[llvm] [VPlan] Implement interleaving as VPlan-to-VPlan transform. (PR #95842)

[via llvm-commits]

================
@@ -1572,3 +1572,421 @@ void VPlanTransforms::dropPoisonGeneratingRecipes(
     }
   }
 }
+
+static VPValue *getInterleavedValue(
+    DenseMap<VPValue *, SmallVector<VPValue *>> &InterleavedValues, VPValue *V,
+    unsigned IC) {
+  if (IC == 0)
+    return V;
+  if (V->isLiveIn())
+    return V;
+  return InterleavedValues[V][IC - 1];
+}
+
+static void interleaveReplicateRegion(
+    VPRegionBlock *VPR, VPlan &Plan, unsigned IC,
+    DenseMap<VPValue *, SmallVector<VPValue *>> &InterleavedValues) {
+  Type *CanIVIntTy = Plan.getCanonicalIV()->getScalarType();
+  VPBlockBase *InsertPt = VPR;
+  for (unsigned I = 1; I != IC; ++I) {
+    auto *Copy = VPR->clone();
+    VPBlockUtils::insertBlockAfter(Copy, InsertPt);
+    InsertPt = Copy;
+
+    ReversePostOrderTraversal<VPBlockShallowTraversalWrapper<VPBlockBase *>>
+        RPOT(Copy->getEntry());
+    ReversePostOrderTraversal<VPBlockShallowTraversalWrapper<VPBlockBase *>>
+        RPOT2(VPR->getEntry());
+    for (const auto &[New, Old] :
+         zip(VPBlockUtils::blocksOnly<VPBasicBlock>(RPOT),
+             VPBlockUtils::blocksOnly<VPBasicBlock>(RPOT2))) {
+      if (New->getParent() != Copy)
+        break;
+      for (const auto &[CopyR, OrigR] : zip(*New, *Old)) {
+        for (unsigned Idx = 0; Idx != CopyR.getNumOperands(); ++Idx) {
+          CopyR.setOperand(Idx, getInterleavedValue(InterleavedValues,
+                                                    CopyR.getOperand(Idx), I));
----------------
ayalz wrote:

This requires recording in InterleavedValues the operands of CopyR before visiting it, i.e., RPOT if thos recording takes place next in the same loop. But if this recording takes place first, updating the operands of CopyR (from Part 0) to Part I can be done by visiting the New recipes in any order.

https://github.com/llvm/llvm-project/pull/95842