[LV] Record memory widening decisions (NFCI)

[gilr]

Summary

Record widening decisions for memory operations within the planned recipes, and use the recorded decisions in code-gen rather than querying the cost model.

[fhahn]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

8781

nit: no () required.

llvm/lib/Transforms/Vectorize/VPlan.h

1517

Should we document both fields?

[Ayal]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

2920

Update above comment? A gather/scatter will be created iff the stride is not consecutive, going either forward or reverse. This includes non-unit strided accesses, which may be part of interleave groups, and non strided accesses.

2927

Above comment, assert, and artificial use of ConsecutiveStride are now redundant; cleanup?

llvm/lib/Transforms/Vectorize/VPlan.h

1534

assert that Reverse implies Consecutive?

Another alternative may be to designate one recipe for Gather/Scatter with arbitrary non-unit strides, and a separate possibly derived recipe for Consecutive/VectorMemoryInstructions, with a Reverse indicator?

1543

ditto

1576

are consecutive and in reverse order

[gilr]

Addressed comments.

[fhahn]

LGTM, thanks

llvm/lib/Transforms/Vectorize/VPlan.h

1517

nit: /// for doc-comment

1520

might be good to mention that this also implies Consecutive

1573

nit: /// for doc-comment

[gilr]

Rebased.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

void vectorizeInterleaveGroup(const InterleaveGroup<Instruction> *Group,
                                 VPValue *BlockInMask = nullptr);
 
   /// Vectorize Load and Store instructions with the base address given in \p
   /// Addr, optionally masking the vector operations if \p BlockInMask is
   /// non-null. Use \p State to translate given VPValues to IR values in the
   /// vectorized loop.
   void vectorizeMemoryInstruction(Instruction *Instr, VPTransformState &State,
                                   VPValue *Def, VPValue *Addr,
-                                  VPValue *StoredValue, VPValue *BlockInMask);
+                                  VPValue *StoredValue, VPValue *BlockInMask,
+                                  bool ConsecutiveStride, bool Reverse);
 
   /// Set the debug location in the builder \p Ptr using the debug location in
   /// \p V. If \p Ptr is None then it uses the class member's Builder.
   void setDebugLocFromInst(const Value *V,
                            Optional<IRBuilder<> *> CustomBuilder = None);
 
   /// Fix the non-induction PHIs in the OrigPHIsToFix vector.
   void fixNonInductionPHIs(VPTransformState &State);
if (BlockInMask || MaskForGaps) {
           Builder.CreateAlignedStore(IVec, AddrParts[Part], Group->getAlign());
 
     Group->addMetadata(NewStoreInstr);
   }
 }
 
 void InnerLoopVectorizer::vectorizeMemoryInstruction(
     Instruction *Instr, VPTransformState &State, VPValue *Def, VPValue *Addr,
-    VPValue *StoredValue, VPValue *BlockInMask) {
+    VPValue *StoredValue, VPValue *BlockInMask, bool ConsecutiveStride,
+    bool Reverse) {
   // Attempt to issue a wide load.
   LoadInst *LI = dyn_cast<LoadInst>(Instr);
   StoreInst *SI = dyn_cast<StoreInst>(Instr);
 
   assert((LI || SI) && "Invalid Load/Store instruction");
   assert((!SI || StoredValue) && "No stored value provided for widened store");
   assert((!LI || !StoredValue) && "Stored value provided for widened load");
 
-  LoopVectorizationCostModel::InstWidening Decision =
-      Cost->getWideningDecision(Instr, VF);
-  assert((Decision == LoopVectorizationCostModel::CM_Widen ||
-          Decision == LoopVectorizationCostModel::CM_Widen_Reverse ||
-          Decision == LoopVectorizationCostModel::CM_GatherScatter) &&
-         "CM decision is not to widen the memory instruction");
-
   Type *ScalarDataTy = getLoadStoreType(Instr);
 
   auto *DataTy = VectorType::get(ScalarDataTy, VF);
   const Align Alignment = getLoadStoreAlignment(Instr);
 
