Home | History | Annotate | Line # | Download | only in Analysis
      1 //===- llvm/Analysis/IVDescriptors.cpp - IndVar Descriptors -----*- C++ -*-===//
      2 //
      3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
      4 // See https://llvm.org/LICENSE.txt for license information.
      5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
      6 //
      7 //===----------------------------------------------------------------------===//
      8 //
      9 // This file "describes" induction and recurrence variables.
     10 //
     11 //===----------------------------------------------------------------------===//
     12 
     13 #include "llvm/Analysis/IVDescriptors.h"
     14 #include "llvm/ADT/ScopeExit.h"
     15 #include "llvm/Analysis/BasicAliasAnalysis.h"
     16 #include "llvm/Analysis/DemandedBits.h"
     17 #include "llvm/Analysis/DomTreeUpdater.h"
     18 #include "llvm/Analysis/GlobalsModRef.h"
     19 #include "llvm/Analysis/InstructionSimplify.h"
     20 #include "llvm/Analysis/LoopInfo.h"
     21 #include "llvm/Analysis/LoopPass.h"
     22 #include "llvm/Analysis/MustExecute.h"
     23 #include "llvm/Analysis/ScalarEvolution.h"
     24 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
     25 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
     26 #include "llvm/Analysis/TargetTransformInfo.h"
     27 #include "llvm/Analysis/ValueTracking.h"
     28 #include "llvm/IR/Dominators.h"
     29 #include "llvm/IR/Instructions.h"
     30 #include "llvm/IR/Module.h"
     31 #include "llvm/IR/PatternMatch.h"
     32 #include "llvm/IR/ValueHandle.h"
     33 #include "llvm/Pass.h"
     34 #include "llvm/Support/Debug.h"
     35 #include "llvm/Support/KnownBits.h"
     36 
     37 using namespace llvm;
     38 using namespace llvm::PatternMatch;
     39 
     40 #define DEBUG_TYPE "iv-descriptors"
     41 
     42 bool RecurrenceDescriptor::areAllUsesIn(Instruction *I,
     43                                         SmallPtrSetImpl<Instruction *> &Set) {
     44   for (User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
     45     if (!Set.count(dyn_cast<Instruction>(*Use)))
     46       return false;
     47   return true;
     48 }
     49 
     50 bool RecurrenceDescriptor::isIntegerRecurrenceKind(RecurKind Kind) {
     51   switch (Kind) {
     52   default:
     53     break;
     54   case RecurKind::Add:
     55   case RecurKind::Mul:
     56   case RecurKind::Or:
     57   case RecurKind::And:
     58   case RecurKind::Xor:
     59   case RecurKind::SMax:
     60   case RecurKind::SMin:
     61   case RecurKind::UMax:
     62   case RecurKind::UMin:
     63     return true;
     64   }
     65   return false;
     66 }
     67 
     68 bool RecurrenceDescriptor::isFloatingPointRecurrenceKind(RecurKind Kind) {
     69   return (Kind != RecurKind::None) && !isIntegerRecurrenceKind(Kind);
     70 }
     71 
     72 bool RecurrenceDescriptor::isArithmeticRecurrenceKind(RecurKind Kind) {
     73   switch (Kind) {
     74   default:
     75     break;
     76   case RecurKind::Add:
     77   case RecurKind::Mul:
     78   case RecurKind::FAdd:
     79   case RecurKind::FMul:
     80     return true;
     81   }
     82   return false;
     83 }
     84 
     85 /// Determines if Phi may have been type-promoted. If Phi has a single user
     86 /// that ANDs the Phi with a type mask, return the user. RT is updated to
     87 /// account for the narrower bit width represented by the mask, and the AND
     88 /// instruction is added to CI.
     89 static Instruction *lookThroughAnd(PHINode *Phi, Type *&RT,
     90                                    SmallPtrSetImpl<Instruction *> &Visited,
     91                                    SmallPtrSetImpl<Instruction *> &CI) {
     92   if (!Phi->hasOneUse())
     93     return Phi;
     94 
     95   const APInt *M = nullptr;
     96   Instruction *I, *J = cast<Instruction>(Phi->use_begin()->getUser());
     97 
     98   // Matches either I & 2^x-1 or 2^x-1 & I. If we find a match, we update RT
     99   // with a new integer type of the corresponding bit width.
    100   if (match(J, m_c_And(m_Instruction(I), m_APInt(M)))) {
    101     int32_t Bits = (*M + 1).exactLogBase2();
    102     if (Bits > 0) {
    103       RT = IntegerType::get(Phi->getContext(), Bits);
    104       Visited.insert(Phi);
    105       CI.insert(J);
    106       return J;
    107     }
    108   }
    109   return Phi;
    110 }
    111 
    112 /// Compute the minimal bit width needed to represent a reduction whose exit
    113 /// instruction is given by Exit.
    114 static std::pair<Type *, bool> computeRecurrenceType(Instruction *Exit,
    115                                                      DemandedBits *DB,
    116                                                      AssumptionCache *AC,
    117                                                      DominatorTree *DT) {
    118   bool IsSigned = false;
    119   const DataLayout &DL = Exit->getModule()->getDataLayout();
    120   uint64_t MaxBitWidth = DL.getTypeSizeInBits(Exit->getType());
    121 
    122   if (DB) {
    123     // Use the demanded bits analysis to determine the bits that are live out
    124     // of the exit instruction, rounding up to the nearest power of two. If the
    125     // use of demanded bits results in a smaller bit width, we know the value
    126     // must be positive (i.e., IsSigned = false), because if this were not the
    127     // case, the sign bit would have been demanded.
    128     auto Mask = DB->getDemandedBits(Exit);
    129     MaxBitWidth = Mask.getBitWidth() - Mask.countLeadingZeros();
    130   }
    131 
    132   if (MaxBitWidth == DL.getTypeSizeInBits(Exit->getType()) && AC && DT) {
    133     // If demanded bits wasn't able to limit the bit width, we can try to use
    134     // value tracking instead. This can be the case, for example, if the value
    135     // may be negative.
    136     auto NumSignBits = ComputeNumSignBits(Exit, DL, 0, AC, nullptr, DT);
    137     auto NumTypeBits = DL.getTypeSizeInBits(Exit->getType());
    138     MaxBitWidth = NumTypeBits - NumSignBits;
    139     KnownBits Bits = computeKnownBits(Exit, DL);
    140     if (!Bits.isNonNegative()) {
    141       // If the value is not known to be non-negative, we set IsSigned to true,
    142       // meaning that we will use sext instructions instead of zext
    143       // instructions to restore the original type.
    144       IsSigned = true;
    145       if (!Bits.isNegative())
    146         // If the value is not known to be negative, we don't known what the
    147         // upper bit is, and therefore, we don't know what kind of extend we
    148         // will need. In this case, just increase the bit width by one bit and
    149         // use sext.
    150         ++MaxBitWidth;
    151     }
    152   }
    153   if (!isPowerOf2_64(MaxBitWidth))
    154     MaxBitWidth = NextPowerOf2(MaxBitWidth);
    155 
    156   return std::make_pair(Type::getIntNTy(Exit->getContext(), MaxBitWidth),
    157                         IsSigned);
    158 }
    159 
    160 /// Collect cast instructions that can be ignored in the vectorizer's cost
    161 /// model, given a reduction exit value and the minimal type in which the
    162 /// reduction can be represented.
