3.1 perfomance problems

[Roman Zippel <zippel at linux-m68k dot org>]

Hi,

While testing 3.1 for the m68k, I've hit an interesting test case:
gcc.c-torture/compile/20001226-1.c. It already needs quite some time on
my i686 machine, but for m68k I had to increase the time out during 3.0
testing.
Anyway, 3.1 times out again and additionally one test fails with out of
memory error. I repeated the test under i686 and it shows the same
behavior, it just doesn't fail because it's faster and has more memory.

First, the test needs more than twice longer to complete, e.g. from
1m13.028s now to 2m52.571s. Adding -ftime-report shows an interesting
entry:

 cfg cleanup           :  81.12 (49%) usr   0.00 ( 0%) sys  81.13 (48%)
wall

This doesn't exists in 3.0 and 3.1 spends half the time here for this
test case.

Second, the test with '-Os' needs more than twice as much memory. top
shows normally peaks of 50MB, but with '-Os' it needs up to 115MB and
nothing of this is visible with -fmem-report.

This behavior is new in 3.1 and I didn't find anything about this in the
archives (at least not such drastic examples), so maybe this helps to
improve this.

bye, Roman