glibc/sysdeps/x86_64/multiarch/memchr-avx2.S
H.J. Lu 2f5d20ac99 x86-64: Optimize memchr/rawmemchr/wmemchr with SSE2/AVX2
SSE2 memchr is extended to support wmemchr.  AVX2 memchr/rawmemchr/wmemchr
are added to search 32 bytes with a single vector compare instruction.
AVX2 memchr/rawmemchr/wmemchr are as fast as SSE2 memchr/rawmemchr/wmemchr
for small sizes and up to 1.5X faster for larger sizes on Haswell and
Skylake.  Select AVX2 memchr/rawmemchr/wmemchr on AVX2 machines where
vzeroupper is preferred and AVX unaligned load is fast.

NB: It uses TZCNT instead of BSF since TZCNT produces the same result
as BSF for non-zero input.  TZCNT is faster than BSF and is executed
as BSF if machine doesn't support TZCNT.

	* sysdeps/x86_64/memchr.S (MEMCHR): New.  Depending on if
	USE_AS_WMEMCHR is defined.
	(PCMPEQ): Likewise.
	(memchr): Renamed to ...
	(MEMCHR): This.  Support wmemchr if USE_AS_WMEMCHR is defined.
	Replace pcmpeqb with PCMPEQ.
	* sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Add
	memchr-sse2, rawmemchr-sse2, memchr-avx2, rawmemchr-avx2,
	wmemchr-sse4_1, wmemchr-avx2 and wmemchr-c.
	* sysdeps/x86_64/multiarch/ifunc-avx2.h: New file.
	* sysdeps/x86_64/multiarch/memchr-avx2.S: Likewise.
	* sysdeps/x86_64/multiarch/memchr-sse2.S: Likewise.
	* sysdeps/x86_64/multiarch/memchr.c: Likewise.
	* sysdeps/x86_64/multiarch/rawmemchr-avx2.S: Likewise.
	* sysdeps/x86_64/multiarch/rawmemchr-sse2.S: Likewise.
	* sysdeps/x86_64/multiarch/rawmemchr.c: Likewise.
	* sysdeps/x86_64/multiarch/wmemchr-avx2.S: Likewise.
	* sysdeps/x86_64/multiarch/wmemchr-sse2.S: Likewise.
	* sysdeps/x86_64/multiarch/wmemchr.c: Likewise.
	* sysdeps/x86_64/multiarch/ifunc-impl-list.c
	(__libc_ifunc_impl_list): Test __memchr_avx2, __memchr_sse2,
	__rawmemchr_avx2, __rawmemchr_sse2, __wmemchr_avx2 and
	__wmemchr_sse2.
2017-06-09 05:13:31 -07:00

341 lines
6.5 KiB
ArmAsm

/* memchr/wmemchr optimized with AVX2.
Copyright (C) 2017 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/>. */
#if IS_IN (libc)
# include <sysdep.h>
# ifndef MEMCHR
# define MEMCHR __memchr_avx2
# endif
# ifdef USE_AS_WMEMCHR
# define VPCMPEQ vpcmpeqd
# else
# define VPCMPEQ vpcmpeqb
# endif
# ifndef VZEROUPPER
# define VZEROUPPER vzeroupper
# endif
# define VEC_SIZE 32
.section .text.avx,"ax",@progbits
ENTRY (MEMCHR)
# ifndef USE_AS_RAWMEMCHR
/* Check for zero length. */
testq %rdx, %rdx
jz L(null)
# endif
movl %edi, %ecx
/* Broadcast CHAR to YMM0. */
vmovd %esi, %xmm0
# ifdef USE_AS_WMEMCHR
shl $2, %rdx
