Re: [tree-ssa] Fix mudflap codegen and allow optimizations [patch]

[Andrew Pinski <pinskia at physics dot uc dot edu>]

Which process is that PowerPC, the timings are totally different 
between processors.
Also the branch predictor are also totally different.  Same Problem in 
x86 but not as drastic.

Thanks,
Andrew Pinski

On Sunday, Apr 27, 2003, at 23:22 US/Eastern, Diego Novillo wrote:

> On Sun, Apr 27, 2003 at 09:56:30PM -0400, Frank Ch. Eigler wrote:
>
> > > - The call to __builtin_expect was using BIT_IOR_EXPR instead of
> > >   TRUTH_ORIF_EXPR.  This prevents the short-circuit evaluation of
> > >   the predicate, producing slightly slower code. [...]
> >
> > Are you sure about that?  Short-circuit evaluation usually requires
> > additional branches.  I recall experimenting with this early on, and
> > it made runs slower, not faster.
> I had tested the attached on x86 and found the TRUTH_ORIF version
> faster than BIT_IOR by about 33%.
>
> But now I've also tested it on x86_64 and PPC and found the exact
> opposite (on x86-64 BIT_IOR is 47% faster, on PPC, BIT_IOR is 17%
> faster).
>
> So, that's 2 out of 3 and you're right that the code is less
> jumpy.  I'll revert to using BIT_IOR.
>
>
> Diego.
> <a.c>