xref: /freebsd/contrib/llvm-project/llvm/lib/Analysis/IVDescriptors.cpp (revision e64bea71c21eb42e97aa615188ba91f6cce0d36d)
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/Analysis/DemandedBits.h"
15 #include "llvm/Analysis/LoopInfo.h"
16 #include "llvm/Analysis/ScalarEvolution.h"
17 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
18 #include "llvm/Analysis/ValueTracking.h"
19 #include "llvm/IR/Dominators.h"
20 #include "llvm/IR/Instructions.h"
21 #include "llvm/IR/PatternMatch.h"
22 #include "llvm/IR/ValueHandle.h"
23 #include "llvm/Support/Debug.h"
24 #include "llvm/Support/KnownBits.h"
25 
26 using namespace llvm;
27 using namespace llvm::PatternMatch;
28 
29 #define DEBUG_TYPE "iv-descriptors"
30 
areAllUsesIn(Instruction * I,SmallPtrSetImpl<Instruction * > & Set)31 bool RecurrenceDescriptor::areAllUsesIn(Instruction *I,
32                                         SmallPtrSetImpl<Instruction *> &Set) {
33   for (const Use &Use : I->operands())
34     if (!Set.count(dyn_cast<Instruction>(Use)))
35       return false;
36   return true;
37 }
38 
isIntegerRecurrenceKind(RecurKind Kind)39 bool RecurrenceDescriptor::isIntegerRecurrenceKind(RecurKind Kind) {
40   switch (Kind) {
41   default:
42     break;
43   case RecurKind::Add:
44   case RecurKind::Mul:
45   case RecurKind::Or:
46   case RecurKind::And:
47   case RecurKind::Xor:
48   case RecurKind::SMax:
49   case RecurKind::SMin:
50   case RecurKind::UMax:
51   case RecurKind::UMin:
52   case RecurKind::AnyOf:
53   case RecurKind::FindFirstIVSMin:
54   case RecurKind::FindFirstIVUMin:
55   case RecurKind::FindLastIVSMax:
56   case RecurKind::FindLastIVUMax:
57     return true;
58   }
59   return false;
60 }
61 
isFloatingPointRecurrenceKind(RecurKind Kind)62 bool RecurrenceDescriptor::isFloatingPointRecurrenceKind(RecurKind Kind) {
63   return (Kind != RecurKind::None) && !isIntegerRecurrenceKind(Kind);
64 }
65 
66 /// Determines if Phi may have been type-promoted. If Phi has a single user
67 /// that ANDs the Phi with a type mask, return the user. RT is updated to
68 /// account for the narrower bit width represented by the mask, and the AND
69 /// instruction is added to CI.
lookThroughAnd(PHINode * Phi,Type * & RT,SmallPtrSetImpl<Instruction * > & Visited,SmallPtrSetImpl<Instruction * > & CI)70 static Instruction *lookThroughAnd(PHINode *Phi, Type *&RT,
71                                    SmallPtrSetImpl<Instruction *> &Visited,
72                                    SmallPtrSetImpl<Instruction *> &CI) {
73   if (!Phi->hasOneUse())
74     return Phi;
75 
76   const APInt *M = nullptr;
77   Instruction *I, *J = cast<Instruction>(Phi->use_begin()->getUser());
78 
79   // Matches either I & 2^x-1 or 2^x-1 & I. If we find a match, we update RT
80   // with a new integer type of the corresponding bit width.
81   if (match(J, m_And(m_Instruction(I), m_APInt(M)))) {
82     int32_t Bits = (*M + 1).exactLogBase2();
83     if (Bits > 0) {
84       RT = IntegerType::get(Phi->getContext(), Bits);
85       Visited.insert(Phi);
86       CI.insert(J);
87       return J;
88     }
89   }
90   return Phi;
91 }
92 
93 /// Compute the minimal bit width needed to represent a reduction whose exit
94 /// instruction is given by Exit.
computeRecurrenceType(Instruction * Exit,DemandedBits * DB,AssumptionCache * AC,DominatorTree * DT)95 static std::pair<Type *, bool> computeRecurrenceType(Instruction *Exit,
96                                                      DemandedBits *DB,
97                                                      AssumptionCache *AC,
98                                                      DominatorTree *DT) {
99   bool IsSigned = false;
100   const DataLayout &DL = Exit->getDataLayout();
101   uint64_t MaxBitWidth = DL.getTypeSizeInBits(Exit->getType());
102 
103   if (DB) {
104     // Use the demanded bits analysis to determine the bits that are live out
105     // of the exit instruction, rounding up to the nearest power of two. If the
106     // use of demanded bits results in a smaller bit width, we know the value
107     // must be positive (i.e., IsSigned = false), because if this were not the
108     // case, the sign bit would have been demanded.
109     auto Mask = DB->getDemandedBits(Exit);
110     MaxBitWidth = Mask.getBitWidth() - Mask.countl_zero();
111   }
112 
113   if (MaxBitWidth == DL.getTypeSizeInBits(Exit->getType()) && AC && DT) {
114     // If demanded bits wasn't able to limit the bit width, we can try to use
115     // value tracking instead. This can be the case, for example, if the value
116     // may be negative.
117     auto NumSignBits = ComputeNumSignBits(Exit, DL, AC, nullptr, DT);
118     auto NumTypeBits = DL.getTypeSizeInBits(Exit->getType());
119     MaxBitWidth = NumTypeBits - NumSignBits;
120     KnownBits Bits = computeKnownBits(Exit, DL);
121     if (!Bits.isNonNegative()) {
122       // If the value is not known to be non-negative, we set IsSigned to true,
123       // meaning that we will use sext instructions instead of zext
124       // instructions to restore the original type.
125       IsSigned = true;
126       // Make sure at least one sign bit is included in the result, so it
127       // will get properly sign-extended.
128       ++MaxBitWidth;
129     }
130   }
131   MaxBitWidth = llvm::bit_ceil(MaxBitWidth);
132 
133   return std::make_pair(Type::getIntNTy(Exit->getContext(), MaxBitWidth),
134                         IsSigned);
135 }
136 
137 /// Collect cast instructions that can be ignored in the vectorizer's cost
138 /// model, given a reduction exit value and the minimal type in which the
139 // reduction can be represented. Also search casts to the recurrence type
140 // to find the minimum width used by the recurrence.
collectCastInstrs(Loop * TheLoop,Instruction * Exit,Type * RecurrenceType,SmallPtrSetImpl<Instruction * > & Casts,unsigned & MinWidthCastToRecurTy)141 static void collectCastInstrs(Loop *TheLoop, Instruction *Exit,
142                               Type *RecurrenceType,
143                               SmallPtrSetImpl<Instruction *> &Casts,
144                               unsigned &MinWidthCastToRecurTy) {
145 
146   SmallVector<Instruction *, 8> Worklist;
147   SmallPtrSet<Instruction *, 8> Visited;
148   Worklist.push_back(Exit);
149   MinWidthCastToRecurTy = -1U;
150 
151   while (!Worklist.empty()) {
152     Instruction *Val = Worklist.pop_back_val();
153     Visited.insert(Val);
154     if (auto *Cast = dyn_cast<CastInst>(Val)) {
155       if (Cast->getSrcTy() == RecurrenceType) {
156         // If the source type of a cast instruction is equal to the recurrence
157         // type, it will be eliminated, and should be ignored in the vectorizer
158         // cost model.
159         Casts.insert(Cast);
160         continue;
161       }
162       if (Cast->getDestTy() == RecurrenceType) {
163         // The minimum width used by the recurrence is found by checking for
164         // casts on its operands. The minimum width is used by the vectorizer
165         // when finding the widest type for in-loop reductions without any
166         // loads/stores.
167         MinWidthCastToRecurTy = std::min<unsigned>(
168             MinWidthCastToRecurTy, Cast->getSrcTy()->getScalarSizeInBits());
169         continue;
170       }
171     }
172     // Add all operands to the work list if they are loop-varying values that
173     // we haven't yet visited.
174     for (Value *O : cast<User>(Val)->operands())
175       if (auto *I = dyn_cast<Instruction>(O))
176         if (TheLoop->contains(I) && !Visited.count(I))
177           Worklist.push_back(I);
178   }
179 }
180 
181 // Check if a given Phi node can be recognized as an ordered reduction for
182 // vectorizing floating point operations without unsafe math.
checkOrderedReduction(RecurKind Kind,Instruction * ExactFPMathInst,Instruction * Exit,PHINode * Phi)183 static bool checkOrderedReduction(RecurKind Kind, Instruction *ExactFPMathInst,
184                                   Instruction *Exit, PHINode *Phi) {
185   // Currently only FAdd and FMulAdd are supported.
186   if (Kind != RecurKind::FAdd && Kind != RecurKind::FMulAdd)
187     return false;
188 
189   if (Kind == RecurKind::FAdd && Exit->getOpcode() != Instruction::FAdd)
190     return false;
191 
192   if (Kind == RecurKind::FMulAdd &&
193       !RecurrenceDescriptor::isFMulAddIntrinsic(Exit))
194     return false;
195 
196   // Ensure the exit instruction has only one user other than the reduction PHI
197   if (Exit != ExactFPMathInst || Exit->hasNUsesOrMore(3))
198     return false;
199 
200   // The only pattern accepted is the one in which the reduction PHI
201   // is used as one of the operands of the exit instruction
202   auto *Op0 = Exit->getOperand(0);
203   auto *Op1 = Exit->getOperand(1);
204   if (Kind == RecurKind::FAdd && Op0 != Phi && Op1 != Phi)
205     return false;
206   if (Kind == RecurKind::FMulAdd && Exit->getOperand(2) != Phi)
207     return false;
208 
209   LLVM_DEBUG(dbgs() << "LV: Found an ordered reduction: Phi: " << *Phi
210                     << ", ExitInst: " << *Exit << "\n");
211 
212   return true;
213 }
214 
AddReductionVar(PHINode * Phi,RecurKind Kind,Loop * TheLoop,FastMathFlags FuncFMF,RecurrenceDescriptor & RedDes,DemandedBits * DB,AssumptionCache * AC,DominatorTree * DT,ScalarEvolution * SE)215 bool RecurrenceDescriptor::AddReductionVar(
216     PHINode *Phi, RecurKind Kind, Loop *TheLoop, FastMathFlags FuncFMF,
217     RecurrenceDescriptor &RedDes, DemandedBits *DB, AssumptionCache *AC,
218     DominatorTree *DT, ScalarEvolution *SE) {
219   if (Phi->getNumIncomingValues() != 2)
220     return false;
221 
222   // Reduction variables are only found in the loop header block.
