|
 |
b40826 |
2010-08-27 Ulrich Drepper <drepper@redhat.com>
|
|
|
b40826 |
|
|
|
b40826 |
* sysdeps/x86_64/multiarch/strlen-no-bsf.S: Move to .text.slow section.
|
|
|
b40826 |
|
|
|
b40826 |
* sysdeps/x86_64/strlen.S: Minimal code improvement.
|
|
|
b40826 |
|
|
|
b40826 |
2010-08-26 H.J. Lu <hongjiu.lu@intel.com>
|
|
|
b40826 |
|
|
|
b40826 |
* sysdeps/x86_64/strlen.S: Unroll the loop.
|
|
|
b40826 |
* sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Add
|
|
|
b40826 |
strlen-sse2 strlen-sse2-bsf.
|
|
|
b40826 |
* sysdeps/x86_64/multiarch/strlen.S ((strlen): Return
|
|
|
b40826 |
__strlen_no_bsf if bit_Slow_BSF is set.
|
|
|
b40826 |
(__strlen_sse42): Removed.
|
|
|
b40826 |
* sysdeps/x86_64/multiarch/strlen-no-bsf.S: New file.
|
|
|
b40826 |
* sysdeps/x86_64/multiarch/strlen-sse4.S: New file.
|
|
|
b40826 |
|
|
|
b40826 |
2010-08-25 H.J. Lu <hongjiu.lu@intel.com>
|
|
|
b40826 |
|
|
|
b40826 |
* sysdeps/i386/i686/multiarch/Makefile (sysdep_routines): Add
|
|
|
b40826 |
strlen-sse2 strlen-sse2-bsf.
|
|
|
b40826 |
* sysdeps/i386/i686/multiarch/strlen.S (strlen): Return
|
|
|
b40826 |
__strlen_sse2_bsf if bit_Slow_BSF is unset.
|
|
|
b40826 |
(__strlen_sse2): Removed.
|
|
|
b40826 |
* sysdeps/i386/i686/multiarch/strlen-sse2-bsf.S: New file.
|
|
|
b40826 |
* sysdeps/i386/i686/multiarch/strlen-sse2.S: New file.
|
|
|
b40826 |
* sysdeps/x86_64/multiarch/init-arch.c (__init_cpu_features): Set
|
|
|
b40826 |
bit_Slow_BSF for Atom.
|
|
|
b40826 |
* sysdeps/x86_64/multiarch/init-arch.h (bit_Slow_BSF): Define.
|
|
|
b40826 |
(index_Slow_BSF): Define.
|
|
|
b40826 |
(HAS_SLOW_BSF): Define.
|
|
|
b40826 |
|
|
|
b40826 |
Index: glibc-2.12-2-gc4ccff1/sysdeps/i386/i686/multiarch/Makefile
|
|
|
b40826 |
===================================================================
|
|
|
b40826 |
--- glibc-2.12-2-gc4ccff1.orig/sysdeps/i386/i686/multiarch/Makefile
|
|
|
b40826 |
+++ glibc-2.12-2-gc4ccff1/sysdeps/i386/i686/multiarch/Makefile
|
|
|
b40826 |
@@ -9,7 +9,8 @@ sysdep_routines += bzero-sse2 memset-sse
|
|
|
b40826 |
memmove-ssse3-rep bcopy-ssse3 bcopy-ssse3-rep \
|
|
|
b40826 |
memset-sse2-rep bzero-sse2-rep strcmp-ssse3 \
|
|
|
b40826 |
strcmp-sse4 strncmp-c strncmp-ssse3 strncmp-sse4 \
|
|
|
b40826 |
- memcmp-ssse3 memcmp-sse4 strcasestr-nonascii
|
|
|
b40826 |
+ memcmp-ssse3 memcmp-sse4 strcasestr-nonascii \
|
|
|
b40826 |
+ strlen-sse2 strlen-sse2-bsf
|
|
|
b40826 |
ifeq (yes,$(config-cflags-sse4))
|
|
|
b40826 |
sysdep_routines += strcspn-c strpbrk-c strspn-c strstr-c strcasestr-c
|
|
|
b40826 |
CFLAGS-strcspn-c.c += -msse4
|
|
|
b40826 |
Index: glibc-2.12-2-gc4ccff1/sysdeps/i386/i686/multiarch/strlen-sse2-bsf.S
|
|
|
b40826 |
===================================================================
|
|
|
b40826 |
--- /dev/null
|
|
|
b40826 |
+++ glibc-2.12-2-gc4ccff1/sysdeps/i386/i686/multiarch/strlen-sse2-bsf.S
|
|
|
b40826 |
@@ -0,0 +1,127 @@
|
|
|
b40826 |
+/* strlen with SSE2 and BSF
|
|
|
b40826 |
+ Copyright (C) 2010 Free Software Foundation, Inc.
|
|
|
b40826 |
+ Contributed by Intel Corporation.
|
|
|
b40826 |
+ This file is part of the GNU C Library.
|
|
|
b40826 |
+
|
|
|
b40826 |
+ The GNU C Library is free software; you can redistribute it and/or
|
|
|
b40826 |
+ modify it under the terms of the GNU Lesser General Public
|
|
|
b40826 |
+ License as published by the Free Software Foundation; either
|
|
|
b40826 |
+ version 2.1 of the License, or (at your option) any later version.
|
|
|
b40826 |
+
|
|
|
b40826 |
+ The GNU C Library is distributed in the hope that it will be useful,
|
|
|
b40826 |
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
b40826 |
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
|
b40826 |
+ Lesser General Public License for more details.
|
|
|
b40826 |
+
|
|
|
b40826 |
+ You should have received a copy of the GNU Lesser General Public
|
|
|
b40826 |
+ License along with the GNU C Library; if not, write to the Free
|
|
|
b40826 |
+ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
|
|
|
b40826 |
+ 02111-1307 USA. */
|
|
|
b40826 |
+
|
|
|
b40826 |
+#if defined SHARED && !defined NOT_IN_libc
|
|
|
b40826 |
+
|
|
|
b40826 |
+#include <sysdep.h>
|
|
|
b40826 |
+#include "asm-syntax.h"
|
|
|
b40826 |
+
|
|
|
b40826 |
+#define CFI_PUSH(REG) \
|
|
|
b40826 |
+ cfi_adjust_cfa_offset (4); \
|
|
|
b40826 |
+ cfi_rel_offset (REG, 0)
|
|
|
b40826 |
+
|
|
|
b40826 |
+#define CFI_POP(REG) \
|
|
|
b40826 |
+ cfi_adjust_cfa_offset (-4); \
|
|
|
b40826 |
+ cfi_restore (REG)
|
|
|
b40826 |
+
|
|
|
b40826 |
+#define PUSH(REG) pushl REG; CFI_PUSH (REG)
|
|
|
b40826 |
+#define POP(REG) popl REG; CFI_POP (REG)
|
|
|
b40826 |
+#define PARMS 4 + 8 /* Preserve ESI and EDI. */
|
|
|
b40826 |
+#define STR PARMS
|
|
|
b40826 |
+#define ENTRANCE PUSH (%esi); PUSH (%edi); cfi_remember_state
|
|
|
b40826 |
+#define RETURN POP (%edi); POP (%esi); ret; \
|
|
|
b40826 |
+ cfi_restore_state; cfi_remember_state
|
|
|
b40826 |
+
|
|
|
b40826 |
+ .text
|
|
|
b40826 |
+ENTRY ( __strlen_sse2_bsf)
|
|
|
b40826 |
+ ENTRANCE
|
|
|
b40826 |
+ mov STR(%esp), %edi
|
|
|
b40826 |
+ xor %eax, %eax
|
|
|
b40826 |
+ mov %edi, %ecx
|
|
|
b40826 |
+ and $0x3f, %ecx
|
|
|
b40826 |
+ pxor %xmm0, %xmm0
|
|
|
b40826 |
+ cmp $0x30, %ecx
|
|
|
b40826 |
+ ja L(next)
|
|
|
b40826 |
+ movdqu (%edi), %xmm1
|
|
|
b40826 |
+ pcmpeqb %xmm1, %xmm0
|
|
|
b40826 |
+ pmovmskb %xmm0, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ jnz L(exit_less16)
|
|
|
b40826 |
+ mov %edi, %eax
|
|
|
