[LoopVectorize][AArch64][SVE] Generate wide active lane masks

This patch makes the LoopVectorize generate lane masks longer than
the VF to allow the target to better utilise the instruction set.
The vectorizer emit one or more wide `llvm.get.active.lane.mask.*`
calls plus several `llvm.vector.extract.*` calls to yield the required
number of VF-wide masks.

The motivating exammple is a vectorised loop with unroll factor 2 that
can use the SVE2.1 `whilelo` instruction with predicate pair result, or
a SVE `whilelo` instruction with smaller element size plus
`punpklo`/`punpkhi`.

How wide is the lane mask that the vectoriser emits is controlled
by a TargetTransformInfo hook `getMaxPredicateLength`.The default
impementation (return the same length as the VF) keeps the
change non-functional for targets that can't or are not prepared
to handle wider lane masks.

Author: momchil-velikov

llvm/include/llvm/Analysis/TargetTransformInfo.h

@@ -1244,6 +1244,8 @@ class TargetTransformInfo {
   /// and the number of execution units in the CPU.
   unsigned getMaxInterleaveFactor(ElementCount VF) const;
 
+  ElementCount getMaxPredicateLength(ElementCount VF) const;
+
   /// Collect properties of V used in cost analysis, e.g. OP_PowerOf2.
   static OperandValueInfo getOperandInfo(const Value *V);
 
@@ -2002,6 +2004,9 @@ class TargetTransformInfo::Concept {
   virtual bool shouldPrefetchAddressSpace(unsigned AS) const = 0;
 
   virtual unsigned getMaxInterleaveFactor(ElementCount VF) = 0;
+
+  virtual ElementCount getMaxPredicateLength(ElementCount VF) const = 0;
+
   virtual InstructionCost getArithmeticInstrCost(
       unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind,
       OperandValueInfo Opd1Info, OperandValueInfo Opd2Info,
@@ -2627,6 +2632,11 @@ class TargetTransformInfo::Model final : public TargetTransformInfo::Concept {
   unsigned getMaxInterleaveFactor(ElementCount VF) override {
     return Impl.getMaxInterleaveFactor(VF);
   }
+
+  ElementCount getMaxPredicateLength(ElementCount VF) const override {
+    return Impl.getMaxPredicateLength(VF);
+  }
+
   unsigned getEstimatedNumberOfCaseClusters(const SwitchInst &SI,
                                             unsigned &JTSize,
                                             ProfileSummaryInfo *PSI,

llvm/include/llvm/Analysis/TargetTransformInfoImpl.h

@@ -537,6 +537,8 @@ class TargetTransformInfoImplBase {
 
   unsigned getMaxInterleaveFactor(ElementCount VF) const { return 1; }
 
+  ElementCount getMaxPredicateLength(ElementCount VF) const { return VF; }
+
   InstructionCost getArithmeticInstrCost(
       unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind,
       TTI::OperandValueInfo Opd1Info, TTI::OperandValueInfo Opd2Info,

llvm/include/llvm/CodeGen/BasicTTIImpl.h

@@ -890,6 +890,8 @@ class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {
 
   unsigned getMaxInterleaveFactor(ElementCount VF) { return 1; }
 
+  ElementCount getMaxPredicateLength(ElementCount VF) const { return VF; }
+
   InstructionCost getArithmeticInstrCost(
       unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind,
       TTI::OperandValueInfo Opd1Info = {TTI::OK_AnyValue, TTI::OP_None},

llvm/lib/Analysis/TargetTransformInfo.cpp

@@ -821,6 +821,10 @@ unsigned TargetTransformInfo::getMaxInterleaveFactor(ElementCount VF) const {
   return TTIImpl->getMaxInterleaveFactor(VF);
 }
 
+ElementCount TargetTransformInfo::getMaxPredicateLength(ElementCount VF) const {
+  return TTIImpl->getMaxPredicateLength(VF);
+}
+
 TargetTransformInfo::OperandValueInfo
 TargetTransformInfo::getOperandInfo(const Value *V) {
   OperandValueKind OpInfo = OK_AnyValue;

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp

@@ -3383,6 +3383,15 @@ unsigned AArch64TTIImpl::getMaxInterleaveFactor(ElementCount VF) {
   return ST->getMaxInterleaveFactor();
 }
 
+ElementCount AArch64TTIImpl::getMaxPredicateLength(ElementCount VF) const {
+  // Do not create masks bigger than `<vscale x 16 x i1>`.
+  unsigned N = ST->hasSVE() ? 16 : 0;
+  // Do not create masks that are more than twice the VF.
+  N = std::min(N, 2 * VF.getKnownMinValue());
+  return VF.isScalable() ? ElementCount::getScalable(N)
+                         : ElementCount::getFixed(N);
+}
+
 // For Falkor, we want to avoid having too many strided loads in a loop since
 // that can exhaust the HW prefetcher resources.  We adjust the unroller
 // MaxCount preference below to attempt to ensure unrolling doesn't create too

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.h

@@ -157,6 +157,8 @@ class AArch64TTIImpl : public BasicTTIImplBase<AArch64TTIImpl> {
 
   unsigned getMaxInterleaveFactor(ElementCount VF);
 
+  ElementCount getMaxPredicateLength(ElementCount VF) const;
+
   bool prefersVectorizedAddressing() const;
 
   InstructionCost getMaskedMemoryOpCost(unsigned Opcode, Type *Src,

llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h

@@ -202,6 +202,14 @@ class VPBuilder {
   VPValue *createICmp(CmpInst::Predicate Pred, VPValue *A, VPValue *B,
                       DebugLoc DL = {}, const Twine &Name = "");
 
