xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Utils/LoopUtils.cpp (revision 18054d0220cfc8df9c9568c437bd6fbb59d53c3c)
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/InstSimplifyFolder.h"
26 #include "llvm/Analysis/InstructionSimplify.h"
27 #include "llvm/Analysis/LoopAccessAnalysis.h"
28 #include "llvm/Analysis/LoopInfo.h"
29 #include "llvm/Analysis/LoopPass.h"
30 #include "llvm/Analysis/MemorySSA.h"
31 #include "llvm/Analysis/MemorySSAUpdater.h"
32 #include "llvm/Analysis/MustExecute.h"
33 #include "llvm/Analysis/ScalarEvolution.h"
34 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
35 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
36 #include "llvm/Analysis/TargetTransformInfo.h"
37 #include "llvm/Analysis/ValueTracking.h"
38 #include "llvm/IR/DIBuilder.h"
39 #include "llvm/IR/Dominators.h"
40 #include "llvm/IR/Instructions.h"
41 #include "llvm/IR/IntrinsicInst.h"
42 #include "llvm/IR/MDBuilder.h"
43 #include "llvm/IR/Module.h"
44 #include "llvm/IR/Operator.h"
45 #include "llvm/IR/PatternMatch.h"
46 #include "llvm/IR/ValueHandle.h"
47 #include "llvm/InitializePasses.h"
48 #include "llvm/Pass.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/KnownBits.h"
51 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
52 #include "llvm/Transforms/Utils/Local.h"
53 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
54 
55 using namespace llvm;
56 using namespace llvm::PatternMatch;
57 
58 #define DEBUG_TYPE "loop-utils"
59 
60 static const char *LLVMLoopDisableNonforced = "llvm.loop.disable_nonforced";
61 static const char *LLVMLoopDisableLICM = "llvm.licm.disable";
62 
63 bool llvm::formDedicatedExitBlocks(Loop *L, DominatorTree *DT, LoopInfo *LI,
64                                    MemorySSAUpdater *MSSAU,
65                                    bool PreserveLCSSA) {
66   bool Changed = false;
67 
68   // We re-use a vector for the in-loop predecesosrs.
69   SmallVector<BasicBlock *, 4> InLoopPredecessors;
70 
71   auto RewriteExit = [&](BasicBlock *BB) {
72     assert(InLoopPredecessors.empty() &&
73            "Must start with an empty predecessors list!");
74     auto Cleanup = make_scope_exit([&] { InLoopPredecessors.clear(); });
75 
76     // See if there are any non-loop predecessors of this exit block and
77     // keep track of the in-loop predecessors.
78     bool IsDedicatedExit = true;
79     for (auto *PredBB : predecessors(BB))
80       if (L->contains(PredBB)) {
81         if (isa<IndirectBrInst>(PredBB->getTerminator()))
82           // We cannot rewrite exiting edges from an indirectbr.
83           return false;
84         if (isa<CallBrInst>(PredBB->getTerminator()))
85           // We cannot rewrite exiting edges from a callbr.
86           return false;
87 
88         InLoopPredecessors.push_back(PredBB);
89       } else {
90         IsDedicatedExit = false;
91       }
92 
93     assert(!InLoopPredecessors.empty() && "Must have *some* loop predecessor!");
94 
95     // Nothing to do if this is already a dedicated exit.
96     if (IsDedicatedExit)
97       return false;
98 
99     auto *NewExitBB = SplitBlockPredecessors(
100         BB, InLoopPredecessors, ".loopexit", DT, LI, MSSAU, PreserveLCSSA);
101 
102     if (!NewExitBB)
103       LLVM_DEBUG(
104           dbgs() << "WARNING: Can't create a dedicated exit block for loop: "
105                  << *L << "\n");
106     else
107       LLVM_DEBUG(dbgs() << "LoopSimplify: Creating dedicated exit block "
108                         << NewExitBB->getName() << "\n");
109     return true;
110   };
111 
112   // Walk the exit blocks directly rather than building up a data structure for
113   // them, but only visit each one once.
114   SmallPtrSet<BasicBlock *, 4> Visited;
115   for (auto *BB : L->blocks())
116     for (auto *SuccBB : successors(BB)) {
117       // We're looking for exit blocks so skip in-loop successors.
118       if (L->contains(SuccBB))
119         continue;
120 
121       // Visit each exit block exactly once.
122       if (!Visited.insert(SuccBB).second)
123         continue;
124 
125       Changed |= RewriteExit(SuccBB);
126     }
127 
128   return Changed;
129 }
130 
131 /// Returns the instructions that use values defined in the loop.
132 SmallVector<Instruction *, 8> llvm::findDefsUsedOutsideOfLoop(Loop *L) {
133   SmallVector<Instruction *, 8> UsedOutside;
134 
135   for (auto *Block : L->getBlocks())
136     // FIXME: I believe that this could use copy_if if the Inst reference could
137     // be adapted into a pointer.
138     for (auto &Inst : *Block) {
139       auto Users = Inst.users();
140       if (any_of(Users, [&](User *U) {
141             auto *Use = cast<Instruction>(U);
142             return !L->contains(Use->getParent());
143           }))
144         UsedOutside.push_back(&Inst);
145     }
146 
147   return UsedOutside;
148 }
149 
150 void llvm::getLoopAnalysisUsage(AnalysisUsage &AU) {
151   // By definition, all loop passes need the LoopInfo analysis and the
152   // Dominator tree it depends on. Because they all participate in the loop
153   // pass manager, they must also preserve these.
154   AU.addRequired<DominatorTreeWrapperPass>();
155   AU.addPreserved<DominatorTreeWrapperPass>();
156   AU.addRequired<LoopInfoWrapperPass>();
157   AU.addPreserved<LoopInfoWrapperPass>();
158 
159   // We must also preserve LoopSimplify and LCSSA. We locally access their IDs
160   // here because users shouldn't directly get them from this header.
161   extern char &LoopSimplifyID;
162   extern char &LCSSAID;
163   AU.addRequiredID(LoopSimplifyID);
164   AU.addPreservedID(LoopSimplifyID);
165   AU.addRequiredID(LCSSAID);
166   AU.addPreservedID(LCSSAID);
167   // This is used in the LPPassManager to perform LCSSA verification on passes
168   // which preserve lcssa form
169   AU.addRequired<LCSSAVerificationPass>();
170   AU.addPreserved<LCSSAVerificationPass>();
171 
172   // Loop passes are designed to run inside of a loop pass manager which means
173   // that any function analyses they require must be required by the first loop
174   // pass in the manager (so that it is computed before the loop pass manager
175   // runs) and preserved by all loop pasess in the manager. To make this
176   // reasonably robust, the set needed for most loop passes is maintained here.
177   // If your loop pass requires an analysis not listed here, you will need to
178   // carefully audit the loop pass manager nesting structure that results.
179   AU.addRequired<AAResultsWrapperPass>();
180   AU.addPreserved<AAResultsWrapperPass>();
181   AU.addPreserved<BasicAAWrapperPass>();
182   AU.addPreserved<GlobalsAAWrapperPass>();
183   AU.addPreserved<SCEVAAWrapperPass>();
184   AU.addRequired<ScalarEvolutionWrapperPass>();
185   AU.addPreserved<ScalarEvolutionWrapperPass>();
186   // FIXME: When all loop passes preserve MemorySSA, it can be required and
187   // preserved here instead of the individual handling in each pass.
188 }
189 
190 /// Manually defined generic "LoopPass" dependency initialization. This is used
191 /// to initialize the exact set of passes from above in \c
192 /// getLoopAnalysisUsage. It can be used within a loop pass's initialization
193 /// with:
194 ///
195 ///   INITIALIZE_PASS_DEPENDENCY(LoopPass)
196 ///
197 /// As-if "LoopPass" were a pass.
198 void llvm::initializeLoopPassPass(PassRegistry &Registry) {
199   INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
200   INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
201   INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
202   INITIALIZE_PASS_DEPENDENCY(LCSSAWrapperPass)
203   INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
204   INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)
205   INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
206   INITIALIZE_PASS_DEPENDENCY(SCEVAAWrapperPass)
207   INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
208   INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
209 }
210 
211 /// Create MDNode for input string.
212 static MDNode *createStringMetadata(Loop *TheLoop, StringRef Name, unsigned V) {
213   LLVMContext &Context = TheLoop->getHeader()->getContext();
214   Metadata *MDs[] = {
215       MDString::get(Context, Name),
216       ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(Context), V))};
217   return MDNode::get(Context, MDs);
218 }
219 
220 /// Set input string into loop metadata by keeping other values intact.
221 /// If the string is already in loop metadata update value if it is
222 /// different.
223 void llvm::addStringMetadataToLoop(Loop *TheLoop, const char *StringMD,
224                                    unsigned V) {
225   SmallVector<Metadata *, 4> MDs(1);
226   // If the loop already has metadata, retain it.
227   MDNode *LoopID = TheLoop->getLoopID();
228   if (LoopID) {
229     for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
230       MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
231       // If it is of form key = value, try to parse it.
232       if (Node->getNumOperands() == 2) {
233         MDString *S = dyn_cast<MDString>(Node->getOperand(0));
234         if (S && S->getString().equals(StringMD)) {
235           ConstantInt *IntMD =
236               mdconst::extract_or_null<ConstantInt>(Node->getOperand(1));
237           if (IntMD && IntMD->getSExtValue() == V)
238             // It is already in place. Do nothing.
239             return;
240           // We need to update the value, so just skip it here and it will
241           // be added after copying other existed nodes.
242           continue;
243         }
244       }
245       MDs.push_back(Node);
246     }
247   }
248   // Add new metadata.
249   MDs.push_back(createStringMetadata(TheLoop, StringMD, V));
250   // Replace current metadata node with new one.
251   LLVMContext &Context = TheLoop->getHeader()->getContext();
252   MDNode *NewLoopID = MDNode::get(Context, MDs);
253   // Set operand 0 to refer to the loop id itself.
