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 cases
87 // 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 switch
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 unsigned BonusInstThreshold;
114 const DataLayout *const DL;
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 SimplifyUnreachable(UnreachableInst *UI);
128 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
129 bool SimplifyIndirectBr(IndirectBrInst *IBI);
130 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder);
131 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder);
134 SimplifyCFGOpt(const TargetTransformInfo &TTI, unsigned BonusInstThreshold,
135 const DataLayout *DL, AssumptionCache *AC)
136 : TTI(TTI), BonusInstThreshold(BonusInstThreshold), DL(DL), AC(AC) {}
137 bool run(BasicBlock *BB);
141 /// SafeToMergeTerminators - Return true if it is safe to merge these two
142 /// 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 /// isProfitableToFoldUnconditional - Return true if it is safe and profitable
168 /// to merge these two terminator instructions together, where SI1 is an
169 /// unconditional branch. PhiNodes will store all PHI nodes in common
172 static bool isProfitableToFoldUnconditional(BranchInst *SI1,
175 SmallVectorImpl<PHINode*> &PhiNodes) {
176 if (SI1 == SI2) return false; // Can't merge with self!
177 assert(SI1->isUnconditional() && SI2->isConditional());
179 // We fold the unconditional branch if we can easily update all PHI nodes in
180 // common successors:
181 // 1> We have a constant incoming value for the conditional branch;
182 // 2> We have "Cond" as the incoming value for the unconditional branch;
183 // 3> SI2->getCondition() and Cond have same operands.
184 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
185 if (!Ci2) return false;
186 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
187 Cond->getOperand(1) == Ci2->getOperand(1)) &&
188 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
189 Cond->getOperand(1) == Ci2->getOperand(0)))
192 BasicBlock *SI1BB = SI1->getParent();
193 BasicBlock *SI2BB = SI2->getParent();
194 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
195 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
196 if (SI1Succs.count(*I))
197 for (BasicBlock::iterator BBI = (*I)->begin();
198 isa<PHINode>(BBI); ++BBI) {
199 PHINode *PN = cast<PHINode>(BBI);
200 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
201 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
203 PhiNodes.push_back(PN);
208 /// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will
209 /// now be entries in it from the 'NewPred' block. The values that will be
210 /// flowing into the PHI nodes will be the same as those coming in from
211 /// ExistPred, an existing predecessor of Succ.
212 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
213 BasicBlock *ExistPred) {
214 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
217 for (BasicBlock::iterator I = Succ->begin();
218 (PN = dyn_cast<PHINode>(I)); ++I)
219 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
222 /// ComputeSpeculationCost - Compute an abstract "cost" of speculating the
223 /// given instruction, which is assumed to be safe to speculate. TCC_Free means
224 /// cheap, TCC_Basic means less cheap, and TCC_Expensive means prohibitively
226 static unsigned ComputeSpeculationCost(const User *I, const DataLayout *DL,
227 const TargetTransformInfo &TTI) {
228 assert(isSafeToSpeculativelyExecute(I, DL) &&
229 "Instruction is not safe to speculatively execute!");
230 return TTI.getUserCost(I);
232 /// DominatesMergePoint - If we have a merge point of an "if condition" as
233 /// accepted above, return true if the specified value dominates the block. We
234 /// don't handle the true generality of domination here, just a special case
235 /// which works well enough for us.
237 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
238 /// see if V (which must be an instruction) and its recursive operands
239 /// that do not dominate BB have a combined cost lower than CostRemaining and
240 /// are non-trapping. If both are true, the instruction is inserted into the
241 /// set and true is returned.
243 /// The cost for most non-trapping instructions is defined as 1 except for
244 /// Select whose cost is 2.
246 /// After this function returns, CostRemaining is decreased by the cost of
247 /// V plus its non-dominating operands. If that cost is greater than
248 /// CostRemaining, false is returned and CostRemaining is undefined.
249 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
250 SmallPtrSetImpl<Instruction*> *AggressiveInsts,
251 unsigned &CostRemaining,
252 const DataLayout *DL,
253 const TargetTransformInfo &TTI) {
254 Instruction *I = dyn_cast<Instruction>(V);
256 // Non-instructions all dominate instructions, but not all constantexprs
257 // can be executed unconditionally.
258 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
263 BasicBlock *PBB = I->getParent();
265 // We don't want to allow weird loops that might have the "if condition" in
266 // the bottom of this block.
267 if (PBB == BB) return false;
269 // If this instruction is defined in a block that contains an unconditional
270 // branch to BB, then it must be in the 'conditional' part of the "if
271 // statement". If not, it definitely dominates the region.
272 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
273 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
276 // If we aren't allowing aggressive promotion anymore, then don't consider
277 // instructions in the 'if region'.
278 if (!AggressiveInsts) return false;
280 // If we have seen this instruction before, don't count it again.
281 if (AggressiveInsts->count(I)) return true;
283 // Okay, it looks like the instruction IS in the "condition". Check to
284 // see if it's a cheap instruction to unconditionally compute, and if it
285 // only uses stuff defined outside of the condition. If so, hoist it out.
286 if (!isSafeToSpeculativelyExecute(I, DL))
289 unsigned Cost = ComputeSpeculationCost(I, DL, TTI);
291 if (Cost > CostRemaining)
294 CostRemaining -= Cost;
296 // Okay, we can only really hoist these out if their operands do
297 // not take us over the cost threshold.
298 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
299 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, DL, TTI))
301 // Okay, it's safe to do this! Remember this instruction.
302 AggressiveInsts->insert(I);
306 /// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr
307 /// and PointerNullValue. Return NULL if value is not a constant int.
308 static ConstantInt *GetConstantInt(Value *V, const DataLayout *DL) {
309 // Normal constant int.
310 ConstantInt *CI = dyn_cast<ConstantInt>(V);
311 if (CI || !DL || !isa<Constant>(V) || !V->getType()->isPointerTy())
314 // This is some kind of pointer constant. Turn it into a pointer-sized
315 // ConstantInt if possible.
316 IntegerType *PtrTy = cast<IntegerType>(DL->getIntPtrType(V->getType()));
318 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
319 if (isa<ConstantPointerNull>(V))
320 return ConstantInt::get(PtrTy, 0);
322 // IntToPtr const int.
323 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
324 if (CE->getOpcode() == Instruction::IntToPtr)
325 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
326 // The constant is very likely to have the right type already.
327 if (CI->getType() == PtrTy)
330 return cast<ConstantInt>
331 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
338 /// Given a chain of or (||) or and (&&) comparison of a value against a
339 /// constant, this will try to recover the information required for a switch
341 /// It will depth-first traverse the chain of comparison, seeking for patterns
342 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
343 /// representing the different cases for the switch.
344 /// Note that if the chain is composed of '||' it will build the set of elements
345 /// that matches the comparisons (i.e. any of this value validate the chain)
346 /// while for a chain of '&&' it will build the set elements that make the test
348 struct ConstantComparesGatherer {
350 Value *CompValue; /// Value found for the switch comparison
351 Value *Extra; /// Extra clause to be checked before the switch
352 SmallVector<ConstantInt *, 8> Vals; /// Set of integers to match in switch
353 unsigned UsedICmps; /// Number of comparisons matched in the and/or chain
355 /// Construct and compute the result for the comparison instruction Cond
356 ConstantComparesGatherer(Instruction *Cond, const DataLayout *DL)
357 : CompValue(nullptr), Extra(nullptr), UsedICmps(0) {
362 ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
363 ConstantComparesGatherer &
364 operator=(const ConstantComparesGatherer &) = delete;
368 /// Try to set the current value used for the comparison, it succeeds only if
369 /// it wasn't set before or if the new value is the same as the old one
370 bool setValueOnce(Value *NewVal) {
371 if(CompValue && CompValue != NewVal) return false;
373 return (CompValue != nullptr);
376 /// Try to match Instruction "I" as a comparison against a constant and
377 /// populates the array Vals with the set of values that match (or do not
378 /// match depending on isEQ).
379 /// Return false on failure. On success, the Value the comparison matched
380 /// against is placed in CompValue.
381 /// If CompValue is already set, the function is expected to fail if a match
382 /// is found but the value compared to is different.
383 bool matchInstruction(Instruction *I, const DataLayout *DL, bool isEQ) {
384 // If this is an icmp against a constant, handle this as one of the cases.
387 if (!((ICI = dyn_cast<ICmpInst>(I)) &&
388 (C = GetConstantInt(I->getOperand(1), DL)))) {
395 // Pattern match a special case
396 // (x & ~2^x) == y --> x == y || x == y|2^x
397 // This undoes a transformation done by instcombine to fuse 2 compares.
398 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
399 if (match(ICI->getOperand(0),
400 m_And(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
401 APInt Not = ~RHSC->getValue();
402 if (Not.isPowerOf2()) {
403 // If we already have a value for the switch, it has to match!
404 if(!setValueOnce(RHSVal))
408 Vals.push_back(ConstantInt::get(C->getContext(),
409 C->getValue() | Not));
415 // If we already have a value for the switch, it has to match!
416 if(!setValueOnce(ICI->getOperand(0)))
421 return ICI->getOperand(0);
424 // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
425 ConstantRange Span = ConstantRange::makeICmpRegion(ICI->getPredicate(),
428 // Shift the range if the compare is fed by an add. This is the range
429 // compare idiom as emitted by instcombine.
430 Value *CandidateVal = I->getOperand(0);
431 if(match(I->getOperand(0), m_Add(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
432 Span = Span.subtract(RHSC->getValue());
433 CandidateVal = RHSVal;
436 // If this is an and/!= check, then we are looking to build the set of
437 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
440 Span = Span.inverse();
442 // If there are a ton of values, we don't want to make a ginormous switch.
443 if (Span.getSetSize().ugt(8) || Span.isEmptySet()) {
447 // If we already have a value for the switch, it has to match!
448 if(!setValueOnce(CandidateVal))
451 // Add all values from the range to the set
452 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
453 Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
460 /// gather - Given a potentially 'or'd or 'and'd together collection of icmp
461 /// eq/ne/lt/gt instructions that compare a value against a constant, extract
462 /// the value being compared, and stick the list constants into the Vals
464 /// One "Extra" case is allowed to differ from the other.
465 void gather(Value *V, const DataLayout *DL) {
466 Instruction *I = dyn_cast<Instruction>(V);
467 bool isEQ = (I->getOpcode() == Instruction::Or);
469 // Keep a stack (SmallVector for efficiency) for depth-first traversal
470 SmallVector<Value *, 8> DFT;
475 while(!DFT.empty()) {
476 V = DFT.pop_back_val();
478 if (Instruction *I = dyn_cast<Instruction>(V)) {
479 // If it is a || (or && depending on isEQ), process the operands.
480 if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
481 DFT.push_back(I->getOperand(1));
482 DFT.push_back(I->getOperand(0));
486 // Try to match the current instruction
487 if (matchInstruction(I, DL, isEQ))
488 // Match succeed, continue the loop
492 // One element of the sequence of || (or &&) could not be match as a
493 // comparison against the same value as the others.
494 // We allow only one "Extra" case to be checked before the switch
499 // Failed to parse a proper sequence, abort now
508 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
509 Instruction *Cond = nullptr;
510 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
511 Cond = dyn_cast<Instruction>(SI->getCondition());
512 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
513 if (BI->isConditional())
514 Cond = dyn_cast<Instruction>(BI->getCondition());
515 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
516 Cond = dyn_cast<Instruction>(IBI->getAddress());
519 TI->eraseFromParent();
520 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
523 /// isValueEqualityComparison - Return true if the specified terminator checks
524 /// to see if a value is equal to constant integer value.
525 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
527 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
528 // Do not permit merging of large switch instructions into their
529 // predecessors unless there is only one predecessor.
530 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()),
531 pred_end(SI->getParent())) <= 128)
532 CV = SI->getCondition();
533 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
534 if (BI->isConditional() && BI->getCondition()->hasOneUse())
535 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition()))
536 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
537 CV = ICI->getOperand(0);
539 // Unwrap any lossless ptrtoint cast.
541 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
542 Value *Ptr = PTII->getPointerOperand();
543 if (PTII->getType() == DL->getIntPtrType(Ptr->getType()))
550 /// GetValueEqualityComparisonCases - Given a value comparison instruction,
551 /// decode all of the 'cases' that it represents and return the 'default' block.
