LLVM  6.0.0svn
LoopUtils.cpp
Go to the documentation of this file.
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 
15 #include "llvm/ADT/ScopeExit.h"
19 #include "llvm/Analysis/LoopInfo.h"
20 #include "llvm/Analysis/LoopPass.h"
26 #include "llvm/IR/Dominators.h"
27 #include "llvm/IR/Instructions.h"
28 #include "llvm/IR/Module.h"
29 #include "llvm/IR/PatternMatch.h"
30 #include "llvm/IR/ValueHandle.h"
31 #include "llvm/Pass.h"
32 #include "llvm/Support/Debug.h"
34 
35 using namespace llvm;
36 using namespace llvm::PatternMatch;
37 
38 #define DEBUG_TYPE "loop-utils"
39 
42  for (User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
43  if (!Set.count(dyn_cast<Instruction>(*Use)))
44  return false;
45  return true;
46 }
47 
49  switch (Kind) {
50  default:
51  break;
52  case RK_IntegerAdd:
53  case RK_IntegerMult:
54  case RK_IntegerOr:
55  case RK_IntegerAnd:
56  case RK_IntegerXor:
57  case RK_IntegerMinMax:
58  return true;
59  }
60  return false;
61 }
62 
64  return (Kind != RK_NoRecurrence) && !isIntegerRecurrenceKind(Kind);
65 }
66 
68  switch (Kind) {
69  default:
70  break;
71  case RK_IntegerAdd:
72  case RK_IntegerMult:
73  case RK_FloatAdd:
74  case RK_FloatMult:
75  return true;
76  }
77  return false;
78 }
79 
84  if (!Phi->hasOneUse())
85  return Phi;
86 
87  const APInt *M = nullptr;
88  Instruction *I, *J = cast<Instruction>(Phi->use_begin()->getUser());
89 
90  // Matches either I & 2^x-1 or 2^x-1 & I. If we find a match, we update RT
91  // with a new integer type of the corresponding bit width.
92  if (match(J, m_c_And(m_Instruction(I), m_APInt(M)))) {
93  int32_t Bits = (*M + 1).exactLogBase2();
94  if (Bits > 0) {
95  RT = IntegerType::get(Phi->getContext(), Bits);
96  Visited.insert(Phi);
97  CI.insert(J);
98  return J;
99  }
100  }
101  return Phi;
102 }
103 
105  Instruction *Start, Instruction *Exit, Type *RT, bool &IsSigned,
108 
110  bool FoundOneOperand = false;
111  unsigned DstSize = RT->getPrimitiveSizeInBits();
112  Worklist.push_back(Exit);
113 
114  // Traverse the instructions in the reduction expression, beginning with the
115  // exit value.
116  while (!Worklist.empty()) {
117  Instruction *I = Worklist.pop_back_val();
118  for (Use &U : I->operands()) {
119 
120  // Terminate the traversal if the operand is not an instruction, or we
121  // reach the starting value.
122  Instruction *J = dyn_cast<Instruction>(U.get());
123  if (!J || J == Start)
124  continue;
125 
126  // Otherwise, investigate the operation if it is also in the expression.
127  if (Visited.count(J)) {
128  Worklist.push_back(J);
129  continue;
130  }
131 
132  // If the operand is not in Visited, it is not a reduction operation, but
133  // it does feed into one. Make sure it is either a single-use sign- or
134  // zero-extend instruction.
135  CastInst *Cast = dyn_cast<CastInst>(J);
136  bool IsSExtInst = isa<SExtInst>(J);
137  if (!Cast || !Cast->hasOneUse() || !(isa<ZExtInst>(J) || IsSExtInst))
138  return false;
139 
140  // Ensure the source type of the extend is no larger than the reduction
141  // type. It is not necessary for the types to be identical.
142  unsigned SrcSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
143  if (SrcSize > DstSize)
144  return false;
145 
146  // Furthermore, ensure that all such extends are of the same kind.
147  if (FoundOneOperand) {
148  if (IsSigned != IsSExtInst)
149  return false;
150  } else {
151  FoundOneOperand = true;
152  IsSigned = IsSExtInst;
153  }
154 
155  // Lastly, if the source type of the extend matches the reduction type,
156  // add the extend to CI so that we can avoid accounting for it in the
157  // cost model.
158  if (SrcSize == DstSize)
159  CI.insert(Cast);
160  }
161  }
162  return true;
163 }
164 
166  Loop *TheLoop, bool HasFunNoNaNAttr,
167  RecurrenceDescriptor &RedDes) {
168  if (Phi->getNumIncomingValues() != 2)
169  return false;
170 
171  // Reduction variables are only found in the loop header block.
172  if (Phi->getParent() != TheLoop->getHeader())
173  return false;
174 
175  // Obtain the reduction start value from the value that comes from the loop
176  // preheader.
177  Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
178 
179  // ExitInstruction is the single value which is used outside the loop.
180  // We only allow for a single reduction value to be used outside the loop.
181  // This includes users of the reduction, variables (which form a cycle
182  // which ends in the phi node).
183  Instruction *ExitInstruction = nullptr;
184  // Indicates that we found a reduction operation in our scan.
185  bool FoundReduxOp = false;
186 
187  // We start with the PHI node and scan for all of the users of this
188  // instruction. All users must be instructions that can be used as reduction
189  // variables (such as ADD). We must have a single out-of-block user. The cycle
190  // must include the original PHI.
191  bool FoundStartPHI = false;
192 
193  // To recognize min/max patterns formed by a icmp select sequence, we store
194  // the number of instruction we saw from the recognized min/max pattern,
195  // to make sure we only see exactly the two instructions.
196  unsigned NumCmpSelectPatternInst = 0;
197  InstDesc ReduxDesc(false, nullptr);
198 
199  // Data used for determining if the recurrence has been type-promoted.
200  Type *RecurrenceType = Phi->getType();
202  Instruction *Start = Phi;
203  bool IsSigned = false;
204 
205  SmallPtrSet<Instruction *, 8> VisitedInsts;
207 
208  // Return early if the recurrence kind does not match the type of Phi. If the
209  // recurrence kind is arithmetic, we attempt to look through AND operations
210  // resulting from the type promotion performed by InstCombine. Vector
211  // operations are not limited to the legal integer widths, so we may be able
212  // to evaluate the reduction in the narrower width.
213  if (RecurrenceType->isFloatingPointTy()) {
214  if (!isFloatingPointRecurrenceKind(Kind))
215  return false;
216  } else {
217  if (!isIntegerRecurrenceKind(Kind))
218  return false;
219  if (isArithmeticRecurrenceKind(Kind))
220  Start = lookThroughAnd(Phi, RecurrenceType, VisitedInsts, CastInsts);
221  }
222 
223  Worklist.push_back(Start);
224  VisitedInsts.insert(Start);
225 
226  // A value in the reduction can be used:
227  // - By the reduction:
228  // - Reduction operation:
229  // - One use of reduction value (safe).
230  // - Multiple use of reduction value (not safe).
231  // - PHI:
232  // - All uses of the PHI must be the reduction (safe).
233  // - Otherwise, not safe.
234  // - By instructions outside of the loop (safe).
235  // * One value may have several outside users, but all outside
236  // uses must be of the same value.
237  // - By an instruction that is not part of the reduction (not safe).
238  // This is either:
239  // * An instruction type other than PHI or the reduction operation.
240  // * A PHI in the header other than the initial PHI.
241  while (!Worklist.empty()) {
242  Instruction *Cur = Worklist.back();
243  Worklist.pop_back();
244 
245  // No Users.
246  // If the instruction has no users then this is a broken chain and can't be
247  // a reduction variable.
248  if (Cur->use_empty())
249  return false;
250 
251  bool IsAPhi = isa<PHINode>(Cur);
252 
253  // A header PHI use other than the original PHI.
254  if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
255  return false;
256 
257  // Reductions of instructions such as Div, and Sub is only possible if the
258  // LHS is the reduction variable.
259  if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
260  !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
261  !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
262  return false;
263 
264  // Any reduction instruction must be of one of the allowed kinds. We ignore
265  // the starting value (the Phi or an AND instruction if the Phi has been
266  // type-promoted).
267  if (Cur != Start) {
268  ReduxDesc = isRecurrenceInstr(Cur, Kind, ReduxDesc, HasFunNoNaNAttr);
269  if (!ReduxDesc.isRecurrence())
270  return false;
271  }
272 
273  // A reduction operation must only have one use of the reduction value.
274  if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
275  hasMultipleUsesOf(Cur, VisitedInsts))
276  return false;
277 
278  // All inputs to a PHI node must be a reduction value.
279  if (IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
280  return false;
281 
282  if (Kind == RK_IntegerMinMax &&
283  (isa<ICmpInst>(Cur) || isa<SelectInst>(Cur)))
284  ++NumCmpSelectPatternInst;
285  if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) || isa<SelectInst>(Cur)))
286  ++NumCmpSelectPatternInst;
287 
288  // Check whether we found a reduction operator.
289  FoundReduxOp |= !IsAPhi && Cur != Start;
290 
291  // Process users of current instruction. Push non-PHI nodes after PHI nodes
292  // onto the stack. This way we are going to have seen all inputs to PHI
293  // nodes once we get to them.
296  for (User *U : Cur->users()) {
297  Instruction *UI = cast<Instruction>(U);
298 
299  // Check if we found the exit user.
300  BasicBlock *Parent = UI->getParent();
301  if (!TheLoop->contains(Parent)) {
302  // If we already know this instruction is used externally, move on to
303  // the next user.
304  if (ExitInstruction == Cur)
305  continue;
306 
307  // Exit if you find multiple values used outside or if the header phi
308  // node is being used. In this case the user uses the value of the
309  // previous iteration, in which case we would loose "VF-1" iterations of
310  // the reduction operation if we vectorize.
311  if (ExitInstruction != nullptr || Cur == Phi)
312  return false;
313 
314  // The instruction used by an outside user must be the last instruction
315  // before we feed back to the reduction phi. Otherwise, we loose VF-1
316  // operations on the value.
317  if (!is_contained(Phi->operands(), Cur))
318  return false;
319 
320  ExitInstruction = Cur;
321  continue;
322  }
323 
324  // Process instructions only once (termination). Each reduction cycle
325  // value must only be used once, except by phi nodes and min/max
326  // reductions which are represented as a cmp followed by a select.
327  InstDesc IgnoredVal(false, nullptr);
328  if (VisitedInsts.insert(UI).second) {
329  if (isa<PHINode>(UI))
330  PHIs.push_back(UI);
331  else
332  NonPHIs.push_back(UI);
333  } else if (!isa<PHINode>(UI) &&
334  ((!isa<FCmpInst>(UI) && !isa<ICmpInst>(UI) &&
335  !isa<SelectInst>(UI)) ||
336  !isMinMaxSelectCmpPattern(UI, IgnoredVal).isRecurrence()))
337  return false;
338 
339  // Remember that we completed the cycle.
