[llvm] SelectionDAG: Support nofpclass(nan/qnan/snan) in arguments (PR #130051)

[YunQiang Su via llvm-commits]

================
@@ -11885,6 +11885,16 @@ void SelectionDAGISel::LowerArguments(const Function &F) {
           AssertOp = ISD::AssertSext;
         else if (Arg.hasAttribute(Attribute::ZExt))
           AssertOp = ISD::AssertZext;
+        if (Arg.hasAttribute(Attribute::NoFPClass)) {
+          SDNodeFlags InValFlags = InVals[i]->getFlags();
+          bool NoSNaN = ((Arg.getNoFPClass() & llvm::fcSNan) == llvm::fcSNan);
+          bool NoQNaN = ((Arg.getNoFPClass() & llvm::fcQNan) == llvm::fcQNan);
+          InValFlags.setNoSNaNs(NoSNaN);
+          InValFlags.setNoQNaNs(NoQNaN);
+          InValFlags.setNoInfs((Arg.getNoFPClass() & llvm::fcInf) ==
+                               llvm::fcInf);
+          InVals[i]->setFlags(InValFlags);
----------------
wzssyqa wrote:

Maybe I misunderstood what you mean.
The flag here is not `FastMathFlags`, and it is not used to an `operators`.
It is `nofpclass` for variables.

Here we just pass the flags into the functions. Such as

 ```
define float @maximumnum_float_nofpclass(float nofpclass(nan) %x, float nofpclass(nan) %y) {
  %z = call float @llvm.maximumnum.f32(float %x, float %y)
  ret float %z
}
```

In `define` we have the information that `%x`, `%y` will never be `nan`, while in
`call`, the information is missing.

I agree that we should something like `ISD::AssertNoFPClass`, while it should be in
```
TargetLowering::expandFMINIMUMNUM_FMAXIMUMNUM
```
Just after we do something operation, and then we can be sure that an SDValue is not NaN, then we can insert a `ISD::AssertNoFPClass` node.

https://github.com/llvm/llvm-project/pull/130051