552 BasicBlock *SimplifyCFGOpt::
553 GetValueEqualityComparisonCases(TerminatorInst *TI,
554 std::vector<ValueEqualityComparisonCase>
556 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
557 Cases.reserve(SI->getNumCases());
558 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
559 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
560 i.getCaseSuccessor()));
561 return SI->getDefaultDest();
564 BranchInst *BI = cast<BranchInst>(TI);
565 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
566 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
567 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
570 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
574 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries
575 /// in the list that match the specified block.
576 static void EliminateBlockCases(BasicBlock *BB,
577 std::vector<ValueEqualityComparisonCase> &Cases) {
578 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
581 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as
584 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
585 std::vector<ValueEqualityComparisonCase > &C2) {
586 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
588 // Make V1 be smaller than V2.
589 if (V1->size() > V2->size())
592 if (V1->size() == 0) return false;
593 if (V1->size() == 1) {
595 ConstantInt *TheVal = (*V1)[0].Value;
596 for (unsigned i = 0, e = V2->size(); i != e; ++i)
597 if (TheVal == (*V2)[i].Value)
601 // Otherwise, just sort both lists and compare element by element.
602 array_pod_sort(V1->begin(), V1->end());
603 array_pod_sort(V2->begin(), V2->end());
604 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
605 while (i1 != e1 && i2 != e2) {
606 if ((*V1)[i1].Value == (*V2)[i2].Value)
608 if ((*V1)[i1].Value < (*V2)[i2].Value)
616 /// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a
617 /// terminator instruction and its block is known to only have a single
618 /// predecessor block, check to see if that predecessor is also a value
619 /// comparison with the same value, and if that comparison determines the
620 /// outcome of this comparison. If so, simplify TI. This does a very limited
621 /// form of jump threading.
622 bool SimplifyCFGOpt::
623 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
625 IRBuilder<> &Builder) {
626 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
627 if (!PredVal) return false; // Not a value comparison in predecessor.
629 Value *ThisVal = isValueEqualityComparison(TI);
630 assert(ThisVal && "This isn't a value comparison!!");
631 if (ThisVal != PredVal) return false; // Different predicates.
633 // TODO: Preserve branch weight metadata, similarly to how
634 // FoldValueComparisonIntoPredecessors preserves it.
636 // Find out information about when control will move from Pred to TI's block.
637 std::vector<ValueEqualityComparisonCase> PredCases;
638 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
640 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
642 // Find information about how control leaves this block.
643 std::vector<ValueEqualityComparisonCase> ThisCases;
644 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
645 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
647 // If TI's block is the default block from Pred's comparison, potentially
648 // simplify TI based on this knowledge.
649 if (PredDef == TI->getParent()) {
650 // If we are here, we know that the value is none of those cases listed in
651 // PredCases. If there are any cases in ThisCases that are in PredCases, we
653 if (!ValuesOverlap(PredCases, ThisCases))
656 if (isa<BranchInst>(TI)) {
657 // Okay, one of the successors of this condbr is dead. Convert it to a
659 assert(ThisCases.size() == 1 && "Branch can only have one case!");
660 // Insert the new branch.
661 Instruction *NI = Builder.CreateBr(ThisDef);
664 // Remove PHI node entries for the dead edge.
665 ThisCases[0].Dest->removePredecessor(TI->getParent());
667 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
668 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
670 EraseTerminatorInstAndDCECond(TI);
674 SwitchInst *SI = cast<SwitchInst>(TI);
675 // Okay, TI has cases that are statically dead, prune them away.
676 SmallPtrSet<Constant*, 16> DeadCases;
677 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
678 DeadCases.insert(PredCases[i].Value);
680 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
681 << "Through successor TI: " << *TI);
683 // Collect branch weights into a vector.
684 SmallVector<uint32_t, 8> Weights;
685 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
686 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
688 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
690 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
691 Weights.push_back(CI->getValue().getZExtValue());
693 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
695 if (DeadCases.count(i.getCaseValue())) {
697 std::swap(Weights[i.getCaseIndex()+1], Weights.back());
700 i.getCaseSuccessor()->removePredecessor(TI->getParent());
704 if (HasWeight && Weights.size() >= 2)
705 SI->setMetadata(LLVMContext::MD_prof,
706 MDBuilder(SI->getParent()->getContext()).
707 createBranchWeights(Weights));
709 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
713 // Otherwise, TI's block must correspond to some matched value. Find out
714 // which value (or set of values) this is.
715 ConstantInt *TIV = nullptr;
716 BasicBlock *TIBB = TI->getParent();
717 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
718 if (PredCases[i].Dest == TIBB) {
720 return false; // Cannot handle multiple values coming to this block.
721 TIV = PredCases[i].Value;
723 assert(TIV && "No edge from pred to succ?");
725 // Okay, we found the one constant that our value can be if we get into TI's
726 // BB. Find out which successor will unconditionally be branched to.
727 BasicBlock *TheRealDest = nullptr;
728 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
729 if (ThisCases[i].Value == TIV) {
730 TheRealDest = ThisCases[i].Dest;
734 // If not handled by any explicit cases, it is handled by the default case.
735 if (!TheRealDest) TheRealDest = ThisDef;
737 // Remove PHI node entries for dead edges.
738 BasicBlock *CheckEdge = TheRealDest;
739 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
740 if (*SI != CheckEdge)
741 (*SI)->removePredecessor(TIBB);
745 // Insert the new branch.
746 Instruction *NI = Builder.CreateBr(TheRealDest);
749 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
750 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
752 EraseTerminatorInstAndDCECond(TI);
757 /// ConstantIntOrdering - This class implements a stable ordering of constant
758 /// integers that does not depend on their address. This is important for
759 /// applications that sort ConstantInt's to ensure uniqueness.
760 struct ConstantIntOrdering {
761 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
762 return LHS->getValue().ult(RHS->getValue());
767 static int ConstantIntSortPredicate(ConstantInt *const *P1,
768 ConstantInt *const *P2) {
769 const ConstantInt *LHS = *P1;
770 const ConstantInt *RHS = *P2;
771 if (LHS->getValue().ult(RHS->getValue()))
773 if (LHS->getValue() == RHS->getValue())
778 static inline bool HasBranchWeights(const Instruction* I) {
779 MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
780 if (ProfMD && ProfMD->getOperand(0))
781 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
782 return MDS->getString().equals("branch_weights");
787 /// Get Weights of a given TerminatorInst, the default weight is at the front
788 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
790 static void GetBranchWeights(TerminatorInst *TI,
791 SmallVectorImpl<uint64_t> &Weights) {
792 MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
794 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
795 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
796 Weights.push_back(CI->getValue().getZExtValue());
799 // If TI is a conditional eq, the default case is the false case,
800 // and the corresponding branch-weight data is at index 2. We swap the
801 // default weight to be the first entry.
802 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
803 assert(Weights.size() == 2);
804 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
805 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
806 std::swap(Weights.front(), Weights.back());
810 /// Keep halving the weights until all can fit in uint32_t.
811 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
812 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
813 if (Max > UINT_MAX) {
814 unsigned Offset = 32 - countLeadingZeros(Max);
815 for (uint64_t &I : Weights)
820 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value
821 /// equality comparison instruction (either a switch or a branch on "X == c").
822 /// See if any of the predecessors of the terminator block are value comparisons
823 /// on the same value. If so, and if safe to do so, fold them together.
824 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
825 IRBuilder<> &Builder) {
826 BasicBlock *BB = TI->getParent();
827 Value *CV = isValueEqualityComparison(TI); // CondVal
828 assert(CV && "Not a comparison?");
829 bool Changed = false;
831 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
832 while (!Preds.empty()) {
833 BasicBlock *Pred = Preds.pop_back_val();
835 // See if the predecessor is a comparison with the same value.
836 TerminatorInst *PTI = Pred->getTerminator();
837 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
839 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
840 // Figure out which 'cases' to copy from SI to PSI.
841 std::vector<ValueEqualityComparisonCase> BBCases;
842 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
844 std::vector<ValueEqualityComparisonCase> PredCases;
845 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
847 // Based on whether the default edge from PTI goes to BB or not, fill in
848 // PredCases and PredDefault with the new switch cases we would like to
850 SmallVector<BasicBlock*, 8> NewSuccessors;
852 // Update the branch weight metadata along the way
853 SmallVector<uint64_t, 8> Weights;
854 bool PredHasWeights = HasBranchWeights(PTI);
855 bool SuccHasWeights = HasBranchWeights(TI);
857 if (PredHasWeights) {
858 GetBranchWeights(PTI, Weights);
859 // branch-weight metadata is inconsistent here.
860 if (Weights.size() != 1 + PredCases.size())
861 PredHasWeights = SuccHasWeights = false;
862 } else if (SuccHasWeights)
863 // If there are no predecessor weights but there are successor weights,
864 // populate Weights with 1, which will later be scaled to the sum of
865 // successor's weights
866 Weights.assign(1 + PredCases.size(), 1);
868 SmallVector<uint64_t, 8> SuccWeights;
869 if (SuccHasWeights) {
870 GetBranchWeights(TI, SuccWeights);
871 // branch-weight metadata is inconsistent here.
872 if (SuccWeights.size() != 1 + BBCases.size())
873 PredHasWeights = SuccHasWeights = false;
874 } else if (PredHasWeights)
875 SuccWeights.assign(1 + BBCases.size(), 1);
877 if (PredDefault == BB) {
878 // If this is the default destination from PTI, only the edges in TI
879 // that don't occur in PTI, or that branch to BB will be activated.
880 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
881 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
882 if (PredCases[i].Dest != BB)
883 PTIHandled.insert(PredCases[i].Value);
885 // The default destination is BB, we don't need explicit targets.
886 std::swap(PredCases[i], PredCases.back());
888 if (PredHasWeights || SuccHasWeights) {
889 // Increase weight for the default case.
890 Weights[0] += Weights[i+1];
891 std::swap(Weights[i+1], Weights.back());
895 PredCases.pop_back();
899 // Reconstruct the new switch statement we will be building.
900 if (PredDefault != BBDefault) {
901 PredDefault->removePredecessor(Pred);
902 PredDefault = BBDefault;
903 NewSuccessors.push_back(BBDefault);
906 unsigned CasesFromPred = Weights.size();
907 uint64_t ValidTotalSuccWeight = 0;
908 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
909 if (!PTIHandled.count(BBCases[i].Value) &&
910 BBCases[i].Dest != BBDefault) {
911 PredCases.push_back(BBCases[i]);
912 NewSuccessors.push_back(BBCases[i].Dest);
913 if (SuccHasWeights || PredHasWeights) {
914 // The default weight is at index 0, so weight for the ith case
915 // should be at index i+1. Scale the cases from successor by
916 // PredDefaultWeight (Weights[0]).
917 Weights.push_back(Weights[0] * SuccWeights[i+1]);
918 ValidTotalSuccWeight += SuccWeights[i+1];
922 if (SuccHasWeights || PredHasWeights) {
923 ValidTotalSuccWeight += SuccWeights[0];
924 // Scale the cases from predecessor by ValidTotalSuccWeight.
925 for (unsigned i = 1; i < CasesFromPred; ++i)
926 Weights[i] *= ValidTotalSuccWeight;
927 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
928 Weights[0] *= SuccWeights[0];
931 // If this is not the default destination from PSI, only the edges
932 // in SI that occur in PSI with a destination of BB will be
934 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
935 std::map<ConstantInt*, uint64_t> WeightsForHandled;
936 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
937 if (PredCases[i].Dest == BB) {
938 PTIHandled.insert(PredCases[i].Value);
940 if (PredHasWeights || SuccHasWeights) {
941 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
942 std::swap(Weights[i+1], Weights.back());
946 std::swap(PredCases[i], PredCases.back());
947 PredCases.pop_back();
951 // Okay, now we know which constants were sent to BB from the
952 // predecessor. Figure out where they will all go now.
953 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
954 if (PTIHandled.count(BBCases[i].Value)) {
955 // If this is one we are capable of getting...
956 if (PredHasWeights || SuccHasWeights)
957 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
958 PredCases.push_back(BBCases[i]);
959 NewSuccessors.push_back(BBCases[i].Dest);
960 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
963 // If there are any constants vectored to BB that TI doesn't handle,
964 // they must go to the default destination of TI.
