1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Peephole optimize the CFG.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Transforms/Utils/Local.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/ADT/SetVector.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/ConstantFolding.h"
22 #include "llvm/Analysis/InstructionSimplify.h"
23 #include "llvm/Analysis/TargetTransformInfo.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/IR/CFG.h"
26 #include "llvm/IR/ConstantRange.h"
27 #include "llvm/IR/Constants.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/DerivedTypes.h"
30 #include "llvm/IR/GlobalVariable.h"
31 #include "llvm/IR/IRBuilder.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/LLVMContext.h"
35 #include "llvm/IR/MDBuilder.h"
36 #include "llvm/IR/Metadata.h"
37 #include "llvm/IR/Module.h"
38 #include "llvm/IR/NoFolder.h"
39 #include "llvm/IR/Operator.h"
40 #include "llvm/IR/PatternMatch.h"
41 #include "llvm/IR/Type.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/Debug.h"
44 #include "llvm/Support/raw_ostream.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Transforms/Utils/ValueMapper.h"
52 using namespace PatternMatch;
54 #define DEBUG_TYPE "simplifycfg"
56 static cl::opt<unsigned>
57 PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1),
58 cl::desc("Control the amount of phi node folding to perform (default = 1)"));
61 DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false),
62 cl::desc("Duplicate return instructions into unconditional branches"));
65 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
66 cl::desc("Sink common instructions down to the end block"));
68 static cl::opt<bool> HoistCondStores(
69 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
70 cl::desc("Hoist conditional stores if an unconditional store precedes"));
72 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
73 STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables");
74 STATISTIC(NumLookupTablesHoles, "Number of switch instructions turned into lookup tables (holes checked)");
75 STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block");
76 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
79 // The first field contains the value that the switch produces when a certain
80 // case group is selected, and the second field is a vector containing the cases
81 // composing the case group.
82 typedef SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>
83 SwitchCaseResultVectorTy;
84 // The first field contains the phi node that generates a result of the switch
85 // and the second field contains the value generated for a certain case in the switch
87 typedef SmallVector<std::pair<PHINode *, Constant *>, 4> SwitchCaseResultsTy;
89 /// ValueEqualityComparisonCase - Represents a case of a switch.
90 struct ValueEqualityComparisonCase {
94 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
95 : Value(Value), Dest(Dest) {}
97 bool operator<(ValueEqualityComparisonCase RHS) const {
98 // Comparing pointers is ok as we only rely on the order for uniquing.
99 return Value < RHS.Value;
102 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
105 class SimplifyCFGOpt {
106 const TargetTransformInfo &TTI;
107 unsigned BonusInstThreshold;
108 const DataLayout *const DL;
109 AssumptionTracker *AT;
110 Value *isValueEqualityComparison(TerminatorInst *TI);
111 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
112 std::vector<ValueEqualityComparisonCase> &Cases);
113 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
115 IRBuilder<> &Builder);
116 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
117 IRBuilder<> &Builder);
119 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
120 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
121 bool SimplifyUnreachable(UnreachableInst *UI);
122 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
123 bool SimplifyIndirectBr(IndirectBrInst *IBI);
124 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder);
125 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder);
128 SimplifyCFGOpt(const TargetTransformInfo &TTI, unsigned BonusInstThreshold,
129 const DataLayout *DL, AssumptionTracker *AT)
130 : TTI(TTI), BonusInstThreshold(BonusInstThreshold), DL(DL), AT(AT) {}
131 bool run(BasicBlock *BB);
135 /// SafeToMergeTerminators - Return true if it is safe to merge these two
136 /// terminator instructions together.
138 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
139 if (SI1 == SI2) return false; // Can't merge with self!
141 // It is not safe to merge these two switch instructions if they have a common
142 // successor, and if that successor has a PHI node, and if *that* PHI node has
143 // conflicting incoming values from the two switch blocks.
144 BasicBlock *SI1BB = SI1->getParent();
145 BasicBlock *SI2BB = SI2->getParent();
146 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
148 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
149 if (SI1Succs.count(*I))
150 for (BasicBlock::iterator BBI = (*I)->begin();
151 isa<PHINode>(BBI); ++BBI) {
152 PHINode *PN = cast<PHINode>(BBI);
153 if (PN->getIncomingValueForBlock(SI1BB) !=
154 PN->getIncomingValueForBlock(SI2BB))
161 /// isProfitableToFoldUnconditional - Return true if it is safe and profitable
162 /// to merge these two terminator instructions together, where SI1 is an
163 /// unconditional branch. PhiNodes will store all PHI nodes in common
166 static bool isProfitableToFoldUnconditional(BranchInst *SI1,
169 SmallVectorImpl<PHINode*> &PhiNodes) {
170 if (SI1 == SI2) return false; // Can't merge with self!
171 assert(SI1->isUnconditional() && SI2->isConditional());
173 // We fold the unconditional branch if we can easily update all PHI nodes in
174 // common successors:
175 // 1> We have a constant incoming value for the conditional branch;
176 // 2> We have "Cond" as the incoming value for the unconditional branch;
177 // 3> SI2->getCondition() and Cond have same operands.
178 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
179 if (!Ci2) return false;
180 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
181 Cond->getOperand(1) == Ci2->getOperand(1)) &&
182 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
183 Cond->getOperand(1) == Ci2->getOperand(0)))
186 BasicBlock *SI1BB = SI1->getParent();
187 BasicBlock *SI2BB = SI2->getParent();
188 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
189 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
190 if (SI1Succs.count(*I))
191 for (BasicBlock::iterator BBI = (*I)->begin();
192 isa<PHINode>(BBI); ++BBI) {
193 PHINode *PN = cast<PHINode>(BBI);
194 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
195 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
197 PhiNodes.push_back(PN);
202 /// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will
203 /// now be entries in it from the 'NewPred' block. The values that will be
204 /// flowing into the PHI nodes will be the same as those coming in from
205 /// ExistPred, an existing predecessor of Succ.
206 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
207 BasicBlock *ExistPred) {
208 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
211 for (BasicBlock::iterator I = Succ->begin();
212 (PN = dyn_cast<PHINode>(I)); ++I)
213 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
216 /// ComputeSpeculationCost - Compute an abstract "cost" of speculating the
217 /// given instruction, which is assumed to be safe to speculate. 1 means
218 /// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive.
219 static unsigned ComputeSpeculationCost(const User *I, const DataLayout *DL) {
220 assert(isSafeToSpeculativelyExecute(I, DL) &&
221 "Instruction is not safe to speculatively execute!");
222 switch (Operator::getOpcode(I)) {
224 // In doubt, be conservative.
226 case Instruction::GetElementPtr:
227 // GEPs are cheap if all indices are constant.
228 if (!cast<GEPOperator>(I)->hasAllConstantIndices())
231 case Instruction::ExtractValue:
232 case Instruction::Load:
233 case Instruction::Add:
234 case Instruction::Sub:
235 case Instruction::And:
236 case Instruction::Or:
237 case Instruction::Xor:
238 case Instruction::Shl:
239 case Instruction::LShr:
240 case Instruction::AShr:
241 case Instruction::ICmp:
242 case Instruction::Trunc:
243 case Instruction::ZExt:
244 case Instruction::SExt:
245 case Instruction::BitCast:
246 case Instruction::ExtractElement:
247 case Instruction::InsertElement:
248 return 1; // These are all cheap.
250 case Instruction::Call:
251 case Instruction::Select:
256 /// DominatesMergePoint - If we have a merge point of an "if condition" as
257 /// accepted above, return true if the specified value dominates the block. We
258 /// don't handle the true generality of domination here, just a special case
259 /// which works well enough for us.
261 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
262 /// see if V (which must be an instruction) and its recursive operands
263 /// that do not dominate BB have a combined cost lower than CostRemaining and
264 /// are non-trapping. If both are true, the instruction is inserted into the
265 /// set and true is returned.
267 /// The cost for most non-trapping instructions is defined as 1 except for
268 /// Select whose cost is 2.
270 /// After this function returns, CostRemaining is decreased by the cost of
271 /// V plus its non-dominating operands. If that cost is greater than
272 /// CostRemaining, false is returned and CostRemaining is undefined.
273 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
274 SmallPtrSetImpl<Instruction*> *AggressiveInsts,
275 unsigned &CostRemaining,
276 const DataLayout *DL) {
277 Instruction *I = dyn_cast<Instruction>(V);
279 // Non-instructions all dominate instructions, but not all constantexprs
280 // can be executed unconditionally.
281 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
286 BasicBlock *PBB = I->getParent();
288 // We don't want to allow weird loops that might have the "if condition" in
289 // the bottom of this block.
290 if (PBB == BB) return false;
292 // If this instruction is defined in a block that contains an unconditional
293 // branch to BB, then it must be in the 'conditional' part of the "if
294 // statement". If not, it definitely dominates the region.
295 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
296 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
299 // If we aren't allowing aggressive promotion anymore, then don't consider
300 // instructions in the 'if region'.
301 if (!AggressiveInsts) return false;
303 // If we have seen this instruction before, don't count it again.
304 if (AggressiveInsts->count(I)) return true;
306 // Okay, it looks like the instruction IS in the "condition". Check to
307 // see if it's a cheap instruction to unconditionally compute, and if it
308 // only uses stuff defined outside of the condition. If so, hoist it out.
309 if (!isSafeToSpeculativelyExecute(I, DL))
312 unsigned Cost = ComputeSpeculationCost(I, DL);
314 if (Cost > CostRemaining)
317 CostRemaining -= Cost;
319 // Okay, we can only really hoist these out if their operands do
320 // not take us over the cost threshold.
321 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
322 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, DL))
324 // Okay, it's safe to do this! Remember this instruction.
325 AggressiveInsts->insert(I);
329 /// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr
330 /// and PointerNullValue. Return NULL if value is not a constant int.
331 static ConstantInt *GetConstantInt(Value *V, const DataLayout *DL) {
332 // Normal constant int.
333 ConstantInt *CI = dyn_cast<ConstantInt>(V);
334 if (CI || !DL || !isa<Constant>(V) || !V->getType()->isPointerTy())
337 // This is some kind of pointer constant. Turn it into a pointer-sized
338 // ConstantInt if possible.
339 IntegerType *PtrTy = cast<IntegerType>(DL->getIntPtrType(V->getType()));
341 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
342 if (isa<ConstantPointerNull>(V))
343 return ConstantInt::get(PtrTy, 0);
345 // IntToPtr const int.
346 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
347 if (CE->getOpcode() == Instruction::IntToPtr)
348 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
349 // The constant is very likely to have the right type already.
350 if (CI->getType() == PtrTy)
353 return cast<ConstantInt>
354 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
359 /// GatherConstantCompares - Given a potentially 'or'd or 'and'd together
360 /// collection of icmp eq/ne instructions that compare a value against a
361 /// constant, return the value being compared, and stick the constant into the
364 GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra,
365 const DataLayout *DL, bool isEQ, unsigned &UsedICmps) {
366 Instruction *I = dyn_cast<Instruction>(V);
367 if (!I) return nullptr;
369 // If this is an icmp against a constant, handle this as one of the cases.
370 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) {
371 if (ConstantInt *C = GetConstantInt(I->getOperand(1), DL)) {
375 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
376 // (x & ~2^x) == y --> x == y || x == y|2^x
377 // This undoes a transformation done by instcombine to fuse 2 compares.
378 if (match(ICI->getOperand(0),
379 m_And(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
380 APInt Not = ~RHSC->getValue();
381 if (Not.isPowerOf2()) {
384 ConstantInt::get(C->getContext(), C->getValue() | Not));
392 return I->getOperand(0);
395 // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to
398 ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue());
400 // Shift the range if the compare is fed by an add. This is the range
401 // compare idiom as emitted by instcombine.
403 match(I->getOperand(0), m_Add(m_Value(RHSVal), m_ConstantInt(RHSC)));
405 Span = Span.subtract(RHSC->getValue());
407 // If this is an and/!= check then we want to optimize "x ugt 2" into
410 Span = Span.inverse();
412 // If there are a ton of values, we don't want to make a ginormous switch.
413 if (Span.getSetSize().ugt(8) || Span.isEmptySet())
416 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
417 Vals.push_back(ConstantInt::get(V->getContext(), Tmp));
419 return hasAdd ? RHSVal : I->getOperand(0);
424 // Otherwise, we can only handle an | or &, depending on isEQ.
425 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And))
428 unsigned NumValsBeforeLHS = Vals.size();
429 unsigned UsedICmpsBeforeLHS = UsedICmps;
430 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, DL,
432 unsigned NumVals = Vals.size();
433 unsigned UsedICmpsBeforeRHS = UsedICmps;
434 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, DL,
438 Vals.resize(NumVals);
439 UsedICmps = UsedICmpsBeforeRHS;
442 // The RHS of the or/and can't be folded in and we haven't used "Extra" yet,
443 // set it and return success.
444 if (Extra == nullptr || Extra == I->getOperand(1)) {
445 Extra = I->getOperand(1);
449 Vals.resize(NumValsBeforeLHS);
450 UsedICmps = UsedICmpsBeforeLHS;
454 // If the LHS can't be folded in, but Extra is available and RHS can, try to
456 if (Extra == nullptr || Extra == I->getOperand(0)) {
457 Value *OldExtra = Extra;
458 Extra = I->getOperand(0);
459 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, DL,
462 assert(Vals.size() == NumValsBeforeLHS);
469 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
470 Instruction *Cond = nullptr;
471 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
472 Cond = dyn_cast<Instruction>(SI->getCondition());
473 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
474 if (BI->isConditional())
475 Cond = dyn_cast<Instruction>(BI->getCondition());
476 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
477 Cond = dyn_cast<Instruction>(IBI->getAddress());
480 TI->eraseFromParent();
481 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
484 /// isValueEqualityComparison - Return true if the specified terminator checks
485 /// to see if a value is equal to constant integer value.
486 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
488 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
489 // Do not permit merging of large switch instructions into their
490 // predecessors unless there is only one predecessor.
491 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()),
492 pred_end(SI->getParent())) <= 128)
493 CV = SI->getCondition();
494 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
495 if (BI->isConditional() && BI->getCondition()->hasOneUse())
496 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition()))
497 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
498 CV = ICI->getOperand(0);
500 // Unwrap any lossless ptrtoint cast.
502 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
503 Value *Ptr = PTII->getPointerOperand();
504 if (PTII->getType() == DL->getIntPtrType(Ptr->getType()))
511 /// GetValueEqualityComparisonCases - Given a value comparison instruction,
512 /// decode all of the 'cases' that it represents and return the 'default' block.
513 BasicBlock *SimplifyCFGOpt::
514 GetValueEqualityComparisonCases(TerminatorInst *TI,
515 std::vector<ValueEqualityComparisonCase>
517 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
518 Cases.reserve(SI->getNumCases());
519 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
520 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
521 i.getCaseSuccessor()));
522 return SI->getDefaultDest();
525 BranchInst *BI = cast<BranchInst>(TI);
526 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
527 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
528 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
531 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
535 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries
536 /// in the list that match the specified block.
537 static void EliminateBlockCases(BasicBlock *BB,
538 std::vector<ValueEqualityComparisonCase> &Cases) {
539 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
542 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as
545 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
546 std::vector<ValueEqualityComparisonCase > &C2) {
547 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
549 // Make V1 be smaller than V2.
550 if (V1->size() > V2->size())
553 if (V1->size() == 0) return false;
554 if (V1->size() == 1) {
556 ConstantInt *TheVal = (*V1)[0].Value;
557 for (unsigned i = 0, e = V2->size(); i != e; ++i)
558 if (TheVal == (*V2)[i].Value)
562 // Otherwise, just sort both lists and compare element by element.
