1 //===- BasicBlockUtils.cpp - BasicBlock Utilities --------------------------==//
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 family of functions perform manipulations on basic blocks, and
10 // instructions contained within basic blocks.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
15 #include "llvm/ADT/ArrayRef.h"
16 #include "llvm/ADT/SmallPtrSet.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/Twine.h"
19 #include "llvm/Analysis/CFG.h"
20 #include "llvm/Analysis/DomTreeUpdater.h"
21 #include "llvm/Analysis/LoopInfo.h"
22 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
23 #include "llvm/Analysis/MemorySSAUpdater.h"
24 #include "llvm/Analysis/PostDominators.h"
25 #include "llvm/IR/BasicBlock.h"
26 #include "llvm/IR/CFG.h"
27 #include "llvm/IR/Constants.h"
28 #include "llvm/IR/DebugInfoMetadata.h"
29 #include "llvm/IR/Dominators.h"
30 #include "llvm/IR/Function.h"
31 #include "llvm/IR/InstrTypes.h"
32 #include "llvm/IR/Instruction.h"
33 #include "llvm/IR/Instructions.h"
34 #include "llvm/IR/IntrinsicInst.h"
35 #include "llvm/IR/LLVMContext.h"
36 #include "llvm/IR/Type.h"
37 #include "llvm/IR/User.h"
38 #include "llvm/IR/Value.h"
39 #include "llvm/IR/ValueHandle.h"
40 #include "llvm/Support/Casting.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Transforms/Utils/Local.h"
44 #include <cassert>
45 #include <cstdint>
46 #include <string>
47 #include <utility>
48 #include <vector>
49 
50 using namespace llvm;
51 
52 #define DEBUG_TYPE "basicblock-utils"
53 
54 void llvm::DetatchDeadBlocks(
55     ArrayRef<BasicBlock *> BBs,
56     SmallVectorImpl<DominatorTree::UpdateType> *Updates,
57     bool KeepOneInputPHIs) {
58   for (auto *BB : BBs) {
59     // Loop through all of our successors and make sure they know that one
60     // of their predecessors is going away.
61     SmallPtrSet<BasicBlock *, 4> UniqueSuccessors;
62     for (BasicBlock *Succ : successors(BB)) {
63       Succ->removePredecessor(BB, KeepOneInputPHIs);
64       if (Updates && UniqueSuccessors.insert(Succ).second)
65         Updates->push_back({DominatorTree::Delete, BB, Succ});
66     }
67 
68     // Zap all the instructions in the block.
69     while (!BB->empty()) {
70       Instruction &I = BB->back();
71       // If this instruction is used, replace uses with an arbitrary value.
72       // Because control flow can't get here, we don't care what we replace the
73       // value with.  Note that since this block is unreachable, and all values
74       // contained within it must dominate their uses, that all uses will
75       // eventually be removed (they are themselves dead).
76       if (!I.use_empty())
77         I.replaceAllUsesWith(UndefValue::get(I.getType()));
78       BB->getInstList().pop_back();
79     }
80     new UnreachableInst(BB->getContext(), BB);
81     assert(BB->getInstList().size() == 1 &&
82            isa<UnreachableInst>(BB->getTerminator()) &&
83            "The successor list of BB isn't empty before "
84            "applying corresponding DTU updates.");
85   }
86 }
87 
88 void llvm::DeleteDeadBlock(BasicBlock *BB, DomTreeUpdater *DTU,
89                            bool KeepOneInputPHIs) {
90   DeleteDeadBlocks({BB}, DTU, KeepOneInputPHIs);
91 }
92 
93 void llvm::DeleteDeadBlocks(ArrayRef <BasicBlock *> BBs, DomTreeUpdater *DTU,
94                             bool KeepOneInputPHIs) {
95 #ifndef NDEBUG
96   // Make sure that all predecessors of each dead block is also dead.
97   SmallPtrSet<BasicBlock *, 4> Dead(BBs.begin(), BBs.end());
98   assert(Dead.size() == BBs.size() && "Duplicating blocks?");
99   for (auto *BB : Dead)
100     for (BasicBlock *Pred : predecessors(BB))
101       assert(Dead.count(Pred) && "All predecessors must be dead!");
102 #endif
103 
104   SmallVector<DominatorTree::UpdateType, 4> Updates;
105   DetatchDeadBlocks(BBs, DTU ? &Updates : nullptr, KeepOneInputPHIs);
106 
107   if (DTU)
108     DTU->applyUpdates(Updates);
109 
110   for (BasicBlock *BB : BBs)
111     if (DTU)
112       DTU->deleteBB(BB);
113     else
114       BB->eraseFromParent();
115 }
116 
117 bool llvm::EliminateUnreachableBlocks(Function &F, DomTreeUpdater *DTU,
118                                       bool KeepOneInputPHIs) {
119   df_iterator_default_set<BasicBlock*> Reachable;
120 
121   // Mark all reachable blocks.
122   for (BasicBlock *BB : depth_first_ext(&F, Reachable))
123     (void)BB/* Mark all reachable blocks */;
124 
125   // Collect all dead blocks.
126   std::vector<BasicBlock*> DeadBlocks;
127   for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
128     if (!Reachable.count(&*I)) {
129       BasicBlock *BB = &*I;
130       DeadBlocks.push_back(BB);
131     }
132 
133   // Delete the dead blocks.
134   DeleteDeadBlocks(DeadBlocks, DTU, KeepOneInputPHIs);
135 
136   return !DeadBlocks.empty();
137 }
138 
139 bool llvm::FoldSingleEntryPHINodes(BasicBlock *BB,
140                                    MemoryDependenceResults *MemDep) {
141   if (!isa<PHINode>(BB->begin()))
142     return false;
143 
144   while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
145     if (PN->getIncomingValue(0) != PN)
146       PN->replaceAllUsesWith(PN->getIncomingValue(0));
147     else
148       PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
149 
150     if (MemDep)
151       MemDep->removeInstruction(PN);  // Memdep updates AA itself.
152 
153     PN->eraseFromParent();
154   }
155   return true;
156 }
157 
158 bool llvm::DeleteDeadPHIs(BasicBlock *BB, const TargetLibraryInfo *TLI,
159                           MemorySSAUpdater *MSSAU) {
160   // Recursively deleting a PHI may cause multiple PHIs to be deleted
161   // or RAUW'd undef, so use an array of WeakTrackingVH for the PHIs to delete.
162   SmallVector<WeakTrackingVH, 8> PHIs;
163   for (PHINode &PN : BB->phis())
164     PHIs.push_back(&PN);
165 
166   bool Changed = false;
167   for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
168     if (PHINode *PN = dyn_cast_or_null<PHINode>(PHIs[i].operator Value*()))
169       Changed |= RecursivelyDeleteDeadPHINode(PN, TLI, MSSAU);
170 
171   return Changed;
172 }
173 
174 bool llvm::MergeBlockIntoPredecessor(BasicBlock *BB, DomTreeUpdater *DTU,
175                                      LoopInfo *LI, MemorySSAUpdater *MSSAU,
176                                      MemoryDependenceResults *MemDep,
177                                      bool PredecessorWithTwoSuccessors) {
178   if (BB->hasAddressTaken())
179     return false;
180 
181   // Can't merge if there are multiple predecessors, or no predecessors.
182   BasicBlock *PredBB = BB->getUniquePredecessor();
183   if (!PredBB) return false;
184 
185   // Don't break self-loops.
186   if (PredBB == BB) return false;
187   // Don't break unwinding instructions.
188   if (PredBB->getTerminator()->isExceptionalTerminator())
189     return false;
190 
191   // Can't merge if there are multiple distinct successors.
192   if (!PredecessorWithTwoSuccessors && PredBB->getUniqueSuccessor() != BB)
193     return false;
194 
195   // Currently only allow PredBB to have two predecessors, one being BB.
196   // Update BI to branch to BB's only successor instead of BB.
197   BranchInst *PredBB_BI;
198   BasicBlock *NewSucc = nullptr;
199   unsigned FallThruPath;
200   if (PredecessorWithTwoSuccessors) {
201     if (!(PredBB_BI = dyn_cast<BranchInst>(PredBB->getTerminator())))
202       return false;
203     BranchInst *BB_JmpI = dyn_cast<BranchInst>(BB->getTerminator());
204     if (!BB_JmpI || !BB_JmpI->isUnconditional())
205       return false;
206     NewSucc = BB_JmpI->getSuccessor(0);
207     FallThruPath = PredBB_BI->getSuccessor(0) == BB ? 0 : 1;
208   }
209 
210   // Can't merge if there is PHI loop.
