Re: gcc 3.2.1 optimizer degradation (strlen, -O2)

["Leonid Pauzner" <uue at pauzner dot dnttm dot ru>]

11-Jan-2003 23:33 Marcel Cox wrote:

> 3) Except for Pentium4, GCC 3.2.1 does not inline the strlen function in
> your code as it considers the gain to be not worthwhile. You can however
> force it to do so anyway by using the option -minline-all-stringops and see
> if that makes a difference.

Apparently, gcc 3.2.1 __builtin_strlen() never inlined on Pentium 90MHz
so the following function hangs:

====== strlen3.c =========
#include <string.h>

size_t
strlen(const char *str)
{
  return __builtin_strlen(str);
}
==========================
gcc -O2 -S -c strlen3.c
====== strlen3.s =========
        .file   "strlen3.c"
        .section .text
        .p2align 4,,15
.globl _strlen
_strlen:
        pushl   %ebp
        movl    %esp, %ebp
        popl    %ebp
        jmp     _strlen
        .ident  "GCC: (GNU) 3.2.1"
==========================

Update my previous tests:
>> 2.95 do inline strlen(), while 3.2.1 is not.
>> Nevertheless I got the following results in my test:
>>
>> gcc321 -O2
>> strlen.c   20 sec
>> strlen2.c  17 sec
   strlen3.c  hangs!
>> (library)  14 sec  (yes, libc compiled with other gcc version/flags)
   strlen3.c   8 sec  (-O2 -minline-all-stringops)
   (builtin)   7 sec  (main.c was compiled with -O2 -minline-all-stringops)
>>
>> gcc295 -O2
>> strlen.c   16 sec  (main.c was compiled without -O)
>> strlen2.c  11 sec  (main.c was compiled without -O)
   strlen3.c   9 sec  (main.c was compiled without -O)
>> (builtin)   8 sec


GCC 321 assembler for the above functions attached.

Attachment: len321.zip
Description: Binary data