|
Packit |
6c4009 |
/* strcmp/wcscmp/strncmp/wcsncmp optimized with AVX2.
|
|
Packit |
6c4009 |
Copyright (C) 2018 Free Software Foundation, Inc.
|
|
Packit |
6c4009 |
This file is part of the GNU C Library.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
The GNU C Library is free software; you can redistribute it and/or
|
|
Packit |
6c4009 |
modify it under the terms of the GNU Lesser General Public
|
|
Packit |
6c4009 |
License as published by the Free Software Foundation; either
|
|
Packit |
6c4009 |
version 2.1 of the License, or (at your option) any later version.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
The GNU C Library is distributed in the hope that it will be useful,
|
|
Packit |
6c4009 |
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
Packit |
6c4009 |
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
Packit |
6c4009 |
Lesser General Public License for more details.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
You should have received a copy of the GNU Lesser General Public
|
|
Packit |
6c4009 |
License along with the GNU C Library; if not, see
|
|
Packit |
6c4009 |
<http://www.gnu.org/licenses/>. */
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
#if IS_IN (libc)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
# include <sysdep.h>
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
# ifndef STRCMP
|
|
Packit |
6c4009 |
# define STRCMP __strcmp_avx2
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
# define PAGE_SIZE 4096
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* VEC_SIZE = Number of bytes in a ymm register */
|
|
Packit |
6c4009 |
# define VEC_SIZE 32
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Shift for dividing by (VEC_SIZE * 4). */
|
|
Packit |
6c4009 |
# define DIVIDE_BY_VEC_4_SHIFT 7
|
|
Packit |
6c4009 |
# if (VEC_SIZE * 4) != (1 << DIVIDE_BY_VEC_4_SHIFT)
|
|
Packit |
6c4009 |
# error (VEC_SIZE * 4) != (1 << DIVIDE_BY_VEC_4_SHIFT)
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
/* Compare packed dwords. */
|
|
Packit |
6c4009 |
# define VPCMPEQ vpcmpeqd
|
|
Packit |
6c4009 |
/* Compare packed dwords and store minimum. */
|
|
Packit |
6c4009 |
# define VPMINU vpminud
|
|
Packit |
6c4009 |
/* 1 dword char == 4 bytes. */
|
|
Packit |
6c4009 |
# define SIZE_OF_CHAR 4
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
/* Compare packed bytes. */
|
|
Packit |
6c4009 |
# define VPCMPEQ vpcmpeqb
|
|
Packit |
6c4009 |
/* Compare packed bytes and store minimum. */
|
|
Packit |
6c4009 |
# define VPMINU vpminub
|
|
Packit |
6c4009 |
/* 1 byte char == 1 byte. */
|
|
Packit |
6c4009 |
# define SIZE_OF_CHAR 1
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
# ifndef VZEROUPPER
|
|
Packit |
6c4009 |
# define VZEROUPPER vzeroupper
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Warning!
|
|
Packit |
6c4009 |
wcscmp/wcsncmp have to use SIGNED comparison for elements.
|
|
Packit |
6c4009 |
strcmp/strncmp have to use UNSIGNED comparison for elements.
|
|
Packit |
6c4009 |
*/
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* The main idea of the string comparison (byte or dword) using AVX2
|
|
Packit |
6c4009 |
consists of comparing (VPCMPEQ) two ymm vectors. The latter can be on
|
|
Packit |
6c4009 |
either packed bytes or dwords depending on USE_AS_WCSCMP. In order
|
|
Packit |
6c4009 |
to check the null char, algorithm keeps the matched bytes/dwords,
|
|
Packit |
6c4009 |
requiring two more AVX2 instructions (VPMINU and VPCMPEQ). In general,
|
|
Packit |
6c4009 |
the costs of comparing VEC_SIZE bytes (32-bytes) are two VPCMPEQ and
|
|
Packit |
6c4009 |
one VPMINU instructions, together with movdqu and testl instructions.
|
|
Packit |
6c4009 |
Main loop (away from from page boundary) compares 4 vectors are a time,
|
|
Packit |
6c4009 |
effectively comparing 4 x VEC_SIZE bytes (128 bytes) on each loop.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
The routine strncmp/wcsncmp (enabled by defining USE_AS_STRNCMP) logic
|
|
Packit |
6c4009 |
is the same as strcmp, except that an a maximum offset is tracked. If
|
|
Packit |
6c4009 |
the maximum offset is reached before a difference is found, zero is
|
|
Packit |
6c4009 |
returned. */
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.section .text.avx,"ax",@progbits
|
|
Packit |
6c4009 |
ENTRY (STRCMP)
