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->getMDNode(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->getMDNode(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->getMDNode(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,
1065 LLVMContext::MD_nonnull
1067 combineMetadata(I1, I2, KnownIDs);
1068 I2->eraseFromParent();
1073 // Skip debug info if it is not identical.
1074 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1075 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1076 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1077 while (isa<DbgInfoIntrinsic>(I1))
1079 while (isa<DbgInfoIntrinsic>(I2))
1082 } while (I1->isIdenticalToWhenDefined(I2));
1087 // It may not be possible to hoist an invoke.
1088 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1091 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1093 for (BasicBlock::iterator BBI = SI->begin();
1094 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1095 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1096 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1100 // Check for passingValueIsAlwaysUndefined here because we would rather
1101 // eliminate undefined control flow then converting it to a select.
1102 if (passingValueIsAlwaysUndefined(BB1V, PN) ||
1103 passingValueIsAlwaysUndefined(BB2V, PN))
1106 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V, DL))
1108 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V, DL))
1113 // Okay, it is safe to hoist the terminator.
1114 Instruction *NT = I1->clone();
1115 BIParent->getInstList().insert(BI, NT);
1116 if (!NT->getType()->isVoidTy()) {
1117 I1->replaceAllUsesWith(NT);
1118 I2->replaceAllUsesWith(NT);
1122 IRBuilder<true, NoFolder> Builder(NT);
1123 // Hoisting one of the terminators from our successor is a great thing.
1124 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1125 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1126 // nodes, so we insert select instruction to compute the final result.
1127 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1128 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1130 for (BasicBlock::iterator BBI = SI->begin();
1131 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1132 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1133 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1134 if (BB1V == BB2V) continue;
1136 // These values do not agree. Insert a select instruction before NT
1137 // that determines the right value.
1138 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1140 SI = cast<SelectInst>
1141 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1142 BB1V->getName()+"."+BB2V->getName()));
1144 // Make the PHI node use the select for all incoming values for BB1/BB2
1145 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1146 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1147 PN->setIncomingValue(i, SI);
1151 // Update any PHI nodes in our new successors.
1152 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1153 AddPredecessorToBlock(*SI, BIParent, BB1);
1155 EraseTerminatorInstAndDCECond(BI);
1159 /// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd,
1160 /// check whether BBEnd has only two predecessors and the other predecessor
1161 /// ends with an unconditional branch. If it is true, sink any common code
1162 /// in the two predecessors to BBEnd.
1163 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1164 assert(BI1->isUnconditional());
1165 BasicBlock *BB1 = BI1->getParent();
1166 BasicBlock *BBEnd = BI1->getSuccessor(0);
1168 // Check that BBEnd has two predecessors and the other predecessor ends with
1169 // an unconditional branch.
1170 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1171 BasicBlock *Pred0 = *PI++;
1172 if (PI == PE) // Only one predecessor.
1174 BasicBlock *Pred1 = *PI++;
1175 if (PI != PE) // More than two predecessors.
1177 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1178 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1179 if (!BI2 || !BI2->isUnconditional())
1182 // Gather the PHI nodes in BBEnd.
1183 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2;
1184 Instruction *FirstNonPhiInBBEnd = nullptr;
1185 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end();
1187 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1188 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1189 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1190 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN);
1192 FirstNonPhiInBBEnd = &*I;
1196 if (!FirstNonPhiInBBEnd)
1200 // This does very trivial matching, with limited scanning, to find identical
1201 // instructions in the two blocks. We scan backward for obviously identical
1202 // instructions in an identical order.
1203 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1204 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(),
1205 RE2 = BB2->getInstList().rend();
1207 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1210 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1213 // Skip the unconditional branches.
1217 bool Changed = false;
1218 while (RI1 != RE1 && RI2 != RE2) {
1220 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1223 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1227 Instruction *I1 = &*RI1, *I2 = &*RI2;
1228 // I1 and I2 should have a single use in the same PHI node, and they
1229 // perform the same operation.
1230 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1231 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1232 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1233 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) ||
1234 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1235 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1236 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1237 !I1->hasOneUse() || !I2->hasOneUse() ||
1238 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() ||
1239 MapValueFromBB1ToBB2[I1].first != I2)
1242 // Check whether we should swap the operands of ICmpInst.
1243 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1244 bool SwapOpnds = false;
1245 if (ICmp1 && ICmp2 &&
1246 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1247 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1248 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1249 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1250 ICmp2->swapOperands();
1253 if (!I1->isSameOperationAs(I2)) {
1255 ICmp2->swapOperands();
1259 // The operands should be either the same or they need to be generated
1260 // with a PHI node after sinking. We only handle the case where there is
1261 // a single pair of different operands.
1262 Value *DifferentOp1 = nullptr, *DifferentOp2 = nullptr;
1263 unsigned Op1Idx = 0;
1264 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1265 if (I1->getOperand(I) == I2->getOperand(I))
1267 // Early exit if we have more-than one pair of different operands or
1268 // the different operand is already in MapValueFromBB1ToBB2.
1269 // Early exit if we need a PHI node to replace a constant.
1271 MapValueFromBB1ToBB2.find(I1->getOperand(I)) !=
1272 MapValueFromBB1ToBB2.end() ||
1273 isa<Constant>(I1->getOperand(I)) ||
1274 isa<Constant>(I2->getOperand(I))) {
1275 // If we can't sink the instructions, undo the swapping.
1277 ICmp2->swapOperands();
1280 DifferentOp1 = I1->getOperand(I);
1282 DifferentOp2 = I2->getOperand(I);
1285 // We insert the pair of different operands to MapValueFromBB1ToBB2 and
1286 // remove (I1, I2) from MapValueFromBB1ToBB2.
1288 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2,
1289 DifferentOp1->getName() + ".sink",
1291 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN);
1292 // I1 should use NewPN instead of DifferentOp1.
1293 I1->setOperand(Op1Idx, NewPN);
1294 NewPN->addIncoming(DifferentOp1, BB1);
1295 NewPN->addIncoming(DifferentOp2, BB2);
1296 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1298 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second;
1299 MapValueFromBB1ToBB2.erase(I1);
1301 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";);
1302 DEBUG(dbgs() << " " << *I2 << "\n";);
1303 // We need to update RE1 and RE2 if we are going to sink the first
1304 // instruction in the basic block down.
1305 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1306 // Sink the instruction.
1307 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1);
1308 if (!OldPN->use_empty())
1309 OldPN->replaceAllUsesWith(I1);
1310 OldPN->eraseFromParent();
1312 if (!I2->use_empty())
1313 I2->replaceAllUsesWith(I1);
1314 I1->intersectOptionalDataWith(I2);
1315 // TODO: Use combineMetadata here to preserve what metadata we can
1316 // (analogous to the hoisting case above).
1317 I2->eraseFromParent();
1320 RE1 = BB1->getInstList().rend();
1322 RE2 = BB2->getInstList().rend();
1323 FirstNonPhiInBBEnd = I1;
1330 /// \brief Determine if we can hoist sink a sole store instruction out of a
1331 /// conditional block.
1333 /// We are looking for code like the following:
1335 /// store i32 %add, i32* %arrayidx2
1336 /// ... // No other stores or function calls (we could be calling a memory
1337 /// ... // function).
1338 /// %cmp = icmp ult %x, %y
1339 /// br i1 %cmp, label %EndBB, label %ThenBB
1341 /// store i32 %add5, i32* %arrayidx2
1345 /// We are going to transform this into:
1347 /// store i32 %add, i32* %arrayidx2
1349 /// %cmp = icmp ult %x, %y
1350 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1351 /// store i32 %add.add5, i32* %arrayidx2
1354 /// \return The pointer to the value of the previous store if the store can be
1355 /// hoisted into the predecessor block. 0 otherwise.
1356 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1357 BasicBlock *StoreBB, BasicBlock *EndBB) {
1358 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1362 // Volatile or atomic.
1363 if (!StoreToHoist->isSimple())
1366 Value *StorePtr = StoreToHoist->getPointerOperand();
1368 // Look for a store to the same pointer in BrBB.
1369 unsigned MaxNumInstToLookAt = 10;
1370 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(),
1371 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) {
1372 Instruction *CurI = &*RI;
1374 // Could be calling an instruction that effects memory like free().
1375 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI))
1378 StoreInst *SI = dyn_cast<StoreInst>(CurI);
1379 // Found the previous store make sure it stores to the same location.
1380 if (SI && SI->getPointerOperand() == StorePtr)
1381 // Found the previous store, return its value operand.
1382 return SI->getValueOperand();
1384 return nullptr; // Unknown store.
1390 /// \brief Speculate a conditional basic block flattening the CFG.
1392 /// Note that this is a very risky transform currently. Speculating
1393 /// instructions like this is most often not desirable. Instead, there is an MI
1394 /// pass which can do it with full awareness of the resource constraints.
1395 /// However, some cases are "obvious" and we should do directly. An example of
1396 /// this is speculating a single, reasonably cheap instruction.
1398 /// There is only one distinct advantage to flattening the CFG at the IR level:
1399 /// it makes very common but simplistic optimizations such as are common in
1400 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1401 /// modeling their effects with easier to reason about SSA value graphs.
1404 /// An illustration of this transform is turning this IR:
1407 /// %cmp = icmp ult %x, %y
1408 /// br i1 %cmp, label %EndBB, label %ThenBB
1410 /// %sub = sub %x, %y
1413 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1420 /// %cmp = icmp ult %x, %y
1421 /// %sub = sub %x, %y
1422 /// %cond = select i1 %cmp, 0, %sub
1426 /// \returns true if the conditional block is removed.
1427 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1428 const DataLayout *DL) {
1429 // Be conservative for now. FP select instruction can often be expensive.
1430 Value *BrCond = BI->getCondition();
1431 if (isa<FCmpInst>(BrCond))
1434 BasicBlock *BB = BI->getParent();
1435 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1437 // If ThenBB is actually on the false edge of the conditional branch, remember
1438 // to swap the select operands later.
1439 bool Invert = false;
1440 if (ThenBB != BI->getSuccessor(0)) {
1441 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1444 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1446 // Keep a count of how many times instructions are used within CondBB when
1447 // they are candidates for sinking into CondBB. Specifically:
1448 // - They are defined in BB, and
1449 // - They have no side effects, and
1450 // - All of their uses are in CondBB.
1451 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1453 unsigned SpeculationCost = 0;
1454 Value *SpeculatedStoreValue = nullptr;
1455 StoreInst *SpeculatedStore = nullptr;
1456 for (BasicBlock::iterator BBI = ThenBB->begin(),
1457 BBE = std::prev(ThenBB->end());
1458 BBI != BBE; ++BBI) {
1459 Instruction *I = BBI;
1461 if (isa<DbgInfoIntrinsic>(I))
1464 // Only speculatively execution a single instruction (not counting the
1465 // terminator) for now.
1467 if (SpeculationCost > 1)
1470 // Don't hoist the instruction if it's unsafe or expensive.
1471 if (!isSafeToSpeculativelyExecute(I, DL) &&
1472 !(HoistCondStores &&
1473 (SpeculatedStoreValue = isSafeToSpeculateStore(I, BB, ThenBB,
1476 if (!SpeculatedStoreValue &&
1477 ComputeSpeculationCost(I, DL) > PHINodeFoldingThreshold)
1480 // Store the store speculation candidate.
1481 if (SpeculatedStoreValue)
1482 SpeculatedStore = cast<StoreInst>(I);
1484 // Do not hoist the instruction if any of its operands are defined but not
1485 // used in BB. The transformation will prevent the operand from
1486 // being sunk into the use block.
1487 for (User::op_iterator i = I->op_begin(), e = I->op_end();
1489 Instruction *OpI = dyn_cast<Instruction>(*i);
1490 if (!OpI || OpI->getParent() != BB ||
1491 OpI->mayHaveSideEffects())
1492 continue; // Not a candidate for sinking.
1494 ++SinkCandidateUseCounts[OpI];
1498 // Consider any sink candidates which are only used in CondBB as costs for
1499 // speculation. Note, while we iterate over a DenseMap here, we are summing
1500 // and so iteration order isn't significant.
