[LV] Vectorize conditional scalar assignments

Based on Michael Maitland's previous work:
#121222

This PR uses the existing recurrences code instead of introducing a
new pass just for CSA autovec. I've also made recipes that are more
generic.

I've enabled it by default to see the impact on tests; if there are
regressions we can put it behind a cli option.

Author: huntergr-arm

llvm/include/llvm/Analysis/IVDescriptors.h

@@ -70,6 +70,9 @@ enum class RecurKind {
   FindLastIVUMax, ///< FindLast reduction with select(cmp(),x,y) where one of
                   ///< (x,y) is increasing loop induction, and both x and y
                   ///< are integer type, producing a UMax reduction.
+  FindLast,       ///< FindLast reduction with select(cmp(),x,y) where x and y
+                  ///< are an integer type, one is the current recurrence value,
+                  ///< and the other is an arbitrary value.
   // clang-format on
   // TODO: Any_of and FindLast reduction need not be restricted to integer type
   // only.
@@ -175,13 +178,12 @@ class RecurrenceDescriptor {
   /// 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 (FindLast) or decreasing (FindFirst)
-  /// loop induction variable, and the other is a PHI value.
-  // TODO: Support non-monotonic variable. FindLast does not need be restricted
-  // to increasing loop induction variables.
-  LLVM_ABI static InstDesc isFindIVPattern(RecurKind Kind, Loop *TheLoop,
-                                           PHINode *OrigPhi, Instruction *I,
-                                           ScalarEvolution &SE);
+  /// where one of (X, Y) is an increasing (FindLastIV) or decreasing
+  /// (FindFirstIV) loop induction variable, or an arbitrary integer value
+  /// (FindLast), and the other is a PHI value.
+  LLVM_ABI static InstDesc isFindPattern(RecurKind Kind, Loop *TheLoop,
+                                         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.
@@ -305,6 +307,13 @@ class RecurrenceDescriptor {
            isFindLastIVRecurrenceKind(Kind);
   }
 
+  /// Returns true if the recurrence kind is of the form
+  ///   select(cmp(),x,y) where one of (x,y) is an arbitrary value and the
+  ///   other is a recurrence.
+  static bool isFindLastRecurrenceKind(RecurKind Kind) {
+    return Kind == RecurKind::FindLast;
+  }
+
   /// 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/lib/Analysis/IVDescriptors.cpp

@@ -56,6 +56,8 @@ bool RecurrenceDescriptor::isIntegerRecurrenceKind(RecurKind Kind) {
   case RecurKind::FindFirstIVUMin:
   case RecurKind::FindLastIVSMax:
   case RecurKind::FindLastIVUMax:
+  // TODO: Make type-agnostic.
+  case RecurKind::FindLast:
     return true;
   }
   return false;
@@ -691,9 +693,9 @@ RecurrenceDescriptor::isAnyOfPattern(Loop *Loop, PHINode *OrigPhi,
 // value of the data type or a non-constant value by using mask and multiple
 // reduction operations.
 RecurrenceDescriptor::InstDesc
-RecurrenceDescriptor::isFindIVPattern(RecurKind Kind, Loop *TheLoop,
-                                      PHINode *OrigPhi, Instruction *I,
-                                      ScalarEvolution &SE) {
+RecurrenceDescriptor::isFindPattern(RecurKind Kind, Loop *TheLoop,
+                                    PHINode *OrigPhi, Instruction *I,
+                                    ScalarEvolution &SE) {
   // TODO: Support the vectorization of FindLastIV when the reduction phi is
   // used by more than one select instruction. This vectorization is only
   // performed when the SCEV of each increasing induction variable used by the
@@ -702,8 +704,10 @@ RecurrenceDescriptor::isFindIVPattern(RecurKind Kind, Loop *TheLoop,
     return InstDesc(false, I);
 
   // We are looking for selects of the form:
-  //   select(cmp(), phi, loop_induction) or
-  //   select(cmp(), loop_induction, phi)
+  //   select(cmp(), phi, value) or
+  //   select(cmp(), value, phi)
+  // where 'value' might be a loop induction variable
+  // (for FindFirstIV/FindLastIV) or an arbitrary value (for FindLast).
   // TODO: Match selects with multi-use cmp conditions.
   Value *NonRdxPhi = nullptr;
   if (!match(I, m_CombineOr(m_Select(m_OneUse(m_Cmp()), m_Value(NonRdxPhi),
@@ -712,6 +716,25 @@ RecurrenceDescriptor::isFindIVPattern(RecurKind Kind, Loop *TheLoop,
                                      m_Value(NonRdxPhi)))))
     return InstDesc(false, I);
 
