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"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Transforms/Utils/ValueMapper.h"
52 using namespace PatternMatch;
54 #define DEBUG_TYPE "simplifycfg"
56 static cl::opt<unsigned>
57 PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1),
58 cl::desc("Control the amount of phi node folding to perform (default = 1)"));
61 DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false),
62 cl::desc("Duplicate return instructions into unconditional branches"));
65 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
66 cl::desc("Sink common instructions down to the end block"));
68 static cl::opt<bool> HoistCondStores(
69 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
70 cl::desc("Hoist conditional stores if an unconditional store precedes"));
72 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
73 STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables");
74 STATISTIC(NumLookupTablesHoles, "Number of switch instructions turned into lookup tables (holes checked)");
75 STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block");
76 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
79 // The first field contains the value that the switch produces when a certain
80 // case group is selected, and the second field is a vector containing the cases
81 // composing the case group.
82 typedef SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>
83 SwitchCaseResultVectorTy;
84 // The first field contains the phi node that generates a result of the switch
85 // and the second field contains the value generated for a certain case in the switch
87 typedef SmallVector<std::pair<PHINode *, Constant *>, 4> SwitchCaseResultsTy;
89 /// ValueEqualityComparisonCase - Represents a case of a switch.
90 struct ValueEqualityComparisonCase {
94 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
95 : Value(Value), Dest(Dest) {}
97 bool operator<(ValueEqualityComparisonCase RHS) const {
98 // Comparing pointers is ok as we only rely on the order for uniquing.
99 return Value < RHS.Value;
102 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
105 class SimplifyCFGOpt {
106 const TargetTransformInfo &TTI;
107 unsigned BonusInstThreshold;
108 const DataLayout *const DL;
109 AssumptionTracker *AT;
110 Value *isValueEqualityComparison(TerminatorInst *TI);
111 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
112 std::vector<ValueEqualityComparisonCase> &Cases);
113 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
115 IRBuilder<> &Builder);
116 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
117 IRBuilder<> &Builder);
119 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
120 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
121 bool SimplifyUnreachable(UnreachableInst *UI);
122 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
123 bool SimplifyIndirectBr(IndirectBrInst *IBI);
124 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder);
125 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder);
128 SimplifyCFGOpt(const TargetTransformInfo &TTI, unsigned BonusInstThreshold,
129 const DataLayout *DL, AssumptionTracker *AT)
130 : TTI(TTI), BonusInstThreshold(BonusInstThreshold), DL(DL), AT(AT) {}
131 bool run(BasicBlock *BB);
135 /// SafeToMergeTerminators - Return true if it is safe to merge these two
136 /// terminator instructions together.
138 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
139 if (SI1 == SI2) return false; // Can't merge with self!
141 // It is not safe to merge these two switch instructions if they have a common
142 // successor, and if that successor has a PHI node, and if *that* PHI node has
143 // conflicting incoming values from the two switch blocks.
144 BasicBlock *SI1BB = SI1->getParent();
145 BasicBlock *SI2BB = SI2->getParent();
146 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
148 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
149 if (SI1Succs.count(*I))
150 for (BasicBlock::iterator BBI = (*I)->begin();
151 isa<PHINode>(BBI); ++BBI) {
152 PHINode *PN = cast<PHINode>(BBI);
153 if (PN->getIncomingValueForBlock(SI1BB) !=
154 PN->getIncomingValueForBlock(SI2BB))
161 /// isProfitableToFoldUnconditional - Return true if it is safe and profitable
162 /// to merge these two terminator instructions together, where SI1 is an
163 /// unconditional branch. PhiNodes will store all PHI nodes in common
166 static bool isProfitableToFoldUnconditional(BranchInst *SI1,
169 SmallVectorImpl<PHINode*> &PhiNodes) {
170 if (SI1 == SI2) return false; // Can't merge with self!
171 assert(SI1->isUnconditional() && SI2->isConditional());
173 // We fold the unconditional branch if we can easily update all PHI nodes in
174 // common successors:
175 // 1> We have a constant incoming value for the conditional branch;
176 // 2> We have "Cond" as the incoming value for the unconditional branch;
177 // 3> SI2->getCondition() and Cond have same operands.
178 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
179 if (!Ci2) return false;
180 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
181 Cond->getOperand(1) == Ci2->getOperand(1)) &&
182 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
183 Cond->getOperand(1) == Ci2->getOperand(0)))
186 BasicBlock *SI1BB = SI1->getParent();
187 BasicBlock *SI2BB = SI2->getParent();
188 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
189 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
190 if (SI1Succs.count(*I))
191 for (BasicBlock::iterator BBI = (*I)->begin();
192 isa<PHINode>(BBI); ++BBI) {
193 PHINode *PN = cast<PHINode>(BBI);
194 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
195 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
197 PhiNodes.push_back(PN);
202 /// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will
203 /// now be entries in it from the 'NewPred' block. The values that will be
204 /// flowing into the PHI nodes will be the same as those coming in from
205 /// ExistPred, an existing predecessor of Succ.
206 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
207 BasicBlock *ExistPred) {
208 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
211 for (BasicBlock::iterator I = Succ->begin();
212 (PN = dyn_cast<PHINode>(I)); ++I)
213 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
216 /// ComputeSpeculationCost - Compute an abstract "cost" of speculating the
217 /// given instruction, which is assumed to be safe to speculate. 1 means
218 /// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive.
219 static unsigned ComputeSpeculationCost(const User *I, const DataLayout *DL) {
220 assert(isSafeToSpeculativelyExecute(I, DL) &&
221 "Instruction is not safe to speculatively execute!");
222 switch (Operator::getOpcode(I)) {
224 // In doubt, be conservative.
226 case Instruction::GetElementPtr:
227 // GEPs are cheap if all indices are constant.
228 if (!cast<GEPOperator>(I)->hasAllConstantIndices())
231 case Instruction::ExtractValue:
232 case Instruction::Load:
233 case Instruction::Add:
234 case Instruction::Sub:
235 case Instruction::And:
236 case Instruction::Or:
237 case Instruction::Xor:
238 case Instruction::Shl:
239 case Instruction::LShr:
240 case Instruction::AShr:
241 case Instruction::ICmp:
242 case Instruction::Trunc:
243 case Instruction::ZExt:
244 case Instruction::SExt:
245 case Instruction::BitCast:
246 case Instruction::ExtractElement:
247 case Instruction::InsertElement:
248 return 1; // These are all cheap.
250 case Instruction::Call:
251 case Instruction::Select:
256 /// DominatesMergePoint - If we have a merge point of an "if condition" as
257 /// accepted above, return true if the specified value dominates the block. We
258 /// don't handle the true generality of domination here, just a special case
259 /// which works well enough for us.
261 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
262 /// see if V (which must be an instruction) and its recursive operands
263 /// that do not dominate BB have a combined cost lower than CostRemaining and
264 /// are non-trapping. If both are true, the instruction is inserted into the
265 /// set and true is returned.
267 /// The cost for most non-trapping instructions is defined as 1 except for
268 /// Select whose cost is 2.
270 /// After this function returns, CostRemaining is decreased by the cost of
271 /// V plus its non-dominating operands. If that cost is greater than
272 /// CostRemaining, false is returned and CostRemaining is undefined.
273 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
274 SmallPtrSetImpl<Instruction*> *AggressiveInsts,
275 unsigned &CostRemaining,
276 const DataLayout *DL) {
277 Instruction *I = dyn_cast<Instruction>(V);
279 // Non-instructions all dominate instructions, but not all constantexprs
280 // can be executed unconditionally.
281 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
286 BasicBlock *PBB = I->getParent();
288 // We don't want to allow weird loops that might have the "if condition" in
289 // the bottom of this block.
290 if (PBB == BB) return false;
292 // If this instruction is defined in a block that contains an unconditional
293 // branch to BB, then it must be in the 'conditional' part of the "if
294 // statement". If not, it definitely dominates the region.
295 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
296 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
299 // If we aren't allowing aggressive promotion anymore, then don't consider
300 // instructions in the 'if region'.
301 if (!AggressiveInsts) return false;
303 // If we have seen this instruction before, don't count it again.
304 if (AggressiveInsts->count(I)) return true;
306 // Okay, it looks like the instruction IS in the "condition". Check to
307 // see if it's a cheap instruction to unconditionally compute, and if it
308 // only uses stuff defined outside of the condition. If so, hoist it out.
309 if (!isSafeToSpeculativelyExecute(I, DL))
312 unsigned Cost = ComputeSpeculationCost(I, DL);
314 if (Cost > CostRemaining)
317 CostRemaining -= Cost;
319 // Okay, we can only really hoist these out if their operands do
320 // not take us over the cost threshold.
321 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
322 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, DL))
324 // Okay, it's safe to do this! Remember this instruction.
325 AggressiveInsts->insert(I);
329 /// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr
330 /// and PointerNullValue. Return NULL if value is not a constant int.
331 static ConstantInt *GetConstantInt(Value *V, const DataLayout *DL) {
332 // Normal constant int.
333 ConstantInt *CI = dyn_cast<ConstantInt>(V);
334 if (CI || !DL || !isa<Constant>(V) || !V->getType()->isPointerTy())
337 // This is some kind of pointer constant. Turn it into a pointer-sized
338 // ConstantInt if possible.
339 IntegerType *PtrTy = cast<IntegerType>(DL->getIntPtrType(V->getType()));
341 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
342 if (isa<ConstantPointerNull>(V))
343 return ConstantInt::get(PtrTy, 0);
345 // IntToPtr const int.
346 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
347 if (CE->getOpcode() == Instruction::IntToPtr)
348 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
349 // The constant is very likely to have the right type already.
350 if (CI->getType() == PtrTy)
353 return cast<ConstantInt>
354 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
359 /// GatherConstantCompares - Given a potentially 'or'd or 'and'd together
360 /// collection of icmp eq/ne instructions that compare a value against a
361 /// constant, return the value being compared, and stick the constant into the
364 GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra,
365 const DataLayout *DL, bool isEQ, unsigned &UsedICmps) {
366 Instruction *I = dyn_cast<Instruction>(V);
367 if (!I) return nullptr;
369 // If this is an icmp against a constant, handle this as one of the cases.
370 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) {
371 if (ConstantInt *C = GetConstantInt(I->getOperand(1), DL)) {
375 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
376 // (x & ~2^x) == y --> x == y || x == y|2^x
377 // This undoes a transformation done by instcombine to fuse 2 compares.
378 if (match(ICI->getOperand(0),
379 m_And(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
380 APInt Not = ~RHSC->getValue();
381 if (Not.isPowerOf2()) {
384 ConstantInt::get(C->getContext(), C->getValue() | Not));
392 return I->getOperand(0);
395 // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to
398 ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue());
400 // Shift the range if the compare is fed by an add. This is the range
401 // compare idiom as emitted by instcombine.
403 match(I->getOperand(0), m_Add(m_Value(RHSVal), m_ConstantInt(RHSC)));
405 Span = Span.subtract(RHSC->getValue());
407 // If this is an and/!= check then we want to optimize "x ugt 2" into
410 Span = Span.inverse();
412 // If there are a ton of values, we don't want to make a ginormous switch.
413 if (Span.getSetSize().ugt(8) || Span.isEmptySet())
416 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
417 Vals.push_back(ConstantInt::get(V->getContext(), Tmp));
419 return hasAdd ? RHSVal : I->getOperand(0);
424 // Otherwise, we can only handle an | or &, depending on isEQ.
425 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And))
428 unsigned NumValsBeforeLHS = Vals.size();
429 unsigned UsedICmpsBeforeLHS = UsedICmps;
430 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, DL,
432 unsigned NumVals = Vals.size();
433 unsigned UsedICmpsBeforeRHS = UsedICmps;
434 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, DL,
438 Vals.resize(NumVals);
439 UsedICmps = UsedICmpsBeforeRHS;
442 // The RHS of the or/and can't be folded in and we haven't used "Extra" yet,
443 // set it and return success.
444 if (Extra == nullptr || Extra == I->getOperand(1)) {
445 Extra = I->getOperand(1);
449 Vals.resize(NumValsBeforeLHS);
450 UsedICmps = UsedICmpsBeforeLHS;
454 // If the LHS can't be folded in, but Extra is available and RHS can, try to
456 if (Extra == nullptr || Extra == I->getOperand(0)) {
457 Value *OldExtra = Extra;
458 Extra = I->getOperand(0);
459 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, DL,
462 assert(Vals.size() == NumValsBeforeLHS);
469 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
470 Instruction *Cond = nullptr;
471 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
472 Cond = dyn_cast<Instruction>(SI->getCondition());
473 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
474 if (BI->isConditional())
475 Cond = dyn_cast<Instruction>(BI->getCondition());
476 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
477 Cond = dyn_cast<Instruction>(IBI->getAddress());
480 TI->eraseFromParent();
481 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
484 /// isValueEqualityComparison - Return true if the specified terminator checks
485 /// to see if a value is equal to constant integer value.
486 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
488 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
489 // Do not permit merging of large switch instructions into their
490 // predecessors unless there is only one predecessor.
491 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()),
492 pred_end(SI->getParent())) <= 128)
493 CV = SI->getCondition();
494 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
495 if (BI->isConditional() && BI->getCondition()->hasOneUse())
496 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition()))
497 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
498 CV = ICI->getOperand(0);
500 // Unwrap any lossless ptrtoint cast.
502 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
503 Value *Ptr = PTII->getPointerOperand();
504 if (PTII->getType() == DL->getIntPtrType(Ptr->getType()))
511 /// GetValueEqualityComparisonCases - Given a value comparison instruction,
512 /// decode all of the 'cases' that it represents and return the 'default' block.
513 BasicBlock *SimplifyCFGOpt::
514 GetValueEqualityComparisonCases(TerminatorInst *TI,
515 std::vector<ValueEqualityComparisonCase>
517 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
518 Cases.reserve(SI->getNumCases());
519 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
520 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
521 i.getCaseSuccessor()));
522 return SI->getDefaultDest();
525 BranchInst *BI = cast<BranchInst>(TI);
526 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
527 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
528 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
531 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
535 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries
536 /// in the list that match the specified block.
537 static void EliminateBlockCases(BasicBlock *BB,
538 std::vector<ValueEqualityComparisonCase> &Cases) {
539 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
542 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as
545 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
546 std::vector<ValueEqualityComparisonCase > &C2) {
547 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
549 // Make V1 be smaller than V2.
550 if (V1->size() > V2->size())
553 if (V1->size() == 0) return false;
554 if (V1->size() == 1) {
556 ConstantInt *TheVal = (*V1)[0].Value;
557 for (unsigned i = 0, e = V2->size(); i != e; ++i)
558 if (TheVal == (*V2)[i].Value)
562 // Otherwise, just sort both lists and compare element by element.
563 array_pod_sort(V1->begin(), V1->end());
564 array_pod_sort(V2->begin(), V2->end());
565 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
566 while (i1 != e1 && i2 != e2) {
567 if ((*V1)[i1].Value == (*V2)[i2].Value)
569 if ((*V1)[i1].Value < (*V2)[i2].Value)
577 /// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a
578 /// terminator instruction and its block is known to only have a single
579 /// predecessor block, check to see if that predecessor is also a value
580 /// comparison with the same value, and if that comparison determines the
581 /// outcome of this comparison. If so, simplify TI. This does a very limited
582 /// form of jump threading.
583 bool SimplifyCFGOpt::
584 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
586 IRBuilder<> &Builder) {
587 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
588 if (!PredVal) return false; // Not a value comparison in predecessor.
590 Value *ThisVal = isValueEqualityComparison(TI);
591 assert(ThisVal && "This isn't a value comparison!!");
592 if (ThisVal != PredVal) return false; // Different predicates.
594 // TODO: Preserve branch weight metadata, similarly to how
595 // FoldValueComparisonIntoPredecessors preserves it.