223   if (Phi->getParent() != TheLoop->getHeader())
224     return false;
225 
226   // Obtain the reduction start value from the value that comes from the loop
227   // preheader.
228   Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
229 
230   // ExitInstruction is the single value which is used outside the loop.
231   // We only allow for a single reduction value to be used outside the loop.
232   // This includes users of the reduction, variables (which form a cycle
233   // which ends in the phi node).
234   Instruction *ExitInstruction = nullptr;
235 
236   // Variable to keep last visited store instruction. By the end of the
237   // algorithm this variable will be either empty or having intermediate
238   // reduction value stored in invariant address.
239   StoreInst *IntermediateStore = nullptr;
240 
241   // Indicates that we found a reduction operation in our scan.
242   bool FoundReduxOp = false;
243 
244   // We start with the PHI node and scan for all of the users of this
245   // instruction. All users must be instructions that can be used as reduction
246   // variables (such as ADD). We must have a single out-of-block user. The cycle
247   // must include the original PHI.
248   bool FoundStartPHI = false;
249 
250   // To recognize min/max patterns formed by a icmp select sequence, we store
251   // the number of instruction we saw from the recognized min/max pattern,
252   //  to make sure we only see exactly the two instructions.
253   unsigned NumCmpSelectPatternInst = 0;
254   InstDesc ReduxDesc(false, nullptr);
255 
256   // Data used for determining if the recurrence has been type-promoted.
257   Type *RecurrenceType = Phi->getType();
258   SmallPtrSet<Instruction *, 4> CastInsts;
259   unsigned MinWidthCastToRecurrenceType;
260   Instruction *Start = Phi;
261   bool IsSigned = false;
262 
263   SmallPtrSet<Instruction *, 8> VisitedInsts;
264   SmallVector<Instruction *, 8> Worklist;
265 
266   // Return early if the recurrence kind does not match the type of Phi. If the
267   // recurrence kind is arithmetic, we attempt to look through AND operations
268   // resulting from the type promotion performed by InstCombine.  Vector
269   // operations are not limited to the legal integer widths, so we may be able
270   // to evaluate the reduction in the narrower width.
271   if (RecurrenceType->isFloatingPointTy()) {
272     if (!isFloatingPointRecurrenceKind(Kind))
273       return false;
274   } else if (RecurrenceType->isIntegerTy()) {
275     if (!isIntegerRecurrenceKind(Kind))
276       return false;
277     if (!isMinMaxRecurrenceKind(Kind))
278       Start = lookThroughAnd(Phi, RecurrenceType, VisitedInsts, CastInsts);
279   } else {
280     // Pointer min/max may exist, but it is not supported as a reduction op.
281     return false;
282   }
283 
284   Worklist.push_back(Start);
285   VisitedInsts.insert(Start);
286 
287   // Start with all flags set because we will intersect this with the reduction
288   // flags from all the reduction operations.
289   FastMathFlags FMF = FastMathFlags::getFast();
290 
291   // The first instruction in the use-def chain of the Phi node that requires
292   // exact floating point operations.
293   Instruction *ExactFPMathInst = nullptr;
294 
295   // A value in the reduction can be used:
296   //  - By the reduction:
297   //      - Reduction operation:
298   //        - One use of reduction value (safe).
299   //        - Multiple use of reduction value (not safe).
300   //      - PHI:
301   //        - All uses of the PHI must be the reduction (safe).
302   //        - Otherwise, not safe.
303   //  - By instructions outside of the loop (safe).
304   //      * One value may have several outside users, but all outside
305   //        uses must be of the same value.
306   //  - By store instructions with a loop invariant address (safe with
307   //    the following restrictions):
308   //      * If there are several stores, all must have the same address.
309   //      * Final value should be stored in that loop invariant address.
310   //  - By an instruction that is not part of the reduction (not safe).
311   //    This is either:
312   //      * An instruction type other than PHI or the reduction operation.
313   //      * A PHI in the header other than the initial PHI.
314   while (!Worklist.empty()) {
315     Instruction *Cur = Worklist.pop_back_val();
316 
317     // Store instructions are allowed iff it is the store of the reduction
318     // value to the same loop invariant memory location.
319     if (auto *SI = dyn_cast<StoreInst>(Cur)) {
320       if (!SE) {
321         LLVM_DEBUG(dbgs() << "Store instructions are not processed without "
322                           << "Scalar Evolution Analysis\n");
323         return false;
324       }
325 
326       const SCEV *PtrScev = SE->getSCEV(SI->getPointerOperand());
327       // Check it is the same address as previous stores
328       if (IntermediateStore) {
329         const SCEV *OtherScev =
330             SE->getSCEV(IntermediateStore->getPointerOperand());
331 
332         if (OtherScev != PtrScev) {
333           LLVM_DEBUG(dbgs() << "Storing reduction value to different addresses "
334                             << "inside the loop: " << *SI->getPointerOperand()
335                             << " and "
336                             << *IntermediateStore->getPointerOperand() << '\n');
337           return false;
338         }
339       }
340 
341       // Check the pointer is loop invariant
342       if (!SE->isLoopInvariant(PtrScev, TheLoop)) {
343         LLVM_DEBUG(dbgs() << "Storing reduction value to non-uniform address "
344                           << "inside the loop: " << *SI->getPointerOperand()
345                           << '\n');
346         return false;
347       }
348 
349       // IntermediateStore is always the last store in the loop.
350       IntermediateStore = SI;
351       continue;
352     }
353 
354     // No Users.
355     // If the instruction has no users then this is a broken chain and can't be
356     // a reduction variable.
357     if (Cur->use_empty())
358       return false;
359 
360     bool IsAPhi = isa<PHINode>(Cur);
361 
362     // A header PHI use other than the original PHI.
363     if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
364       return false;
365 
366     // Reductions of instructions such as Div, and Sub is only possible if the
367     // LHS is the reduction variable.
368     if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
369         !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
370         !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
371       return false;
372 
373     // Any reduction instruction must be of one of the allowed kinds. We ignore
374     // the starting value (the Phi or an AND instruction if the Phi has been
375     // type-promoted).
376     if (Cur != Start) {
377       ReduxDesc =
378           isRecurrenceInstr(TheLoop, Phi, Cur, Kind, ReduxDesc, FuncFMF, SE);
379       ExactFPMathInst = ExactFPMathInst == nullptr
380                             ? ReduxDesc.getExactFPMathInst()
381                             : ExactFPMathInst;
382       if (!ReduxDesc.isRecurrence())
383         return false;
384       // FIXME: FMF is allowed on phi, but propagation is not handled correctly.
385       if (isa<FPMathOperator>(ReduxDesc.getPatternInst()) && !IsAPhi) {
386         FastMathFlags CurFMF = ReduxDesc.getPatternInst()->getFastMathFlags();
387         if (auto *Sel = dyn_cast<SelectInst>(ReduxDesc.getPatternInst())) {
388           // Accept FMF on either fcmp or select of a min/max idiom.
389           // TODO: This is a hack to work-around the fact that FMF may not be
390           //       assigned/propagated correctly. If that problem is fixed or we
391           //       standardize on fmin/fmax via intrinsics, this can be removed.
392           if (auto *FCmp = dyn_cast<FCmpInst>(Sel->getCondition()))
393             CurFMF |= FCmp->getFastMathFlags();
394         }
395         FMF &= CurFMF;
396       }
397       // Update this reduction kind if we matched a new instruction.
398       // TODO: Can we eliminate the need for a 2nd InstDesc by keeping 'Kind'
399       //       state accurate while processing the worklist?
400       if (ReduxDesc.getRecKind() != RecurKind::None)
401         Kind = ReduxDesc.getRecKind();
402     }
403 
404     bool IsASelect = isa<SelectInst>(Cur);
405 
406     // A conditional reduction operation must only have 2 or less uses in
407     // VisitedInsts.
408     if (IsASelect && (Kind == RecurKind::FAdd || Kind == RecurKind::FMul) &&
409         hasMultipleUsesOf(Cur, VisitedInsts, 2))
410       return false;
411 
412     // A reduction operation must only have one use of the reduction value.
413     if (!IsAPhi && !IsASelect && !isMinMaxRecurrenceKind(Kind) &&
414         !isAnyOfRecurrenceKind(Kind) && hasMultipleUsesOf(Cur, VisitedInsts, 1))
415       return false;
416 
417     // All inputs to a PHI node must be a reduction value.
418     if (IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
419       return false;
420 
421     if (isIntMinMaxRecurrenceKind(Kind) && (isa<ICmpInst>(Cur) || IsASelect))
422       ++NumCmpSelectPatternInst;
423     if (isFPMinMaxRecurrenceKind(Kind) && (isa<FCmpInst>(Cur) || IsASelect))
424       ++NumCmpSelectPatternInst;
425     if (isAnyOfRecurrenceKind(Kind) && IsASelect)
426       ++NumCmpSelectPatternInst;
427 
428     // Check  whether we found a reduction operator.
429     FoundReduxOp |= !IsAPhi && Cur != Start;
430 
431     // Process users of current instruction. Push non-PHI nodes after PHI nodes
432     // onto the stack. This way we are going to have seen all inputs to PHI
433     // nodes once we get to them.
