[LoopVectorize] Vectorize select-cmp reduction pattern for increasing

integer induction variable

Consider the following loop:

  int rdx = init;
  for (int i = 0; i < n; ++i)
    rdx = (a[i] > b[i]) ? i : rdx;

We can vectorize this loop if `i` is an increasing induction variable.
The final reduced value will be the maximum of `i` that the condition
`a[i] > b[i]` is satisfied, or the start value `init`.

This patch added new RecurKind enums - IFindLastIV and FFindLastIV.

Author: Mel-Chen

llvm/include/llvm/Analysis/IVDescriptors.h

@@ -51,9 +51,16 @@ enum class RecurKind {
   FMulAdd,  ///< Sum of float products with llvm.fmuladd(a * b + sum).
   IAnyOf,   ///< Any_of reduction with select(icmp(),x,y) where one of (x,y) is
             ///< loop invariant, and both x and y are integer type.
-  FAnyOf    ///< Any_of reduction with select(fcmp(),x,y) where one of (x,y) is
+  FAnyOf,   ///< Any_of reduction with select(fcmp(),x,y) where one of (x,y) is
             ///< loop invariant, and both x and y are integer type.
-  // TODO: Any_of reduction need not be restricted to integer type only.
+  IFindLastIV, ///< FindLast reduction with select(icmp(),x,y) where one of
+               ///< (x,y) is increasing loop induction PHI, and both x and y are
+               ///< integer type.
+  FFindLastIV ///< FindLast reduction with select(fcmp(),x,y) where one of (x,y)
+              ///< is increasing loop induction PHI, and both x and y are
+              ///< integer type.
+  // TODO: Any_of and FindLast reduction need not be restricted to integer type
+  // only.
 };
 
 /// The RecurrenceDescriptor is used to identify recurrences variables in a
@@ -125,7 +132,7 @@ class RecurrenceDescriptor {
   /// the returned struct.
   static InstDesc isRecurrenceInstr(Loop *L, PHINode *Phi, Instruction *I,
                                     RecurKind Kind, InstDesc &Prev,
-                                    FastMathFlags FuncFMF);
+                                    FastMathFlags FuncFMF, ScalarEvolution *SE);
 
   /// Returns true if instruction I has multiple uses in Insts
   static bool hasMultipleUsesOf(Instruction *I,
@@ -152,6 +159,16 @@ class RecurrenceDescriptor {
   static InstDesc isAnyOfPattern(Loop *Loop, PHINode *OrigPhi, Instruction *I,
                                  InstDesc &Prev);
 
+  /// Returns a struct describing whether the instruction is either a
+  ///   Select(ICmp(A, B), X, Y), or
+  ///   Select(FCmp(A, B), X, Y)
+  /// where one of (X, Y) is an increasing loop induction variable, and the
+  /// other is a PHI value.
+  // TODO: FindLast does not need be restricted to increasing loop induction
+  // variables.
+  static InstDesc isFindLastIVPattern(Loop *Loop, PHINode *OrigPhi,
+                                      Instruction *I, ScalarEvolution *SE);
+
   /// Returns a struct describing if the instruction is a
   /// Select(FCmp(X, Y), (Z = X op PHINode), PHINode) instruction pattern.
   static InstDesc isConditionalRdxPattern(RecurKind Kind, Instruction *I);
@@ -240,6 +257,12 @@ class RecurrenceDescriptor {
     return Kind == RecurKind::IAnyOf || Kind == RecurKind::FAnyOf;
   }
 
+  /// Returns true if the recurrence kind is of the form
+  ///   select(cmp(),x,y) where one of (x,y) is increasing loop induction.
+  static bool isFindLastIVRecurrenceKind(RecurKind Kind) {
+    return Kind == RecurKind::IFindLastIV || Kind == RecurKind::FFindLastIV;
+  }
+
   /// Returns the type of the recurrence. This type can be narrower than the
   /// actual type of the Phi if the recurrence has been type-promoted.
   Type *getRecurrenceType() const { return RecurrenceType; }

llvm/include/llvm/Transforms/Utils/LoopUtils.h

@@ -372,6 +372,12 @@ RecurKind getMinMaxReductionRecurKind(Intrinsic::ID RdxID);
 /// Returns the comparison predicate used when expanding a min/max reduction.
 CmpInst::Predicate getMinMaxReductionPredicate(RecurKind RK);
 
