summaryrefslogtreecommitdiff
path: root/arch/sparc/lib/NG2memcpy.S
diff options
context:
space:
mode:
Diffstat (limited to 'arch/sparc/lib/NG2memcpy.S')
-rw-r--r--arch/sparc/lib/NG2memcpy.S118
1 files changed, 62 insertions, 56 deletions
diff --git a/arch/sparc/lib/NG2memcpy.S b/arch/sparc/lib/NG2memcpy.S
index 30eee6e8a..d5f585df2 100644
--- a/arch/sparc/lib/NG2memcpy.S
+++ b/arch/sparc/lib/NG2memcpy.S
@@ -34,10 +34,16 @@
#ifndef EX_LD
#define EX_LD(x) x
#endif
+#ifndef EX_LD_FP
+#define EX_LD_FP(x) x
+#endif
#ifndef EX_ST
#define EX_ST(x) x
#endif
+#ifndef EX_ST_FP
+#define EX_ST_FP(x) x
+#endif
#ifndef EX_RETVAL
#define EX_RETVAL(x) x
@@ -134,40 +140,40 @@
fsrc2 %x6, %f12; \
fsrc2 %x7, %f14;
#define FREG_LOAD_1(base, x0) \
- EX_LD(LOAD(ldd, base + 0x00, %x0))
+ EX_LD_FP(LOAD(ldd, base + 0x00, %x0))
#define FREG_LOAD_2(base, x0, x1) \
- EX_LD(LOAD(ldd, base + 0x00, %x0)); \
- EX_LD(LOAD(ldd, base + 0x08, %x1));
+ EX_LD_FP(LOAD(ldd, base + 0x00, %x0)); \
+ EX_LD_FP(LOAD(ldd, base + 0x08, %x1));
#define FREG_LOAD_3(base, x0, x1, x2) \
- EX_LD(LOAD(ldd, base + 0x00, %x0)); \
- EX_LD(LOAD(ldd, base + 0x08, %x1)); \
- EX_LD(LOAD(ldd, base + 0x10, %x2));
+ EX_LD_FP(LOAD(ldd, base + 0x00, %x0)); \
+ EX_LD_FP(LOAD(ldd, base + 0x08, %x1)); \
+ EX_LD_FP(LOAD(ldd, base + 0x10, %x2));
#define FREG_LOAD_4(base, x0, x1, x2, x3) \
- EX_LD(LOAD(ldd, base + 0x00, %x0)); \
- EX_LD(LOAD(ldd, base + 0x08, %x1)); \
- EX_LD(LOAD(ldd, base + 0x10, %x2)); \
- EX_LD(LOAD(ldd, base + 0x18, %x3));
+ EX_LD_FP(LOAD(ldd, base + 0x00, %x0)); \
+ EX_LD_FP(LOAD(ldd, base + 0x08, %x1)); \
+ EX_LD_FP(LOAD(ldd, base + 0x10, %x2)); \
+ EX_LD_FP(LOAD(ldd, base + 0x18, %x3));
#define FREG_LOAD_5(base, x0, x1, x2, x3, x4) \
- EX_LD(LOAD(ldd, base + 0x00, %x0)); \
- EX_LD(LOAD(ldd, base + 0x08, %x1)); \
- EX_LD(LOAD(ldd, base + 0x10, %x2)); \
- EX_LD(LOAD(ldd, base + 0x18, %x3)); \
- EX_LD(LOAD(ldd, base + 0x20, %x4));
+ EX_LD_FP(LOAD(ldd, base + 0x00, %x0)); \
+ EX_LD_FP(LOAD(ldd, base + 0x08, %x1)); \
+ EX_LD_FP(LOAD(ldd, base + 0x10, %x2)); \
+ EX_LD_FP(LOAD(ldd, base + 0x18, %x3)); \
+ EX_LD_FP(LOAD(ldd, base + 0x20, %x4));
#define FREG_LOAD_6(base, x0, x1, x2, x3, x4, x5) \
- EX_LD(LOAD(ldd, base + 0x00, %x0)); \
- EX_LD(LOAD(ldd, base + 0x08, %x1)); \
- EX_LD(LOAD(ldd, base + 0x10, %x2)); \
- EX_LD(LOAD(ldd, base + 0x18, %x3)); \
- EX_LD(LOAD(ldd, base + 0x20, %x4)); \
- EX_LD(LOAD(ldd, base + 0x28, %x5));
+ EX_LD_FP(LOAD(ldd, base + 0x00, %x0)); \
+ EX_LD_FP(LOAD(ldd, base + 0x08, %x1)); \