211   for (PHINode &PN : BB->phis())
212     for (Value *IncValue : PN.incoming_values())
213       if (IncValue == &PN)
214         return false;
215 
216   LLVM_DEBUG(dbgs() << "Merging: " << BB->getName() << " into "
217                     << PredBB->getName() << "\n");
218 
219   // Begin by getting rid of unneeded PHIs.
220   SmallVector<AssertingVH<Value>, 4> IncomingValues;
221   if (isa<PHINode>(BB->front())) {
222     for (PHINode &PN : BB->phis())
223       if (!isa<PHINode>(PN.getIncomingValue(0)) ||
224           cast<PHINode>(PN.getIncomingValue(0))->getParent() != BB)
225         IncomingValues.push_back(PN.getIncomingValue(0));
226     FoldSingleEntryPHINodes(BB, MemDep);
227   }
228 
229   // DTU update: Collect all the edges that exit BB.
230   // These dominator edges will be redirected from Pred.
231   std::vector<DominatorTree::UpdateType> Updates;
232   if (DTU) {
233     SmallSetVector<BasicBlock *, 2> UniqueSuccessors(succ_begin(BB),
234                                                      succ_end(BB));
235     Updates.reserve(1 + (2 * UniqueSuccessors.size()));
236     // Add insert edges first. Experimentally, for the particular case of two
237     // blocks that can be merged, with a single successor and single predecessor
238     // respectively, it is beneficial to have all insert updates first. Deleting
239     // edges first may lead to unreachable blocks, followed by inserting edges
240     // making the blocks reachable again. Such DT updates lead to high compile
241     // times. We add inserts before deletes here to reduce compile time.
242     for (BasicBlock *UniqueSuccessor : UniqueSuccessors)
243       // This successor of BB may already have PredBB as a predecessor.
244       if (!llvm::is_contained(successors(PredBB), UniqueSuccessor))
245         Updates.push_back({DominatorTree::Insert, PredBB, UniqueSuccessor});
246     for (BasicBlock *UniqueSuccessor : UniqueSuccessors)
247       Updates.push_back({DominatorTree::Delete, BB, UniqueSuccessor});
248     Updates.push_back({DominatorTree::Delete, PredBB, BB});
249   }
250 
251   Instruction *PTI = PredBB->getTerminator();
252   Instruction *STI = BB->getTerminator();
253   Instruction *Start = &*BB->begin();
254   // If there's nothing to move, mark the starting instruction as the last
255   // instruction in the block. Terminator instruction is handled separately.
256   if (Start == STI)
257     Start = PTI;
258 
259   // Move all definitions in the successor to the predecessor...
260   PredBB->getInstList().splice(PTI->getIterator(), BB->getInstList(),
261                                BB->begin(), STI->getIterator());
262 
263   if (MSSAU)
264     MSSAU->moveAllAfterMergeBlocks(BB, PredBB, Start);
265 
266   // Make all PHI nodes that referred to BB now refer to Pred as their
267   // source...
268   BB->replaceAllUsesWith(PredBB);
269 
270   if (PredecessorWithTwoSuccessors) {
271     // Delete the unconditional branch from BB.
272     BB->getInstList().pop_back();
273 
274     // Update branch in the predecessor.
275     PredBB_BI->setSuccessor(FallThruPath, NewSucc);
276   } else {
277     // Delete the unconditional branch from the predecessor.
278     PredBB->getInstList().pop_back();
279 
280     // Move terminator instruction.
281     PredBB->getInstList().splice(PredBB->end(), BB->getInstList());
282 
283     // Terminator may be a memory accessing instruction too.
284     if (MSSAU)
285       if (MemoryUseOrDef *MUD = cast_or_null<MemoryUseOrDef>(
286               MSSAU->getMemorySSA()->getMemoryAccess(PredBB->getTerminator())))
287         MSSAU->moveToPlace(MUD, PredBB, MemorySSA::End);
288   }
289   // Add unreachable to now empty BB.
290   new UnreachableInst(BB->getContext(), BB);
291 
292   // Inherit predecessors name if it exists.
293   if (!PredBB->hasName())
294     PredBB->takeName(BB);
295 
296   if (LI)
297     LI->removeBlock(BB);
298 
299   if (MemDep)
300     MemDep->invalidateCachedPredecessors();
301 
302   // Finally, erase the old block and update dominator info.
303   if (DTU) {
304     assert(BB->getInstList().size() == 1 &&
305            isa<UnreachableInst>(BB->getTerminator()) &&
306            "The successor list of BB isn't empty before "
307            "applying corresponding DTU updates.");
308     DTU->applyUpdates(Updates);
309     DTU->deleteBB(BB);
310   } else {
311     BB->eraseFromParent(); // Nuke BB if DTU is nullptr.
312   }
313 
314   return true;
315 }
316 
317 bool llvm::MergeBlockSuccessorsIntoGivenBlocks(
318     SmallPtrSetImpl<BasicBlock *> &MergeBlocks, Loop *L, DomTreeUpdater *DTU,
319     LoopInfo *LI) {
320   assert(!MergeBlocks.empty() && "MergeBlocks should not be empty");
321 
322   bool BlocksHaveBeenMerged = false;
323   while (!MergeBlocks.empty()) {
324     BasicBlock *BB = *MergeBlocks.begin();
325     BasicBlock *Dest = BB->getSingleSuccessor();
326     if (Dest && (!L || L->contains(Dest))) {
327       BasicBlock *Fold = Dest->getUniquePredecessor();
328       (void)Fold;
329       if (MergeBlockIntoPredecessor(Dest, DTU, LI)) {
330         assert(Fold == BB &&
331                "Expecting BB to be unique predecessor of the Dest block");
332         MergeBlocks.erase(Dest);
333         BlocksHaveBeenMerged = true;
334       } else
335         MergeBlocks.erase(BB);
336     } else
337       MergeBlocks.erase(BB);
338   }
339   return BlocksHaveBeenMerged;
340 }
341 
342 /// Remove redundant instructions within sequences of consecutive dbg.value
343 /// instructions. This is done using a backward scan to keep the last dbg.value
344 /// describing a specific variable/fragment.
345 ///
346 /// BackwardScan strategy:
347 /// ----------------------
348 /// Given a sequence of consecutive DbgValueInst like this
349 ///
350 ///   dbg.value ..., "x", FragmentX1  (*)
351 ///   dbg.value ..., "y", FragmentY1
352 ///   dbg.value ..., "x", FragmentX2
353 ///   dbg.value ..., "x", FragmentX1  (**)
354 ///
355 /// then the instruction marked with (*) can be removed (it is guaranteed to be
356 /// obsoleted by the instruction marked with (**) as the latter instruction is
357 /// describing the same variable using the same fragment info).
358 ///
359 /// Possible improvements:
360 /// - Check fully overlapping fragments and not only identical fragments.
361 /// - Support dbg.addr, dbg.declare. dbg.label, and possibly other meta
362 ///   instructions being part of the sequence of consecutive instructions.
363 static bool removeRedundantDbgInstrsUsingBackwardScan(BasicBlock *BB) {
364   SmallVector<DbgValueInst *, 8> ToBeRemoved;
365   SmallDenseSet<DebugVariable> VariableSet;
366   for (auto &I : reverse(*BB)) {
367     if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(&I)) {
368       DebugVariable Key(DVI->getVariable(),
369                         DVI->getExpression(),
370                         DVI->getDebugLoc()->getInlinedAt());
371       auto R = VariableSet.insert(Key);
372       // If the same variable fragment is described more than once it is enough
373       // to keep the last one (i.e. the first found since we for reverse
374       // iteration).
375       if (!R.second)
376         ToBeRemoved.push_back(DVI);
377       continue;
378     }
379     // Sequence with consecutive dbg.value instrs ended. Clear the map to
380     // restart identifying redundant instructions if case we find another
381     // dbg.value sequence.