563 array_pod_sort(V1->begin(), V1->end());
564 array_pod_sort(V2->begin(), V2->end());
565 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
566 while (i1 != e1 && i2 != e2) {
567 if ((*V1)[i1].Value == (*V2)[i2].Value)
569 if ((*V1)[i1].Value < (*V2)[i2].Value)
577 /// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a
578 /// terminator instruction and its block is known to only have a single
579 /// predecessor block, check to see if that predecessor is also a value
580 /// comparison with the same value, and if that comparison determines the
581 /// outcome of this comparison. If so, simplify TI. This does a very limited
582 /// form of jump threading.
583 bool SimplifyCFGOpt::
584 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
586 IRBuilder<> &Builder) {
587 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
588 if (!PredVal) return false; // Not a value comparison in predecessor.
590 Value *ThisVal = isValueEqualityComparison(TI);
591 assert(ThisVal && "This isn't a value comparison!!");
592 if (ThisVal != PredVal) return false; // Different predicates.
594 // TODO: Preserve branch weight metadata, similarly to how
595 // FoldValueComparisonIntoPredecessors preserves it.
597 // Find out information about when control will move from Pred to TI's block.
598 std::vector<ValueEqualityComparisonCase> PredCases;
599 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
601 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
603 // Find information about how control leaves this block.
604 std::vector<ValueEqualityComparisonCase> ThisCases;
605 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
606 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
608 // If TI's block is the default block from Pred's comparison, potentially
609 // simplify TI based on this knowledge.
610 if (PredDef == TI->getParent()) {
611 // If we are here, we know that the value is none of those cases listed in
612 // PredCases. If there are any cases in ThisCases that are in PredCases, we
614 if (!ValuesOverlap(PredCases, ThisCases))
617 if (isa<BranchInst>(TI)) {
618 // Okay, one of the successors of this condbr is dead. Convert it to a
620 assert(ThisCases.size() == 1 && "Branch can only have one case!");
621 // Insert the new branch.
622 Instruction *NI = Builder.CreateBr(ThisDef);
625 // Remove PHI node entries for the dead edge.
626 ThisCases[0].Dest->removePredecessor(TI->getParent());
628 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
629 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
631 EraseTerminatorInstAndDCECond(TI);
635 SwitchInst *SI = cast<SwitchInst>(TI);
636 // Okay, TI has cases that are statically dead, prune them away.
637 SmallPtrSet<Constant*, 16> DeadCases;
638 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
639 DeadCases.insert(PredCases[i].Value);
641 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
642 << "Through successor TI: " << *TI);
644 // Collect branch weights into a vector.
645 SmallVector<uint32_t, 8> Weights;
646 MDNode* MD = SI->getMetadata(LLVMContext::MD_prof);
647 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
649 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
651 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i));
653 Weights.push_back(CI->getValue().getZExtValue());
655 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
657 if (DeadCases.count(i.getCaseValue())) {
659 std::swap(Weights[i.getCaseIndex()+1], Weights.back());
662 i.getCaseSuccessor()->removePredecessor(TI->getParent());
666 if (HasWeight && Weights.size() >= 2)
667 SI->setMetadata(LLVMContext::MD_prof,
668 MDBuilder(SI->getParent()->getContext()).
669 createBranchWeights(Weights));
671 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
675 // Otherwise, TI's block must correspond to some matched value. Find out
676 // which value (or set of values) this is.
677 ConstantInt *TIV = nullptr;
678 BasicBlock *TIBB = TI->getParent();
679 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
680 if (PredCases[i].Dest == TIBB) {
682 return false; // Cannot handle multiple values coming to this block.
683 TIV = PredCases[i].Value;
685 assert(TIV && "No edge from pred to succ?");
687 // Okay, we found the one constant that our value can be if we get into TI's
688 // BB. Find out which successor will unconditionally be branched to.
689 BasicBlock *TheRealDest = nullptr;
690 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
691 if (ThisCases[i].Value == TIV) {
692 TheRealDest = ThisCases[i].Dest;
696 // If not handled by any explicit cases, it is handled by the default case.
697 if (!TheRealDest) TheRealDest = ThisDef;
699 // Remove PHI node entries for dead edges.
700 BasicBlock *CheckEdge = TheRealDest;
701 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
702 if (*SI != CheckEdge)
703 (*SI)->removePredecessor(TIBB);
707 // Insert the new branch.
708 Instruction *NI = Builder.CreateBr(TheRealDest);
711 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
712 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
714 EraseTerminatorInstAndDCECond(TI);
719 /// ConstantIntOrdering - This class implements a stable ordering of constant
720 /// integers that does not depend on their address. This is important for
721 /// applications that sort ConstantInt's to ensure uniqueness.
722 struct ConstantIntOrdering {
723 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
724 return LHS->getValue().ult(RHS->getValue());
729 static int ConstantIntSortPredicate(ConstantInt *const *P1,
730 ConstantInt *const *P2) {
731 const ConstantInt *LHS = *P1;
732 const ConstantInt *RHS = *P2;
733 if (LHS->getValue().ult(RHS->getValue()))
735 if (LHS->getValue() == RHS->getValue())
740 static inline bool HasBranchWeights(const Instruction* I) {
741 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof);
742 if (ProfMD && ProfMD->getOperand(0))
743 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
744 return MDS->getString().equals("branch_weights");
749 /// Get Weights of a given TerminatorInst, the default weight is at the front
750 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
752 static void GetBranchWeights(TerminatorInst *TI,
753 SmallVectorImpl<uint64_t> &Weights) {
754 MDNode* MD = TI->getMetadata(LLVMContext::MD_prof);
756 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
757 ConstantInt *CI = cast<ConstantInt>(MD->getOperand(i));
758 Weights.push_back(CI->getValue().getZExtValue());
761 // If TI is a conditional eq, the default case is the false case,
762 // and the corresponding branch-weight data is at index 2. We swap the
763 // default weight to be the first entry.
764 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
765 assert(Weights.size() == 2);
766 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
767 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
768 std::swap(Weights.front(), Weights.back());
772 /// Keep halving the weights until all can fit in uint32_t.
773 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
774 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
775 if (Max > UINT_MAX) {
776 unsigned Offset = 32 - countLeadingZeros(Max);
777 for (uint64_t &I : Weights)
782 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value
783 /// equality comparison instruction (either a switch or a branch on "X == c").
784 /// See if any of the predecessors of the terminator block are value comparisons
785 /// on the same value. If so, and if safe to do so, fold them together.
786 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
787 IRBuilder<> &Builder) {
788 BasicBlock *BB = TI->getParent();
789 Value *CV = isValueEqualityComparison(TI); // CondVal
790 assert(CV && "Not a comparison?");
791 bool Changed = false;
793 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
794 while (!Preds.empty()) {
795 BasicBlock *Pred = Preds.pop_back_val();
797 // See if the predecessor is a comparison with the same value.
798 TerminatorInst *PTI = Pred->getTerminator();
799 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
801 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
802 // Figure out which 'cases' to copy from SI to PSI.
803 std::vector<ValueEqualityComparisonCase> BBCases;
804 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
806 std::vector<ValueEqualityComparisonCase> PredCases;
807 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
809 // Based on whether the default edge from PTI goes to BB or not, fill in
810 // PredCases and PredDefault with the new switch cases we would like to
812 SmallVector<BasicBlock*, 8> NewSuccessors;
814 // Update the branch weight metadata along the way
815 SmallVector<uint64_t, 8> Weights;
816 bool PredHasWeights = HasBranchWeights(PTI);
817 bool SuccHasWeights = HasBranchWeights(TI);
819 if (PredHasWeights) {
820 GetBranchWeights(PTI, Weights);
821 // branch-weight metadata is inconsistent here.
822 if (Weights.size() != 1 + PredCases.size())
823 PredHasWeights = SuccHasWeights = false;
824 } else if (SuccHasWeights)
825 // If there are no predecessor weights but there are successor weights,
826 // populate Weights with 1, which will later be scaled to the sum of
827 // successor's weights
828 Weights.assign(1 + PredCases.size(), 1);
830 SmallVector<uint64_t, 8> SuccWeights;
831 if (SuccHasWeights) {
832 GetBranchWeights(TI, SuccWeights);
833 // branch-weight metadata is inconsistent here.
834 if (SuccWeights.size() != 1 + BBCases.size())
835 PredHasWeights = SuccHasWeights = false;
836 } else if (PredHasWeights)
837 SuccWeights.assign(1 + BBCases.size(), 1);
839 if (PredDefault == BB) {
840 // If this is the default destination from PTI, only the edges in TI
841 // that don't occur in PTI, or that branch to BB will be activated.
842 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
843 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
844 if (PredCases[i].Dest != BB)
845 PTIHandled.insert(PredCases[i].Value);
847 // The default destination is BB, we don't need explicit targets.
848 std::swap(PredCases[i], PredCases.back());
850 if (PredHasWeights || SuccHasWeights) {
851 // Increase weight for the default case.
852 Weights[0] += Weights[i+1];
853 std::swap(Weights[i+1], Weights.back());
857 PredCases.pop_back();
861 // Reconstruct the new switch statement we will be building.
862 if (PredDefault != BBDefault) {
863 PredDefault->removePredecessor(Pred);
864 PredDefault = BBDefault;
865 NewSuccessors.push_back(BBDefault);
868 unsigned CasesFromPred = Weights.size();
869 uint64_t ValidTotalSuccWeight = 0;
870 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
871 if (!PTIHandled.count(BBCases[i].Value) &&
872 BBCases[i].Dest != BBDefault) {
873 PredCases.push_back(BBCases[i]);
874 NewSuccessors.push_back(BBCases[i].Dest);
875 if (SuccHasWeights || PredHasWeights) {
876 // The default weight is at index 0, so weight for the ith case
877 // should be at index i+1. Scale the cases from successor by
878 // PredDefaultWeight (Weights[0]).
879 Weights.push_back(Weights[0] * SuccWeights[i+1]);
880 ValidTotalSuccWeight += SuccWeights[i+1];
884 if (SuccHasWeights || PredHasWeights) {
885 ValidTotalSuccWeight += SuccWeights[0];
886 // Scale the cases from predecessor by ValidTotalSuccWeight.
887 for (unsigned i = 1; i < CasesFromPred; ++i)
888 Weights[i] *= ValidTotalSuccWeight;
889 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
890 Weights[0] *= SuccWeights[0];
893 // If this is not the default destination from PSI, only the edges
894 // in SI that occur in PSI with a destination of BB will be
896 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
897 std::map<ConstantInt*, uint64_t> WeightsForHandled;
898 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
899 if (PredCases[i].Dest == BB) {
900 PTIHandled.insert(PredCases[i].Value);
902 if (PredHasWeights || SuccHasWeights) {
903 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
904 std::swap(Weights[i+1], Weights.back());
908 std::swap(PredCases[i], PredCases.back());
909 PredCases.pop_back();
913 // Okay, now we know which constants were sent to BB from the
914 // predecessor. Figure out where they will all go now.
915 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
916 if (PTIHandled.count(BBCases[i].Value)) {
917 // If this is one we are capable of getting...
918 if (PredHasWeights || SuccHasWeights)
919 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
920 PredCases.push_back(BBCases[i]);
921 NewSuccessors.push_back(BBCases[i].Dest);
922 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
925 // If there are any constants vectored to BB that TI doesn't handle,
926 // they must go to the default destination of TI.
927 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
929 E = PTIHandled.end(); I != E; ++I) {
930 if (PredHasWeights || SuccHasWeights)
931 Weights.push_back(WeightsForHandled[*I]);
932 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
933 NewSuccessors.push_back(BBDefault);
937 // Okay, at this point, we know which new successor Pred will get. Make
938 // sure we update the number of entries in the PHI nodes for these
940 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i)
941 AddPredecessorToBlock(NewSuccessors[i], Pred, BB);
943 Builder.SetInsertPoint(PTI);
944 // Convert pointer to int before we switch.
945 if (CV->getType()->isPointerTy()) {
946 assert(DL && "Cannot switch on pointer without DataLayout");
947 CV = Builder.CreatePtrToInt(CV, DL->getIntPtrType(CV->getType()),
951 // Now that the successors are updated, create the new Switch instruction.
952 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
954 NewSI->setDebugLoc(PTI->getDebugLoc());
955 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
956 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest);
958 if (PredHasWeights || SuccHasWeights) {
959 // Halve the weights if any of them cannot fit in an uint32_t
962 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
964 NewSI->setMetadata(LLVMContext::MD_prof,
965 MDBuilder(BB->getContext()).
966 createBranchWeights(MDWeights));
969 EraseTerminatorInstAndDCECond(PTI);
971 // Okay, last check. If BB is still a successor of PSI, then we must
972 // have an infinite loop case. If so, add an infinitely looping block
973 // to handle the case to preserve the behavior of the code.
974 BasicBlock *InfLoopBlock = nullptr;
975 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
976 if (NewSI->getSuccessor(i) == BB) {
978 // Insert it at the end of the function, because it's either code,
979 // or it won't matter if it's hot. :)
980 InfLoopBlock = BasicBlock::Create(BB->getContext(),
981 "infloop", BB->getParent());
982 BranchInst::Create(InfLoopBlock, InfLoopBlock);
984 NewSI->setSuccessor(i, InfLoopBlock);
993 // isSafeToHoistInvoke - If we would need to insert a select that uses the
994 // value of this invoke (comments in HoistThenElseCodeToIf explain why we
995 // would need to do this), we can't hoist the invoke, as there is nowhere
996 // to put the select in this case.
997 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
998 Instruction *I1, Instruction *I2) {
999 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1001 for (BasicBlock::iterator BBI = SI->begin();
1002 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1003 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1004 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1005 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
1013 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
1014 /// BB2, hoist any common code in the two blocks up into the branch block. The
1015 /// caller of this function guarantees that BI's block dominates BB1 and BB2.
1016 static bool HoistThenElseCodeToIf(BranchInst *BI, const DataLayout *DL) {
1017 // This does very trivial matching, with limited scanning, to find identical
1018 // instructions in the two blocks. In particular, we don't want to get into
1019 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1020 // such, we currently just scan for obviously identical instructions in an
1022 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1023 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1025 BasicBlock::iterator BB1_Itr = BB1->begin();
1026 BasicBlock::iterator BB2_Itr = BB2->begin();
1028 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++;
1029 // Skip debug info if it is not identical.
1030 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1031 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1032 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1033 while (isa<DbgInfoIntrinsic>(I1))
1035 while (isa<DbgInfoIntrinsic>(I2))
1038 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1039 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1042 BasicBlock *BIParent = BI->getParent();
1044 bool Changed = false;
1046 // If we are hoisting the terminator instruction, don't move one (making a
1047 // broken BB), instead clone it, and remove BI.
1048 if (isa<TerminatorInst>(I1))
1049 goto HoistTerminator;
1051 // For a normal instruction, we just move one to right before the branch,
1052 // then replace all uses of the other with the first. Finally, we remove
1053 // the now redundant second instruction.
1054 BIParent->getInstList().splice(BI, BB1->getInstList(), I1);
1055 if (!I2->use_empty())
1056 I2->replaceAllUsesWith(I1);
1057 I1->intersectOptionalDataWith(I2);
1058 unsigned KnownIDs[] = {
1059 LLVMContext::MD_tbaa,
1060 LLVMContext::MD_range,
1061 LLVMContext::MD_fpmath,
1062 LLVMContext::MD_invariant_load
1064 combineMetadata(I1, I2, KnownIDs);
1065 I2->eraseFromParent();
1070 // Skip debug info if it is not identical.
1071 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1072 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1073 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1074 while (isa<DbgInfoIntrinsic>(I1))
1076 while (isa<DbgInfoIntrinsic>(I2))
1079 } while (I1->isIdenticalToWhenDefined(I2));
1084 // It may not be possible to hoist an invoke.
1085 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1088 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1090 for (BasicBlock::iterator BBI = SI->begin();
1091 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1092 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1093 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1097 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V, DL))
1099 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V, DL))
1104 // Okay, it is safe to hoist the terminator.
