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

[Matt Arsenault 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);
----------------
arsenm wrote:

This isn't general to all of the fields of nofpclass; if we're going to do something with this I would expect it to look like the AssertSext and AssertZext nodes.

This usage of the fast math flags is kind of abusing the system. At the IR level, only specific operators are permitted to have the flags. The DAG happens to be more permissive, and you're putting this on whatever type of value the argument happens to have been replaced with. I'm not sure we can trust putting flags like this on a random merge, copy from reg, load or arbitrary other operator an argument could have been lowered from 

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