Atom optimized string and memory routines

Change-Id: I27b68bb28551c75c9ac84bb9730e2cd8254d8991

13 files changed, 7342 insertions, 0 deletions

diff --git a/libc/arch-x86/string/bcopy_wrapper.S b/libc/arch-x86/string/bcopy_wrapper.S
new file mode 100644
index 0000000..fa8774c
--- /dev/null
+++ b/libc/arch-x86/string/bcopy_wrapper.S
@@ -0,0 +1,45 @@
+/*
+Copyright (c) 2010, Intel Corporation
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+    * Redistributions of source code must retain the above copyright notice,
+    * this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright notice,
+    * this list of conditions and the following disclaimer in the documentation
+    * and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its contributors
+    * may be used to endorse or promote products derived from this software
+    * without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+
+#if defined(USE_SSSE3)
+
+# include "cache_wrapper.S"
+# undef __i686
+# define MEMCPY	bcopy
+# define USE_AS_MEMMOVE
+# define USE_AS_BCOPY
+# include "ssse3-memcpy5.S"
+
+#else
+
+# include "bcopy.S"
+
+#endif
diff --git a/libc/arch-x86/string/bzero_wrapper.S b/libc/arch-x86/string/bzero_wrapper.S
new file mode 100644
index 0000000..aa1bb9c
--- /dev/null
+++ b/libc/arch-x86/string/bzero_wrapper.S
@@ -0,0 +1,43 @@
+/*
+Copyright (c) 2010, Intel Corporation
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+    * Redistributions of source code must retain the above copyright notice,
+    * this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright notice,
+    * this list of conditions and the following disclaimer in the documentation
+    * and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its contributors
+    * may be used to endorse or promote products derived from this software
+    * without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#if defined(USE_SSE2)
+
+# include "cache_wrapper.S"
+# undef __i686
+# define USE_AS_BZERO
+# define sse2_memset5_atom bzero
+# include "sse2-memset5-atom.S"
+
+#else
+
+# include "bzero.S"
+
+#endif
diff --git a/libc/arch-x86/string/cache_wrapper.S b/libc/arch-x86/string/cache_wrapper.S
new file mode 100644
index 0000000..d9aff5c
--- /dev/null
+++ b/libc/arch-x86/string/cache_wrapper.S
@@ -0,0 +1,35 @@
+/*
+Copyright (c) 2010, Intel Corporation
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+    * Redistributions of source code must retain the above copyright notice,
+    * this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright notice,
+    * this list of conditions and the following disclaimer in the documentation
+    * and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its contributors
+    * may be used to endorse or promote products derived from this software
+    * without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+/* Values are optimized for Atom */
+#define SHARED_CACHE_SIZE	(512*1024)			/* Atom L2 Cache */
+#define DATA_CACHE_SIZE		(24*1024)			/* Atom L1 Data Cache */
+#define SHARED_CACHE_SIZE_HALF	(SHARED_CACHE_SIZE / 2)
+#define DATA_CACHE_SIZE_HALF	(DATA_CACHE_SIZE / 2)
diff --git a/libc/arch-x86/string/memcmp_wrapper.S b/libc/arch-x86/string/memcmp_wrapper.S
new file mode 100644
index 0000000..7e28c1e
--- /dev/null
+++ b/libc/arch-x86/string/memcmp_wrapper.S
@@ -0,0 +1,40 @@
+/*
+Copyright (c) 2010, Intel Corporation
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+    * Redistributions of source code must retain the above copyright notice,
+    * this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright notice,
+    * this list of conditions and the following disclaimer in the documentation
+    * and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its contributors
+    * may be used to endorse or promote products derived from this software
+    * without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#if defined(USE_SSSE3)
+
+# define MEMCMP memcmp
+# include "ssse3-memcmp3.S"
+
+#else
+
+# include "memcmp.S"
+
+#endif
diff --git a/libc/arch-x86/string/memcpy_wrapper.S b/libc/arch-x86/string/memcpy_wrapper.S
new file mode 100644
index 0000000..7e765ea
--- /dev/null
+++ b/libc/arch-x86/string/memcpy_wrapper.S
@@ -0,0 +1,43 @@
+/*
+Copyright (c) 2010, Intel Corporation
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+    * Redistributions of source code must retain the above copyright notice,
+    * this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright notice,
+    * this list of conditions and the following disclaimer in the documentation
+    * and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its contributors
+    * may be used to endorse or promote products derived from this software
+    * without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#if defined(USE_SSSE3)
+
+# include "cache_wrapper.S"
+# undef __i686
+# define MEMCPY	memcpy
+# define USE_AS_MEMMOVE
+# include "ssse3-memcpy5.S"
+
+#else
+
+# include "memcpy.S"
+
+#endif
diff --git a/libc/arch-x86/string/memmove_wrapper.S b/libc/arch-x86/string/memmove_wrapper.S
new file mode 100644
index 0000000..7e83e27
--- /dev/null
+++ b/libc/arch-x86/string/memmove_wrapper.S
@@ -0,0 +1,43 @@
+/*
+Copyright (c) 2010, Intel Corporation
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+    * Redistributions of source code must retain the above copyright notice,
+    * this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright notice,
+    * this list of conditions and the following disclaimer in the documentation
+    * and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its contributors
+    * may be used to endorse or promote products derived from this software
+    * without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#if defined(USE_SSSE3)
+
+# include "cache_wrapper.S"
+# undef __i686
+# define MEMCPY memmove
+# define USE_AS_MEMMOVE
+# include "ssse3-memcpy5.S"
+
+#else
+
+# include "memmove.S"
+
+#endif
diff --git a/libc/arch-x86/string/memset_wrapper.S b/libc/arch-x86/string/memset_wrapper.S
new file mode 100644
index 0000000..d037a50
--- /dev/null
+++ b/libc/arch-x86/string/memset_wrapper.S
@@ -0,0 +1,42 @@
+/*
+Copyright (c) 2010, Intel Corporation
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+    * Redistributions of source code must retain the above copyright notice,
+    * this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright notice,
+    * this list of conditions and the following disclaimer in the documentation
+    * and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its contributors
+    * may be used to endorse or promote products derived from this software
+    * without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#if defined(USE_SSE2)
+
+# include "cache_wrapper.S"
+# undef __i686
+# define sse2_memset5_atom memset
+# include "sse2-memset5-atom.S"
+
+#else
+
+# include "memset.S"
+
+#endif
diff --git a/libc/arch-x86/string/sse2-memset5-atom.S b/libc/arch-x86/string/sse2-memset5-atom.S
new file mode 100644
index 0000000..59a598c
--- /dev/null
+++ b/libc/arch-x86/string/sse2-memset5-atom.S
@@ -0,0 +1,907 @@
+/*
+Copyright (c) 2010, Intel Corporation
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+    * Redistributions of source code must retain the above copyright notice,
+    * this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright notice,
+    * this list of conditions and the following disclaimer in the documentation
+    * and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its contributors
+    * may be used to endorse or promote products derived from this software
+    * without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef L
+# define L(label)	.L##label
+#endif
+
+#ifndef ALIGN
+# define ALIGN(n)	.p2align n
+#endif
+
+#ifndef cfi_startproc
+# define cfi_startproc			.cfi_startproc
+#endif
+
+#ifndef cfi_endproc
+# define cfi_endproc			.cfi_endproc
+#endif
+
+#ifndef cfi_rel_offset
+# define cfi_rel_offset(reg, off)	.cfi_rel_offset reg, off
+#endif
+
+#ifndef cfi_restore
+# define cfi_restore(reg)		.cfi_restore (reg)
+#endif
+
+#ifndef cfi_adjust_cfa_offset
+# define cfi_adjust_cfa_offset(off)	.cfi_adjust_cfa_offset off
+#endif
+
+#ifndef ENTRY
+# define ENTRY(name)			\
+	.type name,  @function; 	\
+	.globl name;			\
+	.p2align 4;			\
+name:					\
+	cfi_startproc
+#endif
+
+#ifndef END
+# define END(name)			\
+	cfi_endproc;			\
+	.size name, .-name
+#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 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
+	ALIGN (4)
+ENTRY (sse2_memset5_atom)
+	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
+	jae	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
+	jb	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
+	jae	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
+	jae	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
+	jae	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
+	jae	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 (sse2_memset5_atom)
diff --git a/libc/arch-x86/string/ssse3-memcmp3.S b/libc/arch-x86/string/ssse3-memcmp3.S
new file mode 100644
index 0000000..a7ce819
--- /dev/null
+++ b/libc/arch-x86/string/ssse3-memcmp3.S
@@ -0,0 +1,2027 @@
+/*
+Copyright (c) 2010, Intel Corporation
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+    * Redistributions of source code must retain the above copyright notice,
+    * this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright notice,
+    * this list of conditions and the following disclaimer in the documentation
+    * and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its contributors
+    * may be used to endorse or promote products derived from this software
+    * without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef MEMCMP
+# define MEMCMP		ssse3_memcmp3_new
+#endif
+
+#ifndef L
+# define L(label)	.L##label
+#endif
+
+#ifndef ALIGN
+# define ALIGN(n)	.p2align n
+#endif
+
+#ifndef cfi_startproc
+# define cfi_startproc			.cfi_startproc
+#endif
+
+#ifndef cfi_endproc
+# define cfi_endproc			.cfi_endproc
+#endif
+
+#ifndef cfi_rel_offset
