[LV] Vectorize FMax w/o fast-math flags.

Add a new recurrence kind for FMax reductions without fast-math flags
and a corresponding VPlan transform tries to vectorize without fast-math
flags.

To do so, a new FindFirstIV reduction is added that tracks the first
indices that contain the maximum values. This serves as tie breaker
if the partial reduction vector contains NaNs or signed zeros.
After the loop, the first index is used to retrieve the final max
value after vectorization from the vector containing the partial
maximum values

Author: fhahn

llvm/include/llvm/Analysis/IVDescriptors.h

@@ -47,6 +47,8 @@ enum class RecurKind {
   FMul,     ///< Product of floats.
   FMin,     ///< FP min implemented in terms of select(cmp()).
   FMax,     ///< FP max implemented in terms of select(cmp()).
+  FMaxNoFMFs, ///< FP max implemented in terms of select(cmp()), but without
+              ///any fast-math flags. Users need to handle NaNs and signed zeros when generating code.
   FMinimum, ///< FP min with llvm.minimum semantics
   FMaximum, ///< FP max with llvm.maximum semantics
   FMinimumNum, ///< FP min with llvm.minimumnum semantics
@@ -250,8 +252,9 @@ class RecurrenceDescriptor {
   /// Returns true if the recurrence kind is a floating-point min/max kind.
   static bool isFPMinMaxRecurrenceKind(RecurKind Kind) {
     return Kind == RecurKind::FMin || Kind == RecurKind::FMax ||
-           Kind == RecurKind::FMinimum || Kind == RecurKind::FMaximum ||
-           Kind == RecurKind::FMinimumNum || Kind == RecurKind::FMaximumNum;
+           Kind == RecurKind::FMaxNoFMFs || Kind == RecurKind::FMinimum ||
+           Kind == RecurKind::FMaximum || Kind == RecurKind::FMinimumNum ||
+           Kind == RecurKind::FMaximumNum;
   }
 
   /// Returns true if the recurrence kind is any min/max kind.

llvm/lib/Analysis/IVDescriptors.cpp

@@ -819,7 +819,8 @@ RecurrenceDescriptor::isMinMaxPattern(Instruction *I, RecurKind Kind,
   if (match(I, m_OrdOrUnordFMin(m_Value(), m_Value())))
     return InstDesc(Kind == RecurKind::FMin, I);
   if (match(I, m_OrdOrUnordFMax(m_Value(), m_Value())))
-    return InstDesc(Kind == RecurKind::FMax, I);
+    return InstDesc(Kind == RecurKind::FMax || Kind == RecurKind::FMaxNoFMFs,
+                    I);
   if (match(I, m_FMinNum(m_Value(), m_Value())))
     return InstDesc(Kind == RecurKind::FMin, I);
   if (match(I, m_FMaxNum(m_Value(), m_Value())))
@@ -941,10 +942,15 @@ RecurrenceDescriptor::InstDesc RecurrenceDescriptor::isRecurrenceInstr(
                   m_Intrinsic<Intrinsic::minimumnum>(m_Value(), m_Value())) ||
             match(I, m_Intrinsic<Intrinsic::maximumnum>(m_Value(), m_Value()));
     };
-    if (isIntMinMaxRecurrenceKind(Kind) ||
-        (HasRequiredFMF() && isFPMinMaxRecurrenceKind(Kind)))
+    if (isIntMinMaxRecurrenceKind(Kind))
       return isMinMaxPattern(I, Kind, Prev);
-    else if (isFMulAddIntrinsic(I))
+    if (isFPMinMaxRecurrenceKind(Kind)) {
+      if (HasRequiredFMF())
+        return isMinMaxPattern(I, Kind, Prev);
+      if ((Kind == RecurKind::FMax || Kind == RecurKind::FMaxNoFMFs) &&
+          isMinMaxPattern(I, Kind, Prev).isRecurrence())
+        return InstDesc(I, RecurKind::FMaxNoFMFs);
+    } else if (isFMulAddIntrinsic(I))
       return InstDesc(Kind == RecurKind::FMulAdd, I,
                       I->hasAllowReassoc() ? nullptr : I);
     return InstDesc(false, I);
@@ -1207,6 +1213,7 @@ unsigned RecurrenceDescriptor::getOpcode(RecurKind Kind) {
   case RecurKind::UMin:
     return Instruction::ICmp;
   case RecurKind::FMax:
+  case RecurKind::FMaxNoFMFs:
   case RecurKind::FMin:
   case RecurKind::FMaximum:
   case RecurKind::FMinimum:

