• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===-- LoopUtils.cpp - Loop Utility functions -------------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file defines common loop utility functions.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/Transforms/Utils/LoopUtils.h"
15 #include "llvm/ADT/ScopeExit.h"
16 #include "llvm/Analysis/AliasAnalysis.h"
17 #include "llvm/Analysis/BasicAliasAnalysis.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/ScalarEvolutionExpander.h"
26 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
27 #include "llvm/Analysis/TargetTransformInfo.h"
28 #include "llvm/Analysis/ValueTracking.h"
29 #include "llvm/IR/Dominators.h"
30 #include "llvm/IR/Instructions.h"
31 #include "llvm/IR/Module.h"
32 #include "llvm/IR/PatternMatch.h"
33 #include "llvm/IR/ValueHandle.h"
34 #include "llvm/Pass.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/KnownBits.h"
37 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
38 
39 using namespace llvm;
40 using namespace llvm::PatternMatch;
41 
42 #define DEBUG_TYPE "loop-utils"
43 
areAllUsesIn(Instruction * I,SmallPtrSetImpl<Instruction * > & Set)44 bool RecurrenceDescriptor::areAllUsesIn(Instruction *I,
45                                         SmallPtrSetImpl<Instruction *> &Set) {
46   for (User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
47     if (!Set.count(dyn_cast<Instruction>(*Use)))
48       return false;
49   return true;
50 }
51 
isIntegerRecurrenceKind(RecurrenceKind Kind)52 bool RecurrenceDescriptor::isIntegerRecurrenceKind(RecurrenceKind Kind) {
53   switch (Kind) {
54   default:
55     break;
56   case RK_IntegerAdd:
57   case RK_IntegerMult:
58   case RK_IntegerOr:
59   case RK_IntegerAnd:
60   case RK_IntegerXor:
61   case RK_IntegerMinMax:
62     return true;
63   }
64   return false;
65 }
66 
isFloatingPointRecurrenceKind(RecurrenceKind Kind)67 bool RecurrenceDescriptor::isFloatingPointRecurrenceKind(RecurrenceKind Kind) {
68   return (Kind != RK_NoRecurrence) && !isIntegerRecurrenceKind(Kind);
69 }
70 
isArithmeticRecurrenceKind(RecurrenceKind Kind)71 bool RecurrenceDescriptor::isArithmeticRecurrenceKind(RecurrenceKind Kind) {
72   switch (Kind) {
73   default:
74     break;
75   case RK_IntegerAdd:
76   case RK_IntegerMult:
77   case RK_FloatAdd:
78   case RK_FloatMult:
79     return true;
80   }
81   return false;
82 }
83 
84 /// Determines if Phi may have been type-promoted. If Phi has a single user
85 /// that ANDs the Phi with a type mask, return the user. RT is updated to
86 /// account for the narrower bit width represented by the mask, and the AND
87 /// instruction is added to CI.
lookThroughAnd(PHINode * Phi,Type * & RT,SmallPtrSetImpl<Instruction * > & Visited,SmallPtrSetImpl<Instruction * > & CI)88 static Instruction *lookThroughAnd(PHINode *Phi, Type *&RT,
89                                    SmallPtrSetImpl<Instruction *> &Visited,
90                                    SmallPtrSetImpl<Instruction *> &CI) {
91   if (!Phi->hasOneUse())
92     return Phi;
93 
94   const APInt *M = nullptr;
95   Instruction *I, *J = cast<Instruction>(Phi->use_begin()->getUser());
96 
97   // Matches either I & 2^x-1 or 2^x-1 & I. If we find a match, we update RT
98   // with a new integer type of the corresponding bit width.
99   if (match(J, m_c_And(m_Instruction(I), m_APInt(M)))) {
100     int32_t Bits = (*M + 1).exactLogBase2();
101     if (Bits > 0) {
102       RT = IntegerType::get(Phi->getContext(), Bits);
103       Visited.insert(Phi);
104       CI.insert(J);
105       return J;
106     }
107   }
108   return Phi;
109 }
110 
111 /// Compute the minimal bit width needed to represent a reduction whose exit
112 /// instruction is given by Exit.
computeRecurrenceType(Instruction * Exit,DemandedBits * DB,AssumptionCache * AC,DominatorTree * DT)113 static std::pair<Type *, bool> computeRecurrenceType(Instruction *Exit,
114                                                      DemandedBits *DB,
115                                                      AssumptionCache *AC,
116                                                      DominatorTree *DT) {
117   bool IsSigned = false;
118   const DataLayout &DL = Exit->getModule()->getDataLayout();
119   uint64_t MaxBitWidth = DL.getTypeSizeInBits(Exit->getType());
120 
121   if (DB) {
122     // Use the demanded bits analysis to determine the bits that are live out
123     // of the exit instruction, rounding up to the nearest power of two. If the
124     // use of demanded bits results in a smaller bit width, we know the value
125     // must be positive (i.e., IsSigned = false), because if this were not the
126     // case, the sign bit would have been demanded.
127     auto Mask = DB->getDemandedBits(Exit);
128     MaxBitWidth = Mask.getBitWidth() - Mask.countLeadingZeros();
129   }
130 
131   if (MaxBitWidth == DL.getTypeSizeInBits(Exit->getType()) && AC && DT) {
132     // If demanded bits wasn't able to limit the bit width, we can try to use
133     // value tracking instead. This can be the case, for example, if the value
134     // may be negative.
135     auto NumSignBits = ComputeNumSignBits(Exit, DL, 0, AC, nullptr, DT);
136     auto NumTypeBits = DL.getTypeSizeInBits(Exit->getType());
137     MaxBitWidth = NumTypeBits - NumSignBits;
138     KnownBits Bits = computeKnownBits(Exit, DL);
139     if (!Bits.isNonNegative()) {
140       // If the value is not known to be non-negative, we set IsSigned to true,
141       // meaning that we will use sext instructions instead of zext
142       // instructions to restore the original type.
143       IsSigned = true;
144       if (!Bits.isNegative())
145         // If the value is not known to be negative, we don't known what the
146         // upper bit is, and therefore, we don't know what kind of extend we
147         // will need. In this case, just increase the bit width by one bit and
148         // use sext.
149         ++MaxBitWidth;
150     }
151   }
152   if (!isPowerOf2_64(MaxBitWidth))
153     MaxBitWidth = NextPowerOf2(MaxBitWidth);
154 
155   return std::make_pair(Type::getIntNTy(Exit->getContext(), MaxBitWidth),
156                         IsSigned);
157 }
158 
159 /// Collect cast instructions that can be ignored in the vectorizer's cost
160 /// model, given a reduction exit value and the minimal type in which the
161 /// reduction can be represented.
collectCastsToIgnore(Loop * TheLoop,Instruction * Exit,Type * RecurrenceType,SmallPtrSetImpl<Instruction * > & Casts)162 static void collectCastsToIgnore(Loop *TheLoop, Instruction *Exit,
163                                  Type *RecurrenceType,
164                                  SmallPtrSetImpl<Instruction *> &Casts) {
165 
166   SmallVector<Instruction *, 8> Worklist;
167   SmallPtrSet<Instruction *, 8> Visited;
168   Worklist.push_back(Exit);
169 
170   while (!Worklist.empty()) {
171     Instruction *Val = Worklist.pop_back_val();
172     Visited.insert(Val);
173     if (auto *Cast = dyn_cast<CastInst>(Val))
174       if (Cast->getSrcTy() == RecurrenceType) {
175         // If the source type of a cast instruction is equal to the recurrence
176         // type, it will be eliminated, and should be ignored in the vectorizer
177         // cost model.
178         Casts.insert(Cast);
179         continue;
180       }
181 
182     // Add all operands to the work list if they are loop-varying values that
183     // we haven't yet visited.
184     for (Value *O : cast<User>(Val)->operands())
185       if (auto *I = dyn_cast<Instruction>(O))
186         if (TheLoop->contains(I) && !Visited.count(I))
187           Worklist.push_back(I);
188   }
189 }
190 
AddReductionVar(PHINode * Phi,RecurrenceKind Kind,Loop * TheLoop,bool HasFunNoNaNAttr,RecurrenceDescriptor & RedDes,DemandedBits * DB,AssumptionCache * AC,DominatorTree * DT)191 bool RecurrenceDescriptor::AddReductionVar(PHINode *Phi, RecurrenceKind Kind,
192                                            Loop *TheLoop, bool HasFunNoNaNAttr,
193                                            RecurrenceDescriptor &RedDes,
194                                            DemandedBits *DB,
195                                            AssumptionCache *AC,
196                                            DominatorTree *DT) {
197   if (Phi->getNumIncomingValues() != 2)
198     return false;
199 
200   // Reduction variables are only found in the loop header block.
201   if (Phi->getParent() != TheLoop->getHeader())
202     return false;
203 
204   // Obtain the reduction start value from the value that comes from the loop
205   // preheader.
206   Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
207 
208   // ExitInstruction is the single value which is used outside the loop.
209   // We only allow for a single reduction value to be used outside the loop.
210   // This includes users of the reduction, variables (which form a cycle
211   // which ends in the phi node).
212   Instruction *ExitInstruction = nullptr;
213   // Indicates that we found a reduction operation in our scan.
214   bool FoundReduxOp = false;
215 
216   // We start with the PHI node and scan for all of the users of this
217   // instruction. All users must be instructions that can be used as reduction
218   // variables (such as ADD). We must have a single out-of-block user. The cycle
219   // must include the original PHI.
220   bool FoundStartPHI = false;
221 
222   // To recognize min/max patterns formed by a icmp select sequence, we store
223   // the number of instruction we saw from the recognized min/max pattern,
224   //  to make sure we only see exactly the two instructions.
225   unsigned NumCmpSelectPatternInst = 0;
226   InstDesc ReduxDesc(false, nullptr);
227 
228   // Data used for determining if the recurrence has been type-promoted.