254   NewLoopID->replaceOperandWith(0, NewLoopID);
255   TheLoop->setLoopID(NewLoopID);
256 }
257 
258 Optional<ElementCount>
259 llvm::getOptionalElementCountLoopAttribute(const Loop *TheLoop) {
260   Optional<int> Width =
261       getOptionalIntLoopAttribute(TheLoop, "llvm.loop.vectorize.width");
262 
263   if (Width.hasValue()) {
264     Optional<int> IsScalable = getOptionalIntLoopAttribute(
265         TheLoop, "llvm.loop.vectorize.scalable.enable");
266     return ElementCount::get(*Width, IsScalable.getValueOr(false));
267   }
268 
269   return None;
270 }
271 
272 Optional<MDNode *> llvm::makeFollowupLoopID(
273     MDNode *OrigLoopID, ArrayRef<StringRef> FollowupOptions,
274     const char *InheritOptionsExceptPrefix, bool AlwaysNew) {
275   if (!OrigLoopID) {
276     if (AlwaysNew)
277       return nullptr;
278     return None;
279   }
280 
281   assert(OrigLoopID->getOperand(0) == OrigLoopID);
282 
283   bool InheritAllAttrs = !InheritOptionsExceptPrefix;
284   bool InheritSomeAttrs =
285       InheritOptionsExceptPrefix && InheritOptionsExceptPrefix[0] != '\0';
286   SmallVector<Metadata *, 8> MDs;
287   MDs.push_back(nullptr);
288 
289   bool Changed = false;
290   if (InheritAllAttrs || InheritSomeAttrs) {
291     for (const MDOperand &Existing : drop_begin(OrigLoopID->operands())) {
292       MDNode *Op = cast<MDNode>(Existing.get());
293 
294       auto InheritThisAttribute = [InheritSomeAttrs,
295                                    InheritOptionsExceptPrefix](MDNode *Op) {
296         if (!InheritSomeAttrs)
297           return false;
298 
299         // Skip malformatted attribute metadata nodes.
300         if (Op->getNumOperands() == 0)
301           return true;
302         Metadata *NameMD = Op->getOperand(0).get();
303         if (!isa<MDString>(NameMD))
304           return true;
305         StringRef AttrName = cast<MDString>(NameMD)->getString();
306 
307         // Do not inherit excluded attributes.
308         return !AttrName.startswith(InheritOptionsExceptPrefix);
309       };
310 
311       if (InheritThisAttribute(Op))
312         MDs.push_back(Op);
313       else
314         Changed = true;
315     }
316   } else {
317     // Modified if we dropped at least one attribute.
318     Changed = OrigLoopID->getNumOperands() > 1;
319   }
320 
321   bool HasAnyFollowup = false;
322   for (StringRef OptionName : FollowupOptions) {
323     MDNode *FollowupNode = findOptionMDForLoopID(OrigLoopID, OptionName);
324     if (!FollowupNode)
325       continue;
326 
327     HasAnyFollowup = true;
328     for (const MDOperand &Option : drop_begin(FollowupNode->operands())) {
329       MDs.push_back(Option.get());
330       Changed = true;
331     }
332   }
333 
334   // Attributes of the followup loop not specified explicity, so signal to the
335   // transformation pass to add suitable attributes.
336   if (!AlwaysNew && !HasAnyFollowup)
337     return None;
338 
339   // If no attributes were added or remove, the previous loop Id can be reused.
340   if (!AlwaysNew && !Changed)
341     return OrigLoopID;
342 
343   // No attributes is equivalent to having no !llvm.loop metadata at all.
344   if (MDs.size() == 1)
345     return nullptr;
346 
347   // Build the new loop ID.
348   MDTuple *FollowupLoopID = MDNode::get(OrigLoopID->getContext(), MDs);
349   FollowupLoopID->replaceOperandWith(0, FollowupLoopID);
350   return FollowupLoopID;
351 }
352 
353 bool llvm::hasDisableAllTransformsHint(const Loop *L) {
354   return getBooleanLoopAttribute(L, LLVMLoopDisableNonforced);
355 }
356 
357 bool llvm::hasDisableLICMTransformsHint(const Loop *L) {
358   return getBooleanLoopAttribute(L, LLVMLoopDisableLICM);
359 }
360 
361 TransformationMode llvm::hasUnrollTransformation(const Loop *L) {
362   if (getBooleanLoopAttribute(L, "llvm.loop.unroll.disable"))
363     return TM_SuppressedByUser;
364 
365   Optional<int> Count =
366       getOptionalIntLoopAttribute(L, "llvm.loop.unroll.count");
367   if (Count.hasValue())
368     return Count.getValue() == 1 ? TM_SuppressedByUser : TM_ForcedByUser;
369 
370   if (getBooleanLoopAttribute(L, "llvm.loop.unroll.enable"))
371     return TM_ForcedByUser;
372 
373   if (getBooleanLoopAttribute(L, "llvm.loop.unroll.full"))
374     return TM_ForcedByUser;
375 
376   if (hasDisableAllTransformsHint(L))
377     return TM_Disable;
378 
379   return TM_Unspecified;
380 }
381 
382 TransformationMode llvm::hasUnrollAndJamTransformation(const Loop *L) {
383   if (getBooleanLoopAttribute(L, "llvm.loop.unroll_and_jam.disable"))
384     return TM_SuppressedByUser;
385 
386   Optional<int> Count =
387       getOptionalIntLoopAttribute(L, "llvm.loop.unroll_and_jam.count");
388   if (Count.hasValue())
389     return Count.getValue() == 1 ? TM_SuppressedByUser : TM_ForcedByUser;
390 
391   if (getBooleanLoopAttribute(L, "llvm.loop.unroll_and_jam.enable"))
392     return TM_ForcedByUser;
393 
394   if (hasDisableAllTransformsHint(L))
395     return TM_Disable;
396 
397   return TM_Unspecified;
398 }
399 
400 TransformationMode llvm::hasVectorizeTransformation(const Loop *L) {
401   Optional<bool> Enable =
402       getOptionalBoolLoopAttribute(L, "llvm.loop.vectorize.enable");
403 
404   if (Enable == false)
405     return TM_SuppressedByUser;
406 
407   Optional<ElementCount> VectorizeWidth =
408       getOptionalElementCountLoopAttribute(L);
409   Optional<int> InterleaveCount =
410       getOptionalIntLoopAttribute(L, "llvm.loop.interleave.count");
411 
412   // 'Forcing' vector width and interleave count to one effectively disables
413   // this tranformation.
414   if (Enable == true && VectorizeWidth && VectorizeWidth->isScalar() &&
415       InterleaveCount == 1)
416     return TM_SuppressedByUser;
417 
418   if (getBooleanLoopAttribute(L, "llvm.loop.isvectorized"))
419     return TM_Disable;
420 
421   if (Enable == true)
422     return TM_ForcedByUser;
423 
424   if ((VectorizeWidth && VectorizeWidth->isScalar()) && InterleaveCount == 1)
425     return TM_Disable;
426 
427   if ((VectorizeWidth && VectorizeWidth->isVector()) || InterleaveCount > 1)
428     return TM_Enable;
429 
430   if (hasDisableAllTransformsHint(L))
431     return TM_Disable;
432 
433   return TM_Unspecified;
434 }
435 
436 TransformationMode llvm::hasDistributeTransformation(const Loop *L) {
437   if (getBooleanLoopAttribute(L, "llvm.loop.distribute.enable"))
438     return TM_ForcedByUser;
439 
440   if (hasDisableAllTransformsHint(L))
441     return TM_Disable;
442 
443   return TM_Unspecified;
444 }
445 
446 TransformationMode llvm::hasLICMVersioningTransformation(const Loop *L) {
447   if (getBooleanLoopAttribute(L, "llvm.loop.licm_versioning.disable"))
448     return TM_SuppressedByUser;
449 
450   if (hasDisableAllTransformsHint(L))
451     return TM_Disable;
452 
453   return TM_Unspecified;
454 }
455 
456 /// Does a BFS from a given node to all of its children inside a given loop.
457 /// The returned vector of nodes includes the starting point.
458 SmallVector<DomTreeNode *, 16>
459 llvm::collectChildrenInLoop(DomTreeNode *N, const Loop *CurLoop) {
460   SmallVector<DomTreeNode *, 16> Worklist;
461   auto AddRegionToWorklist = [&](DomTreeNode *DTN) {
462     // Only include subregions in the top level loop.
463     BasicBlock *BB = DTN->getBlock();
464     if (CurLoop->contains(BB))
465       Worklist.push_back(DTN);
466   };
467 
468   AddRegionToWorklist(N);
469 
470   for (size_t I = 0; I < Worklist.size(); I++) {
471     for (DomTreeNode *Child : Worklist[I]->children())
472       AddRegionToWorklist(Child);
473   }
474 
475   return Worklist;
476 }
477 
478 void llvm::deleteDeadLoop(Loop *L, DominatorTree *DT, ScalarEvolution *SE,
479                           LoopInfo *LI, MemorySSA *MSSA) {
480   assert((!DT || L->isLCSSAForm(*DT)) && "Expected LCSSA!");
481   auto *Preheader = L->getLoopPreheader();
482   assert(Preheader && "Preheader should exist!");
483 
484   std::unique_ptr<MemorySSAUpdater> MSSAU;
485   if (MSSA)
486     MSSAU = std::make_unique<MemorySSAUpdater>(MSSA);
487 
488   // Now that we know the removal is safe, remove the loop by changing the
489   // branch from the preheader to go to the single exit block.
490   //
491   // Because we're deleting a large chunk of code at once, the sequence in which
492   // we remove things is very important to avoid invalidation issues.
493 
494   // Tell ScalarEvolution that the loop is deleted. Do this before
495   // deleting the loop so that ScalarEvolution can look at the loop
496   // to determine what it needs to clean up.
497   if (SE)
498     SE->forgetLoop(L);
499 
500   auto *OldBr = dyn_cast<BranchInst>(Preheader->getTerminator());
501   assert(OldBr && "Preheader must end with a branch");
502   assert(OldBr->isUnconditional() && "Preheader must have a single successor");
503   // Connect the preheader to the exit block. Keep the old edge to the header
504   // around to perform the dominator tree update in two separate steps
505   // -- #1 insertion of the edge preheader -> exit and #2 deletion of the edge
506   // preheader -> header.
