xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Utils/LoopUtils.cpp (revision 5916ae1fb14a28c69818f16ed2903248e08d50b3)
1  //===-- LoopUtils.cpp - Loop Utility functions -------------------------===//
2  //
3  // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4  // See https://llvm.org/LICENSE.txt for license information.
5  // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6  //
7  //===----------------------------------------------------------------------===//
8  //
9  // This file defines common loop utility functions.
10  //
11  //===----------------------------------------------------------------------===//
12  
13  #include "llvm/Transforms/Utils/LoopUtils.h"
14  #include "llvm/ADT/DenseSet.h"
15  #include "llvm/ADT/Optional.h"
16  #include "llvm/ADT/PriorityWorklist.h"
17  #include "llvm/ADT/ScopeExit.h"
18  #include "llvm/ADT/SetVector.h"
19  #include "llvm/ADT/SmallPtrSet.h"
20  #include "llvm/ADT/SmallVector.h"
21  #include "llvm/Analysis/AliasAnalysis.h"
22  #include "llvm/Analysis/BasicAliasAnalysis.h"
23  #include "llvm/Analysis/DomTreeUpdater.h"
24  #include "llvm/Analysis/GlobalsModRef.h"
25  #include "llvm/Analysis/InstructionSimplify.h"
26  #include "llvm/Analysis/LoopAccessAnalysis.h"
27  #include "llvm/Analysis/LoopInfo.h"
28  #include "llvm/Analysis/LoopPass.h"
29  #include "llvm/Analysis/MemorySSA.h"
30  #include "llvm/Analysis/MemorySSAUpdater.h"
31  #include "llvm/Analysis/MustExecute.h"
32  #include "llvm/Analysis/ScalarEvolution.h"
33  #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
34  #include "llvm/Analysis/ScalarEvolutionExpressions.h"
35  #include "llvm/Analysis/TargetTransformInfo.h"
36  #include "llvm/Analysis/ValueTracking.h"
37  #include "llvm/IR/DIBuilder.h"
38  #include "llvm/IR/Dominators.h"
39  #include "llvm/IR/Instructions.h"
40  #include "llvm/IR/IntrinsicInst.h"
41  #include "llvm/IR/MDBuilder.h"
42  #include "llvm/IR/Module.h"
43  #include "llvm/IR/Operator.h"
44  #include "llvm/IR/PatternMatch.h"
45  #include "llvm/IR/ValueHandle.h"
46  #include "llvm/InitializePasses.h"
47  #include "llvm/Pass.h"
48  #include "llvm/Support/Debug.h"
49  #include "llvm/Support/KnownBits.h"
50  #include "llvm/Transforms/Utils/BasicBlockUtils.h"
51  #include "llvm/Transforms/Utils/Local.h"
52  #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
53  
54  using namespace llvm;
55  using namespace llvm::PatternMatch;
56  
57  #define DEBUG_TYPE "loop-utils"
58  
59  static const char *LLVMLoopDisableNonforced = "llvm.loop.disable_nonforced";
60  static const char *LLVMLoopDisableLICM = "llvm.licm.disable";
61  
62  bool llvm::formDedicatedExitBlocks(Loop *L, DominatorTree *DT, LoopInfo *LI,
63                                     MemorySSAUpdater *MSSAU,
64                                     bool PreserveLCSSA) {
65    bool Changed = false;
66  
67    // We re-use a vector for the in-loop predecesosrs.
68    SmallVector<BasicBlock *, 4> InLoopPredecessors;
69  
70    auto RewriteExit = [&](BasicBlock *BB) {
71      assert(InLoopPredecessors.empty() &&
72             "Must start with an empty predecessors list!");
73      auto Cleanup = make_scope_exit([&] { InLoopPredecessors.clear(); });
74  
75      // See if there are any non-loop predecessors of this exit block and
76      // keep track of the in-loop predecessors.
77      bool IsDedicatedExit = true;
78      for (auto *PredBB : predecessors(BB))
79        if (L->contains(PredBB)) {
80          if (isa<IndirectBrInst>(PredBB->getTerminator()))
81            // We cannot rewrite exiting edges from an indirectbr.
82            return false;
83          if (isa<CallBrInst>(PredBB->getTerminator()))
84            // We cannot rewrite exiting edges from a callbr.
85            return false;
86  
87          InLoopPredecessors.push_back(PredBB);
88        } else {
89          IsDedicatedExit = false;
90        }
91  
92      assert(!InLoopPredecessors.empty() && "Must have *some* loop predecessor!");
93  
94      // Nothing to do if this is already a dedicated exit.
95      if (IsDedicatedExit)
96        return false;
97  
98      auto *NewExitBB = SplitBlockPredecessors(
99          BB, InLoopPredecessors, ".loopexit", DT, LI, MSSAU, PreserveLCSSA);
100  
101      if (!NewExitBB)
102        LLVM_DEBUG(
103            dbgs() << "WARNING: Can't create a dedicated exit block for loop: "
104                   << *L << "\n");
105      else
106        LLVM_DEBUG(dbgs() << "LoopSimplify: Creating dedicated exit block "
107                          << NewExitBB->getName() << "\n");
108      return true;
109    };
110  
111    // Walk the exit blocks directly rather than building up a data structure for
112    // them, but only visit each one once.
113    SmallPtrSet<BasicBlock *, 4> Visited;
114    for (auto *BB : L->blocks())
115      for (auto *SuccBB : successors(BB)) {
116        // We're looking for exit blocks so skip in-loop successors.
117        if (L->contains(SuccBB))
118          continue;
119  
120        // Visit each exit block exactly once.
121        if (!Visited.insert(SuccBB).second)
122          continue;
123  
124        Changed |= RewriteExit(SuccBB);
125      }
126  
127    return Changed;
128  }
129  
130  /// Returns the instructions that use values defined in the loop.
131  SmallVector<Instruction *, 8> llvm::findDefsUsedOutsideOfLoop(Loop *L) {
132    SmallVector<Instruction *, 8> UsedOutside;
133  
134    for (auto *Block : L->getBlocks())
135      // FIXME: I believe that this could use copy_if if the Inst reference could
136      // be adapted into a pointer.
137      for (auto &Inst : *Block) {
138        auto Users = Inst.users();
139        if (any_of(Users, [&](User *U) {
140              auto *Use = cast<Instruction>(U);
141              return !L->contains(Use->getParent());
142            }))
143          UsedOutside.push_back(&Inst);
144      }
145  
146    return UsedOutside;
147  }
148  
149  void llvm::getLoopAnalysisUsage(AnalysisUsage &AU) {
150    // By definition, all loop passes need the LoopInfo analysis and the
151    // Dominator tree it depends on. Because they all participate in the loop
152    // pass manager, they must also preserve these.
153    AU.addRequired<DominatorTreeWrapperPass>();
154    AU.addPreserved<DominatorTreeWrapperPass>();
155    AU.addRequired<LoopInfoWrapperPass>();
156    AU.addPreserved<LoopInfoWrapperPass>();
157  
158    // We must also preserve LoopSimplify and LCSSA. We locally access their IDs
159    // here because users shouldn't directly get them from this header.
160    extern char &LoopSimplifyID;
161    extern char &LCSSAID;
162    AU.addRequiredID(LoopSimplifyID);
163    AU.addPreservedID(LoopSimplifyID);
164    AU.addRequiredID(LCSSAID);
165    AU.addPreservedID(LCSSAID);
166    // This is used in the LPPassManager to perform LCSSA verification on passes
167    // which preserve lcssa form
168    AU.addRequired<LCSSAVerificationPass>();
169    AU.addPreserved<LCSSAVerificationPass>();
170  
171    // Loop passes are designed to run inside of a loop pass manager which means
172    // that any function analyses they require must be required by the first loop
173    // pass in the manager (so that it is computed before the loop pass manager
174    // runs) and preserved by all loop pasess in the manager. To make this
175    // reasonably robust, the set needed for most loop passes is maintained here.
176    // If your loop pass requires an analysis not listed here, you will need to
177    // carefully audit the loop pass manager nesting structure that results.
178    AU.addRequired<AAResultsWrapperPass>();
179    AU.addPreserved<AAResultsWrapperPass>();
180    AU.addPreserved<BasicAAWrapperPass>();
181    AU.addPreserved<GlobalsAAWrapperPass>();
182    AU.addPreserved<SCEVAAWrapperPass>();
183    AU.addRequired<ScalarEvolutionWrapperPass>();
184    AU.addPreserved<ScalarEvolutionWrapperPass>();
185    // FIXME: When all loop passes preserve MemorySSA, it can be required and
186    // preserved here instead of the individual handling in each pass.
187  }
188  
189  /// Manually defined generic "LoopPass" dependency initialization. This is used
190  /// to initialize the exact set of passes from above in \c
191  /// getLoopAnalysisUsage. It can be used within a loop pass's initialization
192  /// with:
193  ///
194  ///   INITIALIZE_PASS_DEPENDENCY(LoopPass)
195  ///
196  /// As-if "LoopPass" were a pass.
197  void llvm::initializeLoopPassPass(PassRegistry &Registry) {
198    INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
199    INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
200    INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
201    INITIALIZE_PASS_DEPENDENCY(LCSSAWrapperPass)
202    INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
203    INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)
204    INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
205    INITIALIZE_PASS_DEPENDENCY(SCEVAAWrapperPass)
206    INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
207    INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
208  }
209  
210  /// Create MDNode for input string.
211  static MDNode *createStringMetadata(Loop *TheLoop, StringRef Name, unsigned V) {
212    LLVMContext &Context = TheLoop->getHeader()->getContext();
213    Metadata *MDs[] = {
214        MDString::get(Context, Name),
215        ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(Context), V))};
216    return MDNode::get(Context, MDs);
217  }
218  
219  /// Set input string into loop metadata by keeping other values intact.
220  /// If the string is already in loop metadata update value if it is
221  /// different.
222  void llvm::addStringMetadataToLoop(Loop *TheLoop, const char *StringMD,
223                                     unsigned V) {
224    SmallVector<Metadata *, 4> MDs(1);
225    // If the loop already has metadata, retain it.
226    MDNode *LoopID = TheLoop->getLoopID();
227    if (LoopID) {
228      for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
229        MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
230        // If it is of form key = value, try to parse it.
