1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Peephole optimize the CFG.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Transforms/Utils/Local.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/ADT/SetVector.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/ConstantFolding.h"
22 #include "llvm/Analysis/InstructionSimplify.h"
23 #include "llvm/Analysis/TargetTransformInfo.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/IR/CFG.h"
26 #include "llvm/IR/ConstantRange.h"
27 #include "llvm/IR/Constants.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/DerivedTypes.h"
30 #include "llvm/IR/GlobalVariable.h"
31 #include "llvm/IR/IRBuilder.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/LLVMContext.h"
35 #include "llvm/IR/MDBuilder.h"
36 #include "llvm/IR/Metadata.h"
37 #include "llvm/IR/Module.h"
38 #include "llvm/IR/NoFolder.h"
39 #include "llvm/IR/Operator.h"
40 #include "llvm/IR/PatternMatch.h"
41 #include "llvm/IR/Type.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/Debug.h"
44 #include "llvm/Support/raw_ostream.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
50 using namespace PatternMatch;
52 #define DEBUG_TYPE "simplifycfg"
54 static cl::opt<unsigned>
55 PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1),
56 cl::desc("Control the amount of phi node folding to perform (default = 1)"));
59 DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false),
60 cl::desc("Duplicate return instructions into unconditional branches"));
63 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
64 cl::desc("Sink common instructions down to the end block"));
66 static cl::opt<bool> HoistCondStores(
67 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
68 cl::desc("Hoist conditional stores if an unconditional store precedes"));
70 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
71 STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables");
72 STATISTIC(NumLookupTablesHoles, "Number of switch instructions turned into lookup tables (holes checked)");
73 STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block");
74 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
77 /// ValueEqualityComparisonCase - Represents a case of a switch.
78 struct ValueEqualityComparisonCase {
82 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
83 : Value(Value), Dest(Dest) {}
85 bool operator<(ValueEqualityComparisonCase RHS) const {
86 // Comparing pointers is ok as we only rely on the order for uniquing.
87 return Value < RHS.Value;
90 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
93 class SimplifyCFGOpt {
94 const TargetTransformInfo &TTI;
95 const DataLayout *const DL;
96 Value *isValueEqualityComparison(TerminatorInst *TI);
97 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
98 std::vector<ValueEqualityComparisonCase> &Cases);
99 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
101 IRBuilder<> &Builder);
102 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
103 IRBuilder<> &Builder);
105 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
106 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
107 bool SimplifyUnreachable(UnreachableInst *UI);
108 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
109 bool SimplifyIndirectBr(IndirectBrInst *IBI);
110 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder);
111 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder);
114 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout *DL)
115 : TTI(TTI), DL(DL) {}
116 bool run(BasicBlock *BB);
120 /// SafeToMergeTerminators - Return true if it is safe to merge these two
121 /// terminator instructions together.
123 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
124 if (SI1 == SI2) return false; // Can't merge with self!
126 // It is not safe to merge these two switch instructions if they have a common
127 // successor, and if that successor has a PHI node, and if *that* PHI node has
128 // conflicting incoming values from the two switch blocks.
129 BasicBlock *SI1BB = SI1->getParent();
130 BasicBlock *SI2BB = SI2->getParent();
131 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
133 for (BasicBlock *Succ : successors(SI2BB))
134 if (SI1Succs.count(Succ))
135 for (BasicBlock::iterator BBI = Succ->begin();
136 isa<PHINode>(BBI); ++BBI) {
137 PHINode *PN = cast<PHINode>(BBI);
138 if (PN->getIncomingValueForBlock(SI1BB) !=
139 PN->getIncomingValueForBlock(SI2BB))
146 /// isProfitableToFoldUnconditional - Return true if it is safe and profitable
147 /// to merge these two terminator instructions together, where SI1 is an
148 /// unconditional branch. PhiNodes will store all PHI nodes in common
151 static bool isProfitableToFoldUnconditional(BranchInst *SI1,
154 SmallVectorImpl<PHINode*> &PhiNodes) {
155 if (SI1 == SI2) return false; // Can't merge with self!
156 assert(SI1->isUnconditional() && SI2->isConditional());
158 // We fold the unconditional branch if we can easily update all PHI nodes in
159 // common successors:
160 // 1> We have a constant incoming value for the conditional branch;
161 // 2> We have "Cond" as the incoming value for the unconditional branch;
162 // 3> SI2->getCondition() and Cond have same operands.
163 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
164 if (!Ci2) return false;
165 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
166 Cond->getOperand(1) == Ci2->getOperand(1)) &&
167 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
168 Cond->getOperand(1) == Ci2->getOperand(0)))
171 BasicBlock *SI1BB = SI1->getParent();
172 BasicBlock *SI2BB = SI2->getParent();
173 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
174 for (BasicBlock *Succ : successors(SI2BB))
175 if (SI1Succs.count(Succ))
176 for (BasicBlock::iterator BBI = Succ->begin();
177 isa<PHINode>(BBI); ++BBI) {
178 PHINode *PN = cast<PHINode>(BBI);
179 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
180 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
182 PhiNodes.push_back(PN);
187 /// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will
188 /// now be entries in it from the 'NewPred' block. The values that will be
189 /// flowing into the PHI nodes will be the same as those coming in from
190 /// ExistPred, an existing predecessor of Succ.
191 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
192 BasicBlock *ExistPred) {
193 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
196 for (BasicBlock::iterator I = Succ->begin();
197 (PN = dyn_cast<PHINode>(I)); ++I)
198 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
201 /// ComputeSpeculationCost - Compute an abstract "cost" of speculating the
202 /// given instruction, which is assumed to be safe to speculate. 1 means
203 /// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive.
204 static unsigned ComputeSpeculationCost(const User *I, const DataLayout *DL) {
205 assert(isSafeToSpeculativelyExecute(I, DL) &&
206 "Instruction is not safe to speculatively execute!");
207 switch (Operator::getOpcode(I)) {
209 // In doubt, be conservative.
211 case Instruction::GetElementPtr:
212 // GEPs are cheap if all indices are constant.
213 if (!cast<GEPOperator>(I)->hasAllConstantIndices())
216 case Instruction::ExtractValue:
217 case Instruction::Load:
218 case Instruction::Add:
219 case Instruction::Sub:
220 case Instruction::And:
221 case Instruction::Or:
222 case Instruction::Xor:
223 case Instruction::Shl:
224 case Instruction::LShr:
225 case Instruction::AShr:
226 case Instruction::ICmp:
227 case Instruction::Trunc:
228 case Instruction::ZExt:
229 case Instruction::SExt:
230 case Instruction::BitCast:
231 case Instruction::ExtractElement:
232 case Instruction::InsertElement:
233 return 1; // These are all cheap.
235 case Instruction::Call:
236 case Instruction::Select:
241 /// DominatesMergePoint - If we have a merge point of an "if condition" as
242 /// accepted above, return true if the specified value dominates the block. We
243 /// don't handle the true generality of domination here, just a special case
244 /// which works well enough for us.
246 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
247 /// see if V (which must be an instruction) and its recursive operands
248 /// that do not dominate BB have a combined cost lower than CostRemaining and
249 /// are non-trapping. If both are true, the instruction is inserted into the
250 /// set and true is returned.
252 /// The cost for most non-trapping instructions is defined as 1 except for
253 /// Select whose cost is 2.
255 /// After this function returns, CostRemaining is decreased by the cost of
256 /// V plus its non-dominating operands. If that cost is greater than
257 /// CostRemaining, false is returned and CostRemaining is undefined.
258 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
259 SmallPtrSet<Instruction*, 4> *AggressiveInsts,
260 unsigned &CostRemaining,
261 const DataLayout *DL) {
262 Instruction *I = dyn_cast<Instruction>(V);
264 // Non-instructions all dominate instructions, but not all constantexprs
265 // can be executed unconditionally.
266 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
271 BasicBlock *PBB = I->getParent();
273 // We don't want to allow weird loops that might have the "if condition" in
274 // the bottom of this block.
275 if (PBB == BB) return false;
277 // If this instruction is defined in a block that contains an unconditional
278 // branch to BB, then it must be in the 'conditional' part of the "if
279 // statement". If not, it definitely dominates the region.
280 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
281 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
284 // If we aren't allowing aggressive promotion anymore, then don't consider
285 // instructions in the 'if region'.
286 if (!AggressiveInsts) return false;
288 // If we have seen this instruction before, don't count it again.
289 if (AggressiveInsts->count(I)) return true;
291 // Okay, it looks like the instruction IS in the "condition". Check to
292 // see if it's a cheap instruction to unconditionally compute, and if it
293 // only uses stuff defined outside of the condition. If so, hoist it out.
294 if (!isSafeToSpeculativelyExecute(I, DL))
297 unsigned Cost = ComputeSpeculationCost(I, DL);
299 if (Cost > CostRemaining)
302 CostRemaining -= Cost;
304 // Okay, we can only really hoist these out if their operands do
305 // not take us over the cost threshold.
306 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
307 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, DL))
309 // Okay, it's safe to do this! Remember this instruction.
310 AggressiveInsts->insert(I);
314 /// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr
315 /// and PointerNullValue. Return NULL if value is not a constant int.
316 static ConstantInt *GetConstantInt(Value *V, const DataLayout *DL) {
317 // Normal constant int.
318 ConstantInt *CI = dyn_cast<ConstantInt>(V);
319 if (CI || !DL || !isa<Constant>(V) || !V->getType()->isPointerTy())
322 // This is some kind of pointer constant. Turn it into a pointer-sized
323 // ConstantInt if possible.
324 IntegerType *PtrTy = cast<IntegerType>(DL->getIntPtrType(V->getType()));
326 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
327 if (isa<ConstantPointerNull>(V))
328 return ConstantInt::get(PtrTy, 0);
330 // IntToPtr const int.
331 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
332 if (CE->getOpcode() == Instruction::IntToPtr)
333 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
334 // The constant is very likely to have the right type already.
335 if (CI->getType() == PtrTy)
338 return cast<ConstantInt>
339 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
344 /// GatherConstantCompares - Given a potentially 'or'd or 'and'd together
345 /// collection of icmp eq/ne instructions that compare a value against a
346 /// constant, return the value being compared, and stick the constant into the
349 GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra,
350 const DataLayout *DL, bool isEQ, unsigned &UsedICmps) {
351 Instruction *I = dyn_cast<Instruction>(V);
352 if (!I) return nullptr;
354 // If this is an icmp against a constant, handle this as one of the cases.
355 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) {
356 if (ConstantInt *C = GetConstantInt(I->getOperand(1), DL)) {
360 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
361 // (x & ~2^x) == y --> x == y || x == y|2^x
362 // This undoes a transformation done by instcombine to fuse 2 compares.
363 if (match(ICI->getOperand(0),
364 m_And(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
365 APInt Not = ~RHSC->getValue();
366 if (Not.isPowerOf2()) {
369 ConstantInt::get(C->getContext(), C->getValue() | Not));
377 return I->getOperand(0);
380 // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to
383 ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue());
385 // Shift the range if the compare is fed by an add. This is the range
386 // compare idiom as emitted by instcombine.
388 match(I->getOperand(0), m_Add(m_Value(RHSVal), m_ConstantInt(RHSC)));
390 Span = Span.subtract(RHSC->getValue());
392 // If this is an and/!= check then we want to optimize "x ugt 2" into
395 Span = Span.inverse();
397 // If there are a ton of values, we don't want to make a ginormous switch.
398 if (Span.getSetSize().ugt(8) || Span.isEmptySet())
401 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
402 Vals.push_back(ConstantInt::get(V->getContext(), Tmp));
404 return hasAdd ? RHSVal : I->getOperand(0);
409 // Otherwise, we can only handle an | or &, depending on isEQ.
410 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And))
413 unsigned NumValsBeforeLHS = Vals.size();
414 unsigned UsedICmpsBeforeLHS = UsedICmps;
415 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, DL,
417 unsigned NumVals = Vals.size();
418 unsigned UsedICmpsBeforeRHS = UsedICmps;
419 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, DL,
423 Vals.resize(NumVals);
424 UsedICmps = UsedICmpsBeforeRHS;
427 // The RHS of the or/and can't be folded in and we haven't used "Extra" yet,
428 // set it and return success.
429 if (Extra == nullptr || Extra == I->getOperand(1)) {
430 Extra = I->getOperand(1);
434 Vals.resize(NumValsBeforeLHS);
435 UsedICmps = UsedICmpsBeforeLHS;
439 // If the LHS can't be folded in, but Extra is available and RHS can, try to
441 if (Extra == nullptr || Extra == I->getOperand(0)) {
442 Value *OldExtra = Extra;
443 Extra = I->getOperand(0);
444 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, DL,
447 assert(Vals.size() == NumValsBeforeLHS);
454 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
455 Instruction *Cond = nullptr;
456 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
457 Cond = dyn_cast<Instruction>(SI->getCondition());
458 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
459 if (BI->isConditional())
460 Cond = dyn_cast<Instruction>(BI->getCondition());
461 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
462 Cond = dyn_cast<Instruction>(IBI->getAddress());
465 TI->eraseFromParent();
466 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
469 /// isValueEqualityComparison - Return true if the specified terminator checks
470 /// to see if a value is equal to constant integer value.
471 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
473 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
474 // Do not permit merging of large switch instructions into their
475 // predecessors unless there is only one predecessor.
476 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()),
477 pred_end(SI->getParent())) <= 128)
478 CV = SI->getCondition();
479 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
480 if (BI->isConditional() && BI->getCondition()->hasOneUse())
481 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition()))
482 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
483 CV = ICI->getOperand(0);
485 // Unwrap any lossless ptrtoint cast.
487 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
488 Value *Ptr = PTII->getPointerOperand();
489 if (PTII->getType() == DL->getIntPtrType(Ptr->getType()))
496 /// GetValueEqualityComparisonCases - Given a value comparison instruction,
497 /// decode all of the 'cases' that it represents and return the 'default' block.
498 BasicBlock *SimplifyCFGOpt::
499 GetValueEqualityComparisonCases(TerminatorInst *TI,
500 std::vector<ValueEqualityComparisonCase>
502 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
503 Cases.reserve(SI->getNumCases());
504 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
505 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
506 i.getCaseSuccessor()));
507 return SI->getDefaultDest();
510 BranchInst *BI = cast<BranchInst>(TI);
511 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
512 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
513 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
516 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
520 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries
521 /// in the list that match the specified block.
522 static void EliminateBlockCases(BasicBlock *BB,
523 std::vector<ValueEqualityComparisonCase> &Cases) {
524 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
527 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as
530 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
531 std::vector<ValueEqualityComparisonCase > &C2) {
532 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
534 // Make V1 be smaller than V2.
535 if (V1->size() > V2->size())
538 if (V1->size() == 0) return false;
539 if (V1->size() == 1) {
541 ConstantInt *TheVal = (*V1)[0].Value;
542 for (unsigned i = 0, e = V2->size(); i != e; ++i)
543 if (TheVal == (*V2)[i].Value)
547 // Otherwise, just sort both lists and compare element by element.