507   //
508   //
509   // 0.  Preheader          1.  Preheader           2.  Preheader
510   //        |                    |   |                   |
511   //        V                    |   V                   |
512   //      Header <--\            | Header <--\           | Header <--\
513   //       |  |     |            |  |  |     |           |  |  |     |
514   //       |  V     |            |  |  V     |           |  |  V     |
515   //       | Body --/            |  | Body --/           |  | Body --/
516   //       V                     V  V                    V  V
517   //      Exit                   Exit                    Exit
518   //
519   // By doing this is two separate steps we can perform the dominator tree
520   // update without using the batch update API.
521   //
522   // Even when the loop is never executed, we cannot remove the edge from the
523   // source block to the exit block. Consider the case where the unexecuted loop
524   // branches back to an outer loop. If we deleted the loop and removed the edge
525   // coming to this inner loop, this will break the outer loop structure (by
526   // deleting the backedge of the outer loop). If the outer loop is indeed a
527   // non-loop, it will be deleted in a future iteration of loop deletion pass.
528   IRBuilder<> Builder(OldBr);
529 
530   auto *ExitBlock = L->getUniqueExitBlock();
531   DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);
532   if (ExitBlock) {
533     assert(ExitBlock && "Should have a unique exit block!");
534     assert(L->hasDedicatedExits() && "Loop should have dedicated exits!");
535 
536     Builder.CreateCondBr(Builder.getFalse(), L->getHeader(), ExitBlock);
537     // Remove the old branch. The conditional branch becomes a new terminator.
538     OldBr->eraseFromParent();
539 
540     // Rewrite phis in the exit block to get their inputs from the Preheader
541     // instead of the exiting block.
542     for (PHINode &P : ExitBlock->phis()) {
543       // Set the zero'th element of Phi to be from the preheader and remove all
544       // other incoming values. Given the loop has dedicated exits, all other
545       // incoming values must be from the exiting blocks.
546       int PredIndex = 0;
547       P.setIncomingBlock(PredIndex, Preheader);
548       // Removes all incoming values from all other exiting blocks (including
549       // duplicate values from an exiting block).
550       // Nuke all entries except the zero'th entry which is the preheader entry.
551       // NOTE! We need to remove Incoming Values in the reverse order as done
552       // below, to keep the indices valid for deletion (removeIncomingValues
553       // updates getNumIncomingValues and shifts all values down into the
554       // operand being deleted).
555       for (unsigned i = 0, e = P.getNumIncomingValues() - 1; i != e; ++i)
556         P.removeIncomingValue(e - i, false);
557 
558       assert((P.getNumIncomingValues() == 1 &&
559               P.getIncomingBlock(PredIndex) == Preheader) &&
560              "Should have exactly one value and that's from the preheader!");
561     }
562 
563     if (DT) {
564       DTU.applyUpdates({{DominatorTree::Insert, Preheader, ExitBlock}});
565       if (MSSA) {
566         MSSAU->applyUpdates({{DominatorTree::Insert, Preheader, ExitBlock}},
567                             *DT);
568         if (VerifyMemorySSA)
569           MSSA->verifyMemorySSA();
570       }
571     }
572 
573     // Disconnect the loop body by branching directly to its exit.
574     Builder.SetInsertPoint(Preheader->getTerminator());
575     Builder.CreateBr(ExitBlock);
576     // Remove the old branch.
577     Preheader->getTerminator()->eraseFromParent();
578   } else {
579     assert(L->hasNoExitBlocks() &&
580            "Loop should have either zero or one exit blocks.");
581 
582     Builder.SetInsertPoint(OldBr);
583     Builder.CreateUnreachable();
584     Preheader->getTerminator()->eraseFromParent();
585   }
586 
587   if (DT) {
588     DTU.applyUpdates({{DominatorTree::Delete, Preheader, L->getHeader()}});
589     if (MSSA) {
590       MSSAU->applyUpdates({{DominatorTree::Delete, Preheader, L->getHeader()}},
591                           *DT);
592       SmallSetVector<BasicBlock *, 8> DeadBlockSet(L->block_begin(),
593                                                    L->block_end());
594       MSSAU->removeBlocks(DeadBlockSet);
595       if (VerifyMemorySSA)
596         MSSA->verifyMemorySSA();
597     }
598   }
599 
600   // Use a map to unique and a vector to guarantee deterministic ordering.
601   llvm::SmallDenseSet<std::pair<DIVariable *, DIExpression *>, 4> DeadDebugSet;
602   llvm::SmallVector<DbgVariableIntrinsic *, 4> DeadDebugInst;
603 
604   if (ExitBlock) {
605     // Given LCSSA form is satisfied, we should not have users of instructions
606     // within the dead loop outside of the loop. However, LCSSA doesn't take
607     // unreachable uses into account. We handle them here.
608     // We could do it after drop all references (in this case all users in the
609     // loop will be already eliminated and we have less work to do but according
610     // to API doc of User::dropAllReferences only valid operation after dropping
611     // references, is deletion. So let's substitute all usages of
612     // instruction from the loop with undef value of corresponding type first.
613     for (auto *Block : L->blocks())
614       for (Instruction &I : *Block) {
615         auto *Undef = UndefValue::get(I.getType());
616         for (Use &U : llvm::make_early_inc_range(I.uses())) {
617           if (auto *Usr = dyn_cast<Instruction>(U.getUser()))
618             if (L->contains(Usr->getParent()))
619               continue;
620           // If we have a DT then we can check that uses outside a loop only in
621           // unreachable block.
622           if (DT)
623             assert(!DT->isReachableFromEntry(U) &&
624                    "Unexpected user in reachable block");
625           U.set(Undef);
626         }
627         auto *DVI = dyn_cast<DbgVariableIntrinsic>(&I);
628         if (!DVI)
629           continue;
630         auto Key =
631             DeadDebugSet.find({DVI->getVariable(), DVI->getExpression()});
632         if (Key != DeadDebugSet.end())
633           continue;
634         DeadDebugSet.insert({DVI->getVariable(), DVI->getExpression()});
635         DeadDebugInst.push_back(DVI);
636       }
637 
638     // After the loop has been deleted all the values defined and modified
639     // inside the loop are going to be unavailable.
640     // Since debug values in the loop have been deleted, inserting an undef
641     // dbg.value truncates the range of any dbg.value before the loop where the
642     // loop used to be. This is particularly important for constant values.
643     DIBuilder DIB(*ExitBlock->getModule());
644     Instruction *InsertDbgValueBefore = ExitBlock->getFirstNonPHI();
645     assert(InsertDbgValueBefore &&
646            "There should be a non-PHI instruction in exit block, else these "
647            "instructions will have no parent.");
648     for (auto *DVI : DeadDebugInst)
649       DIB.insertDbgValueIntrinsic(UndefValue::get(Builder.getInt32Ty()),
650                                   DVI->getVariable(), DVI->getExpression(),
651                                   DVI->getDebugLoc(), InsertDbgValueBefore);
652   }
653 
654   // Remove the block from the reference counting scheme, so that we can
655   // delete it freely later.
656   for (auto *Block : L->blocks())
657     Block->dropAllReferences();
658 
659   if (MSSA && VerifyMemorySSA)
660     MSSA->verifyMemorySSA();
661 
662   if (LI) {
663     // Erase the instructions and the blocks without having to worry
664     // about ordering because we already dropped the references.
665     // NOTE: This iteration is safe because erasing the block does not remove
666     // its entry from the loop's block list.  We do that in the next section.
667     for (BasicBlock *BB : L->blocks())
668       BB->eraseFromParent();
669 
670     // Finally, the blocks from loopinfo.  This has to happen late because
671     // otherwise our loop iterators won't work.
672 
673     SmallPtrSet<BasicBlock *, 8> blocks;
674     blocks.insert(L->block_begin(), L->block_end());
675     for (BasicBlock *BB : blocks)
676       LI->removeBlock(BB);
677 
678     // The last step is to update LoopInfo now that we've eliminated this loop.
679     // Note: LoopInfo::erase remove the given loop and relink its subloops with
680     // its parent. While removeLoop/removeChildLoop remove the given loop but
681     // not relink its subloops, which is what we want.
682     if (Loop *ParentLoop = L->getParentLoop()) {
683       Loop::iterator I = find(*ParentLoop, L);
684       assert(I != ParentLoop->end() && "Couldn't find loop");
685       ParentLoop->removeChildLoop(I);
686     } else {
687       Loop::iterator I = find(*LI, L);
688       assert(I != LI->end() && "Couldn't find loop");
689       LI->removeLoop(I);
690     }
691     LI->destroy(L);
692   }
693 }
694 
695 static Loop *getOutermostLoop(Loop *L) {
696   while (Loop *Parent = L->getParentLoop())
697     L = Parent;
698   return L;
699 }
700 
701 void llvm::breakLoopBackedge(Loop *L, DominatorTree &DT, ScalarEvolution &SE,
702                              LoopInfo &LI, MemorySSA *MSSA) {
703   auto *Latch = L->getLoopLatch();
704   assert(Latch && "multiple latches not yet supported");
705   auto *Header = L->getHeader();
706   Loop *OutermostLoop = getOutermostLoop(L);
707 
708   SE.forgetLoop(L);
709 
710   std::unique_ptr<MemorySSAUpdater> MSSAU;
711   if (MSSA)
712     MSSAU = std::make_unique<MemorySSAUpdater>(MSSA);
713 
714   // Update the CFG and domtree.  We chose to special case a couple of
715   // of common cases for code quality and test readability reasons.
716   [&]() -> void {
717     if (auto *BI = dyn_cast<BranchInst>(Latch->getTerminator())) {
718       if (!BI->isConditional()) {
719         DomTreeUpdater DTU(&DT, DomTreeUpdater::UpdateStrategy::Eager);
720         (void)changeToUnreachable(BI, /*PreserveLCSSA*/ true, &DTU,
721                                   MSSAU.get());
722         return;
723       }
724 
725       // Conditional latch/exit - note that latch can be shared by inner
726       // and outer loop so the other target doesn't need to an exit
727       if (L->isLoopExiting(Latch)) {
728         // TODO: Generalize ConstantFoldTerminator so that it can be used
729         // here without invalidating LCSSA or MemorySSA.  (Tricky case for
730         // LCSSA: header is an exit block of a preceeding sibling loop w/o
731         // dedicated exits.)
