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

Explicitly unroll VPReplicateRecipes outside replicate regions by VF,
replacing them by VF single-scalar recipes. Extracts for operands are
added as needed and the scalar results are combined to a vector using a
new BuildVector VPInstruction.

It also adds a few folds to simplify unnecessary extracts/BuildVectors.

It also adds a BuildStructVector opcode for handling of calls that have
struct return types.

VPReplicateRecipe in replicate regions can will be unrolled as follow
up, turing non-single-scalar VPReplicateRecipes into 'abstract', i.e.
not executable.

PR: llvm#142433

Author: fhahn

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -7328,6 +7328,7 @@ DenseMap<const SCEV *, Value *> LoopVectorizationPlanner::executePlan(
   // cost model is complete for better cost estimates.
   VPlanTransforms::runPass(VPlanTransforms::unrollByUF, BestVPlan, BestUF,
                            OrigLoop->getHeader()->getContext());
+  VPlanTransforms::runPass(VPlanTransforms::replicateByVF, BestVPlan, BestVF);
   VPlanTransforms::runPass(VPlanTransforms::materializeBroadcasts, BestVPlan);
   if (hasBranchWeightMD(*OrigLoop->getLoopLatch()->getTerminator()))
     VPlanTransforms::runPass(VPlanTransforms::addBranchWeightToMiddleTerminator,

llvm/lib/Transforms/Vectorize/VPlan.cpp

@@ -261,6 +261,13 @@ Value *VPTransformState::get(const VPValue *Def, const VPLane &Lane) {
     return Data.VPV2Scalars[Def][0];
   }
 
+  // Look through BuildVector to avoid redundant extracts.
+  // TODO: Remove once replicate regions are unrolled explicitly.
+  if (Lane.getKind() == VPLane::Kind::First && match(Def, m_BuildVector())) {
+    auto *BuildVector = cast<VPInstruction>(Def);
+    return get(BuildVector->getOperand(Lane.getKnownLane()), true);
+  }
+
   assert(hasVectorValue(Def));
   auto *VecPart = Data.VPV2Vector[Def];
   if (!VecPart->getType()->isVectorTy()) {

llvm/lib/Transforms/Vectorize/VPlan.h

@@ -936,6 +936,13 @@ class VPInstruction : public VPRecipeWithIRFlags,
     BranchOnCount,
     BranchOnCond,
     Broadcast,
+    /// Given operands of (the same) struct type, creates a struct of fixed-
+    /// width vectors each containing a struct field of all operands. The
+    /// number of operands matches the element count of every vector.
+    BuildStructVector,
+    /// Creates a fixed-width vector containing all operands. The number of
+    /// operands matches the vector element count.
+    BuildVector,
     ComputeAnyOfResult,
     ComputeFindLastIVResult,
     ComputeReductionResult,

llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp

@@ -108,6 +108,8 @@ Type *VPTypeAnalysis::inferScalarTypeForRecipe(const VPInstruction *R) {
   case VPInstruction::CalculateTripCountMinusVF:
   case VPInstruction::CanonicalIVIncrementForPart:
   case VPInstruction::AnyOf:
+  case VPInstruction::BuildStructVector:
+  case VPInstruction::BuildVector:
     return SetResultTyFromOp();
   case VPInstruction::FirstActiveLane:
     return Type::getIntNTy(Ctx, 64);

llvm/lib/Transforms/Vectorize/VPlanPatternMatch.h

@@ -221,6 +221,13 @@ struct Recipe_match {
   }
 
   bool match(const VPRecipeBase *R) const {
+    if (std::tuple_size<Ops_t>::value == 0) {
+      assert(Opcode == VPInstruction::BuildVector &&
+             "can only match BuildVector with empty ops");
+      auto *VPI = dyn_cast<VPInstruction>(R);
+      return VPI && VPI->getOpcode() == VPInstruction::BuildVector;
+    }
+
     if ((!matchRecipeAndOpcode<RecipeTys>(R) && ...))
       return false;
 
