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 // Chosen as 2 so as to be cheap, but still to have enough power to fold
57 // a select, so the "clamp" idiom (of a min followed by a max) will be caught.
58 // To catch this, we need to fold a compare and a select, hence '2' being the
59 // minimum reasonable default.
60 static cl::opt<unsigned>
61 PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(2),
62 cl::desc("Control the amount of phi node folding to perform (default = 2)"));
65 DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false),
66 cl::desc("Duplicate return instructions into unconditional branches"));
69 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
70 cl::desc("Sink common instructions down to the end block"));
72 static cl::opt<bool> HoistCondStores(
73 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
74 cl::desc("Hoist conditional stores if an unconditional store precedes"));
76 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
77 STATISTIC(NumLinearMaps, "Number of switch instructions turned into linear mapping");
78 STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables");
79 STATISTIC(NumLookupTablesHoles, "Number of switch instructions turned into lookup tables (holes checked)");
80 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
81 STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block");
82 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
85 // The first field contains the value that the switch produces when a certain
86 // case group is selected, and the second field is a vector containing the
87 // cases composing the case group.
88 typedef SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>
89 SwitchCaseResultVectorTy;
90 // The first field contains the phi node that generates a result of the switch
91 // and the second field contains the value generated for a certain case in the
92 // switch for that PHI.
93 typedef SmallVector<std::pair<PHINode *, Constant *>, 4> SwitchCaseResultsTy;
95 /// ValueEqualityComparisonCase - Represents a case of a switch.
96 struct ValueEqualityComparisonCase {
100 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
101 : Value(Value), Dest(Dest) {}
103 bool operator<(ValueEqualityComparisonCase RHS) const {
104 // Comparing pointers is ok as we only rely on the order for uniquing.
105 return Value < RHS.Value;
108 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
111 class SimplifyCFGOpt {
112 const TargetTransformInfo &TTI;
113 const DataLayout &DL;
114 unsigned BonusInstThreshold;
116 Value *isValueEqualityComparison(TerminatorInst *TI);
117 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
118 std::vector<ValueEqualityComparisonCase> &Cases);
119 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
121 IRBuilder<> &Builder);
122 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
123 IRBuilder<> &Builder);
125 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
126 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
127 bool SimplifyCleanupReturn(CleanupReturnInst *RI);
128 bool SimplifyUnreachable(UnreachableInst *UI);
129 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
130 bool SimplifyIndirectBr(IndirectBrInst *IBI);
131 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder);
132 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder);
135 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
136 unsigned BonusInstThreshold, AssumptionCache *AC)
137 : TTI(TTI), DL(DL), BonusInstThreshold(BonusInstThreshold), AC(AC) {}
138 bool run(BasicBlock *BB);
142 /// Return true if it is safe to merge these two
143 /// terminator instructions together.
144 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
145 if (SI1 == SI2) return false; // Can't merge with self!
147 // It is not safe to merge these two switch instructions if they have a common
148 // successor, and if that successor has a PHI node, and if *that* PHI node has
149 // conflicting incoming values from the two switch blocks.
150 BasicBlock *SI1BB = SI1->getParent();
151 BasicBlock *SI2BB = SI2->getParent();
152 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
154 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
155 if (SI1Succs.count(*I))
156 for (BasicBlock::iterator BBI = (*I)->begin();
157 isa<PHINode>(BBI); ++BBI) {
158 PHINode *PN = cast<PHINode>(BBI);
159 if (PN->getIncomingValueForBlock(SI1BB) !=
160 PN->getIncomingValueForBlock(SI2BB))
167 /// Return true if it is safe and profitable to merge these two terminator
168 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
169 /// store all PHI nodes in common successors.
170 static bool isProfitableToFoldUnconditional(BranchInst *SI1,
173 SmallVectorImpl<PHINode*> &PhiNodes) {
174 if (SI1 == SI2) return false; // Can't merge with self!
175 assert(SI1->isUnconditional() && SI2->isConditional());
177 // We fold the unconditional branch if we can easily update all PHI nodes in
178 // common successors:
179 // 1> We have a constant incoming value for the conditional branch;
180 // 2> We have "Cond" as the incoming value for the unconditional branch;
181 // 3> SI2->getCondition() and Cond have same operands.
182 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
183 if (!Ci2) return false;
184 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
185 Cond->getOperand(1) == Ci2->getOperand(1)) &&
186 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
187 Cond->getOperand(1) == Ci2->getOperand(0)))
190 BasicBlock *SI1BB = SI1->getParent();
191 BasicBlock *SI2BB = SI2->getParent();
192 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
193 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
194 if (SI1Succs.count(*I))
195 for (BasicBlock::iterator BBI = (*I)->begin();
196 isa<PHINode>(BBI); ++BBI) {
197 PHINode *PN = cast<PHINode>(BBI);
198 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
199 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
201 PhiNodes.push_back(PN);
206 /// Update PHI nodes in Succ to indicate that there will now be entries in it
207 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
208 /// will be the same as those coming in from ExistPred, an existing predecessor
210 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
211 BasicBlock *ExistPred) {
212 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
215 for (BasicBlock::iterator I = Succ->begin();
216 (PN = dyn_cast<PHINode>(I)); ++I)
217 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
220 /// Compute an abstract "cost" of speculating the given instruction,
221 /// which is assumed to be safe to speculate. TCC_Free means cheap,
222 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
224 static unsigned ComputeSpeculationCost(const User *I,
225 const TargetTransformInfo &TTI) {
226 assert(isSafeToSpeculativelyExecute(I) &&
227 "Instruction is not safe to speculatively execute!");
228 return TTI.getUserCost(I);
231 /// If we have a merge point of an "if condition" as accepted above,
232 /// return true if the specified value dominates the block. We
233 /// don't handle the true generality of domination here, just a special case
234 /// which works well enough for us.
236 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
237 /// see if V (which must be an instruction) and its recursive operands
238 /// that do not dominate BB have a combined cost lower than CostRemaining and
239 /// are non-trapping. If both are true, the instruction is inserted into the
240 /// set and true is returned.
242 /// The cost for most non-trapping instructions is defined as 1 except for
243 /// Select whose cost is 2.
245 /// After this function returns, CostRemaining is decreased by the cost of
246 /// V plus its non-dominating operands. If that cost is greater than
247 /// CostRemaining, false is returned and CostRemaining is undefined.
248 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
249 SmallPtrSetImpl<Instruction*> *AggressiveInsts,
250 unsigned &CostRemaining,
251 const TargetTransformInfo &TTI) {
252 Instruction *I = dyn_cast<Instruction>(V);
254 // Non-instructions all dominate instructions, but not all constantexprs
255 // can be executed unconditionally.
256 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
261 BasicBlock *PBB = I->getParent();
263 // We don't want to allow weird loops that might have the "if condition" in
264 // the bottom of this block.
265 if (PBB == BB) return false;
267 // If this instruction is defined in a block that contains an unconditional
268 // branch to BB, then it must be in the 'conditional' part of the "if
269 // statement". If not, it definitely dominates the region.
270 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
271 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
274 // If we aren't allowing aggressive promotion anymore, then don't consider
275 // instructions in the 'if region'.
276 if (!AggressiveInsts) return false;
278 // If we have seen this instruction before, don't count it again.
279 if (AggressiveInsts->count(I)) return true;
281 // Okay, it looks like the instruction IS in the "condition". Check to
282 // see if it's a cheap instruction to unconditionally compute, and if it
283 // only uses stuff defined outside of the condition. If so, hoist it out.
284 if (!isSafeToSpeculativelyExecute(I))
287 unsigned Cost = ComputeSpeculationCost(I, TTI);
289 if (Cost > CostRemaining)
292 CostRemaining -= Cost;
294 // Okay, we can only really hoist these out if their operands do
295 // not take us over the cost threshold.
296 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
297 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, TTI))
299 // Okay, it's safe to do this! Remember this instruction.
300 AggressiveInsts->insert(I);
304 /// Extract ConstantInt from value, looking through IntToPtr
305 /// and PointerNullValue. Return NULL if value is not a constant int.
306 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
307 // Normal constant int.
308 ConstantInt *CI = dyn_cast<ConstantInt>(V);
309 if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
312 // This is some kind of pointer constant. Turn it into a pointer-sized
313 // ConstantInt if possible.
314 IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
316 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
317 if (isa<ConstantPointerNull>(V))
318 return ConstantInt::get(PtrTy, 0);
320 // IntToPtr const int.
321 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
322 if (CE->getOpcode() == Instruction::IntToPtr)
323 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
324 // The constant is very likely to have the right type already.
325 if (CI->getType() == PtrTy)
328 return cast<ConstantInt>
329 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
336 /// Given a chain of or (||) or and (&&) comparison of a value against a
337 /// constant, this will try to recover the information required for a switch
339 /// It will depth-first traverse the chain of comparison, seeking for patterns
340 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
341 /// representing the different cases for the switch.
342 /// Note that if the chain is composed of '||' it will build the set of elements
343 /// that matches the comparisons (i.e. any of this value validate the chain)
344 /// while for a chain of '&&' it will build the set elements that make the test
346 struct ConstantComparesGatherer {
347 const DataLayout &DL;
348 Value *CompValue; /// Value found for the switch comparison
349 Value *Extra; /// Extra clause to be checked before the switch
350 SmallVector<ConstantInt *, 8> Vals; /// Set of integers to match in switch
351 unsigned UsedICmps; /// Number of comparisons matched in the and/or chain
353 /// Construct and compute the result for the comparison instruction Cond
354 ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL)
355 : DL(DL), CompValue(nullptr), Extra(nullptr), UsedICmps(0) {
360 ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
361 ConstantComparesGatherer &
362 operator=(const ConstantComparesGatherer &) = delete;
366 /// Try to set the current value used for the comparison, it succeeds only if
367 /// it wasn't set before or if the new value is the same as the old one
368 bool setValueOnce(Value *NewVal) {
369 if(CompValue && CompValue != NewVal) return false;
371 return (CompValue != nullptr);
374 /// Try to match Instruction "I" as a comparison against a constant and
375 /// populates the array Vals with the set of values that match (or do not
376 /// match depending on isEQ).
377 /// Return false on failure. On success, the Value the comparison matched
378 /// against is placed in CompValue.
379 /// If CompValue is already set, the function is expected to fail if a match
380 /// is found but the value compared to is different.
381 bool matchInstruction(Instruction *I, bool isEQ) {
382 // If this is an icmp against a constant, handle this as one of the cases.
385 if (!((ICI = dyn_cast<ICmpInst>(I)) &&
386 (C = GetConstantInt(I->getOperand(1), DL)))) {
393 // Pattern match a special case
394 // (x & ~2^x) == y --> x == y || x == y|2^x
395 // This undoes a transformation done by instcombine to fuse 2 compares.
396 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
397 if (match(ICI->getOperand(0),
398 m_And(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
399 APInt Not = ~RHSC->getValue();
400 if (Not.isPowerOf2()) {
401 // If we already have a value for the switch, it has to match!
402 if(!setValueOnce(RHSVal))
406 Vals.push_back(ConstantInt::get(C->getContext(),
407 C->getValue() | Not));
413 // If we already have a value for the switch, it has to match!
414 if(!setValueOnce(ICI->getOperand(0)))
419 return ICI->getOperand(0);
422 // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
423 ConstantRange Span = ConstantRange::makeAllowedICmpRegion(
424 ICI->getPredicate(), C->getValue());
426 // Shift the range if the compare is fed by an add. This is the range
427 // compare idiom as emitted by instcombine.
428 Value *CandidateVal = I->getOperand(0);
429 if(match(I->getOperand(0), m_Add(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
430 Span = Span.subtract(RHSC->getValue());
431 CandidateVal = RHSVal;
434 // If this is an and/!= check, then we are looking to build the set of
435 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
438 Span = Span.inverse();
440 // If there are a ton of values, we don't want to make a ginormous switch.
441 if (Span.getSetSize().ugt(8) || Span.isEmptySet()) {
445 // If we already have a value for the switch, it has to match!
446 if(!setValueOnce(CandidateVal))
449 // Add all values from the range to the set
450 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
451 Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
458 /// Given a potentially 'or'd or 'and'd together collection of icmp
459 /// eq/ne/lt/gt instructions that compare a value against a constant, extract
460 /// the value being compared, and stick the list constants into the Vals
462 /// One "Extra" case is allowed to differ from the other.
463 void gather(Value *V) {
464 Instruction *I = dyn_cast<Instruction>(V);
465 bool isEQ = (I->getOpcode() == Instruction::Or);
467 // Keep a stack (SmallVector for efficiency) for depth-first traversal
468 SmallVector<Value *, 8> DFT;
473 while(!DFT.empty()) {
474 V = DFT.pop_back_val();
476 if (Instruction *I = dyn_cast<Instruction>(V)) {
477 // If it is a || (or && depending on isEQ), process the operands.
478 if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
479 DFT.push_back(I->getOperand(1));
480 DFT.push_back(I->getOperand(0));
484 // Try to match the current instruction
485 if (matchInstruction(I, isEQ))
486 // Match succeed, continue the loop
490 // One element of the sequence of || (or &&) could not be match as a
491 // comparison against the same value as the others.
492 // We allow only one "Extra" case to be checked before the switch
497 // Failed to parse a proper sequence, abort now
506 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
507 Instruction *Cond = nullptr;
508 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
509 Cond = dyn_cast<Instruction>(SI->getCondition());
510 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
511 if (BI->isConditional())
512 Cond = dyn_cast<Instruction>(BI->getCondition());
513 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
514 Cond = dyn_cast<Instruction>(IBI->getAddress());
517 TI->eraseFromParent();
518 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
521 /// Return true if the specified terminator checks
522 /// to see if a value is equal to constant integer value.
523 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
525 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
526 // Do not permit merging of large switch instructions into their
527 // predecessors unless there is only one predecessor.
528 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()),
529 pred_end(SI->getParent())) <= 128)
530 CV = SI->getCondition();
531 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
532 if (BI->isConditional() && BI->getCondition()->hasOneUse())
533 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
534 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
535 CV = ICI->getOperand(0);
538 // Unwrap any lossless ptrtoint cast.
540 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
541 Value *Ptr = PTII->getPointerOperand();
542 if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
549 /// Given a value comparison instruction,
550 /// decode all of the 'cases' that it represents and return the 'default' block.
551 BasicBlock *SimplifyCFGOpt::
552 GetValueEqualityComparisonCases(TerminatorInst *TI,
553 std::vector<ValueEqualityComparisonCase>
555 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
556 Cases.reserve(SI->getNumCases());
557 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
558 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
559 i.getCaseSuccessor()));
560 return SI->getDefaultDest();
563 BranchInst *BI = cast<BranchInst>(TI);
564 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
565 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
566 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
569 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
573 /// Given a vector of bb/value pairs, remove any entries
574 /// in the list that match the specified block.
575 static void EliminateBlockCases(BasicBlock *BB,
576 std::vector<ValueEqualityComparisonCase> &Cases) {
577 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
580 /// Return true if there are any keys in C1 that exist in C2 as well.
582 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
583 std::vector<ValueEqualityComparisonCase > &C2) {
584 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
586 // Make V1 be smaller than V2.
587 if (V1->size() > V2->size())
590 if (V1->size() == 0) return false;
591 if (V1->size() == 1) {
593 ConstantInt *TheVal = (*V1)[0].Value;
594 for (unsigned i = 0, e = V2->size(); i != e; ++i)
595 if (TheVal == (*V2)[i].Value)
599 // Otherwise, just sort both lists and compare element by element.
600 array_pod_sort(V1->begin(), V1->end());
601 array_pod_sort(V2->begin(), V2->end());
602 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
603 while (i1 != e1 && i2 != e2) {
604 if ((*V1)[i1].Value == (*V2)[i2].Value)
606 if ((*V1)[i1].Value < (*V2)[i2].Value)
614 /// If TI is known to be a terminator instruction and its block is known to
615 /// only have a single predecessor block, check to see if that predecessor is
616 /// also a value comparison with the same value, and if that comparison
617 /// determines the outcome of this comparison. If so, simplify TI. This does a
618 /// very limited form of jump threading.
619 bool SimplifyCFGOpt::
620 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
622 IRBuilder<> &Builder) {
623 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
624 if (!PredVal) return false; // Not a value comparison in predecessor.
626 Value *ThisVal = isValueEqualityComparison(TI);
627 assert(ThisVal && "This isn't a value comparison!!");
628 if (ThisVal != PredVal) return false; // Different predicates.
630 // TODO: Preserve branch weight metadata, similarly to how
631 // FoldValueComparisonIntoPredecessors preserves it.
633 // Find out information about when control will move from Pred to TI's block.
634 std::vector<ValueEqualityComparisonCase> PredCases;
635 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
637 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
639 // Find information about how control leaves this block.
640 std::vector<ValueEqualityComparisonCase> ThisCases;
641 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
642 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
644 // If TI's block is the default block from Pred's comparison, potentially
645 // simplify TI based on this knowledge.
646 if (PredDef == TI->getParent()) {
647 // If we are here, we know that the value is none of those cases listed in
648 // PredCases. If there are any cases in ThisCases that are in PredCases, we
650 if (!ValuesOverlap(PredCases, ThisCases))
653 if (isa<BranchInst>(TI)) {
654 // Okay, one of the successors of this condbr is dead. Convert it to a
656 assert(ThisCases.size() == 1 && "Branch can only have one case!");
657 // Insert the new branch.
658 Instruction *NI = Builder.CreateBr(ThisDef);
661 // Remove PHI node entries for the dead edge.