231        if (Node->getNumOperands() == 2) {
232          MDString *S = dyn_cast<MDString>(Node->getOperand(0));
233          if (S && S->getString().equals(StringMD)) {
234            ConstantInt *IntMD =
235                mdconst::extract_or_null<ConstantInt>(Node->getOperand(1));
236            if (IntMD && IntMD->getSExtValue() == V)
237              // It is already in place. Do nothing.
238              return;
239            // We need to update the value, so just skip it here and it will
240            // be added after copying other existed nodes.
241            continue;
242          }
243        }
244        MDs.push_back(Node);
245      }
246    }
247    // Add new metadata.
248    MDs.push_back(createStringMetadata(TheLoop, StringMD, V));
249    // Replace current metadata node with new one.
250    LLVMContext &Context = TheLoop->getHeader()->getContext();
251    MDNode *NewLoopID = MDNode::get(Context, MDs);
252    // Set operand 0 to refer to the loop id itself.
253    NewLoopID->replaceOperandWith(0, NewLoopID);
254    TheLoop->setLoopID(NewLoopID);
255  }
256  
257  Optional<ElementCount>
258  llvm::getOptionalElementCountLoopAttribute(const Loop *TheLoop) {
259    Optional<int> Width =
260        getOptionalIntLoopAttribute(TheLoop, "llvm.loop.vectorize.width");
261  
262    if (Width.hasValue()) {
263      Optional<int> IsScalable = getOptionalIntLoopAttribute(
264          TheLoop, "llvm.loop.vectorize.scalable.enable");
265      return ElementCount::get(*Width, IsScalable.getValueOr(false));
266    }
267  
268    return None;
269  }
270  
271  Optional<MDNode *> llvm::makeFollowupLoopID(
272      MDNode *OrigLoopID, ArrayRef<StringRef> FollowupOptions,
273      const char *InheritOptionsExceptPrefix, bool AlwaysNew) {
274    if (!OrigLoopID) {
275      if (AlwaysNew)
276        return nullptr;
277      return None;
278    }
279  
280    assert(OrigLoopID->getOperand(0) == OrigLoopID);
281  
282    bool InheritAllAttrs = !InheritOptionsExceptPrefix;
283    bool InheritSomeAttrs =
284        InheritOptionsExceptPrefix && InheritOptionsExceptPrefix[0] != '\0';
285    SmallVector<Metadata *, 8> MDs;
286    MDs.push_back(nullptr);
287  
288    bool Changed = false;
289    if (InheritAllAttrs || InheritSomeAttrs) {
290      for (const MDOperand &Existing : drop_begin(OrigLoopID->operands())) {
291        MDNode *Op = cast<MDNode>(Existing.get());
292  
293        auto InheritThisAttribute = [InheritSomeAttrs,
294                                     InheritOptionsExceptPrefix](MDNode *Op) {
295          if (!InheritSomeAttrs)
296            return false;
297  
298          // Skip malformatted attribute metadata nodes.
299          if (Op->getNumOperands() == 0)
300            return true;
301          Metadata *NameMD = Op->getOperand(0).get();
302          if (!isa<MDString>(NameMD))
303            return true;
304          StringRef AttrName = cast<MDString>(NameMD)->getString();
305  
306          // Do not inherit excluded attributes.
307          return !AttrName.startswith(InheritOptionsExceptPrefix);
308        };
309  
310        if (InheritThisAttribute(Op))
311          MDs.push_back(Op);
312        else
313          Changed = true;
314      }
315    } else {
316      // Modified if we dropped at least one attribute.
317      Changed = OrigLoopID->getNumOperands() > 1;
318    }
319  
320    bool HasAnyFollowup = false;
321    for (StringRef OptionName : FollowupOptions) {
322      MDNode *FollowupNode = findOptionMDForLoopID(OrigLoopID, OptionName);
323      if (!FollowupNode)
324        continue;
325  
326      HasAnyFollowup = true;
327      for (const MDOperand &Option : drop_begin(FollowupNode->operands())) {
328        MDs.push_back(Option.get());
329        Changed = true;
330      }
331    }
332  
333    // Attributes of the followup loop not specified explicity, so signal to the
334    // transformation pass to add suitable attributes.
335    if (!AlwaysNew && !HasAnyFollowup)
336      return None;
337  
338    // If no attributes were added or remove, the previous loop Id can be reused.
339    if (!AlwaysNew && !Changed)
340      return OrigLoopID;
341  
342    // No attributes is equivalent to having no !llvm.loop metadata at all.
343    if (MDs.size() == 1)
344      return nullptr;
345  
346    // Build the new loop ID.
347    MDTuple *FollowupLoopID = MDNode::get(OrigLoopID->getContext(), MDs);
348    FollowupLoopID->replaceOperandWith(0, FollowupLoopID);
349    return FollowupLoopID;
350  }
351  
352  bool llvm::hasDisableAllTransformsHint(const Loop *L) {
353    return getBooleanLoopAttribute(L, LLVMLoopDisableNonforced);
354  }
355  
356  bool llvm::hasDisableLICMTransformsHint(const Loop *L) {
357    return getBooleanLoopAttribute(L, LLVMLoopDisableLICM);
358  }
359  
360  TransformationMode llvm::hasUnrollTransformation(const Loop *L) {
361    if (getBooleanLoopAttribute(L, "llvm.loop.unroll.disable"))
362      return TM_SuppressedByUser;
363  
364    Optional<int> Count =
365        getOptionalIntLoopAttribute(L, "llvm.loop.unroll.count");
366    if (Count.hasValue())
367      return Count.getValue() == 1 ? TM_SuppressedByUser : TM_ForcedByUser;
368  
369    if (getBooleanLoopAttribute(L, "llvm.loop.unroll.enable"))
370      return TM_ForcedByUser;
371  
372    if (getBooleanLoopAttribute(L, "llvm.loop.unroll.full"))
373      return TM_ForcedByUser;
374  
375    if (hasDisableAllTransformsHint(L))
376      return TM_Disable;
377  
378    return TM_Unspecified;
379  }
380  
381  TransformationMode llvm::hasUnrollAndJamTransformation(const Loop *L) {
382    if (getBooleanLoopAttribute(L, "llvm.loop.unroll_and_jam.disable"))
383      return TM_SuppressedByUser;
384  
385    Optional<int> Count =
386        getOptionalIntLoopAttribute(L, "llvm.loop.unroll_and_jam.count");
387    if (Count.hasValue())
388      return Count.getValue() == 1 ? TM_SuppressedByUser : TM_ForcedByUser;
389  
390    if (getBooleanLoopAttribute(L, "llvm.loop.unroll_and_jam.enable"))
391      return TM_ForcedByUser;
392  
393    if (hasDisableAllTransformsHint(L))
394      return TM_Disable;
395  
396    return TM_Unspecified;
397  }
398  
399  TransformationMode llvm::hasVectorizeTransformation(const Loop *L) {
400    Optional<bool> Enable =
401        getOptionalBoolLoopAttribute(L, "llvm.loop.vectorize.enable");
402  
403    if (Enable == false)
404      return TM_SuppressedByUser;
405  
406    Optional<ElementCount> VectorizeWidth =
407        getOptionalElementCountLoopAttribute(L);
408    Optional<int> InterleaveCount =
409        getOptionalIntLoopAttribute(L, "llvm.loop.interleave.count");
410  
411    // 'Forcing' vector width and interleave count to one effectively disables
412    // this tranformation.
413    if (Enable == true && VectorizeWidth && VectorizeWidth->isScalar() &&
414        InterleaveCount == 1)
415      return TM_SuppressedByUser;
416  
417    if (getBooleanLoopAttribute(L, "llvm.loop.isvectorized"))
418      return TM_Disable;
419  
420    if (Enable == true)
421      return TM_ForcedByUser;
422  
423    if ((VectorizeWidth && VectorizeWidth->isScalar()) && InterleaveCount == 1)
424      return TM_Disable;
425  
426    if ((VectorizeWidth && VectorizeWidth->isVector()) || InterleaveCount > 1)
427      return TM_Enable;
428  
429    if (hasDisableAllTransformsHint(L))
430      return TM_Disable;
431  
432    return TM_Unspecified;
433  }
434  
435  TransformationMode llvm::hasDistributeTransformation(const Loop *L) {
436    if (getBooleanLoopAttribute(L, "llvm.loop.distribute.enable"))
437      return TM_ForcedByUser;
438  
439    if (hasDisableAllTransformsHint(L))
440      return TM_Disable;
441  
442    return TM_Unspecified;
443  }
444  
445  TransformationMode llvm::hasLICMVersioningTransformation(const Loop *L) {
446    if (getBooleanLoopAttribute(L, "llvm.loop.licm_versioning.disable"))
447      return TM_SuppressedByUser;
448  
449    if (hasDisableAllTransformsHint(L))
450      return TM_Disable;
451  
452    return TM_Unspecified;
453  }
454  
455  /// Does a BFS from a given node to all of its children inside a given loop.
456  /// The returned vector of nodes includes the starting point.
457  SmallVector<DomTreeNode *, 16>
458  llvm::collectChildrenInLoop(DomTreeNode *N, const Loop *CurLoop) {
459    SmallVector<DomTreeNode *, 16> Worklist;
460    auto AddRegionToWorklist = [&](DomTreeNode *DTN) {
461      // Only include subregions in the top level loop.
462      BasicBlock *BB = DTN->getBlock();
463      if (CurLoop->contains(BB))
464        Worklist.push_back(DTN);
465    };
466  
467    AddRegionToWorklist(N);
468  
469    for (size_t I = 0; I < Worklist.size(); I++) {
470      for (DomTreeNode *Child : Worklist[I]->children())
471        AddRegionToWorklist(Child);
472    }
473  
474    return Worklist;
475  }
476  
477  void llvm::deleteDeadLoop(Loop *L, DominatorTree *DT, ScalarEvolution *SE,
478                            LoopInfo *LI, MemorySSA *MSSA) {
479    assert((!DT || L->isLCSSAForm(*DT)) && "Expected LCSSA!");
480    auto *Preheader = L->getLoopPreheader();
481    assert(Preheader && "Preheader should exist!");
482  
483    std::unique_ptr<MemorySSAUpdater> MSSAU;
484    if (MSSA)
485      MSSAU = std::make_unique<MemorySSAUpdater>(MSSA);
486  
487    // Now that we know the removal is safe, remove the loop by changing the
488    // branch from the preheader to go to the single exit block.