@@ -263,6 +270,10 @@ struct Recipe_match {
   }
 };
 
+template <unsigned Opcode, typename... RecipeTys>
+using ZeroOpRecipe_match =
+    Recipe_match<std::tuple<>, Opcode, false, RecipeTys...>;
+
 template <typename Op0_t, unsigned Opcode, typename... RecipeTys>
 using UnaryRecipe_match =
     Recipe_match<std::tuple<Op0_t>, Opcode, false, RecipeTys...>;
@@ -271,6 +282,9 @@ template <typename Op0_t, unsigned Opcode>
 using UnaryVPInstruction_match =
     UnaryRecipe_match<Op0_t, Opcode, VPInstruction>;
 
+template <unsigned Opcode>
+using ZeroOpVPInstruction_match = ZeroOpRecipe_match<Opcode, VPInstruction>;
+
 template <typename Op0_t, unsigned Opcode>
 using AllUnaryRecipe_match =
     UnaryRecipe_match<Op0_t, Opcode, VPWidenRecipe, VPReplicateRecipe,
@@ -302,6 +316,12 @@ using AllBinaryRecipe_match =
     BinaryRecipe_match<Op0_t, Op1_t, Opcode, Commutative, VPWidenRecipe,
                        VPReplicateRecipe, VPWidenCastRecipe, VPInstruction>;
 
+/// BuildVector is matches only its opcode, w/o matching its operands as the
+/// number of operands is not fixed.
+inline ZeroOpVPInstruction_match<VPInstruction::BuildVector> m_BuildVector() {
+  return ZeroOpVPInstruction_match<VPInstruction::BuildVector>();
+}
+
 template <unsigned Opcode, typename Op0_t>
 inline UnaryVPInstruction_match<Op0_t, Opcode>
 m_VPInstruction(const Op0_t &Op0) {

llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp

@@ -551,6 +551,11 @@ Value *VPInstruction::generate(VPTransformState &State) {
   }
   case Instruction::ExtractElement: {
     assert(State.VF.isVector() && "Only extract elements from vectors");
+    if (getOperand(1)->isLiveIn()) {
+      unsigned IdxToExtract =
+          cast<ConstantInt>(getOperand(1)->getLiveInIRValue())->getZExtValue();
+      return State.get(getOperand(0), VPLane(IdxToExtract));
+    }
     Value *Vec = State.get(getOperand(0));
     Value *Idx = State.get(getOperand(1), /*IsScalar=*/true);
     return Builder.CreateExtractElement(Vec, Idx, Name);
@@ -664,6 +669,34 @@ Value *VPInstruction::generate(VPTransformState &State) {
     return Builder.CreateVectorSplat(
         State.VF, State.get(getOperand(0), /*IsScalar*/ true), "broadcast");
   }
+  case VPInstruction::BuildStructVector: {
+    // For struct types, we need to build a new 'wide' struct type, where each
+    // element is widened, i.e., we create a struct of vectors.
+    auto *StructTy =
+        cast<StructType>(State.TypeAnalysis.inferScalarType(getOperand(0)));
+    Value *Res = PoisonValue::get(toVectorizedTy(StructTy, State.VF));
+    for (const auto &[LaneIndex, Op] : enumerate(operands())) {
+      for (unsigned FieldIndex = 0; FieldIndex != StructTy->getNumElements();
+           FieldIndex++) {
+        Value *ScalarValue =
+            Builder.CreateExtractValue(State.get(Op, true), FieldIndex);
+        Value *VectorValue = Builder.CreateExtractValue(Res, FieldIndex);
+        VectorValue =
+            Builder.CreateInsertElement(VectorValue, ScalarValue, LaneIndex);
+        Res = Builder.CreateInsertValue(Res, VectorValue, FieldIndex);
+      }
+    }
+    return Res;
+  }
+  case VPInstruction::BuildVector: {
+    auto *ScalarTy = State.TypeAnalysis.inferScalarType(getOperand(0));
+    auto NumOfElements = ElementCount::getFixed(getNumOperands());
+    Value *Res = PoisonValue::get(toVectorizedTy(ScalarTy, NumOfElements));
+    for (const auto &[Idx, Op] : enumerate(operands()))
+      Res = State.Builder.CreateInsertElement(Res, State.get(Op, true),
+                                              State.Builder.getInt32(Idx));
+    return Res;
+  }
   case VPInstruction::ReductionStartVector: {
     if (State.VF.isScalar())
       return State.get(getOperand(0), true);
@@ -953,10 +986,11 @@ void VPInstruction::execute(VPTransformState &State) {
   if (!hasResult())
     return;
   assert(GeneratedValue && "generate must produce a value");
-  assert(
-      (GeneratedValue->getType()->isVectorTy() == !GeneratesPerFirstLaneOnly ||
-       State.VF.isScalar()) &&
-      "scalar value but not only first lane defined");
+  assert((((GeneratedValue->getType()->isVectorTy() ||
+            GeneratedValue->getType()->isStructTy()) ==
+           !GeneratesPerFirstLaneOnly) ||
+          State.VF.isScalar()) &&
+         "scalar value but not only first lane defined");
   State.set(this, GeneratedValue,
             /*IsScalar*/ GeneratesPerFirstLaneOnly);
 }
@@ -970,6 +1004,8 @@ bool VPInstruction::opcodeMayReadOrWriteFromMemory() const {
   case Instruction::ICmp:
   case Instruction::Select:
   case VPInstruction::AnyOf:
+  case VPInstruction::BuildStructVector:
+  case VPInstruction::BuildVector:
   case VPInstruction::CalculateTripCountMinusVF:
   case VPInstruction::CanonicalIVIncrementForPart:
   case VPInstruction::ExtractLastElement:
@@ -1092,6 +1128,12 @@ void VPInstruction::print(raw_ostream &O, const Twine &Indent,
   case VPInstruction::Broadcast:
     O << "broadcast";
     break;
+  case VPInstruction::BuildStructVector:
+    O << "buildstructvector";
+    break;
+  case VPInstruction::BuildVector:
+    O << "buildvector";
+    break;
   case VPInstruction::ExtractLastElement:
     O << "extract-last-element";
     break;
@@ -2686,45 +2728,27 @@ static void scalarizeInstruction(const Instruction *Instr,
 
 void VPReplicateRecipe::execute(VPTransformState &State) {
   Instruction *UI = getUnderlyingInstr();
-  if (State.Lane) { // Generate a single instance.
-    assert((State.VF.isScalar() || !isSingleScalar()) &&
-           "uniform recipe shouldn't be predicated");
-    assert(!State.VF.isScalable() && "Can't scalarize a scalable vector");
-    scalarizeInstruction(UI, this, *State.Lane, State);
-    // Insert scalar instance packing it into a vector.
-    if (State.VF.isVector() && shouldPack()) {
-      Value *WideValue;
-      // If we're constructing lane 0, initialize to start from poison.
-      if (State.Lane->isFirstLane()) {
-        assert(!State.VF.isScalable() && "VF is assumed to be non scalable.");
-        WideValue = PoisonValue::get(VectorType::get(UI->getType(), State.VF));
-      } else {
-        WideValue = State.get(this);
-      }
-      State.set(this, State.packScalarIntoVectorizedValue(this, WideValue,
-                                                          *State.Lane));
-    }
-    return;
-  }
 
-  if (IsSingleScalar) {
-    // Uniform within VL means we need to generate lane 0.
+  if (!State.Lane) {
+    assert(IsSingleScalar && "VPReplicateRecipes outside replicate regions "
+                             "must have already been unrolled");
     scalarizeInstruction(UI, this, VPLane(0), State);
     return;
   }
 