+ EX_LD_FP(LOAD(ldd, base + 0x10, %x2)); \
+ EX_LD_FP(LOAD(ldd, base + 0x18, %x3)); \
+ EX_LD_FP(LOAD(ldd, base + 0x20, %x4)); \
+ EX_LD_FP(LOAD(ldd, base + 0x28, %x5));
#define FREG_LOAD_7(base, x0, x1, x2, x3, x4, x5, x6) \
- EX_LD(LOAD(ldd, base + 0x00, %x0)); \
- EX_LD(LOAD(ldd, base + 0x08, %x1)); \
- EX_LD(LOAD(ldd, base + 0x10, %x2)); \
- EX_LD(LOAD(ldd, base + 0x18, %x3)); \
- EX_LD(LOAD(ldd, base + 0x20, %x4)); \
- EX_LD(LOAD(ldd, base + 0x28, %x5)); \
- EX_LD(LOAD(ldd, base + 0x30, %x6));
+ EX_LD_FP(LOAD(ldd, base + 0x00, %x0)); \
+ EX_LD_FP(LOAD(ldd, base + 0x08, %x1)); \
+ EX_LD_FP(LOAD(ldd, base + 0x10, %x2)); \
+ EX_LD_FP(LOAD(ldd, base + 0x18, %x3)); \
+ EX_LD_FP(LOAD(ldd, base + 0x20, %x4)); \
+ EX_LD_FP(LOAD(ldd, base + 0x28, %x5)); \
+ EX_LD_FP(LOAD(ldd, base + 0x30, %x6));
.register %g2,#scratch
.register %g3,#scratch
@@ -275,11 +281,11 @@ FUNC_NAME: /* %o0=dst, %o1=src, %o2=len */
nop
/* fall through for 0 < low bits < 8 */
110: sub %o4, 64, %g2
- EX_LD(LOAD_BLK(%g2, %f0))
-1: EX_ST(STORE_INIT(%g0, %o4 + %g3))
- EX_LD(LOAD_BLK(%o4, %f16))
+ EX_LD_FP(LOAD_BLK(%g2, %f0))
+1: EX_ST_FP(STORE_INIT(%g0, %o4 + %g3))
+ EX_LD_FP(LOAD_BLK(%o4, %f16))
FREG_FROB(f0, f2, f4, f6, f8, f10, f12, f14, f16)
- EX_ST(STORE_BLK(%f0, %o4 + %g3))
+ EX_ST_FP(STORE_BLK(%f0, %o4 + %g3))
FREG_MOVE_8(f16, f18, f20, f22, f24, f26, f28, f30)
subcc %g1, 64, %g1
add %o4, 64, %o4
@@ -290,10 +296,10 @@ FUNC_NAME: /* %o0=dst, %o1=src, %o2=len */
120: sub %o4, 56, %g2
FREG_LOAD_7(%g2, f0, f2, f4, f6, f8, f10, f12)
-1: EX_ST(STORE_INIT(%g0, %o4 + %g3))
- EX_LD(LOAD_BLK(%o4, %f16))
+1: EX_ST_FP(STORE_INIT(%g0, %o4 + %g3))
+ EX_LD_FP(LOAD_BLK(%o4, %f16))
FREG_FROB(f0, f2, f4, f6, f8, f10, f12, f16, f18)
- EX_ST(STORE_BLK(%f0, %o4 + %g3))
+ EX_ST_FP(STORE_BLK(%f0, %o4 + %g3))
FREG_MOVE_7(f18, f20, f22, f24, f26, f28, f30)
subcc %g1, 64, %g1
add %o4, 64, %o4
@@ -304,10 +310,10 @@ FUNC_NAME: /* %o0=dst, %o1=src, %o2=len */
130: sub %o4, 48, %g2
FREG_LOAD_6(%g2, f0, f2, f4, f6, f8, f10)
-1: EX_ST(STORE_INIT(%g0, %o4 + %g3))
- EX_LD(LOAD_BLK(%o4, %f16))
+1: EX_ST_FP(STORE_INIT(%g0, %o4 + %g3))
+ EX_LD_FP(LOAD_BLK(%o4, %f16))
FREG_FROB(f0, f2, f4, f6, f8, f10, f16, f18, f20)
- EX_ST(STORE_BLK(%f0, %o4 + %g3))
+ EX_ST_FP(STORE_BLK(%f0, %o4 + %g3))
FREG_MOVE_6(f20, f22, f24, f26, f28, f30)
subcc %g1, 64, %g1
add %o4, 64, %o4
@@ -318,10 +324,10 @@ FUNC_NAME: /* %o0=dst, %o1=src, %o2=len */
140: sub %o4, 40, %g2
FREG_LOAD_5(%g2, f0, f2, f4, f6, f8)
-1: EX_ST(STORE_INIT(%g0, %o4 + %g3))