597 // Find out information about when control will move from Pred to TI's block.
598 std::vector<ValueEqualityComparisonCase> PredCases;
599 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
601 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
603 // Find information about how control leaves this block.
604 std::vector<ValueEqualityComparisonCase> ThisCases;
605 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
606 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
608 // If TI's block is the default block from Pred's comparison, potentially
609 // simplify TI based on this knowledge.
610 if (PredDef == TI->getParent()) {
611 // If we are here, we know that the value is none of those cases listed in
612 // PredCases. If there are any cases in ThisCases that are in PredCases, we
614 if (!ValuesOverlap(PredCases, ThisCases))
617 if (isa<BranchInst>(TI)) {
618 // Okay, one of the successors of this condbr is dead. Convert it to a
620 assert(ThisCases.size() == 1 && "Branch can only have one case!");
621 // Insert the new branch.
622 Instruction *NI = Builder.CreateBr(ThisDef);
625 // Remove PHI node entries for the dead edge.
626 ThisCases[0].Dest->removePredecessor(TI->getParent());
628 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
629 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
631 EraseTerminatorInstAndDCECond(TI);
635 SwitchInst *SI = cast<SwitchInst>(TI);
636 // Okay, TI has cases that are statically dead, prune them away.
637 SmallPtrSet<Constant*, 16> DeadCases;
638 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
639 DeadCases.insert(PredCases[i].Value);
641 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
642 << "Through successor TI: " << *TI);
644 // Collect branch weights into a vector.
645 SmallVector<uint32_t, 8> Weights;
646 MDNode* MD = SI->getMetadata(LLVMContext::MD_prof);
647 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
649 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
651 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i));
653 Weights.push_back(CI->getValue().getZExtValue());
655 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
657 if (DeadCases.count(i.getCaseValue())) {
659 std::swap(Weights[i.getCaseIndex()+1], Weights.back());
662 i.getCaseSuccessor()->removePredecessor(TI->getParent());
666 if (HasWeight && Weights.size() >= 2)
667 SI->setMetadata(LLVMContext::MD_prof,
668 MDBuilder(SI->getParent()->getContext()).
669 createBranchWeights(Weights));
671 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
675 // Otherwise, TI's block must correspond to some matched value. Find out
676 // which value (or set of values) this is.
677 ConstantInt *TIV = nullptr;
678 BasicBlock *TIBB = TI->getParent();
679 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
680 if (PredCases[i].Dest == TIBB) {
682 return false; // Cannot handle multiple values coming to this block.
683 TIV = PredCases[i].Value;
685 assert(TIV && "No edge from pred to succ?");
687 // Okay, we found the one constant that our value can be if we get into TI's
688 // BB. Find out which successor will unconditionally be branched to.
689 BasicBlock *TheRealDest = nullptr;
690 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
691 if (ThisCases[i].Value == TIV) {
692 TheRealDest = ThisCases[i].Dest;
696 // If not handled by any explicit cases, it is handled by the default case.
697 if (!TheRealDest) TheRealDest = ThisDef;
699 // Remove PHI node entries for dead edges.
700 BasicBlock *CheckEdge = TheRealDest;
701 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
702 if (*SI != CheckEdge)
703 (*SI)->removePredecessor(TIBB);
707 // Insert the new branch.
708 Instruction *NI = Builder.CreateBr(TheRealDest);
711 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
712 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
714 EraseTerminatorInstAndDCECond(TI);
719 /// ConstantIntOrdering - This class implements a stable ordering of constant
720 /// integers that does not depend on their address. This is important for
721 /// applications that sort ConstantInt's to ensure uniqueness.
722 struct ConstantIntOrdering {
723 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
724 return LHS->getValue().ult(RHS->getValue());
729 static int ConstantIntSortPredicate(ConstantInt *const *P1,
730 ConstantInt *const *P2) {
731 const ConstantInt *LHS = *P1;
732 const ConstantInt *RHS = *P2;
733 if (LHS->getValue().ult(RHS->getValue()))
735 if (LHS->getValue() == RHS->getValue())
740 static inline bool HasBranchWeights(const Instruction* I) {
741 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof);
742 if (ProfMD && ProfMD->getOperand(0))
743 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
744 return MDS->getString().equals("branch_weights");
749 /// Get Weights of a given TerminatorInst, the default weight is at the front
750 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
752 static void GetBranchWeights(TerminatorInst *TI,
753 SmallVectorImpl<uint64_t> &Weights) {
754 MDNode* MD = TI->getMetadata(LLVMContext::MD_prof);
756 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
757 ConstantInt *CI = cast<ConstantInt>(MD->getOperand(i));
758 Weights.push_back(CI->getValue().getZExtValue());
761 // If TI is a conditional eq, the default case is the false case,
762 // and the corresponding branch-weight data is at index 2. We swap the
763 // default weight to be the first entry.
764 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
765 assert(Weights.size() == 2);
766 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
767 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
768 std::swap(Weights.front(), Weights.back());
772 /// Keep halving the weights until all can fit in uint32_t.
773 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
774 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
775 if (Max > UINT_MAX) {
776 unsigned Offset = 32 - countLeadingZeros(Max);
777 for (uint64_t &I : Weights)
782 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value
783 /// equality comparison instruction (either a switch or a branch on "X == c").
784 /// See if any of the predecessors of the terminator block are value comparisons
785 /// on the same value. If so, and if safe to do so, fold them together.
786 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
787 IRBuilder<> &Builder) {
788 BasicBlock *BB = TI->getParent();
789 Value *CV = isValueEqualityComparison(TI); // CondVal
790 assert(CV && "Not a comparison?");
791 bool Changed = false;
793 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
794 while (!Preds.empty()) {
795 BasicBlock *Pred = Preds.pop_back_val();
797 // See if the predecessor is a comparison with the same value.
798 TerminatorInst *PTI = Pred->getTerminator();
799 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
801 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
802 // Figure out which 'cases' to copy from SI to PSI.
803 std::vector<ValueEqualityComparisonCase> BBCases;
804 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
806 std::vector<ValueEqualityComparisonCase> PredCases;
807 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
809 // Based on whether the default edge from PTI goes to BB or not, fill in
810 // PredCases and PredDefault with the new switch cases we would like to
812 SmallVector<BasicBlock*, 8> NewSuccessors;
814 // Update the branch weight metadata along the way
815 SmallVector<uint64_t, 8> Weights;
816 bool PredHasWeights = HasBranchWeights(PTI);
817 bool SuccHasWeights = HasBranchWeights(TI);
819 if (PredHasWeights) {
820 GetBranchWeights(PTI, Weights);
821 // branch-weight metadata is inconsistent here.
822 if (Weights.size() != 1 + PredCases.size())
823 PredHasWeights = SuccHasWeights = false;
824 } else if (SuccHasWeights)
825 // If there are no predecessor weights but there are successor weights,
826 // populate Weights with 1, which will later be scaled to the sum of
827 // successor's weights
828 Weights.assign(1 + PredCases.size(), 1);
830 SmallVector<uint64_t, 8> SuccWeights;
831 if (SuccHasWeights) {
832 GetBranchWeights(TI, SuccWeights);
833 // branch-weight metadata is inconsistent here.
834 if (SuccWeights.size() != 1 + BBCases.size())
835 PredHasWeights = SuccHasWeights = false;
836 } else if (PredHasWeights)
837 SuccWeights.assign(1 + BBCases.size(), 1);
839 if (PredDefault == BB) {
840 // If this is the default destination from PTI, only the edges in TI
841 // that don't occur in PTI, or that branch to BB will be activated.
842 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
843 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
844 if (PredCases[i].Dest != BB)
845 PTIHandled.insert(PredCases[i].Value);
847 // The default destination is BB, we don't need explicit targets.
848 std::swap(PredCases[i], PredCases.back());
850 if (PredHasWeights || SuccHasWeights) {
851 // Increase weight for the default case.
852 Weights[0] += Weights[i+1];
853 std::swap(Weights[i+1], Weights.back());
857 PredCases.pop_back();
861 // Reconstruct the new switch statement we will be building.
862 if (PredDefault != BBDefault) {
863 PredDefault->removePredecessor(Pred);
864 PredDefault = BBDefault;
865 NewSuccessors.push_back(BBDefault);
868 unsigned CasesFromPred = Weights.size();
869 uint64_t ValidTotalSuccWeight = 0;
870 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
871 if (!PTIHandled.count(BBCases[i].Value) &&
872 BBCases[i].Dest != BBDefault) {
873 PredCases.push_back(BBCases[i]);
874 NewSuccessors.push_back(BBCases[i].Dest);
875 if (SuccHasWeights || PredHasWeights) {
876 // The default weight is at index 0, so weight for the ith case
877 // should be at index i+1. Scale the cases from successor by
878 // PredDefaultWeight (Weights[0]).
879 Weights.push_back(Weights[0] * SuccWeights[i+1]);
880 ValidTotalSuccWeight += SuccWeights[i+1];
884 if (SuccHasWeights || PredHasWeights) {
885 ValidTotalSuccWeight += SuccWeights[0];
886 // Scale the cases from predecessor by ValidTotalSuccWeight.
887 for (unsigned i = 1; i < CasesFromPred; ++i)
888 Weights[i] *= ValidTotalSuccWeight;
889 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
890 Weights[0] *= SuccWeights[0];
893 // If this is not the default destination from PSI, only the edges
894 // in SI that occur in PSI with a destination of BB will be
896 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
897 std::map<ConstantInt*, uint64_t> WeightsForHandled;
898 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
899 if (PredCases[i].Dest == BB) {
900 PTIHandled.insert(PredCases[i].Value);
902 if (PredHasWeights || SuccHasWeights) {
903 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
904 std::swap(Weights[i+1], Weights.back());
908 std::swap(PredCases[i], PredCases.back());
909 PredCases.pop_back();
913 // Okay, now we know which constants were sent to BB from the
914 // predecessor. Figure out where they will all go now.
915 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
916 if (PTIHandled.count(BBCases[i].Value)) {
917 // If this is one we are capable of getting...
918 if (PredHasWeights || SuccHasWeights)
919 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
920 PredCases.push_back(BBCases[i]);
921 NewSuccessors.push_back(BBCases[i].Dest);
922 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
925 // If there are any constants vectored to BB that TI doesn't handle,
926 // they must go to the default destination of TI.
927 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
929 E = PTIHandled.end(); I != E; ++I) {
930 if (PredHasWeights || SuccHasWeights)
931 Weights.push_back(WeightsForHandled[*I]);
932 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
933 NewSuccessors.push_back(BBDefault);
937 // Okay, at this point, we know which new successor Pred will get. Make
938 // sure we update the number of entries in the PHI nodes for these
940 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i)
941 AddPredecessorToBlock(NewSuccessors[i], Pred, BB);
943 Builder.SetInsertPoint(PTI);
944 // Convert pointer to int before we switch.
945 if (CV->getType()->isPointerTy()) {
946 assert(DL && "Cannot switch on pointer without DataLayout");
947 CV = Builder.CreatePtrToInt(CV, DL->getIntPtrType(CV->getType()),
951 // Now that the successors are updated, create the new Switch instruction.
952 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
954 NewSI->setDebugLoc(PTI->getDebugLoc());
955 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
956 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest);
958 if (PredHasWeights || SuccHasWeights) {
959 // Halve the weights if any of them cannot fit in an uint32_t
962 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
964 NewSI->setMetadata(LLVMContext::MD_prof,
965 MDBuilder(BB->getContext()).
966 createBranchWeights(MDWeights));
969 EraseTerminatorInstAndDCECond(PTI);
971 // Okay, last check. If BB is still a successor of PSI, then we must
972 // have an infinite loop case. If so, add an infinitely looping block
973 // to handle the case to preserve the behavior of the code.
974 BasicBlock *InfLoopBlock = nullptr;
975 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
976 if (NewSI->getSuccessor(i) == BB) {
978 // Insert it at the end of the function, because it's either code,
979 // or it won't matter if it's hot. :)
980 InfLoopBlock = BasicBlock::Create(BB->getContext(),
981 "infloop", BB->getParent());
982 BranchInst::Create(InfLoopBlock, InfLoopBlock);
984 NewSI->setSuccessor(i, InfLoopBlock);
993 // isSafeToHoistInvoke - If we would need to insert a select that uses the
994 // value of this invoke (comments in HoistThenElseCodeToIf explain why we
995 // would need to do this), we can't hoist the invoke, as there is nowhere
996 // to put the select in this case.
997 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
998 Instruction *I1, Instruction *I2) {
999 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1001 for (BasicBlock::iterator BBI = SI->begin();
1002 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1003 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1004 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1005 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
1013 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I);
1015 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
1016 /// BB2, hoist any common code in the two blocks up into the branch block. The
1017 /// caller of this function guarantees that BI's block dominates BB1 and BB2.
1018 static bool HoistThenElseCodeToIf(BranchInst *BI, const DataLayout *DL) {
1019 // This does very trivial matching, with limited scanning, to find identical
1020 // instructions in the two blocks. In particular, we don't want to get into
1021 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1022 // such, we currently just scan for obviously identical instructions in an
1024 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1025 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1027 BasicBlock::iterator BB1_Itr = BB1->begin();
1028 BasicBlock::iterator BB2_Itr = BB2->begin();
1030 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++;
1031 // Skip debug info if it is not identical.
1032 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1033 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1034 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1035 while (isa<DbgInfoIntrinsic>(I1))
1037 while (isa<DbgInfoIntrinsic>(I2))
1040 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1041 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1044 BasicBlock *BIParent = BI->getParent();
1046 bool Changed = false;
1048 // If we are hoisting the terminator instruction, don't move one (making a
1049 // broken BB), instead clone it, and remove BI.
1050 if (isa<TerminatorInst>(I1))
1051 goto HoistTerminator;
1053 // For a normal instruction, we just move one to right before the branch,
1054 // then replace all uses of the other with the first. Finally, we remove
1055 // the now redundant second instruction.
1056 BIParent->getInstList().splice(BI, BB1->getInstList(), I1);
1057 if (!I2->use_empty())
1058 I2->replaceAllUsesWith(I1);
1059 I1->intersectOptionalDataWith(I2);
1060 unsigned KnownIDs[] = {
1061 LLVMContext::MD_tbaa,
1062 LLVMContext::MD_range,
1063 LLVMContext::MD_fpmath,
1064 LLVMContext::MD_invariant_load
1066 combineMetadata(I1, I2, KnownIDs);
1067 I2->eraseFromParent();
1072 // Skip debug info if it is not identical.
1073 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1074 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1075 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1076 while (isa<DbgInfoIntrinsic>(I1))
1078 while (isa<DbgInfoIntrinsic>(I2))
1081 } while (I1->isIdenticalToWhenDefined(I2));
1086 // It may not be possible to hoist an invoke.
1087 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1090 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1092 for (BasicBlock::iterator BBI = SI->begin();
1093 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1094 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1095 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1099 // Check for passingValueIsAlwaysUndefined here because we would rather
1100 // eliminate undefined control flow then converting it to a select.
1101 if (passingValueIsAlwaysUndefined(BB1V, PN) ||
1102 passingValueIsAlwaysUndefined(BB2V, PN))
1105 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V, DL))
1107 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V, DL))
1112 // Okay, it is safe to hoist the terminator.
1113 Instruction *NT = I1->clone();
1114 BIParent->getInstList().insert(BI, NT);
1115 if (!NT->getType()->isVoidTy()) {
1116 I1->replaceAllUsesWith(NT);
1117 I2->replaceAllUsesWith(NT);
1121 IRBuilder<true, NoFolder> Builder(NT);
1122 // Hoisting one of the terminators from our successor is a great thing.
1123 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1124 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1125 // nodes, so we insert select instruction to compute the final result.
