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/PriorityWorklist.h"
16 #include "llvm/ADT/ScopeExit.h"
17 #include "llvm/ADT/SetVector.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/Analysis/BasicAliasAnalysis.h"
22 #include "llvm/Analysis/DomTreeUpdater.h"
23 #include "llvm/Analysis/GlobalsModRef.h"
24 #include "llvm/Analysis/InstSimplifyFolder.h"
25 #include "llvm/Analysis/LoopAccessAnalysis.h"
26 #include "llvm/Analysis/LoopInfo.h"
27 #include "llvm/Analysis/LoopPass.h"
28 #include "llvm/Analysis/MemorySSA.h"
29 #include "llvm/Analysis/MemorySSAUpdater.h"
30 #include "llvm/Analysis/ScalarEvolution.h"
31 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
32 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
33 #include "llvm/IR/DIBuilder.h"
34 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/Instructions.h"
36 #include "llvm/IR/IntrinsicInst.h"
37 #include "llvm/IR/MDBuilder.h"
38 #include "llvm/IR/Module.h"
39 #include "llvm/IR/PatternMatch.h"
40 #include "llvm/IR/ProfDataUtils.h"
41 #include "llvm/IR/ValueHandle.h"
42 #include "llvm/InitializePasses.h"
43 #include "llvm/Pass.h"
44 #include "llvm/Support/Debug.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
48 
49 using namespace llvm;
50 using namespace llvm::PatternMatch;
51 
52 #define DEBUG_TYPE "loop-utils"
53 
54 static const char *LLVMLoopDisableNonforced = "llvm.loop.disable_nonforced";
55 static const char *LLVMLoopDisableLICM = "llvm.licm.disable";
56 
57 bool llvm::formDedicatedExitBlocks(Loop *L, DominatorTree *DT, LoopInfo *LI,
58                                    MemorySSAUpdater *MSSAU,
59                                    bool PreserveLCSSA) {
60   bool Changed = false;
61 
62   // We re-use a vector for the in-loop predecesosrs.
63   SmallVector<BasicBlock *, 4> InLoopPredecessors;
64 
65   auto RewriteExit = [&](BasicBlock *BB) {
66     assert(InLoopPredecessors.empty() &&
67            "Must start with an empty predecessors list!");
68     auto Cleanup = make_scope_exit([&] { InLoopPredecessors.clear(); });
69 
70     // See if there are any non-loop predecessors of this exit block and
71     // keep track of the in-loop predecessors.
72     bool IsDedicatedExit = true;
73     for (auto *PredBB : predecessors(BB))
74       if (L->contains(PredBB)) {
75         if (isa<IndirectBrInst>(PredBB->getTerminator()))
76           // We cannot rewrite exiting edges from an indirectbr.
77           return false;
78 
79         InLoopPredecessors.push_back(PredBB);
80       } else {
81         IsDedicatedExit = false;
82       }
83 
84     assert(!InLoopPredecessors.empty() && "Must have *some* loop predecessor!");
85 
86     // Nothing to do if this is already a dedicated exit.
87     if (IsDedicatedExit)
88       return false;
89 
90     auto *NewExitBB = SplitBlockPredecessors(
91         BB, InLoopPredecessors, ".loopexit", DT, LI, MSSAU, PreserveLCSSA);
92 
93     if (!NewExitBB)
94       LLVM_DEBUG(
95           dbgs() << "WARNING: Can't create a dedicated exit block for loop: "
96                  << *L << "\n");
97     else
98       LLVM_DEBUG(dbgs() << "LoopSimplify: Creating dedicated exit block "
99                         << NewExitBB->getName() << "\n");
100     return true;
101   };
102 
103   // Walk the exit blocks directly rather than building up a data structure for
104   // them, but only visit each one once.
105   SmallPtrSet<BasicBlock *, 4> Visited;
106   for (auto *BB : L->blocks())
107     for (auto *SuccBB : successors(BB)) {
108       // We're looking for exit blocks so skip in-loop successors.
109       if (L->contains(SuccBB))
110         continue;
111 
112       // Visit each exit block exactly once.
113       if (!Visited.insert(SuccBB).second)
114         continue;
115 
116       Changed |= RewriteExit(SuccBB);
117     }
118 
119   return Changed;
120 }
121 
122 /// Returns the instructions that use values defined in the loop.
123 SmallVector<Instruction *, 8> llvm::findDefsUsedOutsideOfLoop(Loop *L) {
124   SmallVector<Instruction *, 8> UsedOutside;
125 
126   for (auto *Block : L->getBlocks())
127     // FIXME: I believe that this could use copy_if if the Inst reference could
128     // be adapted into a pointer.
129     for (auto &Inst : *Block) {
130       auto Users = Inst.users();
131       if (any_of(Users, [&](User *U) {
132             auto *Use = cast<Instruction>(U);
133             return !L->contains(Use->getParent());
134           }))
135         UsedOutside.push_back(&Inst);
136     }
137 
138   return UsedOutside;
139 }
140 
141 void llvm::getLoopAnalysisUsage(AnalysisUsage &AU) {
142   // By definition, all loop passes need the LoopInfo analysis and the
143   // Dominator tree it depends on. Because they all participate in the loop
144   // pass manager, they must also preserve these.
145   AU.addRequired<DominatorTreeWrapperPass>();
146   AU.addPreserved<DominatorTreeWrapperPass>();
147   AU.addRequired<LoopInfoWrapperPass>();
148   AU.addPreserved<LoopInfoWrapperPass>();
149 
150   // We must also preserve LoopSimplify and LCSSA. We locally access their IDs
151   // here because users shouldn't directly get them from this header.
152   extern char &LoopSimplifyID;
153   extern char &LCSSAID;
154   AU.addRequiredID(LoopSimplifyID);
155   AU.addPreservedID(LoopSimplifyID);
156   AU.addRequiredID(LCSSAID);
157   AU.addPreservedID(LCSSAID);
158   // This is used in the LPPassManager to perform LCSSA verification on passes
159   // which preserve lcssa form
160   AU.addRequired<LCSSAVerificationPass>();
161   AU.addPreserved<LCSSAVerificationPass>();
162 
163   // Loop passes are designed to run inside of a loop pass manager which means
164   // that any function analyses they require must be required by the first loop
165   // pass in the manager (so that it is computed before the loop pass manager
166   // runs) and preserved by all loop pasess in the manager. To make this
167   // reasonably robust, the set needed for most loop passes is maintained here.
168   // If your loop pass requires an analysis not listed here, you will need to
169   // carefully audit the loop pass manager nesting structure that results.
170   AU.addRequired<AAResultsWrapperPass>();
171   AU.addPreserved<AAResultsWrapperPass>();
172   AU.addPreserved<BasicAAWrapperPass>();
173   AU.addPreserved<GlobalsAAWrapperPass>();
174   AU.addPreserved<SCEVAAWrapperPass>();
175   AU.addRequired<ScalarEvolutionWrapperPass>();
176   AU.addPreserved<ScalarEvolutionWrapperPass>();
177   // FIXME: When all loop passes preserve MemorySSA, it can be required and
178   // preserved here instead of the individual handling in each pass.
179 }
180 
181 /// Manually defined generic "LoopPass" dependency initialization. This is used
182 /// to initialize the exact set of passes from above in \c
183 /// getLoopAnalysisUsage. It can be used within a loop pass's initialization
184 /// with:
185 ///
186 ///   INITIALIZE_PASS_DEPENDENCY(LoopPass)
187 ///
188 /// As-if "LoopPass" were a pass.
189 void llvm::initializeLoopPassPass(PassRegistry &Registry) {
190   INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
191   INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
192   INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
193   INITIALIZE_PASS_DEPENDENCY(LCSSAWrapperPass)
194   INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
195   INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)
196   INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
197   INITIALIZE_PASS_DEPENDENCY(SCEVAAWrapperPass)
198   INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
199   INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
200 }
201 
202 /// Create MDNode for input string.
203 static MDNode *createStringMetadata(Loop *TheLoop, StringRef Name, unsigned V) {
204   LLVMContext &Context = TheLoop->getHeader()->getContext();
205   Metadata *MDs[] = {
206       MDString::get(Context, Name),
207       ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(Context), V))};
208   return MDNode::get(Context, MDs);
209 }
210 
211 /// Set input string into loop metadata by keeping other values intact.
212 /// If the string is already in loop metadata update value if it is
213 /// different.
214 void llvm::addStringMetadataToLoop(Loop *TheLoop, const char *StringMD,
215                                    unsigned V) {
216   SmallVector<Metadata *, 4> MDs(1);
217   // If the loop already has metadata, retain it.
218   MDNode *LoopID = TheLoop->getLoopID();
219   if (LoopID) {
220     for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
221       MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
222       // If it is of form key = value, try to parse it.
223       if (Node->getNumOperands() == 2) {
224         MDString *S = dyn_cast<MDString>(Node->getOperand(0));
225         if (S && S->getString().equals(StringMD)) {
226           ConstantInt *IntMD =
227               mdconst::extract_or_null<ConstantInt>(Node->getOperand(1));
228           if (IntMD && IntMD->getSExtValue() == V)
229             // It is already in place. Do nothing.
230             return;
231           // We need to update the value, so just skip it here and it will
232           // be added after copying other existed nodes.
233           continue;
234         }
235       }
236       MDs.push_back(Node);
237     }
238   }
239   // Add new metadata.
240   MDs.push_back(createStringMetadata(TheLoop, StringMD, V));
241   // Replace current metadata node with new one.
242   LLVMContext &Context = TheLoop->getHeader()->getContext();
243   MDNode *NewLoopID = MDNode::get(Context, MDs);
244   // Set operand 0 to refer to the loop id itself.
245   NewLoopID->replaceOperandWith(0, NewLoopID);
246   TheLoop->setLoopID(NewLoopID);
247 }
248 
249 std::optional<ElementCount>
250 llvm::getOptionalElementCountLoopAttribute(const Loop *TheLoop) {
251   std::optional<int> Width =
252       getOptionalIntLoopAttribute(TheLoop, "llvm.loop.vectorize.width");
253 
254   if (Width) {
255     std::optional<int> IsScalable = getOptionalIntLoopAttribute(
256         TheLoop, "llvm.loop.vectorize.scalable.enable");
257     return ElementCount::get(*Width, IsScalable.value_or(false));
258   }
259 
260   return std::nullopt;
261 }
262 
263 std::optional<MDNode *> llvm::makeFollowupLoopID(
264     MDNode *OrigLoopID, ArrayRef<StringRef> FollowupOptions,
265     const char *InheritOptionsExceptPrefix, bool AlwaysNew) {
266   if (!OrigLoopID) {
267     if (AlwaysNew)
268       return nullptr;
269     return std::nullopt;
270   }
271 
272   assert(OrigLoopID->getOperand(0) == OrigLoopID);
273 
274   bool InheritAllAttrs = !InheritOptionsExceptPrefix;
275   bool InheritSomeAttrs =
276       InheritOptionsExceptPrefix && InheritOptionsExceptPrefix[0] != '\0';
277   SmallVector<Metadata *, 8> MDs;
278   MDs.push_back(nullptr);
279 
280   bool Changed = false;
281   if (InheritAllAttrs || InheritSomeAttrs) {
282     for (const MDOperand &Existing : drop_begin(OrigLoopID->operands())) {
283       MDNode *Op = cast<MDNode>(Existing.get());
284 
285       auto InheritThisAttribute = [InheritSomeAttrs,
286                                    InheritOptionsExceptPrefix](MDNode *Op) {
287         if (!InheritSomeAttrs)
288           return false;
289 
290         // Skip malformatted attribute metadata nodes.