965 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
967 E = PTIHandled.end(); I != E; ++I) {
968 if (PredHasWeights || SuccHasWeights)
969 Weights.push_back(WeightsForHandled[*I]);
970 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
971 NewSuccessors.push_back(BBDefault);
975 // Okay, at this point, we know which new successor Pred will get. Make
976 // sure we update the number of entries in the PHI nodes for these
978 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i)
979 AddPredecessorToBlock(NewSuccessors[i], Pred, BB);
981 Builder.SetInsertPoint(PTI);
982 // Convert pointer to int before we switch.
983 if (CV->getType()->isPointerTy()) {
984 assert(DL && "Cannot switch on pointer without DataLayout");
985 CV = Builder.CreatePtrToInt(CV, DL->getIntPtrType(CV->getType()),
989 // Now that the successors are updated, create the new Switch instruction.
990 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
992 NewSI->setDebugLoc(PTI->getDebugLoc());
993 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
994 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest);
996 if (PredHasWeights || SuccHasWeights) {
997 // Halve the weights if any of them cannot fit in an uint32_t
1000 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1002 NewSI->setMetadata(LLVMContext::MD_prof,
1003 MDBuilder(BB->getContext()).
1004 createBranchWeights(MDWeights));
1007 EraseTerminatorInstAndDCECond(PTI);
1009 // Okay, last check. If BB is still a successor of PSI, then we must
1010 // have an infinite loop case. If so, add an infinitely looping block
1011 // to handle the case to preserve the behavior of the code.
1012 BasicBlock *InfLoopBlock = nullptr;
1013 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1014 if (NewSI->getSuccessor(i) == BB) {
1015 if (!InfLoopBlock) {
1016 // Insert it at the end of the function, because it's either code,
1017 // or it won't matter if it's hot. :)
1018 InfLoopBlock = BasicBlock::Create(BB->getContext(),
1019 "infloop", BB->getParent());
1020 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1022 NewSI->setSuccessor(i, InfLoopBlock);
1031 // isSafeToHoistInvoke - If we would need to insert a select that uses the
1032 // value of this invoke (comments in HoistThenElseCodeToIf explain why we
1033 // would need to do this), we can't hoist the invoke, as there is nowhere
1034 // to put the select in this case.
1035 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1036 Instruction *I1, Instruction *I2) {
1037 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1039 for (BasicBlock::iterator BBI = SI->begin();
1040 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1041 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1042 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1043 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
1051 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I);
1053 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
1054 /// BB2, hoist any common code in the two blocks up into the branch block. The
1055 /// caller of this function guarantees that BI's block dominates BB1 and BB2.
1056 static bool HoistThenElseCodeToIf(BranchInst *BI, const DataLayout *DL,
1057 const TargetTransformInfo &TTI) {
1058 // This does very trivial matching, with limited scanning, to find identical
1059 // instructions in the two blocks. In particular, we don't want to get into
1060 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1061 // such, we currently just scan for obviously identical instructions in an
1063 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1064 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1066 BasicBlock::iterator BB1_Itr = BB1->begin();
1067 BasicBlock::iterator BB2_Itr = BB2->begin();
1069 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++;
1070 // Skip debug info if it is not identical.
1071 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1072 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1073 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1074 while (isa<DbgInfoIntrinsic>(I1))
1076 while (isa<DbgInfoIntrinsic>(I2))
1079 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1080 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1083 BasicBlock *BIParent = BI->getParent();
1085 bool Changed = false;
1087 // If we are hoisting the terminator instruction, don't move one (making a
1088 // broken BB), instead clone it, and remove BI.
1089 if (isa<TerminatorInst>(I1))
1090 goto HoistTerminator;
1092 if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1095 // For a normal instruction, we just move one to right before the branch,
1096 // then replace all uses of the other with the first. Finally, we remove
1097 // the now redundant second instruction.
1098 BIParent->getInstList().splice(BI, BB1->getInstList(), I1);
1099 if (!I2->use_empty())
1100 I2->replaceAllUsesWith(I1);
1101 I1->intersectOptionalDataWith(I2);
1102 unsigned KnownIDs[] = {
1103 LLVMContext::MD_tbaa,
1104 LLVMContext::MD_range,
1105 LLVMContext::MD_fpmath,
1106 LLVMContext::MD_invariant_load,
1107 LLVMContext::MD_nonnull
1109 combineMetadata(I1, I2, KnownIDs);
1110 I2->eraseFromParent();
1115 // Skip debug info if it is not identical.
1116 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1117 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1118 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1119 while (isa<DbgInfoIntrinsic>(I1))
1121 while (isa<DbgInfoIntrinsic>(I2))
1124 } while (I1->isIdenticalToWhenDefined(I2));
1129 // It may not be possible to hoist an invoke.
1130 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1133 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1135 for (BasicBlock::iterator BBI = SI->begin();
1136 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1137 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1138 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1142 // Check for passingValueIsAlwaysUndefined here because we would rather
1143 // eliminate undefined control flow then converting it to a select.
1144 if (passingValueIsAlwaysUndefined(BB1V, PN) ||
1145 passingValueIsAlwaysUndefined(BB2V, PN))
1148 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V, DL))
1150 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V, DL))
1155 // Okay, it is safe to hoist the terminator.
1156 Instruction *NT = I1->clone();
1157 BIParent->getInstList().insert(BI, NT);
1158 if (!NT->getType()->isVoidTy()) {
1159 I1->replaceAllUsesWith(NT);
1160 I2->replaceAllUsesWith(NT);
1164 IRBuilder<true, NoFolder> Builder(NT);
1165 // Hoisting one of the terminators from our successor is a great thing.
1166 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1167 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1168 // nodes, so we insert select instruction to compute the final result.
1169 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1170 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1172 for (BasicBlock::iterator BBI = SI->begin();
1173 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1174 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1175 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1176 if (BB1V == BB2V) continue;
1178 // These values do not agree. Insert a select instruction before NT
1179 // that determines the right value.
1180 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1182 SI = cast<SelectInst>
1183 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1184 BB1V->getName()+"."+BB2V->getName()));
1186 // Make the PHI node use the select for all incoming values for BB1/BB2
1187 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1188 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1189 PN->setIncomingValue(i, SI);
1193 // Update any PHI nodes in our new successors.
1194 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1195 AddPredecessorToBlock(*SI, BIParent, BB1);
1197 EraseTerminatorInstAndDCECond(BI);
1201 /// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd,
1202 /// check whether BBEnd has only two predecessors and the other predecessor
1203 /// ends with an unconditional branch. If it is true, sink any common code
1204 /// in the two predecessors to BBEnd.
1205 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1206 assert(BI1->isUnconditional());
1207 BasicBlock *BB1 = BI1->getParent();
1208 BasicBlock *BBEnd = BI1->getSuccessor(0);
1210 // Check that BBEnd has two predecessors and the other predecessor ends with
1211 // an unconditional branch.
1212 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1213 BasicBlock *Pred0 = *PI++;
1214 if (PI == PE) // Only one predecessor.
1216 BasicBlock *Pred1 = *PI++;
1217 if (PI != PE) // More than two predecessors.
1219 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1220 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1221 if (!BI2 || !BI2->isUnconditional())
1224 // Gather the PHI nodes in BBEnd.
1225 SmallDenseMap<std::pair<Value *, Value *>, PHINode *> JointValueMap;
1226 Instruction *FirstNonPhiInBBEnd = nullptr;
1227 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end(); I != E; ++I) {
1228 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1229 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1230 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1231 JointValueMap[std::make_pair(BB1V, BB2V)] = PN;
1233 FirstNonPhiInBBEnd = &*I;
1237 if (!FirstNonPhiInBBEnd)
1240 // This does very trivial matching, with limited scanning, to find identical
1241 // instructions in the two blocks. We scan backward for obviously identical
1242 // instructions in an identical order.
1243 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1244 RE1 = BB1->getInstList().rend(),
1245 RI2 = BB2->getInstList().rbegin(),
1246 RE2 = BB2->getInstList().rend();
1248 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1251 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1254 // Skip the unconditional branches.
1258 bool Changed = false;
1259 while (RI1 != RE1 && RI2 != RE2) {
1261 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1264 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1268 Instruction *I1 = &*RI1, *I2 = &*RI2;
1269 auto InstPair = std::make_pair(I1, I2);
1270 // I1 and I2 should have a single use in the same PHI node, and they
1271 // perform the same operation.
1272 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1273 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1274 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1275 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) ||
1276 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1277 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1278 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1279 !I1->hasOneUse() || !I2->hasOneUse() ||
1280 !JointValueMap.count(InstPair))
1283 // Check whether we should swap the operands of ICmpInst.
1284 // TODO: Add support of communativity.
1285 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1286 bool SwapOpnds = false;
1287 if (ICmp1 && ICmp2 &&
1288 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1289 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1290 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1291 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1292 ICmp2->swapOperands();
1295 if (!I1->isSameOperationAs(I2)) {
1297 ICmp2->swapOperands();
1301 // The operands should be either the same or they need to be generated
1302 // with a PHI node after sinking. We only handle the case where there is
1303 // a single pair of different operands.
1304 Value *DifferentOp1 = nullptr, *DifferentOp2 = nullptr;
1305 unsigned Op1Idx = ~0U;
1306 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1307 if (I1->getOperand(I) == I2->getOperand(I))
1309 // Early exit if we have more-than one pair of different operands or if
1310 // we need a PHI node to replace a constant.
1311 if (Op1Idx != ~0U ||
1312 isa<Constant>(I1->getOperand(I)) ||
1313 isa<Constant>(I2->getOperand(I))) {
1314 // If we can't sink the instructions, undo the swapping.
1316 ICmp2->swapOperands();
1319 DifferentOp1 = I1->getOperand(I);
1321 DifferentOp2 = I2->getOperand(I);
1324 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n");
1325 DEBUG(dbgs() << " " << *I2 << "\n");
1327 // We insert the pair of different operands to JointValueMap and
1328 // remove (I1, I2) from JointValueMap.
1329 if (Op1Idx != ~0U) {
1330 auto &NewPN = JointValueMap[std::make_pair(DifferentOp1, DifferentOp2)];
1333 PHINode::Create(DifferentOp1->getType(), 2,
1334 DifferentOp1->getName() + ".sink", BBEnd->begin());
1335 NewPN->addIncoming(DifferentOp1, BB1);
1336 NewPN->addIncoming(DifferentOp2, BB2);
1337 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1339 // I1 should use NewPN instead of DifferentOp1.
1340 I1->setOperand(Op1Idx, NewPN);
1342 PHINode *OldPN = JointValueMap[InstPair];
1343 JointValueMap.erase(InstPair);
1345 // We need to update RE1 and RE2 if we are going to sink the first
1346 // instruction in the basic block down.
1347 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1348 // Sink the instruction.
1349 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1);
1350 if (!OldPN->use_empty())
1351 OldPN->replaceAllUsesWith(I1);
1352 OldPN->eraseFromParent();
1354 if (!I2->use_empty())
1355 I2->replaceAllUsesWith(I1);
1356 I1->intersectOptionalDataWith(I2);
1357 // TODO: Use combineMetadata here to preserve what metadata we can
1358 // (analogous to the hoisting case above).
1359 I2->eraseFromParent();
1362 RE1 = BB1->getInstList().rend();
1364 RE2 = BB2->getInstList().rend();
1365 FirstNonPhiInBBEnd = I1;
1372 /// \brief Determine if we can hoist sink a sole store instruction out of a
1373 /// conditional block.
1375 /// We are looking for code like the following:
1377 /// store i32 %add, i32* %arrayidx2
1378 /// ... // No other stores or function calls (we could be calling a memory
1379 /// ... // function).
1380 /// %cmp = icmp ult %x, %y
1381 /// br i1 %cmp, label %EndBB, label %ThenBB
1383 /// store i32 %add5, i32* %arrayidx2
1387 /// We are going to transform this into:
1389 /// store i32 %add, i32* %arrayidx2
1391 /// %cmp = icmp ult %x, %y
1392 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1393 /// store i32 %add.add5, i32* %arrayidx2
1396 /// \return The pointer to the value of the previous store if the store can be
1397 /// hoisted into the predecessor block. 0 otherwise.
1398 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1399 BasicBlock *StoreBB, BasicBlock *EndBB) {
1400 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1404 // Volatile or atomic.
