[LV] Vectorize selecting last IV of min/max element.

llvm/include/llvm/Analysis/IVDescriptors.h

@@ -95,12 +95,17 @@ class RecurrenceDescriptor {
                        RecurKind K, FastMathFlags FMF, Instruction *ExactFP,
                        Type *RT, bool Signed, bool Ordered,
                        SmallPtrSetImpl<Instruction *> &CI,
-                       unsigned MinWidthCastToRecurTy)
+                       unsigned MinWidthCastToRecurTy,
+                       bool PhiHasUsesOutsideReductionChain = false)
       : IntermediateStore(Store), StartValue(Start), LoopExitInstr(Exit),
         Kind(K), FMF(FMF), ExactFPMathInst(ExactFP), RecurrenceType(RT),
         IsSigned(Signed), IsOrdered(Ordered),
+        PhiHasUsesOutsideReductionChain(PhiHasUsesOutsideReductionChain),
         MinWidthCastToRecurrenceType(MinWidthCastToRecurTy) {
     CastInsts.insert_range(CI);
+    assert(
+        (!PhiHasUsesOutsideReductionChain || isMinMaxRecurrenceKind(K)) &&
+        "Only min/max recurrences are allowed to have multiple uses currently");
   }
 
   /// This POD struct holds information about a potential recurrence operation.
@@ -339,6 +344,13 @@ class RecurrenceDescriptor {
   /// Expose an ordered FP reduction to the instance users.
   bool isOrdered() const { return IsOrdered; }
 
+  /// Returns true if the reduction PHI has any uses outside the reduction
+  /// chain. This is relevant for min/max reductions that are part of a
+  /// FindLastIV pattern.
+  bool hasUsesOutsideReductionChain() const {
+    return PhiHasUsesOutsideReductionChain;
+  }
+
   /// Attempts to find a chain of operations from Phi to LoopExitInst that can
   /// be treated as a set of reductions instructions for in-loop reductions.
   LLVM_ABI SmallVector<Instruction *, 4> getReductionOpChain(PHINode *Phi,
@@ -376,6 +388,10 @@ class RecurrenceDescriptor {
   // Currently only a non-reassociative FAdd can be considered in-order,
   // if it is also the only FAdd in the PHI's use chain.
   bool IsOrdered = false;
+  // True if the reduction PHI has in-loop users outside the reduction chain.
+  // This is relevant for min/max reductions that are part of a FindLastIV
+  // pattern.
+  bool PhiHasUsesOutsideReductionChain = false;
   // Instructions used for type-promoting the recurrence.
   SmallPtrSet<Instruction *, 8> CastInsts;
   // The minimum width used by the recurrence.

llvm/lib/Analysis/IVDescriptors.cpp

@@ -216,6 +216,52 @@ static bool checkOrderedReduction(RecurKind Kind, Instruction *ExactFPMathInst,
   return true;
 }
 
+/// Returns true if \p Phi is a min/max reduction matching \p Kind where \p Phi
+/// is used outside the reduction chain. This is common for loops selecting the
+/// index of a minimum/maximum value (argmin/argmax).
+static bool isMinMaxReductionPhiWithUsersOutsideReductionChain(
+    PHINode *Phi, RecurKind Kind, Loop *TheLoop, RecurrenceDescriptor &RedDes) {
+  BasicBlock *Latch = TheLoop->getLoopLatch();
+  if (!Latch)
+    return false;
+
+  assert(Phi->getNumIncomingValues() == 2 && "phi must have 2 incoming values");
+  Value *Inc = Phi->getIncomingValueForBlock(Latch);
+  if (Phi->hasOneUse() || !Inc->hasOneUse() ||
+      !RecurrenceDescriptor::isIntMinMaxRecurrenceKind(Kind))
+    return false;
+
+  Value *A, *B;
+  bool IsMinMax = [&]() {
+    switch (Kind) {
+    case RecurKind::UMax:
+      return match(Inc, m_UMax(m_Value(A), m_Value(B)));
+    case RecurKind::UMin:
+      return match(Inc, m_UMin(m_Value(A), m_Value(B)));
+    case RecurKind::SMax:
+      return match(Inc, m_SMax(m_Value(A), m_Value(B)));
+    case RecurKind::SMin:
+      return match(Inc, m_SMin(m_Value(A), m_Value(B)));
+    default:
+      llvm_unreachable("all min/max kinds must be handled");
+    }
+  }();
+  if (!IsMinMax)
+    return false;
+
+  if (A == B || (A != Phi && B != Phi))
+    return false;
+
+  SmallPtrSet<Instruction *, 4> CastInsts;
+  Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
+  RedDes =
+      RecurrenceDescriptor(RdxStart, /*Exit=*/nullptr, /*Store=*/nullptr, Kind,
+                           FastMathFlags(), /*ExactFP=*/nullptr, Phi->getType(),
+                           /*Signed=*/false, /*Ordered=*/false, CastInsts,
+                           /*MinWidthCastToRecurTy=*/-1U, /*PhiMultiUse=*/true);
+  return true;
+}
+
 bool RecurrenceDescriptor::AddReductionVar(
     PHINode *Phi, RecurKind Kind, Loop *TheLoop, FastMathFlags FuncFMF,
     RecurrenceDescriptor &RedDes, DemandedBits *DB, AssumptionCache *AC,
@@ -227,6 +273,11 @@ bool RecurrenceDescriptor::AddReductionVar(
   if (Phi->getParent() != TheLoop->getHeader())
     return false;
 
