Optimize 32bit memset/memcpy with SSE2/SSSE3.

30 files changed, 6229 insertions, 12 deletions

diff --git a/sysdeps/i386/i686/bcopy.S b/sysdeps/i386/i686/bcopy.S
new file mode 100644
index 0000000000..15ef9419a4
--- /dev/null
+++ b/sysdeps/i386/i686/bcopy.S
@@ -0,0 +1,3 @@
+#define USE_AS_BCOPY
+#define memmove bcopy
+#include <sysdeps/i386/i686/memmove.S>
diff --git a/sysdeps/i386/i686/cacheinfo.c b/sysdeps/i386/i686/cacheinfo.c
index 82e4cd223e..f8b7f521ca 100644
--- a/sysdeps/i386/i686/cacheinfo.c
+++ b/sysdeps/i386/i686/cacheinfo.c
@@ -1,3 +1,4 @@
+#define __x86_64_data_cache_size __x86_data_cache_size
 #define __x86_64_data_cache_size_half __x86_data_cache_size_half
 #define __x86_64_shared_cache_size __x86_shared_cache_size
 #define __x86_64_shared_cache_size_half __x86_shared_cache_size_half
diff --git a/sysdeps/i386/i686/memcpy.S b/sysdeps/i386/i686/memcpy.S
index 0b2da1ea27..86ee082beb 100644
--- a/sysdeps/i386/i686/memcpy.S
+++ b/sysdeps/i386/i686/memcpy.S
@@ -32,11 +32,11 @@
 
 	.text
 #if defined PIC && !defined NOT_IN_libc
-ENTRY (__memcpy_chk)
+ENTRY_CHK (__memcpy_chk)
 	movl	12(%esp), %eax
 	cmpl	%eax, 16(%esp)
 	jb	HIDDEN_JUMPTARGET (__chk_fail)
-END (__memcpy_chk)
+END_CHK (__memcpy_chk)
 #endif
 ENTRY (BP_SYM (memcpy))
 	ENTER
diff --git a/sysdeps/i386/i686/memmove.S b/sysdeps/i386/i686/memmove.S
index b93b5c729f..981f14f4e0 100644
--- a/sysdeps/i386/i686/memmove.S
+++ b/sysdeps/i386/i686/memmove.S
@@ -26,18 +26,27 @@
 
 #define PARMS	LINKAGE+4	/* one spilled register */
 #define RTN	PARMS
-#define DEST	RTN+RTN_SIZE
-#define SRC	DEST+PTR_SIZE
-#define LEN	SRC+PTR_SIZE
 
 	.text
-#if defined PIC && !defined NOT_IN_libc
-ENTRY (__memmove_chk)
+
+#ifdef USE_AS_BCOPY
+# define SRC	RTN+RTN_SIZE
+# define DEST	SRC+PTR_SIZE
+# define LEN	DEST+PTR_SIZE
+#else
+# define DEST	RTN+RTN_SIZE
+# define SRC	DEST+PTR_SIZE
+# define LEN	SRC+PTR_SIZE
+
+# if defined PIC && !defined NOT_IN_libc
+ENTRY_CHK (__memmove_chk)
 	movl	12(%esp), %eax
 	cmpl	%eax, 16(%esp)
 	jb	HIDDEN_JUMPTARGET (__chk_fail)
-END (__memmove_chk)
+END_CHK (__memmove_chk)
+# endif
 #endif
+
 ENTRY (BP_SYM (memmove))
 	ENTER
 
@@ -69,8 +78,10 @@ ENTRY (BP_SYM (memmove))
 	movsl
 	movl	%edx, %esi
 	cfi_restore (esi)
+#ifndef USE_AS_BCOPY
 	movl	DEST(%esp), %eax
 	RETURN_BOUNDED_POINTER (DEST(%esp))
+#endif
 
 	popl	%edi
 	cfi_adjust_cfa_offset (-4)
@@ -101,8 +112,10 @@ ENTRY (BP_SYM (memmove))
 	movsl
 	movl	%edx, %esi
 	cfi_restore (esi)
+#ifndef USE_AS_BCOPY
 	movl	DEST(%esp), %eax
 	RETURN_BOUNDED_POINTER (DEST(%esp))
+#endif
 
 	cld
 	popl	%edi
@@ -112,4 +125,6 @@ ENTRY (BP_SYM (memmove))
 	LEAVE
 	RET_PTR
 END (BP_SYM (memmove))
+#ifndef USE_AS_BCOPY
 libc_hidden_builtin_def (memmove)
+#endif
diff --git a/sysdeps/i386/i686/mempcpy.S b/sysdeps/i386/i686/mempcpy.S
index 6437e4a5d4..c10686fb3d 100644
--- a/sysdeps/i386/i686/mempcpy.S
+++ b/sysdeps/i386/i686/mempcpy.S
@@ -32,11 +32,11 @@
 
 	.text
 #if defined PIC && !defined NOT_IN_libc
-ENTRY (__mempcpy_chk)
+ENTRY_CHK (__mempcpy_chk)
 	movl	12(%esp), %eax
 	cmpl	%eax, 16(%esp)
 	jb	HIDDEN_JUMPTARGET (__chk_fail)
-END (__mempcpy_chk)
+END_CHK (__mempcpy_chk)
 #endif
 ENTRY (BP_SYM (__mempcpy))
 	ENTER
diff --git a/sysdeps/i386/i686/memset.S b/sysdeps/i386/i686/memset.S
index dfa1aa7019..b343af7b64 100644
--- a/sysdeps/i386/i686/memset.S
+++ b/sysdeps/i386/i686/memset.S
@@ -40,11 +40,11 @@
 
         .text
 #if defined PIC && !defined NOT_IN_libc && !BZERO_P
-ENTRY (__memset_chk)
+ENTRY_CHK (__memset_chk)
 	movl	12(%esp), %eax
 	cmpl	%eax, 16(%esp)
 	jb	HIDDEN_JUMPTARGET (__chk_fail)
-END (__memset_chk)
+END_CHK (__memset_chk)
 #endif
 ENTRY (BP_SYM (memset))
 	ENTER
diff --git a/sysdeps/i386/i686/multiarch/Makefile b/sysdeps/i386/i686/multiarch/Makefile
index e1553b284e..fbad9ae734 100644
--- a/sysdeps/i386/i686/multiarch/Makefile
+++ b/sysdeps/i386/i686/multiarch/Makefile
@@ -4,6 +4,10 @@ gen-as-const-headers += ifunc-defines.sym
 endif
 
 ifeq ($(subdir),string)
+sysdep_routines += bzero-sse2 memset-sse2 memcpy-ssse3 mempcpy-ssse3 \
+		   memmove-ssse3 memcpy-ssse3-rep mempcpy-ssse3-rep \
+		   memmove-ssse3-rep bcopy-ssse3 bcopy-ssse3-rep \
+		   memset-sse2-rep bzero-sse2-rep
 ifeq (yes,$(config-cflags-sse4))
 sysdep_routines += strcspn-c strpbrk-c strspn-c strstr-c strcasestr-c
 CFLAGS-strcspn-c.c += -msse4
diff --git a/sysdeps/i386/i686/multiarch/bcopy-ssse3-rep.S b/sysdeps/i386/i686/multiarch/bcopy-ssse3-rep.S
new file mode 100644
index 0000000000..cbc8b420e8
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/bcopy-ssse3-rep.S
@@ -0,0 +1,4 @@
+#define USE_AS_MEMMOVE
+#define USE_AS_BCOPY
+#define MEMCPY		__bcopy_ssse3_rep
+#include "memcpy-ssse3-rep.S"
diff --git a/sysdeps/i386/i686/multiarch/bcopy-ssse3.S b/sysdeps/i386/i686/multiarch/bcopy-ssse3.S
new file mode 100644
index 0000000000..36aac44b9c
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/bcopy-ssse3.S
@@ -0,0 +1,4 @@
+#define USE_AS_MEMMOVE
+#define USE_AS_BCOPY
+#define MEMCPY		__bcopy_ssse3
+#include "memcpy-ssse3.S"
diff --git a/sysdeps/i386/i686/multiarch/bcopy.S b/sysdeps/i386/i686/multiarch/bcopy.S
new file mode 100644
index 0000000000..8671bf684e
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/bcopy.S
@@ -0,0 +1,89 @@
+/* Multiple versions of bcopy
+   Copyright (C) 2010 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>
+
+/* Define multiple versions only for the definition in lib.  */
+#ifndef NOT_IN_libc
+# ifdef SHARED
+	.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
+
+	.text
+ENTRY(bcopy)
+	.type	bcopy, @gnu_indirect_function
+	pushl	%ebx
+	cfi_adjust_cfa_offset (4)
+	cfi_rel_offset (ebx, 0)
+	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:	leal	__bcopy_ia32@GOTOFF(%ebx), %eax
+	testl	$bit_SSSE3, CPUID_OFFSET+index_SSSE3+__cpu_features@GOTOFF(%ebx)
+	jz	2f
+	leal	__bcopy_ssse3@GOTOFF(%ebx), %eax
+	testl	$bit_Fast_Rep_String, FEATURE_OFFSET+index_Fast_Rep_String+__cpu_features@GOTOFF(%ebx)
+	jz	2f
+	leal	__bcopy_ssse3_rep@GOTOFF(%ebx), %eax
+2:	popl	%ebx
+	cfi_adjust_cfa_offset (-4)
+	cfi_restore (ebx)
+	ret
+END(bcopy)
+# else
+	.text
+ENTRY(bcopy)
+	.type	bcopy, @gnu_indirect_function
+	cmpl	$0, KIND_OFFSET+__cpu_features
+	jne	1f
+	call	__init_cpu_features
+1:	leal	__bcopy_ia32, %eax
+	testl	$bit_SSSE3, CPUID_OFFSET+index_SSSE3+__cpu_features
+	jz	2f
+	leal	__bcopy_ssse3, %eax
+	testl	$bit_Fast_Rep_String, FEATURE_OFFSET+index_Fast_Rep_String+__cpu_features
+	jz	2f
+	leal	__bcopy_ssse3_rep, %eax
+2:	ret
+END(bcopy)
+# endif
+
+# undef ENTRY
+# define ENTRY(name) \
+	.type __bcopy_ia32, @function; \
+	.p2align 4; \
+	__bcopy_ia32: cfi_startproc; \
+	CALL_MCOUNT
+# undef END
+# define END(name) \
+	cfi_endproc; .size __bcopy_ia32, .-__bcopy_ia32
+
+#endif
+
+#include "../bcopy.S"
diff --git a/sysdeps/i386/i686/multiarch/bzero-sse2-rep.S b/sysdeps/i386/i686/multiarch/bzero-sse2-rep.S
new file mode 100644
index 0000000000..507b288bb3
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/bzero-sse2-rep.S
@@ -0,0 +1,3 @@
+#define USE_AS_BZERO
+#define __memset_sse2_rep __bzero_sse2_rep
+#include "memset-sse2-rep.S"
diff --git a/sysdeps/i386/i686/multiarch/bzero-sse2.S b/sysdeps/i386/i686/multiarch/bzero-sse2.S
new file mode 100644
index 0000000000..8d04512e4e
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/bzero-sse2.S
@@ -0,0 +1,3 @@
+#define USE_AS_BZERO
+#define __memset_sse2 __bzero_sse2
+#include "memset-sse2.S"
diff --git a/sysdeps/i386/i686/multiarch/bzero.S b/sysdeps/i386/i686/multiarch/bzero.S
new file mode 100644
index 0000000000..8c740a42dc
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/bzero.S
@@ -0,0 +1,97 @@
+/* Multiple versions of bzero
+   Copyright (C) 2010 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>
+
+/* Define multiple versions only for the definition in lib.  */
+#ifndef NOT_IN_libc
+# ifdef SHARED
+	.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
+
+	.text
+ENTRY(__bzero)
+	.type	__bzero, @gnu_indirect_function
+	pushl	%ebx
+	cfi_adjust_cfa_offset (4)
+	cfi_rel_offset (ebx, 0)
+	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:	leal	__bzero_ia32@GOTOFF(%ebx), %eax
+	testl	$bit_SSE2, CPUID_OFFSET+index_SSE2+__cpu_features@GOTOFF(%ebx)
+	jz	2f
+	leal	__bzero_sse2@GOTOFF(%ebx), %eax
+	testl	$bit_Fast_Rep_String, FEATURE_OFFSET+index_Fast_Rep_String+__cpu_features@GOTOFF(%ebx)
+	jz	2f
+	leal	__bzero_sse2_rep@GOTOFF(%ebx), %eax
+2:	popl	%ebx
+	cfi_adjust_cfa_offset (-4)
+	cfi_restore (ebx)
+	ret
+END(__bzero)
+# else
+	.text
+ENTRY(__bzero)
+	.type	__bzero, @gnu_indirect_function
+	cmpl	$0, KIND_OFFSET+__cpu_features
+	jne	1f
+	call	__init_cpu_features
+1:	leal	__bzero_ia32, %eax
+	testl	$bit_SSE2, CPUID_OFFSET+index_SSE2+__cpu_features
+	jz	2f
+	leal	__bzero_sse2, %eax
+	testl	$bit_Fast_Rep_String, FEATURE_OFFSET+index_Fast_Rep_String+__cpu_features
+	jz	2f
+	leal	__bzero_sse2_rep, %eax
+2:	ret
+END(__bzero)
+# endif
+
+# undef ENTRY
+# define ENTRY(name) \
+	.type __bzero_ia32, @function; \
+	.p2align 4; \
+	__bzero_ia32: cfi_startproc; \
+	CALL_MCOUNT
+# undef END
+# define END(name) \
+	cfi_endproc; .size __bzero_ia32, .-__bzero_ia32
+
+# ifdef SHARED
+#  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___bzero; __GI___bzero = __bzero_ia32
+# endif
+#endif
+
+#include "../bzero.S"
diff --git a/sysdeps/i386/i686/multiarch/ifunc-defines.sym b/sysdeps/i386/i686/multiarch/ifunc-defines.sym
index e2021cdf87..eb1538abcc 100644
--- a/sysdeps/i386/i686/multiarch/ifunc-defines.sym
+++ b/sysdeps/i386/i686/multiarch/ifunc-defines.sym
@@ -13,5 +13,8 @@ CPUID_ECX_OFFSET	offsetof (struct cpuid_registers, ecx)
 CPUID_EDX_OFFSET	offsetof (struct cpuid_registers, edx)
 FAMILY_OFFSET		offsetof (struct cpu_features, family)
 MODEL_OFFSET		offsetof (struct cpu_features, model)
+FEATURE_OFFSET		offsetof (struct cpu_features, feature)
+FEATURE_SIZE		sizeof (unsigned int)
 
 COMMON_CPUID_INDEX_1
+FEATURE_INDEX_1
diff --git a/sysdeps/i386/i686/multiarch/memcpy-ssse3-rep.S b/sysdeps/i386/i686/multiarch/memcpy-ssse3-rep.S
new file mode 100644
index 0000000000..47ec097f3c
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/memcpy-ssse3-rep.S
@@ -0,0 +1,1785 @@
+/* memcpy with SSSE3 and REP string.
+   Copyright (C) 2010 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>
+
+#if !defined NOT_IN_libc \
+    && (defined SHARED \
+        || defined USE_AS_MEMMOVE \
+	|| !defined USE_MULTIARCH)
+
+#include "asm-syntax.h"
+
+#ifndef MEMCPY
+# define MEMCPY		__memcpy_ssse3_rep
+# define MEMCPY_CHK	__memcpy_chk_ssse3_rep
+#endif
+
+#ifdef USE_AS_BCOPY
+# define SRC		PARMS
+# define DEST		SRC+4
+# define LEN		DEST+4
+#else
+# define DEST		PARMS
+# define SRC		DEST+4
+# define LEN		SRC+4
+#endif
+
+#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)
+
+#ifdef SHARED
+# define PARMS		8		/* Preserve EBX.  */
+# define ENTRANCE	PUSH (%ebx);
+# define RETURN_END	POP (%ebx); ret
+# define RETURN		RETURN_END; CFI_PUSH (%ebx)
+# define JMPTBL(I, B)	I - B
+
+/* Load an entry in a jump table into EBX and branch to it.  TABLE is a
+   jump table with relative offsets.  INDEX is a register contains the
+   index into the jump table.   SCALE is the scale of INDEX. */
+# define BRANCH_TO_JMPTBL_ENTRY(TABLE, INDEX, SCALE)		\
+    /* We first load PC into EBX.  */				\
+    call	__i686.get_pc_thunk.bx;				\
+    /* Get the address of the jump table.  */			\
+    addl	$(TABLE - .), %ebx;				\
+    /* Get the entry and convert the relative offset to the	\
+       absolute address.  */					\
+    addl	(%ebx,INDEX,SCALE), %ebx;			\
+    /* We loaded the jump table.  Go.  */			\
+    jmp		*%ebx
+
+# define BRANCH_TO_JMPTBL_ENTRY_VALUE(TABLE)			\
+    addl	$(TABLE - .), %ebx	
+			
+# define BRANCH_TO_JMPTBL_ENTRY_TAIL(TABLE, INDEX, SCALE)	\
+    addl	(%ebx,INDEX,SCALE), %ebx;			\
+    /* We loaded the jump table.  Go.  */			\
+    jmp		*%ebx
+
+	.section	.gnu.linkonce.t.__i686.get_pc_thunk.bx,"ax",@progbits
+	.globl	__i686.get_pc_thunk.bx
+	.hidden	__i686.get_pc_thunk.bx
+	ALIGN (4)
+	.type	__i686.get_pc_thunk.bx,@function
+__i686.get_pc_thunk.bx:
+	movl	(%esp), %ebx
+	ret
+#else
+# define PARMS		4
+# define ENTRANCE
+# define RETURN_END	ret
+# define RETURN		RETURN_END
+# define JMPTBL(I, B)	I
+
+/* Branch to an entry in a jump table.  TABLE is a jump table with
+   absolute offsets.  INDEX is a register contains the index into the
+   jump table.  SCALE is the scale of INDEX. */
+# define BRANCH_TO_JMPTBL_ENTRY(TABLE, INDEX, SCALE)		\
+    jmp		*TABLE(,INDEX,SCALE)
+
+# define BRANCH_TO_JMPTBL_ENTRY_VALUE(TABLE)			
+
+# define BRANCH_TO_JMPTBL_ENTRY_TAIL(TABLE, INDEX, SCALE)		\
+    jmp		*TABLE(,INDEX,SCALE)
+#endif
+
+	.section .text.ssse3,"ax",@progbits
+#if defined SHARED && !defined NOT_IN_libc && !defined USE_AS_BCOPY
+ENTRY (MEMCPY_CHK)
+	movl	12(%esp), %eax
+	cmpl	%eax, 16(%esp)
+	jb	HIDDEN_JUMPTARGET (__chk_fail)
+END (MEMCPY_CHK)
+#endif
+ENTRY (MEMCPY)
+	ENTRANCE
+	movl	LEN(%esp), %ecx
+	movl	SRC(%esp), %eax
+	movl	DEST(%esp), %edx
+
+#ifdef USE_AS_MEMMOVE
+	cmp	%eax, %edx
+	jb	L(copy_forward)
+	je	L(fwd_write_0bytes)
+	cmp	$32, %ecx
+	jge	L(memmove_bwd)
+	jmp	L(bk_write_less32bytes_2)
+L(memmove_bwd):
+	add	%ecx, %eax
+	cmp	%eax, %edx
+	movl	SRC(%esp), %eax
+	jb	L(copy_backward)
+
+L(copy_forward):
+#endif
+	cmp	$48, %ecx
+	jge	L(48bytesormore)
+
+L(fwd_write_less32bytes):
+#ifndef USE_AS_MEMMOVE
+	cmp	%dl, %al
+	jl	L(bk_write)
+#endif
+	add	%ecx, %edx
+	add	%ecx, %eax
+	BRANCH_TO_JMPTBL_ENTRY (L(table_48bytes_fwd), %ecx, 4)
+#ifndef USE_AS_MEMMOVE
+L(bk_write):
+	BRANCH_TO_JMPTBL_ENTRY (L(table_48_bytes_bwd), %ecx, 4)
+#endif
+
+	ALIGN (4)
+/* ECX > 32 and EDX is 4 byte aligned.  */
+L(48bytesormore):
+	movdqu	(%eax), %xmm0
+	PUSH (%edi)
+	movl	%edx, %edi
+	and	$-16, %edx
+	PUSH (%esi)
+	add	$16, %edx
+	movl	%edi, %esi
+	sub	%edx, %edi
+	add	%edi, %ecx
+	sub	%edi, %eax
+
+#ifdef SHARED_CACHE_SIZE_HALF
+	cmp	$SHARED_CACHE_SIZE_HALF, %ecx
+#else
+# ifdef SHARED
+	call	__i686.get_pc_thunk.bx
+	add	$_GLOBAL_OFFSET_TABLE_, %ebx
+	cmp	__x86_shared_cache_size_half@GOTOFF(%ebx), %ecx
+# else
+	cmp	__x86_shared_cache_size_half, %ecx
+# endif
+#endif
+
+	mov	%eax, %edi
+	jge	L(large_page)
+	and	$0xf, %edi
+	jz	L(shl_0)
+
+	BRANCH_TO_JMPTBL_ENTRY (L(shl_table), %edi, 4)
+
+	ALIGN (4)
+L(shl_0):
+	movdqu	%xmm0, (%esi)
+	xor	%edi, %edi
+	cmp	$127, %ecx
+	ja	L(shl_0_gobble)
+	lea	-32(%ecx), %ecx
+L(shl_0_loop):
+	movdqa	(%eax, %edi), %xmm0
+	movdqa	16(%eax, %edi), %xmm1
+	sub	$32, %ecx
+	movdqa	%xmm0, (%edx, %edi)
+	movdqa	%xmm1, 16(%edx, %edi)
+	lea	32(%edi), %edi
+	jl	L(shl_0_end)
+
+	movdqa	(%eax, %edi), %xmm0
+	movdqa	16(%eax, %edi), %xmm1
+	sub	$32, %ecx
+	movdqa	%xmm0, (%edx, %edi)
+	movdqa	%xmm1, 16(%edx, %edi)
+	lea	32(%edi), %edi
+	jl	L(shl_0_end)
+
+	movdqa	(%eax, %edi), %xmm0
+	movdqa	16(%eax, %edi), %xmm1
+	sub	$32, %ecx
+	movdqa	%xmm0, (%edx, %edi)
+	movdqa	%xmm1, 16(%edx, %edi)
+	lea	32(%edi), %edi
+	jl	L(shl_0_end)
+
+	movdqa	(%eax, %edi), %xmm0
+	movdqa	16(%eax, %edi), %xmm1
+	sub	$32, %ecx
+	movdqa	%xmm0, (%edx, %edi)
+	movdqa	%xmm1, 16(%edx, %edi)
+	lea	32(%edi), %edi
+L(shl_0_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	add	%edi, %eax
+	POP (%esi)
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY (L(table_48bytes_fwd), %ecx, 4)
+
+L(shl_0_gobble):
+
+#ifdef DATA_CACHE_SIZE_HALF
+	cmp	$DATA_CACHE_SIZE_HALF, %ecx
+#else
+# ifdef SHARED
+	call	__i686.get_pc_thunk.bx
+	add	$_GLOBAL_OFFSET_TABLE_, %ebx
+	mov	__x86_data_cache_size_half@GOTOFF(%ebx), %edi
+# else
+	mov	__x86_data_cache_size_half, %edi
+# endif
+#endif
+	mov	%edi, %esi
+	shr	$3, %esi
+	sub	%esi, %edi
+	cmp	%edi, %ecx
+	jge	L(shl_0_gobble_mem_start)
+	lea	-128(%ecx), %ecx
+	ALIGN (4)
+L(shl_0_gobble_cache_loop):
+	movdqa	(%eax), %xmm0
+	movaps	0x10(%eax), %xmm1
+	movaps	0x20(%eax), %xmm2
+	movaps	0x30(%eax), %xmm3
+	movaps	0x40(%eax), %xmm4
+	movaps	0x50(%eax), %xmm5
+	movaps	0x60(%eax), %xmm6
+	movaps	0x70(%eax), %xmm7
+	lea	0x80(%eax), %eax
+	sub	$128, %ecx
+	movdqa	%xmm0, (%edx)
+	movaps	%xmm1, 0x10(%edx)
+	movaps	%xmm2, 0x20(%edx)
+	movaps	%xmm3, 0x30(%edx)
+	movaps	%xmm4, 0x40(%edx)
+	movaps	%xmm5, 0x50(%edx)
+	movaps	%xmm6, 0x60(%edx)
+	movaps	%xmm7, 0x70(%edx)
+	lea	0x80(%edx), %edx
+
+	jge	L(shl_0_gobble_cache_loop)
+L(shl_0_gobble_cache_loop_tail):
+	cmp	$-0x40, %ecx
+	lea	0x80(%ecx), %ecx
+	jl	L(shl_0_cache_less_64bytes)
+
+	movdqa	(%eax), %xmm0
+	sub	$0x40, %ecx
+	movdqa	0x10(%eax), %xmm1
+
+	movdqa	%xmm0, (%edx)
+	movdqa	%xmm1, 0x10(%edx)
+
+	movdqa	0x20(%eax), %xmm0
+	movdqa	0x30(%eax), %xmm1
+	add	$0x40, %eax
+
