[llvm] [VPlan] Unroll VPReplicateRecipe by VF. (PR #142433)

[via llvm-commits]

================
@@ -430,3 +431,83 @@ void VPlanTransforms::unrollByUF(VPlan &Plan, unsigned UF, LLVMContext &Ctx) {
 
   VPlanTransforms::removeDeadRecipes(Plan);
 }
+
+/// Create a single-scalar clone of RepR for lane \p Lane.
+static VPReplicateRecipe *cloneForLane(VPlan &Plan, VPBuilder &Builder,
+                                       Type *IdxTy, VPReplicateRecipe *RepR,
+                                       VPLane Lane) {
+  // Collect the operands at Lane, creating extracts as needed.
+  SmallVector<VPValue *> NewOps;
+  for (VPValue *Op : RepR->operands()) {
+    if (vputils::isSingleScalar(Op)) {
+      NewOps.push_back(Op);
+      continue;
+    }
+    VPValue *Ext;
+    if (Lane.getKind() == VPLane::Kind::ScalableLast) {
+      Ext = Builder.createNaryOp(VPInstruction::ExtractLastElement, {Op});
+    } else {
+      // Look through buildvector to avoid unnecessary extracts.
+      auto *BV = dyn_cast<VPInstruction>(Op);
+      if (BV && BV->getOpcode() == VPInstruction::BuildVector) {
+        NewOps.push_back(BV->getOperand(Lane.getKnownLane()));
+        continue;
+      }
+      VPValue *Idx =
+          Plan.getOrAddLiveIn(ConstantInt::get(IdxTy, Lane.getKnownLane()));
+      Ext = Builder.createNaryOp(Instruction::ExtractElement, {Op, Idx});
+    }
+    NewOps.push_back(Ext);
+  }
+
+  auto *New =
+      new VPReplicateRecipe(RepR->getUnderlyingInstr(), NewOps,
+                            /*IsSingleScalar=*/true, /*Mask=*/nullptr, *RepR);
+  New->insertBefore(RepR);
+  return New;
+}
+
+void VPlanTransforms::unrollByVF(VPlan &Plan, ElementCount VF) {
+  Type *IdxTy = IntegerType::get(
+      Plan.getScalarHeader()->getIRBasicBlock()->getContext(), 32);
+  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
+           vp_depth_first_shallow(Plan.getVectorLoopRegion()->getEntry()))) {
+    for (VPRecipeBase &R : make_early_inc_range(*VPBB)) {
+      auto *RepR = dyn_cast<VPReplicateRecipe>(&R);
+      if (!RepR || RepR->isSingleScalar())
+        continue;
+
+      VPBuilder Builder(RepR);
+      SmallVector<VPValue *> LaneDefs;
+      // Stores to invariant addresses only need to store the last lane.
----------------
ayalz wrote:

```suggestion
      // Stores to invariant addresses need to store the last lane only.
```


https://github.com/llvm/llvm-project/pull/142433