Performance loss between Build #15138 and 15222
|
|
Bugzilla Link |
347 |
Created on |
Feb 27, 2007 19:59 |
Resolution |
FIXED |
Resolved on |
Nov 21, 2007 23:31 |
Version |
1.00beta |
OS |
Linux |
Architecture |
PC |
Extended Description
I tried to rebuild versions of sac2c between the two revision levels noted
> below, but the file structural changes were not cooperating, so I have
> not been able to fault-isolate the problem to any degree more than that
> noted below.
>>> Just as a matter of curiousity, does anyone know why recent versions of
>>> sac2c are generating code that may run an order of magnitude slower
>>> than earlier versions?
> >>> I think I'm looking at current build #15222 vs build #15138, or
> >>> thereabouts.
> >>> apex/histlp.sac is typical. apex/buildv2.sac has similar excitement.
> >>> Both
> >>> of these are scalar-ish looping codes.
>>> The buildv2.sac benchmark constructs a vector
>>> by repeated assigns of other vectors into it.
> >>> histlp.sac computes histograms via a loop over the input data.
>>> I have not taken the time to look into why things are slower, but if you
>>> do know
>>> what's up, perhaps you can enlighten us with what's up, and when we
>>> might expect things to return to normal?
>>>
>>> This is compiling with -dowlfs -dosaacyc -check b -maxwlur 3 -dosaa and
>>> -noPHM.
>>> THe latter should not be relevant, because my earlier timings were also
>>> done with
>>> -noPHM.
>>> ps: Many benchmarks that DID work before now crash sac2c.