732         const unsigned ExitIdx = L->contains(BI->getSuccessor(0)) ? 1 : 0;
733         BasicBlock *ExitBB = BI->getSuccessor(ExitIdx);
734 
735         DomTreeUpdater DTU(&DT, DomTreeUpdater::UpdateStrategy::Eager);
736         Header->removePredecessor(Latch, true);
737 
738         IRBuilder<> Builder(BI);
739         auto *NewBI = Builder.CreateBr(ExitBB);
740         // Transfer the metadata to the new branch instruction (minus the
741         // loop info since this is no longer a loop)
742         NewBI->copyMetadata(*BI, {LLVMContext::MD_dbg,
743                                   LLVMContext::MD_annotation});
744 
745         BI->eraseFromParent();
746         DTU.applyUpdates({{DominatorTree::Delete, Latch, Header}});
747         if (MSSA)
748           MSSAU->applyUpdates({{DominatorTree::Delete, Latch, Header}}, DT);
749         return;
750       }
751     }
752 
753     // General case.  By splitting the backedge, and then explicitly making it
754     // unreachable we gracefully handle corner cases such as switch and invoke
755     // termiantors.
756     auto *BackedgeBB = SplitEdge(Latch, Header, &DT, &LI, MSSAU.get());
757 
758     DomTreeUpdater DTU(&DT, DomTreeUpdater::UpdateStrategy::Eager);
759     (void)changeToUnreachable(BackedgeBB->getTerminator(),
760                               /*PreserveLCSSA*/ true, &DTU, MSSAU.get());
761   }();
762 
763   // Erase (and destroy) this loop instance.  Handles relinking sub-loops
764   // and blocks within the loop as needed.
765   LI.erase(L);
766 
767   // If the loop we broke had a parent, then changeToUnreachable might have
768   // caused a block to be removed from the parent loop (see loop_nest_lcssa
769   // test case in zero-btc.ll for an example), thus changing the parent's
770   // exit blocks.  If that happened, we need to rebuild LCSSA on the outermost
771   // loop which might have a had a block removed.
772   if (OutermostLoop != L)
773     formLCSSARecursively(*OutermostLoop, DT, &LI, &SE);
774 }
775 
776 
777 /// Checks if \p L has an exiting latch branch.  There may also be other
778 /// exiting blocks.  Returns branch instruction terminating the loop
779 /// latch if above check is successful, nullptr otherwise.
780 static BranchInst *getExpectedExitLoopLatchBranch(Loop *L) {
781   BasicBlock *Latch = L->getLoopLatch();
782   if (!Latch)
783     return nullptr;
784 
785   BranchInst *LatchBR = dyn_cast<BranchInst>(Latch->getTerminator());
786   if (!LatchBR || LatchBR->getNumSuccessors() != 2 || !L->isLoopExiting(Latch))
787     return nullptr;
788 
789   assert((LatchBR->getSuccessor(0) == L->getHeader() ||
790           LatchBR->getSuccessor(1) == L->getHeader()) &&
791          "At least one edge out of the latch must go to the header");
792 
793   return LatchBR;
794 }
795 
796 /// Return the estimated trip count for any exiting branch which dominates
797 /// the loop latch.
798 static Optional<uint64_t>
799 getEstimatedTripCount(BranchInst *ExitingBranch, Loop *L,
800                       uint64_t &OrigExitWeight) {
801   // To estimate the number of times the loop body was executed, we want to
802   // know the number of times the backedge was taken, vs. the number of times
803   // we exited the loop.
804   uint64_t LoopWeight, ExitWeight;
805   if (!ExitingBranch->extractProfMetadata(LoopWeight, ExitWeight))
806     return None;
807 
808   if (L->contains(ExitingBranch->getSuccessor(1)))
809     std::swap(LoopWeight, ExitWeight);
810 
811   if (!ExitWeight)
812     // Don't have a way to return predicated infinite
813     return None;
814 
815   OrigExitWeight = ExitWeight;
816 
817   // Estimated exit count is a ratio of the loop weight by the weight of the
818   // edge exiting the loop, rounded to nearest.
819   uint64_t ExitCount = llvm::divideNearest(LoopWeight, ExitWeight);
820   // Estimated trip count is one plus estimated exit count.
821   return ExitCount + 1;
822 }
823 
824 Optional<unsigned>
825 llvm::getLoopEstimatedTripCount(Loop *L,
826                                 unsigned *EstimatedLoopInvocationWeight) {
827   // Currently we take the estimate exit count only from the loop latch,
828   // ignoring other exiting blocks.  This can overestimate the trip count
829   // if we exit through another exit, but can never underestimate it.
830   // TODO: incorporate information from other exits
831   if (BranchInst *LatchBranch = getExpectedExitLoopLatchBranch(L)) {
832     uint64_t ExitWeight;
833     if (Optional<uint64_t> EstTripCount =
834         getEstimatedTripCount(LatchBranch, L, ExitWeight)) {
835       if (EstimatedLoopInvocationWeight)
836         *EstimatedLoopInvocationWeight = ExitWeight;
837       return *EstTripCount;
838     }
839   }
840   return None;
841 }
842 
843 bool llvm::setLoopEstimatedTripCount(Loop *L, unsigned EstimatedTripCount,
844                                      unsigned EstimatedloopInvocationWeight) {
845   // At the moment, we currently support changing the estimate trip count of
846   // the latch branch only.  We could extend this API to manipulate estimated
847   // trip counts for any exit.
848   BranchInst *LatchBranch = getExpectedExitLoopLatchBranch(L);
849   if (!LatchBranch)
850     return false;
851 
852   // Calculate taken and exit weights.
853   unsigned LatchExitWeight = 0;
854   unsigned BackedgeTakenWeight = 0;
855 
856   if (EstimatedTripCount > 0) {
857     LatchExitWeight = EstimatedloopInvocationWeight;
858     BackedgeTakenWeight = (EstimatedTripCount - 1) * LatchExitWeight;
859   }
860 
861   // Make a swap if back edge is taken when condition is "false".
862   if (LatchBranch->getSuccessor(0) != L->getHeader())
863     std::swap(BackedgeTakenWeight, LatchExitWeight);
864 
865   MDBuilder MDB(LatchBranch->getContext());
866 
867   // Set/Update profile metadata.
868   LatchBranch->setMetadata(
869       LLVMContext::MD_prof,
870       MDB.createBranchWeights(BackedgeTakenWeight, LatchExitWeight));
871 
872   return true;
873 }
874 
875 bool llvm::hasIterationCountInvariantInParent(Loop *InnerLoop,
876                                               ScalarEvolution &SE) {
877   Loop *OuterL = InnerLoop->getParentLoop();
878   if (!OuterL)
879     return true;
880 
881   // Get the backedge taken count for the inner loop
882   BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
883   const SCEV *InnerLoopBECountSC = SE.getExitCount(InnerLoop, InnerLoopLatch);
884   if (isa<SCEVCouldNotCompute>(InnerLoopBECountSC) ||
885       !InnerLoopBECountSC->getType()->isIntegerTy())
886     return false;
887 
888   // Get whether count is invariant to the outer loop
889   ScalarEvolution::LoopDisposition LD =
890       SE.getLoopDisposition(InnerLoopBECountSC, OuterL);
891   if (LD != ScalarEvolution::LoopInvariant)
892     return false;
893 
894   return true;
895 }
896 
897 CmpInst::Predicate llvm::getMinMaxReductionPredicate(RecurKind RK) {
898   switch (RK) {
899   default:
900     llvm_unreachable("Unknown min/max recurrence kind");
901   case RecurKind::UMin:
902     return CmpInst::ICMP_ULT;
903   case RecurKind::UMax:
904     return CmpInst::ICMP_UGT;
905   case RecurKind::SMin:
906     return CmpInst::ICMP_SLT;
907   case RecurKind::SMax:
908     return CmpInst::ICMP_SGT;
909   case RecurKind::FMin:
910     return CmpInst::FCMP_OLT;
911   case RecurKind::FMax:
912     return CmpInst::FCMP_OGT;
913   }
914 }
915 
916 Value *llvm::createSelectCmpOp(IRBuilderBase &Builder, Value *StartVal,
917                                RecurKind RK, Value *Left, Value *Right) {
918   if (auto VTy = dyn_cast<VectorType>(Left->getType()))
919     StartVal = Builder.CreateVectorSplat(VTy->getElementCount(), StartVal);
920   Value *Cmp =
921       Builder.CreateCmp(CmpInst::ICMP_NE, Left, StartVal, "rdx.select.cmp");
922   return Builder.CreateSelect(Cmp, Left, Right, "rdx.select");
923 }
924 
925 Value *llvm::createMinMaxOp(IRBuilderBase &Builder, RecurKind RK, Value *Left,
926                             Value *Right) {
927   CmpInst::Predicate Pred = getMinMaxReductionPredicate(RK);
928   Value *Cmp = Builder.CreateCmp(Pred, Left, Right, "rdx.minmax.cmp");
929   Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
930   return Select;
931 }
932 
933 // Helper to generate an ordered reduction.
934 Value *llvm::getOrderedReduction(IRBuilderBase &Builder, Value *Acc, Value *Src,
935                                  unsigned Op, RecurKind RdxKind) {
936   unsigned VF = cast<FixedVectorType>(Src->getType())->getNumElements();
937 
938   // Extract and apply reduction ops in ascending order:
939   // e.g. ((((Acc + Scl[0]) + Scl[1]) + Scl[2]) + ) ... + Scl[VF-1]
940   Value *Result = Acc;
941   for (unsigned ExtractIdx = 0; ExtractIdx != VF; ++ExtractIdx) {
942     Value *Ext =
943         Builder.CreateExtractElement(Src, Builder.getInt32(ExtractIdx));
944 
945     if (Op != Instruction::ICmp && Op != Instruction::FCmp) {
946       Result = Builder.CreateBinOp((Instruction::BinaryOps)Op, Result, Ext,
947                                    "bin.rdx");
948     } else {
949       assert(RecurrenceDescriptor::isMinMaxRecurrenceKind(RdxKind) &&
950              "Invalid min/max");
951       Result = createMinMaxOp(Builder, RdxKind, Result, Ext);
952     }
953   }
954 
955   return Result;
956 }
957 
958 // Helper to generate a log2 shuffle reduction.
