[llvm] [VPlan] Replicate VPScalarIVStepsRecipe by VF outside replicate regions. (PR #170053)

[via llvm-commits]

================
@@ -594,12 +595,28 @@ cloneForLane(VPlan &Plan, VPBuilder &Builder, Type *IdxTy,
                                 /*IsSingleScalar=*/true, /*Mask=*/nullptr,
                                 *RepR, *RepR, RepR->getDebugLoc());
   } else {
-    assert(isa<VPInstruction>(DefR) &&
-           "DefR must be a VPReplicateRecipe or VPInstruction");
     New = DefR->clone();
     for (const auto &[Idx, Op] : enumerate(NewOps)) {
       New->setOperand(Idx, Op);
     }
+    if (auto *Steps = dyn_cast<VPScalarIVStepsRecipe>(New)) {
+      // Skip lane 0: an absent start index is implicitly zero.
+      unsigned KnownLane = Lane.getKnownLane();
+      if (KnownLane != 0) {
+        VPTypeAnalysis TypeInfo(Plan);
+        Type *BaseIVTy = TypeInfo.inferScalarType(DefR->getOperand(0));
+        VPValue *LaneOffset = Plan.getConstantInt(
+            APInt(IdxTy->getScalarSizeInBits(), KnownLane)
+                .zextOrTrunc(BaseIVTy->getScalarSizeInBits()));
+
+        if (VPValue *StartIndex = Steps->getStartIndex()) {
+          // Add lane offset to the existing start index.
+          VPBuilder Builder(DefR);
+          LaneOffset = Builder.createAdd(StartIndex, LaneOffset);
----------------
ayalz wrote:

What if Steps represents floating-point induction rather than integer - is this case tested?
Compare with VPScalarIVStepsRecipe::execute()'s
```StartIdx = Builder.CreateBinOp(AddOp, StartIdx0, LaneValue);```
where `AddOp` is Instruction::Add if base is integer, but an FAdd or FSub otherwise?

https://github.com/llvm/llvm-project/pull/170053