662 ThisCases[0].Dest->removePredecessor(TI->getParent());
664 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
665 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
667 EraseTerminatorInstAndDCECond(TI);
671 SwitchInst *SI = cast<SwitchInst>(TI);
672 // Okay, TI has cases that are statically dead, prune them away.
673 SmallPtrSet<Constant*, 16> DeadCases;
674 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
675 DeadCases.insert(PredCases[i].Value);
677 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
678 << "Through successor TI: " << *TI);
680 // Collect branch weights into a vector.
681 SmallVector<uint32_t, 8> Weights;
682 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
683 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
685 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
687 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
688 Weights.push_back(CI->getValue().getZExtValue());
690 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
692 if (DeadCases.count(i.getCaseValue())) {
694 std::swap(Weights[i.getCaseIndex()+1], Weights.back());
697 i.getCaseSuccessor()->removePredecessor(TI->getParent());
701 if (HasWeight && Weights.size() >= 2)
702 SI->setMetadata(LLVMContext::MD_prof,
703 MDBuilder(SI->getParent()->getContext()).
704 createBranchWeights(Weights));
706 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
710 // Otherwise, TI's block must correspond to some matched value. Find out
711 // which value (or set of values) this is.
712 ConstantInt *TIV = nullptr;
713 BasicBlock *TIBB = TI->getParent();
714 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
715 if (PredCases[i].Dest == TIBB) {
717 return false; // Cannot handle multiple values coming to this block.
718 TIV = PredCases[i].Value;
720 assert(TIV && "No edge from pred to succ?");
722 // Okay, we found the one constant that our value can be if we get into TI's
723 // BB. Find out which successor will unconditionally be branched to.
724 BasicBlock *TheRealDest = nullptr;
725 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
726 if (ThisCases[i].Value == TIV) {
727 TheRealDest = ThisCases[i].Dest;
731 // If not handled by any explicit cases, it is handled by the default case.
732 if (!TheRealDest) TheRealDest = ThisDef;
734 // Remove PHI node entries for dead edges.
735 BasicBlock *CheckEdge = TheRealDest;
736 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
737 if (*SI != CheckEdge)
738 (*SI)->removePredecessor(TIBB);
742 // Insert the new branch.
743 Instruction *NI = Builder.CreateBr(TheRealDest);
746 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
747 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
749 EraseTerminatorInstAndDCECond(TI);
754 /// This class implements a stable ordering of constant
755 /// integers that does not depend on their address. This is important for
756 /// applications that sort ConstantInt's to ensure uniqueness.
757 struct ConstantIntOrdering {
758 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
759 return LHS->getValue().ult(RHS->getValue());
764 static int ConstantIntSortPredicate(ConstantInt *const *P1,
765 ConstantInt *const *P2) {
766 const ConstantInt *LHS = *P1;
767 const ConstantInt *RHS = *P2;
768 if (LHS->getValue().ult(RHS->getValue()))
770 if (LHS->getValue() == RHS->getValue())
775 static inline bool HasBranchWeights(const Instruction* I) {
776 MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
777 if (ProfMD && ProfMD->getOperand(0))
778 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
779 return MDS->getString().equals("branch_weights");
784 /// Get Weights of a given TerminatorInst, the default weight is at the front
785 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
787 static void GetBranchWeights(TerminatorInst *TI,
788 SmallVectorImpl<uint64_t> &Weights) {
789 MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
791 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
792 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
793 Weights.push_back(CI->getValue().getZExtValue());
796 // If TI is a conditional eq, the default case is the false case,
797 // and the corresponding branch-weight data is at index 2. We swap the
798 // default weight to be the first entry.
799 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
800 assert(Weights.size() == 2);
801 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
802 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
803 std::swap(Weights.front(), Weights.back());
807 /// Keep halving the weights until all can fit in uint32_t.
808 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
809 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
810 if (Max > UINT_MAX) {
811 unsigned Offset = 32 - countLeadingZeros(Max);
812 for (uint64_t &I : Weights)
817 /// The specified terminator is a value equality comparison instruction
818 /// (either a switch or a branch on "X == c").
819 /// See if any of the predecessors of the terminator block are value comparisons
820 /// on the same value. If so, and if safe to do so, fold them together.
821 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
822 IRBuilder<> &Builder) {
823 BasicBlock *BB = TI->getParent();
824 Value *CV = isValueEqualityComparison(TI); // CondVal
825 assert(CV && "Not a comparison?");
826 bool Changed = false;
828 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
829 while (!Preds.empty()) {
830 BasicBlock *Pred = Preds.pop_back_val();
832 // See if the predecessor is a comparison with the same value.
833 TerminatorInst *PTI = Pred->getTerminator();
834 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
836 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
837 // Figure out which 'cases' to copy from SI to PSI.
838 std::vector<ValueEqualityComparisonCase> BBCases;
839 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
841 std::vector<ValueEqualityComparisonCase> PredCases;
842 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
844 // Based on whether the default edge from PTI goes to BB or not, fill in
845 // PredCases and PredDefault with the new switch cases we would like to
847 SmallVector<BasicBlock*, 8> NewSuccessors;
849 // Update the branch weight metadata along the way
850 SmallVector<uint64_t, 8> Weights;
851 bool PredHasWeights = HasBranchWeights(PTI);
852 bool SuccHasWeights = HasBranchWeights(TI);
854 if (PredHasWeights) {
855 GetBranchWeights(PTI, Weights);
856 // branch-weight metadata is inconsistent here.
857 if (Weights.size() != 1 + PredCases.size())
858 PredHasWeights = SuccHasWeights = false;
859 } else if (SuccHasWeights)
860 // If there are no predecessor weights but there are successor weights,
861 // populate Weights with 1, which will later be scaled to the sum of
862 // successor's weights
863 Weights.assign(1 + PredCases.size(), 1);
865 SmallVector<uint64_t, 8> SuccWeights;
866 if (SuccHasWeights) {
867 GetBranchWeights(TI, SuccWeights);
868 // branch-weight metadata is inconsistent here.
869 if (SuccWeights.size() != 1 + BBCases.size())
870 PredHasWeights = SuccHasWeights = false;
871 } else if (PredHasWeights)
872 SuccWeights.assign(1 + BBCases.size(), 1);
874 if (PredDefault == BB) {
875 // If this is the default destination from PTI, only the edges in TI
876 // that don't occur in PTI, or that branch to BB will be activated.
877 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
878 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
879 if (PredCases[i].Dest != BB)
880 PTIHandled.insert(PredCases[i].Value);
882 // The default destination is BB, we don't need explicit targets.
883 std::swap(PredCases[i], PredCases.back());
885 if (PredHasWeights || SuccHasWeights) {
886 // Increase weight for the default case.
887 Weights[0] += Weights[i+1];
888 std::swap(Weights[i+1], Weights.back());
892 PredCases.pop_back();
896 // Reconstruct the new switch statement we will be building.
897 if (PredDefault != BBDefault) {
898 PredDefault->removePredecessor(Pred);
899 PredDefault = BBDefault;
900 NewSuccessors.push_back(BBDefault);
903 unsigned CasesFromPred = Weights.size();
904 uint64_t ValidTotalSuccWeight = 0;
905 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
906 if (!PTIHandled.count(BBCases[i].Value) &&
907 BBCases[i].Dest != BBDefault) {
908 PredCases.push_back(BBCases[i]);
909 NewSuccessors.push_back(BBCases[i].Dest);
910 if (SuccHasWeights || PredHasWeights) {
911 // The default weight is at index 0, so weight for the ith case
912 // should be at index i+1. Scale the cases from successor by
913 // PredDefaultWeight (Weights[0]).
914 Weights.push_back(Weights[0] * SuccWeights[i+1]);
915 ValidTotalSuccWeight += SuccWeights[i+1];
919 if (SuccHasWeights || PredHasWeights) {
920 ValidTotalSuccWeight += SuccWeights[0];
921 // Scale the cases from predecessor by ValidTotalSuccWeight.
922 for (unsigned i = 1; i < CasesFromPred; ++i)
923 Weights[i] *= ValidTotalSuccWeight;
924 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
925 Weights[0] *= SuccWeights[0];
928 // If this is not the default destination from PSI, only the edges
929 // in SI that occur in PSI with a destination of BB will be
931 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
932 std::map<ConstantInt*, uint64_t> WeightsForHandled;
933 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
934 if (PredCases[i].Dest == BB) {
935 PTIHandled.insert(PredCases[i].Value);
937 if (PredHasWeights || SuccHasWeights) {
938 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
939 std::swap(Weights[i+1], Weights.back());
943 std::swap(PredCases[i], PredCases.back());
944 PredCases.pop_back();
948 // Okay, now we know which constants were sent to BB from the
949 // predecessor. Figure out where they will all go now.
950 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
951 if (PTIHandled.count(BBCases[i].Value)) {
952 // If this is one we are capable of getting...
953 if (PredHasWeights || SuccHasWeights)
954 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
955 PredCases.push_back(BBCases[i]);
956 NewSuccessors.push_back(BBCases[i].Dest);
957 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
960 // If there are any constants vectored to BB that TI doesn't handle,
961 // they must go to the default destination of TI.
962 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
964 E = PTIHandled.end(); I != E; ++I) {
965 if (PredHasWeights || SuccHasWeights)
966 Weights.push_back(WeightsForHandled[*I]);
967 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
968 NewSuccessors.push_back(BBDefault);
972 // Okay, at this point, we know which new successor Pred will get. Make
973 // sure we update the number of entries in the PHI nodes for these
975 for (BasicBlock *NewSuccessor : NewSuccessors)
976 AddPredecessorToBlock(NewSuccessor, Pred, BB);
978 Builder.SetInsertPoint(PTI);
979 // Convert pointer to int before we switch.
980 if (CV->getType()->isPointerTy()) {
981 CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
985 // Now that the successors are updated, create the new Switch instruction.
986 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
988 NewSI->setDebugLoc(PTI->getDebugLoc());
989 for (ValueEqualityComparisonCase &V : PredCases)
990 NewSI->addCase(V.Value, V.Dest);
992 if (PredHasWeights || SuccHasWeights) {
993 // Halve the weights if any of them cannot fit in an uint32_t
996 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
998 NewSI->setMetadata(LLVMContext::MD_prof,
999 MDBuilder(BB->getContext()).
1000 createBranchWeights(MDWeights));
1003 EraseTerminatorInstAndDCECond(PTI);
1005 // Okay, last check. If BB is still a successor of PSI, then we must
1006 // have an infinite loop case. If so, add an infinitely looping block
1007 // to handle the case to preserve the behavior of the code.
1008 BasicBlock *InfLoopBlock = nullptr;
1009 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1010 if (NewSI->getSuccessor(i) == BB) {
1011 if (!InfLoopBlock) {
1012 // Insert it at the end of the function, because it's either code,
1013 // or it won't matter if it's hot. :)
1014 InfLoopBlock = BasicBlock::Create(BB->getContext(),
1015 "infloop", BB->getParent());
1016 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1018 NewSI->setSuccessor(i, InfLoopBlock);
1027 // If we would need to insert a select that uses the value of this invoke
1028 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1029 // can't hoist the invoke, as there is nowhere to put the select in this case.
1030 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1031 Instruction *I1, Instruction *I2) {
1032 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1034 for (BasicBlock::iterator BBI = SI->begin();
1035 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1036 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1037 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1038 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
1046 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I);
1048 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1049 /// in the two blocks up into the branch block. The caller of this function
1050 /// guarantees that BI's block dominates BB1 and BB2.
1051 static bool HoistThenElseCodeToIf(BranchInst *BI,
1052 const TargetTransformInfo &TTI) {
1053 // This does very trivial matching, with limited scanning, to find identical
1054 // instructions in the two blocks. In particular, we don't want to get into
1055 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1056 // such, we currently just scan for obviously identical instructions in an
1058 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1059 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1061 BasicBlock::iterator BB1_Itr = BB1->begin();
1062 BasicBlock::iterator BB2_Itr = BB2->begin();
1064 Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1065 // Skip debug info if it is not identical.
1066 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1067 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1068 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1069 while (isa<DbgInfoIntrinsic>(I1))
1071 while (isa<DbgInfoIntrinsic>(I2))
1074 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1075 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1078 BasicBlock *BIParent = BI->getParent();
1080 bool Changed = false;
1082 // If we are hoisting the terminator instruction, don't move one (making a
1083 // broken BB), instead clone it, and remove BI.
1084 if (isa<TerminatorInst>(I1))
1085 goto HoistTerminator;
1087 if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1090 // For a normal instruction, we just move one to right before the branch,
1091 // then replace all uses of the other with the first. Finally, we remove
1092 // the now redundant second instruction.
1093 BIParent->getInstList().splice(BI->getIterator(), BB1->getInstList(), I1);
1094 if (!I2->use_empty())
1095 I2->replaceAllUsesWith(I1);
1096 I1->intersectOptionalDataWith(I2);
1097 unsigned KnownIDs[] = {
1098 LLVMContext::MD_tbaa, LLVMContext::MD_range,
1099 LLVMContext::MD_fpmath, LLVMContext::MD_invariant_load,
1100 LLVMContext::MD_nonnull, LLVMContext::MD_invariant_group};
1101 combineMetadata(I1, I2, KnownIDs);
1102 I2->eraseFromParent();
1107 // Skip debug info if it is not identical.
1108 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1109 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1110 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1111 while (isa<DbgInfoIntrinsic>(I1))
1113 while (isa<DbgInfoIntrinsic>(I2))
1116 } while (I1->isIdenticalToWhenDefined(I2));
1121 // It may not be possible to hoist an invoke.
1122 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1125 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1127 for (BasicBlock::iterator BBI = SI->begin();
1128 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1129 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1130 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1134 // Check for passingValueIsAlwaysUndefined here because we would rather
1135 // eliminate undefined control flow then converting it to a select.
1136 if (passingValueIsAlwaysUndefined(BB1V, PN) ||
1137 passingValueIsAlwaysUndefined(BB2V, PN))
1140 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1142 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1147 // Okay, it is safe to hoist the terminator.
1148 Instruction *NT = I1->clone();
1149 BIParent->getInstList().insert(BI->getIterator(), NT);
1150 if (!NT->getType()->isVoidTy()) {
1151 I1->replaceAllUsesWith(NT);
1152 I2->replaceAllUsesWith(NT);
1156 IRBuilder<true, NoFolder> Builder(NT);
1157 // Hoisting one of the terminators from our successor is a great thing.
1158 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1159 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1160 // nodes, so we insert select instruction to compute the final result.
1161 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1162 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1164 for (BasicBlock::iterator BBI = SI->begin();
1165 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1166 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1167 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1168 if (BB1V == BB2V) continue;
1170 // These values do not agree. Insert a select instruction before NT
1171 // that determines the right value.
1172 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1174 SI = cast<SelectInst>
1175 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1176 BB1V->getName()+"."+BB2V->getName()));
1178 // Make the PHI node use the select for all incoming values for BB1/BB2
1179 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1180 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1181 PN->setIncomingValue(i, SI);
1185 // Update any PHI nodes in our new successors.
1186 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1187 AddPredecessorToBlock(*SI, BIParent, BB1);
1189 EraseTerminatorInstAndDCECond(BI);
1193 /// Given an unconditional branch that goes to BBEnd,
1194 /// check whether BBEnd has only two predecessors and the other predecessor
1195 /// ends with an unconditional branch. If it is true, sink any common code
1196 /// in the two predecessors to BBEnd.
1197 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1198 assert(BI1->isUnconditional());
1199 BasicBlock *BB1 = BI1->getParent();
1200 BasicBlock *BBEnd = BI1->getSuccessor(0);
1202 // Check that BBEnd has two predecessors and the other predecessor ends with
1203 // an unconditional branch.
1204 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1205 BasicBlock *Pred0 = *PI++;
1206 if (PI == PE) // Only one predecessor.
1208 BasicBlock *Pred1 = *PI++;
1209 if (PI != PE) // More than two predecessors.
1211 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1212 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1213 if (!BI2 || !BI2->isUnconditional())
1216 // Gather the PHI nodes in BBEnd.
1217 SmallDenseMap<std::pair<Value *, Value *>, PHINode *> JointValueMap;
1218 Instruction *FirstNonPhiInBBEnd = nullptr;
1219 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end(); I != E; ++I) {
1220 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1221 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1222 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1223 JointValueMap[std::make_pair(BB1V, BB2V)] = PN;
1225 FirstNonPhiInBBEnd = &*I;
1229 if (!FirstNonPhiInBBEnd)
1232 // This does very trivial matching, with limited scanning, to find identical
1233 // instructions in the two blocks. We scan backward for obviously identical
1234 // instructions in an identical order.
1235 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1236 RE1 = BB1->getInstList().rend(),
1237 RI2 = BB2->getInstList().rbegin(),
1238 RE2 = BB2->getInstList().rend();
1240 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1243 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1246 // Skip the unconditional branches.
1250 bool Changed = false;
1251 while (RI1 != RE1 && RI2 != RE2) {
1253 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1256 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1260 Instruction *I1 = &*RI1, *I2 = &*RI2;
1261 auto InstPair = std::make_pair(I1, I2);
1262 // I1 and I2 should have a single use in the same PHI node, and they
1263 // perform the same operation.
1264 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1265 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1266 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1267 I1->isEHPad() || I2->isEHPad() ||
1268 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1269 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1270 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1271 !I1->hasOneUse() || !I2->hasOneUse() ||
1272 !JointValueMap.count(InstPair))
1275 // Check whether we should swap the operands of ICmpInst.
1276 // TODO: Add support of communativity.