b40826 |
+ and $-16, %eax
|
|
|
b40826 |
+ jmp L(align16_start)
|
|
|
b40826 |
+L(next):
|
|
|
b40826 |
+
|
|
|
b40826 |
+ mov %edi, %eax
|
|
|
b40826 |
+ and $-16, %eax
|
|
|
b40826 |
+ pcmpeqb (%eax), %xmm0
|
|
|
b40826 |
+ mov $-1, %esi
|
|
|
b40826 |
+ sub %eax, %ecx
|
|
|
b40826 |
+ shl %cl, %esi
|
|
|
b40826 |
+ pmovmskb %xmm0, %edx
|
|
|
b40826 |
+ and %esi, %edx
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+L(align16_start):
|
|
|
b40826 |
+ pxor %xmm0, %xmm0
|
|
|
b40826 |
+ pxor %xmm1, %xmm1
|
|
|
b40826 |
+ pxor %xmm2, %xmm2
|
|
|
b40826 |
+ pxor %xmm3, %xmm3
|
|
|
b40826 |
+ .p2align 4
|
|
|
b40826 |
+L(align16_loop):
|
|
|
b40826 |
+ pcmpeqb 16(%eax), %xmm0
|
|
|
b40826 |
+ pmovmskb %xmm0, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ jnz L(exit16)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb 32(%eax), %xmm1
|
|
|
b40826 |
+ pmovmskb %xmm1, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ jnz L(exit32)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb 48(%eax), %xmm2
|
|
|
b40826 |
+ pmovmskb %xmm2, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ jnz L(exit48)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb 64(%eax), %xmm3
|
|
|
b40826 |
+ pmovmskb %xmm3, %edx
|
|
|
b40826 |
+ lea 64(%eax), %eax
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ jz L(align16_loop)
|
|
|
b40826 |
+L(exit):
|
|
|
b40826 |
+ sub %edi, %eax
|
|
|
b40826 |
+L(exit_less16):
|
|
|
b40826 |
+ bsf %edx, %edx
|
|
|
b40826 |
+ add %edx, %eax
|
|
|
b40826 |
+ RETURN
|
|
|
b40826 |
+L(exit16):
|
|
|
b40826 |
+ sub %edi, %eax
|
|
|
b40826 |
+ bsf %edx, %edx
|
|
|
b40826 |
+ add %edx, %eax
|
|
|
b40826 |
+ add $16, %eax
|
|
|
b40826 |
+ RETURN
|
|
|
b40826 |
+L(exit32):
|
|
|
b40826 |
+ sub %edi, %eax
|
|
|
b40826 |
+ bsf %edx, %edx
|
|
|
b40826 |
+ add %edx, %eax
|
|
|
b40826 |
+ add $32, %eax
|
|
|
b40826 |
+ RETURN
|
|
|
b40826 |
+L(exit48):
|
|
|
b40826 |
+ sub %edi, %eax
|
|
|
b40826 |
+ bsf %edx, %edx
|
|
|
b40826 |
+ add %edx, %eax
|
|
|
b40826 |
+ add $48, %eax
|
|
|
b40826 |
+ POP (%edi)
|
|
|
b40826 |
+ POP (%esi)
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+
|
|
|
b40826 |
+END ( __strlen_sse2_bsf)
|
|
|
b40826 |
+
|
|
|
b40826 |
+#endif
|
|
|
b40826 |
Index: glibc-2.12-2-gc4ccff1/sysdeps/i386/i686/multiarch/strlen-sse2.S
|
|
|
b40826 |
===================================================================
|
|
|
b40826 |
--- /dev/null
|
|
|
b40826 |
+++ glibc-2.12-2-gc4ccff1/sysdeps/i386/i686/multiarch/strlen-sse2.S
|
|
|
b40826 |
@@ -0,0 +1,347 @@
|
|
|
b40826 |
+/* strlen with SSE2
|
|
|
b40826 |
+ Copyright (C) 2010 Free Software Foundation, Inc.
|
|
|
b40826 |
+ Contributed by Intel Corporation.
|
|
|
b40826 |
+ This file is part of the GNU C Library.
|
|
|
b40826 |
+
|
|
|
b40826 |
+ The GNU C Library is free software; you can redistribute it and/or
|
|
|
b40826 |
+ modify it under the terms of the GNU Lesser General Public
|
|
|
b40826 |
+ License as published by the Free Software Foundation; either
|
|
|
b40826 |
+ version 2.1 of the License, or (at your option) any later version.
|
|
|
b40826 |
+
|
|
|
b40826 |
+ The GNU C Library is distributed in the hope that it will be useful,
|
|
|
b40826 |
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
b40826 |
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
|
b40826 |
+ Lesser General Public License for more details.
|
|
|
b40826 |
+
|
|
|
b40826 |
+ You should have received a copy of the GNU Lesser General Public
|
|
|
b40826 |
+ License along with the GNU C Library; if not, write to the Free
|
|
|
b40826 |
+ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
|
|
|
b40826 |
+ 02111-1307 USA. */
|
|
|
b40826 |
+
|
|
|
b40826 |
+#if defined SHARED && !defined NOT_IN_libc
|
|
|
b40826 |
+
|
|
|
b40826 |
+#include <sysdep.h>
|
|
|
b40826 |
+#include "asm-syntax.h"
|
|
|
b40826 |
+
|
|
|
b40826 |
+#define CFI_PUSH(REG) \
|
|
|
b40826 |
+ cfi_adjust_cfa_offset (4); \
|
|
|
b40826 |
+ cfi_rel_offset (REG, 0)
|
|
|
b40826 |
+
|
|
|
b40826 |
+#define CFI_POP(REG) \
|
|
|
b40826 |
+ cfi_adjust_cfa_offset (-4); \
|
|
|
b40826 |
+ cfi_restore (REG)
|
|
|
b40826 |
+
|
|
|
b40826 |
+#define PUSH(REG) pushl REG; CFI_PUSH (REG)
|
|
|
b40826 |
+#define POP(REG) popl REG; CFI_POP (REG)
|
|
|
b40826 |
+#define PARMS 4
|
|
|
b40826 |
+#define STR PARMS
|
|
|
b40826 |
+#define ENTRANCE
|
|
|
b40826 |
+#define RETURN ret
|
|
|
b40826 |
+
|
|
|
b40826 |
+ .text
|
|
|
b40826 |
+ENTRY (__strlen_sse2)
|
|
|
b40826 |
+ ENTRANCE
|
|
|
b40826 |
+ mov STR(%esp), %edx
|
|
|
b40826 |
+ xor %eax, %eax
|
|
|
b40826 |
+ cmpb $0, (%edx)
|
|
|
b40826 |
+ jz L(exit_tail0)
|
|
|
b40826 |
+ cmpb $0, 1(%edx)
|
|
|
b40826 |
+ jz L(exit_tail1)
|
|
|
b40826 |
+ cmpb $0, 2(%edx)
|
|
|
b40826 |
+ jz L(exit_tail2)
|
|
|
b40826 |
+ cmpb $0, 3(%edx)
|
|
|
b40826 |
+ jz L(exit_tail3)
|
|
|
b40826 |
+ cmpb $0, 4(%edx)
|
|
|
b40826 |
+ jz L(exit_tail4)
|
|
|
b40826 |
+ cmpb $0, 5(%edx)
|
|
|
b40826 |
+ jz L(exit_tail5)
|
|
|
b40826 |
+ cmpb $0, 6(%edx)
|
|
|
b40826 |
+ jz L(exit_tail6)
|
|
|
b40826 |
+ cmpb $0, 7(%edx)
|
|
|
b40826 |
+ jz L(exit_tail7)
|
|
|
b40826 |
+ cmpb $0, 8(%edx)
|
|
|
b40826 |
+ jz L(exit_tail8)
|
|
|
b40826 |
+ cmpb $0, 9(%edx)
|
|
|
b40826 |
+ jz L(exit_tail9)
|
|
|
b40826 |
+ cmpb $0, 10(%edx)
|
|
|
b40826 |
+ jz L(exit_tail10)
|
|
|
b40826 |
+ cmpb $0, 11(%edx)
|
|
|
b40826 |
+ jz L(exit_tail11)
|
|
|
b40826 |
+ cmpb $0, 12(%edx)
|
|
|
b40826 |
+ jz L(exit_tail12)
|
|
|
b40826 |
