OpenE2K
/
glibc
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Use IFUNC on x86-64 memset

This commit is contained in:
H.J. Lu 2010-11-08 03:41:34 -05:00 committed by Ulrich Drepper
parent 344d0b545d
commit ff02d5280b
10 changed files with 387 additions and 156 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
ChangeLog
View File

@ -1,3 +1,21 @@
2010-11-07 H.J. Lu <hongjiu.lu@intel.com>
* sysdeps/x86_64/memset.S: Check USE_MULTIARCH and USE_SSE2 for
IFUNC support.
* sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Add
memset-x86-64.
* sysdeps/x86_64/multiarch/bzero.S: New file.
* sysdeps/x86_64/multiarch/cacheinfo.c: New file.
* sysdeps/x86_64/multiarch/memset-x86-64.S: New file.
* sysdeps/x86_64/multiarch/memset.S: New file.
* sysdeps/x86_64/multiarch/memset_chk.S: New file.
* sysdeps/x86_64/multiarch/init-arch.c (__init_cpu_features):
Set bit_Prefer_SSE_for_memop for Intel processors.
* sysdeps/x86_64/multiarch/init-arch.h (bit_Prefer_SSE_for_memop):
Define.
(index_Prefer_SSE_for_memop): Define.
(HAS_PREFER_SSE_FOR_MEMOP): Define.
2010-11-04 Luis Machado <luisgpm@br.ibm.com>
* sysdeps/powerpc/powerpc32/power7/mempcpy.S: New file.