+	movdqa	%xmm0, 0x20(%edx)
+	movdqa	%xmm1, 0x30(%edx)
+	add	$0x40, %edx
+L(shl_0_cache_less_64bytes):
+	cmp	$0x20, %ecx
+	jl	L(shl_0_cache_less_32bytes)
+	movdqa	(%eax), %xmm0
+	sub	$0x20, %ecx
+	movdqa	0x10(%eax), %xmm1
+	add	$0x20, %eax
+	movdqa	%xmm0, (%edx)
+	movdqa	%xmm1, 0x10(%edx)
+	add	$0x20, %edx
+L(shl_0_cache_less_32bytes):
+	cmp	$0x10, %ecx
+	jl	L(shl_0_cache_less_16bytes)
+	sub	$0x10, %ecx
+	movdqa	(%eax), %xmm0
+	add	$0x10, %eax
+	movdqa	%xmm0, (%edx)
+	add	$0x10, %edx
+L(shl_0_cache_less_16bytes):
+	add	%ecx, %edx
+	add	%ecx, %eax
+	POP (%esi)
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY (L(table_48bytes_fwd), %ecx, 4)
+
+
+	ALIGN (4)
+L(shl_0_gobble_mem_start):
+	cmp	%al, %dl
+	je	L(copy_page_by_rep)
+	lea	-128(%ecx), %ecx
+L(shl_0_gobble_mem_loop):
+	prefetchnta 0x1c0(%eax)
+	prefetchnta 0x280(%eax)
+	prefetchnta 0x1c0(%edx)
+	prefetchnta 0x280(%edx)
+
+	movdqa	(%eax), %xmm0
+	movaps	0x10(%eax), %xmm1
+	movaps	0x20(%eax), %xmm2
+	movaps	0x30(%eax), %xmm3
+	movaps	0x40(%eax), %xmm4
+	movaps	0x50(%eax), %xmm5
+	movaps	0x60(%eax), %xmm6
+	movaps	0x70(%eax), %xmm7
+	lea	0x80(%eax), %eax
+	sub	$0x80, %ecx
+	movdqa	%xmm0, (%edx)
+	movaps	%xmm1, 0x10(%edx)
+	movaps	%xmm2, 0x20(%edx)
+	movaps	%xmm3, 0x30(%edx)
+	movaps	%xmm4, 0x40(%edx)
+	movaps	%xmm5, 0x50(%edx)
+	movaps	%xmm6, 0x60(%edx)
+	movaps	%xmm7, 0x70(%edx)
+	lea	0x80(%edx), %edx
+
+	jge	L(shl_0_gobble_mem_loop)
+	cmp	$-0x40, %ecx
+	lea	0x80(%ecx), %ecx
+	jl	L(shl_0_mem_less_64bytes)
+
+	movdqa	(%eax), %xmm0
+	sub	$0x40, %ecx
+	movdqa	0x10(%eax), %xmm1
+
+	movdqa	%xmm0, (%edx)
+	movdqa	%xmm1, 0x10(%edx)
+
+	movdqa	0x20(%eax), %xmm0
+	movdqa	0x30(%eax), %xmm1
+	add	$0x40, %eax
+
+	movdqa	%xmm0, 0x20(%edx)
+	movdqa	%xmm1, 0x30(%edx)
+	add	$0x40, %edx
+L(shl_0_mem_less_64bytes):
+	cmp	$0x20, %ecx
+	jl	L(shl_0_mem_less_32bytes)
+	movdqa	(%eax), %xmm0
+	sub	$0x20, %ecx
+	movdqa	0x10(%eax), %xmm1
+	add	$0x20, %eax
+	movdqa	%xmm0, (%edx)
+	movdqa	%xmm1, 0x10(%edx)
+	add	$0x20, %edx
+L(shl_0_mem_less_32bytes):
+	cmp	$0x10, %ecx
+	jl	L(shl_0_mem_less_16bytes)
+	sub	$0x10, %ecx
+	movdqa	(%eax), %xmm0
+	add	$0x10, %eax
+	movdqa	%xmm0, (%edx)
+	add	$0x10, %edx
+L(shl_0_mem_less_16bytes):
+	add	%ecx, %edx
+	add	%ecx, %eax
+	POP (%esi)
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY (L(table_48bytes_fwd), %ecx, 4)
+
+
+	ALIGN (4)
+L(shl_1):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-1(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_1_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$1, %xmm2, %xmm3
+	palignr	$1, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_1_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$1, %xmm2, %xmm3
+	palignr	$1, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_1_loop)
+
+L(shl_1_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	1(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_2):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-2(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_2_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$2, %xmm2, %xmm3
+	palignr	$2, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_2_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$2, %xmm2, %xmm3
+	palignr	$2, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_2_loop)
+
+L(shl_2_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	2(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_3):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-3(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_3_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$3, %xmm2, %xmm3
+	palignr	$3, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_3_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$3, %xmm2, %xmm3
+	palignr	$3, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_3_loop)
+
+L(shl_3_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	3(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_4):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-4(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_4_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$4, %xmm2, %xmm3
+	palignr	$4, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_4_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$4, %xmm2, %xmm3
+	palignr	$4, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_4_loop)
+
+L(shl_4_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	4(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_5):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-5(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_5_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$5, %xmm2, %xmm3
+	palignr	$5, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_5_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$5, %xmm2, %xmm3
+	palignr	$5, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_5_loop)
+
+L(shl_5_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	5(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+
+	ALIGN (4)
+L(shl_6):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-6(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_6_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$6, %xmm2, %xmm3
+	palignr	$6, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_6_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$6, %xmm2, %xmm3
+	palignr	$6, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_6_loop)
+
+L(shl_6_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	6(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_7):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-7(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_7_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$7, %xmm2, %xmm3
+	palignr	$7, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_7_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$7, %xmm2, %xmm3
+	palignr	$7, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_7_loop)
+
+L(shl_7_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	7(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_8):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-8(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_8_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$8, %xmm2, %xmm3
+	palignr	$8, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_8_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$8, %xmm2, %xmm3
+	palignr	$8, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_8_loop)
+
+L(shl_8_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	8(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_9):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-9(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_9_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$9, %xmm2, %xmm3
+	palignr	$9, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_9_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$9, %xmm2, %xmm3
+	palignr	$9, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_9_loop)
+
+L(shl_9_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	9(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_10):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-10(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_10_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$10, %xmm2, %xmm3
+	palignr	$10, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_10_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$10, %xmm2, %xmm3
+	palignr	$10, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_10_loop)
+
+L(shl_10_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	10(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_11):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-11(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_11_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$11, %xmm2, %xmm3
+	palignr	$11, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_11_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$11, %xmm2, %xmm3
+	palignr	$11, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_11_loop)
+
+L(shl_11_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	11(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_12):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-12(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_12_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$12, %xmm2, %xmm3
+	palignr	$12, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_12_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$12, %xmm2, %xmm3
+	palignr	$12, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_12_loop)
+
+L(shl_12_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	12(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_13):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-13(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_13_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$13, %xmm2, %xmm3
+	palignr	$13, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_13_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$13, %xmm2, %xmm3
+	palignr	$13, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_13_loop)
+
+L(shl_13_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	13(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_14):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-14(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_14_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$14, %xmm2, %xmm3
+	palignr	$14, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_14_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$14, %xmm2, %xmm3
+	palignr	$14, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_14_loop)
+
+L(shl_14_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	14(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+
+	ALIGN (4)
+L(shl_15):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-15(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_15_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$15, %xmm2, %xmm3
+	palignr	$15, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_15_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$15, %xmm2, %xmm3
+	palignr	$15, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_15_loop)
+
+L(shl_15_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	15(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+
+	ALIGN (4)
+L(fwd_write_44bytes):
+	movl	-44(%eax), %ecx
+	movl	%ecx, -44(%edx)
+L(fwd_write_40bytes):
+	movl	-40(%eax), %ecx
+	movl	%ecx, -40(%edx)
+L(fwd_write_36bytes):
+	movl	-36(%eax), %ecx
+	movl	%ecx, -36(%edx)
+L(fwd_write_32bytes):
+	movl	-32(%eax), %ecx
+	movl	%ecx, -32(%edx)
+L(fwd_write_28bytes):
+	movl	-28(%eax), %ecx
+	movl	%ecx, -28(%edx)
+L(fwd_write_24bytes):
+	movl	-24(%eax), %ecx
+	movl	%ecx, -24(%edx)
+L(fwd_write_20bytes):
+	movl	-20(%eax), %ecx
+	movl	%ecx, -20(%edx)
+L(fwd_write_16bytes):
+	movl	-16(%eax), %ecx
+	movl	%ecx, -16(%edx)
+L(fwd_write_12bytes):
+	movl	-12(%eax), %ecx
+	movl	%ecx, -12(%edx)
+L(fwd_write_8bytes):
+	movl	-8(%eax), %ecx
+	movl	%ecx, -8(%edx)
+L(fwd_write_4bytes):
+	movl	-4(%eax), %ecx
+	movl	%ecx, -4(%edx)
+L(fwd_write_0bytes):
+#ifndef USE_AS_BCOPY
+# ifdef USE_AS_MEMPCPY
+	movl	%edx, %eax
+# else
+	movl	DEST(%esp), %eax
+# endif
+#endif
+	RETURN
+
+	ALIGN (4)
+L(fwd_write_5bytes):
+	movl	-5(%eax), %ecx
+	movl	-4(%eax), %eax
+	movl	%ecx, -5(%edx)
+	movl	%eax, -4(%edx)
+#ifndef USE_AS_BCOPY
+# ifdef USE_AS_MEMPCPY
+	movl	%edx, %eax
+# else
+	movl	DEST(%esp), %eax
+# endif
+#endif
+	RETURN
+
+	ALIGN (4)
+L(fwd_write_45bytes):
+	movl	-45(%eax), %ecx
+	movl	%ecx, -45(%edx)
+L(fwd_write_41bytes):
+	movl	-41(%eax), %ecx
+	movl	%ecx, -41(%edx)
+L(fwd_write_37bytes):
+	movl	-37(%eax), %ecx
+	movl	%ecx, -37(%edx)
+L(fwd_write_33bytes):
+	movl	-33(%eax), %ecx
+	movl	%ecx, -33(%edx)
+L(fwd_write_29bytes):
+	movl	-29(%eax), %ecx
+	movl	%ecx, -29(%edx)
+L(fwd_write_25bytes):
+	movl	-25(%eax), %ecx
+	movl	%ecx, -25(%edx)
+L(fwd_write_21bytes):
+	movl	-21(%eax), %ecx
+	movl	%ecx, -21(%edx)
+L(fwd_write_17bytes):
+	movl	-17(%eax), %ecx
+	movl	%ecx, -17(%edx)
+L(fwd_write_13bytes):
+	movl	-13(%eax), %ecx
+	movl	%ecx, -13(%edx)
+L(fwd_write_9bytes):
+	movl	-9(%eax), %ecx
+	movl	%ecx, -9(%edx)
+	movl	-5(%eax), %ecx
+	movl	%ecx, -5(%edx)
+L(fwd_write_1bytes):
+	movzbl	-1(%eax), %ecx
+	movb	%cl, -1(%edx)
+#ifndef USE_AS_BCOPY
+# ifdef USE_AS_MEMPCPY
+	movl	%edx, %eax
+# else
+	movl	DEST(%esp), %eax
+# endif
+#endif
+	RETURN
+
+	ALIGN (4)
+L(fwd_write_46bytes):
+	movl	-46(%eax), %ecx
+	movl	%ecx, -46(%edx)
+L(fwd_write_42bytes):
+	movl	-42(%eax), %ecx
+	movl	%ecx, -42(%edx)
+L(fwd_write_38bytes):
+	movl	-38(%eax), %ecx
+	movl	%ecx, -38(%edx)
+L(fwd_write_34bytes):
+	movl	-34(%eax), %ecx
+	movl	%ecx, -34(%edx)
+L(fwd_write_30bytes):
+	movl	-30(%eax), %ecx
+	movl	%ecx, -30(%edx)
+L(fwd_write_26bytes):
+	movl	-26(%eax), %ecx
+	movl	%ecx, -26(%edx)
+L(fwd_write_22bytes):
+	movl	-22(%eax), %ecx
+	movl	%ecx, -22(%edx)
+L(fwd_write_18bytes):
+	movl	-18(%eax), %ecx
+	movl	%ecx, -18(%edx)
+L(fwd_write_14bytes):
+	movl	-14(%eax), %ecx
+	movl	%ecx, -14(%edx)
+L(fwd_write_10bytes):
+	movl	-10(%eax), %ecx
+	movl	%ecx, -10(%edx)
+L(fwd_write_6bytes):
+	movl	-6(%eax), %ecx
+	movl	%ecx, -6(%edx)
+L(fwd_write_2bytes):
+	movzwl	-2(%eax), %ecx
+	movw	%cx, -2(%edx)
+#ifndef USE_AS_BCOPY
+# ifdef USE_AS_MEMPCPY
+	movl	%edx, %eax
+# else
+	movl	DEST(%esp), %eax
+# endif
+#endif
+	RETURN
+
+	ALIGN (4)
+L(fwd_write_47bytes):
+	movl	-47(%eax), %ecx
+	movl	%ecx, -47(%edx)
+L(fwd_write_43bytes):
+	movl	-43(%eax), %ecx
+	movl	%ecx, -43(%edx)
+L(fwd_write_39bytes):
+	movl	-39(%eax), %ecx
+	movl	%ecx, -39(%edx)
+L(fwd_write_35bytes):
+	movl	-35(%eax), %ecx
+	movl	%ecx, -35(%edx)
+L(fwd_write_31bytes):
+	movl	-31(%eax), %ecx
+	movl	%ecx, -31(%edx)
+L(fwd_write_27bytes):
+	movl	-27(%eax), %ecx
+	movl	%ecx, -27(%edx)
+L(fwd_write_23bytes):
+	movl	-23(%eax), %ecx
+	movl	%ecx, -23(%edx)
+L(fwd_write_19bytes):
+	movl	-19(%eax), %ecx
+	movl	%ecx, -19(%edx)
+L(fwd_write_15bytes):
+	movl	-15(%eax), %ecx
+	movl	%ecx, -15(%edx)
+L(fwd_write_11bytes):
+	movl	-11(%eax), %ecx
+	movl	%ecx, -11(%edx)
+L(fwd_write_7bytes):
+	movl	-7(%eax), %ecx
+	movl	%ecx, -7(%edx)
+L(fwd_write_3bytes):
+	movzwl	-3(%eax), %ecx
+	movzbl	-1(%eax), %eax
+	movw	%cx, -3(%edx)
+	movb	%al, -1(%edx)
+#ifndef USE_AS_BCOPY
+# ifdef USE_AS_MEMPCPY
+	movl	%edx, %eax
+# else
+	movl	DEST(%esp), %eax
+# endif
+#endif
+	RETURN
+
+	ALIGN (4)
+L(large_page):
+	movdqu	(%eax), %xmm1
+	lea	16(%eax), %eax
+	movdqu	%xmm0, (%esi)
+	movntdq	%xmm1, (%edx)
+	lea	16(%edx), %edx
+	cmp	%al, %dl
+	je	L(copy_page_by_rep)
+L(large_page_loop_init):
+	POP (%esi)
+	lea	-0x90(%ecx), %ecx
+	POP (%edi)
+L(large_page_loop):
+	prefetchnta	0x1c0(%eax)
+	prefetchnta	0x280(%eax)
+	movdqu	(%eax), %xmm0
+	movdqu	0x10(%eax), %xmm1
+	movdqu	0x20(%eax), %xmm2
+	movdqu	0x30(%eax), %xmm3
+	movdqu	0x40(%eax), %xmm4
+	movdqu	0x50(%eax), %xmm5
+	movdqu	0x60(%eax), %xmm6
+	movdqu	0x70(%eax), %xmm7
+	lea	0x80(%eax), %eax
+	lfence	
+	sub	$0x80, %ecx
+	movntdq	%xmm0, (%edx)
+	movntdq	%xmm1, 0x10(%edx)
+	movntdq	%xmm2, 0x20(%edx)
+	movntdq	%xmm3, 0x30(%edx)
+	movntdq	%xmm4, 0x40(%edx)
+	movntdq	%xmm5, 0x50(%edx)
+	movntdq	%xmm6, 0x60(%edx)
+	movntdq	%xmm7, 0x70(%edx)
+	lea	0x80(%edx), %edx
+	jae	L(large_page_loop)
+	cmp	$-0x40, %ecx
+	lea	0x80(%ecx), %ecx
+	jl	L(large_page_less_64bytes)
+
+	movdqu	(%eax), %xmm0
+	movdqu	0x10(%eax), %xmm1
+	movdqu	0x20(%eax), %xmm2
+	movdqu	0x30(%eax), %xmm3
+	lea	0x40(%eax), %eax
+
+	movntdq	%xmm0, (%edx)
+	movntdq	%xmm1, 0x10(%edx)
+	movntdq	%xmm2, 0x20(%edx)
+	movntdq	%xmm3, 0x30(%edx)
+	lea	0x40(%edx), %edx
+	sub	$0x40, %ecx
+L(large_page_less_64bytes):
+	cmp	$32, %ecx
+	jl	L(large_page_less_32bytes)
+	movdqu	(%eax), %xmm0
+	movdqu	0x10(%eax), %xmm1
+	lea	0x20(%eax), %eax
+	movntdq	%xmm0, (%edx)
+	movntdq	%xmm1, 0x10(%edx)
+	lea	0x20(%edx), %edx
+	sub	$0x20, %ecx
+L(large_page_less_32bytes):
+	add	%ecx, %edx
+	add	%ecx, %eax
+	sfence
+	BRANCH_TO_JMPTBL_ENTRY (L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(copy_page_by_rep):
+	mov	%eax, %esi
+	mov	%edx, %edi
+	mov	%ecx, %edx
+	shr	$2, %ecx
+	and	$3, %edx
+	rep	movsl
+	jz	L(copy_page_by_rep_exit)
+	cmp	$2, %edx
+	jb	L(copy_page_by_rep_left_1)
+	movzwl	(%esi), %eax
+	movw	%ax, (%edi)
+	add	$2, %esi 
+	add	$2, %edi 
+	sub	$2, %edx
+	jz	L(copy_page_by_rep_exit)
+L(copy_page_by_rep_left_1):
+	movzbl	(%esi), %eax
+	movb	%al, (%edi)
+L(copy_page_by_rep_exit):
+	POP (%esi)
+	POP (%edi)
+#ifndef USE_AS_BCOPY
+	movl	DEST(%esp), %eax
+# ifdef USE_AS_MEMPCPY
+	movl	LEN(%esp), %ecx
+	add	%ecx, %eax
+# endif
+#endif
+	RETURN
