The patch leads to big regression for float operators on target without hard fpu support due to register shuffle.
Please refer https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63743 for more detail. Thanks! -Zhenqiang > -----Original Message----- > From: gcc-patches-ow...@gcc.gnu.org [mailto:gcc-patches- > ow...@gcc.gnu.org] On Behalf Of Jeff Law > Sent: Saturday, October 25, 2014 1:48 AM > To: Richard Biener; gcc-patches@gcc.gnu.org > Cc: Jakub Jelinek > Subject: Re: [PATCH][6/n] Merge from match-and-simplify, make forwprop > fold all stmts > > On 10/24/14 07:16, Richard Biener wrote: > > > > This patch makes GIMPLE forwprop fold all statements, following > > single-use SSA edges only (as suggested by Jeff and certainly how this > > will regress the least until we replace manual simplification code > > that does not restrict itself this way). > > > > forwprop is run up to 4 times at the moment (once only for -Og, not at > > all for -O0), which still seems reasonable. IMHO the forwprop pass > > immediately after inlining is somewhat superfluous, it was added there > > just for its ADDR_EXPR propagation. We should eventually split this > > pass into two. > > > > Note that just folding what we propagated into (like the SSA > > propagators do during substitute-and-fold phase) will miss cases where > > we propagate into a stmt feeding the one we could simplify. Unless we > > always fold all single-use (and their use) stmts we have to fold > > everything from time to time. Changing how / when we fold stuff is > > certainly sth to look after with fold_stmt now being able to follow > > SSA edges. > > > > Bootstrapped on x86_64-unknown-linux-gnu, testing still in progress. > > > > From earlier testing I remember I need to adjust a few testcases that > > don't expect the early folding - notably two strlenopt cases > > (previously XFAILed but then PASSed again). > > > > I also expect to massage the single-use heuristic as I get to merging > > the patterns I added for the various forwprop manual pattern matchings > > to trunk (a lot of them do not restrict themselves this way). > > > > Does this otherwise look ok? > > > > Thanks, > > Richard. > > > > 2014-10-24 Richard Biener <rguent...@suse.de> > > > > * tree-ssa-forwprop.c: Include tree-cfgcleanup.h and tree-into-ssa.h. > > (lattice): New global. > > (fwprop_ssa_val): New function. > > (fold_all_stmts): Likewise. > > (pass_forwprop::execute): Finally fold all stmts. > Seems reasonable. After all, we can iterate on the single-use heuristic. > > jeff >