-  // A store of a loop varying value to a uniform address only needs the last
-  // copy of the store.
-  if (isa<StoreInst>(UI) && vputils::isSingleScalar(getOperand(1))) {
-    auto Lane = VPLane::getLastLaneForVF(State.VF);
-    scalarizeInstruction(UI, this, VPLane(Lane), State);
-    return;
+  assert((State.VF.isScalar() || !isSingleScalar()) &&
+         "uniform recipe shouldn't be predicated");
+  assert(!State.VF.isScalable() && "Can't scalarize a scalable vector");
+  scalarizeInstruction(UI, this, *State.Lane, State);
+  // Insert scalar instance packing it into a vector.
+  if (State.VF.isVector() && shouldPack()) {
+    Value *WideValue =
+        State.Lane->isFirstLane()
+            ? PoisonValue::get(VectorType::get(UI->getType(), State.VF))
+            : State.get(this);
+    State.set(this, State.packScalarIntoVectorizedValue(this, WideValue,
+                                                        *State.Lane));
   }
-
-  // Generate scalar instances for all VF lanes.
-  const unsigned EndLane = State.VF.getFixedValue();
-  for (unsigned Lane = 0; Lane < EndLane; ++Lane)
-    scalarizeInstruction(UI, this, VPLane(Lane), State);
 }
 
 bool VPReplicateRecipe::shouldPack() const {

llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp

@@ -1140,6 +1140,24 @@ static void simplifyRecipe(VPRecipeBase &R, VPTypeAnalysis &TypeInfo) {
     return;
   }
 
+  // Look through ExtractLastElement (BuildVector ....).
+  if (match(&R, m_VPInstruction<VPInstruction::ExtractLastElement>(
+                    m_BuildVector()))) {
+    auto *BuildVector = cast<VPInstruction>(R.getOperand(0));
+    Def->replaceAllUsesWith(
+        BuildVector->getOperand(BuildVector->getNumOperands() - 1));
+    return;
+  }
+
+  // Look through ExtractPenultimateElement (BuildVector ....).
+  if (match(&R, m_VPInstruction<VPInstruction::ExtractPenultimateElement>(
+                    m_BuildVector()))) {
+    auto *BuildVector = cast<VPInstruction>(R.getOperand(0));
+    Def->replaceAllUsesWith(
+        BuildVector->getOperand(BuildVector->getNumOperands() - 2));
+    return;
+  }
+
   // Some simplifications can only be applied after unrolling. Perform them
   // below.
   if (!Plan->isUnrolled())

llvm/lib/Transforms/Vectorize/VPlanTransforms.h

@@ -99,6 +99,12 @@ struct VPlanTransforms {
   /// Explicitly unroll \p Plan by \p UF.
   static void unrollByUF(VPlan &Plan, unsigned UF, LLVMContext &Ctx);
 
+  /// Replace each VPReplicateRecipe outside on any replicate region in \p Plan
+  /// with \p VF single-scalar recipes.
+  /// TODO: Also replicate VPReplicateRecipes inside replicate regions, thereby
+  /// dissolving the latter.
+  static void replicateByVF(VPlan &Plan, ElementCount VF);
+
   /// Optimize \p Plan based on \p BestVF and \p BestUF. This may restrict the
   /// resulting plan to \p BestVF and \p BestUF.
   static void optimizeForVFAndUF(VPlan &Plan, ElementCount BestVF,

llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp

@@ -15,6 +15,7 @@
 #include "VPlan.h"
 #include "VPlanAnalysis.h"
 #include "VPlanCFG.h"
+#include "VPlanHelpers.h"
 #include "VPlanPatternMatch.h"
 #include "VPlanTransforms.h"
 #include "VPlanUtils.h"
@@ -450,3 +451,87 @@ void VPlanTransforms::unrollByUF(VPlan &Plan, unsigned UF, LLVMContext &Ctx) {
 
