[llvm] [RISCV][GISEL] Legalize G_VSCALE (PR #85967)

[Craig Topper via llvm-commits]

================
@@ -1699,6 +1699,36 @@ LegalizerHelper::LegalizeResult LegalizerHelper::narrowScalar(MachineInstr &MI,
   case TargetOpcode::G_FLDEXP:
   case TargetOpcode::G_STRICT_FLDEXP:
     return narrowScalarFLDEXP(MI, TypeIdx, NarrowTy);
+  case TargetOpcode::G_VSCALE: {
+    LLT Ty = MRI.getType(MI.getOperand(0).getReg());
+    const APInt &Val = MI.getOperand(1).getCImm()->getValue();
+    unsigned TotalSize = Ty.getSizeInBits();
+    unsigned NarrowSize = NarrowTy.getSizeInBits();
+    int NumParts = TotalSize / NarrowSize;
+
+    SmallVector<Register, 4> PartRegs;
----------------
topperc wrote:

This doesn't look right to me. Looks like you're just creating a G_VSCALE for each part using just the bits of the constant that belong to that part. G_VSCALE is a multiply, the full multiply result of the first part produce a 2x result. The second half of that full product needs to added to the second part.

https://github.com/llvm/llvm-project/pull/85967