489    //
490    // Because we're deleting a large chunk of code at once, the sequence in which
491    // we remove things is very important to avoid invalidation issues.
492  
493    // Tell ScalarEvolution that the loop is deleted. Do this before
494    // deleting the loop so that ScalarEvolution can look at the loop
495    // to determine what it needs to clean up.
496    if (SE)
497      SE->forgetLoop(L);
498  
499    auto *OldBr = dyn_cast<BranchInst>(Preheader->getTerminator());
500    assert(OldBr && "Preheader must end with a branch");
501    assert(OldBr->isUnconditional() && "Preheader must have a single successor");
502    // Connect the preheader to the exit block. Keep the old edge to the header
503    // around to perform the dominator tree update in two separate steps
504    // -- #1 insertion of the edge preheader -> exit and #2 deletion of the edge
505    // preheader -> header.
506    //
507    //
508    // 0.  Preheader          1.  Preheader           2.  Preheader
509    //        |                    |   |                   |
510    //        V                    |   V                   |
511    //      Header <--\            | Header <--\           | Header <--\
512    //       |  |     |            |  |  |     |           |  |  |     |
513    //       |  V     |            |  |  V     |           |  |  V     |
514    //       | Body --/            |  | Body --/           |  | Body --/
515    //       V                     V  V                    V  V
516    //      Exit                   Exit                    Exit
517    //
518    // By doing this is two separate steps we can perform the dominator tree
519    // update without using the batch update API.
520    //
521    // Even when the loop is never executed, we cannot remove the edge from the
522    // source block to the exit block. Consider the case where the unexecuted loop
523    // branches back to an outer loop. If we deleted the loop and removed the edge
524    // coming to this inner loop, this will break the outer loop structure (by
525    // deleting the backedge of the outer loop). If the outer loop is indeed a
526    // non-loop, it will be deleted in a future iteration of loop deletion pass.
527    IRBuilder<> Builder(OldBr);
528  
529    auto *ExitBlock = L->getUniqueExitBlock();
530    DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);
531    if (ExitBlock) {
532      assert(ExitBlock && "Should have a unique exit block!");
533      assert(L->hasDedicatedExits() && "Loop should have dedicated exits!");
534  
535      Builder.CreateCondBr(Builder.getFalse(), L->getHeader(), ExitBlock);
536      // Remove the old branch. The conditional branch becomes a new terminator.
537      OldBr->eraseFromParent();
538  
539      // Rewrite phis in the exit block to get their inputs from the Preheader
540      // instead of the exiting block.
541      for (PHINode &P : ExitBlock->phis()) {
542        // Set the zero'th element of Phi to be from the preheader and remove all
543        // other incoming values. Given the loop has dedicated exits, all other
544        // incoming values must be from the exiting blocks.
545        int PredIndex = 0;
546        P.setIncomingBlock(PredIndex, Preheader);
547        // Removes all incoming values from all other exiting blocks (including
548        // duplicate values from an exiting block).
549        // Nuke all entries except the zero'th entry which is the preheader entry.
550        // NOTE! We need to remove Incoming Values in the reverse order as done
551        // below, to keep the indices valid for deletion (removeIncomingValues
552        // updates getNumIncomingValues and shifts all values down into the
553        // operand being deleted).
554        for (unsigned i = 0, e = P.getNumIncomingValues() - 1; i != e; ++i)
555          P.removeIncomingValue(e - i, false);
556  
557        assert((P.getNumIncomingValues() == 1 &&
558                P.getIncomingBlock(PredIndex) == Preheader) &&
559               "Should have exactly one value and that's from the preheader!");
560      }
561  
562      if (DT) {
563        DTU.applyUpdates({{DominatorTree::Insert, Preheader, ExitBlock}});
564        if (MSSA) {
565          MSSAU->applyUpdates({{DominatorTree::Insert, Preheader, ExitBlock}},
566                              *DT);
567          if (VerifyMemorySSA)
568            MSSA->verifyMemorySSA();
569        }
570      }
571  
572      // Disconnect the loop body by branching directly to its exit.
573      Builder.SetInsertPoint(Preheader->getTerminator());
574      Builder.CreateBr(ExitBlock);
575      // Remove the old branch.
576      Preheader->getTerminator()->eraseFromParent();
577    } else {
578      assert(L->hasNoExitBlocks() &&
579             "Loop should have either zero or one exit blocks.");
580  
581      Builder.SetInsertPoint(OldBr);
582      Builder.CreateUnreachable();
583      Preheader->getTerminator()->eraseFromParent();
584    }
585  
586    if (DT) {
587      DTU.applyUpdates({{DominatorTree::Delete, Preheader, L->getHeader()}});
588      if (MSSA) {
589        MSSAU->applyUpdates({{DominatorTree::Delete, Preheader, L->getHeader()}},
590                            *DT);
591        SmallSetVector<BasicBlock *, 8> DeadBlockSet(L->block_begin(),
592                                                     L->block_end());
593        MSSAU->removeBlocks(DeadBlockSet);
594        if (VerifyMemorySSA)
595          MSSA->verifyMemorySSA();
596      }
597    }
598  
599    // Use a map to unique and a vector to guarantee deterministic ordering.
600    llvm::SmallDenseSet<std::pair<DIVariable *, DIExpression *>, 4> DeadDebugSet;
601    llvm::SmallVector<DbgVariableIntrinsic *, 4> DeadDebugInst;
602  
603    if (ExitBlock) {
604      // Given LCSSA form is satisfied, we should not have users of instructions
605      // within the dead loop outside of the loop. However, LCSSA doesn't take
606      // unreachable uses into account. We handle them here.
607      // We could do it after drop all references (in this case all users in the
608      // loop will be already eliminated and we have less work to do but according
609      // to API doc of User::dropAllReferences only valid operation after dropping
610      // references, is deletion. So let's substitute all usages of
611      // instruction from the loop with undef value of corresponding type first.
612      for (auto *Block : L->blocks())
613        for (Instruction &I : *Block) {
614          auto *Undef = UndefValue::get(I.getType());
615          for (Value::use_iterator UI = I.use_begin(), E = I.use_end();
616               UI != E;) {
617            Use &U = *UI;
618            ++UI;
619            if (auto *Usr = dyn_cast<Instruction>(U.getUser()))
620              if (L->contains(Usr->getParent()))
621                continue;
622            // If we have a DT then we can check that uses outside a loop only in
623            // unreachable block.
624            if (DT)
625              assert(!DT->isReachableFromEntry(U) &&
626                     "Unexpected user in reachable block");
627            U.set(Undef);
628          }
629          auto *DVI = dyn_cast<DbgVariableIntrinsic>(&I);
630          if (!DVI)
631            continue;
632          auto Key =
633              DeadDebugSet.find({DVI->getVariable(), DVI->getExpression()});
634          if (Key != DeadDebugSet.end())
635            continue;
636          DeadDebugSet.insert({DVI->getVariable(), DVI->getExpression()});
637          DeadDebugInst.push_back(DVI);
638        }
639  
640      // After the loop has been deleted all the values defined and modified
641      // inside the loop are going to be unavailable.
642      // Since debug values in the loop have been deleted, inserting an undef
643      // dbg.value truncates the range of any dbg.value before the loop where the
644      // loop used to be. This is particularly important for constant values.
645      DIBuilder DIB(*ExitBlock->getModule());
646      Instruction *InsertDbgValueBefore = ExitBlock->getFirstNonPHI();
647      assert(InsertDbgValueBefore &&
648             "There should be a non-PHI instruction in exit block, else these "
649             "instructions will have no parent.");
650      for (auto *DVI : DeadDebugInst)
651        DIB.insertDbgValueIntrinsic(UndefValue::get(Builder.getInt32Ty()),
652                                    DVI->getVariable(), DVI->getExpression(),
653                                    DVI->getDebugLoc(), InsertDbgValueBefore);
654    }
655  
656    // Remove the block from the reference counting scheme, so that we can
657    // delete it freely later.
658    for (auto *Block : L->blocks())
659      Block->dropAllReferences();
660  
661    if (MSSA && VerifyMemorySSA)
662      MSSA->verifyMemorySSA();
663  
664    if (LI) {
665      // Erase the instructions and the blocks without having to worry
666      // about ordering because we already dropped the references.
667      // NOTE: This iteration is safe because erasing the block does not remove
668      // its entry from the loop's block list.  We do that in the next section.
669      for (Loop::block_iterator LpI = L->block_begin(), LpE = L->block_end();
670           LpI != LpE; ++LpI)
671        (*LpI)->eraseFromParent();
672  
673      // Finally, the blocks from loopinfo.  This has to happen late because
674      // otherwise our loop iterators won't work.
675  
676      SmallPtrSet<BasicBlock *, 8> blocks;
677      blocks.insert(L->block_begin(), L->block_end());
678      for (BasicBlock *BB : blocks)
679        LI->removeBlock(BB);
680  
681      // The last step is to update LoopInfo now that we've eliminated this loop.
682      // Note: LoopInfo::erase remove the given loop and relink its subloops with
683      // its parent. While removeLoop/removeChildLoop remove the given loop but
684      // not relink its subloops, which is what we want.
685      if (Loop *ParentLoop = L->getParentLoop()) {
686        Loop::iterator I = find(*ParentLoop, L);
687        assert(I != ParentLoop->end() && "Couldn't find loop");
688        ParentLoop->removeChildLoop(I);
689      } else {
690        Loop::iterator I = find(*LI, L);
691        assert(I != LI->end() && "Couldn't find loop");
692        LI->removeLoop(I);
693      }
694      LI->destroy(L);
695    }
696  }
697  
698  static Loop *getOutermostLoop(Loop *L) {
699    while (Loop *Parent = L->getParentLoop())
700      L = Parent;
701    return L;
702  }
703  
704  void llvm::breakLoopBackedge(Loop *L, DominatorTree &DT, ScalarEvolution &SE,
705                               LoopInfo &LI, MemorySSA *MSSA) {
706    auto *Latch = L->getLoopLatch();
707    assert(Latch && "multiple latches not yet supported");
708    auto *Header = L->getHeader();
709    Loop *OutermostLoop = getOutermostLoop(L);
710  
711    SE.forgetLoop(L);
712  
713    // Note: By splitting the backedge, and then explicitly making it unreachable
714    // we gracefully handle corner cases such as non-bottom tested loops and the
715    // like.  We also have the benefit of being able to reuse existing well tested
716    // code.  It might be worth special casing the common bottom tested case at
717    // some point to avoid code churn.