1277 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1278 bool SwapOpnds = false;
1279 if (ICmp1 && ICmp2 &&
1280 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1281 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1282 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1283 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1284 ICmp2->swapOperands();
1287 if (!I1->isSameOperationAs(I2)) {
1289 ICmp2->swapOperands();
1293 // The operands should be either the same or they need to be generated
1294 // with a PHI node after sinking. We only handle the case where there is
1295 // a single pair of different operands.
1296 Value *DifferentOp1 = nullptr, *DifferentOp2 = nullptr;
1297 unsigned Op1Idx = ~0U;
1298 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1299 if (I1->getOperand(I) == I2->getOperand(I))
1301 // Early exit if we have more-than one pair of different operands or if
1302 // we need a PHI node to replace a constant.
1303 if (Op1Idx != ~0U ||
1304 isa<Constant>(I1->getOperand(I)) ||
1305 isa<Constant>(I2->getOperand(I))) {
1306 // If we can't sink the instructions, undo the swapping.
1308 ICmp2->swapOperands();
1311 DifferentOp1 = I1->getOperand(I);
1313 DifferentOp2 = I2->getOperand(I);
1316 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n");
1317 DEBUG(dbgs() << " " << *I2 << "\n");
1319 // We insert the pair of different operands to JointValueMap and
1320 // remove (I1, I2) from JointValueMap.
1321 if (Op1Idx != ~0U) {
1322 auto &NewPN = JointValueMap[std::make_pair(DifferentOp1, DifferentOp2)];
1325 PHINode::Create(DifferentOp1->getType(), 2,
1326 DifferentOp1->getName() + ".sink", &BBEnd->front());
1327 NewPN->addIncoming(DifferentOp1, BB1);
1328 NewPN->addIncoming(DifferentOp2, BB2);
1329 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1331 // I1 should use NewPN instead of DifferentOp1.
1332 I1->setOperand(Op1Idx, NewPN);
1334 PHINode *OldPN = JointValueMap[InstPair];
1335 JointValueMap.erase(InstPair);
1337 // We need to update RE1 and RE2 if we are going to sink the first
1338 // instruction in the basic block down.
1339 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1340 // Sink the instruction.
1341 BBEnd->getInstList().splice(FirstNonPhiInBBEnd->getIterator(),
1342 BB1->getInstList(), I1);
1343 if (!OldPN->use_empty())
1344 OldPN->replaceAllUsesWith(I1);
1345 OldPN->eraseFromParent();
1347 if (!I2->use_empty())
1348 I2->replaceAllUsesWith(I1);
1349 I1->intersectOptionalDataWith(I2);
1350 // TODO: Use combineMetadata here to preserve what metadata we can
1351 // (analogous to the hoisting case above).
1352 I2->eraseFromParent();
1355 RE1 = BB1->getInstList().rend();
1357 RE2 = BB2->getInstList().rend();
1358 FirstNonPhiInBBEnd = &*I1;
1365 /// \brief Determine if we can hoist sink a sole store instruction out of a
1366 /// conditional block.
1368 /// We are looking for code like the following:
1370 /// store i32 %add, i32* %arrayidx2
1371 /// ... // No other stores or function calls (we could be calling a memory
1372 /// ... // function).
1373 /// %cmp = icmp ult %x, %y
1374 /// br i1 %cmp, label %EndBB, label %ThenBB
1376 /// store i32 %add5, i32* %arrayidx2
1380 /// We are going to transform this into:
1382 /// store i32 %add, i32* %arrayidx2
1384 /// %cmp = icmp ult %x, %y
1385 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1386 /// store i32 %add.add5, i32* %arrayidx2
1389 /// \return The pointer to the value of the previous store if the store can be
1390 /// hoisted into the predecessor block. 0 otherwise.
1391 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1392 BasicBlock *StoreBB, BasicBlock *EndBB) {
1393 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1397 // Volatile or atomic.
1398 if (!StoreToHoist->isSimple())
1401 Value *StorePtr = StoreToHoist->getPointerOperand();
1403 // Look for a store to the same pointer in BrBB.
1404 unsigned MaxNumInstToLookAt = 10;
1405 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(),
1406 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) {
1407 Instruction *CurI = &*RI;
1409 // Could be calling an instruction that effects memory like free().
1410 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI))
1413 StoreInst *SI = dyn_cast<StoreInst>(CurI);
1414 // Found the previous store make sure it stores to the same location.
1415 if (SI && SI->getPointerOperand() == StorePtr)
1416 // Found the previous store, return its value operand.
1417 return SI->getValueOperand();
1419 return nullptr; // Unknown store.
1425 /// \brief Speculate a conditional basic block flattening the CFG.
1427 /// Note that this is a very risky transform currently. Speculating
1428 /// instructions like this is most often not desirable. Instead, there is an MI
1429 /// pass which can do it with full awareness of the resource constraints.
1430 /// However, some cases are "obvious" and we should do directly. An example of
1431 /// this is speculating a single, reasonably cheap instruction.
1433 /// There is only one distinct advantage to flattening the CFG at the IR level:
1434 /// it makes very common but simplistic optimizations such as are common in
1435 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1436 /// modeling their effects with easier to reason about SSA value graphs.
1439 /// An illustration of this transform is turning this IR:
1442 /// %cmp = icmp ult %x, %y
1443 /// br i1 %cmp, label %EndBB, label %ThenBB
1445 /// %sub = sub %x, %y
1448 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1455 /// %cmp = icmp ult %x, %y
1456 /// %sub = sub %x, %y
1457 /// %cond = select i1 %cmp, 0, %sub
1461 /// \returns true if the conditional block is removed.
1462 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1463 const TargetTransformInfo &TTI) {
1464 // Be conservative for now. FP select instruction can often be expensive.
1465 Value *BrCond = BI->getCondition();
1466 if (isa<FCmpInst>(BrCond))
1469 BasicBlock *BB = BI->getParent();
1470 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1472 // If ThenBB is actually on the false edge of the conditional branch, remember
1473 // to swap the select operands later.
1474 bool Invert = false;
1475 if (ThenBB != BI->getSuccessor(0)) {
1476 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1479 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1481 // Keep a count of how many times instructions are used within CondBB when
1482 // they are candidates for sinking into CondBB. Specifically:
1483 // - They are defined in BB, and
1484 // - They have no side effects, and
1485 // - All of their uses are in CondBB.
1486 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1488 unsigned SpeculationCost = 0;
1489 Value *SpeculatedStoreValue = nullptr;
1490 StoreInst *SpeculatedStore = nullptr;
1491 for (BasicBlock::iterator BBI = ThenBB->begin(),
1492 BBE = std::prev(ThenBB->end());
1493 BBI != BBE; ++BBI) {
1494 Instruction *I = &*BBI;
1496 if (isa<DbgInfoIntrinsic>(I))
1499 // Only speculatively execute a single instruction (not counting the
1500 // terminator) for now.
1502 if (SpeculationCost > 1)
1505 // Don't hoist the instruction if it's unsafe or expensive.
1506 if (!isSafeToSpeculativelyExecute(I) &&
1507 !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1508 I, BB, ThenBB, EndBB))))
1510 if (!SpeculatedStoreValue &&
1511 ComputeSpeculationCost(I, TTI) >
1512 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
1515 // Store the store speculation candidate.
1516 if (SpeculatedStoreValue)
1517 SpeculatedStore = cast<StoreInst>(I);
1519 // Do not hoist the instruction if any of its operands are defined but not
1520 // used in BB. The transformation will prevent the operand from
1521 // being sunk into the use block.
1522 for (User::op_iterator i = I->op_begin(), e = I->op_end();
1524 Instruction *OpI = dyn_cast<Instruction>(*i);
1525 if (!OpI || OpI->getParent() != BB ||
1526 OpI->mayHaveSideEffects())
1527 continue; // Not a candidate for sinking.
1529 ++SinkCandidateUseCounts[OpI];
1533 // Consider any sink candidates which are only used in CondBB as costs for
1534 // speculation. Note, while we iterate over a DenseMap here, we are summing
1535 // and so iteration order isn't significant.
1536 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I =
1537 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end();
1539 if (I->first->getNumUses() == I->second) {
1541 if (SpeculationCost > 1)
1545 // Check that the PHI nodes can be converted to selects.
1546 bool HaveRewritablePHIs = false;
1547 for (BasicBlock::iterator I = EndBB->begin();
1548 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1549 Value *OrigV = PN->getIncomingValueForBlock(BB);
1550 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1552 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1553 // Skip PHIs which are trivial.
1557 // Don't convert to selects if we could remove undefined behavior instead.
1558 if (passingValueIsAlwaysUndefined(OrigV, PN) ||
1559 passingValueIsAlwaysUndefined(ThenV, PN))
1562 HaveRewritablePHIs = true;
1563 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1564 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1565 if (!OrigCE && !ThenCE)
1566 continue; // Known safe and cheap.
1568 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
1569 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
1571 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
1572 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
1573 unsigned MaxCost = 2 * PHINodeFoldingThreshold *
1574 TargetTransformInfo::TCC_Basic;
1575 if (OrigCost + ThenCost > MaxCost)
1578 // Account for the cost of an unfolded ConstantExpr which could end up
1579 // getting expanded into Instructions.
1580 // FIXME: This doesn't account for how many operations are combined in the
1581 // constant expression.
1583 if (SpeculationCost > 1)
1587 // If there are no PHIs to process, bail early. This helps ensure idempotence
1589 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
1592 // If we get here, we can hoist the instruction and if-convert.
1593 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1595 // Insert a select of the value of the speculated store.
1596 if (SpeculatedStoreValue) {
1597 IRBuilder<true, NoFolder> Builder(BI);
1598 Value *TrueV = SpeculatedStore->getValueOperand();
1599 Value *FalseV = SpeculatedStoreValue;
1601 std::swap(TrueV, FalseV);
1602 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() +
1603 "." + FalseV->getName());
1604 SpeculatedStore->setOperand(0, S);
1607 // Hoist the instructions.
1608 BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
1609 ThenBB->begin(), std::prev(ThenBB->end()));
1611 // Insert selects and rewrite the PHI operands.
1612 IRBuilder<true, NoFolder> Builder(BI);
1613 for (BasicBlock::iterator I = EndBB->begin();
1614 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1615 unsigned OrigI = PN->getBasicBlockIndex(BB);
1616 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1617 Value *OrigV = PN->getIncomingValue(OrigI);
1618 Value *ThenV = PN->getIncomingValue(ThenI);
1620 // Skip PHIs which are trivial.
1624 // Create a select whose true value is the speculatively executed value and
1625 // false value is the preexisting value. Swap them if the branch
1626 // destinations were inverted.
1627 Value *TrueV = ThenV, *FalseV = OrigV;
1629 std::swap(TrueV, FalseV);
1630 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV,
1631 TrueV->getName() + "." + FalseV->getName());
1632 PN->setIncomingValue(OrigI, V);
1633 PN->setIncomingValue(ThenI, V);
1640 /// \returns True if this block contains a CallInst with the NoDuplicate
1642 static bool HasNoDuplicateCall(const BasicBlock *BB) {
1643 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1644 const CallInst *CI = dyn_cast<CallInst>(I);
1647 if (CI->cannotDuplicate())
1653 /// Return true if we can thread a branch across this block.
1654 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1655 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1658 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1659 if (isa<DbgInfoIntrinsic>(BBI))
1661 if (Size > 10) return false; // Don't clone large BB's.
1664 // We can only support instructions that do not define values that are
1665 // live outside of the current basic block.
1666 for (User *U : BBI->users()) {
1667 Instruction *UI = cast<Instruction>(U);
1668 if (UI->getParent() != BB || isa<PHINode>(UI)) return false;
1671 // Looks ok, continue checking.
1677 /// If we have a conditional branch on a PHI node value that is defined in the
1678 /// same block as the branch and if any PHI entries are constants, thread edges
1679 /// corresponding to that entry to be branches to their ultimate destination.
1680 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL) {
1681 BasicBlock *BB = BI->getParent();
1682 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1683 // NOTE: we currently cannot transform this case if the PHI node is used
1684 // outside of the block.
1685 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1688 // Degenerate case of a single entry PHI.
1689 if (PN->getNumIncomingValues() == 1) {
1690 FoldSingleEntryPHINodes(PN->getParent());
1694 // Now we know that this block has multiple preds and two succs.
1695 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1697 if (HasNoDuplicateCall(BB)) return false;
1699 // Okay, this is a simple enough basic block. See if any phi values are
1701 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1702 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1703 if (!CB || !CB->getType()->isIntegerTy(1)) continue;
1705 // Okay, we now know that all edges from PredBB should be revectored to
1706 // branch to RealDest.
1707 BasicBlock *PredBB = PN->getIncomingBlock(i);
1708 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1710 if (RealDest == BB) continue; // Skip self loops.
1711 // Skip if the predecessor's terminator is an indirect branch.
1712 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1714 // The dest block might have PHI nodes, other predecessors and other
1715 // difficult cases. Instead of being smart about this, just insert a new
1716 // block that jumps to the destination block, effectively splitting
1717 // the edge we are about to create.
1718 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1719 RealDest->getName()+".critedge",
1720 RealDest->getParent(), RealDest);
1721 BranchInst::Create(RealDest, EdgeBB);
1723 // Update PHI nodes.
1724 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1726 // BB may have instructions that are being threaded over. Clone these
1727 // instructions into EdgeBB. We know that there will be no uses of the
1728 // cloned instructions outside of EdgeBB.
1729 BasicBlock::iterator InsertPt = EdgeBB->begin();
1730 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1731 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1732 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1733 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1736 // Clone the instruction.
1737 Instruction *N = BBI->clone();
1738 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1740 // Update operands due to translation.
1741 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1743 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1744 if (PI != TranslateMap.end())
1748 // Check for trivial simplification.
1749 if (Value *V = SimplifyInstruction(N, DL)) {
1750 TranslateMap[&*BBI] = V;
1751 delete N; // Instruction folded away, don't need actual inst
1753 // Insert the new instruction into its new home.
1754 EdgeBB->getInstList().insert(InsertPt, N);
1755 if (!BBI->use_empty())
1756 TranslateMap[&*BBI] = N;
1760 // Loop over all of the edges from PredBB to BB, changing them to branch
1761 // to EdgeBB instead.
1762 TerminatorInst *PredBBTI = PredBB->getTerminator();
1763 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1764 if (PredBBTI->getSuccessor(i) == BB) {
1765 BB->removePredecessor(PredBB);
1766 PredBBTI->setSuccessor(i, EdgeBB);
1769 // Recurse, simplifying any other constants.
1770 return FoldCondBranchOnPHI(BI, DL) | true;
1776 /// Given a BB that starts with the specified two-entry PHI node,
1777 /// see if we can eliminate it.
1778 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI,
1779 const DataLayout &DL) {
1780 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1781 // statement", which has a very simple dominance structure. Basically, we
1782 // are trying to find the condition that is being branched on, which
1783 // subsequently causes this merge to happen. We really want control
1784 // dependence information for this check, but simplifycfg can't keep it up
1785 // to date, and this catches most of the cases we care about anyway.
1786 BasicBlock *BB = PN->getParent();
1787 BasicBlock *IfTrue, *IfFalse;
1788 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1790 // Don't bother if the branch will be constant folded trivially.
1791 isa<ConstantInt>(IfCond))
1794 // Okay, we found that we can merge this two-entry phi node into a select.
1795 // Doing so would require us to fold *all* two entry phi nodes in this block.
1796 // At some point this becomes non-profitable (particularly if the target
1797 // doesn't support cmov's). Only do this transformation if there are two or
1798 // fewer PHI nodes in this block.
1799 unsigned NumPhis = 0;
1800 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1804 // Loop over the PHI's seeing if we can promote them all to select
1805 // instructions. While we are at it, keep track of the instructions
1806 // that need to be moved to the dominating block.
1807 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1808 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1809 MaxCostVal1 = PHINodeFoldingThreshold;
1810 MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
1811 MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
1813 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1814 PHINode *PN = cast<PHINode>(II++);
1815 if (Value *V = SimplifyInstruction(PN, DL)) {
1816 PN->replaceAllUsesWith(V);
1817 PN->eraseFromParent();
1821 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1822 MaxCostVal0, TTI) ||
1823 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1828 // If we folded the first phi, PN dangles at this point. Refresh it. If
1829 // we ran out of PHIs then we simplified them all.
1830 PN = dyn_cast<PHINode>(BB->begin());
1831 if (!PN) return true;
1833 // Don't fold i1 branches on PHIs which contain binary operators. These can
1834 // often be turned into switches and other things.
1835 if (PN->getType()->isIntegerTy(1) &&
1836 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1837 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1838 isa<BinaryOperator>(IfCond)))
1841 // If we all PHI nodes are promotable, check to make sure that all
1842 // instructions in the predecessor blocks can be promoted as well. If
1843 // not, we won't be able to get rid of the control flow, so it's not
1844 // worth promoting to select instructions.
1845 BasicBlock *DomBlock = nullptr;
1846 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1847 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1848 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1851 DomBlock = *pred_begin(IfBlock1);
1852 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1853 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
1854 // This is not an aggressive instruction that we can promote.
1855 // Because of this, we won't be able to get rid of the control
1856 // flow, so the xform is not worth it.
1861 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1864 DomBlock = *pred_begin(IfBlock2);
1865 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1866 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
1867 // This is not an aggressive instruction that we can promote.
1868 // Because of this, we won't be able to get rid of the control
1869 // flow, so the xform is not worth it.
1874 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1875 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1877 // If we can still promote the PHI nodes after this gauntlet of tests,
1878 // do all of the PHI's now.
1879 Instruction *InsertPt = DomBlock->getTerminator();
1880 IRBuilder<true, NoFolder> Builder(InsertPt);
1882 // Move all 'aggressive' instructions, which are defined in the
1883 // conditional parts of the if's up to the dominating block.
