xref: /netbsd-src/external/apache2/llvm/dist/llvm/lib/Analysis/IVDescriptors.cpp (revision 82d56013d7b633d116a93943de88e08335357a7c)
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 
areAllUsesIn(Instruction * I,SmallPtrSetImpl<Instruction * > & Set)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 
isIntegerRecurrenceKind(RecurKind Kind)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 
isFloatingPointRecurrenceKind(RecurKind Kind)68 bool RecurrenceDescriptor::isFloatingPointRecurrenceKind(RecurKind Kind) {
69   return (Kind != RecurKind::None) && !isIntegerRecurrenceKind(Kind);
70 }
71 
isArithmeticRecurrenceKind(RecurKind Kind)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.
lookThroughAnd(PHINode * Phi,Type * & RT,SmallPtrSetImpl<Instruction * > & Visited,SmallPtrSetImpl<Instruction * > & 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.
computeRecurrenceType(Instruction * Exit,DemandedBits * DB,AssumptionCache * AC,DominatorTree * DT)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.
collectCastsToIgnore(Loop * TheLoop,Instruction * Exit,Type * RecurrenceType,SmallPtrSetImpl<Instruction * > & Casts)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.
checkOrderedReduction(RecurKind Kind,Instruction * ExactFPMathInst,Instruction * Exit,PHINode * Phi)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 
AddReductionVar(PHINode * Phi,RecurKind Kind,Loop * TheLoop,FastMathFlags FuncFMF,RecurrenceDescriptor & RedDes,DemandedBits * DB,AssumptionCache * AC,DominatorTree * DT)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
isMinMaxSelectCmpPattern(Instruction * I,const InstDesc & Prev)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
isConditionalRdxPattern(RecurKind Kind,Instruction * I)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
isRecurrenceInstr(Instruction * I,RecurKind Kind,InstDesc & Prev,FastMathFlags FMF)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 
hasMultipleUsesOf(Instruction * I,SmallPtrSetImpl<Instruction * > & Insts,unsigned MaxNumUses)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 }
isReductionPHI(PHINode * Phi,Loop * TheLoop,RecurrenceDescriptor & RedDes,DemandedBits * DB,AssumptionCache * AC,DominatorTree * DT)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 
isFirstOrderRecurrence(PHINode * Phi,Loop * TheLoop,DenseMap<Instruction *,Instruction * > & SinkAfter,DominatorTree * DT)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.
getRecurrenceIdentity(RecurKind K,Type * Tp,FastMathFlags FMF)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 
getOpcode(RecurKind Kind)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>
getReductionOpChain(PHINode * Phi,Loop * L) const869 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 
InductionDescriptor(Value * Start,InductionKind K,const SCEV * Step,BinaryOperator * BOp,SmallVectorImpl<Instruction * > * Casts)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 
getConstIntStepValue() const976 ConstantInt *InductionDescriptor::getConstIntStepValue() const {
977   if (isa<SCEVConstant>(Step))
978     return dyn_cast<ConstantInt>(cast<SCEVConstant>(Step)->getValue());
979   return nullptr;
980 }
981 
isFPInductionPHI(PHINode * Phi,const Loop * TheLoop,ScalarEvolution * SE,InductionDescriptor & D)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).
getCastsForInductionPHI(PredicatedScalarEvolution & PSE,const SCEVUnknown * PhiScev,const SCEVAddRecExpr * AR,SmallVectorImpl<Instruction * > & CastInsts)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 
isInductionPHI(PHINode * Phi,const Loop * TheLoop,PredicatedScalarEvolution & PSE,InductionDescriptor & D,bool Assume)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 
isInductionPHI(PHINode * Phi,const Loop * TheLoop,ScalarEvolution * SE,InductionDescriptor & D,const SCEV * Expr,SmallVectorImpl<Instruction * > * CastsToIgnore)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