[llvm] [VPlan] Add ComputeFindLastIVResult opcode (NFC). (PR #132689)

[Mel Chen via llvm-commits]

================
@@ -614,6 +614,27 @@ Value *VPInstruction::generate(VPTransformState &State) {
     return Builder.CreateVectorSplat(
         State.VF, State.get(getOperand(0), /*IsScalar*/ true), "broadcast");
   }
+  case VPInstruction::ComputeFindLastIVResult: {
+    // The recipe's operands are the reduction phi, followed by one operand for
+    // each part of the reduction.
+    unsigned UF = getNumOperands() - 1;
+    Value *ReducedPartRdx = State.get(getOperand(1));
+    for (unsigned Part = 1; Part < UF; ++Part) {
+      ReducedPartRdx = createMinMaxOp(Builder, RecurKind::SMax, ReducedPartRdx,
+                                      State.get(getOperand(1 + Part)));
+    }
+
+    // FIXME: The cross-recipe dependency on VPReductionPHIRecipe is temporary
+    // and will be removed by breaking up the recipe further.
+    auto *PhiR = cast<VPReductionPHIRecipe>(getOperand(0));
+    // Get its reduction variable descriptor.
+    const RecurrenceDescriptor &RdxDesc = PhiR->getRecurrenceDescriptor();
+    RecurKind RK = RdxDesc.getRecurrenceKind();
+
+    assert(RecurrenceDescriptor::isFindLastIVRecurrenceKind(RK));
+    assert(!PhiR->isInLoop());
----------------
Mel-Chen wrote:

```suggestion
    assert(RecurrenceDescriptor::isFindLastIVRecurrenceKind(RK) && "Unexpected reduction kind");
    assert(!PhiR->isInLoop() && "In-loop FindLastIV reduction is not supported yet");
```
and move assertions to the front of the switch case.

https://github.com/llvm/llvm-project/pull/132689