1126 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1127 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1129 for (BasicBlock::iterator BBI = SI->begin();
1130 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1131 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1132 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1133 if (BB1V == BB2V) continue;
1135 // These values do not agree. Insert a select instruction before NT
1136 // that determines the right value.
1137 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1139 SI = cast<SelectInst>
1140 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1141 BB1V->getName()+"."+BB2V->getName()));
1143 // Make the PHI node use the select for all incoming values for BB1/BB2
1144 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1145 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1146 PN->setIncomingValue(i, SI);
1150 // Update any PHI nodes in our new successors.
1151 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1152 AddPredecessorToBlock(*SI, BIParent, BB1);
1154 EraseTerminatorInstAndDCECond(BI);
1158 /// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd,
1159 /// check whether BBEnd has only two predecessors and the other predecessor
1160 /// ends with an unconditional branch. If it is true, sink any common code
1161 /// in the two predecessors to BBEnd.
1162 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1163 assert(BI1->isUnconditional());
1164 BasicBlock *BB1 = BI1->getParent();
1165 BasicBlock *BBEnd = BI1->getSuccessor(0);
1167 // Check that BBEnd has two predecessors and the other predecessor ends with
1168 // an unconditional branch.
1169 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1170 BasicBlock *Pred0 = *PI++;
1171 if (PI == PE) // Only one predecessor.
1173 BasicBlock *Pred1 = *PI++;
1174 if (PI != PE) // More than two predecessors.
1176 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1177 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1178 if (!BI2 || !BI2->isUnconditional())
1181 // Gather the PHI nodes in BBEnd.
1182 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2;
1183 Instruction *FirstNonPhiInBBEnd = nullptr;
1184 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end();
1186 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1187 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1188 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1189 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN);
1191 FirstNonPhiInBBEnd = &*I;
1195 if (!FirstNonPhiInBBEnd)
1199 // This does very trivial matching, with limited scanning, to find identical
1200 // instructions in the two blocks. We scan backward for obviously identical
1201 // instructions in an identical order.
1202 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1203 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(),
1204 RE2 = BB2->getInstList().rend();
1206 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1209 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1212 // Skip the unconditional branches.
1216 bool Changed = false;
1217 while (RI1 != RE1 && RI2 != RE2) {
1219 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1222 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1226 Instruction *I1 = &*RI1, *I2 = &*RI2;
1227 // I1 and I2 should have a single use in the same PHI node, and they
1228 // perform the same operation.
1229 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1230 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1231 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1232 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) ||
1233 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1234 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1235 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1236 !I1->hasOneUse() || !I2->hasOneUse() ||
1237 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() ||
1238 MapValueFromBB1ToBB2[I1].first != I2)
1241 // Check whether we should swap the operands of ICmpInst.
1242 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1243 bool SwapOpnds = false;
1244 if (ICmp1 && ICmp2 &&
1245 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1246 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1247 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1248 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1249 ICmp2->swapOperands();
1252 if (!I1->isSameOperationAs(I2)) {
1254 ICmp2->swapOperands();
1258 // The operands should be either the same or they need to be generated
1259 // with a PHI node after sinking. We only handle the case where there is
1260 // a single pair of different operands.
1261 Value *DifferentOp1 = nullptr, *DifferentOp2 = nullptr;
1262 unsigned Op1Idx = 0;
1263 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1264 if (I1->getOperand(I) == I2->getOperand(I))
1266 // Early exit if we have more-than one pair of different operands or
1267 // the different operand is already in MapValueFromBB1ToBB2.
1268 // Early exit if we need a PHI node to replace a constant.
1270 MapValueFromBB1ToBB2.find(I1->getOperand(I)) !=
1271 MapValueFromBB1ToBB2.end() ||
1272 isa<Constant>(I1->getOperand(I)) ||
1273 isa<Constant>(I2->getOperand(I))) {
1274 // If we can't sink the instructions, undo the swapping.
1276 ICmp2->swapOperands();
1279 DifferentOp1 = I1->getOperand(I);
1281 DifferentOp2 = I2->getOperand(I);
1284 // We insert the pair of different operands to MapValueFromBB1ToBB2 and
1285 // remove (I1, I2) from MapValueFromBB1ToBB2.
1287 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2,
1288 DifferentOp1->getName() + ".sink",
1290 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN);
1291 // I1 should use NewPN instead of DifferentOp1.
1292 I1->setOperand(Op1Idx, NewPN);
1293 NewPN->addIncoming(DifferentOp1, BB1);
1294 NewPN->addIncoming(DifferentOp2, BB2);
1295 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1297 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second;
1298 MapValueFromBB1ToBB2.erase(I1);
1300 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";);
1301 DEBUG(dbgs() << " " << *I2 << "\n";);
1302 // We need to update RE1 and RE2 if we are going to sink the first
1303 // instruction in the basic block down.
1304 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1305 // Sink the instruction.
1306 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1);
1307 if (!OldPN->use_empty())
1308 OldPN->replaceAllUsesWith(I1);
1309 OldPN->eraseFromParent();
1311 if (!I2->use_empty())
1312 I2->replaceAllUsesWith(I1);
1313 I1->intersectOptionalDataWith(I2);
1314 I2->eraseFromParent();
1317 RE1 = BB1->getInstList().rend();
1319 RE2 = BB2->getInstList().rend();
1320 FirstNonPhiInBBEnd = I1;
1327 /// \brief Determine if we can hoist sink a sole store instruction out of a
1328 /// conditional block.
1330 /// We are looking for code like the following:
1332 /// store i32 %add, i32* %arrayidx2
1333 /// ... // No other stores or function calls (we could be calling a memory
1334 /// ... // function).
1335 /// %cmp = icmp ult %x, %y
1336 /// br i1 %cmp, label %EndBB, label %ThenBB
1338 /// store i32 %add5, i32* %arrayidx2
1342 /// We are going to transform this into:
1344 /// store i32 %add, i32* %arrayidx2
1346 /// %cmp = icmp ult %x, %y
1347 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1348 /// store i32 %add.add5, i32* %arrayidx2
1351 /// \return The pointer to the value of the previous store if the store can be
1352 /// hoisted into the predecessor block. 0 otherwise.
1353 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1354 BasicBlock *StoreBB, BasicBlock *EndBB) {
1355 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1359 // Volatile or atomic.
1360 if (!StoreToHoist->isSimple())
1363 Value *StorePtr = StoreToHoist->getPointerOperand();
1365 // Look for a store to the same pointer in BrBB.
1366 unsigned MaxNumInstToLookAt = 10;
1367 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(),
1368 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) {
1369 Instruction *CurI = &*RI;
1371 // Could be calling an instruction that effects memory like free().
1372 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI))
1375 StoreInst *SI = dyn_cast<StoreInst>(CurI);
1376 // Found the previous store make sure it stores to the same location.
1377 if (SI && SI->getPointerOperand() == StorePtr)
1378 // Found the previous store, return its value operand.
1379 return SI->getValueOperand();
1381 return nullptr; // Unknown store.
1387 /// \brief Speculate a conditional basic block flattening the CFG.
1389 /// Note that this is a very risky transform currently. Speculating
1390 /// instructions like this is most often not desirable. Instead, there is an MI
1391 /// pass which can do it with full awareness of the resource constraints.
1392 /// However, some cases are "obvious" and we should do directly. An example of
1393 /// this is speculating a single, reasonably cheap instruction.
1395 /// There is only one distinct advantage to flattening the CFG at the IR level:
1396 /// it makes very common but simplistic optimizations such as are common in
1397 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1398 /// modeling their effects with easier to reason about SSA value graphs.
1401 /// An illustration of this transform is turning this IR:
1404 /// %cmp = icmp ult %x, %y
1405 /// br i1 %cmp, label %EndBB, label %ThenBB
1407 /// %sub = sub %x, %y
1410 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1417 /// %cmp = icmp ult %x, %y
1418 /// %sub = sub %x, %y
1419 /// %cond = select i1 %cmp, 0, %sub
1423 /// \returns true if the conditional block is removed.
1424 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1425 const DataLayout *DL) {
1426 // Be conservative for now. FP select instruction can often be expensive.
1427 Value *BrCond = BI->getCondition();
1428 if (isa<FCmpInst>(BrCond))
1431 BasicBlock *BB = BI->getParent();
1432 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1434 // If ThenBB is actually on the false edge of the conditional branch, remember
1435 // to swap the select operands later.
1436 bool Invert = false;
1437 if (ThenBB != BI->getSuccessor(0)) {
1438 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1441 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1443 // Keep a count of how many times instructions are used within CondBB when
1444 // they are candidates for sinking into CondBB. Specifically:
1445 // - They are defined in BB, and
1446 // - They have no side effects, and
1447 // - All of their uses are in CondBB.
1448 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1450 unsigned SpeculationCost = 0;
1451 Value *SpeculatedStoreValue = nullptr;
1452 StoreInst *SpeculatedStore = nullptr;
1453 for (BasicBlock::iterator BBI = ThenBB->begin(),
1454 BBE = std::prev(ThenBB->end());
1455 BBI != BBE; ++BBI) {
1456 Instruction *I = BBI;
1458 if (isa<DbgInfoIntrinsic>(I))
1461 // Only speculatively execution a single instruction (not counting the
1462 // terminator) for now.
1464 if (SpeculationCost > 1)
1467 // Don't hoist the instruction if it's unsafe or expensive.
1468 if (!isSafeToSpeculativelyExecute(I, DL) &&
1469 !(HoistCondStores &&
1470 (SpeculatedStoreValue = isSafeToSpeculateStore(I, BB, ThenBB,
1473 if (!SpeculatedStoreValue &&
1474 ComputeSpeculationCost(I, DL) > PHINodeFoldingThreshold)
1477 // Store the store speculation candidate.
1478 if (SpeculatedStoreValue)
1479 SpeculatedStore = cast<StoreInst>(I);
1481 // Do not hoist the instruction if any of its operands are defined but not
1482 // used in BB. The transformation will prevent the operand from
1483 // being sunk into the use block.
1484 for (User::op_iterator i = I->op_begin(), e = I->op_end();
1486 Instruction *OpI = dyn_cast<Instruction>(*i);
1487 if (!OpI || OpI->getParent() != BB ||
1488 OpI->mayHaveSideEffects())
1489 continue; // Not a candidate for sinking.
1491 ++SinkCandidateUseCounts[OpI];
1495 // Consider any sink candidates which are only used in CondBB as costs for
1496 // speculation. Note, while we iterate over a DenseMap here, we are summing
1497 // and so iteration order isn't significant.
1498 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I =
1499 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end();
1501 if (I->first->getNumUses() == I->second) {
1503 if (SpeculationCost > 1)
1507 // Check that the PHI nodes can be converted to selects.
1508 bool HaveRewritablePHIs = false;
1509 for (BasicBlock::iterator I = EndBB->begin();
1510 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1511 Value *OrigV = PN->getIncomingValueForBlock(BB);
1512 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1514 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1515 // Skip PHIs which are trivial.
1519 // Don't convert to selects if we could remove undefined behavior instead.
1520 if (passingValueIsAlwaysUndefined(OrigV, PN) ||
1521 passingValueIsAlwaysUndefined(ThenV, PN))
1524 HaveRewritablePHIs = true;
1525 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1526 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1527 if (!OrigCE && !ThenCE)
1528 continue; // Known safe and cheap.
1530 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE, DL)) ||
1531 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE, DL)))
1533 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, DL) : 0;
1534 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, DL) : 0;
1535 if (OrigCost + ThenCost > 2 * PHINodeFoldingThreshold)
1538 // Account for the cost of an unfolded ConstantExpr which could end up
1539 // getting expanded into Instructions.
1540 // FIXME: This doesn't account for how many operations are combined in the
1541 // constant expression.
1543 if (SpeculationCost > 1)
1547 // If there are no PHIs to process, bail early. This helps ensure idempotence
1549 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
1552 // If we get here, we can hoist the instruction and if-convert.
1553 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1555 // Insert a select of the value of the speculated store.
1556 if (SpeculatedStoreValue) {
1557 IRBuilder<true, NoFolder> Builder(BI);
1558 Value *TrueV = SpeculatedStore->getValueOperand();
1559 Value *FalseV = SpeculatedStoreValue;
1561 std::swap(TrueV, FalseV);
1562 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() +
1563 "." + FalseV->getName());
1564 SpeculatedStore->setOperand(0, S);
1567 // Hoist the instructions.
1568 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(),
1569 std::prev(ThenBB->end()));
1571 // Insert selects and rewrite the PHI operands.
1572 IRBuilder<true, NoFolder> Builder(BI);
1573 for (BasicBlock::iterator I = EndBB->begin();
1574 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1575 unsigned OrigI = PN->getBasicBlockIndex(BB);
1576 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1577 Value *OrigV = PN->getIncomingValue(OrigI);
1578 Value *ThenV = PN->getIncomingValue(ThenI);
1580 // Skip PHIs which are trivial.
1584 // Create a select whose true value is the speculatively executed value and
1585 // false value is the preexisting value. Swap them if the branch
1586 // destinations were inverted.
1587 Value *TrueV = ThenV, *FalseV = OrigV;
1589 std::swap(TrueV, FalseV);
1590 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV,
1591 TrueV->getName() + "." + FalseV->getName());
1592 PN->setIncomingValue(OrigI, V);
1593 PN->setIncomingValue(ThenI, V);
1600 /// \returns True if this block contains a CallInst with the NoDuplicate
1602 static bool HasNoDuplicateCall(const BasicBlock *BB) {
1603 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1604 const CallInst *CI = dyn_cast<CallInst>(I);
1607 if (CI->cannotDuplicate())
1613 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1614 /// across this block.
1615 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1616 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1619 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1620 if (isa<DbgInfoIntrinsic>(BBI))
1622 if (Size > 10) return false; // Don't clone large BB's.
1625 // We can only support instructions that do not define values that are
1626 // live outside of the current basic block.
1627 for (User *U : BBI->users()) {
1628 Instruction *UI = cast<Instruction>(U);
1629 if (UI->getParent() != BB || isa<PHINode>(UI)) return false;
1632 // Looks ok, continue checking.
1638 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1639 /// that is defined in the same block as the branch and if any PHI entries are
1640 /// constants, thread edges corresponding to that entry to be branches to their
1641 /// ultimate destination.
1642 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *DL) {
1643 BasicBlock *BB = BI->getParent();
1644 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1645 // NOTE: we currently cannot transform this case if the PHI node is used
1646 // outside of the block.
1647 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1650 // Degenerate case of a single entry PHI.
1651 if (PN->getNumIncomingValues() == 1) {
1652 FoldSingleEntryPHINodes(PN->getParent());
1656 // Now we know that this block has multiple preds and two succs.
1657 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1659 if (HasNoDuplicateCall(BB)) return false;
1661 // Okay, this is a simple enough basic block. See if any phi values are
1663 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1664 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1665 if (!CB || !CB->getType()->isIntegerTy(1)) continue;
1667 // Okay, we now know that all edges from PredBB should be revectored to
1668 // branch to RealDest.
1669 BasicBlock *PredBB = PN->getIncomingBlock(i);
1670 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1672 if (RealDest == BB) continue; // Skip self loops.
1673 // Skip if the predecessor's terminator is an indirect branch.
1674 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1676 // The dest block might have PHI nodes, other predecessors and other
1677 // difficult cases. Instead of being smart about this, just insert a new
1678 // block that jumps to the destination block, effectively splitting
1679 // the edge we are about to create.
1680 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1681 RealDest->getName()+".critedge",
1682 RealDest->getParent(), RealDest);
1683 BranchInst::Create(RealDest, EdgeBB);
1685 // Update PHI nodes.
1686 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1688 // BB may have instructions that are being threaded over. Clone these
1689 // instructions into EdgeBB. We know that there will be no uses of the
1690 // cloned instructions outside of EdgeBB.