+# define cfi_rel_offset(reg, off)	.cfi_rel_offset reg, off
+#endif
+
+#ifndef cfi_restore
+# define cfi_restore(reg)		.cfi_restore (reg)
+#endif
+
+#ifndef cfi_adjust_cfa_offset
+# define cfi_adjust_cfa_offset(off)	.cfi_adjust_cfa_offset off
+#endif
+
+#ifndef ENTRY
+# define ENTRY(name)			\
+	.type name,  @function; 	\
+	.globl name;			\
+	.p2align 4;			\
+name:					\
+	cfi_startproc
+#endif
+
+#ifndef END
+# define END(name)			\
+	cfi_endproc;			\
+	.size name, .-name
+#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)
+
+#define PARMS		4
+#define BLK1		PARMS
+#define BLK2		BLK1+4
+#define LEN		BLK2+4
+#define RETURN_END	POP (%edi); POP (%esi); POP (%ebx); ret
+#define RETURN		RETURN_END; CFI_PUSH (%ebx); CFI_PUSH (%edi); \
+			CFI_PUSH (%esi)
+
+	.section .text.ssse3,"ax",@progbits
+ENTRY (MEMCMP)
+	movl	LEN(%esp), %ecx
+	movl	BLK1(%esp), %eax
+	cmp	$48, %ecx
+	movl	BLK2(%esp), %edx
+	jae	L(48bytesormore)
+	cmp	$1, %ecx
+	jbe	L(less1bytes)
+	PUSH (%ebx)
+	add	%ecx, %edx
+	add	%ecx, %eax
+	jmp	L(less48bytes)
+
+	CFI_POP (%ebx)
+	ALIGN (4)
+L(less1bytes):
+	jb	L(zero)
+	movb	(%eax), %cl
+	cmp	(%edx), %cl
+	je	L(zero)
+	mov	$1, %eax
+	ja	L(1bytesend)
+	neg	%eax
+L(1bytesend):
+	ret
+
+	ALIGN (4)
+L(zero):
+	mov	$0, %eax
+	ret
+
+	ALIGN (4)
+L(48bytesormore):
+	PUSH (%ebx)
+	PUSH (%esi)
+	PUSH (%edi)
+	movdqu    (%eax), %xmm3
+	movdqu    (%edx), %xmm0
+	movl	%eax, %edi
+	movl	%edx, %esi
+	pcmpeqb   %xmm0, %xmm3
+	pmovmskb  %xmm3, %edx
+	lea	16(%edi), %edi
+
+	sub      $0xffff, %edx
+	lea	16(%esi), %esi
+	jnz	  L(less16bytes)
+	mov	%edi, %edx
+	and	$0xf, %edx
+	xor	%edx, %edi
+	sub	%edx, %esi
+	add	%edx, %ecx
+	mov	%esi, %edx
+	and	$0xf, %edx
+	jz	L(shr_0)
+	xor	%edx, %esi
+
+	cmp	$8, %edx
+	jae	L(next_unaligned_table)
+	cmp	$0, %edx
+	je	L(shr_0)
+	cmp	$1, %edx
+	je	L(shr_1)
+	cmp	$2, %edx
+	je	L(shr_2)
+	cmp	$3, %edx
+	je	L(shr_3)
+	cmp	$4, %edx
+	je	L(shr_4)
+	cmp	$5, %edx
+	je	L(shr_5)
+	cmp	$6, %edx
+	je	L(shr_6)
+	jmp	L(shr_7)
+
+	ALIGN (4)
+L(next_unaligned_table):
+	cmp	$8, %edx
+	je	L(shr_8)
+	cmp	$9, %edx
+	je	L(shr_9)
+	cmp	$10, %edx
+	je	L(shr_10)
+	cmp	$11, %edx
+	je	L(shr_11)
+	cmp	$12, %edx
+	je	L(shr_12)
+	cmp	$13, %edx
+	je	L(shr_13)
+	cmp	$14, %edx
+	je	L(shr_14)
+	jmp	L(shr_15)
+
+	ALIGN (4)
+L(shr_0):
+	cmp	$80, %ecx
+	jae	L(shr_0_gobble)
+	lea	-48(%ecx), %ecx
+	xor	%eax, %eax
+	movaps	(%esi), %xmm1
+	pcmpeqb	(%edi), %xmm1
+	movaps	16(%esi), %xmm2
+	pcmpeqb	16(%edi), %xmm2
+	pand	%xmm1, %xmm2
+	pmovmskb %xmm2, %edx
+	add	$32, %edi
+	add	$32, %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+
+	lea	(%ecx, %edi,1), %eax
+	lea	(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_0_gobble):
+	lea	-48(%ecx), %ecx
+	movdqa	(%esi), %xmm0
+	xor	%eax, %eax
+	pcmpeqb	(%edi), %xmm0
+	sub	$32, %ecx
+	movdqa	16(%esi), %xmm2
+	pcmpeqb	16(%edi), %xmm2
+L(shr_0_gobble_loop):
+	pand	%xmm0, %xmm2
+	sub	$32, %ecx
+	pmovmskb %xmm2, %edx
+	movdqa	%xmm0, %xmm1
+	movdqa	32(%esi), %xmm0
+	movdqa	48(%esi), %xmm2
+	sbb	$0xffff, %edx
+	pcmpeqb	32(%edi), %xmm0
+	pcmpeqb	48(%edi), %xmm2
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	jz	L(shr_0_gobble_loop)
+
+	pand	%xmm0, %xmm2
+	cmp	$0, %ecx
+	jge	L(shr_0_gobble_loop_next)
+	inc	%edx
+	add	$32, %ecx
+L(shr_0_gobble_loop_next):
+	test	%edx, %edx
+	jnz	L(exit)
+
+	pmovmskb %xmm2, %edx
+	movdqa	%xmm0, %xmm1
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+	lea	(%ecx, %edi,1), %eax
+	lea	(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_1):
+	cmp	$80, %ecx
+	lea	-48(%ecx), %ecx
+	mov	%edx, %eax
+	jae	L(shr_1_gobble)
+
+	movdqa	16(%esi), %xmm1
+	movdqa	%xmm1, %xmm2
+	palignr	$1,(%esi), %xmm1
+	pcmpeqb	(%edi), %xmm1
+
+	movdqa	32(%esi), %xmm3
+	palignr	$1,%xmm2, %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+	pand	%xmm1, %xmm3
+	pmovmskb %xmm3, %edx
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+	lea	(%ecx, %edi,1), %eax
+	lea	1(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_1_gobble):
+	sub	$32, %ecx
+	movdqa	16(%esi), %xmm0
+	palignr	$1,(%esi), %xmm0
+	pcmpeqb	(%edi), %xmm0
+
+	movdqa	32(%esi), %xmm3
+	palignr	$1,16(%esi), %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+L(shr_1_gobble_loop):
+	pand	%xmm0, %xmm3
+	sub	$32, %ecx
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+
+	movdqa	64(%esi), %xmm3
+	palignr	$1,48(%esi), %xmm3
+	sbb	$0xffff, %edx
+	movdqa	48(%esi), %xmm0
+	palignr	$1,32(%esi), %xmm0
+	pcmpeqb	32(%edi), %xmm0
+	lea	32(%esi), %esi
+	pcmpeqb	48(%edi), %xmm3
+
+	lea	32(%edi), %edi
+	jz	L(shr_1_gobble_loop)
+	pand	%xmm0, %xmm3
+
+	cmp	$0, %ecx
+	jge	L(shr_1_gobble_next)
+	inc	%edx
+	add	$32, %ecx
+L(shr_1_gobble_next):
+	test	%edx, %edx
+	jnz	L(exit)
+
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+
+	lea	(%ecx, %edi,1), %eax
+	lea	1(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_2):
+	cmp	$80, %ecx
+	lea	-48(%ecx), %ecx
+	mov	%edx, %eax
+	jae	L(shr_2_gobble)
+
+	movdqa	16(%esi), %xmm1
+	movdqa	%xmm1, %xmm2
+	palignr	$2,(%esi), %xmm1
+	pcmpeqb	(%edi), %xmm1
+
+	movdqa	32(%esi), %xmm3
+	palignr	$2,%xmm2, %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+	pand	%xmm1, %xmm3
+	pmovmskb %xmm3, %edx
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+	lea	(%ecx, %edi,1), %eax
+	lea	2(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_2_gobble):
+	sub	$32, %ecx
+	movdqa	16(%esi), %xmm0
+	palignr	$2,(%esi), %xmm0
+	pcmpeqb	(%edi), %xmm0
+
+	movdqa	32(%esi), %xmm3
+	palignr	$2,16(%esi), %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+L(shr_2_gobble_loop):
+	pand	%xmm0, %xmm3
+	sub	$32, %ecx
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+
+	movdqa	64(%esi), %xmm3
+	palignr	$2,48(%esi), %xmm3
+	sbb	$0xffff, %edx
+	movdqa	48(%esi), %xmm0
+	palignr	$2,32(%esi), %xmm0
+	pcmpeqb	32(%edi), %xmm0
+	lea	32(%esi), %esi
+	pcmpeqb	48(%edi), %xmm3
+
+	lea	32(%edi), %edi
+	jz	L(shr_2_gobble_loop)
+	pand	%xmm0, %xmm3
+
+	cmp	$0, %ecx
+	jge	L(shr_2_gobble_next)
+	inc	%edx
+	add	$32, %ecx
+L(shr_2_gobble_next):
+	test	%edx, %edx
+	jnz	L(exit)
+
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+
+	lea	(%ecx, %edi,1), %eax
+	lea	2(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_3):
+	cmp	$80, %ecx
+	lea	-48(%ecx), %ecx
+	mov	%edx, %eax
+	jae	L(shr_3_gobble)
+
+	movdqa	16(%esi), %xmm1
+	movdqa	%xmm1, %xmm2
+	palignr	$3,(%esi), %xmm1
+	pcmpeqb	(%edi), %xmm1
+
+	movdqa	32(%esi), %xmm3
+	palignr	$3,%xmm2, %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+	pand	%xmm1, %xmm3
+	pmovmskb %xmm3, %edx
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+	lea	(%ecx, %edi,1), %eax
+	lea	3(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_3_gobble):
+	sub	$32, %ecx
+	movdqa	16(%esi), %xmm0
+	palignr	$3,(%esi), %xmm0
+	pcmpeqb	(%edi), %xmm0
+
+	movdqa	32(%esi), %xmm3
+	palignr	$3,16(%esi), %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+L(shr_3_gobble_loop):
+	pand	%xmm0, %xmm3
+	sub	$32, %ecx
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+
+	movdqa	64(%esi), %xmm3
+	palignr	$3,48(%esi), %xmm3
+	sbb	$0xffff, %edx
+	movdqa	48(%esi), %xmm0
+	palignr	$3,32(%esi), %xmm0
+	pcmpeqb	32(%edi), %xmm0
+	lea	32(%esi), %esi
+	pcmpeqb	48(%edi), %xmm3
+
+	lea	32(%edi), %edi
+	jz	L(shr_3_gobble_loop)
+	pand	%xmm0, %xmm3
+
+	cmp	$0, %ecx
+	jge	L(shr_3_gobble_next)
+	inc	%edx
+	add	$32, %ecx
+L(shr_3_gobble_next):
+	test	%edx, %edx
+	jnz	L(exit)
+
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+
+	lea	(%ecx, %edi,1), %eax
+	lea	3(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_4):
+	cmp	$80, %ecx
+	lea	-48(%ecx), %ecx
+	mov	%edx, %eax
+	jae	L(shr_4_gobble)
+
+	movdqa	16(%esi), %xmm1
+	movdqa	%xmm1, %xmm2
+	palignr	$4,(%esi), %xmm1
+	pcmpeqb	(%edi), %xmm1
+
+	movdqa	32(%esi), %xmm3
+	palignr	$4,%xmm2, %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+	pand	%xmm1, %xmm3
+	pmovmskb %xmm3, %edx
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+	lea	(%ecx, %edi,1), %eax
+	lea	4(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_4_gobble):
+	sub	$32, %ecx
+	movdqa	16(%esi), %xmm0
+	palignr	$4,(%esi), %xmm0
+	pcmpeqb	(%edi), %xmm0
+
+	movdqa	32(%esi), %xmm3
+	palignr	$4,16(%esi), %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+L(shr_4_gobble_loop):
+	pand	%xmm0, %xmm3
+	sub	$32, %ecx
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+
+	movdqa	64(%esi), %xmm3
+	palignr	$4,48(%esi), %xmm3
+	sbb	$0xffff, %edx
+	movdqa	48(%esi), %xmm0
+	palignr	$4,32(%esi), %xmm0
+	pcmpeqb	32(%edi), %xmm0
+	lea	32(%esi), %esi
+	pcmpeqb	48(%edi), %xmm3
+
+	lea	32(%edi), %edi
+	jz	L(shr_4_gobble_loop)
+	pand	%xmm0, %xmm3
+
+	cmp	$0, %ecx
+	jge	L(shr_4_gobble_next)
+	inc	%edx
+	add	$32, %ecx
+L(shr_4_gobble_next):
+	test	%edx, %edx
+	jnz	L(exit)
+
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+
+	lea	(%ecx, %edi,1), %eax
+	lea	4(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_5):
+	cmp	$80, %ecx
+	lea	-48(%ecx), %ecx
+	mov	%edx, %eax
+	jae	L(shr_5_gobble)
+
+	movdqa	16(%esi), %xmm1
+	movdqa	%xmm1, %xmm2
+	palignr	$5,(%esi), %xmm1
+	pcmpeqb	(%edi), %xmm1
+
+	movdqa	32(%esi), %xmm3
+	palignr	$5,%xmm2, %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+	pand	%xmm1, %xmm3
+	pmovmskb %xmm3, %edx
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+	lea	(%ecx, %edi,1), %eax
+	lea	5(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_5_gobble):
+	sub	$32, %ecx
+	movdqa	16(%esi), %xmm0
+	palignr	$5,(%esi), %xmm0
+	pcmpeqb	(%edi), %xmm0
+
+	movdqa	32(%esi), %xmm3
+	palignr	$5,16(%esi), %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+L(shr_5_gobble_loop):
+	pand	%xmm0, %xmm3
+	sub	$32, %ecx
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+
+	movdqa	64(%esi), %xmm3
+	palignr	$5,48(%esi), %xmm3
+	sbb	$0xffff, %edx
+	movdqa	48(%esi), %xmm0
+	palignr	$5,32(%esi), %xmm0
+	pcmpeqb	32(%edi), %xmm0
+	lea	32(%esi), %esi
+	pcmpeqb	48(%edi), %xmm3
+
+	lea	32(%edi), %edi
+	jz	L(shr_5_gobble_loop)
+	pand	%xmm0, %xmm3
+
+	cmp	$0, %ecx
+	jge	L(shr_5_gobble_next)
+	inc	%edx
+	add	$32, %ecx
+L(shr_5_gobble_next):
+	test	%edx, %edx
+	jnz	L(exit)
+
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+
+	lea	(%ecx, %edi,1), %eax
+	lea	5(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_6):
