------- Comment #11 from mark at codesourcery dot com 2005-10-19 22:14 ------- Subject: Re: [3.4/4.0/4.1 Regression] ICE with invert conditional containing throw
rguenth at gcc dot gnu dot org wrote: > ------- Comment #10 from rguenth at gcc dot gnu dot org 2005-10-19 21:51 > ------- > Ok, I'll see how big the middle-end change would get. The easiest way would > be to change invert_truthvalue to ignore void types and do nothing for them. > Like > > + if (VOID_TYPE_P (TREE_TYPE (arg))) > + return arg; > gcc_assert (TREE_CODE (TREE_TYPE (arg)) == BOOLEAN_TYPE); > return build1 (TRUTH_NOT_EXPR, type, arg); > > but auditing all other places we look at COND_EXPR is a big task. Yes, that looks like the right change to invert_truthvalue, or you could even bury that in the COND_EXPR case, since that's the only place we expect that. To be clear, I have no problem whatsoever with adding the *((T*) 0) stuff during gimplification, and that will protect all the optimizers from this case. But, invert_truthvalue has to operate on more than just GENERIC/GIMPLE; it's designed to be called from the front ends, so it has to be tolerant of front-end stuff. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24439