1105 Instruction *NT = I1->clone();
1106 BIParent->getInstList().insert(BI, NT);
1107 if (!NT->getType()->isVoidTy()) {
1108 I1->replaceAllUsesWith(NT);
1109 I2->replaceAllUsesWith(NT);
1113 IRBuilder<true, NoFolder> Builder(NT);
1114 // Hoisting one of the terminators from our successor is a great thing.
1115 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1116 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1117 // nodes, so we insert select instruction to compute the final result.
1118 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1119 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1121 for (BasicBlock::iterator BBI = SI->begin();
1122 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1123 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1124 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1125 if (BB1V == BB2V) continue;
1127 // These values do not agree. Insert a select instruction before NT
1128 // that determines the right value.
1129 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1131 SI = cast<SelectInst>
1132 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1133 BB1V->getName()+"."+BB2V->getName()));
1135 // Make the PHI node use the select for all incoming values for BB1/BB2
1136 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1137 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1138 PN->setIncomingValue(i, SI);
1142 // Update any PHI nodes in our new successors.
1143 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1144 AddPredecessorToBlock(*SI, BIParent, BB1);
1146 EraseTerminatorInstAndDCECond(BI);
1150 /// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd,
1151 /// check whether BBEnd has only two predecessors and the other predecessor
1152 /// ends with an unconditional branch. If it is true, sink any common code
1153 /// in the two predecessors to BBEnd.
1154 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1155 assert(BI1->isUnconditional());
1156 BasicBlock *BB1 = BI1->getParent();
1157 BasicBlock *BBEnd = BI1->getSuccessor(0);
1159 // Check that BBEnd has two predecessors and the other predecessor ends with
1160 // an unconditional branch.
1161 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1162 BasicBlock *Pred0 = *PI++;
1163 if (PI == PE) // Only one predecessor.
1165 BasicBlock *Pred1 = *PI++;
1166 if (PI != PE) // More than two predecessors.
1168 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1169 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1170 if (!BI2 || !BI2->isUnconditional())
1173 // Gather the PHI nodes in BBEnd.
1174 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2;
1175 Instruction *FirstNonPhiInBBEnd = nullptr;
1176 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end();
1178 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1179 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1180 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1181 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN);
1183 FirstNonPhiInBBEnd = &*I;
1187 if (!FirstNonPhiInBBEnd)
1191 // This does very trivial matching, with limited scanning, to find identical
1192 // instructions in the two blocks. We scan backward for obviously identical
1193 // instructions in an identical order.
1194 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1195 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(),
1196 RE2 = BB2->getInstList().rend();
1198 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1201 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1204 // Skip the unconditional branches.
1208 bool Changed = false;
1209 while (RI1 != RE1 && RI2 != RE2) {
1211 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1214 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1218 Instruction *I1 = &*RI1, *I2 = &*RI2;
1219 // I1 and I2 should have a single use in the same PHI node, and they
1220 // perform the same operation.
1221 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1222 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1223 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1224 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) ||
1225 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1226 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1227 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1228 !I1->hasOneUse() || !I2->hasOneUse() ||
1229 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() ||
1230 MapValueFromBB1ToBB2[I1].first != I2)
1233 // Check whether we should swap the operands of ICmpInst.
1234 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1235 bool SwapOpnds = false;
1236 if (ICmp1 && ICmp2 &&
1237 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1238 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1239 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1240 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1241 ICmp2->swapOperands();
1244 if (!I1->isSameOperationAs(I2)) {
1246 ICmp2->swapOperands();
1250 // The operands should be either the same or they need to be generated
1251 // with a PHI node after sinking. We only handle the case where there is
1252 // a single pair of different operands.
1253 Value *DifferentOp1 = nullptr, *DifferentOp2 = nullptr;
1254 unsigned Op1Idx = 0;
1255 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1256 if (I1->getOperand(I) == I2->getOperand(I))
1258 // Early exit if we have more-than one pair of different operands or
1259 // the different operand is already in MapValueFromBB1ToBB2.
1260 // Early exit if we need a PHI node to replace a constant.
1262 MapValueFromBB1ToBB2.find(I1->getOperand(I)) !=
1263 MapValueFromBB1ToBB2.end() ||
1264 isa<Constant>(I1->getOperand(I)) ||
1265 isa<Constant>(I2->getOperand(I))) {
1266 // If we can't sink the instructions, undo the swapping.
1268 ICmp2->swapOperands();
1271 DifferentOp1 = I1->getOperand(I);
1273 DifferentOp2 = I2->getOperand(I);
1276 // We insert the pair of different operands to MapValueFromBB1ToBB2 and
1277 // remove (I1, I2) from MapValueFromBB1ToBB2.
1279 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2,
1280 DifferentOp1->getName() + ".sink",
1282 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN);
1283 // I1 should use NewPN instead of DifferentOp1.
1284 I1->setOperand(Op1Idx, NewPN);
1285 NewPN->addIncoming(DifferentOp1, BB1);
1286 NewPN->addIncoming(DifferentOp2, BB2);
1287 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1289 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second;
1290 MapValueFromBB1ToBB2.erase(I1);
1292 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";);
1293 DEBUG(dbgs() << " " << *I2 << "\n";);
1294 // We need to update RE1 and RE2 if we are going to sink the first
1295 // instruction in the basic block down.
1296 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1297 // Sink the instruction.
1298 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1);
1299 if (!OldPN->use_empty())
1300 OldPN->replaceAllUsesWith(I1);
1301 OldPN->eraseFromParent();
1303 if (!I2->use_empty())
1304 I2->replaceAllUsesWith(I1);
1305 I1->intersectOptionalDataWith(I2);
1306 I2->eraseFromParent();
1309 RE1 = BB1->getInstList().rend();
1311 RE2 = BB2->getInstList().rend();
1312 FirstNonPhiInBBEnd = I1;
1319 /// \brief Determine if we can hoist sink a sole store instruction out of a
1320 /// conditional block.
1322 /// We are looking for code like the following:
1324 /// store i32 %add, i32* %arrayidx2
1325 /// ... // No other stores or function calls (we could be calling a memory
1326 /// ... // function).
1327 /// %cmp = icmp ult %x, %y
1328 /// br i1 %cmp, label %EndBB, label %ThenBB
1330 /// store i32 %add5, i32* %arrayidx2
1334 /// We are going to transform this into:
1336 /// store i32 %add, i32* %arrayidx2
1338 /// %cmp = icmp ult %x, %y
1339 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1340 /// store i32 %add.add5, i32* %arrayidx2
1343 /// \return The pointer to the value of the previous store if the store can be
1344 /// hoisted into the predecessor block. 0 otherwise.
1345 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1346 BasicBlock *StoreBB, BasicBlock *EndBB) {
1347 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1351 // Volatile or atomic.
1352 if (!StoreToHoist->isSimple())
1355 Value *StorePtr = StoreToHoist->getPointerOperand();
1357 // Look for a store to the same pointer in BrBB.
1358 unsigned MaxNumInstToLookAt = 10;
1359 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(),
1360 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) {
1361 Instruction *CurI = &*RI;
1363 // Could be calling an instruction that effects memory like free().
1364 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI))
1367 StoreInst *SI = dyn_cast<StoreInst>(CurI);
1368 // Found the previous store make sure it stores to the same location.
1369 if (SI && SI->getPointerOperand() == StorePtr)
1370 // Found the previous store, return its value operand.
1371 return SI->getValueOperand();
1373 return nullptr; // Unknown store.
1379 /// \brief Speculate a conditional basic block flattening the CFG.
1381 /// Note that this is a very risky transform currently. Speculating
1382 /// instructions like this is most often not desirable. Instead, there is an MI
1383 /// pass which can do it with full awareness of the resource constraints.
1384 /// However, some cases are "obvious" and we should do directly. An example of
1385 /// this is speculating a single, reasonably cheap instruction.
1387 /// There is only one distinct advantage to flattening the CFG at the IR level:
1388 /// it makes very common but simplistic optimizations such as are common in
1389 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1390 /// modeling their effects with easier to reason about SSA value graphs.
1393 /// An illustration of this transform is turning this IR:
1396 /// %cmp = icmp ult %x, %y
1397 /// br i1 %cmp, label %EndBB, label %ThenBB
1399 /// %sub = sub %x, %y
1402 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1409 /// %cmp = icmp ult %x, %y
1410 /// %sub = sub %x, %y
1411 /// %cond = select i1 %cmp, 0, %sub
1415 /// \returns true if the conditional block is removed.
1416 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1417 const DataLayout *DL) {
1418 // Be conservative for now. FP select instruction can often be expensive.
1419 Value *BrCond = BI->getCondition();
1420 if (isa<FCmpInst>(BrCond))
1423 BasicBlock *BB = BI->getParent();
1424 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1426 // If ThenBB is actually on the false edge of the conditional branch, remember
1427 // to swap the select operands later.
1428 bool Invert = false;
1429 if (ThenBB != BI->getSuccessor(0)) {
1430 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1433 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1435 // Keep a count of how many times instructions are used within CondBB when
1436 // they are candidates for sinking into CondBB. Specifically:
1437 // - They are defined in BB, and
1438 // - They have no side effects, and
1439 // - All of their uses are in CondBB.
1440 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1442 unsigned SpeculationCost = 0;
1443 Value *SpeculatedStoreValue = nullptr;
1444 StoreInst *SpeculatedStore = nullptr;
1445 for (BasicBlock::iterator BBI = ThenBB->begin(),
1446 BBE = std::prev(ThenBB->end());
1447 BBI != BBE; ++BBI) {
1448 Instruction *I = BBI;
1450 if (isa<DbgInfoIntrinsic>(I))
1453 // Only speculatively execution a single instruction (not counting the
1454 // terminator) for now.
1456 if (SpeculationCost > 1)
1459 // Don't hoist the instruction if it's unsafe or expensive.
1460 if (!isSafeToSpeculativelyExecute(I, DL) &&
1461 !(HoistCondStores &&
1462 (SpeculatedStoreValue = isSafeToSpeculateStore(I, BB, ThenBB,
1465 if (!SpeculatedStoreValue &&
1466 ComputeSpeculationCost(I, DL) > PHINodeFoldingThreshold)
1469 // Store the store speculation candidate.
1470 if (SpeculatedStoreValue)
1471 SpeculatedStore = cast<StoreInst>(I);
1473 // Do not hoist the instruction if any of its operands are defined but not
1474 // used in BB. The transformation will prevent the operand from
1475 // being sunk into the use block.
1476 for (User::op_iterator i = I->op_begin(), e = I->op_end();
1478 Instruction *OpI = dyn_cast<Instruction>(*i);
1479 if (!OpI || OpI->getParent() != BB ||
1480 OpI->mayHaveSideEffects())
1481 continue; // Not a candidate for sinking.
1483 ++SinkCandidateUseCounts[OpI];
1487 // Consider any sink candidates which are only used in CondBB as costs for
1488 // speculation. Note, while we iterate over a DenseMap here, we are summing
1489 // and so iteration order isn't significant.
1490 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I =
1491 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end();
1493 if (I->first->getNumUses() == I->second) {
1495 if (SpeculationCost > 1)
1499 // Check that the PHI nodes can be converted to selects.
1500 bool HaveRewritablePHIs = false;
1501 for (BasicBlock::iterator I = EndBB->begin();
1502 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1503 Value *OrigV = PN->getIncomingValueForBlock(BB);
1504 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1506 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1507 // Skip PHIs which are trivial.
1511 HaveRewritablePHIs = true;
1512 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1513 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1514 if (!OrigCE && !ThenCE)
1515 continue; // Known safe and cheap.
1517 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE, DL)) ||
1518 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE, DL)))
1520 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, DL) : 0;
1521 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, DL) : 0;
1522 if (OrigCost + ThenCost > 2 * PHINodeFoldingThreshold)
1525 // Account for the cost of an unfolded ConstantExpr which could end up
1526 // getting expanded into Instructions.
1527 // FIXME: This doesn't account for how many operations are combined in the
1528 // constant expression.
1530 if (SpeculationCost > 1)
1534 // If there are no PHIs to process, bail early. This helps ensure idempotence
1536 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
1539 // If we get here, we can hoist the instruction and if-convert.
1540 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1542 // Insert a select of the value of the speculated store.
1543 if (SpeculatedStoreValue) {
1544 IRBuilder<true, NoFolder> Builder(BI);
1545 Value *TrueV = SpeculatedStore->getValueOperand();
1546 Value *FalseV = SpeculatedStoreValue;
1548 std::swap(TrueV, FalseV);
1549 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() +
1550 "." + FalseV->getName());
1551 SpeculatedStore->setOperand(0, S);
1554 // Hoist the instructions.
1555 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(),
1556 std::prev(ThenBB->end()));
1558 // Insert selects and rewrite the PHI operands.
1559 IRBuilder<true, NoFolder> Builder(BI);
1560 for (BasicBlock::iterator I = EndBB->begin();
1561 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1562 unsigned OrigI = PN->getBasicBlockIndex(BB);
1563 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1564 Value *OrigV = PN->getIncomingValue(OrigI);
1565 Value *ThenV = PN->getIncomingValue(ThenI);
1567 // Skip PHIs which are trivial.
1571 // Create a select whose true value is the speculatively executed value and
1572 // false value is the preexisting value. Swap them if the branch
1573 // destinations were inverted.
1574 Value *TrueV = ThenV, *FalseV = OrigV;
1576 std::swap(TrueV, FalseV);
1577 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV,
1578 TrueV->getName() + "." + FalseV->getName());
1579 PN->setIncomingValue(OrigI, V);
1580 PN->setIncomingValue(ThenI, V);
1587 /// \returns True if this block contains a CallInst with the NoDuplicate
1589 static bool HasNoDuplicateCall(const BasicBlock *BB) {
1590 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1591 const CallInst *CI = dyn_cast<CallInst>(I);
1594 if (CI->cannotDuplicate())
1600 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1601 /// across this block.
1602 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1603 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1606 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1607 if (isa<DbgInfoIntrinsic>(BBI))
1609 if (Size > 10) return false; // Don't clone large BB's.
1612 // We can only support instructions that do not define values that are
1613 // live outside of the current basic block.
1614 for (User *U : BBI->users()) {
1615 Instruction *UI = cast<Instruction>(U);
1616 if (UI->getParent() != BB || isa<PHINode>(UI)) return false;
1619 // Looks ok, continue checking.
1625 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1626 /// that is defined in the same block as the branch and if any PHI entries are
1627 /// constants, thread edges corresponding to that entry to be branches to their
1628 /// ultimate destination.
1629 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *DL) {
1630 BasicBlock *BB = BI->getParent();
1631 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1632 // NOTE: we currently cannot transform this case if the PHI node is used
1633 // outside of the block.
1634 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1637 // Degenerate case of a single entry PHI.
1638 if (PN->getNumIncomingValues() == 1) {
1639 FoldSingleEntryPHINodes(PN->getParent());
1643 // Now we know that this block has multiple preds and two succs.
1644 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1646 if (HasNoDuplicateCall(BB)) return false;
1648 // Okay, this is a simple enough basic block. See if any phi values are
1650 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1651 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1652 if (!CB || !CB->getType()->isIntegerTy(1)) continue;
1654 // Okay, we now know that all edges from PredBB should be revectored to
1655 // branch to RealDest.
1656 BasicBlock *PredBB = PN->getIncomingBlock(i);
1657 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1659 if (RealDest == BB) continue; // Skip self loops.
1660 // Skip if the predecessor's terminator is an indirect branch.
1661 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1663 // The dest block might have PHI nodes, other predecessors and other
1664 // difficult cases. Instead of being smart about this, just insert a new
1665 // block that jumps to the destination block, effectively splitting
1666 // the edge we are about to create.
