[llvm] [AMDGPU] Make getDWordFromOffset robust against exotic types (PR #70153)

[Matt Arsenault via llvm-commits]

================
@@ -11230,26 +11230,41 @@ static bool hasNon16BitAccesses(uint64_t PermMask, SDValue &Op,
 static SDValue getDWordFromOffset(SelectionDAG &DAG, SDLoc SL, SDValue Src,
                                   unsigned DWordOffset) {
   SDValue Ret;
-  if (Src.getValueSizeInBits() <= 32)
+
+  auto ValueSize = Src.getValueSizeInBits().getFixedValue();
+
+  if (ValueSize <= 32)
     return DAG.getBitcastedAnyExtOrTrunc(Src, SL, MVT::i32);
 
-  if (Src.getValueSizeInBits() >= 256) {
-    assert(!(Src.getValueSizeInBits() % 32));
-    Ret = DAG.getBitcast(
-        MVT::getVectorVT(MVT::i32, Src.getValueSizeInBits() / 32), Src);
+  if (!(ValueSize % 32)) {
+    Ret = DAG.getBitcast(MVT::getVectorVT(MVT::i32, ValueSize / 32), Src);
----------------
arsenm wrote:

I would expect this to use EVT. What happens if the number of elements isn't a defined MVT?

https://github.com/llvm/llvm-project/pull/70153