[Bug tree-optimization/64084] New: [5 Regression] match-and-simplify prefers complex matches

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64084

            Bug ID: 64084
           Summary: [5 Regression] match-and-simplify prefers complex
                    matches
           Product: gcc
           Version: 5.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: rguenth at gcc dot gnu.org

For

int foo (int x)
{
  int y = 0;
  int z = x + 1;
  return z + y;
}

the first CCP pass ends up producing

  z_3 = x_2(D) + 1;
  _4 = x_2(D) + 1;
  return _4;

because (A + CST) + CST' -> A + (CST + CST') matches before A + 0 -> A.

This is because we build the decision tree depth-first (the idea was to
delay possibly expensive predicate evaluations as much as possible).  I suppose
if we'd build the decision tree breath first we would match the "correct"
one first.

For the testcase FRE of course fixes things up for us.  Similarly we could
drop SSA edge following from the folding done in substitute_and_fold.  But
forwprop would have the same issue (in fact it _did_ have the same issue
in 4.9 if you disable CCP - or rather not, because then nothing propagates
the zero).