718  
719    std::unique_ptr<MemorySSAUpdater> MSSAU;
720    if (MSSA)
721      MSSAU = std::make_unique<MemorySSAUpdater>(MSSA);
722  
723    auto *BackedgeBB = SplitEdge(Latch, Header, &DT, &LI, MSSAU.get());
724  
725    DomTreeUpdater DTU(&DT, DomTreeUpdater::UpdateStrategy::Eager);
726    (void)changeToUnreachable(BackedgeBB->getTerminator(),
727                              /*PreserveLCSSA*/ true, &DTU, MSSAU.get());
728  
729    // Erase (and destroy) this loop instance.  Handles relinking sub-loops
730    // and blocks within the loop as needed.
731    LI.erase(L);
732  
733    // If the loop we broke had a parent, then changeToUnreachable might have
734    // caused a block to be removed from the parent loop (see loop_nest_lcssa
735    // test case in zero-btc.ll for an example), thus changing the parent's
736    // exit blocks.  If that happened, we need to rebuild LCSSA on the outermost
737    // loop which might have a had a block removed.
738    if (OutermostLoop != L)
739      formLCSSARecursively(*OutermostLoop, DT, &LI, &SE);
740  }
741  
742  
743  /// Checks if \p L has single exit through latch block except possibly
744  /// "deoptimizing" exits. Returns branch instruction terminating the loop
745  /// latch if above check is successful, nullptr otherwise.
746  static BranchInst *getExpectedExitLoopLatchBranch(Loop *L) {
747    BasicBlock *Latch = L->getLoopLatch();
748    if (!Latch)
749      return nullptr;
750  
751    BranchInst *LatchBR = dyn_cast<BranchInst>(Latch->getTerminator());
752    if (!LatchBR || LatchBR->getNumSuccessors() != 2 || !L->isLoopExiting(Latch))
753      return nullptr;
754  
755    assert((LatchBR->getSuccessor(0) == L->getHeader() ||
756            LatchBR->getSuccessor(1) == L->getHeader()) &&
757           "At least one edge out of the latch must go to the header");
758  
759    SmallVector<BasicBlock *, 4> ExitBlocks;
760    L->getUniqueNonLatchExitBlocks(ExitBlocks);
761    if (any_of(ExitBlocks, [](const BasicBlock *EB) {
762          return !EB->getTerminatingDeoptimizeCall();
763        }))
764      return nullptr;
765  
766    return LatchBR;
767  }
768  
769  Optional<unsigned>
770  llvm::getLoopEstimatedTripCount(Loop *L,
771                                  unsigned *EstimatedLoopInvocationWeight) {
772    // Support loops with an exiting latch and other existing exists only
773    // deoptimize.
774    BranchInst *LatchBranch = getExpectedExitLoopLatchBranch(L);
775    if (!LatchBranch)
776      return None;
777  
778    // To estimate the number of times the loop body was executed, we want to
779    // know the number of times the backedge was taken, vs. the number of times
780    // we exited the loop.
781    uint64_t BackedgeTakenWeight, LatchExitWeight;
782    if (!LatchBranch->extractProfMetadata(BackedgeTakenWeight, LatchExitWeight))
783      return None;
784  
785    if (LatchBranch->getSuccessor(0) != L->getHeader())
786      std::swap(BackedgeTakenWeight, LatchExitWeight);
787  
788    if (!LatchExitWeight)
789      return None;
790  
791    if (EstimatedLoopInvocationWeight)
792      *EstimatedLoopInvocationWeight = LatchExitWeight;
793  
794    // Estimated backedge taken count is a ratio of the backedge taken weight by
795    // the weight of the edge exiting the loop, rounded to nearest.
796    uint64_t BackedgeTakenCount =
797        llvm::divideNearest(BackedgeTakenWeight, LatchExitWeight);
798    // Estimated trip count is one plus estimated backedge taken count.
799    return BackedgeTakenCount + 1;
800  }
801  
802  bool llvm::setLoopEstimatedTripCount(Loop *L, unsigned EstimatedTripCount,
803                                       unsigned EstimatedloopInvocationWeight) {
804    // Support loops with an exiting latch and other existing exists only
805    // deoptimize.
806    BranchInst *LatchBranch = getExpectedExitLoopLatchBranch(L);
807    if (!LatchBranch)
808      return false;
809  
810    // Calculate taken and exit weights.
811    unsigned LatchExitWeight = 0;
812    unsigned BackedgeTakenWeight = 0;
813  
814    if (EstimatedTripCount > 0) {
815      LatchExitWeight = EstimatedloopInvocationWeight;
816      BackedgeTakenWeight = (EstimatedTripCount - 1) * LatchExitWeight;
817    }
818  
819    // Make a swap if back edge is taken when condition is "false".
820    if (LatchBranch->getSuccessor(0) != L->getHeader())
821      std::swap(BackedgeTakenWeight, LatchExitWeight);
822  
823    MDBuilder MDB(LatchBranch->getContext());
824  
825    // Set/Update profile metadata.
826    LatchBranch->setMetadata(
827        LLVMContext::MD_prof,
828        MDB.createBranchWeights(BackedgeTakenWeight, LatchExitWeight));
829  
830    return true;
831  }
832  
833  bool llvm::hasIterationCountInvariantInParent(Loop *InnerLoop,
834                                                ScalarEvolution &SE) {
835    Loop *OuterL = InnerLoop->getParentLoop();
836    if (!OuterL)
837      return true;
838  
839    // Get the backedge taken count for the inner loop
840    BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
841    const SCEV *InnerLoopBECountSC = SE.getExitCount(InnerLoop, InnerLoopLatch);
842    if (isa<SCEVCouldNotCompute>(InnerLoopBECountSC) ||
843        !InnerLoopBECountSC->getType()->isIntegerTy())
844      return false;
845  
846    // Get whether count is invariant to the outer loop
847    ScalarEvolution::LoopDisposition LD =
848        SE.getLoopDisposition(InnerLoopBECountSC, OuterL);
849    if (LD != ScalarEvolution::LoopInvariant)
850      return false;
851  
852    return true;
853  }
854  
855  Value *llvm::createMinMaxOp(IRBuilderBase &Builder, RecurKind RK, Value *Left,
856                              Value *Right) {
857    CmpInst::Predicate Pred;
858    switch (RK) {
859    default:
860      llvm_unreachable("Unknown min/max recurrence kind");
861    case RecurKind::UMin:
862      Pred = CmpInst::ICMP_ULT;
863      break;
864    case RecurKind::UMax:
865      Pred = CmpInst::ICMP_UGT;
866      break;
867    case RecurKind::SMin:
868      Pred = CmpInst::ICMP_SLT;
869      break;
870    case RecurKind::SMax:
871      Pred = CmpInst::ICMP_SGT;
872      break;
873    case RecurKind::FMin:
874      Pred = CmpInst::FCMP_OLT;
875      break;
876    case RecurKind::FMax:
877      Pred = CmpInst::FCMP_OGT;
878      break;
879    }
880  
881    Value *Cmp = Builder.CreateCmp(Pred, Left, Right, "rdx.minmax.cmp");
882    Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
883    return Select;
884  }
885  
886  // Helper to generate an ordered reduction.
887  Value *llvm::getOrderedReduction(IRBuilderBase &Builder, Value *Acc, Value *Src,
888                                   unsigned Op, RecurKind RdxKind,
889                                   ArrayRef<Value *> RedOps) {
890    unsigned VF = cast<FixedVectorType>(Src->getType())->getNumElements();
891  
892    // Extract and apply reduction ops in ascending order:
893    // e.g. ((((Acc + Scl[0]) + Scl[1]) + Scl[2]) + ) ... + Scl[VF-1]
894    Value *Result = Acc;
895    for (unsigned ExtractIdx = 0; ExtractIdx != VF; ++ExtractIdx) {
896      Value *Ext =
897          Builder.CreateExtractElement(Src, Builder.getInt32(ExtractIdx));
898  
899      if (Op != Instruction::ICmp && Op != Instruction::FCmp) {
900        Result = Builder.CreateBinOp((Instruction::BinaryOps)Op, Result, Ext,
901                                     "bin.rdx");
902      } else {
903        assert(RecurrenceDescriptor::isMinMaxRecurrenceKind(RdxKind) &&
904               "Invalid min/max");
905        Result = createMinMaxOp(Builder, RdxKind, Result, Ext);
906      }
907  
908      if (!RedOps.empty())
909        propagateIRFlags(Result, RedOps);
910    }
911  
912    return Result;
913  }
914  
915  // Helper to generate a log2 shuffle reduction.
916  Value *llvm::getShuffleReduction(IRBuilderBase &Builder, Value *Src,
917                                   unsigned Op, RecurKind RdxKind,
918                                   ArrayRef<Value *> RedOps) {
919    unsigned VF = cast<FixedVectorType>(Src->getType())->getNumElements();
920    // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
921    // and vector ops, reducing the set of values being computed by half each
922    // round.
923    assert(isPowerOf2_32(VF) &&
924           "Reduction emission only supported for pow2 vectors!");
925    Value *TmpVec = Src;
926    SmallVector<int, 32> ShuffleMask(VF);
927    for (unsigned i = VF; i != 1; i >>= 1) {
928      // Move the upper half of the vector to the lower half.
929      for (unsigned j = 0; j != i / 2; ++j)
930        ShuffleMask[j] = i / 2 + j;
931  
932      // Fill the rest of the mask with undef.
933      std::fill(&ShuffleMask[i / 2], ShuffleMask.end(), -1);
934  
935      Value *Shuf = Builder.CreateShuffleVector(TmpVec, ShuffleMask, "rdx.shuf");
936  
937      if (Op != Instruction::ICmp && Op != Instruction::FCmp) {
938        // The builder propagates its fast-math-flags setting.
939        TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
940                                     "bin.rdx");
941      } else {
942        assert(RecurrenceDescriptor::isMinMaxRecurrenceKind(RdxKind) &&
943               "Invalid min/max");
944        TmpVec = createMinMaxOp(Builder, RdxKind, TmpVec, Shuf);
945      }
946      if (!RedOps.empty())
947        propagateIRFlags(TmpVec, RedOps);
948  
949      // We may compute the reassociated scalar ops in a way that does not
950      // preserve nsw/nuw etc. Conservatively, drop those flags.