340  if (UI == Phi)
341  FoundStartPHI = true;
342  }
343  Worklist.append(PHIs.begin(), PHIs.end());
344  Worklist.append(NonPHIs.begin(), NonPHIs.end());
345  }
346 
347  // This means we have seen one but not the other instruction of the
348  // pattern or more than just a select and cmp.
349  if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
350  NumCmpSelectPatternInst != 2)
351  return false;
352 
353  if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
354  return false;
355 
356  // If we think Phi may have been type-promoted, we also need to ensure that
357  // all source operands of the reduction are either SExtInsts or ZEstInsts. If
358  // so, we will be able to evaluate the reduction in the narrower bit width.
359  if (Start != Phi)
360  if (!getSourceExtensionKind(Start, ExitInstruction, RecurrenceType,
361  IsSigned, VisitedInsts, CastInsts))
362  return false;
363 
364  // We found a reduction var if we have reached the original phi node and we
365  // only have a single instruction with out-of-loop users.
366 
367  // The ExitInstruction(Instruction which is allowed to have out-of-loop users)
368  // is saved as part of the RecurrenceDescriptor.
369 
370  // Save the description of this reduction variable.
372  RdxStart, ExitInstruction, Kind, ReduxDesc.getMinMaxKind(),
373  ReduxDesc.getUnsafeAlgebraInst(), RecurrenceType, IsSigned, CastInsts);
374  RedDes = RD;
375 
376  return true;
377 }
378 
379 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
380 /// pattern corresponding to a min(X, Y) or max(X, Y).
383 
384  assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
385  "Expect a select instruction");
386  Instruction *Cmp = nullptr;
387  SelectInst *Select = nullptr;
388 
389  // We must handle the select(cmp()) as a single instruction. Advance to the
390  // select.
391  if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
392  if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
393  return InstDesc(false, I);
394  return InstDesc(Select, Prev.getMinMaxKind());
395  }
396 
397  // Only handle single use cases for now.
398  if (!(Select = dyn_cast<SelectInst>(I)))
399  return InstDesc(false, I);
400  if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
401  !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
402  return InstDesc(false, I);
403  if (!Cmp->hasOneUse())
404  return InstDesc(false, I);
405 
406  Value *CmpLeft;
407  Value *CmpRight;
408 
409  // Look for a min/max pattern.
410  if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
411  return InstDesc(Select, MRK_UIntMin);
412  else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
413  return InstDesc(Select, MRK_UIntMax);
414  else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
415  return InstDesc(Select, MRK_SIntMax);
416  else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
417  return InstDesc(Select, MRK_SIntMin);
418  else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
419  return InstDesc(Select, MRK_FloatMin);
420  else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
421  return InstDesc(Select, MRK_FloatMax);
422  else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
423  return InstDesc(Select, MRK_FloatMin);
424  else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
425  return InstDesc(Select, MRK_FloatMax);
426 
427  return InstDesc(false, I);
428 }
429 
432  InstDesc &Prev, bool HasFunNoNaNAttr) {
433  bool FP = I->getType()->isFloatingPointTy();
434  Instruction *UAI = Prev.getUnsafeAlgebraInst();
435  if (!UAI && FP && !I->hasUnsafeAlgebra())
436  UAI = I; // Found an unsafe (unvectorizable) algebra instruction.
437 
438  switch (I->getOpcode()) {
439  default:
440  return InstDesc(false, I);
441  case Instruction::PHI:
442  return InstDesc(I, Prev.getMinMaxKind(), Prev.getUnsafeAlgebraInst());
443  case Instruction::Sub:
444  case Instruction::Add:
445  return InstDesc(Kind == RK_IntegerAdd, I);
446  case Instruction::Mul:
447  return InstDesc(Kind == RK_IntegerMult, I);
448  case Instruction::And:
449  return InstDesc(Kind == RK_IntegerAnd, I);
450  case Instruction::Or:
451  return InstDesc(Kind == RK_IntegerOr, I);
452  case Instruction::Xor:
453  return InstDesc(Kind == RK_IntegerXor, I);
454  case Instruction::FMul:
455  return InstDesc(Kind == RK_FloatMult, I, UAI);
456  case Instruction::FSub:
457  case Instruction::FAdd:
458  return InstDesc(Kind == RK_FloatAdd, I, UAI);
459  case Instruction::FCmp:
460  case Instruction::ICmp:
461  case Instruction::Select:
462  if (Kind != RK_IntegerMinMax &&
463  (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
464  return InstDesc(false, I);
465  return isMinMaxSelectCmpPattern(I, Prev);
466  }
467 }
468 
471  unsigned NumUses = 0;
472  for (User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E;
473  ++Use) {
474  if (Insts.count(dyn_cast<Instruction>(*Use)))
475  ++NumUses;
476  if (NumUses > 1)
477  return true;
478  }
479 
480  return false;
481 }
483  RecurrenceDescriptor &RedDes) {
484 
485  BasicBlock *Header = TheLoop->getHeader();
486  Function &F = *Header->getParent();
487  bool HasFunNoNaNAttr =
488  F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
489 
490  if (AddReductionVar(Phi, RK_IntegerAdd, TheLoop, HasFunNoNaNAttr, RedDes)) {
491  DEBUG(dbgs() << "Found an ADD reduction PHI." << *Phi << "\n");
492  return true;
493  }
494  if (AddReductionVar(Phi, RK_IntegerMult, TheLoop, HasFunNoNaNAttr, RedDes)) {
495  DEBUG(dbgs() << "Found a MUL reduction PHI." << *Phi << "\n");
496  return true;
497  }
498  if (AddReductionVar(Phi, RK_IntegerOr, TheLoop, HasFunNoNaNAttr, RedDes)) {
499  DEBUG(dbgs() << "Found an OR reduction PHI." << *Phi << "\n");
500  return true;
501  }
502  if (AddReductionVar(Phi, RK_IntegerAnd, TheLoop, HasFunNoNaNAttr, RedDes)) {
503  DEBUG(dbgs() << "Found an AND reduction PHI." << *Phi << "\n");
504  return true;
505  }
506  if (AddReductionVar(Phi, RK_IntegerXor, TheLoop, HasFunNoNaNAttr, RedDes)) {
507  DEBUG(dbgs() << "Found a XOR reduction PHI." << *Phi << "\n");
508  return true;
509  }
510  if (AddReductionVar(Phi, RK_IntegerMinMax, TheLoop, HasFunNoNaNAttr,
511  RedDes)) {
512  DEBUG(dbgs() << "Found a MINMAX reduction PHI." << *Phi << "\n");
513  return true;
514  }
515  if (AddReductionVar(Phi, RK_FloatMult, TheLoop, HasFunNoNaNAttr, RedDes)) {
516  DEBUG(dbgs() << "Found an FMult reduction PHI." << *Phi << "\n");
517  return true;
518  }
519  if (AddReductionVar(Phi, RK_FloatAdd, TheLoop, HasFunNoNaNAttr, RedDes)) {
520  DEBUG(dbgs() << "Found an FAdd reduction PHI." << *Phi << "\n");
521  return true;
522  }
523  if (AddReductionVar(Phi, RK_FloatMinMax, TheLoop, HasFunNoNaNAttr, RedDes)) {
524  DEBUG(dbgs() << "Found an float MINMAX reduction PHI." << *Phi << "\n");
525  return true;
526  }
527  // Not a reduction of known type.
528  return false;
529 }
530 
532  PHINode *Phi, Loop *TheLoop,
534 
535  // Ensure the phi node is in the loop header and has two incoming values.
536  if (Phi->getParent() != TheLoop->getHeader() ||
537  Phi->getNumIncomingValues() != 2)
538  return false;
539 
540  // Ensure the loop has a preheader and a single latch block. The loop
541  // vectorizer will need the latch to set up the next iteration of the loop.
542  auto *Preheader = TheLoop->getLoopPreheader();
543  auto *Latch = TheLoop->getLoopLatch();
544  if (!Preheader || !Latch)
545  return false;
546 
547  // Ensure the phi node's incoming blocks are the loop preheader and latch.
548  if (Phi->getBasicBlockIndex(Preheader) < 0 ||
549  Phi->getBasicBlockIndex(Latch) < 0)
550  return false;
551 
552  // Get the previous value. The previous value comes from the latch edge while
553  // the initial value comes form the preheader edge.
554  auto *Previous = dyn_cast<Instruction>(Phi->getIncomingValueForBlock(Latch));
555  if (!Previous || !TheLoop->contains(Previous) || isa<PHINode>(Previous) ||
556  SinkAfter.count(Previous)) // Cannot rely on dominance due to motion.
557  return false;
558 
559  // Ensure every user of the phi node is dominated by the previous value.
560  // The dominance requirement ensures the loop vectorizer will not need to
561  // vectorize the initial value prior to the first iteration of the loop.
562  // TODO: Consider extending this sinking to handle other kinds of instructions
563  // and expressions, beyond sinking a single cast past Previous.
564  if (Phi->hasOneUse()) {
565  auto *I = Phi->user_back();
566  if (I->isCast() && (I->getParent() == Phi->getParent()) && I->hasOneUse() &&
567  DT->dominates(Previous, I->user_back())) {
568  if (!DT->dominates(Previous, I)) // Otherwise we're good w/o sinking.
569  SinkAfter[I] = Previous;
570  return true;
571  }
572  }
573 
574  for (User *U : Phi->users())
575  if (auto *I = dyn_cast<Instruction>(U)) {
576  if (!DT->dominates(Previous, I))
577  return false;
578  }
579 
580  return true;
581 }
582 
583 /// This function returns the identity element (or neutral element) for
584 /// the operation K.
586  Type *Tp) {
587  switch (K) {
588  case RK_IntegerXor:
589  case RK_IntegerAdd:
590  case RK_IntegerOr:
591  // Adding, Xoring, Oring zero to a number does not change it.
592  return ConstantInt::get(Tp, 0);
593  case RK_IntegerMult:
594  // Multiplying a number by 1 does not change it.
595  return ConstantInt::get(Tp, 1);
596  case RK_IntegerAnd:
597  // AND-ing a number with an all-1 value does not change it.
598  return ConstantInt::get(Tp, -1, true);
599  case RK_FloatMult:
600  // Multiplying a number by 1 does not change it.
601  return ConstantFP::get(Tp, 1.0L);
602  case RK_FloatAdd:
603  // Adding zero to a number does not change it.
604  return ConstantFP::get(Tp, 0.0L);
605  default:
606  llvm_unreachable("Unknown recurrence kind");
607  }
608 }
609 
610 /// This function translates the recurrence kind to an LLVM binary operator.
