Add x86 SSE strlen.

diff --git a/ChangeLog b/ChangeLog
index 149f66f..94cfd06 100644 (file)
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,5 +1,11 @@
 2009-08-03  H.J. Lu  <hongjiu.lu@intel.com>
 
+       * sysdeps/i386/i686/multiarch/strcspn.S: Add comments for no
+       hidden IFUNC functions.
+       * sysdeps/i386/i686/multiarch/strspn.S: Likewise.
+
+       * sysdeps/i386/i686/multiarch/strlen.S: New file.
+
        * sysdeps/i386/i686/multiarch/Makefile [subdir=string]
        (sysdep_routines): Add strcspn-c, strpbrk-c, strspn-c, strstr-c, and
        strcasestr-c.

diff --git a/sysdeps/i386/i686/multiarch/strcspn.S b/sysdeps/i386/i686/multiarch/strcspn.S
index f5ca092..ca9c3f5 100644 (file)
--- a/sysdeps/i386/i686/multiarch/strcspn.S
+++ b/sysdeps/i386/i686/multiarch/strcspn.S
@@ -83,9 +83,9 @@ END(STRCSPN)
 # define END(name) \
        cfi_endproc; .size STRCSPN_IA32, .-STRCSPN_IA32
 # undef libc_hidden_builtin_def
-/* It doesn't make sense to send libc-internal strcspn calls through a PLT.
-   The speedup we get from using SSE4.2 instruction is likely eaten away
-   by the indirect call in the PLT.  */
+/* IFUNC doesn't work with the hidden functions in shared library since
+   they will be called without setting up EBX needed for PLT which is
+   used by IFUNC.  */
 # define libc_hidden_builtin_def(name) \
        .globl __GI_STRCSPN; __GI_STRCSPN = STRCSPN_IA32
 #endif