1885 DomBlock->getInstList().splice(InsertPt->getIterator(),
1886 IfBlock1->getInstList(), IfBlock1->begin(),
1887 IfBlock1->getTerminator()->getIterator());
1889 DomBlock->getInstList().splice(InsertPt->getIterator(),
1890 IfBlock2->getInstList(), IfBlock2->begin(),
1891 IfBlock2->getTerminator()->getIterator());
1893 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1894 // Change the PHI node into a select instruction.
1895 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1896 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1899 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1900 PN->replaceAllUsesWith(NV);
1902 PN->eraseFromParent();
1905 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1906 // has been flattened. Change DomBlock to jump directly to our new block to
1907 // avoid other simplifycfg's kicking in on the diamond.
1908 TerminatorInst *OldTI = DomBlock->getTerminator();
1909 Builder.SetInsertPoint(OldTI);
1910 Builder.CreateBr(BB);
1911 OldTI->eraseFromParent();
1915 /// If we found a conditional branch that goes to two returning blocks,
1916 /// try to merge them together into one return,
1917 /// introducing a select if the return values disagree.
1918 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1919 IRBuilder<> &Builder) {
1920 assert(BI->isConditional() && "Must be a conditional branch");
1921 BasicBlock *TrueSucc = BI->getSuccessor(0);
1922 BasicBlock *FalseSucc = BI->getSuccessor(1);
1923 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1924 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1926 // Check to ensure both blocks are empty (just a return) or optionally empty
1927 // with PHI nodes. If there are other instructions, merging would cause extra
1928 // computation on one path or the other.
1929 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1931 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1934 Builder.SetInsertPoint(BI);
1935 // Okay, we found a branch that is going to two return nodes. If
1936 // there is no return value for this function, just change the
1937 // branch into a return.
1938 if (FalseRet->getNumOperands() == 0) {
1939 TrueSucc->removePredecessor(BI->getParent());
1940 FalseSucc->removePredecessor(BI->getParent());
1941 Builder.CreateRetVoid();
1942 EraseTerminatorInstAndDCECond(BI);
1946 // Otherwise, figure out what the true and false return values are
1947 // so we can insert a new select instruction.
1948 Value *TrueValue = TrueRet->getReturnValue();
1949 Value *FalseValue = FalseRet->getReturnValue();
1951 // Unwrap any PHI nodes in the return blocks.
1952 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1953 if (TVPN->getParent() == TrueSucc)
1954 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1955 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1956 if (FVPN->getParent() == FalseSucc)
1957 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1959 // In order for this transformation to be safe, we must be able to
1960 // unconditionally execute both operands to the return. This is
1961 // normally the case, but we could have a potentially-trapping
1962 // constant expression that prevents this transformation from being
1964 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1967 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1971 // Okay, we collected all the mapped values and checked them for sanity, and
1972 // defined to really do this transformation. First, update the CFG.
1973 TrueSucc->removePredecessor(BI->getParent());
1974 FalseSucc->removePredecessor(BI->getParent());
1976 // Insert select instructions where needed.
1977 Value *BrCond = BI->getCondition();
1979 // Insert a select if the results differ.
1980 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1981 } else if (isa<UndefValue>(TrueValue)) {
1982 TrueValue = FalseValue;
1984 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1985 FalseValue, "retval");
1989 Value *RI = !TrueValue ?
1990 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
1994 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
1995 << "\n " << *BI << "NewRet = " << *RI
1996 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
1998 EraseTerminatorInstAndDCECond(BI);
2003 /// Given a conditional BranchInstruction, retrieve the probabilities of the
2004 /// branch taking each edge. Fills in the two APInt parameters and returns true,
2005 /// or returns false if no or invalid metadata was found.
2006 static bool ExtractBranchMetadata(BranchInst *BI,
2007 uint64_t &ProbTrue, uint64_t &ProbFalse) {
2008 assert(BI->isConditional() &&
2009 "Looking for probabilities on unconditional branch?");
2010 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
2011 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
2012 ConstantInt *CITrue =
2013 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
2014 ConstantInt *CIFalse =
2015 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
2016 if (!CITrue || !CIFalse) return false;
2017 ProbTrue = CITrue->getValue().getZExtValue();
2018 ProbFalse = CIFalse->getValue().getZExtValue();
2022 /// Return true if the given instruction is available
2023 /// in its predecessor block. If yes, the instruction will be removed.
2024 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
2025 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2027 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
2028 Instruction *PBI = &*I;
2029 // Check whether Inst and PBI generate the same value.
2030 if (Inst->isIdenticalTo(PBI)) {
2031 Inst->replaceAllUsesWith(PBI);
2032 Inst->eraseFromParent();
2039 /// If this basic block is simple enough, and if a predecessor branches to us
2040 /// and one of our successors, fold the block into the predecessor and use
2041 /// logical operations to pick the right destination.
2042 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2043 BasicBlock *BB = BI->getParent();
2045 Instruction *Cond = nullptr;
2046 if (BI->isConditional())
2047 Cond = dyn_cast<Instruction>(BI->getCondition());
2049 // For unconditional branch, check for a simple CFG pattern, where
2050 // BB has a single predecessor and BB's successor is also its predecessor's
2051 // successor. If such pattern exisits, check for CSE between BB and its
2053 if (BasicBlock *PB = BB->getSinglePredecessor())
2054 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2055 if (PBI->isConditional() &&
2056 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2057 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2058 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
2060 Instruction *Curr = &*I++;
2061 if (isa<CmpInst>(Curr)) {
2065 // Quit if we can't remove this instruction.
2066 if (!checkCSEInPredecessor(Curr, PB))
2075 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2076 Cond->getParent() != BB || !Cond->hasOneUse())
2079 // Make sure the instruction after the condition is the cond branch.
2080 BasicBlock::iterator CondIt = ++Cond->getIterator();
2082 // Ignore dbg intrinsics.
2083 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
2088 // Only allow this transformation if computing the condition doesn't involve
2089 // too many instructions and these involved instructions can be executed
2090 // unconditionally. We denote all involved instructions except the condition
2091 // as "bonus instructions", and only allow this transformation when the
2092 // number of the bonus instructions does not exceed a certain threshold.
2093 unsigned NumBonusInsts = 0;
2094 for (auto I = BB->begin(); Cond != I; ++I) {
2095 // Ignore dbg intrinsics.
2096 if (isa<DbgInfoIntrinsic>(I))
2098 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2100 // I has only one use and can be executed unconditionally.
2101 Instruction *User = dyn_cast<Instruction>(I->user_back());
2102 if (User == nullptr || User->getParent() != BB)
2104 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2105 // to use any other instruction, User must be an instruction between next(I)
2108 // Early exits once we reach the limit.
2109 if (NumBonusInsts > BonusInstThreshold)
2113 // Cond is known to be a compare or binary operator. Check to make sure that
2114 // neither operand is a potentially-trapping constant expression.
2115 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2118 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2122 // Finally, don't infinitely unroll conditional loops.
2123 BasicBlock *TrueDest = BI->getSuccessor(0);
2124 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2125 if (TrueDest == BB || FalseDest == BB)
2128 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2129 BasicBlock *PredBlock = *PI;
2130 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2132 // Check that we have two conditional branches. If there is a PHI node in
2133 // the common successor, verify that the same value flows in from both
2135 SmallVector<PHINode*, 4> PHIs;
2136 if (!PBI || PBI->isUnconditional() ||
2137 (BI->isConditional() &&
2138 !SafeToMergeTerminators(BI, PBI)) ||
2139 (!BI->isConditional() &&
2140 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2143 // Determine if the two branches share a common destination.
2144 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2145 bool InvertPredCond = false;
2147 if (BI->isConditional()) {
2148 if (PBI->getSuccessor(0) == TrueDest)
2149 Opc = Instruction::Or;
2150 else if (PBI->getSuccessor(1) == FalseDest)
2151 Opc = Instruction::And;
2152 else if (PBI->getSuccessor(0) == FalseDest)
2153 Opc = Instruction::And, InvertPredCond = true;
2154 else if (PBI->getSuccessor(1) == TrueDest)
2155 Opc = Instruction::Or, InvertPredCond = true;
2159 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2163 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2164 IRBuilder<> Builder(PBI);
2166 // If we need to invert the condition in the pred block to match, do so now.
2167 if (InvertPredCond) {
2168 Value *NewCond = PBI->getCondition();
2170 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2171 CmpInst *CI = cast<CmpInst>(NewCond);
2172 CI->setPredicate(CI->getInversePredicate());
2174 NewCond = Builder.CreateNot(NewCond,
2175 PBI->getCondition()->getName()+".not");
2178 PBI->setCondition(NewCond);
2179 PBI->swapSuccessors();
2182 // If we have bonus instructions, clone them into the predecessor block.
2183 // Note that there may be multiple predecessor blocks, so we cannot move
2184 // bonus instructions to a predecessor block.
2185 ValueToValueMapTy VMap; // maps original values to cloned values
2186 // We already make sure Cond is the last instruction before BI. Therefore,
2187 // all instructions before Cond other than DbgInfoIntrinsic are bonus
2189 for (auto BonusInst = BB->begin(); Cond != BonusInst; ++BonusInst) {
2190 if (isa<DbgInfoIntrinsic>(BonusInst))
2192 Instruction *NewBonusInst = BonusInst->clone();
2193 RemapInstruction(NewBonusInst, VMap,
2194 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2195 VMap[&*BonusInst] = NewBonusInst;
2197 // If we moved a load, we cannot any longer claim any knowledge about
2198 // its potential value. The previous information might have been valid
2199 // only given the branch precondition.
2200 // For an analogous reason, we must also drop all the metadata whose
2201 // semantics we don't understand.
2202 NewBonusInst->dropUnknownNonDebugMetadata();
2204 PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2205 NewBonusInst->takeName(&*BonusInst);
2206 BonusInst->setName(BonusInst->getName() + ".old");
2209 // Clone Cond into the predecessor basic block, and or/and the
2210 // two conditions together.
2211 Instruction *New = Cond->clone();
2212 RemapInstruction(New, VMap,
2213 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2214 PredBlock->getInstList().insert(PBI->getIterator(), New);
2215 New->takeName(Cond);
2216 Cond->setName(New->getName() + ".old");
2218 if (BI->isConditional()) {
2219 Instruction *NewCond =
2220 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2222 PBI->setCondition(NewCond);
2224 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2225 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2227 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2229 SmallVector<uint64_t, 8> NewWeights;
2231 if (PBI->getSuccessor(0) == BB) {
2232 if (PredHasWeights && SuccHasWeights) {
2233 // PBI: br i1 %x, BB, FalseDest
2234 // BI: br i1 %y, TrueDest, FalseDest
2235 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2236 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2237 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2238 // TrueWeight for PBI * FalseWeight for BI.
2239 // We assume that total weights of a BranchInst can fit into 32 bits.
2240 // Therefore, we will not have overflow using 64-bit arithmetic.
2241 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2242 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2244 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2245 PBI->setSuccessor(0, TrueDest);
2247 if (PBI->getSuccessor(1) == BB) {
2248 if (PredHasWeights && SuccHasWeights) {
2249 // PBI: br i1 %x, TrueDest, BB
2250 // BI: br i1 %y, TrueDest, FalseDest
2251 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2252 // FalseWeight for PBI * TrueWeight for BI.
2253 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2254 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2255 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2256 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2258 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2259 PBI->setSuccessor(1, FalseDest);
2261 if (NewWeights.size() == 2) {
2262 // Halve the weights if any of them cannot fit in an uint32_t
2263 FitWeights(NewWeights);
2265 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2266 PBI->setMetadata(LLVMContext::MD_prof,
2267 MDBuilder(BI->getContext()).
2268 createBranchWeights(MDWeights));
2270 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2272 // Update PHI nodes in the common successors.
2273 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2274 ConstantInt *PBI_C = cast<ConstantInt>(
2275 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2276 assert(PBI_C->getType()->isIntegerTy(1));
2277 Instruction *MergedCond = nullptr;
2278 if (PBI->getSuccessor(0) == TrueDest) {
2279 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2280 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2281 // is false: !PBI_Cond and BI_Value
2282 Instruction *NotCond =
2283 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2286 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2291 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2292 PBI->getCondition(), MergedCond,
2295 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2296 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2297 // is false: PBI_Cond and BI_Value
2299 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2300 PBI->getCondition(), New,
2302 if (PBI_C->isOne()) {
2303 Instruction *NotCond =
2304 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2307 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2308 NotCond, MergedCond,
2313 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2316 // Change PBI from Conditional to Unconditional.
2317 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2318 EraseTerminatorInstAndDCECond(PBI);
2322 // TODO: If BB is reachable from all paths through PredBlock, then we
2323 // could replace PBI's branch probabilities with BI's.
2325 // Copy any debug value intrinsics into the end of PredBlock.
2326 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2327 if (isa<DbgInfoIntrinsic>(*I))
2328 I->clone()->insertBefore(PBI);
2335 /// If we have a conditional branch as a predecessor of another block,
2336 /// this function tries to simplify it. We know
2337 /// that PBI and BI are both conditional branches, and BI is in one of the
2338 /// successor blocks of PBI - PBI branches to BI.
2339 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
2340 const DataLayout &DL) {
2341 assert(PBI->isConditional() && BI->isConditional());
2342 BasicBlock *BB = BI->getParent();
2344 // If this block ends with a branch instruction, and if there is a
2345 // predecessor that ends on a branch of the same condition, make
2346 // this conditional branch redundant.
2347 if (PBI->getCondition() == BI->getCondition() &&
2348 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2349 // Okay, the outcome of this conditional branch is statically
2350 // knowable. If this block had a single pred, handle specially.
2351 if (BB->getSinglePredecessor()) {
2352 // Turn this into a branch on constant.
2353 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2354 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2356 return true; // Nuke the branch on constant.
2359 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2360 // in the constant and simplify the block result. Subsequent passes of
2361 // simplifycfg will thread the block.
2362 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2363 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2364 PHINode *NewPN = PHINode::Create(
2365 Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
2366 BI->getCondition()->getName() + ".pr", &BB->front());
2367 // Okay, we're going to insert the PHI node. Since PBI is not the only
2368 // predecessor, compute the PHI'd conditional value for all of the preds.
2369 // Any predecessor where the condition is not computable we keep symbolic.
2370 for (pred_iterator PI = PB; PI != PE; ++PI) {
2371 BasicBlock *P = *PI;
2372 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2373 PBI != BI && PBI->isConditional() &&
2374 PBI->getCondition() == BI->getCondition() &&
2375 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2376 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2377 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2380 NewPN->addIncoming(BI->getCondition(), P);
2384 BI->setCondition(NewPN);
2389 if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2393 // If this is a conditional branch in an empty block, and if any
2394 // predecessors are a conditional branch to one of our destinations,
2395 // fold the conditions into logical ops and one cond br.
2396 BasicBlock::iterator BBI = BB->begin();
2397 // Ignore dbg intrinsics.
2398 while (isa<DbgInfoIntrinsic>(BBI))
2404 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2406 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2407 PBIOp = 0, BIOp = 1;
2408 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2409 PBIOp = 1, BIOp = 0;
2410 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2415 // Check to make sure that the other destination of this branch
2416 // isn't BB itself. If so, this is an infinite loop that will
2417 // keep getting unwound.
2418 if (PBI->getSuccessor(PBIOp) == BB)
2421 // Do not perform this transformation if it would require
2422 // insertion of a large number of select instructions. For targets
2423 // without predication/cmovs, this is a big pessimization.
2425 // Also do not perform this transformation if any phi node in the common
2426 // destination block can trap when reached by BB or PBB (PR17073). In that
2427 // case, it would be unsafe to hoist the operation into a select instruction.
2429 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2430 unsigned NumPhis = 0;
2431 for (BasicBlock::iterator II = CommonDest->begin();
2432 isa<PHINode>(II); ++II, ++NumPhis) {
2433 if (NumPhis > 2) // Disable this xform.
2436 PHINode *PN = cast<PHINode>(II);
2437 Value *BIV = PN->getIncomingValueForBlock(BB);
2438 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
2442 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2443 Value *PBIV = PN->getIncomingValue(PBBIdx);
2444 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
2449 // Finally, if everything is ok, fold the branches to logical ops.
2450 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2452 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2453 << "AND: " << *BI->getParent());
2456 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2457 // branch in it, where one edge (OtherDest) goes back to itself but the other
2458 // exits. We don't *know* that the program avoids the infinite loop
2459 // (even though that seems likely). If we do this xform naively, we'll end up
2460 // recursively unpeeling the loop. Since we know that (after the xform is
2461 // done) that the block *is* infinite if reached, we just make it an obviously
2462 // infinite loop with no cond branch.
2463 if (OtherDest == BB) {
2464 // Insert it at the end of the function, because it's either code,
2465 // or it won't matter if it's hot. :)
2466 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2467 "infloop", BB->getParent());
2468 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2469 OtherDest = InfLoopBlock;
2472 DEBUG(dbgs() << *PBI->getParent()->getParent());
2474 // BI may have other predecessors. Because of this, we leave
2475 // it alone, but modify PBI.
2477 // Make sure we get to CommonDest on True&True directions.
2478 Value *PBICond = PBI->getCondition();
2479 IRBuilder<true, NoFolder> Builder(PBI);
2481 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2483 Value *BICond = BI->getCondition();
2485 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2487 // Merge the conditions.
2488 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2490 // Modify PBI to branch on the new condition to the new dests.
