https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68894
--- Comment #12 from Andrew Pinski <pinskia at gcc dot gnu.org> --- Just a quick note on this part of the bug report: (In reply to Richard Biener from comment #3) > Doing this in a classical way in phi-opt might end up being slightly > convoluted. > So I'd propose to try to utilize match-and-simplify by adding > > > (cond (ge @0 @1) (max:s @0 @2) (max:s @1 @2)) > -> (max (max @0 @1) @2) I am going to submit patches for the above later today or tomorrow. > > kind patterns and from phiopt doing exploded queries of this simplification > by seeding the cond expr from the dominating condition plus the PHI operands. > > res = gimple_simplify (COND_EXPR, build2 /* Ick */ (cond-code, > boolean_type, cond-op0, cond-op1), true-phi-arg, false-phi-arg, &seq, > follow_single_use_edges); > > of course you need more than a single pattern or consider swapped ops 2 and 3 > (not yet supported). In the end explicit pattern explosion might make a > manual > implementation in phiopt easier (who knows). This part is fully implemented in r14-204-gf1f5cbaa3f716fcb472dee5 (there was a few pieces of the match-and-simplify phiopt implemented in GCC 12). Note phiopt already does "a ? b : c" -> "!a ? c : b" (where !a folds) (implemented in r12-2040-ga50cecb20a10) specifically because of the explosion issue.