291         if (Op->getNumOperands() == 0)
292           return true;
293         Metadata *NameMD = Op->getOperand(0).get();
294         if (!isa<MDString>(NameMD))
295           return true;
296         StringRef AttrName = cast<MDString>(NameMD)->getString();
297 
298         // Do not inherit excluded attributes.
299         return !AttrName.startswith(InheritOptionsExceptPrefix);
300       };
301 
302       if (InheritThisAttribute(Op))
303         MDs.push_back(Op);
304       else
305         Changed = true;
306     }
307   } else {
308     // Modified if we dropped at least one attribute.
309     Changed = OrigLoopID->getNumOperands() > 1;
310   }
311 
312   bool HasAnyFollowup = false;
313   for (StringRef OptionName : FollowupOptions) {
314     MDNode *FollowupNode = findOptionMDForLoopID(OrigLoopID, OptionName);
315     if (!FollowupNode)
316       continue;
317 
318     HasAnyFollowup = true;
319     for (const MDOperand &Option : drop_begin(FollowupNode->operands())) {
320       MDs.push_back(Option.get());
321       Changed = true;
322     }
323   }
324 
325   // Attributes of the followup loop not specified explicity, so signal to the
326   // transformation pass to add suitable attributes.
327   if (!AlwaysNew && !HasAnyFollowup)
328     return std::nullopt;
329 
330   // If no attributes were added or remove, the previous loop Id can be reused.
331   if (!AlwaysNew && !Changed)
332     return OrigLoopID;
333 
334   // No attributes is equivalent to having no !llvm.loop metadata at all.
335   if (MDs.size() == 1)
336     return nullptr;
337 
338   // Build the new loop ID.
339   MDTuple *FollowupLoopID = MDNode::get(OrigLoopID->getContext(), MDs);
340   FollowupLoopID->replaceOperandWith(0, FollowupLoopID);
341   return FollowupLoopID;
342 }
343 
344 bool llvm::hasDisableAllTransformsHint(const Loop *L) {
345   return getBooleanLoopAttribute(L, LLVMLoopDisableNonforced);
346 }
347 
348 bool llvm::hasDisableLICMTransformsHint(const Loop *L) {
349   return getBooleanLoopAttribute(L, LLVMLoopDisableLICM);
350 }
351 
352 TransformationMode llvm::hasUnrollTransformation(const Loop *L) {
353   if (getBooleanLoopAttribute(L, "llvm.loop.unroll.disable"))
354     return TM_SuppressedByUser;
355 
356   std::optional<int> Count =
357       getOptionalIntLoopAttribute(L, "llvm.loop.unroll.count");
358   if (Count)
359     return *Count == 1 ? TM_SuppressedByUser : TM_ForcedByUser;
360 
361   if (getBooleanLoopAttribute(L, "llvm.loop.unroll.enable"))
362     return TM_ForcedByUser;
363 
364   if (getBooleanLoopAttribute(L, "llvm.loop.unroll.full"))
365     return TM_ForcedByUser;
366 
367   if (hasDisableAllTransformsHint(L))
368     return TM_Disable;
369 
370   return TM_Unspecified;
371 }
372 
373 TransformationMode llvm::hasUnrollAndJamTransformation(const Loop *L) {
374   if (getBooleanLoopAttribute(L, "llvm.loop.unroll_and_jam.disable"))
375     return TM_SuppressedByUser;
376 
377   std::optional<int> Count =
378       getOptionalIntLoopAttribute(L, "llvm.loop.unroll_and_jam.count");
379   if (Count)
380     return *Count == 1 ? TM_SuppressedByUser : TM_ForcedByUser;
381 
382   if (getBooleanLoopAttribute(L, "llvm.loop.unroll_and_jam.enable"))
383     return TM_ForcedByUser;
384 
385   if (hasDisableAllTransformsHint(L))
386     return TM_Disable;
387 
388   return TM_Unspecified;
389 }
390 
391 TransformationMode llvm::hasVectorizeTransformation(const Loop *L) {
392   std::optional<bool> Enable =
393       getOptionalBoolLoopAttribute(L, "llvm.loop.vectorize.enable");
394 
395   if (Enable == false)
396     return TM_SuppressedByUser;
397 
398   std::optional<ElementCount> VectorizeWidth =
399       getOptionalElementCountLoopAttribute(L);
400   std::optional<int> InterleaveCount =
401       getOptionalIntLoopAttribute(L, "llvm.loop.interleave.count");
402 
403   // 'Forcing' vector width and interleave count to one effectively disables
404   // this tranformation.
405   if (Enable == true && VectorizeWidth && VectorizeWidth->isScalar() &&
406       InterleaveCount == 1)
407     return TM_SuppressedByUser;
408 
409   if (getBooleanLoopAttribute(L, "llvm.loop.isvectorized"))
410     return TM_Disable;
411 
412   if (Enable == true)
413     return TM_ForcedByUser;
414 
415   if ((VectorizeWidth && VectorizeWidth->isScalar()) && InterleaveCount == 1)
416     return TM_Disable;
417 
418   if ((VectorizeWidth && VectorizeWidth->isVector()) || InterleaveCount > 1)
419     return TM_Enable;
420 
421   if (hasDisableAllTransformsHint(L))
422     return TM_Disable;
423 
424   return TM_Unspecified;
425 }
426 
427 TransformationMode llvm::hasDistributeTransformation(const Loop *L) {
428   if (getBooleanLoopAttribute(L, "llvm.loop.distribute.enable"))
429     return TM_ForcedByUser;
430 
431   if (hasDisableAllTransformsHint(L))
432     return TM_Disable;
433 
434   return TM_Unspecified;
435 }
436 
437 TransformationMode llvm::hasLICMVersioningTransformation(const Loop *L) {
438   if (getBooleanLoopAttribute(L, "llvm.loop.licm_versioning.disable"))
439     return TM_SuppressedByUser;
440 
441   if (hasDisableAllTransformsHint(L))
442     return TM_Disable;
443 
444   return TM_Unspecified;
445 }
446 
447 /// Does a BFS from a given node to all of its children inside a given loop.
448 /// The returned vector of nodes includes the starting point.
449 SmallVector<DomTreeNode *, 16>
450 llvm::collectChildrenInLoop(DomTreeNode *N, const Loop *CurLoop) {
451   SmallVector<DomTreeNode *, 16> Worklist;
452   auto AddRegionToWorklist = [&](DomTreeNode *DTN) {
453     // Only include subregions in the top level loop.
454     BasicBlock *BB = DTN->getBlock();
455     if (CurLoop->contains(BB))
456       Worklist.push_back(DTN);
457   };
458 
459   AddRegionToWorklist(N);
460 
461   for (size_t I = 0; I < Worklist.size(); I++) {
462     for (DomTreeNode *Child : Worklist[I]->children())
463       AddRegionToWorklist(Child);
464   }
465 
466   return Worklist;
467 }
468 
469 void llvm::deleteDeadLoop(Loop *L, DominatorTree *DT, ScalarEvolution *SE,
470                           LoopInfo *LI, MemorySSA *MSSA) {
471   assert((!DT || L->isLCSSAForm(*DT)) && "Expected LCSSA!");
472   auto *Preheader = L->getLoopPreheader();
473   assert(Preheader && "Preheader should exist!");
474 
475   std::unique_ptr<MemorySSAUpdater> MSSAU;
476   if (MSSA)
477     MSSAU = std::make_unique<MemorySSAUpdater>(MSSA);
478 
479   // Now that we know the removal is safe, remove the loop by changing the
480   // branch from the preheader to go to the single exit block.
481   //
482   // Because we're deleting a large chunk of code at once, the sequence in which
483   // we remove things is very important to avoid invalidation issues.
484 
485   // Tell ScalarEvolution that the loop is deleted. Do this before
486   // deleting the loop so that ScalarEvolution can look at the loop
487   // to determine what it needs to clean up.
488   if (SE) {
489     SE->forgetLoop(L);
490     SE->forgetBlockAndLoopDispositions();
491   }
492 
493   Instruction *OldTerm = Preheader->getTerminator();
494   assert(!OldTerm->mayHaveSideEffects() &&
495          "Preheader must end with a side-effect-free terminator");
496   assert(OldTerm->getNumSuccessors() == 1 &&
497          "Preheader must have a single successor");
498   // Connect the preheader to the exit block. Keep the old edge to the header
499   // around to perform the dominator tree update in two separate steps
500   // -- #1 insertion of the edge preheader -> exit and #2 deletion of the edge
501   // preheader -> header.
502   //
503   //
504   // 0.  Preheader          1.  Preheader           2.  Preheader
505   //        |                    |   |                   |
506   //        V                    |   V                   |
507   //      Header <--\            | Header <--\           | Header <--\
508   //       |  |     |            |  |  |     |           |  |  |     |
509   //       |  V     |            |  |  V     |           |  |  V     |
510   //       | Body --/            |  | Body --/           |  | Body --/
511   //       V                     V  V                    V  V
512   //      Exit                   Exit                    Exit
513   //
514   // By doing this is two separate steps we can perform the dominator tree
515   // update without using the batch update API.
516   //
517   // Even when the loop is never executed, we cannot remove the edge from the
518   // source block to the exit block. Consider the case where the unexecuted loop
519   // branches back to an outer loop. If we deleted the loop and removed the edge
520   // coming to this inner loop, this will break the outer loop structure (by
521   // deleting the backedge of the outer loop). If the outer loop is indeed a
522   // non-loop, it will be deleted in a future iteration of loop deletion pass.