+ cmpb $0, 13(%edx)
|
|
|
b40826 |
+ jz L(exit_tail13)
|
|
|
b40826 |
+ cmpb $0, 14(%edx)
|
|
|
b40826 |
+ jz L(exit_tail14)
|
|
|
b40826 |
+ cmpb $0, 15(%edx)
|
|
|
b40826 |
+ jz L(exit_tail15)
|
|
|
b40826 |
+ pxor %xmm0, %xmm0
|
|
|
b40826 |
+ mov %edx, %eax
|
|
|
b40826 |
+ mov %edx, %ecx
|
|
|
b40826 |
+ and $-16, %eax
|
|
|
b40826 |
+ add $16, %ecx
|
|
|
b40826 |
+ add $16, %eax
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%eax), %xmm0
|
|
|
b40826 |
+ pmovmskb %xmm0, %edx
|
|
|
b40826 |
+ pxor %xmm1, %xmm1
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%eax), %eax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%eax), %xmm1
|
|
|
b40826 |
+ pmovmskb %xmm1, %edx
|
|
|
b40826 |
+ pxor %xmm2, %xmm2
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%eax), %eax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%eax), %xmm2
|
|
|
b40826 |
+ pmovmskb %xmm2, %edx
|
|
|
b40826 |
+ pxor %xmm3, %xmm3
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%eax), %eax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%eax), %xmm3
|
|
|
b40826 |
+ pmovmskb %xmm3, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%eax), %eax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%eax), %xmm0
|
|
|
b40826 |
+ pmovmskb %xmm0, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%eax), %eax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%eax), %xmm1
|
|
|
b40826 |
+ pmovmskb %xmm1, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%eax), %eax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%eax), %xmm2
|
|
|
b40826 |
+ pmovmskb %xmm2, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%eax), %eax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%eax), %xmm3
|
|
|
b40826 |
+ pmovmskb %xmm3, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%eax), %eax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%eax), %xmm0
|
|
|
b40826 |
+ pmovmskb %xmm0, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%eax), %eax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%eax), %xmm1
|
|
|
b40826 |
+ pmovmskb %xmm1, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%eax), %eax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%eax), %xmm2
|
|
|
b40826 |
+ pmovmskb %xmm2, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%eax), %eax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%eax), %xmm3
|
|
|
b40826 |
+ pmovmskb %xmm3, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%eax), %eax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%eax), %xmm0
|
|
|
b40826 |
+ pmovmskb %xmm0, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%eax), %eax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%eax), %xmm1
|
|
|
b40826 |
+ pmovmskb %xmm1, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%eax), %eax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%eax), %xmm2
|
|
|
b40826 |
+ pmovmskb %xmm2, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%eax), %eax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%eax), %xmm3
|
|
|
b40826 |
+ pmovmskb %xmm3, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%eax), %eax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ and $-0x40, %eax
|
|
|
b40826 |
+ PUSH (%esi)
|
|
|
b40826 |
+ PUSH (%edi)
|
|
|
b40826 |
+ PUSH (%ebx)
|
|
|
b40826 |
+ PUSH (%ebp)
|
|
|
b40826 |
+ xor %ebp, %ebp
|
|
|
b40826 |
+L(aligned_64):
|
|
|
b40826 |
+ pcmpeqb (%eax), %xmm0
|
|
|
b40826 |
+ pcmpeqb 16(%eax), %xmm1
|
|
|
b40826 |
+ pcmpeqb 32(%eax), %xmm2
|
|
|
b40826 |
+ pcmpeqb 48(%eax), %xmm3
|
|
|
b40826 |
+ pmovmskb %xmm0, %edx
|
|
|
b40826 |
+ pmovmskb %xmm1, %esi
|
|
|
b40826 |
+ pmovmskb %xmm2, %edi
|
|
|
b40826 |
+ pmovmskb %xmm3, %ebx
|
|
|
b40826 |
+ or %edx, %ebp
|
|
|
b40826 |
+ or %esi, %ebp
|
|
|
b40826 |
+ or %edi, %ebp
|
|
|
b40826 |
+ or %ebx, %ebp
|
|
|
b40826 |
+ lea 64(%eax), %eax
|
|
|
b40826 |
+ jz L(aligned_64)
|
|
|
b40826 |
+L(48leave):
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ jnz L(aligned_64_exit_16)
|
|
|
b40826 |
+ test %esi, %esi
|
|
|
b40826 |
+ jnz L(aligned_64_exit_32)
|
|
|
b40826 |
+ test %edi, %edi
|
|
|
b40826 |
+ jnz L(aligned_64_exit_48)
|
|
|
b40826 |
+ mov %ebx, %edx
|
|
|
b40826 |
+ lea (%eax), %eax
|
|
|
b40826 |
+ jmp L(aligned_64_exit)
|
|
|
b40826 |
+L(aligned_64_exit_48):
|
|
|
b40826 |
+ lea -16(%eax), %eax
|
|
|
b40826 |
+ mov %edi, %edx
|
|
|
b40826 |
+ jmp L(aligned_64_exit)
|
|
|
b40826 |
+L(aligned_64_exit_32):
|
|
|
b40826 |
+ lea -32(%eax), %eax
|
|
|
b40826 |
+ mov %esi, %edx
|
|
|
b40826 |
+ jmp L(aligned_64_exit)
|
|
|
b40826 |
+L(aligned_64_exit_16):
|
|
|
b40826 |
+ lea -48(%eax), %eax
|
|
|
b40826 |
+L(aligned_64_exit):
|
|
|
b40826 |
+ POP (%ebp)
|
|
|
b40826 |
+ POP (%ebx)
|
|
|
b40826 |
+ POP (%edi)
|
|
|
b40826 |
+ POP (%esi)
|
|
|
b40826 |
+L(exit):
|
|
|
b40826 |
+ sub %ecx, %eax
|
|
|
b40826 |
+ test %dl, %dl
|
|
|
b40826 |
+ jz L(exit_high)
|
|
|
b40826 |
+ test $0x01, %dl
|
|
|
b40826 |
+ jnz L(exit_tail0)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x02, %dl
|
|
|
b40826 |
+ jnz L(exit_tail1)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x04, %dl
|
|
|
b40826 |
+ jnz L(exit_tail2)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x08, %dl
|
|
|
b40826 |
+ jnz L(exit_tail3)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x10, %dl
|
|
|
b40826 |
+ jnz L(exit_tail4)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x20, %dl
|
|
|
b40826 |
+ jnz L(exit_tail5)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x40, %dl
