1 //===- DFAJumpThreading.cpp - Threads a switch statement inside a loop ----===//
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 // Transform each threading path to effectively jump thread the DFA. For
10 // example, the CFG below could be transformed as follows, where the cloned
11 // blocks unconditionally branch to the next correct case based on what is
12 // identified in the analysis.
13 //
14 //          sw.bb                        sw.bb
15 //        /   |   \                    /   |   \
16 //   case1  case2  case3          case1  case2  case3
17 //        \   |   /                 |      |      |
18 //       determinator            det.2   det.3  det.1
19 //        br sw.bb                /        |        \
20 //                          sw.bb.2     sw.bb.3     sw.bb.1
21 //                           br case2    br case3    br case1§
22 //
23 // Definitions and Terminology:
24 //
25 // * Threading path:
26 //   a list of basic blocks, the exit state, and the block that determines
27 //   the next state, for which the following notation will be used:
28 //   < path of BBs that form a cycle > [ state, determinator ]
29 //
30 // * Predictable switch:
31 //   The switch variable is always a known constant so that all conditional
32 //   jumps based on switch variable can be converted to unconditional jump.
33 //
34 // * Determinator:
35 //   The basic block that determines the next state of the DFA.
36 //
37 // Representing the optimization in C-like pseudocode: the code pattern on the
38 // left could functionally be transformed to the right pattern if the switch
39 // condition is predictable.
40 //
41 //  X = A                       goto A
42 //  for (...)                   A:
43 //    switch (X)                  ...
44 //      case A                    goto B
45 //        X = B                 B:
46 //      case B                    ...
47 //        X = C                   goto C
48 //
49 // The pass first checks that switch variable X is decided by the control flow
50 // path taken in the loop; for example, in case B, the next value of X is
51 // decided to be C. It then enumerates through all paths in the loop and labels
52 // the basic blocks where the next state is decided.
53 //
54 // Using this information it creates new paths that unconditionally branch to
55 // the next case. This involves cloning code, so it only gets triggered if the
56 // amount of code duplicated is below a threshold.
57 //
58 //===----------------------------------------------------------------------===//
59 
60 #include "llvm/Transforms/Scalar/DFAJumpThreading.h"
61 #include "llvm/ADT/APInt.h"
62 #include "llvm/ADT/DenseMap.h"
63 #include "llvm/ADT/DepthFirstIterator.h"
64 #include "llvm/ADT/SmallSet.h"
65 #include "llvm/ADT/Statistic.h"
66 #include "llvm/Analysis/AssumptionCache.h"
67 #include "llvm/Analysis/CodeMetrics.h"
68 #include "llvm/Analysis/LoopIterator.h"
69 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
70 #include "llvm/Analysis/TargetTransformInfo.h"
71 #include "llvm/IR/CFG.h"
72 #include "llvm/IR/Constants.h"
73 #include "llvm/IR/IntrinsicInst.h"
74 #include "llvm/IR/Verifier.h"
75 #include "llvm/InitializePasses.h"
76 #include "llvm/Pass.h"
77 #include "llvm/Support/CommandLine.h"
78 #include "llvm/Support/Debug.h"
79 #include "llvm/Transforms/Scalar.h"
80 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
81 #include "llvm/Transforms/Utils/Cloning.h"
82 #include "llvm/Transforms/Utils/SSAUpdaterBulk.h"
83 #include "llvm/Transforms/Utils/ValueMapper.h"
84 #include <algorithm>
85 #include <deque>
86 
87 using namespace llvm;
88 
89 #define DEBUG_TYPE "dfa-jump-threading"
90 
91 STATISTIC(NumTransforms, "Number of transformations done");
92 STATISTIC(NumCloned, "Number of blocks cloned");
93 STATISTIC(NumPaths, "Number of individual paths threaded");
94 
95 static cl::opt<bool>
96     ClViewCfgBefore("dfa-jump-view-cfg-before",
97                     cl::desc("View the CFG before DFA Jump Threading"),
98                     cl::Hidden, cl::init(false));
99 
100 static cl::opt<unsigned> MaxPathLength(
101     "dfa-max-path-length",
102     cl::desc("Max number of blocks searched to find a threading path"),
103     cl::Hidden, cl::init(20));
104 
105 static cl::opt<unsigned>
106     CostThreshold("dfa-cost-threshold",
107                   cl::desc("Maximum cost accepted for the transformation"),
108                   cl::Hidden, cl::init(50));
109 
110 namespace {
111 
112 class SelectInstToUnfold {
113   SelectInst *SI;
114   PHINode *SIUse;
115 
116 public:
117   SelectInstToUnfold(SelectInst *SI, PHINode *SIUse) : SI(SI), SIUse(SIUse) {}
118 
119   SelectInst *getInst() { return SI; }
120   PHINode *getUse() { return SIUse; }
121 
122   explicit operator bool() const { return SI && SIUse; }
123 };
124 
125 void unfold(DomTreeUpdater *DTU, SelectInstToUnfold SIToUnfold,
126             std::vector<SelectInstToUnfold> *NewSIsToUnfold,
127             std::vector<BasicBlock *> *NewBBs);
128 
129 class DFAJumpThreading {
130 public:
131   DFAJumpThreading(AssumptionCache *AC, DominatorTree *DT,
132                    TargetTransformInfo *TTI, OptimizationRemarkEmitter *ORE)
133       : AC(AC), DT(DT), TTI(TTI), ORE(ORE) {}
134 
135   bool run(Function &F);
136 
137 private:
138   void
139   unfoldSelectInstrs(DominatorTree *DT,
140                      const SmallVector<SelectInstToUnfold, 4> &SelectInsts) {
141     DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);
142     SmallVector<SelectInstToUnfold, 4> Stack;
143     for (SelectInstToUnfold SIToUnfold : SelectInsts)
144       Stack.push_back(SIToUnfold);
145 
146     while (!Stack.empty()) {
147       SelectInstToUnfold SIToUnfold = Stack.pop_back_val();
148 
149       std::vector<SelectInstToUnfold> NewSIsToUnfold;
150       std::vector<BasicBlock *> NewBBs;
151       unfold(&DTU, SIToUnfold, &NewSIsToUnfold, &NewBBs);
152 
153       // Put newly discovered select instructions into the work list.
