_BGEZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgezt", TILEGX_OPC_BGEZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgtz", TILEGX_OPC_BGTZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bgtzt", TILEGX_OPC_BGTZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blbc", TILEGX_OPC_BLBC, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blbct", TILEGX_OPC_BLBCT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blbs", TILEGX_OPC_BLBS, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blbst", TILEGX_OPC_BLBST, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blez", TILEGX_OPC_BLEZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "blezt", TILEGX_OPC_BLEZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bltz", TILEGX_OPC_BLTZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bltzt", TILEGX_OPC_BLTZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bnez", TILEGX_OPC_BNEZ, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "bnezt", TILEGX_OPC_BNEZT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 20 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "clz", TILEGX_OPC_CLZ, 0x5, 2, TREG_ZERO, 1,
+    { { 6, 7 }, { 0, }, { 10, 11 }, { 0, }, { 0, } },
+  },
+  { "cmoveqz", TILEGX_OPC_CMOVEQZ, 0x5, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 24, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "cmovnez", TILEGX_OPC_CMOVNEZ, 0x5, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 24, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "cmpeq", TILEGX_OPC_CMPEQ, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "cmpeqi", TILEGX_OPC_CMPEQI, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 10, 11, 2 }, { 12, 13, 3 }, { 0, } },
+  },
+  { "cmpexch", TILEGX_OPC_CMPEXCH, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "cmpexch4", TILEGX_OPC_CMPEXCH4, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "cmples", TILEGX_OPC_CMPLES, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "cmpleu", TILEGX_OPC_CMPLEU, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "cmplts", TILEGX_OPC_CMPLTS, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "cmpltsi", TILEGX_OPC_CMPLTSI, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 10, 11, 2 }, { 12, 13, 3 }, { 0, } },
+  },
+  { "cmpltu", TILEGX_OPC_CMPLTU, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "cmpltui", TILEGX_OPC_CMPLTUI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "cmpne", TILEGX_OPC_CMPNE, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "cmul", TILEGX_OPC_CMUL, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "cmula", TILEGX_OPC_CMULA, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "cmulaf", TILEGX_OPC_CMULAF, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "cmulf", TILEGX_OPC_CMULF, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "cmulfr", TILEGX_OPC_CMULFR, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "cmulh", TILEGX_OPC_CMULH, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "cmulhr", TILEGX_OPC_CMULHR, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "crc32_32", TILEGX_OPC_CRC32_32, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "crc32_8", TILEGX_OPC_CRC32_8, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ctz", TILEGX_OPC_CTZ, 0x5, 2, TREG_ZERO, 1,
+    { { 6, 7 }, { 0, }, { 10, 11 }, { 0, }, { 0, } },
+  },
+  { "dblalign", TILEGX_OPC_DBLALIGN, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "dblalign2", TILEGX_OPC_DBLALIGN2, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "dblalign4", TILEGX_OPC_DBLALIGN4, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "dblalign6", TILEGX_OPC_DBLALIGN6, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "drain", TILEGX_OPC_DRAIN, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "dtlbpr", TILEGX_OPC_DTLBPR, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "exch", TILEGX_OPC_EXCH, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "exch4", TILEGX_OPC_EXCH4, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fdouble_add_flags", TILEGX_OPC_FDOUBLE_ADD_FLAGS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fdouble_addsub", TILEGX_OPC_FDOUBLE_ADDSUB, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fdouble_mul_flags", TILEGX_OPC_FDOUBLE_MUL_FLAGS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fdouble_pack1", TILEGX_OPC_FDOUBLE_PACK1, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fdouble_pack2", TILEGX_OPC_FDOUBLE_PACK2, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fdouble_sub_flags", TILEGX_OPC_FDOUBLE_SUB_FLAGS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fdouble_unpack_max", TILEGX_OPC_FDOUBLE_UNPACK_MAX, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fdouble_unpack_min", TILEGX_OPC_FDOUBLE_UNPACK_MIN, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fetchadd", TILEGX_OPC_FETCHADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fetchadd4", TILEGX_OPC_FETCHADD4, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fetchaddgez", TILEGX_OPC_FETCHADDGEZ, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fetchaddgez4", TILEGX_OPC_FETCHADDGEZ4, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fetchand", TILEGX_OPC_FETCHAND, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fetchand4", TILEGX_OPC_FETCHAND4, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fetchor", TILEGX_OPC_FETCHOR, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fetchor4", TILEGX_OPC_FETCHOR4, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "finv", TILEGX_OPC_FINV, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "flush", TILEGX_OPC_FLUSH, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "flushwb", TILEGX_OPC_FLUSHWB, 0x2, 0, TREG_ZERO, 1,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fnop", TILEGX_OPC_FNOP, 0xf, 0, TREG_ZERO, 1,
+    { {  }, {  }, {  }, {  }, { 0, } },
+  },
+  { "fsingle_add1", TILEGX_OPC_FSINGLE_ADD1, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fsingle_addsub2", TILEGX_OPC_FSINGLE_ADDSUB2, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fsingle_mul1", TILEGX_OPC_FSINGLE_MUL1, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fsingle_mul2", TILEGX_OPC_FSINGLE_MUL2, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fsingle_pack1", TILEGX_OPC_FSINGLE_PACK1, 0x5, 2, TREG_ZERO, 1,
+    { { 6, 7 }, { 0, }, { 10, 11 }, { 0, }, { 0, } },
+  },
+  { "fsingle_pack2", TILEGX_OPC_FSINGLE_PACK2, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "fsingle_sub1", TILEGX_OPC_FSINGLE_SUB1, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "icoh", TILEGX_OPC_ICOH, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ill", TILEGX_OPC_ILL, 0xa, 0, TREG_ZERO, 1,
+    { { 0, }, {  }, { 0, }, {  }, { 0, } },
+  },
+  { "inv", TILEGX_OPC_INV, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "iret", TILEGX_OPC_IRET, 0x2, 0, TREG_ZERO, 1,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "j", TILEGX_OPC_J, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 25 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "jal", TILEGX_OPC_JAL, 0x2, 1, TREG_LR, 1,
+    { { 0, }, { 25 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "jalr", TILEGX_OPC_JALR, 0xa, 1, TREG_LR, 1,
+    { { 0, }, { 9 }, { 0, }, { 13 }, { 0, } },
+  },
+  { "jalrp", TILEGX_OPC_JALRP, 0xa, 1, TREG_LR, 1,
+    { { 0, }, { 9 }, { 0, }, { 13 }, { 0, } },
+  },
+  { "jr", TILEGX_OPC_JR, 0xa, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 13 }, { 0, } },
+  },
+  { "jrp", TILEGX_OPC_JRP, 0xa, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 13 }, { 0, } },
+  },
+  { "ld", TILEGX_OPC_LD, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 26, 14 } },
+  },
+  { "ld1s", TILEGX_OPC_LD1S, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 26, 14 } },
+  },
+  { "ld1s_add", TILEGX_OPC_LD1S_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ld1u", TILEGX_OPC_LD1U, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 26, 14 } },
+  },
+  { "ld1u_add", TILEGX_OPC_LD1U_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ld2s", TILEGX_OPC_LD2S, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 26, 14 } },
+  },
+  { "ld2s_add", TILEGX_OPC_LD2S_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ld2u", TILEGX_OPC_LD2U, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 26, 14 } },
+  },
+  { "ld2u_add", TILEGX_OPC_LD2U_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ld4s", TILEGX_OPC_LD4S, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 26, 14 } },
+  },
+  { "ld4s_add", TILEGX_OPC_LD4S_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ld4u", TILEGX_OPC_LD4U, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 26, 14 } },
+  },
+  { "ld4u_add", TILEGX_OPC_LD4U_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ld_add", TILEGX_OPC_LD_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldna", TILEGX_OPC_LDNA, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldna_add", TILEGX_OPC_LDNA_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt", TILEGX_OPC_LDNT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt1s", TILEGX_OPC_LDNT1S, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt1s_add", TILEGX_OPC_LDNT1S_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt1u", TILEGX_OPC_LDNT1U, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt1u_add", TILEGX_OPC_LDNT1U_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt2s", TILEGX_OPC_LDNT2S, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt2s_add", TILEGX_OPC_LDNT2S_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt2u", TILEGX_OPC_LDNT2U, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt2u_add", TILEGX_OPC_LDNT2U_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt4s", TILEGX_OPC_LDNT4S, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt4s_add", TILEGX_OPC_LDNT4S_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt4u", TILEGX_OPC_LDNT4U, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt4u_add", TILEGX_OPC_LDNT4U_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "ldnt_add", TILEGX_OPC_LDNT_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 8, 15, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "lnk", TILEGX_OPC_LNK, 0xa, 1, TREG_ZERO, 1,
+    { { 0, }, { 8 }, { 0, }, { 12 }, { 0, } },
+  },
+  { "mf", TILEGX_OPC_MF, 0x2, 0, TREG_ZERO, 1,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mfspr", TILEGX_OPC_MFSPR, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 8, 27 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mm", TILEGX_OPC_MM, 0x1, 4, TREG_ZERO, 1,
+    { { 23, 7, 21, 22 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mnz", TILEGX_OPC_MNZ, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "mtspr", TILEGX_OPC_MTSPR, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 28, 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mul_hs_hs", TILEGX_OPC_MUL_HS_HS, 0x5, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 10, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "mul_hs_hu", TILEGX_OPC_MUL_HS_HU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mul_hs_ls", TILEGX_OPC_MUL_HS_LS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mul_hs_lu", TILEGX_OPC_MUL_HS_LU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mul_hu_hu", TILEGX_OPC_MUL_HU_HU, 0x5, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 10, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "mul_hu_ls", TILEGX_OPC_MUL_HU_LS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mul_hu_lu", TILEGX_OPC_MUL_HU_LU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mul_ls_ls", TILEGX_OPC_MUL_LS_LS, 0x5, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 10, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "mul_ls_lu", TILEGX_OPC_MUL_LS_LU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mul_lu_lu", TILEGX_OPC_MUL_LU_LU, 0x5, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 10, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "mula_hs_hs", TILEGX_OPC_MULA_HS_HS, 0x5, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 24, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "mula_hs_hu", TILEGX_OPC_MULA_HS_HU, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mula_hs_ls", TILEGX_OPC_MULA_HS_LS, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mula_hs_lu", TILEGX_OPC_MULA_HS_LU, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mula_hu_hu", TILEGX_OPC_MULA_HU_HU, 0x5, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 24, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "mula_hu_ls", TILEGX_OPC_MULA_HU_LS, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mula_hu_lu", TILEGX_OPC_MULA_HU_LU, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mula_ls_ls", TILEGX_OPC_MULA_LS_LS, 0x5, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 24, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "mula_ls_lu", TILEGX_OPC_MULA_LS_LU, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "mula_lu_lu", TILEGX_OPC_MULA_LU_LU, 0x5, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 24, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "mulax", TILEGX_OPC_MULAX, 0x5, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 24, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "mulx", TILEGX_OPC_MULX, 0x5, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 10, 11, 18 }, { 0, }, { 0, } },
+  },
+  { "mz", TILEGX_OPC_MZ, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "nap", TILEGX_OPC_NAP, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "nop", TILEGX_OPC_NOP, 0xf, 0, TREG_ZERO, 1,
+    { {  }, {  }, {  }, {  }, { 0, } },
+  },
+  { "nor", TILEGX_OPC_NOR, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "or", TILEGX_OPC_OR, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "ori", TILEGX_OPC_ORI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "pcnt", TILEGX_OPC_PCNT, 0x5, 2, TREG_ZERO, 1,
+    { { 6, 7 }, { 0, }, { 10, 11 }, { 0, }, { 0, } },
+  },
+  { "revbits", TILEGX_OPC_REVBITS, 0x5, 2, TREG_ZERO, 1,
+    { { 6, 7 }, { 0, }, { 10, 11 }, { 0, }, { 0, } },
+  },
+  { "revbytes", TILEGX_OPC_REVBYTES, 0x5, 2, TREG_ZERO, 1,
+    { { 6, 7 }, { 0, }, { 10, 11 }, { 0, }, { 0, } },
+  },
+  { "rotl", TILEGX_OPC_ROTL, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "rotli", TILEGX_OPC_ROTLI, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 10, 11, 31 }, { 12, 13, 32 }, { 0, } },
+  },
+  { "shl", TILEGX_OPC_SHL, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "shl16insli", TILEGX_OPC_SHL16INSLI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 4 }, { 8, 9, 5 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shl1add", TILEGX_OPC_SHL1ADD, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "shl1addx", TILEGX_OPC_SHL1ADDX, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "shl2add", TILEGX_OPC_SHL2ADD, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "shl2addx", TILEGX_OPC_SHL2ADDX, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "shl3add", TILEGX_OPC_SHL3ADD, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "shl3addx", TILEGX_OPC_SHL3ADDX, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "shli", TILEGX_OPC_SHLI, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 10, 11, 31 }, { 12, 13, 32 }, { 0, } },
+  },
+  { "shlx", TILEGX_OPC_SHLX, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shlxi", TILEGX_OPC_SHLXI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shrs", TILEGX_OPC_SHRS, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "shrsi", TILEGX_OPC_SHRSI, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 10, 11, 31 }, { 12, 13, 32 }, { 0, } },
+  },
+  { "shru", TILEGX_OPC_SHRU, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "shrui", TILEGX_OPC_SHRUI, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 10, 11, 31 }, { 12, 13, 32 }, { 0, } },
+  },
+  { "shrux", TILEGX_OPC_SHRUX, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shruxi", TILEGX_OPC_SHRUXI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "shufflebytes", TILEGX_OPC_SHUFFLEBYTES, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "st", TILEGX_OPC_ST, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 17 }, { 0, }, { 0, }, { 14, 33 } },
+  },
+  { "st1", TILEGX_OPC_ST1, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 17 }, { 0, }, { 0, }, { 14, 33 } },
+  },
+  { "st1_add", TILEGX_OPC_ST1_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 15, 17, 34 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "st2", TILEGX_OPC_ST2, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 17 }, { 0, }, { 0, }, { 14, 33 } },
+  },
+  { "st2_add", TILEGX_OPC_ST2_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 15, 17, 34 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "st4", TILEGX_OPC_ST4, 0x12, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 17 }, { 0, }, { 0, }, { 14, 33 } },
+  },
+  { "st4_add", TILEGX_OPC_ST4_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 15, 17, 34 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "st_add", TILEGX_OPC_ST_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 15, 17, 34 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "stnt", TILEGX_OPC_STNT, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "stnt1", TILEGX_OPC_STNT1, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "stnt1_add", TILEGX_OPC_STNT1_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 15, 17, 34 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "stnt2", TILEGX_OPC_STNT2, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "stnt2_add", TILEGX_OPC_STNT2_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 15, 17, 34 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "stnt4", TILEGX_OPC_STNT4, 0x2, 2, TREG_ZERO, 1,
+    { { 0, }, { 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "stnt4_add", TILEGX_OPC_STNT4_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 15, 17, 34 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "stnt_add", TILEGX_OPC_STNT_ADD, 0x2, 3, TREG_ZERO, 1,
+    { { 0, }, { 15, 17, 34 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "sub", TILEGX_OPC_SUB, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "subx", TILEGX_OPC_SUBX, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "subxsc", TILEGX_OPC_SUBXSC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "swint0", TILEGX_OPC_SWINT0, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "swint1", TILEGX_OPC_SWINT1, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "swint2", TILEGX_OPC_SWINT2, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "swint3", TILEGX_OPC_SWINT3, 0x2, 0, TREG_ZERO, 0,
+    { { 0, }, {  }, { 0, }, { 0, }, { 0, } },
+  },
+  { "tblidxb0", TILEGX_OPC_TBLIDXB0, 0x5, 2, TREG_ZERO, 1,
+    { { 23, 7 }, { 0, }, { 24, 11 }, { 0, }, { 0, } },
+  },
+  { "tblidxb1", TILEGX_OPC_TBLIDXB1, 0x5, 2, TREG_ZERO, 1,
+    { { 23, 7 }, { 0, }, { 24, 11 }, { 0, }, { 0, } },
+  },
+  { "tblidxb2", TILEGX_OPC_TBLIDXB2, 0x5, 2, TREG_ZERO, 1,
+    { { 23, 7 }, { 0, }, { 24, 11 }, { 0, }, { 0, } },
+  },
+  { "tblidxb3", TILEGX_OPC_TBLIDXB3, 0x5, 2, TREG_ZERO, 1,
+    { { 23, 7 }, { 0, }, { 24, 11 }, { 0, }, { 0, } },
+  },
+  { "v1add", TILEGX_OPC_V1ADD, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1addi", TILEGX_OPC_V1ADDI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1adduc", TILEGX_OPC_V1ADDUC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1adiffu", TILEGX_OPC_V1ADIFFU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1avgu", TILEGX_OPC_V1AVGU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1cmpeq", TILEGX_OPC_V1CMPEQ, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1cmpeqi", TILEGX_OPC_V1CMPEQI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1cmples", TILEGX_OPC_V1CMPLES, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1cmpleu", TILEGX_OPC_V1CMPLEU, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1cmplts", TILEGX_OPC_V1CMPLTS, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1cmpltsi", TILEGX_OPC_V1CMPLTSI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1cmpltu", TILEGX_OPC_V1CMPLTU, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1cmpltui", TILEGX_OPC_V1CMPLTUI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1cmpne", TILEGX_OPC_V1CMPNE, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1ddotpu", TILEGX_OPC_V1DDOTPU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1ddotpua", TILEGX_OPC_V1DDOTPUA, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1ddotpus", TILEGX_OPC_V1DDOTPUS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1ddotpusa", TILEGX_OPC_V1DDOTPUSA, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1dotp", TILEGX_OPC_V1DOTP, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1dotpa", TILEGX_OPC_V1DOTPA, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1dotpu", TILEGX_OPC_V1DOTPU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1dotpua", TILEGX_OPC_V1DOTPUA, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1dotpus", TILEGX_OPC_V1DOTPUS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1dotpusa", TILEGX_OPC_V1DOTPUSA, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1int_h", TILEGX_OPC_V1INT_H, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1int_l", TILEGX_OPC_V1INT_L, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1maxu", TILEGX_OPC_V1MAXU, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1maxui", TILEGX_OPC_V1MAXUI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1minu", TILEGX_OPC_V1MINU, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1minui", TILEGX_OPC_V1MINUI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1mnz", TILEGX_OPC_V1MNZ, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1multu", TILEGX_OPC_V1MULTU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1mulu", TILEGX_OPC_V1MULU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1mulus", TILEGX_OPC_V1MULUS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1mz", TILEGX_OPC_V1MZ, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1sadau", TILEGX_OPC_V1SADAU, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1sadu", TILEGX_OPC_V1SADU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1shl", TILEGX_OPC_V1SHL, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1shli", TILEGX_OPC_V1SHLI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1shrs", TILEGX_OPC_V1SHRS, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1shrsi", TILEGX_OPC_V1SHRSI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1shru", TILEGX_OPC_V1SHRU, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1shrui", TILEGX_OPC_V1SHRUI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1sub", TILEGX_OPC_V1SUB, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v1subuc", TILEGX_OPC_V1SUBUC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2add", TILEGX_OPC_V2ADD, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2addi", TILEGX_OPC_V2ADDI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2addsc", TILEGX_OPC_V2ADDSC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2adiffs", TILEGX_OPC_V2ADIFFS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2avgs", TILEGX_OPC_V2AVGS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2cmpeq", TILEGX_OPC_V2CMPEQ, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2cmpeqi", TILEGX_OPC_V2CMPEQI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2cmples", TILEGX_OPC_V2CMPLES, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2cmpleu", TILEGX_OPC_V2CMPLEU, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2cmplts", TILEGX_OPC_V2CMPLTS, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2cmpltsi", TILEGX_OPC_V2CMPLTSI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2cmpltu", TILEGX_OPC_V2CMPLTU, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2cmpltui", TILEGX_OPC_V2CMPLTUI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2cmpne", TILEGX_OPC_V2CMPNE, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2dotp", TILEGX_OPC_V2DOTP, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2dotpa", TILEGX_OPC_V2DOTPA, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2int_h", TILEGX_OPC_V2INT_H, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2int_l", TILEGX_OPC_V2INT_L, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2maxs", TILEGX_OPC_V2MAXS, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2maxsi", TILEGX_OPC_V2MAXSI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2mins", TILEGX_OPC_V2MINS, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2minsi", TILEGX_OPC_V2MINSI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2mnz", TILEGX_OPC_V2MNZ, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2mulfsc", TILEGX_OPC_V2MULFSC, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2muls", TILEGX_OPC_V2MULS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2mults", TILEGX_OPC_V2MULTS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2mz", TILEGX_OPC_V2MZ, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2packh", TILEGX_OPC_V2PACKH, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2packl", TILEGX_OPC_V2PACKL, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2packuc", TILEGX_OPC_V2PACKUC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2sadas", TILEGX_OPC_V2SADAS, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2sadau", TILEGX_OPC_V2SADAU, 0x1, 3, TREG_ZERO, 1,