382     VariableSet.clear();
383   }
384 
385   for (auto &Instr : ToBeRemoved)
386     Instr->eraseFromParent();
387 
388   return !ToBeRemoved.empty();
389 }
390 
391 /// Remove redundant dbg.value instructions using a forward scan. This can
392 /// remove a dbg.value instruction that is redundant due to indicating that a
393 /// variable has the same value as already being indicated by an earlier
394 /// dbg.value.
395 ///
396 /// ForwardScan strategy:
397 /// ---------------------
398 /// Given two identical dbg.value instructions, separated by a block of
399 /// instructions that isn't describing the same variable, like this
400 ///
401 ///   dbg.value X1, "x", FragmentX1  (**)
402 ///   <block of instructions, none being "dbg.value ..., "x", ...">
403 ///   dbg.value X1, "x", FragmentX1  (*)
404 ///
405 /// then the instruction marked with (*) can be removed. Variable "x" is already
406 /// described as being mapped to the SSA value X1.
407 ///
408 /// Possible improvements:
409 /// - Keep track of non-overlapping fragments.
410 static bool removeRedundantDbgInstrsUsingForwardScan(BasicBlock *BB) {
411   SmallVector<DbgValueInst *, 8> ToBeRemoved;
412   DenseMap<DebugVariable, std::pair<Value *, DIExpression *> > VariableMap;
413   for (auto &I : *BB) {
414     if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(&I)) {
415       DebugVariable Key(DVI->getVariable(),
416                         NoneType(),
417                         DVI->getDebugLoc()->getInlinedAt());
418       auto VMI = VariableMap.find(Key);
419       // Update the map if we found a new value/expression describing the
420       // variable, or if the variable wasn't mapped already.
421       if (VMI == VariableMap.end() ||
422           VMI->second.first != DVI->getValue() ||
423           VMI->second.second != DVI->getExpression()) {
424         VariableMap[Key] = { DVI->getValue(), DVI->getExpression() };
425         continue;
426       }
427       // Found an identical mapping. Remember the instruction for later removal.
428       ToBeRemoved.push_back(DVI);
429     }
430   }
431 
432   for (auto &Instr : ToBeRemoved)
433     Instr->eraseFromParent();
434 
435   return !ToBeRemoved.empty();
436 }
437 
438 bool llvm::RemoveRedundantDbgInstrs(BasicBlock *BB) {
439   bool MadeChanges = false;
440   // By using the "backward scan" strategy before the "forward scan" strategy we
441   // can remove both dbg.value (2) and (3) in a situation like this:
442   //
443   //   (1) dbg.value V1, "x", DIExpression()
444   //       ...
445   //   (2) dbg.value V2, "x", DIExpression()
446   //   (3) dbg.value V1, "x", DIExpression()
447   //
448   // The backward scan will remove (2), it is made obsolete by (3). After
449   // getting (2) out of the way, the foward scan will remove (3) since "x"
450   // already is described as having the value V1 at (1).
451   MadeChanges |= removeRedundantDbgInstrsUsingBackwardScan(BB);
452   MadeChanges |= removeRedundantDbgInstrsUsingForwardScan(BB);
453 
454   if (MadeChanges)
455     LLVM_DEBUG(dbgs() << "Removed redundant dbg instrs from: "
456                       << BB->getName() << "\n");
457   return MadeChanges;
458 }
459 
460 void llvm::ReplaceInstWithValue(BasicBlock::InstListType &BIL,
461                                 BasicBlock::iterator &BI, Value *V) {
462   Instruction &I = *BI;
463   // Replaces all of the uses of the instruction with uses of the value
464   I.replaceAllUsesWith(V);
465 
466   // Make sure to propagate a name if there is one already.
467   if (I.hasName() && !V->hasName())
468     V->takeName(&I);
469 
470   // Delete the unnecessary instruction now...
471   BI = BIL.erase(BI);
472 }
473 
474 void llvm::ReplaceInstWithInst(BasicBlock::InstListType &BIL,
475                                BasicBlock::iterator &BI, Instruction *I) {
476   assert(I->getParent() == nullptr &&
477          "ReplaceInstWithInst: Instruction already inserted into basic block!");
478 
479   // Copy debug location to newly added instruction, if it wasn't already set
480   // by the caller.
481   if (!I->getDebugLoc())
482     I->setDebugLoc(BI->getDebugLoc());
483 
484   // Insert the new instruction into the basic block...
485   BasicBlock::iterator New = BIL.insert(BI, I);
486 
487   // Replace all uses of the old instruction, and delete it.
488   ReplaceInstWithValue(BIL, BI, I);
489 
490   // Move BI back to point to the newly inserted instruction
491   BI = New;
492 }
493 
494 void llvm::ReplaceInstWithInst(Instruction *From, Instruction *To) {
495   BasicBlock::iterator BI(From);
496   ReplaceInstWithInst(From->getParent()->getInstList(), BI, To);
497 }
498 
499 BasicBlock *llvm::SplitEdge(BasicBlock *BB, BasicBlock *Succ, DominatorTree *DT,
500                             LoopInfo *LI, MemorySSAUpdater *MSSAU,
501                             const Twine &BBName) {
502   unsigned SuccNum = GetSuccessorNumber(BB, Succ);
503 
504   // If this is a critical edge, let SplitCriticalEdge do it.
505   Instruction *LatchTerm = BB->getTerminator();
506   if (SplitCriticalEdge(
507           LatchTerm, SuccNum,
508           CriticalEdgeSplittingOptions(DT, LI, MSSAU).setPreserveLCSSA(),
509           BBName))
510     return LatchTerm->getSuccessor(SuccNum);
511 
512   // If the edge isn't critical, then BB has a single successor or Succ has a
513   // single pred.  Split the block.
514   if (BasicBlock *SP = Succ->getSinglePredecessor()) {
515     // If the successor only has a single pred, split the top of the successor
516     // block.
517     assert(SP == BB && "CFG broken");
518     SP = nullptr;
519     return SplitBlock(Succ, &Succ->front(), DT, LI, MSSAU, BBName,
520                       /*Before=*/true);
521   }
522 
523   // Otherwise, if BB has a single successor, split it at the bottom of the
524   // block.
525   assert(BB->getTerminator()->getNumSuccessors() == 1 &&
526          "Should have a single succ!");
527   return SplitBlock(BB, BB->getTerminator(), DT, LI, MSSAU, BBName);
528 }
529 
530 unsigned
531 llvm::SplitAllCriticalEdges(Function &F,
532                             const CriticalEdgeSplittingOptions &Options) {
533   unsigned NumBroken = 0;
534   for (BasicBlock &BB : F) {
535     Instruction *TI = BB.getTerminator();
536     if (TI->getNumSuccessors() > 1 && !isa<IndirectBrInst>(TI) &&
537         !isa<CallBrInst>(TI))
538       for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
539         if (SplitCriticalEdge(TI, i, Options))
540           ++NumBroken;
541   }
542   return NumBroken;
543 }
544 
545 BasicBlock *llvm::SplitBlock(BasicBlock *Old, Instruction *SplitPt,
546                              DominatorTree *DT, LoopInfo *LI,
547                              MemorySSAUpdater *MSSAU, const Twine &BBName,
548                              bool Before) {
549   if (Before)
550     return splitBlockBefore(Old, SplitPt, DT, LI, MSSAU, BBName);
551   BasicBlock::iterator SplitIt = SplitPt->getIterator();
552   while (isa<PHINode>(SplitIt) || SplitIt->isEHPad())
553     ++SplitIt;
554   std::string Name = BBName.str();
555   BasicBlock *New = Old->splitBasicBlock(
556       SplitIt, Name.empty() ? Old->getName() + ".split" : Name);
557 
558   // The new block lives in whichever loop the old one did. This preserves
559   // LCSSA as well, because we force the split point to be after any PHI nodes.
560   if (LI)
561     if (Loop *L = LI->getLoopFor(Old))
562       L->addBasicBlockToLoop(New, *LI);
563 
564   if (DT)
565     // Old dominates New. New node dominates all other nodes dominated by Old.
566     if (DomTreeNode *OldNode = DT->getNode(Old)) {
567       std::vector<DomTreeNode *> Children(OldNode->begin(), OldNode->end());
568 
569       DomTreeNode *NewNode = DT->addNewBlock(New, Old);
570       for (DomTreeNode *I : Children)
571         DT->changeImmediateDominator(I, NewNode);
572     }
573 
574   // Move MemoryAccesses still tracked in Old, but part of New now.