1501 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I =
1502 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end();
1504 if (I->first->getNumUses() == I->second) {
1506 if (SpeculationCost > 1)
1510 // Check that the PHI nodes can be converted to selects.
1511 bool HaveRewritablePHIs = false;
1512 for (BasicBlock::iterator I = EndBB->begin();
1513 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1514 Value *OrigV = PN->getIncomingValueForBlock(BB);
1515 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1517 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1518 // Skip PHIs which are trivial.
1522 // Don't convert to selects if we could remove undefined behavior instead.
1523 if (passingValueIsAlwaysUndefined(OrigV, PN) ||
1524 passingValueIsAlwaysUndefined(ThenV, PN))
1527 HaveRewritablePHIs = true;
1528 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1529 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1530 if (!OrigCE && !ThenCE)
1531 continue; // Known safe and cheap.
1533 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE, DL)) ||
1534 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE, DL)))
1536 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, DL) : 0;
1537 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, DL) : 0;
1538 if (OrigCost + ThenCost > 2 * PHINodeFoldingThreshold)
1541 // Account for the cost of an unfolded ConstantExpr which could end up
1542 // getting expanded into Instructions.
1543 // FIXME: This doesn't account for how many operations are combined in the
1544 // constant expression.
1546 if (SpeculationCost > 1)
1550 // If there are no PHIs to process, bail early. This helps ensure idempotence
1552 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
1555 // If we get here, we can hoist the instruction and if-convert.
1556 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1558 // Insert a select of the value of the speculated store.
1559 if (SpeculatedStoreValue) {
1560 IRBuilder<true, NoFolder> Builder(BI);
1561 Value *TrueV = SpeculatedStore->getValueOperand();
1562 Value *FalseV = SpeculatedStoreValue;
1564 std::swap(TrueV, FalseV);
1565 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() +
1566 "." + FalseV->getName());
1567 SpeculatedStore->setOperand(0, S);
1570 // Hoist the instructions.
1571 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(),
1572 std::prev(ThenBB->end()));
1574 // Insert selects and rewrite the PHI operands.
1575 IRBuilder<true, NoFolder> Builder(BI);
1576 for (BasicBlock::iterator I = EndBB->begin();
1577 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1578 unsigned OrigI = PN->getBasicBlockIndex(BB);
1579 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1580 Value *OrigV = PN->getIncomingValue(OrigI);
1581 Value *ThenV = PN->getIncomingValue(ThenI);
1583 // Skip PHIs which are trivial.
1587 // Create a select whose true value is the speculatively executed value and
1588 // false value is the preexisting value. Swap them if the branch
1589 // destinations were inverted.
1590 Value *TrueV = ThenV, *FalseV = OrigV;
1592 std::swap(TrueV, FalseV);
1593 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV,
1594 TrueV->getName() + "." + FalseV->getName());
1595 PN->setIncomingValue(OrigI, V);
1596 PN->setIncomingValue(ThenI, V);
1603 /// \returns True if this block contains a CallInst with the NoDuplicate
1605 static bool HasNoDuplicateCall(const BasicBlock *BB) {
1606 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1607 const CallInst *CI = dyn_cast<CallInst>(I);
1610 if (CI->cannotDuplicate())
1616 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1617 /// across this block.
1618 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1619 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1622 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1623 if (isa<DbgInfoIntrinsic>(BBI))
1625 if (Size > 10) return false; // Don't clone large BB's.
1628 // We can only support instructions that do not define values that are
1629 // live outside of the current basic block.
1630 for (User *U : BBI->users()) {
1631 Instruction *UI = cast<Instruction>(U);
1632 if (UI->getParent() != BB || isa<PHINode>(UI)) return false;
1635 // Looks ok, continue checking.
1641 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1642 /// that is defined in the same block as the branch and if any PHI entries are
1643 /// constants, thread edges corresponding to that entry to be branches to their
1644 /// ultimate destination.
1645 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *DL) {
1646 BasicBlock *BB = BI->getParent();
1647 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1648 // NOTE: we currently cannot transform this case if the PHI node is used
1649 // outside of the block.
1650 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1653 // Degenerate case of a single entry PHI.
1654 if (PN->getNumIncomingValues() == 1) {
1655 FoldSingleEntryPHINodes(PN->getParent());
1659 // Now we know that this block has multiple preds and two succs.
1660 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1662 if (HasNoDuplicateCall(BB)) return false;
1664 // Okay, this is a simple enough basic block. See if any phi values are
1666 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1667 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1668 if (!CB || !CB->getType()->isIntegerTy(1)) continue;
1670 // Okay, we now know that all edges from PredBB should be revectored to
1671 // branch to RealDest.
1672 BasicBlock *PredBB = PN->getIncomingBlock(i);
1673 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1675 if (RealDest == BB) continue; // Skip self loops.
1676 // Skip if the predecessor's terminator is an indirect branch.
1677 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1679 // The dest block might have PHI nodes, other predecessors and other
1680 // difficult cases. Instead of being smart about this, just insert a new
1681 // block that jumps to the destination block, effectively splitting
1682 // the edge we are about to create.
1683 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1684 RealDest->getName()+".critedge",
1685 RealDest->getParent(), RealDest);
1686 BranchInst::Create(RealDest, EdgeBB);
1688 // Update PHI nodes.
1689 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1691 // BB may have instructions that are being threaded over. Clone these
1692 // instructions into EdgeBB. We know that there will be no uses of the
1693 // cloned instructions outside of EdgeBB.
1694 BasicBlock::iterator InsertPt = EdgeBB->begin();
1695 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1696 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1697 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1698 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1701 // Clone the instruction.
1702 Instruction *N = BBI->clone();
1703 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1705 // Update operands due to translation.
1706 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1708 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1709 if (PI != TranslateMap.end())
1713 // Check for trivial simplification.
1714 if (Value *V = SimplifyInstruction(N, DL)) {
1715 TranslateMap[BBI] = V;
1716 delete N; // Instruction folded away, don't need actual inst
1718 // Insert the new instruction into its new home.
1719 EdgeBB->getInstList().insert(InsertPt, N);
1720 if (!BBI->use_empty())
1721 TranslateMap[BBI] = N;
1725 // Loop over all of the edges from PredBB to BB, changing them to branch
1726 // to EdgeBB instead.
1727 TerminatorInst *PredBBTI = PredBB->getTerminator();
1728 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1729 if (PredBBTI->getSuccessor(i) == BB) {
1730 BB->removePredecessor(PredBB);
1731 PredBBTI->setSuccessor(i, EdgeBB);
1734 // Recurse, simplifying any other constants.
1735 return FoldCondBranchOnPHI(BI, DL) | true;
1741 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1742 /// PHI node, see if we can eliminate it.
1743 static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *DL) {
1744 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1745 // statement", which has a very simple dominance structure. Basically, we
1746 // are trying to find the condition that is being branched on, which
1747 // subsequently causes this merge to happen. We really want control
1748 // dependence information for this check, but simplifycfg can't keep it up
1749 // to date, and this catches most of the cases we care about anyway.
1750 BasicBlock *BB = PN->getParent();
1751 BasicBlock *IfTrue, *IfFalse;
1752 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1754 // Don't bother if the branch will be constant folded trivially.
1755 isa<ConstantInt>(IfCond))
1758 // Okay, we found that we can merge this two-entry phi node into a select.
1759 // Doing so would require us to fold *all* two entry phi nodes in this block.
1760 // At some point this becomes non-profitable (particularly if the target
1761 // doesn't support cmov's). Only do this transformation if there are two or
1762 // fewer PHI nodes in this block.
1763 unsigned NumPhis = 0;
1764 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1768 // Loop over the PHI's seeing if we can promote them all to select
1769 // instructions. While we are at it, keep track of the instructions
1770 // that need to be moved to the dominating block.
1771 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1772 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1773 MaxCostVal1 = PHINodeFoldingThreshold;
1775 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1776 PHINode *PN = cast<PHINode>(II++);
1777 if (Value *V = SimplifyInstruction(PN, DL)) {
1778 PN->replaceAllUsesWith(V);
1779 PN->eraseFromParent();
1783 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1785 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1790 // If we folded the first phi, PN dangles at this point. Refresh it. If
1791 // we ran out of PHIs then we simplified them all.
1792 PN = dyn_cast<PHINode>(BB->begin());
1793 if (!PN) return true;
1795 // Don't fold i1 branches on PHIs which contain binary operators. These can
1796 // often be turned into switches and other things.
1797 if (PN->getType()->isIntegerTy(1) &&
1798 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1799 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1800 isa<BinaryOperator>(IfCond)))
1803 // If we all PHI nodes are promotable, check to make sure that all
1804 // instructions in the predecessor blocks can be promoted as well. If
1805 // not, we won't be able to get rid of the control flow, so it's not
1806 // worth promoting to select instructions.
1807 BasicBlock *DomBlock = nullptr;
1808 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1809 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1810 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1813 DomBlock = *pred_begin(IfBlock1);
1814 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1815 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1816 // This is not an aggressive instruction that we can promote.
1817 // Because of this, we won't be able to get rid of the control
1818 // flow, so the xform is not worth it.
1823 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1826 DomBlock = *pred_begin(IfBlock2);
1827 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1828 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1829 // This is not an aggressive instruction that we can promote.
1830 // Because of this, we won't be able to get rid of the control
1831 // flow, so the xform is not worth it.
1836 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1837 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1839 // If we can still promote the PHI nodes after this gauntlet of tests,
1840 // do all of the PHI's now.
1841 Instruction *InsertPt = DomBlock->getTerminator();
1842 IRBuilder<true, NoFolder> Builder(InsertPt);
1844 // Move all 'aggressive' instructions, which are defined in the
1845 // conditional parts of the if's up to the dominating block.
1847 DomBlock->getInstList().splice(InsertPt,
1848 IfBlock1->getInstList(), IfBlock1->begin(),
1849 IfBlock1->getTerminator());
1851 DomBlock->getInstList().splice(InsertPt,
1852 IfBlock2->getInstList(), IfBlock2->begin(),
1853 IfBlock2->getTerminator());
1855 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1856 // Change the PHI node into a select instruction.
1857 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1858 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1861 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1862 PN->replaceAllUsesWith(NV);
1864 PN->eraseFromParent();
1867 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1868 // has been flattened. Change DomBlock to jump directly to our new block to
1869 // avoid other simplifycfg's kicking in on the diamond.
1870 TerminatorInst *OldTI = DomBlock->getTerminator();
1871 Builder.SetInsertPoint(OldTI);
1872 Builder.CreateBr(BB);
1873 OldTI->eraseFromParent();
1877 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1878 /// to two returning blocks, try to merge them together into one return,
1879 /// introducing a select if the return values disagree.
1880 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1881 IRBuilder<> &Builder) {
1882 assert(BI->isConditional() && "Must be a conditional branch");
1883 BasicBlock *TrueSucc = BI->getSuccessor(0);
1884 BasicBlock *FalseSucc = BI->getSuccessor(1);
1885 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1886 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1888 // Check to ensure both blocks are empty (just a return) or optionally empty
1889 // with PHI nodes. If there are other instructions, merging would cause extra
1890 // computation on one path or the other.
1891 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1893 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1896 Builder.SetInsertPoint(BI);
1897 // Okay, we found a branch that is going to two return nodes. If
1898 // there is no return value for this function, just change the
1899 // branch into a return.
1900 if (FalseRet->getNumOperands() == 0) {
1901 TrueSucc->removePredecessor(BI->getParent());
1902 FalseSucc->removePredecessor(BI->getParent());
1903 Builder.CreateRetVoid();
1904 EraseTerminatorInstAndDCECond(BI);
1908 // Otherwise, figure out what the true and false return values are
1909 // so we can insert a new select instruction.
1910 Value *TrueValue = TrueRet->getReturnValue();
1911 Value *FalseValue = FalseRet->getReturnValue();
1913 // Unwrap any PHI nodes in the return blocks.
