[RFC] Imporve 64bit memcpy performance for Haswell CPU with AVX instruction

Message ID 1397480104-5934-1-git-send-email-ling.ma@alibaba-inc.com
State Superseded
Headers

Commit Message

ling.ma.program@gmail.com April 14, 2014, 12:55 p.m. UTC
  From: Ling Ma <ling.ml@alibaba-inc.com>


In this patch we take advantage of HSW memory bandwidth, manage to
reduce miss branch prediction by avoiding using branch instructions and
force destination to be aligned with avx instruction. 

According to Ondra suggestion, the version shrinks code size & improve
destination througput,then re-test all cases compared with sse2 memcpy.
The CPU2006 403.gcc benchmark indicates this patch improves performance
from 6% to 14%. 

---

 ChangeLog                                        |  16 +
 sysdeps/x86_64/multiarch/Makefile                |   1 +
 sysdeps/x86_64/multiarch/ifunc-impl-list.c       |  12 +
 sysdeps/x86_64/multiarch/memcpy-avx-unaligned.S  | 390 +++++++++++++++++++++++
 sysdeps/x86_64/multiarch/memcpy.S                |   4 +
 sysdeps/x86_64/multiarch/memcpy_chk.S            |   3 +
 sysdeps/x86_64/multiarch/memmove-avx-unaligned.S |  22 ++
 sysdeps/x86_64/multiarch/memmove.c               |   7 +-
 sysdeps/x86_64/multiarch/memmove_chk.c           |   6 +-
 sysdeps/x86_64/multiarch/mempcpy-avx-unaligned.S |  22 ++
 sysdeps/x86_64/multiarch/mempcpy.S               |   3 +
 sysdeps/x86_64/multiarch/mempcpy_chk.S           |   3 +
 12 files changed, 485 insertions(+), 4 deletions(-)
 create mode 100644 sysdeps/x86_64/multiarch/memcpy-avx-unaligned.S
 create mode 100644 sysdeps/x86_64/multiarch/memmove-avx-unaligned.S
 create mode 100644 sysdeps/x86_64/multiarch/mempcpy-avx-unaligned.S
  

Patch

diff --git a/ChangeLog b/ChangeLog
index 9e5c3bb..5fad552 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,5 +1,21 @@ 
 2014-04-04  Ling Ma  <ling.ml@alibaba-inc.com>
 
+	* sysdeps/x86_64/multiarch/Makefile: Add avx memcpy/mempcpy/memmove
+	* sysdeps/x86_64/multiarch/ifunc-impl-list.c: Add support for related
+	flies with avx memcpy
+	* sysdeps/x86_64/multiarch/memcpy.S: Add support for avx memcpy
+	* sysdeps/x86_64/multiarch/memcpy_chk.S: Add support for avx memcpy_chk
+	* sysdeps/x86_64/multiarch/memmove.c: Add support for avx memmove
+	* sysdeps/x86_64/multiarch/memmove_chk.c: Add support for avx memmove_chk
+	* sysdeps/x86_64/multiarch/mempcpy.S: Add support for avx mempcpy
+	* sysdeps/x86_64/multiarch/mempcpy_chk.S: Add support for avx mempcpy_chk
+	* sysdeps/x86_64/multiarch/memcpy-avx-unaligned.S: New file for avx memcpy
+	* sysdeps/x86_64/multiarch/mempcpy-avx-unaligned.S: New file for avx mempcpy
+	* sysdeps/x86_64/multiarch/memmove-avx-unaligned.S: New file for avx
+	memmove
+
+2014-04-04  Ling Ma  <ling.ml@alibaba-inc.com>
+
 	* sysdeps/x86_64/multiarch/Makefile: Add memset-avx2
 	* sysdeps/x86_64/multiarch/memset-avx2.S: New file for AVX2 memset
 	* sysdeps/x86_64/multiarch/memset.S: New file for multiple memset
diff --git a/sysdeps/x86_64/multiarch/Makefile b/sysdeps/x86_64/multiarch/Makefile
index 42df96f..5902098 100644
--- a/sysdeps/x86_64/multiarch/Makefile
+++ b/sysdeps/x86_64/multiarch/Makefile
@@ -11,6 +11,7 @@  sysdep_routines += strncat-c stpncpy-c strncpy-c strcmp-ssse3 \
 		   memcmp-sse4 memcpy-ssse3 \
 		   memcpy-sse2-unaligned mempcpy-ssse3 \
 		   memmove-ssse3 memcpy-ssse3-back mempcpy-ssse3-back \
+		   memmove-avx-unaligned memcpy-avx-unaligned mempcpy-avx-unaligned \
 		   memmove-ssse3-back strcasecmp_l-ssse3 \
 		   strncase_l-ssse3 strcat-ssse3 strncat-ssse3\
 		   strcpy-ssse3 strncpy-ssse3 stpcpy-ssse3 stpncpy-ssse3 \
diff --git a/sysdeps/x86_64/multiarch/ifunc-impl-list.c b/sysdeps/x86_64/multiarch/ifunc-impl-list.c
index 6da9be1..34db79e 100644
--- a/sysdeps/x86_64/multiarch/ifunc-impl-list.c
+++ b/sysdeps/x86_64/multiarch/ifunc-impl-list.c
@@ -46,6 +46,8 @@  __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
 