154       for (const SelectInstToUnfold &NewSIToUnfold : NewSIsToUnfold)
155         Stack.push_back(NewSIToUnfold);
156     }
157   }
158 
159   AssumptionCache *AC;
160   DominatorTree *DT;
161   TargetTransformInfo *TTI;
162   OptimizationRemarkEmitter *ORE;
163 };
164 
165 class DFAJumpThreadingLegacyPass : public FunctionPass {
166 public:
167   static char ID; // Pass identification
168   DFAJumpThreadingLegacyPass() : FunctionPass(ID) {}
169 
170   void getAnalysisUsage(AnalysisUsage &AU) const override {
171     AU.addRequired<AssumptionCacheTracker>();
172     AU.addRequired<DominatorTreeWrapperPass>();
173     AU.addPreserved<DominatorTreeWrapperPass>();
174     AU.addRequired<TargetTransformInfoWrapperPass>();
175     AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
176   }
177 
178   bool runOnFunction(Function &F) override {
179     if (skipFunction(F))
180       return false;
181 
182     AssumptionCache *AC =
183         &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
184     DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
185     TargetTransformInfo *TTI =
186         &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
187     OptimizationRemarkEmitter *ORE =
188         &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
189 
190     return DFAJumpThreading(AC, DT, TTI, ORE).run(F);
191   }
192 };
193 } // end anonymous namespace
194 
195 char DFAJumpThreadingLegacyPass::ID = 0;
196 INITIALIZE_PASS_BEGIN(DFAJumpThreadingLegacyPass, "dfa-jump-threading",
197                       "DFA Jump Threading", false, false)
198 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
199 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
200 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
201 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
202 INITIALIZE_PASS_END(DFAJumpThreadingLegacyPass, "dfa-jump-threading",
203                     "DFA Jump Threading", false, false)
204 
205 // Public interface to the DFA Jump Threading pass
206 FunctionPass *llvm::createDFAJumpThreadingPass() {
207   return new DFAJumpThreadingLegacyPass();
208 }
209 
210 namespace {
211 
212 /// Create a new basic block and sink \p SIToSink into it.
213 void createBasicBlockAndSinkSelectInst(
214     DomTreeUpdater *DTU, SelectInst *SI, PHINode *SIUse, SelectInst *SIToSink,
215     BasicBlock *EndBlock, StringRef NewBBName, BasicBlock **NewBlock,
216     BranchInst **NewBranch, std::vector<SelectInstToUnfold> *NewSIsToUnfold,
217     std::vector<BasicBlock *> *NewBBs) {
218   assert(SIToSink->hasOneUse());
219   assert(NewBlock);
220   assert(NewBranch);
221   *NewBlock = BasicBlock::Create(SI->getContext(), NewBBName,
222                                  EndBlock->getParent(), EndBlock);
223   NewBBs->push_back(*NewBlock);
224   *NewBranch = BranchInst::Create(EndBlock, *NewBlock);
225   SIToSink->moveBefore(*NewBranch);
226   NewSIsToUnfold->push_back(SelectInstToUnfold(SIToSink, SIUse));
227   DTU->applyUpdates({{DominatorTree::Insert, *NewBlock, EndBlock}});
228 }
229 
230 /// Unfold the select instruction held in \p SIToUnfold by replacing it with
231 /// control flow.
232 ///
233 /// Put newly discovered select instructions into \p NewSIsToUnfold. Put newly
234 /// created basic blocks into \p NewBBs.
235 ///
236 /// TODO: merge it with CodeGenPrepare::optimizeSelectInst() if possible.
237 void unfold(DomTreeUpdater *DTU, SelectInstToUnfold SIToUnfold,
238             std::vector<SelectInstToUnfold> *NewSIsToUnfold,
239             std::vector<BasicBlock *> *NewBBs) {
240   SelectInst *SI = SIToUnfold.getInst();
241   PHINode *SIUse = SIToUnfold.getUse();
242   BasicBlock *StartBlock = SI->getParent();
243   BasicBlock *EndBlock = SIUse->getParent();
244   BranchInst *StartBlockTerm =
245       dyn_cast<BranchInst>(StartBlock->getTerminator());
246 
247   assert(StartBlockTerm && StartBlockTerm->isUnconditional());
248   assert(SI->hasOneUse());
249 
250   // These are the new basic blocks for the conditional branch.
251   // At least one will become an actual new basic block.
252   BasicBlock *TrueBlock = nullptr;
253   BasicBlock *FalseBlock = nullptr;
254   BranchInst *TrueBranch = nullptr;
255   BranchInst *FalseBranch = nullptr;
256 
257   // Sink select instructions to be able to unfold them later.
258   if (SelectInst *SIOp = dyn_cast<SelectInst>(SI->getTrueValue())) {
259     createBasicBlockAndSinkSelectInst(DTU, SI, SIUse, SIOp, EndBlock,
260                                       "si.unfold.true", &TrueBlock, &TrueBranch,
261                                       NewSIsToUnfold, NewBBs);
262   }
263   if (SelectInst *SIOp = dyn_cast<SelectInst>(SI->getFalseValue())) {
264     createBasicBlockAndSinkSelectInst(DTU, SI, SIUse, SIOp, EndBlock,
265                                       "si.unfold.false", &FalseBlock,
266                                       &FalseBranch, NewSIsToUnfold, NewBBs);
267   }
268 
269   // If there was nothing to sink, then arbitrarily choose the 'false' side
270   // for a new input value to the PHI.
271   if (!TrueBlock && !FalseBlock) {
272     FalseBlock = BasicBlock::Create(SI->getContext(), "si.unfold.false",
273                                     EndBlock->getParent(), EndBlock);
274     NewBBs->push_back(FalseBlock);
275     BranchInst::Create(EndBlock, FalseBlock);
276     DTU->applyUpdates({{DominatorTree::Insert, FalseBlock, EndBlock}});
277   }
278 
279   // Insert the real conditional branch based on the original condition.
280   // If we did not create a new block for one of the 'true' or 'false' paths
281   // of the condition, it means that side of the branch goes to the end block
282   // directly and the path originates from the start block from the point of
283   // view of the new PHI.
284   BasicBlock *TT = EndBlock;
285   BasicBlock *FT = EndBlock;
286   if (TrueBlock && FalseBlock) {
287     // A diamond.
288     TT = TrueBlock;
289     FT = FalseBlock;
290 
291     // Update the phi node of SI.
292     SIUse->removeIncomingValue(StartBlock, /* DeletePHIIfEmpty = */ false);
293     SIUse->addIncoming(SI->getTrueValue(), TrueBlock);
294     SIUse->addIncoming(SI->getFalseValue(), FalseBlock);
295 
296     // Update any other PHI nodes in EndBlock.
297     for (PHINode &Phi : EndBlock->phis()) {
298       if (&Phi != SIUse) {
299         Phi.addIncoming(Phi.getIncomingValueForBlock(StartBlock), TrueBlock);
300         Phi.addIncoming(Phi.getIncomingValueForBlock(StartBlock), FalseBlock);
301       }
302     }
303   } else {
304     BasicBlock *NewBlock = nullptr;
305     Value *SIOp1 = SI->getTrueValue();
306     Value *SIOp2 = SI->getFalseValue();
307 
308     // A triangle pointing right.
309     if (!TrueBlock) {
310       NewBlock = FalseBlock;
311       FT = FalseBlock;
312     }
313     // A triangle pointing left.
314     else {
315       NewBlock = TrueBlock;
316       TT = TrueBlock;
317       std::swap(SIOp1, SIOp2);
318     }
319 
320     // Update the phi node of SI.
321     for (unsigned Idx = 0; Idx < SIUse->getNumIncomingValues(); ++Idx) {
322       if (SIUse->getIncomingBlock(Idx) == StartBlock)
323         SIUse->setIncomingValue(Idx, SIOp1);
324     }
325     SIUse->addIncoming(SIOp2, NewBlock);
326 
327     // Update any other PHI nodes in EndBlock.