229   Type *RecurrenceType = Phi->getType();
230   SmallPtrSet<Instruction *, 4> CastInsts;
231   Instruction *Start = Phi;
232   bool IsSigned = false;
233 
234   SmallPtrSet<Instruction *, 8> VisitedInsts;
235   SmallVector<Instruction *, 8> Worklist;
236 
237   // Return early if the recurrence kind does not match the type of Phi. If the
238   // recurrence kind is arithmetic, we attempt to look through AND operations
239   // resulting from the type promotion performed by InstCombine.  Vector
240   // operations are not limited to the legal integer widths, so we may be able
241   // to evaluate the reduction in the narrower width.
242   if (RecurrenceType->isFloatingPointTy()) {
243     if (!isFloatingPointRecurrenceKind(Kind))
244       return false;
245   } else {
246     if (!isIntegerRecurrenceKind(Kind))
247       return false;
248     if (isArithmeticRecurrenceKind(Kind))
249       Start = lookThroughAnd(Phi, RecurrenceType, VisitedInsts, CastInsts);
250   }
251 
252   Worklist.push_back(Start);
253   VisitedInsts.insert(Start);
254 
255   // A value in the reduction can be used:
256   //  - By the reduction:
257   //      - Reduction operation:
258   //        - One use of reduction value (safe).
259   //        - Multiple use of reduction value (not safe).
260   //      - PHI:
261   //        - All uses of the PHI must be the reduction (safe).
262   //        - Otherwise, not safe.
263   //  - By instructions outside of the loop (safe).
264   //      * One value may have several outside users, but all outside
265   //        uses must be of the same value.
266   //  - By an instruction that is not part of the reduction (not safe).
267   //    This is either:
268   //      * An instruction type other than PHI or the reduction operation.
269   //      * A PHI in the header other than the initial PHI.
270   while (!Worklist.empty()) {
271     Instruction *Cur = Worklist.back();
272     Worklist.pop_back();
273 
274     // No Users.
275     // If the instruction has no users then this is a broken chain and can't be
276     // a reduction variable.
277     if (Cur->use_empty())
278       return false;
279 
280     bool IsAPhi = isa<PHINode>(Cur);
281 
282     // A header PHI use other than the original PHI.
283     if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
284       return false;
285 
286     // Reductions of instructions such as Div, and Sub is only possible if the
287     // LHS is the reduction variable.
288     if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
289         !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
290         !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
291       return false;
292 
293     // Any reduction instruction must be of one of the allowed kinds. We ignore
294     // the starting value (the Phi or an AND instruction if the Phi has been
295     // type-promoted).
296     if (Cur != Start) {
297       ReduxDesc = isRecurrenceInstr(Cur, Kind, ReduxDesc, HasFunNoNaNAttr);
298       if (!ReduxDesc.isRecurrence())
299         return false;
300     }
301 
302     // A reduction operation must only have one use of the reduction value.
303     if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
304         hasMultipleUsesOf(Cur, VisitedInsts))
305       return false;
306 
307     // All inputs to a PHI node must be a reduction value.
308     if (IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
309       return false;
310 
311     if (Kind == RK_IntegerMinMax &&
312         (isa<ICmpInst>(Cur) || isa<SelectInst>(Cur)))
313       ++NumCmpSelectPatternInst;
314     if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) || isa<SelectInst>(Cur)))
315       ++NumCmpSelectPatternInst;
316 
317     // Check  whether we found a reduction operator.
318     FoundReduxOp |= !IsAPhi && Cur != Start;
319 
320     // Process users of current instruction. Push non-PHI nodes after PHI nodes
321     // onto the stack. This way we are going to have seen all inputs to PHI
322     // nodes once we get to them.
323     SmallVector<Instruction *, 8> NonPHIs;
324     SmallVector<Instruction *, 8> PHIs;
325     for (User *U : Cur->users()) {
326       Instruction *UI = cast<Instruction>(U);
327 
328       // Check if we found the exit user.
329       BasicBlock *Parent = UI->getParent();
330       if (!TheLoop->contains(Parent)) {
331         // If we already know this instruction is used externally, move on to
332         // the next user.
333         if (ExitInstruction == Cur)
334           continue;
335 
336         // Exit if you find multiple values used outside or if the header phi
337         // node is being used. In this case the user uses the value of the
338         // previous iteration, in which case we would loose "VF-1" iterations of
339         // the reduction operation if we vectorize.
340         if (ExitInstruction != nullptr || Cur == Phi)
341           return false;
342 
343         // The instruction used by an outside user must be the last instruction
344         // before we feed back to the reduction phi. Otherwise, we loose VF-1
345         // operations on the value.
346         if (!is_contained(Phi->operands(), Cur))
347           return false;
348 
349         ExitInstruction = Cur;
350         continue;
351       }
352 
353       // Process instructions only once (termination). Each reduction cycle
354       // value must only be used once, except by phi nodes and min/max
355       // reductions which are represented as a cmp followed by a select.
356       InstDesc IgnoredVal(false, nullptr);
357       if (VisitedInsts.insert(UI).second) {
358         if (isa<PHINode>(UI))
359           PHIs.push_back(UI);
360         else
361           NonPHIs.push_back(UI);
362       } else if (!isa<PHINode>(UI) &&
363                  ((!isa<FCmpInst>(UI) && !isa<ICmpInst>(UI) &&
364                    !isa<SelectInst>(UI)) ||
365                   !isMinMaxSelectCmpPattern(UI, IgnoredVal).isRecurrence()))
366         return false;
367 
368       // Remember that we completed the cycle.
369       if (UI == Phi)
370         FoundStartPHI = true;
371     }
372     Worklist.append(PHIs.begin(), PHIs.end());
373     Worklist.append(NonPHIs.begin(), NonPHIs.end());
374   }
375 
376   // This means we have seen one but not the other instruction of the
377   // pattern or more than just a select and cmp.
378   if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
379       NumCmpSelectPatternInst != 2)
380     return false;
381 
382   if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
383     return false;
384 
385   if (Start != Phi) {
386     // If the starting value is not the same as the phi node, we speculatively
387     // looked through an 'and' instruction when evaluating a potential
388     // arithmetic reduction to determine if it may have been type-promoted.
389     //
390     // We now compute the minimal bit width that is required to represent the
391     // reduction. If this is the same width that was indicated by the 'and', we
392     // can represent the reduction in the smaller type. The 'and' instruction
393     // will be eliminated since it will essentially be a cast instruction that
394     // can be ignore in the cost model. If we compute a different type than we
395     // did when evaluating the 'and', the 'and' will not be eliminated, and we
396     // will end up with different kinds of operations in the recurrence
397     // expression (e.g., RK_IntegerAND, RK_IntegerADD). We give up if this is
398     // the case.
399     //
400     // The vectorizer relies on InstCombine to perform the actual
401     // type-shrinking. It does this by inserting instructions to truncate the
402     // exit value of the reduction to the width indicated by RecurrenceType and
403     // then extend this value back to the original width. If IsSigned is false,
404     // a 'zext' instruction will be generated; otherwise, a 'sext' will be
405     // used.
406     //
407     // TODO: We should not rely on InstCombine to rewrite the reduction in the
408     //       smaller type. We should just generate a correctly typed expression
409     //       to begin with.
410     Type *ComputedType;
411     std::tie(ComputedType, IsSigned) =
412         computeRecurrenceType(ExitInstruction, DB, AC, DT);
413     if (ComputedType != RecurrenceType)
414       return false;
415 
416     // The recurrence expression will be represented in a narrower type. If
417     // there are any cast instructions that will be unnecessary, collect them
418     // in CastInsts. Note that the 'and' instruction was already included in
419     // this list.
420     //
421     // TODO: A better way to represent this may be to tag in some way all the
422     //       instructions that are a part of the reduction. The vectorizer cost
423     //       model could then apply the recurrence type to these instructions,
424     //       without needing a white list of instructions to ignore.
425     collectCastsToIgnore(TheLoop, ExitInstruction, RecurrenceType, CastInsts);
426   }
427 
428   // We found a reduction var if we have reached the original phi node and we
429   // only have a single instruction with out-of-loop users.
430 
431   // The ExitInstruction(Instruction which is allowed to have out-of-loop users)
432   // is saved as part of the RecurrenceDescriptor.
433 
434   // Save the description of this reduction variable.
435   RecurrenceDescriptor RD(
436       RdxStart, ExitInstruction, Kind, ReduxDesc.getMinMaxKind(),
437       ReduxDesc.getUnsafeAlgebraInst(), RecurrenceType, IsSigned, CastInsts);
438   RedDes = RD;
439 
440   return true;
441 }
442 
443 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
444 /// pattern corresponding to a min(X, Y) or max(X, Y).
445 RecurrenceDescriptor::InstDesc
isMinMaxSelectCmpPattern(Instruction * I,InstDesc & Prev)446 RecurrenceDescriptor::isMinMaxSelectCmpPattern(Instruction *I, InstDesc &Prev) {
447 
448   assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
449          "Expect a select instruction");
450   Instruction *Cmp = nullptr;
451   SelectInst *Select = nullptr;
452 
453   // We must handle the select(cmp()) as a single instruction. Advance to the
454   // select.
455   if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
456     if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
457       return InstDesc(false, I);
458     return InstDesc(Select, Prev.getMinMaxKind());
459   }
460 
461   // Only handle single use cases for now.
462   if (!(Select = dyn_cast<SelectInst>(I)))
463     return InstDesc(false, I);
464   if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
465       !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
466     return InstDesc(false, I);
467   if (!Cmp->hasOneUse())
468     return InstDesc(false, I);
469 
470   Value *CmpLeft;
471   Value *CmpRight;
472 
473   // Look for a min/max pattern.
