[llvm] [NVPTX] support packed f32 instructions for sm_100+ (PR #126337)

[Alex MacLean via llvm-commits]

================
@@ -1425,14 +1466,21 @@ bool NVPTXDAGToDAGISel::tryStore(SDNode *N) {
   // Type Setting: toType + toTypeWidth
   // - for integer type, always use 'u'
   MVT ScalarVT = SimpleVT.getScalarType();
-  unsigned ToTypeWidth = ScalarVT.getSizeInBits();
   if (SimpleVT.isVector()) {
-    assert((Isv2x16VT(StoreVT) || StoreVT == MVT::v4i8) &&
-           "Unexpected vector type");
-    // v2x16 is stored using st.b32
-    ToTypeWidth = 32;
+    switch (StoreVT.getSimpleVT().SimpleTy) {
+    case MVT::v2f16:
+    case MVT::v2bf16:
+    case MVT::v2i16:
+    case MVT::v4i8:
+    case MVT::v2f32:
----------------
AlexMaclean wrote:

I'm seeing something like this in lots of place (`Isv2x16VT(VT) || VT == MVT::v4i8 || VT == MVT::v2f32`). Lets add a helper function in NVPTXUtilities.h and use that throughout.

https://github.com/llvm/llvm-project/pull/126337