llvm/lib/Transforms/Utils/LoopUtils.cpp

@@ -937,6 +937,7 @@ constexpr Intrinsic::ID llvm::getReductionIntrinsicID(RecurKind RK) {
     return Intrinsic::vector_reduce_umax;
   case RecurKind::UMin:
     return Intrinsic::vector_reduce_umin;
+  case RecurKind::FMaxNoFMFs:
   case RecurKind::FMax:
     return Intrinsic::vector_reduce_fmax;
   case RecurKind::FMin:
@@ -1085,6 +1086,7 @@ CmpInst::Predicate llvm::getMinMaxReductionPredicate(RecurKind RK) {
   case RecurKind::FMin:
     return CmpInst::FCMP_OLT;
   case RecurKind::FMax:
+  case RecurKind::FMaxNoFMFs:
     return CmpInst::FCMP_OGT;
   // We do not add FMinimum/FMaximum recurrence kind here since there is no
   // equivalent predicate which compares signed zeroes according to the
@@ -1307,6 +1309,7 @@ Value *llvm::createSimpleReduction(IRBuilderBase &Builder, Value *Src,
   case RecurKind::UMax:
   case RecurKind::UMin:
   case RecurKind::FMax:
+  case RecurKind::FMaxNoFMFs:
   case RecurKind::FMin:
   case RecurKind::FMinimum:
   case RecurKind::FMaximum:

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -4346,8 +4346,13 @@ bool LoopVectorizationPlanner::isCandidateForEpilogueVectorization(
     ElementCount VF) const {
   // Cross iteration phis such as reductions need special handling and are
   // currently unsupported.
-  if (any_of(OrigLoop->getHeader()->phis(),
-             [&](PHINode &Phi) { return Legal->isFixedOrderRecurrence(&Phi); }))
+  if (any_of(OrigLoop->getHeader()->phis(), [&](PHINode &Phi) {
+        return Legal->isFixedOrderRecurrence(&Phi) ||
+               (Legal->isReductionVariable(&Phi) &&
+                Legal->getReductionVars()
+                        .find(&Phi)
+                        ->second.getRecurrenceKind() == RecurKind::FMaxNoFMFs);
+      }))
     return false;
 
   // Phis with uses outside of the loop require special handling and are
@@ -8808,6 +8813,9 @@ VPlanPtr LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(
 
   // Adjust the recipes for any inloop reductions.
   adjustRecipesForReductions(Plan, RecipeBuilder, Range.Start);
+  if (!VPlanTransforms::runPass(
+          VPlanTransforms::handleFMaxReductionsWithoutFastMath, *Plan))
+    return nullptr;
 
   // Transform recipes to abstract recipes if it is legal and beneficial and
   // clamp the range for better cost estimation.

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

@@ -23375,6 +23375,7 @@ class HorizontalReduction {
     case RecurKind::FindFirstIVUMin:
     case RecurKind::FindLastIVSMax:
     case RecurKind::FindLastIVUMax:
+    case RecurKind::FMaxNoFMFs:
     case RecurKind::FMaximumNum:
     case RecurKind::FMinimumNum:
     case RecurKind::None:

llvm/lib/Transforms/Vectorize/VPlan.h

@@ -980,7 +980,10 @@ class VPInstruction : public VPRecipeWithIRFlags,
     ReductionStartVector,
     // Creates a step vector starting from 0 to VF with a step of 1.
     StepVector,
-
+    /// Extracts a single lane (first operand) from a set of vector operands.
+    /// The lane specifies an index into a vector formed by combining all vector
+    /// operands (all operands after the first one).
+    ExtractLane,
   };
 
 private:

llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp

@@ -85,6 +85,7 @@ Type *VPTypeAnalysis::inferScalarTypeForRecipe(const VPInstruction *R) {
     return ResTy;
   }
   case Instruction::ICmp:
+  case Instruction::FCmp:
   case VPInstruction::ActiveLaneMask:
     assert(inferScalarType(R->getOperand(0)) ==
                inferScalarType(R->getOperand(1)) &&
@@ -110,6 +111,8 @@ Type *VPTypeAnalysis::inferScalarTypeForRecipe(const VPInstruction *R) {
   case VPInstruction::BuildStructVector:
   case VPInstruction::BuildVector:
     return SetResultTyFromOp();
+  case VPInstruction::ExtractLane:
+    return inferScalarType(R->getOperand(1));
   case VPInstruction::FirstActiveLane:
     return Type::getIntNTy(Ctx, 64);
   case VPInstruction::ExtractLastElement:

llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp

@@ -25,6 +25,7 @@
 #define DEBUG_TYPE "vplan"
 
 using namespace llvm;
+using namespace VPlanPatternMatch;
 
 namespace {
 // Class that is used to build the plain CFG for the incoming IR.
@@ -427,7 +428,6 @@ static void createLoopRegion(VPlan &Plan, VPBlockBase *HeaderVPB) {
 static void addCanonicalIVRecipes(VPlan &Plan, VPBasicBlock *HeaderVPBB,
                                   VPBasicBlock *LatchVPBB, Type *IdxTy,
                                   DebugLoc DL) {
-  using namespace VPlanPatternMatch;
   Value *StartIdx = ConstantInt::get(IdxTy, 0);
   auto *StartV = Plan.getOrAddLiveIn(StartIdx);
 