+
+	ALIGN (4)
+L(bk_write_44bytes):
+	movl	40(%eax), %ecx
+	movl	%ecx, 40(%edx)
+L(bk_write_40bytes):
+	movl	36(%eax), %ecx
+	movl	%ecx, 36(%edx)
+L(bk_write_36bytes):
+	movl	32(%eax), %ecx
+	movl	%ecx, 32(%edx)
+L(bk_write_32bytes):
+	movl	28(%eax), %ecx
+	movl	%ecx, 28(%edx)
+L(bk_write_28bytes):
+	movl	24(%eax), %ecx
+	movl	%ecx, 24(%edx)
+L(bk_write_24bytes):
+	movl	20(%eax), %ecx
+	movl	%ecx, 20(%edx)
+L(bk_write_20bytes):
+	movl	16(%eax), %ecx
+	movl	%ecx, 16(%edx)
+L(bk_write_16bytes):
+	movl	12(%eax), %ecx
+	movl	%ecx, 12(%edx)
+L(bk_write_12bytes):
+	movl	8(%eax), %ecx
+	movl	%ecx, 8(%edx)
+L(bk_write_8bytes):
+	movl	4(%eax), %ecx
+	movl	%ecx, 4(%edx)
+L(bk_write_4bytes):
+	movl	(%eax), %ecx
+	movl	%ecx, (%edx)
+L(bk_write_0bytes):
+#ifndef USE_AS_BCOPY
+	movl	DEST(%esp), %eax
+# ifdef USE_AS_MEMPCPY
+	movl	LEN(%esp), %ecx
+	add	%ecx, %eax
+# endif
+#endif
+	RETURN
+
+	ALIGN (4)
+L(bk_write_45bytes):
+	movl	41(%eax), %ecx
+	movl	%ecx, 41(%edx)
+L(bk_write_41bytes):
+	movl	37(%eax), %ecx
+	movl	%ecx, 37(%edx)
+L(bk_write_37bytes):
+	movl	33(%eax), %ecx
+	movl	%ecx, 33(%edx)
+L(bk_write_33bytes):
+	movl	29(%eax), %ecx
+	movl	%ecx, 29(%edx)
+L(bk_write_29bytes):
+	movl	25(%eax), %ecx
+	movl	%ecx, 25(%edx)
+L(bk_write_25bytes):
+	movl	21(%eax), %ecx
+	movl	%ecx, 21(%edx)
+L(bk_write_21bytes):
+	movl	17(%eax), %ecx
+	movl	%ecx, 17(%edx)
+L(bk_write_17bytes):
+	movl	13(%eax), %ecx
+	movl	%ecx, 13(%edx)
+L(bk_write_13bytes):
+	movl	9(%eax), %ecx
+	movl	%ecx, 9(%edx)
+L(bk_write_9bytes):
+	movl	5(%eax), %ecx
+	movl	%ecx, 5(%edx)
+L(bk_write_5bytes):
+	movl	1(%eax), %ecx
+	movl	%ecx, 1(%edx)
+L(bk_write_1bytes):
+	movzbl	(%eax), %ecx
+	movb	%cl, (%edx)
+#ifndef USE_AS_BCOPY
+	movl	DEST(%esp), %eax
+# ifdef USE_AS_MEMPCPY
+	movl	LEN(%esp), %ecx
+	add	%ecx, %eax
+# endif
+#endif
+	RETURN
+
+	ALIGN (4)
+L(bk_write_46bytes):
+	movl	42(%eax), %ecx
+	movl	%ecx, 42(%edx)
+L(bk_write_42bytes):
+	movl	38(%eax), %ecx
+	movl	%ecx, 38(%edx)
+L(bk_write_38bytes):
+	movl	34(%eax), %ecx
+	movl	%ecx, 34(%edx)
+L(bk_write_34bytes):
+	movl	30(%eax), %ecx
+	movl	%ecx, 30(%edx)
+L(bk_write_30bytes):
+	movl	26(%eax), %ecx
+	movl	%ecx, 26(%edx)
+L(bk_write_26bytes):
+	movl	22(%eax), %ecx
+	movl	%ecx, 22(%edx)
+L(bk_write_22bytes):
+	movl	18(%eax), %ecx
+	movl	%ecx, 18(%edx)
+L(bk_write_18bytes):
+	movl	14(%eax), %ecx
+	movl	%ecx, 14(%edx)
+L(bk_write_14bytes):
+	movl	10(%eax), %ecx
+	movl	%ecx, 10(%edx)
+L(bk_write_10bytes):
+	movl	6(%eax), %ecx
+	movl	%ecx, 6(%edx)
+L(bk_write_6bytes):
+	movl	2(%eax), %ecx
+	movl	%ecx, 2(%edx)
+L(bk_write_2bytes):
+	movzwl	(%eax), %ecx
+	movw	%cx, (%edx)
+#ifndef USE_AS_BCOPY
+	movl	DEST(%esp), %eax
+# ifdef USE_AS_MEMPCPY
+	movl	LEN(%esp), %ecx
+	add	%ecx, %eax
+# endif
+#endif
+	RETURN
+
+	ALIGN (4)
+L(bk_write_47bytes):
+	movl	43(%eax), %ecx
+	movl	%ecx, 43(%edx)
+L(bk_write_43bytes):
+	movl	39(%eax), %ecx
+	movl	%ecx, 39(%edx)
+L(bk_write_39bytes):
+	movl	35(%eax), %ecx
+	movl	%ecx, 35(%edx)
+L(bk_write_35bytes):
+	movl	31(%eax), %ecx
+	movl	%ecx, 31(%edx)
+L(bk_write_31bytes):
+	movl	27(%eax), %ecx
+	movl	%ecx, 27(%edx)
+L(bk_write_27bytes):
+	movl	23(%eax), %ecx
+	movl	%ecx, 23(%edx)
+L(bk_write_23bytes):
+	movl	19(%eax), %ecx
+	movl	%ecx, 19(%edx)
+L(bk_write_19bytes):
+	movl	15(%eax), %ecx
+	movl	%ecx, 15(%edx)
+L(bk_write_15bytes):
+	movl	11(%eax), %ecx
+	movl	%ecx, 11(%edx)
+L(bk_write_11bytes):
+	movl	7(%eax), %ecx
+	movl	%ecx, 7(%edx)
+L(bk_write_7bytes):
+	movl	3(%eax), %ecx
+	movl	%ecx, 3(%edx)
+L(bk_write_3bytes):
+	movzwl	1(%eax), %ecx
+	movw	%cx, 1(%edx)
+	movzbl	(%eax), %eax
+	movb	%al, (%edx)
+#ifndef USE_AS_BCOPY
+	movl	DEST(%esp), %eax
+# ifdef USE_AS_MEMPCPY
+	movl	LEN(%esp), %ecx
+	add	%ecx, %eax
+# endif
+#endif
+	RETURN_END
+
+
+	.pushsection .rodata.ssse3,"a",@progbits
+	ALIGN (2)
+L(table_48bytes_fwd):
+	.int	JMPTBL (L(fwd_write_0bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_1bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_2bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_3bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_4bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_5bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_6bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_7bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_8bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_9bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_10bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_11bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_12bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_13bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_14bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_15bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_16bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_17bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_18bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_19bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_20bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_21bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_22bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_23bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_24bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_25bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_26bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_27bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_28bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_29bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_30bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_31bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_32bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_33bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_34bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_35bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_36bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_37bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_38bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_39bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_40bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_41bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_42bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_43bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_44bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_45bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_46bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_47bytes), L(table_48bytes_fwd))
+
+	ALIGN (2)
+L(shl_table):
+	.int	JMPTBL (L(shl_0), L(shl_table))
+	.int	JMPTBL (L(shl_1), L(shl_table))
+	.int	JMPTBL (L(shl_2), L(shl_table))
+	.int	JMPTBL (L(shl_3), L(shl_table))
+	.int	JMPTBL (L(shl_4), L(shl_table))
+	.int	JMPTBL (L(shl_5), L(shl_table))
+	.int	JMPTBL (L(shl_6), L(shl_table))
+	.int	JMPTBL (L(shl_7), L(shl_table))
+	.int	JMPTBL (L(shl_8), L(shl_table))
+	.int	JMPTBL (L(shl_9), L(shl_table))
+	.int	JMPTBL (L(shl_10), L(shl_table))
+	.int	JMPTBL (L(shl_11), L(shl_table))
+	.int	JMPTBL (L(shl_12), L(shl_table))
+	.int	JMPTBL (L(shl_13), L(shl_table))
+	.int	JMPTBL (L(shl_14), L(shl_table))
+	.int	JMPTBL (L(shl_15), L(shl_table))
+
+	ALIGN (2)
+L(table_48_bytes_bwd):
+	.int	JMPTBL (L(bk_write_0bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_1bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_2bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_3bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_4bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_5bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_6bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_7bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_8bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_9bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_10bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_11bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_12bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_13bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_14bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_15bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_16bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_17bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_18bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_19bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_20bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_21bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_22bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_23bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_24bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_25bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_26bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_27bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_28bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_29bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_30bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_31bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_32bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_33bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_34bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_35bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_36bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_37bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_38bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_39bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_40bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_41bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_42bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_43bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_44bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_45bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_46bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_47bytes), L(table_48_bytes_bwd))
+
+	.popsection
+
+#ifdef USE_AS_MEMMOVE
+	ALIGN (4)
+L(copy_backward):
+	PUSH (%esi)
+	movl	%eax, %esi
+	lea	(%ecx,%edx,1),%edx
+	lea	(%ecx,%esi,1),%esi
+	testl	$0x3, %edx
+	jnz	L(bk_align)
+
+L(bk_aligned_4):
+	cmp	$64, %ecx
+	jge	L(bk_write_more64bytes)
+
+L(bk_write_64bytesless):
+	cmp	$32, %ecx
+	jl	L(bk_write_less32bytes)
+
+L(bk_write_more32bytes):
+	/* Copy 32 bytes at a time.  */
+	sub	$32, %ecx
+	movl	-4(%esi), %eax
+	movl	%eax, -4(%edx)
+	movl	-8(%esi), %eax
+	movl	%eax, -8(%edx)
+	movl	-12(%esi), %eax
+	movl	%eax, -12(%edx)
+	movl	-16(%esi), %eax
+	movl	%eax, -16(%edx)
+	movl	-20(%esi), %eax
+	movl	%eax, -20(%edx)
+	movl	-24(%esi), %eax
+	movl	%eax, -24(%edx)
+	movl	-28(%esi), %eax
+	movl	%eax, -28(%edx)
+	movl	-32(%esi), %eax
+	movl	%eax, -32(%edx)
+	sub	$32, %edx
+	sub	$32, %esi
+
+L(bk_write_less32bytes):
+	movl	%esi, %eax
+	sub	%ecx, %edx
+	sub	%ecx, %eax
+	POP (%esi)
+L(bk_write_less32bytes_2):
+	BRANCH_TO_JMPTBL_ENTRY (L(table_48_bytes_bwd), %ecx, 4)
+
+	ALIGN (4)
+L(bk_align):
+	cmp	$8, %ecx
+	jle	L(bk_write_less32bytes)
+	testl	$1, %edx
+	/* We get here only if (EDX & 3 ) != 0 so if (EDX & 1) ==0,
+	   then (EDX & 2) must be != 0.  */
+	jz	L(bk_got2)
+	sub	$1, %esi
+	sub	$1, %ecx
+	sub	$1, %edx
+	movzbl	(%esi), %eax
+	movb	%al, (%edx)
+
+	testl	$2, %edx
+	jz	L(bk_aligned_4)
+
+L(bk_got2):
+	sub	$2, %esi
+	sub	$2, %ecx
+	sub	$2, %edx
+	movzwl	(%esi), %eax
+	movw	%ax, (%edx)
+	jmp	L(bk_aligned_4)
+
+	ALIGN (4)
+L(bk_write_more64bytes):
+	/* Check alignment of last byte.  */
+	testl	$15, %edx
+	jz	L(bk_ssse3_cpy_pre)
+
+/* EDX is aligned 4 bytes, but not 16 bytes.  */
+L(bk_ssse3_align):
+	sub	$4, %esi
+	sub	$4, %ecx
+	sub	$4, %edx
+	movl	(%esi), %eax
+	movl	%eax, (%edx)
+
+	testl	$15, %edx
+	jz	L(bk_ssse3_cpy_pre)
+
+	sub	$4, %esi
+	sub	$4, %ecx
+	sub	$4, %edx
+	movl	(%esi), %eax
+	movl	%eax, (%edx)
+
+	testl	$15, %edx
+	jz	L(bk_ssse3_cpy_pre)
+
+	sub	$4, %esi
+	sub	$4, %ecx
+	sub	$4, %edx
+	movl	(%esi), %eax
+	movl	%eax, (%edx)
+
+L(bk_ssse3_cpy_pre):
+	cmp	$64, %ecx
+	jl	L(bk_write_more32bytes)
+
+L(bk_ssse3_cpy):
+	sub	$64, %esi
+	sub	$64, %ecx
+	sub	$64, %edx
+	movdqu	0x30(%esi), %xmm3
+	movdqa	%xmm3, 0x30(%edx)
+	movdqu	0x20(%esi), %xmm2
+	movdqa	%xmm2, 0x20(%edx)
+	movdqu	0x10(%esi), %xmm1
+	movdqa	%xmm1, 0x10(%edx)
+	movdqu	(%esi), %xmm0
+	movdqa	%xmm0, (%edx)
+	cmp	$64, %ecx
+	jge	L(bk_ssse3_cpy)
+	jmp	L(bk_write_64bytesless)
+
+#endif
+
+END (MEMCPY)
+
+#endif
diff --git a/sysdeps/i386/i686/multiarch/memcpy-ssse3.S b/sysdeps/i386/i686/multiarch/memcpy-ssse3.S
new file mode 100644
index 0000000000..3a10de6807
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/memcpy-ssse3.S
@@ -0,0 +1,1737 @@
+/* memcpy with SSSE3
+   Copyright (C) 2010 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>
+
+#if !defined NOT_IN_libc \
+    && (defined SHARED \
+        || defined USE_AS_MEMMOVE \
+	|| !defined USE_MULTIARCH)
+
+#include "asm-syntax.h"
+
+#ifndef MEMCPY
+# define MEMCPY		__memcpy_ssse3
+# define MEMCPY_CHK	__memcpy_chk_ssse3
+#endif
+
+#ifdef USE_AS_BCOPY
+# define SRC		PARMS
+# define DEST		SRC+4
+# define LEN		DEST+4
+#else
+# define DEST		PARMS
+# define SRC		DEST+4
+# define LEN		SRC+4
+#endif
+
+#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)
+
+#ifdef SHARED
+# define PARMS		8		/* Preserve EBX.  */
+# define ENTRANCE	PUSH (%ebx);
+# define RETURN_END	POP (%ebx); ret
+# define RETURN		RETURN_END; CFI_PUSH (%ebx)
+# define JMPTBL(I, B)	I - B
+
+/* Load an entry in a jump table into EBX and branch to it.  TABLE is a
+   jump table with relative offsets.  INDEX is a register contains the
+   index into the jump table.   SCALE is the scale of INDEX. */
+# define BRANCH_TO_JMPTBL_ENTRY(TABLE, INDEX, SCALE)		\
+    /* We first load PC into EBX.  */				\
+    call	__i686.get_pc_thunk.bx;				\
+    /* Get the address of the jump table.  */			\
+    addl	$(TABLE - .), %ebx;				\
+    /* Get the entry and convert the relative offset to the	\
+       absolute address.  */					\
+    addl	(%ebx,INDEX,SCALE), %ebx;			\
+    /* We loaded the jump table.  Go.  */			\
+    jmp		*%ebx
+
+# define BRANCH_TO_JMPTBL_ENTRY_VALUE(TABLE)			\
+    addl	$(TABLE - .), %ebx	
+			
+# define BRANCH_TO_JMPTBL_ENTRY_TAIL(TABLE, INDEX, SCALE)	\
+    addl	(%ebx,INDEX,SCALE), %ebx;			\
+    /* We loaded the jump table.  Go.  */			\
+    jmp		*%ebx
+
+	.section	.gnu.linkonce.t.__i686.get_pc_thunk.bx,"ax",@progbits
+	.globl	__i686.get_pc_thunk.bx
+	.hidden	__i686.get_pc_thunk.bx
+	ALIGN (4)
+	.type	__i686.get_pc_thunk.bx,@function
+__i686.get_pc_thunk.bx:
+	movl	(%esp), %ebx
+	ret
+#else
+# define PARMS		4
+# define ENTRANCE
+# define RETURN_END	ret
+# define RETURN		RETURN_END
+# define JMPTBL(I, B)	I
+
+/* Branch to an entry in a jump table.  TABLE is a jump table with
+   absolute offsets.  INDEX is a register contains the index into the
+   jump table.  SCALE is the scale of INDEX. */
+# define BRANCH_TO_JMPTBL_ENTRY(TABLE, INDEX, SCALE)		\
+    jmp		*TABLE(,INDEX,SCALE)
+
+# define BRANCH_TO_JMPTBL_ENTRY_VALUE(TABLE)			
+
+# define BRANCH_TO_JMPTBL_ENTRY_TAIL(TABLE, INDEX, SCALE)		\
+    jmp		*TABLE(,INDEX,SCALE)
+#endif
+
+	.section .text.ssse3,"ax",@progbits
+#if defined SHARED && !defined NOT_IN_libc && !defined USE_AS_BCOPY
+ENTRY (MEMCPY_CHK)
+	movl	12(%esp), %eax
+	cmpl	%eax, 16(%esp)
+	jb	HIDDEN_JUMPTARGET (__chk_fail)
+END (MEMCPY_CHK)
+#endif
+ENTRY (MEMCPY)
+	ENTRANCE
+	movl	LEN(%esp), %ecx
+	movl	SRC(%esp), %eax
+	movl	DEST(%esp), %edx
+
+#ifdef USE_AS_MEMMOVE
+	cmp	%eax, %edx
+	jb	L(copy_forward)
+	je	L(fwd_write_0bytes)
+	cmp	$32, %ecx
+	jge	L(memmove_bwd)
+	jmp	L(bk_write_less32bytes_2)
+L(memmove_bwd):
+	add	%ecx, %eax
+	cmp	%eax, %edx
+	movl	SRC(%esp), %eax
+	jb	L(copy_backward)
+
+L(copy_forward):
+#endif
+	cmp	$48, %ecx
+	jge	L(48bytesormore)
+
+L(fwd_write_less32bytes):
+#ifndef USE_AS_MEMMOVE
+	cmp	%dl, %al
+	jl	L(bk_write)
+#endif
+	add	%ecx, %edx
+	add	%ecx, %eax
+	BRANCH_TO_JMPTBL_ENTRY (L(table_48bytes_fwd), %ecx, 4)
+#ifndef USE_AS_MEMMOVE
+L(bk_write):
+	BRANCH_TO_JMPTBL_ENTRY (L(table_48_bytes_bwd), %ecx, 4)
+#endif
+
+	ALIGN (4)
+/* ECX > 32 and EDX is 4 byte aligned.  */
+L(48bytesormore):
+	movdqu	(%eax), %xmm0
+	PUSH (%edi)
+	movl	%edx, %edi
+	and	$-16, %edx
+	PUSH (%esi)
+	add	$16, %edx
+	movl	%edi, %esi
+	sub	%edx, %edi
+	add	%edi, %ecx
+	sub	%edi, %eax
+
+#ifdef SHARED_CACHE_SIZE_HALF
+	cmp	$SHARED_CACHE_SIZE_HALF, %ecx
+#else
+# ifdef SHARED
+	call	__i686.get_pc_thunk.bx
+	add	$_GLOBAL_OFFSET_TABLE_, %ebx
+	cmp	__x86_shared_cache_size_half@GOTOFF(%ebx), %ecx
