Optimized memchr, memrchr, rawmemchr for x86-32

14 files changed, 2666 insertions, 4 deletions

diff --git a/ChangeLog b/ChangeLog
index 5dd550cba9..1b7499c805 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,20 @@
+2011-10-12  Liubov Dmitrieva  <liubov.dmitrieva@gmail.com>
+
+	* sysdeps/i386/i686/multiarch/Makefile (sysdep_routines): Add
+	memchr-sse2 memchr-sse2-bsf memrchr-sse2 memrchr-sse2-bsf memrchr-c
+	rawmemchr-sse2 rawmemchr-sse2-bsf.
+	* sysdeps/i386/i686/multiarch/memchr.S: New file.
+	* sysdeps/i386/i686/multiarch/memchr-sse2.S: New file.
+	* sysdeps/i386/i686/multiarch/memchr-sse2-bsf.S: New file.
+	* sysdeps/i386/i686/multiarch/memrchr.S: New file.
+	* sysdeps/i386/i686/multiarch/memrchr-c.c: New file.
+	* sysdeps/i386/i686/multiarch/memrchr-sse2.S: New file.
+	* sysdeps/i386/i686/multiarch/memrchr-sse2-bsf.S: New file.
+	* sysdeps/i386/i686/multiarch/rawmemchr.S: New file.
+	* sysdeps/i386/i686/multiarch/rawmemchr-sse2.S: New file.
+	* sysdeps/i386/i686/multiarch/rawmemchr-sse2-bsf.S: New file.
+	* string/memrchr.c (MEMRCHR): New macro.
+
 2011-10-12  Ulrich Drepper  <drepper@gmail.com>
 
 	Add integration with gcc's -ffinite-math-only and optimize wrapper
diff --git a/NEWS b/NEWS
index 9940957953..e10182b9a8 100644
--- a/NEWS
+++ b/NEWS
@@ -33,7 +33,7 @@ Version 2.15
 * Optimized strchr and strrchr for SSE on x86-32.
   Contributed by Liubov Dmitrieva.
 
-* Optimized memchr, memrchr, rawmemchr for x86-64.
+* Optimized memchr, memrchr, rawmemchr for x86-64 and x86-32.
   Contributed by Liubov Dmitrieva.
 
 * New interfaces: scandirat, scandirat64
diff --git a/string/memrchr.c b/string/memrchr.c
index 21662b1bd7..498a4376ea 100644
--- a/string/memrchr.c
+++ b/string/memrchr.c
@@ -63,7 +63,12 @@
 
 /* Search no more than N bytes of S for C.  */
 __ptr_t
-__memrchr (s, c_in, n)
+#ifndef MEMRCHR
+__memrchr
+#else
+MEMRCHR
+#endif
+     (s, c_in, n)
      const __ptr_t s;
      int c_in;
      size_t n;
@@ -205,6 +210,8 @@ __memrchr (s, c_in, n)
 
   return 0;
 }
-#ifdef weak_alias
+#ifndef MEMRCHR
+# ifdef weak_alias
 weak_alias (__memrchr, memrchr)
+# endif
 #endif
diff --git a/sysdeps/i386/i686/multiarch/Makefile b/sysdeps/i386/i686/multiarch/Makefile
index c89ae92472..8a4c2197b0 100644
--- a/sysdeps/i386/i686/multiarch/Makefile
+++ b/sysdeps/i386/i686/multiarch/Makefile
@@ -15,7 +15,9 @@ sysdep_routines += bzero-sse2 memset-sse2 memcpy-ssse3 mempcpy-ssse3 \
 		   strncpy-sse2 stpcpy-sse2 stpncpy-sse2 strcat-ssse3 \
 		   strcat-sse2 strncat-ssse3 strncat-sse2 strncat-c \
 		   strchr-sse2 strrchr-sse2 strchr-sse2-bsf strrchr-sse2-bsf \
-		   wcscmp-sse2 wcscmp-c
+		   wcscmp-sse2 wcscmp-c memchr-sse2 memchr-sse2-bsf \
+		   memrchr-sse2 memrchr-sse2-bsf memrchr-c \
+		   rawmemchr-sse2 rawmemchr-sse2-bsf
 ifeq (yes,$(config-cflags-sse4))
 sysdep_routines += strcspn-c strpbrk-c strspn-c strstr-c strcasestr-c
 CFLAGS-varshift.c += -msse4
diff --git a/sysdeps/i386/i686/multiarch/memchr-sse2-bsf.S b/sysdeps/i386/i686/multiarch/memchr-sse2-bsf.S
new file mode 100644
index 0000000000..115a2192a8
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/memchr-sse2-bsf.S
@@ -0,0 +1,497 @@
+/* Optimized memchr with sse2
+   Copyright (C) 2011 Free Software Foundation, Inc.
+   Contributed by Intel Corporation.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library 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.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, write to the Free
+   Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+   02111-1307 USA.  */
+
+#ifndef  NOT_IN_libc
+
+# include <sysdep.h>
+
+# define CFI_PUSH(REG)	\
+	cfi_adjust_cfa_offset (4);	\
+	cfi_rel_offset (REG, 0)
+
+# define CFI_POP(REG)	\
+	cfi_adjust_cfa_offset (-4);	\
+	cfi_restore (REG)
+
+# define PUSH(REG) pushl REG; CFI_PUSH (REG)
+# define POP(REG) popl REG; CFI_POP (REG)
+
+# define PARMS  4
+# define STR1  PARMS
+# define STR2  STR1+4
+
+# ifndef USE_AS_RAWMEMCHR
+#  define LEN   STR2+4
+#  define RETURN  POP(%edi); ret; CFI_PUSH(%edi);
+# endif
+
+# ifndef MEMCHR
+#  define MEMCHR __memchr_sse2_bsf
+# endif
+
+	.text
+ENTRY (MEMCHR)
+
+	mov	STR1(%esp), %ecx
+	movd	STR2(%esp), %xmm1
+
+# ifndef USE_AS_RAWMEMCHR
+	mov	LEN(%esp), %edx
+	test	%edx, %edx
+	jz	L(return_null_1)
+# endif
+	mov	%ecx, %eax
+
+	punpcklbw %xmm1, %xmm1
+	punpcklbw %xmm1, %xmm1
+
+	and	$63, %ecx
+	pshufd	$0, %xmm1, %xmm1
+
+	cmp	$48, %ecx
+	ja	L(crosscache)
+
+	movdqu	(%eax), %xmm0
+	pcmpeqb	%xmm1, %xmm0
+/* Check if there is a match.  */
+	pmovmskb %xmm0, %ecx
+	test	%ecx, %ecx
+	je	L(unaligned_no_match_1)
+/* Check which byte is a match.  */
+	bsf	%ecx, %ecx
+
+# ifndef USE_AS_RAWMEMCHR
+	sub	%ecx, %edx
+	jbe	L(return_null_1)
+# endif
+	add	%ecx, %eax
+	ret
+
+	.p2align 4
+L(unaligned_no_match_1):
+# ifndef USE_AS_RAWMEMCHR
+	sub	$16, %edx
+	jbe	L(return_null_1)
+	PUSH	(%edi)
+	lea	16(%eax), %edi
+	and	$15, %eax
+	and	$-16, %edi
+	add	%eax, %edx
+# else
+	lea	16(%eax), %edx
+	and	$-16, %edx
+# endif
+	jmp	L(loop_prolog)
+
+	.p2align 4
+L(return_null_1):
+	xor	%eax, %eax
+	ret
+
+# ifndef USE_AS_RAWMEMCHR
+	CFI_POP	(%edi)
+# endif
+
+	.p2align 4
+L(crosscache):
+/* Handle unaligned string.  */
+
+# ifndef USE_AS_RAWMEMCHR
+	PUSH	(%edi)
+	mov	%eax, %edi
+	and	$15, %ecx
+	and	$-16, %edi
+	movdqa	(%edi), %xmm0
+# else
+	mov	%eax, %edx
+	and	$15, %ecx
+	and	$-16, %edx
+	movdqa	(%edx), %xmm0
+# endif
+	pcmpeqb	%xmm1, %xmm0
+/* Check if there is a match.  */
+	pmovmskb %xmm0, %eax
+/* Remove the leading bytes.  */
+	sar	%cl, %eax
+	test	%eax, %eax
+	je	L(unaligned_no_match)
+/* Check which byte is a match.  */
+	bsf	%eax, %eax
+
+# ifndef USE_AS_RAWMEMCHR
+	sub	%eax, %edx
+	jbe	L(return_null)
+	add	%edi, %eax
+	add	%ecx, %eax
+	RETURN
+# else
+	add	%edx, %eax
+	add	%ecx, %eax
+	ret
+# endif
+
+	.p2align 4
+L(unaligned_no_match):
+# ifndef USE_AS_RAWMEMCHR
+	sub	$16, %edx