   // Determine if the pointer operand of the access is either consecutive or
   // reverse consecutive.
-  bool Reverse = (Decision == LoopVectorizationCostModel::CM_Widen_Reverse);
-  bool ConsecutiveStride =
-      Reverse || (Decision == LoopVectorizationCostModel::CM_Widen);
-  bool CreateGatherScatter =
-      (Decision == LoopVectorizationCostModel::CM_GatherScatter);
+  bool CreateGatherScatter = !ConsecutiveStride;
 
   // Either Ptr feeds a vector load/store, or a vector GEP should feed a vector
   // gather/scatter. Otherwise Decision should have been to Scalarize.
   assert((ConsecutiveStride || CreateGatherScatter) &&
          "The instruction should be scalarized");
   (void)ConsecutiveStride;
 
   VectorParts BlockInMaskParts(UF);
   bool isMaskRequired = BlockInMask;
   if (isMaskRequired)
     for (unsigned Part = 0; Part < UF; ++Part)
       BlockInMaskParts[Part] = State.get(BlockInMask, Part);
 
   const auto CreateVecPtr = [&](unsigned Part, Value *Ptr) -> Value * {
VPRecipeBase *VPRecipeBuilder::tryToWidenMemory(Instruction *I,
 
   if (!LoopVectorizationPlanner::getDecisionAndClampRange(willWiden, Range))
     return nullptr;
 
   VPValue *Mask = nullptr;
   if (Legal->isMaskRequired(I))
     Mask = createBlockInMask(I->getParent(), Plan);
 
+  // Determine if the pointer operand of the access is either consecutive or
+  // reverse consecutive.
+  LoopVectorizationCostModel::InstWidening Decision =
+      CM.getWideningDecision(I, Range.Start);
+  bool Reverse = (Decision == LoopVectorizationCostModel::CM_Widen_Reverse);
+  bool Consecutive =
+      Reverse || (Decision == LoopVectorizationCostModel::CM_Widen);
+
   if (LoadInst *Load = dyn_cast<LoadInst>(I))
-    return new VPWidenMemoryInstructionRecipe(*Load, Operands[0], Mask);
+    return new VPWidenMemoryInstructionRecipe(*Load, Operands[0], Mask,
+                                              Consecutive, Reverse);
 
   StoreInst *Store = cast<StoreInst>(I);
   return new VPWidenMemoryInstructionRecipe(*Store, Operands[1], Operands[0],
-                                            Mask);
+                                            Mask, Consecutive, Reverse);
 }
 
 VPWidenIntOrFpInductionRecipe *
 VPRecipeBuilder::tryToOptimizeInductionPHI(PHINode *Phi,
                                            ArrayRef<VPValue *> Operands) const {
   // Check if this is an integer or fp induction. If so, build the recipe that
   // produces its scalar and vector values.
   InductionDescriptor II = Legal->getInductionVars().lookup(Phi);
if (State.hasVectorValue(getOperand(0), Part)) {
     State.reset(getOperand(0), Phi, *State.Instance);
   }
 }
 
 void VPWidenMemoryInstructionRecipe::execute(VPTransformState &State) {
   VPValue *StoredValue = isStore() ? getStoredValue() : nullptr;
   State.ILV->vectorizeMemoryInstruction(
       &Ingredient, State, StoredValue ? nullptr : getVPSingleValue(), getAddr(),
-      StoredValue, getMask());
+      StoredValue, getMask(), Consecutive, Reverse);
 }
 