2491 PBI->setCondition(Cond);
2492 PBI->setSuccessor(0, CommonDest);
2493 PBI->setSuccessor(1, OtherDest);
2495 // Update branch weight for PBI.
2496 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2497 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2499 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2501 if (PredHasWeights && SuccHasWeights) {
2502 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2503 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2504 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2505 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2506 // The weight to CommonDest should be PredCommon * SuccTotal +
2507 // PredOther * SuccCommon.
2508 // The weight to OtherDest should be PredOther * SuccOther.
2509 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
2510 PredOther * SuccCommon,
2511 PredOther * SuccOther};
2512 // Halve the weights if any of them cannot fit in an uint32_t
2513 FitWeights(NewWeights);
2515 PBI->setMetadata(LLVMContext::MD_prof,
2516 MDBuilder(BI->getContext())
2517 .createBranchWeights(NewWeights[0], NewWeights[1]));
2520 // OtherDest may have phi nodes. If so, add an entry from PBI's
2521 // block that are identical to the entries for BI's block.
2522 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2524 // We know that the CommonDest already had an edge from PBI to
2525 // it. If it has PHIs though, the PHIs may have different
2526 // entries for BB and PBI's BB. If so, insert a select to make
2529 for (BasicBlock::iterator II = CommonDest->begin();
2530 (PN = dyn_cast<PHINode>(II)); ++II) {
2531 Value *BIV = PN->getIncomingValueForBlock(BB);
2532 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2533 Value *PBIV = PN->getIncomingValue(PBBIdx);
2535 // Insert a select in PBI to pick the right value.
2536 Value *NV = cast<SelectInst>
2537 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2538 PN->setIncomingValue(PBBIdx, NV);
2542 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2543 DEBUG(dbgs() << *PBI->getParent()->getParent());
2545 // This basic block is probably dead. We know it has at least
2546 // one fewer predecessor.
2550 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
2551 // true or to FalseBB if Cond is false.
2552 // Takes care of updating the successors and removing the old terminator.
2553 // Also makes sure not to introduce new successors by assuming that edges to
2554 // non-successor TrueBBs and FalseBBs aren't reachable.
2555 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2556 BasicBlock *TrueBB, BasicBlock *FalseBB,
2557 uint32_t TrueWeight,
2558 uint32_t FalseWeight){
2559 // Remove any superfluous successor edges from the CFG.
2560 // First, figure out which successors to preserve.
2561 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2563 BasicBlock *KeepEdge1 = TrueBB;
2564 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
2566 // Then remove the rest.
2567 for (BasicBlock *Succ : OldTerm->successors()) {
2568 // Make sure only to keep exactly one copy of each edge.
2569 if (Succ == KeepEdge1)
2570 KeepEdge1 = nullptr;
2571 else if (Succ == KeepEdge2)
2572 KeepEdge2 = nullptr;
2574 Succ->removePredecessor(OldTerm->getParent(),
2575 /*DontDeleteUselessPHIs=*/true);
2578 IRBuilder<> Builder(OldTerm);
2579 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2581 // Insert an appropriate new terminator.
2582 if (!KeepEdge1 && !KeepEdge2) {
2583 if (TrueBB == FalseBB)
2584 // We were only looking for one successor, and it was present.
2585 // Create an unconditional branch to it.
2586 Builder.CreateBr(TrueBB);
2588 // We found both of the successors we were looking for.
2589 // Create a conditional branch sharing the condition of the select.
2590 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2591 if (TrueWeight != FalseWeight)
2592 NewBI->setMetadata(LLVMContext::MD_prof,
2593 MDBuilder(OldTerm->getContext()).
2594 createBranchWeights(TrueWeight, FalseWeight));
2596 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2597 // Neither of the selected blocks were successors, so this
2598 // terminator must be unreachable.
2599 new UnreachableInst(OldTerm->getContext(), OldTerm);
2601 // One of the selected values was a successor, but the other wasn't.
2602 // Insert an unconditional branch to the one that was found;
2603 // the edge to the one that wasn't must be unreachable.
2605 // Only TrueBB was found.
2606 Builder.CreateBr(TrueBB);
2608 // Only FalseBB was found.
2609 Builder.CreateBr(FalseBB);
2612 EraseTerminatorInstAndDCECond(OldTerm);
2617 // (switch (select cond, X, Y)) on constant X, Y
2618 // with a branch - conditional if X and Y lead to distinct BBs,
2619 // unconditional otherwise.
2620 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2621 // Check for constant integer values in the select.
2622 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2623 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2624 if (!TrueVal || !FalseVal)
2627 // Find the relevant condition and destinations.
2628 Value *Condition = Select->getCondition();
2629 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2630 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2632 // Get weight for TrueBB and FalseBB.
2633 uint32_t TrueWeight = 0, FalseWeight = 0;
2634 SmallVector<uint64_t, 8> Weights;
2635 bool HasWeights = HasBranchWeights(SI);
2637 GetBranchWeights(SI, Weights);
2638 if (Weights.size() == 1 + SI->getNumCases()) {
2639 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2640 getSuccessorIndex()];
2641 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2642 getSuccessorIndex()];
2646 // Perform the actual simplification.
2647 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2648 TrueWeight, FalseWeight);
2652 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2653 // blockaddress(@fn, BlockB)))
2655 // (br cond, BlockA, BlockB).
2656 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2657 // Check that both operands of the select are block addresses.
2658 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2659 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2663 // Extract the actual blocks.
2664 BasicBlock *TrueBB = TBA->getBasicBlock();
2665 BasicBlock *FalseBB = FBA->getBasicBlock();
2667 // Perform the actual simplification.
2668 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2672 /// This is called when we find an icmp instruction
2673 /// (a seteq/setne with a constant) as the only instruction in a
2674 /// block that ends with an uncond branch. We are looking for a very specific
2675 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2676 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2677 /// default value goes to an uncond block with a seteq in it, we get something
2680 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2682 /// %tmp = icmp eq i8 %A, 92
2685 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2687 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2688 /// the PHI, merging the third icmp into the switch.
2689 static bool TryToSimplifyUncondBranchWithICmpInIt(
2690 ICmpInst *ICI, IRBuilder<> &Builder, const DataLayout &DL,
2691 const TargetTransformInfo &TTI, unsigned BonusInstThreshold,
2692 AssumptionCache *AC) {
2693 BasicBlock *BB = ICI->getParent();
2695 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2697 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2699 Value *V = ICI->getOperand(0);
2700 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2702 // The pattern we're looking for is where our only predecessor is a switch on
2703 // 'V' and this block is the default case for the switch. In this case we can
2704 // fold the compared value into the switch to simplify things.
2705 BasicBlock *Pred = BB->getSinglePredecessor();
2706 if (!Pred || !isa<SwitchInst>(Pred->getTerminator())) return false;
2708 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2709 if (SI->getCondition() != V)
2712 // If BB is reachable on a non-default case, then we simply know the value of
2713 // V in this block. Substitute it and constant fold the icmp instruction
2715 if (SI->getDefaultDest() != BB) {
2716 ConstantInt *VVal = SI->findCaseDest(BB);
2717 assert(VVal && "Should have a unique destination value");
2718 ICI->setOperand(0, VVal);
2720 if (Value *V = SimplifyInstruction(ICI, DL)) {
2721 ICI->replaceAllUsesWith(V);
2722 ICI->eraseFromParent();
2724 // BB is now empty, so it is likely to simplify away.
2725 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
2728 // Ok, the block is reachable from the default dest. If the constant we're
2729 // comparing exists in one of the other edges, then we can constant fold ICI
2731 if (SI->findCaseValue(Cst) != SI->case_default()) {
2733 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2734 V = ConstantInt::getFalse(BB->getContext());
2736 V = ConstantInt::getTrue(BB->getContext());
2738 ICI->replaceAllUsesWith(V);
2739 ICI->eraseFromParent();
2740 // BB is now empty, so it is likely to simplify away.
2741 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
2744 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2746 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2747 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
2748 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
2749 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2752 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2754 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2755 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2757 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2758 std::swap(DefaultCst, NewCst);
2760 // Replace ICI (which is used by the PHI for the default value) with true or
2761 // false depending on if it is EQ or NE.
2762 ICI->replaceAllUsesWith(DefaultCst);
2763 ICI->eraseFromParent();
2765 // Okay, the switch goes to this block on a default value. Add an edge from
2766 // the switch to the merge point on the compared value.
2767 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2768 BB->getParent(), BB);
2769 SmallVector<uint64_t, 8> Weights;
2770 bool HasWeights = HasBranchWeights(SI);
2772 GetBranchWeights(SI, Weights);
2773 if (Weights.size() == 1 + SI->getNumCases()) {
2774 // Split weight for default case to case for "Cst".
2775 Weights[0] = (Weights[0]+1) >> 1;
2776 Weights.push_back(Weights[0]);
2778 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2779 SI->setMetadata(LLVMContext::MD_prof,
2780 MDBuilder(SI->getContext()).
2781 createBranchWeights(MDWeights));
2784 SI->addCase(Cst, NewBB);
2786 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2787 Builder.SetInsertPoint(NewBB);
2788 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2789 Builder.CreateBr(SuccBlock);
2790 PHIUse->addIncoming(NewCst, NewBB);
2794 /// The specified branch is a conditional branch.
2795 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2796 /// fold it into a switch instruction if so.
2797 static bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
2798 const DataLayout &DL) {
2799 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2800 if (!Cond) return false;
2802 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2803 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2804 // 'setne's and'ed together, collect them.
2806 // Try to gather values from a chain of and/or to be turned into a switch
2807 ConstantComparesGatherer ConstantCompare(Cond, DL);
2808 // Unpack the result
2809 SmallVectorImpl<ConstantInt*> &Values = ConstantCompare.Vals;
2810 Value *CompVal = ConstantCompare.CompValue;
2811 unsigned UsedICmps = ConstantCompare.UsedICmps;
2812 Value *ExtraCase = ConstantCompare.Extra;
2814 // If we didn't have a multiply compared value, fail.
2815 if (!CompVal) return false;
2817 // Avoid turning single icmps into a switch.
2821 bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
2823 // There might be duplicate constants in the list, which the switch
2824 // instruction can't handle, remove them now.
2825 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2826 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2828 // If Extra was used, we require at least two switch values to do the
2829 // transformation. A switch with one value is just a conditional branch.
2830 if (ExtraCase && Values.size() < 2) return false;
2832 // TODO: Preserve branch weight metadata, similarly to how
2833 // FoldValueComparisonIntoPredecessors preserves it.
2835 // Figure out which block is which destination.
2836 BasicBlock *DefaultBB = BI->getSuccessor(1);
2837 BasicBlock *EdgeBB = BI->getSuccessor(0);
2838 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2840 BasicBlock *BB = BI->getParent();
2842 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2843 << " cases into SWITCH. BB is:\n" << *BB);
2845 // If there are any extra values that couldn't be folded into the switch
2846 // then we evaluate them with an explicit branch first. Split the block
2847 // right before the condbr to handle it.
2850 BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
2851 // Remove the uncond branch added to the old block.
2852 TerminatorInst *OldTI = BB->getTerminator();
2853 Builder.SetInsertPoint(OldTI);
2856 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2858 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2860 OldTI->eraseFromParent();
2862 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2863 // for the edge we just added.
2864 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2866 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2867 << "\nEXTRABB = " << *BB);
2871 Builder.SetInsertPoint(BI);
2872 // Convert pointer to int before we switch.
2873 if (CompVal->getType()->isPointerTy()) {
2874 CompVal = Builder.CreatePtrToInt(
2875 CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
2878 // Create the new switch instruction now.
2879 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2881 // Add all of the 'cases' to the switch instruction.
2882 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2883 New->addCase(Values[i], EdgeBB);
2885 // We added edges from PI to the EdgeBB. As such, if there were any
2886 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2887 // the number of edges added.
2888 for (BasicBlock::iterator BBI = EdgeBB->begin();
2889 isa<PHINode>(BBI); ++BBI) {
2890 PHINode *PN = cast<PHINode>(BBI);
2891 Value *InVal = PN->getIncomingValueForBlock(BB);
2892 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2893 PN->addIncoming(InVal, BB);
2896 // Erase the old branch instruction.
2897 EraseTerminatorInstAndDCECond(BI);
2899 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2903 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2904 // If this is a trivial landing pad that just continues unwinding the caught
2905 // exception then zap the landing pad, turning its invokes into calls.
2906 BasicBlock *BB = RI->getParent();
2907 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2909 // If RI->getValue() is a landing pad, check if the first instruction is
2910 // the same landing pad that caused control to branch here. If RI->getValue
2911 // is a phi of landing pad, check its predecessor blocks to see if any of
2912 // them contains a trivial landing pad.
2913 if (RI->getValue() != LPInst && !isa<PHINode>(RI->getValue()))
2916 // Check that there are no other instructions except for debug intrinsics
2917 // between the landing pad (or phi of landing pad) and resume instruction.
2918 BasicBlock::iterator I = cast<Instruction>(RI->getValue()), E = RI;
2920 if (!isa<DbgInfoIntrinsic>(I))
2923 SmallVector<BasicBlock *, 4> TrivialUnwindBlocks;
2924 if (RI->getValue() == LPInst) {
2925 // Landing pad is in current block, which has already been checked.
2926 TrivialUnwindBlocks.push_back(BB);
2928 // Check incoming blocks to see if any of them are trivial.
2929 auto *PhiLPInst = cast<PHINode>(RI->getValue());
2930 for (unsigned i = 0; i < PhiLPInst->getNumIncomingValues(); i++) {
2931 auto *IncomingBB = PhiLPInst->getIncomingBlock(i);
2932 auto *IncomingValue = PhiLPInst->getIncomingValue(i);
2935 dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
2936 // Not the landing pad that caused the control to branch here.
2937 if (IncomingValue != LandingPad)
2940 bool isTrivial = true;
2942 I = IncomingBB->getFirstNonPHI();
2943 E = IncomingBB->getTerminator();
2945 if (!isa<DbgInfoIntrinsic>(I)) {
2951 TrivialUnwindBlocks.push_back(IncomingBB);
2955 // Turn all invokes that unwind here into calls and delete the basic block.
2956 for (auto *TrivialBB : TrivialUnwindBlocks) {
2957 if (isa<PHINode>(RI->getValue())) {
2958 // Blocks that will be deleted should also be removed
2959 // from the phi node.
2960 BB->removePredecessor(TrivialBB, true);
2963 for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
2965 BasicBlock *Pred = *PI++;
2966 removeUnwindEdge(Pred);
2969 // The landingpad is now unreachable. Zap it.
2970 if (TrivialBB == BB)
2972 TrivialBB->eraseFromParent();
2975 // Delete the resume block if all its predecessors have been deleted, and we
2976 // haven't already deleted it above when deleting the landing pad blocks.
2977 if (BB && pred_begin(BB) == pred_end(BB)) {
2978 BB->eraseFromParent();
2981 return TrivialUnwindBlocks.size() != 0;
2984 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
2985 // If this is a trivial cleanup pad that executes no instructions, it can be
2986 // eliminated. If the cleanup pad continues to the caller, any predecessor
2987 // that is an EH pad will be updated to continue to the caller and any
2988 // predecessor that terminates with an invoke instruction will have its invoke
2989 // instruction converted to a call instruction. If the cleanup pad being
2990 // simplified does not continue to the caller, each predecessor will be
2991 // updated to continue to the unwind destination of the cleanup pad being
2993 BasicBlock *BB = RI->getParent();
2994 Instruction *CPInst = dyn_cast<CleanupPadInst>(BB->getFirstNonPHI());
2996 // This isn't an empty cleanup.
2999 // Check that there are no other instructions except for debug intrinsics.
3000 BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3002 if (!isa<DbgInfoIntrinsic>(I))
3005 // If the cleanup return we are simplifying unwinds to the caller, this
3006 // will set UnwindDest to nullptr.
3007 BasicBlock *UnwindDest = RI->getUnwindDest();
3009 // We're about to remove BB from the control flow. Before we do, sink any
3010 // PHINodes into the unwind destination. Doing this before changing the
3011 // control flow avoids some potentially slow checks, since we can currently
3012 // be certain that UnwindDest and BB have no common predecessors (since they
3013 // are both EH pads).
3015 // First, go through the PHI nodes in UnwindDest and update any nodes that
3016 // reference the block we are removing
3017 for (BasicBlock::iterator I = UnwindDest->begin(),
3018 IE = UnwindDest->getFirstNonPHI()->getIterator();
3020 PHINode *DestPN = cast<PHINode>(I);
3022 int Idx = DestPN->getBasicBlockIndex(BB);
3023 // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3025 // This PHI node has an incoming value that corresponds to a control
3026 // path through the cleanup pad we are removing. If the incoming
3027 // value is in the cleanup pad, it must be a PHINode (because we
3028 // verified above that the block is otherwise empty). Otherwise, the
3029 // value is either a constant or a value that dominates the cleanup
3030 // pad being removed.
3032 // Because BB and UnwindDest are both EH pads, all of their
3033 // predecessors must unwind to these blocks, and since no instruction
3034 // can have multiple unwind destinations, there will be no overlap in
3035 // incoming blocks between SrcPN and DestPN.
3036 Value *SrcVal = DestPN->getIncomingValue(Idx);
3037 PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3039 // Remove the entry for the block we are deleting.
3040 DestPN->removeIncomingValue(Idx, false);
3042 if (SrcPN && SrcPN->getParent() == BB) {
3043 // If the incoming value was a PHI node in the cleanup pad we are
3044 // removing, we need to merge that PHI node's incoming values into
3046 for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3047 SrcIdx != SrcE; ++SrcIdx) {
3048 DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3049 SrcPN->getIncomingBlock(SrcIdx));
3052 // Otherwise, the incoming value came from above BB and
3053 // so we can just reuse it. We must associate all of BB's
3054 // predecessors with this value.