   /* Support sysdeps/x86_64/multiarch/memmove_chk.S.  */
   IFUNC_IMPL (i, name, __memmove_chk,
+	      IFUNC_IMPL_ADD (array, i, __memmove_chk, HAS_AVX,
+			      __memmove_chk_avx_unaligned)
 	      IFUNC_IMPL_ADD (array, i, __memmove_chk, HAS_SSSE3,
 			      __memmove_chk_ssse3_back)
 	      IFUNC_IMPL_ADD (array, i, __memmove_chk, HAS_SSSE3,
@@ -55,6 +57,8 @@  __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
 
   /* Support sysdeps/x86_64/multiarch/memmove.S.  */
   IFUNC_IMPL (i, name, memmove,
+	      IFUNC_IMPL_ADD (array, i, memmove, HAS_AVX,
+			      __memmove_avx_unaligned)
 	      IFUNC_IMPL_ADD (array, i, memmove, HAS_SSSE3,
 			      __memmove_ssse3_back)
 	      IFUNC_IMPL_ADD (array, i, memmove, HAS_SSSE3,
@@ -201,6 +205,8 @@  __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
 #ifdef SHARED
   /* Support sysdeps/x86_64/multiarch/memcpy_chk.S.  */
   IFUNC_IMPL (i, name, __memcpy_chk,
+	      IFUNC_IMPL_ADD (array, i, __memcpy_chk, HAS_AVX,
+			      __memcpy_chk_avx_unaligned)
 	      IFUNC_IMPL_ADD (array, i, __memcpy_chk, HAS_SSSE3,
 			      __memcpy_chk_ssse3_back)
 	      IFUNC_IMPL_ADD (array, i, __memcpy_chk, HAS_SSSE3,
@@ -210,6 +216,8 @@  __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
 
   /* Support sysdeps/x86_64/multiarch/memcpy.S.  */
   IFUNC_IMPL (i, name, memcpy,
+	      IFUNC_IMPL_ADD (array, i, memcpy, HAS_AVX,
+			      __memcpy_avx_unaligned)
 	      IFUNC_IMPL_ADD (array, i, memcpy, HAS_SSSE3,
 			      __memcpy_ssse3_back)
 	      IFUNC_IMPL_ADD (array, i, memcpy, HAS_SSSE3, __memcpy_ssse3)
@@ -218,6 +226,8 @@  __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
 
   /* Support sysdeps/x86_64/multiarch/mempcpy_chk.S.  */
   IFUNC_IMPL (i, name, __mempcpy_chk,
+	      IFUNC_IMPL_ADD (array, i, __mempcpy_chk, HAS_AVX,
+			      __mempcpy_chk_avx_unaligned)
 	      IFUNC_IMPL_ADD (array, i, __mempcpy_chk, HAS_SSSE3,
 			      __mempcpy_chk_ssse3_back)
 	      IFUNC_IMPL_ADD (array, i, __mempcpy_chk, HAS_SSSE3,
@@ -227,6 +237,8 @@  __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
 