1691 BasicBlock::iterator InsertPt = EdgeBB->begin();
1692 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1693 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1694 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1695 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1698 // Clone the instruction.
1699 Instruction *N = BBI->clone();
1700 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1702 // Update operands due to translation.
1703 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1705 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1706 if (PI != TranslateMap.end())
1710 // Check for trivial simplification.
1711 if (Value *V = SimplifyInstruction(N, DL)) {
1712 TranslateMap[BBI] = V;
1713 delete N; // Instruction folded away, don't need actual inst
1715 // Insert the new instruction into its new home.
1716 EdgeBB->getInstList().insert(InsertPt, N);
1717 if (!BBI->use_empty())
1718 TranslateMap[BBI] = N;
1722 // Loop over all of the edges from PredBB to BB, changing them to branch
1723 // to EdgeBB instead.
1724 TerminatorInst *PredBBTI = PredBB->getTerminator();
1725 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1726 if (PredBBTI->getSuccessor(i) == BB) {
1727 BB->removePredecessor(PredBB);
1728 PredBBTI->setSuccessor(i, EdgeBB);
1731 // Recurse, simplifying any other constants.
1732 return FoldCondBranchOnPHI(BI, DL) | true;
1738 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1739 /// PHI node, see if we can eliminate it.
1740 static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *DL) {
1741 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1742 // statement", which has a very simple dominance structure. Basically, we
1743 // are trying to find the condition that is being branched on, which
1744 // subsequently causes this merge to happen. We really want control
1745 // dependence information for this check, but simplifycfg can't keep it up
1746 // to date, and this catches most of the cases we care about anyway.
1747 BasicBlock *BB = PN->getParent();
1748 BasicBlock *IfTrue, *IfFalse;
1749 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1751 // Don't bother if the branch will be constant folded trivially.
1752 isa<ConstantInt>(IfCond))
1755 // Okay, we found that we can merge this two-entry phi node into a select.
1756 // Doing so would require us to fold *all* two entry phi nodes in this block.
1757 // At some point this becomes non-profitable (particularly if the target
1758 // doesn't support cmov's). Only do this transformation if there are two or
1759 // fewer PHI nodes in this block.
1760 unsigned NumPhis = 0;
1761 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1765 // Loop over the PHI's seeing if we can promote them all to select
1766 // instructions. While we are at it, keep track of the instructions
1767 // that need to be moved to the dominating block.
1768 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1769 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1770 MaxCostVal1 = PHINodeFoldingThreshold;
1772 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1773 PHINode *PN = cast<PHINode>(II++);
1774 if (Value *V = SimplifyInstruction(PN, DL)) {
1775 PN->replaceAllUsesWith(V);
1776 PN->eraseFromParent();
1780 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1782 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1787 // If we folded the first phi, PN dangles at this point. Refresh it. If
1788 // we ran out of PHIs then we simplified them all.
1789 PN = dyn_cast<PHINode>(BB->begin());
1790 if (!PN) return true;
1792 // Don't fold i1 branches on PHIs which contain binary operators. These can
1793 // often be turned into switches and other things.
1794 if (PN->getType()->isIntegerTy(1) &&
1795 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1796 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1797 isa<BinaryOperator>(IfCond)))
1800 // If we all PHI nodes are promotable, check to make sure that all
1801 // instructions in the predecessor blocks can be promoted as well. If
1802 // not, we won't be able to get rid of the control flow, so it's not
1803 // worth promoting to select instructions.
1804 BasicBlock *DomBlock = nullptr;
1805 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1806 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1807 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1810 DomBlock = *pred_begin(IfBlock1);
1811 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1812 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1813 // This is not an aggressive instruction that we can promote.
1814 // Because of this, we won't be able to get rid of the control
1815 // flow, so the xform is not worth it.
1820 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1823 DomBlock = *pred_begin(IfBlock2);
1824 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1825 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1826 // This is not an aggressive instruction that we can promote.
1827 // Because of this, we won't be able to get rid of the control
1828 // flow, so the xform is not worth it.
1833 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1834 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1836 // If we can still promote the PHI nodes after this gauntlet of tests,
1837 // do all of the PHI's now.
1838 Instruction *InsertPt = DomBlock->getTerminator();
1839 IRBuilder<true, NoFolder> Builder(InsertPt);
1841 // Move all 'aggressive' instructions, which are defined in the
1842 // conditional parts of the if's up to the dominating block.
1844 DomBlock->getInstList().splice(InsertPt,
1845 IfBlock1->getInstList(), IfBlock1->begin(),
1846 IfBlock1->getTerminator());
1848 DomBlock->getInstList().splice(InsertPt,
1849 IfBlock2->getInstList(), IfBlock2->begin(),
1850 IfBlock2->getTerminator());
1852 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1853 // Change the PHI node into a select instruction.
1854 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1855 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1858 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1859 PN->replaceAllUsesWith(NV);
1861 PN->eraseFromParent();
1864 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1865 // has been flattened. Change DomBlock to jump directly to our new block to
1866 // avoid other simplifycfg's kicking in on the diamond.
1867 TerminatorInst *OldTI = DomBlock->getTerminator();
1868 Builder.SetInsertPoint(OldTI);
1869 Builder.CreateBr(BB);
1870 OldTI->eraseFromParent();
1874 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1875 /// to two returning blocks, try to merge them together into one return,
1876 /// introducing a select if the return values disagree.
1877 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1878 IRBuilder<> &Builder) {
1879 assert(BI->isConditional() && "Must be a conditional branch");
1880 BasicBlock *TrueSucc = BI->getSuccessor(0);
1881 BasicBlock *FalseSucc = BI->getSuccessor(1);
1882 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1883 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1885 // Check to ensure both blocks are empty (just a return) or optionally empty
1886 // with PHI nodes. If there are other instructions, merging would cause extra
1887 // computation on one path or the other.
1888 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1890 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1893 Builder.SetInsertPoint(BI);
1894 // Okay, we found a branch that is going to two return nodes. If
1895 // there is no return value for this function, just change the
1896 // branch into a return.
1897 if (FalseRet->getNumOperands() == 0) {
1898 TrueSucc->removePredecessor(BI->getParent());
1899 FalseSucc->removePredecessor(BI->getParent());
1900 Builder.CreateRetVoid();
1901 EraseTerminatorInstAndDCECond(BI);
1905 // Otherwise, figure out what the true and false return values are
1906 // so we can insert a new select instruction.
1907 Value *TrueValue = TrueRet->getReturnValue();
1908 Value *FalseValue = FalseRet->getReturnValue();
1910 // Unwrap any PHI nodes in the return blocks.
1911 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1912 if (TVPN->getParent() == TrueSucc)
1913 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1914 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1915 if (FVPN->getParent() == FalseSucc)
1916 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1918 // In order for this transformation to be safe, we must be able to
1919 // unconditionally execute both operands to the return. This is
1920 // normally the case, but we could have a potentially-trapping
1921 // constant expression that prevents this transformation from being
1923 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1926 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1930 // Okay, we collected all the mapped values and checked them for sanity, and
1931 // defined to really do this transformation. First, update the CFG.
1932 TrueSucc->removePredecessor(BI->getParent());
1933 FalseSucc->removePredecessor(BI->getParent());
1935 // Insert select instructions where needed.
1936 Value *BrCond = BI->getCondition();
1938 // Insert a select if the results differ.
1939 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1940 } else if (isa<UndefValue>(TrueValue)) {
1941 TrueValue = FalseValue;
1943 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1944 FalseValue, "retval");
1948 Value *RI = !TrueValue ?
1949 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
1953 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
1954 << "\n " << *BI << "NewRet = " << *RI
1955 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
1957 EraseTerminatorInstAndDCECond(BI);
1962 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
1963 /// probabilities of the branch taking each edge. Fills in the two APInt
1964 /// parameters and return true, or returns false if no or invalid metadata was
1966 static bool ExtractBranchMetadata(BranchInst *BI,
1967 uint64_t &ProbTrue, uint64_t &ProbFalse) {
1968 assert(BI->isConditional() &&
1969 "Looking for probabilities on unconditional branch?");
1970 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
1971 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
1972 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
1973 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
1974 if (!CITrue || !CIFalse) return false;
1975 ProbTrue = CITrue->getValue().getZExtValue();
1976 ProbFalse = CIFalse->getValue().getZExtValue();
1980 /// checkCSEInPredecessor - Return true if the given instruction is available
1981 /// in its predecessor block. If yes, the instruction will be removed.
1983 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
1984 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
1986 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
1987 Instruction *PBI = &*I;
1988 // Check whether Inst and PBI generate the same value.
1989 if (Inst->isIdenticalTo(PBI)) {
1990 Inst->replaceAllUsesWith(PBI);
1991 Inst->eraseFromParent();
1998 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
1999 /// predecessor branches to us and one of our successors, fold the block into
2000 /// the predecessor and use logical operations to pick the right destination.
2001 bool llvm::FoldBranchToCommonDest(BranchInst *BI, const DataLayout *DL,
2002 unsigned BonusInstThreshold) {
2003 BasicBlock *BB = BI->getParent();
2005 Instruction *Cond = nullptr;
2006 if (BI->isConditional())
2007 Cond = dyn_cast<Instruction>(BI->getCondition());
2009 // For unconditional branch, check for a simple CFG pattern, where
2010 // BB has a single predecessor and BB's successor is also its predecessor's
2011 // successor. If such pattern exisits, check for CSE between BB and its
2013 if (BasicBlock *PB = BB->getSinglePredecessor())
2014 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2015 if (PBI->isConditional() &&
2016 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2017 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2018 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
2020 Instruction *Curr = I++;
2021 if (isa<CmpInst>(Curr)) {
2025 // Quit if we can't remove this instruction.
2026 if (!checkCSEInPredecessor(Curr, PB))
2035 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2036 Cond->getParent() != BB || !Cond->hasOneUse())
2039 // Make sure the instruction after the condition is the cond branch.
2040 BasicBlock::iterator CondIt = Cond; ++CondIt;
2042 // Ignore dbg intrinsics.
2043 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
2048 // Only allow this transformation if computing the condition doesn't involve
2049 // too many instructions and these involved instructions can be executed
2050 // unconditionally. We denote all involved instructions except the condition
2051 // as "bonus instructions", and only allow this transformation when the
2052 // number of the bonus instructions does not exceed a certain threshold.
2053 unsigned NumBonusInsts = 0;
2054 for (auto I = BB->begin(); Cond != I; ++I) {
2055 // Ignore dbg intrinsics.
2056 if (isa<DbgInfoIntrinsic>(I))
2058 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(I, DL))
2060 // I has only one use and can be executed unconditionally.
2061 Instruction *User = dyn_cast<Instruction>(I->user_back());
2062 if (User == nullptr || User->getParent() != BB)
2064 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2065 // to use any other instruction, User must be an instruction between next(I)
2068 // Early exits once we reach the limit.
2069 if (NumBonusInsts > BonusInstThreshold)
2073 // Cond is known to be a compare or binary operator. Check to make sure that
2074 // neither operand is a potentially-trapping constant expression.
2075 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2078 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2082 // Finally, don't infinitely unroll conditional loops.
2083 BasicBlock *TrueDest = BI->getSuccessor(0);
2084 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2085 if (TrueDest == BB || FalseDest == BB)
2088 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2089 BasicBlock *PredBlock = *PI;
2090 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2092 // Check that we have two conditional branches. If there is a PHI node in
2093 // the common successor, verify that the same value flows in from both
2095 SmallVector<PHINode*, 4> PHIs;
2096 if (!PBI || PBI->isUnconditional() ||
2097 (BI->isConditional() &&
2098 !SafeToMergeTerminators(BI, PBI)) ||
2099 (!BI->isConditional() &&
2100 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2103 // Determine if the two branches share a common destination.
2104 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2105 bool InvertPredCond = false;
2107 if (BI->isConditional()) {
2108 if (PBI->getSuccessor(0) == TrueDest)
2109 Opc = Instruction::Or;
2110 else if (PBI->getSuccessor(1) == FalseDest)
2111 Opc = Instruction::And;
2112 else if (PBI->getSuccessor(0) == FalseDest)
2113 Opc = Instruction::And, InvertPredCond = true;
2114 else if (PBI->getSuccessor(1) == TrueDest)
2115 Opc = Instruction::Or, InvertPredCond = true;
2119 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2123 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2124 IRBuilder<> Builder(PBI);
2126 // If we need to invert the condition in the pred block to match, do so now.
2127 if (InvertPredCond) {
2128 Value *NewCond = PBI->getCondition();
2130 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2131 CmpInst *CI = cast<CmpInst>(NewCond);
2132 CI->setPredicate(CI->getInversePredicate());
2134 NewCond = Builder.CreateNot(NewCond,
2135 PBI->getCondition()->getName()+".not");
2138 PBI->setCondition(NewCond);
2139 PBI->swapSuccessors();
2142 // If we have bonus instructions, clone them into the predecessor block.
2143 // Note that there may be mutliple predecessor blocks, so we cannot move
2144 // bonus instructions to a predecessor block.
2145 ValueToValueMapTy VMap; // maps original values to cloned values
2146 // We already make sure Cond is the last instruction before BI. Therefore,
2147 // every instructions before Cond other than DbgInfoIntrinsic are bonus
2149 for (auto BonusInst = BB->begin(); Cond != BonusInst; ++BonusInst) {
2150 if (isa<DbgInfoIntrinsic>(BonusInst))
2152 Instruction *NewBonusInst = BonusInst->clone();
2153 RemapInstruction(NewBonusInst, VMap,
2154 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2155 VMap[BonusInst] = NewBonusInst;
2157 // If we moved a load, we cannot any longer claim any knowledge about
2158 // its potential value. The previous information might have been valid
2159 // only given the branch precondition.
2160 // For an analogous reason, we must also drop all the metadata whose
2161 // semantics we don't understand.
2162 NewBonusInst->dropUnknownMetadata(LLVMContext::MD_dbg);
2164 PredBlock->getInstList().insert(PBI, NewBonusInst);
2165 NewBonusInst->takeName(BonusInst);
2166 BonusInst->setName(BonusInst->getName() + ".old");
2169 // Clone Cond into the predecessor basic block, and or/and the
2170 // two conditions together.
2171 Instruction *New = Cond->clone();
2172 RemapInstruction(New, VMap,
2173 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2174 PredBlock->getInstList().insert(PBI, New);
2175 New->takeName(Cond);
2176 Cond->setName(New->getName() + ".old");
2178 if (BI->isConditional()) {
2179 Instruction *NewCond =
2180 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2182 PBI->setCondition(NewCond);
2184 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2185 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2187 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2189 SmallVector<uint64_t, 8> NewWeights;
2191 if (PBI->getSuccessor(0) == BB) {
2192 if (PredHasWeights && SuccHasWeights) {
2193 // PBI: br i1 %x, BB, FalseDest
2194 // BI: br i1 %y, TrueDest, FalseDest
2195 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2196 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2197 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2198 // TrueWeight for PBI * FalseWeight for BI.
2199 // We assume that total weights of a BranchInst can fit into 32 bits.
2200 // Therefore, we will not have overflow using 64-bit arithmetic.
2201 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2202 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2204 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2205 PBI->setSuccessor(0, TrueDest);
2207 if (PBI->getSuccessor(1) == BB) {
2208 if (PredHasWeights && SuccHasWeights) {
2209 // PBI: br i1 %x, TrueDest, BB
2210 // BI: br i1 %y, TrueDest, FalseDest
2211 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2212 // FalseWeight for PBI * TrueWeight for BI.