3055 for (auto *pred : predecessors(BB)) {
3056 DestPN->addIncoming(SrcVal, pred);
3061 // Sink any remaining PHI nodes directly into UnwindDest.
3062 Instruction *InsertPt = UnwindDest->getFirstNonPHI();
3063 for (BasicBlock::iterator I = BB->begin(),
3064 IE = BB->getFirstNonPHI()->getIterator();
3066 // The iterator must be incremented here because the instructions are
3067 // being moved to another block.
3068 PHINode *PN = cast<PHINode>(I++);
3069 if (PN->use_empty())
3070 // If the PHI node has no uses, just leave it. It will be erased
3071 // when we erase BB below.
3074 // Otherwise, sink this PHI node into UnwindDest.
3075 // Any predecessors to UnwindDest which are not already represented
3076 // must be back edges which inherit the value from the path through
3077 // BB. In this case, the PHI value must reference itself.
3078 for (auto *pred : predecessors(UnwindDest))
3080 PN->addIncoming(PN, pred);
3081 PN->moveBefore(InsertPt);
3085 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3086 // The iterator must be updated here because we are removing this pred.
3087 BasicBlock *PredBB = *PI++;
3088 if (UnwindDest == nullptr) {
3089 removeUnwindEdge(PredBB);
3091 TerminatorInst *TI = PredBB->getTerminator();
3092 TI->replaceUsesOfWith(BB, UnwindDest);
3096 // The cleanup pad is now unreachable. Zap it.
3097 BB->eraseFromParent();
3101 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
3102 BasicBlock *BB = RI->getParent();
3103 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
3105 // Find predecessors that end with branches.
3106 SmallVector<BasicBlock*, 8> UncondBranchPreds;
3107 SmallVector<BranchInst*, 8> CondBranchPreds;
3108 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
3109 BasicBlock *P = *PI;
3110 TerminatorInst *PTI = P->getTerminator();
3111 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
3112 if (BI->isUnconditional())
3113 UncondBranchPreds.push_back(P);
3115 CondBranchPreds.push_back(BI);
3119 // If we found some, do the transformation!
3120 if (!UncondBranchPreds.empty() && DupRet) {
3121 while (!UncondBranchPreds.empty()) {
3122 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
3123 DEBUG(dbgs() << "FOLDING: " << *BB
3124 << "INTO UNCOND BRANCH PRED: " << *Pred);
3125 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
3128 // If we eliminated all predecessors of the block, delete the block now.
3130 // We know there are no successors, so just nuke the block.
3131 BB->eraseFromParent();
3136 // Check out all of the conditional branches going to this return
3137 // instruction. If any of them just select between returns, change the
3138 // branch itself into a select/return pair.
3139 while (!CondBranchPreds.empty()) {
3140 BranchInst *BI = CondBranchPreds.pop_back_val();
3142 // Check to see if the non-BB successor is also a return block.
3143 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
3144 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
3145 SimplifyCondBranchToTwoReturns(BI, Builder))
3151 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
3152 BasicBlock *BB = UI->getParent();
3154 bool Changed = false;
3156 // If there are any instructions immediately before the unreachable that can
3157 // be removed, do so.
3158 while (UI->getIterator() != BB->begin()) {
3159 BasicBlock::iterator BBI = UI->getIterator();
3161 // Do not delete instructions that can have side effects which might cause
3162 // the unreachable to not be reachable; specifically, calls and volatile
3163 // operations may have this effect.
3164 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
3166 if (BBI->mayHaveSideEffects()) {
3167 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
3168 if (SI->isVolatile())
3170 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
3171 if (LI->isVolatile())
3173 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
3174 if (RMWI->isVolatile())
3176 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
3177 if (CXI->isVolatile())
3179 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
3180 !isa<LandingPadInst>(BBI)) {
3183 // Note that deleting LandingPad's here is in fact okay, although it
3184 // involves a bit of subtle reasoning. If this inst is a LandingPad,
3185 // all the predecessors of this block will be the unwind edges of Invokes,
3186 // and we can therefore guarantee this block will be erased.
3189 // Delete this instruction (any uses are guaranteed to be dead)
3190 if (!BBI->use_empty())
3191 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
3192 BBI->eraseFromParent();
3196 // If the unreachable instruction is the first in the block, take a gander
3197 // at all of the predecessors of this instruction, and simplify them.
3198 if (&BB->front() != UI) return Changed;
3200 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
3201 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
3202 TerminatorInst *TI = Preds[i]->getTerminator();
3203 IRBuilder<> Builder(TI);
3204 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3205 if (BI->isUnconditional()) {
3206 if (BI->getSuccessor(0) == BB) {
3207 new UnreachableInst(TI->getContext(), TI);
3208 TI->eraseFromParent();
3212 if (BI->getSuccessor(0) == BB) {
3213 Builder.CreateBr(BI->getSuccessor(1));
3214 EraseTerminatorInstAndDCECond(BI);
3215 } else if (BI->getSuccessor(1) == BB) {
3216 Builder.CreateBr(BI->getSuccessor(0));
3217 EraseTerminatorInstAndDCECond(BI);
3221 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3222 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3224 if (i.getCaseSuccessor() == BB) {
3225 BB->removePredecessor(SI->getParent());
3230 } else if ((isa<InvokeInst>(TI) &&
3231 cast<InvokeInst>(TI)->getUnwindDest() == BB) ||
3232 isa<CatchEndPadInst>(TI) || isa<TerminatePadInst>(TI)) {
3233 removeUnwindEdge(TI->getParent());
3235 } else if (isa<CleanupReturnInst>(TI) || isa<CleanupEndPadInst>(TI) ||
3236 isa<CatchReturnInst>(TI)) {
3237 new UnreachableInst(TI->getContext(), TI);
3238 TI->eraseFromParent();
3241 // TODO: If TI is a CatchPadInst, then (BB must be its normal dest and)
3242 // we can eliminate it, redirecting its preds to its unwind successor,
3243 // or to the next outer handler if the removed catch is the last for its
3247 // If this block is now dead, remove it.
3248 if (pred_empty(BB) &&
3249 BB != &BB->getParent()->getEntryBlock()) {
3250 // We know there are no successors, so just nuke the block.
3251 BB->eraseFromParent();
3258 static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) {
3259 assert(Cases.size() >= 1);
3261 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3262 for (size_t I = 1, E = Cases.size(); I != E; ++I) {
3263 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
3269 /// Turn a switch with two reachable destinations into an integer range
3270 /// comparison and branch.
3271 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3272 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3275 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
3277 // Partition the cases into two sets with different destinations.
3278 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
3279 BasicBlock *DestB = nullptr;
3280 SmallVector <ConstantInt *, 16> CasesA;
3281 SmallVector <ConstantInt *, 16> CasesB;
3283 for (SwitchInst::CaseIt I : SI->cases()) {
3284 BasicBlock *Dest = I.getCaseSuccessor();
3285 if (!DestA) DestA = Dest;
3286 if (Dest == DestA) {
3287 CasesA.push_back(I.getCaseValue());
3290 if (!DestB) DestB = Dest;
3291 if (Dest == DestB) {
3292 CasesB.push_back(I.getCaseValue());
3295 return false; // More than two destinations.
3298 assert(DestA && DestB && "Single-destination switch should have been folded.");
3299 assert(DestA != DestB);
3300 assert(DestB != SI->getDefaultDest());
3301 assert(!CasesB.empty() && "There must be non-default cases.");
3302 assert(!CasesA.empty() || HasDefault);
3304 // Figure out if one of the sets of cases form a contiguous range.
3305 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
3306 BasicBlock *ContiguousDest = nullptr;
3307 BasicBlock *OtherDest = nullptr;
3308 if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
3309 ContiguousCases = &CasesA;
3310 ContiguousDest = DestA;
3312 } else if (CasesAreContiguous(CasesB)) {
3313 ContiguousCases = &CasesB;
3314 ContiguousDest = DestB;
3319 // Start building the compare and branch.
3321 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
3322 Constant *NumCases = ConstantInt::get(Offset->getType(), ContiguousCases->size());
3324 Value *Sub = SI->getCondition();
3325 if (!Offset->isNullValue())
3326 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
3329 // If NumCases overflowed, then all possible values jump to the successor.
3330 if (NumCases->isNullValue() && !ContiguousCases->empty())
3331 Cmp = ConstantInt::getTrue(SI->getContext());
3333 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3334 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
3336 // Update weight for the newly-created conditional branch.
3337 if (HasBranchWeights(SI)) {
3338 SmallVector<uint64_t, 8> Weights;
3339 GetBranchWeights(SI, Weights);
3340 if (Weights.size() == 1 + SI->getNumCases()) {
3341 uint64_t TrueWeight = 0;
3342 uint64_t FalseWeight = 0;
3343 for (size_t I = 0, E = Weights.size(); I != E; ++I) {
3344 if (SI->getSuccessor(I) == ContiguousDest)
3345 TrueWeight += Weights[I];
3347 FalseWeight += Weights[I];
3349 while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
3353 NewBI->setMetadata(LLVMContext::MD_prof,
3354 MDBuilder(SI->getContext()).createBranchWeights(
3355 (uint32_t)TrueWeight, (uint32_t)FalseWeight));
3359 // Prune obsolete incoming values off the successors' PHI nodes.
3360 for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
3361 unsigned PreviousEdges = ContiguousCases->size();
3362 if (ContiguousDest == SI->getDefaultDest()) ++PreviousEdges;
3363 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
3364 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3366 for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
3367 unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
3368 if (OtherDest == SI->getDefaultDest()) ++PreviousEdges;
3369 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
3370 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3374 SI->eraseFromParent();
3379 /// Compute masked bits for the condition of a switch
3380 /// and use it to remove dead cases.
3381 static bool EliminateDeadSwitchCases(SwitchInst *SI, AssumptionCache *AC,
3382 const DataLayout &DL) {
3383 Value *Cond = SI->getCondition();
3384 unsigned Bits = Cond->getType()->getIntegerBitWidth();
3385 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3386 computeKnownBits(Cond, KnownZero, KnownOne, DL, 0, AC, SI);
3388 // Gather dead cases.
3389 SmallVector<ConstantInt*, 8> DeadCases;
3390 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3391 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3392 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3393 DeadCases.push_back(I.getCaseValue());
3394 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3395 << I.getCaseValue() << "' is dead.\n");
3399 // If we can prove that the cases must cover all possible values, the
3400 // default destination becomes dead and we can remove it. If we know some
3401 // of the bits in the value, we can use that to more precisely compute the
3402 // number of possible unique case values.
3404 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
3405 const unsigned NumUnknownBits = Bits -
3406 (KnownZero.Or(KnownOne)).countPopulation();
3407 assert(NumUnknownBits <= Bits);
3408 if (HasDefault && DeadCases.empty() &&
3409 NumUnknownBits < 64 /* avoid overflow */ &&
3410 SI->getNumCases() == (1ULL << NumUnknownBits)) {
3411 DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
3412 BasicBlock *NewDefault = SplitBlockPredecessors(SI->getDefaultDest(),
3413 SI->getParent(), "");
3414 SI->setDefaultDest(&*NewDefault);
3415 SplitBlock(&*NewDefault, &NewDefault->front());
3416 auto *OldTI = NewDefault->getTerminator();
3417 new UnreachableInst(SI->getContext(), OldTI);
3418 EraseTerminatorInstAndDCECond(OldTI);
3422 SmallVector<uint64_t, 8> Weights;
3423 bool HasWeight = HasBranchWeights(SI);
3425 GetBranchWeights(SI, Weights);
3426 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3429 // Remove dead cases from the switch.
3430 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3431 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3432 assert(Case != SI->case_default() &&
3433 "Case was not found. Probably mistake in DeadCases forming.");
3435 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3439 // Prune unused values from PHI nodes.
3440 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3441 SI->removeCase(Case);
3443 if (HasWeight && Weights.size() >= 2) {
3444 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3445 SI->setMetadata(LLVMContext::MD_prof,
3446 MDBuilder(SI->getParent()->getContext()).
3447 createBranchWeights(MDWeights));
3450 return !DeadCases.empty();
3453 /// If BB would be eligible for simplification by
3454 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3455 /// by an unconditional branch), look at the phi node for BB in the successor
3456 /// block and see if the incoming value is equal to CaseValue. If so, return
3457 /// the phi node, and set PhiIndex to BB's index in the phi node.
3458 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3461 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3462 return nullptr; // BB must be empty to be a candidate for simplification.
3463 if (!BB->getSinglePredecessor())
3464 return nullptr; // BB must be dominated by the switch.
3466 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3467 if (!Branch || !Branch->isUnconditional())
3468 return nullptr; // Terminator must be unconditional branch.
3470 BasicBlock *Succ = Branch->getSuccessor(0);
3472 BasicBlock::iterator I = Succ->begin();
3473 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3474 int Idx = PHI->getBasicBlockIndex(BB);
3475 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3477 Value *InValue = PHI->getIncomingValue(Idx);
3478 if (InValue != CaseValue) continue;
3487 /// Try to forward the condition of a switch instruction to a phi node
3488 /// dominated by the switch, if that would mean that some of the destination
3489 /// blocks of the switch can be folded away.
3490 /// Returns true if a change is made.
3491 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3492 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3493 ForwardingNodesMap ForwardingNodes;
3495 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3496 ConstantInt *CaseValue = I.getCaseValue();
3497 BasicBlock *CaseDest = I.getCaseSuccessor();
3500 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3504 ForwardingNodes[PHI].push_back(PhiIndex);
3507 bool Changed = false;
3509 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3510 E = ForwardingNodes.end(); I != E; ++I) {
3511 PHINode *Phi = I->first;
3512 SmallVectorImpl<int> &Indexes = I->second;
3514 if (Indexes.size() < 2) continue;
3516 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3517 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3524 /// Return true if the backend will be able to handle
3525 /// initializing an array of constants like C.
3526 static bool ValidLookupTableConstant(Constant *C) {
3527 if (C->isThreadDependent())
3529 if (C->isDLLImportDependent())
3532 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3533 return CE->isGEPWithNoNotionalOverIndexing();
3535 return isa<ConstantFP>(C) ||
3536 isa<ConstantInt>(C) ||
3537 isa<ConstantPointerNull>(C) ||
3538 isa<GlobalValue>(C) ||
3542 /// If V is a Constant, return it. Otherwise, try to look up
3543 /// its constant value in ConstantPool, returning 0 if it's not there.
3544 static Constant *LookupConstant(Value *V,
3545 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3546 if (Constant *C = dyn_cast<Constant>(V))
3548 return ConstantPool.lookup(V);
3551 /// Try to fold instruction I into a constant. This works for
3552 /// simple instructions such as binary operations where both operands are
3553 /// constant or can be replaced by constants from the ConstantPool. Returns the
3554 /// resulting constant on success, 0 otherwise.
3556 ConstantFold(Instruction *I, const DataLayout &DL,
3557 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
3558 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3559 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3562 if (A->isAllOnesValue())
3563 return LookupConstant(Select->getTrueValue(), ConstantPool);
3564 if (A->isNullValue())
3565 return LookupConstant(Select->getFalseValue(), ConstantPool);
3569 SmallVector<Constant *, 4> COps;
3570 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
3571 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
3577 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
3578 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
3582 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), COps, DL);
3585 /// Try to determine the resulting constant values in phi nodes
3586 /// at the common destination basic block, *CommonDest, for one of the case
3587 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3588 /// case), of a switch instruction SI.
3590 GetCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest,
3591 BasicBlock **CommonDest,
3592 SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
3593 const DataLayout &DL) {
3594 // The block from which we enter the common destination.
3595 BasicBlock *Pred = SI->getParent();
3597 // If CaseDest is empty except for some side-effect free instructions through
3598 // which we can constant-propagate the CaseVal, continue to its successor.
3599 SmallDenseMap<Value*, Constant*> ConstantPool;
3600 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3601 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3603 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3604 // If the terminator is a simple branch, continue to the next block.
3605 if (T->getNumSuccessors() != 1)
3608 CaseDest = T->getSuccessor(0);
3609 } else if (isa<DbgInfoIntrinsic>(I)) {
3610 // Skip debug intrinsic.
3612 } else if (Constant *C = ConstantFold(&*I, DL, ConstantPool)) {
3613 // Instruction is side-effect free and constant.
3615 // If the instruction has uses outside this block or a phi node slot for
3616 // the block, it is not safe to bypass the instruction since it would then
3617 // no longer dominate all its uses.
3618 for (auto &Use : I->uses()) {
3619 User *User = Use.getUser();
3620 if (Instruction *I = dyn_cast<Instruction>(User))
3621 if (I->getParent() == CaseDest)
3623 if (PHINode *Phi = dyn_cast<PHINode>(User))
3624 if (Phi->getIncomingBlock(Use) == CaseDest)
3629 ConstantPool.insert(std::make_pair(&*I, C));
3635 // If we did not have a CommonDest before, use the current one.
3637 *CommonDest = CaseDest;
3638 // If the destination isn't the common one, abort.
3639 if (CaseDest != *CommonDest)
3642 // Get the values for this case from phi nodes in the destination block.
3643 BasicBlock::iterator I = (*CommonDest)->begin();
3644 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3645 int Idx = PHI->getBasicBlockIndex(Pred);
3649 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3654 // Be conservative about which kinds of constants we support.