+/// See RecurrenceDescriptor::isFindLastIVPattern for a description of the
+/// pattern we are trying to match. In this pattern, since the selected set of
+/// values forms an increasing sequence, we are selecting the maximum value from
+/// \p Left and \p Right.
+Value *createFindLastIVOp(IRBuilderBase &Builder, Value *Left, Value *Right);
+
 /// Returns a Min/Max operation corresponding to MinMaxRecurrenceKind.
 /// The Builder's fast-math-flags must be set to propagate the expected values.
 Value *createMinMaxOp(IRBuilderBase &Builder, RecurKind RK, Value *Left,
@@ -402,6 +408,12 @@ Value *createAnyOfTargetReduction(IRBuilderBase &B, Value *Src,
                                   const RecurrenceDescriptor &Desc,
                                   PHINode *OrigPhi);
 
+/// Create a target reduction of the given vector \p Src for a reduction of the
+/// kind RecurKind::IFindLastIV or RecurKind::FFindLastIV. The reduction
+/// operation is described by \p Desc.
+Value *createFindLastIVTargetReduction(IRBuilderBase &B, Value *Src,
+                                       const RecurrenceDescriptor &Desc);
+
 /// Create a generic target reduction using a recurrence descriptor \p Desc
 /// The target is queried to determine if intrinsics or shuffle sequences are
 /// required to implement the reduction.
@@ -415,6 +427,15 @@ Value *createOrderedReduction(IRBuilderBase &B,
                               const RecurrenceDescriptor &Desc, Value *Src,
                               Value *Start);
 
+/// Returns a set of cmp and select instructions as shown below:
+///   Select(Cmp(NE, Rdx, Iden), Rdx, InitVal)
+/// where \p Rdx is a scalar value generated by target reduction, Iden is the
+/// sentinel value of the recurrence descriptor \p Desc, and InitVal is the
+/// start value of the recurrence descriptor \p Desc.
+Value *createSentinelValueHandling(IRBuilderBase &Builder,
+                                   const RecurrenceDescriptor &Desc,
+                                   Value *Rdx);
+
 /// Get the intersection (logical and) of all of the potential IR flags
 /// of each scalar operation (VL) that will be converted into a vector (I).
 /// If OpValue is non-null, we only consider operations similar to OpValue

llvm/lib/Analysis/IVDescriptors.cpp

@@ -52,6 +52,8 @@ bool RecurrenceDescriptor::isIntegerRecurrenceKind(RecurKind Kind) {
   case RecurKind::UMin:
   case RecurKind::IAnyOf:
   case RecurKind::FAnyOf:
+  case RecurKind::IFindLastIV:
+  case RecurKind::FFindLastIV:
     return true;
   }
   return false;
@@ -373,7 +375,7 @@ bool RecurrenceDescriptor::AddReductionVar(
     // type-promoted).
     if (Cur != Start) {
       ReduxDesc =
-          isRecurrenceInstr(TheLoop, Phi, Cur, Kind, ReduxDesc, FuncFMF);
+          isRecurrenceInstr(TheLoop, Phi, Cur, Kind, ReduxDesc, FuncFMF, SE);
       ExactFPMathInst = ExactFPMathInst == nullptr
                             ? ReduxDesc.getExactFPMathInst()
                             : ExactFPMathInst;
@@ -660,6 +662,87 @@ RecurrenceDescriptor::isAnyOfPattern(Loop *Loop, PHINode *OrigPhi,
                                                      : RecurKind::FAnyOf);
 }
 