+    { { 23, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2sads", TILEGX_OPC_V2SADS, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2sadu", TILEGX_OPC_V2SADU, 0x1, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 0, }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2shl", TILEGX_OPC_V2SHL, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2shli", TILEGX_OPC_V2SHLI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2shlsc", TILEGX_OPC_V2SHLSC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2shrs", TILEGX_OPC_V2SHRS, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2shrsi", TILEGX_OPC_V2SHRSI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2shru", TILEGX_OPC_V2SHRU, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2shrui", TILEGX_OPC_V2SHRUI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 29 }, { 8, 9, 30 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2sub", TILEGX_OPC_V2SUB, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v2subsc", TILEGX_OPC_V2SUBSC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v4add", TILEGX_OPC_V4ADD, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v4addsc", TILEGX_OPC_V4ADDSC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v4int_h", TILEGX_OPC_V4INT_H, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v4int_l", TILEGX_OPC_V4INT_L, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v4packsc", TILEGX_OPC_V4PACKSC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v4shl", TILEGX_OPC_V4SHL, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v4shlsc", TILEGX_OPC_V4SHLSC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v4shrs", TILEGX_OPC_V4SHRS, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v4shru", TILEGX_OPC_V4SHRU, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v4sub", TILEGX_OPC_V4SUB, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "v4subsc", TILEGX_OPC_V4SUBSC, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "wh64", TILEGX_OPC_WH64, 0x2, 1, TREG_ZERO, 1,
+    { { 0, }, { 9 }, { 0, }, { 0, }, { 0, } },
+  },
+  { "xor", TILEGX_OPC_XOR, 0xf, 3, TREG_ZERO, 1,
+    { { 6, 7, 16 }, { 8, 9, 17 }, { 10, 11, 18 }, { 12, 13, 19 }, { 0, } },
+  },
+  { "xori", TILEGX_OPC_XORI, 0x3, 3, TREG_ZERO, 1,
+    { { 6, 7, 0 }, { 8, 9, 1 }, { 0, }, { 0, }, { 0, } },
+  },
+  { NULL, TILEGX_OPC_NONE, 0, 0, TREG_ZERO, 0, { { 0, } },
+  }
+};
+#define BITFIELD(start, size) ((start) | (((1 << (size)) - 1) << 6))
+#define CHILD(array_index) (TILEGX_OPC_NONE + (array_index))
+
+static const unsigned short decode_X0_fsm[936] =
+{
+  BITFIELD(22, 9) /* index 0 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_BFEXTS,
+  TILEGX_OPC_BFEXTS, TILEGX_OPC_BFEXTS, TILEGX_OPC_BFEXTS, TILEGX_OPC_BFEXTU,
+  TILEGX_OPC_BFEXTU, TILEGX_OPC_BFEXTU, TILEGX_OPC_BFEXTU, TILEGX_OPC_BFINS,
+  TILEGX_OPC_BFINS, TILEGX_OPC_BFINS, TILEGX_OPC_BFINS, TILEGX_OPC_MM,
+  TILEGX_OPC_MM, TILEGX_OPC_MM, TILEGX_OPC_MM, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, CHILD(528), CHILD(578),
+  CHILD(583), CHILD(588), CHILD(593), CHILD(598), TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, CHILD(603), CHILD(620), CHILD(637), CHILD(654), CHILD(671),
+  CHILD(703), CHILD(797), CHILD(814), CHILD(831), CHILD(848), CHILD(865),
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, CHILD(889), TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906),
+  CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906),
+  CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906),
+  CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906),
+  CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906),
+  CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906),
+  CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906),
+  CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906),
+  CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906),
+  CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906),
+  CHILD(906), CHILD(906), CHILD(906), CHILD(906), CHILD(906),
+  BITFIELD(6, 2) /* index 513 */,
+  TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, CHILD(518),
+  BITFIELD(8, 2) /* index 518 */,
+  TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, CHILD(523),
+  BITFIELD(10, 2) /* index 523 */,
+  TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, TILEGX_OPC_MOVELI,
+  BITFIELD(20, 2) /* index 528 */,
+  TILEGX_OPC_NONE, CHILD(533), TILEGX_OPC_ADDXI, CHILD(548),
+  BITFIELD(6, 2) /* index 533 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, CHILD(538),
+  BITFIELD(8, 2) /* index 538 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, CHILD(543),
+  BITFIELD(10, 2) /* index 543 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_MOVEI,
+  BITFIELD(0, 2) /* index 548 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(553),
+  BITFIELD(2, 2) /* index 553 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(558),
+  BITFIELD(4, 2) /* index 558 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(563),
+  BITFIELD(6, 2) /* index 563 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(568),
+  BITFIELD(8, 2) /* index 568 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(573),
+  BITFIELD(10, 2) /* index 573 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_INFO,
+  BITFIELD(20, 2) /* index 578 */,
+  TILEGX_OPC_CMPEQI, TILEGX_OPC_CMPLTSI, TILEGX_OPC_CMPLTUI, TILEGX_OPC_ORI,
+  BITFIELD(20, 2) /* index 583 */,
+  TILEGX_OPC_V1ADDI, TILEGX_OPC_V1CMPEQI, TILEGX_OPC_V1CMPLTSI,
+  TILEGX_OPC_V1CMPLTUI,
+  BITFIELD(20, 2) /* index 588 */,
+  TILEGX_OPC_V1MAXUI, TILEGX_OPC_V1MINUI, TILEGX_OPC_V2ADDI,
+  TILEGX_OPC_V2CMPEQI,
+  BITFIELD(20, 2) /* index 593 */,
+  TILEGX_OPC_V2CMPLTSI, TILEGX_OPC_V2CMPLTUI, TILEGX_OPC_V2MAXSI,
+  TILEGX_OPC_V2MINSI,
+  BITFIELD(20, 2) /* index 598 */,
+  TILEGX_OPC_XORI, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(18, 4) /* index 603 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_ADDXSC, TILEGX_OPC_ADDX, TILEGX_OPC_ADD,
+  TILEGX_OPC_AND, TILEGX_OPC_CMOVEQZ, TILEGX_OPC_CMOVNEZ, TILEGX_OPC_CMPEQ,
+  TILEGX_OPC_CMPLES, TILEGX_OPC_CMPLEU, TILEGX_OPC_CMPLTS, TILEGX_OPC_CMPLTU,
+  TILEGX_OPC_CMPNE, TILEGX_OPC_CMULAF, TILEGX_OPC_CMULA, TILEGX_OPC_CMULFR,
+  BITFIELD(18, 4) /* index 620 */,
+  TILEGX_OPC_CMULF, TILEGX_OPC_CMULHR, TILEGX_OPC_CMULH, TILEGX_OPC_CMUL,
+  TILEGX_OPC_CRC32_32, TILEGX_OPC_CRC32_8, TILEGX_OPC_DBLALIGN2,
+  TILEGX_OPC_DBLALIGN4, TILEGX_OPC_DBLALIGN6, TILEGX_OPC_DBLALIGN,
+  TILEGX_OPC_FDOUBLE_ADDSUB, TILEGX_OPC_FDOUBLE_ADD_FLAGS,
+  TILEGX_OPC_FDOUBLE_MUL_FLAGS, TILEGX_OPC_FDOUBLE_PACK1,
+  TILEGX_OPC_FDOUBLE_PACK2, TILEGX_OPC_FDOUBLE_SUB_FLAGS,
+  BITFIELD(18, 4) /* index 637 */,
+  TILEGX_OPC_FDOUBLE_UNPACK_MAX, TILEGX_OPC_FDOUBLE_UNPACK_MIN,
+  TILEGX_OPC_FSINGLE_ADD1, TILEGX_OPC_FSINGLE_ADDSUB2,
+  TILEGX_OPC_FSINGLE_MUL1, TILEGX_OPC_FSINGLE_MUL2, TILEGX_OPC_FSINGLE_PACK2,
+  TILEGX_OPC_FSINGLE_SUB1, TILEGX_OPC_MNZ, TILEGX_OPC_MULAX,
+  TILEGX_OPC_MULA_HS_HS, TILEGX_OPC_MULA_HS_HU, TILEGX_OPC_MULA_HS_LS,
+  TILEGX_OPC_MULA_HS_LU, TILEGX_OPC_MULA_HU_HU, TILEGX_OPC_MULA_HU_LS,
+  BITFIELD(18, 4) /* index 654 */,
+  TILEGX_OPC_MULA_HU_LU, TILEGX_OPC_MULA_LS_LS, TILEGX_OPC_MULA_LS_LU,
+  TILEGX_OPC_MULA_LU_LU, TILEGX_OPC_MULX, TILEGX_OPC_MUL_HS_HS,
+  TILEGX_OPC_MUL_HS_HU, TILEGX_OPC_MUL_HS_LS, TILEGX_OPC_MUL_HS_LU,
+  TILEGX_OPC_MUL_HU_HU, TILEGX_OPC_MUL_HU_LS, TILEGX_OPC_MUL_HU_LU,
+  TILEGX_OPC_MUL_LS_LS, TILEGX_OPC_MUL_LS_LU, TILEGX_OPC_MUL_LU_LU,
+  TILEGX_OPC_MZ,
+  BITFIELD(18, 4) /* index 671 */,
+  TILEGX_OPC_NOR, CHILD(688), TILEGX_OPC_ROTL, TILEGX_OPC_SHL1ADDX,
+  TILEGX_OPC_SHL1ADD, TILEGX_OPC_SHL2ADDX, TILEGX_OPC_SHL2ADD,
+  TILEGX_OPC_SHL3ADDX, TILEGX_OPC_SHL3ADD, TILEGX_OPC_SHLX, TILEGX_OPC_SHL,
+  TILEGX_OPC_SHRS, TILEGX_OPC_SHRUX, TILEGX_OPC_SHRU, TILEGX_OPC_SHUFFLEBYTES,
+  TILEGX_OPC_SUBXSC,
+  BITFIELD(12, 2) /* index 688 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, CHILD(693),
+  BITFIELD(14, 2) /* index 693 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, CHILD(698),
+  BITFIELD(16, 2) /* index 698 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_MOVE,
+  BITFIELD(18, 4) /* index 703 */,
+  TILEGX_OPC_SUBX, TILEGX_OPC_SUB, CHILD(720), TILEGX_OPC_V1ADDUC,
+  TILEGX_OPC_V1ADD, TILEGX_OPC_V1ADIFFU, TILEGX_OPC_V1AVGU,
+  TILEGX_OPC_V1CMPEQ, TILEGX_OPC_V1CMPLES, TILEGX_OPC_V1CMPLEU,
+  TILEGX_OPC_V1CMPLTS, TILEGX_OPC_V1CMPLTU, TILEGX_OPC_V1CMPNE,
+  TILEGX_OPC_V1DDOTPUSA, TILEGX_OPC_V1DDOTPUS, TILEGX_OPC_V1DOTPA,
+  BITFIELD(12, 4) /* index 720 */,
+  TILEGX_OPC_NONE, CHILD(737), CHILD(742), CHILD(747), CHILD(752), CHILD(757),
+  CHILD(762), CHILD(767), CHILD(772), CHILD(777), CHILD(782), CHILD(787),
+  CHILD(792), TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 737 */,
+  TILEGX_OPC_CLZ, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 742 */,
+  TILEGX_OPC_CTZ, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 747 */,
+  TILEGX_OPC_FNOP, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 752 */,
+  TILEGX_OPC_FSINGLE_PACK1, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 757 */,
+  TILEGX_OPC_NOP, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 762 */,
+  TILEGX_OPC_PCNT, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 767 */,
+  TILEGX_OPC_REVBITS, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 772 */,
+  TILEGX_OPC_REVBYTES, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 777 */,
+  TILEGX_OPC_TBLIDXB0, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 782 */,
+  TILEGX_OPC_TBLIDXB1, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 787 */,
+  TILEGX_OPC_TBLIDXB2, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(16, 2) /* index 792 */,
+  TILEGX_OPC_TBLIDXB3, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(18, 4) /* index 797 */,
+  TILEGX_OPC_V1DOTPUSA, TILEGX_OPC_V1DOTPUS, TILEGX_OPC_V1DOTP,
+  TILEGX_OPC_V1INT_H, TILEGX_OPC_V1INT_L, TILEGX_OPC_V1MAXU,
+  TILEGX_OPC_V1MINU, TILEGX_OPC_V1MNZ, TILEGX_OPC_V1MULTU, TILEGX_OPC_V1MULUS,
+  TILEGX_OPC_V1MULU, TILEGX_OPC_V1MZ, TILEGX_OPC_V1SADAU, TILEGX_OPC_V1SADU,
+  TILEGX_OPC_V1SHL, TILEGX_OPC_V1SHRS,
+  BITFIELD(18, 4) /* index 814 */,
+  TILEGX_OPC_V1SHRU, TILEGX_OPC_V1SUBUC, TILEGX_OPC_V1SUB, TILEGX_OPC_V2ADDSC,
+  TILEGX_OPC_V2ADD, TILEGX_OPC_V2ADIFFS, TILEGX_OPC_V2AVGS,
+  TILEGX_OPC_V2CMPEQ, TILEGX_OPC_V2CMPLES, TILEGX_OPC_V2CMPLEU,
+  TILEGX_OPC_V2CMPLTS, TILEGX_OPC_V2CMPLTU, TILEGX_OPC_V2CMPNE,
+  TILEGX_OPC_V2DOTPA, TILEGX_OPC_V2DOTP, TILEGX_OPC_V2INT_H,
+  BITFIELD(18, 4) /* index 831 */,
+  TILEGX_OPC_V2INT_L, TILEGX_OPC_V2MAXS, TILEGX_OPC_V2MINS, TILEGX_OPC_V2MNZ,
+  TILEGX_OPC_V2MULFSC, TILEGX_OPC_V2MULS, TILEGX_OPC_V2MULTS, TILEGX_OPC_V2MZ,
+  TILEGX_OPC_V2PACKH, TILEGX_OPC_V2PACKL, TILEGX_OPC_V2PACKUC,
+  TILEGX_OPC_V2SADAS, TILEGX_OPC_V2SADAU, TILEGX_OPC_V2SADS,
+  TILEGX_OPC_V2SADU, TILEGX_OPC_V2SHLSC,
+  BITFIELD(18, 4) /* index 848 */,
+  TILEGX_OPC_V2SHL, TILEGX_OPC_V2SHRS, TILEGX_OPC_V2SHRU, TILEGX_OPC_V2SUBSC,
+  TILEGX_OPC_V2SUB, TILEGX_OPC_V4ADDSC, TILEGX_OPC_V4ADD, TILEGX_OPC_V4INT_H,
+  TILEGX_OPC_V4INT_L, TILEGX_OPC_V4PACKSC, TILEGX_OPC_V4SHLSC,
+  TILEGX_OPC_V4SHL, TILEGX_OPC_V4SHRS, TILEGX_OPC_V4SHRU, TILEGX_OPC_V4SUBSC,
+  TILEGX_OPC_V4SUB,
+  BITFIELD(18, 3) /* index 865 */,
+  CHILD(874), CHILD(877), CHILD(880), CHILD(883), CHILD(886), TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(21, 1) /* index 874 */,
+  TILEGX_OPC_XOR, TILEGX_OPC_NONE,
+  BITFIELD(21, 1) /* index 877 */,
+  TILEGX_OPC_V1DDOTPUA, TILEGX_OPC_NONE,
+  BITFIELD(21, 1) /* index 880 */,
+  TILEGX_OPC_V1DDOTPU, TILEGX_OPC_NONE,
+  BITFIELD(21, 1) /* index 883 */,
+  TILEGX_OPC_V1DOTPUA, TILEGX_OPC_NONE,
+  BITFIELD(21, 1) /* index 886 */,
+  TILEGX_OPC_V1DOTPU, TILEGX_OPC_NONE,
+  BITFIELD(18, 4) /* index 889 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_ROTLI, TILEGX_OPC_SHLI, TILEGX_OPC_SHLXI,
+  TILEGX_OPC_SHRSI, TILEGX_OPC_SHRUI, TILEGX_OPC_SHRUXI, TILEGX_OPC_V1SHLI,
+  TILEGX_OPC_V1SHRSI, TILEGX_OPC_V1SHRUI, TILEGX_OPC_V2SHLI,
+  TILEGX_OPC_V2SHRSI, TILEGX_OPC_V2SHRUI, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE,
+  BITFIELD(0, 2) /* index 906 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  CHILD(911),
+  BITFIELD(2, 2) /* index 911 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  CHILD(916),
+  BITFIELD(4, 2) /* index 916 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  CHILD(921),
+  BITFIELD(6, 2) /* index 921 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  CHILD(926),
+  BITFIELD(8, 2) /* index 926 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  CHILD(931),
+  BITFIELD(10, 2) /* index 931 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  TILEGX_OPC_INFOL,
+};
+
+static const unsigned short decode_X1_fsm[1206] =
+{
+  BITFIELD(53, 9) /* index 0 */,
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513), CHILD(513),
+  CHILD(513), CHILD(513), CHILD(513), CHILD(513), TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI,
+  TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_ADDXLI, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_BEQZT,
+  TILEGX_OPC_BEQZT, TILEGX_OPC_BEQZ, TILEGX_OPC_BEQZ, TILEGX_OPC_BGEZT,
+  TILEGX_OPC_BGEZT, TILEGX_OPC_BGEZ, TILEGX_OPC_BGEZ, TILEGX_OPC_BGTZT,
+  TILEGX_OPC_BGTZT, TILEGX_OPC_BGTZ, TILEGX_OPC_BGTZ, TILEGX_OPC_BLBCT,
+  TILEGX_OPC_BLBCT, TILEGX_OPC_BLBC, TILEGX_OPC_BLBC, TILEGX_OPC_BLBST,
+  TILEGX_OPC_BLBST, TILEGX_OPC_BLBS, TILEGX_OPC_BLBS, TILEGX_OPC_BLEZT,
+  TILEGX_OPC_BLEZT, TILEGX_OPC_BLEZ, TILEGX_OPC_BLEZ, TILEGX_OPC_BLTZT,
+  TILEGX_OPC_BLTZT, TILEGX_OPC_BLTZ, TILEGX_OPC_BLTZ, TILEGX_OPC_BNEZT,
+  TILEGX_OPC_BNEZT, TILEGX_OPC_BNEZ, TILEGX_OPC_BNEZ, CHILD(528), CHILD(578),
+  CHILD(598), CHILD(663), CHILD(683), CHILD(688), CHILD(693), CHILD(698),
+  CHILD(703), CHILD(708), CHILD(713), CHILD(718), TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_JAL,
+  TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL,
+  TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL,
+  TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL,
+  TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL,
+  TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL,
+  TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL,
+  TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL,
+  TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_JAL, TILEGX_OPC_J, TILEGX_OPC_J,
+  TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J,
+  TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J,
+  TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J,
+  TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J,
+  TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J,
+  TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J, TILEGX_OPC_J,
+  CHILD(723), CHILD(740), CHILD(772), CHILD(789), CHILD(1108), CHILD(1125),
+  CHILD(1142), TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, CHILD(1159), TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176), CHILD(1176),
+  CHILD(1176),
+  BITFIELD(37, 2) /* index 513 */,
+  TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, CHILD(518),
+  BITFIELD(39, 2) /* index 518 */,
+  TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, CHILD(523),
+  BITFIELD(41, 2) /* index 523 */,
+  TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, TILEGX_OPC_ADDLI, TILEGX_OPC_MOVELI,
+  BITFIELD(51, 2) /* index 528 */,
+  TILEGX_OPC_NONE, CHILD(533), TILEGX_OPC_ADDXI, CHILD(548),
+  BITFIELD(37, 2) /* index 533 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, CHILD(538),
+  BITFIELD(39, 2) /* index 538 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, CHILD(543),
+  BITFIELD(41, 2) /* index 543 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_MOVEI,
+  BITFIELD(31, 2) /* index 548 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(553),
+  BITFIELD(33, 2) /* index 553 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(558),
+  BITFIELD(35, 2) /* index 558 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(563),
+  BITFIELD(37, 2) /* index 563 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(568),
+  BITFIELD(39, 2) /* index 568 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(573),
+  BITFIELD(41, 2) /* index 573 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_INFO,
+  BITFIELD(51, 2) /* index 578 */,
+  TILEGX_OPC_CMPEQI, TILEGX_OPC_CMPLTSI, TILEGX_OPC_CMPLTUI, CHILD(583),
+  BITFIELD(31, 2) /* index 583 */,
+  TILEGX_OPC_LD1S_ADD, TILEGX_OPC_LD1S_ADD, TILEGX_OPC_LD1S_ADD, CHILD(588),
+  BITFIELD(33, 2) /* index 588 */,
+  TILEGX_OPC_LD1S_ADD, TILEGX_OPC_LD1S_ADD, TILEGX_OPC_LD1S_ADD, CHILD(593),
+  BITFIELD(35, 2) /* index 593 */,
+  TILEGX_OPC_LD1S_ADD, TILEGX_OPC_LD1S_ADD, TILEGX_OPC_LD1S_ADD,
+  TILEGX_OPC_PREFETCH_ADD_L1_FAULT,
+  BITFIELD(51, 2) /* index 598 */,
+  CHILD(603), CHILD(618), CHILD(633), CHILD(648),
+  BITFIELD(31, 2) /* index 603 */,
+  TILEGX_OPC_LD1U_ADD, TILEGX_OPC_LD1U_ADD, TILEGX_OPC_LD1U_ADD, CHILD(608),
+  BITFIELD(33, 2) /* index 608 */,
+  TILEGX_OPC_LD1U_ADD, TILEGX_OPC_LD1U_ADD, TILEGX_OPC_LD1U_ADD, CHILD(613),
+  BITFIELD(35, 2) /* index 613 */,
+  TILEGX_OPC_LD1U_ADD, TILEGX_OPC_LD1U_ADD, TILEGX_OPC_LD1U_ADD,
+  TILEGX_OPC_PREFETCH_ADD_L1,
+  BITFIELD(31, 2) /* index 618 */,
+  TILEGX_OPC_LD2S_ADD, TILEGX_OPC_LD2S_ADD, TILEGX_OPC_LD2S_ADD, CHILD(623),
+  BITFIELD(33, 2) /* index 623 */,
+  TILEGX_OPC_LD2S_ADD, TILEGX_OPC_LD2S_ADD, TILEGX_OPC_LD2S_ADD, CHILD(628),
+  BITFIELD(35, 2) /* index 628 */,
+  TILEGX_OPC_LD2S_ADD, TILEGX_OPC_LD2S_ADD, TILEGX_OPC_LD2S_ADD,
+  TILEGX_OPC_PREFETCH_ADD_L2_FAULT,
+  BITFIELD(31, 2) /* index 633 */,
+  TILEGX_OPC_LD2U_ADD, TILEGX_OPC_LD2U_ADD, TILEGX_OPC_LD2U_ADD, CHILD(638),
+  BITFIELD(33, 2) /* index 638 */,
+  TILEGX_OPC_LD2U_ADD, TILEGX_OPC_LD2U_ADD, TILEGX_OPC_LD2U_ADD, CHILD(643),
+  BITFIELD(35, 2) /* index 643 */,
+  TILEGX_OPC_LD2U_ADD, TILEGX_OPC_LD2U_ADD, TILEGX_OPC_LD2U_ADD,
+  TILEGX_OPC_PREFETCH_ADD_L2,
+  BITFIELD(31, 2) /* index 648 */,
+  TILEGX_OPC_LD4S_ADD, TILEGX_OPC_LD4S_ADD, TILEGX_OPC_LD4S_ADD, CHILD(653),
+  BITFIELD(33, 2) /* index 653 */,
+  TILEGX_OPC_LD4S_ADD, TILEGX_OPC_LD4S_ADD, TILEGX_OPC_LD4S_ADD, CHILD(658),
+  BITFIELD(35, 2) /* index 658 */,
+  TILEGX_OPC_LD4S_ADD, TILEGX_OPC_LD4S_ADD, TILEGX_OPC_LD4S_ADD,
+  TILEGX_OPC_PREFETCH_ADD_L3_FAULT,
+  BITFIELD(51, 2) /* index 663 */,
+  CHILD(668), TILEGX_OPC_LDNT1S_ADD, TILEGX_OPC_LDNT1U_ADD,
+  TILEGX_OPC_LDNT2S_ADD,
+  BITFIELD(31, 2) /* index 668 */,
+  TILEGX_OPC_LD4U_ADD, TILEGX_OPC_LD4U_ADD, TILEGX_OPC_LD4U_ADD, CHILD(673),
+  BITFIELD(33, 2) /* index 673 */,
+  TILEGX_OPC_LD4U_ADD, TILEGX_OPC_LD4U_ADD, TILEGX_OPC_LD4U_ADD, CHILD(678),
+  BITFIELD(35, 2) /* index 678 */,
+  TILEGX_OPC_LD4U_ADD, TILEGX_OPC_LD4U_ADD, TILEGX_OPC_LD4U_ADD,
+  TILEGX_OPC_PREFETCH_ADD_L3,
+  BITFIELD(51, 2) /* index 683 */,
+  TILEGX_OPC_LDNT2U_ADD, TILEGX_OPC_LDNT4S_ADD, TILEGX_OPC_LDNT4U_ADD,
+  TILEGX_OPC_LDNT_ADD,
+  BITFIELD(51, 2) /* index 688 */,
+  TILEGX_OPC_LD_ADD, TILEGX_OPC_LDNA_ADD, TILEGX_OPC_MFSPR, TILEGX_OPC_MTSPR,
+  BITFIELD(51, 2) /* index 693 */,
+  TILEGX_OPC_ORI, TILEGX_OPC_ST1_ADD, TILEGX_OPC_ST2_ADD, TILEGX_OPC_ST4_ADD,
+  BITFIELD(51, 2) /* index 698 */,
+  TILEGX_OPC_STNT1_ADD, TILEGX_OPC_STNT2_ADD, TILEGX_OPC_STNT4_ADD,
+  TILEGX_OPC_STNT_ADD,
+  BITFIELD(51, 2) /* index 703 */,
+  TILEGX_OPC_ST_ADD, TILEGX_OPC_V1ADDI, TILEGX_OPC_V1CMPEQI,
+  TILEGX_OPC_V1CMPLTSI,
+  BITFIELD(51, 2) /* index 708 */,
+  TILEGX_OPC_V1CMPLTUI, TILEGX_OPC_V1MAXUI, TILEGX_OPC_V1MINUI,
+  TILEGX_OPC_V2ADDI,
+  BITFIELD(51, 2) /* index 713 */,
+  TILEGX_OPC_V2CMPEQI, TILEGX_OPC_V2CMPLTSI, TILEGX_OPC_V2CMPLTUI,
+  TILEGX_OPC_V2MAXSI,
+  BITFIELD(51, 2) /* index 718 */,
+  TILEGX_OPC_V2MINSI, TILEGX_OPC_XORI, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(49, 4) /* index 723 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_ADDXSC, TILEGX_OPC_ADDX, TILEGX_OPC_ADD,
+  TILEGX_OPC_AND, TILEGX_OPC_CMPEQ, TILEGX_OPC_CMPEXCH4, TILEGX_OPC_CMPEXCH,
+  TILEGX_OPC_CMPLES, TILEGX_OPC_CMPLEU, TILEGX_OPC_CMPLTS, TILEGX_OPC_CMPLTU,
+  TILEGX_OPC_CMPNE, TILEGX_OPC_DBLALIGN2, TILEGX_OPC_DBLALIGN4,
+  TILEGX_OPC_DBLALIGN6,
+  BITFIELD(49, 4) /* index 740 */,
+  TILEGX_OPC_EXCH4, TILEGX_OPC_EXCH, TILEGX_OPC_FETCHADD4,
+  TILEGX_OPC_FETCHADDGEZ4, TILEGX_OPC_FETCHADDGEZ, TILEGX_OPC_FETCHADD,
+  TILEGX_OPC_FETCHAND4, TILEGX_OPC_FETCHAND, TILEGX_OPC_FETCHOR4,
+  TILEGX_OPC_FETCHOR, TILEGX_OPC_MNZ, TILEGX_OPC_MZ, TILEGX_OPC_NOR,
+  CHILD(757), TILEGX_OPC_ROTL, TILEGX_OPC_SHL1ADDX,
+  BITFIELD(43, 2) /* index 757 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, CHILD(762),
+  BITFIELD(45, 2) /* index 762 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, CHILD(767),