+	cmp	$80, %ecx
+	lea	-48(%ecx), %ecx
+	mov	%edx, %eax
+	jae	L(shr_6_gobble)
+
+	movdqa	16(%esi), %xmm1
+	movdqa	%xmm1, %xmm2
+	palignr	$6,(%esi), %xmm1
+	pcmpeqb	(%edi), %xmm1
+
+	movdqa	32(%esi), %xmm3
+	palignr	$6,%xmm2, %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+	pand	%xmm1, %xmm3
+	pmovmskb %xmm3, %edx
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+	lea	(%ecx, %edi,1), %eax
+	lea	6(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_6_gobble):
+	sub	$32, %ecx
+	movdqa	16(%esi), %xmm0
+	palignr	$6,(%esi), %xmm0
+	pcmpeqb	(%edi), %xmm0
+
+	movdqa	32(%esi), %xmm3
+	palignr	$6,16(%esi), %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+L(shr_6_gobble_loop):
+	pand	%xmm0, %xmm3
+	sub	$32, %ecx
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+
+	movdqa	64(%esi), %xmm3
+	palignr	$6,48(%esi), %xmm3
+	sbb	$0xffff, %edx
+	movdqa	48(%esi), %xmm0
+	palignr	$6,32(%esi), %xmm0
+	pcmpeqb	32(%edi), %xmm0
+	lea	32(%esi), %esi
+	pcmpeqb	48(%edi), %xmm3
+
+	lea	32(%edi), %edi
+	jz	L(shr_6_gobble_loop)
+	pand	%xmm0, %xmm3
+
+	cmp	$0, %ecx
+	jge	L(shr_6_gobble_next)
+	inc	%edx
+	add	$32, %ecx
+L(shr_6_gobble_next):
+	test	%edx, %edx
+	jnz	L(exit)
+
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+
+	lea	(%ecx, %edi,1), %eax
+	lea	6(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_7):
+	cmp	$80, %ecx
+	lea	-48(%ecx), %ecx
+	mov	%edx, %eax
+	jae	L(shr_7_gobble)
+
+	movdqa	16(%esi), %xmm1
+	movdqa	%xmm1, %xmm2
+	palignr	$7,(%esi), %xmm1
+	pcmpeqb	(%edi), %xmm1
+
+	movdqa	32(%esi), %xmm3
+	palignr	$7,%xmm2, %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+	pand	%xmm1, %xmm3
+	pmovmskb %xmm3, %edx
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+	lea	(%ecx, %edi,1), %eax
+	lea	7(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_7_gobble):
+	sub	$32, %ecx
+	movdqa	16(%esi), %xmm0
+	palignr	$7,(%esi), %xmm0
+	pcmpeqb	(%edi), %xmm0
+
+	movdqa	32(%esi), %xmm3
+	palignr	$7,16(%esi), %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+L(shr_7_gobble_loop):
+	pand	%xmm0, %xmm3
+	sub	$32, %ecx
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+
+	movdqa	64(%esi), %xmm3
+	palignr	$7,48(%esi), %xmm3
+	sbb	$0xffff, %edx
+	movdqa	48(%esi), %xmm0
+	palignr	$7,32(%esi), %xmm0
+	pcmpeqb	32(%edi), %xmm0
+	lea	32(%esi), %esi
+	pcmpeqb	48(%edi), %xmm3
+
+	lea	32(%edi), %edi
+	jz	L(shr_7_gobble_loop)
+	pand	%xmm0, %xmm3
+
+	cmp	$0, %ecx
+	jge	L(shr_7_gobble_next)
+	inc	%edx
+	add	$32, %ecx
+L(shr_7_gobble_next):
+	test	%edx, %edx
+	jnz	L(exit)
+
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+
+	lea	(%ecx, %edi,1), %eax
+	lea	7(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_8):
+	cmp	$80, %ecx
+	lea	-48(%ecx), %ecx
+	mov	%edx, %eax
+	jae	L(shr_8_gobble)
+
+	movdqa	16(%esi), %xmm1
+	movdqa	%xmm1, %xmm2
+	palignr	$8,(%esi), %xmm1
+	pcmpeqb	(%edi), %xmm1
+
+	movdqa	32(%esi), %xmm3
+	palignr	$8,%xmm2, %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+	pand	%xmm1, %xmm3
+	pmovmskb %xmm3, %edx
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+	lea	(%ecx, %edi,1), %eax
+	lea	8(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_8_gobble):
+	sub	$32, %ecx
+	movdqa	16(%esi), %xmm0
+	palignr	$8,(%esi), %xmm0
+	pcmpeqb	(%edi), %xmm0
+
+	movdqa	32(%esi), %xmm3
+	palignr	$8,16(%esi), %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+L(shr_8_gobble_loop):
+	pand	%xmm0, %xmm3
+	sub	$32, %ecx
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+
+	movdqa	64(%esi), %xmm3
+	palignr	$8,48(%esi), %xmm3
+	sbb	$0xffff, %edx
+	movdqa	48(%esi), %xmm0
+	palignr	$8,32(%esi), %xmm0
+	pcmpeqb	32(%edi), %xmm0
+	lea	32(%esi), %esi
+	pcmpeqb	48(%edi), %xmm3
+
+	lea	32(%edi), %edi
+	jz	L(shr_8_gobble_loop)
+	pand	%xmm0, %xmm3
+
+	cmp	$0, %ecx
+	jge	L(shr_8_gobble_next)
+	inc	%edx
+	add	$32, %ecx
+L(shr_8_gobble_next):
+	test	%edx, %edx
+	jnz	L(exit)
+
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+
+	lea	(%ecx, %edi,1), %eax
+	lea	8(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_9):
+	cmp	$80, %ecx
+	lea	-48(%ecx), %ecx
+	mov	%edx, %eax
+	jae	L(shr_9_gobble)
+
+	movdqa	16(%esi), %xmm1
+	movdqa	%xmm1, %xmm2
+	palignr	$9,(%esi), %xmm1
+	pcmpeqb	(%edi), %xmm1
+
+	movdqa	32(%esi), %xmm3
+	palignr	$9,%xmm2, %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+	pand	%xmm1, %xmm3
+	pmovmskb %xmm3, %edx
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+	lea	(%ecx, %edi,1), %eax
+	lea	9(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_9_gobble):
+	sub	$32, %ecx
+	movdqa	16(%esi), %xmm0
+	palignr	$9,(%esi), %xmm0
+	pcmpeqb	(%edi), %xmm0
+
+	movdqa	32(%esi), %xmm3
+	palignr	$9,16(%esi), %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+L(shr_9_gobble_loop):
+	pand	%xmm0, %xmm3
+	sub	$32, %ecx
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+
+	movdqa	64(%esi), %xmm3
+	palignr	$9,48(%esi), %xmm3
+	sbb	$0xffff, %edx
+	movdqa	48(%esi), %xmm0
+	palignr	$9,32(%esi), %xmm0
+	pcmpeqb	32(%edi), %xmm0
+	lea	32(%esi), %esi
+	pcmpeqb	48(%edi), %xmm3
+
+	lea	32(%edi), %edi
+	jz	L(shr_9_gobble_loop)
+	pand	%xmm0, %xmm3
+
+	cmp	$0, %ecx
+	jge	L(shr_9_gobble_next)
+	inc	%edx
+	add	$32, %ecx
+L(shr_9_gobble_next):
+	test	%edx, %edx
+	jnz	L(exit)
+
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+
+	lea	(%ecx, %edi,1), %eax
+	lea	9(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_10):
+	cmp	$80, %ecx
+	lea	-48(%ecx), %ecx
+	mov	%edx, %eax
+	jae	L(shr_10_gobble)
+
+	movdqa	16(%esi), %xmm1
+	movdqa	%xmm1, %xmm2
+	palignr	$10, (%esi), %xmm1
+	pcmpeqb	(%edi), %xmm1
+
+	movdqa	32(%esi), %xmm3
+	palignr	$10,%xmm2, %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+	pand	%xmm1, %xmm3
+	pmovmskb %xmm3, %edx
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+	lea	(%ecx, %edi,1), %eax
+	lea	10(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_10_gobble):
+	sub	$32, %ecx
+	movdqa	16(%esi), %xmm0
+	palignr	$10, (%esi), %xmm0
+	pcmpeqb	(%edi), %xmm0
+
+	movdqa	32(%esi), %xmm3
+	palignr	$10, 16(%esi), %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+L(shr_10_gobble_loop):
+	pand	%xmm0, %xmm3
+	sub	$32, %ecx
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+
+	movdqa	64(%esi), %xmm3
+	palignr	$10,48(%esi), %xmm3
+	sbb	$0xffff, %edx
+	movdqa	48(%esi), %xmm0
+	palignr	$10,32(%esi), %xmm0
+	pcmpeqb	32(%edi), %xmm0
+	lea	32(%esi), %esi
+	pcmpeqb	48(%edi), %xmm3
+
+	lea	32(%edi), %edi
+	jz	L(shr_10_gobble_loop)
+	pand	%xmm0, %xmm3
+
+	cmp	$0, %ecx
+	jge	L(shr_10_gobble_next)
+	inc	%edx
+	add	$32, %ecx
+L(shr_10_gobble_next):
+	test	%edx, %edx
+	jnz	L(exit)
+
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+
+	lea	(%ecx, %edi,1), %eax
+	lea	10(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_11):
+	cmp	$80, %ecx
+	lea	-48(%ecx), %ecx
+	mov	%edx, %eax
+	jae	L(shr_11_gobble)
+
+	movdqa	16(%esi), %xmm1
+	movdqa	%xmm1, %xmm2
+	palignr	$11, (%esi), %xmm1
+	pcmpeqb	(%edi), %xmm1
+
+	movdqa	32(%esi), %xmm3
+	palignr	$11, %xmm2, %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+	pand	%xmm1, %xmm3
+	pmovmskb %xmm3, %edx
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+	lea	(%ecx, %edi,1), %eax
+	lea	11(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_11_gobble):
+	sub	$32, %ecx
+	movdqa	16(%esi), %xmm0
+	palignr	$11, (%esi), %xmm0
+	pcmpeqb	(%edi), %xmm0
+
+	movdqa	32(%esi), %xmm3
+	palignr	$11, 16(%esi), %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+L(shr_11_gobble_loop):
+	pand	%xmm0, %xmm3
+	sub	$32, %ecx
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+
+	movdqa	64(%esi), %xmm3
+	palignr	$11,48(%esi), %xmm3
+	sbb	$0xffff, %edx
+	movdqa	48(%esi), %xmm0
+	palignr	$11,32(%esi), %xmm0
+	pcmpeqb	32(%edi), %xmm0
+	lea	32(%esi), %esi
+	pcmpeqb	48(%edi), %xmm3
+
+	lea	32(%edi), %edi
+	jz	L(shr_11_gobble_loop)
+	pand	%xmm0, %xmm3
+
+	cmp	$0, %ecx
+	jge	L(shr_11_gobble_next)
+	inc	%edx
+	add	$32, %ecx
+L(shr_11_gobble_next):
+	test	%edx, %edx
+	jnz	L(exit)
+
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+
+	lea	(%ecx, %edi,1), %eax
+	lea	11(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_12):
+	cmp	$80, %ecx
+	lea	-48(%ecx), %ecx
+	mov	%edx, %eax
+	jae	L(shr_12_gobble)
+
+	movdqa	16(%esi), %xmm1
+	movdqa	%xmm1, %xmm2
+	palignr	$12, (%esi), %xmm1
+	pcmpeqb	(%edi), %xmm1
+
+	movdqa	32(%esi), %xmm3
+	palignr	$12, %xmm2, %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+	pand	%xmm1, %xmm3
+	pmovmskb %xmm3, %edx
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+	lea	(%ecx, %edi,1), %eax
+	lea	12(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_12_gobble):
+	sub	$32, %ecx
+	movdqa	16(%esi), %xmm0
+	palignr	$12, (%esi), %xmm0
+	pcmpeqb	(%edi), %xmm0
+
+	movdqa	32(%esi), %xmm3
+	palignr	$12, 16(%esi), %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+L(shr_12_gobble_loop):
+	pand	%xmm0, %xmm3
+	sub	$32, %ecx
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+
+	movdqa	64(%esi), %xmm3
+	palignr	$12,48(%esi), %xmm3
+	sbb	$0xffff, %edx
+	movdqa	48(%esi), %xmm0
+	palignr	$12,32(%esi), %xmm0
+	pcmpeqb	32(%edi), %xmm0
+	lea	32(%esi), %esi
+	pcmpeqb	48(%edi), %xmm3
+
+	lea	32(%edi), %edi
+	jz	L(shr_12_gobble_loop)
+	pand	%xmm0, %xmm3
+
+	cmp	$0, %ecx
+	jge	L(shr_12_gobble_next)
+	inc	%edx
+	add	$32, %ecx
+L(shr_12_gobble_next):
+	test	%edx, %edx
+	jnz	L(exit)
+
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+
+	lea	(%ecx, %edi,1), %eax
+	lea	12(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_13):
+	cmp	$80, %ecx
+	lea	-48(%ecx), %ecx
+	mov	%edx, %eax
+	jae	L(shr_13_gobble)
+
+	movdqa	16(%esi), %xmm1
+	movdqa	%xmm1, %xmm2
+	palignr	$13, (%esi), %xmm1
+	pcmpeqb	(%edi), %xmm1
+
+	movdqa	32(%esi), %xmm3
+	palignr	$13, %xmm2, %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+	pand	%xmm1, %xmm3
+	pmovmskb %xmm3, %edx
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+	lea	(%ecx, %edi,1), %eax
+	lea	13(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_13_gobble):
+	sub	$32, %ecx
+	movdqa	16(%esi), %xmm0
+	palignr	$13, (%esi), %xmm0
+	pcmpeqb	(%edi), %xmm0
+
+	movdqa	32(%esi), %xmm3
+	palignr	$13, 16(%esi), %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+L(shr_13_gobble_loop):