+# else
+	cmp	__x86_shared_cache_size_half, %ecx
+# endif
+#endif
+
+	mov	%eax, %edi
+	jge	L(large_page)
+	and	$0xf, %edi
+	jz	L(shl_0)
+
+	BRANCH_TO_JMPTBL_ENTRY (L(shl_table), %edi, 4)
+
+	ALIGN (4)
+L(shl_0):
+	movdqu	%xmm0, (%esi)
+	xor	%edi, %edi
+	POP (%esi)
+	cmp	$127, %ecx
+	ja	L(shl_0_gobble)
+	lea	-32(%ecx), %ecx
+L(shl_0_loop):
+	movdqa	(%eax, %edi), %xmm0
+	movdqa	16(%eax, %edi), %xmm1
+	sub	$32, %ecx
+	movdqa	%xmm0, (%edx, %edi)
+	movdqa	%xmm1, 16(%edx, %edi)
+	lea	32(%edi), %edi
+	jl	L(shl_0_end)
+
+	movdqa	(%eax, %edi), %xmm0
+	movdqa	16(%eax, %edi), %xmm1
+	sub	$32, %ecx
+	movdqa	%xmm0, (%edx, %edi)
+	movdqa	%xmm1, 16(%edx, %edi)
+	lea	32(%edi), %edi
+	jl	L(shl_0_end)
+
+	movdqa	(%eax, %edi), %xmm0
+	movdqa	16(%eax, %edi), %xmm1
+	sub	$32, %ecx
+	movdqa	%xmm0, (%edx, %edi)
+	movdqa	%xmm1, 16(%edx, %edi)
+	lea	32(%edi), %edi
+	jl	L(shl_0_end)
+
+	movdqa	(%eax, %edi), %xmm0
+	movdqa	16(%eax, %edi), %xmm1
+	sub	$32, %ecx
+	movdqa	%xmm0, (%edx, %edi)
+	movdqa	%xmm1, 16(%edx, %edi)
+	lea	32(%edi), %edi
+L(shl_0_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	add	%edi, %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY (L(table_48bytes_fwd), %ecx, 4)
+
+L(shl_0_gobble):
+
+#ifdef DATA_CACHE_SIZE_HALF
+	cmp	$DATA_CACHE_SIZE_HALF, %ecx
+#else
+# ifdef SHARED
+	call	__i686.get_pc_thunk.bx
+	add	$_GLOBAL_OFFSET_TABLE_, %ebx
+	cmp	__x86_data_cache_size_half@GOTOFF(%ebx), %ecx
+# else
+	cmp	__x86_data_cache_size_half, %ecx
+# endif
+#endif
+
+	POP (%edi)
+	lea	-128(%ecx), %ecx
+	jge	L(shl_0_gobble_mem_loop)
+L(shl_0_gobble_cache_loop):
+	movdqa	(%eax), %xmm0
+	movdqa	0x10(%eax), %xmm1
+	movdqa	0x20(%eax), %xmm2
+	movdqa	0x30(%eax), %xmm3
+	movdqa	0x40(%eax), %xmm4
+	movdqa	0x50(%eax), %xmm5
+	movdqa	0x60(%eax), %xmm6
+	movdqa	0x70(%eax), %xmm7
+	lea	0x80(%eax), %eax
+	sub	$128, %ecx
+	movdqa	%xmm0, (%edx)
+	movdqa	%xmm1, 0x10(%edx)
+	movdqa	%xmm2, 0x20(%edx)
+	movdqa	%xmm3, 0x30(%edx)
+	movdqa	%xmm4, 0x40(%edx)
+	movdqa	%xmm5, 0x50(%edx)
+	movdqa	%xmm6, 0x60(%edx)
+	movdqa	%xmm7, 0x70(%edx)
+	lea	0x80(%edx), %edx
+
+	jge	L(shl_0_gobble_cache_loop)
+L(shl_0_gobble_cache_loop_tail):
+	cmp	$-0x40, %ecx
+	lea	0x80(%ecx), %ecx
+	jl	L(shl_0_cache_less_64bytes)
+
+	movdqa	(%eax), %xmm0
+	sub	$0x40, %ecx
+	movdqa	0x10(%eax), %xmm1
+
+	movdqa	%xmm0, (%edx)
+	movdqa	%xmm1, 0x10(%edx)
+
+	movdqa	0x20(%eax), %xmm0
+	movdqa	0x30(%eax), %xmm1
+	add	$0x40, %eax
+
+	movdqa	%xmm0, 0x20(%edx)
+	movdqa	%xmm1, 0x30(%edx)
+	add	$0x40, %edx
+L(shl_0_cache_less_64bytes):
+	cmp	$0x20, %ecx
+	jl	L(shl_0_cache_less_32bytes)
+	movdqa	(%eax), %xmm0
+	sub	$0x20, %ecx
+	movdqa	0x10(%eax), %xmm1
+	add	$0x20, %eax
+	movdqa	%xmm0, (%edx)
+	movdqa	%xmm1, 0x10(%edx)
+	add	$0x20, %edx
+L(shl_0_cache_less_32bytes):
+	cmp	$0x10, %ecx
+	jl	L(shl_0_cache_less_16bytes)
+	sub	$0x10, %ecx
+	movdqa	(%eax), %xmm0
+	add	$0x10, %eax
+	movdqa	%xmm0, (%edx)
+	add	$0x10, %edx
+L(shl_0_cache_less_16bytes):
+	add	%ecx, %edx
+	add	%ecx, %eax
+	BRANCH_TO_JMPTBL_ENTRY (L(table_48bytes_fwd), %ecx, 4)
+
+
+	ALIGN (4)
+L(shl_0_gobble_mem_loop):
+	prefetcht0 0x1c0(%eax)
+	prefetcht0 0x280(%eax)
+	prefetcht0 0x1c0(%edx)
+
+	movdqa	(%eax), %xmm0
+	movdqa	0x10(%eax), %xmm1
+	movdqa	0x20(%eax), %xmm2
+	movdqa	0x30(%eax), %xmm3
+	movdqa	0x40(%eax), %xmm4
+	movdqa	0x50(%eax), %xmm5
+	movdqa	0x60(%eax), %xmm6
+	movdqa	0x70(%eax), %xmm7
+	lea	0x80(%eax), %eax
+	sub	$0x80, %ecx
+	movdqa	%xmm0, (%edx)
+	movdqa	%xmm1, 0x10(%edx)
+	movdqa	%xmm2, 0x20(%edx)
+	movdqa	%xmm3, 0x30(%edx)
+	movdqa	%xmm4, 0x40(%edx)
+	movdqa	%xmm5, 0x50(%edx)
+	movdqa	%xmm6, 0x60(%edx)
+	movdqa	%xmm7, 0x70(%edx)
+	lea	0x80(%edx), %edx
+
+	jge	L(shl_0_gobble_mem_loop)
+	cmp	$-0x40, %ecx
+	lea	0x80(%ecx), %ecx
+	jl	L(shl_0_mem_less_64bytes)
+
+	movdqa	(%eax), %xmm0
+	sub	$0x40, %ecx
+	movdqa	0x10(%eax), %xmm1
+
+	movdqa	%xmm0, (%edx)
+	movdqa	%xmm1, 0x10(%edx)
+
+	movdqa	0x20(%eax), %xmm0
+	movdqa	0x30(%eax), %xmm1
+	add	$0x40, %eax
+
+	movdqa	%xmm0, 0x20(%edx)
+	movdqa	%xmm1, 0x30(%edx)
+	add	$0x40, %edx
+L(shl_0_mem_less_64bytes):
+	cmp	$0x20, %ecx
+	jl	L(shl_0_mem_less_32bytes)
+	movdqa	(%eax), %xmm0
+	sub	$0x20, %ecx
+	movdqa	0x10(%eax), %xmm1
+	add	$0x20, %eax
+	movdqa	%xmm0, (%edx)
+	movdqa	%xmm1, 0x10(%edx)
+	add	$0x20, %edx
+L(shl_0_mem_less_32bytes):
+	cmp	$0x10, %ecx
+	jl	L(shl_0_mem_less_16bytes)
+	sub	$0x10, %ecx
+	movdqa	(%eax), %xmm0
+	add	$0x10, %eax
+	movdqa	%xmm0, (%edx)
+	add	$0x10, %edx
+L(shl_0_mem_less_16bytes):
+	add	%ecx, %edx
+	add	%ecx, %eax
+	BRANCH_TO_JMPTBL_ENTRY (L(table_48bytes_fwd), %ecx, 4)
+
+
+	ALIGN (4)
+L(shl_1):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-1(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_1_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$1, %xmm2, %xmm3
+	palignr	$1, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_1_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$1, %xmm2, %xmm3
+	palignr	$1, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_1_loop)
+
+L(shl_1_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	1(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_2):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-2(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_2_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$2, %xmm2, %xmm3
+	palignr	$2, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_2_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$2, %xmm2, %xmm3
+	palignr	$2, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_2_loop)
+
+L(shl_2_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	2(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_3):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-3(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_3_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$3, %xmm2, %xmm3
+	palignr	$3, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_3_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$3, %xmm2, %xmm3
+	palignr	$3, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_3_loop)
+
+L(shl_3_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	3(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_4):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-4(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_4_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$4, %xmm2, %xmm3
+	palignr	$4, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_4_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$4, %xmm2, %xmm3
+	palignr	$4, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_4_loop)
+
+L(shl_4_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	4(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_5):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-5(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_5_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$5, %xmm2, %xmm3
+	palignr	$5, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_5_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$5, %xmm2, %xmm3
+	palignr	$5, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_5_loop)
+
+L(shl_5_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	5(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+
+	ALIGN (4)
+L(shl_6):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-6(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_6_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$6, %xmm2, %xmm3
+	palignr	$6, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_6_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$6, %xmm2, %xmm3
+	palignr	$6, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_6_loop)
+
+L(shl_6_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	6(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_7):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-7(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_7_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$7, %xmm2, %xmm3
+	palignr	$7, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_7_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$7, %xmm2, %xmm3
+	palignr	$7, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_7_loop)
+
+L(shl_7_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	7(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_8):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-8(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_8_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$8, %xmm2, %xmm3
+	palignr	$8, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_8_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$8, %xmm2, %xmm3
+	palignr	$8, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_8_loop)
+
+L(shl_8_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	8(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_9):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-9(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_9_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$9, %xmm2, %xmm3
+	palignr	$9, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_9_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$9, %xmm2, %xmm3
+	palignr	$9, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_9_loop)
+
+L(shl_9_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	9(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_10):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-10(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_10_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$10, %xmm2, %xmm3
+	palignr	$10, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_10_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$10, %xmm2, %xmm3
+	palignr	$10, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_10_loop)
+
+L(shl_10_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	10(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_11):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-11(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_11_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$11, %xmm2, %xmm3
+	palignr	$11, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_11_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$11, %xmm2, %xmm3
+	palignr	$11, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_11_loop)
+
+L(shl_11_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	11(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_12):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-12(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_12_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$12, %xmm2, %xmm3
+	palignr	$12, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_12_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$12, %xmm2, %xmm3
+	palignr	$12, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_12_loop)
+
+L(shl_12_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	12(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_13):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-13(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_13_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$13, %xmm2, %xmm3
+	palignr	$13, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_13_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$13, %xmm2, %xmm3
+	palignr	$13, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_13_loop)
+
+L(shl_13_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	13(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+	ALIGN (4)
+L(shl_14):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-14(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_14_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$14, %xmm2, %xmm3
+	palignr	$14, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_14_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$14, %xmm2, %xmm3
+	palignr	$14, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_14_loop)
+
+L(shl_14_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	14(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+
+	ALIGN (4)
+L(shl_15):
+	BRANCH_TO_JMPTBL_ENTRY_VALUE(L(table_48bytes_fwd))
+	lea	-15(%eax), %eax
+	movaps	(%eax), %xmm1
+	xor	%edi, %edi
+	lea	-32(%ecx), %ecx
+	movdqu	%xmm0, (%esi)
+	POP (%esi)
+L(shl_15_loop):
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm4
+	palignr	$15, %xmm2, %xmm3
+	palignr	$15, %xmm1, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jl	L(shl_15_end)
+
+	movdqa	16(%eax, %edi), %xmm2
+	sub	$32, %ecx
+	movdqa	32(%eax, %edi), %xmm3
+	movdqa	%xmm3, %xmm1
+	palignr	$15, %xmm2, %xmm3
+	palignr	$15, %xmm4, %xmm2
+	lea	32(%edi), %edi
+	movdqa	%xmm2, -32(%edx, %edi)
+	movdqa	%xmm3, -16(%edx, %edi)
+
+	jae	L(shl_15_loop)
+
+L(shl_15_end):
+	lea	32(%ecx), %ecx
+	add	%ecx, %edi
+	add	%edi, %edx
+	lea	15(%edi, %eax), %eax
+	POP (%edi)
+	BRANCH_TO_JMPTBL_ENTRY_TAIL(L(table_48bytes_fwd), %ecx, 4)
+
+
+	ALIGN (4)
+L(fwd_write_44bytes):
+	movl	-44(%eax), %ecx
+	movl	%ecx, -44(%edx)
+L(fwd_write_40bytes):
+	movl	-40(%eax), %ecx
+	movl	%ecx, -40(%edx)
+L(fwd_write_36bytes):
+	movl	-36(%eax), %ecx
+	movl	%ecx, -36(%edx)
+L(fwd_write_32bytes):
+	movl	-32(%eax), %ecx
+	movl	%ecx, -32(%edx)
+L(fwd_write_28bytes):
+	movl	-28(%eax), %ecx
+	movl	%ecx, -28(%edx)
+L(fwd_write_24bytes):
+	movl	-24(%eax), %ecx
+	movl	%ecx, -24(%edx)
+L(fwd_write_20bytes):
+	movl	-20(%eax), %ecx
+	movl	%ecx, -20(%edx)
+L(fwd_write_16bytes):
+	movl	-16(%eax), %ecx
+	movl	%ecx, -16(%edx)
+L(fwd_write_12bytes):
+	movl	-12(%eax), %ecx
+	movl	%ecx, -12(%edx)
+L(fwd_write_8bytes):
+	movl	-8(%eax), %ecx
+	movl	%ecx, -8(%edx)
+L(fwd_write_4bytes):
+	movl	-4(%eax), %ecx
+	movl	%ecx, -4(%edx)
+L(fwd_write_0bytes):
+#ifndef USE_AS_BCOPY
+# ifdef USE_AS_MEMPCPY
+	movl	%edx, %eax
+# else
+	movl	DEST(%esp), %eax
+# endif
+#endif
+	RETURN
+
+	ALIGN (4)
+L(fwd_write_5bytes):
+	movl	-5(%eax), %ecx
+	movl	-4(%eax), %eax
+	movl	%ecx, -5(%edx)
+	movl	%eax, -4(%edx)
+#ifndef USE_AS_BCOPY
+# ifdef USE_AS_MEMPCPY
+	movl	%edx, %eax
+# else
+	movl	DEST(%esp), %eax
+# endif
+#endif
+	RETURN
+
+	ALIGN (4)
+L(fwd_write_45bytes):
+	movl	-45(%eax), %ecx
+	movl	%ecx, -45(%edx)
+L(fwd_write_41bytes):
+	movl	-41(%eax), %ecx
+	movl	%ecx, -41(%edx)
+L(fwd_write_37bytes):
+	movl	-37(%eax), %ecx
+	movl	%ecx, -37(%edx)
+L(fwd_write_33bytes):
+	movl	-33(%eax), %ecx
+	movl	%ecx, -33(%edx)
+L(fwd_write_29bytes):
+	movl	-29(%eax), %ecx
+	movl	%ecx, -29(%edx)
+L(fwd_write_25bytes):
+	movl	-25(%eax), %ecx
+	movl	%ecx, -25(%edx)
+L(fwd_write_21bytes):
+	movl	-21(%eax), %ecx
+	movl	%ecx, -21(%edx)
+L(fwd_write_17bytes):
+	movl	-17(%eax), %ecx
+	movl	%ecx, -17(%edx)
+L(fwd_write_13bytes):
+	movl	-13(%eax), %ecx
+	movl	%ecx, -13(%edx)
+L(fwd_write_9bytes):
+	movl	-9(%eax), %ecx
+	movl	%ecx, -9(%edx)
+	movl	-5(%eax), %ecx
+	movl	%ecx, -5(%edx)
+L(fwd_write_1bytes):
+	movzbl	-1(%eax), %ecx
+	movb	%cl, -1(%edx)
+#ifndef USE_AS_BCOPY
+# ifdef USE_AS_MEMPCPY
+	movl	%edx, %eax
+# else
+	movl	DEST(%esp), %eax
+# endif
+#endif
+	RETURN
+
+	ALIGN (4)
+L(fwd_write_46bytes):
+	movl	-46(%eax), %ecx
+	movl	%ecx, -46(%edx)
+L(fwd_write_42bytes):
+	movl	-42(%eax), %ecx
+	movl	%ecx, -42(%edx)
+L(fwd_write_38bytes):
+	movl	-38(%eax), %ecx
+	movl	%ecx, -38(%edx)
+L(fwd_write_34bytes):
+	movl	-34(%eax), %ecx
+	movl	%ecx, -34(%edx)
+L(fwd_write_30bytes):
+	movl	-30(%eax), %ecx
+	movl	%ecx, -30(%edx)
+L(fwd_write_26bytes):
+	movl	-26(%eax), %ecx
+	movl	%ecx, -26(%edx)
+L(fwd_write_22bytes):
+	movl	-22(%eax), %ecx
+	movl	%ecx, -22(%edx)
+L(fwd_write_18bytes):
+	movl	-18(%eax), %ecx
+	movl	%ecx, -18(%edx)
+L(fwd_write_14bytes):
+	movl	-14(%eax), %ecx
+	movl	%ecx, -14(%edx)
+L(fwd_write_10bytes):
+	movl	-10(%eax), %ecx
+	movl	%ecx, -10(%edx)
+L(fwd_write_6bytes):
+	movl	-6(%eax), %ecx
+	movl	%ecx, -6(%edx)
+L(fwd_write_2bytes):
+	movzwl	-2(%eax), %ecx
+	movw	%cx, -2(%edx)
+#ifndef USE_AS_BCOPY
+# ifdef USE_AS_MEMPCPY
+	movl	%edx, %eax
+# else
+	movl	DEST(%esp), %eax
+# endif
+#endif
+	RETURN
+
+	ALIGN (4)
+L(fwd_write_47bytes):
+	movl	-47(%eax), %ecx
+	movl	%ecx, -47(%edx)
+L(fwd_write_43bytes):
+	movl	-43(%eax), %ecx
+	movl	%ecx, -43(%edx)
+L(fwd_write_39bytes):
+	movl	-39(%eax), %ecx
+	movl	%ecx, -39(%edx)
+L(fwd_write_35bytes):
+	movl	-35(%eax), %ecx
+	movl	%ecx, -35(%edx)
+L(fwd_write_31bytes):
+	movl	-31(%eax), %ecx
+	movl	%ecx, -31(%edx)
+L(fwd_write_27bytes):
+	movl	-27(%eax), %ecx
+	movl	%ecx, -27(%edx)
+L(fwd_write_23bytes):
+	movl	-23(%eax), %ecx
+	movl	%ecx, -23(%edx)
+L(fwd_write_19bytes):
+	movl	-19(%eax), %ecx
+	movl	%ecx, -19(%edx)
+L(fwd_write_15bytes):
+	movl	-15(%eax), %ecx
+	movl	%ecx, -15(%edx)
+L(fwd_write_11bytes):
