[PATCH] fix i386 memcpy

[Denis Vlasenko]

This patch shortens non-constant memcpy() by two bytes
and fixes spurious out-of-line constant memcpy().

Patch is run-tested (I run on patched kernel right now).

Benchmark and code generation test program will be mailed as reply.

# size vmlinux.org vmlinux
   text    data     bss     dec     hex filename
3954591 1553426  236544 5744561  57a7b1 vmlinux.org
3952615 1553426  236544 5742585  579ff9 vmlinux

Example of changes (part of dump_fpu() body):
old.............................................. new......................
8d 83 40 02 00 00 lea    0x240(%ebx),%eax         8d b3 40 02 00 00 lea    0x240(%ebx),%esi
74 31             je     c0108b27 <dump_fpu+0x9c> 74 2e             je     c0108b1d <dump_fpu+0x92>
6a 1c             push   $0x1c                    8b 7d 0c          mov    0xc(%ebp),%edi
50                push   %eax                     b9 07 00 00 00    mov    $0x7,%ecx
56                push   %esi                     f3 a5             repz movsl %ds:(%esi),%es:(%edi)
e8 49 21 10 00    call   c020ac48 <memcpy>        8b 55 0c          mov    0xc(%ebp),%edx
83 c4 0c          add    $0xc,%esp                83 c2 1c          add    $0x1c,%edx
83 c6 1c          add    $0x1c,%esi               8d 83 60 02 00 00 lea    0x260(%ebx),%eax
81 c3 60 02 00 00 add    $0x260,%ebx              b9 07 00 00 00    mov    $0x7,%ecx
bf 07 00 00 00    mov    $0x7,%edi                89 d7             mov    %edx,%edi
6a 0a             push   $0xa                     89 c6             mov    %eax,%esi
53                push   %ebx                     a5                movsl  %ds:(%esi),%es:(%edi)
56                push   %esi                     a5                movsl  %ds:(%esi),%es:(%edi)
e8 2f 21 10 00    call   c020ac48 <memcpy>        66 a5             movsw  %ds:(%esi),%es:(%edi)
83 c4 0c          add    $0xc,%esp                83 c2 0a          add    $0xa,%edx
83 c6 0a          add    $0xa,%esi                83 c0 10          add    $0x10,%eax
83 c3 10          add    $0x10,%ebx               49                dec    %ecx
4f                dec    %edi                     79 ef             jns    c0108b0a <dump_fpu+0x7f>
79 eb             jns    c0108b10 <dump_fpu+0x85> eb 0a             jmp    c0108b27 <dump_fpu+0x9c>
eb 0c             jmp    c0108b33 <dump_fpu+0xa8> 8b 7d 0c          mov    0xc(%ebp),%edi
6a 6c             push   $0x6c                    b9 1b 00 00 00    mov    $0x1b,%ecx
50                push   %eax                     f3 a5             repz movsl %ds:(%esi),%es:(%edi)
56                push   %esi                     8b 45 f0          mov    0xfffffff0(%ebp),%eax
e8 18 21 10 00    call   c020ac48 <memcpy>        5a                pop    %edx
83 c4 0c          add    $0xc,%esp                5b                pop    %ebx
8b 45 f0          mov    0xfffffff0(%ebp),%eax
8d 65 f4          lea    0xfffffff4(%ebp),%esp
5b                pop    %ebx
5e                pop    %esi

--
vda
-------------- next part --------------
A non-text attachment was scrubbed...
Name: string.memcpy.diff
Type: text/x-diff
Size: 2953 bytes
Desc: not available
URL: <https://gcc.gnu.org/pipermail/gcc/attachments/20050329/9feb557f/attachment.bin>