548 array_pod_sort(V1->begin(), V1->end());
549 array_pod_sort(V2->begin(), V2->end());
550 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
551 while (i1 != e1 && i2 != e2) {
552 if ((*V1)[i1].Value == (*V2)[i2].Value)
554 if ((*V1)[i1].Value < (*V2)[i2].Value)
562 /// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a
563 /// terminator instruction and its block is known to only have a single
564 /// predecessor block, check to see if that predecessor is also a value
565 /// comparison with the same value, and if that comparison determines the
566 /// outcome of this comparison. If so, simplify TI. This does a very limited
567 /// form of jump threading.
568 bool SimplifyCFGOpt::
569 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
571 IRBuilder<> &Builder) {
572 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
573 if (!PredVal) return false; // Not a value comparison in predecessor.
575 Value *ThisVal = isValueEqualityComparison(TI);
576 assert(ThisVal && "This isn't a value comparison!!");
577 if (ThisVal != PredVal) return false; // Different predicates.
579 // TODO: Preserve branch weight metadata, similarly to how
580 // FoldValueComparisonIntoPredecessors preserves it.
582 // Find out information about when control will move from Pred to TI's block.
583 std::vector<ValueEqualityComparisonCase> PredCases;
584 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
586 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
588 // Find information about how control leaves this block.
589 std::vector<ValueEqualityComparisonCase> ThisCases;
590 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
591 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
593 // If TI's block is the default block from Pred's comparison, potentially
594 // simplify TI based on this knowledge.
595 if (PredDef == TI->getParent()) {
596 // If we are here, we know that the value is none of those cases listed in
597 // PredCases. If there are any cases in ThisCases that are in PredCases, we
599 if (!ValuesOverlap(PredCases, ThisCases))
602 if (isa<BranchInst>(TI)) {
603 // Okay, one of the successors of this condbr is dead. Convert it to a
605 assert(ThisCases.size() == 1 && "Branch can only have one case!");
606 // Insert the new branch.
607 Instruction *NI = Builder.CreateBr(ThisDef);
610 // Remove PHI node entries for the dead edge.
611 ThisCases[0].Dest->removePredecessor(TI->getParent());
613 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
614 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
616 EraseTerminatorInstAndDCECond(TI);
620 SwitchInst *SI = cast<SwitchInst>(TI);
621 // Okay, TI has cases that are statically dead, prune them away.
622 SmallPtrSet<Constant*, 16> DeadCases;
623 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
624 DeadCases.insert(PredCases[i].Value);
626 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
627 << "Through successor TI: " << *TI);
629 // Collect branch weights into a vector.
630 SmallVector<uint32_t, 8> Weights;
631 MDNode* MD = SI->getMetadata(LLVMContext::MD_prof);
632 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
634 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
636 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i));
638 Weights.push_back(CI->getValue().getZExtValue());
640 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
642 if (DeadCases.count(i.getCaseValue())) {
644 std::swap(Weights[i.getCaseIndex()+1], Weights.back());
647 i.getCaseSuccessor()->removePredecessor(TI->getParent());
651 if (HasWeight && Weights.size() >= 2)
652 SI->setMetadata(LLVMContext::MD_prof,
653 MDBuilder(SI->getParent()->getContext()).
654 createBranchWeights(Weights));
656 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
660 // Otherwise, TI's block must correspond to some matched value. Find out
661 // which value (or set of values) this is.
662 ConstantInt *TIV = nullptr;
663 BasicBlock *TIBB = TI->getParent();
664 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
665 if (PredCases[i].Dest == TIBB) {
667 return false; // Cannot handle multiple values coming to this block.
668 TIV = PredCases[i].Value;
670 assert(TIV && "No edge from pred to succ?");
672 // Okay, we found the one constant that our value can be if we get into TI's
673 // BB. Find out which successor will unconditionally be branched to.
674 BasicBlock *TheRealDest = nullptr;
675 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
676 if (ThisCases[i].Value == TIV) {
677 TheRealDest = ThisCases[i].Dest;
681 // If not handled by any explicit cases, it is handled by the default case.
682 if (!TheRealDest) TheRealDest = ThisDef;
684 // Remove PHI node entries for dead edges.
685 BasicBlock *CheckEdge = TheRealDest;
686 for (BasicBlock *Succ : successors(TIBB))
687 if (Succ != CheckEdge)
688 Succ->removePredecessor(TIBB);
692 // Insert the new branch.
693 Instruction *NI = Builder.CreateBr(TheRealDest);
696 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
697 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
699 EraseTerminatorInstAndDCECond(TI);
704 /// ConstantIntOrdering - This class implements a stable ordering of constant
705 /// integers that does not depend on their address. This is important for
706 /// applications that sort ConstantInt's to ensure uniqueness.
707 struct ConstantIntOrdering {
708 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
709 return LHS->getValue().ult(RHS->getValue());
714 static int ConstantIntSortPredicate(ConstantInt *const *P1,
715 ConstantInt *const *P2) {
716 const ConstantInt *LHS = *P1;
717 const ConstantInt *RHS = *P2;
718 if (LHS->getValue().ult(RHS->getValue()))
720 if (LHS->getValue() == RHS->getValue())
725 static inline bool HasBranchWeights(const Instruction* I) {
726 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof);
727 if (ProfMD && ProfMD->getOperand(0))
728 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
729 return MDS->getString().equals("branch_weights");
734 /// Get Weights of a given TerminatorInst, the default weight is at the front
735 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
737 static void GetBranchWeights(TerminatorInst *TI,
738 SmallVectorImpl<uint64_t> &Weights) {
739 MDNode* MD = TI->getMetadata(LLVMContext::MD_prof);
741 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
742 ConstantInt *CI = cast<ConstantInt>(MD->getOperand(i));
743 Weights.push_back(CI->getValue().getZExtValue());
746 // If TI is a conditional eq, the default case is the false case,
747 // and the corresponding branch-weight data is at index 2. We swap the
748 // default weight to be the first entry.
749 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
750 assert(Weights.size() == 2);
751 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
752 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
753 std::swap(Weights.front(), Weights.back());
757 /// Keep halving the weights until all can fit in uint32_t.
758 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
759 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
760 if (Max > UINT_MAX) {
761 unsigned Offset = 32 - countLeadingZeros(Max);
762 for (uint64_t &I : Weights)
767 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value
768 /// equality comparison instruction (either a switch or a branch on "X == c").
769 /// See if any of the predecessors of the terminator block are value comparisons
770 /// on the same value. If so, and if safe to do so, fold them together.
771 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
772 IRBuilder<> &Builder) {
773 BasicBlock *BB = TI->getParent();
774 Value *CV = isValueEqualityComparison(TI); // CondVal
775 assert(CV && "Not a comparison?");
776 bool Changed = false;
778 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
779 while (!Preds.empty()) {
780 BasicBlock *Pred = Preds.pop_back_val();
782 // See if the predecessor is a comparison with the same value.
783 TerminatorInst *PTI = Pred->getTerminator();
784 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
786 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
787 // Figure out which 'cases' to copy from SI to PSI.
788 std::vector<ValueEqualityComparisonCase> BBCases;
789 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
791 std::vector<ValueEqualityComparisonCase> PredCases;
792 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
794 // Based on whether the default edge from PTI goes to BB or not, fill in
795 // PredCases and PredDefault with the new switch cases we would like to
797 SmallVector<BasicBlock*, 8> NewSuccessors;
799 // Update the branch weight metadata along the way
800 SmallVector<uint64_t, 8> Weights;
801 bool PredHasWeights = HasBranchWeights(PTI);
802 bool SuccHasWeights = HasBranchWeights(TI);
804 if (PredHasWeights) {
805 GetBranchWeights(PTI, Weights);
806 // branch-weight metadata is inconsistent here.
807 if (Weights.size() != 1 + PredCases.size())
808 PredHasWeights = SuccHasWeights = false;
809 } else if (SuccHasWeights)
810 // If there are no predecessor weights but there are successor weights,
811 // populate Weights with 1, which will later be scaled to the sum of
812 // successor's weights
813 Weights.assign(1 + PredCases.size(), 1);
815 SmallVector<uint64_t, 8> SuccWeights;
816 if (SuccHasWeights) {
817 GetBranchWeights(TI, SuccWeights);
818 // branch-weight metadata is inconsistent here.
819 if (SuccWeights.size() != 1 + BBCases.size())
820 PredHasWeights = SuccHasWeights = false;
821 } else if (PredHasWeights)
822 SuccWeights.assign(1 + BBCases.size(), 1);
824 if (PredDefault == BB) {
825 // If this is the default destination from PTI, only the edges in TI
826 // that don't occur in PTI, or that branch to BB will be activated.
827 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
828 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
829 if (PredCases[i].Dest != BB)
830 PTIHandled.insert(PredCases[i].Value);
832 // The default destination is BB, we don't need explicit targets.
833 std::swap(PredCases[i], PredCases.back());
835 if (PredHasWeights || SuccHasWeights) {
836 // Increase weight for the default case.
837 Weights[0] += Weights[i+1];
838 std::swap(Weights[i+1], Weights.back());
842 PredCases.pop_back();
846 // Reconstruct the new switch statement we will be building.
847 if (PredDefault != BBDefault) {
848 PredDefault->removePredecessor(Pred);
849 PredDefault = BBDefault;
850 NewSuccessors.push_back(BBDefault);
853 unsigned CasesFromPred = Weights.size();
854 uint64_t ValidTotalSuccWeight = 0;
855 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
856 if (!PTIHandled.count(BBCases[i].Value) &&
857 BBCases[i].Dest != BBDefault) {
858 PredCases.push_back(BBCases[i]);
859 NewSuccessors.push_back(BBCases[i].Dest);
860 if (SuccHasWeights || PredHasWeights) {
861 // The default weight is at index 0, so weight for the ith case
862 // should be at index i+1. Scale the cases from successor by
863 // PredDefaultWeight (Weights[0]).
864 Weights.push_back(Weights[0] * SuccWeights[i+1]);
865 ValidTotalSuccWeight += SuccWeights[i+1];
869 if (SuccHasWeights || PredHasWeights) {
870 ValidTotalSuccWeight += SuccWeights[0];
871 // Scale the cases from predecessor by ValidTotalSuccWeight.
872 for (unsigned i = 1; i < CasesFromPred; ++i)
873 Weights[i] *= ValidTotalSuccWeight;
874 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
875 Weights[0] *= SuccWeights[0];
878 // If this is not the default destination from PSI, only the edges
879 // in SI that occur in PSI with a destination of BB will be
881 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
882 std::map<ConstantInt*, uint64_t> WeightsForHandled;
883 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
884 if (PredCases[i].Dest == BB) {
885 PTIHandled.insert(PredCases[i].Value);
887 if (PredHasWeights || SuccHasWeights) {
888 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
889 std::swap(Weights[i+1], Weights.back());
893 std::swap(PredCases[i], PredCases.back());
894 PredCases.pop_back();
898 // Okay, now we know which constants were sent to BB from the
899 // predecessor. Figure out where they will all go now.
900 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
901 if (PTIHandled.count(BBCases[i].Value)) {
902 // If this is one we are capable of getting...
903 if (PredHasWeights || SuccHasWeights)
904 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
905 PredCases.push_back(BBCases[i]);
906 NewSuccessors.push_back(BBCases[i].Dest);
907 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
910 // If there are any constants vectored to BB that TI doesn't handle,
911 // they must go to the default destination of TI.
912 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
914 E = PTIHandled.end(); I != E; ++I) {
915 if (PredHasWeights || SuccHasWeights)
916 Weights.push_back(WeightsForHandled[*I]);
917 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
918 NewSuccessors.push_back(BBDefault);
922 // Okay, at this point, we know which new successor Pred will get. Make
923 // sure we update the number of entries in the PHI nodes for these
925 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i)
926 AddPredecessorToBlock(NewSuccessors[i], Pred, BB);
928 Builder.SetInsertPoint(PTI);
929 // Convert pointer to int before we switch.
930 if (CV->getType()->isPointerTy()) {
931 assert(DL && "Cannot switch on pointer without DataLayout");
932 CV = Builder.CreatePtrToInt(CV, DL->getIntPtrType(CV->getType()),
936 // Now that the successors are updated, create the new Switch instruction.
937 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
939 NewSI->setDebugLoc(PTI->getDebugLoc());
940 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
941 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest);
943 if (PredHasWeights || SuccHasWeights) {
944 // Halve the weights if any of them cannot fit in an uint32_t
947 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
949 NewSI->setMetadata(LLVMContext::MD_prof,
950 MDBuilder(BB->getContext()).
951 createBranchWeights(MDWeights));
954 EraseTerminatorInstAndDCECond(PTI);
956 // Okay, last check. If BB is still a successor of PSI, then we must
957 // have an infinite loop case. If so, add an infinitely looping block
958 // to handle the case to preserve the behavior of the code.
959 BasicBlock *InfLoopBlock = nullptr;
960 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
961 if (NewSI->getSuccessor(i) == BB) {
963 // Insert it at the end of the function, because it's either code,
964 // or it won't matter if it's hot. :)
965 InfLoopBlock = BasicBlock::Create(BB->getContext(),
966 "infloop", BB->getParent());
967 BranchInst::Create(InfLoopBlock, InfLoopBlock);
969 NewSI->setSuccessor(i, InfLoopBlock);
978 // isSafeToHoistInvoke - If we would need to insert a select that uses the
979 // value of this invoke (comments in HoistThenElseCodeToIf explain why we
980 // would need to do this), we can't hoist the invoke, as there is nowhere
981 // to put the select in this case.
982 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
983 Instruction *I1, Instruction *I2) {
984 for (BasicBlock *Succ : successors(BB1)) {
986 for (BasicBlock::iterator BBI = Succ->begin();
987 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
988 Value *BB1V = PN->getIncomingValueForBlock(BB1);
989 Value *BB2V = PN->getIncomingValueForBlock(BB2);
990 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
998 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
999 /// BB2, hoist any common code in the two blocks up into the branch block. The
1000 /// caller of this function guarantees that BI's block dominates BB1 and BB2.
1001 static bool HoistThenElseCodeToIf(BranchInst *BI, const DataLayout *DL) {
1002 // This does very trivial matching, with limited scanning, to find identical
1003 // instructions in the two blocks. In particular, we don't want to get into
1004 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1005 // such, we currently just scan for obviously identical instructions in an
1007 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1008 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1010 BasicBlock::iterator BB1_Itr = BB1->begin();
1011 BasicBlock::iterator BB2_Itr = BB2->begin();
1013 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++;
1014 // Skip debug info if it is not identical.
1015 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1016 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1017 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1018 while (isa<DbgInfoIntrinsic>(I1))
1020 while (isa<DbgInfoIntrinsic>(I2))
1023 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1024 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1027 BasicBlock *BIParent = BI->getParent();
1029 bool Changed = false;
1031 // If we are hoisting the terminator instruction, don't move one (making a
1032 // broken BB), instead clone it, and remove BI.
1033 if (isa<TerminatorInst>(I1))
1034 goto HoistTerminator;
1036 // For a normal instruction, we just move one to right before the branch,
1037 // then replace all uses of the other with the first. Finally, we remove
1038 // the now redundant second instruction.
1039 BIParent->getInstList().splice(BI, BB1->getInstList(), I1);
1040 if (!I2->use_empty())
1041 I2->replaceAllUsesWith(I1);
1042 I1->intersectOptionalDataWith(I2);
1043 I2->eraseFromParent();
1048 // Skip debug info if it is not identical.
