x86: Move wcscmp SSE2 implementation to multiarch/wcscmp-sse2.S

This commit doesn't affect libc.so.6, its just housekeeping to prepare
for adding explicit ISA level support.

Tested build on x86_64 and x86_32 with/without multiarch.

1 files changed, 932 insertions, 4 deletions

diff --git a/sysdeps/x86_64/multiarch/wcscmp-sse2.S b/sysdeps/x86_64/multiarch/wcscmp-sse2.S
index 72a19bd64d..6cb7d9faf9 100644
--- a/sysdeps/x86_64/multiarch/wcscmp-sse2.S
+++ b/sysdeps/x86_64/multiarch/wcscmp-sse2.S
@@ -16,8 +16,936 @@
    License along with the GNU C Library; if not, see
    <https://www.gnu.org/licenses/>.  */
 
-#if IS_IN (libc)
-# define __wcscmp __wcscmp_sse2
-#endif
+#define USE_AS_WCSCMP
+#define STRCMP_ISA	_sse2
+#include "strcmp-naming.h"
 
-#include "../wcscmp.S"
+#include <sysdep.h>
+
+/* Note: wcscmp uses signed comparison, not unsighed as in strcmp function. */
+
+	.text
+ENTRY (STRCMP)
+/*
+	* This implementation uses SSE to compare up to 16 bytes at a time.
+*/
+	mov	%esi, %eax
+	mov	%edi, %edx
+	pxor	%xmm0, %xmm0		/* clear %xmm0 for null char checks */
+	mov	%al, %ch
+	mov	%dl, %cl
+	and	$63, %eax		/* rsi alignment in cache line */
+	and	$63, %edx		/* rdi alignment in cache line */
+	and	$15, %cl
+	jz	L(continue_00)
+	cmp	$16, %edx
+	jb	L(continue_0)
+	cmp	$32, %edx
+	jb	L(continue_16)
+	cmp	$48, %edx
+	jb	L(continue_32)
+
+L(continue_48):
+	and	$15, %ch
+	jz	L(continue_48_00)
+	cmp	$16, %eax
+	jb	L(continue_0_48)
+	cmp	$32, %eax
+	jb	L(continue_16_48)
+	cmp	$48, %eax
+	jb	L(continue_32_48)
+
+	.p2align 4
+L(continue_48_48):
+	mov	(%rsi), %ecx
+	cmp	%ecx, (%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	4(%rsi), %ecx
+	cmp	%ecx, 4(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	8(%rsi), %ecx
+	cmp	%ecx, 8(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	12(%rsi), %ecx
+	cmp	%ecx, 12(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	movdqu	16(%rdi), %xmm1
+	movdqu	16(%rsi), %xmm2
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	%xmm2, %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_16)
+
+	movdqu	32(%rdi), %xmm1
+	movdqu	32(%rsi), %xmm2
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	%xmm2, %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_32)
+
+	movdqu	48(%rdi), %xmm1
+	movdqu	48(%rsi), %xmm2
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	%xmm2, %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_48)
+
+	add	$64, %rsi
+	add	$64, %rdi
+	jmp	L(continue_48_48)
+
+L(continue_0):
+	and	$15, %ch
+	jz	L(continue_0_00)
+	cmp	$16, %eax
+	jb	L(continue_0_0)
+	cmp	$32, %eax
+	jb	L(continue_0_16)
+	cmp	$48, %eax
+	jb	L(continue_0_32)
+
+	.p2align 4
+L(continue_0_48):
+	mov	(%rsi), %ecx
+	cmp	%ecx, (%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	4(%rsi), %ecx
+	cmp	%ecx, 4(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	8(%rsi), %ecx
+	cmp	%ecx, 8(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	12(%rsi), %ecx
+	cmp	%ecx, 12(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	movdqu	16(%rdi), %xmm1
+	movdqu	16(%rsi), %xmm2
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	%xmm2, %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_16)
+
+	movdqu	32(%rdi), %xmm1
+	movdqu	32(%rsi), %xmm2
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	%xmm2, %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_32)
+
+	mov	48(%rsi), %ecx
+	cmp	%ecx, 48(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	52(%rsi), %ecx
+	cmp	%ecx, 52(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	56(%rsi), %ecx
+	cmp	%ecx, 56(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	60(%rsi), %ecx