1405 if (!StoreToHoist->isSimple())
1408 Value *StorePtr = StoreToHoist->getPointerOperand();
1410 // Look for a store to the same pointer in BrBB.
1411 unsigned MaxNumInstToLookAt = 10;
1412 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(),
1413 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) {
1414 Instruction *CurI = &*RI;
1416 // Could be calling an instruction that effects memory like free().
1417 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI))
1420 StoreInst *SI = dyn_cast<StoreInst>(CurI);
1421 // Found the previous store make sure it stores to the same location.
1422 if (SI && SI->getPointerOperand() == StorePtr)
1423 // Found the previous store, return its value operand.
1424 return SI->getValueOperand();
1426 return nullptr; // Unknown store.
1432 /// \brief Speculate a conditional basic block flattening the CFG.
1434 /// Note that this is a very risky transform currently. Speculating
1435 /// instructions like this is most often not desirable. Instead, there is an MI
1436 /// pass which can do it with full awareness of the resource constraints.
1437 /// However, some cases are "obvious" and we should do directly. An example of
1438 /// this is speculating a single, reasonably cheap instruction.
1440 /// There is only one distinct advantage to flattening the CFG at the IR level:
1441 /// it makes very common but simplistic optimizations such as are common in
1442 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1443 /// modeling their effects with easier to reason about SSA value graphs.
1446 /// An illustration of this transform is turning this IR:
1449 /// %cmp = icmp ult %x, %y
1450 /// br i1 %cmp, label %EndBB, label %ThenBB
1452 /// %sub = sub %x, %y
1455 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1462 /// %cmp = icmp ult %x, %y
1463 /// %sub = sub %x, %y
1464 /// %cond = select i1 %cmp, 0, %sub
1468 /// \returns true if the conditional block is removed.
1469 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1470 const DataLayout *DL,
1471 const TargetTransformInfo &TTI) {
1472 // Be conservative for now. FP select instruction can often be expensive.
1473 Value *BrCond = BI->getCondition();
1474 if (isa<FCmpInst>(BrCond))
1477 BasicBlock *BB = BI->getParent();
1478 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1480 // If ThenBB is actually on the false edge of the conditional branch, remember
1481 // to swap the select operands later.
1482 bool Invert = false;
1483 if (ThenBB != BI->getSuccessor(0)) {
1484 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1487 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1489 // Keep a count of how many times instructions are used within CondBB when
1490 // they are candidates for sinking into CondBB. Specifically:
1491 // - They are defined in BB, and
1492 // - They have no side effects, and
1493 // - All of their uses are in CondBB.
1494 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1496 unsigned SpeculationCost = 0;
1497 Value *SpeculatedStoreValue = nullptr;
1498 StoreInst *SpeculatedStore = nullptr;
1499 for (BasicBlock::iterator BBI = ThenBB->begin(),
1500 BBE = std::prev(ThenBB->end());
1501 BBI != BBE; ++BBI) {
1502 Instruction *I = BBI;
1504 if (isa<DbgInfoIntrinsic>(I))
1507 // Only speculatively execution a single instruction (not counting the
1508 // terminator) for now.
1510 if (SpeculationCost > 1)
1513 // Don't hoist the instruction if it's unsafe or expensive.
1514 if (!isSafeToSpeculativelyExecute(I, DL) &&
1515 !(HoistCondStores &&
1516 (SpeculatedStoreValue = isSafeToSpeculateStore(I, BB, ThenBB,
1519 if (!SpeculatedStoreValue &&
1520 ComputeSpeculationCost(I, DL, TTI) > PHINodeFoldingThreshold *
1521 TargetTransformInfo::TCC_Basic)
1524 // Store the store speculation candidate.
1525 if (SpeculatedStoreValue)
1526 SpeculatedStore = cast<StoreInst>(I);
1528 // Do not hoist the instruction if any of its operands are defined but not
1529 // used in BB. The transformation will prevent the operand from
1530 // being sunk into the use block.
1531 for (User::op_iterator i = I->op_begin(), e = I->op_end();
1533 Instruction *OpI = dyn_cast<Instruction>(*i);
1534 if (!OpI || OpI->getParent() != BB ||
1535 OpI->mayHaveSideEffects())
1536 continue; // Not a candidate for sinking.
1538 ++SinkCandidateUseCounts[OpI];
1542 // Consider any sink candidates which are only used in CondBB as costs for
1543 // speculation. Note, while we iterate over a DenseMap here, we are summing
1544 // and so iteration order isn't significant.
1545 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I =
1546 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end();
1548 if (I->first->getNumUses() == I->second) {
1550 if (SpeculationCost > 1)
1554 // Check that the PHI nodes can be converted to selects.
1555 bool HaveRewritablePHIs = false;
1556 for (BasicBlock::iterator I = EndBB->begin();
1557 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1558 Value *OrigV = PN->getIncomingValueForBlock(BB);
1559 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1561 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1562 // Skip PHIs which are trivial.
1566 // Don't convert to selects if we could remove undefined behavior instead.
1567 if (passingValueIsAlwaysUndefined(OrigV, PN) ||
1568 passingValueIsAlwaysUndefined(ThenV, PN))
1571 HaveRewritablePHIs = true;
1572 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1573 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1574 if (!OrigCE && !ThenCE)
1575 continue; // Known safe and cheap.
1577 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE, DL)) ||
1578 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE, DL)))
1580 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, DL, TTI) : 0;
1581 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, DL, TTI) : 0;
1582 unsigned MaxCost = 2 * PHINodeFoldingThreshold *
1583 TargetTransformInfo::TCC_Basic;
1584 if (OrigCost + ThenCost > MaxCost)
1587 // Account for the cost of an unfolded ConstantExpr which could end up
1588 // getting expanded into Instructions.
1589 // FIXME: This doesn't account for how many operations are combined in the
1590 // constant expression.
1592 if (SpeculationCost > 1)
1596 // If there are no PHIs to process, bail early. This helps ensure idempotence
1598 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
1601 // If we get here, we can hoist the instruction and if-convert.
1602 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1604 // Insert a select of the value of the speculated store.
1605 if (SpeculatedStoreValue) {
1606 IRBuilder<true, NoFolder> Builder(BI);
1607 Value *TrueV = SpeculatedStore->getValueOperand();
1608 Value *FalseV = SpeculatedStoreValue;
1610 std::swap(TrueV, FalseV);
1611 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() +
1612 "." + FalseV->getName());
1613 SpeculatedStore->setOperand(0, S);
1616 // Hoist the instructions.
1617 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(),
1618 std::prev(ThenBB->end()));
1620 // Insert selects and rewrite the PHI operands.
1621 IRBuilder<true, NoFolder> Builder(BI);
1622 for (BasicBlock::iterator I = EndBB->begin();
1623 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1624 unsigned OrigI = PN->getBasicBlockIndex(BB);
1625 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1626 Value *OrigV = PN->getIncomingValue(OrigI);
1627 Value *ThenV = PN->getIncomingValue(ThenI);
1629 // Skip PHIs which are trivial.
1633 // Create a select whose true value is the speculatively executed value and
1634 // false value is the preexisting value. Swap them if the branch
1635 // destinations were inverted.
1636 Value *TrueV = ThenV, *FalseV = OrigV;
1638 std::swap(TrueV, FalseV);
1639 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV,
1640 TrueV->getName() + "." + FalseV->getName());
1641 PN->setIncomingValue(OrigI, V);
1642 PN->setIncomingValue(ThenI, V);
1649 /// \returns True if this block contains a CallInst with the NoDuplicate
1651 static bool HasNoDuplicateCall(const BasicBlock *BB) {
1652 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1653 const CallInst *CI = dyn_cast<CallInst>(I);
1656 if (CI->cannotDuplicate())
1662 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1663 /// across this block.
1664 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1665 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1668 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1669 if (isa<DbgInfoIntrinsic>(BBI))
1671 if (Size > 10) return false; // Don't clone large BB's.
1674 // We can only support instructions that do not define values that are
1675 // live outside of the current basic block.
1676 for (User *U : BBI->users()) {
1677 Instruction *UI = cast<Instruction>(U);
1678 if (UI->getParent() != BB || isa<PHINode>(UI)) return false;
1681 // Looks ok, continue checking.
1687 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1688 /// that is defined in the same block as the branch and if any PHI entries are
1689 /// constants, thread edges corresponding to that entry to be branches to their
1690 /// ultimate destination.
1691 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *DL) {
1692 BasicBlock *BB = BI->getParent();
1693 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1694 // NOTE: we currently cannot transform this case if the PHI node is used
1695 // outside of the block.
1696 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1699 // Degenerate case of a single entry PHI.
1700 if (PN->getNumIncomingValues() == 1) {
1701 FoldSingleEntryPHINodes(PN->getParent());
1705 // Now we know that this block has multiple preds and two succs.
1706 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1708 if (HasNoDuplicateCall(BB)) return false;
1710 // Okay, this is a simple enough basic block. See if any phi values are
1712 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1713 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1714 if (!CB || !CB->getType()->isIntegerTy(1)) continue;
1716 // Okay, we now know that all edges from PredBB should be revectored to
1717 // branch to RealDest.
1718 BasicBlock *PredBB = PN->getIncomingBlock(i);
1719 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1721 if (RealDest == BB) continue; // Skip self loops.
1722 // Skip if the predecessor's terminator is an indirect branch.
1723 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1725 // The dest block might have PHI nodes, other predecessors and other
1726 // difficult cases. Instead of being smart about this, just insert a new
1727 // block that jumps to the destination block, effectively splitting
1728 // the edge we are about to create.
1729 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1730 RealDest->getName()+".critedge",
1731 RealDest->getParent(), RealDest);
1732 BranchInst::Create(RealDest, EdgeBB);
1734 // Update PHI nodes.
1735 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1737 // BB may have instructions that are being threaded over. Clone these
1738 // instructions into EdgeBB. We know that there will be no uses of the
1739 // cloned instructions outside of EdgeBB.
1740 BasicBlock::iterator InsertPt = EdgeBB->begin();
1741 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1742 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1743 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1744 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1747 // Clone the instruction.
1748 Instruction *N = BBI->clone();
1749 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1751 // Update operands due to translation.
1752 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1754 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1755 if (PI != TranslateMap.end())
1759 // Check for trivial simplification.
1760 if (Value *V = SimplifyInstruction(N, DL)) {
1761 TranslateMap[BBI] = V;
1762 delete N; // Instruction folded away, don't need actual inst
1764 // Insert the new instruction into its new home.
1765 EdgeBB->getInstList().insert(InsertPt, N);
1766 if (!BBI->use_empty())
1767 TranslateMap[BBI] = N;
1771 // Loop over all of the edges from PredBB to BB, changing them to branch
1772 // to EdgeBB instead.
1773 TerminatorInst *PredBBTI = PredBB->getTerminator();
1774 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1775 if (PredBBTI->getSuccessor(i) == BB) {
1776 BB->removePredecessor(PredBB);
1777 PredBBTI->setSuccessor(i, EdgeBB);
1780 // Recurse, simplifying any other constants.
1781 return FoldCondBranchOnPHI(BI, DL) | true;
1787 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1788 /// PHI node, see if we can eliminate it.
1789 static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *DL,
1790 const TargetTransformInfo &TTI) {
1791 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1792 // statement", which has a very simple dominance structure. Basically, we
1793 // are trying to find the condition that is being branched on, which
1794 // subsequently causes this merge to happen. We really want control
1795 // dependence information for this check, but simplifycfg can't keep it up
1796 // to date, and this catches most of the cases we care about anyway.
1797 BasicBlock *BB = PN->getParent();
1798 BasicBlock *IfTrue, *IfFalse;
1799 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1801 // Don't bother if the branch will be constant folded trivially.
1802 isa<ConstantInt>(IfCond))
1805 // Okay, we found that we can merge this two-entry phi node into a select.
1806 // Doing so would require us to fold *all* two entry phi nodes in this block.
1807 // At some point this becomes non-profitable (particularly if the target
1808 // doesn't support cmov's). Only do this transformation if there are two or
1809 // fewer PHI nodes in this block.
1810 unsigned NumPhis = 0;
1811 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1815 // Loop over the PHI's seeing if we can promote them all to select
1816 // instructions. While we are at it, keep track of the instructions
1817 // that need to be moved to the dominating block.