474   if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
475     return InstDesc(Select, MRK_UIntMin);
476   else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
477     return InstDesc(Select, MRK_UIntMax);
478   else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
479     return InstDesc(Select, MRK_SIntMax);
480   else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
481     return InstDesc(Select, MRK_SIntMin);
482   else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
483     return InstDesc(Select, MRK_FloatMin);
484   else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
485     return InstDesc(Select, MRK_FloatMax);
486   else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
487     return InstDesc(Select, MRK_FloatMin);
488   else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
489     return InstDesc(Select, MRK_FloatMax);
490 
491   return InstDesc(false, I);
492 }
493 
494 RecurrenceDescriptor::InstDesc
isRecurrenceInstr(Instruction * I,RecurrenceKind Kind,InstDesc & Prev,bool HasFunNoNaNAttr)495 RecurrenceDescriptor::isRecurrenceInstr(Instruction *I, RecurrenceKind Kind,
496                                         InstDesc &Prev, bool HasFunNoNaNAttr) {
497   bool FP = I->getType()->isFloatingPointTy();
498   Instruction *UAI = Prev.getUnsafeAlgebraInst();
499   if (!UAI && FP && !I->isFast())
500     UAI = I; // Found an unsafe (unvectorizable) algebra instruction.
501 
502   switch (I->getOpcode()) {
503   default:
504     return InstDesc(false, I);
505   case Instruction::PHI:
506     return InstDesc(I, Prev.getMinMaxKind(), Prev.getUnsafeAlgebraInst());
507   case Instruction::Sub:
508   case Instruction::Add:
509     return InstDesc(Kind == RK_IntegerAdd, I);
510   case Instruction::Mul:
511     return InstDesc(Kind == RK_IntegerMult, I);
512   case Instruction::And:
513     return InstDesc(Kind == RK_IntegerAnd, I);
514   case Instruction::Or:
515     return InstDesc(Kind == RK_IntegerOr, I);
516   case Instruction::Xor:
517     return InstDesc(Kind == RK_IntegerXor, I);
518   case Instruction::FMul:
519     return InstDesc(Kind == RK_FloatMult, I, UAI);
520   case Instruction::FSub:
521   case Instruction::FAdd:
522     return InstDesc(Kind == RK_FloatAdd, I, UAI);
523   case Instruction::FCmp:
524   case Instruction::ICmp:
525   case Instruction::Select:
526     if (Kind != RK_IntegerMinMax &&
527         (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
528       return InstDesc(false, I);
529     return isMinMaxSelectCmpPattern(I, Prev);
530   }
531 }
532 
hasMultipleUsesOf(Instruction * I,SmallPtrSetImpl<Instruction * > & Insts)533 bool RecurrenceDescriptor::hasMultipleUsesOf(
534     Instruction *I, SmallPtrSetImpl<Instruction *> &Insts) {
535   unsigned NumUses = 0;
536   for (User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E;
537        ++Use) {
538     if (Insts.count(dyn_cast<Instruction>(*Use)))
539       ++NumUses;
540     if (NumUses > 1)
541       return true;
542   }
543 
544   return false;
545 }
isReductionPHI(PHINode * Phi,Loop * TheLoop,RecurrenceDescriptor & RedDes,DemandedBits * DB,AssumptionCache * AC,DominatorTree * DT)546 bool RecurrenceDescriptor::isReductionPHI(PHINode *Phi, Loop *TheLoop,
547                                           RecurrenceDescriptor &RedDes,
548                                           DemandedBits *DB, AssumptionCache *AC,
549                                           DominatorTree *DT) {
550 
551   BasicBlock *Header = TheLoop->getHeader();
552   Function &F = *Header->getParent();
553   bool HasFunNoNaNAttr =
554       F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
555 
556   if (AddReductionVar(Phi, RK_IntegerAdd, TheLoop, HasFunNoNaNAttr, RedDes, DB,
557                       AC, DT)) {
558     LLVM_DEBUG(dbgs() << "Found an ADD reduction PHI." << *Phi << "\n");
559     return true;
560   }
561   if (AddReductionVar(Phi, RK_IntegerMult, TheLoop, HasFunNoNaNAttr, RedDes, DB,
562                       AC, DT)) {
563     LLVM_DEBUG(dbgs() << "Found a MUL reduction PHI." << *Phi << "\n");
564     return true;
565   }
566   if (AddReductionVar(Phi, RK_IntegerOr, TheLoop, HasFunNoNaNAttr, RedDes, DB,
567                       AC, DT)) {
568     LLVM_DEBUG(dbgs() << "Found an OR reduction PHI." << *Phi << "\n");
569     return true;
570   }
571   if (AddReductionVar(Phi, RK_IntegerAnd, TheLoop, HasFunNoNaNAttr, RedDes, DB,
572                       AC, DT)) {
573     LLVM_DEBUG(dbgs() << "Found an AND reduction PHI." << *Phi << "\n");
574     return true;
575   }
576   if (AddReductionVar(Phi, RK_IntegerXor, TheLoop, HasFunNoNaNAttr, RedDes, DB,
577                       AC, DT)) {
578     LLVM_DEBUG(dbgs() << "Found a XOR reduction PHI." << *Phi << "\n");
579     return true;
580   }
581   if (AddReductionVar(Phi, RK_IntegerMinMax, TheLoop, HasFunNoNaNAttr, RedDes,
582                       DB, AC, DT)) {
583     LLVM_DEBUG(dbgs() << "Found a MINMAX reduction PHI." << *Phi << "\n");
584     return true;
585   }
586   if (AddReductionVar(Phi, RK_FloatMult, TheLoop, HasFunNoNaNAttr, RedDes, DB,
587                       AC, DT)) {
588     LLVM_DEBUG(dbgs() << "Found an FMult reduction PHI." << *Phi << "\n");
589     return true;
590   }
591   if (AddReductionVar(Phi, RK_FloatAdd, TheLoop, HasFunNoNaNAttr, RedDes, DB,
592                       AC, DT)) {
593     LLVM_DEBUG(dbgs() << "Found an FAdd reduction PHI." << *Phi << "\n");
594     return true;
595   }
596   if (AddReductionVar(Phi, RK_FloatMinMax, TheLoop, HasFunNoNaNAttr, RedDes, DB,
597                       AC, DT)) {
598     LLVM_DEBUG(dbgs() << "Found an float MINMAX reduction PHI." << *Phi
599                       << "\n");
600     return true;
601   }
602   // Not a reduction of known type.
603   return false;
604 }
605 
isFirstOrderRecurrence(PHINode * Phi,Loop * TheLoop,DenseMap<Instruction *,Instruction * > & SinkAfter,DominatorTree * DT)606 bool RecurrenceDescriptor::isFirstOrderRecurrence(
607     PHINode *Phi, Loop *TheLoop,
608     DenseMap<Instruction *, Instruction *> &SinkAfter, DominatorTree *DT) {
609 
610   // Ensure the phi node is in the loop header and has two incoming values.
611   if (Phi->getParent() != TheLoop->getHeader() ||
612       Phi->getNumIncomingValues() != 2)
613     return false;
614 
615   // Ensure the loop has a preheader and a single latch block. The loop
616   // vectorizer will need the latch to set up the next iteration of the loop.
617   auto *Preheader = TheLoop->getLoopPreheader();
618   auto *Latch = TheLoop->getLoopLatch();
619   if (!Preheader || !Latch)
620     return false;
621 
622   // Ensure the phi node's incoming blocks are the loop preheader and latch.
623   if (Phi->getBasicBlockIndex(Preheader) < 0 ||
624       Phi->getBasicBlockIndex(Latch) < 0)
625     return false;
626 
627   // Get the previous value. The previous value comes from the latch edge while
628   // the initial value comes form the preheader edge.
629   auto *Previous = dyn_cast<Instruction>(Phi->getIncomingValueForBlock(Latch));
630   if (!Previous || !TheLoop->contains(Previous) || isa<PHINode>(Previous) ||
631       SinkAfter.count(Previous)) // Cannot rely on dominance due to motion.
632     return false;
633 
634   // Ensure every user of the phi node is dominated by the previous value.
635   // The dominance requirement ensures the loop vectorizer will not need to
636   // vectorize the initial value prior to the first iteration of the loop.
637   // TODO: Consider extending this sinking to handle other kinds of instructions
638   // and expressions, beyond sinking a single cast past Previous.
639   if (Phi->hasOneUse()) {
640     auto *I = Phi->user_back();
641     if (I->isCast() && (I->getParent() == Phi->getParent()) && I->hasOneUse() &&
642         DT->dominates(Previous, I->user_back())) {
643       if (!DT->dominates(Previous, I)) // Otherwise we're good w/o sinking.
644         SinkAfter[I] = Previous;
645       return true;
646     }
647   }
648 
649   for (User *U : Phi->users())
650     if (auto *I = dyn_cast<Instruction>(U)) {
651       if (!DT->dominates(Previous, I))
652         return false;
653     }
654 
655   return true;
656 }
657 
658 /// This function returns the identity element (or neutral element) for
659 /// the operation K.
getRecurrenceIdentity(RecurrenceKind K,Type * Tp)660 Constant *RecurrenceDescriptor::getRecurrenceIdentity(RecurrenceKind K,
661                                                       Type *Tp) {
662   switch (K) {
663   case RK_IntegerXor:
664   case RK_IntegerAdd:
665   case RK_IntegerOr:
666     // Adding, Xoring, Oring zero to a number does not change it.
667     return ConstantInt::get(Tp, 0);
668   case RK_IntegerMult:
669     // Multiplying a number by 1 does not change it.
670     return ConstantInt::get(Tp, 1);
671   case RK_IntegerAnd:
672     // AND-ing a number with an all-1 value does not change it.
673     return ConstantInt::get(Tp, -1, true);
674   case RK_FloatMult:
675     // Multiplying a number by 1 does not change it.
676     return ConstantFP::get(Tp, 1.0L);
677   case RK_FloatAdd:
678     // Adding zero to a number does not change it.