434     SmallVector<Instruction *, 8> NonPHIs;
435     SmallVector<Instruction *, 8> PHIs;
436     for (User *U : Cur->users()) {
437       Instruction *UI = cast<Instruction>(U);
438 
439       // If the user is a call to llvm.fmuladd then the instruction can only be
440       // the final operand.
441       if (isFMulAddIntrinsic(UI))
442         if (Cur == UI->getOperand(0) || Cur == UI->getOperand(1))
443           return false;
444 
445       // Check if we found the exit user.
446       BasicBlock *Parent = UI->getParent();
447       if (!TheLoop->contains(Parent)) {
448         // If we already know this instruction is used externally, move on to
449         // the next user.
450         if (ExitInstruction == Cur)
451           continue;
452 
453         // Exit if you find multiple values used outside or if the header phi
454         // node is being used. In this case the user uses the value of the
455         // previous iteration, in which case we would loose "VF-1" iterations of
456         // the reduction operation if we vectorize.
457         if (ExitInstruction != nullptr || Cur == Phi)
458           return false;
459 
460         // The instruction used by an outside user must be the last instruction
461         // before we feed back to the reduction phi. Otherwise, we loose VF-1
462         // operations on the value.
463         if (!is_contained(Phi->operands(), Cur))
464           return false;
465 
466         ExitInstruction = Cur;
467         continue;
468       }
469 
470       // Process instructions only once (termination). Each reduction cycle
471       // value must only be used once, except by phi nodes and min/max
472       // reductions which are represented as a cmp followed by a select.
473       InstDesc IgnoredVal(false, nullptr);
474       if (VisitedInsts.insert(UI).second) {
475         if (isa<PHINode>(UI)) {
476           PHIs.push_back(UI);
477         } else {
478           StoreInst *SI = dyn_cast<StoreInst>(UI);
479           if (SI && SI->getPointerOperand() == Cur) {
480             // Reduction variable chain can only be stored somewhere but it
481             // can't be used as an address.
482             return false;
483           }
484           NonPHIs.push_back(UI);
485         }
486       } else if (!isa<PHINode>(UI) &&
487                  ((!isa<FCmpInst>(UI) && !isa<ICmpInst>(UI) &&
488                    !isa<SelectInst>(UI)) ||
489                   (!isConditionalRdxPattern(UI).isRecurrence() &&
490                    !isAnyOfPattern(TheLoop, Phi, UI, IgnoredVal)
491                         .isRecurrence() &&
492                    !isMinMaxPattern(UI, Kind, IgnoredVal).isRecurrence())))
493         return false;
494 
495       // Remember that we completed the cycle.
496       if (UI == Phi)
497         FoundStartPHI = true;
498     }
499     Worklist.append(PHIs.begin(), PHIs.end());
500     Worklist.append(NonPHIs.begin(), NonPHIs.end());
501   }
502 
503   // This means we have seen one but not the other instruction of the
504   // pattern or more than just a select and cmp. Zero implies that we saw a
505   // llvm.min/max intrinsic, which is always OK.
506   if (isMinMaxRecurrenceKind(Kind) && NumCmpSelectPatternInst != 2 &&
507       NumCmpSelectPatternInst != 0)
508     return false;
509 
510   if (isAnyOfRecurrenceKind(Kind) && NumCmpSelectPatternInst != 1)
511     return false;
512 
513   if (IntermediateStore) {
514     // Check that stored value goes to the phi node again. This way we make sure
515     // that the value stored in IntermediateStore is indeed the final reduction
516     // value.
517     if (!is_contained(Phi->operands(), IntermediateStore->getValueOperand())) {
518       LLVM_DEBUG(dbgs() << "Not a final reduction value stored: "
519                         << *IntermediateStore << '\n');
520       return false;
521     }
522 
523     // If there is an exit instruction it's value should be stored in
524     // IntermediateStore
525     if (ExitInstruction &&
526         IntermediateStore->getValueOperand() != ExitInstruction) {
527       LLVM_DEBUG(dbgs() << "Last store Instruction of reduction value does not "
528                            "store last calculated value of the reduction: "
529                         << *IntermediateStore << '\n');
530       return false;
531     }
532 
533     // If all uses are inside the loop (intermediate stores), then the
534     // reduction value after the loop will be the one used in the last store.
535     if (!ExitInstruction)
536       ExitInstruction = cast<Instruction>(IntermediateStore->getValueOperand());
537   }
538 
539   if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
540     return false;
541 
542   const bool IsOrdered =
543       checkOrderedReduction(Kind, ExactFPMathInst, ExitInstruction, Phi);
544 
545   if (Start != Phi) {
546     // If the starting value is not the same as the phi node, we speculatively
547     // looked through an 'and' instruction when evaluating a potential
548     // arithmetic reduction to determine if it may have been type-promoted.
549     //
550     // We now compute the minimal bit width that is required to represent the
551     // reduction. If this is the same width that was indicated by the 'and', we
552     // can represent the reduction in the smaller type. The 'and' instruction
553     // will be eliminated since it will essentially be a cast instruction that
554     // can be ignore in the cost model. If we compute a different type than we
555     // did when evaluating the 'and', the 'and' will not be eliminated, and we
556     // will end up with different kinds of operations in the recurrence
557     // expression (e.g., IntegerAND, IntegerADD). We give up if this is
558     // the case.
559     //
560     // The vectorizer relies on InstCombine to perform the actual
561     // type-shrinking. It does this by inserting instructions to truncate the
562     // exit value of the reduction to the width indicated by RecurrenceType and
563     // then extend this value back to the original width. If IsSigned is false,
564     // a 'zext' instruction will be generated; otherwise, a 'sext' will be
565     // used.
566     //
567     // TODO: We should not rely on InstCombine to rewrite the reduction in the
568     //       smaller type. We should just generate a correctly typed expression
569     //       to begin with.
570     Type *ComputedType;
571     std::tie(ComputedType, IsSigned) =
572         computeRecurrenceType(ExitInstruction, DB, AC, DT);
573     if (ComputedType != RecurrenceType)
574       return false;
575   }
576 
577   // Collect cast instructions and the minimum width used by the recurrence.
578   // If the starting value is not the same as the phi node and the computed
579   // recurrence type is equal to the recurrence type, the recurrence expression
580   // will be represented in a narrower or wider type. If there are any cast
581   // instructions that will be unnecessary, collect them in CastsFromRecurTy.
582   // Note that the 'and' instruction was already included in this list.
583   //
584   // TODO: A better way to represent this may be to tag in some way all the
585   //       instructions that are a part of the reduction. The vectorizer cost
586   //       model could then apply the recurrence type to these instructions,
587   //       without needing a white list of instructions to ignore.
588   //       This may also be useful for the inloop reductions, if it can be
589   //       kept simple enough.
590   collectCastInstrs(TheLoop, ExitInstruction, RecurrenceType, CastInsts,
591                     MinWidthCastToRecurrenceType);
592 
593   // We found a reduction var if we have reached the original phi node and we
594   // only have a single instruction with out-of-loop users.
595 
596   // The ExitInstruction(Instruction which is allowed to have out-of-loop users)
597   // is saved as part of the RecurrenceDescriptor.
598 
599   // Save the description of this reduction variable.
600   RecurrenceDescriptor RD(RdxStart, ExitInstruction, IntermediateStore, Kind,
601                           FMF, ExactFPMathInst, RecurrenceType, IsSigned,
602                           IsOrdered, CastInsts, MinWidthCastToRecurrenceType);
603   RedDes = RD;
604 
605   return true;
606 }
607 
608 // We are looking for loops that do something like this:
609 //   int r = 0;
610 //   for (int i = 0; i < n; i++) {
611 //     if (src[i] > 3)
612 //       r = 3;
613 //   }
614 // where the reduction value (r) only has two states, in this example 0 or 3.
615 // The generated LLVM IR for this type of loop will be like this:
616 //   for.body:
617 //     %r = phi i32 [ %spec.select, %for.body ], [ 0, %entry ]
618 //     ...
619 //     %cmp = icmp sgt i32 %5, 3
620 //     %spec.select = select i1 %cmp, i32 3, i32 %r
621 //     ...
622 // In general we can support vectorization of loops where 'r' flips between
623 // any two non-constants, provided they are loop invariant. The only thing
624 // we actually care about at the end of the loop is whether or not any lane
625 // in the selected vector is different from the start value. The final
626 // across-vector reduction after the loop simply involves choosing the start
627 // value if nothing changed (0 in the example above) or the other selected
628 // value (3 in the example above).
629 RecurrenceDescriptor::InstDesc
isAnyOfPattern(Loop * Loop,PHINode * OrigPhi,Instruction * I,InstDesc & Prev)630 RecurrenceDescriptor::isAnyOfPattern(Loop *Loop, PHINode *OrigPhi,
631                                      Instruction *I, InstDesc &Prev) {
632   // We must handle the select(cmp(),x,y) as a single instruction. Advance to
633   // the select.
634   if (match(I, m_OneUse(m_Cmp()))) {
635     if (auto *Select = dyn_cast<SelectInst>(*I->user_begin()))
636       return InstDesc(Select, Prev.getRecKind());
637   }
638 
639   if (!match(I, m_Select(m_Cmp(), m_Value(), m_Value())))
640     return InstDesc(false, I);
641 
642   SelectInst *SI = cast<SelectInst>(I);
643   Value *NonPhi = nullptr;
644 
645   if (OrigPhi == dyn_cast<PHINode>(SI->getTrueValue()))
646     NonPhi = SI->getFalseValue();
647   else if (OrigPhi == dyn_cast<PHINode>(SI->getFalseValue()))
648     NonPhi = SI->getTrueValue();
649   else
650     return InstDesc(false, I);
651 
652   // We are looking for selects of the form:
653   //   select(cmp(), phi, loop_invariant) or
654   //   select(cmp(), loop_invariant, phi)
655   if (!Loop->isLoopInvariant(NonPhi))
656     return InstDesc(false, I);
657 
658   return InstDesc(I, RecurKind::AnyOf);
659 }
660 
661 // We are looking for loops that do something like this:
662 //   int r = 0;
663 //   for (int i = 0; i < n; i++) {
664 //     if (src[i] > 3)
665 //       r = i;
666 //   }
667 // The reduction value (r) is derived from either the values of an induction
668 // variable (i) sequence, or from the start value (0). The LLVM IR generated for
669 // such loops would be as follows:
670 //   for.body:
671 //     %r = phi i32 [ %spec.select, %for.body ], [ 0, %entry ]
672 //     %i = phi i32 [ %inc, %for.body ], [ 0, %entry ]
673 //     ...