951      if (auto *ReductionInst = dyn_cast<Instruction>(TmpVec))
952        ReductionInst->dropPoisonGeneratingFlags();
953    }
954    // The result is in the first element of the vector.
955    return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
956  }
957  
958  Value *llvm::createSimpleTargetReduction(IRBuilderBase &Builder,
959                                           const TargetTransformInfo *TTI,
960                                           Value *Src, RecurKind RdxKind,
961                                           ArrayRef<Value *> RedOps) {
962    TargetTransformInfo::ReductionFlags RdxFlags;
963    RdxFlags.IsMaxOp = RdxKind == RecurKind::SMax || RdxKind == RecurKind::UMax ||
964                       RdxKind == RecurKind::FMax;
965    RdxFlags.IsSigned = RdxKind == RecurKind::SMax || RdxKind == RecurKind::SMin;
966  
967    auto *SrcVecEltTy = cast<VectorType>(Src->getType())->getElementType();
968    switch (RdxKind) {
969    case RecurKind::Add:
970      return Builder.CreateAddReduce(Src);
971    case RecurKind::Mul:
972      return Builder.CreateMulReduce(Src);
973    case RecurKind::And:
974      return Builder.CreateAndReduce(Src);
975    case RecurKind::Or:
976      return Builder.CreateOrReduce(Src);
977    case RecurKind::Xor:
978      return Builder.CreateXorReduce(Src);
979    case RecurKind::FAdd:
980      return Builder.CreateFAddReduce(ConstantFP::getNegativeZero(SrcVecEltTy),
981                                      Src);
982    case RecurKind::FMul:
983      return Builder.CreateFMulReduce(ConstantFP::get(SrcVecEltTy, 1.0), Src);
984    case RecurKind::SMax:
985      return Builder.CreateIntMaxReduce(Src, true);
986    case RecurKind::SMin:
987      return Builder.CreateIntMinReduce(Src, true);
988    case RecurKind::UMax:
989      return Builder.CreateIntMaxReduce(Src, false);
990    case RecurKind::UMin:
991      return Builder.CreateIntMinReduce(Src, false);
992    case RecurKind::FMax:
993      return Builder.CreateFPMaxReduce(Src);
994    case RecurKind::FMin:
995      return Builder.CreateFPMinReduce(Src);
996    default:
997      llvm_unreachable("Unhandled opcode");
998    }
999  }
1000  
1001  Value *llvm::createTargetReduction(IRBuilderBase &B,
1002                                     const TargetTransformInfo *TTI,
1003                                     const RecurrenceDescriptor &Desc,
1004                                     Value *Src) {
1005    // TODO: Support in-order reductions based on the recurrence descriptor.
1006    // All ops in the reduction inherit fast-math-flags from the recurrence
1007    // descriptor.
1008    IRBuilderBase::FastMathFlagGuard FMFGuard(B);
1009    B.setFastMathFlags(Desc.getFastMathFlags());
1010    return createSimpleTargetReduction(B, TTI, Src, Desc.getRecurrenceKind());
1011  }
1012  
1013  Value *llvm::createOrderedReduction(IRBuilderBase &B,
1014                                      const RecurrenceDescriptor &Desc,
1015                                      Value *Src, Value *Start) {
1016    assert(Desc.getRecurrenceKind() == RecurKind::FAdd &&
1017           "Unexpected reduction kind");
1018    assert(Src->getType()->isVectorTy() && "Expected a vector type");
1019    assert(!Start->getType()->isVectorTy() && "Expected a scalar type");
1020  
1021    return B.CreateFAddReduce(Start, Src);
1022  }
1023  
1024  void llvm::propagateIRFlags(Value *I, ArrayRef<Value *> VL, Value *OpValue) {
1025    auto *VecOp = dyn_cast<Instruction>(I);
1026    if (!VecOp)
1027      return;
1028    auto *Intersection = (OpValue == nullptr) ? dyn_cast<Instruction>(VL[0])
1029                                              : dyn_cast<Instruction>(OpValue);
1030    if (!Intersection)
1031      return;
1032    const unsigned Opcode = Intersection->getOpcode();
1033    VecOp->copyIRFlags(Intersection);
1034    for (auto *V : VL) {
1035      auto *Instr = dyn_cast<Instruction>(V);
1036      if (!Instr)
1037        continue;
1038      if (OpValue == nullptr || Opcode == Instr->getOpcode())
1039        VecOp->andIRFlags(V);
1040    }
1041  }
1042  
1043  bool llvm::isKnownNegativeInLoop(const SCEV *S, const Loop *L,
1044                                   ScalarEvolution &SE) {
1045    const SCEV *Zero = SE.getZero(S->getType());
1046    return SE.isAvailableAtLoopEntry(S, L) &&
1047           SE.isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SLT, S, Zero);
1048  }
1049  
1050  bool llvm::isKnownNonNegativeInLoop(const SCEV *S, const Loop *L,
1051                                      ScalarEvolution &SE) {
1052    const SCEV *Zero = SE.getZero(S->getType());
1053    return SE.isAvailableAtLoopEntry(S, L) &&
1054           SE.isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SGE, S, Zero);
1055  }
1056  
1057  bool llvm::cannotBeMinInLoop(const SCEV *S, const Loop *L, ScalarEvolution &SE,
1058                               bool Signed) {
1059    unsigned BitWidth = cast<IntegerType>(S->getType())->getBitWidth();
1060    APInt Min = Signed ? APInt::getSignedMinValue(BitWidth) :
1061      APInt::getMinValue(BitWidth);
1062    auto Predicate = Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1063    return SE.isAvailableAtLoopEntry(S, L) &&
1064           SE.isLoopEntryGuardedByCond(L, Predicate, S,
1065                                       SE.getConstant(Min));
1066  }
1067  
1068  bool llvm::cannotBeMaxInLoop(const SCEV *S, const Loop *L, ScalarEvolution &SE,
1069                               bool Signed) {
1070    unsigned BitWidth = cast<IntegerType>(S->getType())->getBitWidth();
1071    APInt Max = Signed ? APInt::getSignedMaxValue(BitWidth) :
1072      APInt::getMaxValue(BitWidth);
1073    auto Predicate = Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1074    return SE.isAvailableAtLoopEntry(S, L) &&
1075           SE.isLoopEntryGuardedByCond(L, Predicate, S,
1076                                       SE.getConstant(Max));
1077  }
1078  
1079  //===----------------------------------------------------------------------===//
1080  // rewriteLoopExitValues - Optimize IV users outside the loop.
1081  // As a side effect, reduces the amount of IV processing within the loop.
1082  //===----------------------------------------------------------------------===//
1083  
1084  // Return true if the SCEV expansion generated by the rewriter can replace the
1085  // original value. SCEV guarantees that it produces the same value, but the way
1086  // it is produced may be illegal IR.  Ideally, this function will only be
1087  // called for verification.
1088  static bool isValidRewrite(ScalarEvolution *SE, Value *FromVal, Value *ToVal) {
1089    // If an SCEV expression subsumed multiple pointers, its expansion could
1090    // reassociate the GEP changing the base pointer. This is illegal because the
1091    // final address produced by a GEP chain must be inbounds relative to its
1092    // underlying object. Otherwise basic alias analysis, among other things,
1093    // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
1094    // producing an expression involving multiple pointers. Until then, we must
1095    // bail out here.
1096    //
1097    // Retrieve the pointer operand of the GEP. Don't use getUnderlyingObject
1098    // because it understands lcssa phis while SCEV does not.
1099    Value *FromPtr = FromVal;
1100    Value *ToPtr = ToVal;
1101    if (auto *GEP = dyn_cast<GEPOperator>(FromVal))
1102      FromPtr = GEP->getPointerOperand();
1103  
1104    if (auto *GEP = dyn_cast<GEPOperator>(ToVal))
1105      ToPtr = GEP->getPointerOperand();
1106  
1107    if (FromPtr != FromVal || ToPtr != ToVal) {
1108      // Quickly check the common case
1109      if (FromPtr == ToPtr)
1110        return true;
1111  
1112      // SCEV may have rewritten an expression that produces the GEP's pointer
1113      // operand. That's ok as long as the pointer operand has the same base
1114      // pointer. Unlike getUnderlyingObject(), getPointerBase() will find the
1115      // base of a recurrence. This handles the case in which SCEV expansion
1116      // converts a pointer type recurrence into a nonrecurrent pointer base
1117      // indexed by an integer recurrence.
1118  
1119      // If the GEP base pointer is a vector of pointers, abort.
1120      if (!FromPtr->getType()->isPointerTy() || !ToPtr->getType()->isPointerTy())
1121        return false;
1122  
1123      const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
1124      const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
1125      if (FromBase == ToBase)
1126        return true;
1127  
1128      LLVM_DEBUG(dbgs() << "rewriteLoopExitValues: GEP rewrite bail out "
1129                        << *FromBase << " != " << *ToBase << "\n");
1130  
1131      return false;
1132    }
1133    return true;
1134  }
1135  
1136  static bool hasHardUserWithinLoop(const Loop *L, const Instruction *I) {
1137    SmallPtrSet<const Instruction *, 8> Visited;
1138    SmallVector<const Instruction *, 8> WorkList;
1139    Visited.insert(I);
1140    WorkList.push_back(I);
1141    while (!WorkList.empty()) {
1142      const Instruction *Curr = WorkList.pop_back_val();
1143      // This use is outside the loop, nothing to do.
1144      if (!L->contains(Curr))
1145        continue;
1146      // Do we assume it is a "hard" use which will not be eliminated easily?
1147      if (Curr->mayHaveSideEffects())
1148        return true;
1149      // Otherwise, add all its users to worklist.
1150      for (auto U : Curr->users()) {
1151        auto *UI = cast<Instruction>(U);
1152        if (Visited.insert(UI).second)
1153          WorkList.push_back(UI);
1154      }
1155    }
1156    return false;
1157  }
1158  
1159  // Collect information about PHI nodes which can be transformed in
1160  // rewriteLoopExitValues.
1161  struct RewritePhi {
1162    PHINode *PN;               // For which PHI node is this replacement?
1163    unsigned Ith;              // For which incoming value?