679     return ConstantFP::get(Tp, 0.0L);
680   default:
681     llvm_unreachable("Unknown recurrence kind");
682   }
683 }
684 
685 /// This function translates the recurrence kind to an LLVM binary operator.
getRecurrenceBinOp(RecurrenceKind Kind)686 unsigned RecurrenceDescriptor::getRecurrenceBinOp(RecurrenceKind Kind) {
687   switch (Kind) {
688   case RK_IntegerAdd:
689     return Instruction::Add;
690   case RK_IntegerMult:
691     return Instruction::Mul;
692   case RK_IntegerOr:
693     return Instruction::Or;
694   case RK_IntegerAnd:
695     return Instruction::And;
696   case RK_IntegerXor:
697     return Instruction::Xor;
698   case RK_FloatMult:
699     return Instruction::FMul;
700   case RK_FloatAdd:
701     return Instruction::FAdd;
702   case RK_IntegerMinMax:
703     return Instruction::ICmp;
704   case RK_FloatMinMax:
705     return Instruction::FCmp;
706   default:
707     llvm_unreachable("Unknown recurrence operation");
708   }
709 }
710 
createMinMaxOp(IRBuilder<> & Builder,MinMaxRecurrenceKind RK,Value * Left,Value * Right)711 Value *RecurrenceDescriptor::createMinMaxOp(IRBuilder<> &Builder,
712                                             MinMaxRecurrenceKind RK,
713                                             Value *Left, Value *Right) {
714   CmpInst::Predicate P = CmpInst::ICMP_NE;
715   switch (RK) {
716   default:
717     llvm_unreachable("Unknown min/max recurrence kind");
718   case MRK_UIntMin:
719     P = CmpInst::ICMP_ULT;
720     break;
721   case MRK_UIntMax:
722     P = CmpInst::ICMP_UGT;
723     break;
724   case MRK_SIntMin:
725     P = CmpInst::ICMP_SLT;
726     break;
727   case MRK_SIntMax:
728     P = CmpInst::ICMP_SGT;
729     break;
730   case MRK_FloatMin:
731     P = CmpInst::FCMP_OLT;
732     break;
733   case MRK_FloatMax:
734     P = CmpInst::FCMP_OGT;
735     break;
736   }
737 
738   // We only match FP sequences that are 'fast', so we can unconditionally
739   // set it on any generated instructions.
740   IRBuilder<>::FastMathFlagGuard FMFG(Builder);
741   FastMathFlags FMF;
742   FMF.setFast();
743   Builder.setFastMathFlags(FMF);
744 
745   Value *Cmp;
746   if (RK == MRK_FloatMin || RK == MRK_FloatMax)
747     Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
748   else
749     Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
750 
751   Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
752   return Select;
753 }
754 
InductionDescriptor(Value * Start,InductionKind K,const SCEV * Step,BinaryOperator * BOp,SmallVectorImpl<Instruction * > * Casts)755 InductionDescriptor::InductionDescriptor(Value *Start, InductionKind K,
756                                          const SCEV *Step, BinaryOperator *BOp,
757                                          SmallVectorImpl<Instruction *> *Casts)
758   : StartValue(Start), IK(K), Step(Step), InductionBinOp(BOp) {
759   assert(IK != IK_NoInduction && "Not an induction");
760 
761   // Start value type should match the induction kind and the value
762   // itself should not be null.
763   assert(StartValue && "StartValue is null");
764   assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) &&
765          "StartValue is not a pointer for pointer induction");
766   assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) &&
767          "StartValue is not an integer for integer induction");
768 
769   // Check the Step Value. It should be non-zero integer value.
770   assert((!getConstIntStepValue() || !getConstIntStepValue()->isZero()) &&
771          "Step value is zero");
772 
773   assert((IK != IK_PtrInduction || getConstIntStepValue()) &&
774          "Step value should be constant for pointer induction");
775   assert((IK == IK_FpInduction || Step->getType()->isIntegerTy()) &&
776          "StepValue is not an integer");
777 
778   assert((IK != IK_FpInduction || Step->getType()->isFloatingPointTy()) &&
779          "StepValue is not FP for FpInduction");
780   assert((IK != IK_FpInduction || (InductionBinOp &&
781           (InductionBinOp->getOpcode() == Instruction::FAdd ||
782            InductionBinOp->getOpcode() == Instruction::FSub))) &&
783          "Binary opcode should be specified for FP induction");
784 
785   if (Casts) {
786     for (auto &Inst : *Casts) {
787       RedundantCasts.push_back(Inst);
788     }
789   }
790 }
791 
getConsecutiveDirection() const792 int InductionDescriptor::getConsecutiveDirection() const {
793   ConstantInt *ConstStep = getConstIntStepValue();
794   if (ConstStep && (ConstStep->isOne() || ConstStep->isMinusOne()))
795     return ConstStep->getSExtValue();
796   return 0;
797 }
798 
getConstIntStepValue() const799 ConstantInt *InductionDescriptor::getConstIntStepValue() const {
800   if (isa<SCEVConstant>(Step))
801     return dyn_cast<ConstantInt>(cast<SCEVConstant>(Step)->getValue());
802   return nullptr;
803 }
804 
transform(IRBuilder<> & B,Value * Index,ScalarEvolution * SE,const DataLayout & DL) const805 Value *InductionDescriptor::transform(IRBuilder<> &B, Value *Index,
806                                       ScalarEvolution *SE,
807                                       const DataLayout& DL) const {
808 
809   SCEVExpander Exp(*SE, DL, "induction");
810   assert(Index->getType() == Step->getType() &&
811          "Index type does not match StepValue type");
812   switch (IK) {
813   case IK_IntInduction: {
814     assert(Index->getType() == StartValue->getType() &&
815            "Index type does not match StartValue type");
816 
817     // FIXME: Theoretically, we can call getAddExpr() of ScalarEvolution
818     // and calculate (Start + Index * Step) for all cases, without
819     // special handling for "isOne" and "isMinusOne".
820     // But in the real life the result code getting worse. We mix SCEV
821     // expressions and ADD/SUB operations and receive redundant
822     // intermediate values being calculated in different ways and
823     // Instcombine is unable to reduce them all.
824 
825     if (getConstIntStepValue() &&
826         getConstIntStepValue()->isMinusOne())
827       return B.CreateSub(StartValue, Index);
828     if (getConstIntStepValue() &&
829         getConstIntStepValue()->isOne())
830       return B.CreateAdd(StartValue, Index);
831     const SCEV *S = SE->getAddExpr(SE->getSCEV(StartValue),
832                                    SE->getMulExpr(Step, SE->getSCEV(Index)));
833     return Exp.expandCodeFor(S, StartValue->getType(), &*B.GetInsertPoint());
834   }
835   case IK_PtrInduction: {
836     assert(isa<SCEVConstant>(Step) &&
837            "Expected constant step for pointer induction");
838     const SCEV *S = SE->getMulExpr(SE->getSCEV(Index), Step);
839     Index = Exp.expandCodeFor(S, Index->getType(), &*B.GetInsertPoint());
840     return B.CreateGEP(nullptr, StartValue, Index);
841   }
842   case IK_FpInduction: {
843     assert(Step->getType()->isFloatingPointTy() && "Expected FP Step value");
844     assert(InductionBinOp &&
845            (InductionBinOp->getOpcode() == Instruction::FAdd ||
846             InductionBinOp->getOpcode() == Instruction::FSub) &&
847            "Original bin op should be defined for FP induction");
848 
849     Value *StepValue = cast<SCEVUnknown>(Step)->getValue();
850 
851     // Floating point operations had to be 'fast' to enable the induction.
852     FastMathFlags Flags;
853     Flags.setFast();
854 
855     Value *MulExp = B.CreateFMul(StepValue, Index);
856     if (isa<Instruction>(MulExp))
857       // We have to check, the MulExp may be a constant.
858       cast<Instruction>(MulExp)->setFastMathFlags(Flags);
859 
860     Value *BOp = B.CreateBinOp(InductionBinOp->getOpcode() , StartValue,
861                                MulExp, "induction");
862     if (isa<Instruction>(BOp))
863       cast<Instruction>(BOp)->setFastMathFlags(Flags);
864 
865     return BOp;
866   }
867   case IK_NoInduction:
868     return nullptr;
869   }
870   llvm_unreachable("invalid enum");
871 }
872 
isFPInductionPHI(PHINode * Phi,const Loop * TheLoop,ScalarEvolution * SE,InductionDescriptor & D)873 bool InductionDescriptor::isFPInductionPHI(PHINode *Phi, const Loop *TheLoop,
874                                            ScalarEvolution *SE,
875                                            InductionDescriptor &D) {
876 
877   // Here we only handle FP induction variables.
878   assert(Phi->getType()->isFloatingPointTy() && "Unexpected Phi type");
879 
880   if (TheLoop->getHeader() != Phi->getParent())
881     return false;
882 
883   // The loop may have multiple entrances or multiple exits; we can analyze
884   // this phi if it has a unique entry value and a unique backedge value.
885   if (Phi->getNumIncomingValues() != 2)
886     return false;
887   Value *BEValue = nullptr, *StartValue = nullptr;
888   if (TheLoop->contains(Phi->getIncomingBlock(0))) {
889     BEValue = Phi->getIncomingValue(0);
890     StartValue = Phi->getIncomingValue(1);
891   } else {
892     assert(TheLoop->contains(Phi->getIncomingBlock(1)) &&
893            "Unexpected Phi node in the loop");
894     BEValue = Phi->getIncomingValue(1);
895     StartValue = Phi->getIncomingValue(0);
896   }
897 
898   BinaryOperator *BOp = dyn_cast<BinaryOperator>(BEValue);
899   if (!BOp)
900     return false;
901 
902   Value *Addend = nullptr;
903   if (BOp->getOpcode() == Instruction::FAdd) {
904     if (BOp->getOperand(0) == Phi)
905       Addend = BOp->getOperand(1);
906     else if (BOp->getOperand(1) == Phi)
907       Addend = BOp->getOperand(0);
908   } else if (BOp->getOpcode() == Instruction::FSub)
909     if (BOp->getOperand(0) == Phi)
910       Addend = BOp->getOperand(1);
911 
912   if (!Addend)
913     return false;
914 
915   // The addend should be loop invariant
916   if (auto *I = dyn_cast<Instruction>(Addend))
917     if (TheLoop->contains(I))
918       return false;
919 
920   // FP Step has unknown SCEV
921   const SCEV *Step = SE->getUnknown(Addend);
922   D = InductionDescriptor(StartValue, IK_FpInduction, Step, BOp);
923   return true;
924 }
925 
926 /// This function is called when we suspect that the update-chain of a phi node
927 /// (whose symbolic SCEV expression sin \p PhiScev) contains redundant casts,
928 /// that can be ignored. (This can happen when the PSCEV rewriter adds a runtime
929 /// predicate P under which the SCEV expression for the phi can be the
930 /// AddRecurrence \p AR; See createAddRecFromPHIWithCast). We want to find the
931 /// cast instructions that are involved in the update-chain of this induction.