575   // Update accesses in successor blocks accordingly.
576   if (MSSAU)
577     MSSAU->moveAllAfterSpliceBlocks(Old, New, &*(New->begin()));
578 
579   return New;
580 }
581 
582 BasicBlock *llvm::splitBlockBefore(BasicBlock *Old, Instruction *SplitPt,
583                                    DominatorTree *DT, LoopInfo *LI,
584                                    MemorySSAUpdater *MSSAU,
585                                    const Twine &BBName) {
586 
587   BasicBlock::iterator SplitIt = SplitPt->getIterator();
588   while (isa<PHINode>(SplitIt) || SplitIt->isEHPad())
589     ++SplitIt;
590   std::string Name = BBName.str();
591   BasicBlock *New = Old->splitBasicBlock(
592       SplitIt, Name.empty() ? Old->getName() + ".split" : Name,
593       /* Before=*/true);
594 
595   // The new block lives in whichever loop the old one did. This preserves
596   // LCSSA as well, because we force the split point to be after any PHI nodes.
597   if (LI)
598     if (Loop *L = LI->getLoopFor(Old))
599       L->addBasicBlockToLoop(New, *LI);
600 
601   if (DT) {
602     DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy);
603     SmallVector<DominatorTree::UpdateType, 8> DTUpdates;
604     // New dominates Old. The predecessor nodes of the Old node dominate
605     // New node.
606     DTUpdates.push_back({DominatorTree::Insert, New, Old});
607     for (BasicBlock *Pred : predecessors(New))
608       if (DT->getNode(Pred)) {
609         DTUpdates.push_back({DominatorTree::Insert, Pred, New});
610         DTUpdates.push_back({DominatorTree::Delete, Pred, Old});
611       }
612 
613     DTU.applyUpdates(DTUpdates);
614     DTU.flush();
615 
616     // Move MemoryAccesses still tracked in Old, but part of New now.
617     // Update accesses in successor blocks accordingly.
618     if (MSSAU) {
619       MSSAU->applyUpdates(DTUpdates, *DT);
620       if (VerifyMemorySSA)
621         MSSAU->getMemorySSA()->verifyMemorySSA();
622     }
623   }
624   return New;
625 }
626 
627 /// Update DominatorTree, LoopInfo, and LCCSA analysis information.
628 static void UpdateAnalysisInformation(BasicBlock *OldBB, BasicBlock *NewBB,
629                                       ArrayRef<BasicBlock *> Preds,
630                                       DominatorTree *DT, LoopInfo *LI,
631                                       MemorySSAUpdater *MSSAU,
632                                       bool PreserveLCSSA, bool &HasLoopExit) {
633   // Update dominator tree if available.
634   if (DT) {
635     if (OldBB == DT->getRootNode()->getBlock()) {
636       assert(NewBB == &NewBB->getParent()->getEntryBlock());
637       DT->setNewRoot(NewBB);
638     } else {
639       // Split block expects NewBB to have a non-empty set of predecessors.
640       DT->splitBlock(NewBB);
641     }
642   }
643 
644   // Update MemoryPhis after split if MemorySSA is available
645   if (MSSAU)
646     MSSAU->wireOldPredecessorsToNewImmediatePredecessor(OldBB, NewBB, Preds);
647 
648   // The rest of the logic is only relevant for updating the loop structures.
649   if (!LI)
650     return;
651 
652   assert(DT && "DT should be available to update LoopInfo!");
653   Loop *L = LI->getLoopFor(OldBB);
654 
655   // If we need to preserve loop analyses, collect some information about how
656   // this split will affect loops.
657   bool IsLoopEntry = !!L;
658   bool SplitMakesNewLoopHeader = false;
659   for (BasicBlock *Pred : Preds) {
660     // Preds that are not reachable from entry should not be used to identify if
661     // OldBB is a loop entry or if SplitMakesNewLoopHeader. Unreachable blocks
662     // are not within any loops, so we incorrectly mark SplitMakesNewLoopHeader
663     // as true and make the NewBB the header of some loop. This breaks LI.
664     if (!DT->isReachableFromEntry(Pred))
665       continue;
666     // If we need to preserve LCSSA, determine if any of the preds is a loop
667     // exit.
668     if (PreserveLCSSA)
669       if (Loop *PL = LI->getLoopFor(Pred))
670         if (!PL->contains(OldBB))
671           HasLoopExit = true;
672 
673     // If we need to preserve LoopInfo, note whether any of the preds crosses
674     // an interesting loop boundary.
675     if (!L)
676       continue;
677     if (L->contains(Pred))
678       IsLoopEntry = false;
679     else
680       SplitMakesNewLoopHeader = true;
681   }
682 
683   // Unless we have a loop for OldBB, nothing else to do here.
684   if (!L)
685     return;
686 
687   if (IsLoopEntry) {
688     // Add the new block to the nearest enclosing loop (and not an adjacent
689     // loop). To find this, examine each of the predecessors and determine which
690     // loops enclose them, and select the most-nested loop which contains the
691     // loop containing the block being split.
692     Loop *InnermostPredLoop = nullptr;
693     for (BasicBlock *Pred : Preds) {
694       if (Loop *PredLoop = LI->getLoopFor(Pred)) {
695         // Seek a loop which actually contains the block being split (to avoid
696         // adjacent loops).
697         while (PredLoop && !PredLoop->contains(OldBB))
698           PredLoop = PredLoop->getParentLoop();
699 
700         // Select the most-nested of these loops which contains the block.
701         if (PredLoop && PredLoop->contains(OldBB) &&
702             (!InnermostPredLoop ||
703              InnermostPredLoop->getLoopDepth() < PredLoop->getLoopDepth()))
704           InnermostPredLoop = PredLoop;
705       }
706     }
707 
708     if (InnermostPredLoop)
709       InnermostPredLoop->addBasicBlockToLoop(NewBB, *LI);
710   } else {
711     L->addBasicBlockToLoop(NewBB, *LI);
712     if (SplitMakesNewLoopHeader)
713       L->moveToHeader(NewBB);
714   }
715 }
716 
717 /// Update the PHI nodes in OrigBB to include the values coming from NewBB.
718 /// This also updates AliasAnalysis, if available.
719 static void UpdatePHINodes(BasicBlock *OrigBB, BasicBlock *NewBB,
720                            ArrayRef<BasicBlock *> Preds, BranchInst *BI,
721                            bool HasLoopExit) {
722   // Otherwise, create a new PHI node in NewBB for each PHI node in OrigBB.
723   SmallPtrSet<BasicBlock *, 16> PredSet(Preds.begin(), Preds.end());
724   for (BasicBlock::iterator I = OrigBB->begin(); isa<PHINode>(I); ) {
725     PHINode *PN = cast<PHINode>(I++);
726 
727     // Check to see if all of the values coming in are the same.  If so, we
728     // don't need to create a new PHI node, unless it's needed for LCSSA.
729     Value *InVal = nullptr;
730     if (!HasLoopExit) {
731       InVal = PN->getIncomingValueForBlock(Preds[0]);
732       for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
733         if (!PredSet.count(PN->getIncomingBlock(i)))
734           continue;
735         if (!InVal)
736           InVal = PN->getIncomingValue(i);
737         else if (InVal != PN->getIncomingValue(i)) {
738           InVal = nullptr;
739           break;
740         }
741       }
742     }
743 
744     if (InVal) {
745       // If all incoming values for the new PHI would be the same, just don't
746       // make a new PHI.  Instead, just remove the incoming values from the old
747       // PHI.
748 
749       // NOTE! This loop walks backwards for a reason! First off, this minimizes
750       // the cost of removal if we end up removing a large number of values, and
751       // second off, this ensures that the indices for the incoming values
752       // aren't invalidated when we remove one.
753       for (int64_t i = PN->getNumIncomingValues() - 1; i >= 0; --i)
754         if (PredSet.count(PN->getIncomingBlock(i)))
755           PN->removeIncomingValue(i, false);
756 
757       // Add an incoming value to the PHI node in the loop for the preheader
758       // edge.