328     for (auto II = EndBlock->begin(); PHINode *Phi = dyn_cast<PHINode>(II);
329          ++II) {
330       if (Phi != SIUse)
331         Phi->addIncoming(Phi->getIncomingValueForBlock(StartBlock), NewBlock);
332     }
333   }
334   StartBlockTerm->eraseFromParent();
335   BranchInst::Create(TT, FT, SI->getCondition(), StartBlock);
336   DTU->applyUpdates({{DominatorTree::Insert, StartBlock, TT},
337                      {DominatorTree::Insert, StartBlock, FT}});
338 
339   // The select is now dead.
340   SI->eraseFromParent();
341 }
342 
343 struct ClonedBlock {
344   BasicBlock *BB;
345   uint64_t State; ///< \p State corresponds to the next value of a switch stmnt.
346 };
347 
348 typedef std::deque<BasicBlock *> PathType;
349 typedef std::vector<PathType> PathsType;
350 typedef SmallPtrSet<const BasicBlock *, 8> VisitedBlocks;
351 typedef std::vector<ClonedBlock> CloneList;
352 
353 // This data structure keeps track of all blocks that have been cloned.  If two
354 // different ThreadingPaths clone the same block for a certain state it should
355 // be reused, and it can be looked up in this map.
356 typedef DenseMap<BasicBlock *, CloneList> DuplicateBlockMap;
357 
358 // This map keeps track of all the new definitions for an instruction. This
359 // information is needed when restoring SSA form after cloning blocks.
360 typedef MapVector<Instruction *, std::vector<Instruction *>> DefMap;
361 
362 inline raw_ostream &operator<<(raw_ostream &OS, const PathType &Path) {
363   OS << "< ";
364   for (const BasicBlock *BB : Path) {
365     std::string BBName;
366     if (BB->hasName())
367       raw_string_ostream(BBName) << BB->getName();
368     else
369       raw_string_ostream(BBName) << BB;
370     OS << BBName << " ";
371   }
372   OS << ">";
373   return OS;
374 }
375 
376 /// ThreadingPath is a path in the control flow of a loop that can be threaded
377 /// by cloning necessary basic blocks and replacing conditional branches with
378 /// unconditional ones. A threading path includes a list of basic blocks, the
379 /// exit state, and the block that determines the next state.
380 struct ThreadingPath {
381   /// Exit value is DFA's exit state for the given path.
382   uint64_t getExitValue() const { return ExitVal; }
383   void setExitValue(const ConstantInt *V) {
384     ExitVal = V->getZExtValue();
385     IsExitValSet = true;
386   }
387   bool isExitValueSet() const { return IsExitValSet; }
388 
389   /// Determinator is the basic block that determines the next state of the DFA.
390   const BasicBlock *getDeterminatorBB() const { return DBB; }
391   void setDeterminator(const BasicBlock *BB) { DBB = BB; }
392 
393   /// Path is a list of basic blocks.
394   const PathType &getPath() const { return Path; }
395   void setPath(const PathType &NewPath) { Path = NewPath; }
396 
397   void print(raw_ostream &OS) const {
398     OS << Path << " [ " << ExitVal << ", " << DBB->getName() << " ]";
399   }
400 
401 private:
402   PathType Path;
403   uint64_t ExitVal;
404   const BasicBlock *DBB = nullptr;
405   bool IsExitValSet = false;
406 };
407 
408 #ifndef NDEBUG
409 inline raw_ostream &operator<<(raw_ostream &OS, const ThreadingPath &TPath) {
410   TPath.print(OS);
411   return OS;
412 }
413 #endif
414 
415 struct MainSwitch {
416   MainSwitch(SwitchInst *SI, OptimizationRemarkEmitter *ORE) {
417     if (isPredictable(SI)) {
418       Instr = SI;
419     } else {
420       ORE->emit([&]() {
421         return OptimizationRemarkMissed(DEBUG_TYPE, "SwitchNotPredictable", SI)
422                << "Switch instruction is not predictable.";
423       });
424     }
425   }
426 
427   virtual ~MainSwitch() = default;
428 
429   SwitchInst *getInstr() const { return Instr; }
430   const SmallVector<SelectInstToUnfold, 4> getSelectInsts() {
431     return SelectInsts;
432   }
433 
434 private:
435   /// Do a use-def chain traversal. Make sure the value of the switch variable
436   /// is always a known constant. This means that all conditional jumps based on
437   /// switch variable can be converted to unconditional jumps.
438   bool isPredictable(const SwitchInst *SI) {
439     std::deque<Instruction *> Q;
440     SmallSet<Value *, 16> SeenValues;
441     SelectInsts.clear();
442 
443     Value *FirstDef = SI->getOperand(0);
444     auto *Inst = dyn_cast<Instruction>(FirstDef);
445 
446     // If this is a function argument or another non-instruction, then give up.
447     // We are interested in loop local variables.
448     if (!Inst)
449       return false;
450 
451     // Require the first definition to be a PHINode
452     if (!isa<PHINode>(Inst))
453       return false;
454 
455     LLVM_DEBUG(dbgs() << "\tisPredictable() FirstDef: " << *Inst << "\n");
456 
457     Q.push_back(Inst);
458     SeenValues.insert(FirstDef);
459 
460     while (!Q.empty()) {
461       Instruction *Current = Q.front();
462       Q.pop_front();
463 
464       if (auto *Phi = dyn_cast<PHINode>(Current)) {
465         for (Value *Incoming : Phi->incoming_values()) {
466           if (!isPredictableValue(Incoming, SeenValues))
467             return false;
468           addInstToQueue(Incoming, Q, SeenValues);
469         }
470         LLVM_DEBUG(dbgs() << "\tisPredictable() phi: " << *Phi << "\n");
471       } else if (SelectInst *SelI = dyn_cast<SelectInst>(Current)) {
472         if (!isValidSelectInst(SelI))
473           return false;
474         if (!isPredictableValue(SelI->getTrueValue(), SeenValues) ||
475             !isPredictableValue(SelI->getFalseValue(), SeenValues)) {
476           return false;
477         }
478         addInstToQueue(SelI->getTrueValue(), Q, SeenValues);
479         addInstToQueue(SelI->getFalseValue(), Q, SeenValues);
480         LLVM_DEBUG(dbgs() << "\tisPredictable() select: " << *SelI << "\n");
481         if (auto *SelIUse = dyn_cast<PHINode>(SelI->user_back()))
482           SelectInsts.push_back(SelectInstToUnfold(SelI, SelIUse));
483       } else {
484         // If it is neither a phi nor a select, then we give up.
485         return false;
486       }
487     }
488 
489     return true;
490   }
491 
492   bool isPredictableValue(Value *InpVal, SmallSet<Value *, 16> &SeenValues) {
493     if (SeenValues.contains(InpVal))
494       return true;
495 
496     if (isa<ConstantInt>(InpVal))
497       return true;
498 
499     // If this is a function argument or another non-instruction, then give up.
500     if (!isa<Instruction>(InpVal))
501       return false;
502 
503     return true;
504   }
505 
506   void addInstToQueue(Value *Val, std::deque<Instruction *> &Q,
507                       SmallSet<Value *, 16> &SeenValues) {
508     if (SeenValues.contains(Val))
509       return;
510     if (Instruction *I = dyn_cast<Instruction>(Val))
511       Q.push_back(I);
512     SeenValues.insert(Val);
513   }
514 
515   bool isValidSelectInst(SelectInst *SI) {
516     if (!SI->hasOneUse())
517       return false;
518 
519     Instruction *SIUse = dyn_cast<Instruction>(SI->user_back());
520     // The use of the select inst should be either a phi or another select.