1818 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1819 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1820 MaxCostVal1 = PHINodeFoldingThreshold;
1821 MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
1822 MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
1824 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1825 PHINode *PN = cast<PHINode>(II++);
1826 if (Value *V = SimplifyInstruction(PN, DL)) {
1827 PN->replaceAllUsesWith(V);
1828 PN->eraseFromParent();
1832 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1833 MaxCostVal0, DL, TTI) ||
1834 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1835 MaxCostVal1, DL, TTI))
1839 // If we folded the first phi, PN dangles at this point. Refresh it. If
1840 // we ran out of PHIs then we simplified them all.
1841 PN = dyn_cast<PHINode>(BB->begin());
1842 if (!PN) return true;
1844 // Don't fold i1 branches on PHIs which contain binary operators. These can
1845 // often be turned into switches and other things.
1846 if (PN->getType()->isIntegerTy(1) &&
1847 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1848 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1849 isa<BinaryOperator>(IfCond)))
1852 // If we all PHI nodes are promotable, check to make sure that all
1853 // instructions in the predecessor blocks can be promoted as well. If
1854 // not, we won't be able to get rid of the control flow, so it's not
1855 // worth promoting to select instructions.
1856 BasicBlock *DomBlock = nullptr;
1857 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1858 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1859 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1862 DomBlock = *pred_begin(IfBlock1);
1863 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1864 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1865 // This is not an aggressive instruction that we can promote.
1866 // Because of this, we won't be able to get rid of the control
1867 // flow, so the xform is not worth it.
1872 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1875 DomBlock = *pred_begin(IfBlock2);
1876 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1877 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1878 // This is not an aggressive instruction that we can promote.
1879 // Because of this, we won't be able to get rid of the control
1880 // flow, so the xform is not worth it.
1885 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1886 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1888 // If we can still promote the PHI nodes after this gauntlet of tests,
1889 // do all of the PHI's now.
1890 Instruction *InsertPt = DomBlock->getTerminator();
1891 IRBuilder<true, NoFolder> Builder(InsertPt);
1893 // Move all 'aggressive' instructions, which are defined in the
1894 // conditional parts of the if's up to the dominating block.
1896 DomBlock->getInstList().splice(InsertPt,
1897 IfBlock1->getInstList(), IfBlock1->begin(),
1898 IfBlock1->getTerminator());
1900 DomBlock->getInstList().splice(InsertPt,
1901 IfBlock2->getInstList(), IfBlock2->begin(),
1902 IfBlock2->getTerminator());
1904 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1905 // Change the PHI node into a select instruction.
1906 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1907 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1910 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1911 PN->replaceAllUsesWith(NV);
1913 PN->eraseFromParent();
1916 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1917 // has been flattened. Change DomBlock to jump directly to our new block to
1918 // avoid other simplifycfg's kicking in on the diamond.
1919 TerminatorInst *OldTI = DomBlock->getTerminator();
1920 Builder.SetInsertPoint(OldTI);
1921 Builder.CreateBr(BB);
1922 OldTI->eraseFromParent();
1926 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1927 /// to two returning blocks, try to merge them together into one return,
1928 /// introducing a select if the return values disagree.
1929 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1930 IRBuilder<> &Builder) {
1931 assert(BI->isConditional() && "Must be a conditional branch");
1932 BasicBlock *TrueSucc = BI->getSuccessor(0);
1933 BasicBlock *FalseSucc = BI->getSuccessor(1);
1934 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1935 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1937 // Check to ensure both blocks are empty (just a return) or optionally empty
1938 // with PHI nodes. If there are other instructions, merging would cause extra
1939 // computation on one path or the other.
1940 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1942 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1945 Builder.SetInsertPoint(BI);
1946 // Okay, we found a branch that is going to two return nodes. If
1947 // there is no return value for this function, just change the
1948 // branch into a return.
1949 if (FalseRet->getNumOperands() == 0) {
1950 TrueSucc->removePredecessor(BI->getParent());
1951 FalseSucc->removePredecessor(BI->getParent());
1952 Builder.CreateRetVoid();
1953 EraseTerminatorInstAndDCECond(BI);
1957 // Otherwise, figure out what the true and false return values are
1958 // so we can insert a new select instruction.
1959 Value *TrueValue = TrueRet->getReturnValue();
1960 Value *FalseValue = FalseRet->getReturnValue();
1962 // Unwrap any PHI nodes in the return blocks.
1963 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1964 if (TVPN->getParent() == TrueSucc)
1965 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1966 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1967 if (FVPN->getParent() == FalseSucc)
1968 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1970 // In order for this transformation to be safe, we must be able to
1971 // unconditionally execute both operands to the return. This is
1972 // normally the case, but we could have a potentially-trapping
1973 // constant expression that prevents this transformation from being
1975 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1978 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1982 // Okay, we collected all the mapped values and checked them for sanity, and
1983 // defined to really do this transformation. First, update the CFG.
1984 TrueSucc->removePredecessor(BI->getParent());
1985 FalseSucc->removePredecessor(BI->getParent());
1987 // Insert select instructions where needed.
1988 Value *BrCond = BI->getCondition();
1990 // Insert a select if the results differ.
1991 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1992 } else if (isa<UndefValue>(TrueValue)) {
1993 TrueValue = FalseValue;
1995 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1996 FalseValue, "retval");
2000 Value *RI = !TrueValue ?
2001 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2005 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2006 << "\n " << *BI << "NewRet = " << *RI
2007 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
2009 EraseTerminatorInstAndDCECond(BI);
2014 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
2015 /// probabilities of the branch taking each edge. Fills in the two APInt
2016 /// parameters and return true, or returns false if no or invalid metadata was
2018 static bool ExtractBranchMetadata(BranchInst *BI,
2019 uint64_t &ProbTrue, uint64_t &ProbFalse) {
2020 assert(BI->isConditional() &&
2021 "Looking for probabilities on unconditional branch?");
2022 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
2023 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
2024 ConstantInt *CITrue =
2025 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
2026 ConstantInt *CIFalse =
2027 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
2028 if (!CITrue || !CIFalse) return false;
2029 ProbTrue = CITrue->getValue().getZExtValue();
2030 ProbFalse = CIFalse->getValue().getZExtValue();
2034 /// checkCSEInPredecessor - Return true if the given instruction is available
2035 /// in its predecessor block. If yes, the instruction will be removed.
2037 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
2038 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2040 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
2041 Instruction *PBI = &*I;
2042 // Check whether Inst and PBI generate the same value.
2043 if (Inst->isIdenticalTo(PBI)) {
2044 Inst->replaceAllUsesWith(PBI);
2045 Inst->eraseFromParent();
2052 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
2053 /// predecessor branches to us and one of our successors, fold the block into
2054 /// the predecessor and use logical operations to pick the right destination.
2055 bool llvm::FoldBranchToCommonDest(BranchInst *BI, const DataLayout *DL,
2056 unsigned BonusInstThreshold) {
2057 BasicBlock *BB = BI->getParent();
2059 Instruction *Cond = nullptr;
2060 if (BI->isConditional())
2061 Cond = dyn_cast<Instruction>(BI->getCondition());
2063 // For unconditional branch, check for a simple CFG pattern, where
2064 // BB has a single predecessor and BB's successor is also its predecessor's
2065 // successor. If such pattern exisits, check for CSE between BB and its
2067 if (BasicBlock *PB = BB->getSinglePredecessor())
2068 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2069 if (PBI->isConditional() &&
2070 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2071 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2072 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
2074 Instruction *Curr = I++;
2075 if (isa<CmpInst>(Curr)) {
2079 // Quit if we can't remove this instruction.
2080 if (!checkCSEInPredecessor(Curr, PB))
2089 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2090 Cond->getParent() != BB || !Cond->hasOneUse())
2093 // Make sure the instruction after the condition is the cond branch.
2094 BasicBlock::iterator CondIt = Cond; ++CondIt;
2096 // Ignore dbg intrinsics.
2097 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
2102 // Only allow this transformation if computing the condition doesn't involve
2103 // too many instructions and these involved instructions can be executed
2104 // unconditionally. We denote all involved instructions except the condition
2105 // as "bonus instructions", and only allow this transformation when the
2106 // number of the bonus instructions does not exceed a certain threshold.
2107 unsigned NumBonusInsts = 0;
2108 for (auto I = BB->begin(); Cond != I; ++I) {
2109 // Ignore dbg intrinsics.
2110 if (isa<DbgInfoIntrinsic>(I))
2112 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(I, DL))
2114 // I has only one use and can be executed unconditionally.
2115 Instruction *User = dyn_cast<Instruction>(I->user_back());
2116 if (User == nullptr || User->getParent() != BB)
2118 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2119 // to use any other instruction, User must be an instruction between next(I)
2122 // Early exits once we reach the limit.
2123 if (NumBonusInsts > BonusInstThreshold)
2127 // Cond is known to be a compare or binary operator. Check to make sure that
2128 // neither operand is a potentially-trapping constant expression.
2129 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2132 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2136 // Finally, don't infinitely unroll conditional loops.
2137 BasicBlock *TrueDest = BI->getSuccessor(0);
2138 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2139 if (TrueDest == BB || FalseDest == BB)
2142 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2143 BasicBlock *PredBlock = *PI;
2144 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2146 // Check that we have two conditional branches. If there is a PHI node in
2147 // the common successor, verify that the same value flows in from both
2149 SmallVector<PHINode*, 4> PHIs;
2150 if (!PBI || PBI->isUnconditional() ||
2151 (BI->isConditional() &&
2152 !SafeToMergeTerminators(BI, PBI)) ||
2153 (!BI->isConditional() &&
2154 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2157 // Determine if the two branches share a common destination.
2158 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2159 bool InvertPredCond = false;
2161 if (BI->isConditional()) {
2162 if (PBI->getSuccessor(0) == TrueDest)
2163 Opc = Instruction::Or;
2164 else if (PBI->getSuccessor(1) == FalseDest)
2165 Opc = Instruction::And;
2166 else if (PBI->getSuccessor(0) == FalseDest)
2167 Opc = Instruction::And, InvertPredCond = true;
2168 else if (PBI->getSuccessor(1) == TrueDest)
2169 Opc = Instruction::Or, InvertPredCond = true;
2173 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2177 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2178 IRBuilder<> Builder(PBI);
2180 // If we need to invert the condition in the pred block to match, do so now.
2181 if (InvertPredCond) {
2182 Value *NewCond = PBI->getCondition();
2184 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2185 CmpInst *CI = cast<CmpInst>(NewCond);
2186 CI->setPredicate(CI->getInversePredicate());
2188 NewCond = Builder.CreateNot(NewCond,
2189 PBI->getCondition()->getName()+".not");
2192 PBI->setCondition(NewCond);
2193 PBI->swapSuccessors();
2196 // If we have bonus instructions, clone them into the predecessor block.
2197 // Note that there may be mutliple predecessor blocks, so we cannot move
2198 // bonus instructions to a predecessor block.
2199 ValueToValueMapTy VMap; // maps original values to cloned values
2200 // We already make sure Cond is the last instruction before BI. Therefore,
2201 // every instructions before Cond other than DbgInfoIntrinsic are bonus
2203 for (auto BonusInst = BB->begin(); Cond != BonusInst; ++BonusInst) {
2204 if (isa<DbgInfoIntrinsic>(BonusInst))
2206 Instruction *NewBonusInst = BonusInst->clone();
2207 RemapInstruction(NewBonusInst, VMap,
2208 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2209 VMap[BonusInst] = NewBonusInst;
2211 // If we moved a load, we cannot any longer claim any knowledge about
2212 // its potential value. The previous information might have been valid
2213 // only given the branch precondition.
2214 // For an analogous reason, we must also drop all the metadata whose
2215 // semantics we don't understand.
2216 NewBonusInst->dropUnknownMetadata(LLVMContext::MD_dbg);
2218 PredBlock->getInstList().insert(PBI, NewBonusInst);
2219 NewBonusInst->takeName(BonusInst);
2220 BonusInst->setName(BonusInst->getName() + ".old");
2223 // Clone Cond into the predecessor basic block, and or/and the
2224 // two conditions together.