   /* Support sysdeps/x86_64/multiarch/mempcpy.S.  */
   IFUNC_IMPL (i, name, mempcpy,
+	      IFUNC_IMPL_ADD (array, i, mempcpy, HAS_AVX,
+			      __mempcpy_avx_unaligned)
 	      IFUNC_IMPL_ADD (array, i, mempcpy, HAS_SSSE3,
 			      __mempcpy_ssse3_back)
 	      IFUNC_IMPL_ADD (array, i, mempcpy, HAS_SSSE3,
diff --git a/sysdeps/x86_64/multiarch/memcpy-avx-unaligned.S b/sysdeps/x86_64/multiarch/memcpy-avx-unaligned.S
new file mode 100644
index 0000000..b4569b9
--- /dev/null
+++ b/sysdeps/x86_64/multiarch/memcpy-avx-unaligned.S
@@ -0,0 +1,390 @@ 
+/* memcpy with AVX
+   Copyright (C) 2014 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include <sysdep.h>
+
+#if !defined NOT_IN_libc \
+    && (defined SHARED \
+        || defined USE_AS_MEMMOVE \
+	|| !defined USE_MULTIARCH)
+
+#include "asm-syntax.h"
+#ifndef MEMCPY
+# define MEMCPY	__memcpy_avx_unaligned
+# define MEMCPY_CHK	__memcpy_chk_avx_unaligned
+#endif
+
+	.section .text.avx,"ax",@progbits
+#if !defined USE_AS_BCOPY
+ENTRY (MEMCPY_CHK)
+	cmpq	%rdx, %rcx
+	jb	HIDDEN_JUMPTARGET (__chk_fail)
+END (MEMCPY_CHK)
+#endif
+
+ENTRY (MEMCPY)
+	mov	%rdi, %rax
+#ifdef USE_AS_MEMPCPY
+	add	%rdx, %rax
+#endif
+	cmp	$256, %rdx
+	jae	L(256bytesormore)
+	cmp	$128, %dl
+	jb	L(less_128bytes)
+	vmovups (%rsi), %xmm0
+	lea	(%rsi, %rdx), %rcx
+	vmovups 0x10(%rsi), %xmm1
+	vmovups 0x20(%rsi), %xmm2
+	vmovups 0x30(%rsi), %xmm3
+	vmovups 0x40(%rsi), %xmm4
+	vmovups 0x50(%rsi), %xmm5
+	vmovups 0x60(%rsi), %xmm6
+	vmovups 0x70(%rsi), %xmm7
+	vmovups -0x80(%rcx), %xmm8
+	vmovups -0x70(%rcx), %xmm9
+	vmovups -0x60(%rcx), %xmm10
+	vmovups -0x50(%rcx), %xmm11
+	vmovups -0x40(%rcx), %xmm12
+	vmovups -0x30(%rcx), %xmm13
+	vmovups -0x20(%rcx), %xmm14
+	vmovups -0x10(%rcx), %xmm15
+	lea	(%rdi, %rdx), %rdx
+	vmovups %xmm0, (%rdi)
+	vmovups %xmm1, 0x10(%rdi)
+	vmovups %xmm2, 0x20(%rdi)
+	vmovups %xmm3, 0x30(%rdi)
+	vmovups %xmm4, 0x40(%rdi)
+	vmovups %xmm5, 0x50(%rdi)
+	vmovups %xmm6, 0x60(%rdi)
+	vmovups %xmm7, 0x70(%rdi)
+	vmovups %xmm8, -0x80(%rdx)
+	vmovups %xmm9, -0x70(%rdx)
+	vmovups %xmm10, -0x60(%rdx)
+	vmovups %xmm11, -0x50(%rdx)
+	vmovups %xmm12, -0x40(%rdx)
+	vmovups %xmm13, -0x30(%rdx)
+	vmovups %xmm14, -0x20(%rdx)
+	vmovups %xmm15, -0x10(%rdx)
+	ret
+	.p2align 4
+L(less_128bytes):
+	cmp	$64, %dl
+	jb	L(less_64bytes)
+	vmovups (%rsi), %xmm0
+	lea	(%rsi, %rdx), %rcx
+	vmovups 0x10(%rsi), %xmm1
+	vmovups 0x20(%rsi), %xmm2
+	lea	(%rdi, %rdx), %rdx
+	vmovups 0x30(%rsi), %xmm3
+	vmovups -0x40(%rcx), %xmm4
+	vmovups -0x30(%rcx), %xmm5
+	vmovups -0x20(%rcx), %xmm6
+	vmovups -0x10(%rcx), %xmm7
+	vmovups %xmm0, (%rdi)
+	vmovups %xmm1, 0x10(%rdi)
+	vmovups %xmm2, 0x20(%rdi)
+	vmovups %xmm3, 0x30(%rdi)
+	vmovups %xmm4, -0x40(%rdx)
+	vmovups %xmm5, -0x30(%rdx)
+	vmovups %xmm6, -0x20(%rdx)
+	vmovups %xmm7, -0x10(%rdx)
+	ret
+	.p2align 4