    163 static void collectCastsToIgnore(Loop *TheLoop, Instruction *Exit,
    164                                  Type *RecurrenceType,
    165                                  SmallPtrSetImpl<Instruction *> &Casts) {
    166 
    167   SmallVector<Instruction *, 8> Worklist;
    168   SmallPtrSet<Instruction *, 8> Visited;
    169   Worklist.push_back(Exit);
    170 
    171   while (!Worklist.empty()) {
    172     Instruction *Val = Worklist.pop_back_val();
    173     Visited.insert(Val);
    174     if (auto *Cast = dyn_cast<CastInst>(Val))
    175       if (Cast->getSrcTy() == RecurrenceType) {
    176         // If the source type of a cast instruction is equal to the recurrence
    177         // type, it will be eliminated, and should be ignored in the vectorizer
    178         // cost model.
    179         Casts.insert(Cast);
    180         continue;
    181       }
    182 
    183     // Add all operands to the work list if they are loop-varying values that
    184     // we haven't yet visited.
    185     for (Value *O : cast<User>(Val)->operands())
    186       if (auto *I = dyn_cast<Instruction>(O))
    187         if (TheLoop->contains(I) && !Visited.count(I))
    188           Worklist.push_back(I);
    189   }
    190 }
    191 
    192 // Check if a given Phi node can be recognized as an ordered reduction for
    193 // vectorizing floating point operations without unsafe math.
    194 static bool checkOrderedReduction(RecurKind Kind, Instruction *ExactFPMathInst,
    195                                   Instruction *Exit, PHINode *Phi) {
    196   // Currently only FAdd is supported
    197   if (Kind != RecurKind::FAdd)
    198     return false;
    199 
    200   bool IsOrdered =
    201       Exit->getOpcode() == Instruction::FAdd && Exit == ExactFPMathInst;
    202 
    203   // The only pattern accepted is the one in which the reduction PHI
    204   // is used as one of the operands of the exit instruction
    205   auto *LHS = Exit->getOperand(0);
    206   auto *RHS = Exit->getOperand(1);
    207   IsOrdered &= ((LHS == Phi) || (RHS == Phi));
    208 
    209   if (!IsOrdered)
    210     return false;
    211 
    212   LLVM_DEBUG(dbgs() << "LV: Found an ordered reduction: Phi: " << *Phi
    213                     << ", ExitInst: " << *Exit << "\n");
    214 
    215   return true;
    216 }
    217 
    218 bool RecurrenceDescriptor::AddReductionVar(PHINode *Phi, RecurKind Kind,
    219                                            Loop *TheLoop, FastMathFlags FuncFMF,
    220                                            RecurrenceDescriptor &RedDes,
    221                                            DemandedBits *DB,
    222                                            AssumptionCache *AC,
    223                                            DominatorTree *DT) {
    224   if (Phi->getNumIncomingValues() != 2)
    225     return false;
    226 
    227   // Reduction variables are only found in the loop header block.
    228   if (Phi->getParent() != TheLoop->getHeader())
    229     return false;
    230 
    231   // Obtain the reduction start value from the value that comes from the loop
    232   // preheader.
    233   Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
    234 
    235   // ExitInstruction is the single value which is used outside the loop.
    236   // We only allow for a single reduction value to be used outside the loop.
    237   // This includes users of the reduction, variables (which form a cycle
    238   // which ends in the phi node).
    239   Instruction *ExitInstruction = nullptr;
    240   // Indicates that we found a reduction operation in our scan.
    241   bool FoundReduxOp = false;
    242 
    243   // We start with the PHI node and scan for all of the users of this
    244   // instruction. All users must be instructions that can be used as reduction
    245   // variables (such as ADD). We must have a single out-of-block user. The cycle
    246   // must include the original PHI.
    247   bool FoundStartPHI = false;
    248 
    249   // To recognize min/max patterns formed by a icmp select sequence, we store
    250   // the number of instruction we saw from the recognized min/max pattern,
    251   //  to make sure we only see exactly the two instructions.
    252   unsigned NumCmpSelectPatternInst = 0;
    253   InstDesc ReduxDesc(false, nullptr);
    254 
    255   // Data used for determining if the recurrence has been type-promoted.
    256   Type *RecurrenceType = Phi->getType();
    257   SmallPtrSet<Instruction *, 4> CastInsts;
    258   Instruction *Start = Phi;
    259   bool IsSigned = false;
    260 
    261   SmallPtrSet<Instruction *, 8> VisitedInsts;
    262   SmallVector<Instruction *, 8> Worklist;
    263 
    264   // Return early if the recurrence kind does not match the type of Phi. If the
    265   // recurrence kind is arithmetic, we attempt to look through AND operations
    266   // resulting from the type promotion performed by InstCombine.  Vector
    267   // operations are not limited to the legal integer widths, so we may be able
    268   // to evaluate the reduction in the narrower width.
    269   if (RecurrenceType->isFloatingPointTy()) {
    270     if (!isFloatingPointRecurrenceKind(Kind))
    271       return false;
    272   } else if (RecurrenceType->isIntegerTy()) {
    273     if (!isIntegerRecurrenceKind(Kind))
    274       return false;
    275     if (isArithmeticRecurrenceKind(Kind))
    276       Start = lookThroughAnd(Phi, RecurrenceType, VisitedInsts, CastInsts);
    277   } else {
    278     // Pointer min/max may exist, but it is not supported as a reduction op.
    279     return false;
    280   }
    281 
    282   Worklist.push_back(Start);
    283   VisitedInsts.insert(Start);
    284 
    285   // Start with all flags set because we will intersect this with the reduction
    286   // flags from all the reduction operations.
    287   FastMathFlags FMF = FastMathFlags::getFast();
    288 
    289   // A value in the reduction can be used:
    290   //  - By the reduction:
    291   //      - Reduction operation:
    292   //        - One use of reduction value (safe).
    293   //        - Multiple use of reduction value (not safe).
    294   //      - PHI:
    295   //        - All uses of the PHI must be the reduction (safe).
    296   //        - Otherwise, not safe.
    297   //  - By instructions outside of the loop (safe).
    298   //      * One value may have several outside users, but all outside
    299   //        uses must be of the same value.
    300   //  - By an instruction that is not part of the reduction (not safe).
    301   //    This is either:
    302   //      * An instruction type other than PHI or the reduction operation.
    303   //      * A PHI in the header other than the initial PHI.
    304   while (!Worklist.empty()) {
    305     Instruction *Cur = Worklist.pop_back_val();
    306 
    307     // No Users.
    308     // If the instruction has no users then this is a broken chain and can't be
    309     // a reduction variable.
    310     if (Cur->use_empty())
    311       return false;
    312 
    313     bool IsAPhi = isa<PHINode>(Cur);
    314 
    315     // A header PHI use other than the original PHI.
    316     if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
    317       return false;
    318 
    319     // Reductions of instructions such as Div, and Sub is only possible if the
    320     // LHS is the reduction variable.
    321     if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
    322         !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
    323         !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
    324       return false;
    325 
    326     // Any reduction instruction must be of one of the allowed kinds. We ignore
    327     // the starting value (the Phi or an AND instruction if the Phi has been
    328     // type-promoted).
    329     if (Cur != Start) {
    330       ReduxDesc = isRecurrenceInstr(Cur, Kind, ReduxDesc, FuncFMF);
    331       if (!ReduxDesc.isRecurrence())
    332         return false;
    333       // FIXME: FMF is allowed on phi, but propagation is not handled correctly.
