[LV] Vectorize conditional scalar assignments

Based on Michael Maitland's previous work:
llvm#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' is 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())
+      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;
 }
@@ -1243,6 +1270,8 @@ unsigned RecurrenceDescriptor::getOpcode(RecurKind Kind) {
   case RecurKind::FMaximumNum:
   case RecurKind::FMinimumNum:
     return Instruction::FCmp;
+  case RecurKind::FindLast:
+    return Instruction::Select;
   case RecurKind::AnyOf:
   case RecurKind::FindFirstIVSMin:
   case RecurKind::FindFirstIVUMin:

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp

@@ -5471,6 +5471,7 @@ 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

@@ -1293,6 +1293,7 @@ class LoopVectorizationCostModel {
                            "from latch block\n");
       return true;
     }
+
     if (IsVectorizing && InterleaveInfo.requiresScalarEpilogue()) {
       LLVM_DEBUG(dbgs() << "LV: Loop requires scalar epilogue: "
                            "interleaved group requires scalar epilogue\n");
@@ -4084,6 +4085,7 @@ static bool willGenerateVectors(VPlan &Plan, ElementCount VF,
         continue;
       case VPDef::VPReductionSC:
       case VPDef::VPActiveLaneMaskPHISC:
+      case VPDef::VPLastActiveMaskPHISC:
       case VPDef::VPWidenCallSC:
       case VPDef::VPWidenCanonicalIVSC:
       case VPDef::VPWidenCastSC:
@@ -4302,11 +4304,15 @@ bool LoopVectorizationPlanner::isCandidateForEpilogueVectorization(
     ElementCount VF) const {
   // Cross iteration phis such as fixed-order recurrences and FMaxNum/FMinNum
   // reductions need special handling and are currently unsupported.
+  // FindLast reductions also require special handling for the synthesized
+  // mask PHI.
   if (any_of(OrigLoop->getHeader()->phis(), [&](PHINode &Phi) {
         if (!Legal->isReductionVariable(&Phi))
           return Legal->isFixedOrderRecurrence(&Phi);
-        return RecurrenceDescriptor::isFPMinMaxNumRecurrenceKind(
-            Legal->getRecurrenceDescriptor(&Phi).getRecurrenceKind());
+        RecurKind Kind =
+            Legal->getRecurrenceDescriptor(&Phi).getRecurrenceKind();
+        return RecurrenceDescriptor::isFindLastRecurrenceKind(Kind) ||
+               RecurrenceDescriptor::isFPMinMaxNumRecurrenceKind(Kind);
       }))
     return false;
 
@@ -4612,6 +4618,12 @@ 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) {
@@ -8566,6 +8578,10 @@ 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.
@@ -8660,10 +8676,11 @@ void LoopVectorizationPlanner::adjustRecipesForReductions(
       continue;
 
     RecurKind Kind = PhiR->getRecurrenceKind();
-    assert(
-        !RecurrenceDescriptor::isAnyOfRecurrenceKind(Kind) &&
-        !RecurrenceDescriptor::isFindIVRecurrenceKind(Kind) &&
-        "AnyOf and FindIV reductions are not allowed for in-loop reductions");
+    assert(!RecurrenceDescriptor::isFindLastRecurrenceKind(Kind) &&
+           !RecurrenceDescriptor::isAnyOfRecurrenceKind(Kind) &&
+           !RecurrenceDescriptor::isFindIVRecurrenceKind(Kind) &&
+           "AnyOf, FindIV, 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;
@@ -8960,7 +8977,8 @@ 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()));
@@ -9362,7 +9380,7 @@ static void preparePlanForMainVectorLoop(VPlan &MainPlan, VPlan &EpiPlan) {
   SmallPtrSet<PHINode *, 2> EpiWidenedPhis;
   for (VPRecipeBase &R :
        EpiPlan.getVectorLoopRegion()->getEntryBasicBlock()->phis()) {
-    if (isa<VPCanonicalIVPHIRecipe>(&R))
+    if (isa<VPCanonicalIVPHIRecipe, VPLastActiveMaskPHIRecipe>(&R))
       continue;
     EpiWidenedPhis.insert(
         cast<PHINode>(R.getVPSingleValue()->getUnderlyingValue()));
@@ -9559,6 +9577,10 @@ static SmallVector<Instruction *> preparePlanForEpilogueVectorLoop(
           continue;
         }
       }
+    } else if (isa<VPLastActiveMaskPHIRecipe>(R)) {
+      // LastActiveMasks are only used as part of FindLast reductions,
+      // and aren't passed to the scalar loop.
+      continue;
     } else {
       // Retrieve the induction resume values for wide inductions from
       // their original phi nodes in the scalar loop.
@@ -10080,6 +10102,21 @@ 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