759       PN->addIncoming(InVal, NewBB);
760       continue;
761     }
762 
763     // If the values coming into the block are not the same, we need a new
764     // PHI.
765     // Create the new PHI node, insert it into NewBB at the end of the block
766     PHINode *NewPHI =
767         PHINode::Create(PN->getType(), Preds.size(), PN->getName() + ".ph", BI);
768 
769     // NOTE! This loop walks backwards for a reason! First off, this minimizes
770     // the cost of removal if we end up removing a large number of values, and
771     // second off, this ensures that the indices for the incoming values aren't
772     // invalidated when we remove one.
773     for (int64_t i = PN->getNumIncomingValues() - 1; i >= 0; --i) {
774       BasicBlock *IncomingBB = PN->getIncomingBlock(i);
775       if (PredSet.count(IncomingBB)) {
776         Value *V = PN->removeIncomingValue(i, false);
777         NewPHI->addIncoming(V, IncomingBB);
778       }
779     }
780 
781     PN->addIncoming(NewPHI, NewBB);
782   }
783 }
784 
785 BasicBlock *llvm::SplitBlockPredecessors(BasicBlock *BB,
786                                          ArrayRef<BasicBlock *> Preds,
787                                          const char *Suffix, DominatorTree *DT,
788                                          LoopInfo *LI, MemorySSAUpdater *MSSAU,
789                                          bool PreserveLCSSA) {
790   // Do not attempt to split that which cannot be split.
791   if (!BB->canSplitPredecessors())
792     return nullptr;
793 
794   // For the landingpads we need to act a bit differently.
795   // Delegate this work to the SplitLandingPadPredecessors.
796   if (BB->isLandingPad()) {
797     SmallVector<BasicBlock*, 2> NewBBs;
798     std::string NewName = std::string(Suffix) + ".split-lp";
799 
800     SplitLandingPadPredecessors(BB, Preds, Suffix, NewName.c_str(), NewBBs, DT,
801                                 LI, MSSAU, PreserveLCSSA);
802     return NewBBs[0];
803   }
804 
805   // Create new basic block, insert right before the original block.
806   BasicBlock *NewBB = BasicBlock::Create(
807       BB->getContext(), BB->getName() + Suffix, BB->getParent(), BB);
808 
809   // The new block unconditionally branches to the old block.
810   BranchInst *BI = BranchInst::Create(BB, NewBB);
811 
812   Loop *L = nullptr;
813   BasicBlock *OldLatch = nullptr;
814   // Splitting the predecessors of a loop header creates a preheader block.
815   if (LI && LI->isLoopHeader(BB)) {
816     L = LI->getLoopFor(BB);
817     // Using the loop start line number prevents debuggers stepping into the
818     // loop body for this instruction.
819     BI->setDebugLoc(L->getStartLoc());
820 
821     // If BB is the header of the Loop, it is possible that the loop is
822     // modified, such that the current latch does not remain the latch of the
823     // loop. If that is the case, the loop metadata from the current latch needs
824     // to be applied to the new latch.
825     OldLatch = L->getLoopLatch();
826   } else
827     BI->setDebugLoc(BB->getFirstNonPHIOrDbg()->getDebugLoc());
828 
829   // Move the edges from Preds to point to NewBB instead of BB.
830   for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
831     // This is slightly more strict than necessary; the minimum requirement
832     // is that there be no more than one indirectbr branching to BB. And
833     // all BlockAddress uses would need to be updated.
834     assert(!isa<IndirectBrInst>(Preds[i]->getTerminator()) &&
835            "Cannot split an edge from an IndirectBrInst");
836     assert(!isa<CallBrInst>(Preds[i]->getTerminator()) &&
837            "Cannot split an edge from a CallBrInst");
838     Preds[i]->getTerminator()->replaceUsesOfWith(BB, NewBB);
839   }
840 
841   // Insert a new PHI node into NewBB for every PHI node in BB and that new PHI
842   // node becomes an incoming value for BB's phi node.  However, if the Preds
843   // list is empty, we need to insert dummy entries into the PHI nodes in BB to
844   // account for the newly created predecessor.
845   if (Preds.empty()) {
846     // Insert dummy values as the incoming value.
847     for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++I)
848       cast<PHINode>(I)->addIncoming(UndefValue::get(I->getType()), NewBB);
849   }
850 
851   // Update DominatorTree, LoopInfo, and LCCSA analysis information.
852   bool HasLoopExit = false;
853   UpdateAnalysisInformation(BB, NewBB, Preds, DT, LI, MSSAU, PreserveLCSSA,
854                             HasLoopExit);
855 
856   if (!Preds.empty()) {
857     // Update the PHI nodes in BB with the values coming from NewBB.
858     UpdatePHINodes(BB, NewBB, Preds, BI, HasLoopExit);
859   }
860 
861   if (OldLatch) {
862     BasicBlock *NewLatch = L->getLoopLatch();
863     if (NewLatch != OldLatch) {
864       MDNode *MD = OldLatch->getTerminator()->getMetadata("llvm.loop");
865       NewLatch->getTerminator()->setMetadata("llvm.loop", MD);
866       OldLatch->getTerminator()->setMetadata("llvm.loop", nullptr);
867     }
868   }
869 
870   return NewBB;
871 }
872 
873 void llvm::SplitLandingPadPredecessors(BasicBlock *OrigBB,
874                                        ArrayRef<BasicBlock *> Preds,
875                                        const char *Suffix1, const char *Suffix2,
876                                        SmallVectorImpl<BasicBlock *> &NewBBs,
877                                        DominatorTree *DT, LoopInfo *LI,
878                                        MemorySSAUpdater *MSSAU,
879                                        bool PreserveLCSSA) {
880   assert(OrigBB->isLandingPad() && "Trying to split a non-landing pad!");
881 
882   // Create a new basic block for OrigBB's predecessors listed in Preds. Insert
883   // it right before the original block.
884   BasicBlock *NewBB1 = BasicBlock::Create(OrigBB->getContext(),
885                                           OrigBB->getName() + Suffix1,
886                                           OrigBB->getParent(), OrigBB);
887   NewBBs.push_back(NewBB1);
888 
889   // The new block unconditionally branches to the old block.
890   BranchInst *BI1 = BranchInst::Create(OrigBB, NewBB1);
891   BI1->setDebugLoc(OrigBB->getFirstNonPHI()->getDebugLoc());
892 
893   // Move the edges from Preds to point to NewBB1 instead of OrigBB.
894   for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
895     // This is slightly more strict than necessary; the minimum requirement
896     // is that there be no more than one indirectbr branching to BB. And
897     // all BlockAddress uses would need to be updated.
898     assert(!isa<IndirectBrInst>(Preds[i]->getTerminator()) &&
899            "Cannot split an edge from an IndirectBrInst");
900     Preds[i]->getTerminator()->replaceUsesOfWith(OrigBB, NewBB1);
901   }
902 
903   bool HasLoopExit = false;
904   UpdateAnalysisInformation(OrigBB, NewBB1, Preds, DT, LI, MSSAU, PreserveLCSSA,
905                             HasLoopExit);
906 
907   // Update the PHI nodes in OrigBB with the values coming from NewBB1.
908   UpdatePHINodes(OrigBB, NewBB1, Preds, BI1, HasLoopExit);
909 
910   // Move the remaining edges from OrigBB to point to NewBB2.
911   SmallVector<BasicBlock*, 8> NewBB2Preds;
912   for (pred_iterator i = pred_begin(OrigBB), e = pred_end(OrigBB);
913        i != e; ) {
914     BasicBlock *Pred = *i++;
915     if (Pred == NewBB1) continue;
916     assert(!isa<IndirectBrInst>(Pred->getTerminator()) &&
917            "Cannot split an edge from an IndirectBrInst");
918     NewBB2Preds.push_back(Pred);
919     e = pred_end(OrigBB);
920   }
921 
922   BasicBlock *NewBB2 = nullptr;
923   if (!NewBB2Preds.empty()) {
924     // Create another basic block for the rest of OrigBB's predecessors.
925     NewBB2 = BasicBlock::Create(OrigBB->getContext(),
926                                 OrigBB->getName() + Suffix2,
927                                 OrigBB->getParent(), OrigBB);
928     NewBBs.push_back(NewBB2);
929 
930     // The new block unconditionally branches to the old block.
