------- Comment #6 from dominiq at lps dot ens dot fr 2010-07-02 16:53 ------- > Yes, please reduce and lets see if we can discover something more specific > wrong here. Then also consider Mikael's idea.
I don't think there is anything "specific" to discover. The fix for PR44582 is too conservative and creates unneeded temporaries (in channel.f90, test_fpu.f90 and probably others that I have missed). I'll be extremely sad that to fix (ab)uses of the standard (codes you should never write), legitimate "real life" codes get badly penalized. You can also see the effect at http://gcc.opensuse.org/c++bench/polyhedron/polyhedron-summary.txt-2-0.html where the time went from ~16.5s to ~18.7s. The relative effect is less dramatic because channel.f90 is memory bound for caches smaller than ~4Mb: I don't know the cache size of the AMD proc for the above test (~512kb?) compared to the 3Mb on my macbook. So the effect of the new temporaries is to increase the memory access by ~2s, i.e., ~10% on the AMD tests, but ~70% on my machine. So I doubt that a reduced test wil give any new insight in the problem, nevertheless I'll try!-). -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44773
