[RISCV] Teach fillUpExtensionSupportForSplat to handle nxvXi64 VMV_V_X_VL on RV32. (#99251)

Summary:
A nxvXi64 VMV_V_X_VL on RV32 sign extends its 32 bit input to 64 bits.
If that input is positive, the sign extend can also be considered as a
zero extend.

Test Plan: 

Reviewers: 

Subscribers: 

Tasks: 

Tags: 


Differential Revision: https://phabricator.intern.facebook.com/D60250893

Author: topperc

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

@@ -14503,10 +14503,21 @@ struct NodeExtensionHelper {
     // widening opcode by splatting to smaller element size.
     unsigned EltBits = VT.getScalarSizeInBits();
     unsigned ScalarBits = Op.getValueSizeInBits();
-    // Make sure we're getting all element bits from the scalar register.
-    // FIXME: Support implicit sign extension of vmv.v.x?
-    if (ScalarBits < EltBits)
+    // If we're not getting all bits from the element, we need special handling.
+    if (ScalarBits < EltBits) {
+      // This should only occur on RV32.
+      assert(Opc == RISCVISD::VMV_V_X_VL && EltBits == 64 && ScalarBits == 32 &&
+             !Subtarget.is64Bit() && "Unexpected splat");
+      // vmv.v.x sign extends narrow inputs.
+      SupportsSExt = true;
+
+      // If the input is positive, then sign extend is also zero extend.
+      if (DAG.SignBitIsZero(Op))
+        SupportsZExt = true;
+
+      EnforceOneUse = false;
       return;
+    }
 
     unsigned NarrowSize = EltBits / 2;
     // If the narrow type cannot be expressed with a legal VMV,