+  if (isFindLastRecurrenceKind(Kind)) {
+    // Must be an integer scalar.
+    Type *Type = OrigPhi->getType();
+    if (!Type->isIntegerTy() && !Type->isPointerTy())
+      return InstDesc(false, I);
+
+    // FIXME: Support more complex patterns, including multiple selects.
+    // The Select must be used only outside the loop and by the PHI.
+    for (User *U : I->users()) {
+      if (U == OrigPhi)
+        continue;
+      if (auto *UI = dyn_cast<Instruction>(U); UI && !TheLoop->contains(UI))
+        continue;
+      return InstDesc(false, I);
+    }
+
+    return InstDesc(I, RecurKind::FindLast);
+  }
+
   // Returns either FindFirstIV/FindLastIV, if such a pattern is found, or
   // std::nullopt.
   auto GetRecurKind = [&](Value *V) -> std::optional<RecurKind> {
@@ -920,8 +943,8 @@ RecurrenceDescriptor::InstDesc RecurrenceDescriptor::isRecurrenceInstr(
         Kind == RecurKind::Add || Kind == RecurKind::Mul ||
         Kind == RecurKind::Sub || Kind == RecurKind::AddChainWithSubs)
       return isConditionalRdxPattern(I);
-    if (isFindIVRecurrenceKind(Kind) && SE)
-      return isFindIVPattern(Kind, L, OrigPhi, I, *SE);
+    if ((isFindIVRecurrenceKind(Kind) || isFindLastRecurrenceKind(Kind)) && SE)
+      return isFindPattern(Kind, L, OrigPhi, I, *SE);
     [[fallthrough]];
   case Instruction::FCmp:
   case Instruction::ICmp:
@@ -1118,7 +1141,11 @@ bool RecurrenceDescriptor::isReductionPHI(PHINode *Phi, Loop *TheLoop,
                       << "\n");
     return true;
   }
-
+  if (AddReductionVar(Phi, RecurKind::FindLast, TheLoop, FMF, RedDes, DB, AC,
+                      DT, SE)) {
+    LLVM_DEBUG(dbgs() << "Found a FindLast reduction PHI." << *Phi << "\n");
+    return true;
+  }
   // Not a reduction of known type.
   return false;
 }
@@ -1248,6 +1275,8 @@ unsigned RecurrenceDescriptor::getOpcode(RecurKind Kind) {
   case RecurKind::FMaximumNum:
   case RecurKind::FMinimumNum:
     return Instruction::FCmp;
+  case RecurKind::FindLast:
+    return Instruction::Select;
   default:
     llvm_unreachable("Unknown recurrence operation");
   }

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp

@@ -5373,6 +5373,7 @@ bool AArch64TTIImpl::isLegalToVectorizeReduction(
   case RecurKind::FMax:
   case RecurKind::FMulAdd:
   case RecurKind::AnyOf:
+  case RecurKind::FindLast:
     return true;
   default:
     return false;

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -4559,6 +4559,12 @@ LoopVectorizationPlanner::selectInterleaveCount(VPlan &Plan, ElementCount VF,
       any_of(Plan.getVectorLoopRegion()->getEntryBasicBlock()->phis(),
              IsaPred<VPReductionPHIRecipe>);
 
+  // FIXME: implement interleaving for FindLast transform correctly.
+  for (auto &[_, RdxDesc] : Legal->getReductionVars())
+    if (RecurrenceDescriptor::isFindLastRecurrenceKind(
+            RdxDesc.getRecurrenceKind()))
+      return 1;
+
   // If we did not calculate the cost for VF (because the user selected the VF)
   // then we calculate the cost of VF here.
   if (LoopCost == 0) {
@@ -8475,6 +8481,10 @@ VPlanPtr LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(
                                 *Plan, Builder))
     return nullptr;
 