1164    const SCEV *ExpansionSCEV; // The SCEV of the incoming value we are rewriting.
1165    Instruction *ExpansionPoint; // Where we'd like to expand that SCEV?
1166    bool HighCost;               // Is this expansion a high-cost?
1167  
1168    Value *Expansion = nullptr;
1169    bool ValidRewrite = false;
1170  
1171    RewritePhi(PHINode *P, unsigned I, const SCEV *Val, Instruction *ExpansionPt,
1172               bool H)
1173        : PN(P), Ith(I), ExpansionSCEV(Val), ExpansionPoint(ExpansionPt),
1174          HighCost(H) {}
1175  };
1176  
1177  // Check whether it is possible to delete the loop after rewriting exit
1178  // value. If it is possible, ignore ReplaceExitValue and do rewriting
1179  // aggressively.
1180  static bool canLoopBeDeleted(Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet) {
1181    BasicBlock *Preheader = L->getLoopPreheader();
1182    // If there is no preheader, the loop will not be deleted.
1183    if (!Preheader)
1184      return false;
1185  
1186    // In LoopDeletion pass Loop can be deleted when ExitingBlocks.size() > 1.
1187    // We obviate multiple ExitingBlocks case for simplicity.
1188    // TODO: If we see testcase with multiple ExitingBlocks can be deleted
1189    // after exit value rewriting, we can enhance the logic here.
1190    SmallVector<BasicBlock *, 4> ExitingBlocks;
1191    L->getExitingBlocks(ExitingBlocks);
1192    SmallVector<BasicBlock *, 8> ExitBlocks;
1193    L->getUniqueExitBlocks(ExitBlocks);
1194    if (ExitBlocks.size() != 1 || ExitingBlocks.size() != 1)
1195      return false;
1196  
1197    BasicBlock *ExitBlock = ExitBlocks[0];
1198    BasicBlock::iterator BI = ExitBlock->begin();
1199    while (PHINode *P = dyn_cast<PHINode>(BI)) {
1200      Value *Incoming = P->getIncomingValueForBlock(ExitingBlocks[0]);
1201  
1202      // If the Incoming value of P is found in RewritePhiSet, we know it
1203      // could be rewritten to use a loop invariant value in transformation
1204      // phase later. Skip it in the loop invariant check below.
1205      bool found = false;
1206      for (const RewritePhi &Phi : RewritePhiSet) {
1207        if (!Phi.ValidRewrite)
1208          continue;
1209        unsigned i = Phi.Ith;
1210        if (Phi.PN == P && (Phi.PN)->getIncomingValue(i) == Incoming) {
1211          found = true;
1212          break;
1213        }
1214      }
1215  
1216      Instruction *I;
1217      if (!found && (I = dyn_cast<Instruction>(Incoming)))
1218        if (!L->hasLoopInvariantOperands(I))
1219          return false;
1220  
1221      ++BI;
1222    }
1223  
1224    for (auto *BB : L->blocks())
1225      if (llvm::any_of(*BB, [](Instruction &I) {
1226            return I.mayHaveSideEffects();
1227          }))
1228        return false;
1229  
1230    return true;
1231  }
1232  
1233  int llvm::rewriteLoopExitValues(Loop *L, LoopInfo *LI, TargetLibraryInfo *TLI,
1234                                  ScalarEvolution *SE,
1235                                  const TargetTransformInfo *TTI,
1236                                  SCEVExpander &Rewriter, DominatorTree *DT,
1237                                  ReplaceExitVal ReplaceExitValue,
1238                                  SmallVector<WeakTrackingVH, 16> &DeadInsts) {
1239    // Check a pre-condition.
1240    assert(L->isRecursivelyLCSSAForm(*DT, *LI) &&
1241           "Indvars did not preserve LCSSA!");
1242  
1243    SmallVector<BasicBlock*, 8> ExitBlocks;
1244    L->getUniqueExitBlocks(ExitBlocks);
1245  
1246    SmallVector<RewritePhi, 8> RewritePhiSet;
1247    // Find all values that are computed inside the loop, but used outside of it.
1248    // Because of LCSSA, these values will only occur in LCSSA PHI Nodes.  Scan
1249    // the exit blocks of the loop to find them.
1250    for (BasicBlock *ExitBB : ExitBlocks) {
1251      // If there are no PHI nodes in this exit block, then no values defined
1252      // inside the loop are used on this path, skip it.
1253      PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
1254      if (!PN) continue;
1255  
1256      unsigned NumPreds = PN->getNumIncomingValues();
1257  
1258      // Iterate over all of the PHI nodes.
1259      BasicBlock::iterator BBI = ExitBB->begin();
1260      while ((PN = dyn_cast<PHINode>(BBI++))) {
1261        if (PN->use_empty())
1262          continue; // dead use, don't replace it
1263  
1264        if (!SE->isSCEVable(PN->getType()))
1265          continue;
1266  
1267        // It's necessary to tell ScalarEvolution about this explicitly so that
1268        // it can walk the def-use list and forget all SCEVs, as it may not be
1269        // watching the PHI itself. Once the new exit value is in place, there
1270        // may not be a def-use connection between the loop and every instruction
1271        // which got a SCEVAddRecExpr for that loop.
1272        SE->forgetValue(PN);
1273  
1274        // Iterate over all of the values in all the PHI nodes.
1275        for (unsigned i = 0; i != NumPreds; ++i) {
1276          // If the value being merged in is not integer or is not defined
1277          // in the loop, skip it.
1278          Value *InVal = PN->getIncomingValue(i);
1279          if (!isa<Instruction>(InVal))
1280            continue;
1281  
1282          // If this pred is for a subloop, not L itself, skip it.
1283          if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
1284            continue; // The Block is in a subloop, skip it.
1285  
1286          // Check that InVal is defined in the loop.
1287          Instruction *Inst = cast<Instruction>(InVal);
1288          if (!L->contains(Inst))
1289            continue;
1290  
1291          // Okay, this instruction has a user outside of the current loop
1292          // and varies predictably *inside* the loop.  Evaluate the value it
1293          // contains when the loop exits, if possible.  We prefer to start with
1294          // expressions which are true for all exits (so as to maximize
1295          // expression reuse by the SCEVExpander), but resort to per-exit
1296          // evaluation if that fails.
1297          const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
1298          if (isa<SCEVCouldNotCompute>(ExitValue) ||
1299              !SE->isLoopInvariant(ExitValue, L) ||
1300              !isSafeToExpand(ExitValue, *SE)) {
1301            // TODO: This should probably be sunk into SCEV in some way; maybe a
1302            // getSCEVForExit(SCEV*, L, ExitingBB)?  It can be generalized for
1303            // most SCEV expressions and other recurrence types (e.g. shift
1304            // recurrences).  Is there existing code we can reuse?
1305            const SCEV *ExitCount = SE->getExitCount(L, PN->getIncomingBlock(i));
1306            if (isa<SCEVCouldNotCompute>(ExitCount))
1307              continue;
1308            if (auto *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Inst)))
1309              if (AddRec->getLoop() == L)
1310                ExitValue = AddRec->evaluateAtIteration(ExitCount, *SE);
1311            if (isa<SCEVCouldNotCompute>(ExitValue) ||
1312                !SE->isLoopInvariant(ExitValue, L) ||
1313                !isSafeToExpand(ExitValue, *SE))
1314              continue;
1315          }
1316  
1317          // Computing the value outside of the loop brings no benefit if it is
1318          // definitely used inside the loop in a way which can not be optimized
1319          // away. Avoid doing so unless we know we have a value which computes
1320          // the ExitValue already. TODO: This should be merged into SCEV
1321          // expander to leverage its knowledge of existing expressions.
1322          if (ReplaceExitValue != AlwaysRepl && !isa<SCEVConstant>(ExitValue) &&
1323              !isa<SCEVUnknown>(ExitValue) && hasHardUserWithinLoop(L, Inst))
1324            continue;
1325  
1326          // Check if expansions of this SCEV would count as being high cost.
1327          bool HighCost = Rewriter.isHighCostExpansion(
1328              ExitValue, L, SCEVCheapExpansionBudget, TTI, Inst);
1329  
1330          // Note that we must not perform expansions until after
1331          // we query *all* the costs, because if we perform temporary expansion
1332          // inbetween, one that we might not intend to keep, said expansion
1333          // *may* affect cost calculation of the the next SCEV's we'll query,
1334          // and next SCEV may errneously get smaller cost.
1335  
1336          // Collect all the candidate PHINodes to be rewritten.
1337          RewritePhiSet.emplace_back(PN, i, ExitValue, Inst, HighCost);
1338        }
1339      }
1340    }
1341  
1342    // Now that we've done preliminary filtering and billed all the SCEV's,
1343    // we can perform the last sanity check - the expansion must be valid.
1344    for (RewritePhi &Phi : RewritePhiSet) {
1345      Phi.Expansion = Rewriter.expandCodeFor(Phi.ExpansionSCEV, Phi.PN->getType(),
1346                                             Phi.ExpansionPoint);
1347  
1348      LLVM_DEBUG(dbgs() << "rewriteLoopExitValues: AfterLoopVal = "
1349                        << *(Phi.Expansion) << '\n'
1350                        << "  LoopVal = " << *(Phi.ExpansionPoint) << "\n");
1351  
1352      // FIXME: isValidRewrite() is a hack. it should be an assert, eventually.
1353      Phi.ValidRewrite = isValidRewrite(SE, Phi.ExpansionPoint, Phi.Expansion);
1354      if (!Phi.ValidRewrite) {
1355        DeadInsts.push_back(Phi.Expansion);
1356        continue;
1357      }
1358  
1359  #ifndef NDEBUG
1360      // If we reuse an instruction from a loop which is neither L nor one of
1361      // its containing loops, we end up breaking LCSSA form for this loop by
1362      // creating a new use of its instruction.
1363      if (auto *ExitInsn = dyn_cast<Instruction>(Phi.Expansion))
1364        if (auto *EVL = LI->getLoopFor(ExitInsn->getParent()))
1365          if (EVL != L)
1366            assert(EVL->contains(L) && "LCSSA breach detected!");
1367  #endif
1368    }
1369  
1370    // TODO: after isValidRewrite() is an assertion, evaluate whether
1371    // it is beneficial to change how we calculate high-cost:
1372    // if we have SCEV 'A' which we know we will expand, should we calculate
1373    // the cost of other SCEV's after expanding SCEV 'A',
1374    // thus potentially giving cost bonus to those other SCEV's?
