glibc/sysdeps/x86_64/multiarch/memcpy-avx-unaligned.S
Ling Ma 05f3633da4 Improve 64bit memcpy performance for Haswell CPU with AVX instruction
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.

The CPU2006 403.gcc benchmark indicates this patch improves performance
from 2% to 10%.
2014-07-30 08:02:35 -07:00

377 lines
8.3 KiB
ArmAsm

/* 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 $16, %dl
jb L(less_16bytes)
cmp $128, %dl
jb L(less_128bytes)
vmovdqu (%rsi), %xmm0
lea (%rsi, %rdx), %rcx
vmovdqu 0x10(%rsi), %xmm1
vmovdqu 0x20(%rsi), %xmm2
vmovdqu 0x30(%rsi), %xmm3
vmovdqu 0x40(%rsi), %xmm4
vmovdqu 0x50(%rsi), %xmm5
vmovdqu 0x60(%rsi), %xmm6
vmovdqu 0x70(%rsi), %xmm7
vmovdqu -0x80(%rcx), %xmm8
vmovdqu -0x70(%rcx), %xmm9
vmovdqu -0x60(%rcx), %xmm10
vmovdqu -0x50(%rcx), %xmm11
vmovdqu -0x40(%rcx), %xmm12
vmovdqu -0x30(%rcx), %xmm13
vmovdqu -0x20(%rcx), %xmm14
vmovdqu -0x10(%rcx), %xmm15
lea (%rdi, %rdx), %rdx
vmovdqu %xmm0, (%rdi)
vmovdqu %xmm1, 0x10(%rdi)
vmovdqu %xmm2, 0x20(%rdi)
vmovdqu %xmm3, 0x30(%rdi)
vmovdqu %xmm4, 0x40(%rdi)
vmovdqu %xmm5, 0x50(%rdi)
vmovdqu %xmm6, 0x60(%rdi)
vmovdqu %xmm7, 0x70(%rdi)
vmovdqu %xmm8, -0x80(%rdx)
vmovdqu %xmm9, -0x70(%rdx)
vmovdqu %xmm10, -0x60(%rdx)
vmovdqu %xmm11, -0x50(%rdx)
vmovdqu %xmm12, -0x40(%rdx)
vmovdqu %xmm13, -0x30(%rdx)
vmovdqu %xmm14, -0x20(%rdx)
vmovdqu %xmm15, -0x10(%rdx)
ret
.p2align 4
L(less_128bytes):
cmp $64, %dl
jb L(less_64bytes)
vmovdqu (%rsi), %xmm0
lea (%rsi, %rdx), %rcx
vmovdqu 0x10(%rsi), %xmm1
vmovdqu 0x20(%rsi), %xmm2
lea (%rdi, %rdx), %rdx
vmovdqu 0x30(%rsi), %xmm3
vmovdqu -0x40(%rcx), %xmm4
vmovdqu -0x30(%rcx), %xmm5
vmovdqu -0x20(%rcx), %xmm6
vmovdqu -0x10(%rcx), %xmm7
vmovdqu %xmm0, (%rdi)
vmovdqu %xmm1, 0x10(%rdi)
vmovdqu %xmm2, 0x20(%rdi)
vmovdqu %xmm3, 0x30(%rdi)
vmovdqu %xmm4, -0x40(%rdx)
vmovdqu %xmm5, -0x30(%rdx)
vmovdqu %xmm6, -0x20(%rdx)
vmovdqu %xmm7, -0x10(%rdx)
ret
.p2align 4
L(less_64bytes):
cmp $32, %dl
jb L(less_32bytes)
vmovdqu (%rsi), %xmm0
vmovdqu 0x10(%rsi), %xmm1
vmovdqu -0x20(%rsi, %rdx), %xmm6
vmovdqu -0x10(%rsi, %rdx), %xmm7
vmovdqu %xmm0, (%rdi)
vmovdqu %xmm1, 0x10(%rdi)
vmovdqu %xmm6, -0x20(%rdi, %rdx)
vmovdqu %xmm7, -0x10(%rdi, %rdx)
ret
.p2align 4
L(less_32bytes):
vmovdqu (%rsi), %xmm0
vmovdqu -0x10(%rsi, %rdx), %xmm7
vmovdqu %xmm0, (%rdi)
vmovdqu %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
mov %rdi, %rcx
sub %rsi, %rcx
cmp %rdx, %rcx
jc L(copy_backward)
#endif
cmp $2048, %rdx
jae L(gobble_data_movsb)
mov %rax, %r8
lea (%rsi, %rdx), %rcx
mov %rdi, %r10
vmovdqu -0x80(%rcx), %xmm5
vmovdqu -0x70(%rcx), %xmm6
mov $0x80, %rax
and $-32, %rdi
add $32, %rdi
vmovdqu -0x60(%rcx), %xmm7
vmovdqu -0x50(%rcx), %xmm8
mov %rdi, %r11
sub %r10, %r11
vmovdqu -0x40(%rcx), %xmm9
vmovdqu -0x30(%rcx), %xmm10
sub %r11, %rdx
vmovdqu -0x20(%rcx), %xmm11
vmovdqu -0x10(%rcx), %xmm12
vmovdqu (%rsi), %ymm4
add %r11, %rsi
sub %eax, %edx
L(goble_128_loop):
vmovdqu (%rsi), %ymm0
vmovdqu 0x20(%rsi), %ymm1
vmovdqu 0x40(%rsi), %ymm2
vmovdqu 0x60(%rsi), %ymm3
add %rax, %rsi
vmovdqa %ymm0, (%rdi)
vmovdqa %ymm1, 0x20(%rdi)
vmovdqa %ymm2, 0x40(%rdi)