+  BITFIELD(47, 2) /* index 767 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_MOVE,
+  BITFIELD(49, 4) /* index 772 */,
+  TILEGX_OPC_SHL1ADD, TILEGX_OPC_SHL2ADDX, TILEGX_OPC_SHL2ADD,
+  TILEGX_OPC_SHL3ADDX, TILEGX_OPC_SHL3ADD, TILEGX_OPC_SHLX, TILEGX_OPC_SHL,
+  TILEGX_OPC_SHRS, TILEGX_OPC_SHRUX, TILEGX_OPC_SHRU, TILEGX_OPC_ST1,
+  TILEGX_OPC_ST2, TILEGX_OPC_ST4, TILEGX_OPC_STNT1, TILEGX_OPC_STNT2,
+  TILEGX_OPC_STNT4,
+  BITFIELD(46, 7) /* index 789 */,
+  TILEGX_OPC_STNT, TILEGX_OPC_STNT, TILEGX_OPC_STNT, TILEGX_OPC_STNT,
+  TILEGX_OPC_STNT, TILEGX_OPC_STNT, TILEGX_OPC_STNT, TILEGX_OPC_STNT,
+  TILEGX_OPC_ST, TILEGX_OPC_ST, TILEGX_OPC_ST, TILEGX_OPC_ST, TILEGX_OPC_ST,
+  TILEGX_OPC_ST, TILEGX_OPC_ST, TILEGX_OPC_ST, TILEGX_OPC_SUBXSC,
+  TILEGX_OPC_SUBXSC, TILEGX_OPC_SUBXSC, TILEGX_OPC_SUBXSC, TILEGX_OPC_SUBXSC,
+  TILEGX_OPC_SUBXSC, TILEGX_OPC_SUBXSC, TILEGX_OPC_SUBXSC, TILEGX_OPC_SUBX,
+  TILEGX_OPC_SUBX, TILEGX_OPC_SUBX, TILEGX_OPC_SUBX, TILEGX_OPC_SUBX,
+  TILEGX_OPC_SUBX, TILEGX_OPC_SUBX, TILEGX_OPC_SUBX, TILEGX_OPC_SUB,
+  TILEGX_OPC_SUB, TILEGX_OPC_SUB, TILEGX_OPC_SUB, TILEGX_OPC_SUB,
+  TILEGX_OPC_SUB, TILEGX_OPC_SUB, TILEGX_OPC_SUB, CHILD(918), CHILD(927),
+  CHILD(1006), CHILD(1090), CHILD(1099), TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_V1ADDUC, TILEGX_OPC_V1ADDUC, TILEGX_OPC_V1ADDUC,
+  TILEGX_OPC_V1ADDUC, TILEGX_OPC_V1ADDUC, TILEGX_OPC_V1ADDUC,
+  TILEGX_OPC_V1ADDUC, TILEGX_OPC_V1ADDUC, TILEGX_OPC_V1ADD, TILEGX_OPC_V1ADD,
+  TILEGX_OPC_V1ADD, TILEGX_OPC_V1ADD, TILEGX_OPC_V1ADD, TILEGX_OPC_V1ADD,
+  TILEGX_OPC_V1ADD, TILEGX_OPC_V1ADD, TILEGX_OPC_V1CMPEQ, TILEGX_OPC_V1CMPEQ,
+  TILEGX_OPC_V1CMPEQ, TILEGX_OPC_V1CMPEQ, TILEGX_OPC_V1CMPEQ,
+  TILEGX_OPC_V1CMPEQ, TILEGX_OPC_V1CMPEQ, TILEGX_OPC_V1CMPEQ,
+  TILEGX_OPC_V1CMPLES, TILEGX_OPC_V1CMPLES, TILEGX_OPC_V1CMPLES,
+  TILEGX_OPC_V1CMPLES, TILEGX_OPC_V1CMPLES, TILEGX_OPC_V1CMPLES,
+  TILEGX_OPC_V1CMPLES, TILEGX_OPC_V1CMPLES, TILEGX_OPC_V1CMPLEU,
+  TILEGX_OPC_V1CMPLEU, TILEGX_OPC_V1CMPLEU, TILEGX_OPC_V1CMPLEU,
+  TILEGX_OPC_V1CMPLEU, TILEGX_OPC_V1CMPLEU, TILEGX_OPC_V1CMPLEU,
+  TILEGX_OPC_V1CMPLEU, TILEGX_OPC_V1CMPLTS, TILEGX_OPC_V1CMPLTS,
+  TILEGX_OPC_V1CMPLTS, TILEGX_OPC_V1CMPLTS, TILEGX_OPC_V1CMPLTS,
+  TILEGX_OPC_V1CMPLTS, TILEGX_OPC_V1CMPLTS, TILEGX_OPC_V1CMPLTS,
+  TILEGX_OPC_V1CMPLTU, TILEGX_OPC_V1CMPLTU, TILEGX_OPC_V1CMPLTU,
+  TILEGX_OPC_V1CMPLTU, TILEGX_OPC_V1CMPLTU, TILEGX_OPC_V1CMPLTU,
+  TILEGX_OPC_V1CMPLTU, TILEGX_OPC_V1CMPLTU, TILEGX_OPC_V1CMPNE,
+  TILEGX_OPC_V1CMPNE, TILEGX_OPC_V1CMPNE, TILEGX_OPC_V1CMPNE,
+  TILEGX_OPC_V1CMPNE, TILEGX_OPC_V1CMPNE, TILEGX_OPC_V1CMPNE,
+  TILEGX_OPC_V1CMPNE, TILEGX_OPC_V1INT_H, TILEGX_OPC_V1INT_H,
+  TILEGX_OPC_V1INT_H, TILEGX_OPC_V1INT_H, TILEGX_OPC_V1INT_H,
+  TILEGX_OPC_V1INT_H, TILEGX_OPC_V1INT_H, TILEGX_OPC_V1INT_H,
+  TILEGX_OPC_V1INT_L, TILEGX_OPC_V1INT_L, TILEGX_OPC_V1INT_L,
+  TILEGX_OPC_V1INT_L, TILEGX_OPC_V1INT_L, TILEGX_OPC_V1INT_L,
+  TILEGX_OPC_V1INT_L, TILEGX_OPC_V1INT_L,
+  BITFIELD(43, 3) /* index 918 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_DRAIN, TILEGX_OPC_DTLBPR, TILEGX_OPC_FINV,
+  TILEGX_OPC_FLUSHWB, TILEGX_OPC_FLUSH, TILEGX_OPC_FNOP, TILEGX_OPC_ICOH,
+  BITFIELD(43, 3) /* index 927 */,
+  CHILD(936), TILEGX_OPC_INV, TILEGX_OPC_IRET, TILEGX_OPC_JALRP,
+  TILEGX_OPC_JALR, TILEGX_OPC_JRP, TILEGX_OPC_JR, CHILD(991),
+  BITFIELD(31, 2) /* index 936 */,
+  CHILD(941), CHILD(966), TILEGX_OPC_ILL, TILEGX_OPC_ILL,
+  BITFIELD(33, 2) /* index 941 */,
+  TILEGX_OPC_ILL, TILEGX_OPC_ILL, TILEGX_OPC_ILL, CHILD(946),
+  BITFIELD(35, 2) /* index 946 */,
+  TILEGX_OPC_ILL, CHILD(951), TILEGX_OPC_ILL, TILEGX_OPC_ILL,
+  BITFIELD(37, 2) /* index 951 */,
+  TILEGX_OPC_ILL, CHILD(956), TILEGX_OPC_ILL, TILEGX_OPC_ILL,
+  BITFIELD(39, 2) /* index 956 */,
+  TILEGX_OPC_ILL, CHILD(961), TILEGX_OPC_ILL, TILEGX_OPC_ILL,
+  BITFIELD(41, 2) /* index 961 */,
+  TILEGX_OPC_ILL, TILEGX_OPC_ILL, TILEGX_OPC_BPT, TILEGX_OPC_ILL,
+  BITFIELD(33, 2) /* index 966 */,
+  TILEGX_OPC_ILL, TILEGX_OPC_ILL, TILEGX_OPC_ILL, CHILD(971),
+  BITFIELD(35, 2) /* index 971 */,
+  TILEGX_OPC_ILL, CHILD(976), TILEGX_OPC_ILL, TILEGX_OPC_ILL,
+  BITFIELD(37, 2) /* index 976 */,
+  TILEGX_OPC_ILL, CHILD(981), TILEGX_OPC_ILL, TILEGX_OPC_ILL,
+  BITFIELD(39, 2) /* index 981 */,
+  TILEGX_OPC_ILL, CHILD(986), TILEGX_OPC_ILL, TILEGX_OPC_ILL,
+  BITFIELD(41, 2) /* index 986 */,
+  TILEGX_OPC_ILL, TILEGX_OPC_ILL, TILEGX_OPC_RAISE, TILEGX_OPC_ILL,
+  BITFIELD(31, 2) /* index 991 */,
+  TILEGX_OPC_LD1S, TILEGX_OPC_LD1S, TILEGX_OPC_LD1S, CHILD(996),
+  BITFIELD(33, 2) /* index 996 */,
+  TILEGX_OPC_LD1S, TILEGX_OPC_LD1S, TILEGX_OPC_LD1S, CHILD(1001),
+  BITFIELD(35, 2) /* index 1001 */,
+  TILEGX_OPC_LD1S, TILEGX_OPC_LD1S, TILEGX_OPC_LD1S,
+  TILEGX_OPC_PREFETCH_L1_FAULT,
+  BITFIELD(43, 3) /* index 1006 */,
+  CHILD(1015), CHILD(1030), CHILD(1045), CHILD(1060), CHILD(1075),
+  TILEGX_OPC_LDNA, TILEGX_OPC_LDNT1S, TILEGX_OPC_LDNT1U,
+  BITFIELD(31, 2) /* index 1015 */,
+  TILEGX_OPC_LD1U, TILEGX_OPC_LD1U, TILEGX_OPC_LD1U, CHILD(1020),
+  BITFIELD(33, 2) /* index 1020 */,
+  TILEGX_OPC_LD1U, TILEGX_OPC_LD1U, TILEGX_OPC_LD1U, CHILD(1025),
+  BITFIELD(35, 2) /* index 1025 */,
+  TILEGX_OPC_LD1U, TILEGX_OPC_LD1U, TILEGX_OPC_LD1U, TILEGX_OPC_PREFETCH,
+  BITFIELD(31, 2) /* index 1030 */,
+  TILEGX_OPC_LD2S, TILEGX_OPC_LD2S, TILEGX_OPC_LD2S, CHILD(1035),
+  BITFIELD(33, 2) /* index 1035 */,
+  TILEGX_OPC_LD2S, TILEGX_OPC_LD2S, TILEGX_OPC_LD2S, CHILD(1040),
+  BITFIELD(35, 2) /* index 1040 */,
+  TILEGX_OPC_LD2S, TILEGX_OPC_LD2S, TILEGX_OPC_LD2S,
+  TILEGX_OPC_PREFETCH_L2_FAULT,
+  BITFIELD(31, 2) /* index 1045 */,
+  TILEGX_OPC_LD2U, TILEGX_OPC_LD2U, TILEGX_OPC_LD2U, CHILD(1050),
+  BITFIELD(33, 2) /* index 1050 */,
+  TILEGX_OPC_LD2U, TILEGX_OPC_LD2U, TILEGX_OPC_LD2U, CHILD(1055),
+  BITFIELD(35, 2) /* index 1055 */,
+  TILEGX_OPC_LD2U, TILEGX_OPC_LD2U, TILEGX_OPC_LD2U, TILEGX_OPC_PREFETCH_L2,
+  BITFIELD(31, 2) /* index 1060 */,
+  TILEGX_OPC_LD4S, TILEGX_OPC_LD4S, TILEGX_OPC_LD4S, CHILD(1065),
+  BITFIELD(33, 2) /* index 1065 */,
+  TILEGX_OPC_LD4S, TILEGX_OPC_LD4S, TILEGX_OPC_LD4S, CHILD(1070),
+  BITFIELD(35, 2) /* index 1070 */,
+  TILEGX_OPC_LD4S, TILEGX_OPC_LD4S, TILEGX_OPC_LD4S,
+  TILEGX_OPC_PREFETCH_L3_FAULT,
+  BITFIELD(31, 2) /* index 1075 */,
+  TILEGX_OPC_LD4U, TILEGX_OPC_LD4U, TILEGX_OPC_LD4U, CHILD(1080),
+  BITFIELD(33, 2) /* index 1080 */,
+  TILEGX_OPC_LD4U, TILEGX_OPC_LD4U, TILEGX_OPC_LD4U, CHILD(1085),
+  BITFIELD(35, 2) /* index 1085 */,
+  TILEGX_OPC_LD4U, TILEGX_OPC_LD4U, TILEGX_OPC_LD4U, TILEGX_OPC_PREFETCH_L3,
+  BITFIELD(43, 3) /* index 1090 */,
+  TILEGX_OPC_LDNT2S, TILEGX_OPC_LDNT2U, TILEGX_OPC_LDNT4S, TILEGX_OPC_LDNT4U,
+  TILEGX_OPC_LDNT, TILEGX_OPC_LD, TILEGX_OPC_LNK, TILEGX_OPC_MF,
+  BITFIELD(43, 3) /* index 1099 */,
+  TILEGX_OPC_NAP, TILEGX_OPC_NOP, TILEGX_OPC_SWINT0, TILEGX_OPC_SWINT1,
+  TILEGX_OPC_SWINT2, TILEGX_OPC_SWINT3, TILEGX_OPC_WH64, TILEGX_OPC_NONE,
+  BITFIELD(49, 4) /* index 1108 */,
+  TILEGX_OPC_V1MAXU, TILEGX_OPC_V1MINU, TILEGX_OPC_V1MNZ, TILEGX_OPC_V1MZ,
+  TILEGX_OPC_V1SHL, TILEGX_OPC_V1SHRS, TILEGX_OPC_V1SHRU, TILEGX_OPC_V1SUBUC,
+  TILEGX_OPC_V1SUB, TILEGX_OPC_V2ADDSC, TILEGX_OPC_V2ADD, TILEGX_OPC_V2CMPEQ,
+  TILEGX_OPC_V2CMPLES, TILEGX_OPC_V2CMPLEU, TILEGX_OPC_V2CMPLTS,
+  TILEGX_OPC_V2CMPLTU,
+  BITFIELD(49, 4) /* index 1125 */,
+  TILEGX_OPC_V2CMPNE, TILEGX_OPC_V2INT_H, TILEGX_OPC_V2INT_L,
+  TILEGX_OPC_V2MAXS, TILEGX_OPC_V2MINS, TILEGX_OPC_V2MNZ, TILEGX_OPC_V2MZ,
+  TILEGX_OPC_V2PACKH, TILEGX_OPC_V2PACKL, TILEGX_OPC_V2PACKUC,
+  TILEGX_OPC_V2SHLSC, TILEGX_OPC_V2SHL, TILEGX_OPC_V2SHRS, TILEGX_OPC_V2SHRU,
+  TILEGX_OPC_V2SUBSC, TILEGX_OPC_V2SUB,
+  BITFIELD(49, 4) /* index 1142 */,
+  TILEGX_OPC_V4ADDSC, TILEGX_OPC_V4ADD, TILEGX_OPC_V4INT_H,
+  TILEGX_OPC_V4INT_L, TILEGX_OPC_V4PACKSC, TILEGX_OPC_V4SHLSC,
+  TILEGX_OPC_V4SHL, TILEGX_OPC_V4SHRS, TILEGX_OPC_V4SHRU, TILEGX_OPC_V4SUBSC,
+  TILEGX_OPC_V4SUB, TILEGX_OPC_XOR, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(49, 4) /* index 1159 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_ROTLI, TILEGX_OPC_SHLI, TILEGX_OPC_SHLXI,
+  TILEGX_OPC_SHRSI, TILEGX_OPC_SHRUI, TILEGX_OPC_SHRUXI, TILEGX_OPC_V1SHLI,
+  TILEGX_OPC_V1SHRSI, TILEGX_OPC_V1SHRUI, TILEGX_OPC_V2SHLI,
+  TILEGX_OPC_V2SHRSI, TILEGX_OPC_V2SHRUI, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE,
+  BITFIELD(31, 2) /* index 1176 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  CHILD(1181),
+  BITFIELD(33, 2) /* index 1181 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  CHILD(1186),
+  BITFIELD(35, 2) /* index 1186 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  CHILD(1191),
+  BITFIELD(37, 2) /* index 1191 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  CHILD(1196),
+  BITFIELD(39, 2) /* index 1196 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  CHILD(1201),
+  BITFIELD(41, 2) /* index 1201 */,
+  TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI, TILEGX_OPC_SHL16INSLI,
+  TILEGX_OPC_INFOL,
+};
+
+static const unsigned short decode_Y0_fsm[178] =
+{
+  BITFIELD(27, 4) /* index 0 */,
+  CHILD(17), TILEGX_OPC_ADDXI, CHILD(32), TILEGX_OPC_CMPEQI,
+  TILEGX_OPC_CMPLTSI, CHILD(62), CHILD(67), CHILD(118), CHILD(123),
+  CHILD(128), CHILD(133), CHILD(153), CHILD(158), CHILD(163), CHILD(168),
+  CHILD(173),
+  BITFIELD(6, 2) /* index 17 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, CHILD(22),
+  BITFIELD(8, 2) /* index 22 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, CHILD(27),
+  BITFIELD(10, 2) /* index 27 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_MOVEI,
+  BITFIELD(0, 2) /* index 32 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(37),
+  BITFIELD(2, 2) /* index 37 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(42),
+  BITFIELD(4, 2) /* index 42 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(47),
+  BITFIELD(6, 2) /* index 47 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(52),
+  BITFIELD(8, 2) /* index 52 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(57),
+  BITFIELD(10, 2) /* index 57 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_INFO,
+  BITFIELD(18, 2) /* index 62 */,
+  TILEGX_OPC_ADDX, TILEGX_OPC_ADD, TILEGX_OPC_SUBX, TILEGX_OPC_SUB,
+  BITFIELD(15, 5) /* index 67 */,
+  TILEGX_OPC_SHL1ADD, TILEGX_OPC_SHL1ADD, TILEGX_OPC_SHL1ADD,
+  TILEGX_OPC_SHL1ADD, TILEGX_OPC_SHL1ADD, TILEGX_OPC_SHL1ADD,
+  TILEGX_OPC_SHL1ADD, TILEGX_OPC_SHL1ADD, TILEGX_OPC_SHL2ADD,
+  TILEGX_OPC_SHL2ADD, TILEGX_OPC_SHL2ADD, TILEGX_OPC_SHL2ADD,
+  TILEGX_OPC_SHL2ADD, TILEGX_OPC_SHL2ADD, TILEGX_OPC_SHL2ADD,
+  TILEGX_OPC_SHL2ADD, TILEGX_OPC_SHL3ADD, TILEGX_OPC_SHL3ADD,
+  TILEGX_OPC_SHL3ADD, TILEGX_OPC_SHL3ADD, TILEGX_OPC_SHL3ADD,
+  TILEGX_OPC_SHL3ADD, TILEGX_OPC_SHL3ADD, TILEGX_OPC_SHL3ADD, CHILD(100),
+  CHILD(109), TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(12, 3) /* index 100 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_CLZ, TILEGX_OPC_CTZ, TILEGX_OPC_FNOP,
+  TILEGX_OPC_FSINGLE_PACK1, TILEGX_OPC_NOP, TILEGX_OPC_PCNT,
+  TILEGX_OPC_REVBITS,
+  BITFIELD(12, 3) /* index 109 */,
+  TILEGX_OPC_REVBYTES, TILEGX_OPC_TBLIDXB0, TILEGX_OPC_TBLIDXB1,
+  TILEGX_OPC_TBLIDXB2, TILEGX_OPC_TBLIDXB3, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  TILEGX_OPC_NONE,
+  BITFIELD(18, 2) /* index 118 */,
+  TILEGX_OPC_CMPLES, TILEGX_OPC_CMPLEU, TILEGX_OPC_CMPLTS, TILEGX_OPC_CMPLTU,
+  BITFIELD(18, 2) /* index 123 */,
+  TILEGX_OPC_CMPEQ, TILEGX_OPC_CMPNE, TILEGX_OPC_MULAX, TILEGX_OPC_MULX,
+  BITFIELD(18, 2) /* index 128 */,
+  TILEGX_OPC_CMOVEQZ, TILEGX_OPC_CMOVNEZ, TILEGX_OPC_MNZ, TILEGX_OPC_MZ,
+  BITFIELD(18, 2) /* index 133 */,
+  TILEGX_OPC_AND, TILEGX_OPC_NOR, CHILD(138), TILEGX_OPC_XOR,
+  BITFIELD(12, 2) /* index 138 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, CHILD(143),
+  BITFIELD(14, 2) /* index 143 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, CHILD(148),
+  BITFIELD(16, 2) /* index 148 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_MOVE,
+  BITFIELD(18, 2) /* index 153 */,
+  TILEGX_OPC_ROTL, TILEGX_OPC_SHL, TILEGX_OPC_SHRS, TILEGX_OPC_SHRU,
+  BITFIELD(18, 2) /* index 158 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_SHL1ADDX, TILEGX_OPC_SHL2ADDX,
+  TILEGX_OPC_SHL3ADDX,
+  BITFIELD(18, 2) /* index 163 */,
+  TILEGX_OPC_MUL_HS_HS, TILEGX_OPC_MUL_HU_HU, TILEGX_OPC_MUL_LS_LS,
+  TILEGX_OPC_MUL_LU_LU,
+  BITFIELD(18, 2) /* index 168 */,
+  TILEGX_OPC_MULA_HS_HS, TILEGX_OPC_MULA_HU_HU, TILEGX_OPC_MULA_LS_LS,
+  TILEGX_OPC_MULA_LU_LU,
+  BITFIELD(18, 2) /* index 173 */,
+  TILEGX_OPC_ROTLI, TILEGX_OPC_SHLI, TILEGX_OPC_SHRSI, TILEGX_OPC_SHRUI,
+};
+
+static const unsigned short decode_Y1_fsm[167] =
+{
+  BITFIELD(58, 4) /* index 0 */,
+  TILEGX_OPC_NONE, CHILD(17), TILEGX_OPC_ADDXI, CHILD(32), TILEGX_OPC_CMPEQI,
+  TILEGX_OPC_CMPLTSI, CHILD(62), CHILD(67), CHILD(117), CHILD(122),
+  CHILD(127), CHILD(132), CHILD(152), CHILD(157), CHILD(162), TILEGX_OPC_NONE,
+  BITFIELD(37, 2) /* index 17 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, CHILD(22),
+  BITFIELD(39, 2) /* index 22 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, CHILD(27),
+  BITFIELD(41, 2) /* index 27 */,
+  TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_ADDI, TILEGX_OPC_MOVEI,
+  BITFIELD(31, 2) /* index 32 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(37),
+  BITFIELD(33, 2) /* index 37 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(42),
+  BITFIELD(35, 2) /* index 42 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(47),
+  BITFIELD(37, 2) /* index 47 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(52),
+  BITFIELD(39, 2) /* index 52 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, CHILD(57),
+  BITFIELD(41, 2) /* index 57 */,
+  TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_ANDI, TILEGX_OPC_INFO,
+  BITFIELD(49, 2) /* index 62 */,
+  TILEGX_OPC_ADDX, TILEGX_OPC_ADD, TILEGX_OPC_SUBX, TILEGX_OPC_SUB,
+  BITFIELD(47, 4) /* index 67 */,
+  TILEGX_OPC_SHL1ADD, TILEGX_OPC_SHL1ADD, TILEGX_OPC_SHL1ADD,
+  TILEGX_OPC_SHL1ADD, TILEGX_OPC_SHL2ADD, TILEGX_OPC_SHL2ADD,
+  TILEGX_OPC_SHL2ADD, TILEGX_OPC_SHL2ADD, TILEGX_OPC_SHL3ADD,
+  TILEGX_OPC_SHL3ADD, TILEGX_OPC_SHL3ADD, TILEGX_OPC_SHL3ADD, CHILD(84),
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_NONE,
+  BITFIELD(43, 3) /* index 84 */,
+  CHILD(93), CHILD(96), CHILD(99), CHILD(102), CHILD(105), CHILD(108),
+  CHILD(111), CHILD(114),
+  BITFIELD(46, 1) /* index 93 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_FNOP,
+  BITFIELD(46, 1) /* index 96 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_ILL,
+  BITFIELD(46, 1) /* index 99 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_JALRP,
+  BITFIELD(46, 1) /* index 102 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_JALR,
+  BITFIELD(46, 1) /* index 105 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_JRP,
+  BITFIELD(46, 1) /* index 108 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_JR,
+  BITFIELD(46, 1) /* index 111 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_LNK,
+  BITFIELD(46, 1) /* index 114 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_NOP,
+  BITFIELD(49, 2) /* index 117 */,
+  TILEGX_OPC_CMPLES, TILEGX_OPC_CMPLEU, TILEGX_OPC_CMPLTS, TILEGX_OPC_CMPLTU,
+  BITFIELD(49, 2) /* index 122 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_CMPEQ, TILEGX_OPC_CMPNE,
+  BITFIELD(49, 2) /* index 127 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_NONE, TILEGX_OPC_MNZ, TILEGX_OPC_MZ,
+  BITFIELD(49, 2) /* index 132 */,
+  TILEGX_OPC_AND, TILEGX_OPC_NOR, CHILD(137), TILEGX_OPC_XOR,
+  BITFIELD(43, 2) /* index 137 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, CHILD(142),
+  BITFIELD(45, 2) /* index 142 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, CHILD(147),
+  BITFIELD(47, 2) /* index 147 */,
+  TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_OR, TILEGX_OPC_MOVE,
+  BITFIELD(49, 2) /* index 152 */,
+  TILEGX_OPC_ROTL, TILEGX_OPC_SHL, TILEGX_OPC_SHRS, TILEGX_OPC_SHRU,
+  BITFIELD(49, 2) /* index 157 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_SHL1ADDX, TILEGX_OPC_SHL2ADDX,
+  TILEGX_OPC_SHL3ADDX,
+  BITFIELD(49, 2) /* index 162 */,
+  TILEGX_OPC_ROTLI, TILEGX_OPC_SHLI, TILEGX_OPC_SHRSI, TILEGX_OPC_SHRUI,
+};
+
+static const unsigned short decode_Y2_fsm[118] =
+{
+  BITFIELD(62, 2) /* index 0 */,
+  TILEGX_OPC_NONE, CHILD(5), CHILD(66), CHILD(109),
+  BITFIELD(55, 3) /* index 5 */,
+  CHILD(14), CHILD(14), CHILD(14), CHILD(17), CHILD(40), CHILD(40), CHILD(40),
+  CHILD(43),
+  BITFIELD(26, 1) /* index 14 */,
+  TILEGX_OPC_LD1S, TILEGX_OPC_LD1U,
+  BITFIELD(26, 1) /* index 17 */,
+  CHILD(20), CHILD(30),
+  BITFIELD(51, 2) /* index 20 */,
+  TILEGX_OPC_LD1S, TILEGX_OPC_LD1S, TILEGX_OPC_LD1S, CHILD(25),
+  BITFIELD(53, 2) /* index 25 */,
+  TILEGX_OPC_LD1S, TILEGX_OPC_LD1S, TILEGX_OPC_LD1S,
+  TILEGX_OPC_PREFETCH_L1_FAULT,
+  BITFIELD(51, 2) /* index 30 */,
+  TILEGX_OPC_LD1U, TILEGX_OPC_LD1U, TILEGX_OPC_LD1U, CHILD(35),
+  BITFIELD(53, 2) /* index 35 */,
+  TILEGX_OPC_LD1U, TILEGX_OPC_LD1U, TILEGX_OPC_LD1U, TILEGX_OPC_PREFETCH,
+  BITFIELD(26, 1) /* index 40 */,
+  TILEGX_OPC_LD2S, TILEGX_OPC_LD2U,
+  BITFIELD(26, 1) /* index 43 */,
+  CHILD(46), CHILD(56),
+  BITFIELD(51, 2) /* index 46 */,
+  TILEGX_OPC_LD2S, TILEGX_OPC_LD2S, TILEGX_OPC_LD2S, CHILD(51),
+  BITFIELD(53, 2) /* index 51 */,
+  TILEGX_OPC_LD2S, TILEGX_OPC_LD2S, TILEGX_OPC_LD2S,
+  TILEGX_OPC_PREFETCH_L2_FAULT,
+  BITFIELD(51, 2) /* index 56 */,
+  TILEGX_OPC_LD2U, TILEGX_OPC_LD2U, TILEGX_OPC_LD2U, CHILD(61),
+  BITFIELD(53, 2) /* index 61 */,
+  TILEGX_OPC_LD2U, TILEGX_OPC_LD2U, TILEGX_OPC_LD2U, TILEGX_OPC_PREFETCH_L2,
+  BITFIELD(56, 2) /* index 66 */,
+  CHILD(71), CHILD(74), CHILD(90), CHILD(93),
+  BITFIELD(26, 1) /* index 71 */,
+  TILEGX_OPC_NONE, TILEGX_OPC_LD4S,
+  BITFIELD(26, 1) /* index 74 */,
+  TILEGX_OPC_NONE, CHILD(77),
+  BITFIELD(51, 2) /* index 77 */,
+  TILEGX_OPC_LD4S, TILEGX_OPC_LD4S, TILEGX_OPC_LD4S, CHILD(82),
+  BITFIELD(53, 2) /* index 82 */,
+  TILEGX_OPC_LD4S, TILEGX_OPC_LD4S, TILEGX_OPC_LD4S, CHILD(87),
+  BITFIELD(55, 1) /* index 87 */,
+  TILEGX_OPC_LD4S, TILEGX_OPC_PREFETCH_L3_FAULT,
+  BITFIELD(26, 1) /* index 90 */,
+  TILEGX_OPC_LD4U, TILEGX_OPC_LD,
+  BITFIELD(26, 1) /* index 93 */,
+  CHILD(96), TILEGX_OPC_LD,
+  BITFIELD(51, 2) /* index 96 */,
+  TILEGX_OPC_LD4U, TILEGX_OPC_LD4U, TILEGX_OPC_LD4U, CHILD(101),
+  BITFIELD(53, 2) /* index 101 */,
+  TILEGX_OPC_LD4U, TILEGX_OPC_LD4U, TILEGX_OPC_LD4U, CHILD(106),
+  BITFIELD(55, 1) /* index 106 */,
+  TILEGX_OPC_LD4U, TILEGX_OPC_PREFETCH_L3,
+  BITFIELD(26, 1) /* index 109 */,
+  CHILD(112), CHILD(115),
+  BITFIELD(57, 1) /* index 112 */,
+  TILEGX_OPC_ST1, TILEGX_OPC_ST4,
+  BITFIELD(57, 1) /* index 115 */,
+  TILEGX_OPC_ST2, TILEGX_OPC_ST,
+};
+
+#undef BITFIELD
+#undef CHILD
+const unsigned short * const
+tilegx_bundle_decoder_fsms[TILEGX_NUM_PIPELINE_ENCODINGS] =
+{
+  decode_X0_fsm,
+  decode_X1_fsm,
+  decode_Y0_fsm,
+  decode_Y1_fsm,
+  decode_Y2_fsm
+};
+const struct tilegx_operand tilegx_operands[35] =
+{
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_IMM8_X0),
+    8, 1, 0, 0, 0, 0,
+    create_Imm8_X0, get_Imm8_X0
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_IMM8_X1),
+    8, 1, 0, 0, 0, 0,
+    create_Imm8_X1, get_Imm8_X1
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_IMM8_Y0),
+    8, 1, 0, 0, 0, 0,
+    create_Imm8_Y0, get_Imm8_Y0
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_IMM8_Y1),
+    8, 1, 0, 0, 0, 0,
+    create_Imm8_Y1, get_Imm8_Y1
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_IMM16_X0_HW0_LAST),
+    16, 1, 0, 0, 0, 0,
+    create_Imm16_X0, get_Imm16_X0
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_IMM16_X1_HW0_LAST),
+    16, 1, 0, 0, 0, 0,
+    create_Imm16_X1, get_Imm16_X1
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 0, 1, 0, 0,
+    create_Dest_X0, get_Dest_X0
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcA_X0, get_SrcA_X0
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 0, 1, 0, 0,
+    create_Dest_X1, get_Dest_X1
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcA_X1, get_SrcA_X1
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 0, 1, 0, 0,
+    create_Dest_Y0, get_Dest_Y0
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcA_Y0, get_SrcA_Y0
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 0, 1, 0, 0,
+    create_Dest_Y1, get_Dest_Y1
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcA_Y1, get_SrcA_Y1
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcA_Y2, get_SrcA_Y2
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 1, 0, 0,
+    create_SrcA_X1, get_SrcA_X1
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcB_X0, get_SrcB_X0
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcB_X1, get_SrcB_X1
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcB_Y0, get_SrcB_Y0
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcB_Y1, get_SrcB_Y1
+  },
+  {
+    TILEGX_OP_TYPE_ADDRESS, BFD_RELOC(TILEGX_BROFF_X1),
+    17, 1, 0, 0, 1, TILEGX_LOG2_BUNDLE_ALIGNMENT_IN_BYTES,
+    create_BrOff_X1, get_BrOff_X1
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_MMSTART_X0),
+    6, 0, 0, 0, 0, 0,
+    create_BFStart_X0, get_BFStart_X0
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_MMEND_X0),
+    6, 0, 0, 0, 0, 0,
+    create_BFEnd_X0, get_BFEnd_X0
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 1, 0, 0,
+    create_Dest_X0, get_Dest_X0
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 1, 0, 0,
+    create_Dest_Y0, get_Dest_Y0
+  },
+  {
+    TILEGX_OP_TYPE_ADDRESS, BFD_RELOC(TILEGX_JUMPOFF_X1),
+    27, 1, 0, 0, 1, TILEGX_LOG2_BUNDLE_ALIGNMENT_IN_BYTES,
+    create_JumpOff_X1, get_JumpOff_X1
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 0, 1, 0, 0,
+    create_SrcBDest_Y2, get_SrcBDest_Y2
+  },
+  {
+    TILEGX_OP_TYPE_SPR, BFD_RELOC(TILEGX_MF_IMM14_X1),
+    14, 0, 0, 0, 0, 0,
+    create_MF_Imm14_X1, get_MF_Imm14_X1
+  },
+  {
+    TILEGX_OP_TYPE_SPR, BFD_RELOC(TILEGX_MT_IMM14_X1),
+    14, 0, 0, 0, 0, 0,
+    create_MT_Imm14_X1, get_MT_Imm14_X1
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_SHAMT_X0),
+    6, 0, 0, 0, 0, 0,
+    create_ShAmt_X0, get_ShAmt_X0
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_SHAMT_X1),
+    6, 0, 0, 0, 0, 0,
+    create_ShAmt_X1, get_ShAmt_X1
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_SHAMT_Y0),
+    6, 0, 0, 0, 0, 0,
+    create_ShAmt_Y0, get_ShAmt_Y0
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_SHAMT_Y1),
+    6, 0, 0, 0, 0, 0,
+    create_ShAmt_Y1, get_ShAmt_Y1
+  },
+  {
+    TILEGX_OP_TYPE_REGISTER, BFD_RELOC(NONE),
+    6, 0, 1, 0, 0, 0,
+    create_SrcBDest_Y2, get_SrcBDest_Y2
+  },
+  {
+    TILEGX_OP_TYPE_IMMEDIATE, BFD_RELOC(TILEGX_DEST_IMM8_X1),
+    8, 1, 0, 0, 0, 0,
+    create_Dest_Imm8_X1, get_Dest_Imm8_X1
+  }
+};
+
+
+
+
+/* Given a set of bundle bits and the lookup FSM for a specific pipe,
+ * returns which instruction the bundle contains in that pipe.