931     BranchInst *BI2 = BranchInst::Create(OrigBB, NewBB2);
932     BI2->setDebugLoc(OrigBB->getFirstNonPHI()->getDebugLoc());
933 
934     // Move the remaining edges from OrigBB to point to NewBB2.
935     for (BasicBlock *NewBB2Pred : NewBB2Preds)
936       NewBB2Pred->getTerminator()->replaceUsesOfWith(OrigBB, NewBB2);
937 
938     // Update DominatorTree, LoopInfo, and LCCSA analysis information.
939     HasLoopExit = false;
940     UpdateAnalysisInformation(OrigBB, NewBB2, NewBB2Preds, DT, LI, MSSAU,
941                               PreserveLCSSA, HasLoopExit);
942 
943     // Update the PHI nodes in OrigBB with the values coming from NewBB2.
944     UpdatePHINodes(OrigBB, NewBB2, NewBB2Preds, BI2, HasLoopExit);
945   }
946 
947   LandingPadInst *LPad = OrigBB->getLandingPadInst();
948   Instruction *Clone1 = LPad->clone();
949   Clone1->setName(Twine("lpad") + Suffix1);
950   NewBB1->getInstList().insert(NewBB1->getFirstInsertionPt(), Clone1);
951 
952   if (NewBB2) {
953     Instruction *Clone2 = LPad->clone();
954     Clone2->setName(Twine("lpad") + Suffix2);
955     NewBB2->getInstList().insert(NewBB2->getFirstInsertionPt(), Clone2);
956 
957     // Create a PHI node for the two cloned landingpad instructions only
958     // if the original landingpad instruction has some uses.
959     if (!LPad->use_empty()) {
960       assert(!LPad->getType()->isTokenTy() &&
961              "Split cannot be applied if LPad is token type. Otherwise an "
962              "invalid PHINode of token type would be created.");
963       PHINode *PN = PHINode::Create(LPad->getType(), 2, "lpad.phi", LPad);
964       PN->addIncoming(Clone1, NewBB1);
965       PN->addIncoming(Clone2, NewBB2);
966       LPad->replaceAllUsesWith(PN);
967     }
968     LPad->eraseFromParent();
969   } else {
970     // There is no second clone. Just replace the landing pad with the first
971     // clone.
972     LPad->replaceAllUsesWith(Clone1);
973     LPad->eraseFromParent();
974   }
975 }
976 
977 ReturnInst *llvm::FoldReturnIntoUncondBranch(ReturnInst *RI, BasicBlock *BB,
978                                              BasicBlock *Pred,
979                                              DomTreeUpdater *DTU) {
980   Instruction *UncondBranch = Pred->getTerminator();
981   // Clone the return and add it to the end of the predecessor.
982   Instruction *NewRet = RI->clone();
983   Pred->getInstList().push_back(NewRet);
984 
985   // If the return instruction returns a value, and if the value was a
986   // PHI node in "BB", propagate the right value into the return.
987   for (User::op_iterator i = NewRet->op_begin(), e = NewRet->op_end();
988        i != e; ++i) {
989     Value *V = *i;
990     Instruction *NewBC = nullptr;
991     if (BitCastInst *BCI = dyn_cast<BitCastInst>(V)) {
992       // Return value might be bitcasted. Clone and insert it before the
993       // return instruction.
994       V = BCI->getOperand(0);
995       NewBC = BCI->clone();
996       Pred->getInstList().insert(NewRet->getIterator(), NewBC);
997       *i = NewBC;
998     }
999 
1000     Instruction *NewEV = nullptr;
1001     if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(V)) {
1002       V = EVI->getOperand(0);
1003       NewEV = EVI->clone();
1004       if (NewBC) {
1005         NewBC->setOperand(0, NewEV);
1006         Pred->getInstList().insert(NewBC->getIterator(), NewEV);
1007       } else {
1008         Pred->getInstList().insert(NewRet->getIterator(), NewEV);
1009         *i = NewEV;
1010       }
1011     }
1012 
1013     if (PHINode *PN = dyn_cast<PHINode>(V)) {
1014       if (PN->getParent() == BB) {
1015         if (NewEV) {
1016           NewEV->setOperand(0, PN->getIncomingValueForBlock(Pred));
1017         } else if (NewBC)
1018           NewBC->setOperand(0, PN->getIncomingValueForBlock(Pred));
1019         else
1020           *i = PN->getIncomingValueForBlock(Pred);
1021       }
1022     }
1023   }
1024 
1025   // Update any PHI nodes in the returning block to realize that we no
1026   // longer branch to them.
1027   BB->removePredecessor(Pred);
1028   UncondBranch->eraseFromParent();
1029 
1030   if (DTU)
1031     DTU->applyUpdates({{DominatorTree::Delete, Pred, BB}});
1032 
1033   return cast<ReturnInst>(NewRet);
1034 }
1035 
1036 Instruction *llvm::SplitBlockAndInsertIfThen(Value *Cond,
1037                                              Instruction *SplitBefore,
1038                                              bool Unreachable,
1039                                              MDNode *BranchWeights,
1040                                              DominatorTree *DT, LoopInfo *LI,
1041                                              BasicBlock *ThenBlock) {
1042   BasicBlock *Head = SplitBefore->getParent();
1043   BasicBlock *Tail = Head->splitBasicBlock(SplitBefore->getIterator());
1044   Instruction *HeadOldTerm = Head->getTerminator();
1045   LLVMContext &C = Head->getContext();
1046   Instruction *CheckTerm;
1047   bool CreateThenBlock = (ThenBlock == nullptr);
1048   if (CreateThenBlock) {
1049     ThenBlock = BasicBlock::Create(C, "", Head->getParent(), Tail);
1050     if (Unreachable)
1051       CheckTerm = new UnreachableInst(C, ThenBlock);
1052     else
1053       CheckTerm = BranchInst::Create(Tail, ThenBlock);
1054     CheckTerm->setDebugLoc(SplitBefore->getDebugLoc());
1055   } else
1056     CheckTerm = ThenBlock->getTerminator();
1057   BranchInst *HeadNewTerm =
1058     BranchInst::Create(/*ifTrue*/ThenBlock, /*ifFalse*/Tail, Cond);
1059   HeadNewTerm->setMetadata(LLVMContext::MD_prof, BranchWeights);
1060   ReplaceInstWithInst(HeadOldTerm, HeadNewTerm);
1061 
1062   if (DT) {
1063     if (DomTreeNode *OldNode = DT->getNode(Head)) {
1064       std::vector<DomTreeNode *> Children(OldNode->begin(), OldNode->end());
1065 
1066       DomTreeNode *NewNode = DT->addNewBlock(Tail, Head);
1067       for (DomTreeNode *Child : Children)
1068         DT->changeImmediateDominator(Child, NewNode);
1069 
1070       // Head dominates ThenBlock.
