Added memcpy/memmove family optimized with AVX512 for KNL hardware.

Added AVX512 implementations of memcpy, mempcpy, memmove, memcpy_chk,
mempcpy_chk, memmove_chk.
It shows average improvement more than 30% over AVX versions on KNL
hardware (performance results attached in the thread
<https://sourceware.org/ml/libc-alpha/2016-01/msg00258.html>).

    * sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Added new files.
    * sysdeps/x86_64/multiarch/ifunc-impl-list.c: Added new tests.
    * sysdeps/x86_64/multiarch/memcpy-avx512-no-vzeroupper.S: New file.
    * sysdeps/x86_64/multiarch/mempcpy-avx512-no-vzeroupper.S: Likewise.
    * sysdeps/x86_64/multiarch/memmove-avx512-no-vzeroupper.S: Likewise.
    * sysdeps/x86_64/multiarch/memcpy.S: Added new IFUNC branch.
    * sysdeps/x86_64/multiarch/memcpy_chk.S: Likewise.
    * sysdeps/x86_64/multiarch/memmove.c: Likewise.
    * sysdeps/x86_64/multiarch/memmove_chk.c: Likewise.
    * sysdeps/x86_64/multiarch/mempcpy.S: Likewise.
    * sysdeps/x86_64/multiarch/mempcpy_chk.S: Likewise.

1 files changed, 410 insertions, 0 deletions

diff --git a/sysdeps/x86_64/multiarch/memcpy-avx512-no-vzeroupper.S b/sysdeps/x86_64/multiarch/memcpy-avx512-no-vzeroupper.S
new file mode 100644
index 0000000000..cc02934a82
--- /dev/null
+++ b/sysdeps/x86_64/multiarch/memcpy-avx512-no-vzeroupper.S
@@ -0,0 +1,410 @@
+/* memcpy optimized with AVX512 for KNL hardware.

+   Copyright (C) 2016 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 HAVE_AVX512_ASM_SUPPORT && IS_IN (libc) \

+    && (defined SHARED \

+	|| defined USE_AS_MEMMOVE \

+	|| !defined USE_MULTIARCH)

+

+#include "asm-syntax.h"

+#ifndef MEMCPY

+# define MEMCPY		__memcpy_avx512_no_vzeroupper

+# define MEMCPY_CHK	__memcpy_chk_avx512_no_vzeroupper

+#endif

+

+	.section .text,"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

+	lea	(%rsi, %rdx), %rcx

+	lea	(%rdi, %rdx), %r9

+	cmp	$512, %rdx

+	ja	L(512bytesormore)

+

+L(check):

+	cmp	$16, %rdx

+	jbe	L(less_16bytes)

+	cmp	$256, %rdx

+	jb	L(less_256bytes)

+	vmovups	(%rsi), %zmm0

+	vmovups 0x40(%rsi), %zmm1

+	vmovups 0x80(%rsi), %zmm2

+	vmovups 0xC0(%rsi), %zmm3

+	vmovups	-0x100(%rcx), %zmm4

+	vmovups -0xC0(%rcx), %zmm5

+	vmovups -0x80(%rcx), %zmm6

+	vmovups -0x40(%rcx), %zmm7

+	vmovups %zmm0, (%rdi)

+	vmovups %zmm1, 0x40(%rdi)

+	vmovups %zmm2, 0x80(%rdi)

+	vmovups %zmm3, 0xC0(%rdi)

+	vmovups	%zmm4, -0x100(%r9)

+	vmovups %zmm5, -0xC0(%r9)

+	vmovups %zmm6, -0x80(%r9)

+	vmovups %zmm7, -0x40(%r9)

+	ret

+

+L(less_256bytes):

+	cmp	$128, %dl

+	jb	L(less_128bytes)

+	vmovups	(%rsi), %zmm0

+	vmovups 0x40(%rsi), %zmm1

+	vmovups -0x80(%rcx), %zmm2

+	vmovups -0x40(%rcx), %zmm3

+	vmovups	%zmm0, (%rdi)

+	vmovups %zmm1, 0x40(%rdi)

+	vmovups %zmm2, -0x80(%r9)

+	vmovups %zmm3, -0x40(%r9)

+	ret

+

+L(less_128bytes):

+	cmp	$64, %dl

+	jb	L(less_64bytes)

+	vmovdqu (%rsi), %ymm0

+	vmovdqu 0x20(%rsi), %ymm1

+	vmovdqu -0x40(%rcx), %ymm2

+	vmovdqu -0x20(%rcx), %ymm3

+	vmovdqu %ymm0, (%rdi)

+	vmovdqu %ymm1, 0x20(%rdi)

+	vmovdqu %ymm2, -0x40(%r9)

+	vmovdqu %ymm3, -0x20(%r9)

