[PATCH] Variable Expansion in unroller + Register Pressure Utility

Revital Eres ERES@il.ibm.com
Mon Sep 20 19:20:00 GMT 2004


Hello,

Attached is the patch with the changes I was asked to make.

The testing so far on POWER4 showed no significant 
improvement on SPECINT.
On SPECFP I noticed an improvement of ~2.6% on art benchmark 
when running with POWER4 flags.
On i686 the patch showed no significant improvement
on both SPECINT and SPECFP.

The optimization is currently enabled 
with -fvariable-expansion-in-unroller flag.

Bootstrapped (with -funroll-all-loops) & regression tests on POWER4.
(I would also like to report that in the last two days (20-19/9/04)
I received gnucompare error when bootstraping
mainline with -O2 -funroll-all-loops options on both i686 and POWER4.)

Thanks,
Revital


-------------- next part --------------
A non-text attachment was scrubbed...
Name: new_diff_20_9
Type: application/octet-stream
Size: 55269 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20040920/63137cbc/attachment.obj>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: Change_log_20_9
Type: application/octet-stream
Size: 2191 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20040920/63137cbc/attachment-0001.obj>


More information about the Gcc-patches mailing list