This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
PR19039 causes regression of 3.9% in SPEC int
- From: Mostafa Hagog <MUSTAFA at il dot ibm dot com>
- To: gcc at gcc dot gnu dot org
- Cc: stevenb at suse dot de, pinskia at physics dot uc dot edu
- Date: Fri, 24 Dec 2004 10:58:52 +0200
- Subject: PR19039 causes regression of 3.9% in SPEC int
Performing the loop-header-copying is a sort of "work around" to this PR,
I have ported the loop-header-copying from tree's to RTL. This caused a
total 3.9% improvements with the rtl loop header copying (lhc), the
interesting thing is the GCC benchmark which improved by 66%.
The implementation in RTL is aggressive so further tuning may give even
better results. Following are the results on a 2GHz G5 machine,
gcc version 4.0.0 20041222 (experimental)
-O3 -funroll-loops -mtune=G5 -mcpu=G5 -mpowerpc64 -ffast-math
-mdynamic-no-pic -freorder-blocks -fstrict-aliasing
with lhc w/o lhc
164.gzip 780* 780*
175.vpr 559* 556*
176.gcc 1170* 705*
181.mcf 541* 541*
186.crafty 1028* 1060*
197.parser 738* 739*
252.eon 1175* 1178*
253.perlbmk 787* 828*
254.gap 871* 855*
255.vortex 1000* 988*
256.bzip2 643* 645*
300.twolf 615* 616*
Est. SPECint_base2000 798 768
Est. SPECint2000
Mostafa.