959 Value *llvm::getShuffleReduction(IRBuilderBase &Builder, Value *Src,
960                                  unsigned Op, RecurKind RdxKind) {
961   unsigned VF = cast<FixedVectorType>(Src->getType())->getNumElements();
962   // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
963   // and vector ops, reducing the set of values being computed by half each
964   // round.
965   assert(isPowerOf2_32(VF) &&
966          "Reduction emission only supported for pow2 vectors!");
967   // Note: fast-math-flags flags are controlled by the builder configuration
968   // and are assumed to apply to all generated arithmetic instructions.  Other
969   // poison generating flags (nsw/nuw/inbounds/inrange/exact) are not part
970   // of the builder configuration, and since they're not passed explicitly,
971   // will never be relevant here.  Note that it would be generally unsound to
972   // propagate these from an intrinsic call to the expansion anyways as we/
973   // change the order of operations.
974   Value *TmpVec = Src;
975   SmallVector<int, 32> ShuffleMask(VF);
976   for (unsigned i = VF; i != 1; i >>= 1) {
977     // Move the upper half of the vector to the lower half.
978     for (unsigned j = 0; j != i / 2; ++j)
979       ShuffleMask[j] = i / 2 + j;
980 
981     // Fill the rest of the mask with undef.
982     std::fill(&ShuffleMask[i / 2], ShuffleMask.end(), -1);
983 
984     Value *Shuf = Builder.CreateShuffleVector(TmpVec, ShuffleMask, "rdx.shuf");
985 
986     if (Op != Instruction::ICmp && Op != Instruction::FCmp) {
987       TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
988                                    "bin.rdx");
989     } else {
990       assert(RecurrenceDescriptor::isMinMaxRecurrenceKind(RdxKind) &&
991              "Invalid min/max");
992       TmpVec = createMinMaxOp(Builder, RdxKind, TmpVec, Shuf);
993     }
994   }
995   // The result is in the first element of the vector.
996   return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
997 }
998 
999 Value *llvm::createSelectCmpTargetReduction(IRBuilderBase &Builder,
1000                                             const TargetTransformInfo *TTI,
1001                                             Value *Src,
1002                                             const RecurrenceDescriptor &Desc,
1003                                             PHINode *OrigPhi) {
1004   assert(RecurrenceDescriptor::isSelectCmpRecurrenceKind(
1005              Desc.getRecurrenceKind()) &&
1006          "Unexpected reduction kind");
1007   Value *InitVal = Desc.getRecurrenceStartValue();
1008   Value *NewVal = nullptr;
1009 
1010   // First use the original phi to determine the new value we're trying to
1011   // select from in the loop.
1012   SelectInst *SI = nullptr;
1013   for (auto *U : OrigPhi->users()) {
1014     if ((SI = dyn_cast<SelectInst>(U)))
1015       break;
1016   }
1017   assert(SI && "One user of the original phi should be a select");
1018 
1019   if (SI->getTrueValue() == OrigPhi)
1020     NewVal = SI->getFalseValue();
1021   else {
1022     assert(SI->getFalseValue() == OrigPhi &&
1023            "At least one input to the select should be the original Phi");
1024     NewVal = SI->getTrueValue();
1025   }
1026 
1027   // Create a splat vector with the new value and compare this to the vector
1028   // we want to reduce.
1029   ElementCount EC = cast<VectorType>(Src->getType())->getElementCount();
1030   Value *Right = Builder.CreateVectorSplat(EC, InitVal);
1031   Value *Cmp =
1032       Builder.CreateCmp(CmpInst::ICMP_NE, Src, Right, "rdx.select.cmp");
1033 
1034   // If any predicate is true it means that we want to select the new value.
1035   Cmp = Builder.CreateOrReduce(Cmp);
1036   return Builder.CreateSelect(Cmp, NewVal, InitVal, "rdx.select");
1037 }
1038 
1039 Value *llvm::createSimpleTargetReduction(IRBuilderBase &Builder,
1040                                          const TargetTransformInfo *TTI,
1041                                          Value *Src, RecurKind RdxKind) {
1042   auto *SrcVecEltTy = cast<VectorType>(Src->getType())->getElementType();
1043   switch (RdxKind) {
1044   case RecurKind::Add:
1045     return Builder.CreateAddReduce(Src);
1046   case RecurKind::Mul:
1047     return Builder.CreateMulReduce(Src);
1048   case RecurKind::And:
1049     return Builder.CreateAndReduce(Src);
1050   case RecurKind::Or:
1051     return Builder.CreateOrReduce(Src);
1052   case RecurKind::Xor:
1053     return Builder.CreateXorReduce(Src);
1054   case RecurKind::FMulAdd:
1055   case RecurKind::FAdd:
1056     return Builder.CreateFAddReduce(ConstantFP::getNegativeZero(SrcVecEltTy),
1057                                     Src);
1058   case RecurKind::FMul:
1059     return Builder.CreateFMulReduce(ConstantFP::get(SrcVecEltTy, 1.0), Src);
1060   case RecurKind::SMax:
1061     return Builder.CreateIntMaxReduce(Src, true);
1062   case RecurKind::SMin:
1063     return Builder.CreateIntMinReduce(Src, true);
1064   case RecurKind::UMax:
1065     return Builder.CreateIntMaxReduce(Src, false);
1066   case RecurKind::UMin:
1067     return Builder.CreateIntMinReduce(Src, false);
1068   case RecurKind::FMax:
1069     return Builder.CreateFPMaxReduce(Src);
1070   case RecurKind::FMin:
1071     return Builder.CreateFPMinReduce(Src);
1072   default:
1073     llvm_unreachable("Unhandled opcode");
1074   }
1075 }
1076 
1077 Value *llvm::createTargetReduction(IRBuilderBase &B,
1078                                    const TargetTransformInfo *TTI,
1079                                    const RecurrenceDescriptor &Desc, Value *Src,
1080                                    PHINode *OrigPhi) {
1081   // TODO: Support in-order reductions based on the recurrence descriptor.
1082   // All ops in the reduction inherit fast-math-flags from the recurrence
1083   // descriptor.
1084   IRBuilderBase::FastMathFlagGuard FMFGuard(B);
1085   B.setFastMathFlags(Desc.getFastMathFlags());
1086 
1087   RecurKind RK = Desc.getRecurrenceKind();
1088   if (RecurrenceDescriptor::isSelectCmpRecurrenceKind(RK))
1089     return createSelectCmpTargetReduction(B, TTI, Src, Desc, OrigPhi);
1090 
1091   return createSimpleTargetReduction(B, TTI, Src, RK);
1092 }
1093 
1094 Value *llvm::createOrderedReduction(IRBuilderBase &B,
1095                                     const RecurrenceDescriptor &Desc,
1096                                     Value *Src, Value *Start) {
1097   assert((Desc.getRecurrenceKind() == RecurKind::FAdd ||
1098           Desc.getRecurrenceKind() == RecurKind::FMulAdd) &&
1099          "Unexpected reduction kind");
1100   assert(Src->getType()->isVectorTy() && "Expected a vector type");
1101   assert(!Start->getType()->isVectorTy() && "Expected a scalar type");
1102 
1103   return B.CreateFAddReduce(Start, Src);
1104 }
1105 
1106 void llvm::propagateIRFlags(Value *I, ArrayRef<Value *> VL, Value *OpValue) {
1107   auto *VecOp = dyn_cast<Instruction>(I);
1108   if (!VecOp)
1109     return;
1110   auto *Intersection = (OpValue == nullptr) ? dyn_cast<Instruction>(VL[0])
1111                                             : dyn_cast<Instruction>(OpValue);
1112   if (!Intersection)
1113     return;
1114   const unsigned Opcode = Intersection->getOpcode();
1115   VecOp->copyIRFlags(Intersection);
1116   for (auto *V : VL) {
1117     auto *Instr = dyn_cast<Instruction>(V);
1118     if (!Instr)
1119       continue;
1120     if (OpValue == nullptr || Opcode == Instr->getOpcode())
1121       VecOp->andIRFlags(V);
1122   }
1123 }
1124 
1125 bool llvm::isKnownNegativeInLoop(const SCEV *S, const Loop *L,
1126                                  ScalarEvolution &SE) {
1127   const SCEV *Zero = SE.getZero(S->getType());
1128   return SE.isAvailableAtLoopEntry(S, L) &&
1129          SE.isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SLT, S, Zero);
1130 }
1131 
1132 bool llvm::isKnownNonNegativeInLoop(const SCEV *S, const Loop *L,
1133                                     ScalarEvolution &SE) {
1134   const SCEV *Zero = SE.getZero(S->getType());
1135   return SE.isAvailableAtLoopEntry(S, L) &&
1136          SE.isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SGE, S, Zero);
1137 }
1138 
1139 bool llvm::cannotBeMinInLoop(const SCEV *S, const Loop *L, ScalarEvolution &SE,
1140                              bool Signed) {
1141   unsigned BitWidth = cast<IntegerType>(S->getType())->getBitWidth();
1142   APInt Min = Signed ? APInt::getSignedMinValue(BitWidth) :
1143     APInt::getMinValue(BitWidth);
1144   auto Predicate = Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1145   return SE.isAvailableAtLoopEntry(S, L) &&
1146          SE.isLoopEntryGuardedByCond(L, Predicate, S,
1147                                      SE.getConstant(Min));
1148 }
1149 
1150 bool llvm::cannotBeMaxInLoop(const SCEV *S, const Loop *L, ScalarEvolution &SE,
1151                              bool Signed) {
1152   unsigned BitWidth = cast<IntegerType>(S->getType())->getBitWidth();
1153   APInt Max = Signed ? APInt::getSignedMaxValue(BitWidth) :
1154     APInt::getMaxValue(BitWidth);
1155   auto Predicate = Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1156   return SE.isAvailableAtLoopEntry(S, L) &&
1157          SE.isLoopEntryGuardedByCond(L, Predicate, S,
1158                                      SE.getConstant(Max));
1159 }
1160 
1161 //===----------------------------------------------------------------------===//
1162 // rewriteLoopExitValues - Optimize IV users outside the loop.