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
/* Check for simple cases (0 or 1) in offset. */
|
|
Packit |
6c4009 |
cmp $1, %rdx
|
|
Packit |
6c4009 |
je L(char0)
|
|
Packit |
6c4009 |
jb L(zero)
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
/* Convert units: from wide to byte char. */
|
|
Packit |
6c4009 |
shl $2, %rdx
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
/* Register %r11 tracks the maximum offset. */
|
|
Packit |
6c4009 |
movq %rdx, %r11
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
movl %edi, %eax
|
|
Packit |
6c4009 |
xorl %edx, %edx
|
|
Packit |
6c4009 |
/* Make %ymm7 all zeros in this function. */
|
|
Packit |
6c4009 |
vpxor %ymm7, %ymm7, %ymm7
|
|
Packit |
6c4009 |
orl %esi, %eax
|
|
Packit |
6c4009 |
andl $(PAGE_SIZE - 1), %eax
|
|
Packit |
6c4009 |
cmpl $(PAGE_SIZE - (VEC_SIZE * 4)), %eax
|
|
Packit |
6c4009 |
jg L(cross_page)
|
|
Packit |
6c4009 |
/* Start comparing 4 vectors. */
|
|
Packit |
6c4009 |
vmovdqu (%rdi), %ymm1
|
|
Packit |
6c4009 |
VPCMPEQ (%rsi), %ymm1, %ymm0
|
|
Packit |
6c4009 |
VPMINU %ymm1, %ymm0, %ymm0
|
|
Packit |
6c4009 |
VPCMPEQ %ymm7, %ymm0, %ymm0
|
|
Packit |
6c4009 |
vpmovmskb %ymm0, %ecx
|
|
Packit |
6c4009 |
testl %ecx, %ecx
|
|
Packit |
6c4009 |
je L(next_3_vectors)
|
|
Packit |
6c4009 |
tzcntl %ecx, %edx
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
/* Return 0 if the mismatched index (%rdx) is after the maximum
|
|
Packit |
6c4009 |
offset (%r11). */
|
|
Packit |
6c4009 |
cmpq %r11, %rdx
|
|
Packit |
6c4009 |
jae L(zero)
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
movl (%rdi, %rdx), %ecx
|
|
Packit |
6c4009 |
cmpl (%rsi, %rdx), %ecx
|
|
Packit |
6c4009 |
je L(return)
|
|
Packit |
6c4009 |
L(wcscmp_return):
|
|
Packit |
6c4009 |
setl %al
|
|
Packit |
6c4009 |
negl %eax
|
|
Packit |
6c4009 |
orl $1, %eax
|
|
Packit |
6c4009 |
L(return):
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
movzbl (%rdi, %rdx), %eax
|
|
Packit |
6c4009 |
movzbl (%rsi, %rdx), %edx
|
|
Packit |
6c4009 |
subl %edx, %eax
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
VZEROUPPER
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(return_vec_size):
|
|
Packit |
6c4009 |
tzcntl %ecx, %edx
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
/* Return 0 if the mismatched index (%rdx + VEC_SIZE) is after
|
|
Packit |
6c4009 |
the maximum offset (%r11). */
|
|
Packit |
6c4009 |
addq $VEC_SIZE, %rdx
|
|
Packit |
6c4009 |
cmpq %r11, %rdx
|
|
Packit |
6c4009 |
jae L(zero)
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
movl (%rdi, %rdx), %ecx
|
|
Packit |
6c4009 |
cmpl (%rsi, %rdx), %ecx
|
|
Packit |
6c4009 |
jne L(wcscmp_return)
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
movzbl (%rdi, %rdx), %eax
|
|
Packit |
6c4009 |
movzbl (%rsi, %rdx), %edx
|
|
Packit |
6c4009 |
subl %edx, %eax
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
movl VEC_SIZE(%rdi, %rdx), %ecx
|
|
Packit |
6c4009 |
cmpl VEC_SIZE(%rsi, %rdx), %ecx
|
|
Packit |
6c4009 |
jne L(wcscmp_return)
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
movzbl VEC_SIZE(%rdi, %rdx), %eax
|
|
Packit |
6c4009 |
movzbl VEC_SIZE(%rsi, %rdx), %edx
|
|
Packit |
6c4009 |
subl %edx, %eax
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
VZEROUPPER
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(return_2_vec_size):
|
|
Packit |
6c4009 |
tzcntl %ecx, %edx
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
/* Return 0 if the mismatched index (%rdx + 2 * VEC_SIZE) is
|
|
Packit |
6c4009 |
after the maximum offset (%r11). */
|
|
Packit |
6c4009 |
addq $(VEC_SIZE * 2), %rdx
|
|
Packit |
6c4009 |
cmpq %r11, %rdx
|
|
Packit |
6c4009 |
jae L(zero)
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
movl (%rdi, %rdx), %ecx
|
|
Packit |
6c4009 |
cmpl (%rsi, %rdx), %ecx
|
|
Packit |
6c4009 |
jne L(wcscmp_return)
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
movzbl (%rdi, %rdx), %eax
|
|
Packit |
6c4009 |
movzbl (%rsi, %rdx), %edx
|
|
Packit |
6c4009 |
subl %edx, %eax
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
movl (VEC_SIZE * 2)(%rdi, %rdx), %ecx
|
|
Packit |
6c4009 |
cmpl (VEC_SIZE * 2)(%rsi, %rdx), %ecx
|
|
Packit |
6c4009 |
jne L(wcscmp_return)
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
movzbl (VEC_SIZE * 2)(%rdi, %rdx), %eax
|
|
Packit |
6c4009 |
movzbl (VEC_SIZE * 2)(%rsi, %rdx), %edx
|
|
