http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55214
Tobias Burnus <burnus at gcc dot gnu.org> changed: What |Removed |Added ---------------------------------------------------------------------------- Keywords| |wrong-code Status|UNCONFIRMED |NEW Last reconfirmed| |2012-11-05 CC| |burnus at gcc dot gnu.org Ever Confirmed|0 |1 --- Comment #2 from Tobias Burnus <burnus at gcc dot gnu.org> 2012-11-05 14:56:56 UTC --- Confirmed - and not a regression as it also fails with GCC 4.1. >From the C-like dump file (use -fdump-tree-original to create it): For WHERE (L), one has: cond.3 = (logical(kind=1)) l[S.2 + -1]; temp.0[count.1] = cond.3; Which is then used in WHERE (SQRT(X)>0): cond.7 = __builtin_sqrtf (x[S.6 + -1]) > 0.0; mask.8 = temp.0[count.5]; temp.4[count.5] = mask.8 && cond.7; The proper way would be: mask.8 = temp.0[count.5]; temp.4[count.5] = mask.8 ? __builtin_sqrtf (x[S.6 + -1]) > 0.0 : mask.8; That is, the condition only gets evaluated if the outer mask is true.