http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59049
--- Comment #8 from Jorn Wolfgang Rennecke <amylaar at gcc dot gnu.org> --- (In reply to Richard Biener from comment #7) > That is, sth like > > Index: gcc/tree-ssa-ter.c > =================================================================== > --- gcc/tree-ssa-ter.c (revision 204664) > +++ gcc/tree-ssa-ter.c (working copy) > @@ -438,6 +439,12 @@ ter_is_replaceable_p (gimple stmt) > && !is_gimple_val (gimple_assign_rhs1 (stmt))) > return false; > > + /* Do not propagate "modeless" constants - we may end up confusing > the RTL > + expanders. Leave the optimization to RTL CCP. */ > + if (gimple_assign_single_p (stmt) > + && CONSTANT_CLASS_P (gimple_assign_rhs1 (stmt))) > + return false; > + > return true; > } > return false; Constants are often very valuable for rtl expansion, allowing to use cheaper patterns. And some constant propagations are impossible in rtl because of mode oddities. E.g. when you have a have a mulsidi3 pattern, you generally have a sign_extend - you can't have a VOIDmode constant inside that. Therefore, I would rather have the middle-end move the constants to registers only when necessary to preserve the mode, and preferrably fold instead in the first place when optimizing.