Packit |
6c4009 |
subl %edx, %eax
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
VZEROUPPER
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(return_3_vec_size):
|
|
Packit |
6c4009 |
tzcntl %ecx, %edx
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
/* Return 0 if the mismatched index (%rdx + 3 * VEC_SIZE) is
|
|
Packit |
6c4009 |
after the maximum offset (%r11). */
|
|
Packit |
6c4009 |
addq $(VEC_SIZE * 3), %rdx
|
|
Packit |
6c4009 |
cmpq %r11, %rdx
|
|
Packit |
6c4009 |
jae L(zero)
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
movl (%rdi, %rdx), %ecx
|
|
Packit |
6c4009 |
cmpl (%rsi, %rdx), %ecx
|
|
Packit |
6c4009 |
jne L(wcscmp_return)
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
movzbl (%rdi, %rdx), %eax
|
|
Packit |
6c4009 |
movzbl (%rsi, %rdx), %edx
|
|
Packit |
6c4009 |
subl %edx, %eax
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
movl (VEC_SIZE * 3)(%rdi, %rdx), %ecx
|
|
Packit |
6c4009 |
cmpl (VEC_SIZE * 3)(%rsi, %rdx), %ecx
|
|
Packit |
6c4009 |
jne L(wcscmp_return)
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
movzbl (VEC_SIZE * 3)(%rdi, %rdx), %eax
|
|
Packit |
6c4009 |
movzbl (VEC_SIZE * 3)(%rsi, %rdx), %edx
|
|
Packit |
6c4009 |
subl %edx, %eax
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
VZEROUPPER
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(next_3_vectors):
|
|
Packit |
6c4009 |
vmovdqu VEC_SIZE(%rdi), %ymm6
|
|
Packit |
6c4009 |
VPCMPEQ VEC_SIZE(%rsi), %ymm6, %ymm3
|
|
Packit |
6c4009 |
VPMINU %ymm6, %ymm3, %ymm3
|
|
Packit |
6c4009 |
VPCMPEQ %ymm7, %ymm3, %ymm3
|
|
Packit |
6c4009 |
vpmovmskb %ymm3, %ecx
|
|
Packit |
6c4009 |
testl %ecx, %ecx
|
|
Packit |
6c4009 |
jne L(return_vec_size)
|
|
Packit |
6c4009 |
vmovdqu (VEC_SIZE * 2)(%rdi), %ymm5
|
|
Packit |
6c4009 |
vmovdqu (VEC_SIZE * 3)(%rdi), %ymm4
|
|
Packit |
6c4009 |
vmovdqu (VEC_SIZE * 3)(%rsi), %ymm0
|
|
Packit |
6c4009 |
VPCMPEQ (VEC_SIZE * 2)(%rsi), %ymm5, %ymm2
|
|
Packit |
6c4009 |
VPMINU %ymm5, %ymm2, %ymm2
|
|
Packit |
6c4009 |
VPCMPEQ %ymm4, %ymm0, %ymm0
|
|
Packit |
6c4009 |
VPCMPEQ %ymm7, %ymm2, %ymm2
|
|
Packit |
6c4009 |
vpmovmskb %ymm2, %ecx
|
|
Packit |
6c4009 |
testl %ecx, %ecx
|
|
Packit |
6c4009 |
jne L(return_2_vec_size)
|
|
Packit |
6c4009 |
VPMINU %ymm4, %ymm0, %ymm0
|
|
Packit |
6c4009 |
VPCMPEQ %ymm7, %ymm0, %ymm0
|
|
Packit |
6c4009 |
vpmovmskb %ymm0, %ecx
|
|
Packit |
6c4009 |
testl %ecx, %ecx
|
|
Packit |
6c4009 |
jne L(return_3_vec_size)
|
|
Packit |
6c4009 |
L(main_loop_header):
|
|
Packit |
6c4009 |
leaq (VEC_SIZE * 4)(%rdi), %rdx
|
|
Packit |
6c4009 |
movl $PAGE_SIZE, %ecx
|
|
Packit |
6c4009 |
/* Align load via RAX. */
|
|
Packit |
6c4009 |
andq $-(VEC_SIZE * 4), %rdx
|
|
Packit |
6c4009 |
subq %rdi, %rdx
|
|
Packit |
6c4009 |
leaq (%rdi, %rdx), %rax
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
/* Starting from this point, the maximum offset, or simply the
|
|
Packit |
6c4009 |
'offset', DECREASES by the same amount when base pointers are
|
|
Packit |
6c4009 |
moved forward. Return 0 when:
|
|
Packit |
6c4009 |
1) On match: offset <= the matched vector index.
|
|
Packit |
6c4009 |
2) On mistmach, offset is before the mistmatched index.
|
|
Packit |
6c4009 |
*/
|
|
Packit |
6c4009 |
subq %rdx, %r11
|
|
Packit |
6c4009 |
jbe L(zero)
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
addq %rsi, %rdx
|
|
Packit |
6c4009 |
movq %rdx, %rsi
|
|
Packit |
6c4009 |
andl $(PAGE_SIZE - 1), %esi
|
|
Packit |
6c4009 |
/* Number of bytes before page crossing. */
|
|
Packit |
6c4009 |
subq %rsi, %rcx
|
|
Packit |
6c4009 |
/* Number of VEC_SIZE * 4 blocks before page crossing. */
|
|
Packit |
6c4009 |
shrq $DIVIDE_BY_VEC_4_SHIFT, %rcx
|
|
Packit |
6c4009 |
/* ESI: Number of VEC_SIZE * 4 blocks before page crossing. */
|
|
Packit |
6c4009 |
movl %ecx, %esi
|
|
Packit |
6c4009 |
jmp L(loop_start)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(loop):
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
/* Base pointers are moved forward by 4 * VEC_SIZE. Decrease
|
|
Packit |
6c4009 |
the maximum offset (%r11) by the same amount. */
|
|
Packit |
6c4009 |
subq $(VEC_SIZE * 4), %r11
|
|
Packit |
6c4009 |
jbe L(zero)
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
addq $(VEC_SIZE * 4), %rax
|
|
Packit |
6c4009 |
addq $(VEC_SIZE * 4), %rdx
|
|
Packit |
6c4009 |
L(loop_start):
|
|
Packit |