+	ret

+

+L(less_64bytes):

+	cmp	$32, %dl

+	jb	L(less_32bytes)

+	vmovdqu	(%rsi), %ymm0

+	vmovdqu -0x20(%rcx), %ymm1

+	vmovdqu	%ymm0, (%rdi)

+	vmovdqu	%ymm1, -0x20(%r9)

+	ret

+

+L(less_32bytes):

+	vmovdqu (%rsi), %xmm0

+	vmovdqu -0x10(%rcx), %xmm1

+	vmovdqu %xmm0, (%rdi)

+	vmovdqu %xmm1, -0x10(%r9)

+	ret

+

+L(less_16bytes):

+	cmp	$8, %dl

+	jb	L(less_8bytes)

+	movq	(%rsi), %rsi

+	movq	-0x8(%rcx), %rcx

+	movq	%rsi, (%rdi)

+	movq	%rcx, -0x8(%r9)

+	ret

+

+L(less_8bytes):

+	cmp	$4, %dl

+	jb	L(less_4bytes)

+	mov	(%rsi), %esi

+	mov	-0x4(%rcx), %ecx

+	mov	%esi, (%rdi)

+	mov	%ecx, -0x4(%r9)

+	ret

+

+L(less_4bytes):

+	cmp	$2, %dl

+	jb	L(less_2bytes)

+	mov	(%rsi), %si

+	mov	-0x2(%rcx), %cx

+	mov	%si, (%rdi)

+	mov	%cx, -0x2(%r9)

+	ret

+

+L(less_2bytes):

+	cmp	$1, %dl

+	jb	L(less_1bytes)	

+	mov	(%rsi), %cl

+	mov	%cl, (%rdi)

+L(less_1bytes):

+	ret

+

+L(512bytesormore):

+#ifdef SHARED_CACHE_SIZE_HALF

+	mov	$SHARED_CACHE_SIZE_HALF, %r8

+#else

+	mov	__x86_shared_cache_size_half(%rip), %r8

+#endif

+	cmp	%r8, %rdx

+	jae	L(preloop_large)

+	cmp	$1024, %rdx

+	ja	L(1024bytesormore)

+	prefetcht1 (%rsi)

+	prefetcht1 0x40(%rsi)

+	prefetcht1 0x80(%rsi)

+	prefetcht1 0xC0(%rsi)

+	prefetcht1 0x100(%rsi)

+	prefetcht1 0x140(%rsi)

+	prefetcht1 0x180(%rsi)

+	prefetcht1 0x1C0(%rsi)

+	prefetcht1 -0x200(%rcx)

+	prefetcht1 -0x1C0(%rcx)

+	prefetcht1 -0x180(%rcx)

+	prefetcht1 -0x140(%rcx)

+	prefetcht1 -0x100(%rcx)

+	prefetcht1 -0xC0(%rcx)

+	prefetcht1 -0x80(%rcx)

+	prefetcht1 -0x40(%rcx)	

+	vmovups	(%rsi), %zmm0

+	vmovups 0x40(%rsi), %zmm1

+	vmovups 0x80(%rsi), %zmm2

+	vmovups 0xC0(%rsi), %zmm3

+	vmovups	0x100(%rsi), %zmm4

+	vmovups 0x140(%rsi), %zmm5

+	vmovups 0x180(%rsi), %zmm6

+	vmovups 0x1C0(%rsi), %zmm7

+	vmovups	-0x200(%rcx), %zmm8

+	vmovups -0x1C0(%rcx), %zmm9

+	vmovups -0x180(%rcx), %zmm10

+	vmovups -0x140(%rcx), %zmm11

+	vmovups	-0x100(%rcx), %zmm12

+	vmovups -0xC0(%rcx), %zmm13

+	vmovups -0x80(%rcx), %zmm14

+	vmovups -0x40(%rcx), %zmm15

+	vmovups %zmm0, (%rdi)

+	vmovups %zmm1, 0x40(%rdi)

+	vmovups %zmm2, 0x80(%rdi)

+	vmovups %zmm3, 0xC0(%rdi)

+	vmovups %zmm4, 0x100(%rdi)

+	vmovups %zmm5, 0x140(%rdi)

+	vmovups %zmm6, 0x180(%rdi)

+	vmovups %zmm7, 0x1C0(%rdi)

+	vmovups	%zmm8, -0x200(%r9)

+	vmovups %zmm9, -0x1C0(%r9)

+	vmovups %zmm10, -0x180(%r9)

+	vmovups %zmm11, -0x140(%r9)

+	vmovups	%zmm12, -0x100(%r9)

+	vmovups %zmm13, -0xC0(%r9)

+	vmovups %zmm14, -0x80(%r9)

+	vmovups %zmm15, -0x40(%r9)

