[VPlan] Explicitly replicate VPInstructions by VF.

fhahn · fhahn · commit b80eee3a7fea · 2025-08-31T13:33:12.000+01:00
Extend replicateByVF added in #142433 (aa24029) to also explicitly unroll replicating VPInstructions. Now the only remaining case where we replicate for all lanes is VPReplicateRecipes in replicate regions.
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.cpp b/llvm/lib/Transforms/Vectorize/VPlan.cpp
@@ -333,6 +333,9 @@ Value *VPTransformState::get(const VPValue *Def, bool NeedsScalar) {
     LastLane = 0;
   }
 
+  assert(IsSingleScalar && "must be a single-scalar at this point");
+  // We need to construct the vector value for a single-scalar value by
+  // broadcasting the scalar to all lanes.
   auto *LastInst = cast<Instruction>(get(Def, LastLane));
   // Set the insert point after the last scalarized instruction or after the
   // last PHI, if LastInst is a PHI. This ensures the insertelement sequence
@@ -343,27 +346,8 @@ Value *VPTransformState::get(const VPValue *Def, bool NeedsScalar) {
                    : std::next(BasicBlock::iterator(LastInst));
   Builder.SetInsertPoint(&*NewIP);
 
-  // However, if we are vectorizing, we need to construct the vector values.
-  // If the value is known to be uniform after vectorization, we can just
-  // broadcast the scalar value corresponding to lane zero. Otherwise, we
-  // construct the vector values using insertelement instructions. Since the
-  // resulting vectors are stored in State, we will only generate the
-  // insertelements once.
-  Value *VectorValue = nullptr;
-  if (IsSingleScalar) {
-    VectorValue = GetBroadcastInstrs(ScalarValue);
-    set(Def, VectorValue);
-  } else {
-    assert(!VF.isScalable() && "VF is assumed to be non scalable.");
-    assert(isa<VPInstruction>(Def) &&
-           "Explicit BuildVector recipes must have"
-           "handled packing for non-VPInstructions.");
-    // Initialize packing with insertelements to start from poison.
-    VectorValue = PoisonValue::get(toVectorizedTy(LastInst->getType(), VF));
-    for (unsigned Lane = 0; Lane < VF.getFixedValue(); ++Lane)
-      VectorValue = packScalarIntoVectorizedValue(Def, VectorValue, Lane);
-    set(Def, VectorValue);
-  }
+  Value *VectorValue = GetBroadcastInstrs(ScalarValue);
+  set(Def, VectorValue);
   Builder.restoreIP(OldIP);
   return VectorValue;
 }
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.h b/llvm/lib/Transforms/Vectorize/VPlan.h
@@ -901,6 +901,8 @@ struct VPRecipeWithIRFlags : public VPSingleDefRecipe, public VPIRFlags {
     return R && classof(R);
   }
 
+  virtual VPRecipeWithIRFlags *clone() override = 0;
+
   void execute(VPTransformState &State) override = 0;
 
   /// Compute the cost for this recipe for \p VF, using \p Opcode and \p Ctx.
@@ -1045,13 +1047,6 @@ class LLVM_ABI_FOR_TEST VPInstruction : public VPRecipeWithIRFlags,
     VScale,
   };
 
-private:
-  typedef unsigned char OpcodeTy;
-  OpcodeTy Opcode;
-
-  /// An optional name that can be used for the generated IR instruction.
-  const std::string Name;
-
   /// Returns true if this VPInstruction generates scalar values for all lanes.
   /// Most VPInstructions generate a single value per part, either vector or
   /// scalar. VPReplicateRecipe takes care of generating multiple (scalar)
@@ -1060,6 +1055,13 @@ class LLVM_ABI_FOR_TEST VPInstruction : public VPRecipeWithIRFlags,
   /// underlying ingredient.
   bool doesGeneratePerAllLanes() const;
 
+private:
+  typedef unsigned char OpcodeTy;
+  OpcodeTy Opcode;
+
+  /// An optional name that can be used for the generated IR instruction.
+  const std::string Name;
+
   /// Returns true if we can generate a scalar for the first lane only if
   /// needed.
   bool canGenerateScalarForFirstLane() const;
@@ -1069,11 +1071,6 @@ class LLVM_ABI_FOR_TEST VPInstruction : public VPRecipeWithIRFlags,
   /// existing value is returned rather than a generated one.
   Value *generate(VPTransformState &State);
 