+// We are looking for loops that do something like this:
+//   int r = 0;
+//   for (int i = 0; i < n; i++) {
+//     if (src[i] > 3)
+//       r = i;
+//   }
+// The reduction value (r) is derived from either the values of an increasing
+// induction variable (i) sequence, or from the start value (0).
+// The LLVM IR generated for such loops would be as follows:
+//   for.body:
+//     %r = phi i32 [ %spec.select, %for.body ], [ 0, %entry ]
+//     %i = phi i32 [ %inc, %for.body ], [ 0, %entry ]
+//     ...
+//     %cmp = icmp sgt i32 %5, 3
+//     %spec.select = select i1 %cmp, i32 %i, i32 %r
+//     %inc = add nsw i32 %i, 1
+//     ...
+// Since 'i' is an increasing induction variable, the reduction value after the
+// loop will be the maximum value of 'i' that the condition (src[i] > 3) is
+// satisfied, or the start value (0 in the example above). When the start value
+// of the increasing induction variable 'i' is greater than the minimum value of
+// the data type, we can use the minimum value of the data type as a sentinel
+// value to replace the start value. This allows us to perform a single
+// reduction max operation to obtain the final reduction result.
+// TODO: It is possible to solve the case where the start value is the minimum
+// value of the data type or a non-constant value by using mask and multiple
+// reduction operations.
+RecurrenceDescriptor::InstDesc
+RecurrenceDescriptor::isFindLastIVPattern(Loop *Loop, PHINode *OrigPhi,
+                                          Instruction *I, ScalarEvolution *SE) {
+  // Only match select with single use cmp condition.
+  // TODO: Only handle single use for now.
+  CmpInst::Predicate Pred;
+  if (!match(I, m_Select(m_OneUse(m_Cmp(Pred, m_Value(), m_Value())), m_Value(),
+                         m_Value())))
+    return InstDesc(false, I);
+
+  SelectInst *SI = cast<SelectInst>(I);
+  Value *NonRdxPhi = nullptr;
+
+  if (OrigPhi == dyn_cast<PHINode>(SI->getTrueValue()))
+    NonRdxPhi = SI->getFalseValue();
+  else if (OrigPhi == dyn_cast<PHINode>(SI->getFalseValue()))
+    NonRdxPhi = SI->getTrueValue();
+  else
+    return InstDesc(false, I);
+
+  auto IsIncreasingLoopInduction = [&SE, &Loop](Value *V) {
+    auto *Phi = dyn_cast<PHINode>(V);
+    if (!Phi)
+      return false;
+
+    if (!SE)
+      return false;
+
+    InductionDescriptor ID;
+    if (!InductionDescriptor::isInductionPHI(Phi, Loop, SE, ID))
+      return false;
+
+    const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(SE->getSCEV(Phi));
+    if (!AR->hasNoSignedWrap())
+      return false;
+
+    ConstantInt *IVStartValue = dyn_cast<ConstantInt>(ID.getStartValue());
+    if (!IVStartValue || IVStartValue->isMinSignedValue())
+      return false;
+
+    const SCEV *Step = ID.getStep();
+    return SE->isKnownPositive(Step);
+  };
+
+  // We are looking for selects of the form:
+  //   select(cmp(), phi, loop_induction) or
+  //   select(cmp(), loop_induction, phi)
+  if (!IsIncreasingLoopInduction(NonRdxPhi))
+    return InstDesc(false, I);
+
+  return InstDesc(I, isa<ICmpInst>(I->getOperand(0)) ? RecurKind::IFindLastIV
+                                                     : RecurKind::FFindLastIV);
+}
+
 RecurrenceDescriptor::InstDesc
 RecurrenceDescriptor::isMinMaxPattern(Instruction *I, RecurKind Kind,
                                       const InstDesc &Prev) {
@@ -763,10 +846,9 @@ RecurrenceDescriptor::isConditionalRdxPattern(RecurKind Kind, Instruction *I) {
   return InstDesc(true, SI);
 }
 