+L(less_64bytes):
+	cmp	$32, %dl
+	jb	L(less_32bytes)
+	vmovups (%rsi), %xmm0
+	vmovups 0x10(%rsi), %xmm1
+	vmovups -0x20(%rsi, %rdx), %xmm6
+	vmovups -0x10(%rsi, %rdx), %xmm7
+	vmovups %xmm0, (%rdi)
+	vmovups %xmm1, 0x10(%rdi)
+	vmovups %xmm6, -0x20(%rdi, %rdx)
+	vmovups %xmm7, -0x10(%rdi, %rdx)
+	ret
+	.p2align 4
+L(less_32bytes):
+	cmp	$16, %dl
+	jb	L(less_16bytes)
+	vmovups (%rsi), %xmm0
+	vmovups -0x10(%rsi, %rdx), %xmm7
+	vmovups %xmm0, (%rdi)
+	vmovups %xmm7, -0x10(%rdi, %rdx)
+	ret
+	.p2align 4
+L(less_16bytes):
+	cmp	$8, %dl
+	jb	L(less_8bytes)
+	movq -0x08(%rsi, %rdx),	%rcx
+	movq (%rsi),	%rsi
+	movq %rsi, (%rdi)
+	movq %rcx, -0x08(%rdi, %rdx)
+	ret
+	.p2align 4
+L(less_8bytes):
+	cmp	$4, %dl
+	jb	L(less_4bytes)
+	mov -0x04(%rsi, %rdx), %ecx
+	mov (%rsi),	%esi
+	mov %esi, (%rdi)
+	mov %ecx, -0x04(%rdi, %rdx)
+	ret
+L(less_4bytes):
+	cmp	$1, %dl
+	jbe	L(less_2bytes)
+	mov -0x02(%rsi, %rdx),	%cx
+	mov (%rsi),	%si
+	mov %si, (%rdi)
+	mov %cx, -0x02(%rdi, %rdx)
+	ret
+L(less_2bytes):
+	jb	L(less_0bytes)
+	mov	(%rsi), %cl
+	mov	%cl,	(%rdi)
+L(less_0bytes):
+	ret
+
+	.p2align 4
+L(256bytesormore):
+#ifdef USE_AS_MEMMOVE
+	cmp	%rsi, %rdi
+	jae	L(copy_backward)
+#endif
+	cmp	$2048, %rdx
+	jae	L(gobble_data_movsb)
+	mov	%rax, %r8
+	lea	(%rsi, %rdx), %rcx
+	mov	%rdi, %r10
+	vmovups -0x80(%rcx), %xmm5
+	vmovups -0x70(%rcx), %xmm6
+	mov	$0x80, %rax
+	and	$-32, %rdi
+	add	$32, %rdi
+	vmovups -0x60(%rcx), %xmm7
+	vmovups -0x50(%rcx), %xmm8
+	mov	%rdi, %r11
+	sub	%r10, %r11
+	vmovups -0x40(%rcx), %xmm9
+	vmovups -0x30(%rcx), %xmm10
+	sub	%r11, %rdx
+	vmovups -0x20(%rcx), %xmm11
+	vmovups -0x10(%rcx), %xmm12
+	vmovups	(%rsi), %ymm4
+	add	%r11, %rsi
+	sub	%eax, %edx
+L(goble_128_loop):
+	vmovups (%rsi), %ymm0
+	vmovups 0x20(%rsi), %ymm1
+	vmovups 0x40(%rsi), %ymm2
+	vmovups 0x60(%rsi), %ymm3
+	add	%rax, %rsi
+	vmovaps %ymm0, (%rdi)
+	vmovaps %ymm1, 0x20(%rdi)
+	vmovaps %ymm2, 0x40(%rdi)
+	vmovaps %ymm3, 0x60(%rdi)
+	add	%rax, %rdi
+	sub	%eax, %edx
+	jae	L(goble_128_loop)
+	add	%eax, %edx
+	add	%rdi, %rdx
+	vmovups	%ymm4, (%r10)
+	vzeroupper
+	vmovups %xmm5, -0x80(%rdx)
+	vmovups %xmm6, -0x70(%rdx)
+	vmovups %xmm7, -0x60(%rdx)
+	vmovups %xmm8, -0x50(%rdx)
+	vmovups %xmm9, -0x40(%rdx)
+	vmovups %xmm10, -0x30(%rdx)
+	vmovups %xmm11, -0x20(%rdx)
+	vmovups %xmm12, -0x10(%rdx)
+	mov	%r8, %rax
+	ret
+
+	.p2align 4
+L(gobble_data_movsb):
+
+#ifdef SHARED_CACHE_SIZE_HALF
+	mov	$SHARED_CACHE_SIZE_HALF, %rcx
+#else
+	mov	__x86_shared_cache_size_half(%rip), %rcx
+#endif
+	shl	$3, %rcx
+
+#ifdef USE_AS_MEMMOVE
+	mov	%rsi, %r10
+	sub	%rdi, %r10
+	cmp	%rdx, %r10
+	jae	L(memmove_use_memcpy_fwd)
+	cmp	%rcx, %r10
+	jae	L(memmove_use_memcpy_fwd)
+	jmp L(gobble_mem_fwd_llc_start)
+L(memmove_use_memcpy_fwd):
+#endif
+	cmp	%rcx, %rdx
+	jae	L(gobble_big_data_fwd)
+#ifdef USE_AS_MEMMOVE
+L(gobble_mem_fwd_llc_start):