311
sysdeps/x86_64/memset.S
View File

@ -24,7 +24,7 @@
#define __STOS_UPPER_BOUNDARY $65536
.text
#ifndef NOT_IN_libc
#if !defined NOT_IN_libc && !defined USE_MULTIARCH
ENTRY(__bzero)
mov %rsi,%rdx /* Adjust parameter. */
xorl %esi,%esi /* Fill with 0s. */
@ -34,10 +34,10 @@ weak_alias (__bzero, bzero)
#endif
#if defined PIC && !defined NOT_IN_libc
ENTRY (__memset_chk)
ENTRY_CHK (__memset_chk)
cmpq %rdx, %rcx
jb HIDDEN_JUMPTARGET (__chk_fail)
END (__memset_chk)
END_CHK (__memset_chk)
#endif
ENTRY (memset)
L(memset_entry):
@ -591,157 +591,13 @@ L(A6Q1): mov %dx,-0xe(%rdi)
L(A7Q0): mov %dl,-0x7(%rdi)
L(A6Q0): mov %dx,-0x6(%rdi)
mov %edx,-0x4(%rdi)
#ifndef USE_MULTIARCH
jmp L(aligned_now)
.balign 16
L(aligned_now):
cmpl $0x1,__x86_64_preferred_memory_instruction(%rip)
jg L(SSE_pre)
L(8byte_move_try):
cmpq __STOS_LOWER_BOUNDARY,%r8
jae L(8byte_stos_try)
.balign 16
L(8byte_move):
movq %r8,%rcx
shrq $7,%rcx
jz L(8byte_move_skip)
.p2align 4
L(8byte_move_loop):
decq %rcx
movq %rdx, (%rdi)
movq %rdx, 8 (%rdi)
movq %rdx, 16 (%rdi)
movq %rdx, 24 (%rdi)
movq %rdx, 32 (%rdi)
movq %rdx, 40 (%rdi)
movq %rdx, 48 (%rdi)
movq %rdx, 56 (%rdi)
movq %rdx, 64 (%rdi)
movq %rdx, 72 (%rdi)
movq %rdx, 80 (%rdi)
movq %rdx, 88 (%rdi)
movq %rdx, 96 (%rdi)
movq %rdx, 104 (%rdi)
movq %rdx, 112 (%rdi)
movq %rdx, 120 (%rdi)
leaq 128 (%rdi),%rdi
jnz L(8byte_move_loop)
L(8byte_move_skip):
andl $127,%r8d
lea (%rdi,%r8,1),%rdi
#ifndef PIC
lea L(setPxQx)(%rip),%r11
jmpq *(%r11,%r8,8) # old scheme remained for nonPIC
#else
lea L(Got0)(%rip),%r11
lea L(setPxQx)(%rip),%rcx
movswq (%rcx,%r8,2),%rcx
lea (%rcx,%r11,1),%r11
jmpq *%r11
#endif
.balign 16
L(8byte_stos_try):
mov __x86_64_shared_cache_size(%rip),%r9d // ck largest cache size
cmpq %r8,%r9 // calculate the lesser of remaining
cmovaq %r8,%r9 // bytes and largest cache size
jbe L(8byte_stos)
L(8byte_move_reuse_try):
cmp __STOS_UPPER_BOUNDARY,%r8
jae L(8byte_move)
.balign 16
L(8byte_stos):
movq %r9,%rcx
andq $-8,%r9
shrq $3,%rcx
jz L(8byte_stos_skip)
xchgq %rax,%rdx
rep
stosq
xchgq %rax,%rdx
L(8byte_stos_skip):
subq %r9,%r8
ja L(8byte_nt_move)
andl $7,%r8d
lea (%rdi,%r8,1),%rdi
#ifndef PIC
lea L(setPxQx)(%rip),%r11
jmpq *(%r11,%r8,8) # old scheme remained for nonPIC
#else
lea L(Got0)(%rip),%r11
lea L(setPxQx)(%rip),%rcx
movswq (%rcx,%r8,2),%rcx
lea (%rcx,%r11,1),%r11
jmpq *%r11
#endif
.balign 16
L(8byte_nt_move):
movq %r8,%rcx
shrq $7,%rcx
jz L(8byte_nt_move_skip)
.balign 16
L(8byte_nt_move_loop):
decq %rcx
movntiq %rdx, (%rdi)
movntiq %rdx, 8 (%rdi)
movntiq %rdx, 16 (%rdi)
movntiq %rdx, 24 (%rdi)
movntiq %rdx, 32 (%rdi)
movntiq %rdx, 40 (%rdi)
movntiq %rdx, 48 (%rdi)
movntiq %rdx, 56 (%rdi)
movntiq %rdx, 64 (%rdi)
movntiq %rdx, 72 (%rdi)
movntiq %rdx, 80 (%rdi)
movntiq %rdx, 88 (%rdi)
movntiq %rdx, 96 (%rdi)
movntiq %rdx, 104 (%rdi)
movntiq %rdx, 112 (%rdi)
movntiq %rdx, 120 (%rdi)
leaq 128 (%rdi),%rdi
jnz L(8byte_nt_move_loop)
sfence
L(8byte_nt_move_skip):
andl $127,%r8d
lea (%rdi,%r8,1),%rdi
#ifndef PIC
lea L(setPxQx)(%rip),%r11
jmpq *(%r11,%r8,8) # old scheme remained for nonPIC
#else
lea L(Got0)(%rip),%r11
lea L(setPxQx)(%rip),%rcx
movswq (%rcx,%r8,2),%rcx
lea (%rcx,%r11,1),%r11
jmpq *%r11
#endif