523   IRBuilder<> Builder(OldTerm);
524 
525   auto *ExitBlock = L->getUniqueExitBlock();
526   DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);
527   if (ExitBlock) {
528     assert(ExitBlock && "Should have a unique exit block!");
529     assert(L->hasDedicatedExits() && "Loop should have dedicated exits!");
530 
531     Builder.CreateCondBr(Builder.getFalse(), L->getHeader(), ExitBlock);
532     // Remove the old branch. The conditional branch becomes a new terminator.
533     OldTerm->eraseFromParent();
534 
535     // Rewrite phis in the exit block to get their inputs from the Preheader
536     // instead of the exiting block.
537     for (PHINode &P : ExitBlock->phis()) {
538       // Set the zero'th element of Phi to be from the preheader and remove all
539       // other incoming values. Given the loop has dedicated exits, all other
540       // incoming values must be from the exiting blocks.
541       int PredIndex = 0;
542       P.setIncomingBlock(PredIndex, Preheader);
543       // Removes all incoming values from all other exiting blocks (including
544       // duplicate values from an exiting block).
545       // Nuke all entries except the zero'th entry which is the preheader entry.
546       // NOTE! We need to remove Incoming Values in the reverse order as done
547       // below, to keep the indices valid for deletion (removeIncomingValues
548       // updates getNumIncomingValues and shifts all values down into the
549       // operand being deleted).
550       for (unsigned i = 0, e = P.getNumIncomingValues() - 1; i != e; ++i)
551         P.removeIncomingValue(e - i, false);
552 
553       assert((P.getNumIncomingValues() == 1 &&
554               P.getIncomingBlock(PredIndex) == Preheader) &&
555              "Should have exactly one value and that's from the preheader!");
556     }
557 
558     if (DT) {
559       DTU.applyUpdates({{DominatorTree::Insert, Preheader, ExitBlock}});
560       if (MSSA) {
561         MSSAU->applyUpdates({{DominatorTree::Insert, Preheader, ExitBlock}},
562                             *DT);
563         if (VerifyMemorySSA)
564           MSSA->verifyMemorySSA();
565       }
566     }
567 
568     // Disconnect the loop body by branching directly to its exit.
569     Builder.SetInsertPoint(Preheader->getTerminator());
570     Builder.CreateBr(ExitBlock);
571     // Remove the old branch.
572     Preheader->getTerminator()->eraseFromParent();
573   } else {
574     assert(L->hasNoExitBlocks() &&
575            "Loop should have either zero or one exit blocks.");
576 
577     Builder.SetInsertPoint(OldTerm);
578     Builder.CreateUnreachable();
579     Preheader->getTerminator()->eraseFromParent();
580   }
581 
582   if (DT) {
583     DTU.applyUpdates({{DominatorTree::Delete, Preheader, L->getHeader()}});
584     if (MSSA) {
585       MSSAU->applyUpdates({{DominatorTree::Delete, Preheader, L->getHeader()}},
586                           *DT);
587       SmallSetVector<BasicBlock *, 8> DeadBlockSet(L->block_begin(),
588                                                    L->block_end());
589       MSSAU->removeBlocks(DeadBlockSet);
590       if (VerifyMemorySSA)
591         MSSA->verifyMemorySSA();
592     }
593   }
594 
595   // Use a map to unique and a vector to guarantee deterministic ordering.
596   llvm::SmallDenseSet<DebugVariable, 4> DeadDebugSet;
597   llvm::SmallVector<DbgVariableIntrinsic *, 4> DeadDebugInst;
598 
599   if (ExitBlock) {
600     // Given LCSSA form is satisfied, we should not have users of instructions
601     // within the dead loop outside of the loop. However, LCSSA doesn't take
602     // unreachable uses into account. We handle them here.
603     // We could do it after drop all references (in this case all users in the
604     // loop will be already eliminated and we have less work to do but according
605     // to API doc of User::dropAllReferences only valid operation after dropping
606     // references, is deletion. So let's substitute all usages of
607     // instruction from the loop with poison value of corresponding type first.
608     for (auto *Block : L->blocks())
609       for (Instruction &I : *Block) {
610         auto *Poison = PoisonValue::get(I.getType());
611         for (Use &U : llvm::make_early_inc_range(I.uses())) {
612           if (auto *Usr = dyn_cast<Instruction>(U.getUser()))
613             if (L->contains(Usr->getParent()))
614               continue;
615           // If we have a DT then we can check that uses outside a loop only in
616           // unreachable block.
617           if (DT)
618             assert(!DT->isReachableFromEntry(U) &&
619                    "Unexpected user in reachable block");
620           U.set(Poison);
621         }
622         auto *DVI = dyn_cast<DbgVariableIntrinsic>(&I);
623         if (!DVI)
624           continue;
625         if (!DeadDebugSet.insert(DebugVariable(DVI)).second)
626           continue;
627         DeadDebugInst.push_back(DVI);
628       }
629 
630     // After the loop has been deleted all the values defined and modified
631     // inside the loop are going to be unavailable.
632     // Since debug values in the loop have been deleted, inserting an undef
633     // dbg.value truncates the range of any dbg.value before the loop where the
634     // loop used to be. This is particularly important for constant values.
635     Instruction *InsertDbgValueBefore = ExitBlock->getFirstNonPHI();
636     assert(InsertDbgValueBefore &&
637            "There should be a non-PHI instruction in exit block, else these "
638            "instructions will have no parent.");
639     for (auto *DVI : DeadDebugInst) {
640       DVI->setKillLocation();
641       DVI->moveBefore(InsertDbgValueBefore);
642     }
643   }
644 
645   // Remove the block from the reference counting scheme, so that we can
646   // delete it freely later.
647   for (auto *Block : L->blocks())
648     Block->dropAllReferences();
649 
650   if (MSSA && VerifyMemorySSA)
651     MSSA->verifyMemorySSA();
652 
653   if (LI) {
654     // Erase the instructions and the blocks without having to worry
655     // about ordering because we already dropped the references.
656     // NOTE: This iteration is safe because erasing the block does not remove
657     // its entry from the loop's block list.  We do that in the next section.
658     for (BasicBlock *BB : L->blocks())
659       BB->eraseFromParent();
660 
661     // Finally, the blocks from loopinfo.  This has to happen late because
662     // otherwise our loop iterators won't work.
663 
664     SmallPtrSet<BasicBlock *, 8> blocks;
665     blocks.insert(L->block_begin(), L->block_end());
666     for (BasicBlock *BB : blocks)
667       LI->removeBlock(BB);
668 
669     // The last step is to update LoopInfo now that we've eliminated this loop.
670     // Note: LoopInfo::erase remove the given loop and relink its subloops with
671     // its parent. While removeLoop/removeChildLoop remove the given loop but
672     // not relink its subloops, which is what we want.
673     if (Loop *ParentLoop = L->getParentLoop()) {
674       Loop::iterator I = find(*ParentLoop, L);
675       assert(I != ParentLoop->end() && "Couldn't find loop");
676       ParentLoop->removeChildLoop(I);
677     } else {
678       Loop::iterator I = find(*LI, L);
679       assert(I != LI->end() && "Couldn't find loop");
680       LI->removeLoop(I);
681     }
682     LI->destroy(L);
683   }
684 }
685 
686 void llvm::breakLoopBackedge(Loop *L, DominatorTree &DT, ScalarEvolution &SE,
687                              LoopInfo &LI, MemorySSA *MSSA) {
688   auto *Latch = L->getLoopLatch();
689   assert(Latch && "multiple latches not yet supported");
690   auto *Header = L->getHeader();
691   Loop *OutermostLoop = L->getOutermostLoop();
692 
693   SE.forgetLoop(L);
694   SE.forgetBlockAndLoopDispositions();
695 
696   std::unique_ptr<MemorySSAUpdater> MSSAU;
697   if (MSSA)
698     MSSAU = std::make_unique<MemorySSAUpdater>(MSSA);
699 
700   // Update the CFG and domtree.  We chose to special case a couple of
701   // of common cases for code quality and test readability reasons.
702   [&]() -> void {
703     if (auto *BI = dyn_cast<BranchInst>(Latch->getTerminator())) {
704       if (!BI->isConditional()) {
705         DomTreeUpdater DTU(&DT, DomTreeUpdater::UpdateStrategy::Eager);
706         (void)changeToUnreachable(BI, /*PreserveLCSSA*/ true, &DTU,
707                                   MSSAU.get());
708         return;
709       }
710 
711       // Conditional latch/exit - note that latch can be shared by inner
712       // and outer loop so the other target doesn't need to an exit
713       if (L->isLoopExiting(Latch)) {
714         // TODO: Generalize ConstantFoldTerminator so that it can be used
715         // here without invalidating LCSSA or MemorySSA.  (Tricky case for
716         // LCSSA: header is an exit block of a preceeding sibling loop w/o
717         // dedicated exits.)
718         const unsigned ExitIdx = L->contains(BI->getSuccessor(0)) ? 1 : 0;
719         BasicBlock *ExitBB = BI->getSuccessor(ExitIdx);
720 
721         DomTreeUpdater DTU(&DT, DomTreeUpdater::UpdateStrategy::Eager);
722         Header->removePredecessor(Latch, true);
723 
724         IRBuilder<> Builder(BI);
725         auto *NewBI = Builder.CreateBr(ExitBB);
726         // Transfer the metadata to the new branch instruction (minus the
727         // loop info since this is no longer a loop)
728         NewBI->copyMetadata(*BI, {LLVMContext::MD_dbg,
729                                   LLVMContext::MD_annotation});
730 
731         BI->eraseFromParent();
732         DTU.applyUpdates({{DominatorTree::Delete, Latch, Header}});
733         if (MSSA)
734           MSSAU->applyUpdates({{DominatorTree::Delete, Latch, Header}}, DT);
735         return;
736       }
737     }
738 
739     // General case.  By splitting the backedge, and then explicitly making it
740     // unreachable we gracefully handle corner cases such as switch and invoke
741     // termiantors.
742     auto *BackedgeBB = SplitEdge(Latch, Header, &DT, &LI, MSSAU.get());
743 
744     DomTreeUpdater DTU(&DT, DomTreeUpdater::UpdateStrategy::Eager);
745     (void)changeToUnreachable(BackedgeBB->getTerminator(),
746                               /*PreserveLCSSA*/ true, &DTU, MSSAU.get());
747   }();
748 
749   // Erase (and destroy) this loop instance.  Handles relinking sub-loops
750   // and blocks within the loop as needed.
751   LI.erase(L);
752 
753   // If the loop we broke had a parent, then changeToUnreachable might have
754   // caused a block to be removed from the parent loop (see loop_nest_lcssa
755   // test case in zero-btc.ll for an example), thus changing the parent's
756   // exit blocks.  If that happened, we need to rebuild LCSSA on the outermost
757   // loop which might have a had a block removed.