+	pand	%xmm0, %xmm3
+	sub	$32, %ecx
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+
+	movdqa	64(%esi), %xmm3
+	palignr	$13,48(%esi), %xmm3
+	sbb	$0xffff, %edx
+	movdqa	48(%esi), %xmm0
+	palignr	$13,32(%esi), %xmm0
+	pcmpeqb	32(%edi), %xmm0
+	lea	32(%esi), %esi
+	pcmpeqb	48(%edi), %xmm3
+
+	lea	32(%edi), %edi
+	jz	L(shr_13_gobble_loop)
+	pand	%xmm0, %xmm3
+
+	cmp	$0, %ecx
+	jge	L(shr_13_gobble_next)
+	inc	%edx
+	add	$32, %ecx
+L(shr_13_gobble_next):
+	test	%edx, %edx
+	jnz	L(exit)
+
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+
+	lea	(%ecx, %edi,1), %eax
+	lea	13(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_14):
+	cmp	$80, %ecx
+	lea	-48(%ecx), %ecx
+	mov	%edx, %eax
+	jae	L(shr_14_gobble)
+
+	movdqa	16(%esi), %xmm1
+	movdqa	%xmm1, %xmm2
+	palignr	$14, (%esi), %xmm1
+	pcmpeqb	(%edi), %xmm1
+
+	movdqa	32(%esi), %xmm3
+	palignr	$14, %xmm2, %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+	pand	%xmm1, %xmm3
+	pmovmskb %xmm3, %edx
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+	lea	(%ecx, %edi,1), %eax
+	lea	14(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_14_gobble):
+	sub	$32, %ecx
+	movdqa	16(%esi), %xmm0
+	palignr	$14, (%esi), %xmm0
+	pcmpeqb	(%edi), %xmm0
+
+	movdqa	32(%esi), %xmm3
+	palignr	$14, 16(%esi), %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+L(shr_14_gobble_loop):
+	pand	%xmm0, %xmm3
+	sub	$32, %ecx
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+
+	movdqa	64(%esi), %xmm3
+	palignr	$14,48(%esi), %xmm3
+	sbb	$0xffff, %edx
+	movdqa	48(%esi), %xmm0
+	palignr	$14,32(%esi), %xmm0
+	pcmpeqb	32(%edi), %xmm0
+	lea	32(%esi), %esi
+	pcmpeqb	48(%edi), %xmm3
+
+	lea	32(%edi), %edi
+	jz	L(shr_14_gobble_loop)
+	pand	%xmm0, %xmm3
+
+	cmp	$0, %ecx
+	jge	L(shr_14_gobble_next)
+	inc	%edx
+	add	$32, %ecx
+L(shr_14_gobble_next):
+	test	%edx, %edx
+	jnz	L(exit)
+
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+
+	lea	(%ecx, %edi,1), %eax
+	lea	14(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_15):
+	cmp	$80, %ecx
+	lea	-48(%ecx), %ecx
+	mov	%edx, %eax
+	jae	L(shr_15_gobble)
+
+	movdqa	16(%esi), %xmm1
+	movdqa	%xmm1, %xmm2
+	palignr	$15, (%esi), %xmm1
+	pcmpeqb	(%edi), %xmm1
+
+	movdqa	32(%esi), %xmm3
+	palignr	$15, %xmm2, %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+	pand	%xmm1, %xmm3
+	pmovmskb %xmm3, %edx
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+	lea	(%ecx, %edi,1), %eax
+	lea	15(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(shr_15_gobble):
+	sub	$32, %ecx
+	movdqa	16(%esi), %xmm0
+	palignr	$15, (%esi), %xmm0
+	pcmpeqb	(%edi), %xmm0
+
+	movdqa	32(%esi), %xmm3
+	palignr	$15, 16(%esi), %xmm3
+	pcmpeqb	16(%edi), %xmm3
+
+L(shr_15_gobble_loop):
+	pand	%xmm0, %xmm3
+	sub	$32, %ecx
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+
+	movdqa	64(%esi), %xmm3
+	palignr	$15,48(%esi), %xmm3
+	sbb	$0xffff, %edx
+	movdqa	48(%esi), %xmm0
+	palignr	$15,32(%esi), %xmm0
+	pcmpeqb	32(%edi), %xmm0
+	lea	32(%esi), %esi
+	pcmpeqb	48(%edi), %xmm3
+
+	lea	32(%edi), %edi
+	jz	L(shr_15_gobble_loop)
+	pand	%xmm0, %xmm3
+
+	cmp	$0, %ecx
+	jge	L(shr_15_gobble_next)
+	inc	%edx
+	add	$32, %ecx
+L(shr_15_gobble_next):
+	test	%edx, %edx
+	jnz	L(exit)
+
+	pmovmskb %xmm3, %edx
+	movdqa	%xmm0, %xmm1
+	lea	32(%edi), %edi
+	lea	32(%esi), %esi
+	sub	$0xffff, %edx
+	jnz	L(exit)
+
+	lea	(%ecx, %edi,1), %eax
+	lea	15(%ecx, %esi,1), %edx
+	POP (%edi)
+	POP (%esi)
+	jmp	L(less48bytes)
+
+	CFI_PUSH (%esi)
+	CFI_PUSH (%edi)
+	ALIGN (4)
+L(exit):
+	pmovmskb %xmm1, %ebx
+	sub	$0xffff, %ebx
+	jz	L(first16bytes)
+	lea	-16(%esi), %esi
+	lea	-16(%edi), %edi
+	mov	%ebx, %edx
+L(first16bytes):
+	add	%eax, %esi
+L(less16bytes):
+	test	%dl, %dl
+	jz	L(next_24_bytes)
+
+	test	$0x01, %dl
+	jnz	L(Byte16)
+
+	test	$0x02, %dl
+	jnz	L(Byte17)
+
+	test	$0x04, %dl
+	jnz	L(Byte18)
+
+	test	$0x08, %dl
+	jnz	L(Byte19)
+
+	test	$0x10, %dl
+	jnz	L(Byte20)
+
+	test	$0x20, %dl
+	jnz	L(Byte21)
+
+	test	$0x40, %dl
+	jnz	L(Byte22)
+L(Byte23):
+	movzbl	 -9(%edi), %eax
+	movzbl	 -9(%esi), %edx
+	sub	%edx, %eax
+	RETURN
+
+	ALIGN (4)
+L(Byte16):
+	movzbl	 -16(%edi), %eax
+	movzbl	 -16(%esi), %edx
+	sub	%edx, %eax
+	RETURN
+
+	ALIGN (4)
+L(Byte17):
+	movzbl	 -15(%edi), %eax
+	movzbl	 -15(%esi), %edx
+	sub	%edx, %eax
+	RETURN
+
+	ALIGN (4)
+L(Byte18):
+	movzbl	 -14(%edi), %eax
+	movzbl	 -14(%esi), %edx
+	sub	%edx, %eax
+	RETURN
+
+	ALIGN (4)
+L(Byte19):
+	movzbl	 -13(%edi), %eax
+	movzbl	 -13(%esi), %edx
+	sub	%edx, %eax
+	RETURN
+
+	ALIGN (4)
+L(Byte20):
+	movzbl	 -12(%edi), %eax
+	movzbl	 -12(%esi), %edx
+	sub	%edx, %eax
+	RETURN
+
+	ALIGN (4)
+L(Byte21):
+	movzbl	 -11(%edi), %eax
+	movzbl	 -11(%esi), %edx
+	sub	%edx, %eax
+	RETURN
+
+	ALIGN (4)
+L(Byte22):
+	movzbl	 -10(%edi), %eax
+	movzbl	 -10(%esi), %edx
+	sub	%edx, %eax
+	RETURN
+
+	ALIGN (4)
+L(next_24_bytes):
+	lea	8(%edi), %edi
+	lea	8(%esi), %esi
+	test	$0x01, %dh
+	jnz	L(Byte16)
+
+	test	$0x02, %dh
+	jnz	L(Byte17)
+
+	test	$0x04, %dh
+	jnz	L(Byte18)
+
+	test	$0x08, %dh
+	jnz	L(Byte19)
+
+	test	$0x10, %dh
+	jnz	L(Byte20)
+
+	test	$0x20, %dh
+	jnz	L(Byte21)
+
+	test	$0x40, %dh
+	jnz	L(Byte22)
+
+	ALIGN (4)
+L(Byte31):
+	movzbl	 -9(%edi), %eax
+	movzbl	 -9(%esi), %edx
+	sub	%edx, %eax
+	RETURN_END
+	CFI_PUSH (%ebx)
+
+	ALIGN (4)
+L(more8bytes):
+	cmp	$16, %ecx
+	jae	L(more16bytes)
+	cmp	$8, %ecx
+	je	L(8bytes)
+	cmp	$9, %ecx
+	je	L(9bytes)
+	cmp	$10, %ecx
+	je	L(10bytes)
+	cmp	$11, %ecx
+	je	L(11bytes)
+	cmp	$12, %ecx
+	je	L(12bytes)
+	cmp	$13, %ecx
+	je	L(13bytes)
+	cmp	$14, %ecx
+	je	L(14bytes)
+	jmp	L(15bytes)
+
+	ALIGN (4)
+L(more16bytes):
+	cmp	$24, %ecx
+	jae	L(more24bytes)
+	cmp	$16, %ecx
+	je	L(16bytes)
+	cmp	$17, %ecx
+	je	L(17bytes)
+	cmp	$18, %ecx
+	je	L(18bytes)
+	cmp	$19, %ecx
+	je	L(19bytes)
+	cmp	$20, %ecx
+	je	L(20bytes)
+	cmp	$21, %ecx
+	je	L(21bytes)
+	cmp	$22, %ecx
+	je	L(22bytes)
+	jmp	L(23bytes)
+
+	ALIGN (4)
+L(more24bytes):
+	cmp	$32, %ecx
+	jae	L(more32bytes)
+	cmp	$24, %ecx
+	je	L(24bytes)
+	cmp	$25, %ecx
+	je	L(25bytes)
+	cmp	$26, %ecx
+	je	L(26bytes)
+	cmp	$27, %ecx
+	je	L(27bytes)
+	cmp	$28, %ecx
+	je	L(28bytes)
+	cmp	$29, %ecx
+	je	L(29bytes)
+	cmp	$30, %ecx
+	je	L(30bytes)
+	jmp	L(31bytes)
+
+	ALIGN (4)
+L(more32bytes):
+	cmp	$40, %ecx
+	jae	L(more40bytes)
+	cmp	$32, %ecx
+	je	L(32bytes)
+	cmp	$33, %ecx
+	je	L(33bytes)
+	cmp	$34, %ecx
+	je	L(34bytes)
+	cmp	$35, %ecx
+	je	L(35bytes)
+	cmp	$36, %ecx
+	je	L(36bytes)
+	cmp	$37, %ecx
+	je	L(37bytes)
+	cmp	$38, %ecx
+	je	L(38bytes)
+	jmp	L(39bytes)
+
+	ALIGN (4)
+L(more40bytes):
+	cmp	$40, %ecx
+	je	L(40bytes)
+	cmp	$41, %ecx
+	je	L(41bytes)
+	cmp	$42, %ecx
+	je	L(42bytes)
+	cmp	$43, %ecx
+	je	L(43bytes)
+	cmp	$44, %ecx
+	je	L(44bytes)
+	cmp	$45, %ecx
+	je	L(45bytes)
+	cmp	$46, %ecx
+	je	L(46bytes)
+	jmp	L(47bytes)
+
+	ALIGN (4)
+L(less48bytes):
+	cmp	$8, %ecx
+	jae	L(more8bytes)
+	cmp	$2, %ecx
+	je	L(2bytes)
+	cmp	$3, %ecx
+	je	L(3bytes)
+	cmp	$4, %ecx
+	je	L(4bytes)
+	cmp	$5, %ecx
+	je	L(5bytes)
+	cmp	$6, %ecx
+	je	L(6bytes)
+	jmp	L(7bytes)
+
+
+	ALIGN (4)
+L(44bytes):
+	mov	-44(%eax), %ecx
+	mov	-44(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(40bytes):
+	mov	-40(%eax), %ecx
+	mov	-40(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(36bytes):
+	mov	-36(%eax), %ecx
+	mov	-36(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(32bytes):
+	mov	-32(%eax), %ecx
+	mov	-32(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(28bytes):
+	mov	-28(%eax), %ecx
+	mov	-28(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(24bytes):
+	mov	-24(%eax), %ecx
+	mov	-24(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(20bytes):
+	mov	-20(%eax), %ecx
+	mov	-20(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(16bytes):
+	mov	-16(%eax), %ecx
+	mov	-16(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(12bytes):
+	mov	-12(%eax), %ecx
+	mov	-12(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(8bytes):
+	mov	-8(%eax), %ecx
+	mov	-8(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(4bytes):
+	mov	-4(%eax), %ecx
+	mov	-4(%edx), %ebx
+	cmp	%ebx, %ecx
+	mov	$0, %eax
+	jne	L(find_diff)
+	POP (%ebx)
+	ret
+	CFI_PUSH (%ebx)
+
+	ALIGN (4)
+L(45bytes):
+	mov	-45(%eax), %ecx
+	mov	-45(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(41bytes):
+	mov	-41(%eax), %ecx
+	mov	-41(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(37bytes):
+	mov	-37(%eax), %ecx
+	mov	-37(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(33bytes):
+	mov	-33(%eax), %ecx
+	mov	-33(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(29bytes):
+	mov	-29(%eax), %ecx
+	mov	-29(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(25bytes):
+	mov	-25(%eax), %ecx
+	mov	-25(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(21bytes):
+	mov	-21(%eax), %ecx
+	mov	-21(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(17bytes):
+	mov	-17(%eax), %ecx
+	mov	-17(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(13bytes):
+	mov	-13(%eax), %ecx
+	mov	-13(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(9bytes):
+	mov	-9(%eax), %ecx
+	mov	-9(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(5bytes):
+	mov	-5(%eax), %ecx
+	mov	-5(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+	movzbl	-1(%eax), %ecx
+	cmp	-1(%edx), %cl
+	mov	$0, %eax
+	jne	L(end)
+	POP (%ebx)
+	ret
+	CFI_PUSH (%ebx)
+
+	ALIGN (4)
+L(46bytes):
+	mov	-46(%eax), %ecx
+	mov	-46(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(42bytes):
+	mov	-42(%eax), %ecx
+	mov	-42(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(38bytes):
+	mov	-38(%eax), %ecx
+	mov	-38(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(34bytes):
+	mov	-34(%eax), %ecx
+	mov	-34(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(30bytes):
+	mov	-30(%eax), %ecx
+	mov	-30(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(26bytes):
+	mov	-26(%eax), %ecx
+	mov	-26(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(22bytes):
+	mov	-22(%eax), %ecx
+	mov	-22(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(18bytes):
+	mov	-18(%eax), %ecx
+	mov	-18(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(14bytes):
+	mov	-14(%eax), %ecx
