Blame sysdeps/x86_64/strchr.S
|
Packit |
6c4009 |
/* strchr (str, ch) -- Return pointer to first occurrence of CH in STR.
|
|
Packit |
6c4009 |
For AMD x86-64.
|
|
Packit |
6c4009 |
Copyright (C) 2009-2018 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 |
#include <sysdep.h>
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.text
|
|
Packit |
6c4009 |
ENTRY (strchr)
|
|
Packit |
6c4009 |
movd %esi, %xmm1
|
|
Packit |
6c4009 |
movl %edi, %eax
|
|
Packit |
6c4009 |
andl $4095, %eax
|
|
Packit |
6c4009 |
punpcklbw %xmm1, %xmm1
|
|
Packit |
6c4009 |
cmpl $4032, %eax
|
|
Packit |
6c4009 |
punpcklwd %xmm1, %xmm1
|
|
Packit |
6c4009 |
pshufd $0, %xmm1, %xmm1
|
|
Packit |
6c4009 |
jg L(cross_page)
|
|
Packit |
6c4009 |
movdqu (%rdi), %xmm0
|
|
Packit |
6c4009 |
pxor %xmm3, %xmm3
|
|
Packit |
6c4009 |
movdqa %xmm0, %xmm4
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm0
|
|
Packit |
6c4009 |
pcmpeqb %xmm3, %xmm4
|
|
Packit |
6c4009 |
por %xmm4, %xmm0
|
|
Packit |
6c4009 |
pmovmskb %xmm0, %eax
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
je L(next_48_bytes)
|
|
Packit |
6c4009 |
bsf %eax, %eax
|
|
Packit |
6c4009 |
#ifdef AS_STRCHRNUL
|
|
Packit |
6c4009 |
leaq (%rdi,%rax), %rax
|
|
Packit |
6c4009 |
#else
|
|
Packit |
6c4009 |
movl $0, %edx
|
|
Packit |
6c4009 |
leaq (%rdi,%rax), %rax
|
|
Packit |
6c4009 |
cmpb %sil, (%rax)
|
|
Packit |
6c4009 |
cmovne %rdx, %rax
|
|
Packit |
6c4009 |
#endif
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 3
|
|
Packit |
6c4009 |
L(next_48_bytes):
|
|
Packit |
6c4009 |
movdqu 16(%rdi), %xmm0
|
|
Packit |
6c4009 |
movdqa %xmm0, %xmm4
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm0
|
|
Packit |
6c4009 |
pcmpeqb %xmm3, %xmm4
|
|
Packit |
6c4009 |
por %xmm4, %xmm0
|
|
Packit |
6c4009 |
pmovmskb %xmm0, %ecx
|
|
Packit |
6c4009 |
movdqu 32(%rdi), %xmm0
|
|
Packit |
6c4009 |
movdqa %xmm0, %xmm4
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm0
|
|
Packit |
6c4009 |
salq $16, %rcx
|
|
Packit |
6c4009 |
pcmpeqb %xmm3, %xmm4
|
|
Packit |
6c4009 |
por %xmm4, %xmm0
|
|
Packit |
6c4009 |
pmovmskb %xmm0, %eax
|
|
Packit |
6c4009 |
movdqu 48(%rdi), %xmm0
|
|
Packit |
6c4009 |
pcmpeqb %xmm0, %xmm3
|
|
Packit |
6c4009 |
salq $32, %rax
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm0
|
|
Packit |
6c4009 |
orq %rcx, %rax
|
|
Packit |
6c4009 |
por %xmm3, %xmm0
|
|
Packit |
6c4009 |
pmovmskb %xmm0, %ecx
|
|
Packit |
6c4009 |
salq $48, %rcx
|
|
Packit |
6c4009 |
orq %rcx, %rax
|
|
Packit |
6c4009 |
testq %rax, %rax
|
|
Packit |
6c4009 |
jne L(return)
|
|
Packit |
6c4009 |
L(loop_start):
|
|
Packit |
6c4009 |
/* We use this alignment to force loop be aligned to 8 but not
|
|
Packit |
6c4009 |
16 bytes. This gives better sheduling on AMD processors. */
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
pxor %xmm6, %xmm6
|
|
Packit |
6c4009 |
andq $-64, %rdi
|
|
Packit |
6c4009 |
.p2align 3
|
|
Packit |
6c4009 |
L(loop64):
|
|
Packit |
6c4009 |
addq $64, %rdi
|
|
Packit |
6c4009 |
movdqa (%rdi), %xmm5
|
|
Packit |
6c4009 |
movdqa 16(%rdi), %xmm2
|
|
Packit |
6c4009 |
movdqa 32(%rdi), %xmm3
|
|
Packit |
6c4009 |
pxor %xmm1, %xmm5
|
|
Packit |
6c4009 |
movdqa 48(%rdi), %xmm4
|
|
Packit |
6c4009 |
pxor %xmm1, %xmm2
|
|
Packit |
6c4009 |
pxor %xmm1, %xmm3
|
|
Packit |
6c4009 |
pminub (%rdi), %xmm5
|
|
Packit |
6c4009 |
pxor %xmm1, %xmm4
|
|
Packit |
6c4009 |
pminub 16(%rdi), %xmm2
|
|
Packit |
6c4009 |
pminub 32(%rdi), %xmm3
|
|
Packit |
6c4009 |
pminub %xmm2, %xmm5
|
|
Packit |