+  // Check for min/max reduction variables that feed other users in the loop.
+  if (isMinMaxReductionPhiWithUsersOutsideReductionChain(Phi, Kind, TheLoop,
+                                                         RedDes))
+    return true;
+
   // Obtain the reduction start value from the value that comes from the loop
   // preheader.
   Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());

llvm/lib/Transforms/Utils/LoopUnroll.cpp

@@ -1254,6 +1254,8 @@ llvm::canParallelizeReductionWhenUnrolling(PHINode &Phi, Loop *L,
                                             /*DemandedBits=*/nullptr,
                                             /*AC=*/nullptr, /*DT=*/nullptr, SE))
     return std::nullopt;
+  if (RdxDesc.hasUsesOutsideReductionChain())
+    return std::nullopt;
   RecurKind RK = RdxDesc.getRecurrenceKind();
   // Skip unsupported reductions.
   // TODO: Handle additional reductions, including min-max reductions.

llvm/lib/Transforms/Vectorize/LoopVectorizationLegality.cpp

@@ -877,6 +877,11 @@ bool LoopVectorizationLegality::canVectorizeInstr(Instruction &I) {
       Requirements->addExactFPMathInst(RedDes.getExactFPMathInst());
       AllowedExit.insert(RedDes.getLoopExitInstr());
       Reductions[Phi] = RedDes;
+      assert((!RedDes.hasUsesOutsideReductionChain() ||
+              RecurrenceDescriptor::isMinMaxRecurrenceKind(
+                  RedDes.getRecurrenceKind())) &&
+             "Only min/max recurrences are allowed to have multiple uses "
+             "currently");
       return true;
     }
 

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -6593,6 +6593,11 @@ void LoopVectorizationCostModel::collectInLoopReductions() {
     PHINode *Phi = Reduction.first;
     const RecurrenceDescriptor &RdxDesc = Reduction.second;
 
+    // Multi-use reductions (e.g., used in FindLastIV patterns) are handled
+    // separately and should not be considered for in-loop reductions.
+    if (RdxDesc.hasUsesOutsideReductionChain())
+      continue;
+
     // We don't collect reductions that are type promoted (yet).
     if (RdxDesc.getRecurrenceType() != Phi->getType())
       continue;
@@ -7998,9 +8003,10 @@ void VPRecipeBuilder::collectScaledReductions(VFRange &Range) {
   MapVector<Instruction *,
             SmallVector<std::pair<PartialReductionChain, unsigned>>>
       ChainsByPhi;
-  for (const auto &[Phi, RdxDesc] : Legal->getReductionVars())
-    getScaledReductions(Phi, RdxDesc.getLoopExitInstr(), Range,
-                        ChainsByPhi[Phi]);
+  for (const auto &[Phi, RdxDesc] : Legal->getReductionVars()) {
+    if (Instruction *RdxExitInstr = RdxDesc.getLoopExitInstr())
+      getScaledReductions(Phi, RdxExitInstr, Range, ChainsByPhi[Phi]);
+  }
 