+	cmp	%ecx, 60(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	add	$64, %rsi
+	add	$64, %rdi
+	jmp	L(continue_0_48)
+
+	.p2align 4
+L(continue_00):
+	and	$15, %ch
+	jz	L(continue_00_00)
+	cmp	$16, %eax
+	jb	L(continue_00_0)
+	cmp	$32, %eax
+	jb	L(continue_00_16)
+	cmp	$48, %eax
+	jb	L(continue_00_32)
+
+	.p2align 4
+L(continue_00_48):
+	pcmpeqd	(%rdi), %xmm0
+	mov	(%rdi), %eax
+	pmovmskb %xmm0, %ecx
+	test	%ecx, %ecx
+	jnz	L(less4_double_words1)
+
+	cmp	(%rsi), %eax
+	jne	L(nequal)
+
+	mov	4(%rdi), %eax
+	cmp	4(%rsi), %eax
+	jne	L(nequal)
+
+	mov	8(%rdi), %eax
+	cmp	8(%rsi), %eax
+	jne	L(nequal)
+
+	mov	12(%rdi), %eax
+	cmp	12(%rsi), %eax
+	jne	L(nequal)
+
+	movdqu	16(%rsi), %xmm2
+	pcmpeqd	%xmm2, %xmm0		/* Any null double_word? */
+	pcmpeqd	16(%rdi), %xmm2		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm2		/* packed sub of comparison results*/
+	pmovmskb %xmm2, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_16)
+
+	movdqu	32(%rsi), %xmm2
+	pcmpeqd	%xmm2, %xmm0		/* Any null double_word? */
+	pcmpeqd	32(%rdi), %xmm2		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm2		/* packed sub of comparison results*/
+	pmovmskb %xmm2, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_32)
+
+	movdqu	48(%rsi), %xmm2
+	pcmpeqd	%xmm2, %xmm0		/* Any null double_word? */
+	pcmpeqd	48(%rdi), %xmm2		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm2		/* packed sub of comparison results*/
+	pmovmskb %xmm2, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_48)
+
+	add	$64, %rsi
+	add	$64, %rdi
+	jmp	L(continue_00_48)
+
+	.p2align 4
+L(continue_32):
+	and	$15, %ch
+	jz	L(continue_32_00)
+	cmp	$16, %eax
+	jb	L(continue_0_32)
+	cmp	$32, %eax
+	jb	L(continue_16_32)
+	cmp	$48, %eax
+	jb	L(continue_32_32)
+
+	.p2align 4
+L(continue_32_48):
+	mov	(%rsi), %ecx
+	cmp	%ecx, (%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	4(%rsi), %ecx
+	cmp	%ecx, 4(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	8(%rsi), %ecx
+	cmp	%ecx, 8(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	12(%rsi), %ecx
+	cmp	%ecx, 12(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	16(%rsi), %ecx
+	cmp	%ecx, 16(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	20(%rsi), %ecx
+	cmp	%ecx, 20(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	24(%rsi), %ecx
+	cmp	%ecx, 24(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	28(%rsi), %ecx
+	cmp	%ecx, 28(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	movdqu	32(%rdi), %xmm1
+	movdqu	32(%rsi), %xmm2
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	%xmm2, %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_32)
+
+	movdqu	48(%rdi), %xmm1
+	movdqu	48(%rsi), %xmm2
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	%xmm2, %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_48)
+
+	add	$64, %rsi
+	add	$64, %rdi
+	jmp	L(continue_32_48)
+
+	.p2align 4
+L(continue_16):
+	and	$15, %ch
+	jz	L(continue_16_00)
+	cmp	$16, %eax
+	jb	L(continue_0_16)
+	cmp	$32, %eax
+	jb	L(continue_16_16)
+	cmp	$48, %eax
+	jb	L(continue_16_32)
+
+	.p2align 4
+L(continue_16_48):
+	mov	(%rsi), %ecx
+	cmp	%ecx, (%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	4(%rsi), %ecx
+	cmp	%ecx, 4(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	8(%rsi), %ecx
+	cmp	%ecx, 8(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	12(%rsi), %ecx
+	cmp	%ecx, 12(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	movdqu	16(%rdi), %xmm1
+	movdqu	16(%rsi), %xmm2
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	%xmm2, %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_16)
+
+	mov	32(%rsi), %ecx
+	cmp	%ecx, 32(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	36(%rsi), %ecx
