SelectionDAG/expandFMINIMUMNUM_FMAXIMUMNUM: FCANONICALIZE is needed only for sNaN

[wzssyqa]

If we are sure that it is not sNaN, even it may be qNaN, we can use it directly.

[llvmbot]

@llvm/pr-subscribers-llvm-selectiondag

Author: YunQiang Su (wzssyqa)

Changes

If we are sure that it is not sNaN, even it may be qNaN, we can use it directly.

---

Full diff: https://github.com/llvm/llvm-project/pull/108180.diff

1 Files Affected:

• (modified) llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp (+2-2)

diff --git a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index 03010c1df00145..042661af1c8f96 100644
--- a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -8614,8 +8614,8 @@ SDValue TargetLowering::expandFMINIMUMNUM_FMAXIMUMNUM(SDNode *Node,
   SDValue MinMax =
       DAG.getSelectCC(DL, LHS, RHS, LHS, RHS, IsMax ? ISD::SETGT : ISD::SETLT);
   // If MinMax is NaN, let's quiet it.
-  if (!Flags.hasNoNaNs() && !DAG.isKnownNeverNaN(LHS) &&
-      !DAG.isKnownNeverNaN(RHS)) {
+  if (!Flags.hasNoNaNs() && !DAG.isKnownNeverSNaN(LHS) &&
+      !DAG.isKnownNeverSNaN(RHS)) {
     MinMax = DAG.getNode(ISD::FCANONICALIZE, DL, VT, MinMax, Flags);
   }
 

[arsenm]

Needs test changes

[wzssyqa]

Needs test changes

I only see that we have nnan option while there is no similiar one for sNaN.
https://llvm.org/docs/LangRef.html

How to mark a variable to never SNaN while may qNaN?

[wzssyqa]

I guess t hat I am confused by nofpclass in https://llvm.org/docs/LangRef.html#id1932.
Can it used to mark an argument that it will never be a type?

I have a try, while it won't as DAG.isKnownNeverNaN won't return true in this case.
Maybe I misunderstand its meaning.

[wzssyqa]

Needs test changes
Once #108350 is merged, I will fix SelectionDAG::isKnownNeverNaN with sNaN, and add test case for it.

[arsenm]

I guess t hat I am confused by nofpclass in https://llvm.org/docs/LangRef.html#id1932. Can it used to mark an argument that it will never be a type?

I have a try, while it won't as DAG.isKnownNeverNaN won't return true in this case. Maybe I misunderstand its meaning.

isKnownNeverNaN is quite underdeveloped. You can most easily test the no-nan operand case by just having a filler fadd nnan instruction. It would be simpler to do that than pulling in a dependency on some more general nofpclass handling