758   if (OutermostLoop != L)
759     formLCSSARecursively(*OutermostLoop, DT, &LI, &SE);
760 }
761 
762 
763 /// Checks if \p L has an exiting latch branch.  There may also be other
764 /// exiting blocks.  Returns branch instruction terminating the loop
765 /// latch if above check is successful, nullptr otherwise.
766 static BranchInst *getExpectedExitLoopLatchBranch(Loop *L) {
767   BasicBlock *Latch = L->getLoopLatch();
768   if (!Latch)
769     return nullptr;
770 
771   BranchInst *LatchBR = dyn_cast<BranchInst>(Latch->getTerminator());
772   if (!LatchBR || LatchBR->getNumSuccessors() != 2 || !L->isLoopExiting(Latch))
773     return nullptr;
774 
775   assert((LatchBR->getSuccessor(0) == L->getHeader() ||
776           LatchBR->getSuccessor(1) == L->getHeader()) &&
777          "At least one edge out of the latch must go to the header");
778 
779   return LatchBR;
780 }
781 
782 /// Return the estimated trip count for any exiting branch which dominates
783 /// the loop latch.
784 static std::optional<uint64_t> getEstimatedTripCount(BranchInst *ExitingBranch,
785                                                      Loop *L,
786                                                      uint64_t &OrigExitWeight) {
787   // To estimate the number of times the loop body was executed, we want to
788   // know the number of times the backedge was taken, vs. the number of times
789   // we exited the loop.
790   uint64_t LoopWeight, ExitWeight;
791   if (!extractBranchWeights(*ExitingBranch, LoopWeight, ExitWeight))
792     return std::nullopt;
793 
794   if (L->contains(ExitingBranch->getSuccessor(1)))
795     std::swap(LoopWeight, ExitWeight);
796 
797   if (!ExitWeight)
798     // Don't have a way to return predicated infinite
799     return std::nullopt;
800 
801   OrigExitWeight = ExitWeight;
802 
803   // Estimated exit count is a ratio of the loop weight by the weight of the
804   // edge exiting the loop, rounded to nearest.
805   uint64_t ExitCount = llvm::divideNearest(LoopWeight, ExitWeight);
806   // Estimated trip count is one plus estimated exit count.
807   return ExitCount + 1;
808 }
809 
810 std::optional<unsigned>
811 llvm::getLoopEstimatedTripCount(Loop *L,
812                                 unsigned *EstimatedLoopInvocationWeight) {
813   // Currently we take the estimate exit count only from the loop latch,
814   // ignoring other exiting blocks.  This can overestimate the trip count
815   // if we exit through another exit, but can never underestimate it.
816   // TODO: incorporate information from other exits
817   if (BranchInst *LatchBranch = getExpectedExitLoopLatchBranch(L)) {
818     uint64_t ExitWeight;
819     if (std::optional<uint64_t> EstTripCount =
820             getEstimatedTripCount(LatchBranch, L, ExitWeight)) {
821       if (EstimatedLoopInvocationWeight)
822         *EstimatedLoopInvocationWeight = ExitWeight;
823       return *EstTripCount;
824     }
825   }
826   return std::nullopt;
827 }
828 
829 bool llvm::setLoopEstimatedTripCount(Loop *L, unsigned EstimatedTripCount,
830                                      unsigned EstimatedloopInvocationWeight) {
831   // At the moment, we currently support changing the estimate trip count of
832   // the latch branch only.  We could extend this API to manipulate estimated
833   // trip counts for any exit.
834   BranchInst *LatchBranch = getExpectedExitLoopLatchBranch(L);
835   if (!LatchBranch)
836     return false;
837 
838   // Calculate taken and exit weights.
839   unsigned LatchExitWeight = 0;
840   unsigned BackedgeTakenWeight = 0;
841 
842   if (EstimatedTripCount > 0) {
843     LatchExitWeight = EstimatedloopInvocationWeight;
844     BackedgeTakenWeight = (EstimatedTripCount - 1) * LatchExitWeight;
845   }
846 
847   // Make a swap if back edge is taken when condition is "false".
848   if (LatchBranch->getSuccessor(0) != L->getHeader())
849     std::swap(BackedgeTakenWeight, LatchExitWeight);
850 
851   MDBuilder MDB(LatchBranch->getContext());
852 
853   // Set/Update profile metadata.
854   LatchBranch->setMetadata(
855       LLVMContext::MD_prof,
856       MDB.createBranchWeights(BackedgeTakenWeight, LatchExitWeight));
857 
858   return true;
859 }
860 
861 bool llvm::hasIterationCountInvariantInParent(Loop *InnerLoop,
862                                               ScalarEvolution &SE) {
863   Loop *OuterL = InnerLoop->getParentLoop();
864   if (!OuterL)
865     return true;
866 
867   // Get the backedge taken count for the inner loop
868   BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
869   const SCEV *InnerLoopBECountSC = SE.getExitCount(InnerLoop, InnerLoopLatch);
870   if (isa<SCEVCouldNotCompute>(InnerLoopBECountSC) ||
871       !InnerLoopBECountSC->getType()->isIntegerTy())
872     return false;
873 
874   // Get whether count is invariant to the outer loop
875   ScalarEvolution::LoopDisposition LD =
876       SE.getLoopDisposition(InnerLoopBECountSC, OuterL);
877   if (LD != ScalarEvolution::LoopInvariant)
878     return false;
879 
880   return true;
881 }
882 
883 CmpInst::Predicate llvm::getMinMaxReductionPredicate(RecurKind RK) {
884   switch (RK) {
885   default:
886     llvm_unreachable("Unknown min/max recurrence kind");
887   case RecurKind::UMin:
888     return CmpInst::ICMP_ULT;
889   case RecurKind::UMax:
890     return CmpInst::ICMP_UGT;
891   case RecurKind::SMin:
892     return CmpInst::ICMP_SLT;
893   case RecurKind::SMax:
894     return CmpInst::ICMP_SGT;
895   case RecurKind::FMin:
896     return CmpInst::FCMP_OLT;
897   case RecurKind::FMax:
898     return CmpInst::FCMP_OGT;
899   }
900 }
901 
902 Value *llvm::createSelectCmpOp(IRBuilderBase &Builder, Value *StartVal,
903                                RecurKind RK, Value *Left, Value *Right) {
904   if (auto VTy = dyn_cast<VectorType>(Left->getType()))
905     StartVal = Builder.CreateVectorSplat(VTy->getElementCount(), StartVal);
906   Value *Cmp =
907       Builder.CreateCmp(CmpInst::ICMP_NE, Left, StartVal, "rdx.select.cmp");
908   return Builder.CreateSelect(Cmp, Left, Right, "rdx.select");
909 }
910 
911 Value *llvm::createMinMaxOp(IRBuilderBase &Builder, RecurKind RK, Value *Left,
912                             Value *Right) {
913   CmpInst::Predicate Pred = getMinMaxReductionPredicate(RK);
914   Value *Cmp = Builder.CreateCmp(Pred, Left, Right, "rdx.minmax.cmp");
915   Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
916   return Select;
917 }
918 
919 // Helper to generate an ordered reduction.
920 Value *llvm::getOrderedReduction(IRBuilderBase &Builder, Value *Acc, Value *Src,
921                                  unsigned Op, RecurKind RdxKind) {
922   unsigned VF = cast<FixedVectorType>(Src->getType())->getNumElements();
923 
924   // Extract and apply reduction ops in ascending order:
925   // e.g. ((((Acc + Scl[0]) + Scl[1]) + Scl[2]) + ) ... + Scl[VF-1]
926   Value *Result = Acc;
927   for (unsigned ExtractIdx = 0; ExtractIdx != VF; ++ExtractIdx) {
928     Value *Ext =
929         Builder.CreateExtractElement(Src, Builder.getInt32(ExtractIdx));
930 
931     if (Op != Instruction::ICmp && Op != Instruction::FCmp) {
932       Result = Builder.CreateBinOp((Instruction::BinaryOps)Op, Result, Ext,
933                                    "bin.rdx");
934     } else {
935       assert(RecurrenceDescriptor::isMinMaxRecurrenceKind(RdxKind) &&
936              "Invalid min/max");
937       Result = createMinMaxOp(Builder, RdxKind, Result, Ext);
938     }
939   }
940 
941   return Result;
942 }
943 
944 // Helper to generate a log2 shuffle reduction.
945 Value *llvm::getShuffleReduction(IRBuilderBase &Builder, Value *Src,
946                                  unsigned Op, RecurKind RdxKind) {
947   unsigned VF = cast<FixedVectorType>(Src->getType())->getNumElements();
948   // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
949   // and vector ops, reducing the set of values being computed by half each
950   // round.
951   assert(isPowerOf2_32(VF) &&
952          "Reduction emission only supported for pow2 vectors!");
953   // Note: fast-math-flags flags are controlled by the builder configuration
954   // and are assumed to apply to all generated arithmetic instructions.  Other
955   // poison generating flags (nsw/nuw/inbounds/inrange/exact) are not part
956   // of the builder configuration, and since they're not passed explicitly,
957   // will never be relevant here.  Note that it would be generally unsound to
958   // propagate these from an intrinsic call to the expansion anyways as we/
959   // change the order of operations.
960   Value *TmpVec = Src;
961   SmallVector<int, 32> ShuffleMask(VF);
962   for (unsigned i = VF; i != 1; i >>= 1) {
963     // Move the upper half of the vector to the lower half.
964     for (unsigned j = 0; j != i / 2; ++j)
965       ShuffleMask[j] = i / 2 + j;
966 
967     // Fill the rest of the mask with undef.
968     std::fill(&ShuffleMask[i / 2], ShuffleMask.end(), -1);
969 
970     Value *Shuf = Builder.CreateShuffleVector(TmpVec, ShuffleMask, "rdx.shuf");
971 
972     if (Op != Instruction::ICmp && Op != Instruction::FCmp) {
973       TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
974                                    "bin.rdx");
975     } else {
976       assert(RecurrenceDescriptor::isMinMaxRecurrenceKind(RdxKind) &&
977              "Invalid min/max");
978       TmpVec = createMinMaxOp(Builder, RdxKind, TmpVec, Shuf);
979     }
980   }
981   // The result is in the first element of the vector.
982   return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
983 }
984 
985 Value *llvm::createSelectCmpTargetReduction(IRBuilderBase &Builder,
986                                             const TargetTransformInfo *TTI,
987                                             Value *Src,
988                                             const RecurrenceDescriptor &Desc,
989                                             PHINode *OrigPhi) {
990   assert(RecurrenceDescriptor::isSelectCmpRecurrenceKind(
991              Desc.getRecurrenceKind()) &&
992          "Unexpected reduction kind");
993   Value *InitVal = Desc.getRecurrenceStartValue();
994   Value *NewVal = nullptr;
995 
996   // First use the original phi to determine the new value we're trying to
997   // select from in the loop.