3655 if (!ValidLookupTableConstant(ConstVal))
3658 Res.push_back(std::make_pair(PHI, ConstVal));
3661 return Res.size() > 0;
3664 // Helper function used to add CaseVal to the list of cases that generate
3666 static void MapCaseToResult(ConstantInt *CaseVal,
3667 SwitchCaseResultVectorTy &UniqueResults,
3669 for (auto &I : UniqueResults) {
3670 if (I.first == Result) {
3671 I.second.push_back(CaseVal);
3675 UniqueResults.push_back(std::make_pair(Result,
3676 SmallVector<ConstantInt*, 4>(1, CaseVal)));
3679 // Helper function that initializes a map containing
3680 // results for the PHI node of the common destination block for a switch
3681 // instruction. Returns false if multiple PHI nodes have been found or if
3682 // there is not a common destination block for the switch.
3683 static bool InitializeUniqueCases(SwitchInst *SI, PHINode *&PHI,
3684 BasicBlock *&CommonDest,
3685 SwitchCaseResultVectorTy &UniqueResults,
3686 Constant *&DefaultResult,
3687 const DataLayout &DL) {
3688 for (auto &I : SI->cases()) {
3689 ConstantInt *CaseVal = I.getCaseValue();
3691 // Resulting value at phi nodes for this case value.
3692 SwitchCaseResultsTy Results;
3693 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
3697 // Only one value per case is permitted
3698 if (Results.size() > 1)
3700 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
3702 // Check the PHI consistency.
3704 PHI = Results[0].first;
3705 else if (PHI != Results[0].first)
3708 // Find the default result value.
3709 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
3710 BasicBlock *DefaultDest = SI->getDefaultDest();
3711 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
3713 // If the default value is not found abort unless the default destination
3716 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
3717 if ((!DefaultResult &&
3718 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
3724 // Helper function that checks if it is possible to transform a switch with only
3725 // two cases (or two cases + default) that produces a result into a select.
3728 // case 10: %0 = icmp eq i32 %a, 10
3729 // return 10; %1 = select i1 %0, i32 10, i32 4
3730 // case 20: ----> %2 = icmp eq i32 %a, 20
3731 // return 2; %3 = select i1 %2, i32 2, i32 %1
3736 ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
3737 Constant *DefaultResult, Value *Condition,
3738 IRBuilder<> &Builder) {
3739 assert(ResultVector.size() == 2 &&
3740 "We should have exactly two unique results at this point");
3741 // If we are selecting between only two cases transform into a simple
3742 // select or a two-way select if default is possible.
3743 if (ResultVector[0].second.size() == 1 &&
3744 ResultVector[1].second.size() == 1) {
3745 ConstantInt *const FirstCase = ResultVector[0].second[0];
3746 ConstantInt *const SecondCase = ResultVector[1].second[0];
3748 bool DefaultCanTrigger = DefaultResult;
3749 Value *SelectValue = ResultVector[1].first;
3750 if (DefaultCanTrigger) {
3751 Value *const ValueCompare =
3752 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
3753 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
3754 DefaultResult, "switch.select");
3756 Value *const ValueCompare =
3757 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
3758 return Builder.CreateSelect(ValueCompare, ResultVector[0].first, SelectValue,
3765 // Helper function to cleanup a switch instruction that has been converted into
3766 // a select, fixing up PHI nodes and basic blocks.
3767 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
3769 IRBuilder<> &Builder) {
3770 BasicBlock *SelectBB = SI->getParent();
3771 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
3772 PHI->removeIncomingValue(SelectBB);
3773 PHI->addIncoming(SelectValue, SelectBB);
3775 Builder.CreateBr(PHI->getParent());
3777 // Remove the switch.
3778 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
3779 BasicBlock *Succ = SI->getSuccessor(i);
3781 if (Succ == PHI->getParent())
3783 Succ->removePredecessor(SelectBB);
3785 SI->eraseFromParent();
3788 /// If the switch is only used to initialize one or more
3789 /// phi nodes in a common successor block with only two different
3790 /// constant values, replace the switch with select.
3791 static bool SwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
3792 AssumptionCache *AC, const DataLayout &DL) {
3793 Value *const Cond = SI->getCondition();
3794 PHINode *PHI = nullptr;
3795 BasicBlock *CommonDest = nullptr;
3796 Constant *DefaultResult;
3797 SwitchCaseResultVectorTy UniqueResults;
3798 // Collect all the cases that will deliver the same value from the switch.
3799 if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
3802 // Selects choose between maximum two values.
3803 if (UniqueResults.size() != 2)
3805 assert(PHI != nullptr && "PHI for value select not found");
3807 Builder.SetInsertPoint(SI);
3808 Value *SelectValue = ConvertTwoCaseSwitch(
3810 DefaultResult, Cond, Builder);
3812 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
3815 // The switch couldn't be converted into a select.
3820 /// This class represents a lookup table that can be used to replace a switch.
3821 class SwitchLookupTable {
3823 /// Create a lookup table to use as a switch replacement with the contents
3824 /// of Values, using DefaultValue to fill any holes in the table.
3826 Module &M, uint64_t TableSize, ConstantInt *Offset,
3827 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
3828 Constant *DefaultValue, const DataLayout &DL);
3830 /// Build instructions with Builder to retrieve the value at
3831 /// the position given by Index in the lookup table.
3832 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3834 /// Return true if a table with TableSize elements of
3835 /// type ElementType would fit in a target-legal register.
3836 static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
3840 // Depending on the contents of the table, it can be represented in
3843 // For tables where each element contains the same value, we just have to
3844 // store that single value and return it for each lookup.
3847 // For tables where there is a linear relationship between table index
3848 // and values. We calculate the result with a simple multiplication
3849 // and addition instead of a table lookup.
3852 // For small tables with integer elements, we can pack them into a bitmap
3853 // that fits into a target-legal register. Values are retrieved by
3854 // shift and mask operations.
3857 // The table is stored as an array of values. Values are retrieved by load
3858 // instructions from the table.
3862 // For SingleValueKind, this is the single value.
3863 Constant *SingleValue;
3865 // For BitMapKind, this is the bitmap.
3866 ConstantInt *BitMap;
3867 IntegerType *BitMapElementTy;
3869 // For LinearMapKind, these are the constants used to derive the value.
3870 ConstantInt *LinearOffset;
3871 ConstantInt *LinearMultiplier;
3873 // For ArrayKind, this is the array.
3874 GlobalVariable *Array;
3878 SwitchLookupTable::SwitchLookupTable(
3879 Module &M, uint64_t TableSize, ConstantInt *Offset,
3880 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
3881 Constant *DefaultValue, const DataLayout &DL)
3882 : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
3883 LinearOffset(nullptr), LinearMultiplier(nullptr), Array(nullptr) {
3884 assert(Values.size() && "Can't build lookup table without values!");
3885 assert(TableSize >= Values.size() && "Can't fit values in table!");
3887 // If all values in the table are equal, this is that value.
3888 SingleValue = Values.begin()->second;
3890 Type *ValueType = Values.begin()->second->getType();
3892 // Build up the table contents.
3893 SmallVector<Constant*, 64> TableContents(TableSize);
3894 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3895 ConstantInt *CaseVal = Values[I].first;
3896 Constant *CaseRes = Values[I].second;
3897 assert(CaseRes->getType() == ValueType);
3899 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3901 TableContents[Idx] = CaseRes;
3903 if (CaseRes != SingleValue)
3904 SingleValue = nullptr;
3907 // Fill in any holes in the table with the default result.
3908 if (Values.size() < TableSize) {
3909 assert(DefaultValue &&
3910 "Need a default value to fill the lookup table holes.");
3911 assert(DefaultValue->getType() == ValueType);
3912 for (uint64_t I = 0; I < TableSize; ++I) {
3913 if (!TableContents[I])
3914 TableContents[I] = DefaultValue;
3917 if (DefaultValue != SingleValue)
3918 SingleValue = nullptr;
3921 // If each element in the table contains the same value, we only need to store
3922 // that single value.
3924 Kind = SingleValueKind;
3928 // Check if we can derive the value with a linear transformation from the
3930 if (isa<IntegerType>(ValueType)) {
3931 bool LinearMappingPossible = true;
3934 assert(TableSize >= 2 && "Should be a SingleValue table.");
3935 // Check if there is the same distance between two consecutive values.
3936 for (uint64_t I = 0; I < TableSize; ++I) {
3937 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
3939 // This is an undef. We could deal with it, but undefs in lookup tables
3940 // are very seldom. It's probably not worth the additional complexity.
3941 LinearMappingPossible = false;
3944 APInt Val = ConstVal->getValue();
3946 APInt Dist = Val - PrevVal;
3949 } else if (Dist != DistToPrev) {
3950 LinearMappingPossible = false;
3956 if (LinearMappingPossible) {
3957 LinearOffset = cast<ConstantInt>(TableContents[0]);
3958 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
3959 Kind = LinearMapKind;
3965 // If the type is integer and the table fits in a register, build a bitmap.
3966 if (WouldFitInRegister(DL, TableSize, ValueType)) {
3967 IntegerType *IT = cast<IntegerType>(ValueType);
3968 APInt TableInt(TableSize * IT->getBitWidth(), 0);
3969 for (uint64_t I = TableSize; I > 0; --I) {
3970 TableInt <<= IT->getBitWidth();
3971 // Insert values into the bitmap. Undef values are set to zero.
3972 if (!isa<UndefValue>(TableContents[I - 1])) {
3973 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
3974 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
3977 BitMap = ConstantInt::get(M.getContext(), TableInt);
3978 BitMapElementTy = IT;
3984 // Store the table in an array.
3985 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
3986 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3988 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3989 GlobalVariable::PrivateLinkage,
3992 Array->setUnnamedAddr(true);
3996 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
3998 case SingleValueKind:
4000 case LinearMapKind: {
4001 // Derive the result value from the input value.
4002 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
4003 false, "switch.idx.cast");
4004 if (!LinearMultiplier->isOne())
4005 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
4006 if (!LinearOffset->isZero())
4007 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
4011 // Type of the bitmap (e.g. i59).
4012 IntegerType *MapTy = BitMap->getType();
4014 // Cast Index to the same type as the bitmap.
4015 // Note: The Index is <= the number of elements in the table, so
4016 // truncating it to the width of the bitmask is safe.
4017 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
4019 // Multiply the shift amount by the element width.
4020 ShiftAmt = Builder.CreateMul(ShiftAmt,
4021 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
4025 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
4026 "switch.downshift");
4028 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
4032 // Make sure the table index will not overflow when treated as signed.
4033 IntegerType *IT = cast<IntegerType>(Index->getType());
4034 uint64_t TableSize = Array->getInitializer()->getType()
4035 ->getArrayNumElements();
4036 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
4037 Index = Builder.CreateZExt(Index,
4038 IntegerType::get(IT->getContext(),
4039 IT->getBitWidth() + 1),
4040 "switch.tableidx.zext");
4042 Value *GEPIndices[] = { Builder.getInt32(0), Index };
4043 Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
4044 GEPIndices, "switch.gep");
4045 return Builder.CreateLoad(GEP, "switch.load");
4048 llvm_unreachable("Unknown lookup table kind!");
4051 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
4053 Type *ElementType) {
4054 auto *IT = dyn_cast<IntegerType>(ElementType);
4057 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
4058 // are <= 15, we could try to narrow the type.
4060 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
4061 if (TableSize >= UINT_MAX/IT->getBitWidth())
4063 return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
4066 /// Determine whether a lookup table should be built for this switch, based on
4067 /// the number of cases, size of the table, and the types of the results.
4069 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
4070 const TargetTransformInfo &TTI, const DataLayout &DL,
4071 const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
4072 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
4073 return false; // TableSize overflowed, or mul below might overflow.
4075 bool AllTablesFitInRegister = true;
4076 bool HasIllegalType = false;
4077 for (const auto &I : ResultTypes) {
4078 Type *Ty = I.second;
4080 // Saturate this flag to true.
4081 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
4083 // Saturate this flag to false.
4084 AllTablesFitInRegister = AllTablesFitInRegister &&
4085 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
4087 // If both flags saturate, we're done. NOTE: This *only* works with
4088 // saturating flags, and all flags have to saturate first due to the
4089 // non-deterministic behavior of iterating over a dense map.
4090 if (HasIllegalType && !AllTablesFitInRegister)
4094 // If each table would fit in a register, we should build it anyway.
4095 if (AllTablesFitInRegister)
4098 // Don't build a table that doesn't fit in-register if it has illegal types.
4102 // The table density should be at least 40%. This is the same criterion as for
4103 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
4104 // FIXME: Find the best cut-off.
4105 return SI->getNumCases() * 10 >= TableSize * 4;
4108 /// Try to reuse the switch table index compare. Following pattern:
4110 /// if (idx < tablesize)
4111 /// r = table[idx]; // table does not contain default_value
4113 /// r = default_value;
4114 /// if (r != default_value)
4117 /// Is optimized to:
4119 /// cond = idx < tablesize;
4123 /// r = default_value;
4127 /// Jump threading will then eliminate the second if(cond).
4128 static void reuseTableCompare(User *PhiUser, BasicBlock *PhiBlock,
4129 BranchInst *RangeCheckBranch, Constant *DefaultValue,
4130 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values) {
4132 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
4136 // We require that the compare is in the same block as the phi so that jump
4137 // threading can do its work afterwards.
4138 if (CmpInst->getParent() != PhiBlock)
4141 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
4145 Value *RangeCmp = RangeCheckBranch->getCondition();
4146 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
4147 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
4149 // Check if the compare with the default value is constant true or false.
4150 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
4151 DefaultValue, CmpOp1, true);
4152 if (DefaultConst != TrueConst && DefaultConst != FalseConst)
4155 // Check if the compare with the case values is distinct from the default
4157 for (auto ValuePair : Values) {
4158 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
4159 ValuePair.second, CmpOp1, true);
4160 if (!CaseConst || CaseConst == DefaultConst)
4162 assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
4163 "Expect true or false as compare result.");
4166 // Check if the branch instruction dominates the phi node. It's a simple
4167 // dominance check, but sufficient for our needs.
4168 // Although this check is invariant in the calling loops, it's better to do it
4169 // at this late stage. Practically we do it at most once for a switch.
4170 BasicBlock *BranchBlock = RangeCheckBranch->getParent();
4171 for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
4172 BasicBlock *Pred = *PI;
4173 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
4177 if (DefaultConst == FalseConst) {
4178 // The compare yields the same result. We can replace it.
4179 CmpInst->replaceAllUsesWith(RangeCmp);
4180 ++NumTableCmpReuses;
4182 // The compare yields the same result, just inverted. We can replace it.
4183 Value *InvertedTableCmp = BinaryOperator::CreateXor(RangeCmp,
4184 ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
4186 CmpInst->replaceAllUsesWith(InvertedTableCmp);
4187 ++NumTableCmpReuses;
4191 /// If the switch is only used to initialize one or more phi nodes in a common
4192 /// successor block with different constant values, replace the switch with
4194 static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
4195 const DataLayout &DL,
4196 const TargetTransformInfo &TTI) {
4197 assert(SI->getNumCases() > 1 && "Degenerate switch?");
4199 // Only build lookup table when we have a target that supports it.
4200 if (!TTI.shouldBuildLookupTables())
4203 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
4204 // split off a dense part and build a lookup table for that.
4206 // FIXME: This creates arrays of GEPs to constant strings, which means each
4207 // GEP needs a runtime relocation in PIC code. We should just build one big
4208 // string and lookup indices into that.
4210 // Ignore switches with less than three cases. Lookup tables will not make them
4211 // faster, so we don't analyze them.
4212 if (SI->getNumCases() < 3)
4215 // Figure out the corresponding result for each case value and phi node in the
4216 // common destination, as well as the min and max case values.
4217 assert(SI->case_begin() != SI->case_end());
4218 SwitchInst::CaseIt CI = SI->case_begin();
4219 ConstantInt *MinCaseVal = CI.getCaseValue();
4220 ConstantInt *MaxCaseVal = CI.getCaseValue();
4222 BasicBlock *CommonDest = nullptr;
4223 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
4224 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
4225 SmallDenseMap<PHINode*, Constant*> DefaultResults;
4226 SmallDenseMap<PHINode*, Type*> ResultTypes;
4227 SmallVector<PHINode*, 4> PHIs;
4229 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
4230 ConstantInt *CaseVal = CI.getCaseValue();
4231 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
4232 MinCaseVal = CaseVal;
4233 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
4234 MaxCaseVal = CaseVal;
4236 // Resulting value at phi nodes for this case value.
4237 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
4239 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
4243 // Append the result from this case to the list for each phi.
4244 for (const auto &I : Results) {
4245 PHINode *PHI = I.first;
4246 Constant *Value = I.second;
4247 if (!ResultLists.count(PHI))
4248 PHIs.push_back(PHI);
4249 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
4253 // Keep track of the result types.
4254 for (PHINode *PHI : PHIs) {
4255 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
4258 uint64_t NumResults = ResultLists[PHIs[0]].size();
4259 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
4260 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
4261 bool TableHasHoles = (NumResults < TableSize);
4263 // If the table has holes, we need a constant result for the default case
4264 // or a bitmask that fits in a register.
4265 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
4266 bool HasDefaultResults = GetCaseResults(SI, nullptr, SI->getDefaultDest(),
4267 &CommonDest, DefaultResultsList, DL);
4269 bool NeedMask = (TableHasHoles && !HasDefaultResults);
4271 // As an extra penalty for the validity test we require more cases.
4272 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
4274 if (!DL.fitsInLegalInteger(TableSize))
4278 for (const auto &I : DefaultResultsList) {
4279 PHINode *PHI = I.first;
4280 Constant *Result = I.second;
4281 DefaultResults[PHI] = Result;
4284 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
4287 // Create the BB that does the lookups.