vpbroadcastd %xmm0, %ymm0
# else
vpbroadcastb %xmm0, %ymm0
# endif
/* Check if we may cross page boundary with one vector load. */
andl $(2 * VEC_SIZE - 1), %ecx
cmpl $VEC_SIZE, %ecx
ja L(cros_page_boundary)
/* Check the first VEC_SIZE bytes. */
VPCMPEQ (%rdi), %ymm0, %ymm1
vpmovmskb %ymm1, %eax
testl %eax, %eax
# ifndef USE_AS_RAWMEMCHR
jnz L(first_vec_x0_check)
/* Adjust length and check the end of data. */
subq $VEC_SIZE, %rdx
jbe L(zero)
# else
jnz L(first_vec_x0)
# endif
/* Align data for aligned loads in the loop. */
addq $VEC_SIZE, %rdi
andl $(VEC_SIZE - 1), %ecx
andq $-VEC_SIZE, %rdi
# ifndef USE_AS_RAWMEMCHR
/* Adjust length. */
addq %rcx, %rdx
subq $(VEC_SIZE * 4), %rdx
jbe L(last_4x_vec_or_less)
# endif
jmp L(more_4x_vec)
.p2align 4
L(cros_page_boundary):
andl $(VEC_SIZE - 1), %ecx
andq $-VEC_SIZE, %rdi
VPCMPEQ (%rdi), %ymm0, %ymm1
vpmovmskb %ymm1, %eax
/* Remove the leading bytes. */
sarl %cl, %eax
testl %eax, %eax
jz L(aligned_more)
tzcntl %eax, %eax
# ifndef USE_AS_RAWMEMCHR
/* Check the end of data. */
cmpq %rax, %rdx
jbe L(zero)
# endif
addq %rdi, %rax
addq %rcx, %rax
VZEROUPPER
ret
.p2align 4
L(aligned_more):
# ifndef USE_AS_RAWMEMCHR
/* Calculate "rdx + rcx - VEC_SIZE" with "rdx - (VEC_SIZE - rcx)"
instead of "(rdx + rcx) - VEC_SIZE" to void possible addition
overflow. */
negq %rcx
addq $VEC_SIZE, %rcx
/* Check the end of data. */
subq %rcx, %rdx
jbe L(zero)
# endif
addq $VEC_SIZE, %rdi
# ifndef USE_AS_RAWMEMCHR
subq $(VEC_SIZE * 4), %rdx
jbe L(last_4x_vec_or_less)
# endif
L(more_4x_vec):
/* Check the first 4 * VEC_SIZE. Only one VEC_SIZE at a time
since data is only aligned to VEC_SIZE. */
VPCMPEQ (%rdi), %ymm0, %ymm1
vpmovmskb %ymm1, %eax
testl %eax, %eax
jnz L(first_vec_x0)
VPCMPEQ VEC_SIZE(%rdi), %ymm0, %ymm1
vpmovmskb %ymm1, %eax
testl %eax, %eax
jnz L(first_vec_x1)
VPCMPEQ (VEC_SIZE * 2)(%rdi), %ymm0, %ymm1
vpmovmskb %ymm1, %eax
testl %eax, %eax
jnz L(first_vec_x2)
VPCMPEQ (VEC_SIZE * 3)(%rdi), %ymm0, %ymm1
vpmovmskb %ymm1, %eax
testl %eax, %eax
jnz L(first_vec_x3)
addq $(VEC_SIZE * 4), %rdi
# ifndef USE_AS_RAWMEMCHR
subq $(VEC_SIZE * 4), %rdx
jbe L(last_4x_vec_or_less)
# endif
/* Align data to 4 * VEC_SIZE. */
movq %rdi, %rcx
andl $(4 * VEC_SIZE - 1), %ecx
andq $-(4 * VEC_SIZE), %rdi
# ifndef USE_AS_RAWMEMCHR
/* Adjust length. */
addq %rcx, %rdx
# endif
.p2align 4
L(loop_4x_vec):
/* Compare 4 * VEC at a time forward. */
VPCMPEQ (%rdi), %ymm0, %ymm1
VPCMPEQ VEC_SIZE(%rdi), %ymm0, %ymm2
VPCMPEQ (VEC_SIZE * 2)(%rdi), %ymm0, %ymm3
VPCMPEQ (VEC_SIZE * 3)(%rdi), %ymm0, %ymm4