1049 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1050 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1051 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1052 while (isa<DbgInfoIntrinsic>(I1))
1054 while (isa<DbgInfoIntrinsic>(I2))
1057 } while (I1->isIdenticalToWhenDefined(I2));
1062 // It may not be possible to hoist an invoke.
1063 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1066 for (BasicBlock *Succ : successors(BB1)) {
1068 for (BasicBlock::iterator BBI = Succ->begin();
1069 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1070 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1071 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1075 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V, DL))
1077 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V, DL))
1082 // Okay, it is safe to hoist the terminator.
1083 Instruction *NT = I1->clone();
1084 BIParent->getInstList().insert(BI, NT);
1085 if (!NT->getType()->isVoidTy()) {
1086 I1->replaceAllUsesWith(NT);
1087 I2->replaceAllUsesWith(NT);
1091 IRBuilder<true, NoFolder> Builder(NT);
1092 // Hoisting one of the terminators from our successor is a great thing.
1093 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1094 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1095 // nodes, so we insert select instruction to compute the final result.
1096 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1097 for (BasicBlock *Succ : successors(BB1)) {
1099 for (BasicBlock::iterator BBI = Succ->begin();
1100 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1101 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1102 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1103 if (BB1V == BB2V) continue;
1105 // These values do not agree. Insert a select instruction before NT
1106 // that determines the right value.
1107 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1109 SI = cast<SelectInst>
1110 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1111 BB1V->getName()+"."+BB2V->getName()));
1113 // Make the PHI node use the select for all incoming values for BB1/BB2
1114 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1115 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1116 PN->setIncomingValue(i, SI);
1120 // Update any PHI nodes in our new successors.
1121 for (BasicBlock *Succ : successors(BB1))
1122 AddPredecessorToBlock(Succ, BIParent, BB1);
1124 EraseTerminatorInstAndDCECond(BI);
1128 /// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd,
1129 /// check whether BBEnd has only two predecessors and the other predecessor
1130 /// ends with an unconditional branch. If it is true, sink any common code
1131 /// in the two predecessors to BBEnd.
1132 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1133 assert(BI1->isUnconditional());
1134 BasicBlock *BB1 = BI1->getParent();
1135 BasicBlock *BBEnd = BI1->getSuccessor(0);
1137 // Check that BBEnd has two predecessors and the other predecessor ends with
1138 // an unconditional branch.
1139 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1140 BasicBlock *Pred0 = *PI++;
1141 if (PI == PE) // Only one predecessor.
1143 BasicBlock *Pred1 = *PI++;
1144 if (PI != PE) // More than two predecessors.
1146 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1147 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1148 if (!BI2 || !BI2->isUnconditional())
1151 // Gather the PHI nodes in BBEnd.
1152 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2;
1153 Instruction *FirstNonPhiInBBEnd = nullptr;
1154 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end();
1156 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1157 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1158 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1159 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN);
1161 FirstNonPhiInBBEnd = &*I;
1165 if (!FirstNonPhiInBBEnd)
1169 // This does very trivial matching, with limited scanning, to find identical
1170 // instructions in the two blocks. We scan backward for obviously identical
1171 // instructions in an identical order.
1172 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1173 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(),
1174 RE2 = BB2->getInstList().rend();
1176 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1179 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1182 // Skip the unconditional branches.
1186 bool Changed = false;
1187 while (RI1 != RE1 && RI2 != RE2) {
1189 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1192 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1196 Instruction *I1 = &*RI1, *I2 = &*RI2;
1197 // I1 and I2 should have a single use in the same PHI node, and they
1198 // perform the same operation.
1199 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1200 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1201 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1202 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) ||
1203 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1204 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1205 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1206 !I1->hasOneUse() || !I2->hasOneUse() ||
1207 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() ||
1208 MapValueFromBB1ToBB2[I1].first != I2)
1211 // Check whether we should swap the operands of ICmpInst.
1212 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1213 bool SwapOpnds = false;
1214 if (ICmp1 && ICmp2 &&
1215 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1216 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1217 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1218 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1219 ICmp2->swapOperands();
1222 if (!I1->isSameOperationAs(I2)) {
1224 ICmp2->swapOperands();
1228 // The operands should be either the same or they need to be generated
1229 // with a PHI node after sinking. We only handle the case where there is
1230 // a single pair of different operands.
1231 Value *DifferentOp1 = nullptr, *DifferentOp2 = nullptr;
1232 unsigned Op1Idx = 0;
1233 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1234 if (I1->getOperand(I) == I2->getOperand(I))
1236 // Early exit if we have more-than one pair of different operands or
1237 // the different operand is already in MapValueFromBB1ToBB2.
1238 // Early exit if we need a PHI node to replace a constant.
1240 MapValueFromBB1ToBB2.find(I1->getOperand(I)) !=
1241 MapValueFromBB1ToBB2.end() ||
1242 isa<Constant>(I1->getOperand(I)) ||
1243 isa<Constant>(I2->getOperand(I))) {
1244 // If we can't sink the instructions, undo the swapping.
1246 ICmp2->swapOperands();
1249 DifferentOp1 = I1->getOperand(I);
1251 DifferentOp2 = I2->getOperand(I);
1254 // We insert the pair of different operands to MapValueFromBB1ToBB2 and
1255 // remove (I1, I2) from MapValueFromBB1ToBB2.
1257 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2,
1258 DifferentOp1->getName() + ".sink",
1260 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN);
1261 // I1 should use NewPN instead of DifferentOp1.
1262 I1->setOperand(Op1Idx, NewPN);
1263 NewPN->addIncoming(DifferentOp1, BB1);
1264 NewPN->addIncoming(DifferentOp2, BB2);
1265 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1267 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second;
1268 MapValueFromBB1ToBB2.erase(I1);
1270 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";);
1271 DEBUG(dbgs() << " " << *I2 << "\n";);
1272 // We need to update RE1 and RE2 if we are going to sink the first
1273 // instruction in the basic block down.
1274 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1275 // Sink the instruction.
1276 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1);
1277 if (!OldPN->use_empty())
1278 OldPN->replaceAllUsesWith(I1);
1279 OldPN->eraseFromParent();
1281 if (!I2->use_empty())
1282 I2->replaceAllUsesWith(I1);
1283 I1->intersectOptionalDataWith(I2);
1284 I2->eraseFromParent();
1287 RE1 = BB1->getInstList().rend();
1289 RE2 = BB2->getInstList().rend();
1290 FirstNonPhiInBBEnd = I1;
1297 /// \brief Determine if we can hoist sink a sole store instruction out of a
1298 /// conditional block.
1300 /// We are looking for code like the following:
1302 /// store i32 %add, i32* %arrayidx2
1303 /// ... // No other stores or function calls (we could be calling a memory
1304 /// ... // function).
1305 /// %cmp = icmp ult %x, %y
1306 /// br i1 %cmp, label %EndBB, label %ThenBB
1308 /// store i32 %add5, i32* %arrayidx2
1312 /// We are going to transform this into:
1314 /// store i32 %add, i32* %arrayidx2
1316 /// %cmp = icmp ult %x, %y
1317 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1318 /// store i32 %add.add5, i32* %arrayidx2
1321 /// \return The pointer to the value of the previous store if the store can be
1322 /// hoisted into the predecessor block. 0 otherwise.
1323 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1324 BasicBlock *StoreBB, BasicBlock *EndBB) {
1325 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1329 // Volatile or atomic.
1330 if (!StoreToHoist->isSimple())
1333 Value *StorePtr = StoreToHoist->getPointerOperand();
1335 // Look for a store to the same pointer in BrBB.
1336 unsigned MaxNumInstToLookAt = 10;
1337 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(),
1338 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) {
1339 Instruction *CurI = &*RI;
1341 // Could be calling an instruction that effects memory like free().
1342 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI))
1345 StoreInst *SI = dyn_cast<StoreInst>(CurI);
1346 // Found the previous store make sure it stores to the same location.
1347 if (SI && SI->getPointerOperand() == StorePtr)
1348 // Found the previous store, return its value operand.
1349 return SI->getValueOperand();
1351 return nullptr; // Unknown store.
1357 /// \brief Speculate a conditional basic block flattening the CFG.
1359 /// Note that this is a very risky transform currently. Speculating
1360 /// instructions like this is most often not desirable. Instead, there is an MI
1361 /// pass which can do it with full awareness of the resource constraints.
1362 /// However, some cases are "obvious" and we should do directly. An example of
1363 /// this is speculating a single, reasonably cheap instruction.
1365 /// There is only one distinct advantage to flattening the CFG at the IR level:
1366 /// it makes very common but simplistic optimizations such as are common in
1367 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1368 /// modeling their effects with easier to reason about SSA value graphs.
1371 /// An illustration of this transform is turning this IR:
1374 /// %cmp = icmp ult %x, %y
1375 /// br i1 %cmp, label %EndBB, label %ThenBB
1377 /// %sub = sub %x, %y
1380 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1387 /// %cmp = icmp ult %x, %y
1388 /// %sub = sub %x, %y
1389 /// %cond = select i1 %cmp, 0, %sub
1393 /// \returns true if the conditional block is removed.
1394 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1395 const DataLayout *DL) {
1396 // Be conservative for now. FP select instruction can often be expensive.
1397 Value *BrCond = BI->getCondition();
1398 if (isa<FCmpInst>(BrCond))
1401 BasicBlock *BB = BI->getParent();
1402 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1404 // If ThenBB is actually on the false edge of the conditional branch, remember
1405 // to swap the select operands later.
1406 bool Invert = false;
1407 if (ThenBB != BI->getSuccessor(0)) {
1408 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1411 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1413 // Keep a count of how many times instructions are used within CondBB when
1414 // they are candidates for sinking into CondBB. Specifically:
1415 // - They are defined in BB, and
1416 // - They have no side effects, and
1417 // - All of their uses are in CondBB.
1418 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1420 unsigned SpeculationCost = 0;
1421 Value *SpeculatedStoreValue = nullptr;
1422 StoreInst *SpeculatedStore = nullptr;
1423 for (BasicBlock::iterator BBI = ThenBB->begin(),
1424 BBE = std::prev(ThenBB->end());
1425 BBI != BBE; ++BBI) {
1426 Instruction *I = BBI;
1428 if (isa<DbgInfoIntrinsic>(I))
1431 // Only speculatively execution a single instruction (not counting the
1432 // terminator) for now.
1434 if (SpeculationCost > 1)
1437 // Don't hoist the instruction if it's unsafe or expensive.
1438 if (!isSafeToSpeculativelyExecute(I, DL) &&
1439 !(HoistCondStores &&
1440 (SpeculatedStoreValue = isSafeToSpeculateStore(I, BB, ThenBB,
1443 if (!SpeculatedStoreValue &&
1444 ComputeSpeculationCost(I, DL) > PHINodeFoldingThreshold)
1447 // Store the store speculation candidate.
1448 if (SpeculatedStoreValue)
1449 SpeculatedStore = cast<StoreInst>(I);
1451 // Do not hoist the instruction if any of its operands are defined but not
1452 // used in BB. The transformation will prevent the operand from
1453 // being sunk into the use block.
1454 for (User::op_iterator i = I->op_begin(), e = I->op_end();
1456 Instruction *OpI = dyn_cast<Instruction>(*i);
1457 if (!OpI || OpI->getParent() != BB ||
1458 OpI->mayHaveSideEffects())
1459 continue; // Not a candidate for sinking.
1461 ++SinkCandidateUseCounts[OpI];
1465 // Consider any sink candidates which are only used in CondBB as costs for
1466 // speculation. Note, while we iterate over a DenseMap here, we are summing
1467 // and so iteration order isn't significant.
1468 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I =
1469 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end();
1471 if (I->first->getNumUses() == I->second) {
1473 if (SpeculationCost > 1)
1477 // Check that the PHI nodes can be converted to selects.
1478 bool HaveRewritablePHIs = false;
1479 for (BasicBlock::iterator I = EndBB->begin();
1480 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1481 Value *OrigV = PN->getIncomingValueForBlock(BB);
1482 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1484 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1485 // Skip PHIs which are trivial.
1489 HaveRewritablePHIs = true;
1490 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1491 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1492 if (!OrigCE && !ThenCE)
1493 continue; // Known safe and cheap.
1495 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE, DL)) ||
1496 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE, DL)))
1498 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, DL) : 0;
1499 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, DL) : 0;
1500 if (OrigCost + ThenCost > 2 * PHINodeFoldingThreshold)
1503 // Account for the cost of an unfolded ConstantExpr which could end up
1504 // getting expanded into Instructions.
1505 // FIXME: This doesn't account for how many operations are combined in the
1506 // constant expression.
1508 if (SpeculationCost > 1)
1512 // If there are no PHIs to process, bail early. This helps ensure idempotence
1514 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
1517 // If we get here, we can hoist the instruction and if-convert.
1518 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1520 // Insert a select of the value of the speculated store.
1521 if (SpeculatedStoreValue) {
1522 IRBuilder<true, NoFolder> Builder(BI);
1523 Value *TrueV = SpeculatedStore->getValueOperand();
1524 Value *FalseV = SpeculatedStoreValue;
1526 std::swap(TrueV, FalseV);
1527 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() +
1528 "." + FalseV->getName());
1529 SpeculatedStore->setOperand(0, S);
1532 // Hoist the instructions.
1533 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(),
1534 std::prev(ThenBB->end()));
1536 // Insert selects and rewrite the PHI operands.
1537 IRBuilder<true, NoFolder> Builder(BI);
1538 for (BasicBlock::iterator I = EndBB->begin();
1539 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1540 unsigned OrigI = PN->getBasicBlockIndex(BB);
1541 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1542 Value *OrigV = PN->getIncomingValue(OrigI);
1543 Value *ThenV = PN->getIncomingValue(ThenI);
1545 // Skip PHIs which are trivial.
1549 // Create a select whose true value is the speculatively executed value and
1550 // false value is the preexisting value. Swap them if the branch
1551 // destinations were inverted.
1552 Value *TrueV = ThenV, *FalseV = OrigV;
1554 std::swap(TrueV, FalseV);
1555 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV,
1556 TrueV->getName() + "." + FalseV->getName());
1557 PN->setIncomingValue(OrigI, V);
1558 PN->setIncomingValue(ThenI, V);
1565 /// \returns True if this block contains a CallInst with the NoDuplicate
1567 static bool HasNoDuplicateCall(const BasicBlock *BB) {
1568 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1569 const CallInst *CI = dyn_cast<CallInst>(I);
1572 if (CI->cannotDuplicate())
1578 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1579 /// across this block.
1580 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1581 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1584 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1585 if (isa<DbgInfoIntrinsic>(BBI))
1587 if (Size > 10) return false; // Don't clone large BB's.
1590 // We can only support instructions that do not define values that are
1591 // live outside of the current basic block.
1592 for (User *U : BBI->users()) {
1593 Instruction *UI = cast<Instruction>(U);
1594 if (UI->getParent() != BB || isa<PHINode>(UI)) return false;
1597 // Looks ok, continue checking.
1603 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1604 /// that is defined in the same block as the branch and if any PHI entries are
1605 /// constants, thread edges corresponding to that entry to be branches to their
1606 /// ultimate destination.
1607 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *DL) {
1608 BasicBlock *BB = BI->getParent();
1609 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1610 // NOTE: we currently cannot transform this case if the PHI node is used
1611 // outside of the block.
1612 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1615 // Degenerate case of a single entry PHI.