521     if (!SIUse && !(isa<PHINode>(SIUse) || isa<SelectInst>(SIUse)))
522       return false;
523 
524     BasicBlock *SIBB = SI->getParent();
525 
526     // Currently, we can only expand select instructions in basic blocks with
527     // one successor.
528     BranchInst *SITerm = dyn_cast<BranchInst>(SIBB->getTerminator());
529     if (!SITerm || !SITerm->isUnconditional())
530       return false;
531 
532     if (isa<PHINode>(SIUse) &&
533         SIBB->getSingleSuccessor() != cast<Instruction>(SIUse)->getParent())
534       return false;
535 
536     // If select will not be sunk during unfolding, and it is in the same basic
537     // block as another state defining select, then cannot unfold both.
538     for (SelectInstToUnfold SIToUnfold : SelectInsts) {
539       SelectInst *PrevSI = SIToUnfold.getInst();
540       if (PrevSI->getTrueValue() != SI && PrevSI->getFalseValue() != SI &&
541           PrevSI->getParent() == SI->getParent())
542         return false;
543     }
544 
545     return true;
546   }
547 
548   SwitchInst *Instr = nullptr;
549   SmallVector<SelectInstToUnfold, 4> SelectInsts;
550 };
551 
552 struct AllSwitchPaths {
553   AllSwitchPaths(const MainSwitch *MSwitch, OptimizationRemarkEmitter *ORE)
554       : Switch(MSwitch->getInstr()), SwitchBlock(Switch->getParent()),
555         ORE(ORE) {}
556 
557   std::vector<ThreadingPath> &getThreadingPaths() { return TPaths; }
558   unsigned getNumThreadingPaths() { return TPaths.size(); }
559   SwitchInst *getSwitchInst() { return Switch; }
560   BasicBlock *getSwitchBlock() { return SwitchBlock; }
561 
562   void run() {
563     VisitedBlocks Visited;
564     PathsType LoopPaths = paths(SwitchBlock, Visited, /* PathDepth = */ 1);
565     StateDefMap StateDef = getStateDefMap();
566 
567     for (PathType Path : LoopPaths) {
568       ThreadingPath TPath;
569 
570       const BasicBlock *PrevBB = Path.back();
571       for (const BasicBlock *BB : Path) {
572         if (StateDef.count(BB) != 0) {
573           const PHINode *Phi = dyn_cast<PHINode>(StateDef[BB]);
574           assert(Phi && "Expected a state-defining instr to be a phi node.");
575 
576           const Value *V = Phi->getIncomingValueForBlock(PrevBB);
577           if (const ConstantInt *C = dyn_cast<const ConstantInt>(V)) {
578             TPath.setExitValue(C);
579             TPath.setDeterminator(BB);
580             TPath.setPath(Path);
581           }
582         }
583 
584         // Switch block is the determinator, this is the final exit value.
585         if (TPath.isExitValueSet() && BB == Path.front())
586           break;
587 
588         PrevBB = BB;
589       }
590 
591       if (TPath.isExitValueSet() && isSupported(TPath))
592         TPaths.push_back(TPath);
593     }
594   }
595 
596 private:
597   // Value: an instruction that defines a switch state;
598   // Key: the parent basic block of that instruction.
599   typedef DenseMap<const BasicBlock *, const PHINode *> StateDefMap;
600 
601   PathsType paths(BasicBlock *BB, VisitedBlocks &Visited,
602                   unsigned PathDepth) const {
603     PathsType Res;
604 
605     // Stop exploring paths after visiting MaxPathLength blocks
606     if (PathDepth > MaxPathLength) {
607       ORE->emit([&]() {
608         return OptimizationRemarkAnalysis(DEBUG_TYPE, "MaxPathLengthReached",
609                                           Switch)
610                << "Exploration stopped after visiting MaxPathLength="
611                << ore::NV("MaxPathLength", MaxPathLength) << " blocks.";
612       });
613       return Res;
614     }
615 
616     Visited.insert(BB);
617 
618     // Some blocks have multiple edges to the same successor, and this set
619     // is used to prevent a duplicate path from being generated
620     SmallSet<BasicBlock *, 4> Successors;
621     for (BasicBlock *Succ : successors(BB)) {
622       if (!Successors.insert(Succ).second)
623         continue;
624 
625       // Found a cycle through the SwitchBlock
626       if (Succ == SwitchBlock) {
627         Res.push_back({BB});
628         continue;
629       }
630 
631       // We have encountered a cycle, do not get caught in it
632       if (Visited.contains(Succ))
633         continue;
634 
635       PathsType SuccPaths = paths(Succ, Visited, PathDepth + 1);
636       for (PathType Path : SuccPaths) {
637         PathType NewPath(Path);
638         NewPath.push_front(BB);
639         Res.push_back(NewPath);
640       }
641     }
642     // This block could now be visited again from a different predecessor. Note
643     // that this will result in exponential runtime. Subpaths could possibly be
644     // cached but it takes a lot of memory to store them.
645     Visited.erase(BB);
646     return Res;
647   }
648 
649   /// Walk the use-def chain and collect all the state-defining instructions.
650   StateDefMap getStateDefMap() const {
651     StateDefMap Res;
652 
653     Value *FirstDef = Switch->getOperand(0);
654 
655     assert(isa<PHINode>(FirstDef) && "After select unfolding, all state "
656                                      "definitions are expected to be phi "
657                                      "nodes.");
658 
659     SmallVector<PHINode *, 8> Stack;
660     Stack.push_back(dyn_cast<PHINode>(FirstDef));
661     SmallSet<Value *, 16> SeenValues;
662 
663     while (!Stack.empty()) {
664       PHINode *CurPhi = Stack.pop_back_val();
665 
666       Res[CurPhi->getParent()] = CurPhi;
667       SeenValues.insert(CurPhi);
668 
669       for (Value *Incoming : CurPhi->incoming_values()) {
670         if (Incoming == FirstDef || isa<ConstantInt>(Incoming) ||
671             SeenValues.contains(Incoming)) {
672           continue;
673         }
674 
675         assert(isa<PHINode>(Incoming) && "After select unfolding, all state "
676                                          "definitions are expected to be phi "
677                                          "nodes.");
678 
679         Stack.push_back(cast<PHINode>(Incoming));
680       }
681     }
682 
683     return Res;
684   }
685 
686   /// The determinator BB should precede the switch-defining BB.
687   ///
688   /// Otherwise, it is possible that the state defined in the determinator block
689   /// defines the state for the next iteration of the loop, rather than for the
690   /// current one.