vpor %ymm1, %ymm2, %ymm5
vpor %ymm3, %ymm4, %ymm6
vpor %ymm5, %ymm6, %ymm5
vpmovmskb %ymm5, %eax
testl %eax, %eax
jnz L(4x_vec_end)
addq $(VEC_SIZE * 4), %rdi
# ifdef USE_AS_RAWMEMCHR
jmp L(loop_4x_vec)
# else
subq $(VEC_SIZE * 4), %rdx
ja L(loop_4x_vec)
L(last_4x_vec_or_less):
/* Less than 4 * VEC and aligned to VEC_SIZE. */
addl $(VEC_SIZE * 2), %edx
jle L(last_2x_vec)
VPCMPEQ (%rdi), %ymm0, %ymm1
vpmovmskb %ymm1, %eax
testl %eax, %eax
jnz L(first_vec_x0)
VPCMPEQ VEC_SIZE(%rdi), %ymm0, %ymm1
vpmovmskb %ymm1, %eax
testl %eax, %eax
jnz L(first_vec_x1)
VPCMPEQ (VEC_SIZE * 2)(%rdi), %ymm0, %ymm1
vpmovmskb %ymm1, %eax
testl %eax, %eax
jnz L(first_vec_x2_check)
subl $VEC_SIZE, %edx
jle L(zero)
VPCMPEQ (VEC_SIZE * 3)(%rdi), %ymm0, %ymm1
vpmovmskb %ymm1, %eax
testl %eax, %eax
jnz L(first_vec_x3_check)
xorl %eax, %eax
VZEROUPPER
ret
.p2align 4
L(last_2x_vec):
addl $(VEC_SIZE * 2), %edx
VPCMPEQ (%rdi), %ymm0, %ymm1
vpmovmskb %ymm1, %eax
testl %eax, %eax
jnz L(first_vec_x0_check)
subl $VEC_SIZE, %edx
jle L(zero)
VPCMPEQ VEC_SIZE(%rdi), %ymm0, %ymm1
vpmovmskb %ymm1, %eax
testl %eax, %eax
jnz L(first_vec_x1_check)
xorl %eax, %eax
VZEROUPPER
ret
.p2align 4
L(first_vec_x0_check):
tzcntl %eax, %eax
/* Check the end of data. */
cmpq %rax, %rdx
jbe L(zero)
addq %rdi, %rax
VZEROUPPER
ret
.p2align 4
L(first_vec_x1_check):
tzcntl %eax, %eax
/* Check the end of data. */
cmpq %rax, %rdx
jbe L(zero)
addq $VEC_SIZE, %rax
addq %rdi, %rax
VZEROUPPER
ret
.p2align 4
L(first_vec_x2_check):
tzcntl %eax, %eax
/* Check the end of data. */
cmpq %rax, %rdx
jbe L(zero)
addq $(VEC_SIZE * 2), %rax
addq %rdi, %rax
VZEROUPPER
ret
.p2align 4
L(first_vec_x3_check):
tzcntl %eax, %eax
/* Check the end of data. */
cmpq %rax, %rdx
jbe L(zero)
addq $(VEC_SIZE * 3), %rax
addq %rdi, %rax
VZEROUPPER
ret
.p2align 4
L(zero):
VZEROUPPER
L(null):
xorl %eax, %eax
ret
# endif
.p2align 4
L(first_vec_x0):
tzcntl %eax, %eax
addq %rdi, %rax
VZEROUPPER
ret
.p2align 4
L(first_vec_x1):
tzcntl %eax, %eax
addq $VEC_SIZE, %rax
addq %rdi, %rax
VZEROUPPER
ret
.p2align 4
L(first_vec_x2):
tzcntl %eax, %eax
addq $(VEC_SIZE * 2), %rax
addq %rdi, %rax
VZEROUPPER
ret
.p2align 4
L(4x_vec_end):
vpmovmskb %ymm1, %eax
testl %eax, %eax
jnz L(first_vec_x0)
vpmovmskb %ymm2, %eax
testl %eax, %eax
jnz L(first_vec_x1)
vpmovmskb %ymm3, %eax
testl %eax, %eax
jnz L(first_vec_x2)
vpmovmskb %ymm4, %eax
testl %eax, %eax
L(first_vec_x3):
tzcntl %eax, %eax
addq $(VEC_SIZE * 3), %rax
addq %rdi, %rax
VZEROUPPER
ret
END (MEMCHR)
#endif