   VPlanTransforms::removeDeadRecipes(Plan);
 }
+
+/// Create a single-scalar clone of \p 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;
+    }
+    if (Lane.getKind() == VPLane::Kind::ScalableLast) {
+      NewOps.push_back(
+          Builder.createNaryOp(VPInstruction::ExtractLastElement, {Op}));
+      continue;
+    }
+    // Look through buildvector to avoid unnecessary extracts.
+    if (match(Op, m_BuildVector())) {
+      NewOps.push_back(
+          cast<VPInstruction>(Op)->getOperand(Lane.getKnownLane()));
+      continue;
+    }
+    VPValue *Idx =
+        Plan.getOrAddLiveIn(ConstantInt::get(IdxTy, Lane.getKnownLane()));
+    VPValue *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::replicateByVF(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);
+      if (RepR->getNumUsers() == 0) {
+        if (isa<StoreInst>(RepR->getUnderlyingInstr()) &&
+            vputils::isSingleScalar(RepR->getOperand(1))) {
+          // Stores to invariant addresses need to store the last lane only.
+          cloneForLane(Plan, Builder, IdxTy, RepR,
+                       VPLane::getLastLaneForVF(VF));
+        } else {
+          // Create single-scalar version of RepR for all lanes.
+          for (unsigned I = 0; I != VF.getKnownMinValue(); ++I)
+            cloneForLane(Plan, Builder, IdxTy, RepR, VPLane(I));
+        }
+        RepR->eraseFromParent();
+        continue;
+      }
+      /// Create single-scalar version of RepR for all lanes.
+      SmallVector<VPValue *> LaneDefs;
+      for (unsigned I = 0; I != VF.getKnownMinValue(); ++I)
+        LaneDefs.push_back(cloneForLane(Plan, Builder, IdxTy, RepR, VPLane(I)));
+
+      /// Users that only demand the first lane can use the definition for lane
+      /// 0.
+      RepR->replaceUsesWithIf(LaneDefs[0], [RepR](VPUser &U, unsigned) {
+        return U.onlyFirstLaneUsed(RepR);
+      });
+
+      // If needed, create a Build(Struct)Vector recipe to insert the scalar
+      // lane values into a vector.
+      Type *ResTy = RepR->getUnderlyingInstr()->getType();
+      VPValue *VecRes = Builder.createNaryOp(
+          ResTy->isStructTy() ? VPInstruction::BuildStructVector
+                              : VPInstruction::BuildVector,
+          LaneDefs);
+      RepR->replaceAllUsesWith(VecRes);
+      RepR->eraseFromParent();
+    }
+  }
+}

llvm/test/Transforms/LoopVectorize/X86/fixed-order-recurrence.ll

@@ -393,12 +393,6 @@ define void @test_for_tried_to_force_scalar(ptr noalias %A, ptr noalias %B, ptr
 ; CHECK-NEXT:    [[STRIDED_VEC:%.*]] = shufflevector <12 x float> [[WIDE_VEC]], <12 x float> poison, <4 x i32> <i32 0, i32 3, i32 6, i32 9>
 ; CHECK-NEXT:    [[TMP30:%.*]] = extractelement <4 x float> [[STRIDED_VEC]], i32 3
 ; CHECK-NEXT:    store float [[TMP30]], ptr [[C:%.*]], align 4
-; CHECK-NEXT:    [[TMP31:%.*]] = extractelement <4 x ptr> [[TMP29]], i32 0
-; CHECK-NEXT:    [[TMP38:%.*]] = load float, ptr [[TMP31]], align 4
-; CHECK-NEXT:    [[TMP33:%.*]] = extractelement <4 x ptr> [[TMP29]], i32 1
-; CHECK-NEXT:    [[TMP32:%.*]] = load float, ptr [[TMP33]], align 4
-; CHECK-NEXT:    [[TMP35:%.*]] = extractelement <4 x ptr> [[TMP29]], i32 2
-; CHECK-NEXT:    [[TMP34:%.*]] = load float, ptr [[TMP35]], align 4
 ; CHECK-NEXT:    [[TMP37:%.*]] = extractelement <4 x ptr> [[TMP29]], i32 3
 ; CHECK-NEXT:    [[TMP36:%.*]] = load float, ptr [[TMP37]], align 4
 ; CHECK-NEXT:    store float [[TMP36]], ptr [[B:%.*]], align 4