1667 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1668 RealDest->getName()+".critedge",
1669 RealDest->getParent(), RealDest);
1670 BranchInst::Create(RealDest, EdgeBB);
1672 // Update PHI nodes.
1673 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1675 // BB may have instructions that are being threaded over. Clone these
1676 // instructions into EdgeBB. We know that there will be no uses of the
1677 // cloned instructions outside of EdgeBB.
1678 BasicBlock::iterator InsertPt = EdgeBB->begin();
1679 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1680 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1681 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1682 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1685 // Clone the instruction.
1686 Instruction *N = BBI->clone();
1687 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1689 // Update operands due to translation.
1690 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1692 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1693 if (PI != TranslateMap.end())
1697 // Check for trivial simplification.
1698 if (Value *V = SimplifyInstruction(N, DL)) {
1699 TranslateMap[BBI] = V;
1700 delete N; // Instruction folded away, don't need actual inst
1702 // Insert the new instruction into its new home.
1703 EdgeBB->getInstList().insert(InsertPt, N);
1704 if (!BBI->use_empty())
1705 TranslateMap[BBI] = N;
1709 // Loop over all of the edges from PredBB to BB, changing them to branch
1710 // to EdgeBB instead.
1711 TerminatorInst *PredBBTI = PredBB->getTerminator();
1712 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1713 if (PredBBTI->getSuccessor(i) == BB) {
1714 BB->removePredecessor(PredBB);
1715 PredBBTI->setSuccessor(i, EdgeBB);
1718 // Recurse, simplifying any other constants.
1719 return FoldCondBranchOnPHI(BI, DL) | true;
1725 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1726 /// PHI node, see if we can eliminate it.
1727 static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *DL) {
1728 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1729 // statement", which has a very simple dominance structure. Basically, we
1730 // are trying to find the condition that is being branched on, which
1731 // subsequently causes this merge to happen. We really want control
1732 // dependence information for this check, but simplifycfg can't keep it up
1733 // to date, and this catches most of the cases we care about anyway.
1734 BasicBlock *BB = PN->getParent();
1735 BasicBlock *IfTrue, *IfFalse;
1736 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1738 // Don't bother if the branch will be constant folded trivially.
1739 isa<ConstantInt>(IfCond))
1742 // Okay, we found that we can merge this two-entry phi node into a select.
1743 // Doing so would require us to fold *all* two entry phi nodes in this block.
1744 // At some point this becomes non-profitable (particularly if the target
1745 // doesn't support cmov's). Only do this transformation if there are two or
1746 // fewer PHI nodes in this block.
1747 unsigned NumPhis = 0;
1748 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1752 // Loop over the PHI's seeing if we can promote them all to select
1753 // instructions. While we are at it, keep track of the instructions
1754 // that need to be moved to the dominating block.
1755 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1756 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1757 MaxCostVal1 = PHINodeFoldingThreshold;
1759 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1760 PHINode *PN = cast<PHINode>(II++);
1761 if (Value *V = SimplifyInstruction(PN, DL)) {
1762 PN->replaceAllUsesWith(V);
1763 PN->eraseFromParent();
1767 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1769 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1774 // If we folded the first phi, PN dangles at this point. Refresh it. If
1775 // we ran out of PHIs then we simplified them all.
1776 PN = dyn_cast<PHINode>(BB->begin());
1777 if (!PN) return true;
1779 // Don't fold i1 branches on PHIs which contain binary operators. These can
1780 // often be turned into switches and other things.
1781 if (PN->getType()->isIntegerTy(1) &&
1782 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1783 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1784 isa<BinaryOperator>(IfCond)))
1787 // If we all PHI nodes are promotable, check to make sure that all
1788 // instructions in the predecessor blocks can be promoted as well. If
1789 // not, we won't be able to get rid of the control flow, so it's not
1790 // worth promoting to select instructions.
1791 BasicBlock *DomBlock = nullptr;
1792 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1793 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1794 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1797 DomBlock = *pred_begin(IfBlock1);
1798 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1799 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1800 // This is not an aggressive instruction that we can promote.
1801 // Because of this, we won't be able to get rid of the control
1802 // flow, so the xform is not worth it.
1807 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1810 DomBlock = *pred_begin(IfBlock2);
1811 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1812 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1813 // This is not an aggressive instruction that we can promote.
1814 // Because of this, we won't be able to get rid of the control
1815 // flow, so the xform is not worth it.
1820 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1821 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1823 // If we can still promote the PHI nodes after this gauntlet of tests,
1824 // do all of the PHI's now.
1825 Instruction *InsertPt = DomBlock->getTerminator();
1826 IRBuilder<true, NoFolder> Builder(InsertPt);
1828 // Move all 'aggressive' instructions, which are defined in the
1829 // conditional parts of the if's up to the dominating block.
1831 DomBlock->getInstList().splice(InsertPt,
1832 IfBlock1->getInstList(), IfBlock1->begin(),
1833 IfBlock1->getTerminator());
1835 DomBlock->getInstList().splice(InsertPt,
1836 IfBlock2->getInstList(), IfBlock2->begin(),
1837 IfBlock2->getTerminator());
1839 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1840 // Change the PHI node into a select instruction.
1841 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1842 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1845 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1846 PN->replaceAllUsesWith(NV);
1848 PN->eraseFromParent();
1851 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1852 // has been flattened. Change DomBlock to jump directly to our new block to
1853 // avoid other simplifycfg's kicking in on the diamond.
1854 TerminatorInst *OldTI = DomBlock->getTerminator();
1855 Builder.SetInsertPoint(OldTI);
1856 Builder.CreateBr(BB);
1857 OldTI->eraseFromParent();
1861 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1862 /// to two returning blocks, try to merge them together into one return,
1863 /// introducing a select if the return values disagree.
1864 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1865 IRBuilder<> &Builder) {
1866 assert(BI->isConditional() && "Must be a conditional branch");
1867 BasicBlock *TrueSucc = BI->getSuccessor(0);
1868 BasicBlock *FalseSucc = BI->getSuccessor(1);
1869 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1870 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1872 // Check to ensure both blocks are empty (just a return) or optionally empty
1873 // with PHI nodes. If there are other instructions, merging would cause extra
1874 // computation on one path or the other.
1875 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1877 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1880 Builder.SetInsertPoint(BI);
1881 // Okay, we found a branch that is going to two return nodes. If
1882 // there is no return value for this function, just change the
1883 // branch into a return.
1884 if (FalseRet->getNumOperands() == 0) {
1885 TrueSucc->removePredecessor(BI->getParent());
1886 FalseSucc->removePredecessor(BI->getParent());
1887 Builder.CreateRetVoid();
1888 EraseTerminatorInstAndDCECond(BI);
1892 // Otherwise, figure out what the true and false return values are
1893 // so we can insert a new select instruction.
1894 Value *TrueValue = TrueRet->getReturnValue();
1895 Value *FalseValue = FalseRet->getReturnValue();
1897 // Unwrap any PHI nodes in the return blocks.
1898 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1899 if (TVPN->getParent() == TrueSucc)
1900 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1901 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1902 if (FVPN->getParent() == FalseSucc)
1903 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1905 // In order for this transformation to be safe, we must be able to
1906 // unconditionally execute both operands to the return. This is
1907 // normally the case, but we could have a potentially-trapping
1908 // constant expression that prevents this transformation from being
1910 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1913 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1917 // Okay, we collected all the mapped values and checked them for sanity, and
1918 // defined to really do this transformation. First, update the CFG.
1919 TrueSucc->removePredecessor(BI->getParent());
1920 FalseSucc->removePredecessor(BI->getParent());
1922 // Insert select instructions where needed.
1923 Value *BrCond = BI->getCondition();
1925 // Insert a select if the results differ.
1926 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1927 } else if (isa<UndefValue>(TrueValue)) {
1928 TrueValue = FalseValue;
1930 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1931 FalseValue, "retval");
1935 Value *RI = !TrueValue ?
1936 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
1940 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
1941 << "\n " << *BI << "NewRet = " << *RI
1942 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
1944 EraseTerminatorInstAndDCECond(BI);
1949 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
1950 /// probabilities of the branch taking each edge. Fills in the two APInt
1951 /// parameters and return true, or returns false if no or invalid metadata was
1953 static bool ExtractBranchMetadata(BranchInst *BI,
1954 uint64_t &ProbTrue, uint64_t &ProbFalse) {
1955 assert(BI->isConditional() &&
1956 "Looking for probabilities on unconditional branch?");
1957 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
1958 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
1959 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
1960 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
1961 if (!CITrue || !CIFalse) return false;
1962 ProbTrue = CITrue->getValue().getZExtValue();
1963 ProbFalse = CIFalse->getValue().getZExtValue();
1967 /// checkCSEInPredecessor - Return true if the given instruction is available
1968 /// in its predecessor block. If yes, the instruction will be removed.
1970 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
1971 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
1973 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
1974 Instruction *PBI = &*I;
1975 // Check whether Inst and PBI generate the same value.
1976 if (Inst->isIdenticalTo(PBI)) {
1977 Inst->replaceAllUsesWith(PBI);
1978 Inst->eraseFromParent();
1985 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
1986 /// predecessor branches to us and one of our successors, fold the block into
1987 /// the predecessor and use logical operations to pick the right destination.
1988 bool llvm::FoldBranchToCommonDest(BranchInst *BI, const DataLayout *DL,
1989 unsigned BonusInstThreshold) {
1990 BasicBlock *BB = BI->getParent();
1992 Instruction *Cond = nullptr;
1993 if (BI->isConditional())
1994 Cond = dyn_cast<Instruction>(BI->getCondition());
1996 // For unconditional branch, check for a simple CFG pattern, where
1997 // BB has a single predecessor and BB's successor is also its predecessor's
1998 // successor. If such pattern exisits, check for CSE between BB and its
2000 if (BasicBlock *PB = BB->getSinglePredecessor())
2001 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2002 if (PBI->isConditional() &&
2003 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2004 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2005 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
2007 Instruction *Curr = I++;
2008 if (isa<CmpInst>(Curr)) {
2012 // Quit if we can't remove this instruction.
2013 if (!checkCSEInPredecessor(Curr, PB))
2022 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2023 Cond->getParent() != BB || !Cond->hasOneUse())
2026 // Make sure the instruction after the condition is the cond branch.
2027 BasicBlock::iterator CondIt = Cond; ++CondIt;
2029 // Ignore dbg intrinsics.
2030 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
2035 // Only allow this transformation if computing the condition doesn't involve
2036 // too many instructions and these involved instructions can be executed
2037 // unconditionally. We denote all involved instructions except the condition
2038 // as "bonus instructions", and only allow this transformation when the
2039 // number of the bonus instructions does not exceed a certain threshold.
2040 unsigned NumBonusInsts = 0;
2041 for (auto I = BB->begin(); Cond != I; ++I) {
2042 // Ignore dbg intrinsics.
2043 if (isa<DbgInfoIntrinsic>(I))
2045 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(I, DL))
2047 // I has only one use and can be executed unconditionally.
2048 Instruction *User = dyn_cast<Instruction>(I->user_back());
2049 if (User == nullptr || User->getParent() != BB)
2051 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2052 // to use any other instruction, User must be an instruction between next(I)
2055 // Early exits once we reach the limit.
2056 if (NumBonusInsts > BonusInstThreshold)
2060 // Cond is known to be a compare or binary operator. Check to make sure that
2061 // neither operand is a potentially-trapping constant expression.
2062 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2065 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2069 // Finally, don't infinitely unroll conditional loops.
2070 BasicBlock *TrueDest = BI->getSuccessor(0);
2071 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2072 if (TrueDest == BB || FalseDest == BB)
2075 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2076 BasicBlock *PredBlock = *PI;
2077 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2079 // Check that we have two conditional branches. If there is a PHI node in
2080 // the common successor, verify that the same value flows in from both
2082 SmallVector<PHINode*, 4> PHIs;
2083 if (!PBI || PBI->isUnconditional() ||
2084 (BI->isConditional() &&
2085 !SafeToMergeTerminators(BI, PBI)) ||
2086 (!BI->isConditional() &&
2087 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2090 // Determine if the two branches share a common destination.
2091 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2092 bool InvertPredCond = false;
2094 if (BI->isConditional()) {
2095 if (PBI->getSuccessor(0) == TrueDest)
2096 Opc = Instruction::Or;
2097 else if (PBI->getSuccessor(1) == FalseDest)
2098 Opc = Instruction::And;
2099 else if (PBI->getSuccessor(0) == FalseDest)
2100 Opc = Instruction::And, InvertPredCond = true;
2101 else if (PBI->getSuccessor(1) == TrueDest)
2102 Opc = Instruction::Or, InvertPredCond = true;
2106 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2110 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2111 IRBuilder<> Builder(PBI);
2113 // If we need to invert the condition in the pred block to match, do so now.
2114 if (InvertPredCond) {
2115 Value *NewCond = PBI->getCondition();
2117 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2118 CmpInst *CI = cast<CmpInst>(NewCond);
2119 CI->setPredicate(CI->getInversePredicate());
2121 NewCond = Builder.CreateNot(NewCond,
2122 PBI->getCondition()->getName()+".not");
2125 PBI->setCondition(NewCond);
2126 PBI->swapSuccessors();
2129 // If we have bonus instructions, clone them into the predecessor block.
2130 // Note that there may be mutliple predecessor blocks, so we cannot move
2131 // bonus instructions to a predecessor block.
2132 ValueToValueMapTy VMap; // maps original values to cloned values
2133 // We already make sure Cond is the last instruction before BI. Therefore,
2134 // every instructions before Cond other than DbgInfoIntrinsic are bonus
2136 for (auto BonusInst = BB->begin(); Cond != BonusInst; ++BonusInst) {
2137 if (isa<DbgInfoIntrinsic>(BonusInst))
2139 Instruction *NewBonusInst = BonusInst->clone();
2140 RemapInstruction(NewBonusInst, VMap,
2141 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2142 VMap[BonusInst] = NewBonusInst;
2144 // If we moved a load, we cannot any longer claim any knowledge about
2145 // its potential value. The previous information might have been valid
2146 // only given the branch precondition.
2147 // For an analogous reason, we must also drop all the metadata whose
2148 // semantics we don't understand.
2149 NewBonusInst->dropUnknownMetadata(LLVMContext::MD_dbg);
2151 PredBlock->getInstList().insert(PBI, NewBonusInst);
2152 NewBonusInst->takeName(BonusInst);
2153 BonusInst->setName(BonusInst->getName() + ".old");
2156 // Clone Cond into the predecessor basic block, and or/and the
2157 // two conditions together.
2158 Instruction *New = Cond->clone();
2159 RemapInstruction(New, VMap,
2160 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2161 PredBlock->getInstList().insert(PBI, New);
2162 New->takeName(Cond);
2163 Cond->setName(New->getName() + ".old");
2165 if (BI->isConditional()) {
2166 Instruction *NewCond =
2167 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2169 PBI->setCondition(NewCond);
2171 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2172 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2174 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2176 SmallVector<uint64_t, 8> NewWeights;
2178 if (PBI->getSuccessor(0) == BB) {
2179 if (PredHasWeights && SuccHasWeights) {
2180 // PBI: br i1 %x, BB, FalseDest
2181 // BI: br i1 %y, TrueDest, FalseDest
2182 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2183 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2184 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2185 // TrueWeight for PBI * FalseWeight for BI.
2186 // We assume that total weights of a BranchInst can fit into 32 bits.
2187 // Therefore, we will not have overflow using 64-bit arithmetic.
2188 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2189 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2191 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2192 PBI->setSuccessor(0, TrueDest);
2194 if (PBI->getSuccessor(1) == BB) {
2195 if (PredHasWeights && SuccHasWeights) {
2196 // PBI: br i1 %x, TrueDest, BB
2197 // BI: br i1 %y, TrueDest, FalseDest
2198 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2199 // FalseWeight for PBI * TrueWeight for BI.