|
|
|
b40826 |
+ jnz L(exit_tail6)
|
|
|
b40826 |
+ add $7, %eax
|
|
|
b40826 |
+L(exit_tail0):
|
|
|
b40826 |
+ RETURN
|
|
|
b40826 |
+
|
|
|
b40826 |
+L(exit_high):
|
|
|
b40826 |
+ add $8, %eax
|
|
|
b40826 |
+ test $0x01, %dh
|
|
|
b40826 |
+ jnz L(exit_tail0)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x02, %dh
|
|
|
b40826 |
+ jnz L(exit_tail1)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x04, %dh
|
|
|
b40826 |
+ jnz L(exit_tail2)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x08, %dh
|
|
|
b40826 |
+ jnz L(exit_tail3)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x10, %dh
|
|
|
b40826 |
+ jnz L(exit_tail4)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x20, %dh
|
|
|
b40826 |
+ jnz L(exit_tail5)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x40, %dh
|
|
|
b40826 |
+ jnz L(exit_tail6)
|
|
|
b40826 |
+ add $7, %eax
|
|
|
b40826 |
+ RETURN
|
|
|
b40826 |
+
|
|
|
b40826 |
+ .p2align 4
|
|
|
b40826 |
+L(exit_tail1):
|
|
|
b40826 |
+ add $1, %eax
|
|
|
b40826 |
+ RETURN
|
|
|
b40826 |
+
|
|
|
b40826 |
+L(exit_tail2):
|
|
|
b40826 |
+ add $2, %eax
|
|
|
b40826 |
+ RETURN
|
|
|
b40826 |
+
|
|
|
b40826 |
+L(exit_tail3):
|
|
|
b40826 |
+ add $3, %eax
|
|
|
b40826 |
+ RETURN
|
|
|
b40826 |
+
|
|
|
b40826 |
+L(exit_tail4):
|
|
|
b40826 |
+ add $4, %eax
|
|
|
b40826 |
+ RETURN
|
|
|
b40826 |
+
|
|
|
b40826 |
+L(exit_tail5):
|
|
|
b40826 |
+ add $5, %eax
|
|
|
b40826 |
+ RETURN
|
|
|
b40826 |
+
|
|
|
b40826 |
+L(exit_tail6):
|
|
|
b40826 |
+ add $6, %eax
|
|
|
b40826 |
+ RETURN
|
|
|
b40826 |
+
|
|
|
b40826 |
+L(exit_tail7):
|
|
|
b40826 |
+ add $7, %eax
|
|
|
b40826 |
+ RETURN
|
|
|
b40826 |
+
|
|
|
b40826 |
+L(exit_tail8):
|
|
|
b40826 |
+ add $8, %eax
|
|
|
b40826 |
+ RETURN
|
|
|
b40826 |
+
|
|
|
b40826 |
+L(exit_tail9):
|
|
|
b40826 |
+ add $9, %eax
|
|
|
b40826 |
+ RETURN
|
|
|
b40826 |
+
|
|
|
b40826 |
+L(exit_tail10):
|
|
|
b40826 |
+ add $10, %eax
|
|
|
b40826 |
+ RETURN
|
|
|
b40826 |
+
|
|
|
b40826 |
+L(exit_tail11):
|
|
|
b40826 |
+ add $11, %eax
|
|
|
b40826 |
+ RETURN
|
|
|
b40826 |
+
|
|
|
b40826 |
+L(exit_tail12):
|
|
|
b40826 |
+ add $12, %eax
|
|
|
b40826 |
+ RETURN
|
|
|
b40826 |
+
|
|
|
b40826 |
+L(exit_tail13):
|
|
|
b40826 |
+ add $13, %eax
|
|
|
b40826 |
+ RETURN
|
|
|
b40826 |
+
|
|
|
b40826 |
+L(exit_tail14):
|
|
|
b40826 |
+ add $14, %eax
|
|
|
b40826 |
+ RETURN
|
|
|
b40826 |
+
|
|
|
b40826 |
+L(exit_tail15):
|
|
|
b40826 |
+ add $15, %eax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+
|
|
|
b40826 |
+END (__strlen_sse2)
|
|
|
b40826 |
+
|
|
|
b40826 |
+#endif
|
|
|
b40826 |
Index: glibc-2.12-2-gc4ccff1/sysdeps/i386/i686/multiarch/strlen.S
|
|
|
b40826 |
===================================================================
|
|
|
b40826 |
--- glibc-2.12-2-gc4ccff1.orig/sysdeps/i386/i686/multiarch/strlen.S
|
|
|
b40826 |
+++ glibc-2.12-2-gc4ccff1/sysdeps/i386/i686/multiarch/strlen.S
|
|
|
b40826 |
@@ -48,6 +48,9 @@ ENTRY(strlen)
|
|
|
b40826 |
1: leal __strlen_ia32@GOTOFF(%ebx), %eax
|
|
|
b40826 |
testl $bit_SSE2, CPUID_OFFSET+index_SSE2+__cpu_features@GOTOFF(%ebx)
|
|
|
b40826 |
jz 2f
|
|
|
b40826 |
+ leal __strlen_sse2_bsf@GOTOFF(%ebx), %eax
|
|
|
b40826 |
+ testl $bit_Slow_BSF, FEATURE_OFFSET+index_Slow_BSF+__cpu_features@GOTOFF(%ebx)
|
|
|
b40826 |
+ jz 2f
|
|
|
b40826 |
leal __strlen_sse2@GOTOFF(%ebx), %eax
|
|
|
b40826 |
2: popl %ebx
|
|
|
b40826 |
cfi_adjust_cfa_offset (-4);
|
|
|
b40826 |
@@ -55,84 +58,6 @@ ENTRY(strlen)
|
|
|
b40826 |
ret
|
|
|
b40826 |
END(strlen)
|
|
|
b40826 |
|
|
|
b40826 |
-#define CFI_POP(REG) \
|
|
|
b40826 |
- cfi_adjust_cfa_offset (-4); \
|
|
|
b40826 |
- cfi_restore (REG)
|
|
|
b40826 |
-
|
|
|
b40826 |
-#define RETURN popl %esi; CFI_POP (esi); ret
|
|
|
b40826 |
-
|
|
|
b40826 |
- .text
|
|
|
b40826 |
-ENTRY (__strlen_sse2)
|
|
|
b40826 |
-/*
|
|
|
b40826 |
- * This implementation uses SSE instructions to compare up to 16 bytes
|
|
|
b40826 |
- * at a time looking for the end of string (null char).
|
|
|
b40826 |
- */
|
|
|
b40826 |
- pushl %esi
|
|
|
b40826 |
- cfi_adjust_cfa_offset (4)
|
|
|
b40826 |
- cfi_rel_offset (%esi, 0)
|
|
|
b40826 |
- mov 8(%esp), %eax
|
|
|
b40826 |
- mov %eax, %ecx
|
|
|
b40826 |
- pxor %xmm0, %xmm0 /* 16 null chars */
|
|
|
b40826 |
- mov %eax, %esi
|
|
|
b40826 |
- and $15, %ecx
|
|
|
b40826 |
- jz 1f /* string is 16 byte aligned */
|
|
|
b40826 |
-
|
|
|
b40826 |
- /*
|
|
|
b40826 |
- * Unaligned case. Round down to 16-byte boundary before comparing
|
|
|
b40826 |
- * 16 bytes for a null char. The code then compensates for any extra chars
|
|
|
b40826 |
- * preceding the start of the string.
|
|
|
b40826 |
- */
|
|
|
b40826 |
- and $-16, %esi
|
|
|
b40826 |
-
|
|
|
b40826 |
- pcmpeqb (%esi), %xmm0
|
|
|
b40826 |
- lea 16(%eax), %esi
|
|
|
b40826 |
- pmovmskb %xmm0, %edx
|
|
|
b40826 |
-
|
|
|
b40826 |
- shr %cl, %edx /* Compensate for bytes preceding the string */
|
|
|
b40826 |
- test %edx, %edx
|
|
|
b40826 |
- jnz 2f
|
|
|
b40826 |
- sub %ecx, %esi /* no null, adjust to next 16-byte boundary */
|
|
|
b40826 |
- pxor %xmm0, %xmm0 /* clear xmm0, may have been changed... */
|
|
|
b40826 |
-
|
|
|
b40826 |
- .p2align 4
|
|
|
b40826 |
-1: /* 16 byte aligned */
|
|
|
b40826 |
- pcmpeqb (%esi), %xmm0 /* look for null bytes */
|
|
|
b40826 |
- pmovmskb %xmm0, %edx /* move each byte mask of %xmm0 to edx */
|
|
|
b40826 |
-
|
|
|
b40826 |
- add $16, %esi /* prepare to search next 16 bytes */
|
|
|
b40826 |
