This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Alias Analysis Improvement Patch


law@redhat.com wrote:

>
>  >So I think this patch is not for approval in near future but for its
>  >investigation.  I am going to look at this patch for ia64 platform.
> Let me know what you find.
>

I've tried the patch on Itanium2 machine for the mainline gcc.  I've got the
following results (the right column is for gcc with the new aliasing).

   164.gzip          1400   366           383*     1400   367           382*

   175.vpr           1400   303           463*     1400   305           459*

   176.gcc                                   X                             X

   181.mcf           1800   717           251*     1800   715           252*

   186.crafty        1000   199           503*     1000   198           506*

   197.parser        1800   494           364*     1800   496           363*

   252.eon           1300   475           273*
   253.perlbmk       1800   370           487*     1800   372           484*

   254.gap           1100   331           332*     1100   331           333*

   255.vortex        1900   324           586*     1900   325           585*

   256.bzip2         1500   393           382*     1500   393           382*

   300.twolf         3000   561           535*     3000   625           480*

Code for eon generated with the new alias analysis works incorrectly.

In average gcc with the patch is worse especially for twolf.  I think I
should search for the reason of degradation in counteraction of scheduler
and alias analysis.

The compiler is also slower ~25% (890s vs 1140 s) with the new alias
analysis on SPECINT2000.

I did not check the code size (it is not important for Itanium).

Practically the same picture is for SPECFP2000 tests (new aliasing results
in incorrect work of 4 tests).

Although I've found a test when the patch gives a big improvement (2%).  It
is whetstone

Sanjiv, the results are frustrated ones but I am not giving up.  I'll work
on it.  Alias analysis is important thing.  Theoretically it is better
approach then SSA one because it is flow sensitive.  So I am going to find
what is wrong with the results.  Although I am not sure that it will happen
soon.

Another thing is we should speed up it considerably to justify the patch
adoption.

Vlad



Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]