re: c++ performance regressions in gcc > 2.95.3

re: c++ performance regressions in gcc > 2.95.3

[dank]

Anthony wrote:
> We observed that certain large C++ applications perform worse
> in gcc-3.x and gcc-4.x than they did in gcc-2.95.3.
> On the theory that at least some of the cause
> would show up in microbenchmarks, we tried running
> bench++ with both old and new toolchains.
> ...
> http://www.cis.udel.edu/~danalis/OSS/bench_plus_plus/ ...

The biggest unreported regression in in the S000005 tests:

(times are ns per iteration)
==== Fill a buffer using different levels of abstraction
           g++295   g++401   g++410_0723
S000005a 3870.00 20840.00* 21240.00*
S000005b 3878.00 21120.00* 21140.00*
S000005c 3782.00  3894.00* 21320.00*
S000005d 3862.00 21360.00* 21220.00*
S000005e 3916.00  3834.00  19780.00*
S000005f 3818.00  3936.00* 21160.00*
S000005g 3940.00 20280.00* 20640.00*
S000005h 3868.00 21040.00* 21540.00*
S000005i 3928.00 20060.00* 21480.00*
S000005j 3844.00 21840.00* 21140.00*
S000005k 3912.00  3750.00  3964.00*
S000005l 3946.00 21360.00* 3912.00
S000005m 4746.00  3958.00  3904.00

Most of these are slow in both gcc-4.0.1 and gcc-4.1.
S000005e was fine in gcc-4.0.1,
but is suddenly slower in gcc-4.1.

Anthony, can you try submitting a reduced test case for
S000005e?

Thanks,
Dan

c++ performance regressions in gcc > 2.95.3

[Anthony Danalis]

We observed that certain large C++ applications perform worse
in gcc-3.x and gcc-4.x than they did in gcc-2.95.3.
On the theory that at least some of the cause
would show up in microbenchmarks, we tried running
bench++ with both old and new toolchains.
Because we suspect that part of the regression is due to
libstdc++, we also measured performance using stlport.


Here are our results.
A table of nanoseconds-per-iteration for each individual microbenchmark in bench++
for g++-2.95.3, and for g++-4.0.1 and g++-4.1.0-20050723
with and without STLport, is at
http://www.cis.udel.edu/~danalis/OSS/bench_plus_plus/files/report-f15_m2_X_2.6-absolute.txt
A table normalized relative to the gcc-2.95.3 results is at
http://www.cis.udel.edu/~danalis/OSS/bench_plus_plus/files/report-f15_m2_X_2.6.txt


The interesting bits are summarized in a table showing just the performance regressions,
and annotated with descriptions of the microbenchmarks which regressed.
It's at
    http://www.cis.udel.edu/~danalis/OSS/bench_plus_plus/results.html


We reported one of the regressions already as http://gcc.gnu.org/PR22563 .
There seem to be at least ten others that might be worth
reporting.  I'll try to post bug reports for a few, but
my summer internship is running out soon.  If anyone else
has time to look at the data, I'd appreciate suggestions or
criticism; maybe I can fix a few problems in my benchmark scripts
before I turn into a pumpkin.


Anthony Danalis & Dan Kegel