http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58950
--- Comment #8 from Marc Glisse <glisse at gcc dot gnu.org> --- (In reply to Eric Botcazou from comment #5) > > For __builtin_shuffle, the issue is that we now call save_expr, which always > > sets TREE_SIDE_EFFECTS to 1. I don't know if it would make sense to > > introduce a maybe_save_expr that is equivalent to save_expr but does not set > > TREE_SIDE_EFFECTS if its argument doesn't have it. > > No, this would defeat the purpose of the SAVE_EXPR, since you could > duplicate the expression or move it at will, leading to nasty order of > evaluation issues. What I meant was, there are many places in the front-end that do: if (TREE_SIDE_EFFECTS (op0)) op0 = save_expr (op0); and I'd like to replace it with op0 = maybe_save_expr (op0); that would do the same thing to begin with. Now, if op0 is a long expression without side-effects, the code I quoted will not call save_expr, it will duplicate the expression and count on CSE for dedup, whereas we could imagine having a single version to start with, that optimizations would still be allowed to duplicate as usual if they consider it worth it. Of course the uses of SAVE_EXPR that require the stronger protection would be left alone. I may still not be making sense, but at least I think I was a little more precise about what I meant ;-)