2200 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2201 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2202 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2203 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2205 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2206 PBI->setSuccessor(1, FalseDest);
2208 if (NewWeights.size() == 2) {
2209 // Halve the weights if any of them cannot fit in an uint32_t
2210 FitWeights(NewWeights);
2212 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2213 PBI->setMetadata(LLVMContext::MD_prof,
2214 MDBuilder(BI->getContext()).
2215 createBranchWeights(MDWeights));
2217 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2219 // Update PHI nodes in the common successors.
2220 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2221 ConstantInt *PBI_C = cast<ConstantInt>(
2222 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2223 assert(PBI_C->getType()->isIntegerTy(1));
2224 Instruction *MergedCond = nullptr;
2225 if (PBI->getSuccessor(0) == TrueDest) {
2226 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2227 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2228 // is false: !PBI_Cond and BI_Value
2229 Instruction *NotCond =
2230 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2233 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2238 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2239 PBI->getCondition(), MergedCond,
2242 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2243 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2244 // is false: PBI_Cond and BI_Value
2246 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2247 PBI->getCondition(), New,
2249 if (PBI_C->isOne()) {
2250 Instruction *NotCond =
2251 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2254 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2255 NotCond, MergedCond,
2260 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2263 // Change PBI from Conditional to Unconditional.
2264 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2265 EraseTerminatorInstAndDCECond(PBI);
2269 // TODO: If BB is reachable from all paths through PredBlock, then we
2270 // could replace PBI's branch probabilities with BI's.
2272 // Copy any debug value intrinsics into the end of PredBlock.
2273 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2274 if (isa<DbgInfoIntrinsic>(*I))
2275 I->clone()->insertBefore(PBI);
2282 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2283 /// predecessor of another block, this function tries to simplify it. We know
2284 /// that PBI and BI are both conditional branches, and BI is in one of the
2285 /// successor blocks of PBI - PBI branches to BI.
2286 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2287 assert(PBI->isConditional() && BI->isConditional());
2288 BasicBlock *BB = BI->getParent();
2290 // If this block ends with a branch instruction, and if there is a
2291 // predecessor that ends on a branch of the same condition, make
2292 // this conditional branch redundant.
2293 if (PBI->getCondition() == BI->getCondition() &&
2294 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2295 // Okay, the outcome of this conditional branch is statically
2296 // knowable. If this block had a single pred, handle specially.
2297 if (BB->getSinglePredecessor()) {
2298 // Turn this into a branch on constant.
2299 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2300 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2302 return true; // Nuke the branch on constant.
2305 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2306 // in the constant and simplify the block result. Subsequent passes of
2307 // simplifycfg will thread the block.
2308 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2309 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2310 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2311 std::distance(PB, PE),
2312 BI->getCondition()->getName() + ".pr",
2314 // Okay, we're going to insert the PHI node. Since PBI is not the only
2315 // predecessor, compute the PHI'd conditional value for all of the preds.
2316 // Any predecessor where the condition is not computable we keep symbolic.
2317 for (pred_iterator PI = PB; PI != PE; ++PI) {
2318 BasicBlock *P = *PI;
2319 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2320 PBI != BI && PBI->isConditional() &&
2321 PBI->getCondition() == BI->getCondition() &&
2322 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2323 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2324 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2327 NewPN->addIncoming(BI->getCondition(), P);
2331 BI->setCondition(NewPN);
2336 // If this is a conditional branch in an empty block, and if any
2337 // predecessors are a conditional branch to one of our destinations,
2338 // fold the conditions into logical ops and one cond br.
2339 BasicBlock::iterator BBI = BB->begin();
2340 // Ignore dbg intrinsics.
2341 while (isa<DbgInfoIntrinsic>(BBI))
2347 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2352 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2354 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2355 PBIOp = 0, BIOp = 1;
2356 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2357 PBIOp = 1, BIOp = 0;
2358 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2363 // Check to make sure that the other destination of this branch
2364 // isn't BB itself. If so, this is an infinite loop that will
2365 // keep getting unwound.
2366 if (PBI->getSuccessor(PBIOp) == BB)
2369 // Do not perform this transformation if it would require
2370 // insertion of a large number of select instructions. For targets
2371 // without predication/cmovs, this is a big pessimization.
2373 // Also do not perform this transformation if any phi node in the common
2374 // destination block can trap when reached by BB or PBB (PR17073). In that
2375 // case, it would be unsafe to hoist the operation into a select instruction.
2377 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2378 unsigned NumPhis = 0;
2379 for (BasicBlock::iterator II = CommonDest->begin();
2380 isa<PHINode>(II); ++II, ++NumPhis) {
2381 if (NumPhis > 2) // Disable this xform.
2384 PHINode *PN = cast<PHINode>(II);
2385 Value *BIV = PN->getIncomingValueForBlock(BB);
2386 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
2390 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2391 Value *PBIV = PN->getIncomingValue(PBBIdx);
2392 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
2397 // Finally, if everything is ok, fold the branches to logical ops.
2398 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2400 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2401 << "AND: " << *BI->getParent());
2404 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2405 // branch in it, where one edge (OtherDest) goes back to itself but the other
2406 // exits. We don't *know* that the program avoids the infinite loop
2407 // (even though that seems likely). If we do this xform naively, we'll end up
2408 // recursively unpeeling the loop. Since we know that (after the xform is
2409 // done) that the block *is* infinite if reached, we just make it an obviously
2410 // infinite loop with no cond branch.
2411 if (OtherDest == BB) {
2412 // Insert it at the end of the function, because it's either code,
2413 // or it won't matter if it's hot. :)
2414 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2415 "infloop", BB->getParent());
2416 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2417 OtherDest = InfLoopBlock;
2420 DEBUG(dbgs() << *PBI->getParent()->getParent());
2422 // BI may have other predecessors. Because of this, we leave
2423 // it alone, but modify PBI.
2425 // Make sure we get to CommonDest on True&True directions.
2426 Value *PBICond = PBI->getCondition();
2427 IRBuilder<true, NoFolder> Builder(PBI);
2429 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2431 Value *BICond = BI->getCondition();
2433 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2435 // Merge the conditions.
2436 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2438 // Modify PBI to branch on the new condition to the new dests.
2439 PBI->setCondition(Cond);
2440 PBI->setSuccessor(0, CommonDest);
2441 PBI->setSuccessor(1, OtherDest);
2443 // Update branch weight for PBI.
2444 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2445 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2447 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2449 if (PredHasWeights && SuccHasWeights) {
2450 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2451 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2452 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2453 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2454 // The weight to CommonDest should be PredCommon * SuccTotal +
2455 // PredOther * SuccCommon.
2456 // The weight to OtherDest should be PredOther * SuccOther.
2457 SmallVector<uint64_t, 2> NewWeights;
2458 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) +
2459 PredOther * SuccCommon);
2460 NewWeights.push_back(PredOther * SuccOther);
2461 // Halve the weights if any of them cannot fit in an uint32_t
2462 FitWeights(NewWeights);
2464 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end());
2465 PBI->setMetadata(LLVMContext::MD_prof,
2466 MDBuilder(BI->getContext()).
2467 createBranchWeights(MDWeights));
2470 // OtherDest may have phi nodes. If so, add an entry from PBI's
2471 // block that are identical to the entries for BI's block.
2472 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2474 // We know that the CommonDest already had an edge from PBI to
2475 // it. If it has PHIs though, the PHIs may have different
2476 // entries for BB and PBI's BB. If so, insert a select to make
2479 for (BasicBlock::iterator II = CommonDest->begin();
2480 (PN = dyn_cast<PHINode>(II)); ++II) {
2481 Value *BIV = PN->getIncomingValueForBlock(BB);
2482 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2483 Value *PBIV = PN->getIncomingValue(PBBIdx);
2485 // Insert a select in PBI to pick the right value.
2486 Value *NV = cast<SelectInst>
2487 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2488 PN->setIncomingValue(PBBIdx, NV);
2492 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2493 DEBUG(dbgs() << *PBI->getParent()->getParent());
2495 // This basic block is probably dead. We know it has at least
2496 // one fewer predecessor.
2500 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2501 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2502 // Takes care of updating the successors and removing the old terminator.
2503 // Also makes sure not to introduce new successors by assuming that edges to
2504 // non-successor TrueBBs and FalseBBs aren't reachable.
2505 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2506 BasicBlock *TrueBB, BasicBlock *FalseBB,
2507 uint32_t TrueWeight,
2508 uint32_t FalseWeight){
2509 // Remove any superfluous successor edges from the CFG.
2510 // First, figure out which successors to preserve.
2511 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2513 BasicBlock *KeepEdge1 = TrueBB;
2514 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
2516 // Then remove the rest.
2517 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2518 BasicBlock *Succ = OldTerm->getSuccessor(I);
2519 // Make sure only to keep exactly one copy of each edge.
2520 if (Succ == KeepEdge1)
2521 KeepEdge1 = nullptr;
2522 else if (Succ == KeepEdge2)
2523 KeepEdge2 = nullptr;
2525 Succ->removePredecessor(OldTerm->getParent());
2528 IRBuilder<> Builder(OldTerm);
2529 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2531 // Insert an appropriate new terminator.
2532 if (!KeepEdge1 && !KeepEdge2) {
2533 if (TrueBB == FalseBB)
2534 // We were only looking for one successor, and it was present.
2535 // Create an unconditional branch to it.
2536 Builder.CreateBr(TrueBB);
2538 // We found both of the successors we were looking for.
2539 // Create a conditional branch sharing the condition of the select.
2540 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2541 if (TrueWeight != FalseWeight)
2542 NewBI->setMetadata(LLVMContext::MD_prof,
2543 MDBuilder(OldTerm->getContext()).
2544 createBranchWeights(TrueWeight, FalseWeight));
2546 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2547 // Neither of the selected blocks were successors, so this
2548 // terminator must be unreachable.
2549 new UnreachableInst(OldTerm->getContext(), OldTerm);
2551 // One of the selected values was a successor, but the other wasn't.
2552 // Insert an unconditional branch to the one that was found;
2553 // the edge to the one that wasn't must be unreachable.
2555 // Only TrueBB was found.
2556 Builder.CreateBr(TrueBB);
2558 // Only FalseBB was found.
2559 Builder.CreateBr(FalseBB);
2562 EraseTerminatorInstAndDCECond(OldTerm);
2566 // SimplifySwitchOnSelect - Replaces
2567 // (switch (select cond, X, Y)) on constant X, Y
2568 // with a branch - conditional if X and Y lead to distinct BBs,
2569 // unconditional otherwise.
2570 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2571 // Check for constant integer values in the select.
2572 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2573 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2574 if (!TrueVal || !FalseVal)
2577 // Find the relevant condition and destinations.
2578 Value *Condition = Select->getCondition();
2579 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2580 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2582 // Get weight for TrueBB and FalseBB.
2583 uint32_t TrueWeight = 0, FalseWeight = 0;
2584 SmallVector<uint64_t, 8> Weights;
2585 bool HasWeights = HasBranchWeights(SI);
2587 GetBranchWeights(SI, Weights);
2588 if (Weights.size() == 1 + SI->getNumCases()) {
2589 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2590 getSuccessorIndex()];
2591 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2592 getSuccessorIndex()];
2596 // Perform the actual simplification.
2597 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2598 TrueWeight, FalseWeight);
2601 // SimplifyIndirectBrOnSelect - Replaces
2602 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2603 // blockaddress(@fn, BlockB)))
2605 // (br cond, BlockA, BlockB).
2606 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2607 // Check that both operands of the select are block addresses.
2608 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2609 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2613 // Extract the actual blocks.
2614 BasicBlock *TrueBB = TBA->getBasicBlock();
2615 BasicBlock *FalseBB = FBA->getBasicBlock();
2617 // Perform the actual simplification.
2618 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2622 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2623 /// instruction (a seteq/setne with a constant) as the only instruction in a
2624 /// block that ends with an uncond branch. We are looking for a very specific
2625 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2626 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2627 /// default value goes to an uncond block with a seteq in it, we get something
2630 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2632 /// %tmp = icmp eq i8 %A, 92
2635 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2637 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2638 /// the PHI, merging the third icmp into the switch.
2639 static bool TryToSimplifyUncondBranchWithICmpInIt(
2640 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI,
2641 unsigned BonusInstThreshold, const DataLayout *DL, AssumptionTracker *AT) {
2642 BasicBlock *BB = ICI->getParent();
2644 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2646 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2648 Value *V = ICI->getOperand(0);
2649 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2651 // The pattern we're looking for is where our only predecessor is a switch on
2652 // 'V' and this block is the default case for the switch. In this case we can
2653 // fold the compared value into the switch to simplify things.
2654 BasicBlock *Pred = BB->getSinglePredecessor();
2655 if (!Pred || !isa<SwitchInst>(Pred->getTerminator())) return false;
2657 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2658 if (SI->getCondition() != V)
2661 // If BB is reachable on a non-default case, then we simply know the value of
2662 // V in this block. Substitute it and constant fold the icmp instruction
2664 if (SI->getDefaultDest() != BB) {
2665 ConstantInt *VVal = SI->findCaseDest(BB);
2666 assert(VVal && "Should have a unique destination value");
2667 ICI->setOperand(0, VVal);
2669 if (Value *V = SimplifyInstruction(ICI, DL)) {
2670 ICI->replaceAllUsesWith(V);
2671 ICI->eraseFromParent();
2673 // BB is now empty, so it is likely to simplify away.
2674 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
2677 // Ok, the block is reachable from the default dest. If the constant we're
2678 // comparing exists in one of the other edges, then we can constant fold ICI
2680 if (SI->findCaseValue(Cst) != SI->case_default()) {
2682 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2683 V = ConstantInt::getFalse(BB->getContext());
2685 V = ConstantInt::getTrue(BB->getContext());
2687 ICI->replaceAllUsesWith(V);
2688 ICI->eraseFromParent();
2689 // BB is now empty, so it is likely to simplify away.
2690 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
2693 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2695 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2696 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
2697 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
2698 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2701 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2703 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2704 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2706 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2707 std::swap(DefaultCst, NewCst);
2709 // Replace ICI (which is used by the PHI for the default value) with true or
2710 // false depending on if it is EQ or NE.
2711 ICI->replaceAllUsesWith(DefaultCst);
2712 ICI->eraseFromParent();
2714 // Okay, the switch goes to this block on a default value. Add an edge from
2715 // the switch to the merge point on the compared value.
2716 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2717 BB->getParent(), BB);
2718 SmallVector<uint64_t, 8> Weights;
2719 bool HasWeights = HasBranchWeights(SI);
2721 GetBranchWeights(SI, Weights);
2722 if (Weights.size() == 1 + SI->getNumCases()) {
2723 // Split weight for default case to case for "Cst".
2724 Weights[0] = (Weights[0]+1) >> 1;
2725 Weights.push_back(Weights[0]);
2727 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2728 SI->setMetadata(LLVMContext::MD_prof,
2729 MDBuilder(SI->getContext()).
2730 createBranchWeights(MDWeights));
2733 SI->addCase(Cst, NewBB);
2735 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2736 Builder.SetInsertPoint(NewBB);
2737 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2738 Builder.CreateBr(SuccBlock);
2739 PHIUse->addIncoming(NewCst, NewBB);
2743 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2744 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2745 /// fold it into a switch instruction if so.
2746 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *DL,
2747 IRBuilder<> &Builder) {
2748 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2749 if (!Cond) return false;
2752 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2753 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2754 // 'setne's and'ed together, collect them.
2755 Value *CompVal = nullptr;
2756 std::vector<ConstantInt*> Values;
2757 bool TrueWhenEqual = true;
2758 Value *ExtraCase = nullptr;
2759 unsigned UsedICmps = 0;
2761 if (Cond->getOpcode() == Instruction::Or) {
2762 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, DL, true,
2764 } else if (Cond->getOpcode() == Instruction::And) {
2765 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, DL, false,
2767 TrueWhenEqual = false;
2770 // If we didn't have a multiply compared value, fail.
