[Bug tree-optimization/80215] New: VN has trouble with conditional equivalences (DOM/FRE)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80215

            Bug ID: 80215
           Summary: VN has trouble with conditional equivalences (DOM/FRE)
           Product: gcc
           Version: 7.0.1
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: rguenth at gcc dot gnu.org
  Target Milestone: ---

Currently DOM doesn't handle either redundancy in foo or bar while FRE only
handles the first.

int d, e;
int foo (int a, int c)
{
  e = a * c;
  if (a == c)
    return a * c;
  return 0;
}
int bar (int a, int c)
{
  e = a * c;
  if (a == c)
    return c * c;
  return 0;
}

another example is

int e;
int baz (int a, int c)
{
  e = a + c;
  if (a == c)
    return c + c;
  return 0;
}

with the additional complication of c + c -> 2 * c canonicalization.

The issue is unifying of two values (plus eventual followup simplifications
that might be possible).  VN algorithms generally do not handle this well.
For DOM-VRP I played with re-propagating starting from uses of the
newly recorded equivalency.