SelectionDAG: mimic InstCombine when comparing to undef

[mehdi_amini]

Summary

This implements the same folding for SelectionDAG than what is present in InstCombine.

I haven't put a test yet, and it breaks the validation, but before fixing this I seek opinion on NaN handling here. InstCombine does not care and I don't really see why the same IR would be handled differently in the DAG than in InstCombine.

[hfinkel]

I haven't put a test yet, and it breaks the validation, but before fixing this I seek opinion on NaN handling here. InstCombine does not care and I don't really see why the same IR would be handled differently in the DAG than in InstCombine.

I don't know either, but that sounds wrong. This depends on the predicate, I think, for fcmp, but NaN != NaN (or anything else), and so we can't arbitrarily fold fcmp pred x, undef to true unless we know that x cannot be NaN (or we're generically allowing unsafe floating-point transformations).

lib/CodeGen/SelectionDAG/SelectionDAG.cpp

1924

Is the third check supposed to be N2?

[mehdi_amini]

Thanks, this was my impression as well, I’ll try to write a test and a patch for InstCombine, and then I’ll come back to SelectionDAG.

(good catch for N2, I'll try to have a test that stress this as well)

lib/CodeGen/SelectionDAG/SelectionDAG.cpp

     }
   }
 
+  // InstCombine implements:
+  //
+  // icmp X, X -> true/false
+  // X icmp undef -> true/false.  For example, icmp ugt %X, undef -> false
+  // because X could be 0.
+  //
+  // fcmp pred X, undef -> undef
+  //
+  // Do the same here
+  if (N1 == N2 || N1.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
+    return getConstant(isTrueWhenEqual(Cond), VT);
+
   // Could not fold it.
   return SDValue();
 }