612  switch (Kind) {
613  case RK_IntegerAdd:
614  return Instruction::Add;
615  case RK_IntegerMult:
616  return Instruction::Mul;
617  case RK_IntegerOr:
618  return Instruction::Or;
619  case RK_IntegerAnd:
620  return Instruction::And;
621  case RK_IntegerXor:
622  return Instruction::Xor;
623  case RK_FloatMult:
624  return Instruction::FMul;
625  case RK_FloatAdd:
626  return Instruction::FAdd;
627  case RK_IntegerMinMax:
628  return Instruction::ICmp;
629  case RK_FloatMinMax:
630  return Instruction::FCmp;
631  default:
632  llvm_unreachable("Unknown recurrence operation");
633  }
634 }
635 
638  Value *Left, Value *Right) {
640  switch (RK) {
641  default:
642  llvm_unreachable("Unknown min/max recurrence kind");
643  case MRK_UIntMin:
644  P = CmpInst::ICMP_ULT;
645  break;
646  case MRK_UIntMax:
647  P = CmpInst::ICMP_UGT;
648  break;
649  case MRK_SIntMin:
650  P = CmpInst::ICMP_SLT;
651  break;
652  case MRK_SIntMax:
653  P = CmpInst::ICMP_SGT;
654  break;
655  case MRK_FloatMin:
656  P = CmpInst::FCMP_OLT;
657  break;
658  case MRK_FloatMax:
659  P = CmpInst::FCMP_OGT;
660  break;
661  }
662 
663  // We only match FP sequences with unsafe algebra, so we can unconditionally
664  // set it on any generated instructions.
665  IRBuilder<>::FastMathFlagGuard FMFG(Builder);
666  FastMathFlags FMF;
667  FMF.setUnsafeAlgebra();
668  Builder.setFastMathFlags(FMF);
669 
670  Value *Cmp;
671  if (RK == MRK_FloatMin || RK == MRK_FloatMax)
672  Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
673  else
674  Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
675 
676  Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
677  return Select;
678 }
679 
680 InductionDescriptor::InductionDescriptor(Value *Start, InductionKind K,
681  const SCEV *Step, BinaryOperator *BOp)
682  : StartValue(Start), IK(K), Step(Step), InductionBinOp(BOp) {
683  assert(IK != IK_NoInduction && "Not an induction");
684 
685  // Start value type should match the induction kind and the value
686  // itself should not be null.
687  assert(StartValue && "StartValue is null");
688  assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) &&
689  "StartValue is not a pointer for pointer induction");
690  assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) &&
691  "StartValue is not an integer for integer induction");
692 
693  // Check the Step Value. It should be non-zero integer value.
694  assert((!getConstIntStepValue() || !getConstIntStepValue()->isZero()) &&
695  "Step value is zero");
696 
697  assert((IK != IK_PtrInduction || getConstIntStepValue()) &&
698  "Step value should be constant for pointer induction");
699  assert((IK == IK_FpInduction || Step->getType()->isIntegerTy()) &&
700  "StepValue is not an integer");
701 
702  assert((IK != IK_FpInduction || Step->getType()->isFloatingPointTy()) &&
703  "StepValue is not FP for FpInduction");
704  assert((IK != IK_FpInduction || (InductionBinOp &&
705  (InductionBinOp->getOpcode() == Instruction::FAdd ||
706  InductionBinOp->getOpcode() == Instruction::FSub))) &&
707  "Binary opcode should be specified for FP induction");
708 }
709 
711  ConstantInt *ConstStep = getConstIntStepValue();
712  if (ConstStep && (ConstStep->isOne() || ConstStep->isMinusOne()))
713  return ConstStep->getSExtValue();
714  return 0;
715 }
716 
718  if (isa<SCEVConstant>(Step))
719  return dyn_cast<ConstantInt>(cast<SCEVConstant>(Step)->getValue());
720  return nullptr;
721 }
722 
724  ScalarEvolution *SE,
725  const DataLayout& DL) const {
726 
727  SCEVExpander Exp(*SE, DL, "induction");
728  assert(Index->getType() == Step->getType() &&
729  "Index type does not match StepValue type");
730  switch (IK) {
731  case IK_IntInduction: {
732  assert(Index->getType() == StartValue->getType() &&
733  "Index type does not match StartValue type");
734 
735  // FIXME: Theoretically, we can call getAddExpr() of ScalarEvolution
736  // and calculate (Start + Index * Step) for all cases, without
737  // special handling for "isOne" and "isMinusOne".
738  // But in the real life the result code getting worse. We mix SCEV
739  // expressions and ADD/SUB operations and receive redundant
740  // intermediate values being calculated in different ways and
741  // Instcombine is unable to reduce them all.
742 
743  if (getConstIntStepValue() &&
744  getConstIntStepValue()->isMinusOne())
745  return B.CreateSub(StartValue, Index);
746  if (getConstIntStepValue() &&
747  getConstIntStepValue()->isOne())
748  return B.CreateAdd(StartValue, Index);
749  const SCEV *S = SE->getAddExpr(SE->getSCEV(StartValue),
750  SE->getMulExpr(Step, SE->getSCEV(Index)));
751  return Exp.expandCodeFor(S, StartValue->getType(), &*B.GetInsertPoint());
752  }
753  case IK_PtrInduction: {
754  assert(isa<SCEVConstant>(Step) &&
755  "Expected constant step for pointer induction");
756  const SCEV *S = SE->getMulExpr(SE->getSCEV(Index), Step);
757  Index = Exp.expandCodeFor(S, Index->getType(), &*B.GetInsertPoint());
758  return B.CreateGEP(nullptr, StartValue, Index);
759  }
760  case IK_FpInduction: {
761  assert(Step->getType()->isFloatingPointTy() && "Expected FP Step value");
762  assert(InductionBinOp &&
763  (InductionBinOp->getOpcode() == Instruction::FAdd ||
764  InductionBinOp->getOpcode() == Instruction::FSub) &&
765  "Original bin op should be defined for FP induction");
766 
767  Value *StepValue = cast<SCEVUnknown>(Step)->getValue();
768 
769  // Floating point operations had to be 'fast' to enable the induction.
770  FastMathFlags Flags;
771  Flags.setUnsafeAlgebra();
772 
773  Value *MulExp = B.CreateFMul(StepValue, Index);
774  if (isa<Instruction>(MulExp))
775  // We have to check, the MulExp may be a constant.
776  cast<Instruction>(MulExp)->setFastMathFlags(Flags);
777 
778  Value *BOp = B.CreateBinOp(InductionBinOp->getOpcode() , StartValue,
779  MulExp, "induction");
780  if (isa<Instruction>(BOp))
781  cast<Instruction>(BOp)->setFastMathFlags(Flags);
782 
783  return BOp;
784  }
785  case IK_NoInduction:
786  return nullptr;
787  }
788  llvm_unreachable("invalid enum");
789 }
790 
792  ScalarEvolution *SE,
794 
795  // Here we only handle FP induction variables.
796  assert(Phi->getType()->isFloatingPointTy() && "Unexpected Phi type");
797 
798  if (TheLoop->getHeader() != Phi->getParent())
799  return false;
800 
801  // The loop may have multiple entrances or multiple exits; we can analyze
802  // this phi if it has a unique entry value and a unique backedge value.
803  if (Phi->getNumIncomingValues() != 2)
804  return false;
805  Value *BEValue = nullptr, *StartValue = nullptr;
806  if (TheLoop->contains(Phi->getIncomingBlock(0))) {
807  BEValue = Phi->getIncomingValue(0);
808  StartValue = Phi->getIncomingValue(1);
809  } else {
810  assert(TheLoop->contains(Phi->getIncomingBlock(1)) &&
811  "Unexpected Phi node in the loop");
812  BEValue = Phi->getIncomingValue(1);
813  StartValue = Phi->getIncomingValue(0);
814  }
815 
816  BinaryOperator *BOp = dyn_cast<BinaryOperator>(BEValue);
817  if (!BOp)
818  return false;
819 
820  Value *Addend = nullptr;
821  if (BOp->getOpcode() == Instruction::FAdd) {
822  if (BOp->getOperand(0) == Phi)
823  Addend = BOp->getOperand(1);
824  else if (BOp->getOperand(1) == Phi)
825  Addend = BOp->getOperand(0);
826  } else if (BOp->getOpcode() == Instruction::FSub)
827  if (BOp->getOperand(0) == Phi)
828  Addend = BOp->getOperand(1);
829 
830  if (!Addend)
831  return false;
832 
833  // The addend should be loop invariant
834  if (auto *I = dyn_cast<Instruction>(Addend))
835  if (TheLoop->contains(I))
836  return false;
837 
838  // FP Step has unknown SCEV
839  const SCEV *Step = SE->getUnknown(Addend);
840  D = InductionDescriptor(StartValue, IK_FpInduction, Step, BOp);
841  return true;
842 }
843 
847  bool Assume) {
848  Type *PhiTy = Phi->getType();
849 
850  // Handle integer and pointer inductions variables.
851  // Now we handle also FP induction but not trying to make a
852  // recurrent expression from the PHI node in-place.
853 
854  if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy() &&
855  !PhiTy->isFloatTy() && !PhiTy->isDoubleTy() && !PhiTy->isHalfTy())
856  return false;
857 
858  if (PhiTy->isFloatingPointTy())
859  return isFPInductionPHI(Phi, TheLoop, PSE.getSE(), D);
860 
861  const SCEV *PhiScev = PSE.getSCEV(Phi);
862  const auto *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
863 
864  // We need this expression to be an AddRecExpr.
865  if (Assume && !AR)
866  AR = PSE.getAsAddRec(Phi);
867 
868  if (!AR) {
869  DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
870  return false;
871  }
872 
873  return isInductionPHI(Phi, TheLoop, PSE.getSE(), D, AR);
874 }
875 
877  ScalarEvolution *SE,
879  const SCEV *Expr) {
880  Type *PhiTy = Phi->getType();
881  // We only handle integer and pointer inductions variables.
882  if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
883  return false;
884 
885  // Check that the PHI is consecutive.
886  const SCEV *PhiScev = Expr ? Expr : SE->getSCEV(Phi);
887  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
888 
889  if (!AR) {
890  DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
891  return false;
892  }
893 
894  if (AR->getLoop() != TheLoop) {
895  // FIXME: We should treat this as a uniform. Unfortunately, we
896  // don't currently know how to handled uniform PHIs.
897  DEBUG(dbgs() << "LV: PHI is a recurrence with respect to an outer loop.\n");
898  return false;
899  }
900 
901  Value *StartValue =
903  const SCEV *Step = AR->getStepRecurrence(*SE);
904  // Calculate the pointer stride and check if it is consecutive.
905  // The stride may be a constant or a loop invariant integer value.
906  const SCEVConstant *ConstStep = dyn_cast<SCEVConstant>(Step);
907  if (!ConstStep && !SE->isLoopInvariant(Step, TheLoop))
908  return false;
909 
910  if (PhiTy->isIntegerTy()) {
911  D = InductionDescriptor(StartValue, IK_IntInduction, Step);
912  return true;
913  }
914 
915  assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
916  // Pointer induction should be a constant.