932 /// A caller that adds the required runtime predicate can be free to drop these
933 /// cast instructions, and compute the phi using \p AR (instead of some scev
934 /// expression with casts).
935 ///
936 /// For example, without a predicate the scev expression can take the following
937 /// form:
938 ///      (Ext ix (Trunc iy ( Start + i*Step ) to ix) to iy)
939 ///
940 /// It corresponds to the following IR sequence:
941 /// %for.body:
942 ///   %x = phi i64 [ 0, %ph ], [ %add, %for.body ]
943 ///   %casted_phi = "ExtTrunc i64 %x"
944 ///   %add = add i64 %casted_phi, %step
945 ///
946 /// where %x is given in \p PN,
947 /// PSE.getSCEV(%x) is equal to PSE.getSCEV(%casted_phi) under a predicate,
948 /// and the IR sequence that "ExtTrunc i64 %x" represents can take one of
949 /// several forms, for example, such as:
950 ///   ExtTrunc1:    %casted_phi = and  %x, 2^n-1
951 /// or:
952 ///   ExtTrunc2:    %t = shl %x, m
953 ///                 %casted_phi = ashr %t, m
954 ///
955 /// If we are able to find such sequence, we return the instructions
956 /// we found, namely %casted_phi and the instructions on its use-def chain up
957 /// to the phi (not including the phi).
getCastsForInductionPHI(PredicatedScalarEvolution & PSE,const SCEVUnknown * PhiScev,const SCEVAddRecExpr * AR,SmallVectorImpl<Instruction * > & CastInsts)958 static bool getCastsForInductionPHI(PredicatedScalarEvolution &PSE,
959                                     const SCEVUnknown *PhiScev,
960                                     const SCEVAddRecExpr *AR,
961                                     SmallVectorImpl<Instruction *> &CastInsts) {
962 
963   assert(CastInsts.empty() && "CastInsts is expected to be empty.");
964   auto *PN = cast<PHINode>(PhiScev->getValue());
965   assert(PSE.getSCEV(PN) == AR && "Unexpected phi node SCEV expression");
966   const Loop *L = AR->getLoop();
967 
968   // Find any cast instructions that participate in the def-use chain of
969   // PhiScev in the loop.
970   // FORNOW/TODO: We currently expect the def-use chain to include only
971   // two-operand instructions, where one of the operands is an invariant.
972   // createAddRecFromPHIWithCasts() currently does not support anything more
973   // involved than that, so we keep the search simple. This can be
974   // extended/generalized as needed.
975 
976   auto getDef = [&](const Value *Val) -> Value * {
977     const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Val);
978     if (!BinOp)
979       return nullptr;
980     Value *Op0 = BinOp->getOperand(0);
981     Value *Op1 = BinOp->getOperand(1);
982     Value *Def = nullptr;
983     if (L->isLoopInvariant(Op0))
984       Def = Op1;
985     else if (L->isLoopInvariant(Op1))
986       Def = Op0;
987     return Def;
988   };
989 
990   // Look for the instruction that defines the induction via the
991   // loop backedge.
992   BasicBlock *Latch = L->getLoopLatch();
993   if (!Latch)
994     return false;
995   Value *Val = PN->getIncomingValueForBlock(Latch);
996   if (!Val)
997     return false;
998 
999   // Follow the def-use chain until the induction phi is reached.
1000   // If on the way we encounter a Value that has the same SCEV Expr as the
1001   // phi node, we can consider the instructions we visit from that point
1002   // as part of the cast-sequence that can be ignored.
1003   bool InCastSequence = false;
1004   auto *Inst = dyn_cast<Instruction>(Val);
1005   while (Val != PN) {
1006     // If we encountered a phi node other than PN, or if we left the loop,
1007     // we bail out.
1008     if (!Inst || !L->contains(Inst)) {
1009       return false;
1010     }
1011     auto *AddRec = dyn_cast<SCEVAddRecExpr>(PSE.getSCEV(Val));
1012     if (AddRec && PSE.areAddRecsEqualWithPreds(AddRec, AR))
1013       InCastSequence = true;
1014     if (InCastSequence) {
1015       // Only the last instruction in the cast sequence is expected to have
1016       // uses outside the induction def-use chain.
1017       if (!CastInsts.empty())
1018         if (!Inst->hasOneUse())
1019           return false;
1020       CastInsts.push_back(Inst);
1021     }
1022     Val = getDef(Val);
1023     if (!Val)
1024       return false;
1025     Inst = dyn_cast<Instruction>(Val);
1026   }
1027 
1028   return InCastSequence;
1029 }
1030 
isInductionPHI(PHINode * Phi,const Loop * TheLoop,PredicatedScalarEvolution & PSE,InductionDescriptor & D,bool Assume)1031 bool InductionDescriptor::isInductionPHI(PHINode *Phi, const Loop *TheLoop,
1032                                          PredicatedScalarEvolution &PSE,
1033                                          InductionDescriptor &D,
1034                                          bool Assume) {
1035   Type *PhiTy = Phi->getType();
1036 
1037   // Handle integer and pointer inductions variables.
1038   // Now we handle also FP induction but not trying to make a
1039   // recurrent expression from the PHI node in-place.
1040 
1041   if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy() &&
1042       !PhiTy->isFloatTy() && !PhiTy->isDoubleTy() && !PhiTy->isHalfTy())
1043     return false;
1044 
1045   if (PhiTy->isFloatingPointTy())
1046     return isFPInductionPHI(Phi, TheLoop, PSE.getSE(), D);
1047 
1048   const SCEV *PhiScev = PSE.getSCEV(Phi);
1049   const auto *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1050 
1051   // We need this expression to be an AddRecExpr.
1052   if (Assume && !AR)
1053     AR = PSE.getAsAddRec(Phi);
1054 
1055   if (!AR) {
1056     LLVM_DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
1057     return false;
1058   }
1059 
1060   // Record any Cast instructions that participate in the induction update
1061   const auto *SymbolicPhi = dyn_cast<SCEVUnknown>(PhiScev);
1062   // If we started from an UnknownSCEV, and managed to build an addRecurrence
1063   // only after enabling Assume with PSCEV, this means we may have encountered
1064   // cast instructions that required adding a runtime check in order to
1065   // guarantee the correctness of the AddRecurence respresentation of the
1066   // induction.
1067   if (PhiScev != AR && SymbolicPhi) {
1068     SmallVector<Instruction *, 2> Casts;
1069     if (getCastsForInductionPHI(PSE, SymbolicPhi, AR, Casts))
1070       return isInductionPHI(Phi, TheLoop, PSE.getSE(), D, AR, &Casts);
1071   }
1072 
1073   return isInductionPHI(Phi, TheLoop, PSE.getSE(), D, AR);
1074 }
1075 
isInductionPHI(PHINode * Phi,const Loop * TheLoop,ScalarEvolution * SE,InductionDescriptor & D,const SCEV * Expr,SmallVectorImpl<Instruction * > * CastsToIgnore)1076 bool InductionDescriptor::isInductionPHI(
1077     PHINode *Phi, const Loop *TheLoop, ScalarEvolution *SE,
1078     InductionDescriptor &D, const SCEV *Expr,
1079     SmallVectorImpl<Instruction *> *CastsToIgnore) {
1080   Type *PhiTy = Phi->getType();
1081   // We only handle integer and pointer inductions variables.
1082   if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
1083     return false;
1084 
1085   // Check that the PHI is consecutive.
1086   const SCEV *PhiScev = Expr ? Expr : SE->getSCEV(Phi);
1087   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
1088 
1089   if (!AR) {
1090     LLVM_DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
1091     return false;
1092   }
1093 
1094   if (AR->getLoop() != TheLoop) {
1095     // FIXME: We should treat this as a uniform. Unfortunately, we
1096     // don't currently know how to handled uniform PHIs.
1097     LLVM_DEBUG(
1098         dbgs() << "LV: PHI is a recurrence with respect to an outer loop.\n");
1099     return false;
1100   }
1101 
1102   Value *StartValue =
1103     Phi->getIncomingValueForBlock(AR->getLoop()->getLoopPreheader());
1104   const SCEV *Step = AR->getStepRecurrence(*SE);
1105   // Calculate the pointer stride and check if it is consecutive.
1106   // The stride may be a constant or a loop invariant integer value.
1107   const SCEVConstant *ConstStep = dyn_cast<SCEVConstant>(Step);
1108   if (!ConstStep && !SE->isLoopInvariant(Step, TheLoop))
1109     return false;
1110 
1111   if (PhiTy->isIntegerTy()) {
1112     D = InductionDescriptor(StartValue, IK_IntInduction, Step, /*BOp=*/ nullptr,
1113                             CastsToIgnore);
1114     return true;
1115   }
1116 
1117   assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
1118   // Pointer induction should be a constant.
1119   if (!ConstStep)
1120     return false;
1121 
1122   ConstantInt *CV = ConstStep->getValue();
1123   Type *PointerElementType = PhiTy->getPointerElementType();
1124   // The pointer stride cannot be determined if the pointer element type is not
1125   // sized.
