[llvm] [RISCV][ISel] Combine scalable vector fadd/fsub/fmul with fp extend. (PR #88615)

[via llvm-commits]

================
@@ -13597,6 +13599,13 @@ struct NodeExtensionHelper {
     case RISCVISD::VZEXT_VL:
     case RISCVISD::FP_EXTEND_VL:
       return OrigOperand.getOperand(0);
+    case ISD::SPLAT_VECTOR: {
+      SDValue Op = OrigOperand.getOperand(0);
+      if (Op.getOpcode() == ISD::FP_EXTEND)
+        return Op;
----------------
sun-jacobi wrote:

> Why do we need to look through scalar FP_EXTEND here, but we don't need to look throug ZERO_EXTEND or SIGN_EXTEND for integer? 

For integers, I think the `DAG.getSplat` would be an implicit truncation for us. But it seems that the float does not work. 

https://github.com/llvm/llvm-project/pull/88615