1163 // As a side effect, reduces the amount of IV processing within the loop.
1164 //===----------------------------------------------------------------------===//
1165 
1166 static bool hasHardUserWithinLoop(const Loop *L, const Instruction *I) {
1167   SmallPtrSet<const Instruction *, 8> Visited;
1168   SmallVector<const Instruction *, 8> WorkList;
1169   Visited.insert(I);
1170   WorkList.push_back(I);
1171   while (!WorkList.empty()) {
1172     const Instruction *Curr = WorkList.pop_back_val();
1173     // This use is outside the loop, nothing to do.
1174     if (!L->contains(Curr))
1175       continue;
1176     // Do we assume it is a "hard" use which will not be eliminated easily?
1177     if (Curr->mayHaveSideEffects())
1178       return true;
1179     // Otherwise, add all its users to worklist.
1180     for (auto U : Curr->users()) {
1181       auto *UI = cast<Instruction>(U);
1182       if (Visited.insert(UI).second)
1183         WorkList.push_back(UI);
1184     }
1185   }
1186   return false;
1187 }
1188 
1189 // Collect information about PHI nodes which can be transformed in
1190 // rewriteLoopExitValues.
1191 struct RewritePhi {
1192   PHINode *PN;               // For which PHI node is this replacement?
1193   unsigned Ith;              // For which incoming value?
1194   const SCEV *ExpansionSCEV; // The SCEV of the incoming value we are rewriting.
1195   Instruction *ExpansionPoint; // Where we'd like to expand that SCEV?
1196   bool HighCost;               // Is this expansion a high-cost?
1197 
1198   RewritePhi(PHINode *P, unsigned I, const SCEV *Val, Instruction *ExpansionPt,
1199              bool H)
1200       : PN(P), Ith(I), ExpansionSCEV(Val), ExpansionPoint(ExpansionPt),
1201         HighCost(H) {}
1202 };
1203 
1204 // Check whether it is possible to delete the loop after rewriting exit
1205 // value. If it is possible, ignore ReplaceExitValue and do rewriting
1206 // aggressively.
1207 static bool canLoopBeDeleted(Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet) {
1208   BasicBlock *Preheader = L->getLoopPreheader();
1209   // If there is no preheader, the loop will not be deleted.
1210   if (!Preheader)
1211     return false;
1212 
1213   // In LoopDeletion pass Loop can be deleted when ExitingBlocks.size() > 1.
1214   // We obviate multiple ExitingBlocks case for simplicity.
1215   // TODO: If we see testcase with multiple ExitingBlocks can be deleted
1216   // after exit value rewriting, we can enhance the logic here.
1217   SmallVector<BasicBlock *, 4> ExitingBlocks;
1218   L->getExitingBlocks(ExitingBlocks);
1219   SmallVector<BasicBlock *, 8> ExitBlocks;
1220   L->getUniqueExitBlocks(ExitBlocks);
1221   if (ExitBlocks.size() != 1 || ExitingBlocks.size() != 1)
1222     return false;
1223 
1224   BasicBlock *ExitBlock = ExitBlocks[0];
1225   BasicBlock::iterator BI = ExitBlock->begin();
1226   while (PHINode *P = dyn_cast<PHINode>(BI)) {
1227     Value *Incoming = P->getIncomingValueForBlock(ExitingBlocks[0]);
1228 
1229     // If the Incoming value of P is found in RewritePhiSet, we know it
1230     // could be rewritten to use a loop invariant value in transformation
1231     // phase later. Skip it in the loop invariant check below.
1232     bool found = false;
1233     for (const RewritePhi &Phi : RewritePhiSet) {
1234       unsigned i = Phi.Ith;
1235       if (Phi.PN == P && (Phi.PN)->getIncomingValue(i) == Incoming) {
1236         found = true;
1237         break;
1238       }
1239     }
1240 
1241     Instruction *I;
1242     if (!found && (I = dyn_cast<Instruction>(Incoming)))
1243       if (!L->hasLoopInvariantOperands(I))
1244         return false;
1245 
1246     ++BI;
1247   }
1248 
1249   for (auto *BB : L->blocks())
1250     if (llvm::any_of(*BB, [](Instruction &I) {
1251           return I.mayHaveSideEffects();
1252         }))
1253       return false;
1254 
1255   return true;
1256 }
1257 
1258 int llvm::rewriteLoopExitValues(Loop *L, LoopInfo *LI, TargetLibraryInfo *TLI,
1259                                 ScalarEvolution *SE,
1260                                 const TargetTransformInfo *TTI,
1261                                 SCEVExpander &Rewriter, DominatorTree *DT,
1262                                 ReplaceExitVal ReplaceExitValue,
1263                                 SmallVector<WeakTrackingVH, 16> &DeadInsts) {
1264   // Check a pre-condition.
1265   assert(L->isRecursivelyLCSSAForm(*DT, *LI) &&
1266          "Indvars did not preserve LCSSA!");
1267 
1268   SmallVector<BasicBlock*, 8> ExitBlocks;
1269   L->getUniqueExitBlocks(ExitBlocks);
1270 
1271   SmallVector<RewritePhi, 8> RewritePhiSet;
1272   // Find all values that are computed inside the loop, but used outside of it.
1273   // Because of LCSSA, these values will only occur in LCSSA PHI Nodes.  Scan
1274   // the exit blocks of the loop to find them.
1275   for (BasicBlock *ExitBB : ExitBlocks) {
1276     // If there are no PHI nodes in this exit block, then no values defined
1277     // inside the loop are used on this path, skip it.
1278     PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
1279     if (!PN) continue;
1280 
1281     unsigned NumPreds = PN->getNumIncomingValues();
1282 
1283     // Iterate over all of the PHI nodes.
1284     BasicBlock::iterator BBI = ExitBB->begin();
1285     while ((PN = dyn_cast<PHINode>(BBI++))) {
1286       if (PN->use_empty())
1287         continue; // dead use, don't replace it
1288 
1289       if (!SE->isSCEVable(PN->getType()))
1290         continue;
1291 
1292       // Iterate over all of the values in all the PHI nodes.
1293       for (unsigned i = 0; i != NumPreds; ++i) {
1294         // If the value being merged in is not integer or is not defined
1295         // in the loop, skip it.
1296         Value *InVal = PN->getIncomingValue(i);
1297         if (!isa<Instruction>(InVal))
1298           continue;
1299 
1300         // If this pred is for a subloop, not L itself, skip it.
1301         if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
1302           continue; // The Block is in a subloop, skip it.
1303 
1304         // Check that InVal is defined in the loop.
1305         Instruction *Inst = cast<Instruction>(InVal);
1306         if (!L->contains(Inst))
1307           continue;
1308 
1309         // Okay, this instruction has a user outside of the current loop
1310         // and varies predictably *inside* the loop.  Evaluate the value it
1311         // contains when the loop exits, if possible.  We prefer to start with
1312         // expressions which are true for all exits (so as to maximize
1313         // expression reuse by the SCEVExpander), but resort to per-exit
1314         // evaluation if that fails.
1315         const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
1316         if (isa<SCEVCouldNotCompute>(ExitValue) ||
1317             !SE->isLoopInvariant(ExitValue, L) ||
1318             !isSafeToExpand(ExitValue, *SE)) {
1319           // TODO: This should probably be sunk into SCEV in some way; maybe a
1320           // getSCEVForExit(SCEV*, L, ExitingBB)?  It can be generalized for
1321           // most SCEV expressions and other recurrence types (e.g. shift
1322           // recurrences).  Is there existing code we can reuse?
1323           const SCEV *ExitCount = SE->getExitCount(L, PN->getIncomingBlock(i));
1324           if (isa<SCEVCouldNotCompute>(ExitCount))
1325             continue;
1326           if (auto *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Inst)))
1327             if (AddRec->getLoop() == L)
1328               ExitValue = AddRec->evaluateAtIteration(ExitCount, *SE);
1329           if (isa<SCEVCouldNotCompute>(ExitValue) ||
1330               !SE->isLoopInvariant(ExitValue, L) ||
1331               !isSafeToExpand(ExitValue, *SE))
1332             continue;
1333         }
1334 
1335         // Computing the value outside of the loop brings no benefit if it is
1336         // definitely used inside the loop in a way which can not be optimized
1337         // away. Avoid doing so unless we know we have a value which computes
1338         // the ExitValue already. TODO: This should be merged into SCEV
1339         // expander to leverage its knowledge of existing expressions.
1340         if (ReplaceExitValue != AlwaysRepl && !isa<SCEVConstant>(ExitValue) &&
1341             !isa<SCEVUnknown>(ExitValue) && hasHardUserWithinLoop(L, Inst))
1342           continue;
1343 
1344         // Check if expansions of this SCEV would count as being high cost.
1345         bool HighCost = Rewriter.isHighCostExpansion(
1346             ExitValue, L, SCEVCheapExpansionBudget, TTI, Inst);
1347 
1348         // Note that we must not perform expansions until after
1349         // we query *all* the costs, because if we perform temporary expansion
1350         // inbetween, one that we might not intend to keep, said expansion
1351         // *may* affect cost calculation of the the next SCEV's we'll query,
1352         // and next SCEV may errneously get smaller cost.
1353 
1354         // Collect all the candidate PHINodes to be rewritten.
1355         RewritePhiSet.emplace_back(PN, i, ExitValue, Inst, HighCost);
1356       }
1357     }
1358   }
1359 
1360   // TODO: evaluate whether it is beneficial to change how we calculate
1361   // high-cost: if we have SCEV 'A' which we know we will expand, should we
1362   // calculate the cost of other SCEV's after expanding SCEV 'A', thus
1363   // potentially giving cost bonus to those other SCEV's?
1364 
1365   bool LoopCanBeDel = canLoopBeDeleted(L, RewritePhiSet);
1366   int NumReplaced = 0;
1367 
1368   // Transformation.