691   ///
692   /// Currently supported paths:
693   /// \code
694   /// < switch bb1 determ def > [ 42, determ ]
695   /// < switch_and_def bb1 determ > [ 42, determ ]
696   /// < switch_and_def_and_determ bb1 > [ 42, switch_and_def_and_determ ]
697   /// \endcode
698   ///
699   /// Unsupported paths:
700   /// \code
701   /// < switch bb1 def determ > [ 43, determ ]
702   /// < switch_and_determ bb1 def > [ 43, switch_and_determ ]
703   /// \endcode
704   bool isSupported(const ThreadingPath &TPath) {
705     Instruction *SwitchCondI = dyn_cast<Instruction>(Switch->getCondition());
706     assert(SwitchCondI);
707     if (!SwitchCondI)
708       return false;
709 
710     const BasicBlock *SwitchCondDefBB = SwitchCondI->getParent();
711     const BasicBlock *SwitchCondUseBB = Switch->getParent();
712     const BasicBlock *DeterminatorBB = TPath.getDeterminatorBB();
713 
714     assert(
715         SwitchCondUseBB == TPath.getPath().front() &&
716         "The first BB in a threading path should have the switch instruction");
717     if (SwitchCondUseBB != TPath.getPath().front())
718       return false;
719 
720     // Make DeterminatorBB the first element in Path.
721     PathType Path = TPath.getPath();
722     auto ItDet = std::find(Path.begin(), Path.end(), DeterminatorBB);
723     std::rotate(Path.begin(), ItDet, Path.end());
724 
725     bool IsDetBBSeen = false;
726     bool IsDefBBSeen = false;
727     bool IsUseBBSeen = false;
728     for (BasicBlock *BB : Path) {
729       if (BB == DeterminatorBB)
730         IsDetBBSeen = true;
731       if (BB == SwitchCondDefBB)
732         IsDefBBSeen = true;
733       if (BB == SwitchCondUseBB)
734         IsUseBBSeen = true;
735       if (IsDetBBSeen && IsUseBBSeen && !IsDefBBSeen)
736         return false;
737     }
738 
739     return true;
740   }
741 
742   SwitchInst *Switch;
743   BasicBlock *SwitchBlock;
744   OptimizationRemarkEmitter *ORE;
745   std::vector<ThreadingPath> TPaths;
746 };
747 
748 struct TransformDFA {
749   TransformDFA(AllSwitchPaths *SwitchPaths, DominatorTree *DT,
750                AssumptionCache *AC, TargetTransformInfo *TTI,
751                OptimizationRemarkEmitter *ORE,
752                SmallPtrSet<const Value *, 32> EphValues)
753       : SwitchPaths(SwitchPaths), DT(DT), AC(AC), TTI(TTI), ORE(ORE),
754         EphValues(EphValues) {}
755 
756   void run() {
757     if (isLegalAndProfitableToTransform()) {
758       createAllExitPaths();
759       NumTransforms++;
760     }
761   }
762 
763 private:
764   /// This function performs both a legality check and profitability check at
765   /// the same time since it is convenient to do so. It iterates through all
766   /// blocks that will be cloned, and keeps track of the duplication cost. It
767   /// also returns false if it is illegal to clone some required block.
768   bool isLegalAndProfitableToTransform() {
769     CodeMetrics Metrics;
770     SwitchInst *Switch = SwitchPaths->getSwitchInst();
771 
772     // Note that DuplicateBlockMap is not being used as intended here. It is
773     // just being used to ensure (BB, State) pairs are only counted once.
774     DuplicateBlockMap DuplicateMap;
775 
776     for (ThreadingPath &TPath : SwitchPaths->getThreadingPaths()) {
777       PathType PathBBs = TPath.getPath();
778       uint64_t NextState = TPath.getExitValue();
779       const BasicBlock *Determinator = TPath.getDeterminatorBB();
780 
781       // Update Metrics for the Switch block, this is always cloned
782       BasicBlock *BB = SwitchPaths->getSwitchBlock();
783       BasicBlock *VisitedBB = getClonedBB(BB, NextState, DuplicateMap);
784       if (!VisitedBB) {
785         Metrics.analyzeBasicBlock(BB, *TTI, EphValues);
786         DuplicateMap[BB].push_back({BB, NextState});
787       }
788 
789       // If the Switch block is the Determinator, then we can continue since
790       // this is the only block that is cloned and we already counted for it.
791       if (PathBBs.front() == Determinator)
792         continue;
793 
794       // Otherwise update Metrics for all blocks that will be cloned. If any
795       // block is already cloned and would be reused, don't double count it.
796       auto DetIt = std::find(PathBBs.begin(), PathBBs.end(), Determinator);
797       for (auto BBIt = DetIt; BBIt != PathBBs.end(); BBIt++) {
798         BB = *BBIt;
799         VisitedBB = getClonedBB(BB, NextState, DuplicateMap);
800         if (VisitedBB)
801           continue;
802         Metrics.analyzeBasicBlock(BB, *TTI, EphValues);
803         DuplicateMap[BB].push_back({BB, NextState});
804       }
805 
806       if (Metrics.notDuplicatable) {
807         LLVM_DEBUG(dbgs() << "DFA Jump Threading: Not jump threading, contains "
808                           << "non-duplicatable instructions.\n");
809         ORE->emit([&]() {
810           return OptimizationRemarkMissed(DEBUG_TYPE, "NonDuplicatableInst",
811                                           Switch)
812                  << "Contains non-duplicatable instructions.";
813         });
814         return false;
815       }
816 
817       if (Metrics.convergent) {
818         LLVM_DEBUG(dbgs() << "DFA Jump Threading: Not jump threading, contains "
819                           << "convergent instructions.\n");
820         ORE->emit([&]() {
821           return OptimizationRemarkMissed(DEBUG_TYPE, "ConvergentInst", Switch)
822                  << "Contains convergent instructions.";
823         });
824         return false;
825       }
826     }
827 
828     unsigned DuplicationCost = 0;
829 
830     unsigned JumpTableSize = 0;
831     TTI->getEstimatedNumberOfCaseClusters(*Switch, JumpTableSize, nullptr,
832                                           nullptr);
833     if (JumpTableSize == 0) {
834       // Factor in the number of conditional branches reduced from jump
835       // threading. Assume that lowering the switch block is implemented by
836       // using binary search, hence the LogBase2().
837       unsigned CondBranches =
838           APInt(32, Switch->getNumSuccessors()).ceilLogBase2();
839       DuplicationCost = Metrics.NumInsts / CondBranches;
840     } else {
841       // Compared with jump tables, the DFA optimizer removes an indirect branch
842       // on each loop iteration, thus making branch prediction more precise. The
843       // more branch targets there are, the more likely it is for the branch
844       // predictor to make a mistake, and the more benefit there is in the DFA
845       // optimizer. Thus, the more branch targets there are, the lower is the
846       // cost of the DFA opt.
847       DuplicationCost = Metrics.NumInsts / JumpTableSize;
848     }
849 
850     LLVM_DEBUG(dbgs() << "\nDFA Jump Threading: Cost to jump thread block "
851                       << SwitchPaths->getSwitchBlock()->getName()
852                       << " is: " << DuplicationCost << "\n\n");
853 
854     if (DuplicationCost > CostThreshold) {
855       LLVM_DEBUG(dbgs() << "Not jump threading, duplication cost exceeds the "
856                         << "cost threshold.\n");
857       ORE->emit([&]() {
858         return OptimizationRemarkMissed(DEBUG_TYPE, "NotProfitable", Switch)
859                << "Duplication cost exceeds the cost threshold (cost="
860                << ore::NV("Cost", DuplicationCost)
861                << ", threshold=" << ore::NV("Threshold", CostThreshold) << ").";
862       });
863       return false;
864     }
865 
866     ORE->emit([&]() {
867       return OptimizationRemark(DEBUG_TYPE, "JumpThreaded", Switch)
868              << "Switch statement jump-threaded.";
869     });
870 
871     return true;
872   }
873 
874   /// Transform each threading path to effectively jump thread the DFA.