-  /// Utility methods serving execute(): generates a scalar single instance of
-  /// the modeled instruction for a given lane. \returns the scalar generated
-  /// value for lane \p Lane.
-  Value *generatePerLane(VPTransformState &State, const VPLane &Lane);
-
 #if !defined(NDEBUG)
   /// Return the number of operands determined by the opcode of the
   /// VPInstruction. Returns -1u if the number of operands cannot be determined
diff --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -525,16 +525,6 @@ bool VPInstruction::canGenerateScalarForFirstLane() const {
   }
 }
 
-Value *VPInstruction::generatePerLane(VPTransformState &State,
-                                      const VPLane &Lane) {
-  IRBuilderBase &Builder = State.Builder;
-
-  assert(getOpcode() == VPInstruction::PtrAdd &&
-         "only PtrAdd opcodes are supported for now");
-  return Builder.CreatePtrAdd(State.get(getOperand(0), Lane),
-                              State.get(getOperand(1), Lane), Name);
-}
-
 /// Create a conditional branch using \p Cond branching to the successors of \p
 /// VPBB. Note that the first successor is always forward (i.e. not created yet)
 /// while the second successor may already have been created (if it is a header
@@ -1154,24 +1144,13 @@ void VPInstruction::execute(VPTransformState &State) {
          "Set flags not supported for the provided opcode");
   if (hasFastMathFlags())
     State.Builder.setFastMathFlags(getFastMathFlags());
-  bool GeneratesPerFirstLaneOnly = canGenerateScalarForFirstLane() &&
-                                   (vputils::onlyFirstLaneUsed(this) ||
-                                    isVectorToScalar() || isSingleScalar());
-  bool GeneratesPerAllLanes = doesGeneratePerAllLanes();
-  if (GeneratesPerAllLanes) {
-    for (unsigned Lane = 0, NumLanes = State.VF.getFixedValue();
-         Lane != NumLanes; ++Lane) {
-      Value *GeneratedValue = generatePerLane(State, VPLane(Lane));
-      assert(GeneratedValue && "generatePerLane must produce a value");
-      State.set(this, GeneratedValue, VPLane(Lane));
-    }
-    return;
-  }
-
   Value *GeneratedValue = generate(State);
   if (!hasResult())
     return;
   assert(GeneratedValue && "generate must produce a value");
+  bool GeneratesPerFirstLaneOnly = canGenerateScalarForFirstLane() &&
+                                   (vputils::onlyFirstLaneUsed(this) ||
+                                    isVectorToScalar() || isSingleScalar());
   assert((((GeneratedValue->getType()->isVectorTy() ||
             GeneratedValue->getType()->isStructTy()) ==
            !GeneratesPerFirstLaneOnly) ||
@@ -1244,6 +1223,9 @@ bool VPInstruction::onlyFirstLaneUsed(const VPValue *Op) const {
   case VPInstruction::Broadcast:
   case VPInstruction::ReductionStartVector:
     return true;
+  case VPInstruction::BuildStructVector:
+  case VPInstruction::BuildVector:
+    return getNumOperands() > 1;
   case VPInstruction::PtrAdd:
     return Op == getOperand(0) || vputils::onlyFirstLaneUsed(this);
   case VPInstruction::WidePtrAdd:
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -3377,34 +3377,40 @@ void VPlanTransforms::materializeBuildVectors(VPlan &Plan) {
       vp_depth_first_shallow(Plan.getEntry()));
   auto VPBBsInsideLoopRegion = VPBlockUtils::blocksOnly<VPBasicBlock>(
       vp_depth_first_shallow(LoopRegion->getEntry()));
-  // Materialize Build(Struct)Vector for all replicating VPReplicateRecipes,
-  // excluding ones in replicate regions. Those are not materialized explicitly
-  // yet. Those vector users are still handled in VPReplicateRegion::execute(),
-  // via shouldPack().
+  // Materialize Build(Struct)Vector for all replicating VPReplicateRecipes and
+  // VPInstructions, excluding ones in replicate regions. Those are not
+  // materialized explicitly yet. Those vector users are still handled in
+  // VPReplicateRegion::execute(), via shouldPack().
   // TODO: materialize build vectors for replicating recipes in replicating
   // regions.
   // TODO: materialize build vectors for VPInstructions.
   for (VPBasicBlock *VPBB :
        concat<VPBasicBlock *>(VPBBsOutsideLoopRegion, VPBBsInsideLoopRegion)) {
     for (VPRecipeBase &R : make_early_inc_range(*VPBB)) {
-      auto *RepR = dyn_cast<VPReplicateRecipe>(&R);
-      auto UsesVectorOrInsideReplicateRegion = [RepR, LoopRegion](VPUser *U) {
+      auto *DefR = dyn_cast<VPRecipeWithIRFlags>(&R);
+      if (!DefR || !isa<VPReplicateRecipe, VPInstruction>(DefR))
+        continue;
+      auto UsesVectorOrInsideReplicateRegion = [DefR, LoopRegion](VPUser *U) {
         VPRegionBlock *ParentRegion =
             cast<VPRecipeBase>(U)->getParent()->getParent();
-        return !U->usesScalars(RepR) || ParentRegion != LoopRegion;
+        return !U->usesScalars(DefR) || ParentRegion != LoopRegion;
       };
-      if (!RepR || RepR->isSingleScalar() ||
-          none_of(RepR->users(), UsesVectorOrInsideReplicateRegion))
+      if ((isa<VPReplicateRecipe>(DefR) &&
+           cast<VPReplicateRecipe>(DefR)->isSingleScalar()) ||
+          (isa<VPInstruction>(DefR) &&
+           !cast<VPInstruction>(DefR)->doesGeneratePerAllLanes()) ||
+          vputils::onlyFirstLaneUsed(DefR) ||
+          none_of(DefR->users(), UsesVectorOrInsideReplicateRegion))
         continue;
 