1375  
1376    bool LoopCanBeDel = canLoopBeDeleted(L, RewritePhiSet);
1377    int NumReplaced = 0;
1378  
1379    // Transformation.
1380    for (const RewritePhi &Phi : RewritePhiSet) {
1381      if (!Phi.ValidRewrite)
1382        continue;
1383  
1384      PHINode *PN = Phi.PN;
1385      Value *ExitVal = Phi.Expansion;
1386  
1387      // Only do the rewrite when the ExitValue can be expanded cheaply.
1388      // If LoopCanBeDel is true, rewrite exit value aggressively.
1389      if (ReplaceExitValue == OnlyCheapRepl && !LoopCanBeDel && Phi.HighCost) {
1390        DeadInsts.push_back(ExitVal);
1391        continue;
1392      }
1393  
1394      NumReplaced++;
1395      Instruction *Inst = cast<Instruction>(PN->getIncomingValue(Phi.Ith));
1396      PN->setIncomingValue(Phi.Ith, ExitVal);
1397  
1398      // If this instruction is dead now, delete it. Don't do it now to avoid
1399      // invalidating iterators.
1400      if (isInstructionTriviallyDead(Inst, TLI))
1401        DeadInsts.push_back(Inst);
1402  
1403      // Replace PN with ExitVal if that is legal and does not break LCSSA.
1404      if (PN->getNumIncomingValues() == 1 &&
1405          LI->replacementPreservesLCSSAForm(PN, ExitVal)) {
1406        PN->replaceAllUsesWith(ExitVal);
1407        PN->eraseFromParent();
1408      }
1409    }
1410  
1411    // The insertion point instruction may have been deleted; clear it out
1412    // so that the rewriter doesn't trip over it later.
1413    Rewriter.clearInsertPoint();
1414    return NumReplaced;
1415  }
1416  
1417  /// Set weights for \p UnrolledLoop and \p RemainderLoop based on weights for
1418  /// \p OrigLoop.
1419  void llvm::setProfileInfoAfterUnrolling(Loop *OrigLoop, Loop *UnrolledLoop,
1420                                          Loop *RemainderLoop, uint64_t UF) {
1421    assert(UF > 0 && "Zero unrolled factor is not supported");
1422    assert(UnrolledLoop != RemainderLoop &&
1423           "Unrolled and Remainder loops are expected to distinct");
1424  
1425    // Get number of iterations in the original scalar loop.
1426    unsigned OrigLoopInvocationWeight = 0;
1427    Optional<unsigned> OrigAverageTripCount =
1428        getLoopEstimatedTripCount(OrigLoop, &OrigLoopInvocationWeight);
1429    if (!OrigAverageTripCount)
1430      return;
1431  
1432    // Calculate number of iterations in unrolled loop.
1433    unsigned UnrolledAverageTripCount = *OrigAverageTripCount / UF;
1434    // Calculate number of iterations for remainder loop.
1435    unsigned RemainderAverageTripCount = *OrigAverageTripCount % UF;
1436  
1437    setLoopEstimatedTripCount(UnrolledLoop, UnrolledAverageTripCount,
1438                              OrigLoopInvocationWeight);
1439    setLoopEstimatedTripCount(RemainderLoop, RemainderAverageTripCount,
1440                              OrigLoopInvocationWeight);
1441  }
1442  
1443  /// Utility that implements appending of loops onto a worklist.
1444  /// Loops are added in preorder (analogous for reverse postorder for trees),
1445  /// and the worklist is processed LIFO.
1446  template <typename RangeT>
1447  void llvm::appendReversedLoopsToWorklist(
1448      RangeT &&Loops, SmallPriorityWorklist<Loop *, 4> &Worklist) {
1449    // We use an internal worklist to build up the preorder traversal without
1450    // recursion.
1451    SmallVector<Loop *, 4> PreOrderLoops, PreOrderWorklist;
1452  
1453    // We walk the initial sequence of loops in reverse because we generally want
1454    // to visit defs before uses and the worklist is LIFO.
1455    for (Loop *RootL : Loops) {
1456      assert(PreOrderLoops.empty() && "Must start with an empty preorder walk.");
1457      assert(PreOrderWorklist.empty() &&
1458             "Must start with an empty preorder walk worklist.");
1459      PreOrderWorklist.push_back(RootL);
1460      do {
1461        Loop *L = PreOrderWorklist.pop_back_val();
1462        PreOrderWorklist.append(L->begin(), L->end());
1463        PreOrderLoops.push_back(L);
1464      } while (!PreOrderWorklist.empty());
1465  
1466      Worklist.insert(std::move(PreOrderLoops));
1467      PreOrderLoops.clear();
1468    }
1469  }
1470  
1471  template <typename RangeT>
1472  void llvm::appendLoopsToWorklist(RangeT &&Loops,
1473                                   SmallPriorityWorklist<Loop *, 4> &Worklist) {
1474    appendReversedLoopsToWorklist(reverse(Loops), Worklist);
1475  }
1476  
1477  template void llvm::appendLoopsToWorklist<ArrayRef<Loop *> &>(
1478      ArrayRef<Loop *> &Loops, SmallPriorityWorklist<Loop *, 4> &Worklist);
1479  
1480  template void
1481  llvm::appendLoopsToWorklist<Loop &>(Loop &L,
1482                                      SmallPriorityWorklist<Loop *, 4> &Worklist);
1483  
1484  void llvm::appendLoopsToWorklist(LoopInfo &LI,
1485                                   SmallPriorityWorklist<Loop *, 4> &Worklist) {
1486    appendReversedLoopsToWorklist(LI, Worklist);
1487  }
1488  
1489  Loop *llvm::cloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM,
1490                        LoopInfo *LI, LPPassManager *LPM) {
1491    Loop &New = *LI->AllocateLoop();
1492    if (PL)
1493      PL->addChildLoop(&New);
1494    else
1495      LI->addTopLevelLoop(&New);
1496  
1497    if (LPM)
1498      LPM->addLoop(New);
1499  
1500    // Add all of the blocks in L to the new loop.
1501    for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
1502         I != E; ++I)
1503      if (LI->getLoopFor(*I) == L)
1504        New.addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), *LI);
1505  
1506    // Add all of the subloops to the new loop.
1507    for (Loop *I : *L)
1508      cloneLoop(I, &New, VM, LI, LPM);
1509  
1510    return &New;
1511  }
1512  
1513  /// IR Values for the lower and upper bounds of a pointer evolution.  We
1514  /// need to use value-handles because SCEV expansion can invalidate previously
1515  /// expanded values.  Thus expansion of a pointer can invalidate the bounds for
1516  /// a previous one.
1517  struct PointerBounds {
1518    TrackingVH<Value> Start;
1519    TrackingVH<Value> End;
1520  };
1521  
1522  /// Expand code for the lower and upper bound of the pointer group \p CG
1523  /// in \p TheLoop.  \return the values for the bounds.
1524  static PointerBounds expandBounds(const RuntimeCheckingPtrGroup *CG,
1525                                    Loop *TheLoop, Instruction *Loc,
1526                                    SCEVExpander &Exp) {
1527    LLVMContext &Ctx = Loc->getContext();
1528    Type *PtrArithTy = Type::getInt8PtrTy(Ctx, CG->AddressSpace);
1529  
1530    Value *Start = nullptr, *End = nullptr;
1531    LLVM_DEBUG(dbgs() << "LAA: Adding RT check for range:\n");
1532    Start = Exp.expandCodeFor(CG->Low, PtrArithTy, Loc);
1533    End = Exp.expandCodeFor(CG->High, PtrArithTy, Loc);
1534    LLVM_DEBUG(dbgs() << "Start: " << *CG->Low << " End: " << *CG->High << "\n");
1535    return {Start, End};
1536  }
1537  
1538  /// Turns a collection of checks into a collection of expanded upper and
1539  /// lower bounds for both pointers in the check.
1540  static SmallVector<std::pair<PointerBounds, PointerBounds>, 4>
1541  expandBounds(const SmallVectorImpl<RuntimePointerCheck> &PointerChecks, Loop *L,
1542               Instruction *Loc, SCEVExpander &Exp) {
1543    SmallVector<std::pair<PointerBounds, PointerBounds>, 4> ChecksWithBounds;
1544  
1545    // Here we're relying on the SCEV Expander's cache to only emit code for the
1546    // same bounds once.
1547    transform(PointerChecks, std::back_inserter(ChecksWithBounds),
1548              [&](const RuntimePointerCheck &Check) {
1549                PointerBounds First = expandBounds(Check.first, L, Loc, Exp),
1550                              Second = expandBounds(Check.second, L, Loc, Exp);
1551                return std::make_pair(First, Second);
1552              });
1553  
1554    return ChecksWithBounds;
1555  }
1556  
1557  std::pair<Instruction *, Instruction *> llvm::addRuntimeChecks(
1558      Instruction *Loc, Loop *TheLoop,
1559      const SmallVectorImpl<RuntimePointerCheck> &PointerChecks,
1560      SCEVExpander &Exp) {
1561    // TODO: Move noalias annotation code from LoopVersioning here and share with LV if possible.
1562    // TODO: Pass  RtPtrChecking instead of PointerChecks and SE separately, if possible
1563    auto ExpandedChecks = expandBounds(PointerChecks, TheLoop, Loc, Exp);
1564  
1565    LLVMContext &Ctx = Loc->getContext();
1566    Instruction *FirstInst = nullptr;
1567    IRBuilder<> ChkBuilder(Loc);
1568    // Our instructions might fold to a constant.
1569    Value *MemoryRuntimeCheck = nullptr;
1570  
1571    // FIXME: this helper is currently a duplicate of the one in
1572    // LoopVectorize.cpp.