998   SelectInst *SI = nullptr;
999   for (auto *U : OrigPhi->users()) {
1000     if ((SI = dyn_cast<SelectInst>(U)))
1001       break;
1002   }
1003   assert(SI && "One user of the original phi should be a select");
1004 
1005   if (SI->getTrueValue() == OrigPhi)
1006     NewVal = SI->getFalseValue();
1007   else {
1008     assert(SI->getFalseValue() == OrigPhi &&
1009            "At least one input to the select should be the original Phi");
1010     NewVal = SI->getTrueValue();
1011   }
1012 
1013   // Create a splat vector with the new value and compare this to the vector
1014   // we want to reduce.
1015   ElementCount EC = cast<VectorType>(Src->getType())->getElementCount();
1016   Value *Right = Builder.CreateVectorSplat(EC, InitVal);
1017   Value *Cmp =
1018       Builder.CreateCmp(CmpInst::ICMP_NE, Src, Right, "rdx.select.cmp");
1019 
1020   // If any predicate is true it means that we want to select the new value.
1021   Cmp = Builder.CreateOrReduce(Cmp);
1022   return Builder.CreateSelect(Cmp, NewVal, InitVal, "rdx.select");
1023 }
1024 
1025 Value *llvm::createSimpleTargetReduction(IRBuilderBase &Builder,
1026                                          const TargetTransformInfo *TTI,
1027                                          Value *Src, RecurKind RdxKind) {
1028   auto *SrcVecEltTy = cast<VectorType>(Src->getType())->getElementType();
1029   switch (RdxKind) {
1030   case RecurKind::Add:
1031     return Builder.CreateAddReduce(Src);
1032   case RecurKind::Mul:
1033     return Builder.CreateMulReduce(Src);
1034   case RecurKind::And:
1035     return Builder.CreateAndReduce(Src);
1036   case RecurKind::Or:
1037     return Builder.CreateOrReduce(Src);
1038   case RecurKind::Xor:
1039     return Builder.CreateXorReduce(Src);
1040   case RecurKind::FMulAdd:
1041   case RecurKind::FAdd:
1042     return Builder.CreateFAddReduce(ConstantFP::getNegativeZero(SrcVecEltTy),
1043                                     Src);
1044   case RecurKind::FMul:
1045     return Builder.CreateFMulReduce(ConstantFP::get(SrcVecEltTy, 1.0), Src);
1046   case RecurKind::SMax:
1047     return Builder.CreateIntMaxReduce(Src, true);
1048   case RecurKind::SMin:
1049     return Builder.CreateIntMinReduce(Src, true);
1050   case RecurKind::UMax:
1051     return Builder.CreateIntMaxReduce(Src, false);
1052   case RecurKind::UMin:
1053     return Builder.CreateIntMinReduce(Src, false);
1054   case RecurKind::FMax:
1055     return Builder.CreateFPMaxReduce(Src);
1056   case RecurKind::FMin:
1057     return Builder.CreateFPMinReduce(Src);
1058   default:
1059     llvm_unreachable("Unhandled opcode");
1060   }
1061 }
1062 
1063 Value *llvm::createTargetReduction(IRBuilderBase &B,
1064                                    const TargetTransformInfo *TTI,
1065                                    const RecurrenceDescriptor &Desc, Value *Src,
1066                                    PHINode *OrigPhi) {
1067   // TODO: Support in-order reductions based on the recurrence descriptor.
1068   // All ops in the reduction inherit fast-math-flags from the recurrence
1069   // descriptor.
1070   IRBuilderBase::FastMathFlagGuard FMFGuard(B);
1071   B.setFastMathFlags(Desc.getFastMathFlags());
1072 
1073   RecurKind RK = Desc.getRecurrenceKind();
1074   if (RecurrenceDescriptor::isSelectCmpRecurrenceKind(RK))
1075     return createSelectCmpTargetReduction(B, TTI, Src, Desc, OrigPhi);
1076 
1077   return createSimpleTargetReduction(B, TTI, Src, RK);
1078 }
1079 
1080 Value *llvm::createOrderedReduction(IRBuilderBase &B,
1081                                     const RecurrenceDescriptor &Desc,
1082                                     Value *Src, Value *Start) {
1083   assert((Desc.getRecurrenceKind() == RecurKind::FAdd ||
1084           Desc.getRecurrenceKind() == RecurKind::FMulAdd) &&
1085          "Unexpected reduction kind");
1086   assert(Src->getType()->isVectorTy() && "Expected a vector type");
1087   assert(!Start->getType()->isVectorTy() && "Expected a scalar type");
1088 
1089   return B.CreateFAddReduce(Start, Src);
1090 }
1091 
1092 void llvm::propagateIRFlags(Value *I, ArrayRef<Value *> VL, Value *OpValue,
1093                             bool IncludeWrapFlags) {
1094   auto *VecOp = dyn_cast<Instruction>(I);
1095   if (!VecOp)
1096     return;
1097   auto *Intersection = (OpValue == nullptr) ? dyn_cast<Instruction>(VL[0])
1098                                             : dyn_cast<Instruction>(OpValue);
1099   if (!Intersection)
1100     return;
1101   const unsigned Opcode = Intersection->getOpcode();
1102   VecOp->copyIRFlags(Intersection, IncludeWrapFlags);
1103   for (auto *V : VL) {
1104     auto *Instr = dyn_cast<Instruction>(V);
1105     if (!Instr)
1106       continue;
1107     if (OpValue == nullptr || Opcode == Instr->getOpcode())
1108       VecOp->andIRFlags(V);
1109   }
1110 }
1111 
1112 bool llvm::isKnownNegativeInLoop(const SCEV *S, const Loop *L,
1113                                  ScalarEvolution &SE) {
1114   const SCEV *Zero = SE.getZero(S->getType());
1115   return SE.isAvailableAtLoopEntry(S, L) &&
1116          SE.isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SLT, S, Zero);
1117 }
1118 
1119 bool llvm::isKnownNonNegativeInLoop(const SCEV *S, const Loop *L,
1120                                     ScalarEvolution &SE) {
1121   const SCEV *Zero = SE.getZero(S->getType());
1122   return SE.isAvailableAtLoopEntry(S, L) &&
1123          SE.isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SGE, S, Zero);
1124 }
1125 
1126 bool llvm::cannotBeMinInLoop(const SCEV *S, const Loop *L, ScalarEvolution &SE,
1127                              bool Signed) {
1128   unsigned BitWidth = cast<IntegerType>(S->getType())->getBitWidth();
1129   APInt Min = Signed ? APInt::getSignedMinValue(BitWidth) :
1130     APInt::getMinValue(BitWidth);
1131   auto Predicate = Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1132   return SE.isAvailableAtLoopEntry(S, L) &&
1133          SE.isLoopEntryGuardedByCond(L, Predicate, S,
1134                                      SE.getConstant(Min));
1135 }
1136 
1137 bool llvm::cannotBeMaxInLoop(const SCEV *S, const Loop *L, ScalarEvolution &SE,
1138                              bool Signed) {
1139   unsigned BitWidth = cast<IntegerType>(S->getType())->getBitWidth();
1140   APInt Max = Signed ? APInt::getSignedMaxValue(BitWidth) :
1141     APInt::getMaxValue(BitWidth);
1142   auto Predicate = Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1143   return SE.isAvailableAtLoopEntry(S, L) &&
1144          SE.isLoopEntryGuardedByCond(L, Predicate, S,
1145                                      SE.getConstant(Max));
1146 }
1147 
1148 //===----------------------------------------------------------------------===//
1149 // rewriteLoopExitValues - Optimize IV users outside the loop.
1150 // As a side effect, reduces the amount of IV processing within the loop.
1151 //===----------------------------------------------------------------------===//
1152 
1153 static bool hasHardUserWithinLoop(const Loop *L, const Instruction *I) {
1154   SmallPtrSet<const Instruction *, 8> Visited;
1155   SmallVector<const Instruction *, 8> WorkList;
1156   Visited.insert(I);
1157   WorkList.push_back(I);
1158   while (!WorkList.empty()) {
1159     const Instruction *Curr = WorkList.pop_back_val();
1160     // This use is outside the loop, nothing to do.
1161     if (!L->contains(Curr))
1162       continue;
1163     // Do we assume it is a "hard" use which will not be eliminated easily?
1164     if (Curr->mayHaveSideEffects())
1165       return true;
1166     // Otherwise, add all its users to worklist.
1167     for (const auto *U : Curr->users()) {
1168       auto *UI = cast<Instruction>(U);
1169       if (Visited.insert(UI).second)
1170         WorkList.push_back(UI);
1171     }
1172   }
1173   return false;
1174 }
1175 
1176 // Collect information about PHI nodes which can be transformed in
1177 // rewriteLoopExitValues.
1178 struct RewritePhi {
1179   PHINode *PN;               // For which PHI node is this replacement?
1180   unsigned Ith;              // For which incoming value?
1181   const SCEV *ExpansionSCEV; // The SCEV of the incoming value we are rewriting.
1182   Instruction *ExpansionPoint; // Where we'd like to expand that SCEV?
1183   bool HighCost;               // Is this expansion a high-cost?
1184 
1185   RewritePhi(PHINode *P, unsigned I, const SCEV *Val, Instruction *ExpansionPt,
1186              bool H)
1187       : PN(P), Ith(I), ExpansionSCEV(Val), ExpansionPoint(ExpansionPt),
1188         HighCost(H) {}
1189 };
1190 
1191 // Check whether it is possible to delete the loop after rewriting exit
1192 // value. If it is possible, ignore ReplaceExitValue and do rewriting
1193 // aggressively.
1194 static bool canLoopBeDeleted(Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet) {
1195   BasicBlock *Preheader = L->getLoopPreheader();
1196   // If there is no preheader, the loop will not be deleted.
1197   if (!Preheader)
1198     return false;
1199 
1200   // In LoopDeletion pass Loop can be deleted when ExitingBlocks.size() > 1.
1201   // We obviate multiple ExitingBlocks case for simplicity.
1202   // TODO: If we see testcase with multiple ExitingBlocks can be deleted
1203   // after exit value rewriting, we can enhance the logic here.