L(SSE_pre):
#endif
#if !defined USE_MULTIARCH || defined USE_SSE2
# fill RegXMM0 with the pattern
movd %rdx,%xmm0
punpcklqdq %xmm0,%xmm0
@ -1342,11 +1198,162 @@ L(SSExDx):
.short L(SSE15QB)-L(SSE0Q0)
#endif
.popsection
#endif /* !defined USE_MULTIARCH || defined USE_SSE2 */
.balign 16
L(aligned_now):
#ifndef USE_MULTIARCH
cmpl $0x1,__x86_64_preferred_memory_instruction(%rip)
jg L(SSE_pre)
#endif /* USE_MULTIARCH */
L(8byte_move_try):
cmpq __STOS_LOWER_BOUNDARY,%r8
jae L(8byte_stos_try)
.balign 16
L(8byte_move):
movq %r8,%rcx
shrq $7,%rcx
jz L(8byte_move_skip)
.p2align 4
L(8byte_move_loop):
decq %rcx
movq %rdx, (%rdi)
movq %rdx, 8 (%rdi)
movq %rdx, 16 (%rdi)
movq %rdx, 24 (%rdi)
movq %rdx, 32 (%rdi)
movq %rdx, 40 (%rdi)
movq %rdx, 48 (%rdi)
movq %rdx, 56 (%rdi)
movq %rdx, 64 (%rdi)
movq %rdx, 72 (%rdi)
movq %rdx, 80 (%rdi)
movq %rdx, 88 (%rdi)
movq %rdx, 96 (%rdi)
movq %rdx, 104 (%rdi)
movq %rdx, 112 (%rdi)
movq %rdx, 120 (%rdi)
leaq 128 (%rdi),%rdi
jnz L(8byte_move_loop)
L(8byte_move_skip):
andl $127,%r8d
lea (%rdi,%r8,1),%rdi
#ifndef PIC
lea L(setPxQx)(%rip),%r11
jmpq *(%r11,%r8,8) # old scheme remained for nonPIC
#else
lea L(Got0)(%rip),%r11
lea L(setPxQx)(%rip),%rcx
movswq (%rcx,%r8,2),%rcx
lea (%rcx,%r11,1),%r11
jmpq *%r11
#endif
.balign 16
L(8byte_stos_try):
mov __x86_64_shared_cache_size(%rip),%r9d // ck largest cache size
cmpq %r8,%r9 // calculate the lesser of remaining
cmovaq %r8,%r9 // bytes and largest cache size
jbe L(8byte_stos)
L(8byte_move_reuse_try):
cmp __STOS_UPPER_BOUNDARY,%r8
jae L(8byte_move)
.balign 16
L(8byte_stos):
movq %r9,%rcx
andq $-8,%r9
shrq $3,%rcx
jz L(8byte_stos_skip)
xchgq %rax,%rdx
rep
stosq
xchgq %rax,%rdx
L(8byte_stos_skip):
subq %r9,%r8
ja L(8byte_nt_move)
andl $7,%r8d
lea (%rdi,%r8,1),%rdi
#ifndef PIC
lea L(setPxQx)(%rip),%r11
jmpq *(%r11,%r8,8) # old scheme remained for nonPIC
#else
lea L(Got0)(%rip),%r11
lea L(setPxQx)(%rip),%rcx
movswq (%rcx,%r8,2),%rcx
lea (%rcx,%r11,1),%r11
jmpq *%r11
#endif
.balign 16
L(8byte_nt_move):
movq %r8,%rcx
shrq $7,%rcx
jz L(8byte_nt_move_skip)
.balign 16
L(8byte_nt_move_loop):
decq %rcx
movntiq %rdx, (%rdi)
movntiq %rdx, 8 (%rdi)
movntiq %rdx, 16 (%rdi)
movntiq %rdx, 24 (%rdi)
movntiq %rdx, 32 (%rdi)
movntiq %rdx, 40 (%rdi)
movntiq %rdx, 48 (%rdi)
movntiq %rdx, 56 (%rdi)
movntiq %rdx, 64 (%rdi)
movntiq %rdx, 72 (%rdi)
movntiq %rdx, 80 (%rdi)
movntiq %rdx, 88 (%rdi)
movntiq %rdx, 96 (%rdi)
movntiq %rdx, 104 (%rdi)
movntiq %rdx, 112 (%rdi)
movntiq %rdx, 120 (%rdi)
leaq 128 (%rdi),%rdi
jnz L(8byte_nt_move_loop)
sfence
L(8byte_nt_move_skip):
andl $127,%r8d
lea (%rdi,%r8,1),%rdi
#ifndef PIC
lea L(setPxQx)(%rip),%r11
jmpq *(%r11,%r8,8) # old scheme remained for nonPIC
#else
lea L(Got0)(%rip),%r11
lea L(setPxQx)(%rip),%rcx
movswq (%rcx,%r8,2),%rcx
lea (%rcx,%r11,1),%r11
jmpq *%r11
#endif
END (memset)
libc_hidden_builtin_def (memset)
#if defined PIC && !defined NOT_IN_libc
#if defined PIC && !defined NOT_IN_libc && !defined USE_MULTIARCH
strong_alias (__memset_chk, __memset_zero_constant_len_parameter)
.section .gnu.warning.__memset_zero_constant_len_parameter
.string "memset used with constant zero length parameter; this could be due to transposed parameters"