+	add	%ecx, %edx
+	jle	L(return_null)
+	add	$16, %edi
+# else
+	add	$16, %edx
+# endif
+
+	.p2align 4
+/* Loop start on aligned string.  */
+L(loop_prolog):
+# ifndef USE_AS_RAWMEMCHR
+	sub	$64, %edx
+	jbe	L(exit_loop)
+	movdqa	(%edi), %xmm0
+# else
+	movdqa	(%edx), %xmm0
+# endif
+	pcmpeqb	%xmm1, %xmm0
+	pmovmskb %xmm0, %eax
+	test	%eax, %eax
+	jnz	L(matches)
+
+# ifndef USE_AS_RAWMEMCHR
+	movdqa	16(%edi), %xmm2
+# else
+	movdqa	16(%edx), %xmm2
+# endif
+	pcmpeqb	%xmm1, %xmm2
+	pmovmskb %xmm2, %eax
+	test	%eax, %eax
+	jnz	L(matches16)
+
+# ifndef USE_AS_RAWMEMCHR
+	movdqa	32(%edi), %xmm3
+# else
+	movdqa	32(%edx), %xmm3
+# endif
+	pcmpeqb	%xmm1, %xmm3
+	pmovmskb %xmm3, %eax
+	test	%eax, %eax
+	jnz	L(matches32)
+
+# ifndef USE_AS_RAWMEMCHR
+	movdqa	48(%edi), %xmm4
+# else
+	movdqa	48(%edx), %xmm4
+# endif
+	pcmpeqb	%xmm1, %xmm4
+
+# ifndef USE_AS_RAWMEMCHR
+	add	$64, %edi
+# else
+	add	$64, %edx
+# endif
+	pmovmskb %xmm4, %eax
+	test	%eax, %eax
+	jnz	L(matches0)
+
+# ifndef USE_AS_RAWMEMCHR
+	test	$0x3f, %edi
+# else
+	test	$0x3f, %edx
+# endif
+	jz	L(align64_loop)
+
+# ifndef USE_AS_RAWMEMCHR
+	sub	$64, %edx
+	jbe	L(exit_loop)
+	movdqa	(%edi), %xmm0
+# else
+	movdqa	(%edx), %xmm0
+# endif
+	pcmpeqb	%xmm1, %xmm0
+	pmovmskb %xmm0, %eax
+	test	%eax, %eax
+	jnz	L(matches)
+
+# ifndef USE_AS_RAWMEMCHR
+	movdqa	16(%edi), %xmm2
+# else
+	movdqa	16(%edx), %xmm2
+# endif
+	pcmpeqb	%xmm1, %xmm2
+	pmovmskb %xmm2, %eax
+	test	%eax, %eax
+	jnz	L(matches16)
+
+# ifndef USE_AS_RAWMEMCHR
+	movdqa	32(%edi), %xmm3
+# else
+	movdqa	32(%edx), %xmm3
+# endif
+	pcmpeqb	%xmm1, %xmm3
+	pmovmskb %xmm3, %eax
+	test	%eax, %eax
+	jnz	L(matches32)
+
+# ifndef USE_AS_RAWMEMCHR
+	movdqa	48(%edi), %xmm3
+# else
+	movdqa	48(%edx), %xmm3
+# endif
+	pcmpeqb	%xmm1, %xmm3
+	pmovmskb %xmm3, %eax
+
+# ifndef USE_AS_RAWMEMCHR
+	add	$64, %edi
+# else
+	add	$64, %edx
+# endif
+	test	%eax, %eax
+	jnz	L(matches0)
+
+# ifndef USE_AS_RAWMEMCHR
+	mov	%edi, %ecx
+	and	$-64, %edi
+	and	$63, %ecx
+	add	%ecx, %edx
+# else
+	and	$-64, %edx
+# endif
+
+	.p2align 4
+L(align64_loop):
+# ifndef USE_AS_RAWMEMCHR
+	sub	$64, %edx
+	jbe	L(exit_loop)
+	movdqa	(%edi), %xmm0
+	movdqa	16(%edi), %xmm2
+	movdqa	32(%edi), %xmm3
+	movdqa	48(%edi), %xmm4
+# else
+	movdqa	(%edx), %xmm0
+	movdqa	16(%edx), %xmm2
+	movdqa	32(%edx), %xmm3
+	movdqa	48(%edx), %xmm4
+# endif
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm1, %xmm2
+	pcmpeqb	%xmm1, %xmm3
+	pcmpeqb	%xmm1, %xmm4
+
+	pmaxub	%xmm0, %xmm3
+	pmaxub	%xmm2, %xmm4
+	pmaxub	%xmm3, %xmm4
+	pmovmskb %xmm4, %eax
+
+# ifndef USE_AS_RAWMEMCHR
+	add	$64, %edi
+# else
+	add	$64, %edx
+# endif
+
+	test	%eax, %eax
+	jz	L(align64_loop)
+
+# ifndef USE_AS_RAWMEMCHR
+	sub	$64, %edi
+# else
+	sub	$64, %edx
+# endif
+
+	pmovmskb %xmm0, %eax
+	test	%eax, %eax
+	jnz	L(matches)
+
+	pmovmskb %xmm2, %eax
+	test	%eax, %eax
+	jnz	L(matches16)
+
+# ifndef USE_AS_RAWMEMCHR
+	movdqa	32(%edi), %xmm3
+# else
+	movdqa	32(%edx), %xmm3
+# endif
+
+	pcmpeqb	%xmm1, %xmm3
+
+# ifndef USE_AS_RAWMEMCHR
+	pcmpeqb	48(%edi), %xmm1
+# else
+	pcmpeqb	48(%edx), %xmm1
+# endif
+	pmovmskb %xmm3, %eax
+	test	%eax, %eax
+	jnz	L(matches32)
+
+	pmovmskb %xmm1, %eax
+	bsf	%eax, %eax
+
+# ifndef USE_AS_RAWMEMCHR
+	lea	48(%edi, %eax), %eax
+	RETURN
+# else
+	lea	48(%edx, %eax), %eax
+	ret
+# endif
+
+# ifndef USE_AS_RAWMEMCHR
+	.p2align 4
+L(exit_loop):
+	add	$64, %edx
+	cmp	$32, %edx
+	jbe	L(exit_loop_32)
+
+	movdqa	(%edi), %xmm0
+	pcmpeqb	%xmm1, %xmm0
+	pmovmskb %xmm0, %eax
+	test	%eax, %eax
+	jnz	L(matches)
+
+	movdqa	16(%edi), %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	pmovmskb %xmm2, %eax
+	test	%eax, %eax
+	jnz	L(matches16)
+
+	movdqa	32(%edi), %xmm3
+	pcmpeqb	%xmm1, %xmm3
+	pmovmskb %xmm3, %eax
+	test	%eax, %eax
+	jnz	L(matches32_1)
+	cmp	$48, %edx
+	jbe	L(return_null)
+
+	pcmpeqb	48(%edi), %xmm1
+	pmovmskb %xmm1, %eax
+	test	%eax, %eax
+	jnz	L(matches48_1)
+	xor	%eax, %eax
+	RETURN
+
+	.p2align 4
+L(exit_loop_32):
+	movdqa	(%edi), %xmm0
+	pcmpeqb	%xmm1, %xmm0
+	pmovmskb %xmm0, %eax
+	test	%eax, %eax
+	jnz	L(matches_1)
+	cmp	$16, %edx
+	jbe	L(return_null)
+
+	pcmpeqb	16(%edi), %xmm1
+	pmovmskb %xmm1, %eax
+	test	%eax, %eax
+	jnz	L(matches16_1)
+	xor	%eax, %eax
+	RETURN
+# endif
+	.p2align 4
+L(matches0):
+	bsf	%eax, %eax
+# ifndef USE_AS_RAWMEMCHR
+	lea	-16(%eax, %edi), %eax
+	RETURN
+# else
+	lea	-16(%eax, %edx), %eax
+	ret
+# endif
+
+	.p2align 4
+L(matches):
+	bsf	%eax, %eax
+# ifndef USE_AS_RAWMEMCHR
+	add	%edi, %eax
+	RETURN
+# else
+	add	%edx, %eax
+	ret
+# endif
+
+	.p2align 4
+L(matches16):
+	bsf	%eax, %eax
+# ifndef USE_AS_RAWMEMCHR
+	lea	16(%eax, %edi), %eax
+	RETURN
+# else
+	lea	16(%eax, %edx), %eax
+	ret
+# endif
+
+	.p2align 4
+L(matches32):
+	bsf	%eax, %eax
+# ifndef USE_AS_RAWMEMCHR
+	lea	32(%eax, %edi), %eax
+	RETURN
+# else
+	lea	32(%eax, %edx), %eax
+	ret
+# endif
+
+# ifndef USE_AS_RAWMEMCHR
+	.p2align 4
+L(matches_1):
+	bsf	%eax, %eax
+	sub	%eax, %edx
+	jbe	L(return_null)
+
+	add	%edi, %eax
+	RETURN
+
+	.p2align 4
+L(matches16_1):
+	sub	$16, %edx
+	bsf	%eax, %eax
+	sub	%eax, %edx
+	jbe	L(return_null)
+
+	lea	16(%edi, %eax), %eax
+	RETURN
+
+	.p2align 4
+L(matches32_1):
+	sub	$32, %edx
+	bsf	%eax, %eax
+	sub	%eax, %edx
+	jbe	L(return_null)
+
+	lea	32(%edi, %eax), %eax
+	RETURN
+
+	.p2align 4
+L(matches48_1):
+	sub	$48, %edx
+	bsf	%eax, %eax
+	sub	%eax, %edx
+	jbe	L(return_null)
+
+	lea	48(%edi, %eax), %eax
+	RETURN
+# endif
+	.p2align 4
+L(return_null):
+	xor	%eax, %eax
+# ifndef USE_AS_RAWMEMCHR
+	RETURN
+# else
+	ret
+# endif
+
+END (MEMCHR)
+#endif
diff --git a/sysdeps/i386/i686/multiarch/memchr-sse2.S b/sysdeps/i386/i686/multiarch/memchr-sse2.S
new file mode 100644
index 0000000000..63d1d5d7bf
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/memchr-sse2.S
@@ -0,0 +1,706 @@
+/* Optimized memchr with sse2 without bsf
+   Copyright (C) 2011 Free Software Foundation, Inc.