1616 if (PN->getNumIncomingValues() == 1) {
1617 FoldSingleEntryPHINodes(PN->getParent());
1621 // Now we know that this block has multiple preds and two succs.
1622 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1624 if (HasNoDuplicateCall(BB)) return false;
1626 // Okay, this is a simple enough basic block. See if any phi values are
1628 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1629 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1630 if (!CB || !CB->getType()->isIntegerTy(1)) continue;
1632 // Okay, we now know that all edges from PredBB should be revectored to
1633 // branch to RealDest.
1634 BasicBlock *PredBB = PN->getIncomingBlock(i);
1635 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1637 if (RealDest == BB) continue; // Skip self loops.
1638 // Skip if the predecessor's terminator is an indirect branch.
1639 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1641 // The dest block might have PHI nodes, other predecessors and other
1642 // difficult cases. Instead of being smart about this, just insert a new
1643 // block that jumps to the destination block, effectively splitting
1644 // the edge we are about to create.
1645 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1646 RealDest->getName()+".critedge",
1647 RealDest->getParent(), RealDest);
1648 BranchInst::Create(RealDest, EdgeBB);
1650 // Update PHI nodes.
1651 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1653 // BB may have instructions that are being threaded over. Clone these
1654 // instructions into EdgeBB. We know that there will be no uses of the
1655 // cloned instructions outside of EdgeBB.
1656 BasicBlock::iterator InsertPt = EdgeBB->begin();
1657 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1658 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1659 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1660 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1663 // Clone the instruction.
1664 Instruction *N = BBI->clone();
1665 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1667 // Update operands due to translation.
1668 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1670 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1671 if (PI != TranslateMap.end())
1675 // Check for trivial simplification.
1676 if (Value *V = SimplifyInstruction(N, DL)) {
1677 TranslateMap[BBI] = V;
1678 delete N; // Instruction folded away, don't need actual inst
1680 // Insert the new instruction into its new home.
1681 EdgeBB->getInstList().insert(InsertPt, N);
1682 if (!BBI->use_empty())
1683 TranslateMap[BBI] = N;
1687 // Loop over all of the edges from PredBB to BB, changing them to branch
1688 // to EdgeBB instead.
1689 TerminatorInst *PredBBTI = PredBB->getTerminator();
1690 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1691 if (PredBBTI->getSuccessor(i) == BB) {
1692 BB->removePredecessor(PredBB);
1693 PredBBTI->setSuccessor(i, EdgeBB);
1696 // Recurse, simplifying any other constants.
1697 return FoldCondBranchOnPHI(BI, DL) | true;
1703 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1704 /// PHI node, see if we can eliminate it.
1705 static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *DL) {
1706 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1707 // statement", which has a very simple dominance structure. Basically, we
1708 // are trying to find the condition that is being branched on, which
1709 // subsequently causes this merge to happen. We really want control
1710 // dependence information for this check, but simplifycfg can't keep it up
1711 // to date, and this catches most of the cases we care about anyway.
1712 BasicBlock *BB = PN->getParent();
1713 BasicBlock *IfTrue, *IfFalse;
1714 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1716 // Don't bother if the branch will be constant folded trivially.
1717 isa<ConstantInt>(IfCond))
1720 // Okay, we found that we can merge this two-entry phi node into a select.
1721 // Doing so would require us to fold *all* two entry phi nodes in this block.
1722 // At some point this becomes non-profitable (particularly if the target
1723 // doesn't support cmov's). Only do this transformation if there are two or
1724 // fewer PHI nodes in this block.
1725 unsigned NumPhis = 0;
1726 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1730 // Loop over the PHI's seeing if we can promote them all to select
1731 // instructions. While we are at it, keep track of the instructions
1732 // that need to be moved to the dominating block.
1733 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1734 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1735 MaxCostVal1 = PHINodeFoldingThreshold;
1737 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1738 PHINode *PN = cast<PHINode>(II++);
1739 if (Value *V = SimplifyInstruction(PN, DL)) {
1740 PN->replaceAllUsesWith(V);
1741 PN->eraseFromParent();
1745 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1747 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1752 // If we folded the first phi, PN dangles at this point. Refresh it. If
1753 // we ran out of PHIs then we simplified them all.
1754 PN = dyn_cast<PHINode>(BB->begin());
1755 if (!PN) return true;
1757 // Don't fold i1 branches on PHIs which contain binary operators. These can
1758 // often be turned into switches and other things.
1759 if (PN->getType()->isIntegerTy(1) &&
1760 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1761 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1762 isa<BinaryOperator>(IfCond)))
1765 // If we all PHI nodes are promotable, check to make sure that all
1766 // instructions in the predecessor blocks can be promoted as well. If
1767 // not, we won't be able to get rid of the control flow, so it's not
1768 // worth promoting to select instructions.
1769 BasicBlock *DomBlock = nullptr;
1770 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1771 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1772 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1775 DomBlock = *pred_begin(IfBlock1);
1776 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1777 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1778 // This is not an aggressive instruction that we can promote.
1779 // Because of this, we won't be able to get rid of the control
1780 // flow, so the xform is not worth it.
1785 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1788 DomBlock = *pred_begin(IfBlock2);
1789 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1790 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1791 // This is not an aggressive instruction that we can promote.
1792 // Because of this, we won't be able to get rid of the control
1793 // flow, so the xform is not worth it.
1798 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1799 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1801 // If we can still promote the PHI nodes after this gauntlet of tests,
1802 // do all of the PHI's now.
1803 Instruction *InsertPt = DomBlock->getTerminator();
1804 IRBuilder<true, NoFolder> Builder(InsertPt);
1806 // Move all 'aggressive' instructions, which are defined in the
1807 // conditional parts of the if's up to the dominating block.
1809 DomBlock->getInstList().splice(InsertPt,
1810 IfBlock1->getInstList(), IfBlock1->begin(),
1811 IfBlock1->getTerminator());
1813 DomBlock->getInstList().splice(InsertPt,
1814 IfBlock2->getInstList(), IfBlock2->begin(),
1815 IfBlock2->getTerminator());
1817 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1818 // Change the PHI node into a select instruction.
1819 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1820 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1823 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1824 PN->replaceAllUsesWith(NV);
1826 PN->eraseFromParent();
1829 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1830 // has been flattened. Change DomBlock to jump directly to our new block to
1831 // avoid other simplifycfg's kicking in on the diamond.
1832 TerminatorInst *OldTI = DomBlock->getTerminator();
1833 Builder.SetInsertPoint(OldTI);
1834 Builder.CreateBr(BB);
1835 OldTI->eraseFromParent();
1839 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1840 /// to two returning blocks, try to merge them together into one return,
1841 /// introducing a select if the return values disagree.
1842 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1843 IRBuilder<> &Builder) {
1844 assert(BI->isConditional() && "Must be a conditional branch");
1845 BasicBlock *TrueSucc = BI->getSuccessor(0);
1846 BasicBlock *FalseSucc = BI->getSuccessor(1);
1847 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1848 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1850 // Check to ensure both blocks are empty (just a return) or optionally empty
1851 // with PHI nodes. If there are other instructions, merging would cause extra
1852 // computation on one path or the other.
1853 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1855 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1858 Builder.SetInsertPoint(BI);
1859 // Okay, we found a branch that is going to two return nodes. If
1860 // there is no return value for this function, just change the
1861 // branch into a return.
1862 if (FalseRet->getNumOperands() == 0) {
1863 TrueSucc->removePredecessor(BI->getParent());
1864 FalseSucc->removePredecessor(BI->getParent());
1865 Builder.CreateRetVoid();
1866 EraseTerminatorInstAndDCECond(BI);
1870 // Otherwise, figure out what the true and false return values are
1871 // so we can insert a new select instruction.
1872 Value *TrueValue = TrueRet->getReturnValue();
1873 Value *FalseValue = FalseRet->getReturnValue();
1875 // Unwrap any PHI nodes in the return blocks.
1876 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1877 if (TVPN->getParent() == TrueSucc)
1878 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1879 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1880 if (FVPN->getParent() == FalseSucc)
1881 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1883 // In order for this transformation to be safe, we must be able to
1884 // unconditionally execute both operands to the return. This is
1885 // normally the case, but we could have a potentially-trapping
1886 // constant expression that prevents this transformation from being
1888 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1891 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1895 // Okay, we collected all the mapped values and checked them for sanity, and
1896 // defined to really do this transformation. First, update the CFG.
1897 TrueSucc->removePredecessor(BI->getParent());
1898 FalseSucc->removePredecessor(BI->getParent());
1900 // Insert select instructions where needed.
1901 Value *BrCond = BI->getCondition();
1903 // Insert a select if the results differ.
1904 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1905 } else if (isa<UndefValue>(TrueValue)) {
1906 TrueValue = FalseValue;
1908 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1909 FalseValue, "retval");
1913 Value *RI = !TrueValue ?
1914 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
1918 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
1919 << "\n " << *BI << "NewRet = " << *RI
1920 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
1922 EraseTerminatorInstAndDCECond(BI);
1927 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
1928 /// probabilities of the branch taking each edge. Fills in the two APInt
1929 /// parameters and return true, or returns false if no or invalid metadata was
1931 static bool ExtractBranchMetadata(BranchInst *BI,
1932 uint64_t &ProbTrue, uint64_t &ProbFalse) {
1933 assert(BI->isConditional() &&
1934 "Looking for probabilities on unconditional branch?");
1935 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
1936 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
1937 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
1938 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
1939 if (!CITrue || !CIFalse) return false;
1940 ProbTrue = CITrue->getValue().getZExtValue();
1941 ProbFalse = CIFalse->getValue().getZExtValue();
1945 /// checkCSEInPredecessor - Return true if the given instruction is available
1946 /// in its predecessor block. If yes, the instruction will be removed.
1948 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
1949 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
1951 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
1952 Instruction *PBI = &*I;
1953 // Check whether Inst and PBI generate the same value.
1954 if (Inst->isIdenticalTo(PBI)) {
1955 Inst->replaceAllUsesWith(PBI);
1956 Inst->eraseFromParent();
1963 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
1964 /// predecessor branches to us and one of our successors, fold the block into
1965 /// the predecessor and use logical operations to pick the right destination.
1966 bool llvm::FoldBranchToCommonDest(BranchInst *BI, const DataLayout *DL) {
1967 BasicBlock *BB = BI->getParent();
1969 Instruction *Cond = nullptr;
1970 if (BI->isConditional())
1971 Cond = dyn_cast<Instruction>(BI->getCondition());
1973 // For unconditional branch, check for a simple CFG pattern, where
1974 // BB has a single predecessor and BB's successor is also its predecessor's
1975 // successor. If such pattern exisits, check for CSE between BB and its
1977 if (BasicBlock *PB = BB->getSinglePredecessor())
1978 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
1979 if (PBI->isConditional() &&
1980 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
1981 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
1982 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
1984 Instruction *Curr = I++;
1985 if (isa<CmpInst>(Curr)) {
1989 // Quit if we can't remove this instruction.
1990 if (!checkCSEInPredecessor(Curr, PB))
1999 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2000 Cond->getParent() != BB || !Cond->hasOneUse())
2003 // Only allow this if the condition is a simple instruction that can be
2004 // executed unconditionally. It must be in the same block as the branch, and
2005 // must be at the front of the block.
2006 BasicBlock::iterator FrontIt = BB->front();
2008 // Ignore dbg intrinsics.
2009 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
2011 // Allow a single instruction to be hoisted in addition to the compare
2012 // that feeds the branch. We later ensure that any values that _it_ uses
2013 // were also live in the predecessor, so that we don't unnecessarily create
2014 // register pressure or inhibit out-of-order execution.
2015 Instruction *BonusInst = nullptr;
2016 if (&*FrontIt != Cond &&
2017 FrontIt->hasOneUse() && FrontIt->user_back() == Cond &&
2018 isSafeToSpeculativelyExecute(FrontIt, DL)) {
2019 BonusInst = &*FrontIt;
2022 // Ignore dbg intrinsics.
2023 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
2026 // Only a single bonus inst is allowed.
2027 if (&*FrontIt != Cond)
2030 // Make sure the instruction after the condition is the cond branch.
2031 BasicBlock::iterator CondIt = Cond; ++CondIt;
2033 // Ignore dbg intrinsics.
2034 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
2039 // Cond is known to be a compare or binary operator. Check to make sure that
2040 // neither operand is a potentially-trapping constant expression.
2041 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2044 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2048 // Finally, don't infinitely unroll conditional loops.
2049 BasicBlock *TrueDest = BI->getSuccessor(0);
2050 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2051 if (TrueDest == BB || FalseDest == BB)
2054 for (BasicBlock *PredBlock : predecessors(BB)) {
2055 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2057 // Check that we have two conditional branches. If there is a PHI node in
2058 // the common successor, verify that the same value flows in from both
2060 SmallVector<PHINode*, 4> PHIs;
2061 if (!PBI || PBI->isUnconditional() ||
2062 (BI->isConditional() &&
2063 !SafeToMergeTerminators(BI, PBI)) ||
2064 (!BI->isConditional() &&
2065 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2068 // Determine if the two branches share a common destination.
2069 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2070 bool InvertPredCond = false;
2072 if (BI->isConditional()) {
2073 if (PBI->getSuccessor(0) == TrueDest)
2074 Opc = Instruction::Or;
2075 else if (PBI->getSuccessor(1) == FalseDest)
2076 Opc = Instruction::And;
2077 else if (PBI->getSuccessor(0) == FalseDest)
2078 Opc = Instruction::And, InvertPredCond = true;
2079 else if (PBI->getSuccessor(1) == TrueDest)
2080 Opc = Instruction::Or, InvertPredCond = true;
2084 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2088 // Ensure that any values used in the bonus instruction are also used
2089 // by the terminator of the predecessor. This means that those values
2090 // must already have been resolved, so we won't be inhibiting the
2091 // out-of-order core by speculating them earlier. We also allow
2092 // instructions that are used by the terminator's condition because it
2093 // exposes more merging opportunities.
2094 bool UsedByBranch = (BonusInst && BonusInst->hasOneUse() &&
2095 BonusInst->user_back() == Cond);
2097 if (BonusInst && !UsedByBranch) {
2098 // Collect the values used by the bonus inst
2099 SmallPtrSet<Value*, 4> UsedValues;
2100 for (Instruction::op_iterator OI = BonusInst->op_begin(),
2101 OE = BonusInst->op_end(); OI != OE; ++OI) {
2103 if (!isa<Constant>(V) && !isa<Argument>(V))
2104 UsedValues.insert(V);
2107 SmallVector<std::pair<Value*, unsigned>, 4> Worklist;
2108 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0));
2110 // Walk up to four levels back up the use-def chain of the predecessor's
2111 // terminator to see if all those values were used. The choice of four
2112 // levels is arbitrary, to provide a compile-time-cost bound.
2113 while (!Worklist.empty()) {
2114 std::pair<Value*, unsigned> Pair = Worklist.back();
2115 Worklist.pop_back();
2117 if (Pair.second >= 4) continue;
2118 UsedValues.erase(Pair.first);
2119 if (UsedValues.empty()) break;
2121 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) {
2122 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
2124 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1));
2128 if (!UsedValues.empty()) return false;
2131 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2132 IRBuilder<> Builder(PBI);
2134 // If we need to invert the condition in the pred block to match, do so now.