diff --git a/sysdeps/i386/i686/multiarch/strlen.S b/sysdeps/i386/i686/multiarch/strlen.S
new file mode 100644 (file)
index 0000000..3d4d9f1
--- /dev/null
+++ b/sysdeps/i386/i686/multiarch/strlen.S
@@ -0,0 +1,154 @@
+/* Multiple versions of strlen
+   Copyright (C) 2009 Free Software Foundation, Inc.
+   Contributed by Intel Corporation.
+   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 <ifunc-defines.h>
+
+/* Define multiple versions only for the definition in libc and for the
+   DSO.  In static binaries, we need strlen before the initialization
+   happened.  */
+#if defined SHARED && !defined NOT_IN_libc
+       .section        .gnu.linkonce.t.__i686.get_pc_thunk.bx,"ax",@progbits
+       .globl  __i686.get_pc_thunk.bx
+       .hidden __i686.get_pc_thunk.bx
+       .p2align 4
+       .type   __i686.get_pc_thunk.bx,@function
+__i686.get_pc_thunk.bx:
+       movl    (%esp), %ebx
+       ret
+
+       .text
+ENTRY(strlen)
+       .type   strlen, @gnu_indirect_function
+       pushl   %ebx
+       cfi_adjust_cfa_offset (4)
+       cfi_rel_offset (ebx, 0)
+       call    __i686.get_pc_thunk.bx
+       addl    $_GLOBAL_OFFSET_TABLE_, %ebx
+       cmpl    $0, KIND_OFFSET+__cpu_features@GOTOFF(%ebx)
+       jne     1f
+       call    __init_cpu_features
+1:     leal    __strlen_ia32@GOTOFF(%ebx), %eax
+       testl   $(1<<26), CPUID_OFFSET+COMMON_CPUID_INDEX_1*CPUID_SIZE+CPUID_EDX_OFFSET+__cpu_features@GOTOFF(%ebx)
+       jz      2f
+       leal    __strlen_sse2@GOTOFF(%ebx), %eax
+2:     popl    %ebx
+       cfi_adjust_cfa_offset (-4);
+       cfi_restore (ebx)
+       ret
+END(strlen)
+
+#define CFI_POP(REG)                                           \
+  cfi_adjust_cfa_offset (-4);                                  \
+  cfi_restore (REG)
+
+#define RETURN         popl %esi; CFI_POP (esi); ret
+
+       .text
+ENTRY (__strlen_sse2)
+/*
+ * This implementation uses SSE instructions to compare up to 16 bytes
+ * at a time looking for the end of string (null char).
+ */
+       pushl   %esi
+       cfi_adjust_cfa_offset (4)
+       cfi_rel_offset (%esi, 0)
+       mov     8(%esp), %eax
+       mov     %eax, %ecx
+       pxor    %xmm0, %xmm0            /* 16 null chars */
+       mov     %eax, %esi
+       and     $15, %ecx
+       jz      1f                      /* string is 16 byte aligned */
+
+       /*
+       * Unaligned case. Round down to 16-byte boundary before comparing
+       * 16 bytes for a null char. The code then compensates for any extra chars
+       * preceding the start of the string.
+       */
+       and     $-16, %esi
+
+       pcmpeqb (%esi), %xmm0
+       lea     16(%eax), %esi
+       pmovmskb %xmm0, %edx
+
+       shr     %cl, %edx               /* Compensate for bytes preceding the string */
+       test    %edx, %edx
+       jnz     2f
+       sub     %ecx, %esi              /* no null, adjust to next 16-byte boundary */
+       pxor    %xmm0, %xmm0            /* clear xmm0, may have been changed... */
+
+       .p2align 4
+1:                                     /* 16 byte aligned */
+       pcmpeqb (%esi), %xmm0           /* look for null bytes */
+       pmovmskb %xmm0, %edx            /* move each byte mask of %xmm0 to edx */
+
+       add     $16, %esi               /* prepare to search next 16 bytes */
+       test    %edx, %edx              /* if no null byte, %edx must be 0 */
+       jnz     2f                      /* found a null */
+
+       pcmpeqb (%esi), %xmm0
+       pmovmskb %xmm0, %edx
+       add     $16, %esi
+       test    %edx, %edx
+       jnz     2f
+
+       pcmpeqb (%esi), %xmm0
+       pmovmskb %xmm0, %edx
+       add     $16, %esi
+       test    %edx, %edx
+       jnz     2f
+
+       pcmpeqb (%esi), %xmm0
+       pmovmskb %xmm0, %edx
+       add     $16, %esi
+       test    %edx, %edx
+       jz      1b
+
+2:
+       neg     %eax
+       lea     -16(%eax, %esi), %eax   /* calculate exact offset */
+       bsf     %edx, %ecx              /* Least significant 1 bit is index of null */
+       add     %ecx, %eax
+       popl    %esi
+       cfi_adjust_cfa_offset (-4)
+       cfi_restore (%esi)
+       ret
+
+END (__strlen_sse2)
+
+# undef ENTRY
+# define ENTRY(name) \
+       .type __strlen_ia32, @function; \
+       .globl __strlen_ia32; \
+       .p2align 4
+       __strlen_ia32: cfi_startproc; \
+       CALL_MCOUNT
+# undef END
+# define END(name) \
+       cfi_endproc; .size __strlen_ia32, .-__strlen_ia32
+# undef libc_hidden_builtin_def
+/* IFUNC doesn't work with the hidden functions in shared library since
+   they will be called without setting up EBX needed for PLT which is
+   used by IFUNC.  */
+# define libc_hidden_builtin_def(name) \
+       .globl __GI_strlen; __GI_strlen = __strlen_ia32
+#endif
+
+#include "../../i586/strlen.S"

diff --git a/sysdeps/i386/i686/multiarch/strspn.S b/sysdeps/i386/i686/multiarch/strspn.S
index 53db131..f11d3ea 100644 (file)
--- a/sysdeps/i386/i686/multiarch/strspn.S
+++ b/sysdeps/i386/i686/multiarch/strspn.S
@@ -68,9 +68,9 @@ END(strspn)
 # define END(name) \
        cfi_endproc; .size __strspn_ia32, .-__strspn_ia32
 # undef libc_hidden_builtin_def
-/* It doesn't make sense to send libc-internal strspn calls through a PLT.
-   The speedup we get from using SSE4.2 instruction is likely eaten away
-   by the indirect call in the PLT.  */
+/* IFUNC doesn't work with the hidden functions in shared library since
+   they will be called without setting up EBX needed for PLT which is
+   used by IFUNC.  */
 # define libc_hidden_builtin_def(name) \
        .globl __GI_strspn; __GI_strspn = __strspn_ia32
 #endif