+	cmp	%ecx, 36(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	40(%rsi), %ecx
+	cmp	%ecx, 40(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	44(%rsi), %ecx
+	cmp	%ecx, 44(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	movdqu	48(%rdi), %xmm1
+	movdqu	48(%rsi), %xmm2
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	%xmm2, %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_48)
+
+	add	$64, %rsi
+	add	$64, %rdi
+	jmp	L(continue_16_48)
+
+	.p2align 4
+L(continue_00_00):
+	movdqa	(%rdi), %xmm1
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	(%rsi), %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words)
+
+	movdqa	16(%rdi), %xmm3
+	pcmpeqd	%xmm3, %xmm0		/* Any null double_word? */
+	pcmpeqd	16(%rsi), %xmm3		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm3		/* packed sub of comparison results*/
+	pmovmskb %xmm3, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_16)
+
+	movdqa	32(%rdi), %xmm5
+	pcmpeqd	%xmm5, %xmm0		/* Any null double_word? */
+	pcmpeqd	32(%rsi), %xmm5		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm5		/* packed sub of comparison results*/
+	pmovmskb %xmm5, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_32)
+
+	movdqa	48(%rdi), %xmm1
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	48(%rsi), %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_48)
+
+	add	$64, %rsi
+	add	$64, %rdi
+	jmp	L(continue_00_00)
+
+	.p2align 4
+L(continue_00_32):
+	movdqu	(%rsi), %xmm2
+	pcmpeqd	%xmm2, %xmm0		/* Any null double_word? */
+	pcmpeqd	(%rdi), %xmm2		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm2		/* packed sub of comparison results*/
+	pmovmskb %xmm2, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words)
+
+	add	$16, %rsi
+	add	$16, %rdi
+	jmp	L(continue_00_48)
+
+	.p2align 4
+L(continue_00_16):
+	movdqu	(%rsi), %xmm2
+	pcmpeqd	%xmm2, %xmm0		/* Any null double_word? */
+	pcmpeqd	(%rdi), %xmm2		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm2		/* packed sub of comparison results*/
+	pmovmskb %xmm2, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words)
+
+	movdqu	16(%rsi), %xmm2
+	pcmpeqd	%xmm2, %xmm0		/* Any null double_word? */
+	pcmpeqd	16(%rdi), %xmm2		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm2		/* packed sub of comparison results*/
+	pmovmskb %xmm2, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_16)
+
+	add	$32, %rsi
+	add	$32, %rdi
+	jmp	L(continue_00_48)
+
+	.p2align 4
+L(continue_00_0):
+	movdqu	(%rsi), %xmm2
+	pcmpeqd	%xmm2, %xmm0		/* Any null double_word? */
+	pcmpeqd	(%rdi), %xmm2		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm2		/* packed sub of comparison results*/
+	pmovmskb %xmm2, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words)
+
+	movdqu	16(%rsi), %xmm2
+	pcmpeqd	%xmm2, %xmm0		/* Any null double_word? */
+	pcmpeqd	16(%rdi), %xmm2		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm2		/* packed sub of comparison results*/
+	pmovmskb %xmm2, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_16)
+
+	movdqu	32(%rsi), %xmm2
+	pcmpeqd	%xmm2, %xmm0		/* Any null double_word? */
+	pcmpeqd	32(%rdi), %xmm2		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm2		/* packed sub of comparison results*/
+	pmovmskb %xmm2, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_32)
+
+	add	$48, %rsi
+	add	$48, %rdi
+	jmp	L(continue_00_48)
+
+	.p2align 4
+L(continue_48_00):
+	pcmpeqd	(%rsi), %xmm0
+	mov	(%rdi), %eax
+	pmovmskb %xmm0, %ecx
+	test	%ecx, %ecx
+	jnz	L(less4_double_words1)
+
+	cmp	(%rsi), %eax
+	jne	L(nequal)
+
+	mov	4(%rdi), %eax
+	cmp	4(%rsi), %eax
+	jne	L(nequal)
+
+	mov	8(%rdi), %eax
+	cmp	8(%rsi), %eax