-      Type *ScalarTy = TypeInfo.inferScalarType(RepR);
+      Type *ScalarTy = TypeInfo.inferScalarType(DefR);
       unsigned Opcode = ScalarTy->isStructTy()
                             ? VPInstruction::BuildStructVector
                             : VPInstruction::BuildVector;
-      auto *BuildVector = new VPInstruction(Opcode, {RepR});
-      BuildVector->insertAfter(RepR);
+      auto *BuildVector = new VPInstruction(Opcode, {DefR});
+      BuildVector->insertAfter(DefR);
 
-      RepR->replaceUsesWithIf(
+      DefR->replaceUsesWithIf(
           BuildVector, [BuildVector, &UsesVectorOrInsideReplicateRegion](
                            VPUser &U, unsigned) {
             return &U != BuildVector && UsesVectorOrInsideReplicateRegion(&U);
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.h b/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
@@ -158,8 +158,8 @@ struct VPlanTransforms {
   /// Explicitly unroll \p Plan by \p UF.
   static void unrollByUF(VPlan &Plan, unsigned UF);
 
-  /// Replace each VPReplicateRecipe outside on any replicate region in \p Plan
-  /// with \p VF single-scalar recipes.
+  /// Replace each VPReplicateRecipe and replicating VPInstruction 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);
diff --git a/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp b/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp
@@ -463,15 +463,15 @@ void VPlanTransforms::unrollByUF(VPlan &Plan, unsigned UF) {
   VPlanTransforms::removeDeadRecipes(Plan);
 }
 
-/// Create a single-scalar clone of \p RepR for lane \p Lane. Use \p
-/// Def2LaneDefs to look up scalar definitions for operands of \RepR.
-static VPReplicateRecipe *
+/// Create a single-scalar clone of \p DefR for lane \p Lane. Use \p
+/// Def2LaneDefs to look up scalar definitions for operands of \DefR.
+static VPRecipeWithIRFlags *
 cloneForLane(VPlan &Plan, VPBuilder &Builder, Type *IdxTy,
-             VPReplicateRecipe *RepR, VPLane Lane,
+             VPRecipeWithIRFlags *DefR, VPLane Lane,
              const DenseMap<VPValue *, SmallVector<VPValue *>> &Def2LaneDefs) {
   // Collect the operands at Lane, creating extracts as needed.
   SmallVector<VPValue *> NewOps;
-  for (VPValue *Op : RepR->operands()) {
+  for (VPValue *Op : DefR->operands()) {
     // If Op is a definition that has been unrolled, directly use the clone for
     // the corresponding lane.
     auto LaneDefs = Def2LaneDefs.find(Op);
@@ -501,11 +501,19 @@ cloneForLane(VPlan &Plan, VPBuilder &Builder, Type *IdxTy,
     NewOps.push_back(Ext);
   }
 