 // Determine how to lower the scalar epilogue, which depends on 1) optimising
 // for minimum code-size, 2) predicate compiler options, 3) loop hints forcing
 // predication, and 4) a TTI hook that analyses whether the loop is suitable
 // for predication.
 static ScalarEpilogueLowering getScalarEpilogueLowering(
     Function *F, Loop *L, LoopVectorizeHints &Hints, ProfileSummaryInfo *PSI,

llvm/lib/Transforms/Vectorize/VPlan.h

 /// The recipe uses the following VPValues:
 /// - For load: Address, optional mask
 /// - For store: Address, stored value, optional mask
 /// TODO: We currently execute only per-part unless a specific instance is
 /// provided.
 class VPWidenMemoryInstructionRecipe : public VPRecipeBase {
   Instruction &Ingredient;
 
+  bool Consecutive;
+  bool Reverse;
+
   void setMask(VPValue *Mask) {
     if (!Mask)
       return;
     addOperand(Mask);
   }
 
   bool isMasked() const {
     return isStore() ? getNumOperands() == 3 : getNumOperands() == 2;
   }
 
 public:
-  VPWidenMemoryInstructionRecipe(LoadInst &Load, VPValue *Addr, VPValue *Mask)
-      : VPRecipeBase(VPWidenMemoryInstructionSC, {Addr}), Ingredient(Load) {
+  VPWidenMemoryInstructionRecipe(LoadInst &Load, VPValue *Addr, VPValue *Mask,
+                                 bool Consecutive, bool Reverse)
+      : VPRecipeBase(VPWidenMemoryInstructionSC, {Addr}), Ingredient(Load),
+        Consecutive(Consecutive), Reverse(Reverse) {
     new VPValue(VPValue::VPVMemoryInstructionSC, &Load, this);
     setMask(Mask);
   }
 
   VPWidenMemoryInstructionRecipe(StoreInst &Store, VPValue *Addr,
-                                 VPValue *StoredValue, VPValue *Mask)
+                                 VPValue *StoredValue, VPValue *Mask,
+                                 bool Consecutive, bool Reverse)
       : VPRecipeBase(VPWidenMemoryInstructionSC, {Addr, StoredValue}),
-        Ingredient(Store) {
+        Ingredient(Store), Consecutive(Consecutive), Reverse(Reverse) {
     setMask(Mask);
   }
 
   /// Method to support type inquiry through isa, cast, and dyn_cast.
   static inline bool classof(const VPDef *D) {
     return D->getVPDefID() == VPRecipeBase::VPWidenMemoryInstructionSC;
   }
 
public:
   bool isStore() const { return isa<StoreInst>(Ingredient); }
 
   /// Return the address accessed by this recipe.
   VPValue *getStoredValue() const {
     assert(isStore() && "Stored value only available for store instructions");
     return getOperand(1); // Stored value is the 2nd, mandatory operand.
   }
 
+  // Return whether the loaded-from / stored-to addresses are consecutive.
+  bool isConsecutive() const { return Consecutive; }
+
+  // Return whether the loaded/stored addresses are in reverse order.
+  bool isReverse() const { return Reverse; }
+
   /// Generate the wide load/store.
   void execute(VPTransformState &State) override;
 
 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
   /// Print the recipe.
   void print(raw_ostream &O, const Twine &Indent,
              VPSlotTracker &SlotTracker) const override;
 #endif

llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp

for (VPRecipeBase &Ingredient : llvm::make_early_inc_range(*VPBB)) {
       } else {
         assert(isa<VPInstruction>(&Ingredient) &&
                "only VPInstructions expected here");
         assert(!isa<PHINode>(Inst) && "phis should be handled above");
         // Create VPWidenMemoryInstructionRecipe for loads and stores.
         if (LoadInst *Load = dyn_cast<LoadInst>(Inst)) {
           NewRecipe = new VPWidenMemoryInstructionRecipe(
               *Load, Plan->getOrAddVPValue(getLoadStorePointerOperand(Inst)),
-              nullptr /*Mask*/);
+              nullptr /*Mask*/, false /*Consecutive*/, false /*Reverse*/);
         } else if (StoreInst *Store = dyn_cast<StoreInst>(Inst)) {
           NewRecipe = new VPWidenMemoryInstructionRecipe(
               *Store, Plan->getOrAddVPValue(getLoadStorePointerOperand(Inst)),
-              Plan->getOrAddVPValue(Store->getValueOperand()),
-              nullptr /*Mask*/);
+              Plan->getOrAddVPValue(Store->getValueOperand()), nullptr /*Mask*/,
+              false /*Consecutive*/, false /*Reverse*/);
         } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Inst)) {
           NewRecipe = new VPWidenGEPRecipe(
               GEP, Plan->mapToVPValues(GEP->operands()), OrigLoop);
         } else if (CallInst *CI = dyn_cast<CallInst>(Inst)) {
           NewRecipe = new VPWidenCallRecipe(
               *CI, Plan->mapToVPValues(CI->arg_operands()));
         } else if (SelectInst *SI = dyn_cast<SelectInst>(Inst)) {
           bool InvariantCond =

llvm/unittests/Transforms/Vectorize/VPlanTest.cpp

TEST(VPRecipeTest, CastVPWidenMemoryInstructionRecipeToVPUserAndVPDef) {
   LLVMContext C;
 
   IntegerType *Int32 = IntegerType::get(C, 32);
   PointerType *Int32Ptr = PointerType::get(Int32, 0);
   auto *Load =
       new LoadInst(Int32, UndefValue::get(Int32Ptr), "", false, Align(1));
   VPValue Addr;
   VPValue Mask;
-  VPWidenMemoryInstructionRecipe Recipe(*Load, &Addr, &Mask);
+  VPWidenMemoryInstructionRecipe Recipe(*Load, &Addr, &Mask, true, false);
   EXPECT_TRUE(isa<VPUser>(&Recipe));
   VPRecipeBase *BaseR = &Recipe;
   EXPECT_TRUE(isa<VPUser>(BaseR));
   EXPECT_EQ(&Recipe, BaseR);
 
   VPValue *VPV = Recipe.getVPSingleValue();
   EXPECT_TRUE(isa<VPRecipeBase>(VPV->getDef()));
   EXPECT_EQ(&Recipe, dyn_cast<VPRecipeBase>(VPV->getDef()));
PointerType *Int32Ptr = PointerType::get(Int32, 0);
     EXPECT_FALSE(Recipe.mayReadOrWriteMemory());
   }
 
   {
     auto *Load =
         new LoadInst(Int32, UndefValue::get(Int32Ptr), "", false, Align(1));
     VPValue Addr;
     VPValue Mask;
-    VPWidenMemoryInstructionRecipe Recipe(*Load, &Addr, &Mask);
+    VPWidenMemoryInstructionRecipe Recipe(*Load, &Addr, &Mask, true, false);
     EXPECT_TRUE(Recipe.mayHaveSideEffects());
     EXPECT_TRUE(Recipe.mayReadFromMemory());
     EXPECT_FALSE(Recipe.mayWriteToMemory());
     EXPECT_TRUE(Recipe.mayReadOrWriteMemory());
     delete Load;
   }
 
   {
     auto *Store = new StoreInst(UndefValue::get(Int32),
                                 UndefValue::get(Int32Ptr), false, Align(1));
     VPValue Addr;
     VPValue Mask;
     VPValue StoredV;
-    VPWidenMemoryInstructionRecipe Recipe(*Store, &Addr, &StoredV, &Mask);
+    VPWidenMemoryInstructionRecipe Recipe(*Store, &Addr, &StoredV, &Mask, false,
+                                          false);
     EXPECT_TRUE(Recipe.mayHaveSideEffects());
     EXPECT_FALSE(Recipe.mayReadFromMemory());
     EXPECT_TRUE(Recipe.mayWriteToMemory());
     EXPECT_TRUE(Recipe.mayReadOrWriteMemory());
     delete Store;
   }
 
   {