3
sysdeps/x86_64/multiarch/Makefile
View File

@ -8,7 +8,8 @@ sysdep_routines += stpncpy-c strncpy-c strcmp-ssse3 strncmp-ssse3 \
strend-sse4 memcmp-sse4 memcpy-ssse3 mempcpy-ssse3 \
memmove-ssse3 memcpy-ssse3-back mempcpy-ssse3-back \
memmove-ssse3-back strcasestr-nonascii strcasecmp_l-ssse3 \
strncase_l-ssse3 strlen-sse4 strlen-no-bsf
strncase_l-ssse3 strlen-sse4 strlen-no-bsf \
memset-x86-64
ifeq (yes,$(config-cflags-sse4))
sysdep_routines += strcspn-c strpbrk-c strspn-c strstr-c strcasestr-c varshift
CFLAGS-varshift.c += -msse4

56
sysdeps/x86_64/multiarch/bzero.S Normal file
View File

@ -0,0 +1,56 @@
/* Multiple versions of bzero
Copyright (C) 2010 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, write to the Free
Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
02111-1307 USA. */
#include <sysdep.h>
#include <init-arch.h>
.text
ENTRY(__bzero)
.type __bzero, @gnu_indirect_function
cmpl $0, __cpu_features+KIND_OFFSET(%rip)
jne 1f
call __init_cpu_features
1: leaq __bzero_x86_64(%rip), %rax
testl $bit_Prefer_SSE_for_memop, __cpu_features+FEATURE_OFFSET+index_Prefer_SSE_for_memop(%rip)
jz 2f
leaq __bzero_sse2(%rip), %rax
2: ret
END(__bzero)
.type __bzero_sse2, @function
__bzero_sse2:
cfi_startproc
CALL_MCOUNT
mov %rsi,%rdx /* Adjust parameter. */
xorl %esi,%esi /* Fill with 0s. */
jmp __memset_sse2
cfi_endproc
.size __bzero_sse2, .-__bzero_sse2
.type __bzero_x86_64, @function
__bzero_x86_64:
cfi_startproc
CALL_MCOUNT
mov %rsi,%rdx /* Adjust parameter. */
xorl %esi,%esi /* Fill with 0s. */
jmp __memset_x86_64
cfi_endproc
.size __bzero_x86_64, .-__bzero_x86_64
weak_alias (__bzero, bzero)

2
sysdeps/x86_64/multiarch/cacheinfo.c Normal file
View File

@ -0,0 +1,2 @@
#define DISABLE_PREFERRED_MEMORY_INSTRUCTION
#include "../cacheinfo.c"

5
sysdeps/x86_64/multiarch/init-arch.c
View File

@ -59,6 +59,11 @@ __init_cpu_features (void)
get_common_indeces (&family, &model);
/* Intel processors prefer SSE instruction for memory/string
routines if they are avaiable. */
__cpu_features.feature[index_Prefer_SSE_for_memop]
|= bit_Prefer_SSE_for_memop;
unsigned int eax = __cpu_features.cpuid[COMMON_CPUID_INDEX_1].eax;
unsigned int extended_family = (eax >> 20) & 0xff;
unsigned int extended_model = (eax >> 12) & 0xf0;

12
sysdeps/x86_64/multiarch/init-arch.h
View File

@ -16,9 +16,10 @@
Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
02111-1307 USA. */
#define bit_Fast_Rep_String (1 << 0)
#define bit_Fast_Copy_Backward (1 << 1)
#define bit_Slow_BSF (1 << 2)
#define bit_Fast_Rep_String (1 << 0)
#define bit_Fast_Copy_Backward (1 << 1)
#define bit_Slow_BSF (1 << 2)
#define bit_Prefer_SSE_for_memop (1 << 3)
#ifdef __ASSEMBLER__
@ -37,6 +38,7 @@
# define index_Fast_Rep_String FEATURE_INDEX_1*FEATURE_SIZE
# define index_Fast_Copy_Backward FEATURE_INDEX_1*FEATURE_SIZE
# define index_Slow_BSF FEATURE_INDEX_1*FEATURE_SIZE
# define index_Prefer_SSE_for_memop FEATURE_INDEX_1*FEATURE_SIZE
#else /* __ASSEMBLER__ */
@ -109,6 +111,7 @@ extern const struct cpu_features *__get_cpu_features (void)
# define index_Fast_Rep_String FEATURE_INDEX_1
# define index_Fast_Copy_Backward FEATURE_INDEX_1
# define index_Slow_BSF FEATURE_INDEX_1
# define index_Prefer_SSE_for_memop FEATURE_INDEX_1
#define HAS_ARCH_FEATURE(idx, bit) \
((__get_cpu_features ()->feature[idx] & (bit)) != 0)
@ -122,4 +125,7 @@ extern const struct cpu_features *__get_cpu_features (void)
#define HAS_SLOW_BSF \
HAS_ARCH_FEATURE (index_Slow_BSF, bit_Slow_BSF)
#define HAS_PREFER_SSE_FOR_MEMOP \
HAS_ARCH_FEATURE (index_Prefer_SSE_for_memop, bit_Prefer_SSE_for_memop)
#endif /* __ASSEMBLER__ */