1369   for (const RewritePhi &Phi : RewritePhiSet) {
1370     PHINode *PN = Phi.PN;
1371 
1372     // Only do the rewrite when the ExitValue can be expanded cheaply.
1373     // If LoopCanBeDel is true, rewrite exit value aggressively.
1374     if (ReplaceExitValue == OnlyCheapRepl && !LoopCanBeDel && Phi.HighCost)
1375       continue;
1376 
1377     Value *ExitVal = Rewriter.expandCodeFor(
1378         Phi.ExpansionSCEV, Phi.PN->getType(), Phi.ExpansionPoint);
1379 
1380     LLVM_DEBUG(dbgs() << "rewriteLoopExitValues: AfterLoopVal = " << *ExitVal
1381                       << '\n'
1382                       << "  LoopVal = " << *(Phi.ExpansionPoint) << "\n");
1383 
1384 #ifndef NDEBUG
1385     // If we reuse an instruction from a loop which is neither L nor one of
1386     // its containing loops, we end up breaking LCSSA form for this loop by
1387     // creating a new use of its instruction.
1388     if (auto *ExitInsn = dyn_cast<Instruction>(ExitVal))
1389       if (auto *EVL = LI->getLoopFor(ExitInsn->getParent()))
1390         if (EVL != L)
1391           assert(EVL->contains(L) && "LCSSA breach detected!");
1392 #endif
1393 
1394     NumReplaced++;
1395     Instruction *Inst = cast<Instruction>(PN->getIncomingValue(Phi.Ith));
1396     PN->setIncomingValue(Phi.Ith, ExitVal);
1397     // It's necessary to tell ScalarEvolution about this explicitly so that
1398     // it can walk the def-use list and forget all SCEVs, as it may not be
1399     // watching the PHI itself. Once the new exit value is in place, there
1400     // may not be a def-use connection between the loop and every instruction
1401     // which got a SCEVAddRecExpr for that loop.
1402     SE->forgetValue(PN);
1403 
1404     // If this instruction is dead now, delete it. Don't do it now to avoid
1405     // invalidating iterators.
1406     if (isInstructionTriviallyDead(Inst, TLI))
1407       DeadInsts.push_back(Inst);
1408 
1409     // Replace PN with ExitVal if that is legal and does not break LCSSA.
1410     if (PN->getNumIncomingValues() == 1 &&
1411         LI->replacementPreservesLCSSAForm(PN, ExitVal)) {
1412       PN->replaceAllUsesWith(ExitVal);
1413       PN->eraseFromParent();
1414     }
1415   }
1416 
1417   // The insertion point instruction may have been deleted; clear it out
1418   // so that the rewriter doesn't trip over it later.
1419   Rewriter.clearInsertPoint();
1420   return NumReplaced;
1421 }
1422 
1423 /// Set weights for \p UnrolledLoop and \p RemainderLoop based on weights for
1424 /// \p OrigLoop.
1425 void llvm::setProfileInfoAfterUnrolling(Loop *OrigLoop, Loop *UnrolledLoop,
1426                                         Loop *RemainderLoop, uint64_t UF) {
1427   assert(UF > 0 && "Zero unrolled factor is not supported");
1428   assert(UnrolledLoop != RemainderLoop &&
1429          "Unrolled and Remainder loops are expected to distinct");
1430 
1431   // Get number of iterations in the original scalar loop.
1432   unsigned OrigLoopInvocationWeight = 0;
1433   Optional<unsigned> OrigAverageTripCount =
1434       getLoopEstimatedTripCount(OrigLoop, &OrigLoopInvocationWeight);
1435   if (!OrigAverageTripCount)
1436     return;
1437 
1438   // Calculate number of iterations in unrolled loop.
1439   unsigned UnrolledAverageTripCount = *OrigAverageTripCount / UF;
1440   // Calculate number of iterations for remainder loop.
1441   unsigned RemainderAverageTripCount = *OrigAverageTripCount % UF;
1442 
1443   setLoopEstimatedTripCount(UnrolledLoop, UnrolledAverageTripCount,
1444                             OrigLoopInvocationWeight);
1445   setLoopEstimatedTripCount(RemainderLoop, RemainderAverageTripCount,
1446                             OrigLoopInvocationWeight);
1447 }
1448 
1449 /// Utility that implements appending of loops onto a worklist.
1450 /// Loops are added in preorder (analogous for reverse postorder for trees),
1451 /// and the worklist is processed LIFO.
1452 template <typename RangeT>
1453 void llvm::appendReversedLoopsToWorklist(
1454     RangeT &&Loops, SmallPriorityWorklist<Loop *, 4> &Worklist) {
1455   // We use an internal worklist to build up the preorder traversal without
1456   // recursion.
1457   SmallVector<Loop *, 4> PreOrderLoops, PreOrderWorklist;
1458 
1459   // We walk the initial sequence of loops in reverse because we generally want
1460   // to visit defs before uses and the worklist is LIFO.
1461   for (Loop *RootL : Loops) {
1462     assert(PreOrderLoops.empty() && "Must start with an empty preorder walk.");
1463     assert(PreOrderWorklist.empty() &&
1464            "Must start with an empty preorder walk worklist.");
1465     PreOrderWorklist.push_back(RootL);
1466     do {
1467       Loop *L = PreOrderWorklist.pop_back_val();
1468       PreOrderWorklist.append(L->begin(), L->end());
1469       PreOrderLoops.push_back(L);
1470     } while (!PreOrderWorklist.empty());
1471 
1472     Worklist.insert(std::move(PreOrderLoops));
1473     PreOrderLoops.clear();
1474   }
1475 }
1476 
1477 template <typename RangeT>
1478 void llvm::appendLoopsToWorklist(RangeT &&Loops,
1479                                  SmallPriorityWorklist<Loop *, 4> &Worklist) {
1480   appendReversedLoopsToWorklist(reverse(Loops), Worklist);
1481 }
1482 
1483 template void llvm::appendLoopsToWorklist<ArrayRef<Loop *> &>(
1484     ArrayRef<Loop *> &Loops, SmallPriorityWorklist<Loop *, 4> &Worklist);
1485 
1486 template void
1487 llvm::appendLoopsToWorklist<Loop &>(Loop &L,
1488                                     SmallPriorityWorklist<Loop *, 4> &Worklist);
1489 
1490 void llvm::appendLoopsToWorklist(LoopInfo &LI,
1491                                  SmallPriorityWorklist<Loop *, 4> &Worklist) {
1492   appendReversedLoopsToWorklist(LI, Worklist);
1493 }
1494 
1495 Loop *llvm::cloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM,
1496                       LoopInfo *LI, LPPassManager *LPM) {
1497   Loop &New = *LI->AllocateLoop();
1498   if (PL)
1499     PL->addChildLoop(&New);
1500   else
1501     LI->addTopLevelLoop(&New);
1502 
1503   if (LPM)
1504     LPM->addLoop(New);
1505 
1506   // Add all of the blocks in L to the new loop.
1507   for (BasicBlock *BB : L->blocks())
1508     if (LI->getLoopFor(BB) == L)
1509       New.addBasicBlockToLoop(cast<BasicBlock>(VM[BB]), *LI);
1510 
1511   // Add all of the subloops to the new loop.
1512   for (Loop *I : *L)
1513     cloneLoop(I, &New, VM, LI, LPM);
1514 
1515   return &New;
1516 }
1517 
1518 /// IR Values for the lower and upper bounds of a pointer evolution.  We
1519 /// need to use value-handles because SCEV expansion can invalidate previously
1520 /// expanded values.  Thus expansion of a pointer can invalidate the bounds for
1521 /// a previous one.
1522 struct PointerBounds {
1523   TrackingVH<Value> Start;
1524   TrackingVH<Value> End;
1525 };
1526 
1527 /// Expand code for the lower and upper bound of the pointer group \p CG
1528 /// in \p TheLoop.  \return the values for the bounds.
1529 static PointerBounds expandBounds(const RuntimeCheckingPtrGroup *CG,
1530                                   Loop *TheLoop, Instruction *Loc,
1531                                   SCEVExpander &Exp) {
1532   LLVMContext &Ctx = Loc->getContext();
1533   Type *PtrArithTy = Type::getInt8PtrTy(Ctx, CG->AddressSpace);
1534 
1535   Value *Start = nullptr, *End = nullptr;
1536   LLVM_DEBUG(dbgs() << "LAA: Adding RT check for range:\n");
1537   Start = Exp.expandCodeFor(CG->Low, PtrArithTy, Loc);
1538   End = Exp.expandCodeFor(CG->High, PtrArithTy, Loc);
1539   LLVM_DEBUG(dbgs() << "Start: " << *CG->Low << " End: " << *CG->High << "\n");
1540   return {Start, End};
1541 }
1542 
1543 /// Turns a collection of checks into a collection of expanded upper and
1544 /// lower bounds for both pointers in the check.
1545 static SmallVector<std::pair<PointerBounds, PointerBounds>, 4>
1546 expandBounds(const SmallVectorImpl<RuntimePointerCheck> &PointerChecks, Loop *L,
1547              Instruction *Loc, SCEVExpander &Exp) {
1548   SmallVector<std::pair<PointerBounds, PointerBounds>, 4> ChecksWithBounds;
1549 
1550   // Here we're relying on the SCEV Expander's cache to only emit code for the
1551   // same bounds once.
1552   transform(PointerChecks, std::back_inserter(ChecksWithBounds),
1553             [&](const RuntimePointerCheck &Check) {
1554               PointerBounds First = expandBounds(Check.first, L, Loc, Exp),
1555                             Second = expandBounds(Check.second, L, Loc, Exp);
1556               return std::make_pair(First, Second);
1557             });
1558 
1559   return ChecksWithBounds;
1560 }
1561 
1562 Value *llvm::addRuntimeChecks(
1563     Instruction *Loc, Loop *TheLoop,
1564     const SmallVectorImpl<RuntimePointerCheck> &PointerChecks,
1565     SCEVExpander &Exp) {
1566   // TODO: Move noalias annotation code from LoopVersioning here and share with LV if possible.
