Re: Optimising std::find on x86 and PPC

[Daniel Berlin <dberlin at dberlin dot org>]

> | consider the current behavior just a serious (performance) regression
> | (i.e., correctness vs aliasing and other subtleties should not be in
> | the way)
>
> That is a clear regression.
>
> Tree-ssa has been sold on the name of bringing better methodology to
> optimizations; given Theo's analysis, I think this issue no longer
> belongs to V3.  It should be brought to GCC main list


The issue i have with this characterization is that we don't have a 
tree level unroller that does anything but complete unrolling, so i'm not 
sure what tree-ssa has to do with it.

I do agree it's a regression, however, and a serious one (because i trust 
the libstdc++ people when they say this will cause them serious 
problems).

Just so others have context, the trivially broken loop that Paolo had 
sent to me looks like this:

int*
check(int* a, int* b)
{
 for (; a < b; ++a)
   if (*a == 1)
     return a;
 return a;
}


This was unrolled with 3.3.3, but 4.0 (and apparently 3.4) claims the 
number of iterations is not runtime computable.

This is clearly wrong, as there are no stores in the entire function, so b 
couldn't *possible* change, no matter what we do.

:)
--Dan