https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63432
--- Comment #29 from Teresa Johnson <tejohnson at google dot com> --- On Mon, Oct 13, 2014 at 2:32 PM, Teresa Johnson <tejohn...@google.com> wrote: > On Mon, Oct 13, 2014 at 8:53 AM, hjl.tools at gmail dot com > <gcc-bugzi...@gcc.gnu.org> wrote: >> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63432 >> >> --- Comment #27 from H.J. Lu <hjl.tools at gmail dot com> --- >> (In reply to Teresa Johnson from comment #24) >> >>> Arg, looks very similar so maybe another instance of the duplicate >>> threading is slipping through? My own lto profiled bootstrap succeeded >>> with my patch. I will try updating to r216039 and redo it to see if I >>> can provoke the same failure. >>> >> >> I sent you another testcase against r216150. > > Thanks for the testcase, I reproduced it. It is a case of garbage in / > garbage out. The fre2 pass is introducing some big profile count > insanities, leading to the probability insanity being introduced when > we try to use the counts to compute the new probability in > recompute_probabilities. There is already handling for really large > probabilities due to this issue, and we need to add the same thing for > negative probabilities - essentially the patch you had originally > suggested for the first problem which wasn't necessary for that one > since that was an actually jump threading induced issue. Actually, I traced the initial profile insanity back to inlining. FRE merely propagated it further. I have a better overflow test done before the scaling, running it through LTO profiledbootstrap then will send for review. Teresa > > Will test that and send for review. > >> >> -- >> You are receiving this mail because: >> You are on the CC list for the bug. > > > > -- > Teresa Johnson | Software Engineer | tejohn...@google.com | 408-460-2413