+   Contributed by Intel Corporation.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library 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.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, write to the Free
+   Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+   02111-1307 USA.  */
+
+#ifndef  NOT_IN_libc
+
+# include <sysdep.h>
+
+# define CFI_PUSH(REG)	\
+	cfi_adjust_cfa_offset (4);	\
+	cfi_rel_offset (REG, 0)
+
+# define CFI_POP(REG)	\
+	cfi_adjust_cfa_offset (-4);	\
+	cfi_restore (REG)
+
+# define PUSH(REG) pushl REG; CFI_PUSH (REG)
+# define POP(REG) popl REG; CFI_POP (REG)
+
+# ifndef USE_AS_RAWMEMCHR
+#  define ENTRANCE PUSH(%edi);
+#  define PARMS  8
+#  define RETURN  POP(%edi); ret; CFI_PUSH(%edi);
+# else
+#  define ENTRANCE
+#  define PARMS  4
+# endif
+
+# define STR1  PARMS
+# define STR2  STR1+4
+
+# ifndef USE_AS_RAWMEMCHR
+#  define LEN   STR2+4
+# endif
+
+# ifndef MEMCHR
+#  define MEMCHR __memchr_sse2
+# endif
+
+	atom_text_section
+ENTRY (MEMCHR)
+	ENTRANCE
+	mov	STR1(%esp), %ecx
+	movd	STR2(%esp), %xmm1
+# ifndef USE_AS_RAWMEMCHR
+	mov	LEN(%esp), %edx
+	test	%edx, %edx
+	jz	L(return_null)
+# endif
+
+	punpcklbw %xmm1, %xmm1
+# ifndef USE_AS_RAWMEMCHR
+	mov	%ecx, %edi
+# else
+	mov	%ecx, %edx
+# endif
+	punpcklbw %xmm1, %xmm1
+
+	and	$63, %ecx
+	pshufd	$0, %xmm1, %xmm1
+	cmp	$48, %ecx
+	ja	L(crosscache)
+
+# ifndef USE_AS_RAWMEMCHR
+	movdqu	(%edi), %xmm0
+# else
+	movdqu	(%edx), %xmm0
+# endif
+	pcmpeqb	%xmm1, %xmm0
+	pmovmskb %xmm0, %eax
+	test	%eax, %eax
+# ifndef USE_AS_RAWMEMCHR
+	jnz	L(match_case2_prolog)
+
+	sub	$16, %edx
+	jbe	L(return_null)
+	lea	16(%edi), %edi
+	and	$15, %ecx
+	and	$-16, %edi
+	add	%ecx, %edx
+# else
+	jnz	L(match_case1_prolog)
+	lea	16(%edx), %edx
+	and	$-16, %edx
+# endif
+	jmp	L(loop_prolog)
+
+	.p2align 4
+L(crosscache):
+	and	$15, %ecx
+# ifndef USE_AS_RAWMEMCHR
+	and	$-16, %edi
+	movdqa	(%edi), %xmm0
+# else
+	and	$-16, %edx
+	movdqa	(%edx), %xmm0
+# endif
+	pcmpeqb	%xmm1, %xmm0
+	pmovmskb %xmm0, %eax
+	sar	%cl, %eax
+	test	%eax, %eax
+
+# ifndef USE_AS_RAWMEMCHR
+	jnz	L(match_case2_prolog1)
+	lea	-16(%edx), %edx
+	add	%ecx, %edx
+	jle	L(return_null)
+	lea	16(%edi), %edi
+# else
+	jnz	L(match_case1_prolog1)
+	lea	16(%edx), %edx
+# endif
+
+	.p2align 4
+L(loop_prolog):
+# ifndef USE_AS_RAWMEMCHR
+	sub	$64, %edx
+	jbe	L(exit_loop)
+	movdqa	(%edi), %xmm0
+# else
+	movdqa	(%edx), %xmm0
+# endif
+	pcmpeqb	%xmm1, %xmm0
+	xor	%ecx, %ecx
+	pmovmskb %xmm0, %eax
+	test	%eax, %eax
+	jnz	L(match_case1)
+
+# ifndef USE_AS_RAWMEMCHR
+	movdqa	16(%edi), %xmm2
+# else
+	movdqa	16(%edx), %xmm2
+# endif
+	pcmpeqb	%xmm1, %xmm2
+	lea	16(%ecx), %ecx
+	pmovmskb %xmm2, %eax
+	test	%eax, %eax
+	jnz	L(match_case1)
+
+# ifndef USE_AS_RAWMEMCHR
+	movdqa	32(%edi), %xmm3
+# else
+	movdqa	32(%edx), %xmm3
+# endif
+	pcmpeqb	%xmm1, %xmm3
+	lea	16(%ecx), %ecx
+	pmovmskb %xmm3, %eax
+	test	%eax, %eax
+	jnz	L(match_case1)
+
+# ifndef USE_AS_RAWMEMCHR
+	movdqa	48(%edi), %xmm4
+# else
+	movdqa	48(%edx), %xmm4
+# endif
+	pcmpeqb	%xmm1, %xmm4
+	lea	16(%ecx), %ecx
+	pmovmskb %xmm4, %eax
+	test	%eax, %eax
+	jnz	L(match_case1)
+
+# ifndef USE_AS_RAWMEMCHR
+	lea	64(%edi), %edi
+	sub	$64, %edx
+	jbe	L(exit_loop)
+
+	movdqa	(%edi), %xmm0
+# else
+	lea	64(%edx), %edx
+	movdqa	(%edx), %xmm0
+# endif
+	pcmpeqb	%xmm1, %xmm0
+	xor	%ecx, %ecx
+	pmovmskb %xmm0, %eax
+	test	%eax, %eax
+	jnz	L(match_case1)
+
+# ifndef USE_AS_RAWMEMCHR
+	movdqa	16(%edi), %xmm2
+# else
+	movdqa	16(%edx), %xmm2
+# endif
+	pcmpeqb	%xmm1, %xmm2
+	lea	16(%ecx), %ecx
+	pmovmskb %xmm2, %eax
+	test	%eax, %eax
+	jnz	L(match_case1)
+
+# ifndef USE_AS_RAWMEMCHR
+	movdqa	32(%edi), %xmm3
+# else
+	movdqa	32(%edx), %xmm3
+# endif
+	pcmpeqb	%xmm1, %xmm3
+	lea	16(%ecx), %ecx
+	pmovmskb %xmm3, %eax
+	test	%eax, %eax
+	jnz	L(match_case1)
+
+# ifndef USE_AS_RAWMEMCHR
+	movdqa	48(%edi), %xmm4
+# else
+	movdqa	48(%edx), %xmm4
+# endif
+	pcmpeqb	%xmm1, %xmm4
+	lea	16(%ecx), %ecx
+	pmovmskb %xmm4, %eax
+	test	%eax, %eax
+	jnz	L(match_case1)
+
+# ifndef USE_AS_RAWMEMCHR
+	lea	64(%edi), %edi
+	mov	%edi, %ecx
+	and	$-64, %edi
+	and	$63, %ecx
+	add	%ecx, %edx
+# else
+	lea	64(%edx), %edx
+	and	$-64, %edx
+# endif
+
+	.p2align 4
+L(align64_loop):
+
+# ifndef USE_AS_RAWMEMCHR
+	sub	$64, %edx
+	jbe	L(exit_loop)
+	movdqa	(%edi), %xmm0
+	movdqa	16(%edi), %xmm2
+	movdqa	32(%edi), %xmm3
+	movdqa	48(%edi), %xmm4
+# else
+	movdqa	(%edx), %xmm0
+	movdqa	16(%edx), %xmm2
+	movdqa	32(%edx), %xmm3
+	movdqa	48(%edx), %xmm4
+# endif
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm1, %xmm2
+	pcmpeqb	%xmm1, %xmm3
+	pcmpeqb	%xmm1, %xmm4
+
+	pmaxub	%xmm0, %xmm3
+	pmaxub	%xmm2, %xmm4
+	pmaxub	%xmm3, %xmm4
+# ifndef USE_AS_RAWMEMCHR
+	add	$64, %edi
+# else
+	add	$64, %edx
+# endif
+	pmovmskb %xmm4, %eax
+
+	test	%eax, %eax
+	jz	L(align64_loop)
+
+# ifndef USE_AS_RAWMEMCHR
+	sub	$64, %edi
+# else
+	sub	$64, %edx
+# endif
+
+	pmovmskb %xmm0, %eax
+	xor	%ecx, %ecx
+	test	%eax, %eax
+	jnz	L(match_case1)
+
+	pmovmskb %xmm2, %eax
+	lea	16(%ecx), %ecx
+	test	%eax, %eax
+	jnz	L(match_case1)
+
+# ifndef USE_AS_RAWMEMCHR
+	movdqa	32(%edi), %xmm3
+# else
+	movdqa	32(%edx), %xmm3
+# endif
+	pcmpeqb	%xmm1, %xmm3
+	pmovmskb %xmm3, %eax
+	lea	16(%ecx), %ecx
+	test	%eax, %eax
+	jnz	L(match_case1)
+
+# ifndef USE_AS_RAWMEMCHR
+	pcmpeqb	48(%edi), %xmm1
+# else
+	pcmpeqb	48(%edx), %xmm1
+# endif
+	pmovmskb %xmm1, %eax
+	lea	16(%ecx), %ecx
+
+	.p2align 4
+L(match_case1):
+# ifndef USE_AS_RAWMEMCHR
+	add	%ecx, %edi
+# else
+L(match_case1_prolog1):
+	add	%ecx, %edx
+L(match_case1_prolog):
+# endif
+	test	%al, %al
+	jz	L(match_case1_high)
+	mov	%al, %cl
+	and	$15, %cl
+	jz	L(match_case1_8)
+	test	$0x01, %al
+	jnz	L(ExitCase1_1)
+	test	$0x02, %al
+	jnz	L(ExitCase1_2)
+	test	$0x04, %al
+	jnz	L(ExitCase1_3)
+# ifndef USE_AS_RAWMEMCHR
+	lea	3(%edi), %eax
+	RETURN
+# else
+	lea	3(%edx), %eax
+	ret
+# endif
+