+	ret

+

+L(1024bytesormore):

+	cmp	%rsi, %rdi

+	ja	L(1024bytesormore_bkw)

+	sub	$512, %r9

+	vmovups -0x200(%rcx), %zmm8

+	vmovups -0x1C0(%rcx), %zmm9

+	vmovups -0x180(%rcx), %zmm10

+	vmovups -0x140(%rcx), %zmm11

+	vmovups	-0x100(%rcx), %zmm12

+	vmovups -0xC0(%rcx), %zmm13

+	vmovups -0x80(%rcx), %zmm14

+	vmovups -0x40(%rcx), %zmm15

+	prefetcht1 (%rsi)

+	prefetcht1 0x40(%rsi)

+	prefetcht1 0x80(%rsi)

+	prefetcht1 0xC0(%rsi)

+	prefetcht1 0x100(%rsi)

+	prefetcht1 0x140(%rsi)

+	prefetcht1 0x180(%rsi)

+	prefetcht1 0x1C0(%rsi)

+

+/* Loop with unaligned memory access.  */

+L(gobble_512bytes_loop):

+	vmovups	(%rsi), %zmm0

+	vmovups 0x40(%rsi), %zmm1

+	vmovups 0x80(%rsi), %zmm2

+	vmovups 0xC0(%rsi), %zmm3

+	vmovups	0x100(%rsi), %zmm4

+	vmovups 0x140(%rsi), %zmm5

+	vmovups 0x180(%rsi), %zmm6

+	vmovups 0x1C0(%rsi), %zmm7

+	add	$512, %rsi

+	prefetcht1 (%rsi)

+	prefetcht1 0x40(%rsi)

+	prefetcht1 0x80(%rsi)

+	prefetcht1 0xC0(%rsi)

+	prefetcht1 0x100(%rsi)

+	prefetcht1 0x140(%rsi)

+	prefetcht1 0x180(%rsi)

+	prefetcht1 0x1C0(%rsi)

+	vmovups	%zmm0, (%rdi)

+	vmovups %zmm1, 0x40(%rdi)

+	vmovups %zmm2, 0x80(%rdi)

+	vmovups %zmm3, 0xC0(%rdi)

+	vmovups	%zmm4, 0x100(%rdi)

+	vmovups %zmm5, 0x140(%rdi)

+	vmovups %zmm6, 0x180(%rdi)

+	vmovups %zmm7, 0x1C0(%rdi)

+	add	$512, %rdi

+	cmp	%r9, %rdi

+	jb	L(gobble_512bytes_loop)

+	vmovups %zmm8, (%r9)

+	vmovups %zmm9, 0x40(%r9)

+	vmovups %zmm10, 0x80(%r9)

+	vmovups %zmm11, 0xC0(%r9)

+	vmovups %zmm12, 0x100(%r9)

+	vmovups %zmm13, 0x140(%r9)

+	vmovups %zmm14, 0x180(%r9)

+	vmovups %zmm15, 0x1C0(%r9)

+	ret

+

+L(1024bytesormore_bkw):

+	add	$512, %rdi

+	vmovups	0x1C0(%rsi), %zmm8

+	vmovups 0x180(%rsi), %zmm9

+	vmovups 0x140(%rsi), %zmm10

+	vmovups 0x100(%rsi), %zmm11

+	vmovups	0xC0(%rsi), %zmm12

+	vmovups 0x80(%rsi), %zmm13

+	vmovups 0x40(%rsi), %zmm14

+	vmovups (%rsi), %zmm15

+	prefetcht1 -0x40(%rcx)

+	prefetcht1 -0x80(%rcx)

+	prefetcht1 -0xC0(%rcx)

+	prefetcht1 -0x100(%rcx)

+	prefetcht1 -0x140(%rcx)

+	prefetcht1 -0x180(%rcx)

+	prefetcht1 -0x1C0(%rcx)

+	prefetcht1 -0x200(%rcx)

+	

+/* Backward loop with unaligned memory access.  */

+L(gobble_512bytes_loop_bkw):

+	vmovups -0x40(%rcx), %zmm0

+	vmovups -0x80(%rcx), %zmm1

+	vmovups -0xC0(%rcx), %zmm2

+	vmovups	-0x100(%rcx), %zmm3

+	vmovups -0x140(%rcx), %zmm4

+	vmovups -0x180(%rcx), %zmm5

+	vmovups -0x1C0(%rcx), %zmm6

+	vmovups	-0x200(%rcx), %zmm7

+	sub	$512, %rcx

+	prefetcht1 -0x40(%rcx)

+	prefetcht1 -0x80(%rcx)

+	prefetcht1 -0xC0(%rcx)

+	prefetcht1 -0x100(%rcx)

+	prefetcht1 -0x140(%rcx)

