Reapply "[LV] Use ExtractLane(LastActiveLane, V) live outs when tail-folding. (#149042)"

llvm/lib/Transforms/Vectorize/LoopVectorizationLegality.cpp

@@ -2095,24 +2095,6 @@ bool LoopVectorizationLegality::canFoldTailByMasking() const {
   for (const auto &Reduction : getReductionVars())
     ReductionLiveOuts.insert(Reduction.second.getLoopExitInstr());
 
-  // TODO: handle non-reduction outside users when tail is folded by masking.
-  for (auto *AE : AllowedExit) {
-    // Check that all users of allowed exit values are inside the loop or
-    // are the live-out of a reduction.
-    if (ReductionLiveOuts.count(AE))
-      continue;
-    for (User *U : AE->users()) {
-      Instruction *UI = cast<Instruction>(U);
-      if (TheLoop->contains(UI))
-        continue;
-      LLVM_DEBUG(
-          dbgs()
-          << "LV: Cannot fold tail by masking, loop has an outside user for "
-          << *UI << "\n");
-      return false;
-    }
-  }
-
   for (const auto &Entry : getInductionVars()) {
     PHINode *OrigPhi = Entry.first;
     for (User *U : OrigPhi->users()) {

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -8914,7 +8914,8 @@ void LoopVectorizationPlanner::adjustRecipesForReductions(
       if (FinalReductionResult == U || Parent->getParent())
         continue;
       U->replaceUsesOfWith(OrigExitingVPV, FinalReductionResult);
-      if (match(U, m_ExtractLastElement(m_VPValue())))
+      if (match(U, m_CombineOr(m_ExtractLastElement(m_VPValue()),
+                               m_ExtractLane(m_VPValue(), m_VPValue()))))
         cast<VPInstruction>(U)->replaceAllUsesWith(FinalReductionResult);
     }
 

llvm/lib/Transforms/Vectorize/VPlan.h

@@ -1095,8 +1095,19 @@ class LLVM_ABI_FOR_TEST VPInstruction : public VPRecipeWithIRFlags,
     AnyOf,
     // Calculates the first active lane index of the vector predicate operands.
     // It produces the lane index across all unrolled iterations. Unrolling will
-    // add all copies of its original operand as additional operands.
+    // add all copies of its original operand as additional operands. If any of
+    // the vector predicate operands are all zeroes then this returns poison.
     FirstActiveLane,
+    // Similar to FirstActiveLane, but if any of the vector predicate operands
+    // are all zeros then the result is not poison.
+    FirstActiveLaneZeroNotPoison,
+    // Calculates the last active lane index of the vector predicate operands.
+    // The predicates must be prefix-masks (all 1s before all 0s). Used when
+    // tail-folding to extract the correct live-out value from the last active
+    // iteration. It produces the lane index across all unrolled iterations.
+    // Unrolling will add all copies of its original operand as additional
+    // operands.
+    LastActiveLane,
 
     // The opcodes below are used for VPInstructionWithType.
     //

llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp

@@ -115,6 +115,8 @@ Type *VPTypeAnalysis::inferScalarTypeForRecipe(const VPInstruction *R) {
   case VPInstruction::ExtractLane:
     return inferScalarType(R->getOperand(1));
   case VPInstruction::FirstActiveLane:
+  case VPInstruction::FirstActiveLaneZeroNotPoison:
+  case VPInstruction::LastActiveLane:
     return Type::getIntNTy(Ctx, 64);
   case VPInstruction::ExtractLastElement:
   case VPInstruction::ExtractLastLanePerPart:

llvm/lib/Transforms/Vectorize/VPlanPatternMatch.h

@@ -399,12 +399,24 @@ m_ExtractElement(const Op0_t &Op0, const Op1_t &Op1) {
   return m_VPInstruction<Instruction::ExtractElement>(Op0, Op1);
 }
 
+template <typename Op0_t, typename Op1_t>
+inline VPInstruction_match<VPInstruction::ExtractLane, Op0_t, Op1_t>
+m_ExtractLane(const Op0_t &Op0, const Op1_t &Op1) {
+  return m_VPInstruction<VPInstruction::ExtractLane>(Op0, Op1);
+}
+
 template <typename Op0_t>
 inline VPInstruction_match<VPInstruction::ExtractLastLanePerPart, Op0_t>
 m_ExtractLastLanePerPart(const Op0_t &Op0) {
   return m_VPInstruction<VPInstruction::ExtractLastLanePerPart>(Op0);
 }
 