+	.p2align 4
+L(match_case1_8):
+	test	$0x10, %al
+	jnz	L(ExitCase1_5)
+	test	$0x20, %al
+	jnz	L(ExitCase1_6)
+	test	$0x40, %al
+	jnz	L(ExitCase1_7)
+# ifndef USE_AS_RAWMEMCHR
+	lea	7(%edi), %eax
+	RETURN
+# else
+	lea	7(%edx), %eax
+	ret
+# endif
+
+	.p2align 4
+L(match_case1_high):
+	mov	%ah, %ch
+	and	$15, %ch
+	jz	L(match_case1_high_8)
+	test	$0x01, %ah
+	jnz	L(ExitCase1_9)
+	test	$0x02, %ah
+	jnz	L(ExitCase1_10)
+	test	$0x04, %ah
+	jnz	L(ExitCase1_11)
+# ifndef USE_AS_RAWMEMCHR
+	lea	11(%edi), %eax
+	RETURN
+# else
+	lea	11(%edx), %eax
+	ret
+# endif
+
+	.p2align 4
+L(match_case1_high_8):
+	test	$0x10, %ah
+	jnz	L(ExitCase1_13)
+	test	$0x20, %ah
+	jnz	L(ExitCase1_14)
+	test	$0x40, %ah
+	jnz	L(ExitCase1_15)
+# ifndef USE_AS_RAWMEMCHR
+	lea	15(%edi), %eax
+	RETURN
+# else
+	lea	15(%edx), %eax
+	ret
+# endif
+
+# ifndef USE_AS_RAWMEMCHR
+	.p2align 4
+L(exit_loop):
+	add	$64, %edx
+
+	movdqa	(%edi), %xmm0
+	pcmpeqb	%xmm1, %xmm0
+	xor	%ecx, %ecx
+	pmovmskb %xmm0, %eax
+	test	%eax, %eax
+	jnz	L(match_case2)
+	cmp	$16, %edx
+	jbe	L(return_null)
+
+	movdqa	16(%edi), %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	lea	16(%ecx), %ecx
+	pmovmskb %xmm2, %eax
+	test	%eax, %eax
+	jnz	L(match_case2)
+	cmp	$32, %edx
+	jbe	L(return_null)
+
+	movdqa	32(%edi), %xmm3
+	pcmpeqb	%xmm1, %xmm3
+	lea	16(%ecx), %ecx
+	pmovmskb %xmm3, %eax
+	test	%eax, %eax
+	jnz	L(match_case2)
+	cmp	$48, %edx
+	jbe	L(return_null)
+
+	pcmpeqb	48(%edi), %xmm1
+	lea	16(%ecx), %ecx
+	pmovmskb %xmm1, %eax
+	test	%eax, %eax
+	jnz	L(match_case2)
+
+	xor	%eax, %eax
+	RETURN
+# endif
+
+	.p2align 4
+L(ExitCase1_1):
+# ifndef USE_AS_RAWMEMCHR
+	mov	%edi, %eax
+	RETURN
+# else
+	mov	%edx, %eax
+	ret
+# endif
+
+	.p2align 4
+L(ExitCase1_2):
+# ifndef USE_AS_RAWMEMCHR
+	lea	1(%edi), %eax
+	RETURN
+# else
+	lea	1(%edx), %eax
+	ret
+# endif
+
+	.p2align 4
+L(ExitCase1_3):
+# ifndef USE_AS_RAWMEMCHR
+	lea	2(%edi), %eax
+	RETURN
+# else
+	lea	2(%edx), %eax
+	ret
+# endif
+
+	.p2align 4
+L(ExitCase1_5):
+# ifndef USE_AS_RAWMEMCHR
+	lea	4(%edi), %eax
+	RETURN
+# else
+	lea	4(%edx), %eax
+	ret
+# endif
+
+	.p2align 4
+L(ExitCase1_6):
+# ifndef USE_AS_RAWMEMCHR
+	lea	5(%edi), %eax
+	RETURN
+# else
+	lea	5(%edx), %eax
+	ret
+# endif
+
+	.p2align 4
+L(ExitCase1_7):
+# ifndef USE_AS_RAWMEMCHR
+	lea	6(%edi), %eax
+	RETURN
+# else
+	lea	6(%edx), %eax
+	ret
+# endif
+
+	.p2align 4
+L(ExitCase1_9):
+# ifndef USE_AS_RAWMEMCHR
+	lea	8(%edi), %eax
+	RETURN
+# else
+	lea	8(%edx), %eax
+	ret
+# endif
+
+	.p2align 4
+L(ExitCase1_10):
+# ifndef USE_AS_RAWMEMCHR
+	lea	9(%edi), %eax
+	RETURN
+# else
+	lea	9(%edx), %eax
+	ret
+# endif
+
+	.p2align 4
+L(ExitCase1_11):
+# ifndef USE_AS_RAWMEMCHR
+	lea	10(%edi), %eax
+	RETURN
+# else
+	lea	10(%edx), %eax
+	ret
+# endif
+
+	.p2align 4
+L(ExitCase1_13):
+# ifndef USE_AS_RAWMEMCHR
+	lea	12(%edi), %eax
+	RETURN
+# else
+	lea	12(%edx), %eax
+	ret
+# endif
+
+	.p2align 4
+L(ExitCase1_14):
+# ifndef USE_AS_RAWMEMCHR
+	lea	13(%edi), %eax
+	RETURN
+# else
+	lea	13(%edx), %eax
+	ret
+# endif
+
+	.p2align 4
+L(ExitCase1_15):
+# ifndef USE_AS_RAWMEMCHR
+	lea	14(%edi), %eax
+	RETURN
+# else
+	lea	14(%edx), %eax
+	ret
+# endif
+
+# ifndef USE_AS_RAWMEMCHR
+	.p2align 4
+L(match_case2):
+	sub	%ecx, %edx
+L(match_case2_prolog1):
+	add	%ecx, %edi
+L(match_case2_prolog):
+	test	%al, %al
+	jz	L(match_case2_high)
+	mov	%al, %cl
+	and	$15, %cl
+	jz	L(match_case2_8)
+	test	$0x01, %al
+	jnz	L(ExitCase2_1)
+	test	$0x02, %al
+	jnz	L(ExitCase2_2)
+	test	$0x04, %al
+	jnz	L(ExitCase2_3)
+	sub	$4, %edx
+	jb	L(return_null)
+	lea	3(%edi), %eax
+	RETURN
+
+	.p2align 4
+L(match_case2_8):
+	test	$0x10, %al
+	jnz	L(ExitCase2_5)
+	test	$0x20, %al
+	jnz	L(ExitCase2_6)
+	test	$0x40, %al
+	jnz	L(ExitCase2_7)
+	sub	$8, %edx
+	jb	L(return_null)
+	lea	7(%edi), %eax
+	RETURN
+
+	.p2align 4
+L(match_case2_high):
+	mov	%ah, %ch
+	and	$15, %ch
+	jz	L(match_case2_high_8)
+	test	$0x01, %ah
+	jnz	L(ExitCase2_9)
+	test	$0x02, %ah
+	jnz	L(ExitCase2_10)
+	test	$0x04, %ah
+	jnz	L(ExitCase2_11)
+	sub	$12, %edx
+	jb	L(return_null)
+	lea	11(%edi), %eax
+	RETURN
+
+	.p2align 4
+L(match_case2_high_8):
+	test	$0x10, %ah
+	jnz	L(ExitCase2_13)
+	test	$0x20, %ah
+	jnz	L(ExitCase2_14)
+	test	$0x40, %ah
+	jnz	L(ExitCase2_15)
+	sub	$16, %edx
+	jb	L(return_null)
+	lea	15(%edi), %eax
+	RETURN
+
+	.p2align 4
+L(ExitCase2_1):
+	mov	%edi, %eax
+	RETURN
+
+	.p2align 4
+L(ExitCase2_2):
+	sub	$2, %edx
+	jb	L(return_null)
+	lea	1(%edi), %eax
+	RETURN
+
+	.p2align 4
+L(ExitCase2_3):
+	sub	$3, %edx
+	jb	L(return_null)
+	lea	2(%edi), %eax
+	RETURN
+
+	.p2align 4
+L(ExitCase2_5):
+	sub	$5, %edx
+	jb	L(return_null)
+	lea	4(%edi), %eax
+	RETURN
+
+	.p2align 4
+L(ExitCase2_6):
+	sub	$6, %edx
+	jb	L(return_null)
+	lea	5(%edi), %eax
+	RETURN
+
+	.p2align 4
+L(ExitCase2_7):
+	sub	$7, %edx
+	jb	L(return_null)
+	lea	6(%edi), %eax
+	RETURN
+
+	.p2align 4
+L(ExitCase2_9):
+	sub	$9, %edx
+	jb	L(return_null)
+	lea	8(%edi), %eax
+	RETURN
+
+	.p2align 4
+L(ExitCase2_10):
+	sub	$10, %edx
+	jb	L(return_null)
+	lea	9(%edi), %eax
+	RETURN
+
+	.p2align 4
+L(ExitCase2_11):
+	sub	$11, %edx
+	jb	L(return_null)
+	lea	10(%edi), %eax
+	RETURN
+
+	.p2align 4
+L(ExitCase2_13):
+	sub	$13, %edx
+	jb	L(return_null)
+	lea	12(%edi), %eax
+	RETURN
+
+	.p2align 4
+L(ExitCase2_14):
+	sub	$14, %edx
+	jb	L(return_null)
+	lea	13(%edi), %eax
+	RETURN
+
+	.p2align 4
+L(ExitCase2_15):
+	sub	$15, %edx
+	jb	L(return_null)
+	lea	14(%edi), %eax
+	RETURN
+# endif
+
+	.p2align 4
+L(return_null):
+	xor	%eax, %eax
+# ifndef USE_AS_RAWMEMCHR
+	RETURN
+# else
+	ret
+# endif
+
+END (MEMCHR)
+#endif
diff --git a/sysdeps/i386/i686/multiarch/memchr.S b/sysdeps/i386/i686/multiarch/memchr.S
new file mode 100644
index 0000000000..163a83e17b
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/memchr.S
@@ -0,0 +1,99 @@
+/* Multiple versions of memchr
+   Copyright (C) 2011 Free Software Foundation, Inc.