    334       if (isa<FPMathOperator>(ReduxDesc.getPatternInst()) && !IsAPhi) {
    335         FastMathFlags CurFMF = ReduxDesc.getPatternInst()->getFastMathFlags();
    336         if (auto *Sel = dyn_cast<SelectInst>(ReduxDesc.getPatternInst())) {
    337           // Accept FMF on either fcmp or select of a min/max idiom.
    338           // TODO: This is a hack to work-around the fact that FMF may not be
    339           //       assigned/propagated correctly. If that problem is fixed or we
    340           //       standardize on fmin/fmax via intrinsics, this can be removed.
    341           if (auto *FCmp = dyn_cast<FCmpInst>(Sel->getCondition()))
    342             CurFMF |= FCmp->getFastMathFlags();
    343         }
    344         FMF &= CurFMF;
    345       }
    346       // Update this reduction kind if we matched a new instruction.
    347       // TODO: Can we eliminate the need for a 2nd InstDesc by keeping 'Kind'
    348       //       state accurate while processing the worklist?
    349       if (ReduxDesc.getRecKind() != RecurKind::None)
    350         Kind = ReduxDesc.getRecKind();
    351     }
    352 
    353     bool IsASelect = isa<SelectInst>(Cur);
    354 
    355     // A conditional reduction operation must only have 2 or less uses in
    356     // VisitedInsts.
    357     if (IsASelect && (Kind == RecurKind::FAdd || Kind == RecurKind::FMul) &&
    358         hasMultipleUsesOf(Cur, VisitedInsts, 2))
    359       return false;
    360 
    361     // A reduction operation must only have one use of the reduction value.
    362     if (!IsAPhi && !IsASelect && !isMinMaxRecurrenceKind(Kind) &&
    363         hasMultipleUsesOf(Cur, VisitedInsts, 1))
    364       return false;
    365 
    366     // All inputs to a PHI node must be a reduction value.
    367     if (IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
    368       return false;
    369 
    370     if (isIntMinMaxRecurrenceKind(Kind) &&
    371         (isa<ICmpInst>(Cur) || isa<SelectInst>(Cur)))
    372       ++NumCmpSelectPatternInst;
    373     if (isFPMinMaxRecurrenceKind(Kind) &&
    374         (isa<FCmpInst>(Cur) || isa<SelectInst>(Cur)))
    375       ++NumCmpSelectPatternInst;
    376 
    377     // Check  whether we found a reduction operator.
    378     FoundReduxOp |= !IsAPhi && Cur != Start;
    379 
    380     // Process users of current instruction. Push non-PHI nodes after PHI nodes
    381     // onto the stack. This way we are going to have seen all inputs to PHI
    382     // nodes once we get to them.
    383     SmallVector<Instruction *, 8> NonPHIs;
    384     SmallVector<Instruction *, 8> PHIs;
    385     for (User *U : Cur->users()) {
    386       Instruction *UI = cast<Instruction>(U);
    387 
    388       // Check if we found the exit user.
    389       BasicBlock *Parent = UI->getParent();
    390       if (!TheLoop->contains(Parent)) {
    391         // If we already know this instruction is used externally, move on to
    392         // the next user.
    393         if (ExitInstruction == Cur)
    394           continue;
    395 
    396         // Exit if you find multiple values used outside or if the header phi
    397         // node is being used. In this case the user uses the value of the
    398         // previous iteration, in which case we would loose "VF-1" iterations of
    399         // the reduction operation if we vectorize.
    400         if (ExitInstruction != nullptr || Cur == Phi)
    401           return false;
    402 
    403         // The instruction used by an outside user must be the last instruction
    404         // before we feed back to the reduction phi. Otherwise, we loose VF-1
    405         // operations on the value.
    406         if (!is_contained(Phi->operands(), Cur))
    407           return false;
    408 
    409         ExitInstruction = Cur;
    410         continue;
    411       }
    412 
    413       // Process instructions only once (termination). Each reduction cycle
    414       // value must only be used once, except by phi nodes and min/max
    415       // reductions which are represented as a cmp followed by a select.
    416       InstDesc IgnoredVal(false, nullptr);
    417       if (VisitedInsts.insert(UI).second) {
    418         if (isa<PHINode>(UI))
    419           PHIs.push_back(UI);
    420         else
    421           NonPHIs.push_back(UI);
    422       } else if (!isa<PHINode>(UI) &&
    423                  ((!isa<FCmpInst>(UI) && !isa<ICmpInst>(UI) &&
    424                    !isa<SelectInst>(UI)) ||
    425                   (!isConditionalRdxPattern(Kind, UI).isRecurrence() &&
    426                    !isMinMaxSelectCmpPattern(UI, IgnoredVal).isRecurrence())))
    427         return false;
    428 
    429       // Remember that we completed the cycle.
    430       if (UI == Phi)
    431         FoundStartPHI = true;
    432     }
    433     Worklist.append(PHIs.begin(), PHIs.end());
    434     Worklist.append(NonPHIs.begin(), NonPHIs.end());
    435   }
    436 
    437   // This means we have seen one but not the other instruction of the
    438   // pattern or more than just a select and cmp.
    439   if (isMinMaxRecurrenceKind(Kind) && NumCmpSelectPatternInst != 2)
    440     return false;
    441 
    442   if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
    443     return false;
    444 
    445   const bool IsOrdered = checkOrderedReduction(
    446       Kind, ReduxDesc.getExactFPMathInst(), ExitInstruction, Phi);
    447 
    448   if (Start != Phi) {
    449     // If the starting value is not the same as the phi node, we speculatively
    450     // looked through an 'and' instruction when evaluating a potential
    451     // arithmetic reduction to determine if it may have been type-promoted.
    452     //
    453     // We now compute the minimal bit width that is required to represent the
    454     // reduction. If this is the same width that was indicated by the 'and', we
    455     // can represent the reduction in the smaller type. The 'and' instruction
    456     // will be eliminated since it will essentially be a cast instruction that
    457     // can be ignore in the cost model. If we compute a different type than we
    458     // did when evaluating the 'and', the 'and' will not be eliminated, and we
    459     // will end up with different kinds of operations in the recurrence
    460     // expression (e.g., IntegerAND, IntegerADD). We give up if this is
    461     // the case.
    462     //
    463     // The vectorizer relies on InstCombine to perform the actual
    464     // type-shrinking. It does this by inserting instructions to truncate the
    465     // exit value of the reduction to the width indicated by RecurrenceType and
    466     // then extend this value back to the original width. If IsSigned is false,
    467     // a 'zext' instruction will be generated; otherwise, a 'sext' will be
    468     // used.
    469     //
    470     // TODO: We should not rely on InstCombine to rewrite the reduction in the
    471     //       smaller type. We should just generate a correctly typed expression
    472     //       to begin with.
    473     Type *ComputedType;
    474     std::tie(ComputedType, IsSigned) =
    475         computeRecurrenceType(ExitInstruction, DB, AC, DT);
    476     if (ComputedType != RecurrenceType)
    477       return false;
    478 
    479     // The recurrence expression will be represented in a narrower type. If
    480     // there are any cast instructions that will be unnecessary, collect them
    481     // in CastInsts. Note that the 'and' instruction was already included in
    482     // this list.
    483     //
    484     // TODO: A better way to represent this may be to tag in some way all the
    485     //       instructions that are a part of the reduction. The vectorizer cost
    486     //       model could then apply the recurrence type to these instructions,
    487     //       without needing a white list of instructions to ignore.
    488     //       This may also be useful for the inloop reductions, if it can be
    489     //       kept simple enough.