674 //     %cmp = icmp sgt i32 %5, 3
675 //     %spec.select = select i1 %cmp, i32 %i, i32 %r
676 //     %inc = add nsw i32 %i, 1
677 //     ...
678 // Since 'i' is an induction variable, the reduction value after the loop will
679 // be the maximum (increasing induction) or minimum (decreasing induction) value
680 // of 'i' that the condition (src[i] > 3) is satisfied, or the start value (0 in
681 // the example above). When the start value of the induction variable 'i' is
682 // greater than the minimum (increasing induction) or maximum (decreasing
683 // induction) value of the data type, we can use the minimum (increasing
684 // induction) or maximum (decreasing induction) value of the data type as a
685 // sentinel value to replace the start value. This allows us to perform a single
686 // reduction max (increasing induction) or min (decreasing induction) operation
687 // to obtain the final reduction result.
688 // TODO: It is possible to solve the case where the start value is the minimum
689 // value of the data type or a non-constant value by using mask and multiple
690 // reduction operations.
691 RecurrenceDescriptor::InstDesc
isFindIVPattern(RecurKind Kind,Loop * TheLoop,PHINode * OrigPhi,Instruction * I,ScalarEvolution & SE)692 RecurrenceDescriptor::isFindIVPattern(RecurKind Kind, Loop *TheLoop,
693                                       PHINode *OrigPhi, Instruction *I,
694                                       ScalarEvolution &SE) {
695   // TODO: Support the vectorization of FindLastIV when the reduction phi is
696   // used by more than one select instruction. This vectorization is only
697   // performed when the SCEV of each increasing induction variable used by the
698   // select instructions is identical.
699   if (!OrigPhi->hasOneUse())
700     return InstDesc(false, I);
701 
702   // We are looking for selects of the form:
703   //   select(cmp(), phi, loop_induction) or
704   //   select(cmp(), loop_induction, phi)
705   // TODO: Match selects with multi-use cmp conditions.
706   Value *NonRdxPhi = nullptr;
707   if (!match(I, m_CombineOr(m_Select(m_OneUse(m_Cmp()), m_Value(NonRdxPhi),
708                                      m_Specific(OrigPhi)),
709                             m_Select(m_OneUse(m_Cmp()), m_Specific(OrigPhi),
710                                      m_Value(NonRdxPhi)))))
711     return InstDesc(false, I);
712 
713   // Returns either FindFirstIV/FindLastIV, if such a pattern is found, or
714   // std::nullopt.
715   auto GetRecurKind = [&](Value *V) -> std::optional<RecurKind> {
716     Type *Ty = V->getType();
717     if (!SE.isSCEVable(Ty))
718       return std::nullopt;
719 
720     auto *AR = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(V));
721     if (!AR || AR->getLoop() != TheLoop)
722       return std::nullopt;
723 
724     const SCEV *Step = AR->getStepRecurrence(SE);
725     if ((isFindFirstIVRecurrenceKind(Kind) && !SE.isKnownNegative(Step)) ||
726         (isFindLastIVRecurrenceKind(Kind) && !SE.isKnownPositive(Step)))
727       return std::nullopt;
728 
729     // Check if the minimum (FindLast) or maximum (FindFirst) value of the
730     // recurrence type can be used as a sentinel value. The maximum acceptable
731     // range for the induction variable, called the valid range will exclude
732     // <sentinel value>, where <sentinel value> is
733     // [Signed|Unsigned]Min(<recurrence type>) for FindLastIV or
734     // [Signed|Unsigned]Max(<recurrence type>) for FindFirstIV.
735     // TODO: This range restriction can be lifted by adding an additional
736     // virtual OR reduction.
737     auto CheckRange = [&](bool IsSigned) {
738       const ConstantRange IVRange =
739           IsSigned ? SE.getSignedRange(AR) : SE.getUnsignedRange(AR);
740       unsigned NumBits = Ty->getIntegerBitWidth();
741       ConstantRange ValidRange = ConstantRange::getEmpty(NumBits);
742       if (isFindLastIVRecurrenceKind(Kind)) {
743         APInt Sentinel = IsSigned ? APInt::getSignedMinValue(NumBits)
744                                   : APInt::getMinValue(NumBits);
745         ValidRange = ConstantRange::getNonEmpty(Sentinel + 1, Sentinel);
746       } else {
747         if (IsSigned)
748           ValidRange =
749               ConstantRange::getNonEmpty(APInt::getSignedMinValue(NumBits),
750                                          APInt::getSignedMaxValue(NumBits) - 1);
751         else
752           ValidRange = ConstantRange::getNonEmpty(
753               APInt::getMinValue(NumBits), APInt::getMaxValue(NumBits) - 1);
754       }
755 
756       LLVM_DEBUG(dbgs() << "LV: "
757                         << (isFindLastIVRecurrenceKind(Kind) ? "FindLastIV"
758                                                              : "FindFirstIV")
759                         << " valid range is " << ValidRange
760                         << ", and the range of " << *AR << " is " << IVRange
761                         << "\n");
762 
763       // Ensure the induction variable does not wrap around by verifying that
764       // its range is fully contained within the valid range.
765       return ValidRange.contains(IVRange);
766     };
767     if (isFindLastIVRecurrenceKind(Kind)) {
768       if (CheckRange(true))
769         return RecurKind::FindLastIVSMax;
770       if (CheckRange(false))
771         return RecurKind::FindLastIVUMax;
772       return std::nullopt;
773     }
774     assert(isFindFirstIVRecurrenceKind(Kind) &&
775            "Kind must either be a FindLastIV or FindFirstIV");
776 
777     if (CheckRange(true))
778       return RecurKind::FindFirstIVSMin;
779     if (CheckRange(false))
780       return RecurKind::FindFirstIVUMin;
781     return std::nullopt;
782   };
783 
784   if (auto RK = GetRecurKind(NonRdxPhi))
785     return InstDesc(I, *RK);
786 
787   return InstDesc(false, I);
788 }
789 
790 RecurrenceDescriptor::InstDesc
isMinMaxPattern(Instruction * I,RecurKind Kind,const InstDesc & Prev)791 RecurrenceDescriptor::isMinMaxPattern(Instruction *I, RecurKind Kind,
792                                       const InstDesc &Prev) {
793   assert((isa<CmpInst>(I) || isa<SelectInst>(I) || isa<CallInst>(I)) &&
794          "Expected a cmp or select or call instruction");
795   if (!isMinMaxRecurrenceKind(Kind))
796     return InstDesc(false, I);
797 
798   // We must handle the select(cmp()) as a single instruction. Advance to the
799   // select.
800   if (match(I, m_OneUse(m_Cmp()))) {
801     if (auto *Select = dyn_cast<SelectInst>(*I->user_begin()))
802       return InstDesc(Select, Prev.getRecKind());
803   }
804 
805   // Only match select with single use cmp condition, or a min/max intrinsic.
806   if (!isa<IntrinsicInst>(I) &&
807       !match(I, m_Select(m_OneUse(m_Cmp()), m_Value(), m_Value())))
808     return InstDesc(false, I);
809 
810   // Look for a min/max pattern.
811   if (match(I, m_UMin(m_Value(), m_Value())))
812     return InstDesc(Kind == RecurKind::UMin, I);
813   if (match(I, m_UMax(m_Value(), m_Value())))
814     return InstDesc(Kind == RecurKind::UMax, I);
815   if (match(I, m_SMax(m_Value(), m_Value())))
816     return InstDesc(Kind == RecurKind::SMax, I);
817   if (match(I, m_SMin(m_Value(), m_Value())))
818     return InstDesc(Kind == RecurKind::SMin, I);
819   if (match(I, m_OrdOrUnordFMin(m_Value(), m_Value())))
820     return InstDesc(Kind == RecurKind::FMin, I);
821   if (match(I, m_OrdOrUnordFMax(m_Value(), m_Value())))
822     return InstDesc(Kind == RecurKind::FMax, I);
823   if (match(I, m_FMinNum(m_Value(), m_Value())))
824     return InstDesc(Kind == RecurKind::FMin, I);
825   if (match(I, m_FMaxNum(m_Value(), m_Value())))
826     return InstDesc(Kind == RecurKind::FMax, I);
827   if (match(I, m_FMinimumNum(m_Value(), m_Value())))
828     return InstDesc(Kind == RecurKind::FMinimumNum, I);
829   if (match(I, m_FMaximumNum(m_Value(), m_Value())))
830     return InstDesc(Kind == RecurKind::FMaximumNum, I);
831   if (match(I, m_FMinimum(m_Value(), m_Value())))
832     return InstDesc(Kind == RecurKind::FMinimum, I);
833   if (match(I, m_FMaximum(m_Value(), m_Value())))
834     return InstDesc(Kind == RecurKind::FMaximum, I);
835 
836   return InstDesc(false, I);
837 }
838 
839 /// Returns true if the select instruction has users in the compare-and-add
840 /// reduction pattern below. The select instruction argument is the last one
841 /// in the sequence.
842 ///
843 /// %sum.1 = phi ...
844 /// ...
