On Thu, Nov 27, 2014 at 6:48 AM, Richard Biener <rguent...@suse.de> wrote: > > Currently the order in which patterns in match.pd match is pretty > random (well, it's deterministic of course). This is bad if there > are multiple possible matches such as in > > z = x + 1; > w = z + 0; > > where we have both z + 0 -> z and (x + 1) + 0 -> x + (1 + 0). > Currently the latter happens to be applied and thus we simplify to > > z = x + 1; > w = x + 1; > > which is bad for propagation engines which only handle constant > or SSA name results (and also either generates redundant code or leaves > dead code). > > Thus the following patch makes sure that if there are multiple > matches possible then the pattern first appearing in match.pd > is chosen. > > This slightly increases the number of instructions executed during > the matching in the worst case but is certainly the best course > of action here (other possibilities are to make the patterns > disjunct by disabling the latter for the + 0 case). At least > it matches how other generated matchers behave. > > Bootstrap and regtest running on x86_64-unknown-linux-gnu. > > Richard. > > 2014-11-27 Richard Biener <rguent...@suse.de> > > PR middle-end/64084 > * genmatch.c (dt_node::gen_kids_1): New function, split out > from dt_node::gen_kids. > (decision_tree::cmp_node): DT_TRUE are generally not equal. > (decision_tree::find_node): Treat DT_TRUE as barrier for > node CSE on the same level. > (dt_node::append_node): Do not keep DT_TRUE last. > (dt_node::gen_kids): Emit code after each DT_TRUE node seen.
This caused: FAIL: gcc.dg/pr37289.c scan-tree-dump original "-\\(long unsigned int\\) x" on x86. -- H.J.