4288 Module &Mod = *CommonDest->getParent()->getParent();
4289 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
4291 CommonDest->getParent(),
4294 // Compute the table index value.
4295 Builder.SetInsertPoint(SI);
4296 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
4299 // Compute the maximum table size representable by the integer type we are
4301 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
4302 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
4303 assert(MaxTableSize >= TableSize &&
4304 "It is impossible for a switch to have more entries than the max "
4305 "representable value of its input integer type's size.");
4307 // If the default destination is unreachable, or if the lookup table covers
4308 // all values of the conditional variable, branch directly to the lookup table
4309 // BB. Otherwise, check that the condition is within the case range.
4310 const bool DefaultIsReachable =
4311 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4312 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
4313 BranchInst *RangeCheckBranch = nullptr;
4315 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
4316 Builder.CreateBr(LookupBB);
4317 // Note: We call removeProdecessor later since we need to be able to get the
4318 // PHI value for the default case in case we're using a bit mask.
4320 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
4321 MinCaseVal->getType(), TableSize));
4322 RangeCheckBranch = Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
4325 // Populate the BB that does the lookups.
4326 Builder.SetInsertPoint(LookupBB);
4329 // Before doing the lookup we do the hole check.
4330 // The LookupBB is therefore re-purposed to do the hole check
4331 // and we create a new LookupBB.
4332 BasicBlock *MaskBB = LookupBB;
4333 MaskBB->setName("switch.hole_check");
4334 LookupBB = BasicBlock::Create(Mod.getContext(),
4336 CommonDest->getParent(),
4339 // Make the mask's bitwidth at least 8bit and a power-of-2 to avoid
4340 // unnecessary illegal types.
4341 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
4342 APInt MaskInt(TableSizePowOf2, 0);
4343 APInt One(TableSizePowOf2, 1);
4344 // Build bitmask; fill in a 1 bit for every case.
4345 const ResultListTy &ResultList = ResultLists[PHIs[0]];
4346 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
4347 uint64_t Idx = (ResultList[I].first->getValue() -
4348 MinCaseVal->getValue()).getLimitedValue();
4349 MaskInt |= One << Idx;
4351 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
4353 // Get the TableIndex'th bit of the bitmask.
4354 // If this bit is 0 (meaning hole) jump to the default destination,
4355 // else continue with table lookup.
4356 IntegerType *MapTy = TableMask->getType();
4357 Value *MaskIndex = Builder.CreateZExtOrTrunc(TableIndex, MapTy,
4358 "switch.maskindex");
4359 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex,
4361 Value *LoBit = Builder.CreateTrunc(Shifted,
4362 Type::getInt1Ty(Mod.getContext()),
4364 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
4366 Builder.SetInsertPoint(LookupBB);
4367 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
4370 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
4371 // We cached PHINodes in PHIs, to avoid accessing deleted PHINodes later,
4372 // do not delete PHINodes here.
4373 SI->getDefaultDest()->removePredecessor(SI->getParent(),
4374 /*DontDeleteUselessPHIs=*/true);
4377 bool ReturnedEarly = false;
4378 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
4379 PHINode *PHI = PHIs[I];
4380 const ResultListTy &ResultList = ResultLists[PHI];
4382 // If using a bitmask, use any value to fill the lookup table holes.
4383 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
4384 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL);
4386 Value *Result = Table.BuildLookup(TableIndex, Builder);
4388 // If the result is used to return immediately from the function, we want to
4389 // do that right here.
4390 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
4391 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
4392 Builder.CreateRet(Result);
4393 ReturnedEarly = true;
4397 // Do a small peephole optimization: re-use the switch table compare if
4399 if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
4400 BasicBlock *PhiBlock = PHI->getParent();
4401 // Search for compare instructions which use the phi.
4402 for (auto *User : PHI->users()) {
4403 reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
4407 PHI->addIncoming(Result, LookupBB);
4411 Builder.CreateBr(CommonDest);
4413 // Remove the switch.
4414 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4415 BasicBlock *Succ = SI->getSuccessor(i);
4417 if (Succ == SI->getDefaultDest())
4419 Succ->removePredecessor(SI->getParent());
4421 SI->eraseFromParent();
4425 ++NumLookupTablesHoles;
4429 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
4430 BasicBlock *BB = SI->getParent();
4432 if (isValueEqualityComparison(SI)) {
4433 // If we only have one predecessor, and if it is a branch on this value,
4434 // see if that predecessor totally determines the outcome of this switch.
4435 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4436 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
4437 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4439 Value *Cond = SI->getCondition();
4440 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
4441 if (SimplifySwitchOnSelect(SI, Select))
4442 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4444 // If the block only contains the switch, see if we can fold the block
4445 // away into any preds.
4446 BasicBlock::iterator BBI = BB->begin();
4447 // Ignore dbg intrinsics.
4448 while (isa<DbgInfoIntrinsic>(BBI))
4451 if (FoldValueComparisonIntoPredecessors(SI, Builder))
4452 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4455 // Try to transform the switch into an icmp and a branch.
4456 if (TurnSwitchRangeIntoICmp(SI, Builder))
4457 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4459 // Remove unreachable cases.
4460 if (EliminateDeadSwitchCases(SI, AC, DL))
4461 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4463 if (SwitchToSelect(SI, Builder, AC, DL))
4464 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4466 if (ForwardSwitchConditionToPHI(SI))
4467 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4469 if (SwitchToLookupTable(SI, Builder, DL, TTI))
4470 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4475 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
4476 BasicBlock *BB = IBI->getParent();
4477 bool Changed = false;
4479 // Eliminate redundant destinations.
4480 SmallPtrSet<Value *, 8> Succs;
4481 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
4482 BasicBlock *Dest = IBI->getDestination(i);
4483 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
4484 Dest->removePredecessor(BB);
4485 IBI->removeDestination(i);
4491 if (IBI->getNumDestinations() == 0) {
4492 // If the indirectbr has no successors, change it to unreachable.
4493 new UnreachableInst(IBI->getContext(), IBI);
4494 EraseTerminatorInstAndDCECond(IBI);
4498 if (IBI->getNumDestinations() == 1) {
4499 // If the indirectbr has one successor, change it to a direct branch.
4500 BranchInst::Create(IBI->getDestination(0), IBI);
4501 EraseTerminatorInstAndDCECond(IBI);
4505 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
4506 if (SimplifyIndirectBrOnSelect(IBI, SI))
4507 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4512 /// Given an block with only a single landing pad and a unconditional branch
4513 /// try to find another basic block which this one can be merged with. This
4514 /// handles cases where we have multiple invokes with unique landing pads, but
4515 /// a shared handler.
4517 /// We specifically choose to not worry about merging non-empty blocks
4518 /// here. That is a PRE/scheduling problem and is best solved elsewhere. In
4519 /// practice, the optimizer produces empty landing pad blocks quite frequently
4520 /// when dealing with exception dense code. (see: instcombine, gvn, if-else
4521 /// sinking in this file)
4523 /// This is primarily a code size optimization. We need to avoid performing
4524 /// any transform which might inhibit optimization (such as our ability to
4525 /// specialize a particular handler via tail commoning). We do this by not
4526 /// merging any blocks which require us to introduce a phi. Since the same
4527 /// values are flowing through both blocks, we don't loose any ability to
4528 /// specialize. If anything, we make such specialization more likely.
4530 /// TODO - This transformation could remove entries from a phi in the target
4531 /// block when the inputs in the phi are the same for the two blocks being
4532 /// merged. In some cases, this could result in removal of the PHI entirely.
4533 static bool TryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI,
4535 auto Succ = BB->getUniqueSuccessor();
4537 // If there's a phi in the successor block, we'd likely have to introduce
4538 // a phi into the merged landing pad block.
4539 if (isa<PHINode>(*Succ->begin()))
4542 for (BasicBlock *OtherPred : predecessors(Succ)) {
4543 if (BB == OtherPred)
4545 BasicBlock::iterator I = OtherPred->begin();
4546 LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(I);
4547 if (!LPad2 || !LPad2->isIdenticalTo(LPad))
4549 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {}
4550 BranchInst *BI2 = dyn_cast<BranchInst>(I);
4551 if (!BI2 || !BI2->isIdenticalTo(BI))
4554 // We've found an identical block. Update our predeccessors to take that
4555 // path instead and make ourselves dead.
4556 SmallSet<BasicBlock *, 16> Preds;
4557 Preds.insert(pred_begin(BB), pred_end(BB));
4558 for (BasicBlock *Pred : Preds) {
4559 InvokeInst *II = cast<InvokeInst>(Pred->getTerminator());
4560 assert(II->getNormalDest() != BB &&
4561 II->getUnwindDest() == BB && "unexpected successor");
4562 II->setUnwindDest(OtherPred);
4565 // The debug info in OtherPred doesn't cover the merged control flow that
4566 // used to go through BB. We need to delete it or update it.
4567 for (auto I = OtherPred->begin(), E = OtherPred->end();
4569 Instruction &Inst = *I; I++;
4570 if (isa<DbgInfoIntrinsic>(Inst))
4571 Inst.eraseFromParent();
4574 SmallSet<BasicBlock *, 16> Succs;
4575 Succs.insert(succ_begin(BB), succ_end(BB));
4576 for (BasicBlock *Succ : Succs) {
4577 Succ->removePredecessor(BB);
4580 IRBuilder<> Builder(BI);
4581 Builder.CreateUnreachable();
4582 BI->eraseFromParent();
4588 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
4589 BasicBlock *BB = BI->getParent();
4591 if (SinkCommon && SinkThenElseCodeToEnd(BI))
4594 // If the Terminator is the only non-phi instruction, simplify the block.
4595 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg()->getIterator();
4596 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
4597 TryToSimplifyUncondBranchFromEmptyBlock(BB))
4600 // If the only instruction in the block is a seteq/setne comparison
4601 // against a constant, try to simplify the block.
4602 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
4603 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
4604 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
4606 if (I->isTerminator() &&
4607 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, DL, TTI,
4608 BonusInstThreshold, AC))
4612 // See if we can merge an empty landing pad block with another which is
4614 if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(I)) {
4615 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {}
4616 if (I->isTerminator() &&
4617 TryToMergeLandingPad(LPad, BI, BB))
4621 // If this basic block is ONLY a compare and a branch, and if a predecessor
4622 // branches to us and our successor, fold the comparison into the
4623 // predecessor and use logical operations to update the incoming value
4624 // for PHI nodes in common successor.
4625 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
4626 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4631 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
4632 BasicBlock *BB = BI->getParent();
4634 // Conditional branch
4635 if (isValueEqualityComparison(BI)) {
4636 // If we only have one predecessor, and if it is a branch on this value,
4637 // see if that predecessor totally determines the outcome of this
4639 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4640 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
4641 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4643 // This block must be empty, except for the setcond inst, if it exists.
4644 // Ignore dbg intrinsics.
4645 BasicBlock::iterator I = BB->begin();
4646 // Ignore dbg intrinsics.
4647 while (isa<DbgInfoIntrinsic>(I))
4650 if (FoldValueComparisonIntoPredecessors(BI, Builder))
4651 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4652 } else if (&*I == cast<Instruction>(BI->getCondition())){
4654 // Ignore dbg intrinsics.
4655 while (isa<DbgInfoIntrinsic>(I))
4657 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
4658 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4662 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
4663 if (SimplifyBranchOnICmpChain(BI, Builder, DL))
4666 // If this basic block is ONLY a compare and a branch, and if a predecessor
4667 // branches to us and one of our successors, fold the comparison into the
4668 // predecessor and use logical operations to pick the right destination.
4669 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
4670 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4672 // We have a conditional branch to two blocks that are only reachable
4673 // from BI. We know that the condbr dominates the two blocks, so see if
4674 // there is any identical code in the "then" and "else" blocks. If so, we
4675 // can hoist it up to the branching block.
4676 if (BI->getSuccessor(0)->getSinglePredecessor()) {
4677 if (BI->getSuccessor(1)->getSinglePredecessor()) {
4678 if (HoistThenElseCodeToIf(BI, TTI))
4679 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4681 // If Successor #1 has multiple preds, we may be able to conditionally
4682 // execute Successor #0 if it branches to Successor #1.
4683 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
4684 if (Succ0TI->getNumSuccessors() == 1 &&
4685 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
4686 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), TTI))
4687 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4689 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
4690 // If Successor #0 has multiple preds, we may be able to conditionally
4691 // execute Successor #1 if it branches to Successor #0.
4692 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
4693 if (Succ1TI->getNumSuccessors() == 1 &&
4694 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
4695 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), TTI))
4696 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4699 // If this is a branch on a phi node in the current block, thread control
4700 // through this block if any PHI node entries are constants.
4701 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
4702 if (PN->getParent() == BI->getParent())
4703 if (FoldCondBranchOnPHI(BI, DL))
4704 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4706 // Scan predecessor blocks for conditional branches.
4707 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
4708 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
4709 if (PBI != BI && PBI->isConditional())
4710 if (SimplifyCondBranchToCondBranch(PBI, BI, DL))
4711 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4716 /// Check if passing a value to an instruction will cause undefined behavior.
4717 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
4718 Constant *C = dyn_cast<Constant>(V);
4725 if (C->isNullValue()) {
4726 // Only look at the first use, avoid hurting compile time with long uselists
4727 User *Use = *I->user_begin();
4729 // Now make sure that there are no instructions in between that can alter
4730 // control flow (eg. calls)
4731 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
4732 if (i == I->getParent()->end() || i->mayHaveSideEffects())
4735 // Look through GEPs. A load from a GEP derived from NULL is still undefined
4736 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
4737 if (GEP->getPointerOperand() == I)
4738 return passingValueIsAlwaysUndefined(V, GEP);
4740 // Look through bitcasts.
4741 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
4742 return passingValueIsAlwaysUndefined(V, BC);
4744 // Load from null is undefined.
4745 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
4746 if (!LI->isVolatile())
4747 return LI->getPointerAddressSpace() == 0;
4749 // Store to null is undefined.
4750 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
4751 if (!SI->isVolatile())
4752 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
4757 /// If BB has an incoming value that will always trigger undefined behavior
4758 /// (eg. null pointer dereference), remove the branch leading here.
4759 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
4760 for (BasicBlock::iterator i = BB->begin();
4761 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
4762 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
4763 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
4764 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
4765 IRBuilder<> Builder(T);
4766 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
4767 BB->removePredecessor(PHI->getIncomingBlock(i));
4768 // Turn uncoditional branches into unreachables and remove the dead
4769 // destination from conditional branches.
4770 if (BI->isUnconditional())
4771 Builder.CreateUnreachable();
4773 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
4774 BI->getSuccessor(0));
4775 BI->eraseFromParent();
4778 // TODO: SwitchInst.
4784 bool SimplifyCFGOpt::run(BasicBlock *BB) {
4785 bool Changed = false;
4787 assert(BB && BB->getParent() && "Block not embedded in function!");
4788 assert(BB->getTerminator() && "Degenerate basic block encountered!");
4790 // Remove basic blocks that have no predecessors (except the entry block)...
4791 // or that just have themself as a predecessor. These are unreachable.
4792 if ((pred_empty(BB) &&
4793 BB != &BB->getParent()->getEntryBlock()) ||
4794 BB->getSinglePredecessor() == BB) {
4795 DEBUG(dbgs() << "Removing BB: \n" << *BB);
4796 DeleteDeadBlock(BB);
4800 // Check to see if we can constant propagate this terminator instruction
4802 Changed |= ConstantFoldTerminator(BB, true);
4804 // Check for and eliminate duplicate PHI nodes in this block.
4805 Changed |= EliminateDuplicatePHINodes(BB);
4807 // Check for and remove branches that will always cause undefined behavior.
4808 Changed |= removeUndefIntroducingPredecessor(BB);
4810 // Merge basic blocks into their predecessor if there is only one distinct
4811 // pred, and if there is only one distinct successor of the predecessor, and
4812 // if there are no PHI nodes.
4814 if (MergeBlockIntoPredecessor(BB))
4817 IRBuilder<> Builder(BB);
4819 // If there is a trivial two-entry PHI node in this basic block, and we can
4820 // eliminate it, do so now.
4821 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
4822 if (PN->getNumIncomingValues() == 2)
4823 Changed |= FoldTwoEntryPHINode(PN, TTI, DL);
4825 Builder.SetInsertPoint(BB->getTerminator());
4826 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
4827 if (BI->isUnconditional()) {
4828 if (SimplifyUncondBranch(BI, Builder)) return true;
4830 if (SimplifyCondBranch(BI, Builder)) return true;
4832 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
4833 if (SimplifyReturn(RI, Builder)) return true;
4834 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
4835 if (SimplifyResume(RI, Builder)) return true;
4836 } else if (CleanupReturnInst *RI =
4837 dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
4838 if (SimplifyCleanupReturn(RI)) return true;
4839 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
4840 if (SimplifySwitch(SI, Builder)) return true;
4841 } else if (UnreachableInst *UI =
4842 dyn_cast<UnreachableInst>(BB->getTerminator())) {
4843 if (SimplifyUnreachable(UI)) return true;
4844 } else if (IndirectBrInst *IBI =
4845 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
4846 if (SimplifyIndirectBr(IBI)) return true;
4852 /// This function is used to do simplification of a CFG.
4853 /// For example, it adjusts branches to branches to eliminate the extra hop,
4854 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4855 /// of the CFG. It returns true if a modification was made.
4857 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
4858 unsigned BonusInstThreshold, AssumptionCache *AC) {
4859 return SimplifyCFGOpt(TTI, BB->getModule()->getDataLayout(),
4860 BonusInstThreshold, AC).run(BB);