2213 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2214 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2215 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2216 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2218 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2219 PBI->setSuccessor(1, FalseDest);
2221 if (NewWeights.size() == 2) {
2222 // Halve the weights if any of them cannot fit in an uint32_t
2223 FitWeights(NewWeights);
2225 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2226 PBI->setMetadata(LLVMContext::MD_prof,
2227 MDBuilder(BI->getContext()).
2228 createBranchWeights(MDWeights));
2230 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2232 // Update PHI nodes in the common successors.
2233 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2234 ConstantInt *PBI_C = cast<ConstantInt>(
2235 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2236 assert(PBI_C->getType()->isIntegerTy(1));
2237 Instruction *MergedCond = nullptr;
2238 if (PBI->getSuccessor(0) == TrueDest) {
2239 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2240 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2241 // is false: !PBI_Cond and BI_Value
2242 Instruction *NotCond =
2243 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2246 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2251 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2252 PBI->getCondition(), MergedCond,
2255 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2256 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2257 // is false: PBI_Cond and BI_Value
2259 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2260 PBI->getCondition(), New,
2262 if (PBI_C->isOne()) {
2263 Instruction *NotCond =
2264 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2267 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2268 NotCond, MergedCond,
2273 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2276 // Change PBI from Conditional to Unconditional.
2277 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2278 EraseTerminatorInstAndDCECond(PBI);
2282 // TODO: If BB is reachable from all paths through PredBlock, then we
2283 // could replace PBI's branch probabilities with BI's.
2285 // Copy any debug value intrinsics into the end of PredBlock.
2286 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2287 if (isa<DbgInfoIntrinsic>(*I))
2288 I->clone()->insertBefore(PBI);
2295 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2296 /// predecessor of another block, this function tries to simplify it. We know
2297 /// that PBI and BI are both conditional branches, and BI is in one of the
2298 /// successor blocks of PBI - PBI branches to BI.
2299 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2300 assert(PBI->isConditional() && BI->isConditional());
2301 BasicBlock *BB = BI->getParent();
2303 // If this block ends with a branch instruction, and if there is a
2304 // predecessor that ends on a branch of the same condition, make
2305 // this conditional branch redundant.
2306 if (PBI->getCondition() == BI->getCondition() &&
2307 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2308 // Okay, the outcome of this conditional branch is statically
2309 // knowable. If this block had a single pred, handle specially.
2310 if (BB->getSinglePredecessor()) {
2311 // Turn this into a branch on constant.
2312 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2313 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2315 return true; // Nuke the branch on constant.
2318 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2319 // in the constant and simplify the block result. Subsequent passes of
2320 // simplifycfg will thread the block.
2321 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2322 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2323 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2324 std::distance(PB, PE),
2325 BI->getCondition()->getName() + ".pr",
2327 // Okay, we're going to insert the PHI node. Since PBI is not the only
2328 // predecessor, compute the PHI'd conditional value for all of the preds.
2329 // Any predecessor where the condition is not computable we keep symbolic.
2330 for (pred_iterator PI = PB; PI != PE; ++PI) {
2331 BasicBlock *P = *PI;
2332 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2333 PBI != BI && PBI->isConditional() &&
2334 PBI->getCondition() == BI->getCondition() &&
2335 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2336 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2337 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2340 NewPN->addIncoming(BI->getCondition(), P);
2344 BI->setCondition(NewPN);
2349 // If this is a conditional branch in an empty block, and if any
2350 // predecessors are a conditional branch to one of our destinations,
2351 // fold the conditions into logical ops and one cond br.
2352 BasicBlock::iterator BBI = BB->begin();
2353 // Ignore dbg intrinsics.
2354 while (isa<DbgInfoIntrinsic>(BBI))
2360 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2365 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2367 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2368 PBIOp = 0, BIOp = 1;
2369 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2370 PBIOp = 1, BIOp = 0;
2371 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2376 // Check to make sure that the other destination of this branch
2377 // isn't BB itself. If so, this is an infinite loop that will
2378 // keep getting unwound.
2379 if (PBI->getSuccessor(PBIOp) == BB)
2382 // Do not perform this transformation if it would require
2383 // insertion of a large number of select instructions. For targets
2384 // without predication/cmovs, this is a big pessimization.
2386 // Also do not perform this transformation if any phi node in the common
2387 // destination block can trap when reached by BB or PBB (PR17073). In that
2388 // case, it would be unsafe to hoist the operation into a select instruction.
2390 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2391 unsigned NumPhis = 0;
2392 for (BasicBlock::iterator II = CommonDest->begin();
2393 isa<PHINode>(II); ++II, ++NumPhis) {
2394 if (NumPhis > 2) // Disable this xform.
2397 PHINode *PN = cast<PHINode>(II);
2398 Value *BIV = PN->getIncomingValueForBlock(BB);
2399 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
2403 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2404 Value *PBIV = PN->getIncomingValue(PBBIdx);
2405 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
2410 // Finally, if everything is ok, fold the branches to logical ops.
2411 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2413 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2414 << "AND: " << *BI->getParent());
2417 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2418 // branch in it, where one edge (OtherDest) goes back to itself but the other
2419 // exits. We don't *know* that the program avoids the infinite loop
2420 // (even though that seems likely). If we do this xform naively, we'll end up
2421 // recursively unpeeling the loop. Since we know that (after the xform is
2422 // done) that the block *is* infinite if reached, we just make it an obviously
2423 // infinite loop with no cond branch.
2424 if (OtherDest == BB) {
2425 // Insert it at the end of the function, because it's either code,
2426 // or it won't matter if it's hot. :)
2427 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2428 "infloop", BB->getParent());
2429 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2430 OtherDest = InfLoopBlock;
2433 DEBUG(dbgs() << *PBI->getParent()->getParent());
2435 // BI may have other predecessors. Because of this, we leave
2436 // it alone, but modify PBI.
2438 // Make sure we get to CommonDest on True&True directions.
2439 Value *PBICond = PBI->getCondition();
2440 IRBuilder<true, NoFolder> Builder(PBI);
2442 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2444 Value *BICond = BI->getCondition();
2446 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2448 // Merge the conditions.
2449 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2451 // Modify PBI to branch on the new condition to the new dests.
2452 PBI->setCondition(Cond);
2453 PBI->setSuccessor(0, CommonDest);
2454 PBI->setSuccessor(1, OtherDest);
2456 // Update branch weight for PBI.
2457 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2458 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2460 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2462 if (PredHasWeights && SuccHasWeights) {
2463 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2464 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2465 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2466 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2467 // The weight to CommonDest should be PredCommon * SuccTotal +
2468 // PredOther * SuccCommon.
2469 // The weight to OtherDest should be PredOther * SuccOther.
2470 SmallVector<uint64_t, 2> NewWeights;
2471 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) +
2472 PredOther * SuccCommon);
2473 NewWeights.push_back(PredOther * SuccOther);
2474 // Halve the weights if any of them cannot fit in an uint32_t
2475 FitWeights(NewWeights);
2477 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end());
2478 PBI->setMetadata(LLVMContext::MD_prof,
2479 MDBuilder(BI->getContext()).
2480 createBranchWeights(MDWeights));
2483 // OtherDest may have phi nodes. If so, add an entry from PBI's
2484 // block that are identical to the entries for BI's block.
2485 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2487 // We know that the CommonDest already had an edge from PBI to
2488 // it. If it has PHIs though, the PHIs may have different
2489 // entries for BB and PBI's BB. If so, insert a select to make
2492 for (BasicBlock::iterator II = CommonDest->begin();
2493 (PN = dyn_cast<PHINode>(II)); ++II) {
2494 Value *BIV = PN->getIncomingValueForBlock(BB);
2495 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2496 Value *PBIV = PN->getIncomingValue(PBBIdx);
2498 // Insert a select in PBI to pick the right value.
2499 Value *NV = cast<SelectInst>
2500 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2501 PN->setIncomingValue(PBBIdx, NV);
2505 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2506 DEBUG(dbgs() << *PBI->getParent()->getParent());
2508 // This basic block is probably dead. We know it has at least
2509 // one fewer predecessor.
2513 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2514 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2515 // Takes care of updating the successors and removing the old terminator.
2516 // Also makes sure not to introduce new successors by assuming that edges to
2517 // non-successor TrueBBs and FalseBBs aren't reachable.
2518 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2519 BasicBlock *TrueBB, BasicBlock *FalseBB,
2520 uint32_t TrueWeight,
2521 uint32_t FalseWeight){
2522 // Remove any superfluous successor edges from the CFG.
2523 // First, figure out which successors to preserve.
2524 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2526 BasicBlock *KeepEdge1 = TrueBB;
2527 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
2529 // Then remove the rest.
2530 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2531 BasicBlock *Succ = OldTerm->getSuccessor(I);
2532 // Make sure only to keep exactly one copy of each edge.
2533 if (Succ == KeepEdge1)
2534 KeepEdge1 = nullptr;
2535 else if (Succ == KeepEdge2)
2536 KeepEdge2 = nullptr;
2538 Succ->removePredecessor(OldTerm->getParent());
2541 IRBuilder<> Builder(OldTerm);
2542 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2544 // Insert an appropriate new terminator.
2545 if (!KeepEdge1 && !KeepEdge2) {
2546 if (TrueBB == FalseBB)
2547 // We were only looking for one successor, and it was present.
2548 // Create an unconditional branch to it.
2549 Builder.CreateBr(TrueBB);
2551 // We found both of the successors we were looking for.
2552 // Create a conditional branch sharing the condition of the select.
2553 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2554 if (TrueWeight != FalseWeight)
2555 NewBI->setMetadata(LLVMContext::MD_prof,
2556 MDBuilder(OldTerm->getContext()).
2557 createBranchWeights(TrueWeight, FalseWeight));
2559 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2560 // Neither of the selected blocks were successors, so this
2561 // terminator must be unreachable.
2562 new UnreachableInst(OldTerm->getContext(), OldTerm);
2564 // One of the selected values was a successor, but the other wasn't.
2565 // Insert an unconditional branch to the one that was found;
2566 // the edge to the one that wasn't must be unreachable.
2568 // Only TrueBB was found.
2569 Builder.CreateBr(TrueBB);
2571 // Only FalseBB was found.
2572 Builder.CreateBr(FalseBB);
2575 EraseTerminatorInstAndDCECond(OldTerm);
2579 // SimplifySwitchOnSelect - Replaces
2580 // (switch (select cond, X, Y)) on constant X, Y
2581 // with a branch - conditional if X and Y lead to distinct BBs,
2582 // unconditional otherwise.
2583 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2584 // Check for constant integer values in the select.
2585 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2586 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2587 if (!TrueVal || !FalseVal)
2590 // Find the relevant condition and destinations.
2591 Value *Condition = Select->getCondition();
2592 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2593 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2595 // Get weight for TrueBB and FalseBB.
2596 uint32_t TrueWeight = 0, FalseWeight = 0;
2597 SmallVector<uint64_t, 8> Weights;
2598 bool HasWeights = HasBranchWeights(SI);
2600 GetBranchWeights(SI, Weights);
2601 if (Weights.size() == 1 + SI->getNumCases()) {
2602 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2603 getSuccessorIndex()];
2604 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2605 getSuccessorIndex()];
2609 // Perform the actual simplification.
2610 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2611 TrueWeight, FalseWeight);
2614 // SimplifyIndirectBrOnSelect - Replaces
2615 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2616 // blockaddress(@fn, BlockB)))
2618 // (br cond, BlockA, BlockB).
2619 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2620 // Check that both operands of the select are block addresses.
2621 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2622 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2626 // Extract the actual blocks.
2627 BasicBlock *TrueBB = TBA->getBasicBlock();
2628 BasicBlock *FalseBB = FBA->getBasicBlock();
2630 // Perform the actual simplification.
2631 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2635 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2636 /// instruction (a seteq/setne with a constant) as the only instruction in a
2637 /// block that ends with an uncond branch. We are looking for a very specific
2638 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2639 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2640 /// default value goes to an uncond block with a seteq in it, we get something
2643 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2645 /// %tmp = icmp eq i8 %A, 92
2648 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2650 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2651 /// the PHI, merging the third icmp into the switch.
2652 static bool TryToSimplifyUncondBranchWithICmpInIt(
2653 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI,
2654 unsigned BonusInstThreshold, const DataLayout *DL, AssumptionTracker *AT) {
2655 BasicBlock *BB = ICI->getParent();
2657 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2659 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2661 Value *V = ICI->getOperand(0);
2662 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2664 // The pattern we're looking for is where our only predecessor is a switch on
2665 // 'V' and this block is the default case for the switch. In this case we can
2666 // fold the compared value into the switch to simplify things.
2667 BasicBlock *Pred = BB->getSinglePredecessor();
2668 if (!Pred || !isa<SwitchInst>(Pred->getTerminator())) return false;
2670 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2671 if (SI->getCondition() != V)
2674 // If BB is reachable on a non-default case, then we simply know the value of
2675 // V in this block. Substitute it and constant fold the icmp instruction
2677 if (SI->getDefaultDest() != BB) {
2678 ConstantInt *VVal = SI->findCaseDest(BB);
2679 assert(VVal && "Should have a unique destination value");
2680 ICI->setOperand(0, VVal);
2682 if (Value *V = SimplifyInstruction(ICI, DL)) {
2683 ICI->replaceAllUsesWith(V);
2684 ICI->eraseFromParent();
2686 // BB is now empty, so it is likely to simplify away.
2687 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
2690 // Ok, the block is reachable from the default dest. If the constant we're
2691 // comparing exists in one of the other edges, then we can constant fold ICI
2693 if (SI->findCaseValue(Cst) != SI->case_default()) {
2695 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2696 V = ConstantInt::getFalse(BB->getContext());
2698 V = ConstantInt::getTrue(BB->getContext());
2700 ICI->replaceAllUsesWith(V);
2701 ICI->eraseFromParent();
2702 // BB is now empty, so it is likely to simplify away.
2703 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
2706 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2708 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2709 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
2710 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
2711 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2714 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2716 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2717 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2719 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2720 std::swap(DefaultCst, NewCst);
2722 // Replace ICI (which is used by the PHI for the default value) with true or
2723 // false depending on if it is EQ or NE.
2724 ICI->replaceAllUsesWith(DefaultCst);
2725 ICI->eraseFromParent();
2727 // Okay, the switch goes to this block on a default value. Add an edge from
2728 // the switch to the merge point on the compared value.
2729 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2730 BB->getParent(), BB);
2731 SmallVector<uint64_t, 8> Weights;
2732 bool HasWeights = HasBranchWeights(SI);
2734 GetBranchWeights(SI, Weights);
2735 if (Weights.size() == 1 + SI->getNumCases()) {
2736 // Split weight for default case to case for "Cst".
2737 Weights[0] = (Weights[0]+1) >> 1;
2738 Weights.push_back(Weights[0]);
2740 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2741 SI->setMetadata(LLVMContext::MD_prof,
2742 MDBuilder(SI->getContext()).
2743 createBranchWeights(MDWeights));
2746 SI->addCase(Cst, NewBB);
2748 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2749 Builder.SetInsertPoint(NewBB);
2750 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2751 Builder.CreateBr(SuccBlock);
2752 PHIUse->addIncoming(NewCst, NewBB);
2756 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2757 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2758 /// fold it into a switch instruction if so.
2759 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *DL,
2760 IRBuilder<> &Builder) {
2761 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2762 if (!Cond) return false;
2765 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2766 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2767 // 'setne's and'ed together, collect them.
2768 Value *CompVal = nullptr;
2769 std::vector<ConstantInt*> Values;
2770 bool TrueWhenEqual = true;
2771 Value *ExtraCase = nullptr;
2772 unsigned UsedICmps = 0;
2774 if (Cond->getOpcode() == Instruction::Or) {
2775 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, DL, true,
2777 } else if (Cond->getOpcode() == Instruction::And) {
2778 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, DL, false,
2780 TrueWhenEqual = false;
2783 // If we didn't have a multiply compared value, fail.