1071       if (CreateThenBlock)
1072         DT->addNewBlock(ThenBlock, Head);
1073       else
1074         DT->changeImmediateDominator(ThenBlock, Head);
1075     }
1076   }
1077 
1078   if (LI) {
1079     if (Loop *L = LI->getLoopFor(Head)) {
1080       L->addBasicBlockToLoop(ThenBlock, *LI);
1081       L->addBasicBlockToLoop(Tail, *LI);
1082     }
1083   }
1084 
1085   return CheckTerm;
1086 }
1087 
1088 void llvm::SplitBlockAndInsertIfThenElse(Value *Cond, Instruction *SplitBefore,
1089                                          Instruction **ThenTerm,
1090                                          Instruction **ElseTerm,
1091                                          MDNode *BranchWeights) {
1092   BasicBlock *Head = SplitBefore->getParent();
1093   BasicBlock *Tail = Head->splitBasicBlock(SplitBefore->getIterator());
1094   Instruction *HeadOldTerm = Head->getTerminator();
1095   LLVMContext &C = Head->getContext();
1096   BasicBlock *ThenBlock = BasicBlock::Create(C, "", Head->getParent(), Tail);
1097   BasicBlock *ElseBlock = BasicBlock::Create(C, "", Head->getParent(), Tail);
1098   *ThenTerm = BranchInst::Create(Tail, ThenBlock);
1099   (*ThenTerm)->setDebugLoc(SplitBefore->getDebugLoc());
1100   *ElseTerm = BranchInst::Create(Tail, ElseBlock);
1101   (*ElseTerm)->setDebugLoc(SplitBefore->getDebugLoc());
1102   BranchInst *HeadNewTerm =
1103     BranchInst::Create(/*ifTrue*/ThenBlock, /*ifFalse*/ElseBlock, Cond);
1104   HeadNewTerm->setMetadata(LLVMContext::MD_prof, BranchWeights);
1105   ReplaceInstWithInst(HeadOldTerm, HeadNewTerm);
1106 }
1107 
1108 Value *llvm::GetIfCondition(BasicBlock *BB, BasicBlock *&IfTrue,
1109                              BasicBlock *&IfFalse) {
1110   PHINode *SomePHI = dyn_cast<PHINode>(BB->begin());
1111   BasicBlock *Pred1 = nullptr;
1112   BasicBlock *Pred2 = nullptr;
1113 
1114   if (SomePHI) {
1115     if (SomePHI->getNumIncomingValues() != 2)
1116       return nullptr;
1117     Pred1 = SomePHI->getIncomingBlock(0);
1118     Pred2 = SomePHI->getIncomingBlock(1);
1119   } else {
1120     pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
1121     if (PI == PE) // No predecessor
1122       return nullptr;
1123     Pred1 = *PI++;
1124     if (PI == PE) // Only one predecessor
1125       return nullptr;
1126     Pred2 = *PI++;
1127     if (PI != PE) // More than two predecessors
1128       return nullptr;
1129   }
1130 
1131   // We can only handle branches.  Other control flow will be lowered to
1132   // branches if possible anyway.
1133   BranchInst *Pred1Br = dyn_cast<BranchInst>(Pred1->getTerminator());
1134   BranchInst *Pred2Br = dyn_cast<BranchInst>(Pred2->getTerminator());
1135   if (!Pred1Br || !Pred2Br)
1136     return nullptr;
1137 
1138   // Eliminate code duplication by ensuring that Pred1Br is conditional if
1139   // either are.
1140   if (Pred2Br->isConditional()) {
1141     // If both branches are conditional, we don't have an "if statement".  In
1142     // reality, we could transform this case, but since the condition will be
1143     // required anyway, we stand no chance of eliminating it, so the xform is
1144     // probably not profitable.
1145     if (Pred1Br->isConditional())
1146       return nullptr;
1147 
1148     std::swap(Pred1, Pred2);
1149     std::swap(Pred1Br, Pred2Br);
1150   }
1151 
1152   if (Pred1Br->isConditional()) {
1153     // The only thing we have to watch out for here is to make sure that Pred2
1154     // doesn't have incoming edges from other blocks.  If it does, the condition
1155     // doesn't dominate BB.
1156     if (!Pred2->getSinglePredecessor())
1157       return nullptr;
1158 
1159     // If we found a conditional branch predecessor, make sure that it branches
1160     // to BB and Pred2Br.  If it doesn't, this isn't an "if statement".
1161     if (Pred1Br->getSuccessor(0) == BB &&
1162         Pred1Br->getSuccessor(1) == Pred2) {
1163       IfTrue = Pred1;
1164       IfFalse = Pred2;
1165     } else if (Pred1Br->getSuccessor(0) == Pred2 &&
1166                Pred1Br->getSuccessor(1) == BB) {
1167       IfTrue = Pred2;
1168       IfFalse = Pred1;
1169     } else {
1170       // We know that one arm of the conditional goes to BB, so the other must
1171       // go somewhere unrelated, and this must not be an "if statement".
1172       return nullptr;
1173     }
1174 
1175     return Pred1Br->getCondition();
1176   }
1177 
1178   // Ok, if we got here, both predecessors end with an unconditional branch to
1179   // BB.  Don't panic!  If both blocks only have a single (identical)
1180   // predecessor, and THAT is a conditional branch, then we're all ok!
1181   BasicBlock *CommonPred = Pred1->getSinglePredecessor();
1182   if (CommonPred == nullptr || CommonPred != Pred2->getSinglePredecessor())
1183     return nullptr;
1184 
1185   // Otherwise, if this is a conditional branch, then we can use it!
1186   BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator());
1187   if (!BI) return nullptr;
1188 
1189   assert(BI->isConditional() && "Two successors but not conditional?");
1190   if (BI->getSuccessor(0) == Pred1) {
1191     IfTrue = Pred1;
1192     IfFalse = Pred2;
1193   } else {
1194     IfTrue = Pred2;
1195     IfFalse = Pred1;
1196   }
1197   return BI->getCondition();
1198 }
1199 
1200 // After creating a control flow hub, the operands of PHINodes in an outgoing
1201 // block Out no longer match the predecessors of that block. Predecessors of Out
1202 // that are incoming blocks to the hub are now replaced by just one edge from
1203 // the hub. To match this new control flow, the corresponding values from each
1204 // PHINode must now be moved a new PHINode in the first guard block of the hub.
1205 //
1206 // This operation cannot be performed with SSAUpdater, because it involves one
1207 // new use: If the block Out is in the list of Incoming blocks, then the newly
1208 // created PHI in the Hub will use itself along that edge from Out to Hub.
1209 static void reconnectPhis(BasicBlock *Out, BasicBlock *GuardBlock,
1210                           const SetVector<BasicBlock *> &Incoming,
1211                           BasicBlock *FirstGuardBlock) {
1212   auto I = Out->begin();
1213   while (I != Out->end() && isa<PHINode>(I)) {
1214     auto Phi = cast<PHINode>(I);
1215     auto NewPhi =
1216         PHINode::Create(Phi->getType(), Incoming.size(),
1217                         Phi->getName() + ".moved", &FirstGuardBlock->back());
1218     for (auto In : Incoming) {
1219       Value *V = UndefValue::get(Phi->getType());
1220       if (In == Out) {
1221         V = NewPhi;
1222       } else if (Phi->getBasicBlockIndex(In) != -1) {
1223         V = Phi->removeIncomingValue(In, false);
1224       }
1225       NewPhi->addIncoming(V, In);
1226     }
1227     assert(NewPhi->getNumIncomingValues() == Incoming.size());
1228     if (Phi->getNumOperands() == 0) {
1229       Phi->replaceAllUsesWith(NewPhi);
1230       I = Phi->eraseFromParent();
1231       continue;
1232     }
1233     Phi->addIncoming(NewPhi, GuardBlock);
1234     ++I;
1235   }
1236 }
1237 
1238 using BBPredicates = DenseMap<BasicBlock *, PHINode *>;
1239 using BBSetVector = SetVector<BasicBlock *>;
1240 
1241 // Redirects the terminator of the incoming block to the first guard
1242 // block in the hub. The condition of the original terminator (if it
1243 // was conditional) and its original successors are returned as a
1244 // tuple <condition, succ0, succ1>. The function additionally filters
1245 // out successors that are not in the set of outgoing blocks.
1246 //
1247 // - condition is non-null iff the branch is conditional.
1248 // - Succ1 is non-null iff the sole/taken target is an outgoing block.
1249 // - Succ2 is non-null iff condition is non-null and the fallthrough
1250 //         target is an outgoing block.
1251 static std::tuple<Value *, BasicBlock *, BasicBlock *>
1252 redirectToHub(BasicBlock *BB, BasicBlock *FirstGuardBlock,
1253               const BBSetVector &Outgoing) {
1254   auto Branch = cast<BranchInst>(BB->getTerminator());
1255   auto Condition = Branch->isConditional() ? Branch->getCondition() : nullptr;
1256 
1257   BasicBlock *Succ0 = Branch->getSuccessor(0);
1258   BasicBlock *Succ1 = nullptr;
1259   Succ0 = Outgoing.count(Succ0) ? Succ0 : nullptr;
1260 
1261   if (Branch->isUnconditional()) {
1262     Branch->setSuccessor(0, FirstGuardBlock);
1263     assert(Succ0);
1264   } else {
1265     Succ1 = Branch->getSuccessor(1);
1266     Succ1 = Outgoing.count(Succ1) ? Succ1 : nullptr;
1267     assert(Succ0 || Succ1);
1268     if (Succ0 && !Succ1) {
1269       Branch->setSuccessor(0, FirstGuardBlock);
1270     } else if (Succ1 && !Succ0) {
1271       Branch->setSuccessor(1, FirstGuardBlock);
1272     } else {
1273       Branch->eraseFromParent();
1274       BranchInst::Create(FirstGuardBlock, BB);
1275     }
1276   }
1277 
1278   assert(Succ0 || Succ1);
1279   return std::make_tuple(Condition, Succ0, Succ1);
1280 }
1281 
1282 // Capture the existing control flow as guard predicates, and redirect
1283 // control flow from every incoming block to the first guard block in
1284 // the hub.
