https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119903
--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> --- (In reply to Richard Biener from comment #3) > There is likely similar issues with -ftrapv -fnon-call-exceptions and > > _1 = _2 + 3; > > _3 = _1 - 3; > > > where in general forwprop fails to properly replace an internally throwing > stmt with its folded variant. There is a difference though, those will reuse the last gimple statement and end up with the "correct" landing pad. While here we don't reuse that as a trapping comparison can't be inside a GIMPLE_COND. > > With -fno-delete-dead-exceptions your testcase gives the following, > highlighting the issue some more: > > <bb 2> : > _6 = a_5(D) < 1.0e+0; > > <bb 3> : > _13 = a_5(D) < 1.0e+0; > if (_13 != 0) > goto <bb 4>; [INV] > > > I believe that in general the match.pd SSA def following should not follow > to internally throwing defs. Either in the core workers or by means of > the valueization hook of individual passes. Probably the former (though > it makes such lookup a bit expensive and fragile for stmts not in the CFG). Let me see if I can figure that out.