2771 if (!CompVal) return false;
2773 // Avoid turning single icmps into a switch.
2777 // There might be duplicate constants in the list, which the switch
2778 // instruction can't handle, remove them now.
2779 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2780 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2782 // If Extra was used, we require at least two switch values to do the
2783 // transformation. A switch with one value is just an cond branch.
2784 if (ExtraCase && Values.size() < 2) return false;
2786 // TODO: Preserve branch weight metadata, similarly to how
2787 // FoldValueComparisonIntoPredecessors preserves it.
2789 // Figure out which block is which destination.
2790 BasicBlock *DefaultBB = BI->getSuccessor(1);
2791 BasicBlock *EdgeBB = BI->getSuccessor(0);
2792 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2794 BasicBlock *BB = BI->getParent();
2796 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2797 << " cases into SWITCH. BB is:\n" << *BB);
2799 // If there are any extra values that couldn't be folded into the switch
2800 // then we evaluate them with an explicit branch first. Split the block
2801 // right before the condbr to handle it.
2803 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2804 // Remove the uncond branch added to the old block.
2805 TerminatorInst *OldTI = BB->getTerminator();
2806 Builder.SetInsertPoint(OldTI);
2809 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2811 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2813 OldTI->eraseFromParent();
2815 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2816 // for the edge we just added.
2817 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2819 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2820 << "\nEXTRABB = " << *BB);
2824 Builder.SetInsertPoint(BI);
2825 // Convert pointer to int before we switch.
2826 if (CompVal->getType()->isPointerTy()) {
2827 assert(DL && "Cannot switch on pointer without DataLayout");
2828 CompVal = Builder.CreatePtrToInt(CompVal,
2829 DL->getIntPtrType(CompVal->getType()),
2833 // Create the new switch instruction now.
2834 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2836 // Add all of the 'cases' to the switch instruction.
2837 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2838 New->addCase(Values[i], EdgeBB);
2840 // We added edges from PI to the EdgeBB. As such, if there were any
2841 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2842 // the number of edges added.
2843 for (BasicBlock::iterator BBI = EdgeBB->begin();
2844 isa<PHINode>(BBI); ++BBI) {
2845 PHINode *PN = cast<PHINode>(BBI);
2846 Value *InVal = PN->getIncomingValueForBlock(BB);
2847 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2848 PN->addIncoming(InVal, BB);
2851 // Erase the old branch instruction.
2852 EraseTerminatorInstAndDCECond(BI);
2854 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2858 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2859 // If this is a trivial landing pad that just continues unwinding the caught
2860 // exception then zap the landing pad, turning its invokes into calls.
2861 BasicBlock *BB = RI->getParent();
2862 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2863 if (RI->getValue() != LPInst)
2864 // Not a landing pad, or the resume is not unwinding the exception that
2865 // caused control to branch here.
2868 // Check that there are no other instructions except for debug intrinsics.
2869 BasicBlock::iterator I = LPInst, E = RI;
2871 if (!isa<DbgInfoIntrinsic>(I))
2874 // Turn all invokes that unwind here into calls and delete the basic block.
2875 bool InvokeRequiresTableEntry = false;
2876 bool Changed = false;
2877 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2878 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2880 if (II->hasFnAttr(Attribute::UWTable)) {
2881 // Don't remove an `invoke' instruction if the ABI requires an entry into
2883 InvokeRequiresTableEntry = true;
2887 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2889 // Insert a call instruction before the invoke.
2890 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2892 Call->setCallingConv(II->getCallingConv());
2893 Call->setAttributes(II->getAttributes());
2894 Call->setDebugLoc(II->getDebugLoc());
2896 // Anything that used the value produced by the invoke instruction now uses
2897 // the value produced by the call instruction. Note that we do this even
2898 // for void functions and calls with no uses so that the callgraph edge is
2900 II->replaceAllUsesWith(Call);
2901 BB->removePredecessor(II->getParent());
2903 // Insert a branch to the normal destination right before the invoke.
2904 BranchInst::Create(II->getNormalDest(), II);
2906 // Finally, delete the invoke instruction!
2907 II->eraseFromParent();
2911 if (!InvokeRequiresTableEntry)
2912 // The landingpad is now unreachable. Zap it.
2913 BB->eraseFromParent();
2918 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2919 BasicBlock *BB = RI->getParent();
2920 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2922 // Find predecessors that end with branches.
2923 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2924 SmallVector<BranchInst*, 8> CondBranchPreds;
2925 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2926 BasicBlock *P = *PI;
2927 TerminatorInst *PTI = P->getTerminator();
2928 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2929 if (BI->isUnconditional())
2930 UncondBranchPreds.push_back(P);
2932 CondBranchPreds.push_back(BI);
2936 // If we found some, do the transformation!
2937 if (!UncondBranchPreds.empty() && DupRet) {
2938 while (!UncondBranchPreds.empty()) {
2939 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
2940 DEBUG(dbgs() << "FOLDING: " << *BB
2941 << "INTO UNCOND BRANCH PRED: " << *Pred);
2942 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
2945 // If we eliminated all predecessors of the block, delete the block now.
2946 if (pred_begin(BB) == pred_end(BB))
2947 // We know there are no successors, so just nuke the block.
2948 BB->eraseFromParent();
2953 // Check out all of the conditional branches going to this return
2954 // instruction. If any of them just select between returns, change the
2955 // branch itself into a select/return pair.
2956 while (!CondBranchPreds.empty()) {
2957 BranchInst *BI = CondBranchPreds.pop_back_val();
2959 // Check to see if the non-BB successor is also a return block.
2960 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
2961 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
2962 SimplifyCondBranchToTwoReturns(BI, Builder))
2968 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
2969 BasicBlock *BB = UI->getParent();
2971 bool Changed = false;
2973 // If there are any instructions immediately before the unreachable that can
2974 // be removed, do so.
2975 while (UI != BB->begin()) {
2976 BasicBlock::iterator BBI = UI;
2978 // Do not delete instructions that can have side effects which might cause
2979 // the unreachable to not be reachable; specifically, calls and volatile
2980 // operations may have this effect.
2981 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
2983 if (BBI->mayHaveSideEffects()) {
2984 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
2985 if (SI->isVolatile())
2987 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
2988 if (LI->isVolatile())
2990 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
2991 if (RMWI->isVolatile())
2993 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
2994 if (CXI->isVolatile())
2996 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
2997 !isa<LandingPadInst>(BBI)) {
3000 // Note that deleting LandingPad's here is in fact okay, although it
3001 // involves a bit of subtle reasoning. If this inst is a LandingPad,
3002 // all the predecessors of this block will be the unwind edges of Invokes,
3003 // and we can therefore guarantee this block will be erased.
3006 // Delete this instruction (any uses are guaranteed to be dead)
3007 if (!BBI->use_empty())
3008 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
3009 BBI->eraseFromParent();
3013 // If the unreachable instruction is the first in the block, take a gander
3014 // at all of the predecessors of this instruction, and simplify them.
3015 if (&BB->front() != UI) return Changed;
3017 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
3018 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
3019 TerminatorInst *TI = Preds[i]->getTerminator();
3020 IRBuilder<> Builder(TI);
3021 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3022 if (BI->isUnconditional()) {
3023 if (BI->getSuccessor(0) == BB) {
3024 new UnreachableInst(TI->getContext(), TI);
3025 TI->eraseFromParent();
3029 if (BI->getSuccessor(0) == BB) {
3030 Builder.CreateBr(BI->getSuccessor(1));
3031 EraseTerminatorInstAndDCECond(BI);
3032 } else if (BI->getSuccessor(1) == BB) {
3033 Builder.CreateBr(BI->getSuccessor(0));
3034 EraseTerminatorInstAndDCECond(BI);
3038 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3039 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3041 if (i.getCaseSuccessor() == BB) {
3042 BB->removePredecessor(SI->getParent());
3047 // If the default value is unreachable, figure out the most popular
3048 // destination and make it the default.
3049 if (SI->getDefaultDest() == BB) {
3050 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
3051 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3053 std::pair<unsigned, unsigned> &entry =
3054 Popularity[i.getCaseSuccessor()];
3055 if (entry.first == 0) {
3057 entry.second = i.getCaseIndex();
3063 // Find the most popular block.
3064 unsigned MaxPop = 0;
3065 unsigned MaxIndex = 0;
3066 BasicBlock *MaxBlock = nullptr;
3067 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
3068 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
3069 if (I->second.first > MaxPop ||
3070 (I->second.first == MaxPop && MaxIndex > I->second.second)) {
3071 MaxPop = I->second.first;
3072 MaxIndex = I->second.second;
3073 MaxBlock = I->first;
3077 // Make this the new default, allowing us to delete any explicit
3079 SI->setDefaultDest(MaxBlock);
3082 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
3084 if (isa<PHINode>(MaxBlock->begin()))
3085 for (unsigned i = 0; i != MaxPop-1; ++i)
3086 MaxBlock->removePredecessor(SI->getParent());
3088 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3090 if (i.getCaseSuccessor() == MaxBlock) {
3096 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
3097 if (II->getUnwindDest() == BB) {
3098 // Convert the invoke to a call instruction. This would be a good
3099 // place to note that the call does not throw though.
3100 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
3101 II->removeFromParent(); // Take out of symbol table
3103 // Insert the call now...
3104 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
3105 Builder.SetInsertPoint(BI);
3106 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
3107 Args, II->getName());
3108 CI->setCallingConv(II->getCallingConv());
3109 CI->setAttributes(II->getAttributes());
3110 // If the invoke produced a value, the call does now instead.
3111 II->replaceAllUsesWith(CI);
3118 // If this block is now dead, remove it.
3119 if (pred_begin(BB) == pred_end(BB) &&
3120 BB != &BB->getParent()->getEntryBlock()) {
3121 // We know there are no successors, so just nuke the block.
3122 BB->eraseFromParent();
3129 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
3130 /// integer range comparison into a sub, an icmp and a branch.
3131 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3132 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3134 // Make sure all cases point to the same destination and gather the values.
3135 SmallVector<ConstantInt *, 16> Cases;
3136 SwitchInst::CaseIt I = SI->case_begin();
3137 Cases.push_back(I.getCaseValue());
3138 SwitchInst::CaseIt PrevI = I++;
3139 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
3140 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
3142 Cases.push_back(I.getCaseValue());
3144 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
3146 // Sort the case values, then check if they form a range we can transform.
3147 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3148 for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
3149 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
3153 Constant *Offset = ConstantExpr::getNeg(Cases.back());
3154 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
3156 Value *Sub = SI->getCondition();
3157 if (!Offset->isNullValue())
3158 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
3160 // If NumCases overflowed, then all possible values jump to the successor.
3161 if (NumCases->isNullValue() && SI->getNumCases() != 0)
3162 Cmp = ConstantInt::getTrue(SI->getContext());
3164 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3165 BranchInst *NewBI = Builder.CreateCondBr(
3166 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
3168 // Update weight for the newly-created conditional branch.
3169 SmallVector<uint64_t, 8> Weights;
3170 bool HasWeights = HasBranchWeights(SI);
3172 GetBranchWeights(SI, Weights);
3173 if (Weights.size() == 1 + SI->getNumCases()) {
3174 // Combine all weights for the cases to be the true weight of NewBI.
3175 // We assume that the sum of all weights for a Terminator can fit into 32
3177 uint32_t NewTrueWeight = 0;
3178 for (unsigned I = 1, E = Weights.size(); I != E; ++I)
3179 NewTrueWeight += (uint32_t)Weights[I];
3180 NewBI->setMetadata(LLVMContext::MD_prof,
3181 MDBuilder(SI->getContext()).
3182 createBranchWeights(NewTrueWeight,
3183 (uint32_t)Weights[0]));
3187 // Prune obsolete incoming values off the successor's PHI nodes.
3188 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
3189 isa<PHINode>(BBI); ++BBI) {
3190 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
3191 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3193 SI->eraseFromParent();
3198 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
3199 /// and use it to remove dead cases.
3200 static bool EliminateDeadSwitchCases(SwitchInst *SI, const DataLayout *DL,
3201 AssumptionTracker *AT) {
3202 Value *Cond = SI->getCondition();
3203 unsigned Bits = Cond->getType()->getIntegerBitWidth();
3204 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3205 computeKnownBits(Cond, KnownZero, KnownOne, DL, 0, AT, SI);
3207 // Gather dead cases.
3208 SmallVector<ConstantInt*, 8> DeadCases;
3209 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3210 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3211 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3212 DeadCases.push_back(I.getCaseValue());
3213 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3214 << I.getCaseValue() << "' is dead.\n");
3218 SmallVector<uint64_t, 8> Weights;
3219 bool HasWeight = HasBranchWeights(SI);
3221 GetBranchWeights(SI, Weights);
3222 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3225 // Remove dead cases from the switch.
3226 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3227 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3228 assert(Case != SI->case_default() &&
3229 "Case was not found. Probably mistake in DeadCases forming.");
3231 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3235 // Prune unused values from PHI nodes.
3236 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3237 SI->removeCase(Case);
3239 if (HasWeight && Weights.size() >= 2) {
3240 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3241 SI->setMetadata(LLVMContext::MD_prof,
3242 MDBuilder(SI->getParent()->getContext()).
3243 createBranchWeights(MDWeights));
3246 return !DeadCases.empty();
3249 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
3250 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3251 /// by an unconditional branch), look at the phi node for BB in the successor
3252 /// block and see if the incoming value is equal to CaseValue. If so, return
3253 /// the phi node, and set PhiIndex to BB's index in the phi node.
3254 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3257 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3258 return nullptr; // BB must be empty to be a candidate for simplification.
3259 if (!BB->getSinglePredecessor())
3260 return nullptr; // BB must be dominated by the switch.
3262 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3263 if (!Branch || !Branch->isUnconditional())
3264 return nullptr; // Terminator must be unconditional branch.
3266 BasicBlock *Succ = Branch->getSuccessor(0);
3268 BasicBlock::iterator I = Succ->begin();
3269 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3270 int Idx = PHI->getBasicBlockIndex(BB);
3271 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3273 Value *InValue = PHI->getIncomingValue(Idx);
3274 if (InValue != CaseValue) continue;
3283 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
3284 /// instruction to a phi node dominated by the switch, if that would mean that
3285 /// some of the destination blocks of the switch can be folded away.
3286 /// Returns true if a change is made.
3287 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3288 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3289 ForwardingNodesMap ForwardingNodes;
3291 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3292 ConstantInt *CaseValue = I.getCaseValue();
3293 BasicBlock *CaseDest = I.getCaseSuccessor();
3296 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3300 ForwardingNodes[PHI].push_back(PhiIndex);
3303 bool Changed = false;
3305 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3306 E = ForwardingNodes.end(); I != E; ++I) {
3307 PHINode *Phi = I->first;
3308 SmallVectorImpl<int> &Indexes = I->second;
3310 if (Indexes.size() < 2) continue;
3312 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3313 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3320 /// ValidLookupTableConstant - Return true if the backend will be able to handle
3321 /// initializing an array of constants like C.
3322 static bool ValidLookupTableConstant(Constant *C) {
3323 if (C->isThreadDependent())
3325 if (C->isDLLImportDependent())
3328 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3329 return CE->isGEPWithNoNotionalOverIndexing();
3331 return isa<ConstantFP>(C) ||
3332 isa<ConstantInt>(C) ||
3333 isa<ConstantPointerNull>(C) ||
3334 isa<GlobalValue>(C) ||
3338 /// LookupConstant - If V is a Constant, return it. Otherwise, try to look up
3339 /// its constant value in ConstantPool, returning 0 if it's not there.