18
sysdeps/x86_64/multiarch/memset-x86-64.S Normal file
View File

@ -0,0 +1,18 @@
#include <sysdep.h>
#ifndef NOT_IN_libc
# undef ENTRY_CHK
# define ENTRY_CHK(name) \
.type __memset_chk_x86_64, @function; \
.globl __memset_chk_x86_64; \
.p2align 4; \
__memset_chk_x86_64: cfi_startproc; \
CALL_MCOUNT
# undef END_CHK
# define END_CHK(name) \
cfi_endproc; .size __memset_chk_x86_64, .-__memset_chk_x86_64
# define libc_hidden_builtin_def(name)
# define memset __memset_x86_64
# include "../memset.S"
#endif

74
sysdeps/x86_64/multiarch/memset.S Normal file
View File

@ -0,0 +1,74 @@
/* Multiple versions of memset
Copyright (C) 2010 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, write to the Free
Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
02111-1307 USA. */
#include <sysdep.h>
#include <init-arch.h>
/* Define multiple versions only for the definition in lib. */
#ifndef NOT_IN_libc
ENTRY(memset)
.type memset, @gnu_indirect_function
cmpl $0, __cpu_features+KIND_OFFSET(%rip)
jne 1f
call __init_cpu_features
1: leaq __memset_x86_64(%rip), %rax
testl $bit_Prefer_SSE_for_memop, __cpu_features+FEATURE_OFFSET+index_Prefer_SSE_for_memop(%rip)
jz 2f
leaq __memset_sse2(%rip), %rax
2: ret
END(memset)
# define USE_SSE2 1
# undef ENTRY
# define ENTRY(name) \
.type __memset_sse2, @function; \
.globl __memset_sse2; \
.p2align 4; \
__memset_sse2: cfi_startproc; \
CALL_MCOUNT
# undef END
# define END(name) \
cfi_endproc; .size __memset_sse2, .-__memset_sse2
# undef ENTRY_CHK
# define ENTRY_CHK(name) \
.type __memset_chk_sse2, @function; \
.globl __memset_chk_sse2; \
.p2align 4; \
__memset_chk_sse2: cfi_startproc; \
CALL_MCOUNT
# undef END_CHK
# define END_CHK(name) \
cfi_endproc; .size __memset_chk_sse2, .-__memset_chk_sse2
# ifdef SHARED
# undef libc_hidden_builtin_def
/* It doesn't make sense to send libc-internal memset calls through a PLT.
The speedup we get from using GPR instruction is likely eaten away
by the indirect call in the PLT. */
# define libc_hidden_builtin_def(name) \
.globl __GI_memset; __GI_memset = __memset_sse2
# endif
# undef strong_alias
# define strong_alias(original, alias)
#endif
#include "../memset.S"

44
sysdeps/x86_64/multiarch/memset_chk.S Normal file
View File

@ -0,0 +1,44 @@
/* Multiple versions of __memset_chk
Copyright (C) 2010 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, write to the Free
Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
02111-1307 USA. */
#include <sysdep.h>
#include <init-arch.h>
/* Define multiple versions only for the definition in lib. */
#ifndef NOT_IN_libc
# ifdef SHARED
ENTRY(__memset_chk)
.type __memset_chk, @gnu_indirect_function
cmpl $0, __cpu_features+KIND_OFFSET(%rip)
jne 1f
call __init_cpu_features
1: leaq __memset_chk_x86_64(%rip), %rax
testl $bit_Prefer_SSE_for_memop, __cpu_features+FEATURE_OFFSET+index_Prefer_SSE_for_memop(%rip)
jz 2f
leaq __memset_chk_sse2(%rip), %rax
2: ret
END(__memset_chk)
strong_alias (__memset_chk, __memset_zero_constant_len_parameter)
.section .gnu.warning.__memset_zero_constant_len_parameter
.string "memset used with constant zero length parameter; this could be due to transposed parameters"
# else
# include "../memset_chk.S"
# endif
#endif