Compilation performance of GCC 3.4.2 and GCC 4.0.0 (041021) ontramp3d-v3

[Richard Guenther <rguenth at tat dot physik dot uni-tuebingen dot de>]

Just to have another datapoint besides Karels observations.  I'm
using the tramp3d-v3.cpp testcase in place for testing POOMA
application performance.  The testcase can be fetched from

http://www.tat.physik.uni-tuebingen.de/~rguenth/gcc/tramp3d-v3.cpp.gz

detailled logs (including time-report) can be provided on demand.
Testing is done by
g++-3.4 -S tramp3d-v3.cpp -Dleafify=fooblah -ftime-report ...
Note that both compilers are built with checking disabled,
and 3.4.2 is profilebootstrapped, for 4.0.0 profilebootstrap doesnt work.

...                   3.4.2          4.0.0     Delta%
-O0                   24.18s         19.90s    21.5
-O1                   83.75s
-O1 -funit-at-a-time  50.70s         66.38s   -24
-O2                   65.88s         88.18s   -25
-O2 -funroll-loops    69.73s         93.12s   -25

These figures may be distorted because (again) code metrics for
inlining changed between 3.4.2 and 4.0.0 (and we're using unit-at-a-time
for 4.0.0 always), so with leafifying included we get the following
numbers:

...                   3.4.2          4.0.0     Delta%
-O1 -funit-at-a-time  70.26s         117.08s  -40
-O2                   92.77s         150.35s  -38
-O2 -funroll-loops    97.71s         153.48s  -36
-O2 -fno-exceptions   67.90s         110.28s  -38

So while 4.0.0 really shines on -O0, it looses on optimization, typically
about 20%.  Especially with large and a huge number of BBs as resulting
from inlining (or heavy inlining caused from leafifying) where we loose
up to 40% speed!
Note how exceptions cause a huge slow-down (~30%) in compilation time.

Even created code is worse than 3.4.2 (for ia32) - but that's another
story.

Oh yes, measurements were done on a Pentium-4 with 2GB of ram.

Richard.

--
Richard Guenther <richard dot guenther at uni-tuebingen dot de>
WWW: http://www.tat.physik.uni-tuebingen.de/~rguenth/