On Fri, 9 Jan 2015, Jakub Jelinek wrote: > On Fri, Jan 09, 2015 at 03:10:16PM +0100, Richard Biener wrote: > > Well, you have until the end of next week ;) For GIMPLE this is > > a switch with all cases going to the same basic-block, right? > > I think we optimize that in cleanup_control_expr_graph via the > > single_succ_p case? > > No, it is a switch with cases that all look like: > _1 = a; // load > _2 = _1 + 1; > a = _2; // store > So, either if tree-ssa-tail-merge could be tought about loads/stores, > or some other pass would be able to hoist the loads before the switch and > sink the store after the switch, because every switch case does that.
Ah, ok. Indeed code-hoisting on GIMPLE wasn't finished (there is a very old PR with patches still), and sinking has the same issue in that it only exploits partial dead code elimination opportunities. I think that tail-merging already handles some of these cases, just maybe not the one with more than two PHI args or switches. Richard.