-RecurrenceDescriptor::InstDesc
-RecurrenceDescriptor::isRecurrenceInstr(Loop *L, PHINode *OrigPhi,
-                                        Instruction *I, RecurKind Kind,
-                                        InstDesc &Prev, FastMathFlags FuncFMF) {
+RecurrenceDescriptor::InstDesc RecurrenceDescriptor::isRecurrenceInstr(
+    Loop *L, PHINode *OrigPhi, Instruction *I, RecurKind Kind, InstDesc &Prev,
+    FastMathFlags FuncFMF, ScalarEvolution *SE) {
   assert(Prev.getRecKind() == RecurKind::None || Prev.getRecKind() == Kind);
   switch (I->getOpcode()) {
   default:
@@ -796,6 +878,8 @@ RecurrenceDescriptor::isRecurrenceInstr(Loop *L, PHINode *OrigPhi,
     if (Kind == RecurKind::FAdd || Kind == RecurKind::FMul ||
         Kind == RecurKind::Add || Kind == RecurKind::Mul)
       return isConditionalRdxPattern(Kind, I);
+    if (isFindLastIVRecurrenceKind(Kind))
+      return isFindLastIVPattern(L, OrigPhi, I, SE);
     [[fallthrough]];
   case Instruction::FCmp:
   case Instruction::ICmp:
@@ -900,6 +984,11 @@ bool RecurrenceDescriptor::isReductionPHI(PHINode *Phi, Loop *TheLoop,
                       << *Phi << "\n");
     return true;
   }
+  if (AddReductionVar(Phi, RecurKind::IFindLastIV, TheLoop, FMF, RedDes, DB, AC,
+                      DT, SE)) {
+    LLVM_DEBUG(dbgs() << "Found a FindLastIV reduction PHI." << *Phi << "\n");
+    return true;
+  }
   if (AddReductionVar(Phi, RecurKind::FMul, TheLoop, FMF, RedDes, DB, AC, DT,
                       SE)) {
     LLVM_DEBUG(dbgs() << "Found an FMult reduction PHI." << *Phi << "\n");
@@ -1089,6 +1178,9 @@ Value *RecurrenceDescriptor::getRecurrenceIdentity(RecurKind K, Type *Tp,
   case RecurKind::FAnyOf:
     return getRecurrenceStartValue();
     break;
+  case RecurKind::IFindLastIV:
+  case RecurKind::FFindLastIV:
+    return getRecurrenceIdentity(RecurKind::SMax, Tp, FMF);
   default:
     llvm_unreachable("Unknown recurrence kind");
   }
@@ -1116,12 +1208,14 @@ unsigned RecurrenceDescriptor::getOpcode(RecurKind Kind) {
   case RecurKind::UMax:
   case RecurKind::UMin:
   case RecurKind::IAnyOf:
+  case RecurKind::IFindLastIV:
     return Instruction::ICmp;
   case RecurKind::FMax:
   case RecurKind::FMin:
   case RecurKind::FMaximum:
   case RecurKind::FMinimum:
   case RecurKind::FAnyOf:
+  case RecurKind::FFindLastIV:
     return Instruction::FCmp;
   default:
     llvm_unreachable("Unknown recurrence operation");

llvm/lib/Transforms/Utils/LoopUtils.cpp

@@ -1034,6 +1034,11 @@ CmpInst::Predicate llvm::getMinMaxReductionPredicate(RecurKind RK) {
   }
 }
 
+Value *llvm::createFindLastIVOp(IRBuilderBase &Builder, Value *Left,
+                                Value *Right) {
+  return createMinMaxOp(Builder, RecurKind::SMax, Left, Right);
+}
+
 Value *llvm::createMinMaxOp(IRBuilderBase &Builder, RecurKind RK, Value *Left,
                             Value *Right) {
   Type *Ty = Left->getType();
@@ -1151,6 +1156,14 @@ Value *llvm::createAnyOfTargetReduction(IRBuilderBase &Builder, Value *Src,
   return Builder.CreateSelect(AnyOf, NewVal, InitVal, "rdx.select");
 }
 