-  auto *New =
-      new VPReplicateRecipe(RepR->getUnderlyingInstr(), NewOps,
-                            /*IsSingleScalar=*/true, /*Mask=*/nullptr, *RepR);
-  New->transferFlags(*RepR);
-  New->insertBefore(RepR);
+  VPRecipeWithIRFlags *New;
+  if (auto *RepR = dyn_cast<VPReplicateRecipe>(DefR)) {
+    New =
+        new VPReplicateRecipe(RepR->getUnderlyingInstr(), NewOps,
+                              /*IsSingleScalar=*/true, /*Mask=*/nullptr, *RepR);
+  } else {
+    New = DefR->clone();
+    for (const auto &[Idx, Op] : enumerate(NewOps)) {
+      New->setOperand(Idx, Op);
+    }
+  }
+  New->transferFlags(*DefR);
+  New->insertBefore(DefR);
   return New;
 }
 
@@ -530,41 +538,46 @@ void VPlanTransforms::replicateByVF(VPlan &Plan, ElementCount VF) {
   SmallVector<VPRecipeBase *> ToRemove;
   for (VPBasicBlock *VPBB : VPBBsToUnroll) {
     for (VPRecipeBase &R : make_early_inc_range(*VPBB)) {
-      auto *RepR = dyn_cast<VPReplicateRecipe>(&R);
-      if (!RepR || RepR->isSingleScalar())
+      auto *DefR = dyn_cast<VPRecipeWithIRFlags>(&R);
+      if (!DefR || !isa<VPInstruction, VPReplicateRecipe>(DefR))
+        continue;
+      if ((isa<VPReplicateRecipe>(DefR) &&
+           cast<VPReplicateRecipe>(DefR)->isSingleScalar()) ||
+          (isa<VPInstruction>(DefR) &&
+           !cast<VPInstruction>(DefR)->doesGeneratePerAllLanes()))
         continue;
 
-      VPBuilder Builder(RepR);
-      if (RepR->getNumUsers() == 0) {
-        if (isa<StoreInst>(RepR->getUnderlyingInstr()) &&
-            vputils::isSingleScalar(RepR->getOperand(1))) {
+      VPBuilder Builder(DefR);
+      if (DefR->getNumUsers() == 0) {
+        if (isa<StoreInst>(DefR->getUnderlyingInstr()) &&
+            vputils::isSingleScalar(DefR->getOperand(1))) {
           // Stores to invariant addresses need to store the last lane only.
-          cloneForLane(Plan, Builder, IdxTy, RepR, VPLane::getLastLaneForVF(VF),
+          cloneForLane(Plan, Builder, IdxTy, DefR, VPLane::getLastLaneForVF(VF),
                        Def2LaneDefs);
         } else {
-          // Create single-scalar version of RepR for all lanes.
+          // Create single-scalar version of DefR for all lanes.
           for (unsigned I = 0; I != VF.getKnownMinValue(); ++I)
-            cloneForLane(Plan, Builder, IdxTy, RepR, VPLane(I), Def2LaneDefs);
+            cloneForLane(Plan, Builder, IdxTy, DefR, VPLane(I), Def2LaneDefs);
         }
-        RepR->eraseFromParent();
+        DefR->eraseFromParent();
         continue;
       }
-      /// Create single-scalar version of RepR for all lanes.
+      /// Create single-scalar version of DefR for all lanes.
       SmallVector<VPValue *> LaneDefs;
       for (unsigned I = 0; I != VF.getKnownMinValue(); ++I)
         LaneDefs.push_back(
-            cloneForLane(Plan, Builder, IdxTy, RepR, VPLane(I), Def2LaneDefs));
+            cloneForLane(Plan, Builder, IdxTy, DefR, VPLane(I), Def2LaneDefs));
 
-      Def2LaneDefs[RepR] = LaneDefs;
+      Def2LaneDefs[DefR] = LaneDefs;
       /// 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);
+      DefR->replaceUsesWithIf(LaneDefs[0], [DefR](VPUser &U, unsigned) {
+        return U.onlyFirstLaneUsed(DefR);
       });
 