3340 static Constant *LookupConstant(Value *V,
3341 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3342 if (Constant *C = dyn_cast<Constant>(V))
3344 return ConstantPool.lookup(V);
3347 /// ConstantFold - Try to fold instruction I into a constant. This works for
3348 /// simple instructions such as binary operations where both operands are
3349 /// constant or can be replaced by constants from the ConstantPool. Returns the
3350 /// resulting constant on success, 0 otherwise.
3352 ConstantFold(Instruction *I,
3353 const SmallDenseMap<Value *, Constant *> &ConstantPool,
3354 const DataLayout *DL) {
3355 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3356 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3359 if (A->isAllOnesValue())
3360 return LookupConstant(Select->getTrueValue(), ConstantPool);
3361 if (A->isNullValue())
3362 return LookupConstant(Select->getFalseValue(), ConstantPool);
3366 SmallVector<Constant *, 4> COps;
3367 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
3368 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
3374 if (CmpInst *Cmp = dyn_cast<CmpInst>(I))
3375 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
3378 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), COps, DL);
3381 /// GetCaseResults - Try to determine the resulting constant values in phi nodes
3382 /// at the common destination basic block, *CommonDest, for one of the case
3383 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3384 /// case), of a switch instruction SI.
3386 GetCaseResults(SwitchInst *SI,
3387 ConstantInt *CaseVal,
3388 BasicBlock *CaseDest,
3389 BasicBlock **CommonDest,
3390 SmallVectorImpl<std::pair<PHINode *, Constant *> > &Res,
3391 const DataLayout *DL) {
3392 // The block from which we enter the common destination.
3393 BasicBlock *Pred = SI->getParent();
3395 // If CaseDest is empty except for some side-effect free instructions through
3396 // which we can constant-propagate the CaseVal, continue to its successor.
3397 SmallDenseMap<Value*, Constant*> ConstantPool;
3398 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3399 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3401 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3402 // If the terminator is a simple branch, continue to the next block.
3403 if (T->getNumSuccessors() != 1)
3406 CaseDest = T->getSuccessor(0);
3407 } else if (isa<DbgInfoIntrinsic>(I)) {
3408 // Skip debug intrinsic.
3410 } else if (Constant *C = ConstantFold(I, ConstantPool, DL)) {
3411 // Instruction is side-effect free and constant.
3412 ConstantPool.insert(std::make_pair(I, C));
3418 // If we did not have a CommonDest before, use the current one.
3420 *CommonDest = CaseDest;
3421 // If the destination isn't the common one, abort.
3422 if (CaseDest != *CommonDest)
3425 // Get the values for this case from phi nodes in the destination block.
3426 BasicBlock::iterator I = (*CommonDest)->begin();
3427 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3428 int Idx = PHI->getBasicBlockIndex(Pred);
3432 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3437 // Note: If the constant comes from constant-propagating the case value
3438 // through the CaseDest basic block, it will be safe to remove the
3439 // instructions in that block. They cannot be used (except in the phi nodes
3440 // we visit) outside CaseDest, because that block does not dominate its
3441 // successor. If it did, we would not be in this phi node.
3443 // Be conservative about which kinds of constants we support.
3444 if (!ValidLookupTableConstant(ConstVal))
3447 Res.push_back(std::make_pair(PHI, ConstVal));
3450 return Res.size() > 0;
3453 // MapCaseToResult - Helper function used to
3454 // add CaseVal to the list of cases that generate Result.
3455 static void MapCaseToResult(ConstantInt *CaseVal,
3456 SwitchCaseResultVectorTy &UniqueResults,
3458 for (auto &I : UniqueResults) {
3459 if (I.first == Result) {
3460 I.second.push_back(CaseVal);
3464 UniqueResults.push_back(std::make_pair(Result,
3465 SmallVector<ConstantInt*, 4>(1, CaseVal)));
3468 // InitializeUniqueCases - Helper function that initializes a map containing
3469 // results for the PHI node of the common destination block for a switch
3470 // instruction. Returns false if multiple PHI nodes have been found or if
3471 // there is not a common destination block for the switch.
3472 static bool InitializeUniqueCases(
3473 SwitchInst *SI, const DataLayout *DL, PHINode *&PHI,
3474 BasicBlock *&CommonDest,
3475 SwitchCaseResultVectorTy &UniqueResults,
3476 Constant *&DefaultResult) {
3477 for (auto &I : SI->cases()) {
3478 ConstantInt *CaseVal = I.getCaseValue();
3480 // Resulting value at phi nodes for this case value.
3481 SwitchCaseResultsTy Results;
3482 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
3486 // Only one value per case is permitted
3487 if (Results.size() > 1)
3489 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
3491 // Check the PHI consistency.
3493 PHI = Results[0].first;
3494 else if (PHI != Results[0].first)
3497 // Find the default result value.
3498 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
3499 BasicBlock *DefaultDest = SI->getDefaultDest();
3500 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
3502 // If the default value is not found abort unless the default destination
3505 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
3506 if ((!DefaultResult &&
3507 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
3513 // ConvertTwoCaseSwitch - Helper function that checks if it is possible to
3514 // transform a switch with only two cases (or two cases + default)
3515 // that produces a result into a value select.
3518 // case 10: %0 = icmp eq i32 %a, 10
3519 // return 10; %1 = select i1 %0, i32 10, i32 4
3520 // case 20: ----> %2 = icmp eq i32 %a, 20
3521 // return 2; %3 = select i1 %2, i32 2, i32 %1
3526 ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
3527 Constant *DefaultResult, Value *Condition,
3528 IRBuilder<> &Builder) {
3529 assert(ResultVector.size() == 2 &&
3530 "We should have exactly two unique results at this point");
3531 // If we are selecting between only two cases transform into a simple
3532 // select or a two-way select if default is possible.
3533 if (ResultVector[0].second.size() == 1 &&
3534 ResultVector[1].second.size() == 1) {
3535 ConstantInt *const FirstCase = ResultVector[0].second[0];
3536 ConstantInt *const SecondCase = ResultVector[1].second[0];
3538 bool DefaultCanTrigger = DefaultResult;
3539 Value *SelectValue = ResultVector[1].first;
3540 if (DefaultCanTrigger) {
3541 Value *const ValueCompare =
3542 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
3543 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
3544 DefaultResult, "switch.select");
3546 Value *const ValueCompare =
3547 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
3548 return Builder.CreateSelect(ValueCompare, ResultVector[0].first, SelectValue,
3555 // RemoveSwitchAfterSelectConversion - Helper function to cleanup a switch
3556 // instruction that has been converted into a select, fixing up PHI nodes and
3558 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
3560 IRBuilder<> &Builder) {
3561 BasicBlock *SelectBB = SI->getParent();
3562 if (PHI->getBasicBlockIndex(SelectBB) >= 0)
3563 PHI->removeIncomingValue(SelectBB);
3564 PHI->addIncoming(SelectValue, SelectBB);
3566 Builder.CreateBr(PHI->getParent());
3568 // Remove the switch.
3569 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
3570 BasicBlock *Succ = SI->getSuccessor(i);
3572 if (Succ == PHI->getParent())
3574 Succ->removePredecessor(SelectBB);
3576 SI->eraseFromParent();
3579 /// SwitchToSelect - If the switch is only used to initialize one or more
3580 /// phi nodes in a common successor block with only two different
3581 /// constant values, replace the switch with select.
3582 static bool SwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
3583 const DataLayout *DL, AssumptionTracker *AT) {
3584 Value *const Cond = SI->getCondition();
3585 PHINode *PHI = nullptr;
3586 BasicBlock *CommonDest = nullptr;
3587 Constant *DefaultResult;
3588 SwitchCaseResultVectorTy UniqueResults;
3589 // Collect all the cases that will deliver the same value from the switch.
3590 if (!InitializeUniqueCases(SI, DL, PHI, CommonDest, UniqueResults,
3593 // Selects choose between maximum two values.
3594 if (UniqueResults.size() != 2)
3596 assert(PHI != nullptr && "PHI for value select not found");
3598 Builder.SetInsertPoint(SI);
3599 Value *SelectValue = ConvertTwoCaseSwitch(
3601 DefaultResult, Cond, Builder);
3603 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
3606 // The switch couldn't be converted into a select.
3611 /// SwitchLookupTable - This class represents a lookup table that can be used
3612 /// to replace a switch.
3613 class SwitchLookupTable {
3615 /// SwitchLookupTable - Create a lookup table to use as a switch replacement
3616 /// with the contents of Values, using DefaultValue to fill any holes in the
3618 SwitchLookupTable(Module &M,
3620 ConstantInt *Offset,
3621 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3622 Constant *DefaultValue,
3623 const DataLayout *DL);
3625 /// BuildLookup - Build instructions with Builder to retrieve the value at
3626 /// the position given by Index in the lookup table.
3627 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3629 /// WouldFitInRegister - Return true if a table with TableSize elements of
3630 /// type ElementType would fit in a target-legal register.
3631 static bool WouldFitInRegister(const DataLayout *DL,
3633 const Type *ElementType);
3636 // Depending on the contents of the table, it can be represented in
3639 // For tables where each element contains the same value, we just have to
3640 // store that single value and return it for each lookup.
3643 // For small tables with integer elements, we can pack them into a bitmap
3644 // that fits into a target-legal register. Values are retrieved by
3645 // shift and mask operations.
3648 // The table is stored as an array of values. Values are retrieved by load
3649 // instructions from the table.
3653 // For SingleValueKind, this is the single value.
3654 Constant *SingleValue;
3656 // For BitMapKind, this is the bitmap.
3657 ConstantInt *BitMap;
3658 IntegerType *BitMapElementTy;
3660 // For ArrayKind, this is the array.
3661 GlobalVariable *Array;
3665 SwitchLookupTable::SwitchLookupTable(Module &M,
3667 ConstantInt *Offset,
3668 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3669 Constant *DefaultValue,
3670 const DataLayout *DL)
3671 : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
3673 assert(Values.size() && "Can't build lookup table without values!");
3674 assert(TableSize >= Values.size() && "Can't fit values in table!");
3676 // If all values in the table are equal, this is that value.
3677 SingleValue = Values.begin()->second;
3679 Type *ValueType = Values.begin()->second->getType();
3681 // Build up the table contents.
3682 SmallVector<Constant*, 64> TableContents(TableSize);
3683 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3684 ConstantInt *CaseVal = Values[I].first;
3685 Constant *CaseRes = Values[I].second;
3686 assert(CaseRes->getType() == ValueType);
3688 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3690 TableContents[Idx] = CaseRes;
3692 if (CaseRes != SingleValue)
3693 SingleValue = nullptr;
3696 // Fill in any holes in the table with the default result.
3697 if (Values.size() < TableSize) {
3698 assert(DefaultValue &&
3699 "Need a default value to fill the lookup table holes.");
3700 assert(DefaultValue->getType() == ValueType);
3701 for (uint64_t I = 0; I < TableSize; ++I) {
3702 if (!TableContents[I])
3703 TableContents[I] = DefaultValue;
3706 if (DefaultValue != SingleValue)
3707 SingleValue = nullptr;
3710 // If each element in the table contains the same value, we only need to store
3711 // that single value.
3713 Kind = SingleValueKind;
3717 // If the type is integer and the table fits in a register, build a bitmap.
3718 if (WouldFitInRegister(DL, TableSize, ValueType)) {
3719 IntegerType *IT = cast<IntegerType>(ValueType);
3720 APInt TableInt(TableSize * IT->getBitWidth(), 0);
3721 for (uint64_t I = TableSize; I > 0; --I) {
3722 TableInt <<= IT->getBitWidth();
3723 // Insert values into the bitmap. Undef values are set to zero.
3724 if (!isa<UndefValue>(TableContents[I - 1])) {
3725 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
3726 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
3729 BitMap = ConstantInt::get(M.getContext(), TableInt);
3730 BitMapElementTy = IT;
3736 // Store the table in an array.
3737 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
3738 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3740 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3741 GlobalVariable::PrivateLinkage,
3744 Array->setUnnamedAddr(true);
3748 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
3750 case SingleValueKind:
3753 // Type of the bitmap (e.g. i59).
3754 IntegerType *MapTy = BitMap->getType();
3756 // Cast Index to the same type as the bitmap.
3757 // Note: The Index is <= the number of elements in the table, so
3758 // truncating it to the width of the bitmask is safe.
3759 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
3761 // Multiply the shift amount by the element width.
3762 ShiftAmt = Builder.CreateMul(ShiftAmt,
3763 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
3767 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
3768 "switch.downshift");
3770 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
3774 // Make sure the table index will not overflow when treated as signed.
3775 IntegerType *IT = cast<IntegerType>(Index->getType());
3776 uint64_t TableSize = Array->getInitializer()->getType()
3777 ->getArrayNumElements();
3778 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
3779 Index = Builder.CreateZExt(Index,
3780 IntegerType::get(IT->getContext(),
3781 IT->getBitWidth() + 1),
3782 "switch.tableidx.zext");
3784 Value *GEPIndices[] = { Builder.getInt32(0), Index };
3785 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices,
3787 return Builder.CreateLoad(GEP, "switch.load");
3790 llvm_unreachable("Unknown lookup table kind!");
3793 bool SwitchLookupTable::WouldFitInRegister(const DataLayout *DL,
3795 const Type *ElementType) {
3798 const IntegerType *IT = dyn_cast<IntegerType>(ElementType);
3801 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
3802 // are <= 15, we could try to narrow the type.
3804 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
3805 if (TableSize >= UINT_MAX/IT->getBitWidth())
3807 return DL->fitsInLegalInteger(TableSize * IT->getBitWidth());
3810 /// ShouldBuildLookupTable - Determine whether a lookup table should be built
3811 /// for this switch, based on the number of cases, size of the table and the
3812 /// types of the results.
3813 static bool ShouldBuildLookupTable(SwitchInst *SI,
3815 const TargetTransformInfo &TTI,
3816 const DataLayout *DL,
3817 const SmallDenseMap<PHINode*, Type*>& ResultTypes) {
3818 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
3819 return false; // TableSize overflowed, or mul below might overflow.
3821 bool AllTablesFitInRegister = true;
3822 bool HasIllegalType = false;
3823 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(),
3824 E = ResultTypes.end(); I != E; ++I) {
3825 Type *Ty = I->second;
3827 // Saturate this flag to true.
3828 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
3830 // Saturate this flag to false.
3831 AllTablesFitInRegister = AllTablesFitInRegister &&
3832 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
3834 // If both flags saturate, we're done. NOTE: This *only* works with
3835 // saturating flags, and all flags have to saturate first due to the
3836 // non-deterministic behavior of iterating over a dense map.
3837 if (HasIllegalType && !AllTablesFitInRegister)
3841 // If each table would fit in a register, we should build it anyway.
3842 if (AllTablesFitInRegister)
3845 // Don't build a table that doesn't fit in-register if it has illegal types.
3849 // The table density should be at least 40%. This is the same criterion as for
3850 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3851 // FIXME: Find the best cut-off.
3852 return SI->getNumCases() * 10 >= TableSize * 4;
3855 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3856 /// phi nodes in a common successor block with different constant values,
3857 /// replace the switch with lookup tables.
3858 static bool SwitchToLookupTable(SwitchInst *SI,
3859 IRBuilder<> &Builder,
3860 const TargetTransformInfo &TTI,
3861 const DataLayout* DL) {
3862 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3864 // Only build lookup table when we have a target that supports it.
3865 if (!TTI.shouldBuildLookupTables())
3868 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3869 // split off a dense part and build a lookup table for that.
3871 // FIXME: This creates arrays of GEPs to constant strings, which means each
3872 // GEP needs a runtime relocation in PIC code. We should just build one big
3873 // string and lookup indices into that.
3875 // Ignore switches with less than three cases. Lookup tables will not make them
3876 // faster, so we don't analyze them.
3877 if (SI->getNumCases() < 3)
3880 // Figure out the corresponding result for each case value and phi node in the
3881 // common destination, as well as the the min and max case values.