2784 if (!CompVal) return false;
2786 // Avoid turning single icmps into a switch.
2790 // There might be duplicate constants in the list, which the switch
2791 // instruction can't handle, remove them now.
2792 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2793 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2795 // If Extra was used, we require at least two switch values to do the
2796 // transformation. A switch with one value is just an cond branch.
2797 if (ExtraCase && Values.size() < 2) return false;
2799 // TODO: Preserve branch weight metadata, similarly to how
2800 // FoldValueComparisonIntoPredecessors preserves it.
2802 // Figure out which block is which destination.
2803 BasicBlock *DefaultBB = BI->getSuccessor(1);
2804 BasicBlock *EdgeBB = BI->getSuccessor(0);
2805 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2807 BasicBlock *BB = BI->getParent();
2809 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2810 << " cases into SWITCH. BB is:\n" << *BB);
2812 // If there are any extra values that couldn't be folded into the switch
2813 // then we evaluate them with an explicit branch first. Split the block
2814 // right before the condbr to handle it.
2816 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2817 // Remove the uncond branch added to the old block.
2818 TerminatorInst *OldTI = BB->getTerminator();
2819 Builder.SetInsertPoint(OldTI);
2822 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2824 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2826 OldTI->eraseFromParent();
2828 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2829 // for the edge we just added.
2830 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2832 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2833 << "\nEXTRABB = " << *BB);
2837 Builder.SetInsertPoint(BI);
2838 // Convert pointer to int before we switch.
2839 if (CompVal->getType()->isPointerTy()) {
2840 assert(DL && "Cannot switch on pointer without DataLayout");
2841 CompVal = Builder.CreatePtrToInt(CompVal,
2842 DL->getIntPtrType(CompVal->getType()),
2846 // Create the new switch instruction now.
2847 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2849 // Add all of the 'cases' to the switch instruction.
2850 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2851 New->addCase(Values[i], EdgeBB);
2853 // We added edges from PI to the EdgeBB. As such, if there were any
2854 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2855 // the number of edges added.
2856 for (BasicBlock::iterator BBI = EdgeBB->begin();
2857 isa<PHINode>(BBI); ++BBI) {
2858 PHINode *PN = cast<PHINode>(BBI);
2859 Value *InVal = PN->getIncomingValueForBlock(BB);
2860 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2861 PN->addIncoming(InVal, BB);
2864 // Erase the old branch instruction.
2865 EraseTerminatorInstAndDCECond(BI);
2867 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2871 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2872 // If this is a trivial landing pad that just continues unwinding the caught
2873 // exception then zap the landing pad, turning its invokes into calls.
2874 BasicBlock *BB = RI->getParent();
2875 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2876 if (RI->getValue() != LPInst)
2877 // Not a landing pad, or the resume is not unwinding the exception that
2878 // caused control to branch here.
2881 // Check that there are no other instructions except for debug intrinsics.
2882 BasicBlock::iterator I = LPInst, E = RI;
2884 if (!isa<DbgInfoIntrinsic>(I))
2887 // Turn all invokes that unwind here into calls and delete the basic block.
2888 bool InvokeRequiresTableEntry = false;
2889 bool Changed = false;
2890 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2891 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2893 if (II->hasFnAttr(Attribute::UWTable)) {
2894 // Don't remove an `invoke' instruction if the ABI requires an entry into
2896 InvokeRequiresTableEntry = true;
2900 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2902 // Insert a call instruction before the invoke.
2903 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2905 Call->setCallingConv(II->getCallingConv());
2906 Call->setAttributes(II->getAttributes());
2907 Call->setDebugLoc(II->getDebugLoc());
2909 // Anything that used the value produced by the invoke instruction now uses
2910 // the value produced by the call instruction. Note that we do this even
2911 // for void functions and calls with no uses so that the callgraph edge is
2913 II->replaceAllUsesWith(Call);
2914 BB->removePredecessor(II->getParent());
2916 // Insert a branch to the normal destination right before the invoke.
2917 BranchInst::Create(II->getNormalDest(), II);
2919 // Finally, delete the invoke instruction!
2920 II->eraseFromParent();
2924 if (!InvokeRequiresTableEntry)
2925 // The landingpad is now unreachable. Zap it.
2926 BB->eraseFromParent();
2931 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2932 BasicBlock *BB = RI->getParent();
2933 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2935 // Find predecessors that end with branches.
2936 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2937 SmallVector<BranchInst*, 8> CondBranchPreds;
2938 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2939 BasicBlock *P = *PI;
2940 TerminatorInst *PTI = P->getTerminator();
2941 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2942 if (BI->isUnconditional())
2943 UncondBranchPreds.push_back(P);
2945 CondBranchPreds.push_back(BI);
2949 // If we found some, do the transformation!
2950 if (!UncondBranchPreds.empty() && DupRet) {
2951 while (!UncondBranchPreds.empty()) {
2952 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
2953 DEBUG(dbgs() << "FOLDING: " << *BB
2954 << "INTO UNCOND BRANCH PRED: " << *Pred);
2955 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
2958 // If we eliminated all predecessors of the block, delete the block now.
2959 if (pred_begin(BB) == pred_end(BB))
2960 // We know there are no successors, so just nuke the block.
2961 BB->eraseFromParent();
2966 // Check out all of the conditional branches going to this return
2967 // instruction. If any of them just select between returns, change the
2968 // branch itself into a select/return pair.
2969 while (!CondBranchPreds.empty()) {
2970 BranchInst *BI = CondBranchPreds.pop_back_val();
2972 // Check to see if the non-BB successor is also a return block.
2973 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
2974 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
2975 SimplifyCondBranchToTwoReturns(BI, Builder))
2981 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
2982 BasicBlock *BB = UI->getParent();
2984 bool Changed = false;
2986 // If there are any instructions immediately before the unreachable that can
2987 // be removed, do so.
2988 while (UI != BB->begin()) {
2989 BasicBlock::iterator BBI = UI;
2991 // Do not delete instructions that can have side effects which might cause
2992 // the unreachable to not be reachable; specifically, calls and volatile
2993 // operations may have this effect.
2994 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
2996 if (BBI->mayHaveSideEffects()) {
2997 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
2998 if (SI->isVolatile())
3000 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
3001 if (LI->isVolatile())
3003 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
3004 if (RMWI->isVolatile())
3006 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
3007 if (CXI->isVolatile())
3009 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
3010 !isa<LandingPadInst>(BBI)) {
3013 // Note that deleting LandingPad's here is in fact okay, although it
3014 // involves a bit of subtle reasoning. If this inst is a LandingPad,
3015 // all the predecessors of this block will be the unwind edges of Invokes,
3016 // and we can therefore guarantee this block will be erased.
3019 // Delete this instruction (any uses are guaranteed to be dead)
3020 if (!BBI->use_empty())
3021 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
3022 BBI->eraseFromParent();
3026 // If the unreachable instruction is the first in the block, take a gander
3027 // at all of the predecessors of this instruction, and simplify them.
3028 if (&BB->front() != UI) return Changed;
3030 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
3031 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
3032 TerminatorInst *TI = Preds[i]->getTerminator();
3033 IRBuilder<> Builder(TI);
3034 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3035 if (BI->isUnconditional()) {
3036 if (BI->getSuccessor(0) == BB) {
3037 new UnreachableInst(TI->getContext(), TI);
3038 TI->eraseFromParent();
3042 if (BI->getSuccessor(0) == BB) {
3043 Builder.CreateBr(BI->getSuccessor(1));
3044 EraseTerminatorInstAndDCECond(BI);
3045 } else if (BI->getSuccessor(1) == BB) {
3046 Builder.CreateBr(BI->getSuccessor(0));
3047 EraseTerminatorInstAndDCECond(BI);
3051 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3052 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3054 if (i.getCaseSuccessor() == BB) {
3055 BB->removePredecessor(SI->getParent());
3060 // If the default value is unreachable, figure out the most popular
3061 // destination and make it the default.
3062 if (SI->getDefaultDest() == BB) {
3063 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
3064 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3066 std::pair<unsigned, unsigned> &entry =
3067 Popularity[i.getCaseSuccessor()];
3068 if (entry.first == 0) {
3070 entry.second = i.getCaseIndex();
3076 // Find the most popular block.
3077 unsigned MaxPop = 0;
3078 unsigned MaxIndex = 0;
3079 BasicBlock *MaxBlock = nullptr;
3080 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
3081 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
3082 if (I->second.first > MaxPop ||
3083 (I->second.first == MaxPop && MaxIndex > I->second.second)) {
3084 MaxPop = I->second.first;
3085 MaxIndex = I->second.second;
3086 MaxBlock = I->first;
3090 // Make this the new default, allowing us to delete any explicit
3092 SI->setDefaultDest(MaxBlock);
3095 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
3097 if (isa<PHINode>(MaxBlock->begin()))
3098 for (unsigned i = 0; i != MaxPop-1; ++i)
3099 MaxBlock->removePredecessor(SI->getParent());
3101 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3103 if (i.getCaseSuccessor() == MaxBlock) {
3109 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
3110 if (II->getUnwindDest() == BB) {
3111 // Convert the invoke to a call instruction. This would be a good
3112 // place to note that the call does not throw though.
3113 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
3114 II->removeFromParent(); // Take out of symbol table
3116 // Insert the call now...
3117 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
3118 Builder.SetInsertPoint(BI);
3119 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
3120 Args, II->getName());
3121 CI->setCallingConv(II->getCallingConv());
3122 CI->setAttributes(II->getAttributes());
3123 // If the invoke produced a value, the call does now instead.
3124 II->replaceAllUsesWith(CI);
3131 // If this block is now dead, remove it.
3132 if (pred_begin(BB) == pred_end(BB) &&
3133 BB != &BB->getParent()->getEntryBlock()) {
3134 // We know there are no successors, so just nuke the block.
3135 BB->eraseFromParent();
3142 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
3143 /// integer range comparison into a sub, an icmp and a branch.
3144 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3145 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3147 // Make sure all cases point to the same destination and gather the values.
3148 SmallVector<ConstantInt *, 16> Cases;
3149 SwitchInst::CaseIt I = SI->case_begin();
3150 Cases.push_back(I.getCaseValue());
3151 SwitchInst::CaseIt PrevI = I++;
3152 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
3153 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
3155 Cases.push_back(I.getCaseValue());
3157 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
3159 // Sort the case values, then check if they form a range we can transform.
3160 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3161 for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
3162 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
3166 Constant *Offset = ConstantExpr::getNeg(Cases.back());
3167 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
3169 Value *Sub = SI->getCondition();
3170 if (!Offset->isNullValue())
3171 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
3173 // If NumCases overflowed, then all possible values jump to the successor.
3174 if (NumCases->isNullValue() && SI->getNumCases() != 0)
3175 Cmp = ConstantInt::getTrue(SI->getContext());
3177 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3178 BranchInst *NewBI = Builder.CreateCondBr(
3179 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
3181 // Update weight for the newly-created conditional branch.
3182 SmallVector<uint64_t, 8> Weights;
3183 bool HasWeights = HasBranchWeights(SI);
3185 GetBranchWeights(SI, Weights);
3186 if (Weights.size() == 1 + SI->getNumCases()) {
3187 // Combine all weights for the cases to be the true weight of NewBI.
3188 // We assume that the sum of all weights for a Terminator can fit into 32
3190 uint32_t NewTrueWeight = 0;
3191 for (unsigned I = 1, E = Weights.size(); I != E; ++I)
3192 NewTrueWeight += (uint32_t)Weights[I];
3193 NewBI->setMetadata(LLVMContext::MD_prof,
3194 MDBuilder(SI->getContext()).
3195 createBranchWeights(NewTrueWeight,
3196 (uint32_t)Weights[0]));
3200 // Prune obsolete incoming values off the successor's PHI nodes.
3201 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
3202 isa<PHINode>(BBI); ++BBI) {
3203 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
3204 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3206 SI->eraseFromParent();
3211 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
3212 /// and use it to remove dead cases.
3213 static bool EliminateDeadSwitchCases(SwitchInst *SI, const DataLayout *DL,
3214 AssumptionTracker *AT) {
3215 Value *Cond = SI->getCondition();
3216 unsigned Bits = Cond->getType()->getIntegerBitWidth();
3217 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3218 computeKnownBits(Cond, KnownZero, KnownOne, DL, 0, AT, SI);
3220 // Gather dead cases.
3221 SmallVector<ConstantInt*, 8> DeadCases;
3222 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3223 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3224 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3225 DeadCases.push_back(I.getCaseValue());
3226 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3227 << I.getCaseValue() << "' is dead.\n");
3231 SmallVector<uint64_t, 8> Weights;
3232 bool HasWeight = HasBranchWeights(SI);
3234 GetBranchWeights(SI, Weights);
3235 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3238 // Remove dead cases from the switch.
3239 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3240 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3241 assert(Case != SI->case_default() &&
3242 "Case was not found. Probably mistake in DeadCases forming.");
3244 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3248 // Prune unused values from PHI nodes.
3249 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3250 SI->removeCase(Case);
3252 if (HasWeight && Weights.size() >= 2) {
3253 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3254 SI->setMetadata(LLVMContext::MD_prof,
3255 MDBuilder(SI->getParent()->getContext()).
3256 createBranchWeights(MDWeights));
3259 return !DeadCases.empty();
3262 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
3263 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3264 /// by an unconditional branch), look at the phi node for BB in the successor
3265 /// block and see if the incoming value is equal to CaseValue. If so, return
3266 /// the phi node, and set PhiIndex to BB's index in the phi node.
3267 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3270 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3271 return nullptr; // BB must be empty to be a candidate for simplification.
3272 if (!BB->getSinglePredecessor())
3273 return nullptr; // BB must be dominated by the switch.
3275 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3276 if (!Branch || !Branch->isUnconditional())
3277 return nullptr; // Terminator must be unconditional branch.
3279 BasicBlock *Succ = Branch->getSuccessor(0);
3281 BasicBlock::iterator I = Succ->begin();
3282 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3283 int Idx = PHI->getBasicBlockIndex(BB);
3284 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3286 Value *InValue = PHI->getIncomingValue(Idx);
3287 if (InValue != CaseValue) continue;
3296 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
3297 /// instruction to a phi node dominated by the switch, if that would mean that
3298 /// some of the destination blocks of the switch can be folded away.
3299 /// Returns true if a change is made.
3300 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3301 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3302 ForwardingNodesMap ForwardingNodes;
3304 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3305 ConstantInt *CaseValue = I.getCaseValue();
3306 BasicBlock *CaseDest = I.getCaseSuccessor();
3309 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3313 ForwardingNodes[PHI].push_back(PhiIndex);
3316 bool Changed = false;
3318 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3319 E = ForwardingNodes.end(); I != E; ++I) {
3320 PHINode *Phi = I->first;
3321 SmallVectorImpl<int> &Indexes = I->second;
3323 if (Indexes.size() < 2) continue;
3325 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3326 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3333 /// ValidLookupTableConstant - Return true if the backend will be able to handle
3334 /// initializing an array of constants like C.
3335 static bool ValidLookupTableConstant(Constant *C) {
3336 if (C->isThreadDependent())
3338 if (C->isDLLImportDependent())
3341 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3342 return CE->isGEPWithNoNotionalOverIndexing();
3344 return isa<ConstantFP>(C) ||
3345 isa<ConstantInt>(C) ||
3346 isa<ConstantPointerNull>(C) ||
3347 isa<GlobalValue>(C) ||
3351 /// LookupConstant - If V is a Constant, return it. Otherwise, try to look up
3352 /// its constant value in ConstantPool, returning 0 if it's not there.