2135 if (InvertPredCond) {
2136 Value *NewCond = PBI->getCondition();
2138 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2139 CmpInst *CI = cast<CmpInst>(NewCond);
2140 CI->setPredicate(CI->getInversePredicate());
2142 NewCond = Builder.CreateNot(NewCond,
2143 PBI->getCondition()->getName()+".not");
2146 PBI->setCondition(NewCond);
2147 PBI->swapSuccessors();
2150 // If we have a bonus inst, clone it into the predecessor block.
2151 Instruction *NewBonus = nullptr;
2153 NewBonus = BonusInst->clone();
2155 // If we moved a load, we cannot any longer claim any knowledge about
2156 // its potential value. The previous information might have been valid
2157 // only given the branch precondition.
2158 // For an analogous reason, we must also drop all the metadata whose
2159 // semantics we don't understand.
2160 NewBonus->dropUnknownMetadata(LLVMContext::MD_dbg);
2162 PredBlock->getInstList().insert(PBI, NewBonus);
2163 NewBonus->takeName(BonusInst);
2164 BonusInst->setName(BonusInst->getName()+".old");
2167 // Clone Cond into the predecessor basic block, and or/and the
2168 // two conditions together.
2169 Instruction *New = Cond->clone();
2170 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus);
2171 PredBlock->getInstList().insert(PBI, New);
2172 New->takeName(Cond);
2173 Cond->setName(New->getName()+".old");
2175 if (BI->isConditional()) {
2176 Instruction *NewCond =
2177 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2179 PBI->setCondition(NewCond);
2181 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2182 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2184 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2186 SmallVector<uint64_t, 8> NewWeights;
2188 if (PBI->getSuccessor(0) == BB) {
2189 if (PredHasWeights && SuccHasWeights) {
2190 // PBI: br i1 %x, BB, FalseDest
2191 // BI: br i1 %y, TrueDest, FalseDest
2192 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2193 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2194 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2195 // TrueWeight for PBI * FalseWeight for BI.
2196 // We assume that total weights of a BranchInst can fit into 32 bits.
2197 // Therefore, we will not have overflow using 64-bit arithmetic.
2198 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2199 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2201 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2202 PBI->setSuccessor(0, TrueDest);
2204 if (PBI->getSuccessor(1) == BB) {
2205 if (PredHasWeights && SuccHasWeights) {
2206 // PBI: br i1 %x, TrueDest, BB
2207 // BI: br i1 %y, TrueDest, FalseDest
2208 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2209 // FalseWeight for PBI * TrueWeight for BI.
2210 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2211 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2212 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2213 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2215 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2216 PBI->setSuccessor(1, FalseDest);
2218 if (NewWeights.size() == 2) {
2219 // Halve the weights if any of them cannot fit in an uint32_t
2220 FitWeights(NewWeights);
2222 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2223 PBI->setMetadata(LLVMContext::MD_prof,
2224 MDBuilder(BI->getContext()).
2225 createBranchWeights(MDWeights));
2227 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2229 // Update PHI nodes in the common successors.
2230 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2231 ConstantInt *PBI_C = cast<ConstantInt>(
2232 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2233 assert(PBI_C->getType()->isIntegerTy(1));
2234 Instruction *MergedCond = nullptr;
2235 if (PBI->getSuccessor(0) == TrueDest) {
2236 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2237 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2238 // is false: !PBI_Cond and BI_Value
2239 Instruction *NotCond =
2240 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2243 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2248 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2249 PBI->getCondition(), MergedCond,
2252 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2253 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2254 // is false: PBI_Cond and BI_Value
2256 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2257 PBI->getCondition(), New,
2259 if (PBI_C->isOne()) {
2260 Instruction *NotCond =
2261 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2264 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2265 NotCond, MergedCond,
2270 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2273 // Change PBI from Conditional to Unconditional.
2274 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2275 EraseTerminatorInstAndDCECond(PBI);
2279 // TODO: If BB is reachable from all paths through PredBlock, then we
2280 // could replace PBI's branch probabilities with BI's.
2282 // Copy any debug value intrinsics into the end of PredBlock.
2283 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2284 if (isa<DbgInfoIntrinsic>(*I))
2285 I->clone()->insertBefore(PBI);
2292 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2293 /// predecessor of another block, this function tries to simplify it. We know
2294 /// that PBI and BI are both conditional branches, and BI is in one of the
2295 /// successor blocks of PBI - PBI branches to BI.
2296 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2297 assert(PBI->isConditional() && BI->isConditional());
2298 BasicBlock *BB = BI->getParent();
2300 // If this block ends with a branch instruction, and if there is a
2301 // predecessor that ends on a branch of the same condition, make
2302 // this conditional branch redundant.
2303 if (PBI->getCondition() == BI->getCondition() &&
2304 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2305 // Okay, the outcome of this conditional branch is statically
2306 // knowable. If this block had a single pred, handle specially.
2307 if (BB->getSinglePredecessor()) {
2308 // Turn this into a branch on constant.
2309 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2310 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2312 return true; // Nuke the branch on constant.
2315 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2316 // in the constant and simplify the block result. Subsequent passes of
2317 // simplifycfg will thread the block.
2318 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2319 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2320 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2321 std::distance(PB, PE),
2322 BI->getCondition()->getName() + ".pr",
2324 // Okay, we're going to insert the PHI node. Since PBI is not the only
2325 // predecessor, compute the PHI'd conditional value for all of the preds.
2326 // Any predecessor where the condition is not computable we keep symbolic.
2327 for (pred_iterator PI = PB; PI != PE; ++PI) {
2328 BasicBlock *P = *PI;
2329 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2330 PBI != BI && PBI->isConditional() &&
2331 PBI->getCondition() == BI->getCondition() &&
2332 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2333 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2334 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2337 NewPN->addIncoming(BI->getCondition(), P);
2341 BI->setCondition(NewPN);
2346 // If this is a conditional branch in an empty block, and if any
2347 // predecessors are a conditional branch to one of our destinations,
2348 // fold the conditions into logical ops and one cond br.
2349 BasicBlock::iterator BBI = BB->begin();
2350 // Ignore dbg intrinsics.
2351 while (isa<DbgInfoIntrinsic>(BBI))
2357 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2362 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2364 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2365 PBIOp = 0, BIOp = 1;
2366 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2367 PBIOp = 1, BIOp = 0;
2368 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2373 // Check to make sure that the other destination of this branch
2374 // isn't BB itself. If so, this is an infinite loop that will
2375 // keep getting unwound.
2376 if (PBI->getSuccessor(PBIOp) == BB)
2379 // Do not perform this transformation if it would require
2380 // insertion of a large number of select instructions. For targets
2381 // without predication/cmovs, this is a big pessimization.
2383 // Also do not perform this transformation if any phi node in the common
2384 // destination block can trap when reached by BB or PBB (PR17073). In that
2385 // case, it would be unsafe to hoist the operation into a select instruction.
2387 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2388 unsigned NumPhis = 0;
2389 for (BasicBlock::iterator II = CommonDest->begin();
2390 isa<PHINode>(II); ++II, ++NumPhis) {
2391 if (NumPhis > 2) // Disable this xform.
2394 PHINode *PN = cast<PHINode>(II);
2395 Value *BIV = PN->getIncomingValueForBlock(BB);
2396 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
2400 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2401 Value *PBIV = PN->getIncomingValue(PBBIdx);
2402 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
2407 // Finally, if everything is ok, fold the branches to logical ops.
2408 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2410 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2411 << "AND: " << *BI->getParent());
2414 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2415 // branch in it, where one edge (OtherDest) goes back to itself but the other
2416 // exits. We don't *know* that the program avoids the infinite loop
2417 // (even though that seems likely). If we do this xform naively, we'll end up
2418 // recursively unpeeling the loop. Since we know that (after the xform is
2419 // done) that the block *is* infinite if reached, we just make it an obviously
2420 // infinite loop with no cond branch.
2421 if (OtherDest == BB) {
2422 // Insert it at the end of the function, because it's either code,
2423 // or it won't matter if it's hot. :)
2424 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2425 "infloop", BB->getParent());
2426 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2427 OtherDest = InfLoopBlock;
2430 DEBUG(dbgs() << *PBI->getParent()->getParent());
2432 // BI may have other predecessors. Because of this, we leave
2433 // it alone, but modify PBI.
2435 // Make sure we get to CommonDest on True&True directions.
2436 Value *PBICond = PBI->getCondition();
2437 IRBuilder<true, NoFolder> Builder(PBI);
2439 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2441 Value *BICond = BI->getCondition();
2443 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2445 // Merge the conditions.
2446 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2448 // Modify PBI to branch on the new condition to the new dests.
2449 PBI->setCondition(Cond);
2450 PBI->setSuccessor(0, CommonDest);
2451 PBI->setSuccessor(1, OtherDest);
2453 // Update branch weight for PBI.
2454 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2455 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2457 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2459 if (PredHasWeights && SuccHasWeights) {
2460 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2461 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2462 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2463 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2464 // The weight to CommonDest should be PredCommon * SuccTotal +
2465 // PredOther * SuccCommon.
2466 // The weight to OtherDest should be PredOther * SuccOther.
2467 SmallVector<uint64_t, 2> NewWeights;
2468 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) +
2469 PredOther * SuccCommon);
2470 NewWeights.push_back(PredOther * SuccOther);
2471 // Halve the weights if any of them cannot fit in an uint32_t
2472 FitWeights(NewWeights);
2474 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end());
2475 PBI->setMetadata(LLVMContext::MD_prof,
2476 MDBuilder(BI->getContext()).
2477 createBranchWeights(MDWeights));
2480 // OtherDest may have phi nodes. If so, add an entry from PBI's
2481 // block that are identical to the entries for BI's block.
2482 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2484 // We know that the CommonDest already had an edge from PBI to
2485 // it. If it has PHIs though, the PHIs may have different
2486 // entries for BB and PBI's BB. If so, insert a select to make
2489 for (BasicBlock::iterator II = CommonDest->begin();
2490 (PN = dyn_cast<PHINode>(II)); ++II) {
2491 Value *BIV = PN->getIncomingValueForBlock(BB);
2492 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2493 Value *PBIV = PN->getIncomingValue(PBBIdx);
2495 // Insert a select in PBI to pick the right value.
2496 Value *NV = cast<SelectInst>
2497 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2498 PN->setIncomingValue(PBBIdx, NV);
2502 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2503 DEBUG(dbgs() << *PBI->getParent()->getParent());
2505 // This basic block is probably dead. We know it has at least
2506 // one fewer predecessor.
2510 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2511 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2512 // Takes care of updating the successors and removing the old terminator.
2513 // Also makes sure not to introduce new successors by assuming that edges to
2514 // non-successor TrueBBs and FalseBBs aren't reachable.
2515 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2516 BasicBlock *TrueBB, BasicBlock *FalseBB,
2517 uint32_t TrueWeight,
2518 uint32_t FalseWeight){
2519 // Remove any superfluous successor edges from the CFG.
2520 // First, figure out which successors to preserve.
2521 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2523 BasicBlock *KeepEdge1 = TrueBB;
2524 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
2526 // Then remove the rest.
2527 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2528 BasicBlock *Succ = OldTerm->getSuccessor(I);
2529 // Make sure only to keep exactly one copy of each edge.
2530 if (Succ == KeepEdge1)
2531 KeepEdge1 = nullptr;
2532 else if (Succ == KeepEdge2)
2533 KeepEdge2 = nullptr;
2535 Succ->removePredecessor(OldTerm->getParent());
2538 IRBuilder<> Builder(OldTerm);
2539 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2541 // Insert an appropriate new terminator.
2542 if (!KeepEdge1 && !KeepEdge2) {
2543 if (TrueBB == FalseBB)
2544 // We were only looking for one successor, and it was present.
2545 // Create an unconditional branch to it.
2546 Builder.CreateBr(TrueBB);
2548 // We found both of the successors we were looking for.
2549 // Create a conditional branch sharing the condition of the select.
2550 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2551 if (TrueWeight != FalseWeight)
2552 NewBI->setMetadata(LLVMContext::MD_prof,
2553 MDBuilder(OldTerm->getContext()).
2554 createBranchWeights(TrueWeight, FalseWeight));
2556 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2557 // Neither of the selected blocks were successors, so this
2558 // terminator must be unreachable.
2559 new UnreachableInst(OldTerm->getContext(), OldTerm);
2561 // One of the selected values was a successor, but the other wasn't.
2562 // Insert an unconditional branch to the one that was found;
2563 // the edge to the one that wasn't must be unreachable.
2565 // Only TrueBB was found.
2566 Builder.CreateBr(TrueBB);
2568 // Only FalseBB was found.
2569 Builder.CreateBr(FalseBB);
2572 EraseTerminatorInstAndDCECond(OldTerm);
2576 // SimplifySwitchOnSelect - Replaces
2577 // (switch (select cond, X, Y)) on constant X, Y
2578 // with a branch - conditional if X and Y lead to distinct BBs,
2579 // unconditional otherwise.
2580 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2581 // Check for constant integer values in the select.
2582 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2583 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2584 if (!TrueVal || !FalseVal)
2587 // Find the relevant condition and destinations.
2588 Value *Condition = Select->getCondition();
2589 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2590 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2592 // Get weight for TrueBB and FalseBB.
2593 uint32_t TrueWeight = 0, FalseWeight = 0;
2594 SmallVector<uint64_t, 8> Weights;
2595 bool HasWeights = HasBranchWeights(SI);
2597 GetBranchWeights(SI, Weights);
2598 if (Weights.size() == 1 + SI->getNumCases()) {
2599 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2600 getSuccessorIndex()];
2601 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2602 getSuccessorIndex()];
2606 // Perform the actual simplification.
2607 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2608 TrueWeight, FalseWeight);
2611 // SimplifyIndirectBrOnSelect - Replaces
2612 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2613 // blockaddress(@fn, BlockB)))
2615 // (br cond, BlockA, BlockB).
2616 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2617 // Check that both operands of the select are block addresses.
2618 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2619 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2623 // Extract the actual blocks.
2624 BasicBlock *TrueBB = TBA->getBasicBlock();
2625 BasicBlock *FalseBB = FBA->getBasicBlock();
2627 // Perform the actual simplification.
2628 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2632 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2633 /// instruction (a seteq/setne with a constant) as the only instruction in a
2634 /// block that ends with an uncond branch. We are looking for a very specific
2635 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2636 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2637 /// default value goes to an uncond block with a seteq in it, we get something
2640 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2642 /// %tmp = icmp eq i8 %A, 92
2645 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2647 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2648 /// the PHI, merging the third icmp into the switch.
2649 static bool TryToSimplifyUncondBranchWithICmpInIt(
2650 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI,
2651 const DataLayout *DL) {
2652 BasicBlock *BB = ICI->getParent();
2654 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2656 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2658 Value *V = ICI->getOperand(0);
2659 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2661 // The pattern we're looking for is where our only predecessor is a switch on
2662 // 'V' and this block is the default case for the switch. In this case we can
2663 // fold the compared value into the switch to simplify things.