6c4009 |
testl %esi, %esi
|
|
Packit |
6c4009 |
leal -1(%esi), %esi
|
|
Packit |
6c4009 |
je L(loop_cross_page)
|
|
Packit |
6c4009 |
L(back_to_loop):
|
|
Packit |
6c4009 |
/* Main loop, comparing 4 vectors are a time. */
|
|
Packit |
6c4009 |
vmovdqa (%rax), %ymm0
|
|
Packit |
6c4009 |
vmovdqa VEC_SIZE(%rax), %ymm3
|
|
Packit |
6c4009 |
VPCMPEQ (%rdx), %ymm0, %ymm4
|
|
Packit |
6c4009 |
VPCMPEQ VEC_SIZE(%rdx), %ymm3, %ymm1
|
|
Packit |
6c4009 |
VPMINU %ymm0, %ymm4, %ymm4
|
|
Packit |
6c4009 |
VPMINU %ymm3, %ymm1, %ymm1
|
|
Packit |
6c4009 |
vmovdqa (VEC_SIZE * 2)(%rax), %ymm2
|
|
Packit |
6c4009 |
VPMINU %ymm1, %ymm4, %ymm0
|
|
Packit |
6c4009 |
vmovdqa (VEC_SIZE * 3)(%rax), %ymm3
|
|
Packit |
6c4009 |
VPCMPEQ (VEC_SIZE * 2)(%rdx), %ymm2, %ymm5
|
|
Packit |
6c4009 |
VPCMPEQ (VEC_SIZE * 3)(%rdx), %ymm3, %ymm6
|
|
Packit |
6c4009 |
VPMINU %ymm2, %ymm5, %ymm5
|
|
Packit |
6c4009 |
VPMINU %ymm3, %ymm6, %ymm6
|
|
Packit |
6c4009 |
VPMINU %ymm5, %ymm0, %ymm0
|
|
Packit |
6c4009 |
VPMINU %ymm6, %ymm0, %ymm0
|
|
Packit |
6c4009 |
VPCMPEQ %ymm7, %ymm0, %ymm0
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Test each mask (32 bits) individually because for VEC_SIZE
|
|
Packit |
6c4009 |
== 32 is not possible to OR the four masks and keep all bits
|
|
Packit |
6c4009 |
in a 64-bit integer register, differing from SSE2 strcmp
|
|
Packit |
6c4009 |
where ORing is possible. */
|
|
Packit |
6c4009 |
vpmovmskb %ymm0, %ecx
|
|
Packit |
6c4009 |
testl %ecx, %ecx
|
|
Packit |
6c4009 |
je L(loop)
|
|
Packit |
6c4009 |
VPCMPEQ %ymm7, %ymm4, %ymm0
|
|
Packit |
6c4009 |
vpmovmskb %ymm0, %edi
|
|
Packit |
6c4009 |
testl %edi, %edi
|
|
Packit |
6c4009 |
je L(test_vec)
|
|
Packit |
6c4009 |
tzcntl %edi, %ecx
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
cmpq %rcx, %r11
|
|
Packit |
6c4009 |
jbe L(zero)
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
movq %rax, %rsi
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
movl (%rsi, %rcx), %edi
|
|
Packit |
6c4009 |
cmpl (%rdx, %rcx), %edi
|
|
Packit |
6c4009 |
jne L(wcscmp_return)
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
movzbl (%rax, %rcx), %eax
|
|
Packit |
6c4009 |
movzbl (%rdx, %rcx), %edx
|
|
Packit |
6c4009 |
subl %edx, %eax
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
movq %rax, %rsi
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
movl (%rsi, %rcx), %edi
|
|
Packit |
6c4009 |
cmpl (%rdx, %rcx), %edi
|
|
Packit |
6c4009 |
jne L(wcscmp_return)
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
movzbl (%rax, %rcx), %eax
|
|
Packit |
6c4009 |
movzbl (%rdx, %rcx), %edx
|
|
Packit |
6c4009 |
subl %edx, %eax
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
VZEROUPPER
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(test_vec):
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
/* The first vector matched. Return 0 if the maximum offset
|
|
Packit |
6c4009 |
(%r11) <= VEC_SIZE. */
|
|
Packit |
6c4009 |
cmpq $VEC_SIZE, %r11
|
|
Packit |
6c4009 |
jbe L(zero)
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
VPCMPEQ %ymm7, %ymm1, %ymm1
|
|
Packit |
6c4009 |
vpmovmskb %ymm1, %ecx
|
|
Packit |
6c4009 |
testl %ecx, %ecx
|
|
Packit |
6c4009 |
je L(test_2_vec)
|
|
Packit |
6c4009 |
tzcntl %ecx, %edi
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
addq $VEC_SIZE, %rdi
|
|
Packit |
6c4009 |
cmpq %rdi, %r11
|
|
Packit |
6c4009 |
jbe L(zero)
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
movq %rax, %rsi
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
movl (%rsi, %rdi), %ecx
|
|
Packit |
6c4009 |
cmpl (%rdx, %rdi), %ecx
|
|
Packit |
6c4009 |
jne L(wcscmp_return)
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
movzbl (%rax, %rdi), %eax
|
|
Packit |
6c4009 |
movzbl (%rdx, %rdi), %edx
|
|
Packit |
6c4009 |
subl %edx, %eax
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
movq %rax, %rsi
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
movl VEC_SIZE(%rsi, %rdi), %ecx
|
|
Packit |
6c4009 |
cmpl VEC_SIZE(%rdx, %rdi), %ecx
|
|
Packit |
6c4009 |
jne L(wcscmp_return)
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
movzbl VEC_SIZE(%rax, %rdi), %eax
|
|
Packit |
6c4009 |
movzbl VEC_SIZE(%rdx, %rdi), %edx
|
|
Packit |
6c4009 |
subl %edx, %eax
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
VZEROUPPER
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(test_2_vec):