   // A partial reduction is invalid if any of its extends are used by
   // something that isn't another partial reduction. This is because the
@@ -8221,7 +8227,8 @@ VPRecipeBase *VPRecipeBuilder::tryToCreateWidenRecipe(VPSingleDefRecipe *R,
       PhiRecipe = new VPReductionPHIRecipe(
           Phi, RdxDesc.getRecurrenceKind(), *StartV,
           getReductionStyle(UseInLoopReduction, UseOrderedReductions,
-                            ScaleFactor));
+                            ScaleFactor),
+          RdxDesc.hasUsesOutsideReductionChain());
     } else {
       // TODO: Currently fixed-order recurrences are modeled as chains of
       // first-order recurrences. If there are no users of the intermediate
@@ -8555,6 +8562,11 @@ VPlanPtr LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(
   // Adjust the recipes for any inloop reductions.
   adjustRecipesForReductions(Plan, RecipeBuilder, Range.Start);
 
+  // Apply mandatory transformation to handle reductions with multiple in-loop
+  // uses if possible, bail out otherwise.
+  if (!VPlanTransforms::runPass(VPlanTransforms::handleMultiUseReductions,
+                                *Plan))
+    return nullptr;
   // Apply mandatory transformation to handle FP maxnum/minnum reduction with
   // NaNs if possible, bail out otherwise.
   if (!VPlanTransforms::runPass(VPlanTransforms::handleMaxMinNumReductions,

llvm/lib/Transforms/Vectorize/VPlan.h

@@ -2071,6 +2071,9 @@ class LLVM_ABI_FOR_TEST VPHeaderPHIRecipe : public VPSingleDefRecipe,
   static inline bool classof(const VPValue *V) {
     return isa<VPHeaderPHIRecipe>(V->getDefiningRecipe());
   }
+  static inline bool classof(const VPSingleDefRecipe *R) {
+    return isa<VPHeaderPHIRecipe>(static_cast<const VPRecipeBase *>(R));
+  }
 
   /// Generate the phi nodes.
   void execute(VPTransformState &State) override = 0;
@@ -2136,7 +2139,7 @@ class VPWidenInductionRecipe : public VPHeaderPHIRecipe {
     return R && classof(R);
   }
 
-  static inline bool classof(const VPHeaderPHIRecipe *R) {
+  static inline bool classof(const VPSingleDefRecipe *R) {
     return classof(static_cast<const VPRecipeBase *>(R));
   }
 
@@ -2432,19 +2435,27 @@ class VPReductionPHIRecipe : public VPHeaderPHIRecipe,
 
   ReductionStyle Style;
 
+  /// The phi is part of a multi-use reduction (e.g., used in FindLastIV
+  /// patterns for argmin/argmax).
+  /// TODO: Also support cases where the phi itself has a single use, but its
+  /// compare has multiple uses.
+  bool HasUsesOutsideReductionChain;
+
 public:
   /// Create a new VPReductionPHIRecipe for the reduction \p Phi.
   VPReductionPHIRecipe(PHINode *Phi, RecurKind Kind, VPValue &Start,
-                       ReductionStyle Style)
+                       ReductionStyle Style,
+                       bool HasUsesOutsideReductionChain = false)
       : VPHeaderPHIRecipe(VPDef::VPReductionPHISC, Phi, &Start), Kind(Kind),
-        Style(Style) {}
+        Style(Style),
+        HasUsesOutsideReductionChain(HasUsesOutsideReductionChain) {}
 
   ~VPReductionPHIRecipe() override = default;
 
   VPReductionPHIRecipe *clone() override {
     auto *R = new VPReductionPHIRecipe(
         dyn_cast_or_null<PHINode>(getUnderlyingValue()), getRecurrenceKind(),
-        *getOperand(0), Style);
+        *getOperand(0), Style, HasUsesOutsideReductionChain);
     R->addOperand(getBackedgeValue());
     return R;
   }
@@ -2481,6 +2492,11 @@ class VPReductionPHIRecipe : public VPHeaderPHIRecipe,
   /// Returns true if the reduction outputs a vector with a scaled down VF.
   bool isPartialReduction() const { return getVFScaleFactor() > 1; }
 
+  /// Returns true, if the phi is part of a multi-use reduction.
+  bool hasUsesOutsideReductionChain() const {
+    return HasUsesOutsideReductionChain;
+  }
+
   /// Returns true if the recipe only uses the first lane of operand \p Op.
   bool usesFirstLaneOnly(const VPValue *Op) const override {
     assert(is_contained(operands(), Op) &&