Serious performance regression -- some tree optimizer questions

[Ulrich Weigand]

Mircea Namolaru/Haifa/IBM wrote on 01/13/2005 03:12:13 PM:

> >> IBM Haifa has some patches waiting for GCC 4.1 to address sign
> >> extension problems.
> >
> > Zdenek's latest patch addressed all sign-extension problems
> > I've been seeing with my test case -- are the Haifa patches
> > concerned with the same issues or different ones?
>
> Regarding our intended patch you can look at:
>
> - The overview of the algorithm:
> http://gcc.gnu.org/ml/gcc/2004-06/msg00472.html
>
> - The RFC for the implementation:
> http://gcc.gnu.org/ml/gcc/2004-09/msg00377.html

Thanks for the pointers.  It does look that these patches
address a completely different issue; they do not change
the type of a variable (which is what causes the overflow
semantics problems), but solely try to optimize the placement
of sign-extend operations.  Also, the pass apparently will
operate on RTL, while I was disussing issues in the tree
loop optimizer ...

So, your patches look certainly valuable (in fact I think
they'll most likely benefit s390 as well), but are completely
unrelated to the problem I raised in this thread (and which
Zdenek's latest patch solves for me).


Mit freundlichen Gruessen / Best Regards

Ulrich Weigand

--
  Dr. Ulrich Weigand
  Linux for S/390 Design & Development
  IBM Deutschland Entwicklung GmbH, Schoenaicher Str. 220, 71032 Boeblingen
  Phone: +49-7031/16-3727   ---   Email: Ulrich.Weigand@de.ibm.com