1914 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1915 if (TVPN->getParent() == TrueSucc)
1916 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1917 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1918 if (FVPN->getParent() == FalseSucc)
1919 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1921 // In order for this transformation to be safe, we must be able to
1922 // unconditionally execute both operands to the return. This is
1923 // normally the case, but we could have a potentially-trapping
1924 // constant expression that prevents this transformation from being
1926 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1929 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1933 // Okay, we collected all the mapped values and checked them for sanity, and
1934 // defined to really do this transformation. First, update the CFG.
1935 TrueSucc->removePredecessor(BI->getParent());
1936 FalseSucc->removePredecessor(BI->getParent());
1938 // Insert select instructions where needed.
1939 Value *BrCond = BI->getCondition();
1941 // Insert a select if the results differ.
1942 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1943 } else if (isa<UndefValue>(TrueValue)) {
1944 TrueValue = FalseValue;
1946 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1947 FalseValue, "retval");
1951 Value *RI = !TrueValue ?
1952 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
1956 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
1957 << "\n " << *BI << "NewRet = " << *RI
1958 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
1960 EraseTerminatorInstAndDCECond(BI);
1965 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
1966 /// probabilities of the branch taking each edge. Fills in the two APInt
1967 /// parameters and return true, or returns false if no or invalid metadata was
1969 static bool ExtractBranchMetadata(BranchInst *BI,
1970 uint64_t &ProbTrue, uint64_t &ProbFalse) {
1971 assert(BI->isConditional() &&
1972 "Looking for probabilities on unconditional branch?");
1973 MDNode *ProfileData = BI->getMDNode(LLVMContext::MD_prof);
1974 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
1975 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
1976 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
1977 if (!CITrue || !CIFalse) return false;
1978 ProbTrue = CITrue->getValue().getZExtValue();
1979 ProbFalse = CIFalse->getValue().getZExtValue();
1983 /// checkCSEInPredecessor - Return true if the given instruction is available
1984 /// in its predecessor block. If yes, the instruction will be removed.
1986 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
1987 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
1989 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
1990 Instruction *PBI = &*I;
1991 // Check whether Inst and PBI generate the same value.
1992 if (Inst->isIdenticalTo(PBI)) {
1993 Inst->replaceAllUsesWith(PBI);
1994 Inst->eraseFromParent();
2001 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
2002 /// predecessor branches to us and one of our successors, fold the block into
2003 /// the predecessor and use logical operations to pick the right destination.
2004 bool llvm::FoldBranchToCommonDest(BranchInst *BI, const DataLayout *DL,
2005 unsigned BonusInstThreshold) {
2006 BasicBlock *BB = BI->getParent();
2008 Instruction *Cond = nullptr;
2009 if (BI->isConditional())
2010 Cond = dyn_cast<Instruction>(BI->getCondition());
2012 // For unconditional branch, check for a simple CFG pattern, where
2013 // BB has a single predecessor and BB's successor is also its predecessor's
2014 // successor. If such pattern exisits, check for CSE between BB and its
2016 if (BasicBlock *PB = BB->getSinglePredecessor())
2017 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2018 if (PBI->isConditional() &&
2019 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2020 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2021 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
2023 Instruction *Curr = I++;
2024 if (isa<CmpInst>(Curr)) {
2028 // Quit if we can't remove this instruction.
2029 if (!checkCSEInPredecessor(Curr, PB))
2038 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2039 Cond->getParent() != BB || !Cond->hasOneUse())
2042 // Make sure the instruction after the condition is the cond branch.
2043 BasicBlock::iterator CondIt = Cond; ++CondIt;
2045 // Ignore dbg intrinsics.
2046 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
2051 // Only allow this transformation if computing the condition doesn't involve
2052 // too many instructions and these involved instructions can be executed
2053 // unconditionally. We denote all involved instructions except the condition
2054 // as "bonus instructions", and only allow this transformation when the
2055 // number of the bonus instructions does not exceed a certain threshold.
2056 unsigned NumBonusInsts = 0;
2057 for (auto I = BB->begin(); Cond != I; ++I) {
2058 // Ignore dbg intrinsics.
2059 if (isa<DbgInfoIntrinsic>(I))
2061 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(I, DL))
2063 // I has only one use and can be executed unconditionally.
2064 Instruction *User = dyn_cast<Instruction>(I->user_back());
2065 if (User == nullptr || User->getParent() != BB)
2067 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2068 // to use any other instruction, User must be an instruction between next(I)
2071 // Early exits once we reach the limit.
2072 if (NumBonusInsts > BonusInstThreshold)
2076 // Cond is known to be a compare or binary operator. Check to make sure that
2077 // neither operand is a potentially-trapping constant expression.
2078 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2081 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2085 // Finally, don't infinitely unroll conditional loops.
2086 BasicBlock *TrueDest = BI->getSuccessor(0);
2087 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2088 if (TrueDest == BB || FalseDest == BB)
2091 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2092 BasicBlock *PredBlock = *PI;
2093 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2095 // Check that we have two conditional branches. If there is a PHI node in
2096 // the common successor, verify that the same value flows in from both
2098 SmallVector<PHINode*, 4> PHIs;
2099 if (!PBI || PBI->isUnconditional() ||
2100 (BI->isConditional() &&
2101 !SafeToMergeTerminators(BI, PBI)) ||
2102 (!BI->isConditional() &&
2103 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2106 // Determine if the two branches share a common destination.
2107 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2108 bool InvertPredCond = false;
2110 if (BI->isConditional()) {
2111 if (PBI->getSuccessor(0) == TrueDest)
2112 Opc = Instruction::Or;
2113 else if (PBI->getSuccessor(1) == FalseDest)
2114 Opc = Instruction::And;
2115 else if (PBI->getSuccessor(0) == FalseDest)
2116 Opc = Instruction::And, InvertPredCond = true;
2117 else if (PBI->getSuccessor(1) == TrueDest)
2118 Opc = Instruction::Or, InvertPredCond = true;
2122 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2126 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2127 IRBuilder<> Builder(PBI);
2129 // If we need to invert the condition in the pred block to match, do so now.
2130 if (InvertPredCond) {
2131 Value *NewCond = PBI->getCondition();
2133 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2134 CmpInst *CI = cast<CmpInst>(NewCond);
2135 CI->setPredicate(CI->getInversePredicate());
2137 NewCond = Builder.CreateNot(NewCond,
2138 PBI->getCondition()->getName()+".not");
2141 PBI->setCondition(NewCond);
2142 PBI->swapSuccessors();
2145 // If we have bonus instructions, clone them into the predecessor block.
2146 // Note that there may be mutliple predecessor blocks, so we cannot move
2147 // bonus instructions to a predecessor block.
2148 ValueToValueMapTy VMap; // maps original values to cloned values
2149 // We already make sure Cond is the last instruction before BI. Therefore,
2150 // every instructions before Cond other than DbgInfoIntrinsic are bonus
2152 for (auto BonusInst = BB->begin(); Cond != BonusInst; ++BonusInst) {
2153 if (isa<DbgInfoIntrinsic>(BonusInst))
2155 Instruction *NewBonusInst = BonusInst->clone();
2156 RemapInstruction(NewBonusInst, VMap,
2157 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2158 VMap[BonusInst] = NewBonusInst;
2160 // If we moved a load, we cannot any longer claim any knowledge about
2161 // its potential value. The previous information might have been valid
2162 // only given the branch precondition.
2163 // For an analogous reason, we must also drop all the metadata whose
2164 // semantics we don't understand.
2165 NewBonusInst->dropUnknownMetadata(LLVMContext::MD_dbg);
2167 PredBlock->getInstList().insert(PBI, NewBonusInst);
2168 NewBonusInst->takeName(BonusInst);
2169 BonusInst->setName(BonusInst->getName() + ".old");
2172 // Clone Cond into the predecessor basic block, and or/and the
2173 // two conditions together.
2174 Instruction *New = Cond->clone();
2175 RemapInstruction(New, VMap,
2176 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2177 PredBlock->getInstList().insert(PBI, New);
2178 New->takeName(Cond);
2179 Cond->setName(New->getName() + ".old");
2181 if (BI->isConditional()) {
2182 Instruction *NewCond =
2183 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2185 PBI->setCondition(NewCond);
2187 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2188 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2190 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2192 SmallVector<uint64_t, 8> NewWeights;
2194 if (PBI->getSuccessor(0) == BB) {
2195 if (PredHasWeights && SuccHasWeights) {
2196 // PBI: br i1 %x, BB, FalseDest
2197 // BI: br i1 %y, TrueDest, FalseDest
2198 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2199 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2200 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2201 // TrueWeight for PBI * FalseWeight for BI.
2202 // We assume that total weights of a BranchInst can fit into 32 bits.
2203 // Therefore, we will not have overflow using 64-bit arithmetic.
2204 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2205 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2207 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2208 PBI->setSuccessor(0, TrueDest);
2210 if (PBI->getSuccessor(1) == BB) {
2211 if (PredHasWeights && SuccHasWeights) {
2212 // PBI: br i1 %x, TrueDest, BB
2213 // BI: br i1 %y, TrueDest, FalseDest
2214 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2215 // FalseWeight for PBI * TrueWeight for BI.
2216 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2217 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2218 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2219 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2221 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2222 PBI->setSuccessor(1, FalseDest);
2224 if (NewWeights.size() == 2) {
2225 // Halve the weights if any of them cannot fit in an uint32_t
2226 FitWeights(NewWeights);
2228 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2229 PBI->setMetadata(LLVMContext::MD_prof,
2230 MDBuilder(BI->getContext()).
2231 createBranchWeights(MDWeights));
2233 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2235 // Update PHI nodes in the common successors.
2236 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2237 ConstantInt *PBI_C = cast<ConstantInt>(
2238 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2239 assert(PBI_C->getType()->isIntegerTy(1));
2240 Instruction *MergedCond = nullptr;
2241 if (PBI->getSuccessor(0) == TrueDest) {
2242 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2243 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2244 // is false: !PBI_Cond and BI_Value
2245 Instruction *NotCond =
2246 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2249 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2254 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2255 PBI->getCondition(), MergedCond,
2258 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2259 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2260 // is false: PBI_Cond and BI_Value
2262 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2263 PBI->getCondition(), New,
2265 if (PBI_C->isOne()) {
2266 Instruction *NotCond =
2267 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2270 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2271 NotCond, MergedCond,
2276 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2279 // Change PBI from Conditional to Unconditional.
2280 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2281 EraseTerminatorInstAndDCECond(PBI);
2285 // TODO: If BB is reachable from all paths through PredBlock, then we
2286 // could replace PBI's branch probabilities with BI's.
2288 // Copy any debug value intrinsics into the end of PredBlock.
2289 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2290 if (isa<DbgInfoIntrinsic>(*I))
2291 I->clone()->insertBefore(PBI);
2298 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2299 /// predecessor of another block, this function tries to simplify it. We know
2300 /// that PBI and BI are both conditional branches, and BI is in one of the
2301 /// successor blocks of PBI - PBI branches to BI.
2302 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2303 assert(PBI->isConditional() && BI->isConditional());
2304 BasicBlock *BB = BI->getParent();
2306 // If this block ends with a branch instruction, and if there is a
2307 // predecessor that ends on a branch of the same condition, make
2308 // this conditional branch redundant.
2309 if (PBI->getCondition() == BI->getCondition() &&
2310 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2311 // Okay, the outcome of this conditional branch is statically
2312 // knowable. If this block had a single pred, handle specially.
2313 if (BB->getSinglePredecessor()) {
2314 // Turn this into a branch on constant.
2315 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2316 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2318 return true; // Nuke the branch on constant.
2321 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2322 // in the constant and simplify the block result. Subsequent passes of
2323 // simplifycfg will thread the block.
2324 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2325 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2326 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2327 std::distance(PB, PE),
2328 BI->getCondition()->getName() + ".pr",
2330 // Okay, we're going to insert the PHI node. Since PBI is not the only
2331 // predecessor, compute the PHI'd conditional value for all of the preds.
2332 // Any predecessor where the condition is not computable we keep symbolic.