+	mov	-14(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(10bytes):
+	mov	-10(%eax), %ecx
+	mov	-10(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(6bytes):
+	mov	-6(%eax), %ecx
+	mov	-6(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(2bytes):
+	movzwl	-2(%eax), %ecx
+	movzwl	-2(%edx), %ebx
+	cmp	%bl, %cl
+	jne	L(end)
+	cmp	%bh, %ch
+	mov	$0, %eax
+	jne	L(end)
+	POP (%ebx)
+	ret
+	CFI_PUSH (%ebx)
+
+	ALIGN (4)
+L(47bytes):
+	movl	-47(%eax), %ecx
+	movl	-47(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(43bytes):
+	movl	-43(%eax), %ecx
+	movl	-43(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(39bytes):
+	movl	-39(%eax), %ecx
+	movl	-39(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(35bytes):
+	movl	-35(%eax), %ecx
+	movl	-35(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(31bytes):
+	movl	-31(%eax), %ecx
+	movl	-31(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(27bytes):
+	movl	-27(%eax), %ecx
+	movl	-27(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(23bytes):
+	movl	-23(%eax), %ecx
+	movl	-23(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(19bytes):
+	movl	-19(%eax), %ecx
+	movl	-19(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(15bytes):
+	movl	-15(%eax), %ecx
+	movl	-15(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(11bytes):
+	movl	-11(%eax), %ecx
+	movl	-11(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(7bytes):
+	movl	-7(%eax), %ecx
+	movl	-7(%edx), %ebx
+	cmp	%ebx, %ecx
+	jne	L(find_diff)
+L(3bytes):
+	movzwl	-3(%eax), %ecx
+	movzwl	-3(%edx), %ebx
+	cmpb	%bl, %cl
+	jne	L(end)
+	cmp	%bx, %cx
+	jne	L(end)
+	movzbl	-1(%eax), %eax
+	cmpb	-1(%edx), %al
+	mov	$0, %eax
+	jne	L(end)
+	POP (%ebx)
+	ret
+	CFI_PUSH (%ebx)
+
+	ALIGN (4)
+L(find_diff):
+	cmpb	%bl, %cl
+	jne	L(end)
+	cmp	%bx, %cx
+	jne	L(end)
+	shr	$16,%ecx
+	shr	$16,%ebx
+	cmp	%bl, %cl
+	jne	L(end)
+	cmp	%bx, %cx
+L(end):
+	POP (%ebx)
+	mov	$1, %eax
+	ja	L(bigger)
+	neg	%eax
+L(bigger):
+	ret
+
+END (MEMCMP)
diff --git a/libc/arch-x86/string/ssse3-memcpy5.S b/libc/arch-x86/string/ssse3-memcpy5.S
new file mode 100644
index 0000000..6b90402
--- /dev/null
+++ b/libc/arch-x86/string/ssse3-memcpy5.S
@@ -0,0 +1,1770 @@
+/*
+Copyright (c) 2010, Intel Corporation
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+    * Redistributions of source code must retain the above copyright notice,
+    * this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright notice,
+    * this list of conditions and the following disclaimer in the documentation
+    * and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its contributors
+    * may be used to endorse or promote products derived from this software
+    * without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef MEMCPY
+# define MEMCPY         ssse3_memcpy5
+#endif
+
+#ifndef L
+# define L(label)	.L##label
+#endif
+
+#ifndef ALIGN
+# define ALIGN(n)	.p2align n
+#endif
+
+#ifndef cfi_startproc
+# define cfi_startproc			.cfi_startproc
+#endif
+
+#ifndef cfi_endproc
+# define cfi_endproc			.cfi_endproc
+#endif
+
+#ifndef cfi_rel_offset
+# define cfi_rel_offset(reg, off)	.cfi_rel_offset reg, off
+#endif
+
+#ifndef cfi_restore
+# define cfi_restore(reg)		.cfi_restore (reg)
+#endif
+
+#ifndef cfi_adjust_cfa_offset
+# define cfi_adjust_cfa_offset(off)	.cfi_adjust_cfa_offset off
+#endif
+
+#ifndef ENTRY
+# define ENTRY(name)			\
+	.type name,  @function; 	\
+	.globl name;			\
+	.p2align 4;			\
+name:					\
+	cfi_startproc
+#endif
+
+#ifndef END
+# define END(name)			\
+	cfi_endproc;			\
+	.size name, .-name
+#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
+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
+	jae	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
+	jae	L(48bytesormore)
+
+L(fwd_write_less32bytes):
+#ifndef USE_AS_MEMMOVE
+	cmp	%dl, %al
+	jb	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
+	jae	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
+	jb	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
+	jb	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
+	jb	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
+	jae	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
+
+	jae	L(shl_0_gobble_cache_loop)
+	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
+	jb	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
+	jb	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
+
+	jae	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
+	jb	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
+	jb	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)
+
+	jb	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)
+
+	jb	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)
+
+	jb	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)
+
+	jb	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)
+
+	jb	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)
+
+	jb	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)
+
+	jb	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)
+
+	jb	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)
+
+	jb	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)
+
+	jb	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)
+
+	jb	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)
+
+	jb	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)
+
+	jb	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)
+
+	jb	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)
+
+	jb	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
+	jb	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
+	jae	L(bk_write_more64bytes)
+
+L(bk_write_64bytesless):
+	cmp	$32, %ecx
+	jb	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
+	jbe	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
+	jb	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
+	jae	L(bk_ssse3_cpy)
+	jmp	L(bk_write_64bytesless)
+
+#endif
+
+END (MEMCPY)
diff --git a/libc/arch-x86/string/ssse3-strcmp.S b/libc/arch-x86/string/ssse3-strcmp.S
new file mode 100644
index 0000000..cfb2e9f
--- /dev/null
+++ b/libc/arch-x86/string/ssse3-strcmp.S
@@ -0,0 +1,2265 @@
+/*
+Copyright (c) 2010, Intel Corporation
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+    * Redistributions of source code must retain the above copyright notice,
+    * this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright notice,
+    * this list of conditions and the following disclaimer in the documentation
+    * and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its contributors
+    * may be used to endorse or promote products derived from this software
+    * without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef L
+# define L(label)	.L##label
+#endif
+
+#ifndef cfi_startproc
+# define cfi_startproc			.cfi_startproc
+#endif
+
+#ifndef cfi_endproc
+# define cfi_endproc			.cfi_endproc
+#endif
+
+#ifndef cfi_rel_offset
+# define cfi_rel_offset(reg, off)	.cfi_rel_offset reg, off
+#endif
+
+#ifndef cfi_restore
+# define cfi_restore(reg)		.cfi_restore (reg)
+#endif
+
+#ifndef cfi_adjust_cfa_offset
+# define cfi_adjust_cfa_offset(off)	.cfi_adjust_cfa_offset off
+#endif
+
+#ifndef ENTRY
+# define ENTRY(name)			\
+	.type name,  @function; 	\
+	.globl name;			\
+	.p2align 4;			\
+name:					\
+	cfi_startproc
+#endif
+
+#ifndef END
+# define END(name)			\
+	cfi_endproc;			\
+	.size name, .-name
+#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)
+
+#ifndef USE_AS_STRNCMP
+# define STR1		4
+# define STR2		STR1+4
+# define RETURN		ret
+
+# define UPDATE_STRNCMP_COUNTER
+#else
+# define STR1		8
+# define STR2		STR1+4
+# define CNT		STR2+4
+# define RETURN		POP (%ebp); ret; CFI_PUSH (%ebp)
+
+# define UPDATE_STRNCMP_COUNTER				\
+	/* calculate left number to compare */		\
+	mov	$16, %esi;				\
+	sub	%ecx, %esi;				\
+	cmp	%esi, %ebp;				\
+	jbe	L(more8byteseq);			\
+	sub	%esi, %ebp
+#endif
+
+	.section .text.ssse3,"ax",@progbits
+ENTRY (ssse3_strcmp_latest)
+#ifdef USE_AS_STRNCMP
+	PUSH	(%ebp)
+#endif
+	movl	STR1(%esp), %edx
+	movl	STR2(%esp), %eax
+#ifdef USE_AS_STRNCMP
+	movl	CNT(%esp), %ebp
+	cmp	$16, %ebp
+	jb	L(less16bytes_sncmp)
+	jmp	L(more16bytes)
+#endif
+
+	movzbl	(%eax), %ecx
+	cmpb	%cl, (%edx)
+	jne	L(neq)
+	cmpl	$0, %ecx
+	je	L(eq)
+
+	movzbl	1(%eax), %ecx
+	cmpb	%cl, 1(%edx)
+	jne	L(neq)
+	cmpl	$0, %ecx
+	je	L(eq)
+
+	movzbl	2(%eax), %ecx
+	cmpb	%cl, 2(%edx)
+	jne	L(neq)
+	cmpl	$0, %ecx
+	je	L(eq)
+
+	movzbl	3(%eax), %ecx
+	cmpb	%cl, 3(%edx)
+	jne	L(neq)
+	cmpl	$0, %ecx
+	je	L(eq)
+
+	movzbl	4(%eax), %ecx
+	cmpb	%cl, 4(%edx)
+	jne	L(neq)
+	cmpl	$0, %ecx
+	je	L(eq)
+
+	movzbl	5(%eax), %ecx
+	cmpb	%cl, 5(%edx)
+	jne	L(neq)
+	cmpl	$0, %ecx
+	je	L(eq)
+
+	movzbl	6(%eax), %ecx
+	cmpb	%cl, 6(%edx)
+	jne	L(neq)
+	cmpl	$0, %ecx
+	je	L(eq)
+
+	movzbl	7(%eax), %ecx
+	cmpb	%cl, 7(%edx)
+	jne	L(neq)
+	cmpl	$0, %ecx
+	je	L(eq)
+
+	add	$8, %edx
+	add	$8, %eax
+#ifdef USE_AS_STRNCMP
+	cmp	$8, %ebp
+	lea	-8(%ebp), %ebp
+	je	L(eq)
+L(more16bytes):
+#endif
+	movl	%edx, %ecx
+	and	$0xfff, %ecx
+	cmp	$0xff0, %ecx
+	ja	L(crosspage)
+	mov	%eax, %ecx
+	and	$0xfff, %ecx
+	cmp	$0xff0, %ecx
+	ja	L(crosspage)
+	pxor	%xmm0, %xmm0
+	movlpd	(%eax), %xmm1
+	movlpd	(%edx), %xmm2
+	movhpd	8(%eax), %xmm1
+	movhpd	8(%edx), %xmm2
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %ecx
+	sub	$0xffff, %ecx
+	jnz	L(less16bytes)
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(eq)
+#endif
+	add	$16, %eax
+	add	$16, %edx
+
+L(crosspage):
+
+	PUSH	(%ebx)
+	PUSH	(%edi)
+	PUSH	(%esi)
+
+	movl	%edx, %edi
+	movl	%eax, %ecx
+	and	$0xf, %ecx
+	and	$0xf, %edi
+	xor	%ecx, %eax
+	xor	%edi, %edx
+	xor	%ebx, %ebx
+	cmp	%edi, %ecx
+	je	L(ashr_0)
+	ja	L(bigger)
+	or	$0x20, %ebx
+	xchg	%edx, %eax
+	xchg	%ecx, %edi
+L(bigger):
+	lea	15(%edi), %edi
+	sub	%ecx, %edi
+	cmp	$8, %edi
+	jle	L(ashr_less_8)
+	cmp	$14, %edi
+	je	L(ashr_15)
+	cmp	$13, %edi
+	je	L(ashr_14)
+	cmp	$12, %edi
+	je	L(ashr_13)
+	cmp	$11, %edi
+	je	L(ashr_12)
+	cmp	$10, %edi
+	je	L(ashr_11)
+	cmp	$9, %edi
+	je	L(ashr_10)
+L(ashr_less_8):
+	je	L(ashr_9)
+	cmp	$7, %edi
+	je	L(ashr_8)
+	cmp	$6, %edi
+	je	L(ashr_7)
+	cmp	$5, %edi
+	je	L(ashr_6)
+	cmp	$4, %edi
+	je	L(ashr_5)
+	cmp	$3, %edi
+	je	L(ashr_4)