+	jne	L(nequal)
+
+	mov	12(%rdi), %eax
+	cmp	12(%rsi), %eax
+	jne	L(nequal)
+
+	movdqu	16(%rdi), %xmm1
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	16(%rsi), %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_16)
+
+	movdqu	32(%rdi), %xmm1
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	32(%rsi), %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_32)
+
+	movdqu	48(%rdi), %xmm1
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	48(%rsi), %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_48)
+
+	add	$64, %rsi
+	add	$64, %rdi
+	jmp	L(continue_48_00)
+
+	.p2align 4
+L(continue_32_00):
+	movdqu	(%rdi), %xmm1
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	(%rsi), %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words)
+
+	add	$16, %rsi
+	add	$16, %rdi
+	jmp	L(continue_48_00)
+
+	.p2align 4
+L(continue_16_00):
+	movdqu	(%rdi), %xmm1
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	(%rsi), %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words)
+
+	movdqu	16(%rdi), %xmm1
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	16(%rsi), %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_16)
+
+	add	$32, %rsi
+	add	$32, %rdi
+	jmp	L(continue_48_00)
+
+	.p2align 4
+L(continue_0_00):
+	movdqu	(%rdi), %xmm1
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	(%rsi), %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words)
+
+	movdqu	16(%rdi), %xmm1
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	16(%rsi), %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_16)
+
+	movdqu	32(%rdi), %xmm1
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	32(%rsi), %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_32)
+
+	add	$48, %rsi
+	add	$48, %rdi
+	jmp	L(continue_48_00)
+
+	.p2align 4
+L(continue_32_32):
+	movdqu	(%rdi), %xmm1
+	movdqu	(%rsi), %xmm2
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	%xmm2, %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words)
+
+	add	$16, %rsi
+	add	$16, %rdi
+	jmp	L(continue_48_48)
+
+	.p2align 4
+L(continue_16_16):
+	movdqu	(%rdi), %xmm1
+	movdqu	(%rsi), %xmm2
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	%xmm2, %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words)
+
+	movdqu	16(%rdi), %xmm3
+	movdqu	16(%rsi), %xmm4
+	pcmpeqd	%xmm3, %xmm0		/* Any null double_word? */
+	pcmpeqd	%xmm4, %xmm3		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm3		/* packed sub of comparison results*/
+	pmovmskb %xmm3, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_16)
+
+	add	$32, %rsi
+	add	$32, %rdi
+	jmp	L(continue_48_48)
+
+	.p2align 4
+L(continue_0_0):
+	movdqu	(%rdi), %xmm1
+	movdqu	(%rsi), %xmm2
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	%xmm2, %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words)
+
+	movdqu	16(%rdi), %xmm3
+	movdqu	16(%rsi), %xmm4
+	pcmpeqd	%xmm3, %xmm0		/* Any null double_word? */
+	pcmpeqd	%xmm4, %xmm3		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm3		/* packed sub of comparison results*/
+	pmovmskb %xmm3, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_16)
+
+	movdqu	32(%rdi), %xmm1
+	movdqu	32(%rsi), %xmm2
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	%xmm2, %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_32)
+
+	add	$48, %rsi
+	add	$48, %rdi
+	jmp	L(continue_48_48)
+
+	.p2align 4
+L(continue_0_16):
+	movdqu	(%rdi), %xmm1
+	movdqu	(%rsi), %xmm2
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	%xmm2, %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words)
+
+	movdqu	16(%rdi), %xmm1
+	movdqu	16(%rsi), %xmm2
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	%xmm2, %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words_16)