+	movl	-11(%eax), %ecx
+	movl	%ecx, -11(%edx)
+L(fwd_write_7bytes):
+	movl	-7(%eax), %ecx
+	movl	%ecx, -7(%edx)
+L(fwd_write_3bytes):
+	movzwl	-3(%eax), %ecx
+	movzbl	-1(%eax), %eax
+	movw	%cx, -3(%edx)
+	movb	%al, -1(%edx)
+#ifndef USE_AS_BCOPY
+# ifdef USE_AS_MEMPCPY
+	movl	%edx, %eax
+# else
+	movl	DEST(%esp), %eax
+# endif
+#endif
+	RETURN
+
+	ALIGN (4)
+L(large_page):
+	movdqu	(%eax), %xmm1
+	lea	16(%eax), %eax
+	movdqu	%xmm0, (%esi)
+	movntdq	%xmm1, (%edx)
+	lea	16(%edx), %edx
+	POP (%esi)
+	lea	-0x90(%ecx), %ecx
+	POP (%edi)
+L(large_page_loop):
+	movdqu	(%eax), %xmm0
+	movdqu	0x10(%eax), %xmm1
+	movdqu	0x20(%eax), %xmm2
+	movdqu	0x30(%eax), %xmm3
+	movdqu	0x40(%eax), %xmm4
+	movdqu	0x50(%eax), %xmm5
+	movdqu	0x60(%eax), %xmm6
+	movdqu	0x70(%eax), %xmm7
+	lea	0x80(%eax), %eax
+
+	sub	$0x80, %ecx
+	movntdq	%xmm0, (%edx)
+	movntdq	%xmm1, 0x10(%edx)
+	movntdq	%xmm2, 0x20(%edx)
+	movntdq	%xmm3, 0x30(%edx)
+	movntdq	%xmm4, 0x40(%edx)
+	movntdq	%xmm5, 0x50(%edx)
+	movntdq	%xmm6, 0x60(%edx)
+	movntdq	%xmm7, 0x70(%edx)
+	lea	0x80(%edx), %edx
+	jae	L(large_page_loop)
+	cmp	$-0x40, %ecx
+	lea	0x80(%ecx), %ecx
+	jl	L(large_page_less_64bytes)
+
+	movdqu	(%eax), %xmm0
+	movdqu	0x10(%eax), %xmm1
+	movdqu	0x20(%eax), %xmm2
+	movdqu	0x30(%eax), %xmm3
+	lea	0x40(%eax), %eax
+
+	movntdq	%xmm0, (%edx)
+	movntdq	%xmm1, 0x10(%edx)
+	movntdq	%xmm2, 0x20(%edx)
+	movntdq	%xmm3, 0x30(%edx)
+	lea	0x40(%edx), %edx
+	sub	$0x40, %ecx
+L(large_page_less_64bytes):
+	cmp	$32, %ecx
+	jl	L(large_page_less_32bytes)
+	movdqu	(%eax), %xmm0
+	movdqu	0x10(%eax), %xmm1
+	lea	0x20(%eax), %eax
+	movntdq	%xmm0, (%edx)
+	movntdq	%xmm1, 0x10(%edx)
+	lea	0x20(%edx), %edx
+	sub	$0x20, %ecx
+L(large_page_less_32bytes):
+	add	%ecx, %edx
+	add	%ecx, %eax
+	sfence
+	BRANCH_TO_JMPTBL_ENTRY (L(table_48bytes_fwd), %ecx, 4)
+
+
+	ALIGN (4)
+L(bk_write_44bytes):
+	movl	40(%eax), %ecx
+	movl	%ecx, 40(%edx)
+L(bk_write_40bytes):
+	movl	36(%eax), %ecx
+	movl	%ecx, 36(%edx)
+L(bk_write_36bytes):
+	movl	32(%eax), %ecx
+	movl	%ecx, 32(%edx)
+L(bk_write_32bytes):
+	movl	28(%eax), %ecx
+	movl	%ecx, 28(%edx)
+L(bk_write_28bytes):
+	movl	24(%eax), %ecx
+	movl	%ecx, 24(%edx)
+L(bk_write_24bytes):
+	movl	20(%eax), %ecx
+	movl	%ecx, 20(%edx)
+L(bk_write_20bytes):
+	movl	16(%eax), %ecx
+	movl	%ecx, 16(%edx)
+L(bk_write_16bytes):
+	movl	12(%eax), %ecx
+	movl	%ecx, 12(%edx)
+L(bk_write_12bytes):
+	movl	8(%eax), %ecx
+	movl	%ecx, 8(%edx)
+L(bk_write_8bytes):
+	movl	4(%eax), %ecx
+	movl	%ecx, 4(%edx)
+L(bk_write_4bytes):
+	movl	(%eax), %ecx
+	movl	%ecx, (%edx)
+L(bk_write_0bytes):
+#ifndef USE_AS_BCOPY
+	movl	DEST(%esp), %eax
+# ifdef USE_AS_MEMPCPY
+	movl	LEN(%esp), %ecx
+	add	%ecx, %eax
+# endif
+#endif
+	RETURN
+
+	ALIGN (4)
+L(bk_write_45bytes):
+	movl	41(%eax), %ecx
+	movl	%ecx, 41(%edx)
+L(bk_write_41bytes):
+	movl	37(%eax), %ecx
+	movl	%ecx, 37(%edx)
+L(bk_write_37bytes):
+	movl	33(%eax), %ecx
+	movl	%ecx, 33(%edx)
+L(bk_write_33bytes):
+	movl	29(%eax), %ecx
+	movl	%ecx, 29(%edx)
+L(bk_write_29bytes):
+	movl	25(%eax), %ecx
+	movl	%ecx, 25(%edx)
+L(bk_write_25bytes):
+	movl	21(%eax), %ecx
+	movl	%ecx, 21(%edx)
+L(bk_write_21bytes):
+	movl	17(%eax), %ecx
+	movl	%ecx, 17(%edx)
+L(bk_write_17bytes):
+	movl	13(%eax), %ecx
+	movl	%ecx, 13(%edx)
+L(bk_write_13bytes):
+	movl	9(%eax), %ecx
+	movl	%ecx, 9(%edx)
+L(bk_write_9bytes):
+	movl	5(%eax), %ecx
+	movl	%ecx, 5(%edx)
+L(bk_write_5bytes):
+	movl	1(%eax), %ecx
+	movl	%ecx, 1(%edx)
+L(bk_write_1bytes):
+	movzbl	(%eax), %ecx
+	movb	%cl, (%edx)
+#ifndef USE_AS_BCOPY
+	movl	DEST(%esp), %eax
+# ifdef USE_AS_MEMPCPY
+	movl	LEN(%esp), %ecx
+	add	%ecx, %eax
+# endif
+#endif
+	RETURN
+
+	ALIGN (4)
+L(bk_write_46bytes):
+	movl	42(%eax), %ecx
+	movl	%ecx, 42(%edx)
+L(bk_write_42bytes):
+	movl	38(%eax), %ecx
+	movl	%ecx, 38(%edx)
+L(bk_write_38bytes):
+	movl	34(%eax), %ecx
+	movl	%ecx, 34(%edx)
+L(bk_write_34bytes):
+	movl	30(%eax), %ecx
+	movl	%ecx, 30(%edx)
+L(bk_write_30bytes):
+	movl	26(%eax), %ecx
+	movl	%ecx, 26(%edx)
+L(bk_write_26bytes):
+	movl	22(%eax), %ecx
+	movl	%ecx, 22(%edx)
+L(bk_write_22bytes):
+	movl	18(%eax), %ecx
+	movl	%ecx, 18(%edx)
+L(bk_write_18bytes):
+	movl	14(%eax), %ecx
+	movl	%ecx, 14(%edx)
+L(bk_write_14bytes):
+	movl	10(%eax), %ecx
+	movl	%ecx, 10(%edx)
+L(bk_write_10bytes):
+	movl	6(%eax), %ecx
+	movl	%ecx, 6(%edx)
+L(bk_write_6bytes):
+	movl	2(%eax), %ecx
+	movl	%ecx, 2(%edx)
+L(bk_write_2bytes):
+	movzwl	(%eax), %ecx
+	movw	%cx, (%edx)
+#ifndef USE_AS_BCOPY
+	movl	DEST(%esp), %eax
+# ifdef USE_AS_MEMPCPY
+	movl	LEN(%esp), %ecx
+	add	%ecx, %eax
+# endif
+#endif
+	RETURN
+
+	ALIGN (4)
+L(bk_write_47bytes):
+	movl	43(%eax), %ecx
+	movl	%ecx, 43(%edx)
+L(bk_write_43bytes):
+	movl	39(%eax), %ecx
+	movl	%ecx, 39(%edx)
+L(bk_write_39bytes):
+	movl	35(%eax), %ecx
+	movl	%ecx, 35(%edx)
+L(bk_write_35bytes):
+	movl	31(%eax), %ecx
+	movl	%ecx, 31(%edx)
+L(bk_write_31bytes):
+	movl	27(%eax), %ecx
+	movl	%ecx, 27(%edx)
+L(bk_write_27bytes):
+	movl	23(%eax), %ecx
+	movl	%ecx, 23(%edx)
+L(bk_write_23bytes):
+	movl	19(%eax), %ecx
+	movl	%ecx, 19(%edx)
+L(bk_write_19bytes):
+	movl	15(%eax), %ecx
+	movl	%ecx, 15(%edx)
+L(bk_write_15bytes):
+	movl	11(%eax), %ecx
+	movl	%ecx, 11(%edx)
+L(bk_write_11bytes):
+	movl	7(%eax), %ecx
+	movl	%ecx, 7(%edx)
+L(bk_write_7bytes):
+	movl	3(%eax), %ecx
+	movl	%ecx, 3(%edx)
+L(bk_write_3bytes):
+	movzwl	1(%eax), %ecx
+	movw	%cx, 1(%edx)
+	movzbl	(%eax), %eax
+	movb	%al, (%edx)
+#ifndef USE_AS_BCOPY
+	movl	DEST(%esp), %eax
+# ifdef USE_AS_MEMPCPY
+	movl	LEN(%esp), %ecx
+	add	%ecx, %eax
+# endif
+#endif
+	RETURN_END
+
+
+	.pushsection .rodata.ssse3,"a",@progbits
+	ALIGN (2)
+L(table_48bytes_fwd):
+	.int	JMPTBL (L(fwd_write_0bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_1bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_2bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_3bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_4bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_5bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_6bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_7bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_8bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_9bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_10bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_11bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_12bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_13bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_14bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_15bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_16bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_17bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_18bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_19bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_20bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_21bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_22bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_23bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_24bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_25bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_26bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_27bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_28bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_29bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_30bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_31bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_32bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_33bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_34bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_35bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_36bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_37bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_38bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_39bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_40bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_41bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_42bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_43bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_44bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_45bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_46bytes), L(table_48bytes_fwd))
+	.int	JMPTBL (L(fwd_write_47bytes), L(table_48bytes_fwd))
+
+	ALIGN (2)
+L(shl_table):
+	.int	JMPTBL (L(shl_0), L(shl_table))
+	.int	JMPTBL (L(shl_1), L(shl_table))
+	.int	JMPTBL (L(shl_2), L(shl_table))
+	.int	JMPTBL (L(shl_3), L(shl_table))
+	.int	JMPTBL (L(shl_4), L(shl_table))
+	.int	JMPTBL (L(shl_5), L(shl_table))
+	.int	JMPTBL (L(shl_6), L(shl_table))
+	.int	JMPTBL (L(shl_7), L(shl_table))
+	.int	JMPTBL (L(shl_8), L(shl_table))
+	.int	JMPTBL (L(shl_9), L(shl_table))
+	.int	JMPTBL (L(shl_10), L(shl_table))
+	.int	JMPTBL (L(shl_11), L(shl_table))
+	.int	JMPTBL (L(shl_12), L(shl_table))
+	.int	JMPTBL (L(shl_13), L(shl_table))
+	.int	JMPTBL (L(shl_14), L(shl_table))
+	.int	JMPTBL (L(shl_15), L(shl_table))
+
+	ALIGN (2)
+L(table_48_bytes_bwd):
+	.int	JMPTBL (L(bk_write_0bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_1bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_2bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_3bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_4bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_5bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_6bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_7bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_8bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_9bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_10bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_11bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_12bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_13bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_14bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_15bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_16bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_17bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_18bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_19bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_20bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_21bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_22bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_23bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_24bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_25bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_26bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_27bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_28bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_29bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_30bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_31bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_32bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_33bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_34bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_35bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_36bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_37bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_38bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_39bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_40bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_41bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_42bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_43bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_44bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_45bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_46bytes), L(table_48_bytes_bwd))
+	.int	JMPTBL (L(bk_write_47bytes), L(table_48_bytes_bwd))
+
+	.popsection
+
+#ifdef USE_AS_MEMMOVE
+	ALIGN (4)
+L(copy_backward):
+	PUSH (%esi)
+	movl	%eax, %esi
+	lea	(%ecx,%edx,1),%edx
+	lea	(%ecx,%esi,1),%esi
+	testl	$0x3, %edx
+	jnz	L(bk_align)
+
+L(bk_aligned_4):
+	cmp	$64, %ecx
+	jge	L(bk_write_more64bytes)
+
+L(bk_write_64bytesless):
+	cmp	$32, %ecx
+	jl	L(bk_write_less32bytes)
+
+L(bk_write_more32bytes):
+	/* Copy 32 bytes at a time.  */
+	sub	$32, %ecx
+	movl	-4(%esi), %eax
+	movl	%eax, -4(%edx)
+	movl	-8(%esi), %eax
+	movl	%eax, -8(%edx)
+	movl	-12(%esi), %eax
+	movl	%eax, -12(%edx)
+	movl	-16(%esi), %eax
+	movl	%eax, -16(%edx)
+	movl	-20(%esi), %eax
+	movl	%eax, -20(%edx)
+	movl	-24(%esi), %eax
+	movl	%eax, -24(%edx)
+	movl	-28(%esi), %eax
+	movl	%eax, -28(%edx)
+	movl	-32(%esi), %eax
+	movl	%eax, -32(%edx)
+	sub	$32, %edx
+	sub	$32, %esi
+
+L(bk_write_less32bytes):
+	movl	%esi, %eax
+	sub	%ecx, %edx
+	sub	%ecx, %eax
+	POP (%esi)
+L(bk_write_less32bytes_2):
+	BRANCH_TO_JMPTBL_ENTRY (L(table_48_bytes_bwd), %ecx, 4)
+
+	ALIGN (4)
+L(bk_align):
+	cmp	$8, %ecx
+	jle	L(bk_write_less32bytes)
+	testl	$1, %edx
+	/* We get here only if (EDX & 3 ) != 0 so if (EDX & 1) ==0,
+	   then (EDX & 2) must be != 0.  */
+	jz	L(bk_got2)
+	sub	$1, %esi
+	sub	$1, %ecx
+	sub	$1, %edx
+	movzbl	(%esi), %eax
+	movb	%al, (%edx)
+
+	testl	$2, %edx
+	jz	L(bk_aligned_4)
+
+L(bk_got2):
+	sub	$2, %esi
+	sub	$2, %ecx
+	sub	$2, %edx
+	movzwl	(%esi), %eax
+	movw	%ax, (%edx)
+	jmp	L(bk_aligned_4)
+
+	ALIGN (4)
+L(bk_write_more64bytes):
+	/* Check alignment of last byte.  */
+	testl	$15, %edx
+	jz	L(bk_ssse3_cpy_pre)
+
+/* EDX is aligned 4 bytes, but not 16 bytes.  */
+L(bk_ssse3_align):
+	sub	$4, %esi
+	sub	$4, %ecx
+	sub	$4, %edx
+	movl	(%esi), %eax
+	movl	%eax, (%edx)
+
+	testl	$15, %edx
+	jz	L(bk_ssse3_cpy_pre)
+
+	sub	$4, %esi
+	sub	$4, %ecx
+	sub	$4, %edx
+	movl	(%esi), %eax
+	movl	%eax, (%edx)
+
+	testl	$15, %edx
+	jz	L(bk_ssse3_cpy_pre)
+
+	sub	$4, %esi
+	sub	$4, %ecx
+	sub	$4, %edx
+	movl	(%esi), %eax
+	movl	%eax, (%edx)
+
+L(bk_ssse3_cpy_pre):
+	cmp	$64, %ecx
+	jl	L(bk_write_more32bytes)
+
+L(bk_ssse3_cpy):
+	sub	$64, %esi
+	sub	$64, %ecx
+	sub	$64, %edx
+	movdqu	0x30(%esi), %xmm3
+	movdqa	%xmm3, 0x30(%edx)
+	movdqu	0x20(%esi), %xmm2
+	movdqa	%xmm2, 0x20(%edx)
+	movdqu	0x10(%esi), %xmm1
+	movdqa	%xmm1, 0x10(%edx)
+	movdqu	(%esi), %xmm0
+	movdqa	%xmm0, (%edx)
+	cmp	$64, %ecx
+	jge	L(bk_ssse3_cpy)
+	jmp	L(bk_write_64bytesless)
+
+#endif
+
+END (MEMCPY)
+
+#endif
diff --git a/sysdeps/i386/i686/multiarch/memcpy.S b/sysdeps/i386/i686/multiarch/memcpy.S
new file mode 100644
index 0000000000..bf1c7cc2d2
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/memcpy.S
@@ -0,0 +1,90 @@
+/* Multiple versions of memcpy
+   Copyright (C) 2010 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>
+
+/* Define multiple versions only for the definition in lib and for
+   DSO.  In static binaries we need memcpy before the initialization
+   happened.  */
+#if defined SHARED && !defined 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
+
+	.text
+ENTRY(memcpy)
+	.type	memcpy, @gnu_indirect_function
+	pushl	%ebx
+	cfi_adjust_cfa_offset (4)
+	cfi_rel_offset (ebx, 0)
+	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:	leal	__memcpy_ia32@GOTOFF(%ebx), %eax
+	testl	$bit_SSSE3, CPUID_OFFSET+index_SSSE3+__cpu_features@GOTOFF(%ebx)
+	jz	2f
+	leal	__memcpy_ssse3@GOTOFF(%ebx), %eax
+	testl	$bit_Fast_Rep_String, FEATURE_OFFSET+index_Fast_Rep_String+__cpu_features@GOTOFF(%ebx)
+	jz	2f
+	leal	__memcpy_ssse3_rep@GOTOFF(%ebx), %eax
+2:	popl	%ebx
+	cfi_adjust_cfa_offset (-4)
+	cfi_restore (ebx)
+	ret
+END(memcpy)
+
+# undef ENTRY
+# define ENTRY(name) \
+	.type __memcpy_ia32, @function; \
+	.p2align 4; \
+	__memcpy_ia32: cfi_startproc; \
+	CALL_MCOUNT
+# undef END
+# define END(name) \
+	cfi_endproc; .size __memcpy_ia32, .-__memcpy_ia32
+
+# undef ENTRY_CHK
+# define ENTRY_CHK(name) \
+	.type __memcpy_chk_ia32, @function; \
+	.globl __memcpy_chk_ia32; \
+	.p2align 4; \
+	__memcpy_chk_ia32: cfi_startproc; \
+	CALL_MCOUNT
+# undef END_CHK
+# define END_CHK(name) \
+	cfi_endproc; .size __memcpy_chk_ia32, .-__memcpy_chk_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_memcpy; __GI_memcpy = __memcpy_ia32
+#endif
+
+#include "../memcpy.S"
diff --git a/sysdeps/i386/i686/multiarch/memcpy_chk.S b/sysdeps/i386/i686/multiarch/memcpy_chk.S
new file mode 100644
index 0000000000..171ac8adef
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/memcpy_chk.S
@@ -0,0 +1,64 @@
+/* Multiple versions of __memcpy_chk
+   Copyright (C) 2010 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>
+
+/* Define multiple versions only for the definition in lib and for
+   DSO.  There are no multiarch memcpy functions for static binaries.