845 /// %cmp = fcmp pred %0, %CFP
846 /// %add = fadd %0, %sum.1
847 /// %sum.2 = select %cmp, %add, %sum.1
848 RecurrenceDescriptor::InstDesc
isConditionalRdxPattern(Instruction * I)849 RecurrenceDescriptor::isConditionalRdxPattern(Instruction *I) {
850   SelectInst *SI = dyn_cast<SelectInst>(I);
851   if (!SI)
852     return InstDesc(false, I);
853 
854   CmpInst *CI = dyn_cast<CmpInst>(SI->getCondition());
855   // Only handle single use cases for now.
856   if (!CI || !CI->hasOneUse())
857     return InstDesc(false, I);
858 
859   Value *TrueVal = SI->getTrueValue();
860   Value *FalseVal = SI->getFalseValue();
861   // Handle only when either of operands of select instruction is a PHI
862   // node for now.
863   if ((isa<PHINode>(TrueVal) && isa<PHINode>(FalseVal)) ||
864       (!isa<PHINode>(TrueVal) && !isa<PHINode>(FalseVal)))
865     return InstDesc(false, I);
866 
867   Instruction *I1 = isa<PHINode>(TrueVal) ? dyn_cast<Instruction>(FalseVal)
868                                           : dyn_cast<Instruction>(TrueVal);
869   if (!I1 || !I1->isBinaryOp())
870     return InstDesc(false, I);
871 
872   Value *Op1, *Op2;
873   if (!(((m_FAdd(m_Value(Op1), m_Value(Op2)).match(I1) ||
874           m_FSub(m_Value(Op1), m_Value(Op2)).match(I1)) &&
875          I1->isFast()) ||
876         (m_FMul(m_Value(Op1), m_Value(Op2)).match(I1) && (I1->isFast())) ||
877         ((m_Add(m_Value(Op1), m_Value(Op2)).match(I1) ||
878           m_Sub(m_Value(Op1), m_Value(Op2)).match(I1))) ||
879         (m_Mul(m_Value(Op1), m_Value(Op2)).match(I1))))
880     return InstDesc(false, I);
881 
882   Instruction *IPhi = isa<PHINode>(Op1) ? dyn_cast<Instruction>(Op1)
883                                         : dyn_cast<Instruction>(Op2);
884   if (!IPhi || IPhi != FalseVal)
885     return InstDesc(false, I);
886 
887   return InstDesc(true, SI);
888 }
889 
isRecurrenceInstr(Loop * L,PHINode * OrigPhi,Instruction * I,RecurKind Kind,InstDesc & Prev,FastMathFlags FuncFMF,ScalarEvolution * SE)890 RecurrenceDescriptor::InstDesc RecurrenceDescriptor::isRecurrenceInstr(
891     Loop *L, PHINode *OrigPhi, Instruction *I, RecurKind Kind, InstDesc &Prev,
892     FastMathFlags FuncFMF, ScalarEvolution *SE) {
893   assert(Prev.getRecKind() == RecurKind::None || Prev.getRecKind() == Kind);
894   switch (I->getOpcode()) {
895   default:
896     return InstDesc(false, I);
897   case Instruction::PHI:
898     return InstDesc(I, Prev.getRecKind(), Prev.getExactFPMathInst());
899   case Instruction::Sub:
900   case Instruction::Add:
901     return InstDesc(Kind == RecurKind::Add, I);
902   case Instruction::Mul:
903     return InstDesc(Kind == RecurKind::Mul, I);
904   case Instruction::And:
905     return InstDesc(Kind == RecurKind::And, I);
906   case Instruction::Or:
907     return InstDesc(Kind == RecurKind::Or, I);
908   case Instruction::Xor:
909     return InstDesc(Kind == RecurKind::Xor, I);
910   case Instruction::FDiv:
911   case Instruction::FMul:
912     return InstDesc(Kind == RecurKind::FMul, I,
913                     I->hasAllowReassoc() ? nullptr : I);
914   case Instruction::FSub:
915   case Instruction::FAdd:
916     return InstDesc(Kind == RecurKind::FAdd, I,
917                     I->hasAllowReassoc() ? nullptr : I);
918   case Instruction::Select:
919     if (Kind == RecurKind::FAdd || Kind == RecurKind::FMul ||
920         Kind == RecurKind::Add || Kind == RecurKind::Mul)
921       return isConditionalRdxPattern(I);
922     if (isFindIVRecurrenceKind(Kind) && SE)
923       return isFindIVPattern(Kind, L, OrigPhi, I, *SE);
924     [[fallthrough]];
925   case Instruction::FCmp:
926   case Instruction::ICmp:
927   case Instruction::Call:
928     if (isAnyOfRecurrenceKind(Kind))
929       return isAnyOfPattern(L, OrigPhi, I, Prev);
930     auto HasRequiredFMF = [&]() {
931      if (FuncFMF.noNaNs() && FuncFMF.noSignedZeros())
932        return true;
933      if (isa<FPMathOperator>(I) && I->hasNoNaNs() && I->hasNoSignedZeros())
934        return true;
935      // minimum/minnum and maximum/maxnum intrinsics do not require nsz and nnan
936      // flags since NaN and signed zeroes are propagated in the intrinsic
937      // implementation.
938      return match(I, m_Intrinsic<Intrinsic::minimum>(m_Value(), m_Value())) ||
939             match(I, m_Intrinsic<Intrinsic::maximum>(m_Value(), m_Value())) ||
940             match(I,
941                   m_Intrinsic<Intrinsic::minimumnum>(m_Value(), m_Value())) ||
942             match(I, m_Intrinsic<Intrinsic::maximumnum>(m_Value(), m_Value()));
943     };
944     if (isIntMinMaxRecurrenceKind(Kind))
945       return isMinMaxPattern(I, Kind, Prev);
946     if (isFPMinMaxRecurrenceKind(Kind)) {
947       InstDesc Res = isMinMaxPattern(I, Kind, Prev);
948       if (!Res.isRecurrence())
949         return InstDesc(false, I);
950       if (HasRequiredFMF())
951         return Res;
952       // We may be able to vectorize FMax/FMin reductions using maxnum/minnum
953       // intrinsics with extra checks ensuring the vector loop handles only
954       // non-NaN inputs.
955       if (match(I, m_Intrinsic<Intrinsic::maxnum>(m_Value(), m_Value()))) {
956         assert(Kind == RecurKind::FMax &&
957                "unexpected recurrence kind for maxnum");
958         return InstDesc(I, RecurKind::FMaxNum);
959       }
960       if (match(I, m_Intrinsic<Intrinsic::minnum>(m_Value(), m_Value()))) {
961         assert(Kind == RecurKind::FMin &&
962                "unexpected recurrence kind for minnum");
963         return InstDesc(I, RecurKind::FMinNum);
964       }
965       return InstDesc(false, I);
966     }
967     if (isFMulAddIntrinsic(I))
968       return InstDesc(Kind == RecurKind::FMulAdd, I,
969                       I->hasAllowReassoc() ? nullptr : I);
970     return InstDesc(false, I);
971   }
972 }
973 
hasMultipleUsesOf(Instruction * I,SmallPtrSetImpl<Instruction * > & Insts,unsigned MaxNumUses)974 bool RecurrenceDescriptor::hasMultipleUsesOf(
975     Instruction *I, SmallPtrSetImpl<Instruction *> &Insts,
976     unsigned MaxNumUses) {
977   unsigned NumUses = 0;
978   for (const Use &U : I->operands()) {
979     if (Insts.count(dyn_cast<Instruction>(U)))
980       ++NumUses;
981     if (NumUses > MaxNumUses)
982       return true;
983   }
984 
985   return false;
986 }
987 
isReductionPHI(PHINode * Phi,Loop * TheLoop,RecurrenceDescriptor & RedDes,DemandedBits * DB,AssumptionCache * AC,DominatorTree * DT,ScalarEvolution * SE)988 bool RecurrenceDescriptor::isReductionPHI(PHINode *Phi, Loop *TheLoop,
989                                           RecurrenceDescriptor &RedDes,
990                                           DemandedBits *DB, AssumptionCache *AC,
991                                           DominatorTree *DT,
992                                           ScalarEvolution *SE) {
993   BasicBlock *Header = TheLoop->getHeader();
994   Function &F = *Header->getParent();
995   FastMathFlags FMF;
996   FMF.setNoNaNs(
997       F.getFnAttribute("no-nans-fp-math").getValueAsBool());
998   FMF.setNoSignedZeros(
999       F.getFnAttribute("no-signed-zeros-fp-math").getValueAsBool());
1000 
1001   if (AddReductionVar(Phi, RecurKind::Add, TheLoop, FMF, RedDes, DB, AC, DT,
1002                       SE)) {
1003     LLVM_DEBUG(dbgs() << "Found an ADD reduction PHI." << *Phi << "\n");
1004     return true;
1005   }
1006   if (AddReductionVar(Phi, RecurKind::Mul, TheLoop, FMF, RedDes, DB, AC, DT,
1007                       SE)) {
1008     LLVM_DEBUG(dbgs() << "Found a MUL reduction PHI." << *Phi << "\n");
1009     return true;
1010   }
1011   if (AddReductionVar(Phi, RecurKind::Or, TheLoop, FMF, RedDes, DB, AC, DT,
1012                       SE)) {
1013     LLVM_DEBUG(dbgs() << "Found an OR reduction PHI." << *Phi << "\n");
1014     return true;
1015   }
1016   if (AddReductionVar(Phi, RecurKind::And, TheLoop, FMF, RedDes, DB, AC, DT,
1017                       SE)) {
1018     LLVM_DEBUG(dbgs() << "Found an AND reduction PHI." << *Phi << "\n");
1019     return true;
1020   }
1021   if (AddReductionVar(Phi, RecurKind::Xor, TheLoop, FMF, RedDes, DB, AC, DT,
1022                       SE)) {
1023     LLVM_DEBUG(dbgs() << "Found a XOR reduction PHI." << *Phi << "\n");
1024     return true;
1025   }
1026   if (AddReductionVar(Phi, RecurKind::SMax, TheLoop, FMF, RedDes, DB, AC, DT,
1027                       SE)) {
1028     LLVM_DEBUG(dbgs() << "Found a SMAX reduction PHI." << *Phi << "\n");
1029     return true;
1030   }
1031   if (AddReductionVar(Phi, RecurKind::SMin, TheLoop, FMF, RedDes, DB, AC, DT,
1032                       SE)) {
1033     LLVM_DEBUG(dbgs() << "Found a SMIN reduction PHI." << *Phi << "\n");
1034     return true;
1035   }
1036   if (AddReductionVar(Phi, RecurKind::UMax, TheLoop, FMF, RedDes, DB, AC, DT,
1037                       SE)) {