6c4009 |
pminub 48(%rdi), %xmm4
|
|
Packit |
6c4009 |
pminub %xmm3, %xmm5
|
|
Packit |
6c4009 |
pminub %xmm4, %xmm5
|
|
Packit |
6c4009 |
pcmpeqb %xmm6, %xmm5
|
|
Packit |
6c4009 |
pmovmskb %xmm5, %eax
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
testl %eax, %eax
|
|
Packit |
6c4009 |
je L(loop64)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movdqa (%rdi), %xmm5
|
|
Packit |
6c4009 |
movdqa %xmm5, %xmm0
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm5
|
|
Packit |
6c4009 |
pcmpeqb %xmm6, %xmm0
|
|
Packit |
6c4009 |
por %xmm0, %xmm5
|
|
Packit |
6c4009 |
pcmpeqb %xmm6, %xmm2
|
|
Packit |
6c4009 |
pcmpeqb %xmm6, %xmm3
|
|
Packit |
6c4009 |
pcmpeqb %xmm6, %xmm4
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
pmovmskb %xmm5, %ecx
|
|
Packit |
6c4009 |
pmovmskb %xmm2, %eax
|
|
Packit |
6c4009 |
salq $16, %rax
|
|
Packit |
6c4009 |
pmovmskb %xmm3, %r8d
|
|
Packit |
6c4009 |
pmovmskb %xmm4, %edx
|
|
Packit |
6c4009 |
salq $32, %r8
|
|
Packit |
6c4009 |
orq %r8, %rax
|
|
Packit |
6c4009 |
orq %rcx, %rax
|
|
Packit |
6c4009 |
salq $48, %rdx
|
|
Packit |
6c4009 |
orq %rdx, %rax
|
|
Packit |
6c4009 |
.p2align 3
|
|
Packit |
6c4009 |
L(return):
|
|
Packit |
6c4009 |
bsfq %rax, %rax
|
|
Packit |
6c4009 |
#ifdef AS_STRCHRNUL
|
|
Packit |
6c4009 |
leaq (%rdi,%rax), %rax
|
|
Packit |
6c4009 |
#else
|
|
Packit |
6c4009 |
movl $0, %edx
|
|
Packit |
6c4009 |
leaq (%rdi,%rax), %rax
|
|
Packit |
6c4009 |
cmpb %sil, (%rax)
|
|
Packit |
6c4009 |
cmovne %rdx, %rax
|
|
Packit |
6c4009 |
#endif
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
L(cross_page):
|
|
Packit |
6c4009 |
movq %rdi, %rdx
|
|
Packit |
6c4009 |
pxor %xmm2, %xmm2
|
|
Packit |
6c4009 |
andq $-64, %rdx
|
|
Packit |
6c4009 |
movdqa %xmm1, %xmm0
|
|
Packit |
6c4009 |
movdqa (%rdx), %xmm3
|
|
Packit |
6c4009 |
movdqa %xmm3, %xmm4
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm3
|
|
Packit |
6c4009 |
pcmpeqb %xmm2, %xmm4
|
|
Packit |
6c4009 |
por %xmm4, %xmm3
|
|
Packit |
6c4009 |
pmovmskb %xmm3, %r8d
|
|
Packit |
6c4009 |
movdqa 16(%rdx), %xmm3
|
|
Packit |
6c4009 |
movdqa %xmm3, %xmm4
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm3
|
|
Packit |
6c4009 |
pcmpeqb %xmm2, %xmm4
|
|
Packit |
6c4009 |
por %xmm4, %xmm3
|
|
Packit |
6c4009 |
pmovmskb %xmm3, %eax
|
|
Packit |
6c4009 |
movdqa 32(%rdx), %xmm3
|
|
Packit |
6c4009 |
movdqa %xmm3, %xmm4
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm3
|
|
Packit |
6c4009 |
salq $16, %rax
|
|
Packit |
6c4009 |
pcmpeqb %xmm2, %xmm4
|
|
Packit |
6c4009 |
por %xmm4, %xmm3
|
|
Packit |
6c4009 |
pmovmskb %xmm3, %r9d
|
|
Packit |
6c4009 |
movdqa 48(%rdx), %xmm3
|
|
Packit |
6c4009 |
pcmpeqb %xmm3, %xmm2
|
|
Packit |
6c4009 |
salq $32, %r9
|
|
Packit |
6c4009 |
pcmpeqb %xmm3, %xmm0
|
|
Packit |
6c4009 |
orq %r9, %rax
|
|
Packit |
6c4009 |
orq %r8, %rax
|
|
Packit |
6c4009 |
por %xmm2, %xmm0
|
|
Packit |
6c4009 |
pmovmskb %xmm0, %ecx
|
|
Packit |
6c4009 |
salq $48, %rcx
|
|
Packit |
6c4009 |
orq %rcx, %rax
|
|
Packit |
6c4009 |
movl %edi, %ecx
|
|
Packit |
6c4009 |
subb %dl, %cl
|
|
Packit |
6c4009 |
shrq %cl, %rax
|
|
Packit |
6c4009 |
testq %rax, %rax
|
|
Packit |
6c4009 |
jne L(return)
|
|
Packit |
6c4009 |
jmp L(loop_start)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
END (strchr)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
#ifndef AS_STRCHRNUL
|
|
Packit |
6c4009 |
weak_alias (strchr, index)
|
|
Packit |
6c4009 |
libc_hidden_builtin_def (strchr)
|
|
Packit |
6c4009 |
#endif
|