875   void createAllExitPaths() {
876     DomTreeUpdater DTU(*DT, DomTreeUpdater::UpdateStrategy::Eager);
877 
878     // Move the switch block to the end of the path, since it will be duplicated
879     BasicBlock *SwitchBlock = SwitchPaths->getSwitchBlock();
880     for (ThreadingPath &TPath : SwitchPaths->getThreadingPaths()) {
881       LLVM_DEBUG(dbgs() << TPath << "\n");
882       PathType NewPath(TPath.getPath());
883       NewPath.push_back(SwitchBlock);
884       TPath.setPath(NewPath);
885     }
886 
887     // Transform the ThreadingPaths and keep track of the cloned values
888     DuplicateBlockMap DuplicateMap;
889     DefMap NewDefs;
890 
891     SmallSet<BasicBlock *, 16> BlocksToClean;
892     for (BasicBlock *BB : successors(SwitchBlock))
893       BlocksToClean.insert(BB);
894 
895     for (ThreadingPath &TPath : SwitchPaths->getThreadingPaths()) {
896       createExitPath(NewDefs, TPath, DuplicateMap, BlocksToClean, &DTU);
897       NumPaths++;
898     }
899 
900     // After all paths are cloned, now update the last successor of the cloned
901     // path so it skips over the switch statement
902     for (ThreadingPath &TPath : SwitchPaths->getThreadingPaths())
903       updateLastSuccessor(TPath, DuplicateMap, &DTU);
904 
905     // For each instruction that was cloned and used outside, update its uses
906     updateSSA(NewDefs);
907 
908     // Clean PHI Nodes for the newly created blocks
909     for (BasicBlock *BB : BlocksToClean)
910       cleanPhiNodes(BB);
911   }
912 
913   /// For a specific ThreadingPath \p Path, create an exit path starting from
914   /// the determinator block.
915   ///
916   /// To remember the correct destination, we have to duplicate blocks
917   /// corresponding to each state. Also update the terminating instruction of
918   /// the predecessors, and phis in the successor blocks.
919   void createExitPath(DefMap &NewDefs, ThreadingPath &Path,
920                       DuplicateBlockMap &DuplicateMap,
921                       SmallSet<BasicBlock *, 16> &BlocksToClean,
922                       DomTreeUpdater *DTU) {
923     uint64_t NextState = Path.getExitValue();
924     const BasicBlock *Determinator = Path.getDeterminatorBB();
925     PathType PathBBs = Path.getPath();
926 
927     // Don't select the placeholder block in front
928     if (PathBBs.front() == Determinator)
929       PathBBs.pop_front();
930 
931     auto DetIt = std::find(PathBBs.begin(), PathBBs.end(), Determinator);
932     auto Prev = std::prev(DetIt);
933     BasicBlock *PrevBB = *Prev;
934     for (auto BBIt = DetIt; BBIt != PathBBs.end(); BBIt++) {
935       BasicBlock *BB = *BBIt;
936       BlocksToClean.insert(BB);
937 
938       // We already cloned BB for this NextState, now just update the branch
939       // and continue.
940       BasicBlock *NextBB = getClonedBB(BB, NextState, DuplicateMap);
941       if (NextBB) {
942         updatePredecessor(PrevBB, BB, NextBB, DTU);
943         PrevBB = NextBB;
944         continue;
945       }
946 
947       // Clone the BB and update the successor of Prev to jump to the new block
948       BasicBlock *NewBB = cloneBlockAndUpdatePredecessor(
949           BB, PrevBB, NextState, DuplicateMap, NewDefs, DTU);
950       DuplicateMap[BB].push_back({NewBB, NextState});
951       BlocksToClean.insert(NewBB);
952       PrevBB = NewBB;
953     }
954   }
955 
956   /// Restore SSA form after cloning blocks.
957   ///
958   /// Each cloned block creates new defs for a variable, and the uses need to be
959   /// updated to reflect this. The uses may be replaced with a cloned value, or
960   /// some derived phi instruction. Note that all uses of a value defined in the
961   /// same block were already remapped when cloning the block.
962   void updateSSA(DefMap &NewDefs) {
963     SSAUpdaterBulk SSAUpdate;
964     SmallVector<Use *, 16> UsesToRename;
965 
966     for (auto KV : NewDefs) {
967       Instruction *I = KV.first;
968       BasicBlock *BB = I->getParent();
969       std::vector<Instruction *> Cloned = KV.second;
970 
971       // Scan all uses of this instruction to see if it is used outside of its
972       // block, and if so, record them in UsesToRename.
973       for (Use &U : I->uses()) {
974         Instruction *User = cast<Instruction>(U.getUser());
975         if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
976           if (UserPN->getIncomingBlock(U) == BB)
977             continue;
978         } else if (User->getParent() == BB) {
979           continue;
980         }
981 
982         UsesToRename.push_back(&U);
983       }
984 
985       // If there are no uses outside the block, we're done with this
986       // instruction.
987       if (UsesToRename.empty())
988         continue;
989       LLVM_DEBUG(dbgs() << "DFA-JT: Renaming non-local uses of: " << *I
990                         << "\n");
991 
992       // We found a use of I outside of BB.  Rename all uses of I that are
993       // outside its block to be uses of the appropriate PHI node etc.  See
994       // ValuesInBlocks with the values we know.
995       unsigned VarNum = SSAUpdate.AddVariable(I->getName(), I->getType());
996       SSAUpdate.AddAvailableValue(VarNum, BB, I);
997       for (Instruction *New : Cloned)
998         SSAUpdate.AddAvailableValue(VarNum, New->getParent(), New);
999 
1000       while (!UsesToRename.empty())
1001         SSAUpdate.AddUse(VarNum, UsesToRename.pop_back_val());
1002 
1003       LLVM_DEBUG(dbgs() << "\n");
1004     }
1005     // SSAUpdater handles phi placement and renaming uses with the appropriate
1006     // value.
1007     SSAUpdate.RewriteAllUses(DT);
1008   }
1009 
1010   /// Clones a basic block, and adds it to the CFG.
1011   ///
1012   /// This function also includes updating phi nodes in the successors of the
1013   /// BB, and remapping uses that were defined locally in the cloned BB.
1014   BasicBlock *cloneBlockAndUpdatePredecessor(BasicBlock *BB, BasicBlock *PrevBB,
1015                                              uint64_t NextState,
1016                                              DuplicateBlockMap &DuplicateMap,
1017                                              DefMap &NewDefs,
1018                                              DomTreeUpdater *DTU) {
1019     ValueToValueMapTy VMap;
1020     BasicBlock *NewBB = CloneBasicBlock(
1021         BB, VMap, ".jt" + std::to_string(NextState), BB->getParent());
1022     NewBB->moveAfter(BB);
1023     NumCloned++;
1024 
1025     for (Instruction &I : *NewBB) {
1026       // Do not remap operands of PHINode in case a definition in BB is an
1027       // incoming value to a phi in the same block. This incoming value will
1028       // be renamed later while restoring SSA.