2333 for (pred_iterator PI = PB; PI != PE; ++PI) {
2334 BasicBlock *P = *PI;
2335 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2336 PBI != BI && PBI->isConditional() &&
2337 PBI->getCondition() == BI->getCondition() &&
2338 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2339 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2340 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2343 NewPN->addIncoming(BI->getCondition(), P);
2347 BI->setCondition(NewPN);
2352 // If this is a conditional branch in an empty block, and if any
2353 // predecessors are a conditional branch to one of our destinations,
2354 // fold the conditions into logical ops and one cond br.
2355 BasicBlock::iterator BBI = BB->begin();
2356 // Ignore dbg intrinsics.
2357 while (isa<DbgInfoIntrinsic>(BBI))
2363 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2368 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2370 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2371 PBIOp = 0, BIOp = 1;
2372 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2373 PBIOp = 1, BIOp = 0;
2374 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2379 // Check to make sure that the other destination of this branch
2380 // isn't BB itself. If so, this is an infinite loop that will
2381 // keep getting unwound.
2382 if (PBI->getSuccessor(PBIOp) == BB)
2385 // Do not perform this transformation if it would require
2386 // insertion of a large number of select instructions. For targets
2387 // without predication/cmovs, this is a big pessimization.
2389 // Also do not perform this transformation if any phi node in the common
2390 // destination block can trap when reached by BB or PBB (PR17073). In that
2391 // case, it would be unsafe to hoist the operation into a select instruction.
2393 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2394 unsigned NumPhis = 0;
2395 for (BasicBlock::iterator II = CommonDest->begin();
2396 isa<PHINode>(II); ++II, ++NumPhis) {
2397 if (NumPhis > 2) // Disable this xform.
2400 PHINode *PN = cast<PHINode>(II);
2401 Value *BIV = PN->getIncomingValueForBlock(BB);
2402 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
2406 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2407 Value *PBIV = PN->getIncomingValue(PBBIdx);
2408 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
2413 // Finally, if everything is ok, fold the branches to logical ops.
2414 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2416 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2417 << "AND: " << *BI->getParent());
2420 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2421 // branch in it, where one edge (OtherDest) goes back to itself but the other
2422 // exits. We don't *know* that the program avoids the infinite loop
2423 // (even though that seems likely). If we do this xform naively, we'll end up
2424 // recursively unpeeling the loop. Since we know that (after the xform is
2425 // done) that the block *is* infinite if reached, we just make it an obviously
2426 // infinite loop with no cond branch.
2427 if (OtherDest == BB) {
2428 // Insert it at the end of the function, because it's either code,
2429 // or it won't matter if it's hot. :)
2430 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2431 "infloop", BB->getParent());
2432 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2433 OtherDest = InfLoopBlock;
2436 DEBUG(dbgs() << *PBI->getParent()->getParent());
2438 // BI may have other predecessors. Because of this, we leave
2439 // it alone, but modify PBI.
2441 // Make sure we get to CommonDest on True&True directions.
2442 Value *PBICond = PBI->getCondition();
2443 IRBuilder<true, NoFolder> Builder(PBI);
2445 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2447 Value *BICond = BI->getCondition();
2449 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2451 // Merge the conditions.
2452 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2454 // Modify PBI to branch on the new condition to the new dests.
2455 PBI->setCondition(Cond);
2456 PBI->setSuccessor(0, CommonDest);
2457 PBI->setSuccessor(1, OtherDest);
2459 // Update branch weight for PBI.
2460 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2461 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2463 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2465 if (PredHasWeights && SuccHasWeights) {
2466 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2467 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2468 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2469 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2470 // The weight to CommonDest should be PredCommon * SuccTotal +
2471 // PredOther * SuccCommon.
2472 // The weight to OtherDest should be PredOther * SuccOther.
2473 SmallVector<uint64_t, 2> NewWeights;
2474 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) +
2475 PredOther * SuccCommon);
2476 NewWeights.push_back(PredOther * SuccOther);
2477 // Halve the weights if any of them cannot fit in an uint32_t
2478 FitWeights(NewWeights);
2480 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end());
2481 PBI->setMetadata(LLVMContext::MD_prof,
2482 MDBuilder(BI->getContext()).
2483 createBranchWeights(MDWeights));
2486 // OtherDest may have phi nodes. If so, add an entry from PBI's
2487 // block that are identical to the entries for BI's block.
2488 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2490 // We know that the CommonDest already had an edge from PBI to
2491 // it. If it has PHIs though, the PHIs may have different
2492 // entries for BB and PBI's BB. If so, insert a select to make
2495 for (BasicBlock::iterator II = CommonDest->begin();
2496 (PN = dyn_cast<PHINode>(II)); ++II) {
2497 Value *BIV = PN->getIncomingValueForBlock(BB);
2498 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2499 Value *PBIV = PN->getIncomingValue(PBBIdx);
2501 // Insert a select in PBI to pick the right value.
2502 Value *NV = cast<SelectInst>
2503 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2504 PN->setIncomingValue(PBBIdx, NV);
2508 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2509 DEBUG(dbgs() << *PBI->getParent()->getParent());
2511 // This basic block is probably dead. We know it has at least
2512 // one fewer predecessor.
2516 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2517 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2518 // Takes care of updating the successors and removing the old terminator.
2519 // Also makes sure not to introduce new successors by assuming that edges to
2520 // non-successor TrueBBs and FalseBBs aren't reachable.
2521 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2522 BasicBlock *TrueBB, BasicBlock *FalseBB,
2523 uint32_t TrueWeight,
2524 uint32_t FalseWeight){
2525 // Remove any superfluous successor edges from the CFG.
2526 // First, figure out which successors to preserve.
2527 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2529 BasicBlock *KeepEdge1 = TrueBB;
2530 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
2532 // Then remove the rest.
2533 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2534 BasicBlock *Succ = OldTerm->getSuccessor(I);
2535 // Make sure only to keep exactly one copy of each edge.
2536 if (Succ == KeepEdge1)
2537 KeepEdge1 = nullptr;
2538 else if (Succ == KeepEdge2)
2539 KeepEdge2 = nullptr;
2541 Succ->removePredecessor(OldTerm->getParent());
2544 IRBuilder<> Builder(OldTerm);
2545 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2547 // Insert an appropriate new terminator.
2548 if (!KeepEdge1 && !KeepEdge2) {
2549 if (TrueBB == FalseBB)
2550 // We were only looking for one successor, and it was present.
2551 // Create an unconditional branch to it.
2552 Builder.CreateBr(TrueBB);
2554 // We found both of the successors we were looking for.
2555 // Create a conditional branch sharing the condition of the select.
2556 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2557 if (TrueWeight != FalseWeight)
2558 NewBI->setMetadata(LLVMContext::MD_prof,
2559 MDBuilder(OldTerm->getContext()).
2560 createBranchWeights(TrueWeight, FalseWeight));
2562 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2563 // Neither of the selected blocks were successors, so this
2564 // terminator must be unreachable.
2565 new UnreachableInst(OldTerm->getContext(), OldTerm);
2567 // One of the selected values was a successor, but the other wasn't.
2568 // Insert an unconditional branch to the one that was found;
2569 // the edge to the one that wasn't must be unreachable.
2571 // Only TrueBB was found.
2572 Builder.CreateBr(TrueBB);
2574 // Only FalseBB was found.
2575 Builder.CreateBr(FalseBB);
2578 EraseTerminatorInstAndDCECond(OldTerm);
2582 // SimplifySwitchOnSelect - Replaces
2583 // (switch (select cond, X, Y)) on constant X, Y
2584 // with a branch - conditional if X and Y lead to distinct BBs,
2585 // unconditional otherwise.
2586 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2587 // Check for constant integer values in the select.
2588 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2589 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2590 if (!TrueVal || !FalseVal)
2593 // Find the relevant condition and destinations.
2594 Value *Condition = Select->getCondition();
2595 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2596 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2598 // Get weight for TrueBB and FalseBB.
2599 uint32_t TrueWeight = 0, FalseWeight = 0;
2600 SmallVector<uint64_t, 8> Weights;
2601 bool HasWeights = HasBranchWeights(SI);
2603 GetBranchWeights(SI, Weights);
2604 if (Weights.size() == 1 + SI->getNumCases()) {
2605 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2606 getSuccessorIndex()];
2607 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2608 getSuccessorIndex()];
2612 // Perform the actual simplification.
2613 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2614 TrueWeight, FalseWeight);
2617 // SimplifyIndirectBrOnSelect - Replaces
2618 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2619 // blockaddress(@fn, BlockB)))
2621 // (br cond, BlockA, BlockB).
2622 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2623 // Check that both operands of the select are block addresses.
2624 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2625 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2629 // Extract the actual blocks.
2630 BasicBlock *TrueBB = TBA->getBasicBlock();
2631 BasicBlock *FalseBB = FBA->getBasicBlock();
2633 // Perform the actual simplification.
2634 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2638 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2639 /// instruction (a seteq/setne with a constant) as the only instruction in a
2640 /// block that ends with an uncond branch. We are looking for a very specific
2641 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2642 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2643 /// default value goes to an uncond block with a seteq in it, we get something
2646 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2648 /// %tmp = icmp eq i8 %A, 92
2651 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2653 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2654 /// the PHI, merging the third icmp into the switch.
2655 static bool TryToSimplifyUncondBranchWithICmpInIt(
2656 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI,
2657 unsigned BonusInstThreshold, const DataLayout *DL, AssumptionTracker *AT) {
2658 BasicBlock *BB = ICI->getParent();
2660 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2662 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2664 Value *V = ICI->getOperand(0);
2665 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2667 // The pattern we're looking for is where our only predecessor is a switch on
2668 // 'V' and this block is the default case for the switch. In this case we can
2669 // fold the compared value into the switch to simplify things.
2670 BasicBlock *Pred = BB->getSinglePredecessor();
2671 if (!Pred || !isa<SwitchInst>(Pred->getTerminator())) return false;
2673 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2674 if (SI->getCondition() != V)
2677 // If BB is reachable on a non-default case, then we simply know the value of
2678 // V in this block. Substitute it and constant fold the icmp instruction
2680 if (SI->getDefaultDest() != BB) {
2681 ConstantInt *VVal = SI->findCaseDest(BB);
2682 assert(VVal && "Should have a unique destination value");
2683 ICI->setOperand(0, VVal);
2685 if (Value *V = SimplifyInstruction(ICI, DL)) {
2686 ICI->replaceAllUsesWith(V);
2687 ICI->eraseFromParent();
2689 // BB is now empty, so it is likely to simplify away.
2690 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
2693 // Ok, the block is reachable from the default dest. If the constant we're
2694 // comparing exists in one of the other edges, then we can constant fold ICI
2696 if (SI->findCaseValue(Cst) != SI->case_default()) {
2698 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2699 V = ConstantInt::getFalse(BB->getContext());
2701 V = ConstantInt::getTrue(BB->getContext());
2703 ICI->replaceAllUsesWith(V);
2704 ICI->eraseFromParent();
2705 // BB is now empty, so it is likely to simplify away.
2706 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
2709 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2711 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2712 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
2713 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
2714 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2717 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2719 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2720 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2722 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2723 std::swap(DefaultCst, NewCst);
2725 // Replace ICI (which is used by the PHI for the default value) with true or
2726 // false depending on if it is EQ or NE.
2727 ICI->replaceAllUsesWith(DefaultCst);
2728 ICI->eraseFromParent();
2730 // Okay, the switch goes to this block on a default value. Add an edge from
2731 // the switch to the merge point on the compared value.
2732 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2733 BB->getParent(), BB);
2734 SmallVector<uint64_t, 8> Weights;
2735 bool HasWeights = HasBranchWeights(SI);
2737 GetBranchWeights(SI, Weights);
2738 if (Weights.size() == 1 + SI->getNumCases()) {
2739 // Split weight for default case to case for "Cst".
2740 Weights[0] = (Weights[0]+1) >> 1;
2741 Weights.push_back(Weights[0]);
2743 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2744 SI->setMetadata(LLVMContext::MD_prof,
2745 MDBuilder(SI->getContext()).
2746 createBranchWeights(MDWeights));
2749 SI->addCase(Cst, NewBB);
2751 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2752 Builder.SetInsertPoint(NewBB);
2753 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2754 Builder.CreateBr(SuccBlock);
2755 PHIUse->addIncoming(NewCst, NewBB);
2759 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2760 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2761 /// fold it into a switch instruction if so.