+	cmp	$2, %edi
+	je	L(ashr_3)
+	cmp	$1, %edi
+	je	L(ashr_2)
+	cmp	$0, %edi
+	je	L(ashr_1)
+
+/*
+ * The following cases will be handled by ashr_0
+ *  ecx(offset of esi)  eax(offset of edi)  relative offset  corresponding case
+ *        n(0~15)            n(0~15)           15(15+ n-n)         ashr_0
+ */
+	.p2align 4
+L(ashr_0):
+	mov	$0xffff, %esi
+	movdqa	(%eax), %xmm1
+	pxor	%xmm0, %xmm0
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	(%edx), %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %edi
+	shr	%cl, %esi
+	shr	%cl, %edi
+	sub	%edi, %esi
+	mov	%ecx, %edi
+	jne	L(less32bytes)
+	UPDATE_STRNCMP_COUNTER
+	mov	$0x10, %ebx
+	mov	$0x10, %ecx
+	pxor	%xmm0, %xmm0
+	.p2align 4
+L(loop_ashr_0):
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	add	$16, %ecx
+	jmp	L(loop_ashr_0)
+
+/*
+ * The following cases will be handled by ashr_1
+ * ecx(offset of esi)  eax(offset of edi)   relative offset   	corresponding case
+ *        n(15)            n -15            0(15 +(n-15) - n)         ashr_1
+ */
+	.p2align 4
+L(ashr_1):
+	mov	$0xffff, %esi
+	pxor	%xmm0, %xmm0
+	movdqa	(%edx), %xmm2
+	movdqa	(%eax), %xmm1
+	pcmpeqb	%xmm1, %xmm0
+	pslldq	$15, %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	psubb	%xmm0, %xmm2
+	pmovmskb %xmm2, %edi
+	shr	%cl, %esi
+	shr	%cl, %edi
+	sub	%edi, %esi
+	lea	-15(%ecx), %edi
+	jnz	L(less32bytes)
+
+	UPDATE_STRNCMP_COUNTER
+
+	movdqa	(%edx), %xmm3
+	pxor	%xmm0, %xmm0
+	mov	$16, %ecx
+	or	$1, %ebx
+	lea	1(%edx), %edi
+	and	$0xfff, %edi
+	sub	$0x1000, %edi
+
+	.p2align 4
+L(loop_ashr_1):
+	add	$16, %edi
+	jg	L(nibble_ashr_1)
+
+L(gobble_ashr_1):
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$1, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+
+	add	$16, %edi
+	jg	L(nibble_ashr_1)
+
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$1, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+	jmp	L(loop_ashr_1)
+
+	.p2align 4
+L(nibble_ashr_1):
+	pcmpeqb	%xmm3, %xmm0
+	pmovmskb %xmm0, %esi
+	test	$0xfffe, %esi
+	jnz	L(ashr_1_exittail)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$15, %ebp
+	jbe	L(ashr_1_exittail)
+#endif
+	pxor	%xmm0, %xmm0
+	sub	$0x1000, %edi
+	jmp	L(gobble_ashr_1)
+
+	.p2align 4
+L(ashr_1_exittail):
+	movdqa	(%eax, %ecx), %xmm1
+	psrldq	$1, %xmm0
+	psrldq	$1, %xmm3
+	jmp	L(aftertail)
+
+/*
+ * The following cases will be handled by ashr_2
+ * ecx(offset of esi)  eax(offset of edi)   relative offset   	corresponding case
+ *        n(14~15)            n -14            1(15 +(n-14) - n)         ashr_2
+ */
+	.p2align 4
+L(ashr_2):
+	mov	$0xffff, %esi
+	pxor	%xmm0, %xmm0
+	movdqa	(%edx), %xmm2
+	movdqa	(%eax), %xmm1
+	pcmpeqb	%xmm1, %xmm0
+	pslldq	$14, %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	psubb	%xmm0, %xmm2
+	pmovmskb %xmm2, %edi
+	shr	%cl, %esi
+	shr	%cl, %edi
+	sub	%edi, %esi
+	lea	-14(%ecx), %edi
+	jnz	L(less32bytes)
+
+	UPDATE_STRNCMP_COUNTER
+
+	movdqa	(%edx), %xmm3
+	pxor	%xmm0, %xmm0
+	mov	$16, %ecx
+	or	$2, %ebx
+	lea	2(%edx), %edi
+	and	$0xfff, %edi
+	sub	$0x1000, %edi
+
+	.p2align 4
+L(loop_ashr_2):
+	add	$16, %edi
+	jg	L(nibble_ashr_2)
+
+L(gobble_ashr_2):
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$2, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+
+	add	$16, %edi
+	jg	L(nibble_ashr_2)
+
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$2, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+	jmp	L(loop_ashr_2)
+
+	.p2align 4
+L(nibble_ashr_2):
+	pcmpeqb	%xmm3, %xmm0
+	pmovmskb %xmm0, %esi
+	test	$0xfffc, %esi
+	jnz	L(ashr_2_exittail)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$14, %ebp
+	jbe	L(ashr_2_exittail)
+#endif
+
+	pxor	%xmm0, %xmm0
+	sub	$0x1000, %edi
+	jmp	L(gobble_ashr_2)
+
+	.p2align 4
+L(ashr_2_exittail):
+	movdqa	(%eax, %ecx), %xmm1
+	psrldq	$2, %xmm0
+	psrldq	$2, %xmm3
+	jmp	L(aftertail)
+
+/*
+ * The following cases will be handled by ashr_3
+ * ecx(offset of esi)  eax(offset of edi)   relative offset   	corresponding case
+ *        n(13~15)            n -13            2(15 +(n-13) - n)         ashr_3
+ */
+	.p2align 4
+L(ashr_3):
+	mov	$0xffff, %esi
+	pxor	%xmm0, %xmm0
+	movdqa	(%edx), %xmm2
+	movdqa	(%eax), %xmm1
+	pcmpeqb	%xmm1, %xmm0
+	pslldq	$13, %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	psubb	%xmm0, %xmm2
+	pmovmskb %xmm2, %edi
+	shr	%cl, %esi
+	shr	%cl, %edi
+	sub	%edi, %esi
+	lea	-13(%ecx), %edi
+	jnz	L(less32bytes)
+
+	UPDATE_STRNCMP_COUNTER
+
+	movdqa	(%edx), %xmm3
+	pxor	%xmm0, %xmm0
+	mov	$16, %ecx
+	or	$3, %ebx
+	lea	3(%edx), %edi
+	and	$0xfff, %edi
+	sub	$0x1000, %edi
+
+	.p2align 4
+L(loop_ashr_3):
+	add	$16, %edi
+	jg	L(nibble_ashr_3)
+
+L(gobble_ashr_3):
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$3, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+
+	add	$16, %edi
+	jg	L(nibble_ashr_3)
+
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$3, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+	jmp	L(loop_ashr_3)
+
+	.p2align 4
+L(nibble_ashr_3):
+	pcmpeqb	%xmm3, %xmm0
+	pmovmskb %xmm0, %esi
+	test	$0xfff8, %esi
+	jnz	L(ashr_3_exittail)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$13, %ebp
+	jbe	L(ashr_3_exittail)
+#endif
+	pxor	%xmm0, %xmm0
+	sub	$0x1000, %edi
+	jmp	L(gobble_ashr_3)
+
+	.p2align 4
+L(ashr_3_exittail):
+	movdqa	(%eax, %ecx), %xmm1
+	psrldq	$3, %xmm0
+	psrldq	$3, %xmm3
+	jmp	L(aftertail)
+
+/*
+ * The following cases will be handled by ashr_4
+ * ecx(offset of esi)  eax(offset of edi)   relative offset   	corresponding case
+ *        n(12~15)            n -12            3(15 +(n-12) - n)         ashr_4
+ */
+	.p2align 4
+L(ashr_4):
+	mov	$0xffff, %esi
+	pxor	%xmm0, %xmm0
+	movdqa	(%edx), %xmm2
+	movdqa	(%eax), %xmm1
+	pcmpeqb	%xmm1, %xmm0
+	pslldq	$12, %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	psubb	%xmm0, %xmm2
+	pmovmskb %xmm2, %edi
+	shr	%cl, %esi
+	shr	%cl, %edi
+	sub	%edi, %esi
+	lea	-12(%ecx), %edi
+	jnz	L(less32bytes)
+
+	UPDATE_STRNCMP_COUNTER
+
+	movdqa	(%edx), %xmm3
+	pxor	%xmm0, %xmm0
+	mov	$16, %ecx
+	or	$4, %ebx
+	lea	4(%edx), %edi
+	and	$0xfff, %edi
+	sub	$0x1000, %edi
+
+	.p2align 4
+L(loop_ashr_4):
+	add	$16, %edi
+	jg	L(nibble_ashr_4)
+
+L(gobble_ashr_4):
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$4, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+
+	add	$16, %edi
+	jg	L(nibble_ashr_4)
+
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$4, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+	jmp	L(loop_ashr_4)
+
+	.p2align 4
+L(nibble_ashr_4):
+	pcmpeqb	%xmm3, %xmm0
+	pmovmskb %xmm0, %esi
+	test	$0xfff0, %esi
+	jnz	L(ashr_4_exittail)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$12, %ebp
+	jbe	L(ashr_4_exittail)
+#endif
+
+	pxor	%xmm0, %xmm0
+	sub	$0x1000, %edi
+	jmp	L(gobble_ashr_4)
+
+	.p2align 4
+L(ashr_4_exittail):
+	movdqa	(%eax, %ecx), %xmm1
+	psrldq	$4, %xmm0
+	psrldq	$4, %xmm3
+	jmp	L(aftertail)
+
+/*
+ * The following cases will be handled by ashr_5
+ * ecx(offset of esi)  eax(offset of edi)   relative offset   	corresponding case
+ *        n(11~15)            n -11            4(15 +(n-11) - n)         ashr_5
+ */
+	.p2align 4
+L(ashr_5):
+	mov	$0xffff, %esi
+	pxor	%xmm0, %xmm0
+	movdqa	(%edx), %xmm2
+	movdqa	(%eax), %xmm1
+	pcmpeqb	%xmm1, %xmm0
+	pslldq	$11, %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	psubb	%xmm0, %xmm2
+	pmovmskb %xmm2, %edi
+	shr	%cl, %esi
+	shr	%cl, %edi
+	sub	%edi, %esi
+	lea	-11(%ecx), %edi
+	jnz	L(less32bytes)
+
+	UPDATE_STRNCMP_COUNTER
+
+	movdqa	(%edx), %xmm3
+	pxor	%xmm0, %xmm0
+	mov	$16, %ecx
+	or	$5, %ebx
+	lea	5(%edx), %edi
+	and	$0xfff, %edi
+	sub	$0x1000, %edi
+
+	.p2align 4
+L(loop_ashr_5):
+	add	$16, %edi
+	jg	L(nibble_ashr_5)
+
+L(gobble_ashr_5):
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$5, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+
+	add	$16, %edi
+	jg	L(nibble_ashr_5)
+
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$5, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+	jmp	L(loop_ashr_5)
+
+	.p2align 4
+L(nibble_ashr_5):
+	pcmpeqb	%xmm3, %xmm0
+	pmovmskb %xmm0, %esi
+	test	$0xffe0, %esi
+	jnz	L(ashr_5_exittail)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$11, %ebp
+	jbe	L(ashr_5_exittail)
+#endif
+	pxor	%xmm0, %xmm0
+	sub	$0x1000, %edi
+	jmp	L(gobble_ashr_5)
+
+	.p2align 4
+L(ashr_5_exittail):
+	movdqa	(%eax, %ecx), %xmm1
+	psrldq	$5, %xmm0
+	psrldq	$5, %xmm3
+	jmp	L(aftertail)
+
+/*
+ * The following cases will be handled by ashr_6
+ * ecx(offset of esi)  eax(offset of edi)   relative offset   	corresponding case
+ *        n(10~15)            n -10            5(15 +(n-10) - n)         ashr_6
+ */
+
+	.p2align 4
+L(ashr_6):
+	mov	$0xffff, %esi
+	pxor	%xmm0, %xmm0
+	movdqa	(%edx), %xmm2
+	movdqa	(%eax), %xmm1
+	pcmpeqb	%xmm1, %xmm0
+	pslldq	$10, %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	psubb	%xmm0, %xmm2
+	pmovmskb %xmm2, %edi
+	shr	%cl, %esi
+	shr	%cl, %edi
+	sub	%edi, %esi
+	lea	-10(%ecx), %edi
+	jnz	L(less32bytes)
+
+	UPDATE_STRNCMP_COUNTER
+
+	movdqa	(%edx), %xmm3
+	pxor	%xmm0, %xmm0
+	mov	$16, %ecx
+	or	$6, %ebx
+	lea	6(%edx), %edi
+	and	$0xfff, %edi
+	sub	$0x1000, %edi
+
+	.p2align 4
+L(loop_ashr_6):
+	add	$16, %edi
+	jg	L(nibble_ashr_6)
+
+L(gobble_ashr_6):
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$6, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+
+	add	$16, %edi
+	jg	L(nibble_ashr_6)
+
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$6, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+	jmp	L(loop_ashr_6)
+
+	.p2align 4
+L(nibble_ashr_6):
+	pcmpeqb	%xmm3, %xmm0
+	pmovmskb %xmm0, %esi
+	test	$0xffc0, %esi
+	jnz	L(ashr_6_exittail)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$10, %ebp
+	jbe	L(ashr_6_exittail)
+#endif
+	pxor	%xmm0, %xmm0
+	sub	$0x1000, %edi
+	jmp	L(gobble_ashr_6)
+
+	.p2align 4
+L(ashr_6_exittail):
+	movdqa	(%eax, %ecx), %xmm1
+	psrldq	$6, %xmm0
+	psrldq	$6, %xmm3
+	jmp	L(aftertail)
+
+/*
+ * The following cases will be handled by ashr_7
+ * ecx(offset of esi)  eax(offset of edi)   relative offset   	corresponding case
+ *        n(9~15)            n - 9            6(15 +(n-9) - n)         ashr_7
+ */
+
+	.p2align 4
+L(ashr_7):
+	mov	$0xffff, %esi