+  VPValue *createGetActiveLaneMask(VPValue *IV, VPValue *TC, DebugLoc DL,
+                                   const Twine &Name = "") {
+    auto *ALM = new VPActiveLaneMaskRecipe(IV, TC, DL, Name);
+    if (BB)
+      BB->insert(ALM, InsertPt);
+    return ALM;
+  }
+
   //===--------------------------------------------------------------------===//
   // RAII helpers.
   //===--------------------------------------------------------------------===//

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -588,6 +588,10 @@ class InnerLoopVectorizer {
   /// count of the original loop for both main loop and epilogue vectorization.
   void setTripCount(Value *TC) { TripCount = TC; }
 
+  ElementCount getMaxPredicateLength(ElementCount VF) const {
+    return TTI->getMaxPredicateLength(VF);
+  }
+
 protected:
   friend class LoopVectorizationPlanner;
 
@@ -7509,7 +7513,8 @@ LoopVectorizationPlanner::executePlan(
   LLVM_DEBUG(BestVPlan.dump());
 
   // Perform the actual loop transformation.
-  VPTransformState State(BestVF, BestUF, LI, DT, ILV.Builder, &ILV, &BestVPlan,
+  VPTransformState State(BestVF, BestUF, TTI.getMaxPredicateLength(BestVF), LI,
+                         DT, ILV.Builder, &ILV, &BestVPlan,
                          OrigLoop->getHeader()->getContext());
 
   // 0. Generate SCEV-dependent code into the preheader, including TripCount,

llvm/lib/Transforms/Vectorize/VPlan.cpp

@@ -215,12 +215,13 @@ VPBasicBlock::iterator VPBasicBlock::getFirstNonPhi() {
   return It;
 }
 
-VPTransformState::VPTransformState(ElementCount VF, unsigned UF, LoopInfo *LI,
+VPTransformState::VPTransformState(ElementCount VF, unsigned UF,
+                                   ElementCount MaxPred, LoopInfo *LI,
                                    DominatorTree *DT, IRBuilderBase &Builder,
                                    InnerLoopVectorizer *ILV, VPlan *Plan,
                                    LLVMContext &Ctx)
-    : VF(VF), UF(UF), CFG(DT), LI(LI), Builder(Builder), ILV(ILV), Plan(Plan),
-      LVer(nullptr),
+    : VF(VF), UF(UF), MaxPred(MaxPred), CFG(DT), LI(LI), Builder(Builder),
+      ILV(ILV), Plan(Plan), LVer(nullptr),
       TypeAnalysis(Plan->getCanonicalIV()->getScalarType(), Ctx) {}
 
 Value *VPTransformState::get(VPValue *Def, const VPIteration &Instance) {

llvm/lib/Transforms/Vectorize/VPlan.h

@@ -235,13 +235,14 @@ struct VPIteration {
 /// VPTransformState holds information passed down when "executing" a VPlan,
 /// needed for generating the output IR.
 struct VPTransformState {
-  VPTransformState(ElementCount VF, unsigned UF, LoopInfo *LI,
-                   DominatorTree *DT, IRBuilderBase &Builder,
+  VPTransformState(ElementCount VF, unsigned UF, ElementCount MaxPred,
+                   LoopInfo *LI, DominatorTree *DT, IRBuilderBase &Builder,
                    InnerLoopVectorizer *ILV, VPlan *Plan, LLVMContext &Ctx);
 
   /// The chosen Vectorization and Unroll Factors of the loop being vectorized.
   ElementCount VF;
   unsigned UF;
+  ElementCount MaxPred;
 
   /// Hold the indices to generate specific scalar instructions. Null indicates
   /// that all instances are to be generated, using either scalar or vector
@@ -1174,7 +1175,6 @@ class VPInstruction : public VPRecipeWithIRFlags {
     Not,
     SLPLoad,
     SLPStore,
-    ActiveLaneMask,
     ExplicitVectorLength,
     CalculateTripCountMinusVF,
     // Increment the canonical IV separately for each unrolled part.
@@ -1329,6 +1329,50 @@ class VPInstruction : public VPRecipeWithIRFlags {
   }
 };
 
+class VPActiveLaneMaskRecipe : public VPRecipeWithIRFlags {
+  const std::string Name;
+
+public:
+  VPActiveLaneMaskRecipe(VPValue *IV, VPValue *TC, DebugLoc DL = {},
+                         const Twine &Name = "")
+      : VPRecipeWithIRFlags(VPDef::VPActiveLaneMaskSC,
+                            std::initializer_list<VPValue *>{IV, TC}, DL),
+        Name(Name.str()) {}
+
+  VP_CLASSOF_IMPL(VPDef::VPActiveLaneMaskSC)
+
+  VPActiveLaneMaskRecipe *clone() override {
+    SmallVector<VPValue *, 2> Operands(operands());
+    assert(Operands.size() == 2 && "by construction");
+    auto *New = new VPActiveLaneMaskRecipe(Operands[0], Operands[1],
+                                           getDebugLoc(), Name);
+    New->transferFlags(*this);
+    return New;
+  }
+
+  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
+
+  bool onlyFirstLaneUsed(const VPValue *Op) const override {
+    assert(is_contained(operands(), Op) &&
+           "Op must be an operand of the recipe");
+
+    return true;
+  }
+
+  bool onlyFirstPartUsed(const VPValue *Op) const override {
+    assert(is_contained(operands(), Op) &&
+           "Op must be an operand of the recipe");
+
+    return false;
+  }
+};
+
 /// VPWidenRecipe is a recipe for producing a copy of vector type its
 /// ingredient. This recipe covers most of the traditional vectorization cases
 /// where each ingredient transforms into a vectorized version of itself.