On Mon, 4 Aug 2014, Jakub Jelinek wrote: > On Mon, Aug 04, 2014 at 12:26:01PM +0200, Richard Biener wrote: > > I think that tree-ssa-forwprop.c already simplifies this in > > associate_plusminus with (T)(P + A) - (T)P -> (T)A. Well, > > maybe not - but then the code should be massages to handle it. > > > > Can you double-check and do that? > > > > Otherwise I agree with removing this (and other) "premature" > > optimizations scattered throughout the frontends (and convert.c). > > IMHO a precondition of such changes is exactly verifying the middle-end > can perform those optimizations later on and if not, teaching the middle-end > to optimize that.
Yep. Which is why I said "double-check and do that", "do that" in extend what tree-ssa-forwrpop.c does. Richard.