+template <typename Op0_t>
+inline VPInstruction_match<VPInstruction::ExtractPenultimateElement, Op0_t>
+m_ExtractPenultimateElement(const Op0_t &Op0) {
+  return m_VPInstruction<VPInstruction::ExtractPenultimateElement>(Op0);
+}
+
 template <typename Op0_t, typename Op1_t, typename Op2_t>
 inline VPInstruction_match<VPInstruction::ActiveLaneMask, Op0_t, Op1_t, Op2_t>
 m_ActiveLaneMask(const Op0_t &Op0, const Op1_t &Op1, const Op2_t &Op2) {
@@ -433,6 +445,16 @@ m_FirstActiveLane(const Op0_t &Op0) {
   return m_VPInstruction<VPInstruction::FirstActiveLane>(Op0);
 }
 
+template <typename Op0_t>
+inline VPInstruction_match<VPInstruction::LastActiveLane, Op0_t>
+m_LastActiveLane(const Op0_t &Op0) {
+  return m_VPInstruction<VPInstruction::LastActiveLane>(Op0);
+}
+
+inline VPInstruction_match<VPInstruction::StepVector> m_StepVector() {
+  return m_VPInstruction<VPInstruction::StepVector>();
+}
+
 template <unsigned Opcode, typename Op0_t>
 inline AllRecipe_match<Opcode, Op0_t> m_Unary(const Op0_t &Op0) {
   return AllRecipe_match<Opcode, Op0_t>(Op0);

llvm/lib/Transforms/Vectorize/VPlanPredicator.cpp

@@ -44,11 +44,6 @@ class VPPredicator {
   /// possibly inserting new recipes at \p Dst (using Builder's insertion point)
   VPValue *createEdgeMask(VPBasicBlock *Src, VPBasicBlock *Dst);
 
-  /// Returns the *entry* mask for \p VPBB.
-  VPValue *getBlockInMask(VPBasicBlock *VPBB) const {
-    return BlockMaskCache.lookup(VPBB);
-  }
-
   /// Record \p Mask as the *entry* mask of \p VPBB, which is expected to not
   /// already have a mask.
   void setBlockInMask(VPBasicBlock *VPBB, VPValue *Mask) {
@@ -68,6 +63,11 @@ class VPPredicator {
   }
 
 public:
+  /// Returns the *entry* mask for \p VPBB.
+  VPValue *getBlockInMask(VPBasicBlock *VPBB) const {
+    return BlockMaskCache.lookup(VPBB);
+  }
+
   /// Returns the precomputed predicate of the edge from \p Src to \p Dst.
   VPValue *getEdgeMask(const VPBasicBlock *Src, const VPBasicBlock *Dst) const {
     return EdgeMaskCache.lookup({Src, Dst});
@@ -301,5 +301,34 @@ VPlanTransforms::introduceMasksAndLinearize(VPlan &Plan, bool FoldTail) {
 
     PrevVPBB = VPBB;
   }
+
+  // If we folded the tail and introduced a header mask, any extract of the
+  // last element must be updated to extract from the last active lane of the
+  // header mask instead (i.e., the lane corresponding to the last active
+  // iteration).
+  if (FoldTail) {
+    assert(Plan.getExitBlocks().size() == 1 &&
+           "only a single-exit block is supported currently");
+    VPBasicBlock *EB = Plan.getExitBlocks().front();
+    assert(EB->getSinglePredecessor() == Plan.getMiddleBlock() &&
+           "the exit block must have middle block as single predecessor");
+
+    VPBuilder B(Plan.getMiddleBlock()->getTerminator());
+    for (auto &P : EB->phis()) {
+      auto *ExitIRI = cast<VPIRPhi>(&P);
+      VPValue *Inc = ExitIRI->getIncomingValue(0);
+      VPValue *Op;
+      if (!match(Inc, m_ExtractLastElement(m_VPValue(Op))))
+        continue;
+
+      // Compute the index of the last active lane.
+      VPValue *HeaderMask = Predicator.getBlockInMask(Header);
+      VPValue *LastActiveLane =
+          B.createNaryOp(VPInstruction::LastActiveLane, HeaderMask);
+      auto *Ext =
+          B.createNaryOp(VPInstruction::ExtractLane, {LastActiveLane, Op});
+      Inc->replaceAllUsesWith(Ext);
+    }
+  }
   return Predicator.getBlockMaskCache();
 }

llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp

@@ -560,7 +560,6 @@ unsigned VPInstruction::getNumOperandsForOpcode(unsigned Opcode) {
   case VPInstruction::ExtractLastElement:
   case VPInstruction::ExtractLastLanePerPart:
   case VPInstruction::ExtractPenultimateElement:
-  case VPInstruction::FirstActiveLane:
   case VPInstruction::Not:
   case VPInstruction::ResumeForEpilogue:
   case VPInstruction::Unpack:
@@ -591,6 +590,9 @@ unsigned VPInstruction::getNumOperandsForOpcode(unsigned Opcode) {
   case Instruction::Switch:
   case VPInstruction::SLPLoad:
   case VPInstruction::SLPStore:
+  case VPInstruction::FirstActiveLane:
+  case VPInstruction::FirstActiveLaneZeroNotPoison:
+  case VPInstruction::LastActiveLane:
     // Cannot determine the number of operands from the opcode.
     return -1u;
   }
@@ -1001,11 +1003,16 @@ Value *VPInstruction::generate(VPTransformState &State) {
     }
     return Res;
   }
-  case VPInstruction::FirstActiveLane: {
+  case VPInstruction::FirstActiveLane:
+  case VPInstruction::FirstActiveLaneZeroNotPoison: {
+    const bool ZeroIsPoison =
+        getOpcode() != VPInstruction::FirstActiveLaneZeroNotPoison;
     if (getNumOperands() == 1) {
       Value *Mask = State.get(getOperand(0));
+      // LastActiveLane might get expanded to a FirstActiveLane with an all-ones
+      // mask, so make sure zero returns VF and not poison.
       return Builder.CreateCountTrailingZeroElems(Builder.getInt64Ty(), Mask,
-                                                  true, Name);
+                                                  ZeroIsPoison, Name);
     }
     // If there are multiple operands, create a chain of selects to pick the
     // first operand with an active lane and add the number of lanes of the
@@ -1023,7 +1030,7 @@ Value *VPInstruction::generate(VPTransformState &State) {
                     Builder.getInt64Ty())
               : Builder.CreateCountTrailingZeroElems(Builder.getInt64Ty(),
                                                      State.get(getOperand(Idx)),
-                                                     true, Name);
+                                                     ZeroIsPoison, Name);
       Value *Current = Builder.CreateAdd(
           Builder.CreateMul(RuntimeVF, Builder.getInt64(Idx)), TrailingZeros);
       if (Res) {
@@ -1161,7 +1168,8 @@ InstructionCost VPInstruction::computeCost(ElementCount VF,
     return Ctx.TTI.getArithmeticReductionCost(
         Instruction::Or, cast<VectorType>(VecTy), std::nullopt, Ctx.CostKind);
   }
-  case VPInstruction::FirstActiveLane: {
+  case VPInstruction::FirstActiveLane:
+  case VPInstruction::FirstActiveLaneZeroNotPoison: {
     Type *ScalarTy = Ctx.Types.inferScalarType(getOperand(0));
     if (VF.isScalar())
       return Ctx.TTI.getCmpSelInstrCost(Instruction::ICmp, ScalarTy,
@@ -1174,6 +1182,29 @@ InstructionCost VPInstruction::computeCost(ElementCount VF,
                                   {PredTy, Type::getInt1Ty(Ctx.LLVMCtx)});