2762 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *DL,
2763 IRBuilder<> &Builder) {
2764 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2765 if (!Cond) return false;
2768 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2769 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2770 // 'setne's and'ed together, collect them.
2771 Value *CompVal = nullptr;
2772 std::vector<ConstantInt*> Values;
2773 bool TrueWhenEqual = true;
2774 Value *ExtraCase = nullptr;
2775 unsigned UsedICmps = 0;
2777 if (Cond->getOpcode() == Instruction::Or) {
2778 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, DL, true,
2780 } else if (Cond->getOpcode() == Instruction::And) {
2781 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, DL, false,
2783 TrueWhenEqual = false;
2786 // If we didn't have a multiply compared value, fail.
2787 if (!CompVal) return false;
2789 // Avoid turning single icmps into a switch.
2793 // There might be duplicate constants in the list, which the switch
2794 // instruction can't handle, remove them now.
2795 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2796 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2798 // If Extra was used, we require at least two switch values to do the
2799 // transformation. A switch with one value is just an cond branch.
2800 if (ExtraCase && Values.size() < 2) return false;
2802 // TODO: Preserve branch weight metadata, similarly to how
2803 // FoldValueComparisonIntoPredecessors preserves it.
2805 // Figure out which block is which destination.
2806 BasicBlock *DefaultBB = BI->getSuccessor(1);
2807 BasicBlock *EdgeBB = BI->getSuccessor(0);
2808 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2810 BasicBlock *BB = BI->getParent();
2812 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2813 << " cases into SWITCH. BB is:\n" << *BB);
2815 // If there are any extra values that couldn't be folded into the switch
2816 // then we evaluate them with an explicit branch first. Split the block
2817 // right before the condbr to handle it.
2819 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2820 // Remove the uncond branch added to the old block.
2821 TerminatorInst *OldTI = BB->getTerminator();
2822 Builder.SetInsertPoint(OldTI);
2825 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2827 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2829 OldTI->eraseFromParent();
2831 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2832 // for the edge we just added.
2833 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2835 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2836 << "\nEXTRABB = " << *BB);
2840 Builder.SetInsertPoint(BI);
2841 // Convert pointer to int before we switch.
2842 if (CompVal->getType()->isPointerTy()) {
2843 assert(DL && "Cannot switch on pointer without DataLayout");
2844 CompVal = Builder.CreatePtrToInt(CompVal,
2845 DL->getIntPtrType(CompVal->getType()),
2849 // Create the new switch instruction now.
2850 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2852 // Add all of the 'cases' to the switch instruction.
2853 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2854 New->addCase(Values[i], EdgeBB);
2856 // We added edges from PI to the EdgeBB. As such, if there were any
2857 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2858 // the number of edges added.
2859 for (BasicBlock::iterator BBI = EdgeBB->begin();
2860 isa<PHINode>(BBI); ++BBI) {
2861 PHINode *PN = cast<PHINode>(BBI);
2862 Value *InVal = PN->getIncomingValueForBlock(BB);
2863 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2864 PN->addIncoming(InVal, BB);
2867 // Erase the old branch instruction.
2868 EraseTerminatorInstAndDCECond(BI);
2870 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2874 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2875 // If this is a trivial landing pad that just continues unwinding the caught
2876 // exception then zap the landing pad, turning its invokes into calls.
2877 BasicBlock *BB = RI->getParent();
2878 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2879 if (RI->getValue() != LPInst)
2880 // Not a landing pad, or the resume is not unwinding the exception that
2881 // caused control to branch here.
2884 // Check that there are no other instructions except for debug intrinsics.
2885 BasicBlock::iterator I = LPInst, E = RI;
2887 if (!isa<DbgInfoIntrinsic>(I))
2890 // Turn all invokes that unwind here into calls and delete the basic block.
2891 bool InvokeRequiresTableEntry = false;
2892 bool Changed = false;
2893 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2894 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2896 if (II->hasFnAttr(Attribute::UWTable)) {
2897 // Don't remove an `invoke' instruction if the ABI requires an entry into
2899 InvokeRequiresTableEntry = true;
2903 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2905 // Insert a call instruction before the invoke.
2906 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2908 Call->setCallingConv(II->getCallingConv());
2909 Call->setAttributes(II->getAttributes());
2910 Call->setDebugLoc(II->getDebugLoc());
2912 // Anything that used the value produced by the invoke instruction now uses
2913 // the value produced by the call instruction. Note that we do this even
2914 // for void functions and calls with no uses so that the callgraph edge is
2916 II->replaceAllUsesWith(Call);
2917 BB->removePredecessor(II->getParent());
2919 // Insert a branch to the normal destination right before the invoke.
2920 BranchInst::Create(II->getNormalDest(), II);
2922 // Finally, delete the invoke instruction!
2923 II->eraseFromParent();
2927 if (!InvokeRequiresTableEntry)
2928 // The landingpad is now unreachable. Zap it.
2929 BB->eraseFromParent();
2934 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2935 BasicBlock *BB = RI->getParent();
2936 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2938 // Find predecessors that end with branches.
2939 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2940 SmallVector<BranchInst*, 8> CondBranchPreds;
2941 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2942 BasicBlock *P = *PI;
2943 TerminatorInst *PTI = P->getTerminator();
2944 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2945 if (BI->isUnconditional())
2946 UncondBranchPreds.push_back(P);
2948 CondBranchPreds.push_back(BI);
2952 // If we found some, do the transformation!
2953 if (!UncondBranchPreds.empty() && DupRet) {
2954 while (!UncondBranchPreds.empty()) {
2955 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
2956 DEBUG(dbgs() << "FOLDING: " << *BB
2957 << "INTO UNCOND BRANCH PRED: " << *Pred);
2958 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
2961 // If we eliminated all predecessors of the block, delete the block now.
2962 if (pred_begin(BB) == pred_end(BB))
2963 // We know there are no successors, so just nuke the block.
2964 BB->eraseFromParent();
2969 // Check out all of the conditional branches going to this return
2970 // instruction. If any of them just select between returns, change the
2971 // branch itself into a select/return pair.
2972 while (!CondBranchPreds.empty()) {
2973 BranchInst *BI = CondBranchPreds.pop_back_val();
2975 // Check to see if the non-BB successor is also a return block.
2976 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
2977 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
2978 SimplifyCondBranchToTwoReturns(BI, Builder))
2984 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
2985 BasicBlock *BB = UI->getParent();
2987 bool Changed = false;
2989 // If there are any instructions immediately before the unreachable that can
2990 // be removed, do so.
2991 while (UI != BB->begin()) {
2992 BasicBlock::iterator BBI = UI;
2994 // Do not delete instructions that can have side effects which might cause
2995 // the unreachable to not be reachable; specifically, calls and volatile
2996 // operations may have this effect.
2997 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
2999 if (BBI->mayHaveSideEffects()) {
3000 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
3001 if (SI->isVolatile())
3003 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
3004 if (LI->isVolatile())
3006 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
3007 if (RMWI->isVolatile())
3009 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
3010 if (CXI->isVolatile())
3012 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
3013 !isa<LandingPadInst>(BBI)) {
3016 // Note that deleting LandingPad's here is in fact okay, although it
3017 // involves a bit of subtle reasoning. If this inst is a LandingPad,
3018 // all the predecessors of this block will be the unwind edges of Invokes,
3019 // and we can therefore guarantee this block will be erased.
3022 // Delete this instruction (any uses are guaranteed to be dead)
3023 if (!BBI->use_empty())
3024 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
3025 BBI->eraseFromParent();
3029 // If the unreachable instruction is the first in the block, take a gander
3030 // at all of the predecessors of this instruction, and simplify them.
3031 if (&BB->front() != UI) return Changed;
3033 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
3034 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
3035 TerminatorInst *TI = Preds[i]->getTerminator();
3036 IRBuilder<> Builder(TI);
3037 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3038 if (BI->isUnconditional()) {
3039 if (BI->getSuccessor(0) == BB) {
3040 new UnreachableInst(TI->getContext(), TI);
3041 TI->eraseFromParent();
3045 if (BI->getSuccessor(0) == BB) {
3046 Builder.CreateBr(BI->getSuccessor(1));
3047 EraseTerminatorInstAndDCECond(BI);
3048 } else if (BI->getSuccessor(1) == BB) {
3049 Builder.CreateBr(BI->getSuccessor(0));
3050 EraseTerminatorInstAndDCECond(BI);
3054 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3055 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3057 if (i.getCaseSuccessor() == BB) {
3058 BB->removePredecessor(SI->getParent());
3063 // If the default value is unreachable, figure out the most popular
3064 // destination and make it the default.
3065 if (SI->getDefaultDest() == BB) {
3066 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
3067 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3069 std::pair<unsigned, unsigned> &entry =
3070 Popularity[i.getCaseSuccessor()];
3071 if (entry.first == 0) {
3073 entry.second = i.getCaseIndex();
3079 // Find the most popular block.
3080 unsigned MaxPop = 0;
3081 unsigned MaxIndex = 0;
3082 BasicBlock *MaxBlock = nullptr;
3083 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
3084 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
3085 if (I->second.first > MaxPop ||
3086 (I->second.first == MaxPop && MaxIndex > I->second.second)) {
3087 MaxPop = I->second.first;
3088 MaxIndex = I->second.second;
3089 MaxBlock = I->first;
3093 // Make this the new default, allowing us to delete any explicit
3095 SI->setDefaultDest(MaxBlock);
3098 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
3100 if (isa<PHINode>(MaxBlock->begin()))
3101 for (unsigned i = 0; i != MaxPop-1; ++i)
3102 MaxBlock->removePredecessor(SI->getParent());
3104 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3106 if (i.getCaseSuccessor() == MaxBlock) {
3112 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
3113 if (II->getUnwindDest() == BB) {
3114 // Convert the invoke to a call instruction. This would be a good
3115 // place to note that the call does not throw though.
3116 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
3117 II->removeFromParent(); // Take out of symbol table
3119 // Insert the call now...
3120 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
3121 Builder.SetInsertPoint(BI);
3122 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
3123 Args, II->getName());
3124 CI->setCallingConv(II->getCallingConv());
3125 CI->setAttributes(II->getAttributes());
3126 // If the invoke produced a value, the call does now instead.
3127 II->replaceAllUsesWith(CI);
3134 // If this block is now dead, remove it.
3135 if (pred_begin(BB) == pred_end(BB) &&
3136 BB != &BB->getParent()->getEntryBlock()) {
3137 // We know there are no successors, so just nuke the block.
3138 BB->eraseFromParent();
3145 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
3146 /// integer range comparison into a sub, an icmp and a branch.
3147 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3148 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3150 // Make sure all cases point to the same destination and gather the values.
3151 SmallVector<ConstantInt *, 16> Cases;
3152 SwitchInst::CaseIt I = SI->case_begin();
3153 Cases.push_back(I.getCaseValue());
3154 SwitchInst::CaseIt PrevI = I++;
3155 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
3156 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
3158 Cases.push_back(I.getCaseValue());
3160 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
3162 // Sort the case values, then check if they form a range we can transform.
3163 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3164 for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
3165 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
3169 Constant *Offset = ConstantExpr::getNeg(Cases.back());
3170 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
3172 Value *Sub = SI->getCondition();
3173 if (!Offset->isNullValue())
3174 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
3176 // If NumCases overflowed, then all possible values jump to the successor.
3177 if (NumCases->isNullValue() && SI->getNumCases() != 0)
3178 Cmp = ConstantInt::getTrue(SI->getContext());
3180 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3181 BranchInst *NewBI = Builder.CreateCondBr(
3182 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
3184 // Update weight for the newly-created conditional branch.
3185 SmallVector<uint64_t, 8> Weights;
3186 bool HasWeights = HasBranchWeights(SI);
3188 GetBranchWeights(SI, Weights);
3189 if (Weights.size() == 1 + SI->getNumCases()) {
3190 // Combine all weights for the cases to be the true weight of NewBI.