1126   if (!PointerElementType->isSized())
1127     return false;
1128 
1129   const DataLayout &DL = Phi->getModule()->getDataLayout();
1130   int64_t Size = static_cast<int64_t>(DL.getTypeAllocSize(PointerElementType));
1131   if (!Size)
1132     return false;
1133 
1134   int64_t CVSize = CV->getSExtValue();
1135   if (CVSize % Size)
1136     return false;
1137   auto *StepValue = SE->getConstant(CV->getType(), CVSize / Size,
1138                                     true /* signed */);
1139   D = InductionDescriptor(StartValue, IK_PtrInduction, StepValue);
1140   return true;
1141 }
1142 
formDedicatedExitBlocks(Loop * L,DominatorTree * DT,LoopInfo * LI,bool PreserveLCSSA)1143 bool llvm::formDedicatedExitBlocks(Loop *L, DominatorTree *DT, LoopInfo *LI,
1144                                    bool PreserveLCSSA) {
1145   bool Changed = false;
1146 
1147   // We re-use a vector for the in-loop predecesosrs.
1148   SmallVector<BasicBlock *, 4> InLoopPredecessors;
1149 
1150   auto RewriteExit = [&](BasicBlock *BB) {
1151     assert(InLoopPredecessors.empty() &&
1152            "Must start with an empty predecessors list!");
1153     auto Cleanup = make_scope_exit([&] { InLoopPredecessors.clear(); });
1154 
1155     // See if there are any non-loop predecessors of this exit block and
1156     // keep track of the in-loop predecessors.
1157     bool IsDedicatedExit = true;
1158     for (auto *PredBB : predecessors(BB))
1159       if (L->contains(PredBB)) {
1160         if (isa<IndirectBrInst>(PredBB->getTerminator()))
1161           // We cannot rewrite exiting edges from an indirectbr.
1162           return false;
1163 
1164         InLoopPredecessors.push_back(PredBB);
1165       } else {
1166         IsDedicatedExit = false;
1167       }
1168 
1169     assert(!InLoopPredecessors.empty() && "Must have *some* loop predecessor!");
1170 
1171     // Nothing to do if this is already a dedicated exit.
1172     if (IsDedicatedExit)
1173       return false;
1174 
1175     auto *NewExitBB = SplitBlockPredecessors(
1176         BB, InLoopPredecessors, ".loopexit", DT, LI, PreserveLCSSA);
1177 
1178     if (!NewExitBB)
1179       LLVM_DEBUG(
1180           dbgs() << "WARNING: Can't create a dedicated exit block for loop: "
1181                  << *L << "\n");
1182     else
1183       LLVM_DEBUG(dbgs() << "LoopSimplify: Creating dedicated exit block "
1184                         << NewExitBB->getName() << "\n");
1185     return true;
1186   };
1187 
1188   // Walk the exit blocks directly rather than building up a data structure for
1189   // them, but only visit each one once.
1190   SmallPtrSet<BasicBlock *, 4> Visited;
1191   for (auto *BB : L->blocks())
1192     for (auto *SuccBB : successors(BB)) {
1193       // We're looking for exit blocks so skip in-loop successors.
1194       if (L->contains(SuccBB))
1195         continue;
1196 
1197       // Visit each exit block exactly once.
1198       if (!Visited.insert(SuccBB).second)
1199         continue;
1200 
1201       Changed |= RewriteExit(SuccBB);
1202     }
1203 
1204   return Changed;
1205 }
1206 
1207 /// Returns the instructions that use values defined in the loop.
findDefsUsedOutsideOfLoop(Loop * L)1208 SmallVector<Instruction *, 8> llvm::findDefsUsedOutsideOfLoop(Loop *L) {
1209   SmallVector<Instruction *, 8> UsedOutside;
1210 
1211   for (auto *Block : L->getBlocks())
1212     // FIXME: I believe that this could use copy_if if the Inst reference could
1213     // be adapted into a pointer.
1214     for (auto &Inst : *Block) {
1215       auto Users = Inst.users();
1216       if (any_of(Users, [&](User *U) {
1217             auto *Use = cast<Instruction>(U);
1218             return !L->contains(Use->getParent());
1219           }))
1220         UsedOutside.push_back(&Inst);
1221     }
1222 
1223   return UsedOutside;
1224 }
1225 
getLoopAnalysisUsage(AnalysisUsage & AU)1226 void llvm::getLoopAnalysisUsage(AnalysisUsage &AU) {
1227   // By definition, all loop passes need the LoopInfo analysis and the
1228   // Dominator tree it depends on. Because they all participate in the loop
1229   // pass manager, they must also preserve these.
1230   AU.addRequired<DominatorTreeWrapperPass>();
1231   AU.addPreserved<DominatorTreeWrapperPass>();
1232   AU.addRequired<LoopInfoWrapperPass>();
1233   AU.addPreserved<LoopInfoWrapperPass>();
1234 
1235   // We must also preserve LoopSimplify and LCSSA. We locally access their IDs
1236   // here because users shouldn't directly get them from this header.
1237   extern char &LoopSimplifyID;
1238   extern char &LCSSAID;
1239   AU.addRequiredID(LoopSimplifyID);
1240   AU.addPreservedID(LoopSimplifyID);
1241   AU.addRequiredID(LCSSAID);
1242   AU.addPreservedID(LCSSAID);
1243   // This is used in the LPPassManager to perform LCSSA verification on passes
1244   // which preserve lcssa form
1245   AU.addRequired<LCSSAVerificationPass>();
1246   AU.addPreserved<LCSSAVerificationPass>();
1247 
1248   // Loop passes are designed to run inside of a loop pass manager which means
1249   // that any function analyses they require must be required by the first loop
1250   // pass in the manager (so that it is computed before the loop pass manager
1251   // runs) and preserved by all loop pasess in the manager. To make this
1252   // reasonably robust, the set needed for most loop passes is maintained here.
1253   // If your loop pass requires an analysis not listed here, you will need to
1254   // carefully audit the loop pass manager nesting structure that results.
1255   AU.addRequired<AAResultsWrapperPass>();
1256   AU.addPreserved<AAResultsWrapperPass>();
1257   AU.addPreserved<BasicAAWrapperPass>();
1258   AU.addPreserved<GlobalsAAWrapperPass>();
1259   AU.addPreserved<SCEVAAWrapperPass>();
1260   AU.addRequired<ScalarEvolutionWrapperPass>();
1261   AU.addPreserved<ScalarEvolutionWrapperPass>();
1262 }
1263 
1264 /// Manually defined generic "LoopPass" dependency initialization. This is used
1265 /// to initialize the exact set of passes from above in \c
1266 /// getLoopAnalysisUsage. It can be used within a loop pass's initialization
1267 /// with:
1268 ///
1269 ///   INITIALIZE_PASS_DEPENDENCY(LoopPass)
1270 ///
1271 /// As-if "LoopPass" were a pass.
initializeLoopPassPass(PassRegistry & Registry)1272 void llvm::initializeLoopPassPass(PassRegistry &Registry) {
1273   INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1274   INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
1275   INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
1276   INITIALIZE_PASS_DEPENDENCY(LCSSAWrapperPass)
1277   INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
1278   INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)
1279   INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
1280   INITIALIZE_PASS_DEPENDENCY(SCEVAAWrapperPass)
1281   INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
1282 }
1283 
1284 /// Find string metadata for loop
1285 ///
1286 /// If it has a value (e.g. {"llvm.distribute", 1} return the value as an
1287 /// operand or null otherwise.  If the string metadata is not found return
1288 /// Optional's not-a-value.
findStringMetadataForLoop(Loop * TheLoop,StringRef Name)1289 Optional<const MDOperand *> llvm::findStringMetadataForLoop(Loop *TheLoop,
1290                                                             StringRef Name) {
1291   MDNode *LoopID = TheLoop->getLoopID();
1292   // Return none if LoopID is false.
1293   if (!LoopID)
1294     return None;
1295 
1296   // First operand should refer to the loop id itself.
1297   assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1298   assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1299 
1300   // Iterate over LoopID operands and look for MDString Metadata
1301   for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) {
1302     MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
1303     if (!MD)
1304       continue;
1305     MDString *S = dyn_cast<MDString>(MD->getOperand(0));
1306     if (!S)
1307       continue;
1308     // Return true if MDString holds expected MetaData.
1309     if (Name.equals(S->getString()))
1310       switch (MD->getNumOperands()) {
1311       case 1:
1312         return nullptr;
1313       case 2:
1314         return &MD->getOperand(1);
1315       default:
1316         llvm_unreachable("loop metadata has 0 or 1 operand");
1317       }
1318   }
1319   return None;
1320 }
1321 
1322 /// Does a BFS from a given node to all of its children inside a given loop.
1323 /// The returned vector of nodes includes the starting point.
1324 SmallVector<DomTreeNode *, 16>
collectChildrenInLoop(DomTreeNode * N,const Loop * CurLoop)1325 llvm::collectChildrenInLoop(DomTreeNode *N, const Loop *CurLoop) {
1326   SmallVector<DomTreeNode *, 16> Worklist;
1327   auto AddRegionToWorklist = [&](DomTreeNode *DTN) {
1328     // Only include subregions in the top level loop.
1329     BasicBlock *BB = DTN->getBlock();
1330     if (CurLoop->contains(BB))
1331       Worklist.push_back(DTN);
1332   };
1333 
1334   AddRegionToWorklist(N);
1335 
1336   for (size_t I = 0; I < Worklist.size(); I++)
1337     for (DomTreeNode *Child : Worklist[I]->getChildren())
1338       AddRegionToWorklist(Child);
1339 
1340   return Worklist;
1341 }
1342 
deleteDeadLoop(Loop * L,DominatorTree * DT=nullptr,ScalarEvolution * SE=nullptr,LoopInfo * LI=nullptr)1343 void llvm::deleteDeadLoop(Loop *L, DominatorTree *DT = nullptr,
1344                           ScalarEvolution *SE = nullptr,
1345                           LoopInfo *LI = nullptr) {
1346   assert((!DT || L->isLCSSAForm(*DT)) && "Expected LCSSA!");
1347   auto *Preheader = L->getLoopPreheader();
1348   assert(Preheader && "Preheader should exist!");
1349 
1350   // Now that we know the removal is safe, remove the loop by changing the
1351   // branch from the preheader to go to the single exit block.