- test %edx, %edx /* if no null byte, %edx must be 0 */
|
|
|
b40826 |
- jnz 2f /* found a null */
|
|
|
b40826 |
-
|
|
|
b40826 |
- pcmpeqb (%esi), %xmm0
|
|
|
b40826 |
- pmovmskb %xmm0, %edx
|
|
|
b40826 |
- add $16, %esi
|
|
|
b40826 |
- test %edx, %edx
|
|
|
b40826 |
- jnz 2f
|
|
|
b40826 |
-
|
|
|
b40826 |
- pcmpeqb (%esi), %xmm0
|
|
|
b40826 |
- pmovmskb %xmm0, %edx
|
|
|
b40826 |
- add $16, %esi
|
|
|
b40826 |
- test %edx, %edx
|
|
|
b40826 |
- jnz 2f
|
|
|
b40826 |
-
|
|
|
b40826 |
- pcmpeqb (%esi), %xmm0
|
|
|
b40826 |
- pmovmskb %xmm0, %edx
|
|
|
b40826 |
- add $16, %esi
|
|
|
b40826 |
- test %edx, %edx
|
|
|
b40826 |
- jz 1b
|
|
|
b40826 |
-
|
|
|
b40826 |
-2:
|
|
|
b40826 |
- neg %eax
|
|
|
b40826 |
- lea -16(%eax, %esi), %eax /* calculate exact offset */
|
|
|
b40826 |
- bsf %edx, %ecx /* Least significant 1 bit is index of null */
|
|
|
b40826 |
- add %ecx, %eax
|
|
|
b40826 |
- popl %esi
|
|
|
b40826 |
- cfi_adjust_cfa_offset (-4)
|
|
|
b40826 |
- cfi_restore (%esi)
|
|
|
b40826 |
- ret
|
|
|
b40826 |
-
|
|
|
b40826 |
-END (__strlen_sse2)
|
|
|
b40826 |
-
|
|
|
b40826 |
# undef ENTRY
|
|
|
b40826 |
# define ENTRY(name) \
|
|
|
b40826 |
.type __strlen_ia32, @function; \
|
|
|
b40826 |
--- a/sysdeps/x86_64/multiarch/Makefile 2012-03-01 10:43:30.060487726 -0700
|
|
|
b40826 |
+++ b/sysdeps/x86_64/multiarch/Makefile 2012-03-01 10:45:57.894692115 -0700
|
|
|
b40826 |
@@ -7,7 +7,7 @@ ifeq ($(subdir),string)
|
|
|
b40826 |
sysdep_routines += stpncpy-c strncpy-c strcmp-ssse3 strncmp-ssse3 \
|
|
|
b40826 |
strend-sse4 memcmp-sse4 \
|
|
|
b40826 |
strcasestr-nonascii strcasecmp_l-ssse3 \
|
|
|
b40826 |
- strncase_l-ssse3 \
|
|
|
b40826 |
+ strncase_l-ssse3 strlen-sse4 strlen-no-bsf \
|
|
|
b40826 |
memset-x86-64
|
|
|
b40826 |
ifeq (yes,$(config-cflags-sse4))
|
|
|
b40826 |
sysdep_routines += strcspn-c strpbrk-c strspn-c strstr-c strcasestr-c
|
|
|
b40826 |
Index: glibc-2.12-2-gc4ccff1/sysdeps/x86_64/multiarch/init-arch.c
|
|
|
b40826 |
===================================================================
|
|
|
b40826 |
--- glibc-2.12-2-gc4ccff1.orig/sysdeps/x86_64/multiarch/init-arch.c
|
|
|
b40826 |
+++ glibc-2.12-2-gc4ccff1/sysdeps/x86_64/multiarch/init-arch.c
|
|
|
b40826 |
@@ -77,6 +77,12 @@ __init_cpu_features (void)
|
|
|
b40826 |
model += extended_model;
|
|
|
b40826 |
switch (model)
|
|
|
b40826 |
{
|
|
|
b40826 |
+ case 0x1c:
|
|
|
b40826 |
+ case 0x26:
|
|
|
b40826 |
+ /* BSF is slow on Atom. */
|
|
|
b40826 |
+ __cpu_features.feature[index_Slow_BSF] |= bit_Slow_BSF;
|
|
|
b40826 |
+ break;
|
|
|
b40826 |
+
|
|
|
b40826 |
case 0x1a:
|
|
|
b40826 |
case 0x1e:
|
|
|
b40826 |
case 0x1f:
|
|
|
b40826 |
--- a/sysdeps/x86_64/multiarch/init-arch.h 2012-03-01 10:43:30.061487720 -0700
|
|
|
b40826 |
+++ b/sysdeps/x86_64/multiarch/init-arch.h 2012-03-01 10:48:13.371963005 -0700
|
|
|
b40826 |
@@ -17,6 +17,7 @@
|
|
|
b40826 |
02111-1307 USA. */
|
|
|
b40826 |
|
|
|
b40826 |
#define bit_Fast_Rep_String (1 << 0)
|
|
|
b40826 |
+#define bit_Slow_BSF (1 << 2)
|
|
|
b40826 |
#define bit_Prefer_SSE_for_memop (1 << 3)
|
|
|
b40826 |
|
|
|
b40826 |
#ifdef __ASSEMBLER__
|
|
|
b40826 |
@@ -34,6 +35,7 @@
|
|
|
b40826 |
# define index_SSE4_2 COMMON_CPUID_INDEX_1*CPUID_SIZE+CPUID_ECX_OFFSET
|
|
|
b40826 |
|
|
|
b40826 |
#define index_Fast_Rep_String FEATURE_INDEX_1*FEATURE_SIZE
|
|
|
b40826 |
+# define index_Slow_BSF FEATURE_INDEX_1*FEATURE_SIZE
|
|
|
b40826 |
# define index_Prefer_SSE_for_memop FEATURE_INDEX_1*FEATURE_SIZE
|
|
|
b40826 |
|
|
|
b40826 |
#else /* __ASSEMBLER__ */
|
|
|
b40826 |
@@ -105,11 +107,15 @@ extern const struct cpu_features *__get_
|
|
|
b40826 |
# define HAS_FMA HAS_CPU_FEATURE (COMMON_CPUID_INDEX_1, ecx, 12)
|
|
|
b40826 |
|
|
|
b40826 |
# define index_Fast_Rep_String FEATURE_INDEX_1
|
|
|
b40826 |
+# define index_Slow_BSF FEATURE_INDEX_1
|
|
|
b40826 |
# define index_Prefer_SSE_for_memop FEATURE_INDEX_1
|
|
|
b40826 |
|
|
|
b40826 |
#define HAS_ARCH_FEATURE(idx, bit) \
|
|
|
b40826 |
((__get_cpu_features ()->feature[idx] & (bit)) != 0)
|
|
|
b40826 |
|
|
|
b40826 |
+#define HAS_SLOW_BSF \
|
|
|
b40826 |
+ HAS_ARCH_FEATURE (index_Slow_BSF, bit_Slow_BSF)
|
|
|
b40826 |
+
|
|
|
b40826 |
#define HAS_PREFER_SSE_FOR_MEMOP \
|
|
|
b40826 |
HAS_ARCH_FEATURE (index_Prefer_SSE_for_memop, bit_Prefer_SSE_for_memop)
|
|
|
b40826 |
|
|
|
b40826 |
Index: glibc-2.12-2-gc4ccff1/sysdeps/x86_64/multiarch/strlen-no-bsf.S
|
|
|
b40826 |
===================================================================
|
|
|
b40826 |
--- /dev/null
|
|
|
b40826 |
+++ glibc-2.12-2-gc4ccff1/sysdeps/x86_64/multiarch/strlen-no-bsf.S
|
|
|
b40826 |
@@ -0,0 +1,309 @@
|
|
|
b40826 |
+/* strlen without BSF
|
|
|
b40826 |
+ Copyright (C) 2010 Free Software Foundation, Inc.
|
|
|
b40826 |
+ Contributed by Intel Corporation.
|
|
|
b40826 |
+ This file is part of the GNU C Library.
|
|
|
b40826 |
+
|
|
|
b40826 |
+ The GNU C Library is free software; you can redistribute it and/or
|
|
|
b40826 |
+ modify it under the terms of the GNU Lesser General Public
|
|
|
b40826 |
+ License as published by the Free Software Foundation; either
|
|
|
b40826 |
+ version 2.1 of the License, or (at your option) any later version.
|
|
|
b40826 |
+
|
|
|
b40826 |
+ The GNU C Library is distributed in the hope that it will be useful,
|
|
|
b40826 |
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
b40826 |
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
|
b40826 |
+ Lesser General Public License for more details.