+
+	add	$32, %rsi
+	add	$32, %rdi
+	jmp	L(continue_32_48)
+
+	.p2align 4
+L(continue_0_32):
+	movdqu	(%rdi), %xmm1
+	movdqu	(%rsi), %xmm2
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	%xmm2, %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words)
+
+	add	$16, %rsi
+	add	$16, %rdi
+	jmp	L(continue_16_48)
+
+	.p2align 4
+L(continue_16_32):
+	movdqu	(%rdi), %xmm1
+	movdqu	(%rsi), %xmm2
+	pcmpeqd	%xmm1, %xmm0		/* Any null double_word? */
+	pcmpeqd	%xmm2, %xmm1		/* compare first 4 double_words for equality */
+	psubb	%xmm0, %xmm1		/* packed sub of comparison results*/
+	pmovmskb %xmm1, %edx
+	sub	$0xffff, %edx		/* if first 4 double_words are same, edx == 0xffff */
+	jnz	L(less4_double_words)
+
+	add	$16, %rsi
+	add	$16, %rdi
+	jmp	L(continue_32_48)
+
+	.p2align 4
+L(less4_double_words1):
+	cmp	(%rsi), %eax
+	jne	L(nequal)
+	test	%eax, %eax
+	jz	L(equal)
+
+	mov	4(%rsi), %ecx
+	cmp	%ecx, 4(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	8(%rsi), %ecx
+	cmp	%ecx, 8(%rdi)
+	jne	L(nequal)
+	test	%ecx, %ecx
+	jz	L(equal)
+
+	mov	12(%rsi), %ecx
+	cmp	%ecx, 12(%rdi)
+	jne	L(nequal)
+	xor	%eax, %eax
+	ret
+
+	.p2align 4
+L(less4_double_words):
+	xor	%eax, %eax
+	test	%dl, %dl
+	jz	L(next_two_double_words)
+	and	$15, %dl
+	jz	L(second_double_word)
+	mov	(%rdi), %eax
+	cmp	(%rsi), %eax
+	jne	L(nequal)
+	ret
+
+	.p2align 4
+L(second_double_word):
+	mov	4(%rdi), %eax
+	cmp	4(%rsi), %eax
+	jne	L(nequal)
+	ret
+
+	.p2align 4
+L(next_two_double_words):
+	and	$15, %dh
+	jz	L(fourth_double_word)
+	mov	8(%rdi), %eax
+	cmp	8(%rsi), %eax
+	jne	L(nequal)
+	ret
+
+	.p2align 4
+L(fourth_double_word):
+	mov	12(%rdi), %eax
+	cmp	12(%rsi), %eax
+	jne	L(nequal)
+	ret
+
+	.p2align 4
+L(less4_double_words_16):
+	xor	%eax, %eax
+	test	%dl, %dl
+	jz	L(next_two_double_words_16)
+	and	$15, %dl
+	jz	L(second_double_word_16)
+	mov	16(%rdi), %eax
+	cmp	16(%rsi), %eax
+	jne	L(nequal)
+	ret
+
+	.p2align 4
+L(second_double_word_16):
+	mov	20(%rdi), %eax
+	cmp	20(%rsi), %eax
+	jne	L(nequal)
+	ret
+
+	.p2align 4
+L(next_two_double_words_16):
+	and	$15, %dh
+	jz	L(fourth_double_word_16)
+	mov	24(%rdi), %eax
+	cmp	24(%rsi), %eax
+	jne	L(nequal)
+	ret
+
+	.p2align 4
+L(fourth_double_word_16):
+	mov	28(%rdi), %eax
+	cmp	28(%rsi), %eax
+	jne	L(nequal)
+	ret
+
+	.p2align 4
+L(less4_double_words_32):
+	xor	%eax, %eax
+	test	%dl, %dl
+	jz	L(next_two_double_words_32)
+	and	$15, %dl
+	jz	L(second_double_word_32)
+	mov	32(%rdi), %eax
+	cmp	32(%rsi), %eax
+	jne	L(nequal)
+	ret
+
+	.p2align 4
+L(second_double_word_32):
+	mov	36(%rdi), %eax
+	cmp	36(%rsi), %eax
+	jne	L(nequal)
+	ret
+
+	.p2align 4
+L(next_two_double_words_32):
+	and	$15, %dh
+	jz	L(fourth_double_word_32)
+	mov	40(%rdi), %eax
+	cmp	40(%rsi), %eax
+	jne	L(nequal)
+	ret
+
+	.p2align 4
+L(fourth_double_word_32):
+	mov	44(%rdi), %eax
+	cmp	44(%rsi), %eax
+	jne	L(nequal)
+	ret
+
+	.p2align 4
+L(less4_double_words_48):
+	xor	%eax, %eax
+	test	%dl, %dl
+	jz	L(next_two_double_words_48)
+	and	$15, %dl
+	jz	L(second_double_word_48)
+	mov	48(%rdi), %eax
+	cmp	48(%rsi), %eax
+	jne	L(nequal)
+	ret
+
+	.p2align 4
+L(second_double_word_48):
+	mov	52(%rdi), %eax
+	cmp	52(%rsi), %eax
+	jne	L(nequal)
+	ret
+
+	.p2align 4
+L(next_two_double_words_48):
+	and	$15, %dh
+	jz	L(fourth_double_word_48)
+	mov	56(%rdi), %eax
+	cmp	56(%rsi), %eax
+	jne	L(nequal)
+	ret
+
+	.p2align 4
+L(fourth_double_word_48):
+	mov	60(%rdi), %eax
+	cmp	60(%rsi), %eax
+	jne	L(nequal)
+	ret
+
+	.p2align 4
+L(nequal):
+	mov	$1, %eax
+	jg	L(nequal_bigger)
+	neg	%eax
+
+L(nequal_bigger):
+	ret
+
+	.p2align 4
+L(equal):
+	xor	%rax, %rax
+	ret
+
+END (STRCMP)