2664 BasicBlock *Pred = BB->getSinglePredecessor();
2665 if (!Pred || !isa<SwitchInst>(Pred->getTerminator())) return false;
2667 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2668 if (SI->getCondition() != V)
2671 // If BB is reachable on a non-default case, then we simply know the value of
2672 // V in this block. Substitute it and constant fold the icmp instruction
2674 if (SI->getDefaultDest() != BB) {
2675 ConstantInt *VVal = SI->findCaseDest(BB);
2676 assert(VVal && "Should have a unique destination value");
2677 ICI->setOperand(0, VVal);
2679 if (Value *V = SimplifyInstruction(ICI, DL)) {
2680 ICI->replaceAllUsesWith(V);
2681 ICI->eraseFromParent();
2683 // BB is now empty, so it is likely to simplify away.
2684 return SimplifyCFG(BB, TTI, DL) | true;
2687 // Ok, the block is reachable from the default dest. If the constant we're
2688 // comparing exists in one of the other edges, then we can constant fold ICI
2690 if (SI->findCaseValue(Cst) != SI->case_default()) {
2692 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2693 V = ConstantInt::getFalse(BB->getContext());
2695 V = ConstantInt::getTrue(BB->getContext());
2697 ICI->replaceAllUsesWith(V);
2698 ICI->eraseFromParent();
2699 // BB is now empty, so it is likely to simplify away.
2700 return SimplifyCFG(BB, TTI, DL) | true;
2703 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2705 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2706 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
2707 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
2708 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2711 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2713 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2714 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2716 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2717 std::swap(DefaultCst, NewCst);
2719 // Replace ICI (which is used by the PHI for the default value) with true or
2720 // false depending on if it is EQ or NE.
2721 ICI->replaceAllUsesWith(DefaultCst);
2722 ICI->eraseFromParent();
2724 // Okay, the switch goes to this block on a default value. Add an edge from
2725 // the switch to the merge point on the compared value.
2726 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2727 BB->getParent(), BB);
2728 SmallVector<uint64_t, 8> Weights;
2729 bool HasWeights = HasBranchWeights(SI);
2731 GetBranchWeights(SI, Weights);
2732 if (Weights.size() == 1 + SI->getNumCases()) {
2733 // Split weight for default case to case for "Cst".
2734 Weights[0] = (Weights[0]+1) >> 1;
2735 Weights.push_back(Weights[0]);
2737 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2738 SI->setMetadata(LLVMContext::MD_prof,
2739 MDBuilder(SI->getContext()).
2740 createBranchWeights(MDWeights));
2743 SI->addCase(Cst, NewBB);
2745 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2746 Builder.SetInsertPoint(NewBB);
2747 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2748 Builder.CreateBr(SuccBlock);
2749 PHIUse->addIncoming(NewCst, NewBB);
2753 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2754 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2755 /// fold it into a switch instruction if so.
2756 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *DL,
2757 IRBuilder<> &Builder) {
2758 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2759 if (!Cond) return false;
2762 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2763 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2764 // 'setne's and'ed together, collect them.
2765 Value *CompVal = nullptr;
2766 std::vector<ConstantInt*> Values;
2767 bool TrueWhenEqual = true;
2768 Value *ExtraCase = nullptr;
2769 unsigned UsedICmps = 0;
2771 if (Cond->getOpcode() == Instruction::Or) {
2772 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, DL, true,
2774 } else if (Cond->getOpcode() == Instruction::And) {
2775 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, DL, false,
2777 TrueWhenEqual = false;
2780 // If we didn't have a multiply compared value, fail.
2781 if (!CompVal) return false;
2783 // Avoid turning single icmps into a switch.
2787 // There might be duplicate constants in the list, which the switch
2788 // instruction can't handle, remove them now.
2789 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2790 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2792 // If Extra was used, we require at least two switch values to do the
2793 // transformation. A switch with one value is just an cond branch.
2794 if (ExtraCase && Values.size() < 2) return false;
2796 // TODO: Preserve branch weight metadata, similarly to how
2797 // FoldValueComparisonIntoPredecessors preserves it.
2799 // Figure out which block is which destination.
2800 BasicBlock *DefaultBB = BI->getSuccessor(1);
2801 BasicBlock *EdgeBB = BI->getSuccessor(0);
2802 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2804 BasicBlock *BB = BI->getParent();
2806 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2807 << " cases into SWITCH. BB is:\n" << *BB);
2809 // If there are any extra values that couldn't be folded into the switch
2810 // then we evaluate them with an explicit branch first. Split the block
2811 // right before the condbr to handle it.
2813 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2814 // Remove the uncond branch added to the old block.
2815 TerminatorInst *OldTI = BB->getTerminator();
2816 Builder.SetInsertPoint(OldTI);
2819 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2821 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2823 OldTI->eraseFromParent();
2825 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2826 // for the edge we just added.
2827 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2829 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2830 << "\nEXTRABB = " << *BB);
2834 Builder.SetInsertPoint(BI);
2835 // Convert pointer to int before we switch.
2836 if (CompVal->getType()->isPointerTy()) {
2837 assert(DL && "Cannot switch on pointer without DataLayout");
2838 CompVal = Builder.CreatePtrToInt(CompVal,
2839 DL->getIntPtrType(CompVal->getType()),
2843 // Create the new switch instruction now.
2844 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2846 // Add all of the 'cases' to the switch instruction.
2847 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2848 New->addCase(Values[i], EdgeBB);
2850 // We added edges from PI to the EdgeBB. As such, if there were any
2851 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2852 // the number of edges added.
2853 for (BasicBlock::iterator BBI = EdgeBB->begin();
2854 isa<PHINode>(BBI); ++BBI) {
2855 PHINode *PN = cast<PHINode>(BBI);
2856 Value *InVal = PN->getIncomingValueForBlock(BB);
2857 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2858 PN->addIncoming(InVal, BB);
2861 // Erase the old branch instruction.
2862 EraseTerminatorInstAndDCECond(BI);
2864 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2868 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2869 // If this is a trivial landing pad that just continues unwinding the caught
2870 // exception then zap the landing pad, turning its invokes into calls.
2871 BasicBlock *BB = RI->getParent();
2872 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2873 if (RI->getValue() != LPInst)
2874 // Not a landing pad, or the resume is not unwinding the exception that
2875 // caused control to branch here.
2878 // Check that there are no other instructions except for debug intrinsics.
2879 BasicBlock::iterator I = LPInst, E = RI;
2881 if (!isa<DbgInfoIntrinsic>(I))
2884 // Turn all invokes that unwind here into calls and delete the basic block.
2885 bool InvokeRequiresTableEntry = false;
2886 bool Changed = false;
2887 for (BasicBlock *Pred : predecessors(BB)) {
2888 InvokeInst *II = cast<InvokeInst>(Pred->getTerminator());
2890 if (II->hasFnAttr(Attribute::UWTable)) {
2891 // Don't remove an `invoke' instruction if the ABI requires an entry into
2893 InvokeRequiresTableEntry = true;
2897 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2899 // Insert a call instruction before the invoke.
2900 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2902 Call->setCallingConv(II->getCallingConv());
2903 Call->setAttributes(II->getAttributes());
2904 Call->setDebugLoc(II->getDebugLoc());
2906 // Anything that used the value produced by the invoke instruction now uses
2907 // the value produced by the call instruction. Note that we do this even
2908 // for void functions and calls with no uses so that the callgraph edge is
2910 II->replaceAllUsesWith(Call);
2911 BB->removePredecessor(II->getParent());
2913 // Insert a branch to the normal destination right before the invoke.
2914 BranchInst::Create(II->getNormalDest(), II);
2916 // Finally, delete the invoke instruction!
2917 II->eraseFromParent();
2921 if (!InvokeRequiresTableEntry)
2922 // The landingpad is now unreachable. Zap it.
2923 BB->eraseFromParent();
2928 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2929 BasicBlock *BB = RI->getParent();
2930 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2932 // Find predecessors that end with branches.
2933 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2934 SmallVector<BranchInst*, 8> CondBranchPreds;
2935 for (BasicBlock *P : predecessors(BB)) {
2936 TerminatorInst *PTI = P->getTerminator();
2937 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2938 if (BI->isUnconditional())
2939 UncondBranchPreds.push_back(P);
2941 CondBranchPreds.push_back(BI);
2945 // If we found some, do the transformation!
2946 if (!UncondBranchPreds.empty() && DupRet) {
2947 while (!UncondBranchPreds.empty()) {
2948 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
2949 DEBUG(dbgs() << "FOLDING: " << *BB
2950 << "INTO UNCOND BRANCH PRED: " << *Pred);
2951 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
2954 // If we eliminated all predecessors of the block, delete the block now.
2955 if (pred_begin(BB) == pred_end(BB))
2956 // We know there are no successors, so just nuke the block.
2957 BB->eraseFromParent();
2962 // Check out all of the conditional branches going to this return
2963 // instruction. If any of them just select between returns, change the
2964 // branch itself into a select/return pair.
2965 while (!CondBranchPreds.empty()) {
2966 BranchInst *BI = CondBranchPreds.pop_back_val();
2968 // Check to see if the non-BB successor is also a return block.
2969 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
2970 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
2971 SimplifyCondBranchToTwoReturns(BI, Builder))
2977 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
2978 BasicBlock *BB = UI->getParent();
2980 bool Changed = false;
2982 // If there are any instructions immediately before the unreachable that can
2983 // be removed, do so.
2984 while (UI != BB->begin()) {
2985 BasicBlock::iterator BBI = UI;
2987 // Do not delete instructions that can have side effects which might cause
2988 // the unreachable to not be reachable; specifically, calls and volatile
2989 // operations may have this effect.
2990 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
2992 if (BBI->mayHaveSideEffects()) {
2993 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
2994 if (SI->isVolatile())
2996 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
2997 if (LI->isVolatile())
2999 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
3000 if (RMWI->isVolatile())
3002 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
3003 if (CXI->isVolatile())
3005 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
3006 !isa<LandingPadInst>(BBI)) {
3009 // Note that deleting LandingPad's here is in fact okay, although it
3010 // involves a bit of subtle reasoning. If this inst is a LandingPad,
3011 // all the predecessors of this block will be the unwind edges of Invokes,
3012 // and we can therefore guarantee this block will be erased.
3015 // Delete this instruction (any uses are guaranteed to be dead)
3016 if (!BBI->use_empty())
3017 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
3018 BBI->eraseFromParent();
3022 // If the unreachable instruction is the first in the block, take a gander
3023 // at all of the predecessors of this instruction, and simplify them.
3024 if (&BB->front() != UI) return Changed;
3026 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
3027 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
3028 TerminatorInst *TI = Preds[i]->getTerminator();
3029 IRBuilder<> Builder(TI);
3030 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3031 if (BI->isUnconditional()) {
3032 if (BI->getSuccessor(0) == BB) {
3033 new UnreachableInst(TI->getContext(), TI);
3034 TI->eraseFromParent();
3038 if (BI->getSuccessor(0) == BB) {
3039 Builder.CreateBr(BI->getSuccessor(1));
3040 EraseTerminatorInstAndDCECond(BI);
3041 } else if (BI->getSuccessor(1) == BB) {
3042 Builder.CreateBr(BI->getSuccessor(0));
3043 EraseTerminatorInstAndDCECond(BI);
3047 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3048 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3050 if (i.getCaseSuccessor() == BB) {
3051 BB->removePredecessor(SI->getParent());
3056 // If the default value is unreachable, figure out the most popular
3057 // destination and make it the default.
3058 if (SI->getDefaultDest() == BB) {
3059 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
3060 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3062 std::pair<unsigned, unsigned> &entry =
3063 Popularity[i.getCaseSuccessor()];
3064 if (entry.first == 0) {
3066 entry.second = i.getCaseIndex();
3072 // Find the most popular block.
3073 unsigned MaxPop = 0;
3074 unsigned MaxIndex = 0;
3075 BasicBlock *MaxBlock = nullptr;
3076 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
3077 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
3078 if (I->second.first > MaxPop ||
3079 (I->second.first == MaxPop && MaxIndex > I->second.second)) {
3080 MaxPop = I->second.first;
3081 MaxIndex = I->second.second;
3082 MaxBlock = I->first;
3086 // Make this the new default, allowing us to delete any explicit
3088 SI->setDefaultDest(MaxBlock);
3091 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
3093 if (isa<PHINode>(MaxBlock->begin()))
3094 for (unsigned i = 0; i != MaxPop-1; ++i)
3095 MaxBlock->removePredecessor(SI->getParent());
3097 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3099 if (i.getCaseSuccessor() == MaxBlock) {
3105 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
3106 if (II->getUnwindDest() == BB) {
3107 // Convert the invoke to a call instruction. This would be a good
3108 // place to note that the call does not throw though.
3109 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
3110 II->removeFromParent(); // Take out of symbol table
3112 // Insert the call now...
3113 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
3114 Builder.SetInsertPoint(BI);
3115 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
3116 Args, II->getName());
3117 CI->setCallingConv(II->getCallingConv());
3118 CI->setAttributes(II->getAttributes());
3119 // If the invoke produced a value, the call does now instead.
3120 II->replaceAllUsesWith(CI);
3127 // If this block is now dead, remove it.
3128 if (pred_begin(BB) == pred_end(BB) &&
3129 BB != &BB->getParent()->getEntryBlock()) {
3130 // We know there are no successors, so just nuke the block.
3131 BB->eraseFromParent();
3138 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
3139 /// integer range comparison into a sub, an icmp and a branch.
3140 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3141 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3143 // Make sure all cases point to the same destination and gather the values.
3144 SmallVector<ConstantInt *, 16> Cases;
3145 SwitchInst::CaseIt I = SI->case_begin();
3146 Cases.push_back(I.getCaseValue());
3147 SwitchInst::CaseIt PrevI = I++;
3148 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
3149 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
3151 Cases.push_back(I.getCaseValue());
3153 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
3155 // Sort the case values, then check if they form a range we can transform.
3156 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3157 for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
3158 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
3162 Constant *Offset = ConstantExpr::getNeg(Cases.back());
3163 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
3165 Value *Sub = SI->getCondition();
3166 if (!Offset->isNullValue())
3167 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
3169 // If NumCases overflowed, then all possible values jump to the successor.
3170 if (NumCases->isNullValue() && SI->getNumCases() != 0)
3171 Cmp = ConstantInt::getTrue(SI->getContext());
3173 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3174 BranchInst *NewBI = Builder.CreateCondBr(
3175 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
3177 // Update weight for the newly-created conditional branch.
3178 SmallVector<uint64_t, 8> Weights;
3179 bool HasWeights = HasBranchWeights(SI);
3181 GetBranchWeights(SI, Weights);
3182 if (Weights.size() == 1 + SI->getNumCases()) {
3183 // Combine all weights for the cases to be the true weight of NewBI.
3184 // We assume that the sum of all weights for a Terminator can fit into 32
3186 uint32_t NewTrueWeight = 0;
3187 for (unsigned I = 1, E = Weights.size(); I != E; ++I)
3188 NewTrueWeight += (uint32_t)Weights[I];
3189 NewBI->setMetadata(LLVMContext::MD_prof,
3190 MDBuilder(SI->getContext()).
3191 createBranchWeights(NewTrueWeight,
3192 (uint32_t)Weights[0]));
3196 // Prune obsolete incoming values off the successor's PHI nodes.
3197 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
3198 isa<PHINode>(BBI); ++BBI) {
3199 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
3200 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3202 SI->eraseFromParent();
3207 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
3208 /// and use it to remove dead cases.