    490     collectCastsToIgnore(TheLoop, ExitInstruction, RecurrenceType, CastInsts);
    491   }
    492 
    493   // We found a reduction var if we have reached the original phi node and we
    494   // only have a single instruction with out-of-loop users.
    495 
    496   // The ExitInstruction(Instruction which is allowed to have out-of-loop users)
    497   // is saved as part of the RecurrenceDescriptor.
    498 
    499   // Save the description of this reduction variable.
    500   RecurrenceDescriptor RD(RdxStart, ExitInstruction, Kind, FMF,
    501                           ReduxDesc.getExactFPMathInst(), RecurrenceType,
    502                           IsSigned, IsOrdered, CastInsts);
    503   RedDes = RD;
    504 
    505   return true;
    506 }
    507 
    508 RecurrenceDescriptor::InstDesc
    509 RecurrenceDescriptor::isMinMaxSelectCmpPattern(Instruction *I,
    510                                                const InstDesc &Prev) {
    511   assert((isa<CmpInst>(I) || isa<SelectInst>(I)) &&
    512          "Expected a cmp or select instruction");
    513 
    514   // We must handle the select(cmp()) as a single instruction. Advance to the
    515   // select.
    516   CmpInst::Predicate Pred;
    517   if (match(I, m_OneUse(m_Cmp(Pred, m_Value(), m_Value())))) {
    518     if (auto *Select = dyn_cast<SelectInst>(*I->user_begin()))
    519       return InstDesc(Select, Prev.getRecKind());
    520   }
    521 
    522   // Only match select with single use cmp condition.
    523   if (!match(I, m_Select(m_OneUse(m_Cmp(Pred, m_Value(), m_Value())), m_Value(),
    524                          m_Value())))
    525     return InstDesc(false, I);
    526 
    527   // Look for a min/max pattern.
    528   if (match(I, m_UMin(m_Value(), m_Value())))
    529     return InstDesc(I, RecurKind::UMin);
    530   if (match(I, m_UMax(m_Value(), m_Value())))
    531     return InstDesc(I, RecurKind::UMax);
    532   if (match(I, m_SMax(m_Value(), m_Value())))
    533     return InstDesc(I, RecurKind::SMax);
    534   if (match(I, m_SMin(m_Value(), m_Value())))
    535     return InstDesc(I, RecurKind::SMin);
    536   if (match(I, m_OrdFMin(m_Value(), m_Value())))
    537     return InstDesc(I, RecurKind::FMin);
    538   if (match(I, m_OrdFMax(m_Value(), m_Value())))
    539     return InstDesc(I, RecurKind::FMax);
    540   if (match(I, m_UnordFMin(m_Value(), m_Value())))
    541     return InstDesc(I, RecurKind::FMin);
    542   if (match(I, m_UnordFMax(m_Value(), m_Value())))
    543     return InstDesc(I, RecurKind::FMax);
    544 
    545   return InstDesc(false, I);
    546 }
    547 
    548 /// Returns true if the select instruction has users in the compare-and-add
    549 /// reduction pattern below. The select instruction argument is the last one
    550 /// in the sequence.
    551 ///
    552 /// %sum.1 = phi ...
    553 /// ...
    554 /// %cmp = fcmp pred %0, %CFP
    555 /// %add = fadd %0, %sum.1
    556 /// %sum.2 = select %cmp, %add, %sum.1
    557 RecurrenceDescriptor::InstDesc
    558 RecurrenceDescriptor::isConditionalRdxPattern(RecurKind Kind, Instruction *I) {
    559   SelectInst *SI = dyn_cast<SelectInst>(I);
    560   if (!SI)
    561     return InstDesc(false, I);
    562 
    563   CmpInst *CI = dyn_cast<CmpInst>(SI->getCondition());
    564   // Only handle single use cases for now.
    565   if (!CI || !CI->hasOneUse())
    566     return InstDesc(false, I);
    567 
    568   Value *TrueVal = SI->getTrueValue();
    569   Value *FalseVal = SI->getFalseValue();
    570   // Handle only when either of operands of select instruction is a PHI
    571   // node for now.
    572   if ((isa<PHINode>(*TrueVal) && isa<PHINode>(*FalseVal)) ||
    573       (!isa<PHINode>(*TrueVal) && !isa<PHINode>(*FalseVal)))
    574     return InstDesc(false, I);
    575 
    576   Instruction *I1 =
    577       isa<PHINode>(*TrueVal) ? dyn_cast<Instruction>(FalseVal)
    578                              : dyn_cast<Instruction>(TrueVal);
    579   if (!I1 || !I1->isBinaryOp())
    580     return InstDesc(false, I);
    581 
    582   Value *Op1, *Op2;
    583   if ((m_FAdd(m_Value(Op1), m_Value(Op2)).match(I1)  ||
    584        m_FSub(m_Value(Op1), m_Value(Op2)).match(I1)) &&
    585       I1->isFast())
    586     return InstDesc(Kind == RecurKind::FAdd, SI);
    587 
    588   if (m_FMul(m_Value(Op1), m_Value(Op2)).match(I1) && (I1->isFast()))
    589     return InstDesc(Kind == RecurKind::FMul, SI);
    590 
    591   return InstDesc(false, I);
    592 }
    593 
    594 RecurrenceDescriptor::InstDesc
    595 RecurrenceDescriptor::isRecurrenceInstr(Instruction *I, RecurKind Kind,
    596                                         InstDesc &Prev, FastMathFlags FMF) {
    597   switch (I->getOpcode()) {
    598   default:
    599     return InstDesc(false, I);
    600   case Instruction::PHI:
    601     return InstDesc(I, Prev.getRecKind(), Prev.getExactFPMathInst());
    602   case Instruction::Sub:
    603   case Instruction::Add:
    604     return InstDesc(Kind == RecurKind::Add, I);
    605   case Instruction::Mul:
    606     return InstDesc(Kind == RecurKind::Mul, I);
    607   case Instruction::And:
    608     return InstDesc(Kind == RecurKind::And, I);
    609   case Instruction::Or:
    610     return InstDesc(Kind == RecurKind::Or, I);
    611   case Instruction::Xor:
    612     return InstDesc(Kind == RecurKind::Xor, I);
    613   case Instruction::FDiv:
    614   case Instruction::FMul:
    615     return InstDesc(Kind == RecurKind::FMul, I,
    616                     I->hasAllowReassoc() ? nullptr : I);
    617   case Instruction::FSub:
    618   case Instruction::FAdd:
    619     return InstDesc(Kind == RecurKind::FAdd, I,
    620                     I->hasAllowReassoc() ? nullptr : I);
    621   case Instruction::Select:
    622     if (Kind == RecurKind::FAdd || Kind == RecurKind::FMul)
    623       return isConditionalRdxPattern(Kind, I);
    624     LLVM_FALLTHROUGH;
    625   case Instruction::FCmp:
    626   case Instruction::ICmp:
    627     if (isIntMinMaxRecurrenceKind(Kind) ||
    628         (FMF.noNaNs() && FMF.noSignedZeros() && isFPMinMaxRecurrenceKind(Kind)))
    629       return isMinMaxSelectCmpPattern(I, Prev);
    630     return InstDesc(false, I);
    631   }
    632 }
    633 
    634 bool RecurrenceDescriptor::hasMultipleUsesOf(
    635     Instruction *I, SmallPtrSetImpl<Instruction *> &Insts,
    636     unsigned MaxNumUses) {
    637   unsigned NumUses = 0;
    638   for (const Use &U : I->operands()) {
    639     if (Insts.count(dyn_cast<Instruction>(U)))