1204   SmallVector<BasicBlock *, 4> ExitingBlocks;
1205   L->getExitingBlocks(ExitingBlocks);
1206   SmallVector<BasicBlock *, 8> ExitBlocks;
1207   L->getUniqueExitBlocks(ExitBlocks);
1208   if (ExitBlocks.size() != 1 || ExitingBlocks.size() != 1)
1209     return false;
1210 
1211   BasicBlock *ExitBlock = ExitBlocks[0];
1212   BasicBlock::iterator BI = ExitBlock->begin();
1213   while (PHINode *P = dyn_cast<PHINode>(BI)) {
1214     Value *Incoming = P->getIncomingValueForBlock(ExitingBlocks[0]);
1215 
1216     // If the Incoming value of P is found in RewritePhiSet, we know it
1217     // could be rewritten to use a loop invariant value in transformation
1218     // phase later. Skip it in the loop invariant check below.
1219     bool found = false;
1220     for (const RewritePhi &Phi : RewritePhiSet) {
1221       unsigned i = Phi.Ith;
1222       if (Phi.PN == P && (Phi.PN)->getIncomingValue(i) == Incoming) {
1223         found = true;
1224         break;
1225       }
1226     }
1227 
1228     Instruction *I;
1229     if (!found && (I = dyn_cast<Instruction>(Incoming)))
1230       if (!L->hasLoopInvariantOperands(I))
1231         return false;
1232 
1233     ++BI;
1234   }
1235 
1236   for (auto *BB : L->blocks())
1237     if (llvm::any_of(*BB, [](Instruction &I) {
1238           return I.mayHaveSideEffects();
1239         }))
1240       return false;
1241 
1242   return true;
1243 }
1244 
1245 /// Checks if it is safe to call InductionDescriptor::isInductionPHI for \p Phi,
1246 /// and returns true if this Phi is an induction phi in the loop. When
1247 /// isInductionPHI returns true, \p ID will be also be set by isInductionPHI.
1248 static bool checkIsIndPhi(PHINode *Phi, Loop *L, ScalarEvolution *SE,
1249                           InductionDescriptor &ID) {
1250   if (!Phi)
1251     return false;
1252   if (!L->getLoopPreheader())
1253     return false;
1254   if (Phi->getParent() != L->getHeader())
1255     return false;
1256   return InductionDescriptor::isInductionPHI(Phi, L, SE, ID);
1257 }
1258 
1259 int llvm::rewriteLoopExitValues(Loop *L, LoopInfo *LI, TargetLibraryInfo *TLI,
1260                                 ScalarEvolution *SE,
1261                                 const TargetTransformInfo *TTI,
1262                                 SCEVExpander &Rewriter, DominatorTree *DT,
1263                                 ReplaceExitVal ReplaceExitValue,
1264                                 SmallVector<WeakTrackingVH, 16> &DeadInsts) {
1265   // Check a pre-condition.
1266   assert(L->isRecursivelyLCSSAForm(*DT, *LI) &&
1267          "Indvars did not preserve LCSSA!");
1268 
1269   SmallVector<BasicBlock*, 8> ExitBlocks;
1270   L->getUniqueExitBlocks(ExitBlocks);
1271 
1272   SmallVector<RewritePhi, 8> RewritePhiSet;
1273   // Find all values that are computed inside the loop, but used outside of it.
1274   // Because of LCSSA, these values will only occur in LCSSA PHI Nodes.  Scan
1275   // the exit blocks of the loop to find them.
1276   for (BasicBlock *ExitBB : ExitBlocks) {
1277     // If there are no PHI nodes in this exit block, then no values defined
1278     // inside the loop are used on this path, skip it.
1279     PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
1280     if (!PN) continue;
1281 
1282     unsigned NumPreds = PN->getNumIncomingValues();
1283 
1284     // Iterate over all of the PHI nodes.
1285     BasicBlock::iterator BBI = ExitBB->begin();
1286     while ((PN = dyn_cast<PHINode>(BBI++))) {
1287       if (PN->use_empty())
1288         continue; // dead use, don't replace it
1289 
1290       if (!SE->isSCEVable(PN->getType()))
1291         continue;
1292 
1293       // Iterate over all of the values in all the PHI nodes.
1294       for (unsigned i = 0; i != NumPreds; ++i) {
1295         // If the value being merged in is not integer or is not defined
1296         // in the loop, skip it.
1297         Value *InVal = PN->getIncomingValue(i);
1298         if (!isa<Instruction>(InVal))
1299           continue;
1300 
1301         // If this pred is for a subloop, not L itself, skip it.
1302         if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
1303           continue; // The Block is in a subloop, skip it.
1304 
1305         // Check that InVal is defined in the loop.
1306         Instruction *Inst = cast<Instruction>(InVal);
1307         if (!L->contains(Inst))
1308           continue;
1309 
1310         // Find exit values which are induction variables in the loop, and are
1311         // unused in the loop, with the only use being the exit block PhiNode,
1312         // and the induction variable update binary operator.
1313         // The exit value can be replaced with the final value when it is cheap
1314         // to do so.
1315         if (ReplaceExitValue == UnusedIndVarInLoop) {
1316           InductionDescriptor ID;
1317           PHINode *IndPhi = dyn_cast<PHINode>(Inst);
1318           if (IndPhi) {
1319             if (!checkIsIndPhi(IndPhi, L, SE, ID))
1320               continue;
1321             // This is an induction PHI. Check that the only users are PHI
1322             // nodes, and induction variable update binary operators.
1323             if (llvm::any_of(Inst->users(), [&](User *U) {
1324                   if (!isa<PHINode>(U) && !isa<BinaryOperator>(U))
1325                     return true;
1326                   BinaryOperator *B = dyn_cast<BinaryOperator>(U);
1327                   if (B && B != ID.getInductionBinOp())
1328                     return true;
1329                   return false;
1330                 }))
1331               continue;
1332           } else {
1333             // If it is not an induction phi, it must be an induction update
1334             // binary operator with an induction phi user.
1335             BinaryOperator *B = dyn_cast<BinaryOperator>(Inst);
1336             if (!B)
1337               continue;
1338             if (llvm::any_of(Inst->users(), [&](User *U) {
1339                   PHINode *Phi = dyn_cast<PHINode>(U);
1340                   if (Phi != PN && !checkIsIndPhi(Phi, L, SE, ID))
1341                     return true;
1342                   return false;
1343                 }))
1344               continue;
1345             if (B != ID.getInductionBinOp())
1346               continue;
1347           }
1348         }
1349 
1350         // Okay, this instruction has a user outside of the current loop
1351         // and varies predictably *inside* the loop.  Evaluate the value it
1352         // contains when the loop exits, if possible.  We prefer to start with
1353         // expressions which are true for all exits (so as to maximize
1354         // expression reuse by the SCEVExpander), but resort to per-exit
1355         // evaluation if that fails.
1356         const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
1357         if (isa<SCEVCouldNotCompute>(ExitValue) ||
1358             !SE->isLoopInvariant(ExitValue, L) ||
1359             !Rewriter.isSafeToExpand(ExitValue)) {
1360           // TODO: This should probably be sunk into SCEV in some way; maybe a
1361           // getSCEVForExit(SCEV*, L, ExitingBB)?  It can be generalized for
1362           // most SCEV expressions and other recurrence types (e.g. shift
1363           // recurrences).  Is there existing code we can reuse?
1364           const SCEV *ExitCount = SE->getExitCount(L, PN->getIncomingBlock(i));
1365           if (isa<SCEVCouldNotCompute>(ExitCount))
1366             continue;
1367           if (auto *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Inst)))
1368             if (AddRec->getLoop() == L)
1369               ExitValue = AddRec->evaluateAtIteration(ExitCount, *SE);
1370           if (isa<SCEVCouldNotCompute>(ExitValue) ||
1371               !SE->isLoopInvariant(ExitValue, L) ||
1372               !Rewriter.isSafeToExpand(ExitValue))
1373             continue;
1374         }
1375 
1376         // Computing the value outside of the loop brings no benefit if it is
1377         // definitely used inside the loop in a way which can not be optimized
1378         // away. Avoid doing so unless we know we have a value which computes
1379         // the ExitValue already. TODO: This should be merged into SCEV
1380         // expander to leverage its knowledge of existing expressions.
1381         if (ReplaceExitValue != AlwaysRepl && !isa<SCEVConstant>(ExitValue) &&
1382             !isa<SCEVUnknown>(ExitValue) && hasHardUserWithinLoop(L, Inst))
1383           continue;
1384 
1385         // Check if expansions of this SCEV would count as being high cost.
1386         bool HighCost = Rewriter.isHighCostExpansion(
1387             ExitValue, L, SCEVCheapExpansionBudget, TTI, Inst);
1388 
1389         // Note that we must not perform expansions until after
1390         // we query *all* the costs, because if we perform temporary expansion
1391         // inbetween, one that we might not intend to keep, said expansion
1392         // *may* affect cost calculation of the the next SCEV's we'll query,
1393         // and next SCEV may errneously get smaller cost.
1394 
1395         // Collect all the candidate PHINodes to be rewritten.
1396         Instruction *InsertPt =
1397           (isa<PHINode>(Inst) || isa<LandingPadInst>(Inst)) ?
1398           &*Inst->getParent()->getFirstInsertionPt() : Inst;
1399         RewritePhiSet.emplace_back(PN, i, ExitValue, InsertPt, HighCost);
1400       }
1401     }
1402   }
1403 
1404   // TODO: evaluate whether it is beneficial to change how we calculate
1405   // high-cost: if we have SCEV 'A' which we know we will expand, should we
1406   // calculate the cost of other SCEV's after expanding SCEV 'A', thus
1407   // potentially giving cost bonus to those other SCEV's?
1408 
1409   bool LoopCanBeDel = canLoopBeDeleted(L, RewritePhiSet);
1410   int NumReplaced = 0;
1411 
1412   // Transformation.
1413   for (const RewritePhi &Phi : RewritePhiSet) {
1414     PHINode *PN = Phi.PN;
1415 
1416     // Only do the rewrite when the ExitValue can be expanded cheaply.
1417     // If LoopCanBeDel is true, rewrite exit value aggressively.
1418     if ((ReplaceExitValue == OnlyCheapRepl ||
1419          ReplaceExitValue == UnusedIndVarInLoop) &&
1420         !LoopCanBeDel && Phi.HighCost)
1421       continue;
1422 
1423     Value *ExitVal = Rewriter.expandCodeFor(
1424         Phi.ExpansionSCEV, Phi.PN->getType(), Phi.ExpansionPoint);
1425 
1426     LLVM_DEBUG(dbgs() << "rewriteLoopExitValues: AfterLoopVal = " << *ExitVal
1427                       << '\n'
1428                       << "  LoopVal = " << *(Phi.ExpansionPoint) << "\n");
1429 
1430 #ifndef NDEBUG
1431     // If we reuse an instruction from a loop which is neither L nor one of
1432     // its containing loops, we end up breaking LCSSA form for this loop by
1433     // creating a new use of its instruction.