917  if (!ConstStep)
918  return false;
919 
920  ConstantInt *CV = ConstStep->getValue();
921  Type *PointerElementType = PhiTy->getPointerElementType();
922  // The pointer stride cannot be determined if the pointer element type is not
923  // sized.
924  if (!PointerElementType->isSized())
925  return false;
926 
927  const DataLayout &DL = Phi->getModule()->getDataLayout();
928  int64_t Size = static_cast<int64_t>(DL.getTypeAllocSize(PointerElementType));
929  if (!Size)
930  return false;
931 
932  int64_t CVSize = CV->getSExtValue();
933  if (CVSize % Size)
934  return false;
935  auto *StepValue = SE->getConstant(CV->getType(), CVSize / Size,
936  true /* signed */);
937  D = InductionDescriptor(StartValue, IK_PtrInduction, StepValue);
938  return true;
939 }
940 
942  bool PreserveLCSSA) {
943  bool Changed = false;
944 
945  // We re-use a vector for the in-loop predecesosrs.
946  SmallVector<BasicBlock *, 4> InLoopPredecessors;
947 
948  auto RewriteExit = [&](BasicBlock *BB) {
949  assert(InLoopPredecessors.empty() &&
950  "Must start with an empty predecessors list!");
951  auto Cleanup = make_scope_exit([&] { InLoopPredecessors.clear(); });
952 
953  // See if there are any non-loop predecessors of this exit block and
954  // keep track of the in-loop predecessors.
955  bool IsDedicatedExit = true;
956  for (auto *PredBB : predecessors(BB))
957  if (L->contains(PredBB)) {
958  if (isa<IndirectBrInst>(PredBB->getTerminator()))
959  // We cannot rewrite exiting edges from an indirectbr.
960  return false;
961 
962  InLoopPredecessors.push_back(PredBB);
963  } else {
964  IsDedicatedExit = false;
965  }
966 
967  assert(!InLoopPredecessors.empty() && "Must have *some* loop predecessor!");
968 
969  // Nothing to do if this is already a dedicated exit.
970  if (IsDedicatedExit)
971  return false;
972 
973  auto *NewExitBB = SplitBlockPredecessors(
974  BB, InLoopPredecessors, ".loopexit", DT, LI, PreserveLCSSA);
975 
976  if (!NewExitBB)
977  DEBUG(dbgs() << "WARNING: Can't create a dedicated exit block for loop: "
978  << *L << "\n");
979  else
980  DEBUG(dbgs() << "LoopSimplify: Creating dedicated exit block "
981  << NewExitBB->getName() << "\n");
982  return true;
983  };
984 
985  // Walk the exit blocks directly rather than building up a data structure for
986  // them, but only visit each one once.
988  for (auto *BB : L->blocks())
989  for (auto *SuccBB : successors(BB)) {
990  // We're looking for exit blocks so skip in-loop successors.
991  if (L->contains(SuccBB))
992  continue;
993 
994  // Visit each exit block exactly once.
995  if (!Visited.insert(SuccBB).second)
996  continue;
997 
998  Changed |= RewriteExit(SuccBB);
999  }
1000 
1001  return Changed;
1002 }
1003 
1004 /// \brief Returns the instructions that use values defined in the loop.
1006  SmallVector<Instruction *, 8> UsedOutside;
1007 
1008  for (auto *Block : L->getBlocks())
1009  // FIXME: I believe that this could use copy_if if the Inst reference could
1010  // be adapted into a pointer.
1011  for (auto &Inst : *Block) {
1012  auto Users = Inst.users();
1013  if (any_of(Users, [&](User *U) {
1014  auto *Use = cast<Instruction>(U);
1015  return !L->contains(Use->getParent());
1016  }))
1017  UsedOutside.push_back(&Inst);
1018  }
1019 
1020  return UsedOutside;
1021 }
1022 
1024  // By definition, all loop passes need the LoopInfo analysis and the
1025  // Dominator tree it depends on. Because they all participate in the loop
1026  // pass manager, they must also preserve these.
1031 
1032  // We must also preserve LoopSimplify and LCSSA. We locally access their IDs
1033  // here because users shouldn't directly get them from this header.
1034  extern char &LoopSimplifyID;
1035  extern char &LCSSAID;
1036  AU.addRequiredID(LoopSimplifyID);
1037  AU.addPreservedID(LoopSimplifyID);
1038  AU.addRequiredID(LCSSAID);
1039  AU.addPreservedID(LCSSAID);
1040  // This is used in the LPPassManager to perform LCSSA verification on passes
1041  // which preserve lcssa form
1044 
1045  // Loop passes are designed to run inside of a loop pass manager which means
1046  // that any function analyses they require must be required by the first loop
1047  // pass in the manager (so that it is computed before the loop pass manager
1048  // runs) and preserved by all loop pasess in the manager. To make this
1049  // reasonably robust, the set needed for most loop passes is maintained here.
1050  // If your loop pass requires an analysis not listed here, you will need to
1051  // carefully audit the loop pass manager nesting structure that results.
1059 }
1060 
1061 /// Manually defined generic "LoopPass" dependency initialization. This is used
1062 /// to initialize the exact set of passes from above in \c
1063 /// getLoopAnalysisUsage. It can be used within a loop pass's initialization
1064 /// with:
1065 ///
1066 /// INITIALIZE_PASS_DEPENDENCY(LoopPass)
1067 ///
1068 /// As-if "LoopPass" were a pass.
1072  INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
1073  INITIALIZE_PASS_DEPENDENCY(LCSSAWrapperPass)
1079 }
1080 
1081 /// \brief Find string metadata for loop
1082 ///
1083 /// If it has a value (e.g. {"llvm.distribute", 1} return the value as an
1084 /// operand or null otherwise. If the string metadata is not found return
1085 /// Optional's not-a-value.
1087  StringRef Name) {
1088  MDNode *LoopID = TheLoop->getLoopID();
1089  // Return none if LoopID is false.
1090  if (!LoopID)
1091  return None;
1092 
1093  // First operand should refer to the loop id itself.
1094  assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1095  assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1096 
1097  // Iterate over LoopID operands and look for MDString Metadata
1098  for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) {
1099  MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
1100  if (!MD)
1101  continue;
1102  MDString *S = dyn_cast<MDString>(MD->getOperand(0));
1103  if (!S)
1104  continue;
1105  // Return true if MDString holds expected MetaData.
1106  if (Name.equals(S->getString()))
1107  switch (MD->getNumOperands()) {
1108  case 1:
1109  return nullptr;
1110  case 2:
1111  return &MD->getOperand(1);
1112  default:
1113  llvm_unreachable("loop metadata has 0 or 1 operand");
1114  }
1115  }
1116  return None;
1117 }
1118 
1119 /// Does a BFS from a given node to all of its children inside a given loop.
1120 /// The returned vector of nodes includes the starting point.
1124  auto AddRegionToWorklist = [&](DomTreeNode *DTN) {
1125  // Only include subregions in the top level loop.
1126  BasicBlock *BB = DTN->getBlock();
1127  if (CurLoop->contains(BB))
1128  Worklist.push_back(DTN);
1129  };
1130 
1131  AddRegionToWorklist(N);
1132 
1133  for (size_t I = 0; I < Worklist.size(); I++)
1134  for (DomTreeNode *Child : Worklist[I]->getChildren())
1135  AddRegionToWorklist(Child);
1136 
1137  return Worklist;
1138 }
1139 
1140 void llvm::deleteDeadLoop(Loop *L, DominatorTree *DT = nullptr,
1141  ScalarEvolution *SE = nullptr,
1142  LoopInfo *LI = nullptr) {
1143  assert((!DT || L->isLCSSAForm(*DT)) && "Expected LCSSA!");
1144  auto *Preheader = L->getLoopPreheader();
1145  assert(Preheader && "Preheader should exist!");
1146 
1147  // Now that we know the removal is safe, remove the loop by changing the
1148  // branch from the preheader to go to the single exit block.
1149  //
1150  // Because we're deleting a large chunk of code at once, the sequence in which
1151  // we remove things is very important to avoid invalidation issues.
1152 
1153  // Tell ScalarEvolution that the loop is deleted. Do this before
1154  // deleting the loop so that ScalarEvolution can look at the loop
1155  // to determine what it needs to clean up.
1156  if (SE)
1157  SE->forgetLoop(L);
1158 
1159  auto *ExitBlock = L->getUniqueExitBlock();
1160  assert(ExitBlock && "Should have a unique exit block!");
1161  assert(L->hasDedicatedExits() && "Loop should have dedicated exits!");
1162 
1163  auto *OldBr = dyn_cast<BranchInst>(Preheader->getTerminator());
1164  assert(OldBr && "Preheader must end with a branch");
1165  assert(OldBr->isUnconditional() && "Preheader must have a single successor");
1166  // Connect the preheader to the exit block. Keep the old edge to the header
1167  // around to perform the dominator tree update in two separate steps
1168  // -- #1 insertion of the edge preheader -> exit and #2 deletion of the edge
1169  // preheader -> header.
1170  //
1171  //
1172  // 0. Preheader 1. Preheader 2. Preheader
1173  // | | | |
1174  // V | V |
1175  // Header <--\ | Header <--\ | Header <--\
1176  // | | | | | | | | | | |
1177  // | V | | | V | | | V |
1178  // | Body --/ | | Body --/ | | Body --/
1179  // V V V V V
1180  // Exit Exit Exit
1181  //
1182  // By doing this is two separate steps we can perform the dominator tree
1183  // update without using the batch update API.
1184  //
1185  // Even when the loop is never executed, we cannot remove the edge from the
1186  // source block to the exit block. Consider the case where the unexecuted loop
1187  // branches back to an outer loop. If we deleted the loop and removed the edge
1188  // coming to this inner loop, this will break the outer loop structure (by
1189  // deleting the backedge of the outer loop). If the outer loop is indeed a
1190  // non-loop, it will be deleted in a future iteration of loop deletion pass.
1191  IRBuilder<> Builder(OldBr);
1192  Builder.CreateCondBr(Builder.getFalse(), L->getHeader(), ExitBlock);
1193  // Remove the old branch. The conditional branch becomes a new terminator.
1194  OldBr->eraseFromParent();
1195 
1196  // Rewrite phis in the exit block to get their inputs from the Preheader
1197  // instead of the exiting block.
1198  BasicBlock::iterator BI = ExitBlock->begin();
1199  while (PHINode *P = dyn_cast<PHINode>(BI)) {
1200  // Set the zero'th element of Phi to be from the preheader and remove all
1201  // other incoming values. Given the loop has dedicated exits, all other
1202  // incoming values must be from the exiting blocks.
