[llvm] [AMDGPU] Do not treat bitcast across FP types as canonicality-preserving (PR #203560)

[Matt Arsenault via llvm-commits]

================
@@ -15643,8 +15643,17 @@ SDValue SITargetLowering::performRcpCombine(SDNode *N,
 }
 
 bool SITargetLowering::isCanonicalized(SelectionDAG &DAG, SDValue Op,
-                                       SDNodeFlags UserFlags,
-                                       unsigned MaxDepth) const {
+                                       SDNodeFlags UserFlags, unsigned MaxDepth,
+                                       EVT QueryVT) const {
+  EVT VT = Op.getValueType();
+  if (VT.isFloatingPoint()) {
+    EVT ScalarVT = VT.getScalarType();
+    if (QueryVT == EVT())
+      QueryVT = ScalarVT;
----------------
arsenm wrote:

I think I'd prefer a separate overload which omits the QueryVT and assumes the origin value's type 

https://github.com/llvm/llvm-project/pull/203560