1434     if (auto *ExitInsn = dyn_cast<Instruction>(ExitVal))
1435       if (auto *EVL = LI->getLoopFor(ExitInsn->getParent()))
1436         if (EVL != L)
1437           assert(EVL->contains(L) && "LCSSA breach detected!");
1438 #endif
1439 
1440     NumReplaced++;
1441     Instruction *Inst = cast<Instruction>(PN->getIncomingValue(Phi.Ith));
1442     PN->setIncomingValue(Phi.Ith, ExitVal);
1443     // It's necessary to tell ScalarEvolution about this explicitly so that
1444     // it can walk the def-use list and forget all SCEVs, as it may not be
1445     // watching the PHI itself. Once the new exit value is in place, there
1446     // may not be a def-use connection between the loop and every instruction
1447     // which got a SCEVAddRecExpr for that loop.
1448     SE->forgetValue(PN);
1449 
1450     // If this instruction is dead now, delete it. Don't do it now to avoid
1451     // invalidating iterators.
1452     if (isInstructionTriviallyDead(Inst, TLI))
1453       DeadInsts.push_back(Inst);
1454 
1455     // Replace PN with ExitVal if that is legal and does not break LCSSA.
1456     if (PN->getNumIncomingValues() == 1 &&
1457         LI->replacementPreservesLCSSAForm(PN, ExitVal)) {
1458       PN->replaceAllUsesWith(ExitVal);
1459       PN->eraseFromParent();
1460     }
1461   }
1462 
1463   // The insertion point instruction may have been deleted; clear it out
1464   // so that the rewriter doesn't trip over it later.
1465   Rewriter.clearInsertPoint();
1466   return NumReplaced;
1467 }
1468 
1469 /// Set weights for \p UnrolledLoop and \p RemainderLoop based on weights for
1470 /// \p OrigLoop.
1471 void llvm::setProfileInfoAfterUnrolling(Loop *OrigLoop, Loop *UnrolledLoop,
1472                                         Loop *RemainderLoop, uint64_t UF) {
1473   assert(UF > 0 && "Zero unrolled factor is not supported");
1474   assert(UnrolledLoop != RemainderLoop &&
1475          "Unrolled and Remainder loops are expected to distinct");
1476 
1477   // Get number of iterations in the original scalar loop.
1478   unsigned OrigLoopInvocationWeight = 0;
1479   std::optional<unsigned> OrigAverageTripCount =
1480       getLoopEstimatedTripCount(OrigLoop, &OrigLoopInvocationWeight);
1481   if (!OrigAverageTripCount)
1482     return;
1483 
1484   // Calculate number of iterations in unrolled loop.
1485   unsigned UnrolledAverageTripCount = *OrigAverageTripCount / UF;
1486   // Calculate number of iterations for remainder loop.
1487   unsigned RemainderAverageTripCount = *OrigAverageTripCount % UF;
1488 
1489   setLoopEstimatedTripCount(UnrolledLoop, UnrolledAverageTripCount,
1490                             OrigLoopInvocationWeight);
1491   setLoopEstimatedTripCount(RemainderLoop, RemainderAverageTripCount,
1492                             OrigLoopInvocationWeight);
1493 }
1494 
1495 /// Utility that implements appending of loops onto a worklist.
1496 /// Loops are added in preorder (analogous for reverse postorder for trees),
1497 /// and the worklist is processed LIFO.
1498 template <typename RangeT>
1499 void llvm::appendReversedLoopsToWorklist(
1500     RangeT &&Loops, SmallPriorityWorklist<Loop *, 4> &Worklist) {
1501   // We use an internal worklist to build up the preorder traversal without
1502   // recursion.
1503   SmallVector<Loop *, 4> PreOrderLoops, PreOrderWorklist;
1504 
1505   // We walk the initial sequence of loops in reverse because we generally want
1506   // to visit defs before uses and the worklist is LIFO.
1507   for (Loop *RootL : Loops) {
1508     assert(PreOrderLoops.empty() && "Must start with an empty preorder walk.");
1509     assert(PreOrderWorklist.empty() &&
1510            "Must start with an empty preorder walk worklist.");
1511     PreOrderWorklist.push_back(RootL);
1512     do {
1513       Loop *L = PreOrderWorklist.pop_back_val();
1514       PreOrderWorklist.append(L->begin(), L->end());
1515       PreOrderLoops.push_back(L);
1516     } while (!PreOrderWorklist.empty());
1517 
1518     Worklist.insert(std::move(PreOrderLoops));
1519     PreOrderLoops.clear();
1520   }
1521 }
1522 
1523 template <typename RangeT>
1524 void llvm::appendLoopsToWorklist(RangeT &&Loops,
1525                                  SmallPriorityWorklist<Loop *, 4> &Worklist) {
1526   appendReversedLoopsToWorklist(reverse(Loops), Worklist);
1527 }
1528 
1529 template void llvm::appendLoopsToWorklist<ArrayRef<Loop *> &>(
1530     ArrayRef<Loop *> &Loops, SmallPriorityWorklist<Loop *, 4> &Worklist);
1531 
1532 template void
1533 llvm::appendLoopsToWorklist<Loop &>(Loop &L,
1534                                     SmallPriorityWorklist<Loop *, 4> &Worklist);
1535 
1536 void llvm::appendLoopsToWorklist(LoopInfo &LI,
1537                                  SmallPriorityWorklist<Loop *, 4> &Worklist) {
1538   appendReversedLoopsToWorklist(LI, Worklist);
1539 }
1540 
1541 Loop *llvm::cloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM,
1542                       LoopInfo *LI, LPPassManager *LPM) {
1543   Loop &New = *LI->AllocateLoop();
1544   if (PL)
1545     PL->addChildLoop(&New);
1546   else
1547     LI->addTopLevelLoop(&New);
1548 
1549   if (LPM)
1550     LPM->addLoop(New);
1551 
1552   // Add all of the blocks in L to the new loop.
1553   for (BasicBlock *BB : L->blocks())
1554     if (LI->getLoopFor(BB) == L)
1555       New.addBasicBlockToLoop(cast<BasicBlock>(VM[BB]), *LI);
1556 
1557   // Add all of the subloops to the new loop.
1558   for (Loop *I : *L)
1559     cloneLoop(I, &New, VM, LI, LPM);
1560 
1561   return &New;
1562 }
1563 
1564 /// IR Values for the lower and upper bounds of a pointer evolution.  We
1565 /// need to use value-handles because SCEV expansion can invalidate previously
1566 /// expanded values.  Thus expansion of a pointer can invalidate the bounds for
1567 /// a previous one.
1568 struct PointerBounds {
1569   TrackingVH<Value> Start;
1570   TrackingVH<Value> End;
1571 };
1572 
1573 /// Expand code for the lower and upper bound of the pointer group \p CG
1574 /// in \p TheLoop.  \return the values for the bounds.
1575 static PointerBounds expandBounds(const RuntimeCheckingPtrGroup *CG,
1576                                   Loop *TheLoop, Instruction *Loc,
1577                                   SCEVExpander &Exp) {
1578   LLVMContext &Ctx = Loc->getContext();
1579   Type *PtrArithTy = Type::getInt8PtrTy(Ctx, CG->AddressSpace);
1580 
1581   Value *Start = nullptr, *End = nullptr;
1582   LLVM_DEBUG(dbgs() << "LAA: Adding RT check for range:\n");
1583   Start = Exp.expandCodeFor(CG->Low, PtrArithTy, Loc);
1584   End = Exp.expandCodeFor(CG->High, PtrArithTy, Loc);
1585   if (CG->NeedsFreeze) {
1586     IRBuilder<> Builder(Loc);
1587     Start = Builder.CreateFreeze(Start, Start->getName() + ".fr");
1588     End = Builder.CreateFreeze(End, End->getName() + ".fr");
1589   }
1590   LLVM_DEBUG(dbgs() << "Start: " << *CG->Low << " End: " << *CG->High << "\n");
1591   return {Start, End};
1592 }
1593 
1594 /// Turns a collection of checks into a collection of expanded upper and
1595 /// lower bounds for both pointers in the check.
1596 static SmallVector<std::pair<PointerBounds, PointerBounds>, 4>
1597 expandBounds(const SmallVectorImpl<RuntimePointerCheck> &PointerChecks, Loop *L,
1598              Instruction *Loc, SCEVExpander &Exp) {
1599   SmallVector<std::pair<PointerBounds, PointerBounds>, 4> ChecksWithBounds;
1600 
1601   // Here we're relying on the SCEV Expander's cache to only emit code for the
1602   // same bounds once.
1603   transform(PointerChecks, std::back_inserter(ChecksWithBounds),
1604             [&](const RuntimePointerCheck &Check) {
1605               PointerBounds First = expandBounds(Check.first, L, Loc, Exp),
1606                             Second = expandBounds(Check.second, L, Loc, Exp);
1607               return std::make_pair(First, Second);
1608             });
1609 
1610   return ChecksWithBounds;
1611 }
1612 
1613 Value *llvm::addRuntimeChecks(
1614     Instruction *Loc, Loop *TheLoop,
1615     const SmallVectorImpl<RuntimePointerCheck> &PointerChecks,
1616     SCEVExpander &Exp) {
1617   // TODO: Move noalias annotation code from LoopVersioning here and share with LV if possible.
1618   // TODO: Pass  RtPtrChecking instead of PointerChecks and SE separately, if possible
1619   auto ExpandedChecks = expandBounds(PointerChecks, TheLoop, Loc, Exp);
1620 
1621   LLVMContext &Ctx = Loc->getContext();
1622   IRBuilder<InstSimplifyFolder> ChkBuilder(Ctx,
1623                                            Loc->getModule()->getDataLayout());
1624   ChkBuilder.SetInsertPoint(Loc);
1625   // Our instructions might fold to a constant.
1626   Value *MemoryRuntimeCheck = nullptr;
1627 
1628   for (const auto &Check : ExpandedChecks) {
1629     const PointerBounds &A = Check.first, &B = Check.second;
1630     // Check if two pointers (A and B) conflict where conflict is computed as:
1631     // start(A) <= end(B) && start(B) <= end(A)
1632     unsigned AS0 = A.Start->getType()->getPointerAddressSpace();
1633     unsigned AS1 = B.Start->getType()->getPointerAddressSpace();
1634 
1635     assert((AS0 == B.End->getType()->getPointerAddressSpace()) &&
1636            (AS1 == A.End->getType()->getPointerAddressSpace()) &&
1637            "Trying to bounds check pointers with different address spaces");
1638 
1639     Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1640     Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1641 
1642     Value *Start0 = ChkBuilder.CreateBitCast(A.Start, PtrArithTy0, "bc");
1643     Value *Start1 = ChkBuilder.CreateBitCast(B.Start, PtrArithTy1, "bc");
1644     Value *End0 = ChkBuilder.CreateBitCast(A.End, PtrArithTy1, "bc");
1645     Value *End1 = ChkBuilder.CreateBitCast(B.End, PtrArithTy0, "bc");
1646 
1647     // [A|B].Start points to the first accessed byte under base [A|B].
