[llvm] [RISCV] Form VFWMUL.VF when fp_extend is scalar and then splatted. (PR #98590)

[Craig Topper via llvm-commits]

================
@@ -14518,6 +14525,24 @@ struct NodeExtensionHelper {
     case RISCVISD::VMV_V_X_VL:
       fillUpExtensionSupportForSplat(Root, DAG, Subtarget);
       break;
+    case RISCVISD::VFMV_V_F_VL: {
+      MVT VT = OrigOperand.getSimpleValueType();
+
+      if (!OrigOperand.getOperand(0).isUndef())
+        break;
+
+      SDValue Op = OrigOperand.getOperand(1);
+      if (Op.getOpcode() != ISD::FP_EXTEND)
+        break;
+
+      unsigned NarrowSize = VT.getScalarSizeInBits() / 2;
+      unsigned ScalarBits = Op.getOperand(0).getValueSizeInBits();
+      if (NarrowSize != ScalarBits)
+        break;
+
+      SupportsFPExt = true;
----------------
topperc wrote:

I'm not sure. For the integer case we use MaskedValueIsZero and ComputeMaxSignifcant bits so we never directly bypass a scalar zext/sext. That's a bit different than this combine which does bypass a scalar fp_extend.

https://github.com/llvm/llvm-project/pull/98590