@@ -628,3 +628,114 @@ void VPlanTransforms::attachCheckBlock(VPlan &Plan, Value *Cond,
     Term->addMetadata(LLVMContext::MD_prof, BranchWeights);
   }
 }
+
+bool VPlanTransforms::handleFMaxReductionsWithoutFastMath(VPlan &Plan) {
+  VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion();
+  VPReductionPHIRecipe *RedPhiR = nullptr;
+  VPRecipeWithIRFlags *MinMaxOp = nullptr;
+  VPWidenIntOrFpInductionRecipe *WideIV = nullptr;
+
+  // Check if there are any FMaxNoFMFs reductions using wide selects that we can
+  // fix up. To do so, we also need a  wide canonical IV to keep track of the
+  // indices of the max values.
+  for (auto &R : LoopRegion->getEntryBasicBlock()->phis()) {
+    // We need a wide canonical IV
+    if (auto *CurIV = dyn_cast<VPWidenIntOrFpInductionRecipe>(&R)) {
+      if (!CurIV->isCanonical())
+        continue;
+      WideIV = CurIV;
+      continue;
+    }
+
+    // And a single FMaxNoFMFs reduction phi.
+    // TODO: Support FMin reductions as well.
+    auto *CurRedPhiR = dyn_cast<VPReductionPHIRecipe>(&R);
+    if (!CurRedPhiR)
+      continue;
+    if (RedPhiR)
+      return false;
+    if (CurRedPhiR->getRecurrenceKind() != RecurKind::FMaxNoFMFs ||
+        CurRedPhiR->isInLoop() || CurRedPhiR->isOrdered())
+      continue;
+    RedPhiR = CurRedPhiR;
+
+    // MaxOp feeding the reduction phi must be a select (either wide or a
+    // replicate recipe), where the phi is the last operand, and the compare
+    // predicate is strict. This ensures NaNs won't get propagated unless the
+    // initial value is NaN
+    VPRecipeBase *Inc = RedPhiR->getBackedgeValue()->getDefiningRecipe();
+    auto *RepR = dyn_cast<VPReplicateRecipe>(Inc);
+    if (!isa<VPWidenSelectRecipe>(Inc) &&
+        !(RepR && (isa<SelectInst>(RepR->getUnderlyingInstr()))))
+      return false;
+
+    MinMaxOp = cast<VPRecipeWithIRFlags>(Inc);
+    auto *Cmp = cast<VPRecipeWithIRFlags>(MinMaxOp->getOperand(0));
+    if (MinMaxOp->getOperand(1) == RedPhiR ||
+        !CmpInst::isStrictPredicate(Cmp->getPredicate()))
+      return false;
+  }
+
+  // Nothing to do.
+  if (!RedPhiR)
+    return true;
+
+  // A wide canonical IV is currently required.
+  // TODO: Create an induction if no suitable existing one is available.
+  if (!WideIV)
+    return false;
+
+  // Create a reduction that tracks the first indices where the latest maximum
+  // value has been selected. This is later used to select the max value from
+  // the partial reductions in a way that correctly handles signed zeros and
+  // NaNs in the input.
+  // Note that we do not need to check if the induction may hit the sentinel
+  // value. If the sentinel value gets hit, the final reduction value is at the
+  // last index or the maximum was never set and all lanes contain the start
+  // value. In either case, the correct value is selected.
+  unsigned IVWidth =
+      VPTypeAnalysis(Plan).inferScalarType(WideIV)->getScalarSizeInBits();
+  LLVMContext &Ctx = Plan.getScalarHeader()->getIRBasicBlock()->getContext();
+  VPValue *UMinSentinel =
+      Plan.getOrAddLiveIn(ConstantInt::get(Ctx, APInt::getMaxValue(IVWidth)));
+  auto *IdxPhi = new VPReductionPHIRecipe(nullptr, RecurKind::FindFirstIVUMin,
+                                          *UMinSentinel, false, false, 1);
+  IdxPhi->insertBefore(RedPhiR);
+  auto *MinIdxSel = new VPInstruction(
+      Instruction::Select, {MinMaxOp->getOperand(0), WideIV, IdxPhi});
+  MinIdxSel->insertAfter(MinMaxOp);
+  IdxPhi->addOperand(MinIdxSel);
+
+  // Find the first index of with the maximum value. This is used to extract the
+  // lane with the final max value and is needed to handle signed zeros and NaNs
+  // in the input.
+  auto *MiddleVPBB = Plan.getMiddleBlock();
+  auto *OrigRdxResult = cast<VPSingleDefRecipe>(&MiddleVPBB->front());
+  VPBuilder Builder(OrigRdxResult->getParent(),
+                    std::next(OrigRdxResult->getIterator()));
+
+  // Create mask for lanes that have the max value and use it to mask out
+  // indices that don't contain maximum values.
+  auto *MaskFinalMaxValue = Builder.createNaryOp(
+      Instruction::FCmp, {OrigRdxResult->getOperand(1), OrigRdxResult},
+      VPIRFlags(CmpInst::FCMP_OEQ));
+  auto *IndicesWithMaxValue = Builder.createNaryOp(
+      Instruction::Select, {MaskFinalMaxValue, MinIdxSel, UMinSentinel});
+  auto *FirstMaxIdx = Builder.createNaryOp(
+      VPInstruction::ComputeFindIVResult,
+      {IdxPhi, WideIV->getStartValue(), UMinSentinel, IndicesWithMaxValue});
+  // Convert the index of the first max value to an index in the vector lanes of
+  // the partial reduction results. This ensures we select the first max value
+  // and acts as a tie-breaker if the partial reductions contain signed zeros.
+  auto *FirstMaxLane =
+      Builder.createNaryOp(Instruction::URem, {FirstMaxIdx, &Plan.getVFxUF()});
+
+  // Extract the final max value and update the users.
+  auto *Res = Builder.createNaryOp(
+      VPInstruction::ExtractLane, {FirstMaxLane, OrigRdxResult->getOperand(1)});
+  OrigRdxResult->replaceUsesWithIf(Res,
+                                   [MaskFinalMaxValue](VPUser &U, unsigned) {
+                                     return &U != MaskFinalMaxValue;
+                                   });
+  return true;
+}

llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp

@@ -585,6 +585,7 @@ Value *VPInstruction::generate(VPTransformState &State) {
     Value *Op = State.get(getOperand(0), vputils::onlyFirstLaneUsed(this));
     return Builder.CreateFreeze(Op, Name);
   }
+  case Instruction::FCmp:
   case Instruction::ICmp: {
     bool OnlyFirstLaneUsed = vputils::onlyFirstLaneUsed(this);
     Value *A = State.get(getOperand(0), OnlyFirstLaneUsed);
@@ -595,7 +596,8 @@ Value *VPInstruction::generate(VPTransformState &State) {
     llvm_unreachable("should be handled by VPPhi::execute");
   }
   case Instruction::Select: {
-    bool OnlyFirstLaneUsed = vputils::onlyFirstLaneUsed(this);
+    bool OnlyFirstLaneUsed =
+        State.VF.isScalar() || vputils::onlyFirstLaneUsed(this);
     Value *Cond = State.get(getOperand(0), OnlyFirstLaneUsed);
     Value *Op1 = State.get(getOperand(1), OnlyFirstLaneUsed);
     Value *Op2 = State.get(getOperand(2), OnlyFirstLaneUsed);
@@ -858,7 +860,30 @@ Value *VPInstruction::generate(VPTransformState &State) {
     Value *Res = State.get(getOperand(0));
     for (VPValue *Op : drop_begin(operands()))
       Res = Builder.CreateOr(Res, State.get(Op));
-    return Builder.CreateOrReduce(Res);
+    return Res->getType()->isIntegerTy(1) ? Res : Builder.CreateOrReduce(Res);
+  }
+  case VPInstruction::ExtractLane: {
+    Value *LaneToExtract = State.get(getOperand(0), true);
+    Type *IdxTy = State.TypeAnalysis.inferScalarType(getOperand(0));
+    Value *Res = nullptr;
+    Value *RuntimeVF = getRuntimeVF(State.Builder, IdxTy, State.VF);
+
+    for (unsigned Idx = 1; Idx != getNumOperands(); ++Idx) {
+      Value *VectorStart =
+          Builder.CreateMul(RuntimeVF, ConstantInt::get(IdxTy, Idx - 1));
+      Value *VectorIdx = Builder.CreateSub(LaneToExtract, VectorStart);
+      Value *Ext = State.VF.isScalar()
+                       ? State.get(getOperand(Idx))
+                       : Builder.CreateExtractElement(
+                             State.get(getOperand(Idx)), VectorIdx);
+      if (Res) {
+        Value *Cmp = Builder.CreateICmpUGE(LaneToExtract, VectorStart);
+        Res = Builder.CreateSelect(Cmp, Ext, Res);
+      } else {
+        Res = Ext;
+      }
+    }
+    return Res;
   }
   case VPInstruction::FirstActiveLane: {
     if (getNumOperands() == 1) {
@@ -984,7 +1009,8 @@ bool VPInstruction::isVectorToScalar() const {
          getOpcode() == VPInstruction::ComputeAnyOfResult ||
          getOpcode() == VPInstruction::ComputeFindIVResult ||
          getOpcode() == VPInstruction::ComputeReductionResult ||
-         getOpcode() == VPInstruction::AnyOf;
+         getOpcode() == VPInstruction::AnyOf ||
+         getOpcode() == VPInstruction::ExtractLane;
 }
 
 bool VPInstruction::isSingleScalar() const {
@@ -1031,6 +1057,7 @@ bool VPInstruction::opcodeMayReadOrWriteFromMemory() const {
   switch (getOpcode()) {
   case Instruction::ExtractElement:
   case Instruction::Freeze:
+  case Instruction::FCmp:
   case Instruction::ICmp:
   case Instruction::Select:
   case VPInstruction::AnyOf:
@@ -1066,6 +1093,7 @@ bool VPInstruction::onlyFirstLaneUsed(const VPValue *Op) const {
     return Op == getOperand(1);
   case Instruction::PHI:
     return true;
+  case Instruction::FCmp:
   case Instruction::ICmp:
   case Instruction::Select:
   case Instruction::Or:
@@ -1098,6 +1126,7 @@ bool VPInstruction::onlyFirstPartUsed(const VPValue *Op) const {
   switch (getOpcode()) {
   default:
     return false;
+  case Instruction::FCmp:
   case Instruction::ICmp:
   case Instruction::Select:
     return vputils::onlyFirstPartUsed(this);
@@ -1782,7 +1811,7 @@ bool VPIRFlags::flagsValidForOpcode(unsigned Opcode) const {
     return Opcode == Instruction::ZExt;
     break;
   case OperationType::Cmp:
-    return Opcode == Instruction::ICmp;
+    return Opcode == Instruction::FCmp || Opcode == Instruction::ICmp;
   case OperationType::Other:
     return true;
   }

llvm/lib/Transforms/Vectorize/VPlanTransforms.h

@@ -196,6 +196,8 @@ struct VPlanTransforms {
                                          VPBasicBlock *LatchVPBB,
                                          VFRange &Range);
 
+  static bool handleFMaxReductionsWithoutFastMath(VPlan &Plan);
+
   /// Replace loop regions with explicit CFG.
   static void dissolveLoopRegions(VPlan &Plan);