[Bug middle-end/50251] [4.7 Regression] Revision 178353 caused many test failures

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50251

--- Comment #12 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-09-02 09:40:33 UTC ---
(In reply to comment #3)
> Created attachment 25162 [details]
> optimized dump
> 
> 1. The alloca in main is transformed into this declaration:
> 
>   <unnamed-unsigned:8> D.2129[24];
> 
> and accessed like this:
> 
>   iftmp.6_13 = MEM[symbol: D.2129, index: ivtmp.30_31, step: 4, 
>                    offset: 4294967292B];
> 
>   MEM[symbol: D.2129, index: ivtmp.30_31, step: 4, offset: 0B] = D.2105_15;
> 
>   D.2099_18 = MEM[(int *)&D.2129][5]{lb: 0 sz: 4};
> 
>   MEM[symbol: D.2129, index: ivtmp.30_31, step: 4, offset: 0B] = 0;
> 
> 
> 2. __builtin_stack_restore/__builtin_stack_restore in main are not optimized
> away, because the restored value is potentially used. The restored value is the
> same as the current value, so the restore is redundant, but that's not checked
> in optimize_stack_restore.
> 
>    saved_stack.3_5 = __builtin_stack_save ();
> 
>   __builtin_stack_restore (saved_stack.3_5);

It does try it here:

 second_stack_restore:

  /* If there's exactly one use, then zap the call to __builtin_stack_save.
     If there are multiple uses, then the last one should remove the call.
     In any case, whether the call to __builtin_stack_save can be removed
     or not is irrelevant to removing the call to __builtin_stack_restore.  */
  if (has_single_use (gimple_call_arg (call, 0)))

but for some reason it doesn't trigger?