1038     LLVM_DEBUG(dbgs() << "Found a UMAX reduction PHI." << *Phi << "\n");
1039     return true;
1040   }
1041   if (AddReductionVar(Phi, RecurKind::UMin, TheLoop, FMF, RedDes, DB, AC, DT,
1042                       SE)) {
1043     LLVM_DEBUG(dbgs() << "Found a UMIN reduction PHI." << *Phi << "\n");
1044     return true;
1045   }
1046   if (AddReductionVar(Phi, RecurKind::AnyOf, TheLoop, FMF, RedDes, DB, AC, DT,
1047                       SE)) {
1048     LLVM_DEBUG(dbgs() << "Found a conditional select reduction PHI." << *Phi
1049                       << "\n");
1050     return true;
1051   }
1052   if (AddReductionVar(Phi, RecurKind::FindLastIVSMax, TheLoop, FMF, RedDes, DB,
1053                       AC, DT, SE)) {
1054     LLVM_DEBUG(dbgs() << "Found a FindLastIV reduction PHI." << *Phi << "\n");
1055     return true;
1056   }
1057   if (AddReductionVar(Phi, RecurKind::FindFirstIVSMin, TheLoop, FMF, RedDes, DB,
1058                       AC, DT, SE)) {
1059     LLVM_DEBUG(dbgs() << "Found a FindFirstIV reduction PHI." << *Phi << "\n");
1060     return true;
1061   }
1062   if (AddReductionVar(Phi, RecurKind::FMul, TheLoop, FMF, RedDes, DB, AC, DT,
1063                       SE)) {
1064     LLVM_DEBUG(dbgs() << "Found an FMult reduction PHI." << *Phi << "\n");
1065     return true;
1066   }
1067   if (AddReductionVar(Phi, RecurKind::FAdd, TheLoop, FMF, RedDes, DB, AC, DT,
1068                       SE)) {
1069     LLVM_DEBUG(dbgs() << "Found an FAdd reduction PHI." << *Phi << "\n");
1070     return true;
1071   }
1072   if (AddReductionVar(Phi, RecurKind::FMax, TheLoop, FMF, RedDes, DB, AC, DT,
1073                       SE)) {
1074     LLVM_DEBUG(dbgs() << "Found a float MAX reduction PHI." << *Phi << "\n");
1075     return true;
1076   }
1077   if (AddReductionVar(Phi, RecurKind::FMin, TheLoop, FMF, RedDes, DB, AC, DT,
1078                       SE)) {
1079     LLVM_DEBUG(dbgs() << "Found a float MIN reduction PHI." << *Phi << "\n");
1080     return true;
1081   }
1082   if (AddReductionVar(Phi, RecurKind::FMulAdd, TheLoop, FMF, RedDes, DB, AC, DT,
1083                       SE)) {
1084     LLVM_DEBUG(dbgs() << "Found an FMulAdd reduction PHI." << *Phi << "\n");
1085     return true;
1086   }
1087   if (AddReductionVar(Phi, RecurKind::FMaximum, TheLoop, FMF, RedDes, DB, AC, DT,
1088                       SE)) {
1089     LLVM_DEBUG(dbgs() << "Found a float MAXIMUM reduction PHI." << *Phi << "\n");
1090     return true;
1091   }
1092   if (AddReductionVar(Phi, RecurKind::FMinimum, TheLoop, FMF, RedDes, DB, AC, DT,
1093                       SE)) {
1094     LLVM_DEBUG(dbgs() << "Found a float MINIMUM reduction PHI." << *Phi << "\n");
1095     return true;
1096   }
1097   if (AddReductionVar(Phi, RecurKind::FMaximumNum, TheLoop, FMF, RedDes, DB, AC,
1098                       DT, SE)) {
1099     LLVM_DEBUG(dbgs() << "Found a float MAXIMUMNUM reduction PHI." << *Phi
1100                       << "\n");
1101     return true;
1102   }
1103   if (AddReductionVar(Phi, RecurKind::FMinimumNum, TheLoop, FMF, RedDes, DB, AC,
1104                       DT, SE)) {
1105     LLVM_DEBUG(dbgs() << "Found a float MINIMUMNUM reduction PHI." << *Phi
1106                       << "\n");
1107     return true;
1108   }
1109 
1110   // Not a reduction of known type.
1111   return false;
1112 }
1113 
isFixedOrderRecurrence(PHINode * Phi,Loop * TheLoop,DominatorTree * DT)1114 bool RecurrenceDescriptor::isFixedOrderRecurrence(PHINode *Phi, Loop *TheLoop,
1115                                                   DominatorTree *DT) {
1116 
1117   // Ensure the phi node is in the loop header and has two incoming values.
1118   if (Phi->getParent() != TheLoop->getHeader() ||
1119       Phi->getNumIncomingValues() != 2)
1120     return false;
1121 
1122   // Ensure the loop has a preheader and a single latch block. The loop
1123   // vectorizer will need the latch to set up the next iteration of the loop.
1124   auto *Preheader = TheLoop->getLoopPreheader();
1125   auto *Latch = TheLoop->getLoopLatch();
1126   if (!Preheader || !Latch)
1127     return false;
1128 
1129   // Ensure the phi node's incoming blocks are the loop preheader and latch.
1130   if (Phi->getBasicBlockIndex(Preheader) < 0 ||
1131       Phi->getBasicBlockIndex(Latch) < 0)
1132     return false;
1133 
1134   // Get the previous value. The previous value comes from the latch edge while
1135   // the initial value comes from the preheader edge.
1136   auto *Previous = dyn_cast<Instruction>(Phi->getIncomingValueForBlock(Latch));
1137 
1138   // If Previous is a phi in the header, go through incoming values from the
1139   // latch until we find a non-phi value. Use this as the new Previous, all uses
1140   // in the header will be dominated by the original phi, but need to be moved
1141   // after the non-phi previous value.
1142   SmallPtrSet<PHINode *, 4> SeenPhis;
1143   while (auto *PrevPhi = dyn_cast_or_null<PHINode>(Previous)) {
1144     if (PrevPhi->getParent() != Phi->getParent())
1145       return false;
1146     if (!SeenPhis.insert(PrevPhi).second)
1147       return false;
1148     Previous = dyn_cast<Instruction>(PrevPhi->getIncomingValueForBlock(Latch));
1149   }
1150 
1151   if (!Previous || !TheLoop->contains(Previous) || isa<PHINode>(Previous))
1152     return false;
1153 
1154   // Ensure every user of the phi node (recursively) is dominated by the
1155   // previous value. The dominance requirement ensures the loop vectorizer will
1156   // not need to vectorize the initial value prior to the first iteration of the
1157   // loop.
1158   // TODO: Consider extending this sinking to handle memory instructions.
1159 
1160   SmallPtrSet<Value *, 8> Seen;
1161   BasicBlock *PhiBB = Phi->getParent();
1162   SmallVector<Instruction *, 8> WorkList;
1163   auto TryToPushSinkCandidate = [&](Instruction *SinkCandidate) {
1164     // Cyclic dependence.
1165     if (Previous == SinkCandidate)
1166       return false;
1167 
1168     if (!Seen.insert(SinkCandidate).second)
1169       return true;
1170     if (DT->dominates(Previous,
1171                       SinkCandidate)) // We already are good w/o sinking.
1172       return true;
1173 
1174     if (SinkCandidate->getParent() != PhiBB ||
1175         SinkCandidate->mayHaveSideEffects() ||
1176         SinkCandidate->mayReadFromMemory() || SinkCandidate->isTerminator())
1177       return false;
1178 
1179     // If we reach a PHI node that is not dominated by Previous, we reached a
1180     // header PHI. No need for sinking.
1181     if (isa<PHINode>(SinkCandidate))
1182       return true;
1183 
1184     // Sink User tentatively and check its users
1185     WorkList.push_back(SinkCandidate);
1186     return true;
1187   };
1188 
1189   WorkList.push_back(Phi);
1190   // Try to recursively sink instructions and their users after Previous.
1191   while (!WorkList.empty()) {
1192     Instruction *Current = WorkList.pop_back_val();
1193     for (User *User : Current->users()) {
1194       if (!TryToPushSinkCandidate(cast<Instruction>(User)))
1195         return false;
1196     }
1197   }
1198 
1199   return true;
1200 }
1201 
getOpcode(RecurKind Kind)1202 unsigned RecurrenceDescriptor::getOpcode(RecurKind Kind) {
1203   switch (Kind) {
1204   case RecurKind::Add:
1205     return Instruction::Add;
1206   case RecurKind::Mul:
1207     return Instruction::Mul;
1208   case RecurKind::AnyOf:
1209   case RecurKind::FindFirstIVSMin:
1210   case RecurKind::FindFirstIVUMin:
1211   case RecurKind::FindLastIVSMax:
1212   case RecurKind::FindLastIVUMax:
1213   case RecurKind::Or:
1214     return Instruction::Or;
1215   case RecurKind::And:
1216     return Instruction::And;
1217   case RecurKind::Xor:
1218     return Instruction::Xor;
1219   case RecurKind::FMul:
1220     return Instruction::FMul;
1221   case RecurKind::FMulAdd:
1222   case RecurKind::FAdd:
1223     return Instruction::FAdd;
1224   case RecurKind::SMax:
1225   case RecurKind::SMin:
1226   case RecurKind::UMax:
1227   case RecurKind::UMin:
1228     return Instruction::ICmp;
1229   case RecurKind::FMax:
1230   case RecurKind::FMin:
1231   case RecurKind::FMaximum:
1232   case RecurKind::FMinimum:
1233   case RecurKind::FMaximumNum:
1234   case RecurKind::FMinimumNum:
1235     return Instruction::FCmp;
1236   default:
1237     llvm_unreachable("Unknown recurrence operation");
1238   }
1239 }
1240 
1241 SmallVector<Instruction *, 4>
getReductionOpChain(PHINode * Phi,Loop * L) const1242 RecurrenceDescriptor::getReductionOpChain(PHINode *Phi, Loop *L) const {
1243   SmallVector<Instruction *, 4> ReductionOperations;
1244   const bool IsMinMax = isMinMaxRecurrenceKind(Kind);
1245 
1246   // Search down from the Phi to the LoopExitInstr, looking for instructions
1247   // with a single user of the correct type for the reduction.