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
/* The first 2 vectors matched. Return 0 if the maximum offset
|
|
Packit |
6c4009 |
(%r11) <= 2 * VEC_SIZE. */
|
|
Packit |
6c4009 |
cmpq $(VEC_SIZE * 2), %r11
|
|
Packit |
6c4009 |
jbe L(zero)
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
VPCMPEQ %ymm7, %ymm5, %ymm5
|
|
Packit |
6c4009 |
vpmovmskb %ymm5, %ecx
|
|
Packit |
6c4009 |
testl %ecx, %ecx
|
|
Packit |
6c4009 |
je L(test_3_vec)
|
|
Packit |
6c4009 |
tzcntl %ecx, %edi
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
addq $(VEC_SIZE * 2), %rdi
|
|
Packit |
6c4009 |
cmpq %rdi, %r11
|
|
Packit |
6c4009 |
jbe L(zero)
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
movq %rax, %rsi
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
movl (%rsi, %rdi), %ecx
|
|
Packit |
6c4009 |
cmpl (%rdx, %rdi), %ecx
|
|
Packit |
6c4009 |
jne L(wcscmp_return)
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
movzbl (%rax, %rdi), %eax
|
|
Packit |
6c4009 |
movzbl (%rdx, %rdi), %edx
|
|
Packit |
6c4009 |
subl %edx, %eax
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
movq %rax, %rsi
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
movl (VEC_SIZE * 2)(%rsi, %rdi), %ecx
|
|
Packit |
6c4009 |
cmpl (VEC_SIZE * 2)(%rdx, %rdi), %ecx
|
|
Packit |
6c4009 |
jne L(wcscmp_return)
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
movzbl (VEC_SIZE * 2)(%rax, %rdi), %eax
|
|
Packit |
6c4009 |
movzbl (VEC_SIZE * 2)(%rdx, %rdi), %edx
|
|
Packit |
6c4009 |
subl %edx, %eax
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
VZEROUPPER
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(test_3_vec):
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
/* The first 3 vectors matched. Return 0 if the maximum offset
|
|
Packit |
6c4009 |
(%r11) <= 3 * VEC_SIZE. */
|
|
Packit |
6c4009 |
cmpq $(VEC_SIZE * 3), %r11
|
|
Packit |
6c4009 |
jbe L(zero)
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
VPCMPEQ %ymm7, %ymm6, %ymm6
|
|
Packit |
6c4009 |
vpmovmskb %ymm6, %esi
|
|
Packit |
6c4009 |
tzcntl %esi, %ecx
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
addq $(VEC_SIZE * 3), %rcx
|
|
Packit |
6c4009 |
cmpq %rcx, %r11
|
|
Packit |
6c4009 |
jbe L(zero)
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
movq %rax, %rsi
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
movl (%rsi, %rcx), %esi
|
|
Packit |
6c4009 |
cmpl (%rdx, %rcx), %esi
|
|
Packit |
6c4009 |
jne L(wcscmp_return)
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
movzbl (%rax, %rcx), %eax
|
|
Packit |
6c4009 |
movzbl (%rdx, %rcx), %edx
|
|
Packit |
6c4009 |
subl %edx, %eax
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
movq %rax, %rsi
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
movl (VEC_SIZE * 3)(%rsi, %rcx), %esi
|
|
Packit |
6c4009 |
cmpl (VEC_SIZE * 3)(%rdx, %rcx), %esi
|
|
Packit |
6c4009 |
jne L(wcscmp_return)
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
movzbl (VEC_SIZE * 3)(%rax, %rcx), %eax
|
|
Packit |
6c4009 |
movzbl (VEC_SIZE * 3)(%rdx, %rcx), %edx
|
|
Packit |
6c4009 |
subl %edx, %eax
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
VZEROUPPER
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(loop_cross_page):
|
|
Packit |
6c4009 |
xorl %r10d, %r10d
|
|
Packit |
6c4009 |
movq %rdx, %rcx
|
|
Packit |
6c4009 |
/* Align load via RDX. We load the extra ECX bytes which should
|
|
Packit |
6c4009 |
be ignored. */
|
|
Packit |
6c4009 |
andl $((VEC_SIZE * 4) - 1), %ecx
|
|
Packit |
6c4009 |
/* R10 is -RCX. */
|
|
Packit |
6c4009 |
subq %rcx, %r10
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* This works only if VEC_SIZE * 2 == 64. */
|
|
Packit |
6c4009 |
# if (VEC_SIZE * 2) != 64
|
|
Packit |
6c4009 |
# error (VEC_SIZE * 2) != 64
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Check if the first VEC_SIZE * 2 bytes should be ignored. */
|
|
Packit |
6c4009 |
cmpl $(VEC_SIZE * 2), %ecx
|
|
Packit |
6c4009 |
jge L(loop_cross_page_2_vec)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
vmovdqu (%rax, %r10), %ymm2
|
|
Packit |
6c4009 |
vmovdqu VEC_SIZE(%rax, %r10), %ymm3
|
|
Packit |
6c4009 |
VPCMPEQ (%rdx, %r10), %ymm2, %ymm0
|
|
Packit |
6c4009 |
VPCMPEQ VEC_SIZE(%rdx, %r10), %ymm3, %ymm1
|
|
Packit |
6c4009 |
VPMINU %ymm2, %ymm0, %ymm0
|
|
Packit |
6c4009 |
VPMINU %ymm3, %ymm1, %ymm1
|
|
Packit |
6c4009 |