+	pxor	%xmm0, %xmm0
+	movdqa	(%edx), %xmm2
+	movdqa	(%eax), %xmm1
+	pcmpeqb	%xmm1, %xmm0
+	pslldq	$9, %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	psubb	%xmm0, %xmm2
+	pmovmskb %xmm2, %edi
+	shr	%cl, %esi
+	shr	%cl, %edi
+	sub	%edi, %esi
+	lea	-9(%ecx), %edi
+	jnz	L(less32bytes)
+
+	UPDATE_STRNCMP_COUNTER
+
+	movdqa	(%edx), %xmm3
+	pxor	%xmm0, %xmm0
+	mov	$16, %ecx
+	or	$7, %ebx
+	lea	8(%edx), %edi
+	and	$0xfff, %edi
+	sub	$0x1000, %edi
+
+	.p2align 4
+L(loop_ashr_7):
+	add	$16, %edi
+	jg	L(nibble_ashr_7)
+
+L(gobble_ashr_7):
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$7, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+
+	add	$16, %edi
+	jg	L(nibble_ashr_7)
+
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$7, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+	jmp	L(loop_ashr_7)
+
+	.p2align 4
+L(nibble_ashr_7):
+	pcmpeqb	%xmm3, %xmm0
+	pmovmskb %xmm0, %esi
+	test	$0xff80, %esi
+	jnz	L(ashr_7_exittail)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$9, %ebp
+	jbe	L(ashr_7_exittail)
+#endif
+	pxor	%xmm0, %xmm0
+	pxor	%xmm0, %xmm0
+	sub	$0x1000, %edi
+	jmp	L(gobble_ashr_7)
+
+	.p2align 4
+L(ashr_7_exittail):
+	movdqa	(%eax, %ecx), %xmm1
+	psrldq	$7, %xmm0
+	psrldq	$7, %xmm3
+	jmp	L(aftertail)
+
+/*
+ * The following cases will be handled by ashr_8
+ * ecx(offset of esi)  eax(offset of edi)   relative offset   	corresponding case
+ *        n(8~15)            n - 8            7(15 +(n-8) - n)         ashr_8
+ */
+	.p2align 4
+L(ashr_8):
+	mov	$0xffff, %esi
+	pxor	%xmm0, %xmm0
+	movdqa	(%edx), %xmm2
+	movdqa	(%eax), %xmm1
+	pcmpeqb	%xmm1, %xmm0
+	pslldq	$8, %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	psubb	%xmm0, %xmm2
+	pmovmskb %xmm2, %edi
+	shr	%cl, %esi
+	shr	%cl, %edi
+	sub	%edi, %esi
+	lea	-8(%ecx), %edi
+	jnz	L(less32bytes)
+
+	UPDATE_STRNCMP_COUNTER
+
+	movdqa	(%edx), %xmm3
+	pxor	%xmm0, %xmm0
+	mov	$16, %ecx
+	or	$8, %ebx
+	lea	8(%edx), %edi
+	and	$0xfff, %edi
+	sub	$0x1000, %edi
+
+	.p2align 4
+L(loop_ashr_8):
+	add	$16, %edi
+	jg	L(nibble_ashr_8)
+
+L(gobble_ashr_8):
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$8, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+
+	add	$16, %edi
+	jg	L(nibble_ashr_8)
+
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$8, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+	jmp	L(loop_ashr_8)
+
+	.p2align 4
+L(nibble_ashr_8):
+	pcmpeqb	%xmm3, %xmm0
+	pmovmskb %xmm0, %esi
+	test	$0xff00, %esi
+	jnz	L(ashr_8_exittail)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$8, %ebp
+	jbe	L(ashr_8_exittail)
+#endif
+	pxor	%xmm0, %xmm0
+	pxor	%xmm0, %xmm0
+	sub	$0x1000, %edi
+	jmp	L(gobble_ashr_8)
+
+	.p2align 4
+L(ashr_8_exittail):
+	movdqa	(%eax, %ecx), %xmm1
+	psrldq	$8, %xmm0
+	psrldq	$8, %xmm3
+	jmp	L(aftertail)
+
+/*
+ * The following cases will be handled by ashr_9
+ * ecx(offset of esi)  eax(offset of edi)   relative offset   	corresponding case
+ *        n(7~15)            n - 7            8(15 +(n-7) - n)         ashr_9
+ */
+	.p2align 4
+L(ashr_9):
+	mov	$0xffff, %esi
+	pxor	%xmm0, %xmm0
+	movdqa	(%edx), %xmm2
+	movdqa	(%eax), %xmm1
+	pcmpeqb	%xmm1, %xmm0
+	pslldq	$7, %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	psubb	%xmm0, %xmm2
+	pmovmskb %xmm2, %edi
+	shr	%cl, %esi
+	shr	%cl, %edi
+	sub	%edi, %esi
+	lea	-7(%ecx), %edi
+	jnz	L(less32bytes)
+
+	UPDATE_STRNCMP_COUNTER
+
+	movdqa	(%edx), %xmm3
+	pxor	%xmm0, %xmm0
+	mov	$16, %ecx
+	or	$9, %ebx
+	lea	9(%edx), %edi
+	and	$0xfff, %edi
+	sub	$0x1000, %edi
+
+	.p2align 4
+L(loop_ashr_9):
+	add	$16, %edi
+	jg	L(nibble_ashr_9)
+
+L(gobble_ashr_9):
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$9, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+
+	add	$16, %edi
+	jg	L(nibble_ashr_9)
+
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$9, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+	jmp	L(loop_ashr_9)
+
+	.p2align 4
+L(nibble_ashr_9):
+	pcmpeqb	%xmm3, %xmm0
+	pmovmskb %xmm0, %esi
+	test	$0xfe00, %esi
+	jnz	L(ashr_9_exittail)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$7, %ebp
+	jbe	L(ashr_9_exittail)
+#endif
+	pxor	%xmm0, %xmm0
+	sub	$0x1000, %edi
+	jmp	L(gobble_ashr_9)
+
+	.p2align 4
+L(ashr_9_exittail):
+	movdqa	(%eax, %ecx), %xmm1
+	psrldq	$9, %xmm0
+	psrldq	$9, %xmm3
+	jmp	L(aftertail)
+
+/*
+ * The following cases will be handled by ashr_10
+ * ecx(offset of esi)  eax(offset of edi)   relative offset   	corresponding case
+ *        n(6~15)            n - 6            9(15 +(n-6) - n)         ashr_10
+ */
+	.p2align 4
+L(ashr_10):
+	mov	$0xffff, %esi
+	pxor	%xmm0, %xmm0
+	movdqa	(%edx), %xmm2
+	movdqa	(%eax), %xmm1
+	pcmpeqb	%xmm1, %xmm0
+	pslldq	$6, %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	psubb	%xmm0, %xmm2
+	pmovmskb %xmm2, %edi
+	shr	%cl, %esi
+	shr	%cl, %edi
+	sub	%edi, %esi
+	lea	-6(%ecx), %edi
+	jnz	L(less32bytes)
+
+	UPDATE_STRNCMP_COUNTER
+
+	movdqa	(%edx), %xmm3
+	pxor	%xmm0, %xmm0
+	mov	$16, %ecx
+	or	$10, %ebx
+	lea	10(%edx), %edi
+	and	$0xfff, %edi
+	sub	$0x1000, %edi
+
+	.p2align 4
+L(loop_ashr_10):
+	add	$16, %edi
+	jg	L(nibble_ashr_10)
+
+L(gobble_ashr_10):
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$10, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+
+	add	$16, %edi
+	jg	L(nibble_ashr_10)
+
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$10, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+	jmp	L(loop_ashr_10)
+
+	.p2align 4
+L(nibble_ashr_10):
+	pcmpeqb	%xmm3, %xmm0
+	pmovmskb %xmm0, %esi
+	test	$0xfc00, %esi
+	jnz	L(ashr_10_exittail)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$6, %ebp
+	jbe	L(ashr_10_exittail)
+#endif
+	pxor	%xmm0, %xmm0
+	sub	$0x1000, %edi
+	jmp	L(gobble_ashr_10)
+
+	.p2align 4
+L(ashr_10_exittail):
+	movdqa	(%eax, %ecx), %xmm1
+	psrldq	$10, %xmm0
+	psrldq	$10, %xmm3
+	jmp	L(aftertail)
+
+/*
+ * The following cases will be handled by ashr_11
+ * ecx(offset of esi)  eax(offset of edi)   relative offset   	corresponding case
+ *        n(5~15)            n - 5            10(15 +(n-5) - n)         ashr_11
+ */
+	.p2align 4
+L(ashr_11):
+	mov	$0xffff, %esi
+	pxor	%xmm0, %xmm0
+	movdqa	(%edx), %xmm2
+	movdqa	(%eax), %xmm1
+	pcmpeqb	%xmm1, %xmm0
+	pslldq	$5, %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	psubb	%xmm0, %xmm2
+	pmovmskb %xmm2, %edi
+	shr	%cl, %esi
+	shr	%cl, %edi
+	sub	%edi, %esi
+	lea	-5(%ecx), %edi
+	jnz	L(less32bytes)
+
+	UPDATE_STRNCMP_COUNTER
+
+	movdqa	(%edx), %xmm3
+	pxor	%xmm0, %xmm0
+	mov	$16, %ecx
+	or	$11, %ebx
+	lea	11(%edx), %edi
+	and	$0xfff, %edi
+	sub	$0x1000, %edi
+
+	.p2align 4
+L(loop_ashr_11):
+	add	$16, %edi
+	jg	L(nibble_ashr_11)
+
+L(gobble_ashr_11):
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$11, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+
+	add	$16, %edi
+	jg	L(nibble_ashr_11)
+
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$11, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+	jmp	L(loop_ashr_11)
+
+	.p2align 4
+L(nibble_ashr_11):
+	pcmpeqb	%xmm3, %xmm0
+	pmovmskb %xmm0, %esi
+	test	$0xf800, %esi
+	jnz	L(ashr_11_exittail)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$5, %ebp
+	jbe	L(ashr_11_exittail)
+#endif
+	pxor	%xmm0, %xmm0
+	sub	$0x1000, %edi
+	jmp	L(gobble_ashr_11)
+
+	.p2align 4
+L(ashr_11_exittail):
+	movdqa	(%eax, %ecx), %xmm1
+	psrldq	$11, %xmm0
+	psrldq	$11, %xmm3
+	jmp	L(aftertail)
+
+/*
+ * The following cases will be handled by ashr_12
+ * ecx(offset of esi)  eax(offset of edi)   relative offset   	corresponding case
+ *        n(4~15)            n - 4            11(15 +(n-4) - n)         ashr_12
+ */
+	.p2align 4
+L(ashr_12):
+	mov	$0xffff, %esi
+	pxor	%xmm0, %xmm0
+	movdqa	(%edx), %xmm2
+	movdqa	(%eax), %xmm1
+	pcmpeqb	%xmm1, %xmm0
+	pslldq	$4, %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	psubb	%xmm0, %xmm2
+	pmovmskb %xmm2, %edi
+	shr	%cl, %esi
+	shr	%cl, %edi
+	sub	%edi, %esi
+	lea	-4(%ecx), %edi
+	jnz	L(less32bytes)
+
+	UPDATE_STRNCMP_COUNTER
+
+	movdqa	(%edx), %xmm3
+	pxor	%xmm0, %xmm0
+	mov	$16, %ecx
+	or	$12, %ebx
+	lea	12(%edx), %edi
+	and	$0xfff, %edi
+	sub	$0x1000, %edi
+
+	.p2align 4
+L(loop_ashr_12):
+	add	$16, %edi
+	jg	L(nibble_ashr_12)
+
+L(gobble_ashr_12):
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$12, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+
+	add	$16, %edi
+	jg	L(nibble_ashr_12)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$12, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+	jmp	L(loop_ashr_12)
+
+	.p2align 4
+L(nibble_ashr_12):
+	pcmpeqb	%xmm3, %xmm0
+	pmovmskb %xmm0, %esi
+	test	$0xf000, %esi
+	jnz	L(ashr_12_exittail)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$4, %ebp
+	jbe	L(ashr_12_exittail)
+#endif
+	pxor	%xmm0, %xmm0
+	sub	$0x1000, %edi
+	jmp	L(gobble_ashr_12)
+
+	.p2align 4
+L(ashr_12_exittail):
+	movdqa	(%eax, %ecx), %xmm1
+	psrldq	$12, %xmm0
+	psrldq	$12, %xmm3
+	jmp	L(aftertail)
+
+/*
+ * The following cases will be handled by ashr_13
+ * ecx(offset of esi)  eax(offset of edi)   relative offset   	corresponding case
+ *        n(3~15)            n - 3            12(15 +(n-3) - n)         ashr_13
+ */
+	.p2align 4
+L(ashr_13):
+	mov	$0xffff, %esi
+	pxor	%xmm0, %xmm0
+	movdqa	(%edx), %xmm2
+	movdqa	(%eax), %xmm1
+	pcmpeqb	%xmm1, %xmm0
+	pslldq	$3, %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	psubb	%xmm0, %xmm2
+	pmovmskb %xmm2, %edi
+	shr	%cl, %esi
+	shr	%cl, %edi
+	sub	%edi, %esi
+	lea	-3(%ecx), %edi
+	jnz	L(less32bytes)
+
+	UPDATE_STRNCMP_COUNTER
+
+	movdqa	(%edx), %xmm3
+	pxor	%xmm0, %xmm0
+	mov	$16, %ecx
+	or	$13, %ebx
+	lea	13(%edx), %edi
+	and	$0xfff, %edi
+	sub	$0x1000, %edi
+
+	.p2align 4
+L(loop_ashr_13):
+	add	$16, %edi
+	jg	L(nibble_ashr_13)
+
+L(gobble_ashr_13):
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$13, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+
+	add	$16, %edi
+	jg	L(nibble_ashr_13)
+
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$13, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+	jmp	L(loop_ashr_13)
+
+	.p2align 4
+L(nibble_ashr_13):
+	pcmpeqb	%xmm3, %xmm0
+	pmovmskb %xmm0, %esi
+	test	$0xe000, %esi
+	jnz	L(ashr_13_exittail)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$3, %ebp