3353 static Constant *LookupConstant(Value *V,
3354 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3355 if (Constant *C = dyn_cast<Constant>(V))
3357 return ConstantPool.lookup(V);
3360 /// ConstantFold - Try to fold instruction I into a constant. This works for
3361 /// simple instructions such as binary operations where both operands are
3362 /// constant or can be replaced by constants from the ConstantPool. Returns the
3363 /// resulting constant on success, 0 otherwise.
3365 ConstantFold(Instruction *I,
3366 const SmallDenseMap<Value *, Constant *> &ConstantPool,
3367 const DataLayout *DL) {
3368 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3369 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3372 if (A->isAllOnesValue())
3373 return LookupConstant(Select->getTrueValue(), ConstantPool);
3374 if (A->isNullValue())
3375 return LookupConstant(Select->getFalseValue(), ConstantPool);
3379 SmallVector<Constant *, 4> COps;
3380 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
3381 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
3387 if (CmpInst *Cmp = dyn_cast<CmpInst>(I))
3388 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
3391 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), COps, DL);
3394 /// GetCaseResults - Try to determine the resulting constant values in phi nodes
3395 /// at the common destination basic block, *CommonDest, for one of the case
3396 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3397 /// case), of a switch instruction SI.
3399 GetCaseResults(SwitchInst *SI,
3400 ConstantInt *CaseVal,
3401 BasicBlock *CaseDest,
3402 BasicBlock **CommonDest,
3403 SmallVectorImpl<std::pair<PHINode *, Constant *> > &Res,
3404 const DataLayout *DL) {
3405 // The block from which we enter the common destination.
3406 BasicBlock *Pred = SI->getParent();
3408 // If CaseDest is empty except for some side-effect free instructions through
3409 // which we can constant-propagate the CaseVal, continue to its successor.
3410 SmallDenseMap<Value*, Constant*> ConstantPool;
3411 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3412 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3414 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3415 // If the terminator is a simple branch, continue to the next block.
3416 if (T->getNumSuccessors() != 1)
3419 CaseDest = T->getSuccessor(0);
3420 } else if (isa<DbgInfoIntrinsic>(I)) {
3421 // Skip debug intrinsic.
3423 } else if (Constant *C = ConstantFold(I, ConstantPool, DL)) {
3424 // Instruction is side-effect free and constant.
3425 ConstantPool.insert(std::make_pair(I, C));
3431 // If we did not have a CommonDest before, use the current one.
3433 *CommonDest = CaseDest;
3434 // If the destination isn't the common one, abort.
3435 if (CaseDest != *CommonDest)
3438 // Get the values for this case from phi nodes in the destination block.
3439 BasicBlock::iterator I = (*CommonDest)->begin();
3440 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3441 int Idx = PHI->getBasicBlockIndex(Pred);
3445 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3450 // Note: If the constant comes from constant-propagating the case value
3451 // through the CaseDest basic block, it will be safe to remove the
3452 // instructions in that block. They cannot be used (except in the phi nodes
3453 // we visit) outside CaseDest, because that block does not dominate its
3454 // successor. If it did, we would not be in this phi node.
3456 // Be conservative about which kinds of constants we support.
3457 if (!ValidLookupTableConstant(ConstVal))
3460 Res.push_back(std::make_pair(PHI, ConstVal));
3463 return Res.size() > 0;
3466 // MapCaseToResult - Helper function used to
3467 // add CaseVal to the list of cases that generate Result.
3468 static void MapCaseToResult(ConstantInt *CaseVal,
3469 SwitchCaseResultVectorTy &UniqueResults,
3471 for (auto &I : UniqueResults) {
3472 if (I.first == Result) {
3473 I.second.push_back(CaseVal);
3477 UniqueResults.push_back(std::make_pair(Result,
3478 SmallVector<ConstantInt*, 4>(1, CaseVal)));
3481 // InitializeUniqueCases - Helper function that initializes a map containing
3482 // results for the PHI node of the common destination block for a switch
3483 // instruction. Returns false if multiple PHI nodes have been found or if
3484 // there is not a common destination block for the switch.
3485 static bool InitializeUniqueCases(
3486 SwitchInst *SI, const DataLayout *DL, PHINode *&PHI,
3487 BasicBlock *&CommonDest,
3488 SwitchCaseResultVectorTy &UniqueResults,
3489 Constant *&DefaultResult) {
3490 for (auto &I : SI->cases()) {
3491 ConstantInt *CaseVal = I.getCaseValue();
3493 // Resulting value at phi nodes for this case value.
3494 SwitchCaseResultsTy Results;
3495 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
3499 // Only one value per case is permitted
3500 if (Results.size() > 1)
3502 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
3504 // Check the PHI consistency.
3506 PHI = Results[0].first;
3507 else if (PHI != Results[0].first)
3510 // Find the default result value.
3511 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
3512 BasicBlock *DefaultDest = SI->getDefaultDest();
3513 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
3515 // If the default value is not found abort unless the default destination
3518 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
3519 if ((!DefaultResult &&
3520 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
3526 // ConvertTwoCaseSwitch - Helper function that checks if it is possible to
3527 // transform a switch with only two cases (or two cases + default)
3528 // that produces a result into a value select.
3531 // case 10: %0 = icmp eq i32 %a, 10
3532 // return 10; %1 = select i1 %0, i32 10, i32 4
3533 // case 20: ----> %2 = icmp eq i32 %a, 20
3534 // return 2; %3 = select i1 %2, i32 2, i32 %1
3539 ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
3540 Constant *DefaultResult, Value *Condition,
3541 IRBuilder<> &Builder) {
3542 assert(ResultVector.size() == 2 &&
3543 "We should have exactly two unique results at this point");
3544 // If we are selecting between only two cases transform into a simple
3545 // select or a two-way select if default is possible.
3546 if (ResultVector[0].second.size() == 1 &&
3547 ResultVector[1].second.size() == 1) {
3548 ConstantInt *const FirstCase = ResultVector[0].second[0];
3549 ConstantInt *const SecondCase = ResultVector[1].second[0];
3551 bool DefaultCanTrigger = DefaultResult;
3552 Value *SelectValue = ResultVector[1].first;
3553 if (DefaultCanTrigger) {
3554 Value *const ValueCompare =
3555 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
3556 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
3557 DefaultResult, "switch.select");
3559 Value *const ValueCompare =
3560 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
3561 return Builder.CreateSelect(ValueCompare, ResultVector[0].first, SelectValue,
3568 // RemoveSwitchAfterSelectConversion - Helper function to cleanup a switch
3569 // instruction that has been converted into a select, fixing up PHI nodes and
3571 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
3573 IRBuilder<> &Builder) {
3574 BasicBlock *SelectBB = SI->getParent();
3575 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
3576 PHI->removeIncomingValue(SelectBB);
3577 PHI->addIncoming(SelectValue, SelectBB);
3579 Builder.CreateBr(PHI->getParent());
3581 // Remove the switch.
3582 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
3583 BasicBlock *Succ = SI->getSuccessor(i);
3585 if (Succ == PHI->getParent())
3587 Succ->removePredecessor(SelectBB);
3589 SI->eraseFromParent();
3592 /// SwitchToSelect - If the switch is only used to initialize one or more
3593 /// phi nodes in a common successor block with only two different
3594 /// constant values, replace the switch with select.
3595 static bool SwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
3596 const DataLayout *DL, AssumptionTracker *AT) {
3597 Value *const Cond = SI->getCondition();
3598 PHINode *PHI = nullptr;
3599 BasicBlock *CommonDest = nullptr;
3600 Constant *DefaultResult;
3601 SwitchCaseResultVectorTy UniqueResults;
3602 // Collect all the cases that will deliver the same value from the switch.
3603 if (!InitializeUniqueCases(SI, DL, PHI, CommonDest, UniqueResults,
3606 // Selects choose between maximum two values.
3607 if (UniqueResults.size() != 2)
3609 assert(PHI != nullptr && "PHI for value select not found");
3611 Builder.SetInsertPoint(SI);
3612 Value *SelectValue = ConvertTwoCaseSwitch(
3614 DefaultResult, Cond, Builder);
3616 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
3619 // The switch couldn't be converted into a select.
3624 /// SwitchLookupTable - This class represents a lookup table that can be used
3625 /// to replace a switch.
3626 class SwitchLookupTable {
3628 /// SwitchLookupTable - Create a lookup table to use as a switch replacement
3629 /// with the contents of Values, using DefaultValue to fill any holes in the
3631 SwitchLookupTable(Module &M,
3633 ConstantInt *Offset,
3634 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3635 Constant *DefaultValue,
3636 const DataLayout *DL);
3638 /// BuildLookup - Build instructions with Builder to retrieve the value at
3639 /// the position given by Index in the lookup table.
3640 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3642 /// WouldFitInRegister - Return true if a table with TableSize elements of
3643 /// type ElementType would fit in a target-legal register.
3644 static bool WouldFitInRegister(const DataLayout *DL,
3646 const Type *ElementType);
3649 // Depending on the contents of the table, it can be represented in
3652 // For tables where each element contains the same value, we just have to
3653 // store that single value and return it for each lookup.
3656 // For small tables with integer elements, we can pack them into a bitmap
3657 // that fits into a target-legal register. Values are retrieved by
3658 // shift and mask operations.
3661 // The table is stored as an array of values. Values are retrieved by load
3662 // instructions from the table.
3666 // For SingleValueKind, this is the single value.
3667 Constant *SingleValue;
3669 // For BitMapKind, this is the bitmap.
3670 ConstantInt *BitMap;
3671 IntegerType *BitMapElementTy;
3673 // For ArrayKind, this is the array.
3674 GlobalVariable *Array;
3678 SwitchLookupTable::SwitchLookupTable(Module &M,
3680 ConstantInt *Offset,
3681 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3682 Constant *DefaultValue,
3683 const DataLayout *DL)
3684 : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
3686 assert(Values.size() && "Can't build lookup table without values!");
3687 assert(TableSize >= Values.size() && "Can't fit values in table!");
3689 // If all values in the table are equal, this is that value.
3690 SingleValue = Values.begin()->second;
3692 Type *ValueType = Values.begin()->second->getType();
3694 // Build up the table contents.
3695 SmallVector<Constant*, 64> TableContents(TableSize);
3696 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3697 ConstantInt *CaseVal = Values[I].first;
3698 Constant *CaseRes = Values[I].second;
3699 assert(CaseRes->getType() == ValueType);
3701 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3703 TableContents[Idx] = CaseRes;
3705 if (CaseRes != SingleValue)
3706 SingleValue = nullptr;
3709 // Fill in any holes in the table with the default result.
3710 if (Values.size() < TableSize) {
3711 assert(DefaultValue &&
3712 "Need a default value to fill the lookup table holes.");
3713 assert(DefaultValue->getType() == ValueType);
3714 for (uint64_t I = 0; I < TableSize; ++I) {
3715 if (!TableContents[I])
3716 TableContents[I] = DefaultValue;
3719 if (DefaultValue != SingleValue)
3720 SingleValue = nullptr;
3723 // If each element in the table contains the same value, we only need to store
3724 // that single value.
3726 Kind = SingleValueKind;
3730 // If the type is integer and the table fits in a register, build a bitmap.
3731 if (WouldFitInRegister(DL, TableSize, ValueType)) {
3732 IntegerType *IT = cast<IntegerType>(ValueType);
3733 APInt TableInt(TableSize * IT->getBitWidth(), 0);
3734 for (uint64_t I = TableSize; I > 0; --I) {
3735 TableInt <<= IT->getBitWidth();
3736 // Insert values into the bitmap. Undef values are set to zero.
3737 if (!isa<UndefValue>(TableContents[I - 1])) {
3738 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
3739 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
3742 BitMap = ConstantInt::get(M.getContext(), TableInt);
3743 BitMapElementTy = IT;
3749 // Store the table in an array.
3750 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
3751 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3753 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3754 GlobalVariable::PrivateLinkage,
3757 Array->setUnnamedAddr(true);
3761 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
3763 case SingleValueKind:
3766 // Type of the bitmap (e.g. i59).
3767 IntegerType *MapTy = BitMap->getType();
3769 // Cast Index to the same type as the bitmap.
3770 // Note: The Index is <= the number of elements in the table, so
3771 // truncating it to the width of the bitmask is safe.
3772 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
3774 // Multiply the shift amount by the element width.
3775 ShiftAmt = Builder.CreateMul(ShiftAmt,
3776 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
3780 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
3781 "switch.downshift");
3783 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
3787 // Make sure the table index will not overflow when treated as signed.
3788 IntegerType *IT = cast<IntegerType>(Index->getType());
3789 uint64_t TableSize = Array->getInitializer()->getType()
3790 ->getArrayNumElements();
3791 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
3792 Index = Builder.CreateZExt(Index,
3793 IntegerType::get(IT->getContext(),
3794 IT->getBitWidth() + 1),
3795 "switch.tableidx.zext");
3797 Value *GEPIndices[] = { Builder.getInt32(0), Index };
3798 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices,
3800 return Builder.CreateLoad(GEP, "switch.load");
3803 llvm_unreachable("Unknown lookup table kind!");
3806 bool SwitchLookupTable::WouldFitInRegister(const DataLayout *DL,
3808 const Type *ElementType) {
3811 const IntegerType *IT = dyn_cast<IntegerType>(ElementType);
3814 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
3815 // are <= 15, we could try to narrow the type.
3817 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
3818 if (TableSize >= UINT_MAX/IT->getBitWidth())
3820 return DL->fitsInLegalInteger(TableSize * IT->getBitWidth());
3823 /// ShouldBuildLookupTable - Determine whether a lookup table should be built
3824 /// for this switch, based on the number of cases, size of the table and the
3825 /// types of the results.
3826 static bool ShouldBuildLookupTable(SwitchInst *SI,
3828 const TargetTransformInfo &TTI,
3829 const DataLayout *DL,
3830 const SmallDenseMap<PHINode*, Type*>& ResultTypes) {
3831 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
3832 return false; // TableSize overflowed, or mul below might overflow.
3834 bool AllTablesFitInRegister = true;
3835 bool HasIllegalType = false;
3836 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(),
3837 E = ResultTypes.end(); I != E; ++I) {
3838 Type *Ty = I->second;
3840 // Saturate this flag to true.
3841 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
3843 // Saturate this flag to false.
3844 AllTablesFitInRegister = AllTablesFitInRegister &&
3845 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
3847 // If both flags saturate, we're done. NOTE: This *only* works with
3848 // saturating flags, and all flags have to saturate first due to the
3849 // non-deterministic behavior of iterating over a dense map.
3850 if (HasIllegalType && !AllTablesFitInRegister)
3854 // If each table would fit in a register, we should build it anyway.
3855 if (AllTablesFitInRegister)
3858 // Don't build a table that doesn't fit in-register if it has illegal types.
3862 // The table density should be at least 40%. This is the same criterion as for
3863 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3864 // FIXME: Find the best cut-off.
3865 return SI->getNumCases() * 10 >= TableSize * 4;
3868 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3869 /// phi nodes in a common successor block with different constant values,
3870 /// replace the switch with lookup tables.
3871 static bool SwitchToLookupTable(SwitchInst *SI,
3872 IRBuilder<> &Builder,
3873 const TargetTransformInfo &TTI,
3874 const DataLayout* DL) {
3875 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3877 // Only build lookup table when we have a target that supports it.
3878 if (!TTI.shouldBuildLookupTables())
3881 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3882 // split off a dense part and build a lookup table for that.
3884 // FIXME: This creates arrays of GEPs to constant strings, which means each
3885 // GEP needs a runtime relocation in PIC code. We should just build one big
3886 // string and lookup indices into that.
3888 // Ignore switches with less than three cases. Lookup tables will not make them
3889 // faster, so we don't analyze them.
3890 if (SI->getNumCases() < 3)
3893 // Figure out the corresponding result for each case value and phi node in the
3894 // common destination, as well as the the min and max case values.