+ */
+static const struct tilegx_opcode *
+find_opcode(tilegx_bundle_bits bits, const unsigned short *table)
+{
+  int index = 0;
+
+  while (1)
+  {
+    unsigned short bitspec = table[index];
+    unsigned int bitfield =
+      ((unsigned int)(bits >> (bitspec & 63))) & (bitspec >> 6);
+
+    unsigned short next = table[index + 1 + bitfield];
+    if (next <= TILEGX_OPC_NONE)
+      return &tilegx_opcodes[next];
+
+    index = next - TILEGX_OPC_NONE;
+  }
+}
+
+
+int
+parse_insn_tilegx(tilegx_bundle_bits bits,
+                  unsigned long long pc,
+                  struct tilegx_decoded_instruction
+                  decoded[TILEGX_MAX_INSTRUCTIONS_PER_BUNDLE])
+{
+  int num_instructions = 0;
+  int pipe;
+
+  int min_pipe, max_pipe;
+  if ((bits & TILEGX_BUNDLE_MODE_MASK) == 0)
+  {
+    min_pipe = TILEGX_PIPELINE_X0;
+    max_pipe = TILEGX_PIPELINE_X1;
+  }
+  else
+  {
+    min_pipe = TILEGX_PIPELINE_Y0;
+    max_pipe = TILEGX_PIPELINE_Y2;
+  }
+
+  /* For each pipe, find an instruction that fits. */
+  for (pipe = min_pipe; pipe <= max_pipe; pipe++)
+  {
+    const struct tilegx_opcode *opc;
+    struct tilegx_decoded_instruction *d;
+    int i;
+
+    d = &decoded[num_instructions++];
+    opc = find_opcode (bits, tilegx_bundle_decoder_fsms[pipe]);
+    d->opcode = opc;
+
+    /* Decode each operand, sign extending, etc. as appropriate. */
+    for (i = 0; i < opc->num_operands; i++)
+    {
+      const struct tilegx_operand *op =
+        &tilegx_operands[opc->operands[pipe][i]];
+      int raw_opval = op->extract (bits);
+      long long opval;
+
+      if (op->is_signed)
+      {
+        /* Sign-extend the operand. */
+        int shift = (int)((sizeof(int) * 8) - op->num_bits);
+        raw_opval = (raw_opval << shift) >> shift;
+      }
+
+      /* Adjust PC-relative scaled branch offsets. */
+      if (op->type == TILEGX_OP_TYPE_ADDRESS)
+        opval = (raw_opval * TILEGX_BUNDLE_SIZE_IN_BYTES) + pc;
+      else
+        opval = raw_opval;
+
+      /* Record the final value. */
+      d->operands[i] = op;
+      d->operand_values[i] = opval;
+    }
+  }
+
+  return num_instructions;
+}
diff --git a/arch/tile/kernel/time.c b/arch/tile/kernel/time.c
new file mode 100644
index 000000000..00178ecf9
--- /dev/null
+++ b/arch/tile/kernel/time.c
@@ -0,0 +1,300 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Support the cycle counter clocksource and tile timer clock event device.
+ */
+
+#include <linux/time.h>
+#include <linux/timex.h>
+#include <linux/clocksource.h>
+#include <linux/clockchips.h>
+#include <linux/hardirq.h>
+#include <linux/sched.h>
+#include <linux/smp.h>
+#include <linux/delay.h>
+#include <linux/module.h>
+#include <linux/timekeeper_internal.h>
+#include <asm/irq_regs.h>
+#include <asm/traps.h>
+#include <asm/vdso.h>
+#include <hv/hypervisor.h>
+#include <arch/interrupts.h>
+#include <arch/spr_def.h>
+
+
+/*
+ * Define the cycle counter clock source.
+ */
+
+/* How many cycles per second we are running at. */
+static cycles_t cycles_per_sec __write_once;
+
+cycles_t get_clock_rate(void)
+{
+	return cycles_per_sec;
+}
+
+#if CHIP_HAS_SPLIT_CYCLE()
+cycles_t get_cycles(void)
+{
+	unsigned int high = __insn_mfspr(SPR_CYCLE_HIGH);
+	unsigned int low = __insn_mfspr(SPR_CYCLE_LOW);
+	unsigned int high2 = __insn_mfspr(SPR_CYCLE_HIGH);
+
+	while (unlikely(high != high2)) {
+		low = __insn_mfspr(SPR_CYCLE_LOW);
+		high = high2;
+		high2 = __insn_mfspr(SPR_CYCLE_HIGH);
+	}
+
+	return (((cycles_t)high) << 32) | low;
+}
+EXPORT_SYMBOL(get_cycles);
+#endif
+
+/*
+ * We use a relatively small shift value so that sched_clock()
+ * won't wrap around very often.
+ */
+#define SCHED_CLOCK_SHIFT 10
+
+static unsigned long sched_clock_mult __write_once;
+
+static cycles_t clocksource_get_cycles(struct clocksource *cs)
+{
+	return get_cycles();
+}
+
+static struct clocksource cycle_counter_cs = {
+	.name = "cycle counter",
+	.rating = 300,
+	.read = clocksource_get_cycles,
+	.mask = CLOCKSOURCE_MASK(64),
+	.flags = CLOCK_SOURCE_IS_CONTINUOUS,
+};
+
+/*
+ * Called very early from setup_arch() to set cycles_per_sec.
+ * We initialize it early so we can use it to set up loops_per_jiffy.
+ */
+void __init setup_clock(void)
+{
+	cycles_per_sec = hv_sysconf(HV_SYSCONF_CPU_SPEED);
+	sched_clock_mult =
+		clocksource_hz2mult(cycles_per_sec, SCHED_CLOCK_SHIFT);
+}
+
+void __init calibrate_delay(void)
+{
+	loops_per_jiffy = get_clock_rate() / HZ;
+	pr_info("Clock rate yields %lu.%02lu BogoMIPS (lpj=%lu)\n",
+		loops_per_jiffy / (500000 / HZ),
+		(loops_per_jiffy / (5000 / HZ)) % 100, loops_per_jiffy);
+}
+
+/* Called fairly late in init/main.c, but before we go smp. */
+void __init time_init(void)
+{
+	/* Initialize and register the clock source. */
+	clocksource_register_hz(&cycle_counter_cs, cycles_per_sec);
+
+	/* Start up the tile-timer interrupt source on the boot cpu. */
+	setup_tile_timer();
+}
+
+/*
+ * Define the tile timer clock event device.  The timer is driven by
+ * the TILE_TIMER_CONTROL register, which consists of a 31-bit down
+ * counter, plus bit 31, which signifies that the counter has wrapped
+ * from zero to (2**31) - 1.  The INT_TILE_TIMER interrupt will be
+ * raised as long as bit 31 is set.
+ *
+ * The TILE_MINSEC value represents the largest range of real-time
+ * we can possibly cover with the timer, based on MAX_TICK combined
+ * with the slowest reasonable clock rate we might run at.
+ */
+
+#define MAX_TICK 0x7fffffff   /* we have 31 bits of countdown timer */
+#define TILE_MINSEC 5         /* timer covers no more than 5 seconds */
+
+static int tile_timer_set_next_event(unsigned long ticks,
+				     struct clock_event_device *evt)
+{
+	BUG_ON(ticks > MAX_TICK);
+	__insn_mtspr(SPR_TILE_TIMER_CONTROL, ticks);
+	arch_local_irq_unmask_now(INT_TILE_TIMER);
+	return 0;
+}
+
+/*
+ * Whenever anyone tries to change modes, we just mask interrupts
+ * and wait for the next event to get set.
+ */
+static void tile_timer_set_mode(enum clock_event_mode mode,
+				struct clock_event_device *evt)
+{
+	arch_local_irq_mask_now(INT_TILE_TIMER);
+}
+
+/*
+ * Set min_delta_ns to 1 microsecond, since it takes about
+ * that long to fire the interrupt.
+ */
+static DEFINE_PER_CPU(struct clock_event_device, tile_timer) = {
+	.name = "tile timer",
+	.features = CLOCK_EVT_FEAT_ONESHOT,
+	.min_delta_ns = 1000,
+	.rating = 100,
+	.irq = -1,
+	.set_next_event = tile_timer_set_next_event,
+	.set_mode = tile_timer_set_mode,
+};
+
+void setup_tile_timer(void)
+{
+	struct clock_event_device *evt = this_cpu_ptr(&tile_timer);
+
+	/* Fill in fields that are speed-specific. */
+	clockevents_calc_mult_shift(evt, cycles_per_sec, TILE_MINSEC);
+	evt->max_delta_ns = clockevent_delta2ns(MAX_TICK, evt);
+
+	/* Mark as being for this cpu only. */
+	evt->cpumask = cpumask_of(smp_processor_id());
+
+	/* Start out with timer not firing. */
+	arch_local_irq_mask_now(INT_TILE_TIMER);
+
+	/* Register tile timer. */
+	clockevents_register_device(evt);
+}
+
+/* Called from the interrupt vector. */
+void do_timer_interrupt(struct pt_regs *regs, int fault_num)
+{
+	struct pt_regs *old_regs = set_irq_regs(regs);
+	struct clock_event_device *evt = this_cpu_ptr(&tile_timer);
+
+	/*
+	 * Mask the timer interrupt here, since we are a oneshot timer
+	 * and there are now by definition no events pending.
+	 */
+	arch_local_irq_mask(INT_TILE_TIMER);
+
+	/* Track time spent here in an interrupt context */
+	irq_enter();
+
+	/* Track interrupt count. */
+	__this_cpu_inc(irq_stat.irq_timer_count);
+
+	/* Call the generic timer handler */
+	evt->event_handler(evt);
+
+	/*
+	 * Track time spent against the current process again and
+	 * process any softirqs if they are waiting.
+	 */
+	irq_exit();
+
+	set_irq_regs(old_regs);
+}
+
+/*
+ * Scheduler clock - returns current time in nanosec units.
+ * Note that with LOCKDEP, this is called during lockdep_init(), and
+ * we will claim that sched_clock() is zero for a little while, until
+ * we run setup_clock(), above.
+ */
+unsigned long long sched_clock(void)
+{
+	return clocksource_cyc2ns(get_cycles(),
+				  sched_clock_mult, SCHED_CLOCK_SHIFT);
+}
+
+int setup_profiling_timer(unsigned int multiplier)
+{
+	return -EINVAL;
+}
+
+/*
+ * Use the tile timer to convert nsecs to core clock cycles, relying
+ * on it having the same frequency as SPR_CYCLE.
+ */
+cycles_t ns2cycles(unsigned long nsecs)
+{
+	/*
+	 * We do not have to disable preemption here as each core has the same
+	 * clock frequency.
+	 */
+	struct clock_event_device *dev = raw_cpu_ptr(&tile_timer);
+
+	/*
+	 * as in clocksource.h and x86's timer.h, we split the calculation
+	 * into 2 parts to avoid unecessary overflow of the intermediate
+	 * value. This will not lead to any loss of precision.
+	 */
+	u64 quot = (u64)nsecs >> dev->shift;
+	u64 rem  = (u64)nsecs & ((1ULL << dev->shift) - 1);
+	return quot * dev->mult + ((rem * dev->mult) >> dev->shift);
+}
+
+void update_vsyscall_tz(void)
+{
+	write_seqcount_begin(&vdso_data->tz_seq);
+	vdso_data->tz_minuteswest = sys_tz.tz_minuteswest;
+	vdso_data->tz_dsttime = sys_tz.tz_dsttime;
+	write_seqcount_end(&vdso_data->tz_seq);
+}
+
+void update_vsyscall(struct timekeeper *tk)
+{
+	if (tk->tkr_mono.clock != &cycle_counter_cs)
+		return;
+
+	write_seqcount_begin(&vdso_data->tb_seq);
+
+	vdso_data->cycle_last		= tk->tkr_mono.cycle_last;
+	vdso_data->mask			= tk->tkr_mono.mask;
+	vdso_data->mult			= tk->tkr_mono.mult;
+	vdso_data->shift		= tk->tkr_mono.shift;
+
+	vdso_data->wall_time_sec	= tk->xtime_sec;
+	vdso_data->wall_time_snsec	= tk->tkr_mono.xtime_nsec;
+
+	vdso_data->monotonic_time_sec	= tk->xtime_sec
+					+ tk->wall_to_monotonic.tv_sec;
+	vdso_data->monotonic_time_snsec	= tk->tkr_mono.xtime_nsec
+					+ ((u64)tk->wall_to_monotonic.tv_nsec
+						<< tk->tkr_mono.shift);
+	while (vdso_data->monotonic_time_snsec >=
+					(((u64)NSEC_PER_SEC) << tk->tkr_mono.shift)) {
+		vdso_data->monotonic_time_snsec -=
+					((u64)NSEC_PER_SEC) << tk->tkr_mono.shift;
+		vdso_data->monotonic_time_sec++;
+	}
+
+	vdso_data->wall_time_coarse_sec	= tk->xtime_sec;
+	vdso_data->wall_time_coarse_nsec = (long)(tk->tkr_mono.xtime_nsec >>
+						 tk->tkr_mono.shift);
+
+	vdso_data->monotonic_time_coarse_sec =
+		vdso_data->wall_time_coarse_sec + tk->wall_to_monotonic.tv_sec;
+	vdso_data->monotonic_time_coarse_nsec =
+		vdso_data->wall_time_coarse_nsec + tk->wall_to_monotonic.tv_nsec;
+
+	while (vdso_data->monotonic_time_coarse_nsec >= NSEC_PER_SEC) {
+		vdso_data->monotonic_time_coarse_nsec -= NSEC_PER_SEC;
+		vdso_data->monotonic_time_coarse_sec++;
+	}
+
+	write_seqcount_end(&vdso_data->tb_seq);
+}
diff --git a/arch/tile/kernel/tlb.c b/arch/tile/kernel/tlb.c
new file mode 100644
index 000000000..f23b53515
--- /dev/null
+++ b/arch/tile/kernel/tlb.c
@@ -0,0 +1,104 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ */
+
+#include <linux/cpumask.h>
+#include <linux/module.h>
+#include <linux/hugetlb.h>
+#include <asm/tlbflush.h>
+#include <asm/homecache.h>
+#include <hv/hypervisor.h>
+
+/* From tlbflush.h */
+DEFINE_PER_CPU(int, current_asid);
+int min_asid, max_asid;
+
+/*
+ * Note that we flush the L1I (for VM_EXEC pages) as well as the TLB
+ * so that when we are unmapping an executable page, we also flush it.
+ * Combined with flushing the L1I at context switch time, this means
+ * we don't have to do any other icache flushes.
+ */
+
+void flush_tlb_mm(struct mm_struct *mm)
+{
+	HV_Remote_ASID asids[NR_CPUS];
+	int i = 0, cpu;
+	for_each_cpu(cpu, mm_cpumask(mm)) {
+		HV_Remote_ASID *asid = &asids[i++];
+		asid->y = cpu / smp_topology.width;
+		asid->x = cpu % smp_topology.width;
+		asid->asid = per_cpu(current_asid, cpu);
+	}
+	flush_remote(0, HV_FLUSH_EVICT_L1I, mm_cpumask(mm),
+		     0, 0, 0, NULL, asids, i);
+}
+
+void flush_tlb_current_task(void)
+{
+	flush_tlb_mm(current->mm);
+}
+
+void flush_tlb_page_mm(struct vm_area_struct *vma, struct mm_struct *mm,
+		       unsigned long va)
+{
+	unsigned long size = vma_kernel_pagesize(vma);
+	int cache = (vma->vm_flags & VM_EXEC) ? HV_FLUSH_EVICT_L1I : 0;
+	flush_remote(0, cache, mm_cpumask(mm),
+		     va, size, size, mm_cpumask(mm), NULL, 0);
+}
+
+void flush_tlb_page(struct vm_area_struct *vma, unsigned long va)
+{
+	flush_tlb_page_mm(vma, vma->vm_mm, va);
+}
+EXPORT_SYMBOL(flush_tlb_page);
+
+void flush_tlb_range(struct vm_area_struct *vma,
+		     unsigned long start, unsigned long end)
+{
+	unsigned long size = vma_kernel_pagesize(vma);
+	struct mm_struct *mm = vma->vm_mm;
+	int cache = (vma->vm_flags & VM_EXEC) ? HV_FLUSH_EVICT_L1I : 0;
+	flush_remote(0, cache, mm_cpumask(mm), start, end - start, size,
+		     mm_cpumask(mm), NULL, 0);
+}
+
+void flush_tlb_all(void)
+{
+	int i;
+	for (i = 0; ; ++i) {
+		HV_VirtAddrRange r = hv_inquire_virtual(i);
+		if (r.size == 0)
+			break;
+		flush_remote(0, HV_FLUSH_EVICT_L1I, cpu_online_mask,
+			     r.start, r.size, PAGE_SIZE, cpu_online_mask,
+			     NULL, 0);
+		flush_remote(0, 0, NULL,
+			     r.start, r.size, HPAGE_SIZE, cpu_online_mask,
+			     NULL, 0);
+	}
+}
+
+/*
+ * Callers need to flush the L1I themselves if necessary, e.g. for
+ * kernel module unload.  Otherwise we assume callers are not using
+ * executable pgprot_t's.  Using EVICT_L1I means that dataplane cpus
+ * will get an unnecessary interrupt otherwise.
+ */
+void flush_tlb_kernel_range(unsigned long start, unsigned long end)
+{
+	flush_remote(0, 0, NULL,
+		     start, end - start, PAGE_SIZE, cpu_online_mask, NULL, 0);
+}
diff --git a/arch/tile/kernel/traps.c b/arch/tile/kernel/traps.c
new file mode 100644
index 000000000..312fc134c
--- /dev/null
+++ b/arch/tile/kernel/traps.c
@@ -0,0 +1,403 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/sched.h>
+#include <linux/kernel.h>
+#include <linux/kprobes.h>
+#include <linux/kdebug.h>
+#include <linux/module.h>
+#include <linux/reboot.h>
+#include <linux/uaccess.h>
+#include <linux/ptrace.h>
+#include <linux/context_tracking.h>
+#include <asm/stack.h>
+#include <asm/traps.h>
+#include <asm/setup.h>
+
+#include <arch/interrupts.h>
+#include <arch/spr_def.h>
+#include <arch/opcode.h>
+
+void __init trap_init(void)
+{
+	/* Nothing needed here since we link code at .intrpt */
+}
+
+int unaligned_fixup = 1;
+
+static int __init setup_unaligned_fixup(char *str)
+{
+	/*
+	 * Say "=-1" to completely disable it.  If you just do "=0", we
+	 * will still parse the instruction, then fire a SIGBUS with
+	 * the correct address from inside the single_step code.