2225 Instruction *New = Cond->clone();
2226 RemapInstruction(New, VMap,
2227 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2228 PredBlock->getInstList().insert(PBI, New);
2229 New->takeName(Cond);
2230 Cond->setName(New->getName() + ".old");
2232 if (BI->isConditional()) {
2233 Instruction *NewCond =
2234 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2236 PBI->setCondition(NewCond);
2238 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2239 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2241 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2243 SmallVector<uint64_t, 8> NewWeights;
2245 if (PBI->getSuccessor(0) == BB) {
2246 if (PredHasWeights && SuccHasWeights) {
2247 // PBI: br i1 %x, BB, FalseDest
2248 // BI: br i1 %y, TrueDest, FalseDest
2249 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2250 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2251 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2252 // TrueWeight for PBI * FalseWeight for BI.
2253 // We assume that total weights of a BranchInst can fit into 32 bits.
2254 // Therefore, we will not have overflow using 64-bit arithmetic.
2255 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2256 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2258 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2259 PBI->setSuccessor(0, TrueDest);
2261 if (PBI->getSuccessor(1) == BB) {
2262 if (PredHasWeights && SuccHasWeights) {
2263 // PBI: br i1 %x, TrueDest, BB
2264 // BI: br i1 %y, TrueDest, FalseDest
2265 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2266 // FalseWeight for PBI * TrueWeight for BI.
2267 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2268 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2269 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2270 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2272 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2273 PBI->setSuccessor(1, FalseDest);
2275 if (NewWeights.size() == 2) {
2276 // Halve the weights if any of them cannot fit in an uint32_t
2277 FitWeights(NewWeights);
2279 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2280 PBI->setMetadata(LLVMContext::MD_prof,
2281 MDBuilder(BI->getContext()).
2282 createBranchWeights(MDWeights));
2284 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2286 // Update PHI nodes in the common successors.
2287 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2288 ConstantInt *PBI_C = cast<ConstantInt>(
2289 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2290 assert(PBI_C->getType()->isIntegerTy(1));
2291 Instruction *MergedCond = nullptr;
2292 if (PBI->getSuccessor(0) == TrueDest) {
2293 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2294 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2295 // is false: !PBI_Cond and BI_Value
2296 Instruction *NotCond =
2297 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2300 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2305 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2306 PBI->getCondition(), MergedCond,
2309 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2310 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2311 // is false: PBI_Cond and BI_Value
2313 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2314 PBI->getCondition(), New,
2316 if (PBI_C->isOne()) {
2317 Instruction *NotCond =
2318 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2321 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2322 NotCond, MergedCond,
2327 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2330 // Change PBI from Conditional to Unconditional.
2331 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2332 EraseTerminatorInstAndDCECond(PBI);
2336 // TODO: If BB is reachable from all paths through PredBlock, then we
2337 // could replace PBI's branch probabilities with BI's.
2339 // Copy any debug value intrinsics into the end of PredBlock.
2340 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2341 if (isa<DbgInfoIntrinsic>(*I))
2342 I->clone()->insertBefore(PBI);
2349 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2350 /// predecessor of another block, this function tries to simplify it. We know
2351 /// that PBI and BI are both conditional branches, and BI is in one of the
2352 /// successor blocks of PBI - PBI branches to BI.
2353 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2354 assert(PBI->isConditional() && BI->isConditional());
2355 BasicBlock *BB = BI->getParent();
2357 // If this block ends with a branch instruction, and if there is a
2358 // predecessor that ends on a branch of the same condition, make
2359 // this conditional branch redundant.
2360 if (PBI->getCondition() == BI->getCondition() &&
2361 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2362 // Okay, the outcome of this conditional branch is statically
2363 // knowable. If this block had a single pred, handle specially.
2364 if (BB->getSinglePredecessor()) {
2365 // Turn this into a branch on constant.
2366 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2367 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2369 return true; // Nuke the branch on constant.
2372 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2373 // in the constant and simplify the block result. Subsequent passes of
2374 // simplifycfg will thread the block.
2375 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2376 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2377 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2378 std::distance(PB, PE),
2379 BI->getCondition()->getName() + ".pr",
2381 // Okay, we're going to insert the PHI node. Since PBI is not the only
2382 // predecessor, compute the PHI'd conditional value for all of the preds.
2383 // Any predecessor where the condition is not computable we keep symbolic.
2384 for (pred_iterator PI = PB; PI != PE; ++PI) {
2385 BasicBlock *P = *PI;
2386 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2387 PBI != BI && PBI->isConditional() &&
2388 PBI->getCondition() == BI->getCondition() &&
2389 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2390 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2391 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2394 NewPN->addIncoming(BI->getCondition(), P);
2398 BI->setCondition(NewPN);
2403 // If this is a conditional branch in an empty block, and if any
2404 // predecessors are a conditional branch to one of our destinations,
2405 // fold the conditions into logical ops and one cond br.
2406 BasicBlock::iterator BBI = BB->begin();
2407 // Ignore dbg intrinsics.
2408 while (isa<DbgInfoIntrinsic>(BBI))
2414 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2419 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2421 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2422 PBIOp = 0, BIOp = 1;
2423 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2424 PBIOp = 1, BIOp = 0;
2425 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2430 // Check to make sure that the other destination of this branch
2431 // isn't BB itself. If so, this is an infinite loop that will
2432 // keep getting unwound.
2433 if (PBI->getSuccessor(PBIOp) == BB)
2436 // Do not perform this transformation if it would require
2437 // insertion of a large number of select instructions. For targets
2438 // without predication/cmovs, this is a big pessimization.
2440 // Also do not perform this transformation if any phi node in the common
2441 // destination block can trap when reached by BB or PBB (PR17073). In that
2442 // case, it would be unsafe to hoist the operation into a select instruction.
2444 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2445 unsigned NumPhis = 0;
2446 for (BasicBlock::iterator II = CommonDest->begin();
2447 isa<PHINode>(II); ++II, ++NumPhis) {
2448 if (NumPhis > 2) // Disable this xform.
2451 PHINode *PN = cast<PHINode>(II);
2452 Value *BIV = PN->getIncomingValueForBlock(BB);
2453 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
2457 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2458 Value *PBIV = PN->getIncomingValue(PBBIdx);
2459 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
2464 // Finally, if everything is ok, fold the branches to logical ops.
2465 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2467 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2468 << "AND: " << *BI->getParent());
2471 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2472 // branch in it, where one edge (OtherDest) goes back to itself but the other
2473 // exits. We don't *know* that the program avoids the infinite loop
2474 // (even though that seems likely). If we do this xform naively, we'll end up
2475 // recursively unpeeling the loop. Since we know that (after the xform is
2476 // done) that the block *is* infinite if reached, we just make it an obviously
2477 // infinite loop with no cond branch.
2478 if (OtherDest == BB) {
2479 // Insert it at the end of the function, because it's either code,
2480 // or it won't matter if it's hot. :)
2481 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2482 "infloop", BB->getParent());
2483 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2484 OtherDest = InfLoopBlock;
2487 DEBUG(dbgs() << *PBI->getParent()->getParent());
2489 // BI may have other predecessors. Because of this, we leave
2490 // it alone, but modify PBI.
2492 // Make sure we get to CommonDest on True&True directions.
2493 Value *PBICond = PBI->getCondition();
2494 IRBuilder<true, NoFolder> Builder(PBI);
2496 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2498 Value *BICond = BI->getCondition();
2500 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2502 // Merge the conditions.
2503 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2505 // Modify PBI to branch on the new condition to the new dests.
2506 PBI->setCondition(Cond);
2507 PBI->setSuccessor(0, CommonDest);
2508 PBI->setSuccessor(1, OtherDest);
2510 // Update branch weight for PBI.
2511 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2512 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2514 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2516 if (PredHasWeights && SuccHasWeights) {
2517 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2518 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2519 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2520 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2521 // The weight to CommonDest should be PredCommon * SuccTotal +
2522 // PredOther * SuccCommon.
2523 // The weight to OtherDest should be PredOther * SuccOther.
2524 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
2525 PredOther * SuccCommon,
2526 PredOther * SuccOther};
2527 // Halve the weights if any of them cannot fit in an uint32_t
2528 FitWeights(NewWeights);
2530 PBI->setMetadata(LLVMContext::MD_prof,
2531 MDBuilder(BI->getContext())
2532 .createBranchWeights(NewWeights[0], NewWeights[1]));
2535 // OtherDest may have phi nodes. If so, add an entry from PBI's
2536 // block that are identical to the entries for BI's block.
2537 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2539 // We know that the CommonDest already had an edge from PBI to
2540 // it. If it has PHIs though, the PHIs may have different
2541 // entries for BB and PBI's BB. If so, insert a select to make
2544 for (BasicBlock::iterator II = CommonDest->begin();
2545 (PN = dyn_cast<PHINode>(II)); ++II) {
2546 Value *BIV = PN->getIncomingValueForBlock(BB);
2547 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2548 Value *PBIV = PN->getIncomingValue(PBBIdx);
2550 // Insert a select in PBI to pick the right value.
2551 Value *NV = cast<SelectInst>
2552 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2553 PN->setIncomingValue(PBBIdx, NV);
2557 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2558 DEBUG(dbgs() << *PBI->getParent()->getParent());
2560 // This basic block is probably dead. We know it has at least
2561 // one fewer predecessor.
2565 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2566 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2567 // Takes care of updating the successors and removing the old terminator.
2568 // Also makes sure not to introduce new successors by assuming that edges to
2569 // non-successor TrueBBs and FalseBBs aren't reachable.
2570 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2571 BasicBlock *TrueBB, BasicBlock *FalseBB,
2572 uint32_t TrueWeight,
2573 uint32_t FalseWeight){
2574 // Remove any superfluous successor edges from the CFG.
2575 // First, figure out which successors to preserve.
2576 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2578 BasicBlock *KeepEdge1 = TrueBB;
2579 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
2581 // Then remove the rest.
2582 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2583 BasicBlock *Succ = OldTerm->getSuccessor(I);
2584 // Make sure only to keep exactly one copy of each edge.
2585 if (Succ == KeepEdge1)
2586 KeepEdge1 = nullptr;
2587 else if (Succ == KeepEdge2)
2588 KeepEdge2 = nullptr;
2590 Succ->removePredecessor(OldTerm->getParent());
2593 IRBuilder<> Builder(OldTerm);
2594 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2596 // Insert an appropriate new terminator.
2597 if (!KeepEdge1 && !KeepEdge2) {
2598 if (TrueBB == FalseBB)
2599 // We were only looking for one successor, and it was present.
2600 // Create an unconditional branch to it.
2601 Builder.CreateBr(TrueBB);
2603 // We found both of the successors we were looking for.
2604 // Create a conditional branch sharing the condition of the select.
2605 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2606 if (TrueWeight != FalseWeight)
2607 NewBI->setMetadata(LLVMContext::MD_prof,
2608 MDBuilder(OldTerm->getContext()).
2609 createBranchWeights(TrueWeight, FalseWeight));
2611 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2612 // Neither of the selected blocks were successors, so this
2613 // terminator must be unreachable.
2614 new UnreachableInst(OldTerm->getContext(), OldTerm);
2616 // One of the selected values was a successor, but the other wasn't.
2617 // Insert an unconditional branch to the one that was found;
2618 // the edge to the one that wasn't must be unreachable.
2620 // Only TrueBB was found.
2621 Builder.CreateBr(TrueBB);
2623 // Only FalseBB was found.
2624 Builder.CreateBr(FalseBB);
2627 EraseTerminatorInstAndDCECond(OldTerm);
2631 // SimplifySwitchOnSelect - Replaces
2632 // (switch (select cond, X, Y)) on constant X, Y
2633 // with a branch - conditional if X and Y lead to distinct BBs,
2634 // unconditional otherwise.
2635 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2636 // Check for constant integer values in the select.
2637 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2638 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2639 if (!TrueVal || !FalseVal)
2642 // Find the relevant condition and destinations.
2643 Value *Condition = Select->getCondition();
2644 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2645 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2647 // Get weight for TrueBB and FalseBB.
2648 uint32_t TrueWeight = 0, FalseWeight = 0;
2649 SmallVector<uint64_t, 8> Weights;
2650 bool HasWeights = HasBranchWeights(SI);
2652 GetBranchWeights(SI, Weights);
2653 if (Weights.size() == 1 + SI->getNumCases()) {
2654 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2655 getSuccessorIndex()];
2656 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2657 getSuccessorIndex()];
2661 // Perform the actual simplification.
2662 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2663 TrueWeight, FalseWeight);
2666 // SimplifyIndirectBrOnSelect - Replaces
2667 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2668 // blockaddress(@fn, BlockB)))
2670 // (br cond, BlockA, BlockB).