1648     // [A|B].End points to the last accessed byte, plus one.
1649     // There is no conflict when the intervals are disjoint:
1650     // NoConflict = (B.Start >= A.End) || (A.Start >= B.End)
1651     //
1652     // bound0 = (B.Start < A.End)
1653     // bound1 = (A.Start < B.End)
1654     //  IsConflict = bound0 & bound1
1655     Value *Cmp0 = ChkBuilder.CreateICmpULT(Start0, End1, "bound0");
1656     Value *Cmp1 = ChkBuilder.CreateICmpULT(Start1, End0, "bound1");
1657     Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1658     if (MemoryRuntimeCheck) {
1659       IsConflict =
1660           ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict, "conflict.rdx");
1661     }
1662     MemoryRuntimeCheck = IsConflict;
1663   }
1664 
1665   return MemoryRuntimeCheck;
1666 }
1667 
1668 Value *llvm::addDiffRuntimeChecks(
1669     Instruction *Loc, ArrayRef<PointerDiffInfo> Checks, SCEVExpander &Expander,
1670     function_ref<Value *(IRBuilderBase &, unsigned)> GetVF, unsigned IC) {
1671 
1672   LLVMContext &Ctx = Loc->getContext();
1673   IRBuilder<InstSimplifyFolder> ChkBuilder(Ctx,
1674                                            Loc->getModule()->getDataLayout());
1675   ChkBuilder.SetInsertPoint(Loc);
1676   // Our instructions might fold to a constant.
1677   Value *MemoryRuntimeCheck = nullptr;
1678 
1679   for (const auto &C : Checks) {
1680     Type *Ty = C.SinkStart->getType();
1681     // Compute VF * IC * AccessSize.
1682     auto *VFTimesUFTimesSize =
1683         ChkBuilder.CreateMul(GetVF(ChkBuilder, Ty->getScalarSizeInBits()),
1684                              ConstantInt::get(Ty, IC * C.AccessSize));
1685     Value *Sink = Expander.expandCodeFor(C.SinkStart, Ty, Loc);
1686     Value *Src = Expander.expandCodeFor(C.SrcStart, Ty, Loc);
1687     if (C.NeedsFreeze) {
1688       IRBuilder<> Builder(Loc);
1689       Sink = Builder.CreateFreeze(Sink, Sink->getName() + ".fr");
1690       Src = Builder.CreateFreeze(Src, Src->getName() + ".fr");
1691     }
1692     Value *Diff = ChkBuilder.CreateSub(Sink, Src);
1693     Value *IsConflict =
1694         ChkBuilder.CreateICmpULT(Diff, VFTimesUFTimesSize, "diff.check");
1695 
1696     if (MemoryRuntimeCheck) {
1697       IsConflict =
1698           ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict, "conflict.rdx");
1699     }
1700     MemoryRuntimeCheck = IsConflict;
1701   }
1702 
1703   return MemoryRuntimeCheck;
1704 }
1705 
1706 std::optional<IVConditionInfo>
1707 llvm::hasPartialIVCondition(const Loop &L, unsigned MSSAThreshold,
1708                             const MemorySSA &MSSA, AAResults &AA) {
1709   auto *TI = dyn_cast<BranchInst>(L.getHeader()->getTerminator());
1710   if (!TI || !TI->isConditional())
1711     return {};
1712 
1713   auto *CondI = dyn_cast<CmpInst>(TI->getCondition());
1714   // The case with the condition outside the loop should already be handled
1715   // earlier.
1716   if (!CondI || !L.contains(CondI))
1717     return {};
1718 
1719   SmallVector<Instruction *> InstToDuplicate;
1720   InstToDuplicate.push_back(CondI);
1721 
1722   SmallVector<Value *, 4> WorkList;
1723   WorkList.append(CondI->op_begin(), CondI->op_end());
1724 
1725   SmallVector<MemoryAccess *, 4> AccessesToCheck;
1726   SmallVector<MemoryLocation, 4> AccessedLocs;
1727   while (!WorkList.empty()) {
1728     Instruction *I = dyn_cast<Instruction>(WorkList.pop_back_val());
1729     if (!I || !L.contains(I))
1730       continue;
1731 
1732     // TODO: support additional instructions.
1733     if (!isa<LoadInst>(I) && !isa<GetElementPtrInst>(I))
1734       return {};
1735 
1736     // Do not duplicate volatile and atomic loads.
1737     if (auto *LI = dyn_cast<LoadInst>(I))
1738       if (LI->isVolatile() || LI->isAtomic())
1739         return {};
1740 
1741     InstToDuplicate.push_back(I);
1742     if (MemoryAccess *MA = MSSA.getMemoryAccess(I)) {
1743       if (auto *MemUse = dyn_cast_or_null<MemoryUse>(MA)) {
1744         // Queue the defining access to check for alias checks.
1745         AccessesToCheck.push_back(MemUse->getDefiningAccess());
1746         AccessedLocs.push_back(MemoryLocation::get(I));
1747       } else {
1748         // MemoryDefs may clobber the location or may be atomic memory
1749         // operations. Bail out.
1750         return {};
1751       }
1752     }
1753     WorkList.append(I->op_begin(), I->op_end());
1754   }
1755 
1756   if (InstToDuplicate.empty())
1757     return {};
1758 
1759   SmallVector<BasicBlock *, 4> ExitingBlocks;
1760   L.getExitingBlocks(ExitingBlocks);
1761   auto HasNoClobbersOnPath =
1762       [&L, &AA, &AccessedLocs, &ExitingBlocks, &InstToDuplicate,
1763        MSSAThreshold](BasicBlock *Succ, BasicBlock *Header,
1764                       SmallVector<MemoryAccess *, 4> AccessesToCheck)
1765       -> std::optional<IVConditionInfo> {
1766     IVConditionInfo Info;
1767     // First, collect all blocks in the loop that are on a patch from Succ
1768     // to the header.
1769     SmallVector<BasicBlock *, 4> WorkList;
1770     WorkList.push_back(Succ);
1771     WorkList.push_back(Header);
1772     SmallPtrSet<BasicBlock *, 4> Seen;
1773     Seen.insert(Header);
1774     Info.PathIsNoop &=
1775         all_of(*Header, [](Instruction &I) { return !I.mayHaveSideEffects(); });
1776 
1777     while (!WorkList.empty()) {
1778       BasicBlock *Current = WorkList.pop_back_val();
1779       if (!L.contains(Current))
1780         continue;
1781       const auto &SeenIns = Seen.insert(Current);
1782       if (!SeenIns.second)
1783         continue;
1784 
1785       Info.PathIsNoop &= all_of(
1786           *Current, [](Instruction &I) { return !I.mayHaveSideEffects(); });
1787       WorkList.append(succ_begin(Current), succ_end(Current));
1788     }
1789 
1790     // Require at least 2 blocks on a path through the loop. This skips
1791     // paths that directly exit the loop.
1792     if (Seen.size() < 2)
1793       return {};
1794 
1795     // Next, check if there are any MemoryDefs that are on the path through
1796     // the loop (in the Seen set) and they may-alias any of the locations in
1797     // AccessedLocs. If that is the case, they may modify the condition and
1798     // partial unswitching is not possible.
1799     SmallPtrSet<MemoryAccess *, 4> SeenAccesses;
1800     while (!AccessesToCheck.empty()) {
1801       MemoryAccess *Current = AccessesToCheck.pop_back_val();
1802       auto SeenI = SeenAccesses.insert(Current);
1803       if (!SeenI.second || !Seen.contains(Current->getBlock()))
1804         continue;
1805 
1806       // Bail out if exceeded the threshold.
1807       if (SeenAccesses.size() >= MSSAThreshold)
1808         return {};
1809 
1810       // MemoryUse are read-only accesses.
1811       if (isa<MemoryUse>(Current))
1812         continue;
1813 
1814       // For a MemoryDef, check if is aliases any of the location feeding
1815       // the original condition.
1816       if (auto *CurrentDef = dyn_cast<MemoryDef>(Current)) {
1817         if (any_of(AccessedLocs, [&AA, CurrentDef](MemoryLocation &Loc) {
1818               return isModSet(
1819                   AA.getModRefInfo(CurrentDef->getMemoryInst(), Loc));
1820             }))
1821           return {};
1822       }
1823 
1824       for (Use &U : Current->uses())
1825         AccessesToCheck.push_back(cast<MemoryAccess>(U.getUser()));
1826     }
1827 
1828     // We could also allow loops with known trip counts without mustprogress,
1829     // but ScalarEvolution may not be available.
1830     Info.PathIsNoop &= isMustProgress(&L);
1831 
1832     // If the path is considered a no-op so far, check if it reaches a
1833     // single exit block without any phis. This ensures no values from the
1834     // loop are used outside of the loop.
1835     if (Info.PathIsNoop) {
1836       for (auto *Exiting : ExitingBlocks) {
1837         if (!Seen.contains(Exiting))
1838           continue;
1839         for (auto *Succ : successors(Exiting)) {
1840           if (L.contains(Succ))
1841             continue;
1842 
1843           Info.PathIsNoop &= Succ->phis().empty() &&
1844                              (!Info.ExitForPath || Info.ExitForPath == Succ);
1845           if (!Info.PathIsNoop)
1846             break;
1847           assert((!Info.ExitForPath || Info.ExitForPath == Succ) &&
1848                  "cannot have multiple exit blocks");
1849           Info.ExitForPath = Succ;
1850         }
1851       }
1852     }
1853     if (!Info.ExitForPath)
1854       Info.PathIsNoop = false;
1855 
1856     Info.InstToDuplicate = InstToDuplicate;
1857     return Info;
1858   };
1859 
1860   // If we branch to the same successor, partial unswitching will not be
1861   // beneficial.
1862   if (TI->getSuccessor(0) == TI->getSuccessor(1))
1863     return {};
1864 
1865   if (auto Info = HasNoClobbersOnPath(TI->getSuccessor(0), L.getHeader(),
1866                                       AccessesToCheck)) {
1867     Info->KnownValue = ConstantInt::getTrue(TI->getContext());
1868     return Info;
1869   }
1870   if (auto Info = HasNoClobbersOnPath(TI->getSuccessor(1), L.getHeader(),
1871                                       AccessesToCheck)) {
1872     Info->KnownValue = ConstantInt::getFalse(TI->getContext());
1873     return Info;
1874   }
1875 
1876   return {};
1877 }
1878