VPCMPEQ %ymm7, %ymm0, %ymm0
|
|
Packit |
6c4009 |
VPCMPEQ %ymm7, %ymm1, %ymm1
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
vpmovmskb %ymm0, %edi
|
|
Packit |
6c4009 |
vpmovmskb %ymm1, %esi
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
salq $32, %rsi
|
|
Packit |
6c4009 |
xorq %rsi, %rdi
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Since ECX < VEC_SIZE * 2, simply skip the first ECX bytes. */
|
|
Packit |
6c4009 |
shrq %cl, %rdi
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
testq %rdi, %rdi
|
|
Packit |
6c4009 |
je L(loop_cross_page_2_vec)
|
|
Packit |
6c4009 |
tzcntq %rdi, %rcx
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
cmpq %rcx, %r11
|
|
Packit |
6c4009 |
jbe L(zero)
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
movq %rax, %rsi
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
movl (%rsi, %rcx), %edi
|
|
Packit |
6c4009 |
cmpl (%rdx, %rcx), %edi
|
|
Packit |
6c4009 |
jne L(wcscmp_return)
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
movzbl (%rax, %rcx), %eax
|
|
Packit |
6c4009 |
movzbl (%rdx, %rcx), %edx
|
|
Packit |
6c4009 |
subl %edx, %eax
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
movq %rax, %rsi
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
movl (%rsi, %rcx), %edi
|
|
Packit |
6c4009 |
cmpl (%rdx, %rcx), %edi
|
|
Packit |
6c4009 |
jne L(wcscmp_return)
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
movzbl (%rax, %rcx), %eax
|
|
Packit |
6c4009 |
movzbl (%rdx, %rcx), %edx
|
|
Packit |
6c4009 |
subl %edx, %eax
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
VZEROUPPER
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(loop_cross_page_2_vec):
|
|
Packit |
6c4009 |
/* The first VEC_SIZE * 2 bytes match or are ignored. */
|
|
Packit |
6c4009 |
vmovdqu (VEC_SIZE * 2)(%rax, %r10), %ymm2
|
|
Packit |
6c4009 |
vmovdqu (VEC_SIZE * 3)(%rax, %r10), %ymm3
|
|
Packit |
6c4009 |
VPCMPEQ (VEC_SIZE * 2)(%rdx, %r10), %ymm2, %ymm5
|
|
Packit |
6c4009 |
VPMINU %ymm2, %ymm5, %ymm5
|
|
Packit |
6c4009 |
VPCMPEQ (VEC_SIZE * 3)(%rdx, %r10), %ymm3, %ymm6
|
|
Packit |
6c4009 |
VPCMPEQ %ymm7, %ymm5, %ymm5
|
|
Packit |
6c4009 |
VPMINU %ymm3, %ymm6, %ymm6
|
|
Packit |
6c4009 |
VPCMPEQ %ymm7, %ymm6, %ymm6
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
vpmovmskb %ymm5, %edi
|
|
Packit |
6c4009 |
vpmovmskb %ymm6, %esi
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
salq $32, %rsi
|
|
Packit |
6c4009 |
xorq %rsi, %rdi
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
xorl %r8d, %r8d
|
|
Packit |
6c4009 |
/* If ECX > VEC_SIZE * 2, skip ECX - (VEC_SIZE * 2) bytes. */
|
|
Packit |
6c4009 |
subl $(VEC_SIZE * 2), %ecx
|
|
Packit |
6c4009 |
jle 1f
|
|
Packit |
6c4009 |
/* Skip ECX bytes. */
|
|
Packit |
6c4009 |
shrq %cl, %rdi
|
|
Packit |
6c4009 |
/* R8 has number of bytes skipped. */
|
|
Packit |
6c4009 |
movl %ecx, %r8d
|
|
Packit |
6c4009 |
1:
|
|
Packit |
6c4009 |
/* Before jumping back to the loop, set ESI to the number of
|
|
Packit |
6c4009 |
VEC_SIZE * 4 blocks before page crossing. */
|
|
Packit |
6c4009 |
movl $(PAGE_SIZE / (VEC_SIZE * 4) - 1), %esi
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
testq %rdi, %rdi
|
|
Packit Service |
5962e3 |
# ifdef USE_AS_STRNCMP
|
|
Packit Service |
5962e3 |
/* At this point, if %rdi value is 0, it already tested
|
|
Packit Service |
5962e3 |
VEC_SIZE*4+%r10 byte starting from %rax. This label
|
|
Packit Service |
5962e3 |
checks whether strncmp maximum offset reached or not. */
|
|
Packit Service |
5962e3 |
je L(string_nbyte_offset_check)
|
|
Packit Service |
5962e3 |
# else
|
|
Packit |
6c4009 |
je L(back_to_loop)
|
|
Packit Service |
5962e3 |
# endif
|
|
Packit |
6c4009 |
tzcntq %rdi, %rcx
|
|
Packit |
6c4009 |
addq %r10, %rcx
|
|
Packit |
6c4009 |
/* Adjust for number of bytes skipped. */
|
|
Packit |
6c4009 |
addq %r8, %rcx
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
addq $(VEC_SIZE * 2), %rcx
|
|
Packit |
6c4009 |
subq %rcx, %r11
|
|
Packit |
6c4009 |
jbe L(zero)
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
movq %rax, %rsi
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
movl (%rsi, %rcx), %edi
|
|
Packit |
6c4009 |
cmpl (%rdx, %rcx), %edi
|
|
Packit |
6c4009 |
jne L(wcscmp_return)
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
movzbl (%rax, %rcx), %eax
|
|
Packit |
6c4009 |