3191 // We assume that the sum of all weights for a Terminator can fit into 32
3193 uint32_t NewTrueWeight = 0;
3194 for (unsigned I = 1, E = Weights.size(); I != E; ++I)
3195 NewTrueWeight += (uint32_t)Weights[I];
3196 NewBI->setMetadata(LLVMContext::MD_prof,
3197 MDBuilder(SI->getContext()).
3198 createBranchWeights(NewTrueWeight,
3199 (uint32_t)Weights[0]));
3203 // Prune obsolete incoming values off the successor's PHI nodes.
3204 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
3205 isa<PHINode>(BBI); ++BBI) {
3206 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
3207 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3209 SI->eraseFromParent();
3214 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
3215 /// and use it to remove dead cases.
3216 static bool EliminateDeadSwitchCases(SwitchInst *SI, const DataLayout *DL,
3217 AssumptionTracker *AT) {
3218 Value *Cond = SI->getCondition();
3219 unsigned Bits = Cond->getType()->getIntegerBitWidth();
3220 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3221 computeKnownBits(Cond, KnownZero, KnownOne, DL, 0, AT, SI);
3223 // Gather dead cases.
3224 SmallVector<ConstantInt*, 8> DeadCases;
3225 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3226 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3227 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3228 DeadCases.push_back(I.getCaseValue());
3229 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3230 << I.getCaseValue() << "' is dead.\n");
3234 SmallVector<uint64_t, 8> Weights;
3235 bool HasWeight = HasBranchWeights(SI);
3237 GetBranchWeights(SI, Weights);
3238 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3241 // Remove dead cases from the switch.
3242 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3243 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3244 assert(Case != SI->case_default() &&
3245 "Case was not found. Probably mistake in DeadCases forming.");
3247 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3251 // Prune unused values from PHI nodes.
3252 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3253 SI->removeCase(Case);
3255 if (HasWeight && Weights.size() >= 2) {
3256 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3257 SI->setMetadata(LLVMContext::MD_prof,
3258 MDBuilder(SI->getParent()->getContext()).
3259 createBranchWeights(MDWeights));
3262 return !DeadCases.empty();
3265 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
3266 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3267 /// by an unconditional branch), look at the phi node for BB in the successor
3268 /// block and see if the incoming value is equal to CaseValue. If so, return
3269 /// the phi node, and set PhiIndex to BB's index in the phi node.
3270 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3273 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3274 return nullptr; // BB must be empty to be a candidate for simplification.
3275 if (!BB->getSinglePredecessor())
3276 return nullptr; // BB must be dominated by the switch.
3278 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3279 if (!Branch || !Branch->isUnconditional())
3280 return nullptr; // Terminator must be unconditional branch.
3282 BasicBlock *Succ = Branch->getSuccessor(0);
3284 BasicBlock::iterator I = Succ->begin();
3285 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3286 int Idx = PHI->getBasicBlockIndex(BB);
3287 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3289 Value *InValue = PHI->getIncomingValue(Idx);
3290 if (InValue != CaseValue) continue;
3299 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
3300 /// instruction to a phi node dominated by the switch, if that would mean that
3301 /// some of the destination blocks of the switch can be folded away.
3302 /// Returns true if a change is made.
3303 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3304 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3305 ForwardingNodesMap ForwardingNodes;
3307 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3308 ConstantInt *CaseValue = I.getCaseValue();
3309 BasicBlock *CaseDest = I.getCaseSuccessor();
3312 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3316 ForwardingNodes[PHI].push_back(PhiIndex);
3319 bool Changed = false;
3321 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3322 E = ForwardingNodes.end(); I != E; ++I) {
3323 PHINode *Phi = I->first;
3324 SmallVectorImpl<int> &Indexes = I->second;
3326 if (Indexes.size() < 2) continue;
3328 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3329 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3336 /// ValidLookupTableConstant - Return true if the backend will be able to handle
3337 /// initializing an array of constants like C.
3338 static bool ValidLookupTableConstant(Constant *C) {
3339 if (C->isThreadDependent())
3341 if (C->isDLLImportDependent())
3344 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3345 return CE->isGEPWithNoNotionalOverIndexing();
3347 return isa<ConstantFP>(C) ||
3348 isa<ConstantInt>(C) ||
3349 isa<ConstantPointerNull>(C) ||
3350 isa<GlobalValue>(C) ||
3354 /// LookupConstant - If V is a Constant, return it. Otherwise, try to look up
3355 /// its constant value in ConstantPool, returning 0 if it's not there.
3356 static Constant *LookupConstant(Value *V,
3357 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3358 if (Constant *C = dyn_cast<Constant>(V))
3360 return ConstantPool.lookup(V);
3363 /// ConstantFold - Try to fold instruction I into a constant. This works for
3364 /// simple instructions such as binary operations where both operands are
3365 /// constant or can be replaced by constants from the ConstantPool. Returns the
3366 /// resulting constant on success, 0 otherwise.
3368 ConstantFold(Instruction *I,
3369 const SmallDenseMap<Value *, Constant *> &ConstantPool,
3370 const DataLayout *DL) {
3371 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3372 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3375 if (A->isAllOnesValue())
3376 return LookupConstant(Select->getTrueValue(), ConstantPool);
3377 if (A->isNullValue())
3378 return LookupConstant(Select->getFalseValue(), ConstantPool);
3382 SmallVector<Constant *, 4> COps;
3383 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
3384 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
3390 if (CmpInst *Cmp = dyn_cast<CmpInst>(I))
3391 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
3394 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), COps, DL);
3397 /// GetCaseResults - Try to determine the resulting constant values in phi nodes
3398 /// at the common destination basic block, *CommonDest, for one of the case
3399 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3400 /// case), of a switch instruction SI.
3402 GetCaseResults(SwitchInst *SI,
3403 ConstantInt *CaseVal,
3404 BasicBlock *CaseDest,
3405 BasicBlock **CommonDest,
3406 SmallVectorImpl<std::pair<PHINode *, Constant *> > &Res,
3407 const DataLayout *DL) {
3408 // The block from which we enter the common destination.
3409 BasicBlock *Pred = SI->getParent();
3411 // If CaseDest is empty except for some side-effect free instructions through
3412 // which we can constant-propagate the CaseVal, continue to its successor.
3413 SmallDenseMap<Value*, Constant*> ConstantPool;
3414 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3415 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3417 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3418 // If the terminator is a simple branch, continue to the next block.
3419 if (T->getNumSuccessors() != 1)
3422 CaseDest = T->getSuccessor(0);
3423 } else if (isa<DbgInfoIntrinsic>(I)) {
3424 // Skip debug intrinsic.
3426 } else if (Constant *C = ConstantFold(I, ConstantPool, DL)) {
3427 // Instruction is side-effect free and constant.
3428 ConstantPool.insert(std::make_pair(I, C));
3434 // If we did not have a CommonDest before, use the current one.
3436 *CommonDest = CaseDest;
3437 // If the destination isn't the common one, abort.
3438 if (CaseDest != *CommonDest)
3441 // Get the values for this case from phi nodes in the destination block.
3442 BasicBlock::iterator I = (*CommonDest)->begin();
3443 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3444 int Idx = PHI->getBasicBlockIndex(Pred);
3448 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3453 // Note: If the constant comes from constant-propagating the case value
3454 // through the CaseDest basic block, it will be safe to remove the
3455 // instructions in that block. They cannot be used (except in the phi nodes
3456 // we visit) outside CaseDest, because that block does not dominate its
3457 // successor. If it did, we would not be in this phi node.
3459 // Be conservative about which kinds of constants we support.
3460 if (!ValidLookupTableConstant(ConstVal))
3463 Res.push_back(std::make_pair(PHI, ConstVal));
3466 return Res.size() > 0;
3469 // MapCaseToResult - Helper function used to
3470 // add CaseVal to the list of cases that generate Result.
3471 static void MapCaseToResult(ConstantInt *CaseVal,
3472 SwitchCaseResultVectorTy &UniqueResults,
3474 for (auto &I : UniqueResults) {
3475 if (I.first == Result) {
3476 I.second.push_back(CaseVal);
3480 UniqueResults.push_back(std::make_pair(Result,
3481 SmallVector<ConstantInt*, 4>(1, CaseVal)));
3484 // InitializeUniqueCases - Helper function that initializes a map containing
3485 // results for the PHI node of the common destination block for a switch
3486 // instruction. Returns false if multiple PHI nodes have been found or if
3487 // there is not a common destination block for the switch.
3488 static bool InitializeUniqueCases(
3489 SwitchInst *SI, const DataLayout *DL, PHINode *&PHI,
3490 BasicBlock *&CommonDest,
3491 SwitchCaseResultVectorTy &UniqueResults,
3492 Constant *&DefaultResult) {
3493 for (auto &I : SI->cases()) {
3494 ConstantInt *CaseVal = I.getCaseValue();
3496 // Resulting value at phi nodes for this case value.
3497 SwitchCaseResultsTy Results;
3498 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
3502 // Only one value per case is permitted
3503 if (Results.size() > 1)
3505 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
3507 // Check the PHI consistency.
3509 PHI = Results[0].first;
3510 else if (PHI != Results[0].first)
3513 // Find the default result value.
3514 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
3515 BasicBlock *DefaultDest = SI->getDefaultDest();
3516 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
3518 // If the default value is not found abort unless the default destination
3521 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
3522 if ((!DefaultResult &&
3523 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
3529 // ConvertTwoCaseSwitch - Helper function that checks if it is possible to
3530 // transform a switch with only two cases (or two cases + default)
3531 // that produces a result into a value select.
3534 // case 10: %0 = icmp eq i32 %a, 10
3535 // return 10; %1 = select i1 %0, i32 10, i32 4
3536 // case 20: ----> %2 = icmp eq i32 %a, 20
3537 // return 2; %3 = select i1 %2, i32 2, i32 %1
3542 ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
3543 Constant *DefaultResult, Value *Condition,
3544 IRBuilder<> &Builder) {
3545 assert(ResultVector.size() == 2 &&
3546 "We should have exactly two unique results at this point");
3547 // If we are selecting between only two cases transform into a simple
3548 // select or a two-way select if default is possible.
3549 if (ResultVector[0].second.size() == 1 &&
3550 ResultVector[1].second.size() == 1) {
3551 ConstantInt *const FirstCase = ResultVector[0].second[0];
3552 ConstantInt *const SecondCase = ResultVector[1].second[0];
3554 bool DefaultCanTrigger = DefaultResult;
3555 Value *SelectValue = ResultVector[1].first;
3556 if (DefaultCanTrigger) {
3557 Value *const ValueCompare =
3558 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
3559 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
3560 DefaultResult, "switch.select");
3562 Value *const ValueCompare =
3563 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
3564 return Builder.CreateSelect(ValueCompare, ResultVector[0].first, SelectValue,
3571 // RemoveSwitchAfterSelectConversion - Helper function to cleanup a switch
3572 // instruction that has been converted into a select, fixing up PHI nodes and
3574 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
3576 IRBuilder<> &Builder) {
3577 BasicBlock *SelectBB = SI->getParent();
3578 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
3579 PHI->removeIncomingValue(SelectBB);
3580 PHI->addIncoming(SelectValue, SelectBB);
3582 Builder.CreateBr(PHI->getParent());
3584 // Remove the switch.
3585 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
3586 BasicBlock *Succ = SI->getSuccessor(i);
3588 if (Succ == PHI->getParent())
3590 Succ->removePredecessor(SelectBB);
3592 SI->eraseFromParent();
3595 /// SwitchToSelect - If the switch is only used to initialize one or more
3596 /// phi nodes in a common successor block with only two different
3597 /// constant values, replace the switch with select.
3598 static bool SwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
3599 const DataLayout *DL, AssumptionTracker *AT) {
3600 Value *const Cond = SI->getCondition();
3601 PHINode *PHI = nullptr;
3602 BasicBlock *CommonDest = nullptr;
3603 Constant *DefaultResult;
3604 SwitchCaseResultVectorTy UniqueResults;
3605 // Collect all the cases that will deliver the same value from the switch.
3606 if (!InitializeUniqueCases(SI, DL, PHI, CommonDest, UniqueResults,
3609 // Selects choose between maximum two values.