1029       if (isa<PHINode>(&I))
1030         continue;
1031       RemapInstruction(&I, VMap,
1032                        RF_IgnoreMissingLocals | RF_NoModuleLevelChanges);
1033       if (AssumeInst *II = dyn_cast<AssumeInst>(&I))
1034         AC->registerAssumption(II);
1035     }
1036 
1037     updateSuccessorPhis(BB, NewBB, NextState, VMap, DuplicateMap);
1038     updatePredecessor(PrevBB, BB, NewBB, DTU);
1039     updateDefMap(NewDefs, VMap);
1040 
1041     // Add all successors to the DominatorTree
1042     SmallPtrSet<BasicBlock *, 4> SuccSet;
1043     for (auto *SuccBB : successors(NewBB)) {
1044       if (SuccSet.insert(SuccBB).second)
1045         DTU->applyUpdates({{DominatorTree::Insert, NewBB, SuccBB}});
1046     }
1047     SuccSet.clear();
1048     return NewBB;
1049   }
1050 
1051   /// Update the phi nodes in BB's successors.
1052   ///
1053   /// This means creating a new incoming value from NewBB with the new
1054   /// instruction wherever there is an incoming value from BB.
1055   void updateSuccessorPhis(BasicBlock *BB, BasicBlock *ClonedBB,
1056                            uint64_t NextState, ValueToValueMapTy &VMap,
1057                            DuplicateBlockMap &DuplicateMap) {
1058     std::vector<BasicBlock *> BlocksToUpdate;
1059 
1060     // If BB is the last block in the path, we can simply update the one case
1061     // successor that will be reached.
1062     if (BB == SwitchPaths->getSwitchBlock()) {
1063       SwitchInst *Switch = SwitchPaths->getSwitchInst();
1064       BasicBlock *NextCase = getNextCaseSuccessor(Switch, NextState);
1065       BlocksToUpdate.push_back(NextCase);
1066       BasicBlock *ClonedSucc = getClonedBB(NextCase, NextState, DuplicateMap);
1067       if (ClonedSucc)
1068         BlocksToUpdate.push_back(ClonedSucc);
1069     }
1070     // Otherwise update phis in all successors.
1071     else {
1072       for (BasicBlock *Succ : successors(BB)) {
1073         BlocksToUpdate.push_back(Succ);
1074 
1075         // Check if a successor has already been cloned for the particular exit
1076         // value. In this case if a successor was already cloned, the phi nodes
1077         // in the cloned block should be updated directly.
1078         BasicBlock *ClonedSucc = getClonedBB(Succ, NextState, DuplicateMap);
1079         if (ClonedSucc)
1080           BlocksToUpdate.push_back(ClonedSucc);
1081       }
1082     }
1083 
1084     // If there is a phi with an incoming value from BB, create a new incoming
1085     // value for the new predecessor ClonedBB. The value will either be the same
1086     // value from BB or a cloned value.
1087     for (BasicBlock *Succ : BlocksToUpdate) {
1088       for (auto II = Succ->begin(); PHINode *Phi = dyn_cast<PHINode>(II);
1089            ++II) {
1090         Value *Incoming = Phi->getIncomingValueForBlock(BB);
1091         if (Incoming) {
1092           if (isa<Constant>(Incoming)) {
1093             Phi->addIncoming(Incoming, ClonedBB);
1094             continue;
1095           }
1096           Value *ClonedVal = VMap[Incoming];
1097           if (ClonedVal)
1098             Phi->addIncoming(ClonedVal, ClonedBB);
1099           else
1100             Phi->addIncoming(Incoming, ClonedBB);
1101         }
1102       }
1103     }
1104   }
1105 
1106   /// Sets the successor of PrevBB to be NewBB instead of OldBB. Note that all
1107   /// other successors are kept as well.
1108   void updatePredecessor(BasicBlock *PrevBB, BasicBlock *OldBB,
1109                          BasicBlock *NewBB, DomTreeUpdater *DTU) {
1110     // When a path is reused, there is a chance that predecessors were already
1111     // updated before. Check if the predecessor needs to be updated first.
1112     if (!isPredecessor(OldBB, PrevBB))
1113       return;
1114 
1115     Instruction *PrevTerm = PrevBB->getTerminator();
1116     for (unsigned Idx = 0; Idx < PrevTerm->getNumSuccessors(); Idx++) {
1117       if (PrevTerm->getSuccessor(Idx) == OldBB) {
1118         OldBB->removePredecessor(PrevBB, /* KeepOneInputPHIs = */ true);
1119         PrevTerm->setSuccessor(Idx, NewBB);
1120       }
1121     }
1122     DTU->applyUpdates({{DominatorTree::Delete, PrevBB, OldBB},
1123                        {DominatorTree::Insert, PrevBB, NewBB}});
1124   }
1125 
1126   /// Add new value mappings to the DefMap to keep track of all new definitions
1127   /// for a particular instruction. These will be used while updating SSA form.
1128   void updateDefMap(DefMap &NewDefs, ValueToValueMapTy &VMap) {
1129     SmallVector<std::pair<Instruction *, Instruction *>> NewDefsVector;
1130     NewDefsVector.reserve(VMap.size());
1131 
1132     for (auto Entry : VMap) {
1133       Instruction *Inst =
1134           dyn_cast<Instruction>(const_cast<Value *>(Entry.first));
1135       if (!Inst || !Entry.second || isa<BranchInst>(Inst) ||
1136           isa<SwitchInst>(Inst)) {
1137         continue;
1138       }
1139 
1140       Instruction *Cloned = dyn_cast<Instruction>(Entry.second);
1141       if (!Cloned)
1142         continue;
1143 
1144       NewDefsVector.push_back({Inst, Cloned});
1145     }
1146 
1147     // Sort the defs to get deterministic insertion order into NewDefs.
1148     sort(NewDefsVector, [](const auto &LHS, const auto &RHS) {
1149       if (LHS.first == RHS.first)
1150         return LHS.second->comesBefore(RHS.second);
1151       return LHS.first->comesBefore(RHS.first);
1152     });
1153 
1154     for (const auto &KV : NewDefsVector)
1155       NewDefs[KV.first].push_back(KV.second);
1156   }
1157 
1158   /// Update the last branch of a particular cloned path to point to the correct
1159   /// case successor.
1160   ///
1161   /// Note that this is an optional step and would have been done in later
1162   /// optimizations, but it makes the CFG significantly easier to work with.
1163   void updateLastSuccessor(ThreadingPath &TPath,
1164                            DuplicateBlockMap &DuplicateMap,
1165                            DomTreeUpdater *DTU) {
1166     uint64_t NextState = TPath.getExitValue();
1167     BasicBlock *BB = TPath.getPath().back();
1168     BasicBlock *LastBlock = getClonedBB(BB, NextState, DuplicateMap);
1169 
1170     // Note multiple paths can end at the same block so check that it is not
1171     // updated yet
1172     if (!isa<SwitchInst>(LastBlock->getTerminator()))
1173       return;
1174     SwitchInst *Switch = cast<SwitchInst>(LastBlock->getTerminator());
1175     BasicBlock *NextCase = getNextCaseSuccessor(Switch, NextState);
1176 
1177     std::vector<DominatorTree::UpdateType> DTUpdates;
1178     SmallPtrSet<BasicBlock *, 4> SuccSet;
1179     for (BasicBlock *Succ : successors(LastBlock)) {
1180       if (Succ != NextCase && SuccSet.insert(Succ).second)
1181         DTUpdates.push_back({DominatorTree::Delete, LastBlock, Succ});
1182     }
1183 
1184     Switch->eraseFromParent();
1185     BranchInst::Create(NextCase, LastBlock);
1186 
1187     DTU->applyUpdates(DTUpdates);
1188   }
1189 
1190   /// After cloning blocks, some of the phi nodes have extra incoming values
1191   /// that are no longer used. This function removes them.
