[llvm] [RFC] IR: Support atomicrmw FP ops with vector types (PR #86796)

[Matt Arsenault via llvm-commits]

================
@@ -8222,7 +8222,7 @@ int LLParser::parseAtomicRMW(Instruction *&Inst, PerFunctionState &PFS) {
               " operand must be an integer, floating point, or pointer type");
     }
   } else if (IsFP) {
-    if (!Val->getType()->isFloatingPointTy()) {
+    if (!Val->getType()->isFPOrFPVectorTy()) {
----------------
arsenm wrote:

Do any scalable architectures support vector atomics? I assume it would fail in codegen somewhere. Is that worse than having it be considered invalid IR?

https://github.com/llvm/llvm-project/pull/86796