1203  int PredIndex = 0;
1204  P->setIncomingBlock(PredIndex, Preheader);
1205  // Removes all incoming values from all other exiting blocks (including
1206  // duplicate values from an exiting block).
1207  // Nuke all entries except the zero'th entry which is the preheader entry.
1208  // NOTE! We need to remove Incoming Values in the reverse order as done
1209  // below, to keep the indices valid for deletion (removeIncomingValues
1210  // updates getNumIncomingValues and shifts all values down into the operand
1211  // being deleted).
1212  for (unsigned i = 0, e = P->getNumIncomingValues() - 1; i != e; ++i)
1213  P->removeIncomingValue(e - i, false);
1214 
1215  assert((P->getNumIncomingValues() == 1 &&
1216  P->getIncomingBlock(PredIndex) == Preheader) &&
1217  "Should have exactly one value and that's from the preheader!");
1218  ++BI;
1219  }
1220 
1221  // Disconnect the loop body by branching directly to its exit.
1222  Builder.SetInsertPoint(Preheader->getTerminator());
1223  Builder.CreateBr(ExitBlock);
1224  // Remove the old branch.
1225  Preheader->getTerminator()->eraseFromParent();
1226 
1227  if (DT) {
1228  // Update the dominator tree by informing it about the new edge from the
1229  // preheader to the exit.
1230  DT->insertEdge(Preheader, ExitBlock);
1231  // Inform the dominator tree about the removed edge.
1232  DT->deleteEdge(Preheader, L->getHeader());
1233  }
1234 
1235  // Remove the block from the reference counting scheme, so that we can
1236  // delete it freely later.
1237  for (auto *Block : L->blocks())
1238  Block->dropAllReferences();
1239 
1240  if (LI) {
1241  // Erase the instructions and the blocks without having to worry
1242  // about ordering because we already dropped the references.
1243  // NOTE: This iteration is safe because erasing the block does not remove
1244  // its entry from the loop's block list. We do that in the next section.
1245  for (Loop::block_iterator LpI = L->block_begin(), LpE = L->block_end();
1246  LpI != LpE; ++LpI)
1247  (*LpI)->eraseFromParent();
1248 
1249  // Finally, the blocks from loopinfo. This has to happen late because
1250  // otherwise our loop iterators won't work.
1251 
1253  blocks.insert(L->block_begin(), L->block_end());
1254  for (BasicBlock *BB : blocks)
1255  LI->removeBlock(BB);
1256 
1257  // The last step is to update LoopInfo now that we've eliminated this loop.
1258  LI->erase(L);
1259  }
1260 }
1261 
1262 /// Returns true if the instruction in a loop is guaranteed to execute at least
1263 /// once.
1265  const DominatorTree *DT, const Loop *CurLoop,
1266  const LoopSafetyInfo *SafetyInfo) {
1267  // We have to check to make sure that the instruction dominates all
1268  // of the exit blocks. If it doesn't, then there is a path out of the loop
1269  // which does not execute this instruction, so we can't hoist it.
1270 
1271  // If the instruction is in the header block for the loop (which is very
1272  // common), it is always guaranteed to dominate the exit blocks. Since this
1273  // is a common case, and can save some work, check it now.
1274  if (Inst.getParent() == CurLoop->getHeader())
1275  // If there's a throw in the header block, we can't guarantee we'll reach
1276  // Inst.
1277  return !SafetyInfo->HeaderMayThrow;
1278 
1279  // Somewhere in this loop there is an instruction which may throw and make us
1280  // exit the loop.
1281  if (SafetyInfo->MayThrow)
1282  return false;
1283 
1284  // Get the exit blocks for the current loop.
1285  SmallVector<BasicBlock *, 8> ExitBlocks;
1286  CurLoop->getExitBlocks(ExitBlocks);
1287 
1288  // Verify that the block dominates each of the exit blocks of the loop.
1289  for (BasicBlock *ExitBlock : ExitBlocks)
1290  if (!DT->dominates(Inst.getParent(), ExitBlock))
1291  return false;
1292 
1293  // As a degenerate case, if the loop is statically infinite then we haven't
1294  // proven anything since there are no exit blocks.
1295  if (ExitBlocks.empty())
1296  return false;
1297 
1298  // FIXME: In general, we have to prove that the loop isn't an infinite loop.
1299  // See http::llvm.org/PR24078 . (The "ExitBlocks.empty()" check above is
1300  // just a special case of this.)
1301  return true;
1302 }
1303 
1305  // Only support loops with a unique exiting block, and a latch.
1306  if (!L->getExitingBlock())
1307  return None;
1308 
1309  // Get the branch weights for the the loop's backedge.
1310  BranchInst *LatchBR =
1312  if (!LatchBR || LatchBR->getNumSuccessors() != 2)
1313  return None;
1314 
1315  assert((LatchBR->getSuccessor(0) == L->getHeader() ||
1316  LatchBR->getSuccessor(1) == L->getHeader()) &&
1317  "At least one edge out of the latch must go to the header");
1318 
1319  // To estimate the number of times the loop body was executed, we want to
1320  // know the number of times the backedge was taken, vs. the number of times
1321  // we exited the loop.
1322  uint64_t TrueVal, FalseVal;
1323  if (!LatchBR->extractProfMetadata(TrueVal, FalseVal))
1324  return None;
1325 
1326  if (!TrueVal || !FalseVal)
1327  return 0;
1328 
1329  // Divide the count of the backedge by the count of the edge exiting the loop,
1330  // rounding to nearest.
1331  if (LatchBR->getSuccessor(0) == L->getHeader())
1332  return (TrueVal + (FalseVal / 2)) / FalseVal;
1333  else
1334  return (FalseVal + (TrueVal / 2)) / TrueVal;
1335 }
1336 
1337 /// \brief Adds a 'fast' flag to floating point operations.
1339  if (isa<FPMathOperator>(V)) {
1340  FastMathFlags Flags;
1341  Flags.setUnsafeAlgebra();
1342  cast<Instruction>(V)->setFastMathFlags(Flags);
1343  }
1344  return V;
1345 }
1346 
1347 // Helper to generate a log2 shuffle reduction.
1348 Value *
1351  ArrayRef<Value *> RedOps) {
1352  unsigned VF = Src->getType()->getVectorNumElements();
1353  // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
1354  // and vector ops, reducing the set of values being computed by half each
1355  // round.
1356  assert(isPowerOf2_32(VF) &&
1357  "Reduction emission only supported for pow2 vectors!");
1358  Value *TmpVec = Src;
1359  SmallVector<Constant *, 32> ShuffleMask(VF, nullptr);
1360  for (unsigned i = VF; i != 1; i >>= 1) {
1361  // Move the upper half of the vector to the lower half.
1362  for (unsigned j = 0; j != i / 2; ++j)
1363  ShuffleMask[j] = Builder.getInt32(i / 2 + j);
1364 
1365  // Fill the rest of the mask with undef.
1366  std::fill(&ShuffleMask[i / 2], ShuffleMask.end(),
1367  UndefValue::get(Builder.getInt32Ty()));
1368 
1369  Value *Shuf = Builder.CreateShuffleVector(
1370  TmpVec, UndefValue::get(TmpVec->getType()),
1371  ConstantVector::get(ShuffleMask), "rdx.shuf");
1372 
1373  if (Op != Instruction::ICmp && Op != Instruction::FCmp) {
1374  // Floating point operations had to be 'fast' to enable the reduction.
1375  TmpVec = addFastMathFlag(Builder.CreateBinOp((Instruction::BinaryOps)Op,
1376  TmpVec, Shuf, "bin.rdx"));
1377  } else {
1378  assert(MinMaxKind != RecurrenceDescriptor::MRK_Invalid &&
1379  "Invalid min/max");
1380  TmpVec = RecurrenceDescriptor::createMinMaxOp(Builder, MinMaxKind, TmpVec,
1381  Shuf);
1382  }
1383  if (!RedOps.empty())
1384  propagateIRFlags(TmpVec, RedOps);
1385  }
1386  // The result is in the first element of the vector.
1387  return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
1388 }
1389 
1390 /// Create a simple vector reduction specified by an opcode and some
1391 /// flags (if generating min/max reductions).
1393  IRBuilder<> &Builder, const TargetTransformInfo *TTI, unsigned Opcode,
1395  ArrayRef<Value *> RedOps) {
1396  assert(isa<VectorType>(Src->getType()) && "Type must be a vector");
1397 
1398  Value *ScalarUdf = UndefValue::get(Src->getType()->getVectorElementType());
1399  std::function<Value*()> BuildFunc;
1400  using RD = RecurrenceDescriptor;
1401  RD::MinMaxRecurrenceKind MinMaxKind = RD::MRK_Invalid;
1402  // TODO: Support creating ordered reductions.
1403  FastMathFlags FMFUnsafe;
1404  FMFUnsafe.setUnsafeAlgebra();
1405 
1406  switch (Opcode) {
1407  case Instruction::Add:
1408  BuildFunc = [&]() { return Builder.CreateAddReduce(Src); };
1409  break;
1410  case Instruction::Mul:
1411  BuildFunc = [&]() { return Builder.CreateMulReduce(Src); };
1412  break;
1413  case Instruction::And:
1414  BuildFunc = [&]() { return Builder.CreateAndReduce(Src); };
1415  break;
1416  case Instruction::Or:
1417  BuildFunc = [&]() { return Builder.CreateOrReduce(Src); };
1418  break;
1419  case Instruction::Xor:
1420  BuildFunc = [&]() { return Builder.CreateXorReduce(Src); };
1421  break;
1422  case Instruction::FAdd:
1423  BuildFunc = [&]() {
1424  auto Rdx = Builder.CreateFAddReduce(ScalarUdf, Src);
1425  cast<CallInst>(Rdx)->setFastMathFlags(FMFUnsafe);
1426  return Rdx;
1427  };
1428  break;
1429  case Instruction::FMul:
1430  BuildFunc = [&]() {
1431  auto Rdx = Builder.CreateFMulReduce(ScalarUdf, Src);
1432  cast<CallInst>(Rdx)->setFastMathFlags(FMFUnsafe);
1433  return Rdx;
1434  };
1435  break;
1436  case Instruction::ICmp:
1437  if (Flags.IsMaxOp) {
1438  MinMaxKind = Flags.IsSigned ? RD::MRK_SIntMax : RD::MRK_UIntMax;
1439  BuildFunc = [&]() {
1440  return Builder.CreateIntMaxReduce(Src, Flags.IsSigned);
1441  };
1442  } else {
1443  MinMaxKind = Flags.IsSigned ? RD::MRK_SIntMin : RD::MRK_UIntMin;
1444  BuildFunc = [&]() {
1445  return Builder.CreateIntMinReduce(Src, Flags.IsSigned);
1446  };
1447  }
1448  break;
1449  case Instruction::FCmp:
1450  if (Flags.IsMaxOp) {
1451  MinMaxKind = RD::MRK_FloatMax;
1452  BuildFunc = [&]() { return Builder.CreateFPMaxReduce(Src, Flags.NoNaN); };
1453  } else {
1454  MinMaxKind = RD::MRK_FloatMin;
1455  BuildFunc = [&]() { return Builder.CreateFPMinReduce(Src, Flags.NoNaN); };
1456  }
1457  break;
1458  default:
1459  llvm_unreachable("Unhandled opcode");
1460  break;
1461  }
1462  if (TTI->useReductionIntrinsic(Opcode, Src->getType(), Flags))
1463  return BuildFunc();
1464  return getShuffleReduction(Builder, Src, Opcode, MinMaxKind, RedOps);
1465 }
1466 
1467 /// Create a vector reduction using a given recurrence descriptor.