+   Contributed by Intel Corporation.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library 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.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, write to the Free
+   Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+   02111-1307 USA.  */
+
+#include <sysdep.h>
+#include <init-arch.h>
+
+#ifndef  NOT_IN_libc
+	.section	.gnu.linkonce.t.__i686.get_pc_thunk.bx,"ax",@progbits
+	.globl	__i686.get_pc_thunk.bx
+	.hidden	__i686.get_pc_thunk.bx
+	.p2align 4
+	.type	__i686.get_pc_thunk.bx,@function
+__i686.get_pc_thunk.bx:
+	movl	(%esp), %ebx
+	ret
+
+# define CFI_POP(REG) \
+	cfi_adjust_cfa_offset (-4); \
+	cfi_restore (REG)
+
+# define CFI_PUSH(REG) \
+	cfi_adjust_cfa_offset (4); \
+	cfi_rel_offset (REG, 0)
+
+	.text
+ENTRY(__memchr)
+	.type	__memchr, @gnu_indirect_function
+	pushl	%ebx
+	CFI_PUSH (%ebx)
+	call	__i686.get_pc_thunk.bx
+	addl	$_GLOBAL_OFFSET_TABLE_, %ebx
+	cmpl	$0, KIND_OFFSET+__cpu_features@GOTOFF(%ebx)
+	jne	1f
+	call	__init_cpu_features
+
+1:	testl	$bit_SSE2, CPUID_OFFSET+index_SSE2+__cpu_features@GOTOFF(%ebx)
+	jz	2f
+	testl	$bit_Slow_BSF, FEATURE_OFFSET+index_Slow_BSF+__cpu_features@GOTOFF(%ebx)
+	jz	3f
+
+	leal	__memchr_sse2@GOTOFF(%ebx), %eax
+	popl	%ebx
+	CFI_POP	(%ebx)
+	ret
+
+	CFI_PUSH (%ebx)
+
+2:	leal	__memchr_ia32@GOTOFF(%ebx), %eax
+	popl	%ebx
+	CFI_POP	(%ebx)
+	ret
+
+	CFI_PUSH (%ebx)
+
+3:	leal	__memchr_sse2_bsf@GOTOFF(%ebx), %eax
+	popl	%ebx
+	CFI_POP	(%ebx)
+	ret
+END(__memchr)
+
+weak_alias(__memchr, memchr)
+
+# undef ENTRY
+# define ENTRY(name) \
+	.type __memchr_ia32, @function; \
+	.globl __memchr_ia32; \
+	.p2align 4; \
+	__memchr_ia32: cfi_startproc; \
+	CALL_MCOUNT
+# undef END
+# define END(name) \
+	cfi_endproc; .size __memchr_ia32, .-__memchr_ia32
+
+# undef libc_hidden_builtin_def
+/* IFUNC doesn't work with the hidden functions in shared library since
+   they will be called without setting up EBX needed for PLT which is
+   used by IFUNC.  */
+# define libc_hidden_builtin_def(name) \
+	.globl __GI_memchr; __GI_memchr = __memchr_ia32
+
+#endif
+#include "../../memchr.S"
diff --git a/sysdeps/i386/i686/multiarch/memrchr-c.c b/sysdeps/i386/i686/multiarch/memrchr-c.c
new file mode 100644
index 0000000000..44ec1a6ed9
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/memrchr-c.c
@@ -0,0 +1,7 @@
+#ifndef NOT_IN_libc
+# define MEMRCHR  __memrchr_ia32
+# include <string.h>
+extern void *__memrchr_ia32 (const void *, int, size_t);
+#endif
+
+#include "string/memrchr.c"
diff --git a/sysdeps/i386/i686/multiarch/memrchr-sse2-bsf.S b/sysdeps/i386/i686/multiarch/memrchr-sse2-bsf.S
new file mode 100644
index 0000000000..1c2a867ca2
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/memrchr-sse2-bsf.S
@@ -0,0 +1,418 @@
+/* Optimized memrchr with sse2
+   Copyright (C) 2011 Free Software Foundation, Inc.
+   Contributed by Intel Corporation.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library 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.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, write to the Free
+   Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+   02111-1307 USA.  */
+
+#ifndef  NOT_IN_libc
+
+# include <sysdep.h>
+
+# define CFI_PUSH(REG)	\
+	cfi_adjust_cfa_offset (4);	\
+	cfi_rel_offset (REG, 0)
+
+# define CFI_POP(REG)	\
+	cfi_adjust_cfa_offset (-4);	\
+	cfi_restore (REG)
+
+# define PUSH(REG) pushl REG; CFI_PUSH (REG)
+# define POP(REG) popl REG; CFI_POP (REG)
+
+# define PARMS  4
+# define STR1  PARMS
+# define STR2  STR1+4
+# define LEN   STR2+4
+
+# define MEMCHR __memrchr_sse2_bsf
+
+	.text
+ENTRY (MEMCHR)
+	mov	STR1(%esp), %ecx
+	movd	STR2(%esp), %xmm1
+	mov	LEN(%esp), %edx
+
+	sub	$16, %edx
+	jbe	L(length_less16)
+
+	punpcklbw %xmm1, %xmm1
+	add	%edx, %ecx
+	punpcklbw %xmm1, %xmm1
+
+	movdqu	(%ecx), %xmm0
+	pshufd	$0, %xmm1, %xmm1
+	pcmpeqb	%xmm1, %xmm0
+
+/* Check if there is a match.  */
+	pmovmskb %xmm0, %eax
+	test	%eax, %eax
+	jnz	L(matches0)
+
+	sub	$64, %ecx
+	mov	%ecx, %eax
+	and	$15, %eax
+	jz	L(loop_prolog)
+
+	add	$16, %ecx
+	add	$16, %edx
+	sub	%eax, %ecx
+	sub	%eax, %edx
+
+	.p2align 4
+/* Loop start on aligned string.  */
+L(loop_prolog):
+	sub	$64, %edx
+	jbe	L(exit_loop)
+
+	movdqa	48(%ecx), %xmm0
+	pcmpeqb	%xmm1, %xmm0
+	pmovmskb %xmm0, %eax
+	test	%eax, %eax
+	jnz	L(matches48)
+
+	movdqa	32(%ecx), %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	pmovmskb %xmm2, %eax
+	test	%eax, %eax
+	jnz	L(matches32)
+
+	movdqa	16(%ecx), %xmm3
+	pcmpeqb	%xmm1, %xmm3
+	pmovmskb %xmm3, %eax
+	test	%eax, %eax
+	jnz	L(matches16)
+
+	movdqa	(%ecx), %xmm4
+	pcmpeqb	%xmm1, %xmm4
+	pmovmskb %xmm4, %eax
+	test	%eax, %eax
+	jnz	L(matches0)
+
+	sub	$64, %ecx
+	sub	$64, %edx
+	jbe	L(exit_loop)
+
+	movdqa	48(%ecx), %xmm0
+	pcmpeqb	%xmm1, %xmm0
+	pmovmskb %xmm0, %eax
+	test	%eax, %eax
+	jnz	L(matches48)
+
+	movdqa	32(%ecx), %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	pmovmskb %xmm2, %eax
+	test	%eax, %eax
+	jnz	L(matches32)
+
+	movdqa	16(%ecx), %xmm3
+	pcmpeqb	%xmm1, %xmm3
+	pmovmskb %xmm3, %eax
+	test	%eax, %eax
+	jnz	L(matches16)
+
+	movdqa	(%ecx), %xmm3
+	pcmpeqb	%xmm1, %xmm3
+	pmovmskb %xmm3, %eax
+	test	%eax, %eax
+	jnz	L(matches0)
+
+	mov	%ecx, %eax
+	and	$63, %eax
+	test	%eax, %eax
+	jz	L(align64_loop)    
+
+	add	$64, %ecx
+	add	$64, %edx
+	sub	%eax, %ecx
+	sub	%eax, %edx
+
+	.p2align 4
+L(align64_loop):
+	sub	$64, %ecx
+	sub	$64, %edx
+	jbe	L(exit_loop)
+
+	movdqa	(%ecx), %xmm0
+	movdqa	16(%ecx), %xmm2
+	movdqa	32(%ecx), %xmm3
+	movdqa	48(%ecx), %xmm4
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm1, %xmm2
+	pcmpeqb	%xmm1, %xmm3
+	pcmpeqb	%xmm1, %xmm4
+
+	pmaxub	%xmm3, %xmm0
+	pmaxub	%xmm4, %xmm2
+	pmaxub	%xmm0, %xmm2
+	pmovmskb %xmm2, %eax
+
+	test	%eax, %eax
+	jz	L(align64_loop)
+
+	pmovmskb %xmm4, %eax
+	test	%eax, %eax
+	jnz	L(matches48)
+
+	pmovmskb %xmm3, %eax
+	test	%eax, %eax
+	jnz	L(matches32)
+
+	movdqa	16(%ecx), %xmm2
+
+	pcmpeqb	%xmm1, %xmm2
+	pcmpeqb	(%ecx), %xmm1
+
+	pmovmskb %xmm2, %eax
+	test	%eax, %eax
+	jnz	L(matches16)
+
+	pmovmskb %xmm1, %eax
+	bsr	%eax, %eax
+
+	add	%ecx, %eax
+	ret
+
+	.p2align 4
+L(exit_loop):
+	add	$64, %edx
+	cmp	$32, %edx
+	jbe	L(exit_loop_32)
+
+	movdqa	48(%ecx), %xmm0
+	pcmpeqb	%xmm1, %xmm0
+	pmovmskb %xmm0, %eax
+	test	%eax, %eax
+	jnz	L(matches48)
+
+	movdqa	32(%ecx), %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	pmovmskb %xmm2, %eax
+	test	%eax, %eax
+	jnz	L(matches32)
+
+	movdqa	16(%ecx), %xmm3
+	pcmpeqb	%xmm1, %xmm3
+	pmovmskb %xmm3, %eax
+	test	%eax, %eax
+	jnz	L(matches16_1)
+	cmp	$48, %edx
+	jbe	L(return_null)
+
+	pcmpeqb	(%ecx), %xmm1
+	pmovmskb %xmm1, %eax