+	jbe	L(ashr_13_exittail)
+#endif
+	pxor	%xmm0, %xmm0
+	sub	$0x1000, %edi
+	jmp	L(gobble_ashr_13)
+
+	.p2align 4
+L(ashr_13_exittail):
+	movdqa	(%eax, %ecx), %xmm1
+	psrldq	$13, %xmm0
+	psrldq	$13, %xmm3
+	jmp	L(aftertail)
+
+/*
+ * The following cases will be handled by ashr_14
+ * ecx(offset of esi)  eax(offset of edi)   relative offset   	corresponding case
+ *        n(2~15)            n - 2            13(15 +(n-2) - n)         ashr_14
+ */
+	.p2align 4
+L(ashr_14):
+	mov	$0xffff, %esi
+	pxor	%xmm0, %xmm0
+	movdqa	(%edx), %xmm2
+	movdqa	(%eax), %xmm1
+	pcmpeqb	%xmm1, %xmm0
+	pslldq	$2, %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	psubb	%xmm0, %xmm2
+	pmovmskb %xmm2, %edi
+	shr	%cl, %esi
+	shr	%cl, %edi
+	sub	%edi, %esi
+	lea	-2(%ecx), %edi
+	jnz	L(less32bytes)
+
+	UPDATE_STRNCMP_COUNTER
+
+	movdqa	(%edx), %xmm3
+	pxor	%xmm0, %xmm0
+	mov	$16, %ecx
+	or	$14, %ebx
+	lea	14(%edx), %edi
+	and	$0xfff, %edi
+	sub	$0x1000, %edi
+
+	.p2align 4
+L(loop_ashr_14):
+	add	$16, %edi
+	jg	L(nibble_ashr_14)
+
+L(gobble_ashr_14):
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$14, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+
+	add	$16, %edi
+	jg	L(nibble_ashr_14)
+
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$14, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+	jmp	L(loop_ashr_14)
+
+	.p2align 4
+L(nibble_ashr_14):
+	pcmpeqb	%xmm3, %xmm0
+	pmovmskb %xmm0, %esi
+	test	$0xc000, %esi
+	jnz	L(ashr_14_exittail)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$2, %ebp
+	jbe	L(ashr_14_exittail)
+#endif
+	pxor	%xmm0, %xmm0
+	sub	$0x1000, %edi
+	jmp	L(gobble_ashr_14)
+
+	.p2align 4
+L(ashr_14_exittail):
+	movdqa	(%eax, %ecx), %xmm1
+	psrldq	$14, %xmm0
+	psrldq	$14, %xmm3
+	jmp	L(aftertail)
+
+/*
+ * The following cases will be handled by ashr_14
+ * ecx(offset of esi)  eax(offset of edi)   relative offset   	corresponding case
+ *        n(1~15)            n - 1            14(15 +(n-1) - n)         ashr_15
+ */
+
+	.p2align 4
+L(ashr_15):
+	mov	$0xffff, %esi
+	pxor	%xmm0, %xmm0
+	movdqa	(%edx), %xmm2
+	movdqa	(%eax), %xmm1
+	pcmpeqb	%xmm1, %xmm0
+	pslldq	$1, %xmm2
+	pcmpeqb	%xmm1, %xmm2
+	psubb	%xmm0, %xmm2
+	pmovmskb %xmm2, %edi
+	shr	%cl, %esi
+	shr	%cl, %edi
+	sub	%edi, %esi
+	lea	-1(%ecx), %edi
+	jnz	L(less32bytes)
+
+	UPDATE_STRNCMP_COUNTER
+
+	movdqa	(%edx), %xmm3
+	pxor	%xmm0, %xmm0
+	mov	$16, %ecx
+	or	$15, %ebx
+	lea	15(%edx), %edi
+	and	$0xfff, %edi
+	sub	$0x1000, %edi
+
+	.p2align 4
+L(loop_ashr_15):
+	add	$16, %edi
+	jg	L(nibble_ashr_15)
+
+L(gobble_ashr_15):
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$15, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+
+	add	$16, %edi
+	jg	L(nibble_ashr_15)
+
+	movdqa	(%eax, %ecx), %xmm1
+	movdqa	(%edx, %ecx), %xmm2
+	movdqa	%xmm2, %xmm4
+
+	palignr	$15, %xmm3, %xmm2
+
+	pcmpeqb	%xmm1, %xmm0
+	pcmpeqb	%xmm2, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	sub	$0xffff, %esi
+	jnz	L(exit)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$16, %ebp
+	lea	-16(%ebp), %ebp
+	jbe	L(more8byteseq)
+#endif
+	add	$16, %ecx
+	movdqa	%xmm4, %xmm3
+	jmp	L(loop_ashr_15)
+
+	.p2align 4
+L(nibble_ashr_15):
+	pcmpeqb	%xmm3, %xmm0
+	pmovmskb %xmm0, %esi
+	test	$0x8000, %esi
+	jnz	L(ashr_15_exittail)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$1, %ebp
+	jbe	L(ashr_15_exittail)
+#endif
+	pxor	%xmm0, %xmm0
+	sub	$0x1000, %edi
+	jmp	L(gobble_ashr_15)
+
+	.p2align 4
+L(ashr_15_exittail):
+	movdqa	(%eax, %ecx), %xmm1
+	psrldq	$15, %xmm0
+	psrldq	$15, %xmm3
+	jmp	L(aftertail)
+
+	.p2align 4
+L(aftertail):
+	pcmpeqb	%xmm3, %xmm1
+	psubb	%xmm0, %xmm1
+	pmovmskb %xmm1, %esi
+	not	%esi
+L(exit):
+	mov	%ebx, %edi
+	and	$0x1f, %edi
+	lea	-16(%edi, %ecx), %edi
+L(less32bytes):
+	add	%edi, %edx
+	add	%ecx, %eax
+	test	$0x20, %ebx
+	jz	L(ret2)
+	xchg	%eax, %edx
+
+	.p2align 4
+L(ret2):
+	mov	%esi, %ecx
+	POP	(%esi)
+	POP	(%edi)
+	POP	(%ebx)
+L(less16bytes):
+	test	%cl, %cl
+	jz	L(2next_8_bytes)
+
+	test	$0x01, %cl
+	jnz	L(Byte0)
+
+	test	$0x02, %cl
+	jnz	L(Byte1)
+
+	test	$0x04, %cl
+	jnz	L(Byte2)
+
+	test	$0x08, %cl
+	jnz	L(Byte3)
+
+	test	$0x10, %cl
+	jnz	L(Byte4)
+
+	test	$0x20, %cl
+	jnz	L(Byte5)
+
+	test	$0x40, %cl
+	jnz	L(Byte6)
+#ifdef USE_AS_STRNCMP
+	cmp	$7, %ebp
+	jbe	L(eq)
+#endif
+
+	movzx	7(%eax), %ecx
+	movzx	7(%edx), %eax
+
+	sub	%ecx, %eax
+	RETURN
+
+	.p2align 4
+L(Byte0):
+#ifdef USE_AS_STRNCMP
+	cmp	$0, %ebp
+	jbe	L(eq)
+#endif
+	movzx	(%eax), %ecx
+	movzx	(%edx), %eax
+
+	sub	%ecx, %eax
+	RETURN
+
+	.p2align 4
+L(Byte1):
+#ifdef USE_AS_STRNCMP
+	cmp	$1, %ebp
+	jbe	L(eq)
+#endif
+	movzx	1(%eax), %ecx
+	movzx	1(%edx), %eax
+
+	sub	%ecx, %eax
+	RETURN
+
+	.p2align 4
+L(Byte2):
+#ifdef USE_AS_STRNCMP
+	cmp	$2, %ebp
+	jbe	L(eq)
+#endif
+	movzx	2(%eax), %ecx
+	movzx	2(%edx), %eax
+
+	sub	%ecx, %eax
+	RETURN
+
+	.p2align 4
+L(Byte3):
+#ifdef USE_AS_STRNCMP
+	cmp	$3, %ebp
+	jbe	L(eq)
+#endif
+	movzx	3(%eax), %ecx
+	movzx	3(%edx), %eax
+
+	sub	%ecx, %eax
+	RETURN
+
+	.p2align 4
+L(Byte4):
+#ifdef USE_AS_STRNCMP
+	cmp	$4, %ebp
+	jbe	L(eq)
+#endif
+	movzx	4(%eax), %ecx
+	movzx	4(%edx), %eax
+
+	sub	%ecx, %eax
+	RETURN
+
+	.p2align 4
+L(Byte5):
+#ifdef USE_AS_STRNCMP
+	cmp	$5, %ebp
+	jbe	L(eq)
+#endif
+	movzx	5(%eax), %ecx
+	movzx	5(%edx), %eax
+
+	sub	%ecx, %eax
+	RETURN
+
+	.p2align 4
+L(Byte6):
+#ifdef USE_AS_STRNCMP
+	cmp	$6, %ebp
+	jbe	L(eq)
+#endif
+	movzx	6(%eax), %ecx
+	movzx	6(%edx), %eax
+
+	sub	%ecx, %eax
+	RETURN
+
+	.p2align 4
+L(2next_8_bytes):
+	add	$8, %eax
+	add	$8, %edx
+#ifdef USE_AS_STRNCMP
+	cmp	$8, %ebp
+	lea	-8(%ebp), %ebp
+	jbe	L(eq)
+#endif
+
+	test	$0x01, %ch
+	jnz	L(Byte0)
+
+	test	$0x02, %ch
+	jnz	L(Byte1)
+
+	test	$0x04, %ch
+	jnz	L(Byte2)
+
+	test	$0x08, %ch
+	jnz	L(Byte3)
+
+	test	$0x10, %ch
+	jnz	L(Byte4)
+
+	test	$0x20, %ch
+	jnz	L(Byte5)
+
+	test	$0x40, %ch
+	jnz	L(Byte6)
+
+#ifdef USE_AS_STRNCMP
+	cmp	$7, %ebp
+	jbe	L(eq)
+#endif
+	movzx	7(%eax), %ecx
+	movzx	7(%edx), %eax
+
+	sub	%ecx, %eax
+	RETURN
+
+	.p2align 4
+L(neq):
+	mov	$1, %eax
+	ja	L(neq_bigger)
+	neg	%eax
+L(neq_bigger):
+	RETURN
+
+#ifdef USE_AS_STRNCMP
+	CFI_PUSH (%ebx)
+	CFI_PUSH (%edi)
+	CFI_PUSH (%esi)
+
+	.p2align 4
+L(more8byteseq):
+	POP	(%esi)
+	POP	(%edi)
+	POP	(%ebx)
+#endif
+
+L(eq):
+
+#ifdef USE_AS_STRNCMP
+	POP	(%ebp)
+#endif
+	xorl	%eax, %eax
+	ret
+
+#ifdef USE_AS_STRNCMP
+	CFI_PUSH (%ebp)
+
+	.p2align 4
+L(less16bytes_sncmp):
+	test	%ebp, %ebp
+	jz	L(eq)
+
+	movzbl	(%eax), %ecx
+	cmpb	%cl, (%edx)
+	jne	L(neq)
+	test	%cl, %cl
+	je	L(eq)
+
+	cmp	$1, %ebp
+	je	L(eq)
+
+	movzbl	1(%eax), %ecx
+	cmpb	%cl, 1(%edx)
+	jne	L(neq)
+	test	%cl, %cl
+	je	L(eq)
+
+	cmp	$2, %ebp
+	je	L(eq)
+
+	movzbl	2(%eax), %ecx
+	cmpb	%cl, 2(%edx)
+	jne	L(neq)
+	test	%cl, %cl
+	je	L(eq)
+
+	cmp	$3, %ebp
+	je	L(eq)
+
+	movzbl	3(%eax), %ecx
+	cmpb	%cl, 3(%edx)
+	jne	L(neq)
+	test	%cl, %cl
+	je	L(eq)
+
+	cmp	$4, %ebp
+	je	L(eq)
+
+	movzbl	4(%eax), %ecx
+	cmpb	%cl, 4(%edx)
+	jne	L(neq)
+	test	%cl, %cl
+	je	L(eq)
+
+	cmp	$5, %ebp
+	je	L(eq)
+
+	movzbl	5(%eax), %ecx
+	cmpb	%cl, 5(%edx)
+	jne	L(neq)
+	test	%cl, %cl
+	je	L(eq)
+
+	cmp	$6, %ebp
+	je	L(eq)
+
+	movzbl	6(%eax), %ecx
+	cmpb	%cl, 6(%edx)
+	jne	L(neq)
+	test	%cl, %cl
+	je	L(eq)
+
+	cmp	$7, %ebp
+	je	L(eq)
+
+	movzbl	7(%eax), %ecx
+	cmpb	%cl, 7(%edx)
+	jne	L(neq)
+	test	%cl, %cl
+	je	L(eq)
+
+
+	cmp	$8, %ebp
+	je	L(eq)
+
+	movzbl	8(%eax), %ecx
+	cmpb	%cl, 8(%edx)
+	jne	L(neq)
+	test	%cl, %cl
+	je	L(eq)
+
+	cmp	$9, %ebp
+	je	L(eq)
+
+	movzbl	9(%eax), %ecx
+	cmpb	%cl, 9(%edx)
+	jne	L(neq)
+	test	%cl, %cl
+	je	L(eq)
+
+	cmp	$10, %ebp
+	je	L(eq)
+
+	movzbl	10(%eax), %ecx
+	cmpb	%cl, 10(%edx)
+	jne	L(neq)
+	test	%cl, %cl
+	je	L(eq)
+
+	cmp	$11, %ebp
+	je	L(eq)
+
+	movzbl	11(%eax), %ecx
+	cmpb	%cl, 11(%edx)
+	jne	L(neq)
+	test	%cl, %cl
+	je	L(eq)
+
+
+	cmp	$12, %ebp
+	je	L(eq)
+
+	movzbl	12(%eax), %ecx
+	cmpb	%cl, 12(%edx)
+	jne	L(neq)
+	test	%cl, %cl
+	je	L(eq)
+
+	cmp	$13, %ebp
+	je	L(eq)
+
+	movzbl	13(%eax), %ecx
+	cmpb	%cl, 13(%edx)
+	jne	L(neq)
+	test	%cl, %cl
+	je	L(eq)
+
+	cmp	$14, %ebp
+	je	L(eq)
+
+	movzbl	14(%eax), %ecx
+	cmpb	%cl, 14(%edx)
+	jne	L(neq)
+	test	%cl, %cl
+	je	L(eq)
+
+	cmp	$15, %ebp
+	je	L(eq)
+
+	movzbl	15(%eax), %ecx
+	cmpb	%cl, 15(%edx)
+	jne	L(neq)
+	test	%cl, %cl
+	je	L(eq)
+
+	POP	(%ebp)
+	xor	%eax, %eax
+	ret
+#endif
+
+END (ssse3_strcmp_latest)
diff --git a/libc/arch-x86/string/strcmp_wrapper.S b/libc/arch-x86/string/strcmp_wrapper.S
new file mode 100644
index 0000000..69b7f0b
--- /dev/null
+++ b/libc/arch-x86/string/strcmp_wrapper.S
@@ -0,0 +1,40 @@
+/*
+Copyright (c) 2010, Intel Corporation
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+    * Redistributions of source code must retain the above copyright notice,
+    * this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright notice,
+    * this list of conditions and the following disclaimer in the documentation
+    * and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its contributors
+    * may be used to endorse or promote products derived from this software
+    * without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#if defined(USE_SSSE3)
+
+# define ssse3_strcmp_latest strcmp
+# include "ssse3-strcmp.S"
+
+#else
+
+# include "strcmp.S"
+
+#endif
diff --git a/libc/arch-x86/string/strncmp_wrapper.S b/libc/arch-x86/string/strncmp_wrapper.S
new file mode 100644
index 0000000..2050184
--- /dev/null
+++ b/libc/arch-x86/string/strncmp_wrapper.S
@@ -0,0 +1,42 @@
+/*
+Copyright (c) 2010, Intel Corporation
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+    * Redistributions of source code must retain the above copyright notice,
+    * this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright notice,
+    * this list of conditions and the following disclaimer in the documentation
+    * and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its contributors
+    * may be used to endorse or promote products derived from this software
+    * without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#if defined(USE_SSSE3)
+
+# define USE_AS_STRNCMP
+# define ssse3_strcmp_latest strncmp
+# include "ssse3-strcmp.S"
+
+#else
+
+# include "strncmp.S"
+
+#endif
+