https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104280
--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> --- (In reply to Andrew Pinski from comment #3) > (In reply to Jakub Jelinek from comment #2) > > To me this looks like a bug in gimple-fold.cc. > > The r12-6924-gc2b610e7c6c89fd4 simplifies > > _5 = 1 / c.0_4; > > into > > _5 = c.0_4 == 1; > > No this is not valid gimple as the comparisons always return a boolean type. > There needs to be a cast added. I don't think we ever say that, nor we enforce that. It is true that we don't fold _Bool _1; int _4; <bb 2> : _1 = a_2(D) == b_3(D); _4 = (int) _1; return _4; to just _4 = a_2(D) == b_3(D); return _4; though. And it is true that e.g. truth_valued_p would misbehave if the comparison/truth_* result is precision 1 signed integer (unless truth valued is defined as 0 or non-zero rather than 0 and 1). Certainly e.g. in Fortran EQ_EXPR etc. result doesn't need to be boolean_type_node, can be also any of the logical(kind=8) etc. types.