|
|
|
b40826 |
+
|
|
|
b40826 |
+ You should have received a copy of the GNU Lesser General Public
|
|
|
b40826 |
+ License along with the GNU C Library; if not, write to the Free
|
|
|
b40826 |
+ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
|
|
|
b40826 |
+ 02111-1307 USA. */
|
|
|
b40826 |
+
|
|
|
b40826 |
+#if defined SHARED && !defined NOT_IN_libc
|
|
|
b40826 |
+
|
|
|
b40826 |
+#include <sysdep.h>
|
|
|
b40826 |
+
|
|
|
b40826 |
+ .section .text.slow,"ax",@progbits
|
|
|
b40826 |
+ENTRY (__strlen_no_bsf)
|
|
|
b40826 |
+ xor %eax, %eax
|
|
|
b40826 |
+ cmpb $0, (%rdi)
|
|
|
b40826 |
+ jz L(exit_tail0)
|
|
|
b40826 |
+ cmpb $0, 1(%rdi)
|
|
|
b40826 |
+ jz L(exit_tail1)
|
|
|
b40826 |
+ cmpb $0, 2(%rdi)
|
|
|
b40826 |
+ jz L(exit_tail2)
|
|
|
b40826 |
+ cmpb $0, 3(%rdi)
|
|
|
b40826 |
+ jz L(exit_tail3)
|
|
|
b40826 |
+ cmpb $0, 4(%rdi)
|
|
|
b40826 |
+ jz L(exit_tail4)
|
|
|
b40826 |
+ cmpb $0, 5(%rdi)
|
|
|
b40826 |
+ jz L(exit_tail5)
|
|
|
b40826 |
+ cmpb $0, 6(%rdi)
|
|
|
b40826 |
+ jz L(exit_tail6)
|
|
|
b40826 |
+ cmpb $0, 7(%rdi)
|
|
|
b40826 |
+ jz L(exit_tail7)
|
|
|
b40826 |
+ cmpb $0, 8(%rdi)
|
|
|
b40826 |
+ jz L(exit_tail8)
|
|
|
b40826 |
+ cmpb $0, 9(%rdi)
|
|
|
b40826 |
+ jz L(exit_tail9)
|
|
|
b40826 |
+ cmpb $0, 10(%rdi)
|
|
|
b40826 |
+ jz L(exit_tail10)
|
|
|
b40826 |
+ cmpb $0, 11(%rdi)
|
|
|
b40826 |
+ jz L(exit_tail11)
|
|
|
b40826 |
+ cmpb $0, 12(%rdi)
|
|
|
b40826 |
+ jz L(exit_tail12)
|
|
|
b40826 |
+ cmpb $0, 13(%rdi)
|
|
|
b40826 |
+ jz L(exit_tail13)
|
|
|
b40826 |
+ cmpb $0, 14(%rdi)
|
|
|
b40826 |
+ jz L(exit_tail14)
|
|
|
b40826 |
+ cmpb $0, 15(%rdi)
|
|
|
b40826 |
+ jz L(exit_tail15)
|
|
|
b40826 |
+ pxor %xmm0, %xmm0
|
|
|
b40826 |
+ mov %rdi, %rcx
|
|
|
b40826 |
+ mov %rdi, %rax
|
|
|
b40826 |
+ and $-16, %rax
|
|
|
b40826 |
+ add $16, %rax
|
|
|
b40826 |
+ add $16, %rcx
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%rax), %xmm0
|
|
|
b40826 |
+ pmovmskb %xmm0, %edx
|
|
|
b40826 |
+ pxor %xmm1, %xmm1
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%rax), %rax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%rax), %xmm1
|
|
|
b40826 |
+ pmovmskb %xmm1, %edx
|
|
|
b40826 |
+ pxor %xmm2, %xmm2
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%rax), %rax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%rax), %xmm2
|
|
|
b40826 |
+ pmovmskb %xmm2, %edx
|
|
|
b40826 |
+ pxor %xmm3, %xmm3
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%rax), %rax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%rax), %xmm3
|
|
|
b40826 |
+ pmovmskb %xmm3, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%rax), %rax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%rax), %xmm0
|
|
|
b40826 |
+ pmovmskb %xmm0, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%rax), %rax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%rax), %xmm1
|
|
|
b40826 |
+ pmovmskb %xmm1, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%rax), %rax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%rax), %xmm2
|
|
|
b40826 |
+ pmovmskb %xmm2, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%rax), %rax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%rax), %xmm3
|
|
|
b40826 |
+ pmovmskb %xmm3, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%rax), %rax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%rax), %xmm0
|
|
|
b40826 |
+ pmovmskb %xmm0, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%rax), %rax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%rax), %xmm1
|
|
|
b40826 |
+ pmovmskb %xmm1, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%rax), %rax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%rax), %xmm2
|
|
|
b40826 |
+ pmovmskb %xmm2, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%rax), %rax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%rax), %xmm3
|
|
|
b40826 |
+ pmovmskb %xmm3, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%rax), %rax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%rax), %xmm0
|
|
|
b40826 |
+ pmovmskb %xmm0, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%rax), %rax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%rax), %xmm1
|
|
|
b40826 |
+ pmovmskb %xmm1, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%rax), %rax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%rax), %xmm2
|
|
|
b40826 |
+ pmovmskb %xmm2, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%rax), %rax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb (%rax), %xmm3
|
|
|
b40826 |
+ pmovmskb %xmm3, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ lea 16(%rax), %rax
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ and $-0x40, %rax
|
|
|
b40826 |
+ xor %r8d, %r8d
|
|
|
b40826 |
+L(aligned_64):
|
|
|
b40826 |
+ pcmpeqb (%rax), %xmm0
|
|
|
b40826 |
+ pcmpeqb 16(%rax), %xmm1
|
|
|
b40826 |
+ pcmpeqb 32(%rax), %xmm2
|
|
|
b40826 |
+ pcmpeqb 48(%rax), %xmm3
|
|
|
b40826 |
+ pmovmskb %xmm0, %edx
|
|
|
b40826 |
+ pmovmskb %xmm1, %esi
|
|
|
b40826 |
+ pmovmskb %xmm2, %edi
|
|
|
b40826 |
+ pmovmskb %xmm3, %r9d
|
|
|
b40826 |
+ or %edx, %r8d
|
|
|
b40826 |
+ or %esi, %r8d
|
|
|
b40826 |
+ or %edi, %r8d
|
|
|
b40826 |
+ or %r9d, %r8d
|
|
|
b40826 |
+ lea 64(%rax), %rax
|
|
|
b40826 |
+ jz L(aligned_64)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ jnz L(aligned_64_exit_16)
|
|
|
b40826 |
+ test %esi, %esi
|
|
|
b40826 |
+ jnz L(aligned_64_exit_32)
|
|
|
b40826 |
+ test %edi, %edi
|
|
|
b40826 |
+ jnz L(aligned_64_exit_48)
|
|
|
b40826 |
+L(aligned_64_exit_64):
|
|
|
b40826 |
+ mov %r9d, %edx
|
|
|
b40826 |
+ jmp L(aligned_64_exit)
|
|
|
b40826 |
+L(aligned_64_exit_48):
|
|
|
b40826 |
+ lea -16(%rax), %rax
|
|
|
b40826 |
+ mov %edi, %edx
|
|
|
b40826 |
+ jmp L(aligned_64_exit)
|
|
|
b40826 |
+L(aligned_64_exit_32):
|
|
|
b40826 |
+ lea -32(%rax), %rax
|
|
|
b40826 |
+ mov %esi, %edx
|
|
|
b40826 |
+ jmp L(aligned_64_exit)
|
|
|
b40826 |
+L(aligned_64_exit_16):
|
|
|
b40826 |
+ lea -48(%rax), %rax
|
|
|
b40826 |
+L(aligned_64_exit):
|
|
|
b40826 |
+L(exit):
|
|
|
b40826 |
+ sub %rcx, %rax
|
|
|
b40826 |
+ test %dl, %dl
|
|
|
b40826 |
+ jz L(exit_high)
|
|
|
b40826 |
+ test $0x01, %dl
|
|
|
b40826 |
+ jnz L(exit_tail0)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x02, %dl
|
|
|
b40826 |
+ jnz L(exit_tail1)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x04, %dl
|
|
|
b40826 |
+ jnz L(exit_tail2)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x08, %dl
|
|
|
b40826 |
+ jnz L(exit_tail3)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x10, %dl
|
|
|
b40826 |
+ jnz L(exit_tail4)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x20, %dl
|
|
|
b40826 |
+ jnz L(exit_tail5)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x40, %dl
|
|
|
b40826 |
+ jnz L(exit_tail6)
|
|
|
b40826 |
+ add $7, %eax
|
|
|
b40826 |
+L(exit_tail0):
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+
|
|
|
b40826 |
+L(exit_high):
|
|
|
b40826 |
+ add $8, %eax
|
|
|
b40826 |
+ test $0x01, %dh
|
|
|
b40826 |
+ jnz L(exit_tail0)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x02, %dh
|
|
|
b40826 |
+ jnz L(exit_tail1)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x04, %dh
|
|
|
b40826 |
+ jnz L(exit_tail2)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x08, %dh
|
|
|
b40826 |
+ jnz L(exit_tail3)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x10, %dh
|
|
|
b40826 |
+ jnz L(exit_tail4)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x20, %dh
|
|
|
b40826 |
+ jnz L(exit_tail5)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ test $0x40, %dh
|
|
|
b40826 |
+ jnz L(exit_tail6)
|
|
|
b40826 |
+ add $7, %eax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+ .p2align 4
|
|
|
b40826 |
+L(exit_tail1):
|
|
|
b40826 |
+ add $1, %eax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+
|
|
|
b40826 |
+L(exit_tail2):
|
|
|
b40826 |
+ add $2, %eax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+
|
|
|
b40826 |
+L(exit_tail3):
|
|
|
b40826 |
+ add $3, %eax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+
|
|
|
b40826 |
+L(exit_tail4):
|
|
|
b40826 |
+ add $4, %eax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+
|
|
|
b40826 |
+L(exit_tail5):
|
|
|
b40826 |
+ add $5, %eax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+L(exit_tail6):
|
|
|
b40826 |
+ add $6, %eax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+L(exit_tail7):
|
|
|
b40826 |
+ add $7, %eax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+L(exit_tail8):
|
|
|
b40826 |
+ add $8, %eax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+L(exit_tail9):
|
|
|
b40826 |
+ add $9, %eax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+L(exit_tail10):
|
|
|
b40826 |
+ add $10, %eax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+L(exit_tail11):
|
|
|
b40826 |
+ add $11, %eax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+L(exit_tail12):
|
|
|
b40826 |
+ add $12, %eax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+L(exit_tail13):
|
|
|
b40826 |
+ add $13, %eax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+L(exit_tail14):
|
|
|
b40826 |
+ add $14, %eax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+L(exit_tail15):
|
|
|
b40826 |
+ add $15, %eax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+END (__strlen_no_bsf)
|
|
|
b40826 |
+
|
|
|
b40826 |
+#endif
|
|
|
b40826 |
Index: glibc-2.12-2-gc4ccff1/sysdeps/x86_64/multiarch/strlen-sse4.S
|
|
|
b40826 |
===================================================================
|
|
|
b40826 |
--- /dev/null
|
|
|
b40826 |
+++ glibc-2.12-2-gc4ccff1/sysdeps/x86_64/multiarch/strlen-sse4.S
|
|
|
b40826 |
@@ -0,0 +1,85 @@
|
|
|
b40826 |
+/* strlen with SSE4
|
|
|
b40826 |
+ Copyright (C) 2009, 2010 Free Software Foundation, Inc.