3610 if (UniqueResults.size() != 2)
3612 assert(PHI != nullptr && "PHI for value select not found");
3614 Builder.SetInsertPoint(SI);
3615 Value *SelectValue = ConvertTwoCaseSwitch(
3617 DefaultResult, Cond, Builder);
3619 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
3622 // The switch couldn't be converted into a select.
3627 /// SwitchLookupTable - This class represents a lookup table that can be used
3628 /// to replace a switch.
3629 class SwitchLookupTable {
3631 /// SwitchLookupTable - Create a lookup table to use as a switch replacement
3632 /// with the contents of Values, using DefaultValue to fill any holes in the
3634 SwitchLookupTable(Module &M,
3636 ConstantInt *Offset,
3637 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3638 Constant *DefaultValue,
3639 const DataLayout *DL);
3641 /// BuildLookup - Build instructions with Builder to retrieve the value at
3642 /// the position given by Index in the lookup table.
3643 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3645 /// WouldFitInRegister - Return true if a table with TableSize elements of
3646 /// type ElementType would fit in a target-legal register.
3647 static bool WouldFitInRegister(const DataLayout *DL,
3649 const Type *ElementType);
3652 // Depending on the contents of the table, it can be represented in
3655 // For tables where each element contains the same value, we just have to
3656 // store that single value and return it for each lookup.
3659 // For small tables with integer elements, we can pack them into a bitmap
3660 // that fits into a target-legal register. Values are retrieved by
3661 // shift and mask operations.
3664 // The table is stored as an array of values. Values are retrieved by load
3665 // instructions from the table.
3669 // For SingleValueKind, this is the single value.
3670 Constant *SingleValue;
3672 // For BitMapKind, this is the bitmap.
3673 ConstantInt *BitMap;
3674 IntegerType *BitMapElementTy;
3676 // For ArrayKind, this is the array.
3677 GlobalVariable *Array;
3681 SwitchLookupTable::SwitchLookupTable(Module &M,
3683 ConstantInt *Offset,
3684 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3685 Constant *DefaultValue,
3686 const DataLayout *DL)
3687 : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
3689 assert(Values.size() && "Can't build lookup table without values!");
3690 assert(TableSize >= Values.size() && "Can't fit values in table!");
3692 // If all values in the table are equal, this is that value.
3693 SingleValue = Values.begin()->second;
3695 Type *ValueType = Values.begin()->second->getType();
3697 // Build up the table contents.
3698 SmallVector<Constant*, 64> TableContents(TableSize);
3699 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3700 ConstantInt *CaseVal = Values[I].first;
3701 Constant *CaseRes = Values[I].second;
3702 assert(CaseRes->getType() == ValueType);
3704 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3706 TableContents[Idx] = CaseRes;
3708 if (CaseRes != SingleValue)
3709 SingleValue = nullptr;
3712 // Fill in any holes in the table with the default result.
3713 if (Values.size() < TableSize) {
3714 assert(DefaultValue &&
3715 "Need a default value to fill the lookup table holes.");
3716 assert(DefaultValue->getType() == ValueType);
3717 for (uint64_t I = 0; I < TableSize; ++I) {
3718 if (!TableContents[I])
3719 TableContents[I] = DefaultValue;
3722 if (DefaultValue != SingleValue)
3723 SingleValue = nullptr;
3726 // If each element in the table contains the same value, we only need to store
3727 // that single value.
3729 Kind = SingleValueKind;
3733 // If the type is integer and the table fits in a register, build a bitmap.
3734 if (WouldFitInRegister(DL, TableSize, ValueType)) {
3735 IntegerType *IT = cast<IntegerType>(ValueType);
3736 APInt TableInt(TableSize * IT->getBitWidth(), 0);
3737 for (uint64_t I = TableSize; I > 0; --I) {
3738 TableInt <<= IT->getBitWidth();
3739 // Insert values into the bitmap. Undef values are set to zero.
3740 if (!isa<UndefValue>(TableContents[I - 1])) {
3741 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
3742 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
3745 BitMap = ConstantInt::get(M.getContext(), TableInt);
3746 BitMapElementTy = IT;
3752 // Store the table in an array.
3753 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
3754 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3756 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3757 GlobalVariable::PrivateLinkage,
3760 Array->setUnnamedAddr(true);
3764 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
3766 case SingleValueKind:
3769 // Type of the bitmap (e.g. i59).
3770 IntegerType *MapTy = BitMap->getType();
3772 // Cast Index to the same type as the bitmap.
3773 // Note: The Index is <= the number of elements in the table, so
3774 // truncating it to the width of the bitmask is safe.
3775 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
3777 // Multiply the shift amount by the element width.
3778 ShiftAmt = Builder.CreateMul(ShiftAmt,
3779 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
3783 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
3784 "switch.downshift");
3786 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
3790 // Make sure the table index will not overflow when treated as signed.
3791 IntegerType *IT = cast<IntegerType>(Index->getType());
3792 uint64_t TableSize = Array->getInitializer()->getType()
3793 ->getArrayNumElements();
3794 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
3795 Index = Builder.CreateZExt(Index,
3796 IntegerType::get(IT->getContext(),
3797 IT->getBitWidth() + 1),
3798 "switch.tableidx.zext");
3800 Value *GEPIndices[] = { Builder.getInt32(0), Index };
3801 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices,
3803 return Builder.CreateLoad(GEP, "switch.load");
3806 llvm_unreachable("Unknown lookup table kind!");
3809 bool SwitchLookupTable::WouldFitInRegister(const DataLayout *DL,
3811 const Type *ElementType) {
3814 const IntegerType *IT = dyn_cast<IntegerType>(ElementType);
3817 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
3818 // are <= 15, we could try to narrow the type.
3820 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
3821 if (TableSize >= UINT_MAX/IT->getBitWidth())
3823 return DL->fitsInLegalInteger(TableSize * IT->getBitWidth());
3826 /// ShouldBuildLookupTable - Determine whether a lookup table should be built
3827 /// for this switch, based on the number of cases, size of the table and the
3828 /// types of the results.
3829 static bool ShouldBuildLookupTable(SwitchInst *SI,
3831 const TargetTransformInfo &TTI,
3832 const DataLayout *DL,
3833 const SmallDenseMap<PHINode*, Type*>& ResultTypes) {
3834 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
3835 return false; // TableSize overflowed, or mul below might overflow.
3837 bool AllTablesFitInRegister = true;
3838 bool HasIllegalType = false;
3839 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(),
3840 E = ResultTypes.end(); I != E; ++I) {
3841 Type *Ty = I->second;
3843 // Saturate this flag to true.
3844 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
3846 // Saturate this flag to false.
3847 AllTablesFitInRegister = AllTablesFitInRegister &&
3848 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
3850 // If both flags saturate, we're done. NOTE: This *only* works with
3851 // saturating flags, and all flags have to saturate first due to the
3852 // non-deterministic behavior of iterating over a dense map.
3853 if (HasIllegalType && !AllTablesFitInRegister)
3857 // If each table would fit in a register, we should build it anyway.
3858 if (AllTablesFitInRegister)
3861 // Don't build a table that doesn't fit in-register if it has illegal types.
3865 // The table density should be at least 40%. This is the same criterion as for
3866 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3867 // FIXME: Find the best cut-off.
3868 return SI->getNumCases() * 10 >= TableSize * 4;
3871 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3872 /// phi nodes in a common successor block with different constant values,
3873 /// replace the switch with lookup tables.
3874 static bool SwitchToLookupTable(SwitchInst *SI,
3875 IRBuilder<> &Builder,
3876 const TargetTransformInfo &TTI,
3877 const DataLayout* DL) {
3878 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3880 // Only build lookup table when we have a target that supports it.
3881 if (!TTI.shouldBuildLookupTables())
3884 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3885 // split off a dense part and build a lookup table for that.
3887 // FIXME: This creates arrays of GEPs to constant strings, which means each
3888 // GEP needs a runtime relocation in PIC code. We should just build one big
3889 // string and lookup indices into that.
3891 // Ignore switches with less than three cases. Lookup tables will not make them
3892 // faster, so we don't analyze them.
3893 if (SI->getNumCases() < 3)
3896 // Figure out the corresponding result for each case value and phi node in the
3897 // common destination, as well as the the min and max case values.
3898 assert(SI->case_begin() != SI->case_end());
3899 SwitchInst::CaseIt CI = SI->case_begin();
3900 ConstantInt *MinCaseVal = CI.getCaseValue();
3901 ConstantInt *MaxCaseVal = CI.getCaseValue();
3903 BasicBlock *CommonDest = nullptr;
3904 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
3905 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
3906 SmallDenseMap<PHINode*, Constant*> DefaultResults;
3907 SmallDenseMap<PHINode*, Type*> ResultTypes;
3908 SmallVector<PHINode*, 4> PHIs;
3910 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
3911 ConstantInt *CaseVal = CI.getCaseValue();
3912 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
3913 MinCaseVal = CaseVal;
3914 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
3915 MaxCaseVal = CaseVal;
3917 // Resulting value at phi nodes for this case value.
3918 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
3920 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
3924 // Append the result from this case to the list for each phi.
3925 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) {
3926 if (!ResultLists.count(I->first))
3927 PHIs.push_back(I->first);
3928 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second));
3932 // Keep track of the result types.
3933 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
3934 PHINode *PHI = PHIs[I];
3935 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
3938 uint64_t NumResults = ResultLists[PHIs[0]].size();
3939 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
3940 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
3941 bool TableHasHoles = (NumResults < TableSize);
3943 // If the table has holes, we need a constant result for the default case
3944 // or a bitmask that fits in a register.
3945 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
3946 bool HasDefaultResults = false;
3947 if (TableHasHoles) {
3948 HasDefaultResults = GetCaseResults(SI, nullptr, SI->getDefaultDest(),
3949 &CommonDest, DefaultResultsList, DL);
3951 bool NeedMask = (TableHasHoles && !HasDefaultResults);
3953 // As an extra penalty for the validity test we require more cases.
3954 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
3956 if (!(DL && DL->fitsInLegalInteger(TableSize)))
3960 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) {
3961 PHINode *PHI = DefaultResultsList[I].first;
3962 Constant *Result = DefaultResultsList[I].second;
3963 DefaultResults[PHI] = Result;
3966 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
3969 // Create the BB that does the lookups.
3970 Module &Mod = *CommonDest->getParent()->getParent();
3971 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
3973 CommonDest->getParent(),
3976 // Compute the table index value.
3977 Builder.SetInsertPoint(SI);
3978 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
3981 // Compute the maximum table size representable by the integer type we are
3983 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
3984 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
3985 assert(MaxTableSize >= TableSize &&
3986 "It is impossible for a switch to have more entries than the max "
3987 "representable value of its input integer type's size.");
3989 // If we have a fully covered lookup table, unconditionally branch to the
3990 // lookup table BB. Otherwise, check if the condition value is within the case
3991 // range. If it is so, branch to the new BB. Otherwise branch to SI's default
3993 const bool GeneratingCoveredLookupTable = MaxTableSize == TableSize;
3994 if (GeneratingCoveredLookupTable) {
3995 Builder.CreateBr(LookupBB);
3996 // We cached PHINodes in PHIs, to avoid accessing deleted PHINodes later,
3997 // do not delete PHINodes here.
3998 SI->getDefaultDest()->removePredecessor(SI->getParent(),
3999 true/*DontDeleteUselessPHIs*/);
4001 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
4002 MinCaseVal->getType(), TableSize));
4003 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
4006 // Populate the BB that does the lookups.
4007 Builder.SetInsertPoint(LookupBB);
4010 // Before doing the lookup we do the hole check.
4011 // The LookupBB is therefore re-purposed to do the hole check
4012 // and we create a new LookupBB.
4013 BasicBlock *MaskBB = LookupBB;
4014 MaskBB->setName("switch.hole_check");
4015 LookupBB = BasicBlock::Create(Mod.getContext(),
4017 CommonDest->getParent(),
4020 // Build bitmask; fill in a 1 bit for every case.
