http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43716
--- Comment #49 from richard.guenther at gmail dot com <richard.guenther at gmail dot com> 2013-01-09 08:52:21 UTC --- On Tue, Jan 8, 2013 at 8:52 PM, dominiq at lps dot ens.fr <gcc-bugzi...@gcc.gnu.org> wrote: > > http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43716 > > --- Comment #48 from Dominique d'Humieres <dominiq at lps dot ens.fr> > 2013-01-08 19:52:39 UTC --- > From comment #40: > >> with -ffast-math, so for example >> >> if (x != 0) >> tem = y / x; >> else >> tem = 0.; >> ... do sth with tem ... >> >> will execute y / x unconditionally based on the fact that it cannot trap. > > This optimization generates an exception trapped when using -ffpe-trap=invalid > along with -ffast-math. > This unfortunately prevents any debugging based -ffpe-trap=invalid for > miscompilations occurring with -ffast-math. Well - that's maybe unfortunate but expected. You can't have both ;) > One thing I hope, though I am not > sure about it, is that the above block is still compiled as > > tem=y/x > if (x==0) tem=0. Yes, it's basically turned into an unconditional divide plus a conditional move based on the fact that we cannot vectorize non-straight-line-code (so it's really only a vectorization enabler). > My original report was for '-O3 -funsafe-math-optimizations > -ffinite-math-only' > without -ffpe-trap=invalid. The segmentation fault resulted from the fact that > some variables were used to access a table and were out of bound when the > miscompilation generated some NAN (see comment #13). Yes, that's another common issue - with FP indexing even slight rounding differences can cause bogus accesses (consider producing a[0.99] instead of a[1.0]). That mostly happens with FP loop induction variables that are also used for indexing (a really bad practice the frontend should warn about - at least when -funsafe-math-optimizations is in effect). Richard. > -- > Configure bugmail: http://gcc.gnu.org/bugzilla/userprefs.cgi?tab=email > ------- You are receiving this mail because: ------- > You are on the CC list for the bug.