movzbl (%rdx, %rcx), %edx
|
|
Packit |
6c4009 |
subl %edx, %eax
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
movq %rax, %rsi
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
movl (VEC_SIZE * 2)(%rsi, %rcx), %edi
|
|
Packit |
6c4009 |
cmpl (VEC_SIZE * 2)(%rdx, %rcx), %edi
|
|
Packit |
6c4009 |
jne L(wcscmp_return)
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
movzbl (VEC_SIZE * 2)(%rax, %rcx), %eax
|
|
Packit |
6c4009 |
movzbl (VEC_SIZE * 2)(%rdx, %rcx), %edx
|
|
Packit |
6c4009 |
subl %edx, %eax
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
VZEROUPPER
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit Service |
5962e3 |
# ifdef USE_AS_STRNCMP
|
|
Packit Service |
5962e3 |
L(string_nbyte_offset_check):
|
|
Packit Service |
5962e3 |
leaq (VEC_SIZE * 4)(%r10), %r10
|
|
Packit Service |
5962e3 |
cmpq %r10, %r11
|
|
Packit Service |
5962e3 |
jbe L(zero)
|
|
Packit Service |
5962e3 |
jmp L(back_to_loop)
|
|
Packit Service |
5962e3 |
# endif
|
|
Packit Service |
5962e3 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(cross_page_loop):
|
|
Packit |
6c4009 |
/* Check one byte/dword at a time. */
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
cmpl %ecx, %eax
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
subl %ecx, %eax
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
jne L(different)
|
|
Packit |
6c4009 |
addl $SIZE_OF_CHAR, %edx
|
|
Packit |
6c4009 |
cmpl $(VEC_SIZE * 4), %edx
|
|
Packit |
6c4009 |
je L(main_loop_header)
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
cmpq %r11, %rdx
|
|
Packit |
6c4009 |
jae L(zero)
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
movl (%rdi, %rdx), %eax
|
|
Packit |
6c4009 |
movl (%rsi, %rdx), %ecx
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
movzbl (%rdi, %rdx), %eax
|
|
Packit |
6c4009 |
movzbl (%rsi, %rdx), %ecx
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
/* Check null char. */
|
|
Packit |
6c4009 |
testl %eax, %eax
|
|
Packit |
6c4009 |
jne L(cross_page_loop)
|
|
Packit |
6c4009 |
/* Since %eax == 0, subtract is OK for both SIGNED and UNSIGNED
|
|
Packit |
6c4009 |
comparisons. */
|
|
Packit |
6c4009 |
subl %ecx, %eax
|
|
Packit |
6c4009 |
# ifndef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
L(different):
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
VZEROUPPER
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(different):
|
|
Packit |
6c4009 |
/* Use movl to avoid modifying EFLAGS. */
|
|
Packit |
6c4009 |
movl $0, %eax
|
|
Packit |
6c4009 |
setl %al
|
|
Packit |
6c4009 |
negl %eax
|
|
Packit |
6c4009 |
orl $1, %eax
|
|
Packit |
6c4009 |
VZEROUPPER
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(zero):
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
VZEROUPPER
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(char0):
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
movl (%rdi), %ecx
|
|
Packit |
6c4009 |
cmpl (%rsi), %ecx
|
|
Packit |
6c4009 |
jne L(wcscmp_return)
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
movzbl (%rsi), %ecx
|
|
Packit |
6c4009 |
movzbl (%rdi), %eax
|
|
Packit |
6c4009 |
subl %ecx, %eax
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
VZEROUPPER
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(last_vector):
|
|
Packit |
6c4009 |
addq %rdx, %rdi
|
|
Packit |
6c4009 |
addq %rdx, %rsi
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
subq %rdx, %r11
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
tzcntl %ecx, %edx
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
cmpq %r11, %rdx
|
|
Packit |
6c4009 |
jae L(zero)
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
movl (%rdi, %rdx), %ecx
|
|
Packit |
6c4009 |
cmpl (%rsi, %rdx), %ecx
|
|
Packit |
6c4009 |
jne L(wcscmp_return)
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
movzbl (%rdi, %rdx), %eax
|
|
Packit |
6c4009 |
movzbl (%rsi, %rdx), %edx
|
|
Packit |
6c4009 |
subl %edx, %eax
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
VZEROUPPER
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Comparing on page boundary region requires special treatment:
|
|
Packit |
6c4009 |
It must done one vector at the time, starting with the wider
|
|
Packit |
6c4009 |
ymm vector if possible, if not, with xmm. If fetching 16 bytes
|
|
Packit |
6c4009 |
(xmm) still passes the boundary, byte comparison must be done.