+	 */
+	if (kstrtoint(str, 0, &unaligned_fixup) != 0)
+		return 0;
+
+	pr_info("Fixups for unaligned data accesses are %s\n",
+		unaligned_fixup >= 0 ?
+		(unaligned_fixup ? "enabled" : "disabled") :
+		"completely disabled");
+	return 1;
+}
+__setup("unaligned_fixup=", setup_unaligned_fixup);
+
+#if CHIP_HAS_TILE_DMA()
+
+static int dma_disabled;
+
+static int __init nodma(char *str)
+{
+	pr_info("User-space DMA is disabled\n");
+	dma_disabled = 1;
+	return 1;
+}
+__setup("nodma", nodma);
+
+/* How to decode SPR_GPV_REASON */
+#define IRET_ERROR (1U << 31)
+#define MT_ERROR   (1U << 30)
+#define MF_ERROR   (1U << 29)
+#define SPR_INDEX  ((1U << 15) - 1)
+#define SPR_MPL_SHIFT  9  /* starting bit position for MPL encoded in SPR */
+
+/*
+ * See if this GPV is just to notify the kernel of SPR use and we can
+ * retry the user instruction after adjusting some MPLs suitably.
+ */
+static int retry_gpv(unsigned int gpv_reason)
+{
+	int mpl;
+
+	if (gpv_reason & IRET_ERROR)
+		return 0;
+
+	BUG_ON((gpv_reason & (MT_ERROR|MF_ERROR)) == 0);
+	mpl = (gpv_reason & SPR_INDEX) >> SPR_MPL_SHIFT;
+	if (mpl == INT_DMA_NOTIFY && !dma_disabled) {
+		/* User is turning on DMA. Allow it and retry. */
+		printk(KERN_DEBUG "Process %d/%s is now enabled for DMA\n",
+		       current->pid, current->comm);
+		BUG_ON(current->thread.tile_dma_state.enabled);
+		current->thread.tile_dma_state.enabled = 1;
+		grant_dma_mpls();
+		return 1;
+	}
+
+	return 0;
+}
+
+#endif /* CHIP_HAS_TILE_DMA() */
+
+extern tile_bundle_bits bpt_code;
+
+asm(".pushsection .rodata.bpt_code,\"a\";"
+    ".align 8;"
+    "bpt_code: bpt;"
+    ".size bpt_code,.-bpt_code;"
+    ".popsection");
+
+static int special_ill(tile_bundle_bits bundle, int *sigp, int *codep)
+{
+	int sig, code, maxcode;
+
+	if (bundle == bpt_code) {
+		*sigp = SIGTRAP;
+		*codep = TRAP_BRKPT;
+		return 1;
+	}
+
+	/* If it's a "raise" bundle, then "ill" must be in pipe X1. */
+#ifdef __tilegx__
+	if ((bundle & TILEGX_BUNDLE_MODE_MASK) != 0)
+		return 0;
+	if (get_Opcode_X1(bundle) != RRR_0_OPCODE_X1)
+		return 0;
+	if (get_RRROpcodeExtension_X1(bundle) != UNARY_RRR_0_OPCODE_X1)
+		return 0;
+	if (get_UnaryOpcodeExtension_X1(bundle) != ILL_UNARY_OPCODE_X1)
+		return 0;
+#else
+	if (bundle & TILEPRO_BUNDLE_Y_ENCODING_MASK)
+		return 0;
+	if (get_Opcode_X1(bundle) != SHUN_0_OPCODE_X1)
+		return 0;
+	if (get_UnShOpcodeExtension_X1(bundle) != UN_0_SHUN_0_OPCODE_X1)
+		return 0;
+	if (get_UnOpcodeExtension_X1(bundle) != ILL_UN_0_SHUN_0_OPCODE_X1)
+		return 0;
+#endif
+
+	/* Check that the magic distinguishers are set to mean "raise". */
+	if (get_Dest_X1(bundle) != 29 || get_SrcA_X1(bundle) != 37)
+		return 0;
+
+	/* There must be an "addli zero, zero, VAL" in X0. */
+	if (get_Opcode_X0(bundle) != ADDLI_OPCODE_X0)
+		return 0;
+	if (get_Dest_X0(bundle) != TREG_ZERO)
+		return 0;
+	if (get_SrcA_X0(bundle) != TREG_ZERO)
+		return 0;
+
+	/*
+	 * Validate the proposed signal number and si_code value.
+	 * Note that we embed these in the static instruction itself
+	 * so that we perturb the register state as little as possible
+	 * at the time of the actual fault; it's unlikely you'd ever
+	 * need to dynamically choose which kind of fault to raise
+	 * from user space.
+	 */
+	sig = get_Imm16_X0(bundle) & 0x3f;
+	switch (sig) {
+	case SIGILL:
+		maxcode = NSIGILL;
+		break;
+	case SIGFPE:
+		maxcode = NSIGFPE;
+		break;
+	case SIGSEGV:
+		maxcode = NSIGSEGV;
+		break;
+	case SIGBUS:
+		maxcode = NSIGBUS;
+		break;
+	case SIGTRAP:
+		maxcode = NSIGTRAP;
+		break;
+	default:
+		return 0;
+	}
+	code = (get_Imm16_X0(bundle) >> 6) & 0xf;
+	if (code <= 0 || code > maxcode)
+		return 0;
+
+	/* Make it the requested signal. */
+	*sigp = sig;
+	*codep = code | __SI_FAULT;
+	return 1;
+}
+
+static const char *const int_name[] = {
+	[INT_MEM_ERROR] = "Memory error",
+	[INT_ILL] = "Illegal instruction",
+	[INT_GPV] = "General protection violation",
+	[INT_UDN_ACCESS] = "UDN access",
+	[INT_IDN_ACCESS] = "IDN access",
+#if CHIP_HAS_SN()
+	[INT_SN_ACCESS] = "SN access",
+#endif
+	[INT_SWINT_3] = "Software interrupt 3",
+	[INT_SWINT_2] = "Software interrupt 2",
+	[INT_SWINT_0] = "Software interrupt 0",
+	[INT_UNALIGN_DATA] = "Unaligned data",
+	[INT_DOUBLE_FAULT] = "Double fault",
+#ifdef __tilegx__
+	[INT_ILL_TRANS] = "Illegal virtual address",
+#endif
+};
+
+static int do_bpt(struct pt_regs *regs)
+{
+	unsigned long bundle, bcode, bpt;
+
+	bundle = *(unsigned long *)instruction_pointer(regs);
+
+	/*
+	 * bpt shoule be { bpt; nop }, which is 0x286a44ae51485000ULL.
+	 * we encode the unused least significant bits for other purpose.
+	 */
+	bpt = bundle & ~((1ULL << 12) - 1);
+	if (bpt != TILE_BPT_BUNDLE)
+		return 0;
+
+	bcode = bundle & ((1ULL << 12) - 1);
+	/*
+	 * notify the kprobe handlers, if instruction is likely to
+	 * pertain to them.
+	 */
+	switch (bcode) {
+	/* breakpoint_insn */
+	case 0:
+		notify_die(DIE_BREAK, "debug", regs, bundle,
+			INT_ILL, SIGTRAP);
+		break;
+	/* compiled_bpt */
+	case DIE_COMPILED_BPT:
+		notify_die(DIE_COMPILED_BPT, "debug", regs, bundle,
+			INT_ILL, SIGTRAP);
+		break;
+	/* breakpoint2_insn */
+	case DIE_SSTEPBP:
+		notify_die(DIE_SSTEPBP, "single_step", regs, bundle,
+			INT_ILL, SIGTRAP);
+		break;
+	default:
+		return 0;
+	}
+
+	return 1;
+}
+
+void __kprobes do_trap(struct pt_regs *regs, int fault_num,
+		       unsigned long reason)
+{
+	enum ctx_state prev_state = exception_enter();
+	siginfo_t info = { 0 };
+	int signo, code;
+	unsigned long address = 0;
+	tile_bundle_bits instr;
+	int is_kernel = !user_mode(regs);
+
+	/* Handle breakpoints, etc. */
+	if (is_kernel && fault_num == INT_ILL && do_bpt(regs))
+		goto done;
+
+	/* Re-enable interrupts, if they were previously enabled. */
+	if (!(regs->flags & PT_FLAGS_DISABLE_IRQ))
+		local_irq_enable();
+
+	/*
+	 * If it hits in kernel mode and we can't fix it up, just exit the
+	 * current process and hope for the best.
+	 */
+	if (is_kernel) {
+		const char *name;
+		char buf[100];
+		if (fixup_exception(regs))  /* ILL_TRANS or UNALIGN_DATA */
+			goto done;
+		if (fault_num >= 0 &&
+		    fault_num < ARRAY_SIZE(int_name) &&
+		    int_name[fault_num] != NULL)
+			name = int_name[fault_num];
+		else
+			name = "Unknown interrupt";
+		if (fault_num == INT_GPV)
+			snprintf(buf, sizeof(buf), "; GPV_REASON %#lx", reason);
+#ifdef __tilegx__
+		else if (fault_num == INT_ILL_TRANS)
+			snprintf(buf, sizeof(buf), "; address %#lx", reason);
+#endif
+		else
+			buf[0] = '\0';
+		pr_alert("Kernel took bad trap %d (%s) at PC %#lx%s\n",
+			 fault_num, name, regs->pc, buf);
+		show_regs(regs);
+		do_exit(SIGKILL);  /* FIXME: implement i386 die() */
+	}
+
+	switch (fault_num) {
+	case INT_MEM_ERROR:
+		signo = SIGBUS;
+		code = BUS_OBJERR;
+		break;
+	case INT_ILL:
+		if (copy_from_user(&instr, (void __user *)regs->pc,
+				   sizeof(instr))) {
+			pr_err("Unreadable instruction for INT_ILL: %#lx\n",
+			       regs->pc);
+			do_exit(SIGKILL);
+		}
+		if (!special_ill(instr, &signo, &code)) {
+			signo = SIGILL;
+			code = ILL_ILLOPC;
+		}
+		address = regs->pc;
+		break;
+	case INT_GPV:
+#if CHIP_HAS_TILE_DMA()
+		if (retry_gpv(reason))
+			goto done;
+#endif
+		/*FALLTHROUGH*/
+	case INT_UDN_ACCESS:
+	case INT_IDN_ACCESS:
+#if CHIP_HAS_SN()
+	case INT_SN_ACCESS:
+#endif
+		signo = SIGILL;
+		code = ILL_PRVREG;
+		address = regs->pc;
+		break;
+	case INT_SWINT_3:
+	case INT_SWINT_2:
+	case INT_SWINT_0:
+		signo = SIGILL;
+		code = ILL_ILLTRP;
+		address = regs->pc;
+		break;
+	case INT_UNALIGN_DATA:
+#ifndef __tilegx__  /* Emulated support for single step debugging */
+		if (unaligned_fixup >= 0) {
+			struct single_step_state *state =
+				current_thread_info()->step_state;
+			if (!state ||
+			    (void __user *)(regs->pc) != state->buffer) {
+				single_step_once(regs);
+				goto done;
+			}
+		}
+#endif
+		signo = SIGBUS;
+		code = BUS_ADRALN;
+		address = 0;
+		break;
+	case INT_DOUBLE_FAULT:
+		/*
+		 * For double fault, "reason" is actually passed as
+		 * SYSTEM_SAVE_K_2, the hypervisor's double-fault info, so
+		 * we can provide the original fault number rather than
+		 * the uninteresting "INT_DOUBLE_FAULT" so the user can
+		 * learn what actually struck while PL0 ICS was set.
+		 */
+		fault_num = reason;
+		signo = SIGILL;
+		code = ILL_DBLFLT;
+		address = regs->pc;
+		break;
+#ifdef __tilegx__
+	case INT_ILL_TRANS: {
+		/* Avoid a hardware erratum with the return address stack. */
+		fill_ra_stack();
+
+		signo = SIGSEGV;
+		address = reason;
+		code = SEGV_MAPERR;
+		break;
+	}
+#endif
+	default:
+		panic("Unexpected do_trap interrupt number %d", fault_num);
+	}
+
+	info.si_signo = signo;
+	info.si_code = code;
+	info.si_addr = (void __user *)address;
+	if (signo == SIGILL)
+		info.si_trapno = fault_num;
+	if (signo != SIGTRAP)
+		trace_unhandled_signal("trap", regs, address, signo);
+	force_sig_info(signo, &info, current);
+
+done:
+	exception_exit(prev_state);
+}
+
+void kernel_double_fault(int dummy, ulong pc, ulong lr, ulong sp, ulong r52)
+{
+	_dump_stack(dummy, pc, lr, sp, r52);
+	pr_emerg("Double fault: exiting\n");
+	machine_halt();
+}
diff --git a/arch/tile/kernel/unaligned.c b/arch/tile/kernel/unaligned.c
new file mode 100644
index 000000000..d075f92cc
--- /dev/null
+++ b/arch/tile/kernel/unaligned.c
@@ -0,0 +1,1600 @@
+/*
+ * Copyright 2013 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * A code-rewriter that handles unaligned exception.
+ */
+
+#include <linux/smp.h>
+#include <linux/ptrace.h>
+#include <linux/slab.h>
+#include <linux/thread_info.h>
+#include <linux/uaccess.h>
+#include <linux/mman.h>
+#include <linux/types.h>
+#include <linux/err.h>
+#include <linux/module.h>
+#include <linux/compat.h>
+#include <linux/prctl.h>
+#include <linux/context_tracking.h>
+#include <asm/cacheflush.h>
+#include <asm/traps.h>
+#include <asm/uaccess.h>
+#include <asm/unaligned.h>
+#include <arch/abi.h>
+#include <arch/spr_def.h>
+#include <arch/opcode.h>
+
+
+/*
+ * This file handles unaligned exception for tile-Gx. The tilepro's unaligned
+ * exception is supported out of single_step.c
+ */
+
+int unaligned_printk;
+
+static int __init setup_unaligned_printk(char *str)
+{
+	long val;
+	if (kstrtol(str, 0, &val) != 0)
+		return 0;
+	unaligned_printk = val;
+	pr_info("Printk for each unaligned data accesses is %s\n",
+		unaligned_printk ? "enabled" : "disabled");
+	return 1;
+}
+__setup("unaligned_printk=", setup_unaligned_printk);
+
+unsigned int unaligned_fixup_count;
+
+#ifdef __tilegx__
+
+/*
+ * Unalign data jit fixup code fragement. Reserved space is 128 bytes.
+ * The 1st 64-bit word saves fault PC address, 2nd word is the fault
+ * instruction bundle followed by 14 JIT bundles.
+ */
+
+struct unaligned_jit_fragment {
+	unsigned long       pc;
+	tilegx_bundle_bits  bundle;
+	tilegx_bundle_bits  insn[14];
+};
+
+/*
+ * Check if a nop or fnop at bundle's pipeline X0.
+ */
+
+static bool is_bundle_x0_nop(tilegx_bundle_bits bundle)
+{
+	return (((get_UnaryOpcodeExtension_X0(bundle) ==
+		  NOP_UNARY_OPCODE_X0) &&
+		 (get_RRROpcodeExtension_X0(bundle) ==
+		  UNARY_RRR_0_OPCODE_X0) &&
+		 (get_Opcode_X0(bundle) ==
+		  RRR_0_OPCODE_X0)) ||
+		((get_UnaryOpcodeExtension_X0(bundle) ==
+		  FNOP_UNARY_OPCODE_X0) &&
+		 (get_RRROpcodeExtension_X0(bundle) ==
+		  UNARY_RRR_0_OPCODE_X0) &&
+		 (get_Opcode_X0(bundle) ==
+		  RRR_0_OPCODE_X0)));
+}
+
+/*
+ * Check if nop or fnop at bundle's pipeline X1.
+ */
+
+static bool is_bundle_x1_nop(tilegx_bundle_bits bundle)
+{
+	return (((get_UnaryOpcodeExtension_X1(bundle) ==
+		  NOP_UNARY_OPCODE_X1) &&
+		 (get_RRROpcodeExtension_X1(bundle) ==
+		  UNARY_RRR_0_OPCODE_X1) &&
+		 (get_Opcode_X1(bundle) ==
+		  RRR_0_OPCODE_X1)) ||
+		((get_UnaryOpcodeExtension_X1(bundle) ==
+		  FNOP_UNARY_OPCODE_X1) &&
+		 (get_RRROpcodeExtension_X1(bundle) ==
+		  UNARY_RRR_0_OPCODE_X1) &&
+		 (get_Opcode_X1(bundle) ==
+		  RRR_0_OPCODE_X1)));
+}
+
+/*
+ * Check if nop or fnop at bundle's Y0 pipeline.
+ */
+
+static bool is_bundle_y0_nop(tilegx_bundle_bits bundle)
+{
+	return (((get_UnaryOpcodeExtension_Y0(bundle) ==
+		  NOP_UNARY_OPCODE_Y0) &&
+		 (get_RRROpcodeExtension_Y0(bundle) ==
+		  UNARY_RRR_1_OPCODE_Y0) &&
+		 (get_Opcode_Y0(bundle) ==
+		  RRR_1_OPCODE_Y0)) ||
+		((get_UnaryOpcodeExtension_Y0(bundle) ==
+		  FNOP_UNARY_OPCODE_Y0) &&
+		 (get_RRROpcodeExtension_Y0(bundle) ==
+		  UNARY_RRR_1_OPCODE_Y0) &&
+		 (get_Opcode_Y0(bundle) ==
+		  RRR_1_OPCODE_Y0)));
+}
+
+/*
+ * Check if nop or fnop at bundle's pipeline Y1.
+ */
+
+static bool is_bundle_y1_nop(tilegx_bundle_bits bundle)
+{
+	return (((get_UnaryOpcodeExtension_Y1(bundle) ==
+		  NOP_UNARY_OPCODE_Y1) &&
+		 (get_RRROpcodeExtension_Y1(bundle) ==
+		  UNARY_RRR_1_OPCODE_Y1) &&
+		 (get_Opcode_Y1(bundle) ==
+		  RRR_1_OPCODE_Y1)) ||
+		((get_UnaryOpcodeExtension_Y1(bundle) ==
+		  FNOP_UNARY_OPCODE_Y1) &&
+		 (get_RRROpcodeExtension_Y1(bundle) ==
+		  UNARY_RRR_1_OPCODE_Y1) &&
+		 (get_Opcode_Y1(bundle) ==
+		  RRR_1_OPCODE_Y1)));
+}
+
+/*
+ * Test if a bundle's y0 and y1 pipelines are both nop or fnop.
+ */
+
+static bool is_y0_y1_nop(tilegx_bundle_bits bundle)
+{
+	return is_bundle_y0_nop(bundle) && is_bundle_y1_nop(bundle);
+}
+
+/*
+ * Test if a bundle's x0 and x1 pipelines are both nop or fnop.
+ */
+
+static bool is_x0_x1_nop(tilegx_bundle_bits bundle)
+{
+	return is_bundle_x0_nop(bundle) && is_bundle_x1_nop(bundle);
+}
+
+/*
+ * Find the destination, source registers of fault unalign access instruction
+ * at X1 or Y2. Also, allocate up to 3 scratch registers clob1, clob2 and
+ * clob3, which are guaranteed different from any register used in the fault
+ * bundle. r_alias is used to return if the other instructions other than the
+ * unalign load/store shares same register with ra, rb and rd.
+ */
+
+static void find_regs(tilegx_bundle_bits bundle, uint64_t *rd, uint64_t *ra,
+		      uint64_t *rb, uint64_t *clob1, uint64_t *clob2,
+		      uint64_t *clob3, bool *r_alias)
+{
+	int i;
+	uint64_t reg;
+	uint64_t reg_map = 0, alias_reg_map = 0, map;
+	bool alias = false;
+
+	/*
+	 * Parse fault bundle, find potential used registers and mark
+	 * corresponding bits in reg_map and alias_map. These 2 bit maps
+	 * are used to find the scratch registers and determine if there
+	 * is register alais.
+	 */
+	if (bundle & TILEGX_BUNDLE_MODE_MASK) {  /* Y Mode Bundle. */
+
+		reg = get_SrcA_Y2(bundle);
+		reg_map |= 1ULL << reg;
+		*ra = reg;
+		reg = get_SrcBDest_Y2(bundle);
+		reg_map |= 1ULL << reg;
+
+		if (rd) {
+			/* Load. */
+			*rd = reg;
+			alias_reg_map = (1ULL << *rd) | (1ULL << *ra);
+		} else {
+			/* Store. */
+			*rb = reg;
+			alias_reg_map = (1ULL << *ra) | (1ULL << *rb);
+		}
+
+		if (!is_bundle_y1_nop(bundle)) {
+			reg = get_SrcA_Y1(bundle);
+			reg_map |= (1ULL << reg);
+			map = (1ULL << reg);
+
+			reg = get_SrcB_Y1(bundle);
+			reg_map |= (1ULL << reg);
+			map |= (1ULL << reg);
+
+			reg = get_Dest_Y1(bundle);
+			reg_map |= (1ULL << reg);
+			map |= (1ULL << reg);
+
+			if (map & alias_reg_map)
+				alias = true;
+		}
+
+		if (!is_bundle_y0_nop(bundle)) {
+			reg = get_SrcA_Y0(bundle);
+			reg_map |= (1ULL << reg);
+			map = (1ULL << reg);
+
+			reg = get_SrcB_Y0(bundle);
+			reg_map |= (1ULL << reg);
+			map |= (1ULL << reg);
+
+			reg = get_Dest_Y0(bundle);
+			reg_map |= (1ULL << reg);
+			map |= (1ULL << reg);
+
+			if (map & alias_reg_map)
+				alias = true;
+		}
+	} else	{ /* X Mode Bundle. */
+
+		reg = get_SrcA_X1(bundle);
+		reg_map |= (1ULL << reg);
+		*ra = reg;
+		if (rd)	{
+			/* Load. */
+			reg = get_Dest_X1(bundle);
+			reg_map |= (1ULL << reg);
+			*rd = reg;
+			alias_reg_map = (1ULL << *rd) | (1ULL << *ra);
+		} else {
+			/* Store. */
+			reg = get_SrcB_X1(bundle);
+			reg_map |= (1ULL << reg);
+			*rb = reg;
+			alias_reg_map = (1ULL << *ra) | (1ULL << *rb);
+		}
+
+		if (!is_bundle_x0_nop(bundle)) {
+			reg = get_SrcA_X0(bundle);
+			reg_map |= (1ULL << reg);
+			map = (1ULL << reg);
+
+			reg = get_SrcB_X0(bundle);
+			reg_map |= (1ULL << reg);
+			map |= (1ULL << reg);
+
+			reg = get_Dest_X0(bundle);
+			reg_map |= (1ULL << reg);
+			map |= (1ULL << reg);
+
+			if (map & alias_reg_map)
+				alias = true;
+		}
+	}
+
+	/*
+	 * "alias" indicates if the unalign access registers have collision
+	 * with others in the same bundle. We jsut simply test all register
+	 * operands case (RRR), ignored the case with immidate. If a bundle
+	 * has no register alias, we may do fixup in a simple or fast manner.
+	 * So if an immidata field happens to hit with a register, we may end
+	 * up fall back to the generic handling.
+	 */
+
+	*r_alias = alias;
+
+	/* Flip bits on reg_map. */
+	reg_map ^= -1ULL;
+
+	/* Scan reg_map lower 54(TREG_SP) bits to find 3 set bits. */
+	for (i = 0; i < TREG_SP; i++) {
+		if (reg_map & (0x1ULL << i)) {
+			if (*clob1 == -1) {
+				*clob1 = i;
+			} else if (*clob2 == -1) {
+				*clob2 = i;
+			} else if (*clob3 == -1) {
+				*clob3 = i;
+				return;
+			}
+		}
+	}
+}
+
+/*
+ * Sanity check for register ra, rb, rd, clob1/2/3. Return true if any of them
+ * is unexpected.
+ */
+
+static bool check_regs(uint64_t rd, uint64_t ra, uint64_t rb,
+		       uint64_t clob1, uint64_t clob2,  uint64_t clob3)
+{
+	bool unexpected = false;
+	if ((ra >= 56) && (ra != TREG_ZERO))
+		unexpected = true;
+
+	if ((clob1 >= 56) || (clob2 >= 56) || (clob3 >= 56))
+		unexpected = true;
+
+	if (rd != -1) {
+		if ((rd >= 56) && (rd != TREG_ZERO))
+			unexpected = true;
+	} else {
+		if ((rb >= 56) && (rb != TREG_ZERO))
+			unexpected = true;
+	}
+	return unexpected;
+}
+
+
+#define  GX_INSN_X0_MASK   ((1ULL << 31) - 1)
+#define  GX_INSN_X1_MASK   (((1ULL << 31) - 1) << 31)
+#define  GX_INSN_Y0_MASK   ((0xFULL << 27) | (0xFFFFFULL))
+#define  GX_INSN_Y1_MASK   (GX_INSN_Y0_MASK << 31)
+#define  GX_INSN_Y2_MASK   ((0x7FULL << 51) | (0x7FULL << 20))
+
+#ifdef __LITTLE_ENDIAN
+#define  GX_INSN_BSWAP(_bundle_)    (_bundle_)
+#else
+#define  GX_INSN_BSWAP(_bundle_)    swab64(_bundle_)
+#endif /* __LITTLE_ENDIAN */
+
+/*
+ * __JIT_CODE(.) creates template bundles in .rodata.unalign_data section.
+ * The corresponding static function jix_x#_###(.) generates partial or
+ * whole bundle based on the template and given arguments.
