------- Additional Comments From jgrimm2 at us dot ibm dot com 2004-11-23 01:43 ------- Finally got SPEC set up w/profiling to do a little comparison. Its not done yet (and only running c,c++ tests) but first glance through the initials results, it appears that -fspeculative-prefetching is damaging to spec results at least on ppc64 (-m32). I wanted to do this to guage whether this bug entry was some pathological testcase or this would also show up in regularly looked at tests.
I'm currently comparing -O2 -funroll-loops -fno-speculative-prefetching vs -O2 -funroll-loops -fspeculative-prefetching and using profiling for both runs. I'll run again to see if I can get those results. Additionally, I'd like to see what the behavior is when I limit the prefetches (don't emit prefetch if just prefetched from near address + crudely limit maximum number of prefetches), I just uploaded a patch for that, but haven't tested how it affects spec yet. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=17950