[Bug tree-optimization/94589] Optimize (i<=>0)>0 to i>0

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94589

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2020-04-14
     Ever confirmed|0                           |1
                 CC|                            |rguenth at gcc dot gnu.org

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Hmm.  I would have said phiopt but then there's the missing opportunity to
handle PHIs as COND_EXPRs and simplify them via match.pd patterns.  In this
case simplify

 i_1(D) != 0 ? i_1(D) >= 0 : 0

that could be forwprops job then.  The only pecularity is when the PHI
isn't the only thing controlled by the condition, thus the control flow
not vanishing with replacing the PHI.  In this case we'd turn the code into

  <bb 2> [local count: 1073741824]:
  if (i_1(D) != 0)
    goto <bb 3>; [50.00%]
  else
    goto <bb 4>; [50.00%]

  <bb 3> [local count: 536870913]:
  _2 = i_1(D) >= 0;

  <bb 4> [local count: 1073741824]:
  prephitmp_6 = i_1(D) > 0;
  return prephitmp_6;

and leave the rest to DCE.  Obviously if the other two compares are not
dead afterwards this isn't going to be a good idea.

And CFG pattern matching is done by phiopt - which for the actual
simplification could resort to match.pd simplifying a COND_EXPR as well.
And yeah, I think there's a dup about PHIs and "folding".