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

[Princeton Ferro via llvm-commits]

================
@@ -903,6 +921,11 @@ NVPTXTargetLowering::NVPTXTargetLowering(const NVPTXTargetMachine &TM,
     }
   }
 
+  // Expand v2f32 = fp_extend
+  setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Expand);
+  // Expand v2[b]f16 = fp_round v2f32
+  setOperationAction(ISD::FP_ROUND, {MVT::v2bf16, MVT::v2f16}, Expand);
----------------
Prince781 wrote:

This wasn't needed before because `fp_round v2f32` was already being expanded since `v2f32` was an invalid type. Now we need it, otherwise this node will persist to isel.

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