    640       ++NumUses;
    641     if (NumUses > MaxNumUses)
    642       return true;
    643   }
    644 
    645   return false;
    646 }
    647 bool RecurrenceDescriptor::isReductionPHI(PHINode *Phi, Loop *TheLoop,
    648                                           RecurrenceDescriptor &RedDes,
    649                                           DemandedBits *DB, AssumptionCache *AC,
    650                                           DominatorTree *DT) {
    651 
    652   BasicBlock *Header = TheLoop->getHeader();
    653   Function &F = *Header->getParent();
    654   FastMathFlags FMF;
    655   FMF.setNoNaNs(
    656       F.getFnAttribute("no-nans-fp-math").getValueAsBool());
    657   FMF.setNoSignedZeros(
    658       F.getFnAttribute("no-signed-zeros-fp-math").getValueAsBool());
    659 
    660   if (AddReductionVar(Phi, RecurKind::Add, TheLoop, FMF, RedDes, DB, AC, DT)) {
    661     LLVM_DEBUG(dbgs() << "Found an ADD reduction PHI." << *Phi << "\n");
    662     return true;
    663   }
    664   if (AddReductionVar(Phi, RecurKind::Mul, TheLoop, FMF, RedDes, DB, AC, DT)) {
    665     LLVM_DEBUG(dbgs() << "Found a MUL reduction PHI." << *Phi << "\n");
    666     return true;
    667   }
    668   if (AddReductionVar(Phi, RecurKind::Or, TheLoop, FMF, RedDes, DB, AC, DT)) {
    669     LLVM_DEBUG(dbgs() << "Found an OR reduction PHI." << *Phi << "\n");
    670     return true;
    671   }
    672   if (AddReductionVar(Phi, RecurKind::And, TheLoop, FMF, RedDes, DB, AC, DT)) {
    673     LLVM_DEBUG(dbgs() << "Found an AND reduction PHI." << *Phi << "\n");
    674     return true;
    675   }
    676   if (AddReductionVar(Phi, RecurKind::Xor, TheLoop, FMF, RedDes, DB, AC, DT)) {
    677     LLVM_DEBUG(dbgs() << "Found a XOR reduction PHI." << *Phi << "\n");
    678     return true;
    679   }
    680   if (AddReductionVar(Phi, RecurKind::SMax, TheLoop, FMF, RedDes, DB, AC, DT)) {
    681     LLVM_DEBUG(dbgs() << "Found a SMAX reduction PHI." << *Phi << "\n");
    682     return true;
    683   }
    684   if (AddReductionVar(Phi, RecurKind::SMin, TheLoop, FMF, RedDes, DB, AC, DT)) {
    685     LLVM_DEBUG(dbgs() << "Found a SMIN reduction PHI." << *Phi << "\n");
    686     return true;
    687   }
    688   if (AddReductionVar(Phi, RecurKind::UMax, TheLoop, FMF, RedDes, DB, AC, DT)) {
    689     LLVM_DEBUG(dbgs() << "Found a UMAX reduction PHI." << *Phi << "\n");
    690     return true;
    691   }
    692   if (AddReductionVar(Phi, RecurKind::UMin, TheLoop, FMF, RedDes, DB, AC, DT)) {
    693     LLVM_DEBUG(dbgs() << "Found a UMIN reduction PHI." << *Phi << "\n");
    694     return true;
    695   }
    696   if (AddReductionVar(Phi, RecurKind::FMul, TheLoop, FMF, RedDes, DB, AC, DT)) {
    697     LLVM_DEBUG(dbgs() << "Found an FMult reduction PHI." << *Phi << "\n");
    698     return true;
    699   }
    700   if (AddReductionVar(Phi, RecurKind::FAdd, TheLoop, FMF, RedDes, DB, AC, DT)) {
    701     LLVM_DEBUG(dbgs() << "Found an FAdd reduction PHI." << *Phi << "\n");
    702     return true;
    703   }
    704   if (AddReductionVar(Phi, RecurKind::FMax, TheLoop, FMF, RedDes, DB, AC, DT)) {
    705     LLVM_DEBUG(dbgs() << "Found a float MAX reduction PHI." << *Phi << "\n");
    706     return true;
    707   }
    708   if (AddReductionVar(Phi, RecurKind::FMin, TheLoop, FMF, RedDes, DB, AC, DT)) {
    709     LLVM_DEBUG(dbgs() << "Found a float MIN reduction PHI." << *Phi << "\n");
    710     return true;
    711   }
    712   // Not a reduction of known type.
    713   return false;
    714 }
    715 
    716 bool RecurrenceDescriptor::isFirstOrderRecurrence(
    717     PHINode *Phi, Loop *TheLoop,
    718     DenseMap<Instruction *, Instruction *> &SinkAfter, DominatorTree *DT) {
    719 
    720   // Ensure the phi node is in the loop header and has two incoming values.
    721   if (Phi->getParent() != TheLoop->getHeader() ||
    722       Phi->getNumIncomingValues() != 2)
    723     return false;
    724 
    725   // Ensure the loop has a preheader and a single latch block. The loop
    726   // vectorizer will need the latch to set up the next iteration of the loop.
    727   auto *Preheader = TheLoop->getLoopPreheader();
    728   auto *Latch = TheLoop->getLoopLatch();
    729   if (!Preheader || !Latch)
    730     return false;
    731 
    732   // Ensure the phi node's incoming blocks are the loop preheader and latch.
    733   if (Phi->getBasicBlockIndex(Preheader) < 0 ||
    734       Phi->getBasicBlockIndex(Latch) < 0)
    735     return false;
    736 
    737   // Get the previous value. The previous value comes from the latch edge while
    738   // the initial value comes form the preheader edge.
    739   auto *Previous = dyn_cast<Instruction>(Phi->getIncomingValueForBlock(Latch));
    740   if (!Previous || !TheLoop->contains(Previous) || isa<PHINode>(Previous) ||
    741       SinkAfter.count(Previous)) // Cannot rely on dominance due to motion.
    742     return false;
    743 
    744   // Ensure every user of the phi node is dominated by the previous value.
    745   // The dominance requirement ensures the loop vectorizer will not need to
    746   // vectorize the initial value prior to the first iteration of the loop.
    747   // TODO: Consider extending this sinking to handle memory instructions and
    748   // phis with multiple users.
    749 
    750   // Returns true, if all users of I are dominated by DominatedBy.
    751   auto allUsesDominatedBy = [DT](Instruction *I, Instruction *DominatedBy) {
    752     return all_of(I->uses(), [DT, DominatedBy](Use &U) {
    753       return DT->dominates(DominatedBy, U);
    754     });
    755   };
    756 
    757   if (Phi->hasOneUse()) {
    758     Instruction *I = Phi->user_back();
    759 
    760     // If the user of the PHI is also the incoming value, we potentially have a
    761     // reduction and which cannot be handled by sinking.
    762     if (Previous == I)
    763       return false;
    764 
    765     // We cannot sink terminator instructions.
    766     if (I->getParent()->getTerminator() == I)
    767       return false;
    768 
    769     // Do not try to sink an instruction multiple times (if multiple operands
    770     // are first order recurrences).
    771     // TODO: We can support this case, by sinking the instruction after the
    772     // 'deepest' previous instruction.
    773     if (SinkAfter.find(I) != SinkAfter.end())
    774       return false;
    775 
    776     if (DT->dominates(Previous, I)) // We already are good w/o sinking.