+ */
+#ifndef NOT_IN_libc
+# ifdef SHARED
+	.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
+
+	.text
+ENTRY(__memcpy_chk)
+	.type	__memcpy_chk, @gnu_indirect_function
+	pushl	%ebx
+	cfi_adjust_cfa_offset (4)
+	cfi_rel_offset (ebx, 0)
+	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:	leal	__memcpy_chk_ia32@GOTOFF(%ebx), %eax
+	testl	$bit_SSSE3, CPUID_OFFSET+index_SSSE3+__cpu_features@GOTOFF(%ebx)
+	jz	2f
+	leal	__memcpy_chk_ssse3@GOTOFF(%ebx), %eax
+	testl	$bit_Fast_Rep_String, FEATURE_OFFSET+index_Fast_Rep_String+__cpu_features@GOTOFF(%ebx)
+	jz	2f
+	leal	__memcpy_chk_ssse3_rep@GOTOFF(%ebx), %eax
+2:	popl	%ebx
+	cfi_adjust_cfa_offset (-4)
+	cfi_restore (ebx)
+	ret
+END(__memcpy_chk)
+# else
+#  include "../memcpy_chk.S"
+# endif
+#endif
diff --git a/sysdeps/i386/i686/multiarch/memmove-ssse3-rep.S b/sysdeps/i386/i686/multiarch/memmove-ssse3-rep.S
new file mode 100644
index 0000000000..d202fc4a13
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/memmove-ssse3-rep.S
@@ -0,0 +1,4 @@
+#define USE_AS_MEMMOVE
+#define MEMCPY		__memmove_ssse3_rep
+#define MEMCPY_CHK	__memmove_chk_ssse3_rep
+#include "memcpy-ssse3-rep.S"
diff --git a/sysdeps/i386/i686/multiarch/memmove-ssse3.S b/sysdeps/i386/i686/multiarch/memmove-ssse3.S
new file mode 100644
index 0000000000..295430b1ef
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/memmove-ssse3.S
@@ -0,0 +1,4 @@
+#define USE_AS_MEMMOVE
+#define MEMCPY		__memmove_ssse3
+#define MEMCPY_CHK	__memmove_chk_ssse3
+#include "memcpy-ssse3.S"
diff --git a/sysdeps/i386/i686/multiarch/memmove.S b/sysdeps/i386/i686/multiarch/memmove.S
new file mode 100644
index 0000000000..d3d8d3675c
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/memmove.S
@@ -0,0 +1,114 @@
+/* Multiple versions of memmove
+   Copyright (C) 2010 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>
+
+/* Define multiple versions only for the definition in lib.  */
+#ifndef NOT_IN_libc
+# ifdef SHARED
+	.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
+
+	.text
+ENTRY(memmove)
+	.type	memmove, @gnu_indirect_function
+	pushl	%ebx
+	cfi_adjust_cfa_offset (4)
+	cfi_rel_offset (ebx, 0)
+	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:	leal	__memmove_ia32@GOTOFF(%ebx), %eax
+	testl	$bit_SSSE3, CPUID_OFFSET+index_SSSE3+__cpu_features@GOTOFF(%ebx)
+	jz	2f
+	leal	__memmove_ssse3@GOTOFF(%ebx), %eax
+	testl	$bit_Fast_Rep_String, FEATURE_OFFSET+index_Fast_Rep_String+__cpu_features@GOTOFF(%ebx)
+	jz	2f
+	leal	__memmove_ssse3_rep@GOTOFF(%ebx), %eax
+2:	popl	%ebx
+	cfi_adjust_cfa_offset (-4)
+	cfi_restore (ebx)
+	ret
+END(memmove)
+
+# undef ENTRY
+# define ENTRY(name) \
+	.type __memmove_ia32, @function; \
+	.p2align 4; \
+	__memmove_ia32: cfi_startproc; \
+	CALL_MCOUNT
+# else
+	.text
+ENTRY(memmove)
+	.type	memmove, @gnu_indirect_function
+	cmpl	$0, KIND_OFFSET+__cpu_features
+	jne	1f
+	call	__init_cpu_features
+1:	leal	__memmove_ia32, %eax
+	testl	$bit_SSSE3, CPUID_OFFSET+index_SSSE3+__cpu_features
+	jz	2f
+	leal	__memmove_ssse3, %eax
+2:	ret
+END(memmove)
+
+# undef ENTRY
+# define ENTRY(name) \
+	.type __memmove_ia32, @function; \
+	.globl __memmove_ia32; \
+	.p2align 4; \
+	__memmove_ia32: cfi_startproc; \
+	CALL_MCOUNT
+# endif
+
+# undef END
+# define END(name) \
+	cfi_endproc; .size __memmove_ia32, .-__memmove_ia32
+
+# undef ENTRY_CHK
+# define ENTRY_CHK(name) \
+	.type __memmove_chk_ia32, @function; \
+	.globl __memmove_chk_ia32; \
+	.p2align 4; \
+	__memmove_chk_ia32: cfi_startproc; \
+	CALL_MCOUNT
+# undef END_CHK
+# define END_CHK(name) \
+	cfi_endproc; .size __memmove_chk_ia32, .-__memmove_chk_ia32
+
+# ifdef SHARED
+#  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_memmove; __GI_memmove = __memmove_ia32
+# endif
+#endif
+
+#include "../memmove.S"
diff --git a/sysdeps/i386/i686/multiarch/memmove_chk.S b/sysdeps/i386/i686/multiarch/memmove_chk.S
new file mode 100644
index 0000000000..6c80dfb8f8
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/memmove_chk.S
@@ -0,0 +1,112 @@
+/* Multiple versions of __memmove_chk
+   Copyright (C) 2010 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>
+
+/* Define multiple versions only for the definition in lib.  */
+#ifndef NOT_IN_libc
+# ifdef SHARED
+	.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
+
+	.text
+ENTRY(__memmove_chk)
+	.type	__memmove_chk, @gnu_indirect_function
+	pushl	%ebx
+	cfi_adjust_cfa_offset (4)
+	cfi_rel_offset (ebx, 0)
+	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:	leal	__memmove_chk_ia32@GOTOFF(%ebx), %eax
+	testl	$bit_SSSE3, CPUID_OFFSET+index_SSSE3+__cpu_features@GOTOFF(%ebx)
+	jz	2f
+	leal	__memmove_chk_ssse3@GOTOFF(%ebx), %eax
+	testl	$bit_Fast_Rep_String, FEATURE_OFFSET+index_Fast_Rep_String+__cpu_features@GOTOFF(%ebx)
+	jz	2f
+	leal	__memmove_chk_ssse3_rep@GOTOFF(%ebx), %eax
+2:	popl	%ebx
+	cfi_adjust_cfa_offset (-4)
+	cfi_restore (ebx)
+	ret
+END(__memmove_chk)
+# else
+	.text
+ENTRY(__memmove_chk)
+	.type	__memmove_chk, @gnu_indirect_function
+	cmpl	$0, KIND_OFFSET+__cpu_features
+	jne	1f
+	call	__init_cpu_features
+1:	leal	__memmove_chk_ia32, %eax
+	testl	$bit_SSSE3, CPUID_OFFSET+index_SSSE3+__cpu_features
+	jz	2f
+	leal	__memmove_chk_ssse3, %eax
+	testl	$bit_Fast_Rep_String, FEATURE_OFFSET+index_Fast_Rep_String+__cpu_features
+	jz	2f
+	leal	__memmove_chk_ssse3_rep, %eax
+2:	ret
+END(__memmove_chk)
+
+	.type __memmove_chk_ssse3, @function
+	.p2align 4;
+__memmove_chk_ssse3:
+	cfi_startproc
+	CALL_MCOUNT
+	movl	12(%esp), %eax
+	cmpl	%eax, 16(%esp)
+	jb	__chk_fail
+	jmp	__memmove_ssse3
+	cfi_endproc 
+	.size __memmove_chk_ssse3, .-__memmove_chk_ssse3
+
+	.type __memmove_chk_ssse3_rep, @function
+	.p2align 4;
+__memmove_chk_ssse3_rep:
+	cfi_startproc
+	CALL_MCOUNT
+	movl	12(%esp), %eax
+	cmpl	%eax, 16(%esp)
+	jb	__chk_fail
+	jmp	__memmove_ssse3_rep
+	cfi_endproc 
+	.size __memmove_chk_ssse3_rep, .-__memmove_chk_ssse3_rep
+
+	.type __memmove_chk_ia32, @function
+	.p2align 4;
+__memmove_chk_ia32:
+	cfi_startproc
+	CALL_MCOUNT
+	movl	12(%esp), %eax
+	cmpl	%eax, 16(%esp)
+	jb	__chk_fail
+	jmp	__memmove_ia32
+	cfi_endproc 
+	.size __memmove_chk_ia32, .-__memmove_chk_ia32
+# endif
+#endif
diff --git a/sysdeps/i386/i686/multiarch/mempcpy-ssse3-rep.S b/sysdeps/i386/i686/multiarch/mempcpy-ssse3-rep.S
new file mode 100644
index 0000000000..5357b33e18
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/mempcpy-ssse3-rep.S
@@ -0,0 +1,4 @@
+#define USE_AS_MEMPCPY
+#define MEMCPY		__mempcpy_ssse3_rep
+#define MEMCPY_CHK	__mempcpy_chk_ssse3_rep
+#include "memcpy-ssse3-rep.S"
diff --git a/sysdeps/i386/i686/multiarch/mempcpy-ssse3.S b/sysdeps/i386/i686/multiarch/mempcpy-ssse3.S
new file mode 100644
index 0000000000..822d98e954
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/mempcpy-ssse3.S
@@ -0,0 +1,4 @@
+#define USE_AS_MEMPCPY
+#define MEMCPY		__mempcpy_ssse3
+#define MEMCPY_CHK	__mempcpy_chk_ssse3
+#include "memcpy-ssse3.S"
diff --git a/sysdeps/i386/i686/multiarch/mempcpy.S b/sysdeps/i386/i686/multiarch/mempcpy.S
new file mode 100644
index 0000000000..df830d2e63
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/mempcpy.S
@@ -0,0 +1,93 @@
+/* Multiple versions of mempcpy
+   Copyright (C) 2010 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>
+
+/* Define multiple versions only for the definition in lib and for
+   DSO.  In static binaries we need mempcpy before the initialization
+   happened.  */
+#if defined SHARED && !defined 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
+
+	.text
+ENTRY(__mempcpy)
+	.type	__mempcpy, @gnu_indirect_function
+	pushl	%ebx
+	cfi_adjust_cfa_offset (4)
+	cfi_rel_offset (ebx, 0)
+	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:	leal	__mempcpy_ia32@GOTOFF(%ebx), %eax
+	testl	$bit_SSSE3, CPUID_OFFSET+index_SSSE3+__cpu_features@GOTOFF(%ebx)
+	jz	2f
+	leal	__mempcpy_ssse3@GOTOFF(%ebx), %eax
+	testl	$bit_Fast_Rep_String, FEATURE_OFFSET+index_Fast_Rep_String+__cpu_features@GOTOFF(%ebx)
+	jz	2f
+	leal	__mempcpy_ssse3_rep@GOTOFF(%ebx), %eax
+2:	popl	%ebx
+	cfi_adjust_cfa_offset (-4)
+	cfi_restore (ebx)
+	ret
+END(__mempcpy)
+
+# undef ENTRY
+# define ENTRY(name) \
+	.type __mempcpy_ia32, @function; \
+	.p2align 4; \
+	__mempcpy_ia32: cfi_startproc; \
+	CALL_MCOUNT
+# undef END
+# define END(name) \
+	cfi_endproc; .size __mempcpy_ia32, .-__mempcpy_ia32
+
+# undef ENTRY_CHK
+# define ENTRY_CHK(name) \
+	.type __mempcpy_chk_ia32, @function; \
+	.globl __mempcpy_chk_ia32; \
+	.p2align 4; \
+	__mempcpy_chk_ia32: cfi_startproc; \
+	CALL_MCOUNT
+# undef END_CHK
+# define END_CHK(name) \
+	cfi_endproc; .size __mempcpy_chk_ia32, .-__mempcpy_chk_ia32
+
+# undef libc_hidden_def
+# 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_def(name) \
+	.globl __GI_mempcpy; __GI_mempcpy = __mempcpy_ia32
+# define libc_hidden_builtin_def(name) \
+	.globl __GI___mempcpy; __GI___mempcpy = __mempcpy_ia32
+#endif
+
+#include "../mempcpy.S"
diff --git a/sysdeps/i386/i686/multiarch/mempcpy_chk.S b/sysdeps/i386/i686/multiarch/mempcpy_chk.S
new file mode 100644
index 0000000000..828fb5e608
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/mempcpy_chk.S
@@ -0,0 +1,64 @@
+/* Multiple versions of __mempcpy_chk
+   Copyright (C) 2010 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>
+
+/* Define multiple versions only for the definition in lib and for
+   DSO.  There are no multiarch mempcpy functions for static binaries.
+ */
+#ifndef NOT_IN_libc
+# ifdef SHARED
+	.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
+
+	.text
+ENTRY(__mempcpy_chk)
+	.type	__mempcpy_chk, @gnu_indirect_function
+	pushl	%ebx
+	cfi_adjust_cfa_offset (4)
+	cfi_rel_offset (ebx, 0)
+	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:	leal	__mempcpy_chk_ia32@GOTOFF(%ebx), %eax
+	testl	$bit_SSSE3, CPUID_OFFSET+index_SSSE3+__cpu_features@GOTOFF(%ebx)
+	jz	2f
+	leal	__mempcpy_chk_ssse3@GOTOFF(%ebx), %eax
+	testl	$bit_Fast_Rep_String, FEATURE_OFFSET+index_Fast_Rep_String+__cpu_features@GOTOFF(%ebx)
+	jz	2f
+	leal	__mempcpy_chk_ssse3_rep@GOTOFF(%ebx), %eax
+2:	popl	%ebx
+	cfi_adjust_cfa_offset (-4)
+	cfi_restore (ebx)
+	ret
+END(__mempcpy_chk)
+# else
+#  include "../mempcpy_chk.S"
+# endif
+#endif
diff --git a/sysdeps/i386/i686/multiarch/memset-sse2-rep.S b/sysdeps/i386/i686/multiarch/memset-sse2-rep.S
new file mode 100644
index 0000000000..1e7ba11917
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/memset-sse2-rep.S
@@ -0,0 +1,821 @@
+/* memset with SSE2 and REP string.
+   Copyright (C) 2010 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>
+#include "asm-syntax.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)
+
+#ifdef USE_AS_BZERO
+# define DEST		PARMS
+# define LEN		DEST+4
+# define SETRTNVAL
+#else
+# define DEST		PARMS
+# define CHR		DEST+4
+# define LEN		CHR+4
+# define SETRTNVAL	movl DEST(%esp), %eax
+#endif
+
+#ifdef SHARED
+# define ENTRANCE	PUSH (%ebx);
+# define RETURN_END	POP (%ebx); ret
+# define RETURN		RETURN_END; CFI_PUSH (%ebx)
+# define PARMS		8		/* Preserve EBX.  */
+# define JMPTBL(I, B)	I - B
+
+/* Load an entry in a jump table into EBX and branch to it.  TABLE is a
+   jump table with relative offsets.   */
+# define BRANCH_TO_JMPTBL_ENTRY(TABLE)				\
+    /* We first load PC into EBX.  */				\
+    call	__i686.get_pc_thunk.bx;				\
+    /* Get the address of the jump table.  */			\
+    add		$(TABLE - .), %ebx;				\
+    /* Get the entry and convert the relative offset to the	\
+       absolute address.  */					\
+    add		(%ebx,%ecx,4), %ebx;				\
+    add		%ecx, %edx;					\
+    /* We loaded the jump table and adjuested EDX. Go.  */	\
+    jmp		*%ebx
+
+	.section	.gnu.linkonce.t.__i686.get_pc_thunk.bx,"ax",@progbits
+	.globl	__i686.get_pc_thunk.bx
+	.hidden	__i686.get_pc_thunk.bx
+	ALIGN (4)
+	.type	__i686.get_pc_thunk.bx,@function
+__i686.get_pc_thunk.bx:
+	movl	(%esp), %ebx
+	ret
+#else
+# define ENTRANCE
+# define RETURN_END	ret
+# define RETURN		RETURN_END
+# define PARMS		4
+# define JMPTBL(I, B)	I
+
+/* Branch to an entry in a jump table.  TABLE is a jump table with
+   absolute offsets.  */
+# define BRANCH_TO_JMPTBL_ENTRY(TABLE)				\
+    add		%ecx, %edx;					\
+    jmp		*TABLE(,%ecx,4)
+#endif
+
+	.section .text.sse2,"ax",@progbits
+#if defined SHARED && !defined NOT_IN_libc && !defined USE_AS_BZERO
+ENTRY (__memset_chk_sse2_rep)
+	movl	12(%esp), %eax
+	cmpl	%eax, 16(%esp)
+	jb	HIDDEN_JUMPTARGET (__chk_fail)
+END (__memset_chk_sse2_rep)
+#endif
+ENTRY (__memset_sse2_rep)
+	ENTRANCE
+
+	movl	LEN(%esp), %ecx
+#ifdef USE_AS_BZERO
+	xor	%eax, %eax
+#else
+	movzbl	CHR(%esp), %eax
+	movb	%al, %ah
+	/* Fill the whole EAX with pattern.  */
+	movl	%eax, %edx
+	shl	$16, %eax
+	or	%edx, %eax
+#endif
+	movl	DEST(%esp), %edx
+	cmp	$32, %ecx
+	jae	L(32bytesormore)
+
+L(write_less32bytes):
+	BRANCH_TO_JMPTBL_ENTRY (L(table_less_32bytes))
+
+
+	.pushsection .rodata.sse2,"a",@progbits
+	ALIGN (2)
+L(table_less_32bytes):
+	.int	JMPTBL (L(write_0bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_1bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_2bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_3bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_4bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_5bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_6bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_7bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_8bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_9bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_10bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_11bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_12bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_13bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_14bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_15bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_16bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_17bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_18bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_19bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_20bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_21bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_22bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_23bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_24bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_25bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_26bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_27bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_28bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_29bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_30bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_31bytes), L(table_less_32bytes))
+	.popsection
+
+	ALIGN (4)
+L(write_28bytes):
+	movl	%eax, -28(%edx)
+L(write_24bytes):
+	movl	%eax, -24(%edx)
+L(write_20bytes):
+	movl	%eax, -20(%edx)
+L(write_16bytes):
+	movl	%eax, -16(%edx)
+L(write_12bytes):
+	movl	%eax, -12(%edx)
+L(write_8bytes):
+	movl	%eax, -8(%edx)
+L(write_4bytes):
+	movl	%eax, -4(%edx)
+L(write_0bytes):
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(write_29bytes):
+	movl	%eax, -29(%edx)
+L(write_25bytes):
+	movl	%eax, -25(%edx)
+L(write_21bytes):
+	movl	%eax, -21(%edx)
+L(write_17bytes):
+	movl	%eax, -17(%edx)
+L(write_13bytes):
+	movl	%eax, -13(%edx)
+L(write_9bytes):
+	movl	%eax, -9(%edx)
+L(write_5bytes):
+	movl	%eax, -5(%edx)
+L(write_1bytes):
+	movb	%al, -1(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(write_30bytes):
+	movl	%eax, -30(%edx)
+L(write_26bytes):
+	movl	%eax, -26(%edx)
+L(write_22bytes):
+	movl	%eax, -22(%edx)
+L(write_18bytes):
+	movl	%eax, -18(%edx)
+L(write_14bytes):
+	movl	%eax, -14(%edx)
+L(write_10bytes):
+	movl	%eax, -10(%edx)
+L(write_6bytes):
+	movl	%eax, -6(%edx)
+L(write_2bytes):
+	movw	%ax, -2(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(write_31bytes):
+	movl	%eax, -31(%edx)
+L(write_27bytes):
+	movl	%eax, -27(%edx)
+L(write_23bytes):
+	movl	%eax, -23(%edx)
+L(write_19bytes):
+	movl	%eax, -19(%edx)
+L(write_15bytes):
+	movl	%eax, -15(%edx)
+L(write_11bytes):
+	movl	%eax, -11(%edx)
+L(write_7bytes):
+	movl	%eax, -7(%edx)
+L(write_3bytes):
+	movw	%ax, -3(%edx)
+	movb	%al, -1(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+/* ECX > 32 and EDX is 4 byte aligned.  */
+L(32bytesormore):
+	/* Fill xmm0 with the pattern.  */
+#ifdef USE_AS_BZERO
+	pxor	%xmm0, %xmm0
+#else
+	movd	%eax, %xmm0
+	punpcklbw %xmm0, %xmm0
+	pshufd	$0, %xmm0, %xmm0
+#endif
+	testl	$0xf, %edx
+	jz	L(aligned_16)
+/* ECX > 32 and EDX is not 16 byte aligned.  */
+L(not_aligned_16):
+	movdqu	%xmm0, (%edx)
+	movl	%edx, %eax
+	and	$-16, %edx
+	add	$16, %edx
+	sub	%edx, %eax
+	add	%eax, %ecx
+	movd	%xmm0, %eax
+
+	ALIGN (4)
+L(aligned_16):
+	cmp	$128, %ecx
+	jge	L(128bytesormore)
+
+L(aligned_16_less128bytes):
+	BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes))
+
+	ALIGN (4)
+L(128bytesormore):
+	PUSH (%edi)
+#ifdef DATA_CACHE_SIZE
+	PUSH (%ebx)
+	mov	$DATA_CACHE_SIZE, %ebx
+#else
+# ifdef SHARED
+	call	__i686.get_pc_thunk.bx
+	add	$_GLOBAL_OFFSET_TABLE_, %ebx
+	mov	__x86_data_cache_size@GOTOFF(%ebx), %ebx
+# else
+	PUSH (%ebx)
+	mov	__x86_data_cache_size, %ebx
+# endif
+#endif
+	mov	%ebx, %edi
+	shr	$4, %ebx
+	sub	%ebx, %edi
+#if defined DATA_CACHE_SIZE || !defined SHARED
+	POP (%ebx)
+#endif
+/*
+ * When data size approximate the end of L1 cache,
+ * fast string will prefetch and combine data efficiently.