1567   // TODO: Pass  RtPtrChecking instead of PointerChecks and SE separately, if possible
1568   auto ExpandedChecks = expandBounds(PointerChecks, TheLoop, Loc, Exp);
1569 
1570   LLVMContext &Ctx = Loc->getContext();
1571   IRBuilder<InstSimplifyFolder> ChkBuilder(Ctx,
1572                                            Loc->getModule()->getDataLayout());
1573   ChkBuilder.SetInsertPoint(Loc);
1574   // Our instructions might fold to a constant.
1575   Value *MemoryRuntimeCheck = nullptr;
1576 
1577   for (const auto &Check : ExpandedChecks) {
1578     const PointerBounds &A = Check.first, &B = Check.second;
1579     // Check if two pointers (A and B) conflict where conflict is computed as:
1580     // start(A) <= end(B) && start(B) <= end(A)
1581     unsigned AS0 = A.Start->getType()->getPointerAddressSpace();
1582     unsigned AS1 = B.Start->getType()->getPointerAddressSpace();
1583 
1584     assert((AS0 == B.End->getType()->getPointerAddressSpace()) &&
1585            (AS1 == A.End->getType()->getPointerAddressSpace()) &&
1586            "Trying to bounds check pointers with different address spaces");
1587 
1588     Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
1589     Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
1590 
1591     Value *Start0 = ChkBuilder.CreateBitCast(A.Start, PtrArithTy0, "bc");
1592     Value *Start1 = ChkBuilder.CreateBitCast(B.Start, PtrArithTy1, "bc");
1593     Value *End0 = ChkBuilder.CreateBitCast(A.End, PtrArithTy1, "bc");
1594     Value *End1 = ChkBuilder.CreateBitCast(B.End, PtrArithTy0, "bc");
1595 
1596     // [A|B].Start points to the first accessed byte under base [A|B].
1597     // [A|B].End points to the last accessed byte, plus one.
1598     // There is no conflict when the intervals are disjoint:
1599     // NoConflict = (B.Start >= A.End) || (A.Start >= B.End)
1600     //
1601     // bound0 = (B.Start < A.End)
1602     // bound1 = (A.Start < B.End)
1603     //  IsConflict = bound0 & bound1
1604     Value *Cmp0 = ChkBuilder.CreateICmpULT(Start0, End1, "bound0");
1605     Value *Cmp1 = ChkBuilder.CreateICmpULT(Start1, End0, "bound1");
1606     Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
1607     if (MemoryRuntimeCheck) {
1608       IsConflict =
1609           ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict, "conflict.rdx");
1610     }
1611     MemoryRuntimeCheck = IsConflict;
1612   }
1613 
1614   return MemoryRuntimeCheck;
1615 }
1616 
1617 Optional<IVConditionInfo> llvm::hasPartialIVCondition(Loop &L,
1618                                                       unsigned MSSAThreshold,
1619                                                       MemorySSA &MSSA,
1620                                                       AAResults &AA) {
1621   auto *TI = dyn_cast<BranchInst>(L.getHeader()->getTerminator());
1622   if (!TI || !TI->isConditional())
1623     return {};
1624 
1625   auto *CondI = dyn_cast<CmpInst>(TI->getCondition());
1626   // The case with the condition outside the loop should already be handled
1627   // earlier.
1628   if (!CondI || !L.contains(CondI))
1629     return {};
1630 
1631   SmallVector<Instruction *> InstToDuplicate;
1632   InstToDuplicate.push_back(CondI);
1633 
1634   SmallVector<Value *, 4> WorkList;
1635   WorkList.append(CondI->op_begin(), CondI->op_end());
1636 
1637   SmallVector<MemoryAccess *, 4> AccessesToCheck;
1638   SmallVector<MemoryLocation, 4> AccessedLocs;
1639   while (!WorkList.empty()) {
1640     Instruction *I = dyn_cast<Instruction>(WorkList.pop_back_val());
1641     if (!I || !L.contains(I))
1642       continue;
1643 
1644     // TODO: support additional instructions.
1645     if (!isa<LoadInst>(I) && !isa<GetElementPtrInst>(I))
1646       return {};
1647 
1648     // Do not duplicate volatile and atomic loads.
1649     if (auto *LI = dyn_cast<LoadInst>(I))
1650       if (LI->isVolatile() || LI->isAtomic())
1651         return {};
1652 
1653     InstToDuplicate.push_back(I);
1654     if (MemoryAccess *MA = MSSA.getMemoryAccess(I)) {
1655       if (auto *MemUse = dyn_cast_or_null<MemoryUse>(MA)) {
1656         // Queue the defining access to check for alias checks.
1657         AccessesToCheck.push_back(MemUse->getDefiningAccess());
1658         AccessedLocs.push_back(MemoryLocation::get(I));
1659       } else {
1660         // MemoryDefs may clobber the location or may be atomic memory
1661         // operations. Bail out.
1662         return {};
1663       }
1664     }
1665     WorkList.append(I->op_begin(), I->op_end());
1666   }
1667 
1668   if (InstToDuplicate.empty())
1669     return {};
1670 
1671   SmallVector<BasicBlock *, 4> ExitingBlocks;
1672   L.getExitingBlocks(ExitingBlocks);
1673   auto HasNoClobbersOnPath =
1674       [&L, &AA, &AccessedLocs, &ExitingBlocks, &InstToDuplicate,
1675        MSSAThreshold](BasicBlock *Succ, BasicBlock *Header,
1676                       SmallVector<MemoryAccess *, 4> AccessesToCheck)
1677       -> Optional<IVConditionInfo> {
1678     IVConditionInfo Info;
1679     // First, collect all blocks in the loop that are on a patch from Succ
1680     // to the header.
1681     SmallVector<BasicBlock *, 4> WorkList;
1682     WorkList.push_back(Succ);
1683     WorkList.push_back(Header);
1684     SmallPtrSet<BasicBlock *, 4> Seen;
1685     Seen.insert(Header);
1686     Info.PathIsNoop &=
1687         all_of(*Header, [](Instruction &I) { return !I.mayHaveSideEffects(); });
1688 
1689     while (!WorkList.empty()) {
1690       BasicBlock *Current = WorkList.pop_back_val();
1691       if (!L.contains(Current))
1692         continue;
1693       const auto &SeenIns = Seen.insert(Current);
1694       if (!SeenIns.second)
1695         continue;
1696 
1697       Info.PathIsNoop &= all_of(
1698           *Current, [](Instruction &I) { return !I.mayHaveSideEffects(); });
1699       WorkList.append(succ_begin(Current), succ_end(Current));
1700     }
1701 
1702     // Require at least 2 blocks on a path through the loop. This skips
1703     // paths that directly exit the loop.
1704     if (Seen.size() < 2)
1705       return {};
1706 
1707     // Next, check if there are any MemoryDefs that are on the path through
1708     // the loop (in the Seen set) and they may-alias any of the locations in
1709     // AccessedLocs. If that is the case, they may modify the condition and
1710     // partial unswitching is not possible.
1711     SmallPtrSet<MemoryAccess *, 4> SeenAccesses;
1712     while (!AccessesToCheck.empty()) {
1713       MemoryAccess *Current = AccessesToCheck.pop_back_val();
1714       auto SeenI = SeenAccesses.insert(Current);
1715       if (!SeenI.second || !Seen.contains(Current->getBlock()))
1716         continue;
1717 
1718       // Bail out if exceeded the threshold.
1719       if (SeenAccesses.size() >= MSSAThreshold)
1720         return {};
1721 
1722       // MemoryUse are read-only accesses.
1723       if (isa<MemoryUse>(Current))
1724         continue;
1725 
1726       // For a MemoryDef, check if is aliases any of the location feeding
1727       // the original condition.
1728       if (auto *CurrentDef = dyn_cast<MemoryDef>(Current)) {
1729         if (any_of(AccessedLocs, [&AA, CurrentDef](MemoryLocation &Loc) {
1730               return isModSet(
1731                   AA.getModRefInfo(CurrentDef->getMemoryInst(), Loc));
1732             }))
1733           return {};
1734       }
1735 
1736       for (Use &U : Current->uses())
1737         AccessesToCheck.push_back(cast<MemoryAccess>(U.getUser()));
1738     }
1739 
1740     // We could also allow loops with known trip counts without mustprogress,
1741     // but ScalarEvolution may not be available.
1742     Info.PathIsNoop &= isMustProgress(&L);
1743 
1744     // If the path is considered a no-op so far, check if it reaches a
1745     // single exit block without any phis. This ensures no values from the
1746     // loop are used outside of the loop.
1747     if (Info.PathIsNoop) {
1748       for (auto *Exiting : ExitingBlocks) {
1749         if (!Seen.contains(Exiting))
1750           continue;
1751         for (auto *Succ : successors(Exiting)) {
1752           if (L.contains(Succ))
1753             continue;
1754 
1755           Info.PathIsNoop &= llvm::empty(Succ->phis()) &&
1756                              (!Info.ExitForPath || Info.ExitForPath == Succ);
1757           if (!Info.PathIsNoop)
1758             break;
1759           assert((!Info.ExitForPath || Info.ExitForPath == Succ) &&
1760                  "cannot have multiple exit blocks");
1761           Info.ExitForPath = Succ;
1762         }
1763       }
1764     }
1765     if (!Info.ExitForPath)
1766       Info.PathIsNoop = false;
1767 
1768     Info.InstToDuplicate = InstToDuplicate;
1769     return Info;
1770   };
1771 
1772   // If we branch to the same successor, partial unswitching will not be
1773   // beneficial.
1774   if (TI->getSuccessor(0) == TI->getSuccessor(1))
1775     return {};
1776 
1777   if (auto Info = HasNoClobbersOnPath(TI->getSuccessor(0), L.getHeader(),
1778                                       AccessesToCheck)) {
1779     Info->KnownValue = ConstantInt::getTrue(TI->getContext());
1780     return Info;
1781   }
1782   if (auto Info = HasNoClobbersOnPath(TI->getSuccessor(1), L.getHeader(),
1783                                       AccessesToCheck)) {
1784     Info->KnownValue = ConstantInt::getFalse(TI->getContext());
1785     return Info;
1786   }
1787 
1788   return {};
1789 }
1790