+ */
+
+#define __JIT_CODE(_X_)						\
+	asm (".pushsection .rodata.unalign_data, \"a\"\n"	\
+	     _X_"\n"						\
+	     ".popsection\n")
+
+__JIT_CODE("__unalign_jit_x1_mtspr:   {mtspr 0,  r0}");
+static tilegx_bundle_bits jit_x1_mtspr(int spr, int reg)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x1_mtspr;
+	return (GX_INSN_BSWAP(__unalign_jit_x1_mtspr) & GX_INSN_X1_MASK) |
+		create_MT_Imm14_X1(spr) | create_SrcA_X1(reg);
+}
+
+__JIT_CODE("__unalign_jit_x1_mfspr:   {mfspr r0, 0}");
+static tilegx_bundle_bits  jit_x1_mfspr(int reg, int spr)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x1_mfspr;
+	return (GX_INSN_BSWAP(__unalign_jit_x1_mfspr) & GX_INSN_X1_MASK) |
+		create_MF_Imm14_X1(spr) | create_Dest_X1(reg);
+}
+
+__JIT_CODE("__unalign_jit_x0_addi:   {addi  r0, r0, 0; iret}");
+static tilegx_bundle_bits  jit_x0_addi(int rd, int ra, int imm8)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x0_addi;
+	return (GX_INSN_BSWAP(__unalign_jit_x0_addi) & GX_INSN_X0_MASK) |
+		create_Dest_X0(rd) | create_SrcA_X0(ra) |
+		create_Imm8_X0(imm8);
+}
+
+__JIT_CODE("__unalign_jit_x1_ldna:   {ldna  r0, r0}");
+static tilegx_bundle_bits  jit_x1_ldna(int rd, int ra)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x1_ldna;
+	return (GX_INSN_BSWAP(__unalign_jit_x1_ldna) &  GX_INSN_X1_MASK) |
+		create_Dest_X1(rd) | create_SrcA_X1(ra);
+}
+
+__JIT_CODE("__unalign_jit_x0_dblalign:   {dblalign r0, r0 ,r0}");
+static tilegx_bundle_bits  jit_x0_dblalign(int rd, int ra, int rb)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x0_dblalign;
+	return (GX_INSN_BSWAP(__unalign_jit_x0_dblalign) & GX_INSN_X0_MASK) |
+		create_Dest_X0(rd) | create_SrcA_X0(ra) |
+		create_SrcB_X0(rb);
+}
+
+__JIT_CODE("__unalign_jit_x1_iret:   {iret}");
+static tilegx_bundle_bits  jit_x1_iret(void)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x1_iret;
+	return GX_INSN_BSWAP(__unalign_jit_x1_iret) & GX_INSN_X1_MASK;
+}
+
+__JIT_CODE("__unalign_jit_x01_fnop:   {fnop;fnop}");
+static tilegx_bundle_bits  jit_x0_fnop(void)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x01_fnop;
+	return GX_INSN_BSWAP(__unalign_jit_x01_fnop) & GX_INSN_X0_MASK;
+}
+
+static tilegx_bundle_bits  jit_x1_fnop(void)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x01_fnop;
+	return GX_INSN_BSWAP(__unalign_jit_x01_fnop) & GX_INSN_X1_MASK;
+}
+
+__JIT_CODE("__unalign_jit_y2_dummy:   {fnop; fnop; ld zero, sp}");
+static tilegx_bundle_bits  jit_y2_dummy(void)
+{
+	extern  tilegx_bundle_bits __unalign_jit_y2_dummy;
+	return GX_INSN_BSWAP(__unalign_jit_y2_dummy) & GX_INSN_Y2_MASK;
+}
+
+static tilegx_bundle_bits  jit_y1_fnop(void)
+{
+	extern  tilegx_bundle_bits __unalign_jit_y2_dummy;
+	return GX_INSN_BSWAP(__unalign_jit_y2_dummy) & GX_INSN_Y1_MASK;
+}
+
+__JIT_CODE("__unalign_jit_x1_st1_add:  {st1_add r1, r0, 0}");
+static tilegx_bundle_bits  jit_x1_st1_add(int ra, int rb, int imm8)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x1_st1_add;
+	return (GX_INSN_BSWAP(__unalign_jit_x1_st1_add) &
+		(~create_SrcA_X1(-1)) &
+		GX_INSN_X1_MASK) | create_SrcA_X1(ra) |
+		create_SrcB_X1(rb) | create_Dest_Imm8_X1(imm8);
+}
+
+__JIT_CODE("__unalign_jit_x1_st:  {crc32_8 r1, r0, r0; st  r0, r0}");
+static tilegx_bundle_bits  jit_x1_st(int ra, int rb)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x1_st;
+	return (GX_INSN_BSWAP(__unalign_jit_x1_st) & GX_INSN_X1_MASK) |
+		create_SrcA_X1(ra) | create_SrcB_X1(rb);
+}
+
+__JIT_CODE("__unalign_jit_x1_st_add:  {st_add  r1, r0, 0}");
+static tilegx_bundle_bits  jit_x1_st_add(int ra, int rb, int imm8)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x1_st_add;
+	return (GX_INSN_BSWAP(__unalign_jit_x1_st_add) &
+		(~create_SrcA_X1(-1)) &
+		GX_INSN_X1_MASK) | create_SrcA_X1(ra) |
+		create_SrcB_X1(rb) | create_Dest_Imm8_X1(imm8);
+}
+
+__JIT_CODE("__unalign_jit_x1_ld:  {crc32_8 r1, r0, r0; ld  r0, r0}");
+static tilegx_bundle_bits  jit_x1_ld(int rd, int ra)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x1_ld;
+	return (GX_INSN_BSWAP(__unalign_jit_x1_ld) & GX_INSN_X1_MASK) |
+		create_Dest_X1(rd) | create_SrcA_X1(ra);
+}
+
+__JIT_CODE("__unalign_jit_x1_ld_add:  {ld_add  r1, r0, 0}");
+static tilegx_bundle_bits  jit_x1_ld_add(int rd, int ra, int imm8)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x1_ld_add;
+	return (GX_INSN_BSWAP(__unalign_jit_x1_ld_add) &
+		(~create_Dest_X1(-1)) &
+		GX_INSN_X1_MASK) | create_Dest_X1(rd) |
+		create_SrcA_X1(ra) | create_Imm8_X1(imm8);
+}
+
+__JIT_CODE("__unalign_jit_x0_bfexts:  {bfexts r0, r0, 0, 0}");
+static tilegx_bundle_bits  jit_x0_bfexts(int rd, int ra, int bfs, int bfe)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x0_bfexts;
+	return (GX_INSN_BSWAP(__unalign_jit_x0_bfexts) &
+		GX_INSN_X0_MASK) |
+		create_Dest_X0(rd) | create_SrcA_X0(ra) |
+		create_BFStart_X0(bfs) | create_BFEnd_X0(bfe);
+}
+
+__JIT_CODE("__unalign_jit_x0_bfextu:  {bfextu r0, r0, 0, 0}");
+static tilegx_bundle_bits  jit_x0_bfextu(int rd, int ra, int bfs, int bfe)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x0_bfextu;
+	return (GX_INSN_BSWAP(__unalign_jit_x0_bfextu) &
+		GX_INSN_X0_MASK) |
+		create_Dest_X0(rd) | create_SrcA_X0(ra) |
+		create_BFStart_X0(bfs) | create_BFEnd_X0(bfe);
+}
+
+__JIT_CODE("__unalign_jit_x1_addi:  {bfextu r1, r1, 0, 0; addi r0, r0, 0}");
+static tilegx_bundle_bits  jit_x1_addi(int rd, int ra, int imm8)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x1_addi;
+	return (GX_INSN_BSWAP(__unalign_jit_x1_addi) & GX_INSN_X1_MASK) |
+		create_Dest_X1(rd) | create_SrcA_X1(ra) |
+		create_Imm8_X1(imm8);
+}
+
+__JIT_CODE("__unalign_jit_x0_shrui:  {shrui r0, r0, 0; iret}");
+static tilegx_bundle_bits  jit_x0_shrui(int rd, int ra, int imm6)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x0_shrui;
+	return (GX_INSN_BSWAP(__unalign_jit_x0_shrui) &
+		GX_INSN_X0_MASK) |
+		create_Dest_X0(rd) | create_SrcA_X0(ra) |
+		create_ShAmt_X0(imm6);
+}
+
+__JIT_CODE("__unalign_jit_x0_rotli:  {rotli r0, r0, 0; iret}");
+static tilegx_bundle_bits  jit_x0_rotli(int rd, int ra, int imm6)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x0_rotli;
+	return (GX_INSN_BSWAP(__unalign_jit_x0_rotli) &
+		GX_INSN_X0_MASK) |
+		create_Dest_X0(rd) | create_SrcA_X0(ra) |
+		create_ShAmt_X0(imm6);
+}
+
+__JIT_CODE("__unalign_jit_x1_bnezt:  {bnezt r0, __unalign_jit_x1_bnezt}");
+static tilegx_bundle_bits  jit_x1_bnezt(int ra, int broff)
+{
+	extern  tilegx_bundle_bits __unalign_jit_x1_bnezt;
+	return (GX_INSN_BSWAP(__unalign_jit_x1_bnezt) &
+		GX_INSN_X1_MASK) |
+		create_SrcA_X1(ra) | create_BrOff_X1(broff);
+}
+
+#undef __JIT_CODE
+
+/*
+ * This function generates unalign fixup JIT.
+ *
+ * We first find unalign load/store instruction's destination, source
+ * registers: ra, rb and rd. and 3 scratch registers by calling
+ * find_regs(...). 3 scratch clobbers should not alias with any register
+ * used in the fault bundle. Then analyze the fault bundle to determine
+ * if it's a load or store, operand width, branch or address increment etc.
+ * At last generated JIT is copied into JIT code area in user space.
+ */
+
+static
+void jit_bundle_gen(struct pt_regs *regs, tilegx_bundle_bits bundle,
+		    int align_ctl)
+{
+	struct thread_info *info = current_thread_info();
+	struct unaligned_jit_fragment frag;
+	struct unaligned_jit_fragment *jit_code_area;
+	tilegx_bundle_bits bundle_2 = 0;
+	/* If bundle_2_enable = false, bundle_2 is fnop/nop operation. */
+	bool     bundle_2_enable = true;
+	uint64_t ra = -1, rb = -1, rd = -1, clob1 = -1, clob2 = -1, clob3 = -1;
+	/*
+	 * Indicate if the unalign access
+	 * instruction's registers hit with
+	 * others in the same bundle.
+	 */
+	bool     alias = false;
+	bool     load_n_store = true;
+	bool     load_store_signed = false;
+	unsigned int  load_store_size = 8;
+	bool     y1_br = false;  /* True, for a branch in same bundle at Y1.*/
+	int      y1_br_reg = 0;
+	/* True for link operation. i.e. jalr or lnk at Y1 */
+	bool     y1_lr = false;
+	int      y1_lr_reg = 0;
+	bool     x1_add = false;/* True, for load/store ADD instruction at X1*/
+	int      x1_add_imm8 = 0;
+	bool     unexpected = false;
+	int      n = 0, k;
+
+	jit_code_area =
+		(struct unaligned_jit_fragment *)(info->unalign_jit_base);
+
+	memset((void *)&frag, 0, sizeof(frag));
+
+	/* 0: X mode, Otherwise: Y mode. */
+	if (bundle & TILEGX_BUNDLE_MODE_MASK) {
+		unsigned int mod, opcode;
+
+		if (get_Opcode_Y1(bundle) == RRR_1_OPCODE_Y1 &&
+		    get_RRROpcodeExtension_Y1(bundle) ==
+		    UNARY_RRR_1_OPCODE_Y1) {
+
+			opcode = get_UnaryOpcodeExtension_Y1(bundle);
+
+			/*
+			 * Test "jalr", "jalrp", "jr", "jrp" instruction at Y1
+			 * pipeline.
+			 */
+			switch (opcode) {
+			case JALR_UNARY_OPCODE_Y1:
+			case JALRP_UNARY_OPCODE_Y1:
+				y1_lr = true;
+				y1_lr_reg = 55; /* Link register. */
+				/* FALLTHROUGH */
+			case JR_UNARY_OPCODE_Y1:
+			case JRP_UNARY_OPCODE_Y1:
+				y1_br = true;
+				y1_br_reg = get_SrcA_Y1(bundle);
+				break;
+			case LNK_UNARY_OPCODE_Y1:
+				/* "lnk" at Y1 pipeline. */
+				y1_lr = true;
+				y1_lr_reg = get_Dest_Y1(bundle);
+				break;
+			}
+		}
+
+		opcode = get_Opcode_Y2(bundle);
+		mod = get_Mode(bundle);
+
+		/*
+		 *  bundle_2 is bundle after making Y2 as a dummy operation
+		 *  - ld zero, sp
+		 */
+		bundle_2 = (bundle & (~GX_INSN_Y2_MASK)) | jit_y2_dummy();
+
+		/* Make Y1 as fnop if Y1 is a branch or lnk operation. */
+		if (y1_br || y1_lr) {
+			bundle_2 &= ~(GX_INSN_Y1_MASK);
+			bundle_2 |= jit_y1_fnop();
+		}
+
+		if (is_y0_y1_nop(bundle_2))
+			bundle_2_enable = false;
+
+		if (mod == MODE_OPCODE_YC2) {
+			/* Store. */
+			load_n_store = false;
+			load_store_size = 1 << opcode;
+			load_store_signed = false;
+			find_regs(bundle, 0, &ra, &rb, &clob1, &clob2,
+				  &clob3, &alias);
+			if (load_store_size > 8)
+				unexpected = true;
+		} else {
+			/* Load. */
+			load_n_store = true;
+			if (mod == MODE_OPCODE_YB2) {
+				switch (opcode) {
+				case LD_OPCODE_Y2:
+					load_store_signed = false;
+					load_store_size = 8;
+					break;
+				case LD4S_OPCODE_Y2:
+					load_store_signed = true;
+					load_store_size = 4;
+					break;
+				case LD4U_OPCODE_Y2:
+					load_store_signed = false;
+					load_store_size = 4;
+					break;
+				default:
+					unexpected = true;
+				}
+			} else if (mod == MODE_OPCODE_YA2) {
+				if (opcode == LD2S_OPCODE_Y2) {
+					load_store_signed = true;
+					load_store_size = 2;
+				} else if (opcode == LD2U_OPCODE_Y2) {
+					load_store_signed = false;
+					load_store_size = 2;
+				} else
+					unexpected = true;
+			} else
+				unexpected = true;
+			find_regs(bundle, &rd, &ra, &rb, &clob1, &clob2,
+				  &clob3, &alias);
+		}
+	} else {
+		unsigned int opcode;
+
+		/* bundle_2 is bundle after making X1 as "fnop". */
+		bundle_2 = (bundle & (~GX_INSN_X1_MASK)) | jit_x1_fnop();
+
+		if (is_x0_x1_nop(bundle_2))
+			bundle_2_enable = false;
+
+		if (get_Opcode_X1(bundle) == RRR_0_OPCODE_X1) {
+			opcode = get_UnaryOpcodeExtension_X1(bundle);
+
+			if (get_RRROpcodeExtension_X1(bundle) ==
+			    UNARY_RRR_0_OPCODE_X1) {
+				load_n_store = true;
+				find_regs(bundle, &rd, &ra, &rb, &clob1,
+					  &clob2, &clob3, &alias);
+
+				switch (opcode) {
+				case LD_UNARY_OPCODE_X1:
+					load_store_signed = false;
+					load_store_size = 8;
+					break;
+				case LD4S_UNARY_OPCODE_X1:
+					load_store_signed = true;
+					/* FALLTHROUGH */
+				case LD4U_UNARY_OPCODE_X1:
+					load_store_size = 4;
+					break;
+
+				case LD2S_UNARY_OPCODE_X1:
+					load_store_signed = true;
+					/* FALLTHROUGH */
+				case LD2U_UNARY_OPCODE_X1:
+					load_store_size = 2;
+					break;
+				default:
+					unexpected = true;
+				}
+			} else {
+				load_n_store = false;
+				load_store_signed = false;
+				find_regs(bundle, 0, &ra, &rb,
+					  &clob1, &clob2, &clob3,
+					  &alias);
+
+				opcode = get_RRROpcodeExtension_X1(bundle);
+				switch (opcode)	{
+				case ST_RRR_0_OPCODE_X1:
+					load_store_size = 8;
+					break;
+				case ST4_RRR_0_OPCODE_X1:
+					load_store_size = 4;
+					break;
+				case ST2_RRR_0_OPCODE_X1:
+					load_store_size = 2;
+					break;
+				default:
+					unexpected = true;
+				}
+			}
+		} else if (get_Opcode_X1(bundle) == IMM8_OPCODE_X1) {
+			load_n_store = true;
+			opcode = get_Imm8OpcodeExtension_X1(bundle);
+			switch (opcode)	{
+			case LD_ADD_IMM8_OPCODE_X1:
+				load_store_size = 8;
+				break;
+
+			case LD4S_ADD_IMM8_OPCODE_X1:
+				load_store_signed = true;
+				/* FALLTHROUGH */
+			case LD4U_ADD_IMM8_OPCODE_X1:
+				load_store_size = 4;
+				break;
+
+			case LD2S_ADD_IMM8_OPCODE_X1:
+				load_store_signed = true;
+				/* FALLTHROUGH */
+			case LD2U_ADD_IMM8_OPCODE_X1:
+				load_store_size = 2;
+				break;
+
+			case ST_ADD_IMM8_OPCODE_X1:
+				load_n_store = false;
+				load_store_size = 8;
+				break;
+			case ST4_ADD_IMM8_OPCODE_X1:
+				load_n_store = false;
+				load_store_size = 4;
+				break;
+			case ST2_ADD_IMM8_OPCODE_X1:
+				load_n_store = false;
+				load_store_size = 2;
+				break;
+			default:
+				unexpected = true;
+			}
+
+			if (!unexpected) {
+				x1_add = true;
+				if (load_n_store)
+					x1_add_imm8 = get_Imm8_X1(bundle);
+				else
+					x1_add_imm8 = get_Dest_Imm8_X1(bundle);
+			}
+
+			find_regs(bundle, load_n_store ? (&rd) : NULL,
+				  &ra, &rb, &clob1, &clob2, &clob3, &alias);
+		} else
+			unexpected = true;
+	}
+
+	/*
+	 * Some sanity check for register numbers extracted from fault bundle.
+	 */
+	if (check_regs(rd, ra, rb, clob1, clob2, clob3) == true)
+		unexpected = true;
+
+	/* Give warning if register ra has an aligned address. */
+	if (!unexpected)
+		WARN_ON(!((load_store_size - 1) & (regs->regs[ra])));
+
+
+	/*
+	 * Fault came from kernel space, here we only need take care of
+	 * unaligned "get_user/put_user" macros defined in "uaccess.h".
+	 * Basically, we will handle bundle like this:
+	 * {ld/2u/4s rd, ra; movei rx, 0} or {st/2/4 ra, rb; movei rx, 0}
+	 * (Refer to file "arch/tile/include/asm/uaccess.h" for details).
+	 * For either load or store, byte-wise operation is performed by calling
+	 * get_user() or put_user(). If the macro returns non-zero value,
+	 * set the value to rx, otherwise set zero to rx. Finally make pc point
+	 * to next bundle and return.
+	 */
+
+	if (EX1_PL(regs->ex1) != USER_PL) {
+
+		unsigned long rx = 0;
+		unsigned long x = 0, ret = 0;
+
+		if (y1_br || y1_lr || x1_add ||
+		    (load_store_signed !=
+		     (load_n_store && load_store_size == 4))) {
+			/* No branch, link, wrong sign-ext or load/store add. */
+			unexpected = true;
+		} else if (!unexpected) {
+			if (bundle & TILEGX_BUNDLE_MODE_MASK) {
+				/*
+				 * Fault bundle is Y mode.
+				 * Check if the Y1 and Y0 is the form of
+				 * { movei rx, 0; nop/fnop }, if yes,
+				 * find the rx.
+				 */
+
+				if ((get_Opcode_Y1(bundle) == ADDI_OPCODE_Y1)
+				    && (get_SrcA_Y1(bundle) == TREG_ZERO) &&
+				    (get_Imm8_Y1(bundle) == 0) &&
+				    is_bundle_y0_nop(bundle)) {
+					rx = get_Dest_Y1(bundle);
+				} else if ((get_Opcode_Y0(bundle) ==
+					    ADDI_OPCODE_Y0) &&
+					   (get_SrcA_Y0(bundle) == TREG_ZERO) &&
+					   (get_Imm8_Y0(bundle) == 0) &&
+					   is_bundle_y1_nop(bundle)) {
+					rx = get_Dest_Y0(bundle);
+				} else {
+					unexpected = true;
+				}
+			} else {
+				/*
+				 * Fault bundle is X mode.
+				 * Check if the X0 is 'movei rx, 0',
+				 * if yes, find the rx.
+				 */
+
+				if ((get_Opcode_X0(bundle) == IMM8_OPCODE_X0)
+				    && (get_Imm8OpcodeExtension_X0(bundle) ==
+					ADDI_IMM8_OPCODE_X0) &&
+				    (get_SrcA_X0(bundle) == TREG_ZERO) &&
+				    (get_Imm8_X0(bundle) == 0)) {
+					rx = get_Dest_X0(bundle);
+				} else {
+					unexpected = true;
+				}
+			}
+
+			/* rx should be less than 56. */
+			if (!unexpected && (rx >= 56))
+				unexpected = true;
+		}
+
+		if (!search_exception_tables(regs->pc)) {
+			/* No fixup in the exception tables for the pc. */
+			unexpected = true;
+		}
+
+		if (unexpected) {
+			/* Unexpected unalign kernel fault. */
+			struct task_struct *tsk = validate_current();
+
+			bust_spinlocks(1);
+
+			show_regs(regs);
+
+			if (unlikely(tsk->pid < 2)) {
+				panic("Kernel unalign fault running %s!",
+				      tsk->pid ? "init" : "the idle task");
+			}
+#ifdef SUPPORT_DIE
+			die("Oops", regs);
+#endif
+			bust_spinlocks(1);
+
+			do_group_exit(SIGKILL);
+
+		} else {
+			unsigned long i, b = 0;
+			unsigned char *ptr =
+				(unsigned char *)regs->regs[ra];
+			if (load_n_store) {
+				/* handle get_user(x, ptr) */
+				for (i = 0; i < load_store_size; i++) {
+					ret = get_user(b, ptr++);
+					if (!ret) {
+						/* Success! update x. */
+#ifdef __LITTLE_ENDIAN
+						x |= (b << (8 * i));
+#else
+						x <<= 8;
+						x |= b;
+#endif /* __LITTLE_ENDIAN */
+					} else {
+						x = 0;
+						break;
+					}
+				}
+
+				/* Sign-extend 4-byte loads. */
+				if (load_store_size == 4)
+					x = (long)(int)x;
+
+				/* Set register rd. */
+				regs->regs[rd] = x;
+
+				/* Set register rx. */
+				regs->regs[rx] = ret;
+
+				/* Bump pc. */
+				regs->pc += 8;
+
+			} else {
+				/* Handle put_user(x, ptr) */
+				x = regs->regs[rb];
+#ifdef __LITTLE_ENDIAN
+				b = x;
+#else
+				/*
+				 * Swap x in order to store x from low
+				 * to high memory same as the
+				 * little-endian case.
+				 */
+				switch (load_store_size) {
+				case 8:
+					b = swab64(x);
+					break;
+				case 4:
+					b = swab32(x);
+					break;
+				case 2:
+					b = swab16(x);
+					break;
+				}
+#endif /* __LITTLE_ENDIAN */
+				for (i = 0; i < load_store_size; i++) {
+					ret = put_user(b, ptr++);
+					if (ret)
+						break;
+					/* Success! shift 1 byte. */
+					b >>= 8;
+				}
+				/* Set register rx. */
+				regs->regs[rx] = ret;
+
+				/* Bump pc. */
+				regs->pc += 8;
+			}
+		}
+
+		unaligned_fixup_count++;
+
+		if (unaligned_printk) {
+			pr_info("%s/%d - Unalign fixup for kernel access to userspace %lx\n",
+				current->comm, current->pid, regs->regs[ra]);
+		}
+
+		/* Done! Return to the exception handler. */
+		return;
+	}
+
+	if ((align_ctl == 0) || unexpected) {
+		siginfo_t info = {
+			.si_signo = SIGBUS,
+			.si_code = BUS_ADRALN,
+			.si_addr = (unsigned char __user *)0
+		};
+		if (unaligned_printk)
+			pr_info("Unalign bundle: unexp @%llx, %llx\n",
+				(unsigned long long)regs->pc,
+				(unsigned long long)bundle);
+
+		if (ra < 56) {
+			unsigned long uaa = (unsigned long)regs->regs[ra];
+			/* Set bus Address. */
+			info.si_addr = (unsigned char __user *)uaa;
+		}
+
+		unaligned_fixup_count++;
+
+		trace_unhandled_signal("unaligned fixup trap", regs,
+				       (unsigned long)info.si_addr, SIGBUS);
+		force_sig_info(info.si_signo, &info, current);
+		return;
+	}
+
+#ifdef __LITTLE_ENDIAN
+#define UA_FIXUP_ADDR_DELTA          1
+#define UA_FIXUP_BFEXT_START(_B_)    0
+#define UA_FIXUP_BFEXT_END(_B_)     (8 * (_B_) - 1)
+#else /* __BIG_ENDIAN */
+#define UA_FIXUP_ADDR_DELTA          -1
+#define UA_FIXUP_BFEXT_START(_B_)   (64 - 8 * (_B_))
+#define UA_FIXUP_BFEXT_END(_B_)      63
+#endif /* __LITTLE_ENDIAN */
+
+
+
+	if ((ra != rb) && (rd != TREG_SP) && !alias &&
+	    !y1_br && !y1_lr && !x1_add) {
+		/*
+		 * Simple case: ra != rb and no register alias found,
+		 * and no branch or link. This will be the majority.
+		 * We can do a little better for simplae case than the
+		 * generic scheme below.
+		 */
+		if (!load_n_store) {
+			/*
+			 * Simple store: ra != rb, no need for scratch register.