|
|
|
b40826 |
+ Contributed by Ulrich Drepper <drepper@redhat.com>.
|
|
|
b40826 |
+ This file is part of the GNU C Library.
|
|
|
b40826 |
+
|
|
|
b40826 |
+ The GNU C Library is free software; you can redistribute it and/or
|
|
|
b40826 |
+ modify it under the terms of the GNU Lesser General Public
|
|
|
b40826 |
+ License as published by the Free Software Foundation; either
|
|
|
b40826 |
+ version 2.1 of the License, or (at your option) any later version.
|
|
|
b40826 |
+
|
|
|
b40826 |
+ The GNU C Library is distributed in the hope that it will be useful,
|
|
|
b40826 |
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
b40826 |
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
|
b40826 |
+ Lesser General Public License for more details.
|
|
|
b40826 |
+
|
|
|
b40826 |
+ You should have received a copy of the GNU Lesser General Public
|
|
|
b40826 |
+ License along with the GNU C Library; if not, write to the Free
|
|
|
b40826 |
+ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
|
|
|
b40826 |
+ 02111-1307 USA. */
|
|
|
b40826 |
+
|
|
|
b40826 |
+#if defined SHARED && !defined NOT_IN_libc
|
|
|
b40826 |
+
|
|
|
b40826 |
+#include <sysdep.h>
|
|
|
b40826 |
+
|
|
|
b40826 |
+ .section .text.sse4.2,"ax",@progbits
|
|
|
b40826 |
+ENTRY (__strlen_sse42)
|
|
|
b40826 |
+ pxor %xmm1, %xmm1
|
|
|
b40826 |
+ movl %edi, %ecx
|
|
|
b40826 |
+ movq %rdi, %r8
|
|
|
b40826 |
+ andq $~15, %rdi
|
|
|
b40826 |
+ xor %edi, %ecx
|
|
|
b40826 |
+ pcmpeqb (%rdi), %xmm1
|
|
|
b40826 |
+ pmovmskb %xmm1, %edx
|
|
|
b40826 |
+ shrl %cl, %edx
|
|
|
b40826 |
+ shll %cl, %edx
|
|
|
b40826 |
+ andl %edx, %edx
|
|
|
b40826 |
+ jnz L(less16bytes)
|
|
|
b40826 |
+ pxor %xmm1, %xmm1
|
|
|
b40826 |
+
|
|
|
b40826 |
+ .p2align 4
|
|
|
b40826 |
+L(more64bytes_loop):
|
|
|
b40826 |
+ pcmpistri $0x08, 16(%rdi), %xmm1
|
|
|
b40826 |
+ jz L(more32bytes)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpistri $0x08, 32(%rdi), %xmm1
|
|
|
b40826 |
+ jz L(more48bytes)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpistri $0x08, 48(%rdi), %xmm1
|
|
|
b40826 |
+ jz L(more64bytes)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ add $64, %rdi
|
|
|
b40826 |
+ pcmpistri $0x08, (%rdi), %xmm1
|
|
|
b40826 |
+ jnz L(more64bytes_loop)
|
|
|
b40826 |
+ leaq (%rdi,%rcx), %rax
|
|
|
b40826 |
+ subq %r8, %rax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+
|
|
|
b40826 |
+ .p2align 4
|
|
|
b40826 |
+L(more32bytes):
|
|
|
b40826 |
+ leaq 16(%rdi,%rcx, 1), %rax
|
|
|
b40826 |
+ subq %r8, %rax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+
|
|
|
b40826 |
+ .p2align 4
|
|
|
b40826 |
+L(more48bytes):
|
|
|
b40826 |
+ leaq 32(%rdi,%rcx, 1), %rax
|
|
|
b40826 |
+ subq %r8, %rax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+
|
|
|
b40826 |
+ .p2align 4
|
|
|
b40826 |
+L(more64bytes):
|
|
|
b40826 |
+ leaq 48(%rdi,%rcx, 1), %rax
|
|
|
b40826 |
+ subq %r8, %rax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+
|
|
|
b40826 |
+ .p2align 4
|
|
|
b40826 |
+L(less16bytes):
|
|
|
b40826 |
+ subq %r8, %rdi
|
|
|
b40826 |
+ bsfl %edx, %eax
|
|
|
b40826 |
+ addq %rdi, %rax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+
|
|
|
b40826 |
+END (__strlen_sse42)
|
|
|
b40826 |
+
|
|
|
b40826 |
+#endif
|
|
|
b40826 |
Index: glibc-2.12-2-gc4ccff1/sysdeps/x86_64/multiarch/strlen.S
|
|
|
b40826 |
===================================================================
|
|
|
b40826 |
--- glibc-2.12-2-gc4ccff1.orig/sysdeps/x86_64/multiarch/strlen.S
|
|
|
b40826 |
+++ glibc-2.12-2-gc4ccff1/sysdeps/x86_64/multiarch/strlen.S
|
|
|
b40826 |
@@ -36,74 +36,12 @@ ENTRY(strlen)
|
|
|
b40826 |
testl $bit_SSE4_2, __cpu_features+CPUID_OFFSET+index_SSE4_2(%rip)
|
|
|
b40826 |
jz 2f
|
|
|
b40826 |
leaq __strlen_sse42(%rip), %rax
|
|
|
b40826 |
-2: ret
|
|
|
b40826 |
-END(strlen)
|
|
|
b40826 |
-
|
|
|
b40826 |
-
|
|
|
b40826 |
- .section .text.sse4.2,"ax",@progbits
|
|
|
b40826 |
- .align 16
|
|
|
b40826 |
- .type __strlen_sse42, @function
|
|
|
b40826 |
-__strlen_sse42:
|
|
|
b40826 |
- cfi_startproc
|
|
|
b40826 |
- CALL_MCOUNT
|
|
|
b40826 |
- pxor %xmm1, %xmm1
|
|
|
b40826 |
- movl %edi, %ecx
|
|
|
b40826 |
- movq %rdi, %r8
|
|
|
b40826 |
- andq $~15, %rdi
|
|
|
b40826 |
- xor %edi, %ecx
|
|
|
b40826 |
- pcmpeqb (%rdi), %xmm1
|
|
|
b40826 |
- pmovmskb %xmm1, %edx
|
|
|
b40826 |
- shrl %cl, %edx
|
|
|
b40826 |
- shll %cl, %edx
|
|
|
b40826 |
- andl %edx, %edx
|
|
|
b40826 |
- jnz L(less16bytes)
|
|
|
b40826 |
- pxor %xmm1, %xmm1
|
|
|
b40826 |
-
|
|
|
b40826 |
- .p2align 4
|
|
|
b40826 |
-L(more64bytes_loop):
|
|
|
b40826 |
- pcmpistri $0x08, 16(%rdi), %xmm1
|
|
|
b40826 |
- jz L(more32bytes)
|
|
|
b40826 |
-
|
|
|
b40826 |
- pcmpistri $0x08, 32(%rdi), %xmm1
|
|
|
b40826 |
- jz L(more48bytes)
|
|
|
b40826 |
-
|
|
|
b40826 |
- pcmpistri $0x08, 48(%rdi), %xmm1
|
|
|
b40826 |
- jz L(more64bytes)
|
|
|
b40826 |
-
|
|
|
b40826 |
- add $64, %rdi