+#endif
+	mov	%rdx, %rcx
+	rep	movsb
+	ret
+
+	.p2align 4
+L(gobble_big_data_fwd):
+	lea	(%rsi, %rdx), %rcx
+	vmovups	(%rsi), %ymm4
+	vmovups -0x80(%rsi,%rdx), %xmm5
+	vmovups -0x70(%rcx), %xmm6
+	vmovups -0x60(%rcx), %xmm7
+	vmovups -0x50(%rcx), %xmm8
+	vmovups -0x40(%rcx), %xmm9
+	vmovups -0x30(%rcx), %xmm10
+	vmovups -0x20(%rcx), %xmm11
+	vmovups -0x10(%rcx), %xmm12
+	mov	%rdi, %r8
+	and	$-32, %rdi
+	add	$32, %rdi
+	mov	%rdi, %r10
+	sub	%r8, %r10
+	sub	%r10, %rdx
+	add	%r10, %rsi
+	lea	(%rdi, %rdx), %rcx
+	sub	$0x80, %rdx
+L(gobble_mem_fwd_loop):
+	prefetchnta 0x1c0(%rsi)
+	prefetchnta 0x280(%rsi)
+	vmovups	(%rsi), %ymm0
+	vmovups	0x20(%rsi), %ymm1
+	vmovups	0x40(%rsi), %ymm2
+	vmovups	0x60(%rsi), %ymm3
+	lea	0x80(%rsi), %rsi
+	vmovntdq	%ymm0, (%rdi)
+	vmovntdq	%ymm1, 0x20(%rdi)
+	vmovntdq	%ymm2, 0x40(%rdi)
+	vmovntdq	%ymm3, 0x60(%rdi)
+	lea	0x80(%rdi), %rdi
+	sub	$0x80, %rdx
+	jae	L(gobble_mem_fwd_loop)
+	sfence
+	vmovups	%ymm4, (%r8)
+	vzeroupper
+	vmovups %xmm5, -0x80(%rcx)
+	vmovups %xmm6, -0x70(%rcx)
+	vmovups %xmm7, -0x60(%rcx)
+	vmovups %xmm8, -0x50(%rcx)
+	vmovups %xmm9, -0x40(%rcx)
+	vmovups %xmm10, -0x30(%rcx)
+	vmovups %xmm11, -0x20(%rcx)
+	vmovups %xmm12, -0x10(%rcx)
+	ret
+
+#ifdef USE_AS_MEMMOVE
+	.p2align 4
+L(copy_backward):
+#ifdef SHARED_CACHE_SIZE_HALF
+	mov	$SHARED_CACHE_SIZE_HALF, %rcx
+#else
+	mov	__x86_shared_cache_size_half(%rip), %rcx
+#endif
+	shl	$3, %rcx
+	mov	%rdi, %r9
+	vmovups (%rsi), %xmm8
+	vmovups 0x10(%rsi), %xmm9
+	add	%rdx, %rdi
+	vmovups 0x20(%rsi), %xmm10
+	vmovups 0x30(%rsi), %xmm11
+	lea	-0x20(%rdi), %r10
+	mov %rdi, %r11
+	vmovups 0x40(%rsi), %xmm12
+	vmovups 0x50(%rsi), %xmm13
+	and	$0x1f, %r11
+	vmovups 0x60(%rsi), %xmm14
+	vmovups 0x70(%rsi), %xmm15
+	xor	%r11, %rdi
+	add	%rdx, %rsi
+	vmovups	-0x20(%rsi), %ymm4
+	sub	%r11, %rsi
+	sub	%r11, %rdx
+	mov	%rdi, %r11
+	sub	%rsi, %r11
+	cmp	%rdx, %r11
+	jae	L(memmove_use_memcpy_bwd)
+	cmp	%rcx, %r11
+	jae	L(memmove_use_memcpy_bwd)
+	jmp L(gobble_mem_bwd_llc_start)
+L(memmove_use_memcpy_bwd):
+	cmp	%rcx, %rdx
+	ja	L(gobble_big_data_bwd)
+L(gobble_mem_bwd_llc_start):
+	sub	$0x80, %rdx
+L(gobble_mem_bwd_llc):
+	vmovups	-0x20(%rsi), %ymm0
+	vmovups	-0x40(%rsi), %ymm1
+	vmovups	-0x60(%rsi), %ymm2
+	vmovups	-0x80(%rsi), %ymm3
+	lea	-0x80(%rsi), %rsi
+	vmovaps	%ymm0, -0x20(%rdi)
+	vmovaps	%ymm1, -0x40(%rdi)
+	vmovaps	%ymm2, -0x60(%rdi)
+	vmovaps	%ymm3, -0x80(%rdi)
+	lea	-0x80(%rdi), %rdi
+	sub	$0x80, %rdx
+	jae	L(gobble_mem_bwd_llc)
+	vmovups	%ymm4, (%r10)
+	vzeroupper
+	vmovups %xmm8, (%r9)
+	vmovups %xmm9, 0x10(%r9)
+	vmovups %xmm10, 0x20(%r9)
+	vmovups %xmm11, 0x30(%r9)
+	vmovups %xmm12, 0x40(%r9)
+	vmovups %xmm13, 0x50(%r9)
+	vmovups %xmm14, 0x60(%r9)
+	vmovups %xmm15, 0x70(%r9)
+	ret
+L(gobble_big_data_bwd):
+	sub	$0x80, %rdx
+L(gobble_mem_bwd_loop):
+	prefetchnta -0x1c0(%rsi)