1192   void cleanPhiNodes(BasicBlock *BB) {
1193     // If BB is no longer reachable, remove any remaining phi nodes
1194     if (pred_empty(BB)) {
1195       std::vector<PHINode *> PhiToRemove;
1196       for (auto II = BB->begin(); PHINode *Phi = dyn_cast<PHINode>(II); ++II) {
1197         PhiToRemove.push_back(Phi);
1198       }
1199       for (PHINode *PN : PhiToRemove) {
1200         PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
1201         PN->eraseFromParent();
1202       }
1203       return;
1204     }
1205 
1206     // Remove any incoming values that come from an invalid predecessor
1207     for (auto II = BB->begin(); PHINode *Phi = dyn_cast<PHINode>(II); ++II) {
1208       std::vector<BasicBlock *> BlocksToRemove;
1209       for (BasicBlock *IncomingBB : Phi->blocks()) {
1210         if (!isPredecessor(BB, IncomingBB))
1211           BlocksToRemove.push_back(IncomingBB);
1212       }
1213       for (BasicBlock *BB : BlocksToRemove)
1214         Phi->removeIncomingValue(BB);
1215     }
1216   }
1217 
1218   /// Checks if BB was already cloned for a particular next state value. If it
1219   /// was then it returns this cloned block, and otherwise null.
1220   BasicBlock *getClonedBB(BasicBlock *BB, uint64_t NextState,
1221                           DuplicateBlockMap &DuplicateMap) {
1222     CloneList ClonedBBs = DuplicateMap[BB];
1223 
1224     // Find an entry in the CloneList with this NextState. If it exists then
1225     // return the corresponding BB
1226     auto It = llvm::find_if(ClonedBBs, [NextState](const ClonedBlock &C) {
1227       return C.State == NextState;
1228     });
1229     return It != ClonedBBs.end() ? (*It).BB : nullptr;
1230   }
1231 
1232   /// Helper to get the successor corresponding to a particular case value for
1233   /// a switch statement.
1234   BasicBlock *getNextCaseSuccessor(SwitchInst *Switch, uint64_t NextState) {
1235     BasicBlock *NextCase = nullptr;
1236     for (auto Case : Switch->cases()) {
1237       if (Case.getCaseValue()->getZExtValue() == NextState) {
1238         NextCase = Case.getCaseSuccessor();
1239         break;
1240       }
1241     }
1242     if (!NextCase)
1243       NextCase = Switch->getDefaultDest();
1244     return NextCase;
1245   }
1246 
1247   /// Returns true if IncomingBB is a predecessor of BB.
1248   bool isPredecessor(BasicBlock *BB, BasicBlock *IncomingBB) {
1249     return llvm::find(predecessors(BB), IncomingBB) != pred_end(BB);
1250   }
1251 
1252   AllSwitchPaths *SwitchPaths;
1253   DominatorTree *DT;
1254   AssumptionCache *AC;
1255   TargetTransformInfo *TTI;
1256   OptimizationRemarkEmitter *ORE;
1257   SmallPtrSet<const Value *, 32> EphValues;
1258   std::vector<ThreadingPath> TPaths;
1259 };
1260 
1261 bool DFAJumpThreading::run(Function &F) {
1262   LLVM_DEBUG(dbgs() << "\nDFA Jump threading: " << F.getName() << "\n");
1263 
1264   if (F.hasOptSize()) {
1265     LLVM_DEBUG(dbgs() << "Skipping due to the 'minsize' attribute\n");
1266     return false;
1267   }
1268 
1269   if (ClViewCfgBefore)
1270     F.viewCFG();
1271 
1272   SmallVector<AllSwitchPaths, 2> ThreadableLoops;
1273   bool MadeChanges = false;
1274 
1275   for (BasicBlock &BB : F) {
1276     auto *SI = dyn_cast<SwitchInst>(BB.getTerminator());
1277     if (!SI)
1278       continue;
1279 
1280     LLVM_DEBUG(dbgs() << "\nCheck if SwitchInst in BB " << BB.getName()
1281                       << " is predictable\n");
1282     MainSwitch Switch(SI, ORE);
1283 
1284     if (!Switch.getInstr())
1285       continue;
1286 
1287     LLVM_DEBUG(dbgs() << "\nSwitchInst in BB " << BB.getName() << " is a "
1288                       << "candidate for jump threading\n");
1289     LLVM_DEBUG(SI->dump());
1290 
1291     unfoldSelectInstrs(DT, Switch.getSelectInsts());
1292     if (!Switch.getSelectInsts().empty())
1293       MadeChanges = true;
1294 
1295     AllSwitchPaths SwitchPaths(&Switch, ORE);
1296     SwitchPaths.run();
1297 
1298     if (SwitchPaths.getNumThreadingPaths() > 0) {
1299       ThreadableLoops.push_back(SwitchPaths);
1300 
1301       // For the time being limit this optimization to occurring once in a
1302       // function since it can change the CFG significantly. This is not a
1303       // strict requirement but it can cause buggy behavior if there is an
1304       // overlap of blocks in different opportunities. There is a lot of room to
1305       // experiment with catching more opportunities here.
1306       break;
1307     }
1308   }
1309 
1310   SmallPtrSet<const Value *, 32> EphValues;
1311   if (ThreadableLoops.size() > 0)
1312     CodeMetrics::collectEphemeralValues(&F, AC, EphValues);
1313 
1314   for (AllSwitchPaths SwitchPaths : ThreadableLoops) {
1315     TransformDFA Transform(&SwitchPaths, DT, AC, TTI, ORE, EphValues);
1316     Transform.run();
1317     MadeChanges = true;
1318   }
1319 
1320 #ifdef EXPENSIVE_CHECKS
1321   assert(DT->verify(DominatorTree::VerificationLevel::Full));
1322   verifyFunction(F, &dbgs());
1323 #endif
1324 
1325   return MadeChanges;
1326 }
1327 
1328 } // end anonymous namespace
1329 
1330 /// Integrate with the new Pass Manager
1331 PreservedAnalyses DFAJumpThreadingPass::run(Function &F,
1332                                             FunctionAnalysisManager &AM) {
1333   AssumptionCache &AC = AM.getResult<AssumptionAnalysis>(F);
1334   DominatorTree &DT = AM.getResult<DominatorTreeAnalysis>(F);
1335   TargetTransformInfo &TTI = AM.getResult<TargetIRAnalysis>(F);
1336   OptimizationRemarkEmitter ORE(&F);
1337 
1338   if (!DFAJumpThreading(&AC, &DT, &TTI, &ORE).run(F))
1339     return PreservedAnalyses::all();
1340 
1341   PreservedAnalyses PA;
1342   PA.preserve<DominatorTreeAnalysis>();
1343   return PA;
1344 }
1345