     return Ctx.TTI.getIntrinsicInstrCost(Attrs, Ctx.CostKind);
   }
+  case VPInstruction::LastActiveLane: {
+    Type *ScalarTy = Ctx.Types.inferScalarType(getOperand(0));
+    if (VF.isScalar())
+      return Ctx.TTI.getCmpSelInstrCost(Instruction::ICmp, ScalarTy,
+                                        CmpInst::makeCmpResultType(ScalarTy),
+                                        CmpInst::ICMP_EQ, Ctx.CostKind);
+    // Calculate the cost of determining the lane index: NOT + cttz_elts + SUB.
+    auto *PredTy = toVectorTy(ScalarTy, VF);
+    IntrinsicCostAttributes Attrs(Intrinsic::experimental_cttz_elts,
+                                  Type::getInt64Ty(Ctx.LLVMCtx),
+                                  {PredTy, Type::getInt1Ty(Ctx.LLVMCtx)});
+    InstructionCost Cost = Ctx.TTI.getIntrinsicInstrCost(Attrs, Ctx.CostKind);
+    // Add cost of NOT operation on the predicate.
+    Cost += Ctx.TTI.getArithmeticInstrCost(
+        Instruction::Xor, PredTy, Ctx.CostKind,
+        {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None},
+        {TargetTransformInfo::OK_UniformConstantValue,
+         TargetTransformInfo::OP_None});
+    // Add cost of SUB operation on the index.
+    Cost += Ctx.TTI.getArithmeticInstrCost(
+        Instruction::Sub, Type::getInt64Ty(Ctx.LLVMCtx), Ctx.CostKind);
+    return Cost;
+  }
   case VPInstruction::FirstOrderRecurrenceSplice: {
     assert(VF.isVector() && "Scalar FirstOrderRecurrenceSplice?");
     SmallVector<int> Mask(VF.getKnownMinValue());
@@ -1228,6 +1259,8 @@ bool VPInstruction::isVectorToScalar() const {
          getOpcode() == Instruction::ExtractElement ||
          getOpcode() == VPInstruction::ExtractLane ||
          getOpcode() == VPInstruction::FirstActiveLane ||
+         getOpcode() == VPInstruction::FirstActiveLaneZeroNotPoison ||
+         getOpcode() == VPInstruction::LastActiveLane ||
          getOpcode() == VPInstruction::ComputeAnyOfResult ||
          getOpcode() == VPInstruction::ComputeFindIVResult ||
          getOpcode() == VPInstruction::ComputeReductionResult ||
@@ -1294,6 +1327,8 @@ bool VPInstruction::opcodeMayReadOrWriteFromMemory() const {
   case VPInstruction::ActiveLaneMask:
   case VPInstruction::ExplicitVectorLength:
   case VPInstruction::FirstActiveLane:
+  case VPInstruction::FirstActiveLaneZeroNotPoison:
+  case VPInstruction::LastActiveLane:
   case VPInstruction::FirstOrderRecurrenceSplice:
   case VPInstruction::LogicalAnd:
   case VPInstruction::Not:
@@ -1470,6 +1505,12 @@ void VPInstruction::printRecipe(raw_ostream &O, const Twine &Indent,
   case VPInstruction::FirstActiveLane:
     O << "first-active-lane";
     break;
+  case VPInstruction::FirstActiveLaneZeroNotPoison:
+    O << "first-active-lane-zero-not-poison";
+    break;
+  case VPInstruction::LastActiveLane:
+    O << "last-active-lane";
+    break;
   case VPInstruction::ReductionStartVector:
     O << "reduction-start-vector";
     break;

llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp

@@ -826,8 +826,8 @@ static VPValue *optimizeEarlyExitInductionUser(VPlan &Plan,
                                                VPValue *Op,
                                                ScalarEvolution &SE) {
   VPValue *Incoming, *Mask;
-  if (!match(Op, m_VPInstruction<VPInstruction::ExtractLane>(
-                     m_FirstActiveLane(m_VPValue(Mask)), m_VPValue(Incoming))))
+  if (!match(Op, m_ExtractLane(m_FirstActiveLane(m_VPValue(Mask)),
+                               m_VPValue(Incoming))))
     return nullptr;
 
   auto *WideIV = getOptimizableIVOf(Incoming, SE);
@@ -1295,8 +1295,7 @@ static void simplifyRecipe(VPSingleDefRecipe *Def, VPTypeAnalysis &TypeInfo) {
   }
 
   // Look through ExtractPenultimateElement (BuildVector ....).
-  if (match(Def, m_VPInstruction<VPInstruction::ExtractPenultimateElement>(
-                     m_BuildVector()))) {
+  if (match(Def, m_ExtractPenultimateElement(m_BuildVector()))) {
     auto *BuildVector = cast<VPInstruction>(Def->getOperand(0));
     Def->replaceAllUsesWith(
         BuildVector->getOperand(BuildVector->getNumOperands() - 2));
@@ -2106,6 +2105,32 @@ bool VPlanTransforms::adjustFixedOrderRecurrences(VPlan &Plan,
     // Set the first operand of RecurSplice to FOR again, after replacing
     // all users.
     RecurSplice->setOperand(0, FOR);
+
+    // Check for users extracting at the penultimate active lane of the FOR.
+    // If only a single lane is active in the current iteration, we need to
+    // select the last element from the previous iteration (from the FOR phi
+    // directly).
+    for (VPUser *U : RecurSplice->users()) {
+      if (!match(U, m_ExtractLane(m_LastActiveLane(m_VPValue()),
+                                  m_Specific(RecurSplice))))
+        continue;
+
+      VPBuilder B(cast<VPInstruction>(U));
+      VPValue *LastActiveLane = cast<VPInstruction>(U)->getOperand(0);
+      Type *I64Ty = Type::getInt64Ty(Plan.getContext());
+      VPValue *Zero = Plan.getOrAddLiveIn(ConstantInt::get(I64Ty, 0));
+      VPValue *One = Plan.getOrAddLiveIn(ConstantInt::get(I64Ty, 1));
+      VPValue *PenultimateIndex =
+          B.createNaryOp(Instruction::Sub, {LastActiveLane, One});
+      VPValue *PenultimateLastIter =
+          B.createNaryOp(VPInstruction::ExtractLane,
+                         {PenultimateIndex, FOR->getBackedgeValue()});
+      VPValue *LastPrevIter =
+          B.createNaryOp(VPInstruction::ExtractLastElement, FOR);
+      VPValue *Cmp = B.createICmp(CmpInst::ICMP_EQ, LastActiveLane, Zero);
+      VPValue *Sel = B.createSelect(Cmp, LastPrevIter, PenultimateLastIter);
+      cast<VPInstruction>(U)->replaceAllUsesWith(Sel);
+    }
   }
   return true;
 }
@@ -3492,6 +3517,34 @@ void VPlanTransforms::convertToConcreteRecipes(VPlan &Plan) {
         ToRemove.push_back(Expr);
       }
 
+      // Expand LastActiveLane into Not + FirstActiveLaneZeroNotPoison + Sub.
+      auto *LastActiveL = dyn_cast<VPInstruction>(&R);
+      if (LastActiveL &&
+          LastActiveL->getOpcode() == VPInstruction::LastActiveLane) {
+        // Create Not(Mask) for all operands.
+        SmallVector<VPValue *, 2> NotMasks;
+        for (VPValue *Op : LastActiveL->operands()) {
+          VPValue *NotMask = Builder.createNot(Op, LastActiveL->getDebugLoc());
+          NotMasks.push_back(NotMask);
+        }
+
+        // Create FirstActiveLaneZeroNotPoison on the inverted masks.
+        VPValue *FirstInactiveLane = Builder.createNaryOp(
+            VPInstruction::FirstActiveLaneZeroNotPoison, NotMasks,
+            LastActiveL->getDebugLoc(), "first.inactive.lane");
+
+        // Subtract 1 to get the last active lane.
+        VPValue *One = Plan.getOrAddLiveIn(
+            ConstantInt::get(Type::getInt64Ty(Plan.getContext()), 1));
+        VPValue *LastLane = Builder.createNaryOp(
+            Instruction::Sub, {FirstInactiveLane, One},
+            LastActiveL->getDebugLoc(), "last.active.lane");
+
+        LastActiveL->replaceAllUsesWith(LastLane);
+        ToRemove.push_back(LastActiveL);
+        continue;
+      }
+
       VPValue *VectorStep;
       VPValue *ScalarStep;
       if (!match(&R, m_VPInstruction<VPInstruction::WideIVStep>(