3209 static bool EliminateDeadSwitchCases(SwitchInst *SI) {
3210 Value *Cond = SI->getCondition();
3211 unsigned Bits = Cond->getType()->getIntegerBitWidth();
3212 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3213 computeKnownBits(Cond, KnownZero, KnownOne);
3215 // Gather dead cases.
3216 SmallVector<ConstantInt*, 8> DeadCases;
3217 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3218 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3219 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3220 DeadCases.push_back(I.getCaseValue());
3221 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3222 << I.getCaseValue() << "' is dead.\n");
3226 SmallVector<uint64_t, 8> Weights;
3227 bool HasWeight = HasBranchWeights(SI);
3229 GetBranchWeights(SI, Weights);
3230 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3233 // Remove dead cases from the switch.
3234 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3235 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3236 assert(Case != SI->case_default() &&
3237 "Case was not found. Probably mistake in DeadCases forming.");
3239 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3243 // Prune unused values from PHI nodes.
3244 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3245 SI->removeCase(Case);
3247 if (HasWeight && Weights.size() >= 2) {
3248 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3249 SI->setMetadata(LLVMContext::MD_prof,
3250 MDBuilder(SI->getParent()->getContext()).
3251 createBranchWeights(MDWeights));
3254 return !DeadCases.empty();
3257 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
3258 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3259 /// by an unconditional branch), look at the phi node for BB in the successor
3260 /// block and see if the incoming value is equal to CaseValue. If so, return
3261 /// the phi node, and set PhiIndex to BB's index in the phi node.
3262 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3265 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3266 return nullptr; // BB must be empty to be a candidate for simplification.
3267 if (!BB->getSinglePredecessor())
3268 return nullptr; // BB must be dominated by the switch.
3270 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3271 if (!Branch || !Branch->isUnconditional())
3272 return nullptr; // Terminator must be unconditional branch.
3274 BasicBlock *Succ = Branch->getSuccessor(0);
3276 BasicBlock::iterator I = Succ->begin();
3277 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3278 int Idx = PHI->getBasicBlockIndex(BB);
3279 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3281 Value *InValue = PHI->getIncomingValue(Idx);
3282 if (InValue != CaseValue) continue;
3291 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
3292 /// instruction to a phi node dominated by the switch, if that would mean that
3293 /// some of the destination blocks of the switch can be folded away.
3294 /// Returns true if a change is made.
3295 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3296 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3297 ForwardingNodesMap ForwardingNodes;
3299 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3300 ConstantInt *CaseValue = I.getCaseValue();
3301 BasicBlock *CaseDest = I.getCaseSuccessor();
3304 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3308 ForwardingNodes[PHI].push_back(PhiIndex);
3311 bool Changed = false;
3313 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3314 E = ForwardingNodes.end(); I != E; ++I) {
3315 PHINode *Phi = I->first;
3316 SmallVectorImpl<int> &Indexes = I->second;
3318 if (Indexes.size() < 2) continue;
3320 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3321 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3328 /// ValidLookupTableConstant - Return true if the backend will be able to handle
3329 /// initializing an array of constants like C.
3330 static bool ValidLookupTableConstant(Constant *C) {
3331 if (C->isThreadDependent())
3333 if (C->isDLLImportDependent())
3336 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3337 return CE->isGEPWithNoNotionalOverIndexing();
3339 return isa<ConstantFP>(C) ||
3340 isa<ConstantInt>(C) ||
3341 isa<ConstantPointerNull>(C) ||
3342 isa<GlobalValue>(C) ||
3346 /// LookupConstant - If V is a Constant, return it. Otherwise, try to look up
3347 /// its constant value in ConstantPool, returning 0 if it's not there.
3348 static Constant *LookupConstant(Value *V,
3349 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3350 if (Constant *C = dyn_cast<Constant>(V))
3352 return ConstantPool.lookup(V);
3355 /// ConstantFold - Try to fold instruction I into a constant. This works for
3356 /// simple instructions such as binary operations where both operands are
3357 /// constant or can be replaced by constants from the ConstantPool. Returns the
3358 /// resulting constant on success, 0 otherwise.
3360 ConstantFold(Instruction *I,
3361 const SmallDenseMap<Value *, Constant *> &ConstantPool,
3362 const DataLayout *DL) {
3363 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3364 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3367 if (A->isAllOnesValue())
3368 return LookupConstant(Select->getTrueValue(), ConstantPool);
3369 if (A->isNullValue())
3370 return LookupConstant(Select->getFalseValue(), ConstantPool);
3374 SmallVector<Constant *, 4> COps;
3375 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
3376 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
3382 if (CmpInst *Cmp = dyn_cast<CmpInst>(I))
3383 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
3386 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), COps, DL);
3389 /// GetCaseResults - Try to determine the resulting constant values in phi nodes
3390 /// at the common destination basic block, *CommonDest, for one of the case
3391 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3392 /// case), of a switch instruction SI.
3394 GetCaseResults(SwitchInst *SI,
3395 ConstantInt *CaseVal,
3396 BasicBlock *CaseDest,
3397 BasicBlock **CommonDest,
3398 SmallVectorImpl<std::pair<PHINode *, Constant *> > &Res,
3399 const DataLayout *DL) {
3400 // The block from which we enter the common destination.
3401 BasicBlock *Pred = SI->getParent();
3403 // If CaseDest is empty except for some side-effect free instructions through
3404 // which we can constant-propagate the CaseVal, continue to its successor.
3405 SmallDenseMap<Value*, Constant*> ConstantPool;
3406 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3407 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3409 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3410 // If the terminator is a simple branch, continue to the next block.
3411 if (T->getNumSuccessors() != 1)
3414 CaseDest = T->getSuccessor(0);
3415 } else if (isa<DbgInfoIntrinsic>(I)) {
3416 // Skip debug intrinsic.
3418 } else if (Constant *C = ConstantFold(I, ConstantPool, DL)) {
3419 // Instruction is side-effect free and constant.
3420 ConstantPool.insert(std::make_pair(I, C));
3426 // If we did not have a CommonDest before, use the current one.
3428 *CommonDest = CaseDest;
3429 // If the destination isn't the common one, abort.
3430 if (CaseDest != *CommonDest)
3433 // Get the values for this case from phi nodes in the destination block.
3434 BasicBlock::iterator I = (*CommonDest)->begin();
3435 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3436 int Idx = PHI->getBasicBlockIndex(Pred);
3440 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3445 // Note: If the constant comes from constant-propagating the case value
3446 // through the CaseDest basic block, it will be safe to remove the
3447 // instructions in that block. They cannot be used (except in the phi nodes
3448 // we visit) outside CaseDest, because that block does not dominate its
3449 // successor. If it did, we would not be in this phi node.
3451 // Be conservative about which kinds of constants we support.
3452 if (!ValidLookupTableConstant(ConstVal))
3455 Res.push_back(std::make_pair(PHI, ConstVal));
3458 return Res.size() > 0;
3462 /// SwitchLookupTable - This class represents a lookup table that can be used
3463 /// to replace a switch.
3464 class SwitchLookupTable {
3466 /// SwitchLookupTable - Create a lookup table to use as a switch replacement
3467 /// with the contents of Values, using DefaultValue to fill any holes in the
3469 SwitchLookupTable(Module &M,
3471 ConstantInt *Offset,
3472 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3473 Constant *DefaultValue,
3474 const DataLayout *DL);
3476 /// BuildLookup - Build instructions with Builder to retrieve the value at
3477 /// the position given by Index in the lookup table.
3478 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3480 /// WouldFitInRegister - Return true if a table with TableSize elements of
3481 /// type ElementType would fit in a target-legal register.
3482 static bool WouldFitInRegister(const DataLayout *DL,
3484 const Type *ElementType);
3487 // Depending on the contents of the table, it can be represented in
3490 // For tables where each element contains the same value, we just have to
3491 // store that single value and return it for each lookup.
3494 // For small tables with integer elements, we can pack them into a bitmap
3495 // that fits into a target-legal register. Values are retrieved by
3496 // shift and mask operations.
3499 // The table is stored as an array of values. Values are retrieved by load
3500 // instructions from the table.
3504 // For SingleValueKind, this is the single value.
3505 Constant *SingleValue;
3507 // For BitMapKind, this is the bitmap.
3508 ConstantInt *BitMap;
3509 IntegerType *BitMapElementTy;
3511 // For ArrayKind, this is the array.
3512 GlobalVariable *Array;
3516 SwitchLookupTable::SwitchLookupTable(Module &M,
3518 ConstantInt *Offset,
3519 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3520 Constant *DefaultValue,
3521 const DataLayout *DL)
3522 : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
3524 assert(Values.size() && "Can't build lookup table without values!");
3525 assert(TableSize >= Values.size() && "Can't fit values in table!");
3527 // If all values in the table are equal, this is that value.
3528 SingleValue = Values.begin()->second;
3530 Type *ValueType = Values.begin()->second->getType();
3532 // Build up the table contents.
3533 SmallVector<Constant*, 64> TableContents(TableSize);
3534 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3535 ConstantInt *CaseVal = Values[I].first;
3536 Constant *CaseRes = Values[I].second;
3537 assert(CaseRes->getType() == ValueType);
3539 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3541 TableContents[Idx] = CaseRes;
3543 if (CaseRes != SingleValue)
3544 SingleValue = nullptr;
3547 // Fill in any holes in the table with the default result.
3548 if (Values.size() < TableSize) {
3549 assert(DefaultValue &&
3550 "Need a default value to fill the lookup table holes.");
3551 assert(DefaultValue->getType() == ValueType);
3552 for (uint64_t I = 0; I < TableSize; ++I) {
3553 if (!TableContents[I])
3554 TableContents[I] = DefaultValue;
3557 if (DefaultValue != SingleValue)
3558 SingleValue = nullptr;
3561 // If each element in the table contains the same value, we only need to store
3562 // that single value.
3564 Kind = SingleValueKind;
3568 // If the type is integer and the table fits in a register, build a bitmap.
3569 if (WouldFitInRegister(DL, TableSize, ValueType)) {
3570 IntegerType *IT = cast<IntegerType>(ValueType);
3571 APInt TableInt(TableSize * IT->getBitWidth(), 0);
3572 for (uint64_t I = TableSize; I > 0; --I) {
3573 TableInt <<= IT->getBitWidth();
3574 // Insert values into the bitmap. Undef values are set to zero.
3575 if (!isa<UndefValue>(TableContents[I - 1])) {
3576 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
3577 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
3580 BitMap = ConstantInt::get(M.getContext(), TableInt);
3581 BitMapElementTy = IT;
3587 // Store the table in an array.
3588 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
3589 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3591 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3592 GlobalVariable::PrivateLinkage,
3595 Array->setUnnamedAddr(true);
3599 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
3601 case SingleValueKind:
3604 // Type of the bitmap (e.g. i59).
3605 IntegerType *MapTy = BitMap->getType();
3607 // Cast Index to the same type as the bitmap.
3608 // Note: The Index is <= the number of elements in the table, so
3609 // truncating it to the width of the bitmask is safe.
3610 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
3612 // Multiply the shift amount by the element width.
3613 ShiftAmt = Builder.CreateMul(ShiftAmt,
3614 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
3618 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
3619 "switch.downshift");
3621 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
3625 Value *GEPIndices[] = { Builder.getInt32(0), Index };
3626 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices,
3628 return Builder.CreateLoad(GEP, "switch.load");
3631 llvm_unreachable("Unknown lookup table kind!");
3634 bool SwitchLookupTable::WouldFitInRegister(const DataLayout *DL,
3636 const Type *ElementType) {
3639 const IntegerType *IT = dyn_cast<IntegerType>(ElementType);
3642 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
3643 // are <= 15, we could try to narrow the type.
3645 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
3646 if (TableSize >= UINT_MAX/IT->getBitWidth())
3648 return DL->fitsInLegalInteger(TableSize * IT->getBitWidth());
3651 /// ShouldBuildLookupTable - Determine whether a lookup table should be built
3652 /// for this switch, based on the number of cases, size of the table and the
3653 /// types of the results.
3654 static bool ShouldBuildLookupTable(SwitchInst *SI,
3656 const TargetTransformInfo &TTI,
3657 const DataLayout *DL,
3658 const SmallDenseMap<PHINode*, Type*>& ResultTypes) {
3659 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
3660 return false; // TableSize overflowed, or mul below might overflow.
3662 bool AllTablesFitInRegister = true;
3663 bool HasIllegalType = false;
3664 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(),
3665 E = ResultTypes.end(); I != E; ++I) {
3666 Type *Ty = I->second;
3668 // Saturate this flag to true.
3669 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
3671 // Saturate this flag to false.
3672 AllTablesFitInRegister = AllTablesFitInRegister &&
3673 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
3675 // If both flags saturate, we're done. NOTE: This *only* works with
3676 // saturating flags, and all flags have to saturate first due to the
3677 // non-deterministic behavior of iterating over a dense map.
3678 if (HasIllegalType && !AllTablesFitInRegister)
3682 // If each table would fit in a register, we should build it anyway.
3683 if (AllTablesFitInRegister)
3686 // Don't build a table that doesn't fit in-register if it has illegal types.
3690 // The table density should be at least 40%. This is the same criterion as for
3691 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3692 // FIXME: Find the best cut-off.
3693 return SI->getNumCases() * 10 >= TableSize * 4;
3696 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3697 /// phi nodes in a common successor block with different constant values,
3698 /// replace the switch with lookup tables.
3699 static bool SwitchToLookupTable(SwitchInst *SI,
3700 IRBuilder<> &Builder,
3701 const TargetTransformInfo &TTI,
3702 const DataLayout* DL) {
3703 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3705 // Only build lookup table when we have a target that supports it.
3706 if (!TTI.shouldBuildLookupTables())
3709 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3710 // split off a dense part and build a lookup table for that.
3712 // FIXME: This creates arrays of GEPs to constant strings, which means each
3713 // GEP needs a runtime relocation in PIC code. We should just build one big
3714 // string and lookup indices into that.
3716 // Ignore switches with less than three cases. Lookup tables will not make them
3717 // faster, so we don't analyze them.
3718 if (SI->getNumCases() < 3)
3721 // Figure out the corresponding result for each case value and phi node in the
3722 // common destination, as well as the the min and max case values.
3723 assert(SI->case_begin() != SI->case_end());
3724 SwitchInst::CaseIt CI = SI->case_begin();
3725 ConstantInt *MinCaseVal = CI.getCaseValue();
3726 ConstantInt *MaxCaseVal = CI.getCaseValue();
3728 BasicBlock *CommonDest = nullptr;
3729 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
3730 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
3731 SmallDenseMap<PHINode*, Constant*> DefaultResults;
3732 SmallDenseMap<PHINode*, Type*> ResultTypes;
3733 SmallVector<PHINode*, 4> PHIs;
3735 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
3736 ConstantInt *CaseVal = CI.getCaseValue();
3737 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
3738 MinCaseVal = CaseVal;
3739 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
3740 MaxCaseVal = CaseVal;
3742 // Resulting value at phi nodes for this case value.
3743 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
3745 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
3749 // Append the result from this case to the list for each phi.
3750 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) {
3751 if (!ResultLists.count(I->first))
3752 PHIs.push_back(I->first);
3753 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second));
3757 // Keep track of the result types.
3758 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
3759 PHINode *PHI = PHIs[I];
3760 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
3763 uint64_t NumResults = ResultLists[PHIs[0]].size();
3764 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
3765 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
3766 bool TableHasHoles = (NumResults < TableSize);
3768 // If the table has holes, we need a constant result for the default case
3769 // or a bitmask that fits in a register.