+			 * Just store and rotate to right bytewise.
+			 */
+#ifdef __BIG_ENDIAN
+			frag.insn[n++] =
+				jit_x0_addi(ra, ra, load_store_size - 1) |
+				jit_x1_fnop();
+#endif /* __BIG_ENDIAN */
+			for (k = 0; k < load_store_size; k++) {
+				/* Store a byte. */
+				frag.insn[n++] =
+					jit_x0_rotli(rb, rb, 56) |
+					jit_x1_st1_add(ra, rb,
+						       UA_FIXUP_ADDR_DELTA);
+			}
+#ifdef __BIG_ENDIAN
+			frag.insn[n] = jit_x1_addi(ra, ra, 1);
+#else
+			frag.insn[n] = jit_x1_addi(ra, ra,
+						   -1 * load_store_size);
+#endif /* __LITTLE_ENDIAN */
+
+			if (load_store_size == 8) {
+				frag.insn[n] |= jit_x0_fnop();
+			} else if (load_store_size == 4) {
+				frag.insn[n] |= jit_x0_rotli(rb, rb, 32);
+			} else { /* = 2 */
+				frag.insn[n] |= jit_x0_rotli(rb, rb, 16);
+			}
+			n++;
+			if (bundle_2_enable)
+				frag.insn[n++] = bundle_2;
+			frag.insn[n++] = jit_x0_fnop() | jit_x1_iret();
+		} else {
+			if (rd == ra) {
+				/* Use two clobber registers: clob1/2. */
+				frag.insn[n++] =
+					jit_x0_addi(TREG_SP, TREG_SP, -16) |
+					jit_x1_fnop();
+				frag.insn[n++] =
+					jit_x0_addi(clob1, ra, 7) |
+					jit_x1_st_add(TREG_SP, clob1, -8);
+				frag.insn[n++] =
+					jit_x0_addi(clob2, ra, 0) |
+					jit_x1_st(TREG_SP, clob2);
+				frag.insn[n++] =
+					jit_x0_fnop() |
+					jit_x1_ldna(rd, ra);
+				frag.insn[n++] =
+					jit_x0_fnop() |
+					jit_x1_ldna(clob1, clob1);
+				/*
+				 * Note: we must make sure that rd must not
+				 * be sp. Recover clob1/2 from stack.
+				 */
+				frag.insn[n++] =
+					jit_x0_dblalign(rd, clob1, clob2) |
+					jit_x1_ld_add(clob2, TREG_SP, 8);
+				frag.insn[n++] =
+					jit_x0_fnop() |
+					jit_x1_ld_add(clob1, TREG_SP, 16);
+			} else {
+				/* Use one clobber register: clob1 only. */
+				frag.insn[n++] =
+					jit_x0_addi(TREG_SP, TREG_SP, -16) |
+					jit_x1_fnop();
+				frag.insn[n++] =
+					jit_x0_addi(clob1, ra, 7) |
+					jit_x1_st(TREG_SP, clob1);
+				frag.insn[n++] =
+					jit_x0_fnop() |
+					jit_x1_ldna(rd, ra);
+				frag.insn[n++] =
+					jit_x0_fnop() |
+					jit_x1_ldna(clob1, clob1);
+				/*
+				 * Note: we must make sure that rd must not
+				 * be sp. Recover clob1 from stack.
+				 */
+				frag.insn[n++] =
+					jit_x0_dblalign(rd, clob1, ra) |
+					jit_x1_ld_add(clob1, TREG_SP, 16);
+			}
+
+			if (bundle_2_enable)
+				frag.insn[n++] = bundle_2;
+			/*
+			 * For non 8-byte load, extract corresponding bytes and
+			 * signed extension.
+			 */
+			if (load_store_size == 4) {
+				if (load_store_signed)
+					frag.insn[n++] =
+						jit_x0_bfexts(
+							rd, rd,
+							UA_FIXUP_BFEXT_START(4),
+							UA_FIXUP_BFEXT_END(4)) |
+						jit_x1_fnop();
+				else
+					frag.insn[n++] =
+						jit_x0_bfextu(
+							rd, rd,
+							UA_FIXUP_BFEXT_START(4),
+							UA_FIXUP_BFEXT_END(4)) |
+						jit_x1_fnop();
+			} else if (load_store_size == 2) {
+				if (load_store_signed)
+					frag.insn[n++] =
+						jit_x0_bfexts(
+							rd, rd,
+							UA_FIXUP_BFEXT_START(2),
+							UA_FIXUP_BFEXT_END(2)) |
+						jit_x1_fnop();
+				else
+					frag.insn[n++] =
+						jit_x0_bfextu(
+							rd, rd,
+							UA_FIXUP_BFEXT_START(2),
+							UA_FIXUP_BFEXT_END(2)) |
+						jit_x1_fnop();
+			}
+
+			frag.insn[n++] =
+				jit_x0_fnop()  |
+				jit_x1_iret();
+		}
+	} else if (!load_n_store) {
+
+		/*
+		 * Generic memory store cases: use 3 clobber registers.
+		 *
+		 * Alloc space for saveing clob2,1,3 on user's stack.
+		 * register clob3 points to where clob2 saved, followed by
+		 * clob1 and 3 from high to low memory.
+		 */
+		frag.insn[n++] =
+			jit_x0_addi(TREG_SP, TREG_SP, -32)    |
+			jit_x1_fnop();
+		frag.insn[n++] =
+			jit_x0_addi(clob3, TREG_SP, 16)  |
+			jit_x1_st_add(TREG_SP, clob3, 8);
+#ifdef __LITTLE_ENDIAN
+		frag.insn[n++] =
+			jit_x0_addi(clob1, ra, 0)   |
+			jit_x1_st_add(TREG_SP, clob1, 8);
+#else
+		frag.insn[n++] =
+			jit_x0_addi(clob1, ra, load_store_size - 1)   |
+			jit_x1_st_add(TREG_SP, clob1, 8);
+#endif
+		if (load_store_size == 8) {
+			/*
+			 * We save one byte a time, not for fast, but compact
+			 * code. After each store, data source register shift
+			 * right one byte. unchanged after 8 stores.
+			 */
+			frag.insn[n++] =
+				jit_x0_addi(clob2, TREG_ZERO, 7)     |
+				jit_x1_st_add(TREG_SP, clob2, 16);
+			frag.insn[n++] =
+				jit_x0_rotli(rb, rb, 56)      |
+				jit_x1_st1_add(clob1, rb, UA_FIXUP_ADDR_DELTA);
+			frag.insn[n++] =
+				jit_x0_addi(clob2, clob2, -1) |
+				jit_x1_bnezt(clob2, -1);
+			frag.insn[n++] =
+				jit_x0_fnop()                 |
+				jit_x1_addi(clob2, y1_br_reg, 0);
+		} else if (load_store_size == 4) {
+			frag.insn[n++] =
+				jit_x0_addi(clob2, TREG_ZERO, 3)     |
+				jit_x1_st_add(TREG_SP, clob2, 16);
+			frag.insn[n++] =
+				jit_x0_rotli(rb, rb, 56)      |
+				jit_x1_st1_add(clob1, rb, UA_FIXUP_ADDR_DELTA);
+			frag.insn[n++] =
+				jit_x0_addi(clob2, clob2, -1) |
+				jit_x1_bnezt(clob2, -1);
+			/*
+			 * same as 8-byte case, but need shift another 4
+			 * byte to recover rb for 4-byte store.
+			 */
+			frag.insn[n++] = jit_x0_rotli(rb, rb, 32)      |
+				jit_x1_addi(clob2, y1_br_reg, 0);
+		} else { /* =2 */
+			frag.insn[n++] =
+				jit_x0_addi(clob2, rb, 0)     |
+				jit_x1_st_add(TREG_SP, clob2, 16);
+			for (k = 0; k < 2; k++) {
+				frag.insn[n++] =
+					jit_x0_shrui(rb, rb, 8)  |
+					jit_x1_st1_add(clob1, rb,
+						       UA_FIXUP_ADDR_DELTA);
+			}
+			frag.insn[n++] =
+				jit_x0_addi(rb, clob2, 0)       |
+				jit_x1_addi(clob2, y1_br_reg, 0);
+		}
+
+		if (bundle_2_enable)
+			frag.insn[n++] = bundle_2;
+
+		if (y1_lr) {
+			frag.insn[n++] =
+				jit_x0_fnop()                    |
+				jit_x1_mfspr(y1_lr_reg,
+					     SPR_EX_CONTEXT_0_0);
+		}
+		if (y1_br) {
+			frag.insn[n++] =
+				jit_x0_fnop()                    |
+				jit_x1_mtspr(SPR_EX_CONTEXT_0_0,
+					     clob2);
+		}
+		if (x1_add) {
+			frag.insn[n++] =
+				jit_x0_addi(ra, ra, x1_add_imm8) |
+				jit_x1_ld_add(clob2, clob3, -8);
+		} else {
+			frag.insn[n++] =
+				jit_x0_fnop()                    |
+				jit_x1_ld_add(clob2, clob3, -8);
+		}
+		frag.insn[n++] =
+			jit_x0_fnop()   |
+			jit_x1_ld_add(clob1, clob3, -8);
+		frag.insn[n++] = jit_x0_fnop()   | jit_x1_ld(clob3, clob3);
+		frag.insn[n++] = jit_x0_fnop()   | jit_x1_iret();
+
+	} else {
+		/*
+		 * Generic memory load cases.
+		 *
+		 * Alloc space for saveing clob1,2,3 on user's stack.
+		 * register clob3 points to where clob1 saved, followed
+		 * by clob2 and 3 from high to low memory.
+		 */
+
+		frag.insn[n++] =
+			jit_x0_addi(TREG_SP, TREG_SP, -32) |
+			jit_x1_fnop();
+		frag.insn[n++] =
+			jit_x0_addi(clob3, TREG_SP, 16) |
+			jit_x1_st_add(TREG_SP, clob3, 8);
+		frag.insn[n++] =
+			jit_x0_addi(clob2, ra, 0) |
+			jit_x1_st_add(TREG_SP, clob2, 8);
+
+		if (y1_br) {
+			frag.insn[n++] =
+				jit_x0_addi(clob1, y1_br_reg, 0) |
+				jit_x1_st_add(TREG_SP, clob1, 16);
+		} else {
+			frag.insn[n++] =
+				jit_x0_fnop() |
+				jit_x1_st_add(TREG_SP, clob1, 16);
+		}
+
+		if (bundle_2_enable)
+			frag.insn[n++] = bundle_2;
+
+		if (y1_lr) {
+			frag.insn[n++] =
+				jit_x0_fnop()  |
+				jit_x1_mfspr(y1_lr_reg,
+					     SPR_EX_CONTEXT_0_0);
+		}
+
+		if (y1_br) {
+			frag.insn[n++] =
+				jit_x0_fnop() |
+				jit_x1_mtspr(SPR_EX_CONTEXT_0_0,
+					     clob1);
+		}
+
+		frag.insn[n++] =
+			jit_x0_addi(clob1, clob2, 7)      |
+			jit_x1_ldna(rd, clob2);
+		frag.insn[n++] =
+			jit_x0_fnop()                     |
+			jit_x1_ldna(clob1, clob1);
+		frag.insn[n++] =
+			jit_x0_dblalign(rd, clob1, clob2) |
+			jit_x1_ld_add(clob1, clob3, -8);
+		if (x1_add) {
+			frag.insn[n++] =
+				jit_x0_addi(ra, ra, x1_add_imm8) |
+				jit_x1_ld_add(clob2, clob3, -8);
+		} else {
+			frag.insn[n++] =
+				jit_x0_fnop()  |
+				jit_x1_ld_add(clob2, clob3, -8);
+		}
+
+		frag.insn[n++] =
+			jit_x0_fnop() |
+			jit_x1_ld(clob3, clob3);
+
+		if (load_store_size == 4) {
+			if (load_store_signed)
+				frag.insn[n++] =
+					jit_x0_bfexts(
+						rd, rd,
+						UA_FIXUP_BFEXT_START(4),
+						UA_FIXUP_BFEXT_END(4)) |
+					jit_x1_fnop();
+			else
+				frag.insn[n++] =
+					jit_x0_bfextu(
+						rd, rd,
+						UA_FIXUP_BFEXT_START(4),
+						UA_FIXUP_BFEXT_END(4)) |
+					jit_x1_fnop();
+		} else if (load_store_size == 2) {
+			if (load_store_signed)
+				frag.insn[n++] =
+					jit_x0_bfexts(
+						rd, rd,
+						UA_FIXUP_BFEXT_START(2),
+						UA_FIXUP_BFEXT_END(2)) |
+					jit_x1_fnop();
+			else
+				frag.insn[n++] =
+					jit_x0_bfextu(
+						rd, rd,
+						UA_FIXUP_BFEXT_START(2),
+						UA_FIXUP_BFEXT_END(2)) |
+					jit_x1_fnop();
+		}
+
+		frag.insn[n++] = jit_x0_fnop() | jit_x1_iret();
+	}
+
+	/* Max JIT bundle count is 14. */
+	WARN_ON(n > 14);
+
+	if (!unexpected) {
+		int status = 0;
+		int idx = (regs->pc >> 3) &
+			((1ULL << (PAGE_SHIFT - UNALIGN_JIT_SHIFT)) - 1);
+
+		frag.pc = regs->pc;
+		frag.bundle = bundle;
+
+		if (unaligned_printk) {
+			pr_info("%s/%d, Unalign fixup: pc=%lx bundle=%lx %d %d %d %d %d %d %d %d\n",
+				current->comm, current->pid,
+				(unsigned long)frag.pc,
+				(unsigned long)frag.bundle,
+				(int)alias, (int)rd, (int)ra,
+				(int)rb, (int)bundle_2_enable,
+				(int)y1_lr, (int)y1_br, (int)x1_add);
+
+			for (k = 0; k < n; k += 2)
+				pr_info("[%d] %016llx %016llx\n",
+					k, (unsigned long long)frag.insn[k],
+					(unsigned long long)frag.insn[k+1]);
+		}
+
+		/* Swap bundle byte order for big endian sys. */
+#ifdef __BIG_ENDIAN
+		frag.bundle = GX_INSN_BSWAP(frag.bundle);
+		for (k = 0; k < n; k++)
+			frag.insn[k] = GX_INSN_BSWAP(frag.insn[k]);
+#endif /* __BIG_ENDIAN */
+
+		status = copy_to_user((void __user *)&jit_code_area[idx],
+				      &frag, sizeof(frag));
+		if (status) {
+			/* Fail to copy JIT into user land. send SIGSEGV. */
+			siginfo_t info = {
+				.si_signo = SIGSEGV,
+				.si_code = SEGV_MAPERR,
+				.si_addr = (void __user *)&jit_code_area[idx]
+			};
+
+			pr_warn("Unalign fixup: pid=%d %s jit_code_area=%llx\n",
+				current->pid, current->comm,
+				(unsigned long long)&jit_code_area[idx]);
+
+			trace_unhandled_signal("segfault in unalign fixup",
+					       regs,
+					       (unsigned long)info.si_addr,
+					       SIGSEGV);
+			force_sig_info(info.si_signo, &info, current);
+			return;
+		}
+
+
+		/* Do a cheaper increment, not accurate. */
+		unaligned_fixup_count++;
+		__flush_icache_range((unsigned long)&jit_code_area[idx],
+				     (unsigned long)&jit_code_area[idx] +
+				     sizeof(frag));
+
+		/* Setup SPR_EX_CONTEXT_0_0/1 for returning to user program.*/
+		__insn_mtspr(SPR_EX_CONTEXT_0_0, regs->pc + 8);
+		__insn_mtspr(SPR_EX_CONTEXT_0_1, PL_ICS_EX1(USER_PL, 0));
+
+		/* Modify pc at the start of new JIT. */
+		regs->pc = (unsigned long)&jit_code_area[idx].insn[0];
+		/* Set ICS in SPR_EX_CONTEXT_K_1. */
+		regs->ex1 = PL_ICS_EX1(USER_PL, 1);
+	}
+}
+
+
+/*
+ * C function to generate unalign data JIT. Called from unalign data
+ * interrupt handler.
+ *
+ * First check if unalign fix is disabled or exception did not not come from
+ * user space or sp register points to unalign address, if true, generate a
+ * SIGBUS. Then map a page into user space as JIT area if it is not mapped
+ * yet. Genenerate JIT code by calling jit_bundle_gen(). After that return
+ * back to exception handler.
+ *
+ * The exception handler will "iret" to new generated JIT code after
+ * restoring caller saved registers. In theory, the JIT code will perform
+ * another "iret" to resume user's program.
+ */
+
+void do_unaligned(struct pt_regs *regs, int vecnum)
+{
+	enum ctx_state prev_state = exception_enter();
+	tilegx_bundle_bits __user  *pc;
+	tilegx_bundle_bits bundle;
+	struct thread_info *info = current_thread_info();
+	int align_ctl;
+
+	/* Checks the per-process unaligned JIT flags */
+	align_ctl = unaligned_fixup;
+	switch (task_thread_info(current)->align_ctl) {
+	case PR_UNALIGN_NOPRINT:
+		align_ctl = 1;
+		break;
+	case PR_UNALIGN_SIGBUS:
+		align_ctl = 0;
+		break;
+	}
+
+	/* Enable iterrupt in order to access user land. */
+	local_irq_enable();
+
+	/*
+	 * The fault came from kernel space. Two choices:
+	 * (a) unaligned_fixup < 1, we will first call get/put_user fixup
+	 *     to return -EFAULT. If no fixup, simply panic the kernel.
+	 * (b) unaligned_fixup >=1, we will try to fix the unaligned access
+	 *     if it was triggered by get_user/put_user() macros. Panic the
+	 *     kernel if it is not fixable.
+	 */
+
+	if (EX1_PL(regs->ex1) != USER_PL) {
+
+		if (align_ctl < 1) {
+			unaligned_fixup_count++;
+			/* If exception came from kernel, try fix it up. */
+			if (fixup_exception(regs)) {
+				if (unaligned_printk)
+					pr_info("Unalign fixup: %d %llx @%llx\n",
+						(int)unaligned_fixup,
+						(unsigned long long)regs->ex1,
+						(unsigned long long)regs->pc);
+			} else {
+				/* Not fixable. Go panic. */
+				panic("Unalign exception in Kernel. pc=%lx",
+				      regs->pc);
+			}
+		} else {
+			/*
+			 * Try to fix the exception. If we can't, panic the
+			 * kernel.
+			 */
+			bundle = GX_INSN_BSWAP(
+				*((tilegx_bundle_bits *)(regs->pc)));
+			jit_bundle_gen(regs, bundle, align_ctl);
+		}
+		goto done;
+	}
+
+	/*
+	 * Fault came from user with ICS or stack is not aligned.
+	 * If so, we will trigger SIGBUS.
+	 */
+	if ((regs->sp & 0x7) || (regs->ex1) || (align_ctl < 0)) {
+		siginfo_t info = {
+			.si_signo = SIGBUS,
+			.si_code = BUS_ADRALN,
+			.si_addr = (unsigned char __user *)0
+		};
+
+		if (unaligned_printk)
+			pr_info("Unalign fixup: %d %llx @%llx\n",
+				(int)unaligned_fixup,
+				(unsigned long long)regs->ex1,
+				(unsigned long long)regs->pc);
+
+		unaligned_fixup_count++;
+
+		trace_unhandled_signal("unaligned fixup trap", regs, 0, SIGBUS);
+		force_sig_info(info.si_signo, &info, current);
+		goto done;
+	}
+
+
+	/* Read the bundle casued the exception! */
+	pc = (tilegx_bundle_bits __user *)(regs->pc);
+	if (get_user(bundle, pc) != 0) {
+		/* Probably never be here since pc is valid user address.*/
+		siginfo_t info = {
+			.si_signo = SIGSEGV,
+			.si_code = SEGV_MAPERR,
+			.si_addr = (void __user *)pc
+		};
+		pr_err("Couldn't read instruction at %p trying to step\n", pc);
+		trace_unhandled_signal("segfault in unalign fixup", regs,
+				       (unsigned long)info.si_addr, SIGSEGV);
+		force_sig_info(info.si_signo, &info, current);
+		goto done;
+	}
+
+	if (!info->unalign_jit_base) {
+		void __user *user_page;
+
+		/*
+		 * Allocate a page in userland.
+		 * For 64-bit processes we try to place the mapping far
+		 * from anything else that might be going on (specifically
+		 * 64 GB below the top of the user address space).  If it
+		 * happens not to be possible to put it there, it's OK;
+		 * the kernel will choose another location and we'll
+		 * remember it for later.
+		 */
+		if (is_compat_task())
+			user_page = NULL;
+		else
+			user_page = (void __user *)(TASK_SIZE - (1UL << 36)) +
+				(current->pid << PAGE_SHIFT);
+
+		user_page = (void __user *) vm_mmap(NULL,
+						    (unsigned long)user_page,
+						    PAGE_SIZE,
+						    PROT_EXEC | PROT_READ |
+						    PROT_WRITE,
+#ifdef CONFIG_HOMECACHE
+						    MAP_CACHE_HOME_TASK |
+#endif
+						    MAP_PRIVATE |
+						    MAP_ANONYMOUS,
+						    0);
+
+		if (IS_ERR((void __force *)user_page)) {
+			pr_err("Out of kernel pages trying do_mmap\n");
+			goto done;
+		}
+
+		/* Save the address in the thread_info struct */
+		info->unalign_jit_base = user_page;
+		if (unaligned_printk)
+			pr_info("Unalign bundle: %d:%d, allocate page @%llx\n",
+				raw_smp_processor_id(), current->pid,
+				(unsigned long long)user_page);
+	}
+
+	/* Generate unalign JIT */
+	jit_bundle_gen(regs, GX_INSN_BSWAP(bundle), align_ctl);
+
+done:
+	exception_exit(prev_state);
+}
+
+#endif /* __tilegx__ */
diff --git a/arch/tile/kernel/usb.c b/arch/tile/kernel/usb.c
new file mode 100644
index 000000000..5af8debc6
--- /dev/null
+++ b/arch/tile/kernel/usb.c
@@ -0,0 +1,69 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ * Register the Tile-Gx USB interfaces as platform devices.
+ *
+ * The actual USB driver is just some glue (in
+ * drivers/usb/host/[eo]hci-tilegx.c) which makes the registers available
+ * to the standard kernel EHCI and OHCI drivers.
+ */
+
+#include <linux/dma-mapping.h>
+#include <linux/platform_device.h>
+#include <linux/usb/tilegx.h>
+#include <linux/types.h>
+
+static u64 ehci_dmamask = DMA_BIT_MASK(32);
+
+#define USB_HOST_DEF(unit, type, dmamask) \
+	static struct \
+	    tilegx_usb_platform_data tilegx_usb_platform_data_ ## type ## \
+		hci ## unit = { \
+		.dev_index = unit, \
+	}; \
+	\
+	static struct platform_device tilegx_usb_ ## type ## hci ## unit = { \
+		.name		= "tilegx-" #type "hci", \
+		.id		= unit, \
+		.dev = { \
+			.dma_mask		= dmamask, \
+			.coherent_dma_mask	= DMA_BIT_MASK(32), \
+			.platform_data = \
+				&tilegx_usb_platform_data_ ## type ## hci ## \
+				unit, \
+		}, \
+	};
+
+USB_HOST_DEF(0, e, &ehci_dmamask)
+USB_HOST_DEF(0, o, NULL)
+USB_HOST_DEF(1, e, &ehci_dmamask)
+USB_HOST_DEF(1, o, NULL)
+
+#undef USB_HOST_DEF
+
+static struct platform_device *tilegx_usb_devices[] __initdata = {
+	&tilegx_usb_ehci0,
+	&tilegx_usb_ehci1,
+	&tilegx_usb_ohci0,
+	&tilegx_usb_ohci1,
+};
+
+/** Add our set of possible USB devices. */
+static int __init tilegx_usb_init(void)
+{
+	platform_add_devices(tilegx_usb_devices,
+			     ARRAY_SIZE(tilegx_usb_devices));
+
+	return 0;
+}
+arch_initcall(tilegx_usb_init);
diff --git a/arch/tile/kernel/vdso.c b/arch/tile/kernel/vdso.c
new file mode 100644
index 000000000..5bc51d7df
--- /dev/null
+++ b/arch/tile/kernel/vdso.c
@@ -0,0 +1,197 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/binfmts.h>
+#include <linux/compat.h>
+#include <linux/elf.h>
+#include <linux/mm.h>
+#include <linux/pagemap.h>
+
+#include <asm/vdso.h>
+#include <asm/mman.h>
+#include <asm/sections.h>
+
+#include <arch/sim.h>
+
+/* The alignment of the vDSO. */
+#define VDSO_ALIGNMENT  PAGE_SIZE
+
+
+static unsigned int vdso_pages;
+static struct page **vdso_pagelist;
+
+#ifdef CONFIG_COMPAT
+static unsigned int vdso32_pages;
+static struct page **vdso32_pagelist;
+#endif
+static int vdso_ready;
+
+/*
+ * The vdso data page.
+ */
+static union {
+	struct vdso_data	data;
+	u8			page[PAGE_SIZE];
+} vdso_data_store __page_aligned_data;
+
+struct vdso_data *vdso_data = &vdso_data_store.data;
+
+static unsigned int __read_mostly vdso_enabled = 1;
+
+static struct page **vdso_setup(void *vdso_kbase, unsigned int pages)
+{
+	int i;
+	struct page **pagelist;
+
+	pagelist = kzalloc(sizeof(struct page *) * (pages + 1), GFP_KERNEL);
+	BUG_ON(pagelist == NULL);
+	for (i = 0; i < pages - 1; i++) {
+		struct page *pg = virt_to_page(vdso_kbase + i*PAGE_SIZE);
+		ClearPageReserved(pg);
+		pagelist[i] = pg;
+	}
+	pagelist[pages - 1] = virt_to_page(vdso_data);
+	pagelist[pages] = NULL;
+
+	return pagelist;
+}
+
+static int __init vdso_init(void)
+{
+	int data_pages = sizeof(vdso_data_store) >> PAGE_SHIFT;
+
+	/*
+	 * We can disable vDSO support generally, but we need to retain
+	 * one page to support the two-bundle (16-byte) rt_sigreturn path.