+	prefetchnta -0x280(%rsi)
+	vmovups	-0x20(%rsi), %ymm0
+	vmovups	-0x40(%rsi), %ymm1
+	vmovups	-0x60(%rsi), %ymm2
+	vmovups	-0x80(%rsi), %ymm3
+	lea	-0x80(%rsi), %rsi
+	vmovntdq	%ymm0, -0x20(%rdi)
+	vmovntdq	%ymm1, -0x40(%rdi)
+	vmovntdq	%ymm2, -0x60(%rdi)
+	vmovntdq	%ymm3, -0x80(%rdi)
+	lea	-0x80(%rdi), %rdi
+	sub	$0x80, %rdx
+	jae	L(gobble_mem_bwd_loop)
+	sfence
+	vmovups	%ymm4, (%r10)
+	vzeroupper
+	vmovups %xmm8, (%r9)
+	vmovups %xmm9, 0x10(%r9)
+	vmovups %xmm10, 0x20(%r9)
+	vmovups %xmm11, 0x30(%r9)
+	vmovups %xmm12, 0x40(%r9)
+	vmovups %xmm13, 0x50(%r9)
+	vmovups %xmm14, 0x60(%r9)
+	vmovups %xmm15, 0x70(%r9)
+	ret
+#endif
+END (MEMCPY)
+#endif
diff --git a/sysdeps/x86_64/multiarch/memcpy.S b/sysdeps/x86_64/multiarch/memcpy.S
index 40ae926..b92f986 100644
--- a/sysdeps/x86_64/multiarch/memcpy.S
+++ b/sysdeps/x86_64/multiarch/memcpy.S
@@ -32,6 +32,10 @@  ENTRY(__new_memcpy)
 	cmpl	$0, KIND_OFFSET+__cpu_features(%rip)
 	jne	1f
 	call	__init_cpu_features
+1:	leaq	__memcpy_avx_unaligned(%rip), %rax
+	testl	$bit_AVX_Usable, __cpu_features+FEATURE_OFFSET+index_AVX_Usable(%rip)	
+	jz 1f
+	ret
 1:	leaq	__memcpy_sse2(%rip), %rax
 	testl	$bit_Slow_BSF, __cpu_features+FEATURE_OFFSET+index_Slow_BSF(%rip)
 	jnz	2f
diff --git a/sysdeps/x86_64/multiarch/memcpy_chk.S b/sysdeps/x86_64/multiarch/memcpy_chk.S
index 3c0270f..076b19a 100644
--- a/sysdeps/x86_64/multiarch/memcpy_chk.S
+++ b/sysdeps/x86_64/multiarch/memcpy_chk.S
@@ -39,6 +39,9 @@  ENTRY(__memcpy_chk)
 	testl	$bit_Fast_Copy_Backward, __cpu_features+FEATURE_OFFSET+index_Fast_Copy_Backward(%rip)
 	jz	2f
 	leaq	__memcpy_chk_ssse3_back(%rip), %rax
+	testl   $bit_AVX_Usable, __cpu_features+FEATURE_OFFSET+index_AVX_Usable(%rip)
+	jz  2f
+	leaq    __memcpy_chk_avx_unaligned(%rip), %rax
 2:	ret
 END(__memcpy_chk)
 # else
diff --git a/sysdeps/x86_64/multiarch/memmove-avx-unaligned.S b/sysdeps/x86_64/multiarch/memmove-avx-unaligned.S
new file mode 100644
index 0000000..faed9fb
--- /dev/null
+++ b/sysdeps/x86_64/multiarch/memmove-avx-unaligned.S
@@ -0,0 +1,22 @@ 
+/* memmove with AVX
+   Copyright (C) 2014 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#define USE_AS_MEMMOVE
+#define MEMCPY		__memmove_avx_unaligned
+#define MEMCPY_CHK	__memmove_chk_avx_unaligned
+#include "memcpy-avx-unaligned.S"
diff --git a/sysdeps/x86_64/multiarch/memmove.c b/sysdeps/x86_64/multiarch/memmove.c
index ba86e7b..ee73f23 100644
--- a/sysdeps/x86_64/multiarch/memmove.c
+++ b/sysdeps/x86_64/multiarch/memmove.c
@@ -35,6 +35,8 @@ 
 extern __typeof (__redirect_memmove) __memmove_sse2 attribute_hidden;
 extern __typeof (__redirect_memmove) __memmove_ssse3 attribute_hidden;
 extern __typeof (__redirect_memmove) __memmove_ssse3_back attribute_hidden;
+extern __typeof (__redirect_memmove) __memmove_avx_unaligned attribute_hidden;
+
 #endif
 