2671 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2672 // Check that both operands of the select are block addresses.
2673 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2674 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2678 // Extract the actual blocks.
2679 BasicBlock *TrueBB = TBA->getBasicBlock();
2680 BasicBlock *FalseBB = FBA->getBasicBlock();
2682 // Perform the actual simplification.
2683 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2687 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2688 /// instruction (a seteq/setne with a constant) as the only instruction in a
2689 /// block that ends with an uncond branch. We are looking for a very specific
2690 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2691 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2692 /// default value goes to an uncond block with a seteq in it, we get something
2695 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2697 /// %tmp = icmp eq i8 %A, 92
2700 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2702 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2703 /// the PHI, merging the third icmp into the switch.
2704 static bool TryToSimplifyUncondBranchWithICmpInIt(
2705 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI,
2706 unsigned BonusInstThreshold, const DataLayout *DL, AssumptionCache *AC) {
2707 BasicBlock *BB = ICI->getParent();
2709 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2711 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2713 Value *V = ICI->getOperand(0);
2714 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2716 // The pattern we're looking for is where our only predecessor is a switch on
2717 // 'V' and this block is the default case for the switch. In this case we can
2718 // fold the compared value into the switch to simplify things.
2719 BasicBlock *Pred = BB->getSinglePredecessor();
2720 if (!Pred || !isa<SwitchInst>(Pred->getTerminator())) return false;
2722 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2723 if (SI->getCondition() != V)
2726 // If BB is reachable on a non-default case, then we simply know the value of
2727 // V in this block. Substitute it and constant fold the icmp instruction
2729 if (SI->getDefaultDest() != BB) {
2730 ConstantInt *VVal = SI->findCaseDest(BB);
2731 assert(VVal && "Should have a unique destination value");
2732 ICI->setOperand(0, VVal);
2734 if (Value *V = SimplifyInstruction(ICI, DL)) {
2735 ICI->replaceAllUsesWith(V);
2736 ICI->eraseFromParent();
2738 // BB is now empty, so it is likely to simplify away.
2739 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AC) | true;
2742 // Ok, the block is reachable from the default dest. If the constant we're
2743 // comparing exists in one of the other edges, then we can constant fold ICI
2745 if (SI->findCaseValue(Cst) != SI->case_default()) {
2747 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2748 V = ConstantInt::getFalse(BB->getContext());
2750 V = ConstantInt::getTrue(BB->getContext());
2752 ICI->replaceAllUsesWith(V);
2753 ICI->eraseFromParent();
2754 // BB is now empty, so it is likely to simplify away.
2755 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AC) | true;
2758 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2760 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2761 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
2762 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
2763 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2766 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2768 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2769 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2771 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2772 std::swap(DefaultCst, NewCst);
2774 // Replace ICI (which is used by the PHI for the default value) with true or
2775 // false depending on if it is EQ or NE.
2776 ICI->replaceAllUsesWith(DefaultCst);
2777 ICI->eraseFromParent();
2779 // Okay, the switch goes to this block on a default value. Add an edge from
2780 // the switch to the merge point on the compared value.
2781 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2782 BB->getParent(), BB);
2783 SmallVector<uint64_t, 8> Weights;
2784 bool HasWeights = HasBranchWeights(SI);
2786 GetBranchWeights(SI, Weights);
2787 if (Weights.size() == 1 + SI->getNumCases()) {
2788 // Split weight for default case to case for "Cst".
2789 Weights[0] = (Weights[0]+1) >> 1;
2790 Weights.push_back(Weights[0]);
2792 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2793 SI->setMetadata(LLVMContext::MD_prof,
2794 MDBuilder(SI->getContext()).
2795 createBranchWeights(MDWeights));
2798 SI->addCase(Cst, NewBB);
2800 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2801 Builder.SetInsertPoint(NewBB);
2802 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2803 Builder.CreateBr(SuccBlock);
2804 PHIUse->addIncoming(NewCst, NewBB);
2808 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2809 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2810 /// fold it into a switch instruction if so.
2811 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *DL,
2812 IRBuilder<> &Builder) {
2813 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2814 if (!Cond) return false;
2816 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2817 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2818 // 'setne's and'ed together, collect them.
2820 // Try to gather values from a chain of and/or to be turned into a switch
2821 ConstantComparesGatherer ConstantCompare(Cond, DL);
2822 // Unpack the result
2823 SmallVectorImpl<ConstantInt*> &Values = ConstantCompare.Vals;
2824 Value *CompVal = ConstantCompare.CompValue;
2825 unsigned UsedICmps = ConstantCompare.UsedICmps;
2826 Value *ExtraCase = ConstantCompare.Extra;
2828 // If we didn't have a multiply compared value, fail.
2829 if (!CompVal) return false;
2831 // Avoid turning single icmps into a switch.
2835 bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
2837 // There might be duplicate constants in the list, which the switch
2838 // instruction can't handle, remove them now.
2839 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2840 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2842 // If Extra was used, we require at least two switch values to do the
2843 // transformation. A switch with one value is just an cond branch.
2844 if (ExtraCase && Values.size() < 2) return false;
2846 // TODO: Preserve branch weight metadata, similarly to how
2847 // FoldValueComparisonIntoPredecessors preserves it.
2849 // Figure out which block is which destination.
2850 BasicBlock *DefaultBB = BI->getSuccessor(1);
2851 BasicBlock *EdgeBB = BI->getSuccessor(0);
2852 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2854 BasicBlock *BB = BI->getParent();
2856 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2857 << " cases into SWITCH. BB is:\n" << *BB);
2859 // If there are any extra values that couldn't be folded into the switch
2860 // then we evaluate them with an explicit branch first. Split the block
2861 // right before the condbr to handle it.
2863 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2864 // Remove the uncond branch added to the old block.
2865 TerminatorInst *OldTI = BB->getTerminator();
2866 Builder.SetInsertPoint(OldTI);
2869 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2871 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2873 OldTI->eraseFromParent();
2875 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2876 // for the edge we just added.
2877 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2879 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2880 << "\nEXTRABB = " << *BB);
2884 Builder.SetInsertPoint(BI);
2885 // Convert pointer to int before we switch.
2886 if (CompVal->getType()->isPointerTy()) {
2887 assert(DL && "Cannot switch on pointer without DataLayout");
2888 CompVal = Builder.CreatePtrToInt(CompVal,
2889 DL->getIntPtrType(CompVal->getType()),
2893 // Create the new switch instruction now.
2894 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2896 // Add all of the 'cases' to the switch instruction.
2897 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2898 New->addCase(Values[i], EdgeBB);
2900 // We added edges from PI to the EdgeBB. As such, if there were any
2901 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2902 // the number of edges added.
2903 for (BasicBlock::iterator BBI = EdgeBB->begin();
2904 isa<PHINode>(BBI); ++BBI) {
2905 PHINode *PN = cast<PHINode>(BBI);
2906 Value *InVal = PN->getIncomingValueForBlock(BB);
2907 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2908 PN->addIncoming(InVal, BB);
2911 // Erase the old branch instruction.
2912 EraseTerminatorInstAndDCECond(BI);
2914 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2918 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2919 // If this is a trivial landing pad that just continues unwinding the caught
2920 // exception then zap the landing pad, turning its invokes into calls.
2921 BasicBlock *BB = RI->getParent();
2922 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2923 if (RI->getValue() != LPInst)
2924 // Not a landing pad, or the resume is not unwinding the exception that
2925 // caused control to branch here.
2928 // Check that there are no other instructions except for debug intrinsics.
2929 BasicBlock::iterator I = LPInst, E = RI;
2931 if (!isa<DbgInfoIntrinsic>(I))
2934 // Turn all invokes that unwind here into calls and delete the basic block.
2935 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2936 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2937 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2938 // Insert a call instruction before the invoke.
2939 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2941 Call->setCallingConv(II->getCallingConv());
2942 Call->setAttributes(II->getAttributes());
2943 Call->setDebugLoc(II->getDebugLoc());
2945 // Anything that used the value produced by the invoke instruction now uses
2946 // the value produced by the call instruction. Note that we do this even
2947 // for void functions and calls with no uses so that the callgraph edge is
2949 II->replaceAllUsesWith(Call);
2950 BB->removePredecessor(II->getParent());
2952 // Insert a branch to the normal destination right before the invoke.
2953 BranchInst::Create(II->getNormalDest(), II);
2955 // Finally, delete the invoke instruction!
2956 II->eraseFromParent();
2959 // The landingpad is now unreachable. Zap it.
2960 BB->eraseFromParent();
2964 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2965 BasicBlock *BB = RI->getParent();
2966 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2968 // Find predecessors that end with branches.
2969 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2970 SmallVector<BranchInst*, 8> CondBranchPreds;
2971 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2972 BasicBlock *P = *PI;
2973 TerminatorInst *PTI = P->getTerminator();
2974 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2975 if (BI->isUnconditional())
2976 UncondBranchPreds.push_back(P);
2978 CondBranchPreds.push_back(BI);
2982 // If we found some, do the transformation!
2983 if (!UncondBranchPreds.empty() && DupRet) {
2984 while (!UncondBranchPreds.empty()) {
2985 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
2986 DEBUG(dbgs() << "FOLDING: " << *BB
2987 << "INTO UNCOND BRANCH PRED: " << *Pred);
2988 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
2991 // If we eliminated all predecessors of the block, delete the block now.
2993 // We know there are no successors, so just nuke the block.
2994 BB->eraseFromParent();
2999 // Check out all of the conditional branches going to this return
3000 // instruction. If any of them just select between returns, change the
3001 // branch itself into a select/return pair.
3002 while (!CondBranchPreds.empty()) {
3003 BranchInst *BI = CondBranchPreds.pop_back_val();
3005 // Check to see if the non-BB successor is also a return block.
3006 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
3007 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
3008 SimplifyCondBranchToTwoReturns(BI, Builder))
3014 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
3015 BasicBlock *BB = UI->getParent();
3017 bool Changed = false;
3019 // If there are any instructions immediately before the unreachable that can
3020 // be removed, do so.
3021 while (UI != BB->begin()) {
3022 BasicBlock::iterator BBI = UI;
3024 // Do not delete instructions that can have side effects which might cause
3025 // the unreachable to not be reachable; specifically, calls and volatile
3026 // operations may have this effect.
3027 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
3029 if (BBI->mayHaveSideEffects()) {
3030 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
3031 if (SI->isVolatile())
3033 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
3034 if (LI->isVolatile())
3036 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
3037 if (RMWI->isVolatile())
3039 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
3040 if (CXI->isVolatile())
3042 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
3043 !isa<LandingPadInst>(BBI)) {
3046 // Note that deleting LandingPad's here is in fact okay, although it
3047 // involves a bit of subtle reasoning. If this inst is a LandingPad,
3048 // all the predecessors of this block will be the unwind edges of Invokes,
3049 // and we can therefore guarantee this block will be erased.
3052 // Delete this instruction (any uses are guaranteed to be dead)
3053 if (!BBI->use_empty())
3054 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
3055 BBI->eraseFromParent();
3059 // If the unreachable instruction is the first in the block, take a gander
3060 // at all of the predecessors of this instruction, and simplify them.
3061 if (&BB->front() != UI) return Changed;
3063 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
3064 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
3065 TerminatorInst *TI = Preds[i]->getTerminator();
3066 IRBuilder<> Builder(TI);
3067 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3068 if (BI->isUnconditional()) {
3069 if (BI->getSuccessor(0) == BB) {
3070 new UnreachableInst(TI->getContext(), TI);
3071 TI->eraseFromParent();
3075 if (BI->getSuccessor(0) == BB) {
3076 Builder.CreateBr(BI->getSuccessor(1));
3077 EraseTerminatorInstAndDCECond(BI);
3078 } else if (BI->getSuccessor(1) == BB) {
3079 Builder.CreateBr(BI->getSuccessor(0));
3080 EraseTerminatorInstAndDCECond(BI);
3084 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3085 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3087 if (i.getCaseSuccessor() == BB) {
3088 BB->removePredecessor(SI->getParent());
3093 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
3094 if (II->getUnwindDest() == BB) {
3095 // Convert the invoke to a call instruction. This would be a good
3096 // place to note that the call does not throw though.
3097 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
3098 II->removeFromParent(); // Take out of symbol table
3100 // Insert the call now...