@@ -25284,6 +25284,7 @@ class HorizontalReduction {
         case RecurKind::FindFirstIVUMin:
         case RecurKind::FindLastIVSMax:
         case RecurKind::FindLastIVUMax:
+        case RecurKind::FindLast:
         case RecurKind::FMaxNum:
         case RecurKind::FMinNum:
         case RecurKind::FMaximumNum:
@@ -25425,6 +25426,7 @@ class HorizontalReduction {
     case RecurKind::FindFirstIVUMin:
     case RecurKind::FindLastIVSMax:
     case RecurKind::FindLastIVUMax:
+    case RecurKind::FindLast:
     case RecurKind::FMaxNum:
     case RecurKind::FMinNum:
     case RecurKind::FMaximumNum:
@@ -25531,6 +25533,7 @@ 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

@@ -561,6 +561,7 @@ class VPSingleDefRecipe : public VPRecipeBase, public VPValue {
     case VPRecipeBase::VPPredInstPHISC:
     case VPRecipeBase::VPCanonicalIVPHISC:
     case VPRecipeBase::VPActiveLaneMaskPHISC:
+    case VPRecipeBase::VPLastActiveMaskPHISC:
     case VPRecipeBase::VPFirstOrderRecurrencePHISC:
     case VPRecipeBase::VPWidenPHISC:
     case VPRecipeBase::VPWidenIntOrFpInductionSC:
@@ -1119,6 +1120,8 @@ 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.
@@ -3652,6 +3655,40 @@ class VPActiveLaneMaskPHIRecipe : public VPHeaderPHIRecipe {
 #endif
 };
 
+// TODO: Can we unify the PHI recipe hierarchy a bit? VPPredInstPHISC is close
+//       to this (just a PHI of a predicate), but isn't a header phi so can't
+//       be used for the mask of FindLastActive reductions.
+//
+// This is basically a clone of VPActiveLaneMaskPHIRecipe, but won't run into
+// problems with transforms that expect there to only be a single ALM PHI, and
+// can be ignored by other code looking for a (non-existent) underlying value.
+class VPLastActiveMaskPHIRecipe : public VPHeaderPHIRecipe {
+public:
+  VPLastActiveMaskPHIRecipe(VPValue *StartMask, DebugLoc DL)
+      : VPHeaderPHIRecipe(VPDef::VPLastActiveMaskPHISC, nullptr, StartMask,
+                          DL) {}
+
+  ~VPLastActiveMaskPHIRecipe() override = default;
+
+  VPLastActiveMaskPHIRecipe *clone() override {
+    auto *R = new VPLastActiveMaskPHIRecipe(getOperand(0), getDebugLoc());
+    if (getNumOperands() == 2)
+      R->addOperand(getOperand(1));
+    return R;
+  }
+
+  VP_CLASSOF_IMPL(VPDef::VPLastActiveMaskPHISC);
+
+  /// Generate the mask phi
+  void execute(VPTransformState &State) override;
+
+#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
+  /// Print the recipe
+  void print(raw_ostream &O, const Twine &Indent,
+             VPSlotTracker &SlotTracker) const override;
+#endif
+};
+
 /// A recipe for generating the phi node for the current index of elements,
 /// adjusted in accordance with EVL value. It starts at the start value of the
 /// canonical induction and gets incremented by EVL in each iteration of the

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();
@@ -276,14 +277,14 @@ Type *VPTypeAnalysis::inferScalarType(const VPValue *V) {
       TypeSwitch<const VPRecipeBase *, Type *>(V->getDefiningRecipe())
           .Case<VPActiveLaneMaskPHIRecipe, VPCanonicalIVPHIRecipe,
                 VPFirstOrderRecurrencePHIRecipe, VPReductionPHIRecipe,
-                VPWidenPointerInductionRecipe, VPEVLBasedIVPHIRecipe>(
-              [this](const auto *R) {
-                // Handle header phi recipes, except VPWidenIntOrFpInduction
-                // which needs special handling due it being possibly truncated.
-                // TODO: consider inferring/caching type of siblings, e.g.,
-                // backedge value, here and in cases below.
-                return inferScalarType(R->getStartValue());
-              })
+                VPWidenPointerInductionRecipe, VPEVLBasedIVPHIRecipe,
+                VPLastActiveMaskPHIRecipe>([this](const auto *R) {
+            // Handle header phi recipes, except VPWidenIntOrFpInduction
+            // which needs special handling due it being possibly truncated.
+            // TODO: consider inferring/caching type of siblings, e.g.,
+            // backedge value, here and in cases below.
+            return inferScalarType(R->getStartValue());
+          })
           .Case<VPWidenIntOrFpInductionRecipe, VPDerivedIVRecipe>(
               [](const auto *R) { return R->getScalarType(); })
           .Case<VPReductionRecipe, VPPredInstPHIRecipe, VPWidenPHIRecipe,