1469  const TargetTransformInfo *TTI,
1470  RecurrenceDescriptor &Desc, Value *Src,
1471  bool NoNaN) {
1472  // TODO: Support in-order reductions based on the recurrence descriptor.
1475  Flags.NoNaN = NoNaN;
1476  auto getSimpleRdx = [&](unsigned Opc) {
1477  return createSimpleTargetReduction(Builder, TTI, Opc, Src, Flags);
1478  };
1479  switch (RecKind) {
1481  return getSimpleRdx(Instruction::FAdd);
1483  return getSimpleRdx(Instruction::FMul);
1485  return getSimpleRdx(Instruction::Add);
1487  return getSimpleRdx(Instruction::Mul);
1489  return getSimpleRdx(Instruction::And);
1491  return getSimpleRdx(Instruction::Or);
1493  return getSimpleRdx(Instruction::Xor);
1495  switch (Desc.getMinMaxRecurrenceKind()) {
1497  Flags.IsSigned = true;
1498  Flags.IsMaxOp = true;
1499  break;
1501  Flags.IsMaxOp = true;
1502  break;
1504  Flags.IsSigned = true;
1505  break;
1507  break;
1508  default:
1509  llvm_unreachable("Unhandled MRK");
1510  }
1511  return getSimpleRdx(Instruction::ICmp);
1512  }
1514  Flags.IsMaxOp =
1516  return getSimpleRdx(Instruction::FCmp);
1517  }
1518  default:
1519  llvm_unreachable("Unhandled RecKind");
1520  }
1521 }
1522 
1524  auto *VecOp = dyn_cast<Instruction>(I);
1525  if (!VecOp)
1526  return;
1527  auto *Intersection = (OpValue == nullptr) ? dyn_cast<Instruction>(VL[0])
1528  : dyn_cast<Instruction>(OpValue);
1529  if (!Intersection)
1530  return;
1531  const unsigned Opcode = Intersection->getOpcode();
1532  VecOp->copyIRFlags(Intersection);
1533  for (auto *V : VL) {
1534  auto *Instr = dyn_cast<Instruction>(V);
1535  if (!Instr)
1536  continue;
1537  if (OpValue == nullptr || Opcode == Instr->getOpcode())
1538  VecOp->andIRFlags(V);
1539  }
1540 }
Legacy wrapper pass to provide the GlobalsAAResult object.
MaxMin_match< ICmpInst, LHS, RHS, umin_pred_ty > m_UMin(const LHS &L, const RHS &R)
Bitwise or logical XOR of numbers.
Definition: LoopUtils.h:82
Type * getVectorElementType() const
Definition: Type.h:368
const NoneType None
Definition: None.h:24
static bool isArithmeticRecurrenceKind(RecurrenceKind Kind)
Returns true if the recurrence kind is an arithmetic kind.
Definition: LoopUtils.cpp:67
IntegerType * getType() const
getType - Specialize the getType() method to always return an IntegerType, which reduces the amount o...
Definition: Constants.h:172
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:109
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:72
Value * CreateICmp(CmpInst::Predicate P, Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1634
Type * getSrcTy() const
Return the source type, as a convenience.
Definition: InstrTypes.h:818
AnalysisUsage & addPreserved()
Add the specified Pass class to the set of analyses preserved by this pass.
BlockT * getLoopLatch() const
If there is a single latch block for this loop, return it.
Definition: LoopInfoImpl.h:157
Value * CreateBinOp(Instruction::BinaryOps Opc, Value *LHS, Value *RHS, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1108
static bool isFPInductionPHI(PHINode *Phi, const Loop *L, ScalarEvolution *SE, InductionDescriptor &D)
Returns true if Phi is a floating point induction in the loop L.
Definition: LoopUtils.cpp:791
InductionDescriptor()=default
Default constructor - creates an invalid induction.
const SCEV * getConstant(ConstantInt *V)
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
BinaryOps getOpcode() const
Definition: InstrTypes.h:523
This is the interface for a simple mod/ref and alias analysis over globals.
bool isSized(SmallPtrSetImpl< Type *> *Visited=nullptr) const
Return true if it makes sense to take the size of this type.
Definition: Type.h:262
LLVM_ATTRIBUTE_ALWAYS_INLINE size_type size() const
Definition: SmallVector.h:136
int getConsecutiveDirection() const
Get the consecutive direction.
Definition: LoopUtils.cpp:710
ConstantInt * getConstIntStepValue() const
Definition: LoopUtils.cpp:717
bool isLCSSAForm(DominatorTree &DT) const
Return true if the Loop is in LCSSA form.
Definition: LoopInfo.cpp:175
Min/max implemented in terms of select(cmp()).
Definition: LoopUtils.h:83
The main scalar evolution driver.
A global registry used in conjunction with static constructors to make pluggable components (like tar...
Definition: Registry.h:45
static Instruction * lookThroughAnd(PHINode *Phi, Type *&RT, SmallPtrSetImpl< Instruction *> &Visited, SmallPtrSetImpl< Instruction *> &CI)
Determines if Phi may have been type-promoted.
Definition: LoopUtils.cpp:81
BlockT * getLoopPreheader() const
If there is a preheader for this loop, return it.
Definition: LoopInfoImpl.h:106
unsigned less than
Definition: InstrTypes.h:878
static bool isReductionPHI(PHINode *Phi, Loop *TheLoop, RecurrenceDescriptor &RedDes)
Returns true if Phi is a reduction in TheLoop.
Definition: LoopUtils.cpp:482
BasicBlock * getUniqueExitBlock() const
If getUniqueExitBlocks would return exactly one block, return that block.
Definition: LoopInfo.cpp:435
0 1 0 0 True if ordered and less than
Definition: InstrTypes.h:859
LLVM_NODISCARD detail::scope_exit< typename std::decay< Callable >::type > make_scope_exit(Callable &&F)
Definition: ScopeExit.h:47
LLVMContext & getContext() const
All values hold a context through their type.
Definition: Value.cpp:697
bool isLoopInvariant(const SCEV *S, const Loop *L)
Return true if the value of the given SCEV is unchanging in the specified loop.
BasicBlock * getSuccessor(unsigned i) const
MinMaxRecurrenceKind getMinMaxRecurrenceKind()
Definition: LoopUtils.h:199
Metadata node.
Definition: Metadata.h:862
F(f)
const MDOperand & getOperand(unsigned I) const
Definition: Metadata.h:1067
MaxMin_match< FCmpInst, LHS, RHS, ufmax_pred_ty > m_UnordFMax(const LHS &L, const RHS &R)
Match an &#39;unordered&#39; floating point maximum function.
iv Induction Variable Users
Definition: IVUsers.cpp:51
static InstDesc isRecurrenceInstr(Instruction *I, RecurrenceKind Kind, InstDesc &Prev, bool HasFunNoNaNAttr)
Returns a struct describing if the instruction &#39;I&#39; can be a recurrence variable of type &#39;Kind&#39;...
Definition: LoopUtils.cpp:431
op_iterator op_begin()
Definition: User.h:214
static bool hasMultipleUsesOf(Instruction *I, SmallPtrSetImpl< Instruction *> &Insts)
Returns true if instruction I has multiple uses in Insts.
Definition: LoopUtils.cpp:469
IntegerType * getInt32Ty()
Fetch the type representing a 32-bit integer.
Definition: IRBuilder.h:348
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:49
AnalysisUsage & addRequired()
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:51
This is the interface for a SCEV-based alias analysis.
void initializeLoopPassPass(PassRegistry &)
Manually defined generic "LoopPass" dependency initialization.
Definition: LoopUtils.cpp:1069
This class represents the LLVM &#39;select&#39; instruction.
Type * getPointerElementType() const
Definition: Type.h:373
const DataLayout & getDataLayout() const
Get the data layout for the module&#39;s target platform.
Definition: Module.cpp:361
int getBasicBlockIndex(const BasicBlock *BB) const
Return the first index of the specified basic block in the value list for this PHI.
This is the base class for all instructions that perform data casts.
Definition: InstrTypes.h:560
bool isFloatingPointTy() const
Return true if this is one of the six floating-point types.
Definition: Type.h:162
Value * getShuffleReduction(IRBuilder<> &Builder, Value *Src, unsigned Op, RecurrenceDescriptor::MinMaxRecurrenceKind MinMaxKind=RecurrenceDescriptor::MRK_Invalid, ArrayRef< Value *> RedOps=ArrayRef< Value *>())
Generates a vector reduction using shufflevectors to reduce the value.
Definition: LoopUtils.cpp:1349
unsigned getNumSuccessors() const
CallInst * CreateFAddReduce(Value *Acc, Value *Src)
Create a vector fadd reduction intrinsic of the source vector.
Definition: IRBuilder.cpp:204
A Use represents the edge between a Value definition and its users.
Definition: Use.h:56
bool isIntegerTy() const
True if this is an instance of IntegerType.
Definition: Type.h:197
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Definition: IRBuilder.h:664
bool hasUnsafeAlgebra() const
Determine whether the unsafe-algebra flag is set.
static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl< Instruction *> &Set)
Returns true if all uses of the instruction I is within the Set.
Definition: LoopUtils.cpp:40
static InstDesc isMinMaxSelectCmpPattern(Instruction *I, InstDesc &Prev)
Returns a struct describing if the instruction if the instruction is a Select(ICmp(X, Y), X, Y) instruction pattern corresponding to a min(X, Y) or max(X, Y).
Definition: LoopUtils.cpp:382
Value * CreateAdd(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:889
MaxMin_match< FCmpInst, LHS, RHS, ufmin_pred_ty > m_UnordFMin(const LHS &L, const RHS &R)
Match an &#39;unordered&#39; floating point minimum function.
SmallVector< Instruction *, 8 > findDefsUsedOutsideOfLoop(Loop *L)
Returns the instructions that use values defined in the loop.
Definition: LoopUtils.cpp:1005
BlockT * getHeader() const
Definition: LoopInfo.h:100
CallInst * CreateFPMinReduce(Value *Src, bool NoNaN=false)
Create a vector float min reduction intrinsic of the source vector.