-      // Update each build vector user that currently has RepR as its only
+      // Update each build vector user that currently has DefR as its only
       // operand, to have all LaneDefs as its operands.
-      for (VPUser *U : to_vector(RepR->users())) {
+      for (VPUser *U : to_vector(DefR->users())) {
         auto *VPI = dyn_cast<VPInstruction>(U);
         if (!VPI || (VPI->getOpcode() != VPInstruction::BuildVector &&
                      VPI->getOpcode() != VPInstruction::BuildStructVector))
@@ -576,7 +589,7 @@ void VPlanTransforms::replicateByVF(VPlan &Plan, ElementCount VF) {
         for (VPValue *LaneDef : drop_begin(LaneDefs))
           VPI->addOperand(LaneDef);
       }
-      ToRemove.push_back(RepR);
+      ToRemove.push_back(DefR);
     }
   }
   for (auto *R : reverse(ToRemove))
diff --git a/llvm/test/Transforms/LoopVectorize/pointer-induction.ll b/llvm/test/Transforms/LoopVectorize/pointer-induction.ll
@@ -33,6 +33,10 @@ define void @a(ptr readnone %b) {
 ; CHECK-NEXT:    [[NEXT_GEP2:%.*]] = getelementptr i8, ptr null, i64 [[TMP11]]
 ; CHECK-NEXT:    [[NEXT_GEP3:%.*]] = getelementptr i8, ptr null, i64 [[TMP14]]
 ; CHECK-NEXT:    [[NEXT_GEP4:%.*]] = getelementptr i8, ptr null, i64 [[TMP17]]
+; CHECK-NEXT:    [[TMP21:%.*]] = insertelement <4 x ptr> poison, ptr [[NEXT_GEP]], i32 0
+; CHECK-NEXT:    [[TMP22:%.*]] = insertelement <4 x ptr> [[TMP21]], ptr [[NEXT_GEP2]], i32 1
+; CHECK-NEXT:    [[TMP23:%.*]] = insertelement <4 x ptr> [[TMP22]], ptr [[NEXT_GEP3]], i32 2
+; CHECK-NEXT:    [[TMP24:%.*]] = insertelement <4 x ptr> [[TMP23]], ptr [[NEXT_GEP4]], i32 3
 ; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i8, ptr [[NEXT_GEP]], i64 -1
 ; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i8, ptr [[TMP3]], i32 0
 ; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr inbounds i8, ptr [[TMP4]], i32 -3
@@ -649,9 +653,6 @@ define i64 @ivopt_widen_ptr_indvar_3(ptr noalias %a, i64 %stride, i64 %n) {
 ; STRIDED-NEXT:    [[TMP9:%.*]] = add i64 [[OFFSET_IDX]], [[TMP8]]
 ; STRIDED-NEXT:    [[TMP10:%.*]] = mul i64 3, [[TMP1]]
 ; STRIDED-NEXT:    [[TMP11:%.*]] = add i64 [[OFFSET_IDX]], [[TMP10]]
-; STRIDED-NEXT:    [[NEXT_GEP:%.*]] = getelementptr i8, ptr null, i64 [[TMP5]]
-; STRIDED-NEXT:    [[NEXT_GEP1:%.*]] = getelementptr i8, ptr null, i64 [[TMP7]]
-; STRIDED-NEXT:    [[NEXT_GEP2:%.*]] = getelementptr i8, ptr null, i64 [[TMP9]]
 ; STRIDED-NEXT:    [[NEXT_GEP3:%.*]] = getelementptr i8, ptr null, i64 [[TMP11]]
 ; STRIDED-NEXT:    [[TMP12:%.*]] = getelementptr i64, ptr [[A:%.*]], i64 [[INDEX]]
 ; STRIDED-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i64>, ptr [[TMP12]], align 8
diff --git a/llvm/test/Transforms/LoopVectorize/predicate-switch.ll b/llvm/test/Transforms/LoopVectorize/predicate-switch.ll
diff --git a/llvm/test/Transforms/LoopVectorize/vplan-predicate-switch.ll b/llvm/test/Transforms/LoopVectorize/vplan-predicate-switch.ll
