Re: Compilation time has more than doubled on some Polyhedron tests

[Tobias dot Schlueter at Physik dot Uni-Muenchen dot DE]

[ Added gcc@gcc.gnu.org to the CC list and included Dominique's mail in full, as
the problem might well be in the common code.  ]

Quoting Dominique Dhumieres <dominiq@lps.ens.fr>:

> I have the following timing for the compilation of the polyhedron tests
> on a 1.8Ghz G5 with:
>
> gfortran -w -O3 -ftree-loop-linear -ffast-math -funroll-loops
>
> for the following snapshots patched with the Paul Thomas' optimization of
> dot_product.
>
> snapshot     12/24     12/31     01/07     01/14
>
> ac            5.8s      6.0s     24.7s     23.0s    +
> aermod      223.9s    227.3s    485.8s    467.6s    +
> air          21.0s     21.6s      8.2s      8.0s    -
> capacita      4.6s      4.7s      5.8s      5.6s
> channel       2.5s      2.5s     17.6s     17.7s    +
> doduc        19.7s     20.0s     38.6s     37.0s    +
> fatigue      15.0s     15.4s      7.2s      6.8s    -
> gas_dyn       8.6s      8.8s     19.6s     18.5s    +
> induct       28.1s     28.7s     66.1s     61.6s    +
> kepler        0.6s      0.6s      1.1s      1.0s    +
> linpk         1.4s      1.4s      1.4s      1.4s
> mdbx          5.3s      5.4s      6.9s      6.5s
> nf            3.6s      3.6s      4.4s      4.3s
> protein      15.5s     15.9s     15.9s     14.5s
> rnflow       11.4s     11.6s     12.6s     12.7s
> test_fpu      8.4s      8.5s      8.8s      8.5s
> tfft          1.3s      1.3s      1.8s      1.7s
>
> I only noticed it today, but there has been a noticable change in the
> compilation time between the 20051231 and 20060107 snapshots:
> air.f90 and fatigue.f90 compile twice faster, but aermode.f90,
> doduc.f90, gas_dyn.f90, induct.f90, and kepler.f90 compile twice slower,
> while ac.f90 takes four more time and channel.f90 takes seven more time.
>
> Am I the only one to see that? If not, does any one has any idea of what
> caused this slow down in compilation?
>
> Using the François-Xavier's builds I have narrowed the problem between the
> 05/01 and the 06/01:
>
> [scala] gfortran/2006> time irun/bin/gfortran -v -O3 -ffast-math
> -funroll-loops induct.f90
> ...
> GNU F95 version 4.2.0 20060105 (experimental) (x86_64-unknown-linux-gnu)
> ...
> 19.375u 0.250s 0:20.39 96.2%    0+0k 0+0io 0pf+0w
>
> [scala] gfortran/2006> time irun/bin/gfortran -v -O3 -ffast-math
> -funroll-loops induct.f90
> ...
> GNU F95 version 4.2.0 20060106 (experimental) (x86_64-unknown-linux-gnu)
> ...
> 33.865u 0.362s 0:35.09 97.5%    0+0k 0+0io 0pf+0w
>
> TIA
>
> Dominique
>
> On the induct.f90 test I get with the latest snapshot:
>
>                      options                        compile   exec
>
> -O3 -ftree-loop-linear -ffast-math -funroll-loops     62.7    66.7
> -O3 -ffast-math -funroll-loops                        61.0    66.6
> -O3 -ffast-math                                       59.5    72.0
> -O3 -funroll-loops                                    60.1    81.1
> -O3                                                   59.5    81.4
> -O2 -ffast-math -funroll-loops                        60.1    67.0
> -O2 -ffast-math                                       54.3   154.5
> -O2 -funroll-loops                                    58.7    81.0
> -O2                                                   55.2   112.3
> -O1 -ffast-math -funroll-loops                        14.6    80.5
> -O1 -ffast-math                                       12.6   160.3
> -O1 -funroll-loops                                    14.4   105.7
> -O1                                                   12.7   164.4
> -O0 -ffast-math -funroll-loops                         3.3   638.0
> -O0                                                    3.4   634.4

Are you building with --enable-checking (the default)?  Can you try compiling
some of the most-affected files with -ftime-report, so that we can see where
the slowdown occured?

- Tobi