1285 //
1286 // There is one guard predicate for each outgoing block OutBB. The
1287 // predicate is a PHINode with one input for each InBB which
1288 // represents whether the hub should transfer control flow to OutBB if
1289 // it arrived from InBB. These predicates are NOT ORTHOGONAL. The Hub
1290 // evaluates them in the same order as the Outgoing set-vector, and
1291 // control branches to the first outgoing block whose predicate
1292 // evaluates to true.
1293 static void convertToGuardPredicates(
1294     BasicBlock *FirstGuardBlock, BBPredicates &GuardPredicates,
1295     SmallVectorImpl<WeakVH> &DeletionCandidates, const BBSetVector &Incoming,
1296     const BBSetVector &Outgoing) {
1297   auto &Context = Incoming.front()->getContext();
1298   auto BoolTrue = ConstantInt::getTrue(Context);
1299   auto BoolFalse = ConstantInt::getFalse(Context);
1300 
1301   // The predicate for the last outgoing is trivially true, and so we
1302   // process only the first N-1 successors.
1303   for (int i = 0, e = Outgoing.size() - 1; i != e; ++i) {
1304     auto Out = Outgoing[i];
1305     LLVM_DEBUG(dbgs() << "Creating guard for " << Out->getName() << "\n");
1306     auto Phi =
1307         PHINode::Create(Type::getInt1Ty(Context), Incoming.size(),
1308                         StringRef("Guard.") + Out->getName(), FirstGuardBlock);
1309     GuardPredicates[Out] = Phi;
1310   }
1311 
1312   for (auto In : Incoming) {
1313     Value *Condition;
1314     BasicBlock *Succ0;
1315     BasicBlock *Succ1;
1316     std::tie(Condition, Succ0, Succ1) =
1317         redirectToHub(In, FirstGuardBlock, Outgoing);
1318 
1319     // Optimization: Consider an incoming block A with both successors
1320     // Succ0 and Succ1 in the set of outgoing blocks. The predicates
1321     // for Succ0 and Succ1 complement each other. If Succ0 is visited
1322     // first in the loop below, control will branch to Succ0 using the
1323     // corresponding predicate. But if that branch is not taken, then
1324     // control must reach Succ1, which means that the predicate for
1325     // Succ1 is always true.
1326     bool OneSuccessorDone = false;
1327     for (int i = 0, e = Outgoing.size() - 1; i != e; ++i) {
1328       auto Out = Outgoing[i];
1329       auto Phi = GuardPredicates[Out];
1330       if (Out != Succ0 && Out != Succ1) {
1331         Phi->addIncoming(BoolFalse, In);
1332         continue;
1333       }
1334       // Optimization: When only one successor is an outgoing block,
1335       // the predicate is always true.
1336       if (!Succ0 || !Succ1 || OneSuccessorDone) {
1337         Phi->addIncoming(BoolTrue, In);
1338         continue;
1339       }
1340       assert(Succ0 && Succ1);
1341       OneSuccessorDone = true;
1342       if (Out == Succ0) {
1343         Phi->addIncoming(Condition, In);
1344         continue;
1345       }
1346       auto Inverted = invertCondition(Condition);
1347       DeletionCandidates.push_back(Condition);
1348       Phi->addIncoming(Inverted, In);
1349     }
1350   }
1351 }
1352 
1353 // For each outgoing block OutBB, create a guard block in the Hub. The
1354 // first guard block was already created outside, and available as the
1355 // first element in the vector of guard blocks.
1356 //
1357 // Each guard block terminates in a conditional branch that transfers
1358 // control to the corresponding outgoing block or the next guard
1359 // block. The last guard block has two outgoing blocks as successors
1360 // since the condition for the final outgoing block is trivially
1361 // true. So we create one less block (including the first guard block)
1362 // than the number of outgoing blocks.
1363 static void createGuardBlocks(SmallVectorImpl<BasicBlock *> &GuardBlocks,
1364                               Function *F, const BBSetVector &Outgoing,
1365                               BBPredicates &GuardPredicates, StringRef Prefix) {
1366   for (int i = 0, e = Outgoing.size() - 2; i != e; ++i) {
1367     GuardBlocks.push_back(
1368         BasicBlock::Create(F->getContext(), Prefix + ".guard", F));
1369   }
1370   assert(GuardBlocks.size() == GuardPredicates.size());
1371 
1372   // To help keep the loop simple, temporarily append the last
1373   // outgoing block to the list of guard blocks.
1374   GuardBlocks.push_back(Outgoing.back());
1375 
1376   for (int i = 0, e = GuardBlocks.size() - 1; i != e; ++i) {
1377     auto Out = Outgoing[i];
1378     assert(GuardPredicates.count(Out));
1379     BranchInst::Create(Out, GuardBlocks[i + 1], GuardPredicates[Out],
1380                        GuardBlocks[i]);
1381   }
1382 
1383   // Remove the last block from the guard list.
1384   GuardBlocks.pop_back();
1385 }
1386 
1387 BasicBlock *llvm::CreateControlFlowHub(
1388     DomTreeUpdater *DTU, SmallVectorImpl<BasicBlock *> &GuardBlocks,
1389     const BBSetVector &Incoming, const BBSetVector &Outgoing,
1390     const StringRef Prefix) {
1391   auto F = Incoming.front()->getParent();
1392   auto FirstGuardBlock =
1393       BasicBlock::Create(F->getContext(), Prefix + ".guard", F);
1394 
1395   SmallVector<DominatorTree::UpdateType, 16> Updates;
1396   if (DTU) {
1397     for (auto In : Incoming) {
1398       Updates.push_back({DominatorTree::Insert, In, FirstGuardBlock});
1399       for (auto Succ : successors(In)) {
1400         if (Outgoing.count(Succ))
1401           Updates.push_back({DominatorTree::Delete, In, Succ});
1402       }
1403     }
1404   }
1405 
1406   BBPredicates GuardPredicates;
1407   SmallVector<WeakVH, 8> DeletionCandidates;
1408   convertToGuardPredicates(FirstGuardBlock, GuardPredicates, DeletionCandidates,
1409                            Incoming, Outgoing);
1410 
1411   GuardBlocks.push_back(FirstGuardBlock);
1412   createGuardBlocks(GuardBlocks, F, Outgoing, GuardPredicates, Prefix);
1413 
1414   // Update the PHINodes in each outgoing block to match the new control flow.
1415   for (int i = 0, e = GuardBlocks.size(); i != e; ++i) {
1416     reconnectPhis(Outgoing[i], GuardBlocks[i], Incoming, FirstGuardBlock);
1417   }
1418   reconnectPhis(Outgoing.back(), GuardBlocks.back(), Incoming, FirstGuardBlock);
1419 
1420   if (DTU) {
1421     int NumGuards = GuardBlocks.size();
1422     assert((int)Outgoing.size() == NumGuards + 1);
1423     for (int i = 0; i != NumGuards - 1; ++i) {
1424       Updates.push_back({DominatorTree::Insert, GuardBlocks[i], Outgoing[i]});
1425       Updates.push_back(
1426           {DominatorTree::Insert, GuardBlocks[i], GuardBlocks[i + 1]});
1427     }
1428     Updates.push_back({DominatorTree::Insert, GuardBlocks[NumGuards - 1],
1429                        Outgoing[NumGuards - 1]});
1430     Updates.push_back({DominatorTree::Insert, GuardBlocks[NumGuards - 1],
1431                        Outgoing[NumGuards]});
1432     DTU->applyUpdates(Updates);
1433   }
1434 
1435   for (auto I : DeletionCandidates) {
1436     if (I->use_empty())
1437       if (auto Inst = dyn_cast_or_null<Instruction>(I))
1438         Inst->eraseFromParent();
1439   }
1440 
1441   return FirstGuardBlock;
1442 }
1443