- EX_LD(LOAD_BLK(%o4, %f16))
+1: EX_ST_FP(STORE_INIT(%g0, %o4 + %g3))
+ EX_LD_FP(LOAD_BLK(%o4, %f16))
FREG_FROB(f0, f2, f4, f6, f8, f16, f18, f20, f22)
- EX_ST(STORE_BLK(%f0, %o4 + %g3))
+ EX_ST_FP(STORE_BLK(%f0, %o4 + %g3))
FREG_MOVE_5(f22, f24, f26, f28, f30)
subcc %g1, 64, %g1
add %o4, 64, %o4
@@ -332,10 +338,10 @@ FUNC_NAME: /* %o0=dst, %o1=src, %o2=len */
150: sub %o4, 32, %g2
FREG_LOAD_4(%g2, f0, f2, f4, f6)
-1: EX_ST(STORE_INIT(%g0, %o4 + %g3))
- EX_LD(LOAD_BLK(%o4, %f16))
+1: EX_ST_FP(STORE_INIT(%g0, %o4 + %g3))
+ EX_LD_FP(LOAD_BLK(%o4, %f16))
FREG_FROB(f0, f2, f4, f6, f16, f18, f20, f22, f24)
- EX_ST(STORE_BLK(%f0, %o4 + %g3))
+ EX_ST_FP(STORE_BLK(%f0, %o4 + %g3))
FREG_MOVE_4(f24, f26, f28, f30)
subcc %g1, 64, %g1
add %o4, 64, %o4
@@ -346,10 +352,10 @@ FUNC_NAME: /* %o0=dst, %o1=src, %o2=len */
160: sub %o4, 24, %g2
FREG_LOAD_3(%g2, f0, f2, f4)
-1: EX_ST(STORE_INIT(%g0, %o4 + %g3))
- EX_LD(LOAD_BLK(%o4, %f16))
+1: EX_ST_FP(STORE_INIT(%g0, %o4 + %g3))
+ EX_LD_FP(LOAD_BLK(%o4, %f16))
FREG_FROB(f0, f2, f4, f16, f18, f20, f22, f24, f26)
- EX_ST(STORE_BLK(%f0, %o4 + %g3))
+ EX_ST_FP(STORE_BLK(%f0, %o4 + %g3))
FREG_MOVE_3(f26, f28, f30)
subcc %g1, 64, %g1
add %o4, 64, %o4
@@ -360,10 +366,10 @@ FUNC_NAME: /* %o0=dst, %o1=src, %o2=len */
170: sub %o4, 16, %g2
FREG_LOAD_2(%g2, f0, f2)
-1: EX_ST(STORE_INIT(%g0, %o4 + %g3))
- EX_LD(LOAD_BLK(%o4, %f16))
+1: EX_ST_FP(STORE_INIT(%g0, %o4 + %g3))
+ EX_LD_FP(LOAD_BLK(%o4, %f16))
FREG_FROB(f0, f2, f16, f18, f20, f22, f24, f26, f28)
- EX_ST(STORE_BLK(%f0, %o4 + %g3))
+ EX_ST_FP(STORE_BLK(%f0, %o4 + %g3))
FREG_MOVE_2(f28, f30)
subcc %g1, 64, %g1
add %o4, 64, %o4
@@ -374,10 +380,10 @@ FUNC_NAME: /* %o0=dst, %o1=src, %o2=len */
180: sub %o4, 8, %g2
FREG_LOAD_1(%g2, f0)
-1: EX_ST(STORE_INIT(%g0, %o4 + %g3))
- EX_LD(LOAD_BLK(%o4, %f16))
+1: EX_ST_FP(STORE_INIT(%g0, %o4 + %g3))
+ EX_LD_FP(LOAD_BLK(%o4, %f16))
FREG_FROB(f0, f16, f18, f20, f22, f24, f26, f28, f30)
- EX_ST(STORE_BLK(%f0, %o4 + %g3))
+ EX_ST_FP(STORE_BLK(%f0, %o4 + %g3))
FREG_MOVE_1(f30)
subcc %g1, 64, %g1
add %o4, 64, %o4
@@ -387,10 +393,10 @@ FUNC_NAME: /* %o0=dst, %o1=src, %o2=len */
nop
190:
-1: EX_ST(STORE_INIT(%g0, %o4 + %g3))
+1: EX_ST_FP(STORE_INIT(%g0, %o4 + %g3))
subcc %g1, 64, %g1
- EX_LD(LOAD_BLK(%o4, %f0))
- EX_ST(STORE_BLK(%f0, %o4 + %g3))
+ EX_LD_FP(LOAD_BLK(%o4, %f0))
+ EX_ST_FP(STORE_BLK(%f0, %o4 + %g3))
add %o4, 64, %o4
bne,pt %xcc, 1b
LOAD(prefetch, %o4 + 64, #one_read)
generated by cgit v1.2.3-54-g00ecf (git 2.45.2) at 2024-11-19 06:00:13 +0000