 #include "string/memmove.c"
@@ -47,10 +49,11 @@  extern __typeof (__redirect_memmove) __memmove_ssse3_back attribute_hidden;
    ifunc symbol properly.  */
 extern __typeof (__redirect_memmove) __libc_memmove;
 libc_ifunc (__libc_memmove,
-	    HAS_SSSE3
+	    HAS_AVX ? __memmove_avx_unaligned : 
+	    (HAS_SSSE3
 	    ? (HAS_FAST_COPY_BACKWARD
 	       ? __memmove_ssse3_back : __memmove_ssse3)
-	    : __memmove_sse2)
+	    : __memmove_sse2));
 
 strong_alias (__libc_memmove, memmove)
 
diff --git a/sysdeps/x86_64/multiarch/memmove_chk.c b/sysdeps/x86_64/multiarch/memmove_chk.c
index cb1acb6..44344f2 100644
--- a/sysdeps/x86_64/multiarch/memmove_chk.c
+++ b/sysdeps/x86_64/multiarch/memmove_chk.c
@@ -25,11 +25,13 @@ 
 extern __typeof (__memmove_chk) __memmove_chk_sse2 attribute_hidden;
 extern __typeof (__memmove_chk) __memmove_chk_ssse3 attribute_hidden;
 extern __typeof (__memmove_chk) __memmove_chk_ssse3_back attribute_hidden;
+extern __typeof (__memmove_chk) __memmove_chk_avx_unaligned attribute_hidden;
 