+	test	%eax, %eax
+	jnz	L(matches0_1)
+	xor	%eax, %eax
+	ret
+
+	.p2align 4
+L(exit_loop_32):
+	movdqa	48(%ecx), %xmm0
+	pcmpeqb	%xmm1, %xmm0
+	pmovmskb %xmm0, %eax
+	test	%eax, %eax
+	jnz	L(matches48_1)
+	cmp	$16, %edx
+	jbe	L(return_null)
+
+	pcmpeqb	32(%ecx), %xmm1
+	pmovmskb %xmm1, %eax
+	test	%eax, %eax
+	jnz	L(matches32_1)
+	xor	%eax, %eax
+	ret
+
+	.p2align 4
+L(matches0):
+	bsr	%eax, %eax
+	add	%ecx, %eax
+	ret
+
+	.p2align 4
+L(matches16):
+	bsr	%eax, %eax
+	lea	16(%eax, %ecx), %eax
+	ret
+
+	.p2align 4
+L(matches32):
+	bsr	%eax, %eax
+	lea	32(%eax, %ecx), %eax
+	ret
+
+	.p2align 4
+L(matches48):
+	bsr	%eax, %eax
+	lea	48(%eax, %ecx), %eax
+	ret
+
+	.p2align 4
+L(matches0_1):
+	bsr	%eax, %eax
+	sub	$64, %edx
+	add	%eax, %edx
+	jl	L(return_null)
+	add	%ecx, %eax
+	ret
+
+	.p2align 4
+L(matches16_1):
+	bsr	%eax, %eax
+	sub	$48, %edx
+	add	%eax, %edx
+	jl	L(return_null)
+	lea	16(%ecx, %eax), %eax
+	ret
+
+	.p2align 4
+L(matches32_1):
+	bsr	%eax, %eax
+	sub	$32, %edx
+	add	%eax, %edx
+	jl	L(return_null)
+	lea	32(%ecx, %eax), %eax
+	ret
+
+	.p2align 4
+L(matches48_1):
+	bsr	%eax, %eax
+	sub	$16, %edx
+	add	%eax, %edx
+	jl	L(return_null)
+	lea	48(%ecx, %eax), %eax
+	ret
+
+	.p2align 4
+L(return_null):
+	xor	%eax, %eax
+	ret
+
+	.p2align 4
+L(length_less16_offset0):
+	mov	%dl, %cl
+	pcmpeqb	(%eax), %xmm1
+
+	mov	$1, %edx
+	sal	%cl, %edx
+	sub	$1, %edx
+	mov	%edx, %ecx
+
+	pmovmskb %xmm1, %edx
+
+	and	%ecx, %edx
+	test	%edx, %edx
+	jz	L(return_null)
+
+	bsr	%edx, %ecx
+	add	%ecx, %eax
+	ret
+
+	.p2align 4
+L(length_less16):
+	punpcklbw %xmm1, %xmm1
+	mov	%ecx, %eax
+	punpcklbw %xmm1, %xmm1
+	add	$16, %edx
+	jz	L(return_null)
+
+	pshufd	$0, %xmm1, %xmm1
+	and	$15, %ecx
+	jz	L(length_less16_offset0)
+	
+	PUSH	(%edi)
+	mov	%cl, %dh
+	add	%dl, %dh
+	and	$-16, %eax
+
+	sub	$16, %dh
+	ja	L(length_less16_part2)
+
+	pcmpeqb	(%eax), %xmm1
+	pmovmskb %xmm1, %edi
+
+	sar	%cl, %edi
+	add	%ecx, %eax
+	mov	%dl, %cl
+
+	mov	$1, %edx
+	sal	%cl, %edx
+	sub	$1, %edx
+
+	and	%edx, %edi
+	test	%edi, %edi
+	jz	L(ret_null)
+
+	bsr	%edi, %edi
+	add	%edi, %eax
+	POP	(%edi)
+	ret
+
+	CFI_PUSH     (%edi)
+
+	.p2align 4
+L(length_less16_part2):
+	movdqa	16(%eax), %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	pmovmskb %xmm2, %edi
+
+	mov	%cl, %ch
+
+	mov	%dh, %cl
+	mov	$1, %edx
+	sal	%cl, %edx
+	sub	$1, %edx
+
+	and	%edx, %edi
+
+	test	%edi, %edi
+	jnz	L(length_less16_part2_return)
+
+	pcmpeqb	(%eax), %xmm1
+	pmovmskb %xmm1, %edi
+
+	mov	%ch, %cl
+	sar	%cl, %edi
+	test	%edi, %edi
+	jz	L(ret_null)
+
+	bsr	%edi, %edi
+	add	%edi, %eax
+	xor	%ch, %ch
+	add	%ecx, %eax
+	POP	(%edi)
+	ret
+
+	CFI_PUSH     (%edi)
+
+	.p2align 4
+L(length_less16_part2_return):
+	bsr	%edi, %edi
+	lea	16(%eax, %edi), %eax
+	POP	(%edi)
+	ret
+
+	CFI_PUSH     (%edi)
+
+	.p2align 4
+L(ret_null):
+	xor	%eax, %eax
+	POP	(%edi)
+	ret
+
+END (MEMCHR)
+#endif
diff --git a/sysdeps/i386/i686/multiarch/memrchr-sse2.S b/sysdeps/i386/i686/multiarch/memrchr-sse2.S
new file mode 100644
index 0000000000..68f4bc7b26
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/memrchr-sse2.S
@@ -0,0 +1,725 @@
+/* Optimized memrchr with sse2 without bsf
+   Copyright (C) 2011 Free Software Foundation, Inc.
+   Contributed by Intel Corporation.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library 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.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, write to the Free
+   Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+   02111-1307 USA.  */
+
+#ifndef  NOT_IN_libc
+
+# include <sysdep.h>
+# define CFI_PUSH(REG)	\
+	cfi_adjust_cfa_offset (4);	\
+	cfi_rel_offset (REG, 0)
+
+# define CFI_POP(REG)	\
+	cfi_adjust_cfa_offset (-4);	\
+	cfi_restore (REG)
+
+# define PUSH(REG) pushl REG; CFI_PUSH (REG)
+# define POP(REG) popl REG; CFI_POP (REG)
+
+# define PARMS  4
+# define STR1  PARMS
+# define STR2  STR1+4
+# define LEN   STR2+4
+
+	atom_text_section
+ENTRY (__memrchr_sse2)
+	mov	STR1(%esp), %ecx
+	movd	STR2(%esp), %xmm1
+	mov	LEN(%esp), %edx
+
+	sub	$16, %edx
+	jbe	L(length_less16)
+
+	punpcklbw %xmm1, %xmm1
+	add	%edx, %ecx
+	punpcklbw %xmm1, %xmm1
+
+	movdqu	(%ecx), %xmm0
+	pshufd	$0, %xmm1, %xmm1
+	pcmpeqb	%xmm1, %xmm0
+
+	pmovmskb %xmm0, %eax
+	test	%eax, %eax
+	jnz	L(exit_dispatch)  
+
+	sub	$64, %ecx
+	mov	%ecx, %eax
+	and	$15, %eax
+	jz	L(loop_prolog)
+
+	lea	16(%ecx), %ecx
+	lea	16(%edx), %edx
+	sub	%eax, %edx
+	and	$-16, %ecx
+
+	.p2align 4
+/* Loop start on aligned string.  */
+L(loop_prolog):
+	sub	$64, %edx
+	jbe	L(exit_loop)
+
+	movdqa	48(%ecx), %xmm0
+	pcmpeqb	%xmm1, %xmm0
+	pmovmskb %xmm0, %eax
+	test	%eax, %eax
+	jnz	L(matches48)
+
+	movdqa	32(%ecx), %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	pmovmskb %xmm2, %eax
+	test	%eax, %eax
+	jnz	L(matches32)
+
+	movdqa	16(%ecx), %xmm3
+	pcmpeqb	%xmm1, %xmm3
+	pmovmskb %xmm3, %eax
+	test	%eax, %eax
+	jnz	L(matches16)
+
+	movdqa	(%ecx), %xmm4
+	pcmpeqb	%xmm1, %xmm4
+	pmovmskb %xmm4, %eax
+	test	%eax, %eax
+	jnz	L(exit_dispatch)
+
+	sub	$64, %ecx
+	sub	$64, %edx
+	jbe	L(exit_loop)
+
+	movdqa	48(%ecx), %xmm0
+	pcmpeqb	%xmm1, %xmm0
+	pmovmskb %xmm0, %eax
+	test	%eax, %eax
+	jnz	L(matches48)
+
+	movdqa	32(%ecx), %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	pmovmskb %xmm2, %eax
+	test	%eax, %eax
+	jnz	L(matches32)
+
+	movdqa	16(%ecx), %xmm3
+	pcmpeqb	%xmm1, %xmm3
+	pmovmskb %xmm3, %eax
+	test	%eax, %eax
+	jnz	L(matches16)
+
+	movdqa	(%ecx), %xmm3
+	pcmpeqb	%xmm1, %xmm3
+	pmovmskb %xmm3, %eax
+	test	%eax, %eax
+	jnz	L(exit_dispatch)
+
+	mov	%ecx, %eax
+	and	$63, %eax
+	test	%eax, %eax
+	jz	L(align64_loop)    
+
+	lea	64(%ecx), %ecx
+	lea	64(%edx), %edx
+	and	$-64, %ecx
+	sub	%eax, %edx
+
+	.p2align 4
+L(align64_loop):
+	sub	$64, %ecx
+	sub	$64, %edx
+	jbe	L(exit_loop)
+
+	movdqa	(%ecx), %xmm0
+	movdqa	16(%ecx), %xmm2
+	movdqa	32(%ecx), %xmm3
+	movdqa	48(%ecx), %xmm4
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm1, %xmm2
+	pcmpeqb	%xmm1, %xmm3
+	pcmpeqb	%xmm1, %xmm4
+
+	pmaxub	%xmm3, %xmm0
+	pmaxub	%xmm4, %xmm2
+	pmaxub	%xmm0, %xmm2
+	pmovmskb %xmm2, %eax
+
+	test	%eax, %eax
+	jz	L(align64_loop)
+
+	pmovmskb %xmm4, %eax
+	test	%eax, %eax
+	jnz	L(matches48)
+
+	pmovmskb %xmm3, %eax
+	test	%eax, %eax
+	jnz	L(matches32)
+
+	movdqa	16(%ecx), %xmm2
+
+	pcmpeqb	%xmm1, %xmm2
+	pcmpeqb	(%ecx), %xmm1
+
+	pmovmskb %xmm2, %eax
+	test	%eax, %eax
+	jnz	L(matches16)
+
+	pmovmskb %xmm1, %eax
+	test	%ah, %ah
+	jnz	L(exit_dispatch_high)
+	mov	%al, %dl
+	and	$15 << 4, %dl
+	jnz	L(exit_dispatch_8)
+	test	$0x08, %al
+	jnz	L(exit_4)
+	test	$0x04, %al
+	jnz	L(exit_3)
+	test	$0x02, %al
+	jnz	L(exit_2)
+	mov	%ecx, %eax
+	ret
+