|
|
Packit |
6c4009 |
*/
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(cross_page):
|
|
Packit |
6c4009 |
/* Try one ymm vector at a time. */
|
|
Packit |
6c4009 |
cmpl $(PAGE_SIZE - VEC_SIZE), %eax
|
|
Packit |
6c4009 |
jg L(cross_page_1_vector)
|
|
Packit |
6c4009 |
L(loop_1_vector):
|
|
Packit |
6c4009 |
vmovdqu (%rdi, %rdx), %ymm1
|
|
Packit |
6c4009 |
VPCMPEQ (%rsi, %rdx), %ymm1, %ymm0
|
|
Packit |
6c4009 |
VPMINU %ymm1, %ymm0, %ymm0
|
|
Packit |
6c4009 |
VPCMPEQ %ymm7, %ymm0, %ymm0
|
|
Packit |
6c4009 |
vpmovmskb %ymm0, %ecx
|
|
Packit |
6c4009 |
testl %ecx, %ecx
|
|
Packit |
6c4009 |
jne L(last_vector)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
addl $VEC_SIZE, %edx
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
addl $VEC_SIZE, %eax
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
/* Return 0 if the current offset (%rdx) >= the maximum offset
|
|
Packit |
6c4009 |
(%r11). */
|
|
Packit |
6c4009 |
cmpq %r11, %rdx
|
|
Packit |
6c4009 |
jae L(zero)
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
cmpl $(PAGE_SIZE - VEC_SIZE), %eax
|
|
Packit |
6c4009 |
jle L(loop_1_vector)
|
|
Packit |
6c4009 |
L(cross_page_1_vector):
|
|
Packit |
6c4009 |
/* Less than 32 bytes to check, try one xmm vector. */
|
|
Packit |
6c4009 |
cmpl $(PAGE_SIZE - 16), %eax
|
|
Packit |
6c4009 |
jg L(cross_page_1_xmm)
|
|
Packit |
6c4009 |
vmovdqu (%rdi, %rdx), %xmm1
|
|
Packit |
6c4009 |
VPCMPEQ (%rsi, %rdx), %xmm1, %xmm0
|
|
Packit |
6c4009 |
VPMINU %xmm1, %xmm0, %xmm0
|
|
Packit |
6c4009 |
VPCMPEQ %xmm7, %xmm0, %xmm0
|
|
Packit |
6c4009 |
vpmovmskb %xmm0, %ecx
|
|
Packit |
6c4009 |
testl %ecx, %ecx
|
|
Packit |
6c4009 |
jne L(last_vector)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
addl $16, %edx
|
|
Packit |
6c4009 |
# ifndef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
addl $16, %eax
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
/* Return 0 if the current offset (%rdx) >= the maximum offset
|
|
Packit |
6c4009 |
(%r11). */
|
|
Packit |
6c4009 |
cmpq %r11, %rdx
|
|
Packit |
6c4009 |
jae L(zero)
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
L(cross_page_1_xmm):
|
|
Packit |
6c4009 |
# ifndef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
/* Less than 16 bytes to check, try 8 byte vector. NB: No need
|
|
Packit |
6c4009 |
for wcscmp nor wcsncmp since wide char is 4 bytes. */
|
|
Packit |
6c4009 |
cmpl $(PAGE_SIZE - 8), %eax
|
|
Packit |
6c4009 |
jg L(cross_page_8bytes)
|
|
Packit |
6c4009 |
vmovq (%rdi, %rdx), %xmm1
|
|
Packit |
6c4009 |
vmovq (%rsi, %rdx), %xmm0
|
|
Packit |
6c4009 |
VPCMPEQ %xmm0, %xmm1, %xmm0
|
|
Packit |
6c4009 |
VPMINU %xmm1, %xmm0, %xmm0
|
|
Packit |
6c4009 |
VPCMPEQ %xmm7, %xmm0, %xmm0
|
|
Packit |
6c4009 |
vpmovmskb %xmm0, %ecx
|
|
Packit |
6c4009 |
/* Only last 8 bits are valid. */
|
|
Packit |
6c4009 |
andl $0xff, %ecx
|
|
Packit |
6c4009 |
testl %ecx, %ecx
|
|
Packit |
6c4009 |
jne L(last_vector)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
addl $8, %edx
|
|
Packit |
6c4009 |
addl $8, %eax
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
/* Return 0 if the current offset (%rdx) >= the maximum offset
|
|
Packit |
6c4009 |
(%r11). */
|
|
Packit |
6c4009 |
cmpq %r11, %rdx
|
|
Packit |
6c4009 |
jae L(zero)
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
L(cross_page_8bytes):
|
|
Packit |
6c4009 |
/* Less than 8 bytes to check, try 4 byte vector. */
|
|
Packit |
6c4009 |
cmpl $(PAGE_SIZE - 4), %eax
|
|
Packit |
6c4009 |
jg L(cross_page_4bytes)
|
|
Packit |
6c4009 |
vmovd (%rdi, %rdx), %xmm1
|
|
Packit |
6c4009 |
vmovd (%rsi, %rdx), %xmm0
|
|
Packit |
6c4009 |
VPCMPEQ %xmm0, %xmm1, %xmm0
|
|
Packit |
6c4009 |
VPMINU %xmm1, %xmm0, %xmm0
|
|
Packit |
6c4009 |
VPCMPEQ %xmm7, %xmm0, %xmm0
|
|
Packit |
6c4009 |
vpmovmskb %xmm0, %ecx
|
|
Packit |
6c4009 |
/* Only last 4 bits are valid. */
|
|
Packit |
6c4009 |
andl $0xf, %ecx
|
|
Packit |
6c4009 |
testl %ecx, %ecx
|
|
Packit |
6c4009 |
jne L(last_vector)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
addl $4, %edx
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
/* Return 0 if the current offset (%rdx) >= the maximum offset
|
|
Packit |
6c4009 |
(%r11). */
|
|
Packit |
6c4009 |
cmpq %r11, %rdx
|
|
Packit |
6c4009 |
jae L(zero)
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
L(cross_page_4bytes):
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
/* Less than 4 bytes to check, try one byte/dword at a time. */
|
|
Packit |
6c4009 |
# ifdef USE_AS_STRNCMP
|
|
Packit |
6c4009 |
cmpq %r11, %rdx
|
|
Packit |
6c4009 |
jae L(zero)
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
# ifdef USE_AS_WCSCMP
|
|
Packit |
6c4009 |
movl (%rdi, %rdx), %eax
|
|
Packit |
6c4009 |
movl (%rsi, %rdx), %ecx
|
|
Packit |
6c4009 |
# else
|
|
Packit |
6c4009 |
movzbl (%rdi, %rdx), %eax
|
|
Packit |
6c4009 |
movzbl (%rsi, %rdx), %ecx
|
|
Packit |
6c4009 |
# endif
|
|
Packit |
6c4009 |
testl %eax, %eax
|
|
Packit |
6c4009 |
jne L(cross_page_loop)
|
|
Packit |
6c4009 |
subl %ecx, %eax
|
|
Packit |
6c4009 |
VZEROUPPER
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
END (STRCMP)
|
|
Packit |
6c4009 |
#endif
|