 #include "debug/memmove_chk.c"
 
 libc_ifunc (__memmove_chk,
-	    HAS_SSSE3
+	    HAS_AVX ? __memmove_chk_avx_unaligned :
+	    (HAS_SSSE3
 	    ? (HAS_FAST_COPY_BACKWARD
 	       ? __memmove_chk_ssse3_back : __memmove_chk_ssse3)
-	    : __memmove_chk_sse2);
+	    : __memmove_chk_sse2));
diff --git a/sysdeps/x86_64/multiarch/mempcpy-avx-unaligned.S b/sysdeps/x86_64/multiarch/mempcpy-avx-unaligned.S
new file mode 100644
index 0000000..438bda3
--- /dev/null
+++ b/sysdeps/x86_64/multiarch/mempcpy-avx-unaligned.S
@@ -0,0 +1,22 @@ 
+/* mempcpy with AVX
+   Copyright (C) 2014 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#define USE_AS_MEMPCPY
+#define MEMCPY		__mempcpy_avx_unaligned
+#define MEMCPY_CHK	__mempcpy_chk_avx_unaligned
+#include "memcpy-avx-unaligned.S"
diff --git a/sysdeps/x86_64/multiarch/mempcpy.S b/sysdeps/x86_64/multiarch/mempcpy.S
index b9f04c2..7589d8c 100644
--- a/sysdeps/x86_64/multiarch/mempcpy.S
+++ b/sysdeps/x86_64/multiarch/mempcpy.S
@@ -37,6 +37,9 @@  ENTRY(__mempcpy)
 	testl	$bit_Fast_Copy_Backward, __cpu_features+FEATURE_OFFSET+index_Fast_Copy_Backward(%rip)
 	jz	2f
 	leaq	__mempcpy_ssse3_back(%rip), %rax
+	testl	$bit_AVX_Usable, __cpu_features+FEATURE_OFFSET+index_AVX_Usable(%rip)
+	jz	2f
+	leaq	__mempcpy_avx_unaligned(%rip), %rax
 2:	ret
 END(__mempcpy)
 
diff --git a/sysdeps/x86_64/multiarch/mempcpy_chk.S b/sysdeps/x86_64/multiarch/mempcpy_chk.S
index c28473a..88e0b74 100644
--- a/sysdeps/x86_64/multiarch/mempcpy_chk.S
+++ b/sysdeps/x86_64/multiarch/mempcpy_chk.S
@@ -39,6 +39,9 @@  ENTRY(__mempcpy_chk)
 	testl	$bit_Fast_Copy_Backward, __cpu_features+FEATURE_OFFSET+index_Fast_Copy_Backward(%rip)
 	jz	2f
 	leaq	__mempcpy_chk_ssse3_back(%rip), %rax
+	testl	$bit_AVX_Usable, __cpu_features+FEATURE_OFFSET+index_AVX_Usable(%rip)
+	jz	2f
+	leaq	__mempcpy_chk_avx_unaligned(%rip), %rax
 2:	ret
 END(__mempcpy_chk)
 # else