+	.p2align 4
+L(exit_loop):
+	add	$64, %edx
+	cmp	$32, %edx
+	jbe	L(exit_loop_32)
+
+	movdqa	48(%ecx), %xmm0
+	pcmpeqb	%xmm1, %xmm0
+	pmovmskb %xmm0, %eax
+	test	%eax, %eax
+	jnz	L(matches48)
+
+	movdqa	32(%ecx), %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	pmovmskb %xmm2, %eax
+	test	%eax, %eax
+	jnz	L(matches32)
+
+	movdqa	16(%ecx), %xmm3
+	pcmpeqb	%xmm1, %xmm3
+	pmovmskb %xmm3, %eax
+	test	%eax, %eax
+	jnz	L(matches16_1)
+	cmp	$48, %edx
+	jbe	L(return_null)
+
+	pcmpeqb	(%ecx), %xmm1
+	pmovmskb %xmm1, %eax
+	test	%eax, %eax
+	jnz	L(matches0_1)
+	xor	%eax, %eax
+	ret
+
+	.p2align 4
+L(exit_loop_32):
+	movdqa	48(%ecx), %xmm0
+	pcmpeqb	%xmm1, %xmm0
+	pmovmskb %xmm0, %eax
+	test	%eax, %eax
+	jnz	L(matches48_1)
+	cmp	$16, %edx
+	jbe	L(return_null)
+
+	pcmpeqb	32(%ecx), %xmm1
+	pmovmskb %xmm1, %eax
+	test	%eax, %eax
+	jnz	L(matches32_1)
+	xor	%eax, %eax
+	ret
+
+	.p2align 4
+L(matches16):
+	lea	16(%ecx), %ecx	
+	test	%ah, %ah
+	jnz	L(exit_dispatch_high)
+	mov	%al, %dl
+	and	$15 << 4, %dl
+	jnz	L(exit_dispatch_8)
+	test	$0x08, %al
+	jnz	L(exit_4)
+	test	$0x04, %al
+	jnz	L(exit_3)
+	test	$0x02, %al
+	jnz	L(exit_2)
+	mov	%ecx, %eax
+	ret
+
+	.p2align 4
+L(matches32):
+	lea	32(%ecx), %ecx	
+	test	%ah, %ah
+	jnz	L(exit_dispatch_high)
+	mov	%al, %dl
+	and	$15 << 4, %dl
+	jnz	L(exit_dispatch_8)
+	test	$0x08, %al
+	jnz	L(exit_4)
+	test	$0x04, %al
+	jnz	L(exit_3)
+	test	$0x02, %al
+	jnz	L(exit_2)
+	mov	%ecx, %eax
+	ret
+
+	.p2align 4
+L(matches48):
+	lea	48(%ecx), %ecx
+
+	.p2align 4
+L(exit_dispatch):
+	test	%ah, %ah
+	jnz	L(exit_dispatch_high)
+	mov	%al, %dl
+	and	$15 << 4, %dl
+	jnz	L(exit_dispatch_8)
+	test	$0x08, %al
+	jnz	L(exit_4)
+	test	$0x04, %al
+	jnz	L(exit_3)
+	test	$0x02, %al
+	jnz	L(exit_2)
+	mov	%ecx, %eax
+	ret
+
+	.p2align 4
+L(exit_dispatch_8):        
+	test	$0x80, %al
+	jnz	L(exit_8)
+	test	$0x40, %al
+	jnz	L(exit_7)
+	test	$0x20, %al
+	jnz	L(exit_6)
+	lea	4(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_dispatch_high):
+	mov	%ah, %dh
+	and	$15 << 4, %dh
+	jnz	L(exit_dispatch_high_8)
+	test	$0x08, %ah
+	jnz	L(exit_12)
+	test	$0x04, %ah
+	jnz	L(exit_11)
+	test	$0x02, %ah
+	jnz	L(exit_10)
+	lea	8(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_dispatch_high_8):
+	test	$0x80, %ah
+	jnz	L(exit_16)
+	test	$0x40, %ah
+	jnz	L(exit_15)
+	test	$0x20, %ah
+	jnz	L(exit_14)
+	lea	12(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_2):
+	lea	1(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_3):
+	lea	2(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_4):
+	lea	3(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_6):
+	lea	5(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_7):
+	lea	6(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_8):
+	lea	7(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_10):
+	lea	9(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_11):
+	lea	10(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_12):
+	lea	11(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_14):
+	lea	13(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_15):
+	lea	14(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_16):
+	lea	15(%ecx), %eax
+	ret
+
+	.p2align 4
+L(matches0_1):
+	lea	-64(%edx), %edx
+
+	test	%ah, %ah
+	jnz	L(exit_dispatch_1_high)
+	mov	%al, %ah
+	and	$15 << 4, %ah
+	jnz	L(exit_dispatch_1_8)
+	test	$0x08, %al
+	jnz	L(exit_1_4)
+	test	$0x04, %al
+	jnz	L(exit_1_3)
+	test	$0x02, %al
+	jnz	L(exit_1_2)
+	add	$0, %edx
+	jl	L(return_null)
+	mov	%ecx, %eax
+	ret
+
+	.p2align 4
+L(matches16_1):
+	lea	-48(%edx), %edx
+	lea	16(%ecx), %ecx
+
+	test	%ah, %ah
+	jnz	L(exit_dispatch_1_high)
+	mov	%al, %ah
+	and	$15 << 4, %ah
+	jnz	L(exit_dispatch_1_8)
+	test	$0x08, %al
+	jnz	L(exit_1_4)
+	test	$0x04, %al
+	jnz	L(exit_1_3)
+	test	$0x02, %al
+	jnz	L(exit_1_2)
+	add	$0, %edx
+	jl	L(return_null)
+	mov	%ecx, %eax
+	ret
+
+	.p2align 4
+L(matches32_1):
+	lea	-32(%edx), %edx
+	lea	32(%ecx), %ecx
+	
+	test	%ah, %ah
+	jnz	L(exit_dispatch_1_high)
+	mov	%al, %ah
+	and	$15 << 4, %ah
+	jnz	L(exit_dispatch_1_8)
+	test	$0x08, %al
+	jnz	L(exit_1_4)
+	test	$0x04, %al
+	jnz	L(exit_1_3)
+	test	$0x02, %al
+	jnz	L(exit_1_2)
+	add	$0, %edx
+	jl	L(return_null)
+	mov	%ecx, %eax
+	ret
+
+	.p2align 4
+L(matches48_1):
+	lea	-16(%edx), %edx
+	lea	48(%ecx), %ecx
+
+	.p2align 4
+L(exit_dispatch_1):
+	test	%ah, %ah
+	jnz	L(exit_dispatch_1_high)
+	mov	%al, %ah
+	and	$15 << 4, %ah
+	jnz	L(exit_dispatch_1_8)
+	test	$0x08, %al
+	jnz	L(exit_1_4)
+	test	$0x04, %al
+	jnz	L(exit_1_3)
+	test	$0x02, %al
+	jnz	L(exit_1_2)
+	add	$0, %edx
+	jl	L(return_null)
+	mov	%ecx, %eax
+	ret
+
+	.p2align 4
+L(exit_dispatch_1_8):        
+	test	$0x80, %al
+	jnz	L(exit_1_8)
+	test	$0x40, %al
+	jnz	L(exit_1_7)
+	test	$0x20, %al
+	jnz	L(exit_1_6)
+	add	$4, %edx
+	jl	L(return_null)
+	lea	4(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_dispatch_1_high):
+	mov	%ah, %al
+	and	$15 << 4, %al
+	jnz	L(exit_dispatch_1_high_8)
+	test	$0x08, %ah
+	jnz	L(exit_1_12)
+	test	$0x04, %ah
+	jnz	L(exit_1_11)
+	test	$0x02, %ah
+	jnz	L(exit_1_10)
+	add	$8, %edx
+	jl	L(return_null)
+	lea	8(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_dispatch_1_high_8):
+	test	$0x80, %ah
+	jnz	L(exit_1_16)
+	test	$0x40, %ah
+	jnz	L(exit_1_15)
+	test	$0x20, %ah
+	jnz	L(exit_1_14)
+	add	$12, %edx
+	jl	L(return_null)
+	lea	12(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_1_2):
+	add	$1, %edx
+	jl	L(return_null)
+	lea	1(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_1_3):
+	add	$2, %edx
+	jl	L(return_null)
+	lea	2(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_1_4):
+	add	$3, %edx
+	jl	L(return_null)
+	lea	3(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_1_6):
+	add	$5, %edx
+	jl	L(return_null)
+	lea	5(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_1_7):
+	add	$6, %edx
+	jl	L(return_null)
+	lea	6(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_1_8):
+	add	$7, %edx
+	jl	L(return_null)
+	lea	7(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_1_10):
+	add	$9, %edx
+	jl	L(return_null)
+	lea	9(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_1_11):
+	add	$10, %edx
+	jl	L(return_null)
+	lea	10(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_1_12):
+	add	$11, %edx
+	jl	L(return_null)
+	lea	11(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_1_14):
+	add	$13, %edx
+	jl	L(return_null)
+	lea	13(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_1_15):
+	add	$14, %edx
+	jl	L(return_null)
+	lea	14(%ecx), %eax
+	ret
+
+	.p2align 4
+L(exit_1_16):
+	add	$15, %edx
+	jl	L(return_null)
+	lea	15(%ecx), %eax
+	ret
+
+	.p2align 4
+L(return_null):
+	xor	%eax, %eax