4021 APInt MaskInt(TableSize, 0);
4022 APInt One(TableSize, 1);
4023 const ResultListTy &ResultList = ResultLists[PHIs[0]];
4024 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
4025 uint64_t Idx = (ResultList[I].first->getValue() -
4026 MinCaseVal->getValue()).getLimitedValue();
4027 MaskInt |= One << Idx;
4029 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
4031 // Get the TableIndex'th bit of the bitmask.
4032 // If this bit is 0 (meaning hole) jump to the default destination,
4033 // else continue with table lookup.
4034 IntegerType *MapTy = TableMask->getType();
4035 Value *MaskIndex = Builder.CreateZExtOrTrunc(TableIndex, MapTy,
4036 "switch.maskindex");
4037 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex,
4039 Value *LoBit = Builder.CreateTrunc(Shifted,
4040 Type::getInt1Ty(Mod.getContext()),
4042 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
4044 Builder.SetInsertPoint(LookupBB);
4045 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
4048 bool ReturnedEarly = false;
4049 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
4050 PHINode *PHI = PHIs[I];
4052 // If using a bitmask, use any value to fill the lookup table holes.
4053 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
4054 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI],
4057 Value *Result = Table.BuildLookup(TableIndex, Builder);
4059 // If the result is used to return immediately from the function, we want to
4060 // do that right here.
4061 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
4062 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
4063 Builder.CreateRet(Result);
4064 ReturnedEarly = true;
4068 PHI->addIncoming(Result, LookupBB);
4072 Builder.CreateBr(CommonDest);
4074 // Remove the switch.
4075 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4076 BasicBlock *Succ = SI->getSuccessor(i);
4078 if (Succ == SI->getDefaultDest())
4080 Succ->removePredecessor(SI->getParent());
4082 SI->eraseFromParent();
4086 ++NumLookupTablesHoles;
4090 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
4091 BasicBlock *BB = SI->getParent();
4093 if (isValueEqualityComparison(SI)) {
4094 // If we only have one predecessor, and if it is a branch on this value,
4095 // see if that predecessor totally determines the outcome of this switch.
4096 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4097 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
4098 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4100 Value *Cond = SI->getCondition();
4101 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
4102 if (SimplifySwitchOnSelect(SI, Select))
4103 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4105 // If the block only contains the switch, see if we can fold the block
4106 // away into any preds.
4107 BasicBlock::iterator BBI = BB->begin();
4108 // Ignore dbg intrinsics.
4109 while (isa<DbgInfoIntrinsic>(BBI))
4112 if (FoldValueComparisonIntoPredecessors(SI, Builder))
4113 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4116 // Try to transform the switch into an icmp and a branch.
4117 if (TurnSwitchRangeIntoICmp(SI, Builder))
4118 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4120 // Remove unreachable cases.
4121 if (EliminateDeadSwitchCases(SI, DL, AT))
4122 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4124 if (SwitchToSelect(SI, Builder, DL, AT))
4125 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4127 if (ForwardSwitchConditionToPHI(SI))
4128 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4130 if (SwitchToLookupTable(SI, Builder, TTI, DL))
4131 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4136 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
4137 BasicBlock *BB = IBI->getParent();
4138 bool Changed = false;
4140 // Eliminate redundant destinations.
4141 SmallPtrSet<Value *, 8> Succs;
4142 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
4143 BasicBlock *Dest = IBI->getDestination(i);
4144 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) {
4145 Dest->removePredecessor(BB);
4146 IBI->removeDestination(i);
4152 if (IBI->getNumDestinations() == 0) {
4153 // If the indirectbr has no successors, change it to unreachable.
4154 new UnreachableInst(IBI->getContext(), IBI);
4155 EraseTerminatorInstAndDCECond(IBI);
4159 if (IBI->getNumDestinations() == 1) {
4160 // If the indirectbr has one successor, change it to a direct branch.
4161 BranchInst::Create(IBI->getDestination(0), IBI);
4162 EraseTerminatorInstAndDCECond(IBI);
4166 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
4167 if (SimplifyIndirectBrOnSelect(IBI, SI))
4168 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4173 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
4174 BasicBlock *BB = BI->getParent();
4176 if (SinkCommon && SinkThenElseCodeToEnd(BI))
4179 // If the Terminator is the only non-phi instruction, simplify the block.
4180 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg();
4181 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
4182 TryToSimplifyUncondBranchFromEmptyBlock(BB))
4185 // If the only instruction in the block is a seteq/setne comparison
4186 // against a constant, try to simplify the block.
4187 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
4188 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
4189 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
4191 if (I->isTerminator() &&
4192 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI,
4193 BonusInstThreshold, DL, AT))
4197 // If this basic block is ONLY a compare and a branch, and if a predecessor
4198 // branches to us and our successor, fold the comparison into the
4199 // predecessor and use logical operations to update the incoming value
4200 // for PHI nodes in common successor.
4201 if (FoldBranchToCommonDest(BI, DL, BonusInstThreshold))
4202 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4207 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
4208 BasicBlock *BB = BI->getParent();
4210 // Conditional branch
4211 if (isValueEqualityComparison(BI)) {
4212 // If we only have one predecessor, and if it is a branch on this value,
4213 // see if that predecessor totally determines the outcome of this
4215 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4216 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
4217 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4219 // This block must be empty, except for the setcond inst, if it exists.
4220 // Ignore dbg intrinsics.
4221 BasicBlock::iterator I = BB->begin();
4222 // Ignore dbg intrinsics.
4223 while (isa<DbgInfoIntrinsic>(I))
4226 if (FoldValueComparisonIntoPredecessors(BI, Builder))
4227 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4228 } else if (&*I == cast<Instruction>(BI->getCondition())){
4230 // Ignore dbg intrinsics.
4231 while (isa<DbgInfoIntrinsic>(I))
4233 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
4234 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4238 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
4239 if (SimplifyBranchOnICmpChain(BI, DL, Builder))
4242 // If this basic block is ONLY a compare and a branch, and if a predecessor
4243 // branches to us and one of our successors, fold the comparison into the
4244 // predecessor and use logical operations to pick the right destination.
4245 if (FoldBranchToCommonDest(BI, DL, BonusInstThreshold))
4246 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4248 // We have a conditional branch to two blocks that are only reachable
4249 // from BI. We know that the condbr dominates the two blocks, so see if
4250 // there is any identical code in the "then" and "else" blocks. If so, we
4251 // can hoist it up to the branching block.
4252 if (BI->getSuccessor(0)->getSinglePredecessor()) {
4253 if (BI->getSuccessor(1)->getSinglePredecessor()) {
4254 if (HoistThenElseCodeToIf(BI, DL))
4255 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4257 // If Successor #1 has multiple preds, we may be able to conditionally
4258 // execute Successor #0 if it branches to Successor #1.
4259 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
4260 if (Succ0TI->getNumSuccessors() == 1 &&
4261 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
4262 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), DL))
4263 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4265 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
4266 // If Successor #0 has multiple preds, we may be able to conditionally
4267 // execute Successor #1 if it branches to Successor #0.
4268 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
4269 if (Succ1TI->getNumSuccessors() == 1 &&
4270 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
4271 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), DL))
4272 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4275 // If this is a branch on a phi node in the current block, thread control
4276 // through this block if any PHI node entries are constants.
4277 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
4278 if (PN->getParent() == BI->getParent())
4279 if (FoldCondBranchOnPHI(BI, DL))
4280 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4282 // Scan predecessor blocks for conditional branches.
4283 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
4284 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
4285 if (PBI != BI && PBI->isConditional())
4286 if (SimplifyCondBranchToCondBranch(PBI, BI))
4287 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4292 /// Check if passing a value to an instruction will cause undefined behavior.
4293 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
4294 Constant *C = dyn_cast<Constant>(V);
4301 if (C->isNullValue()) {
4302 // Only look at the first use, avoid hurting compile time with long uselists
4303 User *Use = *I->user_begin();
4305 // Now make sure that there are no instructions in between that can alter
4306 // control flow (eg. calls)
4307 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
4308 if (i == I->getParent()->end() || i->mayHaveSideEffects())
4311 // Look through GEPs. A load from a GEP derived from NULL is still undefined
4312 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
4313 if (GEP->getPointerOperand() == I)
4314 return passingValueIsAlwaysUndefined(V, GEP);
4316 // Look through bitcasts.
4317 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
4318 return passingValueIsAlwaysUndefined(V, BC);
4320 // Load from null is undefined.
4321 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
4322 if (!LI->isVolatile())
4323 return LI->getPointerAddressSpace() == 0;
4325 // Store to null is undefined.
4326 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
4327 if (!SI->isVolatile())
4328 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
4333 /// If BB has an incoming value that will always trigger undefined behavior
4334 /// (eg. null pointer dereference), remove the branch leading here.
4335 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
4336 for (BasicBlock::iterator i = BB->begin();
4337 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
4338 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
4339 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
4340 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
4341 IRBuilder<> Builder(T);
4342 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
4343 BB->removePredecessor(PHI->getIncomingBlock(i));
4344 // Turn uncoditional branches into unreachables and remove the dead
4345 // destination from conditional branches.
4346 if (BI->isUnconditional())
4347 Builder.CreateUnreachable();
4349 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
4350 BI->getSuccessor(0));
4351 BI->eraseFromParent();
4354 // TODO: SwitchInst.
4360 bool SimplifyCFGOpt::run(BasicBlock *BB) {
4361 bool Changed = false;
4363 assert(BB && BB->getParent() && "Block not embedded in function!");
4364 assert(BB->getTerminator() && "Degenerate basic block encountered!");
4366 // Remove basic blocks that have no predecessors (except the entry block)...
4367 // or that just have themself as a predecessor. These are unreachable.
4368 if ((pred_begin(BB) == pred_end(BB) &&
4369 BB != &BB->getParent()->getEntryBlock()) ||
4370 BB->getSinglePredecessor() == BB) {
4371 DEBUG(dbgs() << "Removing BB: \n" << *BB);
4372 DeleteDeadBlock(BB);
4376 // Check to see if we can constant propagate this terminator instruction
4378 Changed |= ConstantFoldTerminator(BB, true);
4380 // Check for and eliminate duplicate PHI nodes in this block.
4381 Changed |= EliminateDuplicatePHINodes(BB);
4383 // Check for and remove branches that will always cause undefined behavior.
4384 Changed |= removeUndefIntroducingPredecessor(BB);
4386 // Merge basic blocks into their predecessor if there is only one distinct
4387 // pred, and if there is only one distinct successor of the predecessor, and
4388 // if there are no PHI nodes.
4390 if (MergeBlockIntoPredecessor(BB))
4393 IRBuilder<> Builder(BB);
4395 // If there is a trivial two-entry PHI node in this basic block, and we can
4396 // eliminate it, do so now.
4397 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
4398 if (PN->getNumIncomingValues() == 2)
4399 Changed |= FoldTwoEntryPHINode(PN, DL);
4401 Builder.SetInsertPoint(BB->getTerminator());
4402 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
4403 if (BI->isUnconditional()) {
4404 if (SimplifyUncondBranch(BI, Builder)) return true;
4406 if (SimplifyCondBranch(BI, Builder)) return true;
4408 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
4409 if (SimplifyReturn(RI, Builder)) return true;
4410 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
4411 if (SimplifyResume(RI, Builder)) return true;
4412 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
4413 if (SimplifySwitch(SI, Builder)) return true;
4414 } else if (UnreachableInst *UI =
4415 dyn_cast<UnreachableInst>(BB->getTerminator())) {
4416 if (SimplifyUnreachable(UI)) return true;
4417 } else if (IndirectBrInst *IBI =
4418 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
4419 if (SimplifyIndirectBr(IBI)) return true;
4425 /// SimplifyCFG - This function is used to do simplification of a CFG. For
4426 /// example, it adjusts branches to branches to eliminate the extra hop, it
4427 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4428 /// of the CFG. It returns true if a modification was made.
4430 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
4431 unsigned BonusInstThreshold,
4432 const DataLayout *DL, AssumptionTracker *AT) {
4433 return SimplifyCFGOpt(TTI, BonusInstThreshold, DL, AT).run(BB);