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: Fix -fwhole-program on LTO -- Another SPEC2000 benchmakring round

From: Vladimir Makarov <vmakarov at redhat dot com>
To: Jan Hubicka <hubicka at ucw dot cz>
Cc: Richard Guenther <rguenther at suse dot de>, gcc-patches at gcc dot gnu dot org, Diego Novillo <dnovillo at google dot com>, Martin Jambor <mjambor at suse dot de>
Date: Thu, 08 Oct 2009 10:05:25 -0400
Subject: Re: Fix -fwhole-program on LTO -- Another SPEC2000 benchmakring round
References: <20091006200322.GB19212@kam.mff.cuni.cz> <alpine.LNX.2.00.0910071355240.4520@zhemvz.fhfr.qr> <20091007120217.GG9046@atrey.karlin.mff.cuni.cz> <alpine.LNX.2.00.0910071604160.4520@zhemvz.fhfr.qr> <20091007142848.GI9046@atrey.karlin.mff.cuni.cz> <20091007163829.GJ9046@atrey.karlin.mff.cuni.cz>

Jan Hubicka wrote:

On Wed, 7 Oct 2009, Jan Hubicka wrote:

I am testing patch fixing the ipa-sra issue.
and commited after x86_64-linux testing.

With this Jan's fix, almost all SPEC2000 problems are gone.  The
current results (for -O3 vs -O3 -flto -fwhole-program) are

x86:
 o Int2000:
   - LTO crashes the compiler on vortex (lto1 segmentation fault).
   - Compiler is 1.85 times slower in pcu time with LTO (compiler
     was built with --enable-checking=release).
   - Average code size is more than 5% smaller:

4.389% 44383 46331 164.gzip -2.898% 144261 140081 175.vpr 0.549% 1568494 1577105 176.gcc -12.299% 12359 10839 181.mcf 11.018% 210196 233356 186.crafty -29.809% 155454 109114 197.parser -16.658% 497977 415024 252.eon 9.025% 552707 602591 253.perlbmk 1.492% 503534 511046 254.gap -20.175% 47801 38157 256.bzip2 -3.038% 198941 192897 300.twolf Average = -5.30941% - Performance is improved almost by 3%

     164.gzip     1670   1631 -2.33%
     175.vpr      2112   2145  1.56%
     176.gcc      3205   3266  1.90%
     181.mcf      5027   5030  0.05%
     186.crafty   2254   2261  0.31%
     197.parser   1926   1920 -0.31%
     252.eon      2471   2945 19.18%
     253.perlbmk  2871   2897  0.90%
     254.gap      2700   2738  1.40%
     256.bzip2    1863   1934  3.81%
     300.twolf    2827   3025  7.00%
     GeoMean      2510   2583  2.90%

 o FP2000
   - Compiler is 2.16 times slower with LTO
   - Average code size is almost 1.4% smaller:

       -8.718%          27528          25128 168.wupwise
        2.499%           9124           9352 171.swim
        5.149%          18081          19012 172.mgrid
        0.450%          77144          77491 173.applu
       -5.735%         581390         548049 177.mesa
       -1.223%         183955         181705 178.galgel
      -10.635%          15881          14192 179.art
      -16.431%          28860          24118 183.equake
       -2.197%          67367          65887 187.facerec
       11.101%         125305         139215 188.ammp
       -1.002%          49201          48708 189.lucas
       -0.464%        1144470        1139156 191.fma3d
       11.387%         938725        1045616 200.sixtrack
       -3.301%         126947         122757 301.apsi
      Average = -1.36578%

- Performance is improved by more 3%

168.wupwise 2330 3245 39.27% 171.swim 3521 3526 0.14% 172.mgrid 1079 1084 0.46% 173.applu 2658 2619 -1.46% 177.mesa 1956 2017 3.11% 178.galgel 5687 5623 -1.12% 179.art 7075 7143 0.96% 183.equake 3837 4136 7.79% 187.facerec 2620 2593 -1.03% 188.ammp 2369 2399 1.26% 189.lucas 2827 2816 -0.38% 191.fma3d 2397 2462 2.71% 200.sixtrack 914 916 0.21% 301.apsi 2344 2325 -0.81% GeoMean 2599 2683 3.23% x86_64: o Int2000: - LTO crashes the compiler on gcc (explow:copy_to_mode_reg). - Compiler is 1.8 times slower with LTO - Average code size is almost 8% smaller:

        1.276%          49231          49859 164.gzip
       -4.298%         158357         151551 175.vpr
      -17.025%          15037          12477 181.mcf
       12.651%         195394         220114 186.crafty
      -29.741%         180852         127064 197.parser
      -19.558%         522443         420265 252.eon
        9.643%         645861         708141 253.perlbmk
        6.476%         585876         623820 254.gap
      -22.703%         660134         510262 255.vortex
      -18.687%          55825          45393 256.bzip2
       -5.444%         213071         201471 300.twolf
       Average = -7.94634%

- Performance is improved by 3.6%

     164.gzip     1806   1777 -1.60%
     175.vpr      2397   2403  0.25%
     181.mcf      3483   3474 -0.25%
     186.crafty   3387   3403  0.47%
     197.parser   1848   1802 -2.48%
     252.eon      4054   4512 11.29%
     253.perlbmk  3029   3035  0.19%
     254.gap      2960   3044  2.83%
     255.vortex   3514   4306 22.53%
     256.bzip2    2185   2246  2.79%
     300.twolf    2860   3037  6.18%
     GeoMean      2776   2876  3.60%

 o FP2000
   - Compiler is 2.2 times slower with LTO
   - Average code size is 2.6% smaller:

      27.459%          33934          43252 168.wupwise
      -2.785%          15800          15360 171.swim
       1.893%          22977          23412 172.mgrid
      -0.791%         103792         102971 173.applu
      -8.280%         680036         623727 177.mesa
      -4.232%         249941         239363 178.galgel
     -25.801%          20375          15118 179.art
     -28.926%          37578          26708 183.equake
      -1.347%          76885          75849 187.facerec
       1.570%         168283         170925 188.ammp
      -0.748%          57383          56954 189.lucas
      -0.828%        1279124        1268538 191.fma3d
      10.837%        1108531        1228666 200.sixtrack
      -4.150%         173625         166419 301.apsi
      Average = -2.58067%

- Performance is improved by 1.7%

     168.wupwise      2532     3732  47.39%
     171.swim         3726     3742   0.42%
     172.mgrid        2202     2145  -2.58%
     173.applu        3113     3150   1.18%
     177.mesa         2893     2961   2.35%
     178.galgel       5692     4779 -16.04%
     179.art          7276     7340   0.87%
     183.equake       3970     4210   6.04%
     187.facerec      2484     2497   0.52%
     188.ammp         2534     2556   0.86%
     189.lucas        3900     3828  -1.84%
     191.fma3d        2044     1985  -2.88%
     200.sixtrack     1156     1156   0%
     301.apsi         2867     2810  -1.98%
     GeoMean          3014     3066   1.72%

Follow-Ups:
- Re: Fix -fwhole-program on LTO -- Another SPEC2000 benchmakring round
  - From: Steven Bosscher

References:
- Fix -fwhole-program on LTO
  - From: Jan Hubicka
- Re: Fix -fwhole-program on LTO
  - From: Richard Guenther
- Re: Fix -fwhole-program on LTO
  - From: Jan Hubicka
- Re: Fix -fwhole-program on LTO
  - From: Richard Guenther
- Re: Fix -fwhole-program on LTO
  - From: Jan Hubicka
- Re: Fix -fwhole-program on LTO
  - From: Jan Hubicka

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