1352   //
1353   // Because we're deleting a large chunk of code at once, the sequence in which
1354   // we remove things is very important to avoid invalidation issues.
1355 
1356   // Tell ScalarEvolution that the loop is deleted. Do this before
1357   // deleting the loop so that ScalarEvolution can look at the loop
1358   // to determine what it needs to clean up.
1359   if (SE)
1360     SE->forgetLoop(L);
1361 
1362   auto *ExitBlock = L->getUniqueExitBlock();
1363   assert(ExitBlock && "Should have a unique exit block!");
1364   assert(L->hasDedicatedExits() && "Loop should have dedicated exits!");
1365 
1366   auto *OldBr = dyn_cast<BranchInst>(Preheader->getTerminator());
1367   assert(OldBr && "Preheader must end with a branch");
1368   assert(OldBr->isUnconditional() && "Preheader must have a single successor");
1369   // Connect the preheader to the exit block. Keep the old edge to the header
1370   // around to perform the dominator tree update in two separate steps
1371   // -- #1 insertion of the edge preheader -> exit and #2 deletion of the edge
1372   // preheader -> header.
1373   //
1374   //
1375   // 0.  Preheader          1.  Preheader           2.  Preheader
1376   //        |                    |   |                   |
1377   //        V                    |   V                   |
1378   //      Header <--\            | Header <--\           | Header <--\
1379   //       |  |     |            |  |  |     |           |  |  |     |
1380   //       |  V     |            |  |  V     |           |  |  V     |
1381   //       | Body --/            |  | Body --/           |  | Body --/
1382   //       V                     V  V                    V  V
1383   //      Exit                   Exit                    Exit
1384   //
1385   // By doing this is two separate steps we can perform the dominator tree
1386   // update without using the batch update API.
1387   //
1388   // Even when the loop is never executed, we cannot remove the edge from the
1389   // source block to the exit block. Consider the case where the unexecuted loop
1390   // branches back to an outer loop. If we deleted the loop and removed the edge
1391   // coming to this inner loop, this will break the outer loop structure (by
1392   // deleting the backedge of the outer loop). If the outer loop is indeed a
1393   // non-loop, it will be deleted in a future iteration of loop deletion pass.
1394   IRBuilder<> Builder(OldBr);
1395   Builder.CreateCondBr(Builder.getFalse(), L->getHeader(), ExitBlock);
1396   // Remove the old branch. The conditional branch becomes a new terminator.
1397   OldBr->eraseFromParent();
1398 
1399   // Rewrite phis in the exit block to get their inputs from the Preheader
1400   // instead of the exiting block.
1401   for (PHINode &P : ExitBlock->phis()) {
1402     // Set the zero'th element of Phi to be from the preheader and remove all
1403     // other incoming values. Given the loop has dedicated exits, all other
1404     // incoming values must be from the exiting blocks.
1405     int PredIndex = 0;
1406     P.setIncomingBlock(PredIndex, Preheader);
1407     // Removes all incoming values from all other exiting blocks (including
1408     // duplicate values from an exiting block).
1409     // Nuke all entries except the zero'th entry which is the preheader entry.
1410     // NOTE! We need to remove Incoming Values in the reverse order as done
1411     // below, to keep the indices valid for deletion (removeIncomingValues
1412     // updates getNumIncomingValues and shifts all values down into the operand
1413     // being deleted).
1414     for (unsigned i = 0, e = P.getNumIncomingValues() - 1; i != e; ++i)
1415       P.removeIncomingValue(e - i, false);
1416 
1417     assert((P.getNumIncomingValues() == 1 &&
1418             P.getIncomingBlock(PredIndex) == Preheader) &&
1419            "Should have exactly one value and that's from the preheader!");
1420   }
1421 
1422   // Disconnect the loop body by branching directly to its exit.
1423   Builder.SetInsertPoint(Preheader->getTerminator());
1424   Builder.CreateBr(ExitBlock);
1425   // Remove the old branch.
1426   Preheader->getTerminator()->eraseFromParent();
1427 
1428   if (DT) {
1429     // Update the dominator tree by informing it about the new edge from the
1430     // preheader to the exit.
1431     DT->insertEdge(Preheader, ExitBlock);
1432     // Inform the dominator tree about the removed edge.
1433     DT->deleteEdge(Preheader, L->getHeader());
1434   }
1435 
1436   // Given LCSSA form is satisfied, we should not have users of instructions
1437   // within the dead loop outside of the loop. However, LCSSA doesn't take
1438   // unreachable uses into account. We handle them here.
1439   // We could do it after drop all references (in this case all users in the
1440   // loop will be already eliminated and we have less work to do but according
1441   // to API doc of User::dropAllReferences only valid operation after dropping
1442   // references, is deletion. So let's substitute all usages of
1443   // instruction from the loop with undef value of corresponding type first.
1444   for (auto *Block : L->blocks())
1445     for (Instruction &I : *Block) {
1446       auto *Undef = UndefValue::get(I.getType());
1447       for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E;) {
1448         Use &U = *UI;
1449         ++UI;
1450         if (auto *Usr = dyn_cast<Instruction>(U.getUser()))
1451           if (L->contains(Usr->getParent()))
1452             continue;
1453         // If we have a DT then we can check that uses outside a loop only in
1454         // unreachable block.
1455         if (DT)
1456           assert(!DT->isReachableFromEntry(U) &&
1457                  "Unexpected user in reachable block");
1458         U.set(Undef);
1459       }
1460     }
1461 
1462   // Remove the block from the reference counting scheme, so that we can
1463   // delete it freely later.
1464   for (auto *Block : L->blocks())
1465     Block->dropAllReferences();
1466 
1467   if (LI) {
1468     // Erase the instructions and the blocks without having to worry
1469     // about ordering because we already dropped the references.
1470     // NOTE: This iteration is safe because erasing the block does not remove
1471     // its entry from the loop's block list.  We do that in the next section.
1472     for (Loop::block_iterator LpI = L->block_begin(), LpE = L->block_end();
1473          LpI != LpE; ++LpI)
1474       (*LpI)->eraseFromParent();
1475 
1476     // Finally, the blocks from loopinfo.  This has to happen late because
1477     // otherwise our loop iterators won't work.
1478 
1479     SmallPtrSet<BasicBlock *, 8> blocks;
1480     blocks.insert(L->block_begin(), L->block_end());
1481     for (BasicBlock *BB : blocks)
1482       LI->removeBlock(BB);
1483 
1484     // The last step is to update LoopInfo now that we've eliminated this loop.
1485     LI->erase(L);
1486   }
1487 }
1488 
getLoopEstimatedTripCount(Loop * L)1489 Optional<unsigned> llvm::getLoopEstimatedTripCount(Loop *L) {
1490   // Only support loops with a unique exiting block, and a latch.
1491   if (!L->getExitingBlock())
1492     return None;
1493 
1494   // Get the branch weights for the loop's backedge.
1495   BranchInst *LatchBR =
1496       dyn_cast<BranchInst>(L->getLoopLatch()->getTerminator());
1497   if (!LatchBR || LatchBR->getNumSuccessors() != 2)
1498     return None;
1499 
1500   assert((LatchBR->getSuccessor(0) == L->getHeader() ||
1501           LatchBR->getSuccessor(1) == L->getHeader()) &&
1502          "At least one edge out of the latch must go to the header");
1503 
1504   // To estimate the number of times the loop body was executed, we want to
1505   // know the number of times the backedge was taken, vs. the number of times
1506   // we exited the loop.
1507   uint64_t TrueVal, FalseVal;
1508   if (!LatchBR->extractProfMetadata(TrueVal, FalseVal))
1509     return None;
1510 
1511   if (!TrueVal || !FalseVal)
1512     return 0;
1513 
1514   // Divide the count of the backedge by the count of the edge exiting the loop,
1515   // rounding to nearest.
1516   if (LatchBR->getSuccessor(0) == L->getHeader())
1517     return (TrueVal + (FalseVal / 2)) / FalseVal;
1518   else
1519     return (FalseVal + (TrueVal / 2)) / TrueVal;
1520 }
1521 
1522 /// Adds a 'fast' flag to floating point operations.
addFastMathFlag(Value * V)1523 static Value *addFastMathFlag(Value *V) {
1524   if (isa<FPMathOperator>(V)) {
1525     FastMathFlags Flags;
1526     Flags.setFast();
1527     cast<Instruction>(V)->setFastMathFlags(Flags);
1528   }
1529   return V;
1530 }
1531 
1532 // Helper to generate an ordered reduction.
1533 Value *
getOrderedReduction(IRBuilder<> & Builder,Value * Acc,Value * Src,unsigned Op,RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind,ArrayRef<Value * > RedOps)1534 llvm::getOrderedReduction(IRBuilder<> &Builder, Value *Acc, Value *Src,
1535                           unsigned Op,
1536                           RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind,
1537                           ArrayRef<Value *> RedOps) {
1538   unsigned VF = Src->getType()->getVectorNumElements();
1539 
1540   // Extract and apply reduction ops in ascending order:
1541   // e.g. ((((Acc + Scl[0]) + Scl[1]) + Scl[2]) + ) ... + Scl[VF-1]
1542   Value *Result = Acc;
1543   for (unsigned ExtractIdx = 0; ExtractIdx != VF; ++ExtractIdx) {
1544     Value *Ext =
1545         Builder.CreateExtractElement(Src, Builder.getInt32(ExtractIdx));
1546 
1547     if (Op != Instruction::ICmp && Op != Instruction::FCmp) {
1548       Result = Builder.CreateBinOp((Instruction::BinaryOps)Op, Result, Ext,
1549                                    "bin.rdx");
1550     } else {
1551       assert(MinMaxKind != RecurrenceDescriptor::MRK_Invalid &&
1552              "Invalid min/max");
1553       Result = RecurrenceDescriptor::createMinMaxOp(Builder, MinMaxKind, Result,
1554                                                     Ext);
1555     }
1556 
1557     if (!RedOps.empty())
1558       propagateIRFlags(Result, RedOps);
1559   }
1560 
1561   return Result;
1562 }
1563 
1564 // Helper to generate a log2 shuffle reduction.
