[BZ #4314]

	* sysdeps/i386/i686/memcpy.S: Optimize copying of aligned buffers.

diff --git a/ChangeLog b/ChangeLog
index 76ec21b..38efd89 100644 (file)
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,5 +1,8 @@
 2008-04-09  Ulrich Drepper  <drepper@redhat.com>
 
+       [BZ #4314]
+       * sysdeps/i386/i686/memcpy.S: Optimize copying of aligned buffers.
+
        [BZ #5209]
        * sysdeps/unix/sysv/syscalls.list: The times syscall doesn't return
        an error value.

diff --git a/sysdeps/i386/i686/memcpy.S b/sysdeps/i386/i686/memcpy.S
index 00e84ec..ff5c66e 100644 (file)
--- a/sysdeps/i386/i686/memcpy.S
+++ b/sysdeps/i386/i686/memcpy.S
@@ -1,7 +1,7 @@
 /* Copy memory block and return pointer to beginning of destination block
    For Intel 80x86, x>=6.
    This file is part of the GNU C Library.
-   Copyright (C) 1999, 2000, 2003, 2004 Free Software Foundation, Inc.
+   Copyright (C) 1999, 2000, 2003, 2004, 2008 Free Software Foundation, Inc.
    Contributed by Ulrich Drepper <drepper@cygnus.com>, 1999.
 
    The GNU C Library is free software; you can redistribute it and/or
@@ -41,29 +41,64 @@ END (__memcpy_chk)
 ENTRY (BP_SYM (memcpy))
        ENTER
 
-       movl    LEN(%esp), %ecx
        movl    %edi, %eax
        movl    DEST(%esp), %edi
        movl    %esi, %edx
        movl    SRC(%esp), %esi
-       CHECK_BOUNDS_BOTH_WIDE (%edi, DEST(%esp), %ecx)
-       CHECK_BOUNDS_BOTH_WIDE (%esi, SRC(%esp), %ecx)
 
+       movl    %edi, %ecx
+       xorl    %esi, %ecx
+       andl    $3, %ecx
+       movl    LEN(%esp), %ecx
        cld
-       shrl    $1, %ecx
-       jnc     1f
+       jne     .Lunaligned
+
+       cmpl    $3, %ecx
+       jbe     .Lunaligned
+
+       testl   $3, %esi
+       je      1f
        movsb
-1:     shrl    $1, %ecx
-       jnc     2f
-       movsw
-2:     rep
+       decl    %ecx
+       testl   $3, %esi
+       je      1f
+       movsb
+       decl    %ecx
+       testl   $3, %esi
+       je      1f
+       movsb
+       decl    %ecx
+1:     pushl   %eax
+       movl    %ecx, %eax
+       shrl    $2, %ecx
+       rep
        movsl
-       movl    %eax, %edi
+       movl    %eax, %ecx
+       andl    $3, %ecx
+       rep
+       movsb
+       popl    %eax
+
+.Lend: movl    %eax, %edi
        movl    %edx, %esi
        movl    DEST(%esp), %eax
        RETURN_BOUNDED_POINTER (DEST(%esp))
 
        LEAVE
        RET_PTR
+
+       /* When we come here the pointers do not have the same
+          alignment or the length is too short.  No need to optimize for
+          aligned memory accesses. */
+.Lunaligned:
+       shrl    $1, %ecx
+       jnc     1f
+       movsb
+1:     shrl    $1, %ecx
+       jnc     2f
+       movsw
+2:     rep
+       movsl
+       jmp     .Lend
 END (BP_SYM (memcpy))
 libc_hidden_builtin_def (memcpy)