1 files changed, 88 insertions, 0 deletions
diff --git a/arch/cris/arch-v32/lib/checksum.S b/arch/cris/arch-v32/lib/checksum.S
new file mode 100644
index 000000000..4a72a94a4
--- /dev/null
+++ b/arch/cris/arch-v32/lib/checksum.S
@@ -0,0 +1,88 @@
+/*
+ * A fast checksum routine using movem
+ * Copyright (c) 1998-2007 Axis Communications AB
+ *
+ * csum_partial(const unsigned char * buff, int len, unsigned int sum)
+ */
+
+	.globl	csum_partial
+	.type   csum_partial,@function
+csum_partial:
+
+	;; r10 - src
+	;; r11 - length
+	;; r12 - checksum
+
+	;; Optimized for large packets
+	subq	10*4, $r11
+	blt	_word_loop
+	move.d	$r11, $acr
+
+	subq	9*4,$sp
+	clearf	c
+	movem	$r8,[$sp]
+
+	;; do a movem checksum
+
+_mloop:	movem	[$r10+],$r9	; read 10 longwords
+	;; Loop count without touching the c flag.
+	addoq	-10*4, $acr, $acr
+	;; perform dword checksumming on the 10 longwords
+
+	addc	$r0,$r12
+	addc	$r1,$r12
+	addc	$r2,$r12
+	addc	$r3,$r12
+	addc	$r4,$r12
+	addc	$r5,$r12
+	addc	$r6,$r12
+	addc	$r7,$r12
+	addc	$r8,$r12
+	addc	$r9,$r12
+
+	;; test $acr without trashing carry.
+	move.d	$acr, $acr
+	bpl	_mloop
+	;; r11 <= acr  is not really needed in the mloop, just using the dslot
+	;; to prepare for what is needed after mloop.
+	move.d	$acr, $r11
+
+	;; fold the last carry into r13
+	addc	0, $r12
+	movem	[$sp+],$r8	; restore regs
+
+_word_loop:
+	addq	10*4,$r11	; compensate for last loop underflowing length
+
+	moveq	-1,$r9		; put 0xffff in r9, faster than move.d 0xffff,r9
+	lsrq	16,$r9
+
+	move.d	$r12,$r13
+	lsrq	16,$r13		; r13 = checksum >> 16
+	and.d	$r9,$r12	; checksum = checksum & 0xffff
+
+_no_fold:
+	subq	2,$r11
+	blt	_no_words
+	add.d	$r13,$r12	; checksum += r13
+
+	;; checksum the rest of the words
+_wloop:	subq	2,$r11
+	bge	_wloop
+	addu.w	[$r10+],$r12
+
+_no_words:
+	addq	2,$r11
+	;; see if we have one odd byte more
+	bne	_do_byte
+	nop
+	ret
+	move.d	$r12,$r10
+
+_do_byte:
+	;; copy and checksum the last byte
+	addu.b	[$r10],$r12
+	ret
+	move.d	$r12,$r10
+
+	.size   csum_partial, .-csum_partial