3770 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
3771 bool HasDefaultResults = false;
3772 if (TableHasHoles) {
3773 HasDefaultResults = GetCaseResults(SI, nullptr, SI->getDefaultDest(),
3774 &CommonDest, DefaultResultsList, DL);
3776 bool NeedMask = (TableHasHoles && !HasDefaultResults);
3778 // As an extra penalty for the validity test we require more cases.
3779 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
3781 if (!(DL && DL->fitsInLegalInteger(TableSize)))
3785 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) {
3786 PHINode *PHI = DefaultResultsList[I].first;
3787 Constant *Result = DefaultResultsList[I].second;
3788 DefaultResults[PHI] = Result;
3791 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
3794 // Create the BB that does the lookups.
3795 Module &Mod = *CommonDest->getParent()->getParent();
3796 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
3798 CommonDest->getParent(),
3801 // Compute the table index value.
3802 Builder.SetInsertPoint(SI);
3803 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
3806 // Compute the maximum table size representable by the integer type we are
3808 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
3809 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
3810 assert(MaxTableSize >= TableSize &&
3811 "It is impossible for a switch to have more entries than the max "
3812 "representable value of its input integer type's size.");
3814 // If we have a fully covered lookup table, unconditionally branch to the
3815 // lookup table BB. Otherwise, check if the condition value is within the case
3816 // range. If it is so, branch to the new BB. Otherwise branch to SI's default
3818 const bool GeneratingCoveredLookupTable = MaxTableSize == TableSize;
3819 if (GeneratingCoveredLookupTable) {
3820 Builder.CreateBr(LookupBB);
3821 SI->getDefaultDest()->removePredecessor(SI->getParent());
3823 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
3824 MinCaseVal->getType(), TableSize));
3825 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
3828 // Populate the BB that does the lookups.
3829 Builder.SetInsertPoint(LookupBB);
3832 // Before doing the lookup we do the hole check.
3833 // The LookupBB is therefore re-purposed to do the hole check
3834 // and we create a new LookupBB.
3835 BasicBlock *MaskBB = LookupBB;
3836 MaskBB->setName("switch.hole_check");
3837 LookupBB = BasicBlock::Create(Mod.getContext(),
3839 CommonDest->getParent(),
3842 // Build bitmask; fill in a 1 bit for every case.
3843 APInt MaskInt(TableSize, 0);
3844 APInt One(TableSize, 1);
3845 const ResultListTy &ResultList = ResultLists[PHIs[0]];
3846 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
3847 uint64_t Idx = (ResultList[I].first->getValue() -
3848 MinCaseVal->getValue()).getLimitedValue();
3849 MaskInt |= One << Idx;
3851 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
3853 // Get the TableIndex'th bit of the bitmask.
3854 // If this bit is 0 (meaning hole) jump to the default destination,
3855 // else continue with table lookup.
3856 IntegerType *MapTy = TableMask->getType();
3857 Value *MaskIndex = Builder.CreateZExtOrTrunc(TableIndex, MapTy,
3858 "switch.maskindex");
3859 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex,
3861 Value *LoBit = Builder.CreateTrunc(Shifted,
3862 Type::getInt1Ty(Mod.getContext()),
3864 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
3866 Builder.SetInsertPoint(LookupBB);
3867 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
3870 bool ReturnedEarly = false;
3871 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
3872 PHINode *PHI = PHIs[I];
3874 // If using a bitmask, use any value to fill the lookup table holes.
3875 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
3876 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI],
3879 Value *Result = Table.BuildLookup(TableIndex, Builder);
3881 // If the result is used to return immediately from the function, we want to
3882 // do that right here.
3883 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
3884 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
3885 Builder.CreateRet(Result);
3886 ReturnedEarly = true;
3890 PHI->addIncoming(Result, LookupBB);
3894 Builder.CreateBr(CommonDest);
3896 // Remove the switch.
3897 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
3898 BasicBlock *Succ = SI->getSuccessor(i);
3900 if (Succ == SI->getDefaultDest())
3902 Succ->removePredecessor(SI->getParent());
3904 SI->eraseFromParent();
3908 ++NumLookupTablesHoles;
3912 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
3913 BasicBlock *BB = SI->getParent();
3915 if (isValueEqualityComparison(SI)) {
3916 // If we only have one predecessor, and if it is a branch on this value,
3917 // see if that predecessor totally determines the outcome of this switch.
3918 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3919 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
3920 return SimplifyCFG(BB, TTI, DL) | true;
3922 Value *Cond = SI->getCondition();
3923 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
3924 if (SimplifySwitchOnSelect(SI, Select))
3925 return SimplifyCFG(BB, TTI, DL) | true;
3927 // If the block only contains the switch, see if we can fold the block
3928 // away into any preds.
3929 BasicBlock::iterator BBI = BB->begin();
3930 // Ignore dbg intrinsics.
3931 while (isa<DbgInfoIntrinsic>(BBI))
3934 if (FoldValueComparisonIntoPredecessors(SI, Builder))
3935 return SimplifyCFG(BB, TTI, DL) | true;
3938 // Try to transform the switch into an icmp and a branch.
3939 if (TurnSwitchRangeIntoICmp(SI, Builder))
3940 return SimplifyCFG(BB, TTI, DL) | true;
3942 // Remove unreachable cases.
3943 if (EliminateDeadSwitchCases(SI))
3944 return SimplifyCFG(BB, TTI, DL) | true;
3946 if (ForwardSwitchConditionToPHI(SI))
3947 return SimplifyCFG(BB, TTI, DL) | true;
3949 if (SwitchToLookupTable(SI, Builder, TTI, DL))
3950 return SimplifyCFG(BB, TTI, DL) | true;
3955 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
3956 BasicBlock *BB = IBI->getParent();
3957 bool Changed = false;
3959 // Eliminate redundant destinations.
3960 SmallPtrSet<Value *, 8> Succs;
3961 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
3962 BasicBlock *Dest = IBI->getDestination(i);
3963 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) {
3964 Dest->removePredecessor(BB);
3965 IBI->removeDestination(i);
3971 if (IBI->getNumDestinations() == 0) {
3972 // If the indirectbr has no successors, change it to unreachable.
3973 new UnreachableInst(IBI->getContext(), IBI);
3974 EraseTerminatorInstAndDCECond(IBI);
3978 if (IBI->getNumDestinations() == 1) {
3979 // If the indirectbr has one successor, change it to a direct branch.
3980 BranchInst::Create(IBI->getDestination(0), IBI);
3981 EraseTerminatorInstAndDCECond(IBI);
3985 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
3986 if (SimplifyIndirectBrOnSelect(IBI, SI))
3987 return SimplifyCFG(BB, TTI, DL) | true;
3992 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
3993 BasicBlock *BB = BI->getParent();
3995 if (SinkCommon && SinkThenElseCodeToEnd(BI))
3998 // If the Terminator is the only non-phi instruction, simplify the block.
3999 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime();
4000 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
4001 TryToSimplifyUncondBranchFromEmptyBlock(BB))
4004 // If the only instruction in the block is a seteq/setne comparison
4005 // against a constant, try to simplify the block.
4006 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
4007 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
4008 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
4010 if (I->isTerminator() &&
4011 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI, DL))
4015 // If this basic block is ONLY a compare and a branch, and if a predecessor
4016 // branches to us and our successor, fold the comparison into the
4017 // predecessor and use logical operations to update the incoming value
4018 // for PHI nodes in common successor.
4019 if (FoldBranchToCommonDest(BI, DL))
4020 return SimplifyCFG(BB, TTI, DL) | true;
4025 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
4026 BasicBlock *BB = BI->getParent();
4028 // Conditional branch
4029 if (isValueEqualityComparison(BI)) {
4030 // If we only have one predecessor, and if it is a branch on this value,
4031 // see if that predecessor totally determines the outcome of this
4033 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4034 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
4035 return SimplifyCFG(BB, TTI, DL) | true;
4037 // This block must be empty, except for the setcond inst, if it exists.
4038 // Ignore dbg intrinsics.
4039 BasicBlock::iterator I = BB->begin();
4040 // Ignore dbg intrinsics.
4041 while (isa<DbgInfoIntrinsic>(I))
4044 if (FoldValueComparisonIntoPredecessors(BI, Builder))
4045 return SimplifyCFG(BB, TTI, DL) | true;
4046 } else if (&*I == cast<Instruction>(BI->getCondition())){
4048 // Ignore dbg intrinsics.
4049 while (isa<DbgInfoIntrinsic>(I))
4051 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
4052 return SimplifyCFG(BB, TTI, DL) | true;
4056 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
4057 if (SimplifyBranchOnICmpChain(BI, DL, Builder))
4060 // If this basic block is ONLY a compare and a branch, and if a predecessor
4061 // branches to us and one of our successors, fold the comparison into the
4062 // predecessor and use logical operations to pick the right destination.
4063 if (FoldBranchToCommonDest(BI, DL))
4064 return SimplifyCFG(BB, TTI, DL) | true;
4066 // We have a conditional branch to two blocks that are only reachable
4067 // from BI. We know that the condbr dominates the two blocks, so see if
4068 // there is any identical code in the "then" and "else" blocks. If so, we
4069 // can hoist it up to the branching block.
4070 if (BI->getSuccessor(0)->getSinglePredecessor()) {
4071 if (BI->getSuccessor(1)->getSinglePredecessor()) {
4072 if (HoistThenElseCodeToIf(BI, DL))
4073 return SimplifyCFG(BB, TTI, DL) | true;
4075 // If Successor #1 has multiple preds, we may be able to conditionally
4076 // execute Successor #0 if it branches to Successor #1.
4077 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
4078 if (Succ0TI->getNumSuccessors() == 1 &&
4079 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
4080 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), DL))
4081 return SimplifyCFG(BB, TTI, DL) | true;
4083 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
4084 // If Successor #0 has multiple preds, we may be able to conditionally
4085 // execute Successor #1 if it branches to Successor #0.
4086 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
4087 if (Succ1TI->getNumSuccessors() == 1 &&
4088 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
4089 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), DL))
4090 return SimplifyCFG(BB, TTI, DL) | true;
4093 // If this is a branch on a phi node in the current block, thread control
4094 // through this block if any PHI node entries are constants.
4095 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
4096 if (PN->getParent() == BI->getParent())
4097 if (FoldCondBranchOnPHI(BI, DL))
4098 return SimplifyCFG(BB, TTI, DL) | true;
4100 // Scan predecessor blocks for conditional branches.
4101 for (BasicBlock *Pred : predecessors(BB))
4102 if (BranchInst *PBI = dyn_cast<BranchInst>(Pred->getTerminator()))
4103 if (PBI != BI && PBI->isConditional())
4104 if (SimplifyCondBranchToCondBranch(PBI, BI))
4105 return SimplifyCFG(BB, TTI, DL) | true;
4110 /// Check if passing a value to an instruction will cause undefined behavior.
4111 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
4112 Constant *C = dyn_cast<Constant>(V);
4119 if (C->isNullValue()) {
4120 // Only look at the first use, avoid hurting compile time with long uselists
4121 User *Use = *I->user_begin();
4123 // Now make sure that there are no instructions in between that can alter
4124 // control flow (eg. calls)
4125 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
4126 if (i == I->getParent()->end() || i->mayHaveSideEffects())
4129 // Look through GEPs. A load from a GEP derived from NULL is still undefined
4130 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
4131 if (GEP->getPointerOperand() == I)
4132 return passingValueIsAlwaysUndefined(V, GEP);
4134 // Look through bitcasts.
4135 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
4136 return passingValueIsAlwaysUndefined(V, BC);
4138 // Load from null is undefined.
4139 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
4140 if (!LI->isVolatile())
4141 return LI->getPointerAddressSpace() == 0;
4143 // Store to null is undefined.
4144 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
4145 if (!SI->isVolatile())
4146 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
4151 /// If BB has an incoming value that will always trigger undefined behavior
4152 /// (eg. null pointer dereference), remove the branch leading here.
4153 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
4154 for (BasicBlock::iterator i = BB->begin();
4155 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
4156 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
4157 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
4158 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
4159 IRBuilder<> Builder(T);
4160 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
4161 BB->removePredecessor(PHI->getIncomingBlock(i));
4162 // Turn uncoditional branches into unreachables and remove the dead
4163 // destination from conditional branches.
4164 if (BI->isUnconditional())
4165 Builder.CreateUnreachable();
4167 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
4168 BI->getSuccessor(0));
4169 BI->eraseFromParent();
4172 // TODO: SwitchInst.
4178 bool SimplifyCFGOpt::run(BasicBlock *BB) {
4179 bool Changed = false;
4181 assert(BB && BB->getParent() && "Block not embedded in function!");
4182 assert(BB->getTerminator() && "Degenerate basic block encountered!");
4184 // Remove basic blocks that have no predecessors (except the entry block)...
4185 // or that just have themself as a predecessor. These are unreachable.
4186 if ((pred_begin(BB) == pred_end(BB) &&
4187 BB != &BB->getParent()->getEntryBlock()) ||
4188 BB->getSinglePredecessor() == BB) {
4189 DEBUG(dbgs() << "Removing BB: \n" << *BB);
4190 DeleteDeadBlock(BB);
4194 // Check to see if we can constant propagate this terminator instruction
4196 Changed |= ConstantFoldTerminator(BB, true);
4198 // Check for and eliminate duplicate PHI nodes in this block.
4199 Changed |= EliminateDuplicatePHINodes(BB);
4201 // Check for and remove branches that will always cause undefined behavior.
4202 Changed |= removeUndefIntroducingPredecessor(BB);
4204 // Merge basic blocks into their predecessor if there is only one distinct
4205 // pred, and if there is only one distinct successor of the predecessor, and
4206 // if there are no PHI nodes.
4208 if (MergeBlockIntoPredecessor(BB))
4211 IRBuilder<> Builder(BB);
4213 // If there is a trivial two-entry PHI node in this basic block, and we can
4214 // eliminate it, do so now.
4215 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
4216 if (PN->getNumIncomingValues() == 2)
4217 Changed |= FoldTwoEntryPHINode(PN, DL);
4219 Builder.SetInsertPoint(BB->getTerminator());
4220 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
4221 if (BI->isUnconditional()) {
4222 if (SimplifyUncondBranch(BI, Builder)) return true;
4224 if (SimplifyCondBranch(BI, Builder)) return true;
4226 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
4227 if (SimplifyReturn(RI, Builder)) return true;
4228 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
4229 if (SimplifyResume(RI, Builder)) return true;
4230 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
4231 if (SimplifySwitch(SI, Builder)) return true;
4232 } else if (UnreachableInst *UI =
4233 dyn_cast<UnreachableInst>(BB->getTerminator())) {
4234 if (SimplifyUnreachable(UI)) return true;
4235 } else if (IndirectBrInst *IBI =
4236 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
4237 if (SimplifyIndirectBr(IBI)) return true;
4243 /// SimplifyCFG - This function is used to do simplification of a CFG. For
4244 /// example, it adjusts branches to branches to eliminate the extra hop, it
4245 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4246 /// of the CFG. It returns true if a modification was made.
4248 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
4249 const DataLayout *DL) {
4250 return SimplifyCFGOpt(TTI, DL).run(BB);