    777       return true;
    778 
    779     // We can sink any instruction without side effects, as long as all users
    780     // are dominated by the instruction we are sinking after.
    781     if (I->getParent() == Phi->getParent() && !I->mayHaveSideEffects() &&
    782         allUsesDominatedBy(I, Previous)) {
    783       SinkAfter[I] = Previous;
    784       return true;
    785     }
    786   }
    787 
    788   return allUsesDominatedBy(Phi, Previous);
    789 }
    790 
    791 /// This function returns the identity element (or neutral element) for
    792 /// the operation K.
    793 Constant *RecurrenceDescriptor::getRecurrenceIdentity(RecurKind K, Type *Tp,
    794                                                       FastMathFlags FMF) {
    795   switch (K) {
    796   case RecurKind::Xor:
    797   case RecurKind::Add:
    798   case RecurKind::Or:
    799     // Adding, Xoring, Oring zero to a number does not change it.
    800     return ConstantInt::get(Tp, 0);
    801   case RecurKind::Mul:
    802     // Multiplying a number by 1 does not change it.
    803     return ConstantInt::get(Tp, 1);
    804   case RecurKind::And:
    805     // AND-ing a number with an all-1 value does not change it.
    806     return ConstantInt::get(Tp, -1, true);
    807   case RecurKind::FMul:
    808     // Multiplying a number by 1 does not change it.
    809     return ConstantFP::get(Tp, 1.0L);
    810   case RecurKind::FAdd:
    811     // Adding zero to a number does not change it.
    812     // FIXME: Ideally we should not need to check FMF for FAdd and should always
    813     // use -0.0. However, this will currently result in mixed vectors of 0.0/-0.0.
    814     // Instead, we should ensure that 1) the FMF from FAdd are propagated to the PHI
    815     // nodes where possible, and 2) PHIs with the nsz flag + -0.0 use 0.0. This would
    816     // mean we can then remove the check for noSignedZeros() below (see D98963).
    817     if (FMF.noSignedZeros())
    818       return ConstantFP::get(Tp, 0.0L);
    819     return ConstantFP::get(Tp, -0.0L);
    820   case RecurKind::UMin:
    821     return ConstantInt::get(Tp, -1);
    822   case RecurKind::UMax:
    823     return ConstantInt::get(Tp, 0);
    824   case RecurKind::SMin:
    825     return ConstantInt::get(Tp,
    826                             APInt::getSignedMaxValue(Tp->getIntegerBitWidth()));
    827   case RecurKind::SMax:
    828     return ConstantInt::get(Tp,
    829                             APInt::getSignedMinValue(Tp->getIntegerBitWidth()));
    830   case RecurKind::FMin:
    831     return ConstantFP::getInfinity(Tp, true);
    832   case RecurKind::FMax:
    833     return ConstantFP::getInfinity(Tp, false);
    834   default:
    835     llvm_unreachable("Unknown recurrence kind");
    836   }
    837 }
    838 
    839 unsigned RecurrenceDescriptor::getOpcode(RecurKind Kind) {
    840   switch (Kind) {
    841   case RecurKind::Add:
    842     return Instruction::Add;
    843   case RecurKind::Mul:
    844     return Instruction::Mul;
    845   case RecurKind::Or:
    846     return Instruction::Or;
    847   case RecurKind::And:
    848     return Instruction::And;
    849   case RecurKind::Xor:
    850     return Instruction::Xor;
    851   case RecurKind::FMul:
    852     return Instruction::FMul;
    853   case RecurKind::FAdd:
    854     return Instruction::FAdd;
    855   case RecurKind::SMax:
    856   case RecurKind::SMin:
    857   case RecurKind::UMax:
    858   case RecurKind::UMin:
    859     return Instruction::ICmp;
    860   case RecurKind::FMax:
    861   case RecurKind::FMin:
    862     return Instruction::FCmp;
    863   default:
    864     llvm_unreachable("Unknown recurrence operation");
    865   }
    866 }
    867 
    868 SmallVector<Instruction *, 4>
    869 RecurrenceDescriptor::getReductionOpChain(PHINode *Phi, Loop *L) const {
    870   SmallVector<Instruction *, 4> ReductionOperations;
    871   unsigned RedOp = getOpcode(Kind);
    872 
    873   // Search down from the Phi to the LoopExitInstr, looking for instructions
    874   // with a single user of the correct type for the reduction.
    875 
    876   // Note that we check that the type of the operand is correct for each item in
    877   // the chain, including the last (the loop exit value). This can come up from
    878   // sub, which would otherwise be treated as an add reduction. MinMax also need
    879   // to check for a pair of icmp/select, for which we use getNextInstruction and
    880   // isCorrectOpcode functions to step the right number of instruction, and
    881   // check the icmp/select pair.
    882   // FIXME: We also do not attempt to look through Phi/Select's yet, which might
    883   // be part of the reduction chain, or attempt to looks through And's to find a
    884   // smaller bitwidth. Subs are also currently not allowed (which are usually
    885   // treated as part of a add reduction) as they are expected to generally be
    886   // more expensive than out-of-loop reductions, and need to be costed more
    887   // carefully.
    888   unsigned ExpectedUses = 1;
    889   if (RedOp == Instruction::ICmp || RedOp == Instruction::FCmp)
    890     ExpectedUses = 2;
    891 
    892   auto getNextInstruction = [&](Instruction *Cur) {
    893     if (RedOp == Instruction::ICmp || RedOp == Instruction::FCmp) {
    894       // We are expecting a icmp/select pair, which we go to the next select
    895       // instruction if we can. We already know that Cur has 2 uses.
    896       if (isa<SelectInst>(*Cur->user_begin()))
    897         return cast<Instruction>(*Cur->user_begin());
    898       else
    899         return cast<Instruction>(*std::next(Cur->user_begin()));
    900     }
    901     return cast<Instruction>(*Cur->user_begin());
    902   };
    903   auto isCorrectOpcode = [&](Instruction *Cur) {
    904     if (RedOp == Instruction::ICmp || RedOp == Instruction::FCmp) {
    905       Value *LHS, *RHS;
    906       return SelectPatternResult::isMinOrMax(
    907           matchSelectPattern(Cur, LHS, RHS).Flavor);
    908     }
    909     return Cur->getOpcode() == RedOp;
    910   };
    911 
    912   // The loop exit instruction we check first (as a quick test) but add last. We
    913   // check the opcode is correct (and dont allow them to be Subs) and that they
    914   // have expected to have the expected number of uses. They will have one use
    915   // from the phi and one from a LCSSA value, no matter the type.
    916   if (!isCorrectOpcode(LoopExitInstr) || !LoopExitInstr->hasNUses(2))
    917     return {};
    918 
    919   // Check that the Phi has one (or two for min/max) uses.
    920   if (!Phi->hasNUses(ExpectedUses))
    921     return {};
    922   Instruction *Cur = getNextInstruction(Phi);
    923 
    924   // Each other instruction in the chain should have the expected number of uses
    925   // and be the correct opcode.
    926   while (Cur != LoopExitInstr) {
    927     if (!isCorrectOpcode(Cur) || !Cur->hasNUses(ExpectedUses))
    928       return {};
    929 
    930     ReductionOperations.push_back(Cur);
    931     Cur = getNextInstruction(Cur);
    932   }
    933 
    934   ReductionOperations.push_back(Cur);
    935   return ReductionOperations;
    936 }
    937 
    938 InductionDescriptor::InductionDescriptor(Value *Start, InductionKind K,
    939                                          const SCEV *Step, BinaryOperator *BOp,
    940                                          SmallVectorImpl<Instruction *> *Casts)
    941     : StartValue(Start), IK(K), Step(Step), InductionBinOp(BOp) {
    942   assert(IK != IK_NoInduction && "Not an induction");
    943 
    944   // Start value type should match the induction kind and the value
    945   // itself should not be null.