Definition: IRBuilder.cpp:272
void deleteDeadLoop(Loop *L, DominatorTree *DT, ScalarEvolution *SE, LoopInfo *LI)
This function deletes dead loops.
Definition: LoopUtils.cpp:1140
void getExitBlocks(SmallVectorImpl< BlockT *> &ExitBlocks) const
Return all of the successor blocks of this loop.
Definition: LoopInfoImpl.h:63
bool hasDedicatedExits() const
Return true if no exit block for the loop has a predecessor that is outside the loop.
Definition: LoopInfo.cpp:380
static bool getSourceExtensionKind(Instruction *Start, Instruction *Exit, Type *RT, bool &IsSigned, SmallPtrSetImpl< Instruction *> &Visited, SmallPtrSetImpl< Instruction *> &CI)
Returns true if all the source operands of a recurrence are either SExtInsts or ZExtInsts.
Definition: LoopUtils.cpp:104
ConstantInt * getValue() const
bool isOne() const
This is just a convenience method to make client code smaller for a common case.
Definition: Constants.h:201
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:245
This node represents a polynomial recurrence on the trip count of the specified loop.
This instruction compares its operands according to the predicate given to the constructor.
void andIRFlags(const Value *V)
Logical &#39;and&#39; of any supported wrapping, exact, and fast-math flags of V and this instruction...
MaxMin_match< FCmpInst, LHS, RHS, ofmin_pred_ty > m_OrdFMin(const LHS &L, const RHS &R)
Match an &#39;ordered&#39; floating point minimum function.
MaxMin_match< ICmpInst, LHS, RHS, smin_pred_ty > m_SMin(const LHS &L, const RHS &R)
const SCEVAddRecExpr * getAsAddRec(Value *V)
Attempts to produce an AddRecExpr for V by adding additional SCEV predicates.
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory)...
Definition: APInt.h:33
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:125
AnalysisUsage & addPreservedID(const void *ID)
Value * CreateSub(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:911
bool isMinusOne() const
This function will return true iff every bit in this constant is set to true.
Definition: Constants.h:209
This POD struct holds information about a potential recurrence operation.
Definition: LoopUtils.h:111
const SCEV * getStepRecurrence(ScalarEvolution &SE) const
Constructs and returns the recurrence indicating how much this expression steps by.
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:140
Bitwise or logical AND of numbers.
Definition: LoopUtils.h:81
Value * getOperand(unsigned i) const
Definition: User.h:154
const SCEV * getSCEV(Value *V)
Returns the SCEV expression of V, in the context of the current SCEV predicate.
CallInst * CreateXorReduce(Value *Src)
Create a vector int XOR reduction intrinsic of the source vector.
Definition: IRBuilder.cpp:244
StringRef getString() const
Definition: Metadata.cpp:456
static bool AddReductionVar(PHINode *Phi, RecurrenceKind Kind, Loop *TheLoop, bool HasFunNoNaNAttr, RecurrenceDescriptor &RedDes)
Returns true if Phi is a reduction of type Kind and adds it to the RecurrenceDescriptor.
Definition: LoopUtils.cpp:165
Value * CreateFCmp(CmpInst::Predicate P, Value *LHS, Value *RHS, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1641
bool isFloatTy() const
Return true if this is &#39;float&#39;, a 32-bit IEEE fp type.
Definition: Type.h:147
#define P(N)
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
MaxMin_match< ICmpInst, LHS, RHS, umax_pred_ty > m_UMax(const LHS &L, const RHS &R)
* if(!EatIfPresent(lltok::kw_thread_local)) return false
ParseOptionalThreadLocal := /*empty.
apint_match m_APInt(const APInt *&Res)
Match a ConstantInt or splatted ConstantVector, binding the specified pointer to the contained APInt...
Definition: PatternMatch.h:260
constexpr bool isPowerOf2_32(uint32_t Value)
Return true if the argument is a power of two > 0.
Definition: MathExtras.h:421
LLVM Basic Block Representation.
Definition: BasicBlock.h:59
Flags describing the kind of vector reduction.
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:46
Value * createSimpleTargetReduction(IRBuilder<> &B, const TargetTransformInfo *TTI, unsigned Opcode, Value *Src, TargetTransformInfo::ReductionFlags Flags=TargetTransformInfo::ReductionFlags(), ArrayRef< Value *> RedOps=ArrayRef< Value *>())
Create a target reduction of the given vector.
Definition: LoopUtils.cpp:1392
Conditional or Unconditional Branch instruction.
Min/max implemented in terms of select(cmp()).
Definition: LoopUtils.h:86
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
This is an important base class in LLVM.
Definition: Constant.h:42
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator begin()
Definition: SmallVector.h:116
Value * getIncomingValueForBlock(const BasicBlock *BB) const
Value * CreateSelect(Value *C, Value *True, Value *False, const Twine &Name="", Instruction *MDFrom=nullptr)
Definition: IRBuilder.h:1689
bool isPointerTy() const
True if this is an instance of PointerType.
Definition: Type.h:221
BinaryOp_match< LHS, RHS, Instruction::And, true > m_c_And(const LHS &L, const RHS &R)
Matches an And with LHS and RHS in either order.
char & LCSSAID
Definition: LCSSA.cpp:412
std::pair< iterator, bool > insert(PtrType Ptr)
Inserts Ptr if and only if there is no element in the container equal to Ptr.
Definition: SmallPtrSet.h:371
const SCEV * getAddExpr(SmallVectorImpl< const SCEV *> &Ops, SCEV::NoWrapFlags Flags=SCEV::FlagAnyWrap, unsigned Depth=0)
Get a canonical add expression, or something simpler if possible.
const std::vector< BlockT * > & getBlocks() const
Get a list of the basic blocks which make up this loop.
Definition: LoopInfo.h:149
void setUnsafeAlgebra()
Definition: Operator.h:204
Represent the analysis usage information of a pass.
op_iterator op_end()
Definition: User.h:216
bool any_of(R &&Range, UnaryPredicate P)
Provide wrappers to std::any_of which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:823
bool isHalfTy() const
Return true if this is &#39;half&#39;, a 16-bit IEEE fp type.
Definition: Type.h:144
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:853
Value * expandCodeFor(const SCEV *SH, Type *Ty, Instruction *I)
Insert code to directly compute the specified SCEV expression into the program.
std::vector< BasicBlock *>::const_iterator block_iterator
Definition: LoopInfo.h:153
op_range operands()
Definition: User.h:222
BasicBlock * SplitBlockPredecessors(BasicBlock *BB, ArrayRef< BasicBlock *> Preds, const char *Suffix, DominatorTree *DT=nullptr, LoopInfo *LI=nullptr, bool PreserveLCSSA=false)
This method introduces at least one new basic block into the function and moves some of the predecess...
size_type count(ConstPtrType Ptr) const
count - Return 1 if the specified pointer is in the set, 0 otherwise.
Definition: SmallPtrSet.h:382
Optional< unsigned > getLoopEstimatedTripCount(Loop *L)
Get a loop&#39;s estimated trip count based on branch weight metadata.
Definition: LoopUtils.cpp:1304
Value * CreateExtractElement(Value *Vec, Value *Idx, const Twine &Name="")
Definition: IRBuilder.h:1709
static UndefValue * get(Type *T)
Static factory methods - Return an &#39;undef&#39; object of the specified type.
Definition: Constants.cpp:1320
ScalarEvolution * getSE() const
Returns the ScalarEvolution analysis used.
bool IsMaxOp
If the op a min/max kind, true if it&#39;s a max operation.
RecurrenceKind getRecurrenceKind()
Definition: LoopUtils.h:197
Value * getIncomingValue(unsigned i) const
Return incoming value number x.
Optional< const MDOperand * > findStringMetadataForLoop(Loop *TheLoop, StringRef Name)
Find string metadata for loop.
Definition: LoopUtils.cpp:1086
static unsigned getRecurrenceBinOp(RecurrenceKind Kind)
Returns the opcode of binary operation corresponding to the RecurrenceKind.
Definition: LoopUtils.cpp:611
const SCEV * getMulExpr(SmallVectorImpl< const SCEV *> &Ops, SCEV::NoWrapFlags Flags=SCEV::FlagAnyWrap, unsigned Depth=0)
Get a canonical multiply expression, or something simpler if possible.
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
void propagateIRFlags(Value *I, ArrayRef< Value *> VL, Value *OpValue=nullptr)
Get the intersection (logical and) of all of the potential IR flags of each scalar operation (VL) tha...
Definition: LoopUtils.cpp:1523
static bool isFirstOrderRecurrence(PHINode *Phi, Loop *TheLoop, DenseMap< Instruction *, Instruction *> &SinkAfter, DominatorTree *DT)
Returns true if Phi is a first-order recurrence.
Definition: LoopUtils.cpp:531
signed greater than
Definition: InstrTypes.h:880
char & LoopSimplifyID
The RecurrenceDescriptor is used to identify recurrences variables in a loop.
Definition: LoopUtils.h:73
Value * CreateGEP(Value *Ptr, ArrayRef< Value *> IdxList, const Twine &Name="")
Definition: IRBuilder.h:1223
0 0 1 0 True if ordered and greater than
Definition: InstrTypes.h:857
Value * transform(IRBuilder<> &B, Value *Index, ScalarEvolution *SE, const DataLayout &DL) const
Compute the transformed value of Index at offset StartValue using step StepValue. ...
Definition: LoopUtils.cpp:723
static IntegerType * get(LLVMContext &C, unsigned NumBits)
This static method is the primary way of constructing an IntegerType.
Definition: Type.cpp:240
bool contains(const LoopT *L) const
Return true if the specified loop is contained within in this loop.
Definition: LoopInfo.h:110
Iterator for intrusive lists based on ilist_node.
MinMaxRecurrenceKind getMinMaxKind()
Definition: LoopUtils.h:127
CallInst * CreateAddReduce(Value *Src)
Create a vector int add reduction intrinsic of the source vector.
Definition: IRBuilder.cpp:224
Legacy wrapper pass to provide the SCEVAAResult object.
This is the shared class of boolean and integer constants.
Definition: Constants.h:84
Type * getType() const
Return the LLVM type of this SCEV expression.
A struct for saving information about induction variables.
Definition: LoopUtils.h:271
Bitwise or logical OR of numbers.
Definition: LoopUtils.h:80
This pass provides access to the codegen interfaces that are needed for IR-level transformations.
CallInst * CreateIntMaxReduce(Value *Src, bool IsSigned=false)
Create a vector integer max reduction intrinsic of the source vector.
Definition: IRBuilder.cpp:249
AnalysisUsage & addRequiredID(const void *ID)
Definition: Pass.cpp:298
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:864
bool dominates(const Instruction *Def, const Use &U) const
Return true if Def dominates a use in User.