1565 Value *
getShuffleReduction(IRBuilder<> & Builder,Value * Src,unsigned Op,RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind,ArrayRef<Value * > RedOps)1566 llvm::getShuffleReduction(IRBuilder<> &Builder, Value *Src, unsigned Op,
1567                           RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind,
1568                           ArrayRef<Value *> RedOps) {
1569   unsigned VF = Src->getType()->getVectorNumElements();
1570   // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
1571   // and vector ops, reducing the set of values being computed by half each
1572   // round.
1573   assert(isPowerOf2_32(VF) &&
1574          "Reduction emission only supported for pow2 vectors!");
1575   Value *TmpVec = Src;
1576   SmallVector<Constant *, 32> ShuffleMask(VF, nullptr);
1577   for (unsigned i = VF; i != 1; i >>= 1) {
1578     // Move the upper half of the vector to the lower half.
1579     for (unsigned j = 0; j != i / 2; ++j)
1580       ShuffleMask[j] = Builder.getInt32(i / 2 + j);
1581 
1582     // Fill the rest of the mask with undef.
1583     std::fill(&ShuffleMask[i / 2], ShuffleMask.end(),
1584               UndefValue::get(Builder.getInt32Ty()));
1585 
1586     Value *Shuf = Builder.CreateShuffleVector(
1587         TmpVec, UndefValue::get(TmpVec->getType()),
1588         ConstantVector::get(ShuffleMask), "rdx.shuf");
1589 
1590     if (Op != Instruction::ICmp && Op != Instruction::FCmp) {
1591       // Floating point operations had to be 'fast' to enable the reduction.
1592       TmpVec = addFastMathFlag(Builder.CreateBinOp((Instruction::BinaryOps)Op,
1593                                                    TmpVec, Shuf, "bin.rdx"));
1594     } else {
1595       assert(MinMaxKind != RecurrenceDescriptor::MRK_Invalid &&
1596              "Invalid min/max");
1597       TmpVec = RecurrenceDescriptor::createMinMaxOp(Builder, MinMaxKind, TmpVec,
1598                                                     Shuf);
1599     }
1600     if (!RedOps.empty())
1601       propagateIRFlags(TmpVec, RedOps);
1602   }
1603   // The result is in the first element of the vector.
1604   return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
1605 }
1606 
1607 /// Create a simple vector reduction specified by an opcode and some
1608 /// flags (if generating min/max reductions).
createSimpleTargetReduction(IRBuilder<> & Builder,const TargetTransformInfo * TTI,unsigned Opcode,Value * Src,TargetTransformInfo::ReductionFlags Flags,ArrayRef<Value * > RedOps)1609 Value *llvm::createSimpleTargetReduction(
1610     IRBuilder<> &Builder, const TargetTransformInfo *TTI, unsigned Opcode,
1611     Value *Src, TargetTransformInfo::ReductionFlags Flags,
1612     ArrayRef<Value *> RedOps) {
1613   assert(isa<VectorType>(Src->getType()) && "Type must be a vector");
1614 
1615   Value *ScalarUdf = UndefValue::get(Src->getType()->getVectorElementType());
1616   std::function<Value*()> BuildFunc;
1617   using RD = RecurrenceDescriptor;
1618   RD::MinMaxRecurrenceKind MinMaxKind = RD::MRK_Invalid;
1619   // TODO: Support creating ordered reductions.
1620   FastMathFlags FMFFast;
1621   FMFFast.setFast();
1622 
1623   switch (Opcode) {
1624   case Instruction::Add:
1625     BuildFunc = [&]() { return Builder.CreateAddReduce(Src); };
1626     break;
1627   case Instruction::Mul:
1628     BuildFunc = [&]() { return Builder.CreateMulReduce(Src); };
1629     break;
1630   case Instruction::And:
1631     BuildFunc = [&]() { return Builder.CreateAndReduce(Src); };
1632     break;
1633   case Instruction::Or:
1634     BuildFunc = [&]() { return Builder.CreateOrReduce(Src); };
1635     break;
1636   case Instruction::Xor:
1637     BuildFunc = [&]() { return Builder.CreateXorReduce(Src); };
1638     break;
1639   case Instruction::FAdd:
1640     BuildFunc = [&]() {
1641       auto Rdx = Builder.CreateFAddReduce(ScalarUdf, Src);
1642       cast<CallInst>(Rdx)->setFastMathFlags(FMFFast);
1643       return Rdx;
1644     };
1645     break;
1646   case Instruction::FMul:
1647     BuildFunc = [&]() {
1648       auto Rdx = Builder.CreateFMulReduce(ScalarUdf, Src);
1649       cast<CallInst>(Rdx)->setFastMathFlags(FMFFast);
1650       return Rdx;
1651     };
1652     break;
1653   case Instruction::ICmp:
1654     if (Flags.IsMaxOp) {
1655       MinMaxKind = Flags.IsSigned ? RD::MRK_SIntMax : RD::MRK_UIntMax;
1656       BuildFunc = [&]() {
1657         return Builder.CreateIntMaxReduce(Src, Flags.IsSigned);
1658       };
1659     } else {
1660       MinMaxKind = Flags.IsSigned ? RD::MRK_SIntMin : RD::MRK_UIntMin;
1661       BuildFunc = [&]() {
1662         return Builder.CreateIntMinReduce(Src, Flags.IsSigned);
1663       };
1664     }
1665     break;
1666   case Instruction::FCmp:
1667     if (Flags.IsMaxOp) {
1668       MinMaxKind = RD::MRK_FloatMax;
1669       BuildFunc = [&]() { return Builder.CreateFPMaxReduce(Src, Flags.NoNaN); };
1670     } else {
1671       MinMaxKind = RD::MRK_FloatMin;
1672       BuildFunc = [&]() { return Builder.CreateFPMinReduce(Src, Flags.NoNaN); };
1673     }
1674     break;
1675   default:
1676     llvm_unreachable("Unhandled opcode");
1677     break;
1678   }
1679   if (TTI->useReductionIntrinsic(Opcode, Src->getType(), Flags))
1680     return BuildFunc();
1681   return getShuffleReduction(Builder, Src, Opcode, MinMaxKind, RedOps);
1682 }
1683 
1684 /// Create a vector reduction using a given recurrence descriptor.
createTargetReduction(IRBuilder<> & B,const TargetTransformInfo * TTI,RecurrenceDescriptor & Desc,Value * Src,bool NoNaN)1685 Value *llvm::createTargetReduction(IRBuilder<> &B,
1686                                    const TargetTransformInfo *TTI,
1687                                    RecurrenceDescriptor &Desc, Value *Src,
1688                                    bool NoNaN) {
1689   // TODO: Support in-order reductions based on the recurrence descriptor.
1690   using RD = RecurrenceDescriptor;
1691   RD::RecurrenceKind RecKind = Desc.getRecurrenceKind();
1692   TargetTransformInfo::ReductionFlags Flags;
1693   Flags.NoNaN = NoNaN;
1694   switch (RecKind) {
1695   case RD::RK_FloatAdd:
1696     return createSimpleTargetReduction(B, TTI, Instruction::FAdd, Src, Flags);
1697   case RD::RK_FloatMult:
1698     return createSimpleTargetReduction(B, TTI, Instruction::FMul, Src, Flags);
1699   case RD::RK_IntegerAdd:
1700     return createSimpleTargetReduction(B, TTI, Instruction::Add, Src, Flags);
1701   case RD::RK_IntegerMult:
1702     return createSimpleTargetReduction(B, TTI, Instruction::Mul, Src, Flags);
1703   case RD::RK_IntegerAnd:
1704     return createSimpleTargetReduction(B, TTI, Instruction::And, Src, Flags);
1705   case RD::RK_IntegerOr:
1706     return createSimpleTargetReduction(B, TTI, Instruction::Or, Src, Flags);
1707   case RD::RK_IntegerXor:
1708     return createSimpleTargetReduction(B, TTI, Instruction::Xor, Src, Flags);
1709   case RD::RK_IntegerMinMax: {
1710     RD::MinMaxRecurrenceKind MMKind = Desc.getMinMaxRecurrenceKind();
1711     Flags.IsMaxOp = (MMKind == RD::MRK_SIntMax || MMKind == RD::MRK_UIntMax);
1712     Flags.IsSigned = (MMKind == RD::MRK_SIntMax || MMKind == RD::MRK_SIntMin);
1713     return createSimpleTargetReduction(B, TTI, Instruction::ICmp, Src, Flags);
1714   }
1715   case RD::RK_FloatMinMax: {
1716     Flags.IsMaxOp = Desc.getMinMaxRecurrenceKind() == RD::MRK_FloatMax;
1717     return createSimpleTargetReduction(B, TTI, Instruction::FCmp, Src, Flags);
1718   }
1719   default:
1720     llvm_unreachable("Unhandled RecKind");
1721   }
1722 }
1723 
propagateIRFlags(Value * I,ArrayRef<Value * > VL,Value * OpValue)1724 void llvm::propagateIRFlags(Value *I, ArrayRef<Value *> VL, Value *OpValue) {
1725   auto *VecOp = dyn_cast<Instruction>(I);
1726   if (!VecOp)
1727     return;
1728   auto *Intersection = (OpValue == nullptr) ? dyn_cast<Instruction>(VL[0])
1729                                             : dyn_cast<Instruction>(OpValue);
1730   if (!Intersection)
1731     return;
1732   const unsigned Opcode = Intersection->getOpcode();
1733   VecOp->copyIRFlags(Intersection);
1734   for (auto *V : VL) {
1735     auto *Instr = dyn_cast<Instruction>(V);
1736     if (!Instr)
1737       continue;
1738     if (OpValue == nullptr || Opcode == Instr->getOpcode())
1739       VecOp->andIRFlags(V);
1740   }
1741 }
1742