|
|
|
b40826 |
- pcmpistri $0x08, (%rdi), %xmm1
|
|
|
b40826 |
- jnz L(more64bytes_loop)
|
|
|
b40826 |
- leaq (%rdi,%rcx), %rax
|
|
|
b40826 |
- subq %r8, %rax
|
|
|
b40826 |
- ret
|
|
|
b40826 |
-
|
|
|
b40826 |
- .p2align 4
|
|
|
b40826 |
-L(more32bytes):
|
|
|
b40826 |
- leaq 16(%rdi,%rcx, 1), %rax
|
|
|
b40826 |
- subq %r8, %rax
|
|
|
b40826 |
- ret
|
|
|
b40826 |
-
|
|
|
b40826 |
- .p2align 4
|
|
|
b40826 |
-L(more48bytes):
|
|
|
b40826 |
- leaq 32(%rdi,%rcx, 1), %rax
|
|
|
b40826 |
- subq %r8, %rax
|
|
|
b40826 |
- ret
|
|
|
b40826 |
-
|
|
|
b40826 |
- .p2align 4
|
|
|
b40826 |
-L(more64bytes):
|
|
|
b40826 |
- leaq 48(%rdi,%rcx, 1), %rax
|
|
|
b40826 |
- subq %r8, %rax
|
|
|
b40826 |
ret
|
|
|
b40826 |
-
|
|
|
b40826 |
- .p2align 4
|
|
|
b40826 |
-L(less16bytes):
|
|
|
b40826 |
- subq %r8, %rdi
|
|
|
b40826 |
- bsfl %edx, %eax
|
|
|
b40826 |
- addq %rdi, %rax
|
|
|
b40826 |
- ret
|
|
|
b40826 |
- cfi_endproc
|
|
|
b40826 |
- .size __strlen_sse42, .-__strlen_sse42
|
|
|
b40826 |
-
|
|
|
b40826 |
+2: testl $bit_Slow_BSF, __cpu_features+FEATURE_OFFSET+index_Slow_BSF(%rip)
|
|
|
b40826 |
+ jz 3f
|
|
|
b40826 |
+ leaq __strlen_no_bsf(%rip), %rax
|
|
|
b40826 |
+3: ret
|
|
|
b40826 |
+END(strlen)
|
|
|
b40826 |
|
|
|
b40826 |
# undef ENTRY
|
|
|
b40826 |
# define ENTRY(name) \
|
|
|
b40826 |
Index: glibc-2.12-2-gc4ccff1/sysdeps/x86_64/strlen.S
|
|
|
b40826 |
===================================================================
|
|
|
b40826 |
--- glibc-2.12-2-gc4ccff1.orig/sysdeps/x86_64/strlen.S
|
|
|
b40826 |
+++ glibc-2.12-2-gc4ccff1/sysdeps/x86_64/strlen.S
|
|
|
b40826 |
@@ -23,29 +23,80 @@
|
|
|
b40826 |
|
|
|
b40826 |
.text
|
|
|
b40826 |
ENTRY(strlen)
|
|
|
b40826 |
- pxor %xmm2, %xmm2
|
|
|
b40826 |
- movq %rdi, %rcx
|
|
|
b40826 |
- movq %rdi, %r8
|
|
|
b40826 |
- andq $~15, %rdi
|
|
|
b40826 |
- movdqa %xmm2, %xmm1
|
|
|
b40826 |
- pcmpeqb (%rdi), %xmm2
|
|
|
b40826 |
- orl $0xffffffff, %esi
|
|
|
b40826 |
- subq %rdi, %rcx
|
|
|
b40826 |
- shll %cl, %esi
|
|
|
b40826 |
- pmovmskb %xmm2, %edx
|
|
|
b40826 |
- andl %esi, %edx
|
|
|
b40826 |
- jnz 1f
|
|
|
b40826 |
-
|
|
|
b40826 |
-2: movdqa 16(%rdi), %xmm0
|
|
|
b40826 |
- leaq 16(%rdi), %rdi
|
|
|
b40826 |
+ xor %rax, %rax
|
|
|
b40826 |
+ mov %edi, %ecx
|
|
|
b40826 |
+ and $0x3f, %ecx
|
|
|
b40826 |
+ pxor %xmm0, %xmm0
|
|
|
b40826 |
+ cmp $0x30, %ecx
|
|
|
b40826 |
+ ja L(next)
|
|
|
b40826 |
+ movdqu (%rdi), %xmm1
|
|
|
b40826 |
pcmpeqb %xmm1, %xmm0
|
|
|
b40826 |
pmovmskb %xmm0, %edx
|
|
|
b40826 |
- testl %edx, %edx
|
|
|
b40826 |
- jz 2b
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ jnz L(exit_less16)
|
|
|
b40826 |
+ mov %rdi, %rax
|
|
|
b40826 |
+ and $-16, %rax
|
|
|
b40826 |
+ jmp L(align16_start)
|
|
|
b40826 |
+L(next):
|
|
|
b40826 |
+ mov %rdi, %rax
|
|
|
b40826 |
+ and $-16, %rax
|
|
|
b40826 |
+ pcmpeqb (%rax), %xmm0
|
|
|
b40826 |
+ mov $-1, %esi
|
|
|
b40826 |
+ sub %rax, %rcx
|
|
|
b40826 |
+ shl %cl, %esi
|
|
|
b40826 |
+ pmovmskb %xmm0, %edx
|
|
|
b40826 |
+ and %esi, %edx
|
|
|
b40826 |
+ jnz L(exit)
|
|
|
b40826 |
+L(align16_start):
|
|
|
b40826 |
+ pxor %xmm0, %xmm0
|
|
|
b40826 |
+ pxor %xmm1, %xmm1
|
|
|
b40826 |
+ pxor %xmm2, %xmm2
|
|
|
b40826 |
+ pxor %xmm3, %xmm3
|
|
|
b40826 |
+ .p2align 4
|
|
|
b40826 |
+L(align16_loop):
|
|
|
b40826 |
+ pcmpeqb 16(%rax), %xmm0
|
|
|
b40826 |
+ pmovmskb %xmm0, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ jnz L(exit16)
|
|
|
b40826 |
|
|
|
b40826 |
-1: subq %r8, %rdi
|
|
|
b40826 |
- bsfl %edx, %eax
|
|
|
b40826 |
- addq %rdi, %rax
|
|
|
b40826 |
+ pcmpeqb 32(%rax), %xmm1
|
|
|
b40826 |
+ pmovmskb %xmm1, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ jnz L(exit32)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb 48(%rax), %xmm2
|
|
|
b40826 |
+ pmovmskb %xmm2, %edx
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ jnz L(exit48)
|
|
|
b40826 |
+
|
|
|
b40826 |
+ pcmpeqb 64(%rax), %xmm3
|
|
|
b40826 |
+ pmovmskb %xmm3, %edx
|
|
|
b40826 |
+ lea 64(%rax), %rax
|
|
|
b40826 |
+ test %edx, %edx
|
|
|
b40826 |
+ jz L(align16_loop)
|
|
|
b40826 |
+L(exit):
|
|
|
b40826 |
+ sub %rdi, %rax
|
|
|
b40826 |
+L(exit_less16):
|
|
|
b40826 |
+ bsf %rdx, %rdx
|
|
|
b40826 |
+ add %rdx, %rax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+ .p2align 4
|
|
|
b40826 |
+L(exit16):
|
|
|
b40826 |
+ sub %rdi, %rax
|
|
|
b40826 |
+ bsf %rdx, %rdx
|
|
|
b40826 |
+ lea 16(%rdx,%rax), %rax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+ .p2align 4
|
|
|
b40826 |
+L(exit32):
|
|
|
b40826 |
+ sub %rdi, %rax
|
|
|
b40826 |
+ bsf %rdx, %rdx
|
|
|
b40826 |
+ lea 32(%rdx,%rax), %rax
|
|
|
b40826 |
+ ret
|
|
|
b40826 |
+ .p2align 4
|
|
|
b40826 |
+L(exit48):
|
|
|
b40826 |
+ sub %rdi, %rax
|
|
|
b40826 |
+ bsf %rdx, %rdx
|
|
|
b40826 |
+ lea 48(%rdx,%rax), %rax
|
|
|
b40826 |
ret
|
|
|
b40826 |
END(strlen)
|
|
|
b40826 |
libc_hidden_builtin_def (strlen)
|