    946   assert(StartValue && "StartValue is null");
    947   assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) &&
    948          "StartValue is not a pointer for pointer induction");
    949   assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) &&
    950          "StartValue is not an integer for integer induction");
    951 
    952   // Check the Step Value. It should be non-zero integer value.
    953   assert((!getConstIntStepValue() || !getConstIntStepValue()->isZero()) &&
    954          "Step value is zero");
    955 
    956   assert((IK != IK_PtrInduction || getConstIntStepValue()) &&
    957          "Step value should be constant for pointer induction");
    958   assert((IK == IK_FpInduction || Step->getType()->isIntegerTy()) &&
    959          "StepValue is not an integer");
    960 
    961   assert((IK != IK_FpInduction || Step->getType()->isFloatingPointTy()) &&
    962          "StepValue is not FP for FpInduction");
    963   assert((IK != IK_FpInduction ||
    964           (InductionBinOp &&
    965            (InductionBinOp->getOpcode() == Instruction::FAdd ||
    966             InductionBinOp->getOpcode() == Instruction::FSub))) &&
    967          "Binary opcode should be specified for FP induction");
    968 
    969   if (Casts) {
    970     for (auto &Inst : *Casts) {
    971       RedundantCasts.push_back(Inst);
    972     }
    973   }
    974 }
    975 
    976 ConstantInt *InductionDescriptor::getConstIntStepValue() const {
    977   if (isa<SCEVConstant>(Step))
    978     return dyn_cast<ConstantInt>(cast<SCEVConstant>(Step)->getValue());
    979   return nullptr;
    980 }
    981 
    982 bool InductionDescriptor::isFPInductionPHI(PHINode *Phi, const Loop *TheLoop,
    983                                            ScalarEvolution *SE,
    984                                            InductionDescriptor &D) {
    985 
    986   // Here we only handle FP induction variables.
    987   assert(Phi->getType()->isFloatingPointTy() && "Unexpected Phi type");
    988 
    989   if (TheLoop->getHeader() != Phi->getParent())
    990     return false;
    991 
    992   // The loop may have multiple entrances or multiple exits; we can analyze
    993   // this phi if it has a unique entry value and a unique backedge value.
    994   if (Phi->getNumIncomingValues() != 2)
    995     return false;
    996   Value *BEValue = nullptr, *StartValue = nullptr;
    997   if (TheLoop->contains(Phi->getIncomingBlock(0))) {
    998     BEValue = Phi->getIncomingValue(0);
    999     StartValue = Phi->getIncomingValue(1);
   1000   } else {
   1001     assert(TheLoop->contains(Phi->getIncomingBlock(1)) &&
   1002            "Unexpected Phi node in the loop");
   1003     BEValue = Phi->getIncomingValue(1);
   1004     StartValue = Phi->getIncomingValue(0);
   1005   }
   1006 
   1007   BinaryOperator *BOp = dyn_cast<BinaryOperator>(BEValue);
   1008   if (!BOp)
   1009     return false;
   1010 
   1011   Value *Addend = nullptr;
   1012   if (BOp->getOpcode() == Instruction::FAdd) {
   1013     if (BOp->getOperand(0) == Phi)
   1014       Addend = BOp->getOperand(1);
   1015     else if (BOp->getOperand(1) == Phi)
   1016       Addend = BOp->getOperand(0);
   1017   } else if (BOp->getOpcode() == Instruction::FSub)
   1018     if (BOp->getOperand(0) == Phi)
   1019       Addend = BOp->getOperand(1);
   1020 
   1021   if (!Addend)
   1022     return false;
   1023 
   1024   // The addend should be loop invariant
   1025   if (auto *I = dyn_cast<Instruction>(Addend))
   1026     if (TheLoop->contains(I))
   1027       return false;
   1028 
   1029   // FP Step has unknown SCEV
   1030   const SCEV *Step = SE->getUnknown(Addend);
   1031   D = InductionDescriptor(StartValue, IK_FpInduction, Step, BOp);
   1032   return true;
   1033 }
   1034 
   1035 /// This function is called when we suspect that the update-chain of a phi node
   1036 /// (whose symbolic SCEV expression sin \p PhiScev) contains redundant casts,
   1037 /// that can be ignored. (This can happen when the PSCEV rewriter adds a runtime
   1038 /// predicate P under which the SCEV expression for the phi can be the
   1039 /// AddRecurrence \p AR; See createAddRecFromPHIWithCast). We want to find the
   1040 /// cast instructions that are involved in the update-chain of this induction.
   1041 /// A caller that adds the required runtime predicate can be free to drop these
   1042 /// cast instructions, and compute the phi using \p AR (instead of some scev
   1043 /// expression with casts).
   1044 ///
   1045 /// For example, without a predicate the scev expression can take the following
   1046 /// form:
   1047 ///      (Ext ix (Trunc iy ( Start + i*Step ) to ix) to iy)
   1048 ///
   1049 /// It corresponds to the following IR sequence:
   1050 /// %for.body:
   1051 ///   %x = phi i64 [ 0, %ph ], [ %add, %for.body ]
   1052 ///   %casted_phi = "ExtTrunc i64 %x"
   1053 ///   %add = add i64 %casted_phi, %step
   1054 ///
   1055 /// where %x is given in \p PN,
   1056 /// PSE.getSCEV(%x) is equal to PSE.getSCEV(%casted_phi) under a predicate,
   1057 /// and the IR sequence that "ExtTrunc i64 %x" represents can take one of
   1058 /// several forms, for example, such as:
   1059 ///   ExtTrunc1:    %casted_phi = and  %x, 2^n-1
   1060 /// or:
   1061 ///   ExtTrunc2:    %t = shl %x, m
   1062 ///                 %casted_phi = ashr %t, m
   1063 ///
   1064 /// If we are able to find such sequence, we return the instructions
   1065 /// we found, namely %casted_phi and the instructions on its use-def chain up
   1066 /// to the phi (not including the phi).
   1067 static bool getCastsForInductionPHI(PredicatedScalarEvolution &PSE,
   1068                                     const SCEVUnknown *PhiScev,
   1069                                     const SCEVAddRecExpr *AR,
   1070                                     SmallVectorImpl<Instruction *> &CastInsts) {
   1071 
   1072   assert(CastInsts.empty() && "CastInsts is expected to be empty.");
   1073   auto *PN = cast<PHINode>(PhiScev->getValue());
   1074   assert(PSE.getSCEV(PN) == AR && "Unexpected phi node SCEV expression");
   1075   const Loop *L = AR->getLoop();
   1076 
   1077   // Find any cast instructions that participate in the def-use chain of
   1078   // PhiScev in the loop.
   1079   // FORNOW/TODO: We currently expect the def-use chain to include only
   1080   // two-operand instructions, where one of the operands is an invariant.
   1081   // createAddRecFromPHIWithCasts() currently does not support anything more
   1082   // involved than that, so we keep the search simple. This can be
   1083   // extended/generalized as needed.
