LoongArch: Add ifunc support for memchr{aligned, lsx, lasx}

According to glibc memchr microbenchmark, this implementation could reduce
the runtime as following:

Name               Percent of runtime reduced
memchr-lasx        37%-83%
memchr-lsx         30%-66%
memchr-aligned     0%-15%
This commit is contained in:
dengjianbo 2023-08-28 10:08:36 +08:00 committed by caiyinyu
parent f8664fe215
commit 60bcb9acbf
7 changed files with 401 additions and 0 deletions

View file

@ -24,5 +24,8 @@ sysdep_routines += \
rawmemchr-aligned \
rawmemchr-lsx \
rawmemchr-lasx \
memchr-aligned \
memchr-lsx \
memchr-lasx \
# sysdep_routines
endif

View file

@ -102,5 +102,12 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
IFUNC_IMPL_ADD (array, i, rawmemchr, 1, __rawmemchr_aligned)
)
IFUNC_IMPL (i, name, memchr,
#if !defined __loongarch_soft_float
IFUNC_IMPL_ADD (array, i, memchr, SUPPORT_LASX, __memchr_lasx)
IFUNC_IMPL_ADD (array, i, memchr, SUPPORT_LSX, __memchr_lsx)
#endif
IFUNC_IMPL_ADD (array, i, memchr, 1, __memchr_aligned)
)
return i;
}

View file

@ -0,0 +1,40 @@
/* Common definition for memchr ifunc selections.
All versions must be listed in ifunc-impl-list.c.
Copyright (C) 2023 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
<https://www.gnu.org/licenses/>. */
#include <ldsodefs.h>
#include <ifunc-init.h>
#if !defined __loongarch_soft_float
extern __typeof (REDIRECT_NAME) OPTIMIZE (lasx) attribute_hidden;
extern __typeof (REDIRECT_NAME) OPTIMIZE (lsx) attribute_hidden;
#endif
extern __typeof (REDIRECT_NAME) OPTIMIZE (aligned) attribute_hidden;
static inline void *
IFUNC_SELECTOR (void)
{
#if !defined __loongarch_soft_float
if (SUPPORT_LASX)
return OPTIMIZE (lasx);
else if (SUPPORT_LSX)
return OPTIMIZE (lsx);
else
#endif
return OPTIMIZE (aligned);
}

View file

@ -0,0 +1,95 @@
/* Optimized memchr implementation using basic LoongArch instructions.
Copyright (C) 2023 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
<https://www.gnu.org/licenses/>. */
#include <sysdep.h>
#include <sys/regdef.h>
#include <sys/asm.h>
#if IS_IN (libc)
# define MEMCHR_NAME __memchr_aligned
#else
# define MEMCHR_NAME memchr
#endif
LEAF(MEMCHR_NAME, 6)
beqz a2, L(out)
andi t1, a0, 0x7
add.d a5, a0, a2
bstrins.d a0, zero, 2, 0
ld.d t0, a0, 0
bstrins.d a1, a1, 15, 8
lu12i.w a3, 0x01010
slli.d t2, t1, 03
bstrins.d a1, a1, 31, 16
ori a3, a3, 0x101
li.d t7, -1
li.d t8, 8
bstrins.d a1, a1, 63, 32
bstrins.d a3, a3, 63, 32
sll.d t2, t7, t2
xor t0, t0, a1
addi.d a6, a5, -1
slli.d a4, a3, 7
sub.d t1, t8, t1
orn t0, t0, t2
sub.d t2, t0, a3
andn t3, a4, t0
bstrins.d a6, zero, 2, 0
and t0, t2, t3
bgeu t1, a2, L(end)
L(loop):
bnez t0, L(found)
ld.d t1, a0, 8
xor t0, t1, a1
addi.d a0, a0, 8
sub.d t2, t0, a3
andn t3, a4, t0
and t0, t2, t3
bne a0, a6, L(loop)
L(end):
sub.d t1, a5, a6
ctz.d t0, t0
srli.d t0, t0, 3
sltu t1, t0, t1
add.d a0, a0, t0
maskeqz a0, a0, t1
jr ra
L(found):
ctz.d t0, t0
srli.d t0, t0, 3
add.d a0, a0, t0
jr ra
L(out):
move a0, zero
jr ra
END(MEMCHR_NAME)
libc_hidden_builtin_def (MEMCHR_NAME)