1248 
1249   // Note that we check that the type of the operand is correct for each item in
1250   // the chain, including the last (the loop exit value). This can come up from
1251   // sub, which would otherwise be treated as an add reduction. MinMax also need
1252   // to check for a pair of icmp/select, for which we use getNextInstruction and
1253   // isCorrectOpcode functions to step the right number of instruction, and
1254   // check the icmp/select pair.
1255   // FIXME: We also do not attempt to look through Select's yet, which might
1256   // be part of the reduction chain, or attempt to looks through And's to find a
1257   // smaller bitwidth. Subs are also currently not allowed (which are usually
1258   // treated as part of a add reduction) as they are expected to generally be
1259   // more expensive than out-of-loop reductions, and need to be costed more
1260   // carefully.
1261   unsigned ExpectedUses = 1;
1262   if (IsMinMax)
1263     ExpectedUses = 2;
1264 
1265   auto getNextInstruction = [&](Instruction *Cur) -> Instruction * {
1266     for (auto *User : Cur->users()) {
1267       Instruction *UI = cast<Instruction>(User);
1268       if (isa<PHINode>(UI))
1269         continue;
1270       if (IsMinMax) {
1271         // We are expecting a icmp/select pair, which we go to the next select
1272         // instruction if we can. We already know that Cur has 2 uses.
1273         if (isa<SelectInst>(UI))
1274           return UI;
1275         continue;
1276       }
1277       return UI;
1278     }
1279     return nullptr;
1280   };
1281   auto isCorrectOpcode = [&](Instruction *Cur) {
1282     if (IsMinMax) {
1283       Value *LHS, *RHS;
1284       return SelectPatternResult::isMinOrMax(
1285           matchSelectPattern(Cur, LHS, RHS).Flavor);
1286     }
1287     // Recognize a call to the llvm.fmuladd intrinsic.
1288     if (isFMulAddIntrinsic(Cur))
1289       return true;
1290 
1291     return Cur->getOpcode() == getOpcode();
1292   };
1293 
1294   // Attempt to look through Phis which are part of the reduction chain
1295   unsigned ExtraPhiUses = 0;
1296   Instruction *RdxInstr = LoopExitInstr;
1297   if (auto ExitPhi = dyn_cast<PHINode>(LoopExitInstr)) {
1298     if (ExitPhi->getNumIncomingValues() != 2)
1299       return {};
1300 
1301     Instruction *Inc0 = dyn_cast<Instruction>(ExitPhi->getIncomingValue(0));
1302     Instruction *Inc1 = dyn_cast<Instruction>(ExitPhi->getIncomingValue(1));
1303 
1304     Instruction *Chain = nullptr;
1305     if (Inc0 == Phi)
1306       Chain = Inc1;
1307     else if (Inc1 == Phi)
1308       Chain = Inc0;
1309     else
1310       return {};
1311 
1312     RdxInstr = Chain;
1313     ExtraPhiUses = 1;
1314   }
1315 
1316   // The loop exit instruction we check first (as a quick test) but add last. We
1317   // check the opcode is correct (and dont allow them to be Subs) and that they
1318   // have expected to have the expected number of uses. They will have one use
1319   // from the phi and one from a LCSSA value, no matter the type.
1320   if (!isCorrectOpcode(RdxInstr) || !LoopExitInstr->hasNUses(2))
1321     return {};
1322 
1323   // Check that the Phi has one (or two for min/max) uses, plus an extra use
1324   // for conditional reductions.
1325   if (!Phi->hasNUses(ExpectedUses + ExtraPhiUses))
1326     return {};
1327 
1328   Instruction *Cur = getNextInstruction(Phi);
1329 
1330   // Each other instruction in the chain should have the expected number of uses
1331   // and be the correct opcode.
1332   while (Cur != RdxInstr) {
1333     if (!Cur || !isCorrectOpcode(Cur) || !Cur->hasNUses(ExpectedUses))
1334       return {};
1335 
1336     ReductionOperations.push_back(Cur);
1337     Cur = getNextInstruction(Cur);
1338   }
1339 
1340   ReductionOperations.push_back(Cur);
1341   return ReductionOperations;
1342 }
1343 
InductionDescriptor(Value * Start,InductionKind K,const SCEV * Step,BinaryOperator * BOp,SmallVectorImpl<Instruction * > * Casts)1344 InductionDescriptor::InductionDescriptor(Value *Start, InductionKind K,
1345                                          const SCEV *Step, BinaryOperator *BOp,
1346                                          SmallVectorImpl<Instruction *> *Casts)
1347     : StartValue(Start), IK(K), Step(Step), InductionBinOp(BOp) {
1348   assert(IK != IK_NoInduction && "Not an induction");
1349 
1350   // Start value type should match the induction kind and the value
1351   // itself should not be null.
1352   assert(StartValue && "StartValue is null");
1353   assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) &&
1354          "StartValue is not a pointer for pointer induction");
1355   assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) &&
1356          "StartValue is not an integer for integer induction");
1357 
1358   // Check the Step Value. It should be non-zero integer value.
1359   assert((!getConstIntStepValue() || !getConstIntStepValue()->isZero()) &&
1360          "Step value is zero");
1361 
1362   assert((IK == IK_FpInduction || Step->getType()->isIntegerTy()) &&
1363          "StepValue is not an integer");
1364 
1365   assert((IK != IK_FpInduction || Step->getType()->isFloatingPointTy()) &&
1366          "StepValue is not FP for FpInduction");
1367   assert((IK != IK_FpInduction ||
1368           (InductionBinOp &&
1369            (InductionBinOp->getOpcode() == Instruction::FAdd ||
1370             InductionBinOp->getOpcode() == Instruction::FSub))) &&
1371          "Binary opcode should be specified for FP induction");
1372 
1373   if (Casts)
1374     llvm::append_range(RedundantCasts, *Casts);
1375 }
1376 
getConstIntStepValue() const1377 ConstantInt *InductionDescriptor::getConstIntStepValue() const {
1378   if (isa<SCEVConstant>(Step))
1379     return dyn_cast<ConstantInt>(cast<SCEVConstant>(Step)->getValue());
1380   return nullptr;
1381 }
1382 
isFPInductionPHI(PHINode * Phi,const Loop * TheLoop,ScalarEvolution * SE,InductionDescriptor & D)1383 bool InductionDescriptor::isFPInductionPHI(PHINode *Phi, const Loop *TheLoop,
1384                                            ScalarEvolution *SE,
1385                                            InductionDescriptor &D) {
1386 
1387   // Here we only handle FP induction variables.
1388   assert(Phi->getType()->isFloatingPointTy() && "Unexpected Phi type");
1389 
1390   if (TheLoop->getHeader() != Phi->getParent())
1391     return false;
1392 
1393   // The loop may have multiple entrances or multiple exits; we can analyze
1394   // this phi if it has a unique entry value and a unique backedge value.
1395   if (Phi->getNumIncomingValues() != 2)
1396     return false;
1397   Value *BEValue = nullptr, *StartValue = nullptr;
1398   if (TheLoop->contains(Phi->getIncomingBlock(0))) {
1399     BEValue = Phi->getIncomingValue(0);
1400     StartValue = Phi->getIncomingValue(1);
1401   } else {
1402     assert(TheLoop->contains(Phi->getIncomingBlock(1)) &&
1403            "Unexpected Phi node in the loop");
1404     BEValue = Phi->getIncomingValue(1);
1405     StartValue = Phi->getIncomingValue(0);
1406   }
1407 
1408   BinaryOperator *BOp = dyn_cast<BinaryOperator>(BEValue);
1409   if (!BOp)
1410     return false;
1411 
1412   Value *Addend = nullptr;
1413   if (BOp->getOpcode() == Instruction::FAdd) {
1414     if (BOp->getOperand(0) == Phi)
1415       Addend = BOp->getOperand(1);
1416     else if (BOp->getOperand(1) == Phi)
1417       Addend = BOp->getOperand(0);
1418   } else if (BOp->getOpcode() == Instruction::FSub)
1419     if (BOp->getOperand(0) == Phi)
1420       Addend = BOp->getOperand(1);
1421 
1422   if (!Addend)
1423     return false;
1424 
1425   // The addend should be loop invariant
1426   if (auto *I = dyn_cast<Instruction>(Addend))
1427     if (TheLoop->contains(I))
1428       return false;
1429 
1430   // FP Step has unknown SCEV
1431   const SCEV *Step = SE->getUnknown(Addend);
1432   D = InductionDescriptor(StartValue, IK_FpInduction, Step, BOp);
1433   return true;
1434 }
1435 
1436 /// This function is called when we suspect that the update-chain of a phi node
1437 /// (whose symbolic SCEV expression sin \p PhiScev) contains redundant casts,
1438 /// that can be ignored. (This can happen when the PSCEV rewriter adds a runtime
1439 /// predicate P under which the SCEV expression for the phi can be the
1440 /// AddRecurrence \p AR; See createAddRecFromPHIWithCast). We want to find the
1441 /// cast instructions that are involved in the update-chain of this induction.