Definition: Dominators.cpp:239
Module.h This file contains the declarations for the Module class.
Instruction * user_back()
Specialize the methods defined in Value, as we know that an instruction can only be used by other ins...
Definition: Instruction.h:63
static bool isFloatingPointRecurrenceKind(RecurrenceKind Kind)
Returns true if the recurrence kind is a floating point kind.
Definition: LoopUtils.cpp:63
An interface layer with SCEV used to manage how we see SCEV expressions for values in the context of ...
signed less than
Definition: InstrTypes.h:882
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:385
ConstantInt * getInt32(uint32_t C)
Get a constant 32-bit value.
Definition: IRBuilder.h:308
Value * CreateShuffleVector(Value *V1, Value *V2, Value *Mask, const Twine &Name="")
Definition: IRBuilder.h:1736
static GCRegistry::Add< StatepointGC > D("statepoint-example", "an example strategy for statepoint")
static Constant * get(Type *Ty, uint64_t V, bool isSigned=false)
If Ty is a vector type, return a Constant with a splat of the given value.
Definition: Constants.cpp:560
pred_range predecessors(BasicBlock *BB)
Definition: CFG.h:110
static Constant * get(Type *Ty, double V)
This returns a ConstantFP, or a vector containing a splat of a ConstantFP, for the specified value in...
Definition: Constants.cpp:623
CallInst * CreateFMulReduce(Value *Acc, Value *Src)
Create a vector fmul reduction intrinsic of the source vector.
Definition: IRBuilder.cpp:214
unsigned getNumIncomingValues() const
Return the number of incoming edges.
bool isCommutative() const
Return true if the instruction is commutative:
Definition: Instruction.h:435
Value * createTargetReduction(IRBuilder<> &B, const TargetTransformInfo *TTI, RecurrenceDescriptor &Desc, Value *Src, bool NoNaN=false)
Create a generic target reduction using a recurrence descriptor Desc The target is queried to determi...
Definition: LoopUtils.cpp:1468
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:132
unsigned getVectorNumElements() const
Definition: DerivedTypes.h:462
static Constant * getRecurrenceIdentity(RecurrenceKind K, Type *Tp)
Returns identity corresponding to the RecurrenceKind.
Definition: LoopUtils.cpp:585
const Module * getModule() const
Return the module owning the function this instruction belongs to or nullptr it the function does not...
Definition: Instruction.cpp:57
Class for arbitrary precision integers.
Definition: APInt.h:69
iterator_range< user_iterator > users()
Definition: Value.h:395
static Value * createMinMaxOp(IRBuilder<> &Builder, MinMaxRecurrenceKind RK, Value *Left, Value *Right)
Returns a Min/Max operation corresponding to MinMaxRecurrenceKind.
Definition: LoopUtils.cpp:636
This class uses information about analyze scalars to rewrite expressions in canonical form...
LLVM_NODISCARD LLVM_ATTRIBUTE_ALWAYS_INLINE bool equals(StringRef RHS) const
equals - Check for string equality, this is more efficient than compare() when the relative ordering ...
Definition: StringRef.h:169
void append(in_iter in_start, in_iter in_end)
Add the specified range to the end of the SmallVector.
Definition: SmallVector.h:398
uint64_t getTypeAllocSize(Type *Ty) const
Returns the offset in bytes between successive objects of the specified type, including alignment pad...
Definition: DataLayout.h:405
use_iterator use_begin()
Definition: Value.h:334
CallInst * CreateAndReduce(Value *Src)
Create a vector int AND reduction intrinsic of the source vector.
Definition: IRBuilder.cpp:234
static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT, AssumptionCache *AC)
Definition: Lint.cpp:529
MDNode * getLoopID() const
Return the llvm.loop loop id metadata node for this loop if it is present.
Definition: LoopInfo.cpp:213
CallInst * CreateOrReduce(Value *Src)
Create a vector int OR reduction intrinsic of the source vector.
Definition: IRBuilder.cpp:239
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator end()
Definition: SmallVector.h:120
CallInst * CreateIntMinReduce(Value *Src, bool IsSigned=false)
Create a vector integer min reduction intrinsic of the source vector.
Definition: IRBuilder.cpp:255
Captures loop safety information.
Definition: LoopUtils.h:51
This class represents an analyzed expression in the program.
static bool isIntegerRecurrenceKind(RecurrenceKind Kind)
Returns true if the recurrence kind is an integer kind.
Definition: LoopUtils.cpp:48
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:61
StringRef getValueAsString() const
Return the attribute&#39;s value as a string.
Definition: Attributes.cpp:195
Represents a single loop in the control flow graph.
Definition: LoopInfo.h:420
BasicBlock * getIncomingBlock(unsigned i) const
Return incoming basic block number i.
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:108
#define I(x, y, z)
Definition: MD5.cpp:58
#define N
MaxMin_match< ICmpInst, LHS, RHS, smax_pred_ty > m_SMax(const LHS &L, const RHS &R)
void getLoopAnalysisUsage(AnalysisUsage &AU)
Helper to consistently add the set of standard passes to a loop pass&#39;s AnalysisUsage.
Definition: LoopUtils.cpp:1023
CallInst * CreateMulReduce(Value *Src)
Create a vector int mul reduction intrinsic of the source vector.
Definition: IRBuilder.cpp:229
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:323
block_iterator block_end() const
Definition: LoopInfo.h:155
size_type count(const_arg_type_t< KeyT > Val) const
Return 1 if the specified key is in the map, 0 otherwise.
Definition: DenseMap.h:141
SmallVector< DomTreeNode *, 16 > collectChildrenInLoop(DomTreeNode *N, const Loop *CurLoop)
Does a BFS from a given node to all of its children inside a given loop.
Definition: LoopUtils.cpp:1122
const unsigned Kind
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
user_iterator user_begin()
Definition: Value.h:371
MaxMin_match< FCmpInst, LHS, RHS, ofmax_pred_ty > m_OrdFMax(const LHS &L, const RHS &R)
Match an &#39;ordered&#39; floating point maximum function.
unsigned getPrimitiveSizeInBits() const LLVM_READONLY
Return the basic size of this type if it is a primitive type.
Definition: Type.cpp:115
LLVM Value Representation.
Definition: Value.h:73
succ_range successors(BasicBlock *BB)
Definition: CFG.h:143
const SCEV * getSCEV(Value *V)
Return a SCEV expression for the full generality of the specified expression.
bool useReductionIntrinsic(unsigned Opcode, Type *Ty, ReductionFlags Flags) const
constexpr char Size[]
Key for Kernel::Arg::Metadata::mSize.
void setFastMathFlags(FastMathFlags NewFMF)
Set the fast-math flags to be used with generated fp-math operators.
Definition: IRBuilder.h:221
BasicBlock::iterator GetInsertPoint() const
Definition: IRBuilder.h:123
Attribute getFnAttribute(Attribute::AttrKind Kind) const
Return the attribute for the given attribute kind.
Definition: Function.h:270
bool formDedicatedExitBlocks(Loop *L, DominatorTree *DT, LoopInfo *LI, bool PreserveLCSSA)
Ensure that all exit blocks of the loop are dedicated exits.
Definition: LoopUtils.cpp:941
#define DEBUG(X)
Definition: Debug.h:118
const SCEV * getUnknown(Value *V)
The legacy pass manager&#39;s analysis pass to compute loop information.
Definition: LoopInfo.h:932
bool hasOneUse() const
Return true if there is exactly one user of this value.
Definition: Value.h:408
Convenience struct for specifying and reasoning about fast-math flags.
Definition: Operator.h:160
unsigned greater than
Definition: InstrTypes.h:876
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:49
This is the interface for LLVM&#39;s primary stateless and local alias analysis.
PassRegistry - This class manages the registration and intitialization of the pass subsystem as appli...
Definition: PassRegistry.h:39
A single uniqued string.
Definition: Metadata.h:602
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:267
This pass exposes codegen information to IR-level passes.
A wrapper pass to provide the legacy pass manager access to a suitably prepared AAResults object...
int64_t getSExtValue() const
Return the constant as a 64-bit integer value after it has been sign extended as appropriate for the ...
Definition: Constants.h:157
const TerminatorInst * getTerminator() const LLVM_READONLY
Returns the terminator instruction if the block is well formed or null if the block is not well forme...
Definition: BasicBlock.cpp:120
CallInst * CreateFPMaxReduce(Value *Src, bool NoNaN=false)
Create a vector float max reduction intrinsic of the source vector.
Definition: IRBuilder.cpp:261
unsigned getNumOperands() const
Return number of MDNode operands.
Definition: Metadata.h:1073
bool extractProfMetadata(uint64_t &TrueVal, uint64_t &FalseVal) const
Retrieve the raw weight values of a conditional branch or select.
Definition: Metadata.cpp:1303
Value * CreateFMul(Value *LHS, Value *RHS, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:947
iterator_range< block_iterator > blocks() const
Definition: LoopInfo.h:156
bool isGuaranteedToExecute(const Instruction &Inst, const DominatorTree *DT, const Loop *CurLoop, const LoopSafetyInfo *SafetyInfo)
Returns true if the instruction in a loop is guaranteed to execute at least once. ...
Definition: LoopUtils.cpp:1264
bool isDoubleTy() const
Return true if this is &#39;double&#39;, a 64-bit IEEE fp type.
Definition: Type.h:150
static Value * addFastMathFlag(Value *V)
Adds a &#39;fast&#39; flag to floating point operations.
Definition: LoopUtils.cpp:1338
block_iterator block_begin() const
Definition: LoopInfo.h:154
BlockT * getExitingBlock() const
If getExitingBlocks would return exactly one block, return that block.
Definition: LoopInfoImpl.h:50
bool NoNaN
If op is an fp min/max, whether NaNs may be present.
bool use_empty() const
Definition: Value.h:322
static Constant * get(ArrayRef< Constant *> V)
Definition: Constants.cpp:984
static bool isInductionPHI(PHINode *Phi, const Loop *L, ScalarEvolution *SE, InductionDescriptor &D, const SCEV *Expr=nullptr)
Returns true if Phi is an induction in the loop L.
Definition: LoopUtils.cpp:876
bind_ty< Instruction > m_Instruction(Instruction *&I)
Match an instruction, capturing it if we match.
Definition: PatternMatch.h:359
bool empty() const
empty - Check if the array is empty.
Definition: ArrayRef.h:144
RecurrenceKind
This enum represents the kinds of recurrences that we support.
Definition: LoopUtils.h:76
const BasicBlock * getParent() const
Definition: Instruction.h:66
This class represents a constant integer value.
Legacy wrapper pass to provide the BasicAAResult object.
bool IsSigned
Whether the operation is a signed int reduction.
bool is_contained(R &&Range, const E &Element)
Wrapper function around std::find to detect if an element exists in a container.
Definition: STLExtras.h:870