+ */
+	cmp	%edi, %ecx
+	jae	L(128bytesormore_nt)
+	subl	$128, %ecx
+L(128bytesormore_normal):
+	sub	$128, %ecx
+	movdqa	%xmm0, (%edx)
+	movdqa	%xmm0, 0x10(%edx)
+	movdqa	%xmm0, 0x20(%edx)
+	movdqa	%xmm0, 0x30(%edx)
+	movdqa	%xmm0, 0x40(%edx)
+	movdqa	%xmm0, 0x50(%edx)
+	movdqa	%xmm0, 0x60(%edx)
+	movdqa	%xmm0, 0x70(%edx)
+	lea	128(%edx), %edx
+	jl	L(128bytesless_normal)
+
+
+	sub	$128, %ecx
+	movdqa	%xmm0, (%edx)
+	movdqa	%xmm0, 0x10(%edx)
+	movdqa	%xmm0, 0x20(%edx)
+	movdqa	%xmm0, 0x30(%edx)
+	movdqa	%xmm0, 0x40(%edx)
+	movdqa	%xmm0, 0x50(%edx)
+	movdqa	%xmm0, 0x60(%edx)
+	movdqa	%xmm0, 0x70(%edx)
+	lea	128(%edx), %edx
+	jge	L(128bytesormore_normal)
+
+L(128bytesless_normal):
+	POP (%edi)
+	lea	128(%ecx), %ecx
+	BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes))
+
+	ALIGN (4)
+L(128bytesormore_nt):
+	mov	%edx, %edi
+	mov	%ecx, %edx
+	shr	$2, %ecx
+	and	$3, %edx
+	rep stosl
+	jz	L(copy_page_by_rep_exit)
+	cmp	$2, %edx
+	jb	L(copy_page_by_rep_left_1)
+	movw	%ax, (%edi)
+	add	$2, %edi 
+	sub	$2, %edx
+	jz	L(copy_page_by_rep_exit)
+L(copy_page_by_rep_left_1):
+	movb	%al, (%edi)
+L(copy_page_by_rep_exit):
+	POP (%edi)
+	SETRTNVAL
+	RETURN
+
+	.pushsection .rodata.sse2,"a",@progbits
+	ALIGN (2)
+L(table_16_128bytes):
+	.int	JMPTBL (L(aligned_16_0bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_1bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_2bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_3bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_4bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_5bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_6bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_7bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_8bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_9bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_10bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_11bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_12bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_13bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_14bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_15bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_16bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_17bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_18bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_19bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_20bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_21bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_22bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_23bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_24bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_25bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_26bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_27bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_28bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_29bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_30bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_31bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_32bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_33bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_34bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_35bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_36bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_37bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_38bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_39bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_40bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_41bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_42bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_43bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_44bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_45bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_46bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_47bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_48bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_49bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_50bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_51bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_52bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_53bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_54bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_55bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_56bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_57bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_58bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_59bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_60bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_61bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_62bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_63bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_64bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_65bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_66bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_67bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_68bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_69bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_70bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_71bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_72bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_73bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_74bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_75bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_76bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_77bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_78bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_79bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_80bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_81bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_82bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_83bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_84bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_85bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_86bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_87bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_88bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_89bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_90bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_91bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_92bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_93bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_94bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_95bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_96bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_97bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_98bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_99bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_100bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_101bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_102bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_103bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_104bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_105bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_106bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_107bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_108bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_109bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_110bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_111bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_112bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_113bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_114bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_115bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_116bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_117bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_118bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_119bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_120bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_121bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_122bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_123bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_124bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_125bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_126bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_127bytes), L(table_16_128bytes))
+	.popsection
+
+	ALIGN (4)
+L(aligned_16_112bytes):
+	movdqa	%xmm0, -112(%edx)
+L(aligned_16_96bytes):
+	movdqa	%xmm0, -96(%edx)
+L(aligned_16_80bytes):
+	movdqa	%xmm0, -80(%edx)
+L(aligned_16_64bytes):
+	movdqa	%xmm0, -64(%edx)
+L(aligned_16_48bytes):
+	movdqa	%xmm0, -48(%edx)
+L(aligned_16_32bytes):
+	movdqa	%xmm0, -32(%edx)
+L(aligned_16_16bytes):
+	movdqa	%xmm0, -16(%edx)
+L(aligned_16_0bytes):
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_113bytes):
+	movdqa	%xmm0, -113(%edx)
+L(aligned_16_97bytes):
+	movdqa	%xmm0, -97(%edx)
+L(aligned_16_81bytes):
+	movdqa	%xmm0, -81(%edx)
+L(aligned_16_65bytes):
+	movdqa	%xmm0, -65(%edx)
+L(aligned_16_49bytes):
+	movdqa	%xmm0, -49(%edx)
+L(aligned_16_33bytes):
+	movdqa	%xmm0, -33(%edx)
+L(aligned_16_17bytes):
+	movdqa	%xmm0, -17(%edx)
+L(aligned_16_1bytes):
+	movb	%al, -1(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_114bytes):
+	movdqa	%xmm0, -114(%edx)
+L(aligned_16_98bytes):
+	movdqa	%xmm0, -98(%edx)
+L(aligned_16_82bytes):
+	movdqa	%xmm0, -82(%edx)
+L(aligned_16_66bytes):
+	movdqa	%xmm0, -66(%edx)
+L(aligned_16_50bytes):
+	movdqa	%xmm0, -50(%edx)
+L(aligned_16_34bytes):
+	movdqa	%xmm0, -34(%edx)
+L(aligned_16_18bytes):
+	movdqa	%xmm0, -18(%edx)
+L(aligned_16_2bytes):
+	movw	%ax, -2(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_115bytes):
+	movdqa	%xmm0, -115(%edx)
+L(aligned_16_99bytes):
+	movdqa	%xmm0, -99(%edx)
+L(aligned_16_83bytes):
+	movdqa	%xmm0, -83(%edx)
+L(aligned_16_67bytes):
+	movdqa	%xmm0, -67(%edx)
+L(aligned_16_51bytes):
+	movdqa	%xmm0, -51(%edx)
+L(aligned_16_35bytes):
+	movdqa	%xmm0, -35(%edx)
+L(aligned_16_19bytes):
+	movdqa	%xmm0, -19(%edx)
+L(aligned_16_3bytes):
+	movw	%ax, -3(%edx)
+	movb	%al, -1(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_116bytes):
+	movdqa	%xmm0, -116(%edx)
+L(aligned_16_100bytes):
+	movdqa	%xmm0, -100(%edx)
+L(aligned_16_84bytes):
+	movdqa	%xmm0, -84(%edx)
+L(aligned_16_68bytes):
+	movdqa	%xmm0, -68(%edx)
+L(aligned_16_52bytes):
+	movdqa	%xmm0, -52(%edx)
+L(aligned_16_36bytes):
+	movdqa	%xmm0, -36(%edx)
+L(aligned_16_20bytes):
+	movdqa	%xmm0, -20(%edx)
+L(aligned_16_4bytes):
+	movl	%eax, -4(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_117bytes):
+	movdqa	%xmm0, -117(%edx)
+L(aligned_16_101bytes):
+	movdqa	%xmm0, -101(%edx)
+L(aligned_16_85bytes):
+	movdqa	%xmm0, -85(%edx)
+L(aligned_16_69bytes):
+	movdqa	%xmm0, -69(%edx)
+L(aligned_16_53bytes):
+	movdqa	%xmm0, -53(%edx)
+L(aligned_16_37bytes):
+	movdqa	%xmm0, -37(%edx)
+L(aligned_16_21bytes):
+	movdqa	%xmm0, -21(%edx)
+L(aligned_16_5bytes):
+	movl	%eax, -5(%edx)
+	movb	%al, -1(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_118bytes):
+	movdqa	%xmm0, -118(%edx)
+L(aligned_16_102bytes):
+	movdqa	%xmm0, -102(%edx)
+L(aligned_16_86bytes):
+	movdqa	%xmm0, -86(%edx)
+L(aligned_16_70bytes):
+	movdqa	%xmm0, -70(%edx)
+L(aligned_16_54bytes):
+	movdqa	%xmm0, -54(%edx)
+L(aligned_16_38bytes):
+	movdqa	%xmm0, -38(%edx)
+L(aligned_16_22bytes):
+	movdqa	%xmm0, -22(%edx)
+L(aligned_16_6bytes):
+	movl	%eax, -6(%edx)
+	movw	%ax, -2(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_119bytes):
+	movdqa	%xmm0, -119(%edx)
+L(aligned_16_103bytes):
+	movdqa	%xmm0, -103(%edx)
+L(aligned_16_87bytes):
+	movdqa	%xmm0, -87(%edx)
+L(aligned_16_71bytes):
+	movdqa	%xmm0, -71(%edx)
+L(aligned_16_55bytes):
+	movdqa	%xmm0, -55(%edx)
+L(aligned_16_39bytes):
+	movdqa	%xmm0, -39(%edx)
+L(aligned_16_23bytes):
+	movdqa	%xmm0, -23(%edx)
+L(aligned_16_7bytes):
+	movl	%eax, -7(%edx)
+	movw	%ax, -3(%edx)
+	movb	%al, -1(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_120bytes):
+	movdqa	%xmm0, -120(%edx)
+L(aligned_16_104bytes):
+	movdqa	%xmm0, -104(%edx)
+L(aligned_16_88bytes):
+	movdqa	%xmm0, -88(%edx)
+L(aligned_16_72bytes):
+	movdqa	%xmm0, -72(%edx)
+L(aligned_16_56bytes):
+	movdqa	%xmm0, -56(%edx)
+L(aligned_16_40bytes):
+	movdqa	%xmm0, -40(%edx)
+L(aligned_16_24bytes):
+	movdqa	%xmm0, -24(%edx)
+L(aligned_16_8bytes):
+	movq	%xmm0, -8(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_121bytes):
+	movdqa	%xmm0, -121(%edx)
+L(aligned_16_105bytes):
+	movdqa	%xmm0, -105(%edx)
+L(aligned_16_89bytes):
+	movdqa	%xmm0, -89(%edx)
+L(aligned_16_73bytes):
+	movdqa	%xmm0, -73(%edx)
+L(aligned_16_57bytes):
+	movdqa	%xmm0, -57(%edx)
+L(aligned_16_41bytes):
+	movdqa	%xmm0, -41(%edx)
+L(aligned_16_25bytes):
+	movdqa	%xmm0, -25(%edx)
+L(aligned_16_9bytes):
+	movq	%xmm0, -9(%edx)
+	movb	%al, -1(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_122bytes):
+	movdqa	%xmm0, -122(%edx)
+L(aligned_16_106bytes):
+	movdqa	%xmm0, -106(%edx)
+L(aligned_16_90bytes):
+	movdqa	%xmm0, -90(%edx)
+L(aligned_16_74bytes):
+	movdqa	%xmm0, -74(%edx)
+L(aligned_16_58bytes):
+	movdqa	%xmm0, -58(%edx)
+L(aligned_16_42bytes):
+	movdqa	%xmm0, -42(%edx)
+L(aligned_16_26bytes):
+	movdqa	%xmm0, -26(%edx)
+L(aligned_16_10bytes):
+	movq	%xmm0, -10(%edx)
+	movw	%ax, -2(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_123bytes):
+	movdqa	%xmm0, -123(%edx)
+L(aligned_16_107bytes):
+	movdqa	%xmm0, -107(%edx)
+L(aligned_16_91bytes):
+	movdqa	%xmm0, -91(%edx)
+L(aligned_16_75bytes):
+	movdqa	%xmm0, -75(%edx)
+L(aligned_16_59bytes):
+	movdqa	%xmm0, -59(%edx)
+L(aligned_16_43bytes):
+	movdqa	%xmm0, -43(%edx)
+L(aligned_16_27bytes):
+	movdqa	%xmm0, -27(%edx)
+L(aligned_16_11bytes):
+	movq	%xmm0, -11(%edx)
+	movw	%ax, -3(%edx)
+	movb	%al, -1(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_124bytes):
+	movdqa	%xmm0, -124(%edx)
+L(aligned_16_108bytes):
+	movdqa	%xmm0, -108(%edx)
+L(aligned_16_92bytes):
+	movdqa	%xmm0, -92(%edx)
+L(aligned_16_76bytes):
+	movdqa	%xmm0, -76(%edx)
+L(aligned_16_60bytes):
+	movdqa	%xmm0, -60(%edx)
+L(aligned_16_44bytes):
+	movdqa	%xmm0, -44(%edx)
+L(aligned_16_28bytes):
+	movdqa	%xmm0, -28(%edx)
+L(aligned_16_12bytes):
+	movq	%xmm0, -12(%edx)
+	movl	%eax, -4(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_125bytes):
+	movdqa	%xmm0, -125(%edx)
+L(aligned_16_109bytes):
+	movdqa	%xmm0, -109(%edx)
+L(aligned_16_93bytes):
+	movdqa	%xmm0, -93(%edx)
+L(aligned_16_77bytes):
+	movdqa	%xmm0, -77(%edx)
+L(aligned_16_61bytes):
+	movdqa	%xmm0, -61(%edx)
+L(aligned_16_45bytes):
+	movdqa	%xmm0, -45(%edx)
+L(aligned_16_29bytes):
+	movdqa	%xmm0, -29(%edx)
+L(aligned_16_13bytes):
+	movq	%xmm0, -13(%edx)
+	movl	%eax, -5(%edx)
+	movb	%al, -1(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_126bytes):
+	movdqa	%xmm0, -126(%edx)
+L(aligned_16_110bytes):
+	movdqa	%xmm0, -110(%edx)
+L(aligned_16_94bytes):
+	movdqa	%xmm0, -94(%edx)
+L(aligned_16_78bytes):
+	movdqa	%xmm0, -78(%edx)
+L(aligned_16_62bytes):
+	movdqa	%xmm0, -62(%edx)
+L(aligned_16_46bytes):
+	movdqa	%xmm0, -46(%edx)
+L(aligned_16_30bytes):
+	movdqa	%xmm0, -30(%edx)
+L(aligned_16_14bytes):
+	movq	%xmm0, -14(%edx)
+	movl	%eax, -6(%edx)
+	movw	%ax, -2(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_127bytes):
+	movdqa	%xmm0, -127(%edx)
+L(aligned_16_111bytes):
+	movdqa	%xmm0, -111(%edx)
+L(aligned_16_95bytes):
+	movdqa	%xmm0, -95(%edx)
+L(aligned_16_79bytes):
+	movdqa	%xmm0, -79(%edx)
+L(aligned_16_63bytes):
+	movdqa	%xmm0, -63(%edx)
+L(aligned_16_47bytes):
+	movdqa	%xmm0, -47(%edx)
+L(aligned_16_31bytes):
+	movdqa	%xmm0, -31(%edx)
+L(aligned_16_15bytes):
+	movq	%xmm0, -15(%edx)
+	movl	%eax, -7(%edx)
+	movw	%ax, -3(%edx)
+	movb	%al, -1(%edx)
+	SETRTNVAL
+	RETURN_END
+
+END (__memset_sse2_rep)
+
+#endif
diff --git a/sysdeps/i386/i686/multiarch/memset-sse2.S b/sysdeps/i386/i686/multiarch/memset-sse2.S
new file mode 100644
index 0000000000..02fe57f6b0
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/memset-sse2.S
@@ -0,0 +1,867 @@
+/* memset with SSE2
+   Copyright (C) 2010 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>
+#include "asm-syntax.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)
+
+#ifdef USE_AS_BZERO
+# define DEST		PARMS
+# define LEN		DEST+4
+# define SETRTNVAL
+#else
+# define DEST		PARMS
+# define CHR		DEST+4
+# define LEN		CHR+4
+# define SETRTNVAL	movl DEST(%esp), %eax
+#endif
+
+#ifdef SHARED
+# define ENTRANCE	PUSH (%ebx);
+# define RETURN_END	POP (%ebx); ret
+# define RETURN		RETURN_END; CFI_PUSH (%ebx)
+# define PARMS		8		/* Preserve EBX.  */
+# define JMPTBL(I, B)	I - B
+
+/* Load an entry in a jump table into EBX and branch to it.  TABLE is a
+   jump table with relative offsets.   */
+# define BRANCH_TO_JMPTBL_ENTRY(TABLE)				\
+    /* We first load PC into EBX.  */				\
+    call	__i686.get_pc_thunk.bx;				\
+    /* Get the address of the jump table.  */			\
+    add		$(TABLE - .), %ebx;				\
+    /* Get the entry and convert the relative offset to the	\
+       absolute address.  */					\
+    add		(%ebx,%ecx,4), %ebx;				\
+    add		%ecx, %edx;					\
+    /* We loaded the jump table and adjuested EDX. Go.  */	\
+    jmp		*%ebx
+
+	.section	.gnu.linkonce.t.__i686.get_pc_thunk.bx,"ax",@progbits
+	.globl	__i686.get_pc_thunk.bx
+	.hidden	__i686.get_pc_thunk.bx
+	ALIGN (4)
+	.type	__i686.get_pc_thunk.bx,@function
+__i686.get_pc_thunk.bx:
+	movl	(%esp), %ebx
+	ret
+#else
+# define ENTRANCE
+# define RETURN_END	ret
+# define RETURN		RETURN_END
+# define PARMS		4
+# define JMPTBL(I, B)	I
+
+/* Branch to an entry in a jump table.  TABLE is a jump table with
+   absolute offsets.  */
+# define BRANCH_TO_JMPTBL_ENTRY(TABLE)				\
+    add		%ecx, %edx;					\
+    jmp		*TABLE(,%ecx,4)
+#endif
+
+	.section .text.sse2,"ax",@progbits
+#if defined SHARED && !defined NOT_IN_libc && !defined USE_AS_BZERO
+ENTRY (__memset_chk_sse2)
+	movl	12(%esp), %eax
+	cmpl	%eax, 16(%esp)
+	jb	HIDDEN_JUMPTARGET (__chk_fail)
+END (__memset_chk_sse2)
+#endif
+ENTRY (__memset_sse2)
+	ENTRANCE
+
+	movl	LEN(%esp), %ecx
+#ifdef USE_AS_BZERO
+	xor	%eax, %eax
+#else
+	movzbl	CHR(%esp), %eax
+	movb	%al, %ah
+	/* Fill the whole EAX with pattern.  */
+	movl	%eax, %edx
+	shl	$16, %eax
+	or	%edx, %eax
+#endif
+	movl	DEST(%esp), %edx
+	cmp	$32, %ecx
+	jae	L(32bytesormore)
+
+L(write_less32bytes):
+	BRANCH_TO_JMPTBL_ENTRY (L(table_less_32bytes))
+
+
+	.pushsection .rodata.sse2,"a",@progbits
+	ALIGN (2)
+L(table_less_32bytes):
+	.int	JMPTBL (L(write_0bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_1bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_2bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_3bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_4bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_5bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_6bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_7bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_8bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_9bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_10bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_11bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_12bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_13bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_14bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_15bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_16bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_17bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_18bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_19bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_20bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_21bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_22bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_23bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_24bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_25bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_26bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_27bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_28bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_29bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_30bytes), L(table_less_32bytes))
+	.int	JMPTBL (L(write_31bytes), L(table_less_32bytes))
+	.popsection
+
+	ALIGN (4)
+L(write_28bytes):
+	movl	%eax, -28(%edx)
+L(write_24bytes):
+	movl	%eax, -24(%edx)
+L(write_20bytes):
+	movl	%eax, -20(%edx)
+L(write_16bytes):
+	movl	%eax, -16(%edx)
+L(write_12bytes):
+	movl	%eax, -12(%edx)
+L(write_8bytes):
+	movl	%eax, -8(%edx)
+L(write_4bytes):
+	movl	%eax, -4(%edx)
+L(write_0bytes):
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(write_29bytes):
+	movl	%eax, -29(%edx)
+L(write_25bytes):
+	movl	%eax, -25(%edx)
+L(write_21bytes):
+	movl	%eax, -21(%edx)
+L(write_17bytes):
+	movl	%eax, -17(%edx)
+L(write_13bytes):
+	movl	%eax, -13(%edx)
+L(write_9bytes):
+	movl	%eax, -9(%edx)
+L(write_5bytes):
+	movl	%eax, -5(%edx)
+L(write_1bytes):
+	movb	%al, -1(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(write_30bytes):
+	movl	%eax, -30(%edx)
+L(write_26bytes):
+	movl	%eax, -26(%edx)
+L(write_22bytes):
+	movl	%eax, -22(%edx)
+L(write_18bytes):
+	movl	%eax, -18(%edx)
+L(write_14bytes):
+	movl	%eax, -14(%edx)
+L(write_10bytes):
+	movl	%eax, -10(%edx)
+L(write_6bytes):
+	movl	%eax, -6(%edx)
+L(write_2bytes):
+	movw	%ax, -2(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(write_31bytes):
+	movl	%eax, -31(%edx)
+L(write_27bytes):
+	movl	%eax, -27(%edx)
+L(write_23bytes):
+	movl	%eax, -23(%edx)
+L(write_19bytes):
+	movl	%eax, -19(%edx)
+L(write_15bytes):
+	movl	%eax, -15(%edx)
+L(write_11bytes):
+	movl	%eax, -11(%edx)
+L(write_7bytes):
+	movl	%eax, -7(%edx)
+L(write_3bytes):
+	movw	%ax, -3(%edx)
+	movb	%al, -1(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+/* ECX > 32 and EDX is 4 byte aligned.  */
+L(32bytesormore):
+	/* Fill xmm0 with the pattern.  */
+#ifdef USE_AS_BZERO
+	pxor	%xmm0, %xmm0
+#else
+	movd	%eax, %xmm0
+	punpcklbw %xmm0, %xmm0
+	pshufd	$0, %xmm0, %xmm0
+#endif
+	testl	$0xf, %edx
+	jz	L(aligned_16)
+/* ECX > 32 and EDX is not 16 byte aligned.  */
+L(not_aligned_16):
+	movdqu	%xmm0, (%edx)
+	movl	%edx, %eax
+	and	$-16, %edx
+	add	$16, %edx
+	sub	%edx, %eax
+	add	%eax, %ecx
+	movd	%xmm0, %eax
+
+	ALIGN (4)
+L(aligned_16):
+	cmp	$128, %ecx
+	jge	L(128bytesormore)
+
+L(aligned_16_less128bytes):