1442 /// A caller that adds the required runtime predicate can be free to drop these
1443 /// cast instructions, and compute the phi using \p AR (instead of some scev
1444 /// expression with casts).
1445 ///
1446 /// For example, without a predicate the scev expression can take the following
1447 /// form:
1448 ///      (Ext ix (Trunc iy ( Start + i*Step ) to ix) to iy)
1449 ///
1450 /// It corresponds to the following IR sequence:
1451 /// %for.body:
1452 ///   %x = phi i64 [ 0, %ph ], [ %add, %for.body ]
1453 ///   %casted_phi = "ExtTrunc i64 %x"
1454 ///   %add = add i64 %casted_phi, %step
1455 ///
1456 /// where %x is given in \p PN,
1457 /// PSE.getSCEV(%x) is equal to PSE.getSCEV(%casted_phi) under a predicate,
1458 /// and the IR sequence that "ExtTrunc i64 %x" represents can take one of
1459 /// several forms, for example, such as:
1460 ///   ExtTrunc1:    %casted_phi = and  %x, 2^n-1
1461 /// or:
1462 ///   ExtTrunc2:    %t = shl %x, m
1463 ///                 %casted_phi = ashr %t, m
1464 ///
1465 /// If we are able to find such sequence, we return the instructions
1466 /// we found, namely %casted_phi and the instructions on its use-def chain up
1467 /// to the phi (not including the phi).
getCastsForInductionPHI(PredicatedScalarEvolution & PSE,const SCEVUnknown * PhiScev,const SCEVAddRecExpr * AR,SmallVectorImpl<Instruction * > & CastInsts)1468 static bool getCastsForInductionPHI(PredicatedScalarEvolution &PSE,
1469                                     const SCEVUnknown *PhiScev,
1470                                     const SCEVAddRecExpr *AR,
1471                                     SmallVectorImpl<Instruction *> &CastInsts) {
1472 
1473   assert(CastInsts.empty() && "CastInsts is expected to be empty.");
1474   auto *PN = cast<PHINode>(PhiScev->getValue());
1475   assert(PSE.getSCEV(PN) == AR && "Unexpected phi node SCEV expression");
1476   const Loop *L = AR->getLoop();
1477 
1478   // Find any cast instructions that participate in the def-use chain of
1479   // PhiScev in the loop.
1480   // FORNOW/TODO: We currently expect the def-use chain to include only
1481   // two-operand instructions, where one of the operands is an invariant.
1482   // createAddRecFromPHIWithCasts() currently does not support anything more
1483   // involved than that, so we keep the search simple. This can be
1484   // extended/generalized as needed.
1485 
1486   auto getDef = [&](const Value *Val) -> Value * {
1487     const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Val);
1488     if (!BinOp)
1489       return nullptr;
1490     Value *Op0 = BinOp->getOperand(0);
1491     Value *Op1 = BinOp->getOperand(1);
1492     Value *Def = nullptr;
1493     if (L->isLoopInvariant(Op0))
1494       Def = Op1;
1495     else if (L->isLoopInvariant(Op1))
1496       Def = Op0;
1497     return Def;
1498   };
1499 
1500   // Look for the instruction that defines the induction via the
1501   // loop backedge.
1502   BasicBlock *Latch = L->getLoopLatch();
1503   if (!Latch)
1504     return false;
1505   Value *Val = PN->getIncomingValueForBlock(Latch);
1506   if (!Val)
1507     return false;
1508 
1509   // Follow the def-use chain until the induction phi is reached.
1510   // If on the way we encounter a Value that has the same SCEV Expr as the
1511   // phi node, we can consider the instructions we visit from that point
1512   // as part of the cast-sequence that can be ignored.
1513   bool InCastSequence = false;
1514   auto *Inst = dyn_cast<Instruction>(Val);
1515   while (Val != PN) {
1516     // If we encountered a phi node other than PN, or if we left the loop,
1517     // we bail out.
1518     if (!Inst || !L->contains(Inst)) {
1519       return false;
1520     }
1521     auto *AddRec = dyn_cast<SCEVAddRecExpr>(PSE.getSCEV(Val));
1522     if (AddRec && PSE.areAddRecsEqualWithPreds(AddRec, AR))
1523       InCastSequence = true;
1524     if (InCastSequence) {
1525       // Only the last instruction in the cast sequence is expected to have
1526       // uses outside the induction def-use chain.
1527       if (!CastInsts.empty())
1528         if (!Inst->hasOneUse())
1529           return false;
1530       CastInsts.push_back(Inst);
1531     }
1532     Val = getDef(Val);
1533     if (!Val)
1534       return false;
1535     Inst = dyn_cast<Instruction>(Val);
1536   }
1537 
1538   return InCastSequence;
1539 }
1540 
isInductionPHI(PHINode * Phi,const Loop * TheLoop,PredicatedScalarEvolution & PSE,InductionDescriptor & D,bool Assume)1541 bool InductionDescriptor::isInductionPHI(PHINode *Phi, const Loop *TheLoop,
1542                                          PredicatedScalarEvolution &PSE,
1543                                          InductionDescriptor &D, bool Assume) {
1544   Type *PhiTy = Phi->getType();
1545 
1546   // Handle integer and pointer inductions variables.
1547   // Now we handle also FP induction but not trying to make a
1548   // recurrent expression from the PHI node in-place.
1549 
1550   if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy() && !PhiTy->isFloatTy() &&
1551       !PhiTy->isDoubleTy() && !PhiTy->isHalfTy())
1552     return false;
1553 
1554   if (PhiTy->isFloatingPointTy())
1555     return isFPInductionPHI(Phi, TheLoop, PSE.getSE(), D);
1556 
1557   const SCEV *PhiScev = PSE.getSCEV(Phi);
1558   const auto *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1559 
1560   // We need this expression to be an AddRecExpr.
1561   if (Assume && !AR)
1562     AR = PSE.getAsAddRec(Phi);
1563 
1564   if (!AR) {
1565     LLVM_DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
1566     return false;
1567   }
1568 
1569   // Record any Cast instructions that participate in the induction update
1570   const auto *SymbolicPhi = dyn_cast<SCEVUnknown>(PhiScev);
1571   // If we started from an UnknownSCEV, and managed to build an addRecurrence
1572   // only after enabling Assume with PSCEV, this means we may have encountered
1573   // cast instructions that required adding a runtime check in order to
1574   // guarantee the correctness of the AddRecurrence respresentation of the
1575   // induction.
1576   if (PhiScev != AR && SymbolicPhi) {
1577     SmallVector<Instruction *, 2> Casts;
1578     if (getCastsForInductionPHI(PSE, SymbolicPhi, AR, Casts))
1579       return isInductionPHI(Phi, TheLoop, PSE.getSE(), D, AR, &Casts);
1580   }
1581 
1582   return isInductionPHI(Phi, TheLoop, PSE.getSE(), D, AR);
1583 }
1584 
isInductionPHI(PHINode * Phi,const Loop * TheLoop,ScalarEvolution * SE,InductionDescriptor & D,const SCEV * Expr,SmallVectorImpl<Instruction * > * CastsToIgnore)1585 bool InductionDescriptor::isInductionPHI(
1586     PHINode *Phi, const Loop *TheLoop, ScalarEvolution *SE,
1587     InductionDescriptor &D, const SCEV *Expr,
1588     SmallVectorImpl<Instruction *> *CastsToIgnore) {
1589   Type *PhiTy = Phi->getType();
1590   // isSCEVable returns true for integer and pointer types.
1591   if (!SE->isSCEVable(PhiTy))
1592     return false;
1593 
1594   // Check that the PHI is consecutive.
1595   const SCEV *PhiScev = Expr ? Expr : SE->getSCEV(Phi);
1596   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1597 
1598   if (!AR) {
1599     LLVM_DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
1600     return false;
1601   }
1602 
1603   if (AR->getLoop() != TheLoop) {
1604     // FIXME: We should treat this as a uniform. Unfortunately, we
1605     // don't currently know how to handled uniform PHIs.
1606     LLVM_DEBUG(
1607         dbgs() << "LV: PHI is a recurrence with respect to an outer loop.\n");
1608     return false;
1609   }
1610 
1611   // This function assumes that InductionPhi is called only on Phi nodes
1612   // present inside loop headers. Check for the same, and throw an assert if
1613   // the current Phi is not present inside the loop header.
1614   assert(Phi->getParent() == AR->getLoop()->getHeader()
1615     && "Invalid Phi node, not present in loop header");
1616 
1617   Value *StartValue =
1618       Phi->getIncomingValueForBlock(AR->getLoop()->getLoopPreheader());
1619 
1620   BasicBlock *Latch = AR->getLoop()->getLoopLatch();
1621   if (!Latch)
1622     return false;
1623 
1624   const SCEV *Step = AR->getStepRecurrence(*SE);
1625   // Calculate the pointer stride and check if it is consecutive.
1626   // The stride may be a constant or a loop invariant integer value.
1627   const SCEVConstant *ConstStep = dyn_cast<SCEVConstant>(Step);
1628   if (!ConstStep && !SE->isLoopInvariant(Step, TheLoop))
1629     return false;
1630 
1631   if (PhiTy->isIntegerTy()) {
1632     BinaryOperator *BOp =
1633         dyn_cast<BinaryOperator>(Phi->getIncomingValueForBlock(Latch));
1634     D = InductionDescriptor(StartValue, IK_IntInduction, Step, BOp,
1635                             CastsToIgnore);
1636     return true;
1637   }
1638 
1639   assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
1640 
1641   // This allows induction variables w/non-constant steps.
1642   D = InductionDescriptor(StartValue, IK_PtrInduction, Step);
1643   return true;
1644 }
1645