+Value *llvm::createFindLastIVTargetReduction(IRBuilderBase &Builder, Value *Src,
+                                             const RecurrenceDescriptor &Desc) {
+  assert(RecurrenceDescriptor::isFindLastIVRecurrenceKind(
+             Desc.getRecurrenceKind()) &&
+         "Unexpected reduction kind");
+  return Builder.CreateIntMaxReduce(Src, true);
+}
+
 Value *llvm::createSimpleTargetReduction(IRBuilderBase &Builder, Value *Src,
                                          RecurKind RdxKind) {
   auto *SrcVecEltTy = cast<VectorType>(Src->getType())->getElementType();
@@ -1204,6 +1217,8 @@ Value *llvm::createTargetReduction(IRBuilderBase &B,
   RecurKind RK = Desc.getRecurrenceKind();
   if (RecurrenceDescriptor::isAnyOfRecurrenceKind(RK))
     return createAnyOfTargetReduction(B, Src, Desc, OrigPhi);
+  if (RecurrenceDescriptor::isFindLastIVRecurrenceKind(RK))
+    return createFindLastIVTargetReduction(B, Src, Desc);
 
   return createSimpleTargetReduction(B, Src, RK);
 }
@@ -1220,6 +1235,16 @@ Value *llvm::createOrderedReduction(IRBuilderBase &B,
   return B.CreateFAddReduce(Start, Src);
 }
 
+Value *llvm::createSentinelValueHandling(IRBuilderBase &Builder,
+                                         const RecurrenceDescriptor &Desc,
+                                         Value *Rdx) {
+  Value *InitVal = Desc.getRecurrenceStartValue();
+  Value *Iden = Desc.getRecurrenceIdentity(
+      Desc.getRecurrenceKind(), Rdx->getType(), Desc.getFastMathFlags());
+  Value *Cmp = Builder.CreateCmp(CmpInst::ICMP_NE, Rdx, Iden, "rdx.select.cmp");
+  return Builder.CreateSelect(Cmp, Rdx, InitVal, "rdx.select");
+}
+
 void llvm::propagateIRFlags(Value *I, ArrayRef<Value *> VL, Value *OpValue,
                             bool IncludeWrapFlags) {
   auto *VecOp = dyn_cast<Instruction>(I);

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -5510,8 +5510,9 @@ LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF,
         HasReductions &&
         any_of(Legal->getReductionVars(), [&](auto &Reduction) -> bool {
           const RecurrenceDescriptor &RdxDesc = Reduction.second;
-          return RecurrenceDescriptor::isAnyOfRecurrenceKind(
-              RdxDesc.getRecurrenceKind());
+          RecurKind RK = RdxDesc.getRecurrenceKind();
+          return RecurrenceDescriptor::isAnyOfRecurrenceKind(RK) ||
+                 RecurrenceDescriptor::isFindLastIVRecurrenceKind(RK);
         });
     if (HasSelectCmpReductions) {
       LLVM_DEBUG(dbgs() << "LV: Not interleaving select-cmp reductions.\n");
@@ -8948,8 +8949,10 @@ void LoopVectorizationPlanner::adjustRecipesForReductions(
 
     const RecurrenceDescriptor &RdxDesc = PhiR->getRecurrenceDescriptor();
     RecurKind Kind = RdxDesc.getRecurrenceKind();
-    assert(!RecurrenceDescriptor::isAnyOfRecurrenceKind(Kind) &&
-           "AnyOf reductions are not allowed for in-loop reductions");
+    assert(
+        (!RecurrenceDescriptor::isAnyOfRecurrenceKind(Kind) &&
+         !RecurrenceDescriptor::isFindLastIVRecurrenceKind(Kind)) &&
+        "AnyOf and FindLast reductions are not allowed for in-loop reductions");
 
     // Collect the chain of "link" recipes for the reduction starting at PhiR.
     SetVector<VPSingleDefRecipe *> Worklist;

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

@@ -16781,6 +16781,8 @@ class HorizontalReduction {
     case RecurKind::FMulAdd:
     case RecurKind::IAnyOf:
     case RecurKind::FAnyOf:
+    case RecurKind::IFindLastIV:
+    case RecurKind::FFindLastIV:
     case RecurKind::None:
       llvm_unreachable("Unexpected reduction kind for repeated scalar.");
     }
@@ -16883,6 +16885,8 @@ class HorizontalReduction {
     case RecurKind::FMulAdd:
     case RecurKind::IAnyOf:
     case RecurKind::FAnyOf:
+    case RecurKind::IFindLastIV:
+    case RecurKind::FFindLastIV:
     case RecurKind::None:
       llvm_unreachable("Unexpected reduction kind for reused scalars.");
     }