+	prefetcht1 -0x180(%rcx)

+	prefetcht1 -0x1C0(%rcx)

+	prefetcht1 -0x200(%rcx)

+	vmovups %zmm0, -0x40(%r9)	

+	vmovups %zmm1, -0x80(%r9)	

+	vmovups %zmm2, -0xC0(%r9)	

+	vmovups	%zmm3, -0x100(%r9)	

+	vmovups %zmm4, -0x140(%r9)	

+	vmovups %zmm5, -0x180(%r9)	

+	vmovups %zmm6, -0x1C0(%r9)

+	vmovups	%zmm7, -0x200(%r9)	

+	sub	$512, %r9

+	cmp	%rdi, %r9

+	ja	L(gobble_512bytes_loop_bkw)

+	vmovups %zmm8, -0x40(%rdi)

+	vmovups %zmm9, -0x80(%rdi)

+	vmovups %zmm10, -0xC0(%rdi)

+	vmovups %zmm11, -0x100(%rdi)

+	vmovups %zmm12, -0x140(%rdi)

+	vmovups %zmm13, -0x180(%rdi)

+	vmovups %zmm14, -0x1C0(%rdi)

+	vmovups %zmm15, -0x200(%rdi)

+	ret

+

+L(preloop_large):

+	cmp	%rsi, %rdi

+	ja	L(preloop_large_bkw)

+	vmovups	(%rsi), %zmm4

+	vmovups	0x40(%rsi), %zmm5

+

+/* Align destination for access with non-temporal stores in the loop.  */

+	mov	%rdi, %r8

+	and	$-0x80, %rdi

+	add	$0x80, %rdi

+	sub	%rdi, %r8	

+	sub	%r8, %rsi

+	add	%r8, %rdx

+	prefetcht1 (%rsi)

+	prefetcht1 0x40(%rsi)

+	prefetcht1 0x80(%rsi)

+	prefetcht1 0xC0(%rsi)

+L(gobble_256bytes_nt_loop):

+	vmovups	(%rsi), %zmm0

+	prefetcht1 0x100(%rsi)

+	vmovups	0x40(%rsi), %zmm1

+	prefetcht1 0x140(%rsi)

+	vmovups	0x80(%rsi), %zmm2

+	prefetcht1 0x180(%rsi)

+	vmovups	0xC0(%rsi), %zmm3

+	prefetcht1 0x1C0(%rsi)

+	vmovntdq %zmm0, (%rdi)

+	vmovntdq %zmm1, 0x40(%rdi)

+	vmovntdq %zmm2, 0x80(%rdi)

+	vmovntdq %zmm3, 0xC0(%rdi)

+	sub	$256, %rdx

+	add	$256, %rsi

+	add	$256, %rdi

+	cmp	$256, %rdx

+	ja	L(gobble_256bytes_nt_loop)

+	sfence

+	vmovups	%zmm4, (%rax)

+	vmovups	%zmm5, 0x40(%rax)

+	jmp	L(check)

+

+L(preloop_large_bkw):

+	vmovups -0x80(%rcx), %zmm4

+	vmovups -0x40(%rcx), %zmm5

+

+/* Align end of destination for access with non-temporal stores.  */

+	mov	%r9, %r8

+	and	$-0x80, %r9

+	sub	%r9, %r8

+	sub	%r8, %rcx

+	sub	%r8, %rdx

+	add	%r9, %r8

+ 	prefetcht1 -0x100(%rcx)

+	prefetcht1 -0xC0(%rcx)

+	prefetcht1 -0x80(%rcx)

+	prefetcht1 -0x40(%rcx)

+L(gobble_256bytes_nt_loop_bkw):

+	vmovups	-0x100(%rcx), %zmm0

+	prefetcht1 -0x200(%rcx)

+	vmovups	-0xC0(%rcx), %zmm1

+	prefetcht1 -0x1C0(%rcx)

+	vmovups	-0x80(%rcx), %zmm2

+	prefetcht1 -0x180(%rcx)

+	vmovups	-0x40(%rcx), %zmm3

+	prefetcht1 -0x140(%rcx)

+	vmovntdq %zmm0,	-0x100(%r9)

+	vmovntdq %zmm1,	-0xC0(%r9)

+	vmovntdq %zmm2,	-0x80(%r9)

+	vmovntdq %zmm3,	-0x40(%r9)

+	sub	$256, %rdx

+	sub	$256, %rcx

+	sub	$256, %r9

+	cmp	$256, %rdx

+	ja	L(gobble_256bytes_nt_loop_bkw)

+	sfence

+	vmovups	%zmm4, -0x80(%r8)

+	vmovups	%zmm5, -0x40(%r8)

+	jmp	L(check)

+

+END (MEMCPY)

+

+#endif