3882 assert(SI->case_begin() != SI->case_end());
3883 SwitchInst::CaseIt CI = SI->case_begin();
3884 ConstantInt *MinCaseVal = CI.getCaseValue();
3885 ConstantInt *MaxCaseVal = CI.getCaseValue();
3887 BasicBlock *CommonDest = nullptr;
3888 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
3889 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
3890 SmallDenseMap<PHINode*, Constant*> DefaultResults;
3891 SmallDenseMap<PHINode*, Type*> ResultTypes;
3892 SmallVector<PHINode*, 4> PHIs;
3894 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
3895 ConstantInt *CaseVal = CI.getCaseValue();
3896 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
3897 MinCaseVal = CaseVal;
3898 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
3899 MaxCaseVal = CaseVal;
3901 // Resulting value at phi nodes for this case value.
3902 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
3904 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
3908 // Append the result from this case to the list for each phi.
3909 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) {
3910 if (!ResultLists.count(I->first))
3911 PHIs.push_back(I->first);
3912 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second));
3916 // Keep track of the result types.
3917 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
3918 PHINode *PHI = PHIs[I];
3919 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
3922 uint64_t NumResults = ResultLists[PHIs[0]].size();
3923 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
3924 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
3925 bool TableHasHoles = (NumResults < TableSize);
3927 // If the table has holes, we need a constant result for the default case
3928 // or a bitmask that fits in a register.
3929 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
3930 bool HasDefaultResults = false;
3931 if (TableHasHoles) {
3932 HasDefaultResults = GetCaseResults(SI, nullptr, SI->getDefaultDest(),
3933 &CommonDest, DefaultResultsList, DL);
3935 bool NeedMask = (TableHasHoles && !HasDefaultResults);
3937 // As an extra penalty for the validity test we require more cases.
3938 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
3940 if (!(DL && DL->fitsInLegalInteger(TableSize)))
3944 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) {
3945 PHINode *PHI = DefaultResultsList[I].first;
3946 Constant *Result = DefaultResultsList[I].second;
3947 DefaultResults[PHI] = Result;
3950 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
3953 // Create the BB that does the lookups.
3954 Module &Mod = *CommonDest->getParent()->getParent();
3955 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
3957 CommonDest->getParent(),
3960 // Compute the table index value.
3961 Builder.SetInsertPoint(SI);
3962 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
3965 // Compute the maximum table size representable by the integer type we are
3967 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
3968 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
3969 assert(MaxTableSize >= TableSize &&
3970 "It is impossible for a switch to have more entries than the max "
3971 "representable value of its input integer type's size.");
3973 // If we have a fully covered lookup table, unconditionally branch to the
3974 // lookup table BB. Otherwise, check if the condition value is within the case
3975 // range. If it is so, branch to the new BB. Otherwise branch to SI's default
3977 const bool GeneratingCoveredLookupTable = MaxTableSize == TableSize;
3978 if (GeneratingCoveredLookupTable) {
3979 Builder.CreateBr(LookupBB);
3980 // We cached PHINodes in PHIs, to avoid accessing deleted PHINodes later,
3981 // do not delete PHINodes here.
3982 SI->getDefaultDest()->removePredecessor(SI->getParent(),
3983 true/*DontDeleteUselessPHIs*/);
3985 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
3986 MinCaseVal->getType(), TableSize));
3987 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
3990 // Populate the BB that does the lookups.
3991 Builder.SetInsertPoint(LookupBB);
3994 // Before doing the lookup we do the hole check.
3995 // The LookupBB is therefore re-purposed to do the hole check
3996 // and we create a new LookupBB.
3997 BasicBlock *MaskBB = LookupBB;
3998 MaskBB->setName("switch.hole_check");
3999 LookupBB = BasicBlock::Create(Mod.getContext(),
4001 CommonDest->getParent(),
4004 // Build bitmask; fill in a 1 bit for every case.
4005 APInt MaskInt(TableSize, 0);
4006 APInt One(TableSize, 1);
4007 const ResultListTy &ResultList = ResultLists[PHIs[0]];
4008 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
4009 uint64_t Idx = (ResultList[I].first->getValue() -
4010 MinCaseVal->getValue()).getLimitedValue();
4011 MaskInt |= One << Idx;
4013 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
4015 // Get the TableIndex'th bit of the bitmask.
4016 // If this bit is 0 (meaning hole) jump to the default destination,
4017 // else continue with table lookup.
4018 IntegerType *MapTy = TableMask->getType();
4019 Value *MaskIndex = Builder.CreateZExtOrTrunc(TableIndex, MapTy,
4020 "switch.maskindex");
4021 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex,
4023 Value *LoBit = Builder.CreateTrunc(Shifted,
4024 Type::getInt1Ty(Mod.getContext()),
4026 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
4028 Builder.SetInsertPoint(LookupBB);
4029 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
4032 bool ReturnedEarly = false;
4033 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
4034 PHINode *PHI = PHIs[I];
4036 // If using a bitmask, use any value to fill the lookup table holes.
4037 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
4038 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI],
4041 Value *Result = Table.BuildLookup(TableIndex, Builder);
4043 // If the result is used to return immediately from the function, we want to
4044 // do that right here.
4045 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
4046 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
4047 Builder.CreateRet(Result);
4048 ReturnedEarly = true;
4052 PHI->addIncoming(Result, LookupBB);
4056 Builder.CreateBr(CommonDest);
4058 // Remove the switch.
4059 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4060 BasicBlock *Succ = SI->getSuccessor(i);
4062 if (Succ == SI->getDefaultDest())
4064 Succ->removePredecessor(SI->getParent());
4066 SI->eraseFromParent();
4070 ++NumLookupTablesHoles;
4074 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
4075 BasicBlock *BB = SI->getParent();
4077 if (isValueEqualityComparison(SI)) {
4078 // If we only have one predecessor, and if it is a branch on this value,
4079 // see if that predecessor totally determines the outcome of this switch.
4080 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4081 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
4082 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4084 Value *Cond = SI->getCondition();
4085 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
4086 if (SimplifySwitchOnSelect(SI, Select))
4087 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4089 // If the block only contains the switch, see if we can fold the block
4090 // away into any preds.
4091 BasicBlock::iterator BBI = BB->begin();
4092 // Ignore dbg intrinsics.
4093 while (isa<DbgInfoIntrinsic>(BBI))
4096 if (FoldValueComparisonIntoPredecessors(SI, Builder))
4097 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4100 // Try to transform the switch into an icmp and a branch.
4101 if (TurnSwitchRangeIntoICmp(SI, Builder))
4102 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4104 // Remove unreachable cases.
4105 if (EliminateDeadSwitchCases(SI, DL, AT))
4106 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4108 if (SwitchToSelect(SI, Builder, DL, AT))
4109 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4111 if (ForwardSwitchConditionToPHI(SI))
4112 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4114 if (SwitchToLookupTable(SI, Builder, TTI, DL))
4115 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4120 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
4121 BasicBlock *BB = IBI->getParent();
4122 bool Changed = false;
4124 // Eliminate redundant destinations.
4125 SmallPtrSet<Value *, 8> Succs;
4126 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
4127 BasicBlock *Dest = IBI->getDestination(i);
4128 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) {
4129 Dest->removePredecessor(BB);
4130 IBI->removeDestination(i);
4136 if (IBI->getNumDestinations() == 0) {
4137 // If the indirectbr has no successors, change it to unreachable.
4138 new UnreachableInst(IBI->getContext(), IBI);
4139 EraseTerminatorInstAndDCECond(IBI);
4143 if (IBI->getNumDestinations() == 1) {
4144 // If the indirectbr has one successor, change it to a direct branch.
4145 BranchInst::Create(IBI->getDestination(0), IBI);
4146 EraseTerminatorInstAndDCECond(IBI);
4150 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
4151 if (SimplifyIndirectBrOnSelect(IBI, SI))
4152 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4157 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
4158 BasicBlock *BB = BI->getParent();
4160 if (SinkCommon && SinkThenElseCodeToEnd(BI))
4163 // If the Terminator is the only non-phi instruction, simplify the block.
4164 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg();
4165 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
4166 TryToSimplifyUncondBranchFromEmptyBlock(BB))
4169 // If the only instruction in the block is a seteq/setne comparison
4170 // against a constant, try to simplify the block.
4171 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
4172 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
4173 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
4175 if (I->isTerminator() &&
4176 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI,
4177 BonusInstThreshold, DL, AT))
4181 // If this basic block is ONLY a compare and a branch, and if a predecessor
4182 // branches to us and our successor, fold the comparison into the
4183 // predecessor and use logical operations to update the incoming value
4184 // for PHI nodes in common successor.
4185 if (FoldBranchToCommonDest(BI, DL, BonusInstThreshold))
4186 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4191 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
4192 BasicBlock *BB = BI->getParent();
4194 // Conditional branch
4195 if (isValueEqualityComparison(BI)) {
4196 // If we only have one predecessor, and if it is a branch on this value,
4197 // see if that predecessor totally determines the outcome of this
4199 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4200 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
4201 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4203 // This block must be empty, except for the setcond inst, if it exists.
4204 // Ignore dbg intrinsics.
4205 BasicBlock::iterator I = BB->begin();
4206 // Ignore dbg intrinsics.
4207 while (isa<DbgInfoIntrinsic>(I))
4210 if (FoldValueComparisonIntoPredecessors(BI, Builder))
4211 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4212 } else if (&*I == cast<Instruction>(BI->getCondition())){
4214 // Ignore dbg intrinsics.
4215 while (isa<DbgInfoIntrinsic>(I))
4217 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
4218 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4222 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
4223 if (SimplifyBranchOnICmpChain(BI, DL, Builder))
4226 // If this basic block is ONLY a compare and a branch, and if a predecessor
4227 // branches to us and one of our successors, fold the comparison into the
4228 // predecessor and use logical operations to pick the right destination.
4229 if (FoldBranchToCommonDest(BI, DL, BonusInstThreshold))
4230 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4232 // We have a conditional branch to two blocks that are only reachable
4233 // from BI. We know that the condbr dominates the two blocks, so see if
4234 // there is any identical code in the "then" and "else" blocks. If so, we
4235 // can hoist it up to the branching block.
4236 if (BI->getSuccessor(0)->getSinglePredecessor()) {
4237 if (BI->getSuccessor(1)->getSinglePredecessor()) {
4238 if (HoistThenElseCodeToIf(BI, DL))
4239 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4241 // If Successor #1 has multiple preds, we may be able to conditionally
4242 // execute Successor #0 if it branches to Successor #1.
4243 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
4244 if (Succ0TI->getNumSuccessors() == 1 &&
4245 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
4246 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), DL))
4247 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4249 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
4250 // If Successor #0 has multiple preds, we may be able to conditionally
4251 // execute Successor #1 if it branches to Successor #0.
4252 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
4253 if (Succ1TI->getNumSuccessors() == 1 &&
4254 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
4255 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), DL))
4256 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4259 // If this is a branch on a phi node in the current block, thread control
4260 // through this block if any PHI node entries are constants.
4261 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
4262 if (PN->getParent() == BI->getParent())
4263 if (FoldCondBranchOnPHI(BI, DL))
4264 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4266 // Scan predecessor blocks for conditional branches.
4267 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
4268 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
4269 if (PBI != BI && PBI->isConditional())
4270 if (SimplifyCondBranchToCondBranch(PBI, BI))
4271 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4276 /// Check if passing a value to an instruction will cause undefined behavior.
4277 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
4278 Constant *C = dyn_cast<Constant>(V);
4285 if (C->isNullValue()) {
4286 // Only look at the first use, avoid hurting compile time with long uselists
4287 User *Use = *I->user_begin();
4289 // Now make sure that there are no instructions in between that can alter
4290 // control flow (eg. calls)
4291 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
4292 if (i == I->getParent()->end() || i->mayHaveSideEffects())
4295 // Look through GEPs. A load from a GEP derived from NULL is still undefined
4296 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
4297 if (GEP->getPointerOperand() == I)
4298 return passingValueIsAlwaysUndefined(V, GEP);
4300 // Look through bitcasts.
4301 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
4302 return passingValueIsAlwaysUndefined(V, BC);
4304 // Load from null is undefined.
4305 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
4306 if (!LI->isVolatile())
4307 return LI->getPointerAddressSpace() == 0;
4309 // Store to null is undefined.
4310 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
4311 if (!SI->isVolatile())
4312 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
4317 /// If BB has an incoming value that will always trigger undefined behavior
4318 /// (eg. null pointer dereference), remove the branch leading here.
4319 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
4320 for (BasicBlock::iterator i = BB->begin();
4321 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
4322 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
4323 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
4324 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
4325 IRBuilder<> Builder(T);
4326 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
4327 BB->removePredecessor(PHI->getIncomingBlock(i));
4328 // Turn uncoditional branches into unreachables and remove the dead
4329 // destination from conditional branches.
4330 if (BI->isUnconditional())
4331 Builder.CreateUnreachable();
4333 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
4334 BI->getSuccessor(0));
4335 BI->eraseFromParent();
4338 // TODO: SwitchInst.
4344 bool SimplifyCFGOpt::run(BasicBlock *BB) {
4345 bool Changed = false;
4347 assert(BB && BB->getParent() && "Block not embedded in function!");
4348 assert(BB->getTerminator() && "Degenerate basic block encountered!");
4350 // Remove basic blocks that have no predecessors (except the entry block)...
4351 // or that just have themself as a predecessor. These are unreachable.
4352 if ((pred_begin(BB) == pred_end(BB) &&
4353 BB != &BB->getParent()->getEntryBlock()) ||
4354 BB->getSinglePredecessor() == BB) {
4355 DEBUG(dbgs() << "Removing BB: \n" << *BB);
4356 DeleteDeadBlock(BB);
4360 // Check to see if we can constant propagate this terminator instruction
4362 Changed |= ConstantFoldTerminator(BB, true);
4364 // Check for and eliminate duplicate PHI nodes in this block.
4365 Changed |= EliminateDuplicatePHINodes(BB);
4367 // Check for and remove branches that will always cause undefined behavior.
4368 Changed |= removeUndefIntroducingPredecessor(BB);
4370 // Merge basic blocks into their predecessor if there is only one distinct
4371 // pred, and if there is only one distinct successor of the predecessor, and
4372 // if there are no PHI nodes.
4374 if (MergeBlockIntoPredecessor(BB))
4377 IRBuilder<> Builder(BB);
4379 // If there is a trivial two-entry PHI node in this basic block, and we can
4380 // eliminate it, do so now.
4381 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
4382 if (PN->getNumIncomingValues() == 2)
4383 Changed |= FoldTwoEntryPHINode(PN, DL);
4385 Builder.SetInsertPoint(BB->getTerminator());
4386 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
4387 if (BI->isUnconditional()) {
4388 if (SimplifyUncondBranch(BI, Builder)) return true;
4390 if (SimplifyCondBranch(BI, Builder)) return true;
4392 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
4393 if (SimplifyReturn(RI, Builder)) return true;
4394 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
4395 if (SimplifyResume(RI, Builder)) return true;
4396 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
4397 if (SimplifySwitch(SI, Builder)) return true;
4398 } else if (UnreachableInst *UI =
4399 dyn_cast<UnreachableInst>(BB->getTerminator())) {
4400 if (SimplifyUnreachable(UI)) return true;
4401 } else if (IndirectBrInst *IBI =
4402 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
4403 if (SimplifyIndirectBr(IBI)) return true;
4409 /// SimplifyCFG - This function is used to do simplification of a CFG. For
4410 /// example, it adjusts branches to branches to eliminate the extra hop, it
4411 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4412 /// of the CFG. It returns true if a modification was made.
4414 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
4415 unsigned BonusInstThreshold,
4416 const DataLayout *DL, AssumptionTracker *AT) {
4417 return SimplifyCFGOpt(TTI, BonusInstThreshold, DL, AT).run(BB);