View file

@ -0,0 +1,117 @@
/* Optimized memchr implementation using LoongArch LASX instructions.
Copyright (C) 2023 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
<https://www.gnu.org/licenses/>. */
#include <sysdep.h>
#include <sys/regdef.h>
#include <sys/asm.h>
#if IS_IN (libc) && !defined __loongarch_soft_float
# define MEMCHR __memchr_lasx
LEAF(MEMCHR, 6)
beqz a2, L(ret0)
add.d a3, a0, a2
andi t0, a0, 0x3f
bstrins.d a0, zero, 5, 0
xvld xr0, a0, 0
xvld xr1, a0, 32
li.d t1, -1
li.d t2, 64
xvreplgr2vr.b xr2, a1
sll.d t3, t1, t0
sub.d t2, t2, t0
xvseq.b xr0, xr0, xr2
xvseq.b xr1, xr1, xr2
xvmsknz.b xr0, xr0
xvmsknz.b xr1, xr1
xvpickve.w xr3, xr0, 4
xvpickve.w xr4, xr1, 4
vilvl.h vr0, vr3, vr0
vilvl.h vr1, vr4, vr1
vilvl.w vr0, vr1, vr0
movfr2gr.d t0, fa0
and t0, t0, t3
bgeu t2, a2, L(end)
bnez t0, L(found)
addi.d a4, a3, -1
bstrins.d a4, zero, 5, 0
L(loop):
xvld xr0, a0, 64
xvld xr1, a0, 96
addi.d a0, a0, 64
xvseq.b xr0, xr0, xr2
xvseq.b xr1, xr1, xr2
beq a0, a4, L(out)
xvmax.bu xr3, xr0, xr1
xvseteqz.v fcc0, xr3
bcnez fcc0, L(loop)
xvmsknz.b xr0, xr0
xvmsknz.b xr1, xr1
xvpickve.w xr3, xr0, 4
xvpickve.w xr4, xr1, 4
vilvl.h vr0, vr3, vr0
vilvl.h vr1, vr4, vr1
vilvl.w vr0, vr1, vr0
movfr2gr.d t0, fa0
L(found):
ctz.d t1, t0
add.d a0, a0, t1
jr ra
L(ret0):
move a0, zero
jr ra
L(out):
xvmsknz.b xr0, xr0
xvmsknz.b xr1, xr1
xvpickve.w xr3, xr0, 4
xvpickve.w xr4, xr1, 4
vilvl.h vr0, vr3, vr0
vilvl.h vr1, vr4, vr1
vilvl.w vr0, vr1, vr0
movfr2gr.d t0, fa0
L(end):
sub.d t2, zero, a3
srl.d t1, t1, t2
and t0, t0, t1
ctz.d t1, t0
add.d a0, a0, t1
maskeqz a0, a0, t0
jr ra
END(MEMCHR)
libc_hidden_builtin_def (MEMCHR)
#endif

View file

@ -0,0 +1,102 @@
/* Optimized memchr implementation using LoongArch LSX instructions.
Copyright (C) 2023 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
<https://www.gnu.org/licenses/>. */
#include <sysdep.h>
#include <sys/regdef.h>
#include <sys/asm.h>
#if IS_IN (libc) && !defined __loongarch_soft_float
# define MEMCHR __memchr_lsx
LEAF(MEMCHR, 6)
beqz a2, L(ret0)
add.d a3, a0, a2
andi t0, a0, 0x1f
bstrins.d a0, zero, 4, 0
vld vr0, a0, 0
vld vr1, a0, 16
li.d t1, -1
li.d t2, 32
vreplgr2vr.b vr2, a1
sll.d t3, t1, t0
sub.d t2, t2, t0
vseq.b vr0, vr0, vr2
vseq.b vr1, vr1, vr2
vmsknz.b vr0, vr0
vmsknz.b vr1, vr1
vilvl.h vr0, vr1, vr0
movfr2gr.s t0, fa0
and t0, t0, t3
bgeu t2, a2, L(end)
bnez t0, L(found)
addi.d a4, a3, -1
bstrins.d a4, zero, 4, 0
L(loop):
vld vr0, a0, 32
vld vr1, a0, 48
addi.d a0, a0, 32
vseq.b vr0, vr0, vr2
vseq.b vr1, vr1, vr2
beq a0, a4, L(out)
vmax.bu vr3, vr0, vr1
vseteqz.v fcc0, vr3
bcnez fcc0, L(loop)
vmsknz.b vr0, vr0
vmsknz.b vr1, vr1
vilvl.h vr0, vr1, vr0
movfr2gr.s t0, fa0
L(found):
ctz.w t0, t0
add.d a0, a0, t0
jr ra
L(ret0):
move a0, zero
jr ra
L(out):
vmsknz.b vr0, vr0
vmsknz.b vr1, vr1
vilvl.h vr0, vr1, vr0
movfr2gr.s t0, fa0
L(end):
sub.d t2, zero, a3
srl.w t1, t1, t2
and t0, t0, t1
ctz.w t1, t0
add.d a0, a0, t1
maskeqz a0, a0, t0
jr ra
END(MEMCHR)
libc_hidden_builtin_def (MEMCHR)
#endif

View file

@ -0,0 +1,37 @@
/* Multiple versions of memchr.
All versions must be listed in ifunc-impl-list.c.
Copyright (C) 2023 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
<https://www.gnu.org/licenses/>. */
/* Define multiple versions only for the definition in libc. */
#if IS_IN (libc)
# define memchr __redirect_memchr
# include <string.h>
# undef memchr
# define SYMBOL_NAME memchr
# include "ifunc-memchr.h"
libc_ifunc_redirected (__redirect_memchr, memchr,
IFUNC_SELECTOR ());
# ifdef SHARED
__hidden_ver1 (memchr, __GI_memchr, __redirect_memchr)
__attribute__ ((visibility ("hidden"))) __attribute_copy__ (memchr);
# endif
#endif