tree-ssa vs. tree-ssa comparison

[Steven Bosscher]

Hi,

Following Paolo's suggestion in a mail to gcc-patches, I benchmarked
tree-ssa against itself, but with GCSE disabled, and path following
in CSE also disabled.  These are SPECint2k scores of 3 runs on an
otherwise not loaded machine.  The tree-ssa compiler I used is just
CVS, checking disabled.  I did not apply Andrew Pinski's cast killer
patch, which should further improve both compile time and performance.

base: -O2 -mtune=athlon-xp
peak: -O2 -mtune=athlon-xp -fno-gcse --param max-cse-path-length=0

Total benchmarks compile time: peak is 13% faster than base.

Results of peak scores relative to base:
164.gzip	+3.5%
175.vpr		-0.0%
181.mcf		+0.0%
186.crafty	+1.7%
197.parser	-0.0%
253.perlbmk	-2.4%
254.gap		-0.2%
255.vortex	-5.6%
256.bzip2	-0.0%
300.twolf	+0.0%

So other than the perlbmk and vortex benchmarks, both gcse and path
following code are already subsumed by the tree-ssa passes on this
benchmark.  I don't know why gzip is faster, but it was so on all
the runs I did.

Diego, perhaps this is also what your mainline vs. tree-ssa should
compare in your daily SPEC tester.  Same for Richard ;-)

Gr.
Steven