Re: GCC 3.1 Prerelease

Re: GCC 3.1 Prerelease

[Paolo Carlini]

Gerald Pfeifer wrote:
> As promised, I performed tests comparing GCC 2.95.3, GCC 3.0, GCC 3.0.3,
> the 3.1-branch as of yesterday, the 3.1-branch with -finline-limit=800
> and 1200, respectively, and the 3.2-branch as of yesterday.

Hi,

even if it's not immediately useful in order to deal with the performance problems of 3.1-branch, I think it would be interesting to see what the 3.2-branch numbers would become with the following patch reverted:

	http://gcc.gnu.org/ml/gcc/2002-03/msg00216.html

Any chance you can do that?

(I would guess that the improvement on 3.1 comes from the better alias analysis)

Ciao,
Paolo.

C++ aliasing changes on mainline (was: GCC 3.1 Prerelease)

[Gerald Pfeifer]

[ Subject change, as this indeed isn't a 3.1 issue ]

On Tue, 23 Apr 2002, Paolo Carlini wrote:
> even if it's not immediately useful in order to deal with the
> performance problems of 3.1-branch, I think it would be interesting to
> see what the 3.2-branch numbers would become with the following patch
> reverted:
>
> 	http://gcc.gnu.org/ml/gcc/2002-03/msg00216.html
>
> Any chance you can do that?

Okay, I performed `cvs update -j 1.167 -j 1.166 gcc/alias.c` and made the
following change in gcc/cp/cp-lang.c:

  static HOST_WIDE_INT
  cxx_get_alias_set (t)
     tree t;
  {
    return 0;
  }

> (I would guess that the improvement on 3.1 comes from the better alias
> analysis)

I got mixed results. The binary with the changes above (and all CVS
changes between Monday and Tuesday)  was somewhat smaller than the
original. Performance didn't change too much, either way.

Gerald


(Times in seconds,   |
stddev of three runs |
in parentheses.)     |      3.1      |      3.2      | 3.2-patched   |
---------------------+---------------+---------------+---------------+
      STRATCOMP1-ALL |  96.62 (0.10) |  91.96 (0.14) |  81.64 (0.03) |
   STRATCOMP-770.2-Q |   0.94 (0.01) |   0.93 (0.00) |   0.93 (0.00) |
               2QBF1 |  22.25 (0.01) |  20.98 (0.03) |  21.96 (0.01) |
          PRIMEIMPL2 |  12.87 (0.00) |  12.35 (0.02) |  12.33 (0.00) |
            ANCESTOR |   9.53 (0.00) |   9.44 (0.02) |   9.67 (0.00) |
       3COL-SIMPLEX1 |   7.14 (0.01) |   7.29 (0.02) |   7.19 (0.01) |
        3COL-LADDER1 |  42.35 (0.01) |  40.94 (0.05) |  41.77 (0.01) |
      3COL-N-LADDER1 |  22.79 (0.03) |  21.31 (0.04) |  22.77 (0.09) |
        3COL-RANDOM1 |  12.21 (0.00) |  11.50 (0.01) |  11.53 (0.02) |
          HP-RANDOM1 |  14.77 (0.02) |  14.42 (0.02) |  14.30 (0.00) |
       HAMCYCLE-FREE |   1.70 (0.00) |   1.59 (0.00) |   1.58 (0.00) |
             DECOMP2 |  24.07 (0.06) |  24.33 (0.02) |  24.72 (0.00) |
        BW-P4-Esra-a |  99.21 (0.07) |  96.45 (0.03) |  98.48 (0.01) |
        BW-P5-nopush |   7.42 (0.01) |   7.24 (0.01) |   7.36 (0.02) |
       BW-P5-pushbin |   6.49 (0.01) |   6.29 (0.00) |   6.36 (0.00) |
     BW-P5-nopushbin |   2.08 (0.00) |   2.04 (0.00) |   2.06 (0.00) |
              3SAT-1 |  38.42 (0.02) |  36.79 (0.02) |  36.93 (0.03) |
   3SAT-1-CONSTRAINT |  20.61 (0.03) |  19.24 (0.05) |  19.47 (0.03) |
        HANOI-Towers |   4.94 (0.01) |   5.06 (0.01) |   5.04 (0.01) |
              RAMSEY |   8.69 (0.00) |   8.66 (0.01) |   8.53 (0.01) |
             CRISTAL |  13.42 (0.01) |  13.30 (0.01) |  13.10 (0.03) |
             HANOI-K |  38.55 (0.08) |  36.66 (0.01) |  37.53 (0.03) |
           21-QUEENS |  10.40 (0.00) |   9.21 (0.01) |   9.24 (0.01) |
   MSTDir[V=13,A=40] |  21.07 (0.00) |  20.44 (0.01) |  20.45 (0.00) |
   MSTDir[V=15,A=40] |  21.16 (0.00) |  20.50 (0.01) |  20.49 (0.02) |
 MSTUndir[V=13,A=40] |  11.46 (0.01) |  11.05 (0.00) |  10.86 (0.01) |
 MSTUndir[V=15,A=40] | 188.59 (0.02) | 182.72 (0.05) | 178.36 (0.05) |
         TIMETABLING |  10.62 (0.01) |  10.81 (0.01) |  10.51 (0.00) |
---------------------+---------------+---------------+---------------+

Re: C++ aliasing changes on mainline (was: GCC 3.1 Prerelease)

[Paolo Carlini]

Gerald Pfeifer wrote:

>I got mixed results. The binary with the changes above (and all CVS
>changes between Monday and Tuesday)  was somewhat smaller than the
>original. Performance didn't change too much, either way.
>
Thank you very much Gerald.
Indeed, I find those numbers very interesting and very puzzling! 
Definitely I did expect a clear positive effect of the C++ aliasing 
patch... First, I have to learn more about the main features of such C++ 
code.

Ciao, Paolo.

Re: C++ aliasing changes on mainline (was: GCC 3.1 Prerelease)

[Jan Hubicka]

> Gerald Pfeifer wrote:
> 
> >I got mixed results. The binary with the changes above (and all CVS
> >changes between Monday and Tuesday)  was somewhat smaller than the
> >original. Performance didn't change too much, either way.
> >
> Thank you very much Gerald.
> Indeed, I find those numbers very interesting and very puzzling! 
> Definitely I did expect a clear positive effect of the C++ aliasing 
> patch... First, I have to learn more about the main features of such C++ 
> code.
Just note that there has been at least positive change for eon benchmark
on Andreas' SPECint pages (www.suse.de/~aj/SPEC)

Honza
> 
> Ciao, Paolo.
>