Patch for 3.4.4. Infinite memory allocation on -O3 (PR #18081).

[Mark Mitchell]

Ulrich Weigand wrote:

> However, I've found a definite bug in 3.4 caused by this particular 
> backport, which is also very likely the cause of the performance
> regression reported here:
> http://gcc.gnu.org/ml/gcc/2005-05/msg00523.html

> This didn't matter until now as the code calls set_mem_attribute on
> the MEM generated by PUT_CODE, which used to recompute the value of
> the MEM_VOLATILE_P flag.  However, after your patch, such MEMs will
> remain marked as volatile ...

Thank you very much for looking into these issues.

> I'd suggest to either revert the MEM_VOLATILE backport(s) for 3.4.4
> or else fix the regressions by something like the below patch.

I'm not going to to be tricked into chasing through a never-ending patch 
sequence here.  I have every confidence that your changes are correct, 
but I think what we've proven is that there's too many non-local impacts 
from the patch.

I will revert the backport shortly.

Thanks,

-- 
Mark Mitchell
CodeSourcery, LLC
mark@codesourcery.com
(916) 791-8304