+	 */
+	if (!vdso_enabled) {
+		size_t offset = (unsigned long)&__vdso_rt_sigreturn;
+		static struct page *sigret_page;
+		sigret_page = alloc_page(GFP_KERNEL | __GFP_ZERO);
+		BUG_ON(sigret_page == NULL);
+		vdso_pagelist = &sigret_page;
+		vdso_pages = 1;
+		BUG_ON(offset >= PAGE_SIZE);
+		memcpy(page_address(sigret_page) + offset,
+		       vdso_start + offset, 16);
+#ifdef CONFIG_COMPAT
+		vdso32_pages = vdso_pages;
+		vdso32_pagelist = vdso_pagelist;
+#endif
+		vdso_ready = 1;
+		return 0;
+	}
+
+	vdso_pages = (vdso_end - vdso_start) >> PAGE_SHIFT;
+	vdso_pages += data_pages;
+	vdso_pagelist = vdso_setup(vdso_start, vdso_pages);
+
+#ifdef CONFIG_COMPAT
+	vdso32_pages = (vdso32_end - vdso32_start) >> PAGE_SHIFT;
+	vdso32_pages += data_pages;
+	vdso32_pagelist = vdso_setup(vdso32_start, vdso32_pages);
+#endif
+
+	smp_wmb();
+	vdso_ready = 1;
+
+	return 0;
+}
+arch_initcall(vdso_init);
+
+const char *arch_vma_name(struct vm_area_struct *vma)
+{
+	if (vma->vm_mm && vma->vm_start == VDSO_BASE)
+		return "[vdso]";
+#ifndef __tilegx__
+	if (vma->vm_start == MEM_USER_INTRPT)
+		return "[intrpt]";
+#endif
+	return NULL;
+}
+
+int setup_vdso_pages(void)
+{
+	struct page **pagelist;
+	unsigned long pages;
+	struct mm_struct *mm = current->mm;
+	unsigned long vdso_base = 0;
+	int retval = 0;
+
+	if (!vdso_ready)
+		return 0;
+
+	mm->context.vdso_base = 0;
+
+	pagelist = vdso_pagelist;
+	pages = vdso_pages;
+#ifdef CONFIG_COMPAT
+	if (is_compat_task()) {
+		pagelist = vdso32_pagelist;
+		pages = vdso32_pages;
+	}
+#endif
+
+	/*
+	 * vDSO has a problem and was disabled, just don't "enable" it for the
+	 * process.
+	 */
+	if (pages == 0)
+		return 0;
+
+	vdso_base = get_unmapped_area(NULL, vdso_base,
+				      (pages << PAGE_SHIFT) +
+				      ((VDSO_ALIGNMENT - 1) & PAGE_MASK),
+				      0, 0);
+	if (IS_ERR_VALUE(vdso_base)) {
+		retval = vdso_base;
+		return retval;
+	}
+
+	/* Add required alignment. */
+	vdso_base = ALIGN(vdso_base, VDSO_ALIGNMENT);
+
+	/*
+	 * Put vDSO base into mm struct. We need to do this before calling
+	 * install_special_mapping or the perf counter mmap tracking code
+	 * will fail to recognise it as a vDSO (since arch_vma_name fails).
+	 */
+	mm->context.vdso_base = vdso_base;
+
+	/*
+	 * our vma flags don't have VM_WRITE so by default, the process isn't
+	 * allowed to write those pages.
+	 * gdb can break that with ptrace interface, and thus trigger COW on
+	 * those pages but it's then your responsibility to never do that on
+	 * the "data" page of the vDSO or you'll stop getting kernel updates
+	 * and your nice userland gettimeofday will be totally dead.
+	 * It's fine to use that for setting breakpoints in the vDSO code
+	 * pages though
+	 */
+	retval = install_special_mapping(mm, vdso_base,
+					 pages << PAGE_SHIFT,
+					 VM_READ|VM_EXEC |
+					 VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC,
+					 pagelist);
+	if (retval)
+		mm->context.vdso_base = 0;
+
+	return retval;
+}
+
+static __init int vdso_func(char *s)
+{
+	return kstrtouint(s, 0, &vdso_enabled);
+}
+__setup("vdso=", vdso_func);
diff --git a/arch/tile/kernel/vdso/Makefile b/arch/tile/kernel/vdso/Makefile
new file mode 100644
index 000000000..a025f63d5
--- /dev/null
+++ b/arch/tile/kernel/vdso/Makefile
@@ -0,0 +1,118 @@
+# Symbols present in the vdso
+vdso-syms = rt_sigreturn gettimeofday
+
+# Files to link into the vdso
+obj-vdso = $(patsubst %, v%.o, $(vdso-syms))
+
+# Build rules
+targets := $(obj-vdso) vdso.so vdso.so.dbg vdso.lds
+obj-vdso := $(addprefix $(obj)/, $(obj-vdso))
+
+# vdso32 is only for tilegx -m32 compat task.
+VDSO32-$(CONFIG_COMPAT) := y
+
+obj-y += vdso.o
+obj-$(VDSO32-y) += vdso32.o
+extra-y += vdso.lds
+CPPFLAGS_vdso.lds += -P -C -U$(ARCH)
+
+# vDSO code runs in userspace and -pg doesn't help with profiling anyway.
+CFLAGS_REMOVE_vdso.o = -pg
+CFLAGS_REMOVE_vdso32.o = -pg
+CFLAGS_REMOVE_vrt_sigreturn.o = -pg
+CFLAGS_REMOVE_vrt_sigreturn32.o = -pg
+CFLAGS_REMOVE_vgettimeofday.o = -pg
+CFLAGS_REMOVE_vgettimeofday32.o = -pg
+
+ifdef CONFIG_FEEDBACK_COLLECT
+# vDSO code runs in userspace, not collecting feedback data.
+CFLAGS_REMOVE_vdso.o = -ffeedback-generate
+CFLAGS_REMOVE_vdso32.o = -ffeedback-generate
+CFLAGS_REMOVE_vrt_sigreturn.o = -ffeedback-generate
+CFLAGS_REMOVE_vrt_sigreturn32.o = -ffeedback-generate
+CFLAGS_REMOVE_vgettimeofday.o = -ffeedback-generate
+CFLAGS_REMOVE_vgettimeofday32.o = -ffeedback-generate
+endif
+
+# Disable gcov profiling for VDSO code
+GCOV_PROFILE := n
+
+# Force dependency
+$(obj)/vdso.o: $(obj)/vdso.so
+
+# link rule for the .so file, .lds has to be first
+SYSCFLAGS_vdso.so.dbg = $(c_flags)
+$(obj)/vdso.so.dbg: $(src)/vdso.lds $(obj-vdso)
+	$(call if_changed,vdsold)
+
+
+# We also create a special relocatable object that should mirror the symbol
+# table and layout of the linked DSO.  With ld -R we can then refer to
+# these symbols in the kernel code rather than hand-coded addresses.
+extra-y += vdso-syms.o
+$(obj)/built-in.o: $(obj)/vdso-syms.o
+$(obj)/built-in.o: ld_flags += -R $(obj)/vdso-syms.o
+
+SYSCFLAGS_vdso.so.dbg = -shared -s -Wl,-soname=linux-vdso.so.1 \
+                            $(call cc-ldoption, -Wl$(comma)--hash-style=sysv)
+SYSCFLAGS_vdso_syms.o = -r
+$(obj)/vdso-syms.o: $(src)/vdso.lds $(obj)/vrt_sigreturn.o FORCE
+	$(call if_changed,vdsold)
+
+
+# strip rule for the .so file
+$(obj)/%.so: OBJCOPYFLAGS := -S
+$(obj)/%.so: $(obj)/%.so.dbg FORCE
+	$(call if_changed,objcopy)
+
+# actual build commands
+# The DSO images are built using a special linker script
+# Add -lgcc so tilepro gets static muldi3 and lshrdi3 definitions.
+# Make sure only to export the intended __vdso_xxx symbol offsets.
+quiet_cmd_vdsold = VDSOLD  $@
+      cmd_vdsold = $(CC) $(KCFLAGS) -nostdlib $(SYSCFLAGS_$(@F)) \
+                           -Wl,-T,$(filter-out FORCE,$^) -o $@.tmp -lgcc && \
+                   $(CROSS_COMPILE)objcopy \
+                           $(patsubst %, -G __vdso_%, $(vdso-syms)) $@.tmp $@
+
+# install commands for the unstripped file
+quiet_cmd_vdso_install = INSTALL $@
+      cmd_vdso_install = cp $(obj)/$@.dbg $(MODLIB)/vdso/$@
+
+vdso.so: $(obj)/vdso.so.dbg
+	@mkdir -p $(MODLIB)/vdso
+	$(call cmd,vdso_install)
+
+vdso32.so: $(obj)/vdso32.so.dbg
+	$(call cmd,vdso_install)
+
+vdso_install: vdso.so
+vdso32_install: vdso32.so
+
+
+KBUILD_AFLAGS_32 := $(filter-out -m64,$(KBUILD_AFLAGS))
+KBUILD_AFLAGS_32 += -m32 -s
+KBUILD_CFLAGS_32 := $(filter-out -m64,$(KBUILD_CFLAGS))
+KBUILD_CFLAGS_32 += -m32 -fPIC -shared
+
+obj-vdso32 = $(patsubst %, v%32.o, $(vdso-syms))
+obj-vdso32 := $(addprefix $(obj)/, $(obj-vdso32))
+
+targets += $(obj-vdso32) vdso32.so vdso32.so.dbg
+
+$(obj-vdso32:%=%): KBUILD_AFLAGS = $(KBUILD_AFLAGS_32)
+$(obj-vdso32:%=%): KBUILD_CFLAGS = $(KBUILD_CFLAGS_32)
+
+$(obj)/vgettimeofday32.o: $(obj)/vgettimeofday.c
+	$(call if_changed_rule,cc_o_c)
+
+$(obj)/vrt_sigreturn32.o: $(obj)/vrt_sigreturn.S
+	$(call if_changed,as_o_S)
+
+# Force dependency
+$(obj)/vdso32.o: $(obj)/vdso32.so
+
+SYSCFLAGS_vdso32.so.dbg = -m32 -shared -s -Wl,-soname=linux-vdso32.so.1 \
+			    $(call cc-ldoption, -Wl$(comma)--hash-style=sysv)
+$(obj)/vdso32.so.dbg: $(src)/vdso.lds $(obj-vdso32)
+	$(call if_changed,vdsold)
diff --git a/arch/tile/kernel/vdso/vdso.S b/arch/tile/kernel/vdso/vdso.S
new file mode 100644
index 000000000..3467adb41
--- /dev/null
+++ b/arch/tile/kernel/vdso/vdso.S
@@ -0,0 +1,28 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/init.h>
+#include <linux/linkage.h>
+#include <asm/page.h>
+
+	__PAGE_ALIGNED_DATA
+
+	.global vdso_start, vdso_end
+	.align PAGE_SIZE
+vdso_start:
+	.incbin "arch/tile/kernel/vdso/vdso.so"
+	.align PAGE_SIZE
+vdso_end:
+
+	.previous
diff --git a/arch/tile/kernel/vdso/vdso.lds.S b/arch/tile/kernel/vdso/vdso.lds.S
new file mode 100644
index 000000000..731529f3f
--- /dev/null
+++ b/arch/tile/kernel/vdso/vdso.lds.S
@@ -0,0 +1,89 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#define VDSO_VERSION_STRING	LINUX_2.6
+
+
+OUTPUT_ARCH(tile)
+
+/* The ELF entry point can be used to set the AT_SYSINFO value. */
+ENTRY(__vdso_rt_sigreturn);
+
+
+SECTIONS
+{
+	. = SIZEOF_HEADERS;
+
+	.hash		: { *(.hash) }			:text
+	.gnu.hash	: { *(.gnu.hash) }
+	.dynsym		: { *(.dynsym) }
+	.dynstr		: { *(.dynstr) }
+	.gnu.version	: { *(.gnu.version) }
+	.gnu.version_d	: { *(.gnu.version_d) }
+	.gnu.version_r	: { *(.gnu.version_r) }
+
+	.note		: { *(.note.*) }		:text	:note
+	.dynamic	: { *(.dynamic) }		:text	:dynamic
+
+	.eh_frame_hdr	: { *(.eh_frame_hdr) }		:text	:eh_frame_hdr
+	.eh_frame	: { KEEP (*(.eh_frame)) }	:text
+
+	.rodata	 : { *(.rodata .rodata.* .gnu.linkonce.r.*) }
+
+	/*
+	 * This linker script is used both with -r and with -shared.
+	 * For the layouts to match, we need to skip more than enough
+	 * space for the dynamic symbol table et al. If this amount
+	 * is insufficient, ld -shared will barf. Just increase it here.
+	 */
+	. = 0x1000;
+	.text		: { *(.text .text.*) }		:text
+
+	.data		: {
+		*(.got.plt) *(.got)
+		*(.data .data.* .gnu.linkonce.d.*)
+		*(.dynbss)
+		*(.bss .bss.* .gnu.linkonce.b.*)
+	}
+}
+
+
+/*
+ * We must supply the ELF program headers explicitly to get just one
+ * PT_LOAD segment, and set the flags explicitly to make segments read-only.
+ */
+PHDRS
+{
+	text		PT_LOAD		FLAGS(5) FILEHDR PHDRS;	/* PF_R|PF_X */
+	dynamic		PT_DYNAMIC	FLAGS(4);		/* PF_R */
+	note		PT_NOTE		FLAGS(4);		/* PF_R */
+	eh_frame_hdr	PT_GNU_EH_FRAME;
+}
+
+
+/*
+ * This controls what userland symbols we export from the vDSO.
+ */
+VERSION
+{
+	VDSO_VERSION_STRING {
+	global:
+		__vdso_rt_sigreturn;
+		__vdso_gettimeofday;
+		gettimeofday;
+		__vdso_clock_gettime;
+		clock_gettime;
+	local:*;
+	};
+}
diff --git a/arch/tile/kernel/vdso/vdso32.S b/arch/tile/kernel/vdso/vdso32.S
new file mode 100644
index 000000000..1d1ac3257
--- /dev/null
+++ b/arch/tile/kernel/vdso/vdso32.S
@@ -0,0 +1,28 @@
+/*
+ * Copyright 2013 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/init.h>
+#include <linux/linkage.h>
+#include <asm/page.h>
+
+	__PAGE_ALIGNED_DATA
+
+	.global vdso32_start, vdso32_end
+	.align PAGE_SIZE
+vdso32_start:
+	.incbin "arch/tile/kernel/vdso/vdso32.so"
+	.align PAGE_SIZE
+vdso32_end:
+
+	.previous
diff --git a/arch/tile/kernel/vdso/vgettimeofday.c b/arch/tile/kernel/vdso/vgettimeofday.c
new file mode 100644
index 000000000..8bb21eda0
--- /dev/null
+++ b/arch/tile/kernel/vdso/vgettimeofday.c
@@ -0,0 +1,198 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#define VDSO_BUILD  /* avoid some shift warnings for -m32 in <asm/page.h> */
+#include <linux/time.h>
+#include <asm/timex.h>
+#include <asm/unistd.h>
+#include <asm/vdso.h>
+
+#if CHIP_HAS_SPLIT_CYCLE()
+static inline cycles_t get_cycles_inline(void)
+{
+	unsigned int high = __insn_mfspr(SPR_CYCLE_HIGH);
+	unsigned int low = __insn_mfspr(SPR_CYCLE_LOW);
+	unsigned int high2 = __insn_mfspr(SPR_CYCLE_HIGH);
+
+	while (unlikely(high != high2)) {
+		low = __insn_mfspr(SPR_CYCLE_LOW);
+		high = high2;
+		high2 = __insn_mfspr(SPR_CYCLE_HIGH);
+	}
+
+	return (((cycles_t)high) << 32) | low;
+}
+#define get_cycles get_cycles_inline
+#endif
+
+struct syscall_return_value {
+	long value;
+	long error;
+};
+
+/*
+ * Find out the vDSO data page address in the process address space.
+ */
+inline unsigned long get_datapage(void)
+{
+	unsigned long ret;
+
+	/* vdso data page located in the 2nd vDSO page. */
+	asm volatile ("lnk %0" : "=r"(ret));
+	ret &= ~(PAGE_SIZE - 1);
+	ret += PAGE_SIZE;
+
+	return ret;
+}
+
+static inline u64 vgetsns(struct vdso_data *vdso)
+{
+	return ((get_cycles() - vdso->cycle_last) & vdso->mask) * vdso->mult;
+}
+
+static inline int do_realtime(struct vdso_data *vdso, struct timespec *ts)
+{
+	unsigned count;
+	u64 ns;
+
+	do {
+		count = read_seqcount_begin(&vdso->tb_seq);
+		ts->tv_sec = vdso->wall_time_sec;
+		ns = vdso->wall_time_snsec;
+		ns += vgetsns(vdso);
+		ns >>= vdso->shift;
+	} while (unlikely(read_seqcount_retry(&vdso->tb_seq, count)));
+
+	ts->tv_sec += __iter_div_u64_rem(ns, NSEC_PER_SEC, &ns);
+	ts->tv_nsec = ns;
+
+	return 0;
+}
+
+static inline int do_monotonic(struct vdso_data *vdso, struct timespec *ts)
+{
+	unsigned count;
+	u64 ns;
+
+	do {
+		count = read_seqcount_begin(&vdso->tb_seq);
+		ts->tv_sec = vdso->monotonic_time_sec;
+		ns = vdso->monotonic_time_snsec;
+		ns += vgetsns(vdso);
+		ns >>= vdso->shift;
+	} while (unlikely(read_seqcount_retry(&vdso->tb_seq, count)));
+
+	ts->tv_sec += __iter_div_u64_rem(ns, NSEC_PER_SEC, &ns);
+	ts->tv_nsec = ns;
+
+	return 0;
+}
+
+static inline int do_realtime_coarse(struct vdso_data *vdso,
+				     struct timespec *ts)
+{
+	unsigned count;
+
+	do {
+		count = read_seqcount_begin(&vdso->tb_seq);
+		ts->tv_sec = vdso->wall_time_coarse_sec;
+		ts->tv_nsec = vdso->wall_time_coarse_nsec;
+	} while (unlikely(read_seqcount_retry(&vdso->tb_seq, count)));
+
+	return 0;
+}
+
+static inline int do_monotonic_coarse(struct vdso_data *vdso,
+				      struct timespec *ts)
+{
+	unsigned count;
+
+	do {
+		count = read_seqcount_begin(&vdso->tb_seq);
+		ts->tv_sec = vdso->monotonic_time_coarse_sec;
+		ts->tv_nsec = vdso->monotonic_time_coarse_nsec;
+	} while (unlikely(read_seqcount_retry(&vdso->tb_seq, count)));
+
+	return 0;
+}
+
+struct syscall_return_value __vdso_gettimeofday(struct timeval *tv,
+						struct timezone *tz)
+{
+	struct syscall_return_value ret = { 0, 0 };
+	unsigned count;
+	struct vdso_data *vdso = (struct vdso_data *)get_datapage();
+
+	/* The use of the timezone is obsolete, normally tz is NULL. */
+	if (unlikely(tz != NULL)) {
+		do {
+			count = read_seqcount_begin(&vdso->tz_seq);
+			tz->tz_minuteswest = vdso->tz_minuteswest;
+			tz->tz_dsttime = vdso->tz_dsttime;
+		} while (unlikely(read_seqcount_retry(&vdso->tz_seq, count)));
+	}
+
+	if (unlikely(tv == NULL))
+		return ret;
+
+	do_realtime(vdso, (struct timespec *)tv);
+	tv->tv_usec /= 1000;
+
+	return ret;
+}
+
+int gettimeofday(struct timeval *tv, struct timezone *tz)
+	__attribute__((weak, alias("__vdso_gettimeofday")));
+
+static struct syscall_return_value vdso_fallback_gettime(long clock,
+							 struct timespec *ts)
+{
+	struct syscall_return_value ret;
+	__asm__ __volatile__ (
+		"swint1"
+		: "=R00" (ret.value), "=R01" (ret.error)
+		: "R10" (__NR_clock_gettime), "R00" (clock), "R01" (ts)
+		: "r2", "r3", "r4", "r5", "r6", "r7",
+		"r8",  "r9", "r11", "r12", "r13", "r14", "r15",
+		"r16", "r17", "r18", "r19", "r20", "r21", "r22", "r23",
+		"r24", "r25", "r26", "r27", "r28", "r29", "memory");
+	return ret;
+}
+
+struct syscall_return_value __vdso_clock_gettime(clockid_t clock,
+						 struct timespec *ts)
+{
+	struct vdso_data *vdso = (struct vdso_data *)get_datapage();
+	struct syscall_return_value ret = { 0, 0 };
+
+	switch (clock) {
+	case CLOCK_REALTIME:
+		do_realtime(vdso, ts);
+		return ret;
+	case CLOCK_MONOTONIC:
+		do_monotonic(vdso, ts);
+		return ret;
+	case CLOCK_REALTIME_COARSE:
+		do_realtime_coarse(vdso, ts);
+		return ret;
+	case CLOCK_MONOTONIC_COARSE:
+		do_monotonic_coarse(vdso, ts);
+		return ret;
+	default:
+		return vdso_fallback_gettime(clock, ts);
+	}
+}
+
+int clock_gettime(clockid_t clock, struct timespec *ts)
+	__attribute__((weak, alias("__vdso_clock_gettime")));
diff --git a/arch/tile/kernel/vdso/vrt_sigreturn.S b/arch/tile/kernel/vdso/vrt_sigreturn.S
new file mode 100644
index 000000000..6326caf4a
--- /dev/null
+++ b/arch/tile/kernel/vdso/vrt_sigreturn.S
@@ -0,0 +1,30 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/linkage.h>
+#include <arch/abi.h>
+#include <asm/unistd.h>
+
+/*
+ * Note that libc has a copy of this function that it uses to compare
+ * against the PC when a stack backtrace ends, so if this code is
+ * changed, the libc implementation(s) should also be updated.
+ */
+ENTRY(__vdso_rt_sigreturn)
+	moveli TREG_SYSCALL_NR_NAME, __NR_rt_sigreturn
+	swint1
+	/* We don't use ENDPROC to avoid tagging this symbol as FUNC,
+	 * which confuses the perf tool.
+	 */
+	END(__vdso_rt_sigreturn)
diff --git a/arch/tile/kernel/vmlinux.lds.S b/arch/tile/kernel/vmlinux.lds.S
new file mode 100644
index 000000000..0e059a010
--- /dev/null
+++ b/arch/tile/kernel/vmlinux.lds.S
@@ -0,0 +1,90 @@
+#include <asm-generic/vmlinux.lds.h>
+#include <asm/page.h>
+#include <asm/cache.h>
+#include <asm/thread_info.h>
+#include <hv/hypervisor.h>
+
+/* Text loads starting from the supervisor interrupt vector address. */
+#define TEXT_OFFSET MEM_SV_START
+
+OUTPUT_ARCH(tile)
+ENTRY(_start)
+jiffies = jiffies_64;
+
+PHDRS
+{
+  intrpt PT_LOAD ;
+  text PT_LOAD ;
+  data PT_LOAD ;
+}
+SECTIONS
+{
+  /* Text is loaded with a different VA than data; start with text. */
+  #undef LOAD_OFFSET
+  #define LOAD_OFFSET TEXT_OFFSET
+
+  /* Interrupt vectors */
+  .intrpt (LOAD_OFFSET) : AT ( 0 )   /* put at the start of physical memory */
+  {
+    _text = .;
+    *(.intrpt)
+  } :intrpt =0
+
+  /* Hypervisor call vectors */
+  . = ALIGN(0x10000);
+  .hvglue : AT (ADDR(.hvglue) - LOAD_OFFSET) {
+    *(.hvglue)
+  } :NONE
+
+  /* Now the real code */
+  . = ALIGN(0x20000);
+  _stext = .;
+  .text : AT (ADDR(.text) - LOAD_OFFSET) {
+    HEAD_TEXT
+    SCHED_TEXT
+    LOCK_TEXT
+    KPROBES_TEXT
+    IRQENTRY_TEXT
+    __fix_text_end = .;   /* tile-cpack won't rearrange before this */
+    ALIGN_FUNCTION();
+    *(.hottext*)
+    TEXT_TEXT
+    *(.text.*)
+    *(.coldtext*)
+    *(.fixup)
+    *(.gnu.warning)
+  } :text =0
+  _etext = .;
+
+  /* "Init" is divided into two areas with very different virtual addresses. */
+  INIT_TEXT_SECTION(PAGE_SIZE)
+
+  /* Now we skip back to PAGE_OFFSET for the data. */
+  . = (. - TEXT_OFFSET + PAGE_OFFSET);
+  #undef LOAD_OFFSET
+  #define LOAD_OFFSET PAGE_OFFSET
+
+  . = ALIGN(PAGE_SIZE);
+  __init_begin = .;
+  INIT_DATA_SECTION(16) :data =0
+  PERCPU_SECTION(L2_CACHE_BYTES)
+  . = ALIGN(PAGE_SIZE);
+  __init_end = .;
+
+  _sdata = .;                   /* Start of data section */
+  RO_DATA_SECTION(PAGE_SIZE)
+  RW_DATA_SECTION(L2_CACHE_BYTES, PAGE_SIZE, THREAD_SIZE)
+  _edata = .;
+
+  EXCEPTION_TABLE(L2_CACHE_BYTES)
+  NOTES
+
+
+  BSS_SECTION(8, PAGE_SIZE, 1)
+  _end = . ;
+
+  STABS_DEBUG
+  DWARF_DEBUG
+
+  DISCARDS
+}