+	ret
+
+	.p2align 4
+L(length_less16_offset0):
+	mov	%dl, %cl
+	pcmpeqb	(%eax), %xmm1
+
+	mov	$1, %edx
+	sal	%cl, %edx
+	sub	$1, %edx
+
+	mov	%eax, %ecx
+	pmovmskb %xmm1, %eax
+
+	and	%edx, %eax
+	test	%eax, %eax
+	jnz	L(exit_dispatch)
+
+	xor	%eax, %eax
+	ret
+
+	.p2align 4
+L(length_less16):
+	punpcklbw %xmm1, %xmm1
+	add	$16, %edx
+	je	L(return_null)
+	punpcklbw %xmm1, %xmm1
+
+	mov	%ecx, %eax
+	pshufd	$0, %xmm1, %xmm1
+
+	and	$15, %ecx
+	jz	L(length_less16_offset0)
+	
+	PUSH	(%edi)
+
+	mov	%cl, %dh
+	add	%dl, %dh
+	and	$-16, %eax
+
+	sub	$16, %dh
+	ja	L(length_less16_part2)
+
+	pcmpeqb	(%eax), %xmm1
+	pmovmskb %xmm1, %edi
+
+	sar	%cl, %edi
+	add	%ecx, %eax
+	mov	%dl, %cl
+
+	mov	$1, %edx
+	sal	%cl, %edx
+	sub	$1, %edx
+
+	and	%edx, %edi
+	test	%edi, %edi
+	jz	L(ret_null)
+
+	bsr	%edi, %edi
+	add	%edi, %eax
+	POP	(%edi)
+	ret
+
+	CFI_PUSH     (%edi)
+
+	.p2align 4
+L(length_less16_part2):
+	movdqa	16(%eax), %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	pmovmskb %xmm2, %edi
+
+	mov	%cl, %ch
+
+	mov	%dh, %cl
+	mov	$1, %edx
+	sal	%cl, %edx
+	sub	$1, %edx
+
+	and	%edx, %edi
+
+	test	%edi, %edi
+	jnz	L(length_less16_part2_return)
+
+	pcmpeqb	(%eax), %xmm1
+	pmovmskb %xmm1, %edi
+
+	mov	%ch, %cl
+	sar	%cl, %edi
+	test	%edi, %edi
+	jz	L(ret_null)
+
+	bsr	%edi, %edi
+	add	%edi, %eax
+	xor	%ch, %ch
+	add	%ecx, %eax
+	POP	(%edi)
+	ret
+
+	CFI_PUSH     (%edi)
+
+	.p2align 4
+L(length_less16_part2_return):
+	bsr	%edi, %edi
+	lea	16(%eax, %edi), %eax
+	POP	(%edi)
+	ret
+
+	CFI_PUSH     (%edi)
+
+	.p2align 4
+L(ret_null):
+	xor	%eax, %eax
+	POP	(%edi)
+	ret
+
+END (__memrchr_sse2)
+#endif
diff --git a/sysdeps/i386/i686/multiarch/memrchr.S b/sysdeps/i386/i686/multiarch/memrchr.S
new file mode 100644
index 0000000000..8e5b2c50a2
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/memrchr.S
@@ -0,0 +1,79 @@
+/* Multiple versions of memrchr
+   Copyright (C) 2011 Free Software Foundation, Inc.
+   Contributed by Intel Corporation.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library 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.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, write to the Free
+   Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+   02111-1307 USA.  */
+
+#include <sysdep.h>
+#include <init-arch.h>
+
+#ifndef  NOT_IN_libc
+	.section	.gnu.linkonce.t.__i686.get_pc_thunk.bx,"ax",@progbits
+	.globl	__i686.get_pc_thunk.bx
+	.hidden	__i686.get_pc_thunk.bx
+	.p2align 4
+	.type	__i686.get_pc_thunk.bx,@function
+__i686.get_pc_thunk.bx:
+	movl	(%esp), %ebx
+	ret
+
+# define CFI_POP(REG) \
+	cfi_adjust_cfa_offset (-4); \
+	cfi_restore (REG)
+
+# define CFI_PUSH(REG) \
+	cfi_adjust_cfa_offset (4); \
+	cfi_rel_offset (REG, 0)
+
+	.text
+ENTRY(__memrchr)
+	.type	__memrchr, @gnu_indirect_function
+	pushl	%ebx
+	CFI_PUSH (%ebx)
+	call	__i686.get_pc_thunk.bx
+	addl	$_GLOBAL_OFFSET_TABLE_, %ebx
+	cmpl	$0, KIND_OFFSET+__cpu_features@GOTOFF(%ebx)
+	jne	1f
+	call	__init_cpu_features
+
+1:	testl	$bit_SSE2, CPUID_OFFSET+index_SSE2+__cpu_features@GOTOFF(%ebx)
+	jz	2f
+	testl	$bit_Slow_BSF, FEATURE_OFFSET+index_Slow_BSF+__cpu_features@GOTOFF(%ebx)
+	jz	3f
+
+	leal	__memrchr_sse2@GOTOFF(%ebx), %eax
+	popl	%ebx
+	CFI_POP	(%ebx)
+	ret
+
+	CFI_PUSH (%ebx)
+
+2:	leal	__memrchr_ia32@GOTOFF(%ebx), %eax
+	popl	%ebx
+	CFI_POP	(%ebx)
+	ret
+
+	CFI_PUSH (%ebx)
+
+3:	leal	__memrchr_sse2_bsf@GOTOFF(%ebx), %eax
+	popl	%ebx
+	CFI_POP	(%ebx)
+	ret
+END(__memrchr)
+
+weak_alias(__memrchr, memrchr)
+#endif
diff --git a/sysdeps/i386/i686/multiarch/rawmemchr-sse2-bsf.S b/sysdeps/i386/i686/multiarch/rawmemchr-sse2-bsf.S
new file mode 100644
index 0000000000..88c0e5776c
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/rawmemchr-sse2-bsf.S
@@ -0,0 +1,3 @@
+#define USE_AS_RAWMEMCHR
+#define MEMCHR __rawmemchr_sse2_bsf
+#include "memchr-sse2-bsf.S"
diff --git a/sysdeps/i386/i686/multiarch/rawmemchr-sse2.S b/sysdeps/i386/i686/multiarch/rawmemchr-sse2.S
new file mode 100644
index 0000000000..038c74896b
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/rawmemchr-sse2.S
@@ -0,0 +1,3 @@
+#define USE_AS_RAWMEMCHR
+#define MEMCHR __rawmemchr_sse2
+#include "memchr-sse2.S"
diff --git a/sysdeps/i386/i686/multiarch/rawmemchr.S b/sysdeps/i386/i686/multiarch/rawmemchr.S
new file mode 100644
index 0000000000..111f0dcf63
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/rawmemchr.S
@@ -0,0 +1,99 @@
+/* Multiple versions of rawmemchr
+   Copyright (C) 2011 Free Software Foundation, Inc.
+   Contributed by Intel Corporation.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library 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.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, write to the Free
+   Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+   02111-1307 USA.  */
+
+#include <sysdep.h>
+#include <init-arch.h>
+
+#ifndef  NOT_IN_libc
+	.section	.gnu.linkonce.t.__i686.get_pc_thunk.bx,"ax",@progbits
+	.globl	__i686.get_pc_thunk.bx
+	.hidden	__i686.get_pc_thunk.bx
+	.p2align 4
+	.type	__i686.get_pc_thunk.bx,@function
+__i686.get_pc_thunk.bx:
+	movl	(%esp), %ebx
+	ret
+
+# define CFI_POP(REG) \
+	cfi_adjust_cfa_offset (-4); \
+	cfi_restore (REG)
+
+# define CFI_PUSH(REG) \
+	cfi_adjust_cfa_offset (4); \
+	cfi_rel_offset (REG, 0)
+
+	.text
+ENTRY(__rawmemchr)
+	.type	__rawmemchr, @gnu_indirect_function
+	pushl	%ebx
+	CFI_PUSH (%ebx)
+	call	__i686.get_pc_thunk.bx
+	addl	$_GLOBAL_OFFSET_TABLE_, %ebx
+	cmpl	$0, KIND_OFFSET+__cpu_features@GOTOFF(%ebx)
+	jne	1f
+	call	__init_cpu_features
+
+1:	testl	$bit_SSE2, CPUID_OFFSET+index_SSE2+__cpu_features@GOTOFF(%ebx)
+	jz	2f
+	testl	$bit_Slow_BSF, FEATURE_OFFSET+index_Slow_BSF+__cpu_features@GOTOFF(%ebx)
+	jz	3f
+
+	leal	__rawmemchr_sse2@GOTOFF(%ebx), %eax
+	popl	%ebx
+	CFI_POP	(%ebx)
+	ret
+
+	CFI_PUSH (%ebx)
+
+2:	leal	__rawmemchr_ia32@GOTOFF(%ebx), %eax
+	popl	%ebx
+	CFI_POP	(%ebx)
+	ret
+
+	CFI_PUSH (%ebx)
+
+3:	leal	__rawmemchr_sse2_bsf@GOTOFF(%ebx), %eax
+	popl	%ebx
+	CFI_POP	(%ebx)
+	ret
+END(__rawmemchr)
+
+weak_alias(__rawmemchr, rawmemchr)
+
+# undef ENTRY
+# define ENTRY(name) \
+	.type __rawmemchr_ia32, @function; \
+	.globl __rawmemchr_ia32; \
+	.p2align 4; \
+	__rawmemchr_ia32: cfi_startproc; \
+	CALL_MCOUNT
+# undef END
+# define END(name) \
+	cfi_endproc; .size __rawmemchr_ia32, .-__rawmemchr_ia32
+
+# undef libc_hidden_def
+/* IFUNC doesn't work with the hidden functions in shared library since
+   they will be called without setting up EBX needed for PLT which is
+   used by IFUNC.  */
+# define libc_hidden_def(name) \
+	.globl __GI___rawmemchr; __GI___rawmemchr = __rawmemchr_ia32
+
+#endif
+#include "../../rawmemchr.S"