   1084 
   1085   auto getDef = [&](const Value *Val) -> Value * {
   1086     const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Val);
   1087     if (!BinOp)
   1088       return nullptr;
   1089     Value *Op0 = BinOp->getOperand(0);
   1090     Value *Op1 = BinOp->getOperand(1);
   1091     Value *Def = nullptr;
   1092     if (L->isLoopInvariant(Op0))
   1093       Def = Op1;
   1094     else if (L->isLoopInvariant(Op1))
   1095       Def = Op0;
   1096     return Def;
   1097   };
   1098 
   1099   // Look for the instruction that defines the induction via the
   1100   // loop backedge.
   1101   BasicBlock *Latch = L->getLoopLatch();
   1102   if (!Latch)
   1103     return false;
   1104   Value *Val = PN->getIncomingValueForBlock(Latch);
   1105   if (!Val)
   1106     return false;
   1107 
   1108   // Follow the def-use chain until the induction phi is reached.
   1109   // If on the way we encounter a Value that has the same SCEV Expr as the
   1110   // phi node, we can consider the instructions we visit from that point
   1111   // as part of the cast-sequence that can be ignored.
   1112   bool InCastSequence = false;
   1113   auto *Inst = dyn_cast<Instruction>(Val);
   1114   while (Val != PN) {
   1115     // If we encountered a phi node other than PN, or if we left the loop,
   1116     // we bail out.
   1117     if (!Inst || !L->contains(Inst)) {
   1118       return false;
   1119     }
   1120     auto *AddRec = dyn_cast<SCEVAddRecExpr>(PSE.getSCEV(Val));
   1121     if (AddRec && PSE.areAddRecsEqualWithPreds(AddRec, AR))
   1122       InCastSequence = true;
   1123     if (InCastSequence) {
   1124       // Only the last instruction in the cast sequence is expected to have
   1125       // uses outside the induction def-use chain.
   1126       if (!CastInsts.empty())
   1127         if (!Inst->hasOneUse())
   1128           return false;
   1129       CastInsts.push_back(Inst);
   1130     }
   1131     Val = getDef(Val);
   1132     if (!Val)
   1133       return false;
   1134     Inst = dyn_cast<Instruction>(Val);
   1135   }
   1136 
   1137   return InCastSequence;
   1138 }
   1139 
   1140 bool InductionDescriptor::isInductionPHI(PHINode *Phi, const Loop *TheLoop,
   1141                                          PredicatedScalarEvolution &PSE,
   1142                                          InductionDescriptor &D, bool Assume) {
   1143   Type *PhiTy = Phi->getType();
   1144 
   1145   // Handle integer and pointer inductions variables.
   1146   // Now we handle also FP induction but not trying to make a
   1147   // recurrent expression from the PHI node in-place.
   1148 
   1149   if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy() && !PhiTy->isFloatTy() &&
   1150       !PhiTy->isDoubleTy() && !PhiTy->isHalfTy())
   1151     return false;
   1152 
   1153   if (PhiTy->isFloatingPointTy())
   1154     return isFPInductionPHI(Phi, TheLoop, PSE.getSE(), D);
   1155 
   1156   const SCEV *PhiScev = PSE.getSCEV(Phi);
   1157   const auto *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
   1158 
   1159   // We need this expression to be an AddRecExpr.
   1160   if (Assume && !AR)
   1161     AR = PSE.getAsAddRec(Phi);
   1162 
   1163   if (!AR) {
   1164     LLVM_DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
   1165     return false;
   1166   }
   1167 
   1168   // Record any Cast instructions that participate in the induction update
   1169   const auto *SymbolicPhi = dyn_cast<SCEVUnknown>(PhiScev);
   1170   // If we started from an UnknownSCEV, and managed to build an addRecurrence
   1171   // only after enabling Assume with PSCEV, this means we may have encountered
   1172   // cast instructions that required adding a runtime check in order to
   1173   // guarantee the correctness of the AddRecurrence respresentation of the
   1174   // induction.
   1175   if (PhiScev != AR && SymbolicPhi) {
   1176     SmallVector<Instruction *, 2> Casts;
   1177     if (getCastsForInductionPHI(PSE, SymbolicPhi, AR, Casts))
   1178       return isInductionPHI(Phi, TheLoop, PSE.getSE(), D, AR, &Casts);
   1179   }
   1180 
   1181   return isInductionPHI(Phi, TheLoop, PSE.getSE(), D, AR);
   1182 }
   1183 
   1184 bool InductionDescriptor::isInductionPHI(
   1185     PHINode *Phi, const Loop *TheLoop, ScalarEvolution *SE,
   1186     InductionDescriptor &D, const SCEV *Expr,
   1187     SmallVectorImpl<Instruction *> *CastsToIgnore) {
   1188   Type *PhiTy = Phi->getType();
   1189   // We only handle integer and pointer inductions variables.
   1190   if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
   1191     return false;
   1192 
   1193   // Check that the PHI is consecutive.
   1194   const SCEV *PhiScev = Expr ? Expr : SE->getSCEV(Phi);
   1195   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
   1196 
   1197   if (!AR) {
   1198     LLVM_DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
   1199     return false;
   1200   }
   1201 
   1202   if (AR->getLoop() != TheLoop) {
   1203     // FIXME: We should treat this as a uniform. Unfortunately, we
   1204     // don't currently know how to handled uniform PHIs.
   1205     LLVM_DEBUG(
   1206         dbgs() << "LV: PHI is a recurrence with respect to an outer loop.\n");
   1207     return false;
   1208   }
   1209 
   1210   Value *StartValue =
   1211       Phi->getIncomingValueForBlock(AR->getLoop()->getLoopPreheader());
   1212 
   1213   BasicBlock *Latch = AR->getLoop()->getLoopLatch();
   1214   if (!Latch)
   1215     return false;
   1216   BinaryOperator *BOp =
   1217       dyn_cast<BinaryOperator>(Phi->getIncomingValueForBlock(Latch));
   1218 
   1219   const SCEV *Step = AR->getStepRecurrence(*SE);
   1220   // Calculate the pointer stride and check if it is consecutive.
   1221   // The stride may be a constant or a loop invariant integer value.
   1222   const SCEVConstant *ConstStep = dyn_cast<SCEVConstant>(Step);
   1223   if (!ConstStep && !SE->isLoopInvariant(Step, TheLoop))
   1224     return false;
   1225 
   1226   if (PhiTy->isIntegerTy()) {
   1227     D = InductionDescriptor(StartValue, IK_IntInduction, Step, BOp,
   1228                             CastsToIgnore);
   1229     return true;
   1230   }
   1231 
   1232   assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
   1233   // Pointer induction should be a constant.
   1234   if (!ConstStep)
   1235     return false;
   1236 
   1237   ConstantInt *CV = ConstStep->getValue();
   1238   Type *PointerElementType = PhiTy->getPointerElementType();
   1239   // The pointer stride cannot be determined if the pointer element type is not
   1240   // sized.
   1241   if (!PointerElementType->isSized())
   1242     return false;
   1243 
   1244   const DataLayout &DL = Phi->getModule()->getDataLayout();
   1245   int64_t Size = static_cast<int64_t>(DL.getTypeAllocSize(PointerElementType));
   1246   if (!Size)
   1247     return false;
   1248 
   1249   int64_t CVSize = CV->getSExtValue();
   1250   if (CVSize % Size)
   1251     return false;
   1252   auto *StepValue =
   1253       SE->getConstant(CV->getType(), CVSize / Size, true /* signed */);
   1254   D = InductionDescriptor(StartValue, IK_PtrInduction, StepValue, BOp);
   1255   return true;
   1256 }
   1257