+  // Create whole-vector selects for find-last recurrences.
+  VPlanTransforms::runPass(VPlanTransforms::convertFindLastRecurrences, *Plan,
+                           RecipeBuilder, Legal);
+
   if (useActiveLaneMask(Style)) {
     // TODO: Move checks to VPlanTransforms::addActiveLaneMask once
     // TailFoldingStyle is visible there.
@@ -8569,6 +8579,7 @@ void LoopVectorizationPlanner::adjustRecipesForReductions(
 
     RecurKind Kind = PhiR->getRecurrenceKind();
     assert(
+        !RecurrenceDescriptor::isFindLastRecurrenceKind(Kind) &&
         !RecurrenceDescriptor::isAnyOfRecurrenceKind(Kind) &&
         !RecurrenceDescriptor::isFindIVRecurrenceKind(Kind) &&
         "AnyOf and FindIV reductions are not allowed for in-loop reductions");
@@ -8872,7 +8883,8 @@ void LoopVectorizationPlanner::adjustRecipesForReductions(
     RecurKind RK = RdxDesc.getRecurrenceKind();
     if ((!RecurrenceDescriptor::isAnyOfRecurrenceKind(RK) &&
          !RecurrenceDescriptor::isFindIVRecurrenceKind(RK) &&
-         !RecurrenceDescriptor::isMinMaxRecurrenceKind(RK))) {
+         !RecurrenceDescriptor::isMinMaxRecurrenceKind(RK) &&
+         !RecurrenceDescriptor::isFindLastRecurrenceKind(RK))) {
       VPBuilder PHBuilder(Plan->getVectorPreheader());
       VPValue *Iden = Plan->getOrAddLiveIn(
           getRecurrenceIdentity(RK, PhiTy, RdxDesc.getFastMathFlags()));
@@ -9996,6 +10008,21 @@ bool LoopVectorizePass::processLoop(Loop *L) {
   // Override IC if user provided an interleave count.
   IC = UserIC > 0 ? UserIC : IC;
 
+  // FIXME: Enable interleaving for last_active reductions.
+  if (any_of(make_second_range(LVL.getReductionVars()), [&](auto &RdxDesc) {
+        return RecurrenceDescriptor::isFindLastRecurrenceKind(
+            RdxDesc.getRecurrenceKind());
+      })) {
+    LLVM_DEBUG(dbgs() << "LV: Not interleaving without vectorization due "
+                      << "to conditional scalar assignments.\n");
+    IntDiagMsg = {
+        "ConditionalAssignmentPreventsScalarInterleaving",
+        "Unable to interleave without vectorization due to conditional "
+        "assignments"};
+    InterleaveLoop = false;
+    IC = 1;
+  }
+
   // Emit diagnostic messages, if any.
   const char *VAPassName = Hints.vectorizeAnalysisPassName();
   if (!VectorizeLoop && !InterleaveLoop) {

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

@@ -25123,6 +25123,7 @@ class HorizontalReduction {
         case RecurKind::FindFirstIVUMin:
         case RecurKind::FindLastIVSMax:
         case RecurKind::FindLastIVUMax:
+        case RecurKind::FindLast:
         case RecurKind::FMaxNum:
         case RecurKind::FMinNum:
         case RecurKind::FMaximumNum:
@@ -25264,6 +25265,7 @@ class HorizontalReduction {
     case RecurKind::FindFirstIVUMin:
     case RecurKind::FindLastIVSMax:
     case RecurKind::FindLastIVUMax:
+    case RecurKind::FindLast:
     case RecurKind::FMaxNum:
     case RecurKind::FMinNum:
     case RecurKind::FMaximumNum:
@@ -25370,6 +25372,7 @@ class HorizontalReduction {
     case RecurKind::FindFirstIVUMin:
     case RecurKind::FindLastIVSMax:
     case RecurKind::FindLastIVUMax:
+    case RecurKind::FindLast:
     case RecurKind::FMaxNum:
     case RecurKind::FMinNum:
     case RecurKind::FMaximumNum:

llvm/lib/Transforms/Vectorize/VPlan.h

@@ -1067,6 +1067,8 @@ class LLVM_ABI_FOR_TEST VPInstruction : public VPRecipeWithIRFlags,
     /// Returns the value for vscale.
     VScale,
     OpsEnd = VScale,
+    /// Extracts the last active lane based on a predicate vector operand.
+    ExtractLastActive,
   };
 
   /// Returns true if this VPInstruction generates scalar values for all lanes.

llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp

@@ -118,7 +118,8 @@ Type *VPTypeAnalysis::inferScalarTypeForRecipe(const VPInstruction *R) {
     return Type::getIntNTy(Ctx, 64);
   case VPInstruction::ExtractLastElement:
   case VPInstruction::ExtractLastLanePerPart:
-  case VPInstruction::ExtractPenultimateElement: {
+  case VPInstruction::ExtractPenultimateElement:
+  case VPInstruction::ExtractLastActive: {
     Type *BaseTy = inferScalarType(R->getOperand(0));
     if (auto *VecTy = dyn_cast<VectorType>(BaseTy))
       return VecTy->getElementType();

llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp

@@ -546,6 +546,7 @@ unsigned VPInstruction::getNumOperandsForOpcode(unsigned Opcode) {
   case VPInstruction::ActiveLaneMask:
   case VPInstruction::ComputeAnyOfResult:
   case VPInstruction::ReductionStartVector:
+  case VPInstruction::ExtractLastActive:
     return 3;
   case VPInstruction::ComputeFindIVResult:
     return 4;
@@ -999,6 +1000,17 @@ Value *VPInstruction::generate(VPTransformState &State) {
   }
   case VPInstruction::ResumeForEpilogue:
     return State.get(getOperand(0), true);
+  case VPInstruction::ExtractLastActive: {
+    Value *Data = State.get(getOperand(0));
+    Value *Mask = State.get(getOperand(1));
+    Value *Default = State.get(getOperand(2), /*IsScalar=*/true);
+    Type *VTy = Data->getType();
+
+    Module *M = State.Builder.GetInsertBlock()->getModule();
+    Function *ExtractLast = Intrinsic::getOrInsertDeclaration(
+        M, Intrinsic::experimental_vector_extract_last_active, {VTy});
+    return Builder.CreateCall(ExtractLast, {Data, Mask, Default});
+  }
   default:
     llvm_unreachable("Unsupported opcode for instruction");
   }
@@ -1135,6 +1147,15 @@ InstructionCost VPInstruction::computeCost(ElementCount VF,
                                   {PredTy, Type::getInt1Ty(Ctx.LLVMCtx)});
     return Ctx.TTI.getIntrinsicInstrCost(Attrs, Ctx.CostKind);
   }
+  case VPInstruction::ExtractLastActive: {
+    Type *ScalarTy = Ctx.Types.inferScalarType(this);
+    Type *VecTy = toVectorTy(ScalarTy, VF);
+    Type *MaskTy = toVectorTy(Type::getInt1Ty(Ctx.LLVMCtx), VF);
+    IntrinsicCostAttributes ICA(
+        Intrinsic::experimental_vector_extract_last_active, ScalarTy,
+        {VecTy, MaskTy, ScalarTy});
+    return Ctx.TTI.getIntrinsicInstrCost(ICA, Ctx.CostKind);
+  }
   case VPInstruction::FirstOrderRecurrenceSplice: {
     assert(VF.isVector() && "Scalar FirstOrderRecurrenceSplice?");
     SmallVector<int> Mask(VF.getKnownMinValue());
@@ -1191,6 +1212,7 @@ bool VPInstruction::isVectorToScalar() const {
          getOpcode() == VPInstruction::FirstActiveLane ||
          getOpcode() == VPInstruction::ComputeAnyOfResult ||
          getOpcode() == VPInstruction::ComputeFindIVResult ||
+         getOpcode() == VPInstruction::ExtractLastActive ||
          getOpcode() == VPInstruction::ComputeReductionResult ||
          getOpcode() == VPInstruction::AnyOf;
 }
@@ -1252,6 +1274,7 @@ bool VPInstruction::opcodeMayReadOrWriteFromMemory() const {
   case VPInstruction::ExtractPenultimateElement:
   case VPInstruction::ActiveLaneMask:
   case VPInstruction::FirstActiveLane:
+  case VPInstruction::ExtractLastActive:
   case VPInstruction::FirstOrderRecurrenceSplice:
   case VPInstruction::LogicalAnd:
   case VPInstruction::Not:
@@ -1437,6 +1460,9 @@ void VPInstruction::print(raw_ostream &O, const Twine &Indent,
   case VPInstruction::Unpack:
     O << "unpack";
     break;
+  case VPInstruction::ExtractLastActive:
+    O << "extract-last-active";
+    break;
   default:
     O << Instruction::getOpcodeName(getOpcode());
   }