vmovdqa %ymm3, 0x60(%rdi)
add %rax, %rdi
sub %eax, %edx
jae L(goble_128_loop)
add %eax, %edx
add %rdi, %rdx
vmovdqu %ymm4, (%r10)
vzeroupper
vmovdqu %xmm5, -0x80(%rdx)
vmovdqu %xmm6, -0x70(%rdx)
vmovdqu %xmm7, -0x60(%rdx)
vmovdqu %xmm8, -0x50(%rdx)
vmovdqu %xmm9, -0x40(%rdx)
vmovdqu %xmm10, -0x30(%rdx)
vmovdqu %xmm11, -0x20(%rdx)
vmovdqu %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
cmp %rcx, %rdx
jae L(gobble_big_data_fwd)
mov %rdx, %rcx
mov %rdx, %rcx
rep movsb
ret
.p2align 4
L(gobble_big_data_fwd):
lea (%rsi, %rdx), %rcx
vmovdqu (%rsi), %ymm4
vmovdqu -0x80(%rsi,%rdx), %xmm5
vmovdqu -0x70(%rcx), %xmm6
vmovdqu -0x60(%rcx), %xmm7
vmovdqu -0x50(%rcx), %xmm8
vmovdqu -0x40(%rcx), %xmm9
vmovdqu -0x30(%rcx), %xmm10
vmovdqu -0x20(%rcx), %xmm11
vmovdqu -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
add $-0x80, %rdx
L(gobble_mem_fwd_loop):
prefetchnta 0x1c0(%rsi)
prefetchnta 0x280(%rsi)
vmovdqu (%rsi), %ymm0
vmovdqu 0x20(%rsi), %ymm1
vmovdqu 0x40(%rsi), %ymm2
vmovdqu 0x60(%rsi), %ymm3
sub $-0x80, %rsi
vmovntdq %ymm0, (%rdi)
vmovntdq %ymm1, 0x20(%rdi)
vmovntdq %ymm2, 0x40(%rdi)
vmovntdq %ymm3, 0x60(%rdi)
sub $-0x80, %rdi
add $-0x80, %rdx
jb L(gobble_mem_fwd_loop)
sfence
vmovdqu %ymm4, (%r8)
vzeroupper
vmovdqu %xmm5, -0x80(%rcx)
vmovdqu %xmm6, -0x70(%rcx)
vmovdqu %xmm7, -0x60(%rcx)
vmovdqu %xmm8, -0x50(%rcx)
vmovdqu %xmm9, -0x40(%rcx)
vmovdqu %xmm10, -0x30(%rcx)
vmovdqu %xmm11, -0x20(%rcx)
vmovdqu %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
vmovdqu (%rsi), %xmm5
vmovdqu 0x10(%rsi), %xmm6
add %rdx, %rdi
vmovdqu 0x20(%rsi), %xmm7
vmovdqu 0x30(%rsi), %xmm8
lea -0x20(%rdi), %r10
mov %rdi, %r11
vmovdqu 0x40(%rsi), %xmm9
vmovdqu 0x50(%rsi), %xmm10
and $0x1f, %r11
vmovdqu 0x60(%rsi), %xmm11
vmovdqu 0x70(%rsi), %xmm12
xor %r11, %rdi
add %rdx, %rsi
vmovdqu -0x20(%rsi), %ymm4
sub %r11, %rsi
sub %r11, %rdx
cmp %rcx, %rdx
ja L(gobble_big_data_bwd)
add $-0x80, %rdx
L(gobble_mem_bwd_llc):
vmovdqu -0x20(%rsi), %ymm0
vmovdqu -0x40(%rsi), %ymm1
vmovdqu -0x60(%rsi), %ymm2
vmovdqu -0x80(%rsi), %ymm3
lea -0x80(%rsi), %rsi
vmovdqa %ymm0, -0x20(%rdi)
vmovdqa %ymm1, -0x40(%rdi)
vmovdqa %ymm2, -0x60(%rdi)
vmovdqa %ymm3, -0x80(%rdi)
lea -0x80(%rdi), %rdi
add $-0x80, %rdx
jb L(gobble_mem_bwd_llc)
vmovdqu %ymm4, (%r10)
vzeroupper
vmovdqu %xmm5, (%rax)
vmovdqu %xmm6, 0x10(%rax)
vmovdqu %xmm7, 0x20(%rax)
vmovdqu %xmm8, 0x30(%rax)
vmovdqu %xmm9, 0x40(%rax)
vmovdqu %xmm10, 0x50(%rax)
vmovdqu %xmm11, 0x60(%rax)
vmovdqu %xmm12, 0x70(%rax)
ret
.p2align 4
L(gobble_big_data_bwd):
add $-0x80, %rdx
L(gobble_mem_bwd_loop):
prefetchnta -0x1c0(%rsi)
prefetchnta -0x280(%rsi)
vmovdqu -0x20(%rsi), %ymm0
vmovdqu -0x40(%rsi), %ymm1
vmovdqu -0x60(%rsi), %ymm2
vmovdqu -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
add $-0x80, %rdx
jb L(gobble_mem_bwd_loop)
sfence
vmovdqu %ymm4, (%r10)
vzeroupper
vmovdqu %xmm5, (%rax)
vmovdqu %xmm6, 0x10(%rax)
vmovdqu %xmm7, 0x20(%rax)
vmovdqu %xmm8, 0x30(%rax)
vmovdqu %xmm9, 0x40(%rax)
vmovdqu %xmm10, 0x50(%rax)
vmovdqu %xmm11, 0x60(%rax)
vmovdqu %xmm12, 0x70(%rax)
ret
#endif
END (MEMCPY)
#endif