+	BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes))
+
+	ALIGN (4)
+L(128bytesormore):
+#ifdef SHARED_CACHE_SIZE
+	PUSH (%ebx)
+	mov	$SHARED_CACHE_SIZE, %ebx
+#else
+# ifdef SHARED
+	call	__i686.get_pc_thunk.bx
+	add	$_GLOBAL_OFFSET_TABLE_, %ebx
+	mov	__x86_shared_cache_size@GOTOFF(%ebx), %ebx
+# else
+	PUSH (%ebx)
+	mov	__x86_shared_cache_size, %ebx
+# endif
+#endif
+	cmp	%ebx, %ecx
+	jae	L(128bytesormore_nt_start)
+
+	
+#ifdef DATA_CACHE_SIZE
+	POP (%ebx)
+	cmp	$DATA_CACHE_SIZE, %ecx
+#else
+# ifdef SHARED
+	call	__i686.get_pc_thunk.bx
+	add	$_GLOBAL_OFFSET_TABLE_, %ebx
+	cmp	__x86_data_cache_size@GOTOFF(%ebx), %ecx
+# else
+	POP (%ebx)
+	cmp	__x86_data_cache_size, %ecx
+# endif
+#endif
+
+	jae	L(128bytes_L2_normal)
+	subl	$128, %ecx
+L(128bytesormore_normal):
+	sub	$128, %ecx
+	movdqa	%xmm0, (%edx)
+	movdqa	%xmm0, 0x10(%edx)
+	movdqa	%xmm0, 0x20(%edx)
+	movdqa	%xmm0, 0x30(%edx)
+	movdqa	%xmm0, 0x40(%edx)
+	movdqa	%xmm0, 0x50(%edx)
+	movdqa	%xmm0, 0x60(%edx)
+	movdqa	%xmm0, 0x70(%edx)
+	lea	128(%edx), %edx
+	jl	L(128bytesless_normal)
+
+
+	sub	$128, %ecx
+	movdqa	%xmm0, (%edx)
+	movdqa	%xmm0, 0x10(%edx)
+	movdqa	%xmm0, 0x20(%edx)
+	movdqa	%xmm0, 0x30(%edx)
+	movdqa	%xmm0, 0x40(%edx)
+	movdqa	%xmm0, 0x50(%edx)
+	movdqa	%xmm0, 0x60(%edx)
+	movdqa	%xmm0, 0x70(%edx)
+	lea	128(%edx), %edx
+	jge	L(128bytesormore_normal)
+
+L(128bytesless_normal):
+	lea	128(%ecx), %ecx
+	BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes))
+
+	ALIGN (4)
+L(128bytes_L2_normal):
+	prefetcht0	0x380(%edx)
+	prefetcht0	0x3c0(%edx)
+	sub	$128, %ecx
+	movdqa	%xmm0, (%edx)
+	movaps	%xmm0, 0x10(%edx)
+	movaps	%xmm0, 0x20(%edx)
+	movaps	%xmm0, 0x30(%edx)
+	movaps	%xmm0, 0x40(%edx)
+	movaps	%xmm0, 0x50(%edx)
+	movaps	%xmm0, 0x60(%edx)
+	movaps	%xmm0, 0x70(%edx)
+	add	$128, %edx
+	cmp	$128, %ecx 	
+	jge	L(128bytes_L2_normal)
+
+L(128bytesless_L2_normal):
+	BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes))
+
+L(128bytesormore_nt_start):
+	sub	%ebx, %ecx
+	ALIGN (4)
+L(128bytesormore_shared_cache_loop):
+	prefetcht0	0x3c0(%edx)
+	prefetcht0	0x380(%edx)
+	sub	$0x80, %ebx
+	movdqa	%xmm0, (%edx)
+	movdqa	%xmm0, 0x10(%edx)
+	movdqa	%xmm0, 0x20(%edx)
+	movdqa	%xmm0, 0x30(%edx)
+	movdqa	%xmm0, 0x40(%edx)
+	movdqa	%xmm0, 0x50(%edx)
+	movdqa	%xmm0, 0x60(%edx)
+	movdqa	%xmm0, 0x70(%edx)
+	add	$0x80, %edx
+	cmp	$0x80, %ebx
+	jge	L(128bytesormore_shared_cache_loop)
+	cmp	$0x80, %ecx
+	jb	L(shared_cache_loop_end)
+	ALIGN (4)
+L(128bytesormore_nt):
+	sub	$0x80, %ecx
+	movntdq	%xmm0, (%edx)
+	movntdq	%xmm0, 0x10(%edx)
+	movntdq	%xmm0, 0x20(%edx)
+	movntdq	%xmm0, 0x30(%edx)
+	movntdq	%xmm0, 0x40(%edx)
+	movntdq	%xmm0, 0x50(%edx)
+	movntdq	%xmm0, 0x60(%edx)
+	movntdq	%xmm0, 0x70(%edx)
+	add	$0x80, %edx
+	cmp	$0x80, %ecx
+	jge	L(128bytesormore_nt)
+	sfence
+L(shared_cache_loop_end):
+#if defined DATA_CACHE_SIZE || !defined SHARED
+	POP (%ebx)
+#endif
+	BRANCH_TO_JMPTBL_ENTRY (L(table_16_128bytes))
+
+
+	.pushsection .rodata.sse2,"a",@progbits
+	ALIGN (2)
+L(table_16_128bytes):
+	.int	JMPTBL (L(aligned_16_0bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_1bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_2bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_3bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_4bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_5bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_6bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_7bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_8bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_9bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_10bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_11bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_12bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_13bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_14bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_15bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_16bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_17bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_18bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_19bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_20bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_21bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_22bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_23bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_24bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_25bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_26bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_27bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_28bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_29bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_30bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_31bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_32bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_33bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_34bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_35bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_36bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_37bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_38bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_39bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_40bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_41bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_42bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_43bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_44bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_45bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_46bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_47bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_48bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_49bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_50bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_51bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_52bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_53bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_54bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_55bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_56bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_57bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_58bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_59bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_60bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_61bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_62bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_63bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_64bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_65bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_66bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_67bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_68bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_69bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_70bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_71bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_72bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_73bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_74bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_75bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_76bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_77bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_78bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_79bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_80bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_81bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_82bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_83bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_84bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_85bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_86bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_87bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_88bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_89bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_90bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_91bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_92bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_93bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_94bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_95bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_96bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_97bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_98bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_99bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_100bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_101bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_102bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_103bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_104bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_105bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_106bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_107bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_108bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_109bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_110bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_111bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_112bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_113bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_114bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_115bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_116bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_117bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_118bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_119bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_120bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_121bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_122bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_123bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_124bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_125bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_126bytes), L(table_16_128bytes))
+	.int	JMPTBL (L(aligned_16_127bytes), L(table_16_128bytes))
+	.popsection
+
+	ALIGN (4)
+L(aligned_16_112bytes):
+	movdqa	%xmm0, -112(%edx)
+L(aligned_16_96bytes):
+	movdqa	%xmm0, -96(%edx)
+L(aligned_16_80bytes):
+	movdqa	%xmm0, -80(%edx)
+L(aligned_16_64bytes):
+	movdqa	%xmm0, -64(%edx)
+L(aligned_16_48bytes):
+	movdqa	%xmm0, -48(%edx)
+L(aligned_16_32bytes):
+	movdqa	%xmm0, -32(%edx)
+L(aligned_16_16bytes):
+	movdqa	%xmm0, -16(%edx)
+L(aligned_16_0bytes):
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_113bytes):
+	movdqa	%xmm0, -113(%edx)
+L(aligned_16_97bytes):
+	movdqa	%xmm0, -97(%edx)
+L(aligned_16_81bytes):
+	movdqa	%xmm0, -81(%edx)
+L(aligned_16_65bytes):
+	movdqa	%xmm0, -65(%edx)
+L(aligned_16_49bytes):
+	movdqa	%xmm0, -49(%edx)
+L(aligned_16_33bytes):
+	movdqa	%xmm0, -33(%edx)
+L(aligned_16_17bytes):
+	movdqa	%xmm0, -17(%edx)
+L(aligned_16_1bytes):
+	movb	%al, -1(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_114bytes):
+	movdqa	%xmm0, -114(%edx)
+L(aligned_16_98bytes):
+	movdqa	%xmm0, -98(%edx)
+L(aligned_16_82bytes):
+	movdqa	%xmm0, -82(%edx)
+L(aligned_16_66bytes):
+	movdqa	%xmm0, -66(%edx)
+L(aligned_16_50bytes):
+	movdqa	%xmm0, -50(%edx)
+L(aligned_16_34bytes):
+	movdqa	%xmm0, -34(%edx)
+L(aligned_16_18bytes):
+	movdqa	%xmm0, -18(%edx)
+L(aligned_16_2bytes):
+	movw	%ax, -2(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_115bytes):
+	movdqa	%xmm0, -115(%edx)
+L(aligned_16_99bytes):
+	movdqa	%xmm0, -99(%edx)
+L(aligned_16_83bytes):
+	movdqa	%xmm0, -83(%edx)
+L(aligned_16_67bytes):
+	movdqa	%xmm0, -67(%edx)
+L(aligned_16_51bytes):
+	movdqa	%xmm0, -51(%edx)
+L(aligned_16_35bytes):
+	movdqa	%xmm0, -35(%edx)
+L(aligned_16_19bytes):
+	movdqa	%xmm0, -19(%edx)
+L(aligned_16_3bytes):
+	movw	%ax, -3(%edx)
+	movb	%al, -1(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_116bytes):
+	movdqa	%xmm0, -116(%edx)
+L(aligned_16_100bytes):
+	movdqa	%xmm0, -100(%edx)
+L(aligned_16_84bytes):
+	movdqa	%xmm0, -84(%edx)
+L(aligned_16_68bytes):
+	movdqa	%xmm0, -68(%edx)
+L(aligned_16_52bytes):
+	movdqa	%xmm0, -52(%edx)
+L(aligned_16_36bytes):
+	movdqa	%xmm0, -36(%edx)
+L(aligned_16_20bytes):
+	movdqa	%xmm0, -20(%edx)
+L(aligned_16_4bytes):
+	movl	%eax, -4(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_117bytes):
+	movdqa	%xmm0, -117(%edx)
+L(aligned_16_101bytes):
+	movdqa	%xmm0, -101(%edx)
+L(aligned_16_85bytes):
+	movdqa	%xmm0, -85(%edx)
+L(aligned_16_69bytes):
+	movdqa	%xmm0, -69(%edx)
+L(aligned_16_53bytes):
+	movdqa	%xmm0, -53(%edx)
+L(aligned_16_37bytes):
+	movdqa	%xmm0, -37(%edx)
+L(aligned_16_21bytes):
+	movdqa	%xmm0, -21(%edx)
+L(aligned_16_5bytes):
+	movl	%eax, -5(%edx)
+	movb	%al, -1(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_118bytes):
+	movdqa	%xmm0, -118(%edx)
+L(aligned_16_102bytes):
+	movdqa	%xmm0, -102(%edx)
+L(aligned_16_86bytes):
+	movdqa	%xmm0, -86(%edx)
+L(aligned_16_70bytes):
+	movdqa	%xmm0, -70(%edx)
+L(aligned_16_54bytes):
+	movdqa	%xmm0, -54(%edx)
+L(aligned_16_38bytes):
+	movdqa	%xmm0, -38(%edx)
+L(aligned_16_22bytes):
+	movdqa	%xmm0, -22(%edx)
+L(aligned_16_6bytes):
+	movl	%eax, -6(%edx)
+	movw	%ax, -2(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_119bytes):
+	movdqa	%xmm0, -119(%edx)
+L(aligned_16_103bytes):
+	movdqa	%xmm0, -103(%edx)
+L(aligned_16_87bytes):
+	movdqa	%xmm0, -87(%edx)
+L(aligned_16_71bytes):
+	movdqa	%xmm0, -71(%edx)
+L(aligned_16_55bytes):
+	movdqa	%xmm0, -55(%edx)
+L(aligned_16_39bytes):
+	movdqa	%xmm0, -39(%edx)
+L(aligned_16_23bytes):
+	movdqa	%xmm0, -23(%edx)
+L(aligned_16_7bytes):
+	movl	%eax, -7(%edx)
+	movw	%ax, -3(%edx)
+	movb	%al, -1(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_120bytes):
+	movdqa	%xmm0, -120(%edx)
+L(aligned_16_104bytes):
+	movdqa	%xmm0, -104(%edx)
+L(aligned_16_88bytes):
+	movdqa	%xmm0, -88(%edx)
+L(aligned_16_72bytes):
+	movdqa	%xmm0, -72(%edx)
+L(aligned_16_56bytes):
+	movdqa	%xmm0, -56(%edx)
+L(aligned_16_40bytes):
+	movdqa	%xmm0, -40(%edx)
+L(aligned_16_24bytes):
+	movdqa	%xmm0, -24(%edx)
+L(aligned_16_8bytes):
+	movq	%xmm0, -8(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_121bytes):
+	movdqa	%xmm0, -121(%edx)
+L(aligned_16_105bytes):
+	movdqa	%xmm0, -105(%edx)
+L(aligned_16_89bytes):
+	movdqa	%xmm0, -89(%edx)
+L(aligned_16_73bytes):
+	movdqa	%xmm0, -73(%edx)
+L(aligned_16_57bytes):
+	movdqa	%xmm0, -57(%edx)
+L(aligned_16_41bytes):
+	movdqa	%xmm0, -41(%edx)
+L(aligned_16_25bytes):
+	movdqa	%xmm0, -25(%edx)
+L(aligned_16_9bytes):
+	movq	%xmm0, -9(%edx)
+	movb	%al, -1(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_122bytes):
+	movdqa	%xmm0, -122(%edx)
+L(aligned_16_106bytes):
+	movdqa	%xmm0, -106(%edx)
+L(aligned_16_90bytes):
+	movdqa	%xmm0, -90(%edx)
+L(aligned_16_74bytes):
+	movdqa	%xmm0, -74(%edx)
+L(aligned_16_58bytes):
+	movdqa	%xmm0, -58(%edx)
+L(aligned_16_42bytes):
+	movdqa	%xmm0, -42(%edx)
+L(aligned_16_26bytes):
+	movdqa	%xmm0, -26(%edx)
+L(aligned_16_10bytes):
+	movq	%xmm0, -10(%edx)
+	movw	%ax, -2(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_123bytes):
+	movdqa	%xmm0, -123(%edx)
+L(aligned_16_107bytes):
+	movdqa	%xmm0, -107(%edx)
+L(aligned_16_91bytes):
+	movdqa	%xmm0, -91(%edx)
+L(aligned_16_75bytes):
+	movdqa	%xmm0, -75(%edx)
+L(aligned_16_59bytes):
+	movdqa	%xmm0, -59(%edx)
+L(aligned_16_43bytes):
+	movdqa	%xmm0, -43(%edx)
+L(aligned_16_27bytes):
+	movdqa	%xmm0, -27(%edx)
+L(aligned_16_11bytes):
+	movq	%xmm0, -11(%edx)
+	movw	%ax, -3(%edx)
+	movb	%al, -1(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_124bytes):
+	movdqa	%xmm0, -124(%edx)
+L(aligned_16_108bytes):
+	movdqa	%xmm0, -108(%edx)
+L(aligned_16_92bytes):
+	movdqa	%xmm0, -92(%edx)
+L(aligned_16_76bytes):
+	movdqa	%xmm0, -76(%edx)
+L(aligned_16_60bytes):
+	movdqa	%xmm0, -60(%edx)
+L(aligned_16_44bytes):
+	movdqa	%xmm0, -44(%edx)
+L(aligned_16_28bytes):
+	movdqa	%xmm0, -28(%edx)
+L(aligned_16_12bytes):
+	movq	%xmm0, -12(%edx)
+	movl	%eax, -4(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_125bytes):
+	movdqa	%xmm0, -125(%edx)
+L(aligned_16_109bytes):
+	movdqa	%xmm0, -109(%edx)
+L(aligned_16_93bytes):
+	movdqa	%xmm0, -93(%edx)
+L(aligned_16_77bytes):
+	movdqa	%xmm0, -77(%edx)
+L(aligned_16_61bytes):
+	movdqa	%xmm0, -61(%edx)
+L(aligned_16_45bytes):
+	movdqa	%xmm0, -45(%edx)
+L(aligned_16_29bytes):
+	movdqa	%xmm0, -29(%edx)
+L(aligned_16_13bytes):
+	movq	%xmm0, -13(%edx)
+	movl	%eax, -5(%edx)
+	movb	%al, -1(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_126bytes):
+	movdqa	%xmm0, -126(%edx)
+L(aligned_16_110bytes):
+	movdqa	%xmm0, -110(%edx)
+L(aligned_16_94bytes):
+	movdqa	%xmm0, -94(%edx)
+L(aligned_16_78bytes):
+	movdqa	%xmm0, -78(%edx)
+L(aligned_16_62bytes):
+	movdqa	%xmm0, -62(%edx)
+L(aligned_16_46bytes):
+	movdqa	%xmm0, -46(%edx)
+L(aligned_16_30bytes):
+	movdqa	%xmm0, -30(%edx)
+L(aligned_16_14bytes):
+	movq	%xmm0, -14(%edx)
+	movl	%eax, -6(%edx)
+	movw	%ax, -2(%edx)
+	SETRTNVAL
+	RETURN
+
+	ALIGN (4)
+L(aligned_16_127bytes):
+	movdqa	%xmm0, -127(%edx)
+L(aligned_16_111bytes):
+	movdqa	%xmm0, -111(%edx)
+L(aligned_16_95bytes):
+	movdqa	%xmm0, -95(%edx)
+L(aligned_16_79bytes):
+	movdqa	%xmm0, -79(%edx)
+L(aligned_16_63bytes):
+	movdqa	%xmm0, -63(%edx)
+L(aligned_16_47bytes):
+	movdqa	%xmm0, -47(%edx)
+L(aligned_16_31bytes):
+	movdqa	%xmm0, -31(%edx)
+L(aligned_16_15bytes):
+	movq	%xmm0, -15(%edx)
+	movl	%eax, -7(%edx)
+	movw	%ax, -3(%edx)
+	movb	%al, -1(%edx)
+	SETRTNVAL
+	RETURN_END
+
+END (__memset_sse2)
+
+#endif
diff --git a/sysdeps/i386/i686/multiarch/memset.S b/sysdeps/i386/i686/multiarch/memset.S
new file mode 100644
index 0000000000..34dddcef7b
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/memset.S
@@ -0,0 +1,112 @@
+/* Multiple versions of memset
+   Copyright (C) 2010 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>
+
+/* Define multiple versions only for the definition in lib.  */
+#ifndef NOT_IN_libc
+# ifdef SHARED
+	.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
+
+	.text
+ENTRY(memset)
+	.type	memset, @gnu_indirect_function
+	pushl	%ebx
+	cfi_adjust_cfa_offset (4)
+	cfi_rel_offset (ebx, 0)
+	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:	leal	__memset_ia32@GOTOFF(%ebx), %eax
+	testl	$bit_SSE2, CPUID_OFFSET+index_SSE2+__cpu_features@GOTOFF(%ebx)
+	jz	2f
+	leal	__memset_sse2@GOTOFF(%ebx), %eax
+	testl	$bit_Fast_Rep_String, FEATURE_OFFSET+index_Fast_Rep_String+__cpu_features@GOTOFF(%ebx)
+	jz	2f
+	leal	__memset_sse2_rep@GOTOFF(%ebx), %eax
+2:	popl	%ebx
+	cfi_adjust_cfa_offset (-4)
+	cfi_restore (ebx)
+	ret
+END(memset)
+# else
+	.text
+ENTRY(memset)
+	.type	memset, @gnu_indirect_function
+	cmpl	$0, KIND_OFFSET+__cpu_features
+	jne	1f
+	call	__init_cpu_features
+1:	leal	__memset_ia32, %eax
+	testl	$bit_SSE2, CPUID_OFFSET+index_SSE2+__cpu_features
+	jz	2f
+	leal	__memset_sse2, %eax
+	testl	$bit_Fast_Rep_String, FEATURE_OFFSET+index_Fast_Rep_String+__cpu_features
+	jz	2f
+	leal	__memset_sse2_rep, %eax
+2:	ret
+END(memset)
+# endif
+
+# undef ENTRY
+# define ENTRY(name) \
+	.type __memset_ia32, @function; \
+	.globl __memset_ia32; \
+	.p2align 4; \
+	__memset_ia32: cfi_startproc; \
+	CALL_MCOUNT
+# undef END
+# define END(name) \
+	cfi_endproc; .size __memset_ia32, .-__memset_ia32
+
+# undef ENTRY_CHK
+# define ENTRY_CHK(name) \
+	.type __memset_chk_ia32, @function; \
+	.globl __memset_chk_ia32; \
+	.p2align 4; \
+	__memset_chk_ia32: cfi_startproc; \
+	CALL_MCOUNT
+# undef END_CHK
+# define END_CHK(name) \
+	cfi_endproc; .size __memset_chk_ia32, .-__memset_chk_ia32
+
+# ifdef SHARED
+#  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_memset; __GI_memset = __memset_ia32
+# endif
+
+# undef strong_alias
+# define strong_alias(original, alias)
+#endif
+
+#include "../memset.S"
diff --git a/sysdeps/i386/i686/multiarch/memset_chk.S b/sysdeps/i386/i686/multiarch/memset_chk.S
new file mode 100644
index 0000000000..e3665be7b2
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/memset_chk.S
@@ -0,0 +1,116 @@
+/* Multiple versions of __memset_chk
+   Copyright (C) 2010 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>
+
+/* Define multiple versions only for the definition in lib.  */
+#ifndef NOT_IN_libc
+# ifdef SHARED
+	.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
+
+	.text
+ENTRY(__memset_chk)
+	.type	__memset_chk, @gnu_indirect_function
+	pushl	%ebx
+	cfi_adjust_cfa_offset (4)
+	cfi_rel_offset (ebx, 0)
+	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:	leal	__memset_chk_ia32@GOTOFF(%ebx), %eax
+	testl	$bit_SSE2, CPUID_OFFSET+index_SSE2+__cpu_features@GOTOFF(%ebx)
+	jz	2f
+	leal	__memset_chk_sse2@GOTOFF(%ebx), %eax
+	testl	$bit_Fast_Rep_String, FEATURE_OFFSET+index_Fast_Rep_String+__cpu_features@GOTOFF(%ebx)
+	jz	2f
+	leal	__memset_chk_sse2_rep@GOTOFF(%ebx), %eax
+2:	popl	%ebx
+	cfi_adjust_cfa_offset (-4)
+	cfi_restore (ebx)
+	ret
+END(__memset_chk)
+
+strong_alias (__memset_chk, __memset_zero_constant_len_parameter)
+	.section .gnu.warning.__memset_zero_constant_len_parameter
+	.string "memset used with constant zero length parameter; this could be due to transposed parameters"
+# else
+	.text
+ENTRY(__memset_chk)
+	.type	__memset_chk, @gnu_indirect_function
+	cmpl	$0, KIND_OFFSET+__cpu_features
+	jne	1f
+	call	__init_cpu_features
+1:	leal	__memset_chk_ia32, %eax
+	testl	$bit_SSE2, CPUID_OFFSET+index_SSE2+__cpu_features
+	jz	2f
+	leal	__memset_chk_sse2, %eax
+	testl	$bit_Fast_Rep_String, FEATURE_OFFSET+index_Fast_Rep_String+__cpu_features
+	jz	2f
+	leal	__memset_chk_sse2_rep, %eax
+2:	ret
+END(__memset_chk)
+
+	.type __memset_chk_sse2, @function
+	.p2align 4;
+__memset_chk_sse2:
+	cfi_startproc
+	CALL_MCOUNT
+	movl	12(%esp), %eax
+	cmpl	%eax, 16(%esp)
+	jb	__chk_fail
+	jmp	__memset_sse2
+	cfi_endproc 
+	.size __memset_chk_sse2, .-__memset_chk_sse2
+
+	.type __memset_chk_sse2_rep, @function
+	.p2align 4;
+__memset_chk_sse2_rep:
+	cfi_startproc
+	CALL_MCOUNT
+	movl	12(%esp), %eax
+	cmpl	%eax, 16(%esp)
+	jb	__chk_fail
+	jmp	__memset_sse2_rep
+	cfi_endproc 
+	.size __memset_chk_sse2_rep, .-__memset_chk_sse2_rep
+
+	.type __memset_chk_ia32, @function
+	.p2align 4;
+__memset_chk_ia32:
+	cfi_startproc
+	CALL_MCOUNT
+	movl	12(%esp), %eax
+	cmpl	%eax, 16(%esp)
+	jb	__chk_fail
+	jmp	__memset_ia32
+	cfi_endproc 
+	.size __memset_chk_ia32, .-__memset_chk_ia32
+# endif
+#endif