1573    auto GetFirstInst = [](Instruction *FirstInst, Value *V,
1574                           Instruction *Loc) -> Instruction * {
1575      if (FirstInst)
1576        return FirstInst;
1577      if (Instruction *I = dyn_cast<Instruction>(V))
1578        return I->getParent() == Loc->getParent() ? I : nullptr;
1579      return nullptr;
1580    };
1581  
1582    for (const auto &Check : ExpandedChecks) {
1583      const PointerBounds &A = Check.first, &B = Check.second;
1584      // Check if two pointers (A and B) conflict where conflict is computed as:
1585      // start(A) <= end(B) && start(B) <= end(A)
1586      unsigned AS0 = A.Start->getType()->getPointerAddressSpace();
1587      unsigned AS1 = B.Start->getType()->getPointerAddressSpace();
1588  
1589      assert((AS0 == B.End->getType()->getPointerAddressSpace()) &&
1590             (AS1 == A.End->getType()->getPointerAddressSpace()) &&
1591             "Trying to bounds check pointers with different address spaces");
1592  
1593      Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1594      Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1595  
1596      Value *Start0 = ChkBuilder.CreateBitCast(A.Start, PtrArithTy0, "bc");
1597      Value *Start1 = ChkBuilder.CreateBitCast(B.Start, PtrArithTy1, "bc");
1598      Value *End0 = ChkBuilder.CreateBitCast(A.End, PtrArithTy1, "bc");
1599      Value *End1 = ChkBuilder.CreateBitCast(B.End, PtrArithTy0, "bc");
1600  
1601      // [A|B].Start points to the first accessed byte under base [A|B].
1602      // [A|B].End points to the last accessed byte, plus one.
1603      // There is no conflict when the intervals are disjoint:
1604      // NoConflict = (B.Start >= A.End) || (A.Start >= B.End)
1605      //
1606      // bound0 = (B.Start < A.End)
1607      // bound1 = (A.Start < B.End)
1608      //  IsConflict = bound0 & bound1
1609      Value *Cmp0 = ChkBuilder.CreateICmpULT(Start0, End1, "bound0");
1610      FirstInst = GetFirstInst(FirstInst, Cmp0, Loc);
1611      Value *Cmp1 = ChkBuilder.CreateICmpULT(Start1, End0, "bound1");
1612      FirstInst = GetFirstInst(FirstInst, Cmp1, Loc);
1613      Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1614      FirstInst = GetFirstInst(FirstInst, IsConflict, Loc);
1615      if (MemoryRuntimeCheck) {
1616        IsConflict =
1617            ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict, "conflict.rdx");
1618        FirstInst = GetFirstInst(FirstInst, IsConflict, Loc);
1619      }
1620      MemoryRuntimeCheck = IsConflict;
1621    }
1622  
1623    if (!MemoryRuntimeCheck)
1624      return std::make_pair(nullptr, nullptr);
1625  
1626    // We have to do this trickery because the IRBuilder might fold the check to a
1627    // constant expression in which case there is no Instruction anchored in a
1628    // the block.
1629    Instruction *Check =
1630        BinaryOperator::CreateAnd(MemoryRuntimeCheck, ConstantInt::getTrue(Ctx));
1631    ChkBuilder.Insert(Check, "memcheck.conflict");
1632    FirstInst = GetFirstInst(FirstInst, Check, Loc);
1633    return std::make_pair(FirstInst, Check);
1634  }
1635  
1636  Optional<IVConditionInfo> llvm::hasPartialIVCondition(Loop &L,
1637                                                        unsigned MSSAThreshold,
1638                                                        MemorySSA &MSSA,
1639                                                        AAResults &AA) {
1640    auto *TI = dyn_cast<BranchInst>(L.getHeader()->getTerminator());
1641    if (!TI || !TI->isConditional())
1642      return {};
1643  
1644    auto *CondI = dyn_cast<CmpInst>(TI->getCondition());
1645    // The case with the condition outside the loop should already be handled
1646    // earlier.
1647    if (!CondI || !L.contains(CondI))
1648      return {};
1649  
1650    SmallVector<Instruction *> InstToDuplicate;
1651    InstToDuplicate.push_back(CondI);
1652  
1653    SmallVector<Value *, 4> WorkList;
1654    WorkList.append(CondI->op_begin(), CondI->op_end());
1655  
1656    SmallVector<MemoryAccess *, 4> AccessesToCheck;
1657    SmallVector<MemoryLocation, 4> AccessedLocs;
1658    while (!WorkList.empty()) {
1659      Instruction *I = dyn_cast<Instruction>(WorkList.pop_back_val());
1660      if (!I || !L.contains(I))
1661        continue;
1662  
1663      // TODO: support additional instructions.
1664      if (!isa<LoadInst>(I) && !isa<GetElementPtrInst>(I))
1665        return {};
1666  
1667      // Do not duplicate volatile and atomic loads.
1668      if (auto *LI = dyn_cast<LoadInst>(I))
1669        if (LI->isVolatile() || LI->isAtomic())
1670          return {};
1671  
1672      InstToDuplicate.push_back(I);
1673      if (MemoryAccess *MA = MSSA.getMemoryAccess(I)) {
1674        if (auto *MemUse = dyn_cast_or_null<MemoryUse>(MA)) {
1675          // Queue the defining access to check for alias checks.
1676          AccessesToCheck.push_back(MemUse->getDefiningAccess());
1677          AccessedLocs.push_back(MemoryLocation::get(I));
1678        } else {
1679          // MemoryDefs may clobber the location or may be atomic memory
1680          // operations. Bail out.
1681          return {};
1682        }
1683      }
1684      WorkList.append(I->op_begin(), I->op_end());
1685    }
1686  
1687    if (InstToDuplicate.empty())
1688      return {};
1689  
1690    SmallVector<BasicBlock *, 4> ExitingBlocks;
1691    L.getExitingBlocks(ExitingBlocks);
1692    auto HasNoClobbersOnPath =
1693        [&L, &AA, &AccessedLocs, &ExitingBlocks, &InstToDuplicate,
1694         MSSAThreshold](BasicBlock *Succ, BasicBlock *Header,
1695                        SmallVector<MemoryAccess *, 4> AccessesToCheck)
1696        -> Optional<IVConditionInfo> {
1697      IVConditionInfo Info;
1698      // First, collect all blocks in the loop that are on a patch from Succ
1699      // to the header.
1700      SmallVector<BasicBlock *, 4> WorkList;
1701      WorkList.push_back(Succ);
1702      WorkList.push_back(Header);
1703      SmallPtrSet<BasicBlock *, 4> Seen;
1704      Seen.insert(Header);
1705      Info.PathIsNoop &=
1706          all_of(*Header, [](Instruction &I) { return !I.mayHaveSideEffects(); });
1707  
1708      while (!WorkList.empty()) {
1709        BasicBlock *Current = WorkList.pop_back_val();
1710        if (!L.contains(Current))
1711          continue;
1712        const auto &SeenIns = Seen.insert(Current);
1713        if (!SeenIns.second)
1714          continue;
1715  
1716        Info.PathIsNoop &= all_of(
1717            *Current, [](Instruction &I) { return !I.mayHaveSideEffects(); });
1718        WorkList.append(succ_begin(Current), succ_end(Current));
1719      }
1720  
1721      // Require at least 2 blocks on a path through the loop. This skips
1722      // paths that directly exit the loop.
1723      if (Seen.size() < 2)
1724        return {};
1725  
1726      // Next, check if there are any MemoryDefs that are on the path through
1727      // the loop (in the Seen set) and they may-alias any of the locations in
1728      // AccessedLocs. If that is the case, they may modify the condition and
1729      // partial unswitching is not possible.
1730      SmallPtrSet<MemoryAccess *, 4> SeenAccesses;
1731      while (!AccessesToCheck.empty()) {
1732        MemoryAccess *Current = AccessesToCheck.pop_back_val();
1733        auto SeenI = SeenAccesses.insert(Current);
1734        if (!SeenI.second || !Seen.contains(Current->getBlock()))
1735          continue;
1736  
1737        // Bail out if exceeded the threshold.
1738        if (SeenAccesses.size() >= MSSAThreshold)
1739          return {};
1740  
1741        // MemoryUse are read-only accesses.
1742        if (isa<MemoryUse>(Current))
1743          continue;
1744  
1745        // For a MemoryDef, check if is aliases any of the location feeding
1746        // the original condition.
1747        if (auto *CurrentDef = dyn_cast<MemoryDef>(Current)) {
1748          if (any_of(AccessedLocs, [&AA, CurrentDef](MemoryLocation &Loc) {
1749                return isModSet(
1750                    AA.getModRefInfo(CurrentDef->getMemoryInst(), Loc));
1751              }))
1752            return {};
1753        }
1754  
1755        for (Use &U : Current->uses())
1756          AccessesToCheck.push_back(cast<MemoryAccess>(U.getUser()));
1757      }
1758  
1759      // We could also allow loops with known trip counts without mustprogress,
1760      // but ScalarEvolution may not be available.
1761      Info.PathIsNoop &= isMustProgress(&L);
1762  
1763      // If the path is considered a no-op so far, check if it reaches a
1764      // single exit block without any phis. This ensures no values from the
1765      // loop are used outside of the loop.
1766      if (Info.PathIsNoop) {
1767        for (auto *Exiting : ExitingBlocks) {
1768          if (!Seen.contains(Exiting))
1769            continue;
1770          for (auto *Succ : successors(Exiting)) {
1771            if (L.contains(Succ))
1772              continue;
1773  
1774            Info.PathIsNoop &= llvm::empty(Succ->phis()) &&
1775                               (!Info.ExitForPath || Info.ExitForPath == Succ);
1776            if (!Info.PathIsNoop)
1777              break;
1778            assert((!Info.ExitForPath || Info.ExitForPath == Succ) &&
1779                   "cannot have multiple exit blocks");
1780            Info.ExitForPath = Succ;
1781          }
1782        }
1783      }
1784      if (!Info.ExitForPath)
1785        Info.PathIsNoop = false;
1786  
1787      Info.InstToDuplicate = InstToDuplicate;
1788      return Info;
1789    };
1790  
1791    // If we branch to the same successor, partial unswitching will not be
1792    // beneficial.
1793    if (TI->getSuccessor(0) == TI->getSuccessor(1))
1794      return {};
1795  
1796    if (auto Info = HasNoClobbersOnPath(TI->getSuccessor(0), L.getHeader(),
1797                                        AccessesToCheck)) {
1798      Info->KnownValue = ConstantInt::getTrue(TI->getContext());
1799      return Info;
1800    }
1801    if (auto Info = HasNoClobbersOnPath(TI->getSuccessor(1), L.getHeader(),
1802                                        AccessesToCheck)) {
1803      Info->KnownValue = ConstantInt::getFalse(TI->getContext());
1804      return Info;
1805    }
1806  
1807    return {};
1808  }
1809