[llvm] [VPlan] Add transformation to narrow interleave groups. (PR #106441)

[Alexey Bataev via llvm-commits]

================
@@ -1623,3 +1661,129 @@ void VPlanTransforms::createInterleaveGroups(
       }
   }
 }
+
+static bool supportedLoad(VPWidenRecipe *R0, VPValue *V, unsigned Idx) {
+  if (auto *W = dyn_cast_or_null<VPWidenLoadRecipe>(V->getDefiningRecipe())) {
+    if (W->getMask())
+      return false;
+    return !W->getMask() && (R0->getOperand(0) == V || R0->getOperand(1) == V);
+  }
+
+  if (auto *IR = dyn_cast_or_null<VPInterleaveRecipe>(V->getDefiningRecipe())) {
+    return IR->getInterleaveGroup()->getFactor() ==
+               IR->getInterleaveGroup()->getNumMembers() &&
+           IR->getVPValue(Idx) == V;
+  }
+  return false;
+}
+
+/// Returns true if \p IR is a full interleave group with factor and number of
+/// members both equal to \p VF.
+static bool isConsecutiveInterleaveGroup(VPInterleaveRecipe *IR,
+                                         ElementCount VF) {
+  if (!IR)
+    return false;
+  auto IG = IR->getInterleaveGroup();
+  return IG->getFactor() == IG->getNumMembers() &&
+         IG->getNumMembers() == VF.getFixedValue();
+}
+
+void VPlanTransforms::narrowInterleaveGroups(VPlan &Plan, ElementCount VF) {
+  using namespace llvm::VPlanPatternMatch;
+  if (VF.isScalable())
+    return;
+
+  SmallVector<VPInterleaveRecipe *> StoreGroups;
+  for (auto &R : *Plan.getVectorLoopRegion()->getEntryBasicBlock()) {
+    if (match(&R, m_BranchOnCount(m_VPValue(), m_VPValue())) ||
+        isa<VPCanonicalIVPHIRecipe>(&R))
+      continue;
+
+    // Bail out on recipes not supported at the moment:
+    //  * phi recipes other than the canonical induction
+    //  * recipes writing to memory except interleave groups
+    // Only support plans with a canonical induction phi.
+    if (R.isPhi())
+      return;
+
+    auto *IR = dyn_cast<VPInterleaveRecipe>(&R);
+    if (R.mayWriteToMemory() && !IR)
+      return;
+
+    if (!IR)
+      continue;
+
+    if (!isConsecutiveInterleaveGroup(IR, VF))
+      return;
+    if (IR->getStoredValues().empty())
+      continue;
+
+    auto *Lane0 = dyn_cast_or_null<VPWidenRecipe>(
+        IR->getStoredValues()[0]->getDefiningRecipe());
+    if (!Lane0)
+      return;
+    for (const auto &[I, V] : enumerate(IR->getStoredValues())) {
+      auto *R = dyn_cast<VPWidenRecipe>(V->getDefiningRecipe());
+      if (!R || R->getOpcode() != Lane0->getOpcode())
+        return;
+      // Work around captured structured bindings being a C++20 extension.
+      auto Idx = I;
+      if (any_of(R->operands(), [Lane0, Idx](VPValue *V) {
----------------
alexey-bataev wrote:

```suggestion
      if (any_of(R->operands(), [Lane0, Idx=I](VPValue *V) {
```


https://github.com/llvm/llvm-project/pull/106441