3895 assert(SI->case_begin() != SI->case_end());
3896 SwitchInst::CaseIt CI = SI->case_begin();
3897 ConstantInt *MinCaseVal = CI.getCaseValue();
3898 ConstantInt *MaxCaseVal = CI.getCaseValue();
3900 BasicBlock *CommonDest = nullptr;
3901 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
3902 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
3903 SmallDenseMap<PHINode*, Constant*> DefaultResults;
3904 SmallDenseMap<PHINode*, Type*> ResultTypes;
3905 SmallVector<PHINode*, 4> PHIs;
3907 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
3908 ConstantInt *CaseVal = CI.getCaseValue();
3909 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
3910 MinCaseVal = CaseVal;
3911 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
3912 MaxCaseVal = CaseVal;
3914 // Resulting value at phi nodes for this case value.
3915 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
3917 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
3921 // Append the result from this case to the list for each phi.
3922 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) {
3923 if (!ResultLists.count(I->first))
3924 PHIs.push_back(I->first);
3925 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second));
3929 // Keep track of the result types.
3930 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
3931 PHINode *PHI = PHIs[I];
3932 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
3935 uint64_t NumResults = ResultLists[PHIs[0]].size();
3936 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
3937 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
3938 bool TableHasHoles = (NumResults < TableSize);
3940 // If the table has holes, we need a constant result for the default case
3941 // or a bitmask that fits in a register.
3942 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
3943 bool HasDefaultResults = false;
3944 if (TableHasHoles) {
3945 HasDefaultResults = GetCaseResults(SI, nullptr, SI->getDefaultDest(),
3946 &CommonDest, DefaultResultsList, DL);
3948 bool NeedMask = (TableHasHoles && !HasDefaultResults);
3950 // As an extra penalty for the validity test we require more cases.
3951 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
3953 if (!(DL && DL->fitsInLegalInteger(TableSize)))
3957 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) {
3958 PHINode *PHI = DefaultResultsList[I].first;
3959 Constant *Result = DefaultResultsList[I].second;
3960 DefaultResults[PHI] = Result;
3963 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
3966 // Create the BB that does the lookups.
3967 Module &Mod = *CommonDest->getParent()->getParent();
3968 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
3970 CommonDest->getParent(),
3973 // Compute the table index value.
3974 Builder.SetInsertPoint(SI);
3975 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
3978 // Compute the maximum table size representable by the integer type we are
3980 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
3981 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
3982 assert(MaxTableSize >= TableSize &&
3983 "It is impossible for a switch to have more entries than the max "
3984 "representable value of its input integer type's size.");
3986 // If we have a fully covered lookup table, unconditionally branch to the
3987 // lookup table BB. Otherwise, check if the condition value is within the case
3988 // range. If it is so, branch to the new BB. Otherwise branch to SI's default
3990 const bool GeneratingCoveredLookupTable = MaxTableSize == TableSize;
3991 if (GeneratingCoveredLookupTable) {
3992 Builder.CreateBr(LookupBB);
3993 // We cached PHINodes in PHIs, to avoid accessing deleted PHINodes later,
3994 // do not delete PHINodes here.
3995 SI->getDefaultDest()->removePredecessor(SI->getParent(),
3996 true/*DontDeleteUselessPHIs*/);
3998 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
3999 MinCaseVal->getType(), TableSize));
4000 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
4003 // Populate the BB that does the lookups.
4004 Builder.SetInsertPoint(LookupBB);
4007 // Before doing the lookup we do the hole check.
4008 // The LookupBB is therefore re-purposed to do the hole check
4009 // and we create a new LookupBB.
4010 BasicBlock *MaskBB = LookupBB;
4011 MaskBB->setName("switch.hole_check");
4012 LookupBB = BasicBlock::Create(Mod.getContext(),
4014 CommonDest->getParent(),
4017 // Build bitmask; fill in a 1 bit for every case.
4018 APInt MaskInt(TableSize, 0);
4019 APInt One(TableSize, 1);
4020 const ResultListTy &ResultList = ResultLists[PHIs[0]];
4021 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
4022 uint64_t Idx = (ResultList[I].first->getValue() -
4023 MinCaseVal->getValue()).getLimitedValue();
4024 MaskInt |= One << Idx;
4026 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
4028 // Get the TableIndex'th bit of the bitmask.
4029 // If this bit is 0 (meaning hole) jump to the default destination,
4030 // else continue with table lookup.
4031 IntegerType *MapTy = TableMask->getType();
4032 Value *MaskIndex = Builder.CreateZExtOrTrunc(TableIndex, MapTy,
4033 "switch.maskindex");
4034 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex,
4036 Value *LoBit = Builder.CreateTrunc(Shifted,
4037 Type::getInt1Ty(Mod.getContext()),
4039 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
4041 Builder.SetInsertPoint(LookupBB);
4042 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
4045 bool ReturnedEarly = false;
4046 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
4047 PHINode *PHI = PHIs[I];
4049 // If using a bitmask, use any value to fill the lookup table holes.
4050 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
4051 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI],
4054 Value *Result = Table.BuildLookup(TableIndex, Builder);
4056 // If the result is used to return immediately from the function, we want to
4057 // do that right here.
4058 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
4059 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
4060 Builder.CreateRet(Result);
4061 ReturnedEarly = true;
4065 PHI->addIncoming(Result, LookupBB);
4069 Builder.CreateBr(CommonDest);
4071 // Remove the switch.
4072 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4073 BasicBlock *Succ = SI->getSuccessor(i);
4075 if (Succ == SI->getDefaultDest())
4077 Succ->removePredecessor(SI->getParent());
4079 SI->eraseFromParent();
4083 ++NumLookupTablesHoles;
4087 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
4088 BasicBlock *BB = SI->getParent();
4090 if (isValueEqualityComparison(SI)) {
4091 // If we only have one predecessor, and if it is a branch on this value,
4092 // see if that predecessor totally determines the outcome of this switch.
4093 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4094 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
4095 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4097 Value *Cond = SI->getCondition();
4098 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
4099 if (SimplifySwitchOnSelect(SI, Select))
4100 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4102 // If the block only contains the switch, see if we can fold the block
4103 // away into any preds.
4104 BasicBlock::iterator BBI = BB->begin();
4105 // Ignore dbg intrinsics.
4106 while (isa<DbgInfoIntrinsic>(BBI))
4109 if (FoldValueComparisonIntoPredecessors(SI, Builder))
4110 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4113 // Try to transform the switch into an icmp and a branch.
4114 if (TurnSwitchRangeIntoICmp(SI, Builder))
4115 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4117 // Remove unreachable cases.
4118 if (EliminateDeadSwitchCases(SI, DL, AT))
4119 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4121 if (SwitchToSelect(SI, Builder, DL, AT))
4122 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4124 if (ForwardSwitchConditionToPHI(SI))
4125 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4127 if (SwitchToLookupTable(SI, Builder, TTI, DL))
4128 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4133 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
4134 BasicBlock *BB = IBI->getParent();
4135 bool Changed = false;
4137 // Eliminate redundant destinations.
4138 SmallPtrSet<Value *, 8> Succs;
4139 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
4140 BasicBlock *Dest = IBI->getDestination(i);
4141 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) {
4142 Dest->removePredecessor(BB);
4143 IBI->removeDestination(i);
4149 if (IBI->getNumDestinations() == 0) {
4150 // If the indirectbr has no successors, change it to unreachable.
4151 new UnreachableInst(IBI->getContext(), IBI);
4152 EraseTerminatorInstAndDCECond(IBI);
4156 if (IBI->getNumDestinations() == 1) {
4157 // If the indirectbr has one successor, change it to a direct branch.
4158 BranchInst::Create(IBI->getDestination(0), IBI);
4159 EraseTerminatorInstAndDCECond(IBI);
4163 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
4164 if (SimplifyIndirectBrOnSelect(IBI, SI))
4165 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4170 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
4171 BasicBlock *BB = BI->getParent();
4173 if (SinkCommon && SinkThenElseCodeToEnd(BI))
4176 // If the Terminator is the only non-phi instruction, simplify the block.
4177 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg();
4178 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
4179 TryToSimplifyUncondBranchFromEmptyBlock(BB))
4182 // If the only instruction in the block is a seteq/setne comparison
4183 // against a constant, try to simplify the block.
4184 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
4185 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
4186 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
4188 if (I->isTerminator() &&
4189 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI,
4190 BonusInstThreshold, DL, AT))
4194 // If this basic block is ONLY a compare and a branch, and if a predecessor
4195 // branches to us and our successor, fold the comparison into the
4196 // predecessor and use logical operations to update the incoming value
4197 // for PHI nodes in common successor.
4198 if (FoldBranchToCommonDest(BI, DL, BonusInstThreshold))
4199 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4204 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
4205 BasicBlock *BB = BI->getParent();
4207 // Conditional branch
4208 if (isValueEqualityComparison(BI)) {
4209 // If we only have one predecessor, and if it is a branch on this value,
4210 // see if that predecessor totally determines the outcome of this
4212 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4213 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
4214 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4216 // This block must be empty, except for the setcond inst, if it exists.
4217 // Ignore dbg intrinsics.
4218 BasicBlock::iterator I = BB->begin();
4219 // Ignore dbg intrinsics.
4220 while (isa<DbgInfoIntrinsic>(I))
4223 if (FoldValueComparisonIntoPredecessors(BI, Builder))
4224 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4225 } else if (&*I == cast<Instruction>(BI->getCondition())){
4227 // Ignore dbg intrinsics.
4228 while (isa<DbgInfoIntrinsic>(I))
4230 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
4231 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4235 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
4236 if (SimplifyBranchOnICmpChain(BI, DL, Builder))
4239 // If this basic block is ONLY a compare and a branch, and if a predecessor
4240 // branches to us and one of our successors, fold the comparison into the
4241 // predecessor and use logical operations to pick the right destination.
4242 if (FoldBranchToCommonDest(BI, DL, BonusInstThreshold))
4243 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4245 // We have a conditional branch to two blocks that are only reachable
4246 // from BI. We know that the condbr dominates the two blocks, so see if
4247 // there is any identical code in the "then" and "else" blocks. If so, we
4248 // can hoist it up to the branching block.
4249 if (BI->getSuccessor(0)->getSinglePredecessor()) {
4250 if (BI->getSuccessor(1)->getSinglePredecessor()) {
4251 if (HoistThenElseCodeToIf(BI, DL))
4252 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4254 // If Successor #1 has multiple preds, we may be able to conditionally
4255 // execute Successor #0 if it branches to Successor #1.
4256 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
4257 if (Succ0TI->getNumSuccessors() == 1 &&
4258 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
4259 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), DL))
4260 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4262 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
4263 // If Successor #0 has multiple preds, we may be able to conditionally
4264 // execute Successor #1 if it branches to Successor #0.
4265 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
4266 if (Succ1TI->getNumSuccessors() == 1 &&
4267 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
4268 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), DL))
4269 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4272 // If this is a branch on a phi node in the current block, thread control
4273 // through this block if any PHI node entries are constants.
4274 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
4275 if (PN->getParent() == BI->getParent())
4276 if (FoldCondBranchOnPHI(BI, DL))
4277 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4279 // Scan predecessor blocks for conditional branches.
4280 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
4281 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
4282 if (PBI != BI && PBI->isConditional())
4283 if (SimplifyCondBranchToCondBranch(PBI, BI))
4284 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4289 /// Check if passing a value to an instruction will cause undefined behavior.
4290 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
4291 Constant *C = dyn_cast<Constant>(V);
4298 if (C->isNullValue()) {
4299 // Only look at the first use, avoid hurting compile time with long uselists
4300 User *Use = *I->user_begin();
4302 // Now make sure that there are no instructions in between that can alter
4303 // control flow (eg. calls)
4304 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
4305 if (i == I->getParent()->end() || i->mayHaveSideEffects())
4308 // Look through GEPs. A load from a GEP derived from NULL is still undefined
4309 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
4310 if (GEP->getPointerOperand() == I)
4311 return passingValueIsAlwaysUndefined(V, GEP);
4313 // Look through bitcasts.
4314 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
4315 return passingValueIsAlwaysUndefined(V, BC);
4317 // Load from null is undefined.
4318 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
4319 if (!LI->isVolatile())
4320 return LI->getPointerAddressSpace() == 0;
4322 // Store to null is undefined.
4323 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
4324 if (!SI->isVolatile())
4325 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
4330 /// If BB has an incoming value that will always trigger undefined behavior
4331 /// (eg. null pointer dereference), remove the branch leading here.
4332 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
4333 for (BasicBlock::iterator i = BB->begin();
4334 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
4335 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
4336 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
4337 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
4338 IRBuilder<> Builder(T);
4339 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
4340 BB->removePredecessor(PHI->getIncomingBlock(i));
4341 // Turn uncoditional branches into unreachables and remove the dead
4342 // destination from conditional branches.
4343 if (BI->isUnconditional())
4344 Builder.CreateUnreachable();
4346 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
4347 BI->getSuccessor(0));
4348 BI->eraseFromParent();
4351 // TODO: SwitchInst.
4357 bool SimplifyCFGOpt::run(BasicBlock *BB) {
4358 bool Changed = false;
4360 assert(BB && BB->getParent() && "Block not embedded in function!");
4361 assert(BB->getTerminator() && "Degenerate basic block encountered!");
4363 // Remove basic blocks that have no predecessors (except the entry block)...
4364 // or that just have themself as a predecessor. These are unreachable.
4365 if ((pred_begin(BB) == pred_end(BB) &&
4366 BB != &BB->getParent()->getEntryBlock()) ||
4367 BB->getSinglePredecessor() == BB) {
4368 DEBUG(dbgs() << "Removing BB: \n" << *BB);
4369 DeleteDeadBlock(BB);
4373 // Check to see if we can constant propagate this terminator instruction
4375 Changed |= ConstantFoldTerminator(BB, true);
4377 // Check for and eliminate duplicate PHI nodes in this block.
4378 Changed |= EliminateDuplicatePHINodes(BB);
4380 // Check for and remove branches that will always cause undefined behavior.
4381 Changed |= removeUndefIntroducingPredecessor(BB);
4383 // Merge basic blocks into their predecessor if there is only one distinct
4384 // pred, and if there is only one distinct successor of the predecessor, and
4385 // if there are no PHI nodes.
4387 if (MergeBlockIntoPredecessor(BB))
4390 IRBuilder<> Builder(BB);
4392 // If there is a trivial two-entry PHI node in this basic block, and we can
4393 // eliminate it, do so now.
4394 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
4395 if (PN->getNumIncomingValues() == 2)
4396 Changed |= FoldTwoEntryPHINode(PN, DL);
4398 Builder.SetInsertPoint(BB->getTerminator());
4399 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
4400 if (BI->isUnconditional()) {
4401 if (SimplifyUncondBranch(BI, Builder)) return true;
4403 if (SimplifyCondBranch(BI, Builder)) return true;
4405 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
4406 if (SimplifyReturn(RI, Builder)) return true;
4407 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
4408 if (SimplifyResume(RI, Builder)) return true;
4409 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
4410 if (SimplifySwitch(SI, Builder)) return true;
4411 } else if (UnreachableInst *UI =
4412 dyn_cast<UnreachableInst>(BB->getTerminator())) {
4413 if (SimplifyUnreachable(UI)) return true;
4414 } else if (IndirectBrInst *IBI =
4415 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
4416 if (SimplifyIndirectBr(IBI)) return true;
4422 /// SimplifyCFG - This function is used to do simplification of a CFG. For
4423 /// example, it adjusts branches to branches to eliminate the extra hop, it
4424 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4425 /// of the CFG. It returns true if a modification was made.
4427 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
4428 unsigned BonusInstThreshold,
4429 const DataLayout *DL, AssumptionTracker *AT) {
4430 return SimplifyCFGOpt(TTI, BonusInstThreshold, DL, AT).run(BB);