3101 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
3102 Builder.SetInsertPoint(BI);
3103 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
3104 Args, II->getName());
3105 CI->setCallingConv(II->getCallingConv());
3106 CI->setAttributes(II->getAttributes());
3107 // If the invoke produced a value, the call does now instead.
3108 II->replaceAllUsesWith(CI);
3115 // If this block is now dead, remove it.
3116 if (pred_empty(BB) &&
3117 BB != &BB->getParent()->getEntryBlock()) {
3118 // We know there are no successors, so just nuke the block.
3119 BB->eraseFromParent();
3126 static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) {
3127 assert(Cases.size() >= 1);
3129 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3130 for (size_t I = 1, E = Cases.size(); I != E; ++I) {
3131 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
3137 /// Turn a switch with two reachable destinations into an integer range
3138 /// comparison and branch.
3139 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3140 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3143 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
3145 // Partition the cases into two sets with different destinations.
3146 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
3147 BasicBlock *DestB = nullptr;
3148 SmallVector <ConstantInt *, 16> CasesA;
3149 SmallVector <ConstantInt *, 16> CasesB;
3151 for (SwitchInst::CaseIt I : SI->cases()) {
3152 BasicBlock *Dest = I.getCaseSuccessor();
3153 if (!DestA) DestA = Dest;
3154 if (Dest == DestA) {
3155 CasesA.push_back(I.getCaseValue());
3158 if (!DestB) DestB = Dest;
3159 if (Dest == DestB) {
3160 CasesB.push_back(I.getCaseValue());
3163 return false; // More than two destinations.
3166 assert(DestA && DestB && "Single-destination switch should have been folded.");
3167 assert(DestA != DestB);
3168 assert(DestB != SI->getDefaultDest());
3169 assert(!CasesB.empty() && "There must be non-default cases.");
3170 assert(!CasesA.empty() || HasDefault);
3172 // Figure out if one of the sets of cases form a contiguous range.
3173 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
3174 BasicBlock *ContiguousDest = nullptr;
3175 BasicBlock *OtherDest = nullptr;
3176 if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
3177 ContiguousCases = &CasesA;
3178 ContiguousDest = DestA;
3180 } else if (CasesAreContiguous(CasesB)) {
3181 ContiguousCases = &CasesB;
3182 ContiguousDest = DestB;
3187 // Start building the compare and branch.
3189 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
3190 Constant *NumCases = ConstantInt::get(Offset->getType(), ContiguousCases->size());
3192 Value *Sub = SI->getCondition();
3193 if (!Offset->isNullValue())
3194 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
3197 // If NumCases overflowed, then all possible values jump to the successor.
3198 if (NumCases->isNullValue() && !ContiguousCases->empty())
3199 Cmp = ConstantInt::getTrue(SI->getContext());
3201 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3202 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
3204 // Update weight for the newly-created conditional branch.
3205 if (HasBranchWeights(SI)) {
3206 SmallVector<uint64_t, 8> Weights;
3207 GetBranchWeights(SI, Weights);
3208 if (Weights.size() == 1 + SI->getNumCases()) {
3209 uint64_t TrueWeight = 0;
3210 uint64_t FalseWeight = 0;
3211 for (size_t I = 0, E = Weights.size(); I != E; ++I) {
3212 if (SI->getSuccessor(I) == ContiguousDest)
3213 TrueWeight += Weights[I];
3215 FalseWeight += Weights[I];
3217 while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
3221 NewBI->setMetadata(LLVMContext::MD_prof,
3222 MDBuilder(SI->getContext()).createBranchWeights(
3223 (uint32_t)TrueWeight, (uint32_t)FalseWeight));
3227 // Prune obsolete incoming values off the successors' PHI nodes.
3228 for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
3229 unsigned PreviousEdges = ContiguousCases->size();
3230 if (ContiguousDest == SI->getDefaultDest()) ++PreviousEdges;
3231 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
3232 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3234 for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
3235 unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
3236 if (OtherDest == SI->getDefaultDest()) ++PreviousEdges;
3237 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
3238 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3242 SI->eraseFromParent();
3247 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
3248 /// and use it to remove dead cases.
3249 static bool EliminateDeadSwitchCases(SwitchInst *SI, const DataLayout *DL,
3250 AssumptionCache *AC) {
3251 Value *Cond = SI->getCondition();
3252 unsigned Bits = Cond->getType()->getIntegerBitWidth();
3253 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3254 computeKnownBits(Cond, KnownZero, KnownOne, DL, 0, AC, SI);
3256 // Gather dead cases.
3257 SmallVector<ConstantInt*, 8> DeadCases;
3258 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3259 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3260 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3261 DeadCases.push_back(I.getCaseValue());
3262 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3263 << I.getCaseValue() << "' is dead.\n");
3267 SmallVector<uint64_t, 8> Weights;
3268 bool HasWeight = HasBranchWeights(SI);
3270 GetBranchWeights(SI, Weights);
3271 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3274 // Remove dead cases from the switch.
3275 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3276 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3277 assert(Case != SI->case_default() &&
3278 "Case was not found. Probably mistake in DeadCases forming.");
3280 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3284 // Prune unused values from PHI nodes.
3285 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3286 SI->removeCase(Case);
3288 if (HasWeight && Weights.size() >= 2) {
3289 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3290 SI->setMetadata(LLVMContext::MD_prof,
3291 MDBuilder(SI->getParent()->getContext()).
3292 createBranchWeights(MDWeights));
3295 return !DeadCases.empty();
3298 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
3299 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3300 /// by an unconditional branch), look at the phi node for BB in the successor
3301 /// block and see if the incoming value is equal to CaseValue. If so, return
3302 /// the phi node, and set PhiIndex to BB's index in the phi node.
3303 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3306 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3307 return nullptr; // BB must be empty to be a candidate for simplification.
3308 if (!BB->getSinglePredecessor())
3309 return nullptr; // BB must be dominated by the switch.
3311 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3312 if (!Branch || !Branch->isUnconditional())
3313 return nullptr; // Terminator must be unconditional branch.
3315 BasicBlock *Succ = Branch->getSuccessor(0);
3317 BasicBlock::iterator I = Succ->begin();
3318 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3319 int Idx = PHI->getBasicBlockIndex(BB);
3320 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3322 Value *InValue = PHI->getIncomingValue(Idx);
3323 if (InValue != CaseValue) continue;
3332 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
3333 /// instruction to a phi node dominated by the switch, if that would mean that
3334 /// some of the destination blocks of the switch can be folded away.
3335 /// Returns true if a change is made.
3336 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3337 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3338 ForwardingNodesMap ForwardingNodes;
3340 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3341 ConstantInt *CaseValue = I.getCaseValue();
3342 BasicBlock *CaseDest = I.getCaseSuccessor();
3345 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3349 ForwardingNodes[PHI].push_back(PhiIndex);
3352 bool Changed = false;
3354 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3355 E = ForwardingNodes.end(); I != E; ++I) {
3356 PHINode *Phi = I->first;
3357 SmallVectorImpl<int> &Indexes = I->second;
3359 if (Indexes.size() < 2) continue;
3361 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3362 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3369 /// ValidLookupTableConstant - Return true if the backend will be able to handle
3370 /// initializing an array of constants like C.
3371 static bool ValidLookupTableConstant(Constant *C) {
3372 if (C->isThreadDependent())
3374 if (C->isDLLImportDependent())
3377 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3378 return CE->isGEPWithNoNotionalOverIndexing();
3380 return isa<ConstantFP>(C) ||
3381 isa<ConstantInt>(C) ||
3382 isa<ConstantPointerNull>(C) ||
3383 isa<GlobalValue>(C) ||
3387 /// LookupConstant - If V is a Constant, return it. Otherwise, try to look up
3388 /// its constant value in ConstantPool, returning 0 if it's not there.
3389 static Constant *LookupConstant(Value *V,
3390 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3391 if (Constant *C = dyn_cast<Constant>(V))
3393 return ConstantPool.lookup(V);
3396 /// ConstantFold - Try to fold instruction I into a constant. This works for
3397 /// simple instructions such as binary operations where both operands are
3398 /// constant or can be replaced by constants from the ConstantPool. Returns the
3399 /// resulting constant on success, 0 otherwise.
3401 ConstantFold(Instruction *I,
3402 const SmallDenseMap<Value *, Constant *> &ConstantPool,
3403 const DataLayout *DL) {
3404 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3405 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3408 if (A->isAllOnesValue())
3409 return LookupConstant(Select->getTrueValue(), ConstantPool);
3410 if (A->isNullValue())
3411 return LookupConstant(Select->getFalseValue(), ConstantPool);
3415 SmallVector<Constant *, 4> COps;
3416 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
3417 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
3423 if (CmpInst *Cmp = dyn_cast<CmpInst>(I))
3424 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
3427 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), COps, DL);
3430 /// GetCaseResults - Try to determine the resulting constant values in phi nodes
3431 /// at the common destination basic block, *CommonDest, for one of the case
3432 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3433 /// case), of a switch instruction SI.
3435 GetCaseResults(SwitchInst *SI,
3436 ConstantInt *CaseVal,
3437 BasicBlock *CaseDest,
3438 BasicBlock **CommonDest,
3439 SmallVectorImpl<std::pair<PHINode *, Constant *> > &Res,
3440 const DataLayout *DL) {
3441 // The block from which we enter the common destination.
3442 BasicBlock *Pred = SI->getParent();
3444 // If CaseDest is empty except for some side-effect free instructions through
3445 // which we can constant-propagate the CaseVal, continue to its successor.
3446 SmallDenseMap<Value*, Constant*> ConstantPool;
3447 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3448 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3450 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3451 // If the terminator is a simple branch, continue to the next block.
3452 if (T->getNumSuccessors() != 1)
3455 CaseDest = T->getSuccessor(0);
3456 } else if (isa<DbgInfoIntrinsic>(I)) {
3457 // Skip debug intrinsic.
3459 } else if (Constant *C = ConstantFold(I, ConstantPool, DL)) {
3460 // Instruction is side-effect free and constant.
3462 // If the instruction has uses outside this block or a phi node slot for
3463 // the block, it is not safe to bypass the instruction since it would then
3464 // no longer dominate all its uses.
3465 for (auto &Use : I->uses()) {
3466 User *User = Use.getUser();
3467 if (Instruction *I = dyn_cast<Instruction>(User))
3468 if (I->getParent() == CaseDest)
3470 if (PHINode *Phi = dyn_cast<PHINode>(User))
3471 if (Phi->getIncomingBlock(Use) == CaseDest)
3476 ConstantPool.insert(std::make_pair(I, C));
3482 // If we did not have a CommonDest before, use the current one.
3484 *CommonDest = CaseDest;
3485 // If the destination isn't the common one, abort.
3486 if (CaseDest != *CommonDest)
3489 // Get the values for this case from phi nodes in the destination block.
3490 BasicBlock::iterator I = (*CommonDest)->begin();
3491 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3492 int Idx = PHI->getBasicBlockIndex(Pred);
3496 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3501 // Be conservative about which kinds of constants we support.
3502 if (!ValidLookupTableConstant(ConstVal))
3505 Res.push_back(std::make_pair(PHI, ConstVal));
3508 return Res.size() > 0;
3511 // MapCaseToResult - Helper function used to
3512 // add CaseVal to the list of cases that generate Result.
3513 static void MapCaseToResult(ConstantInt *CaseVal,
3514 SwitchCaseResultVectorTy &UniqueResults,
3516 for (auto &I : UniqueResults) {
3517 if (I.first == Result) {
3518 I.second.push_back(CaseVal);
3522 UniqueResults.push_back(std::make_pair(Result,
3523 SmallVector<ConstantInt*, 4>(1, CaseVal)));
3526 // InitializeUniqueCases - Helper function that initializes a map containing
3527 // results for the PHI node of the common destination block for a switch
3528 // instruction. Returns false if multiple PHI nodes have been found or if
3529 // there is not a common destination block for the switch.
3530 static bool InitializeUniqueCases(
3531 SwitchInst *SI, const DataLayout *DL, PHINode *&PHI,
3532 BasicBlock *&CommonDest,
3533 SwitchCaseResultVectorTy &UniqueResults,
3534 Constant *&DefaultResult) {
3535 for (auto &I : SI->cases()) {
3536 ConstantInt *CaseVal = I.getCaseValue();
3538 // Resulting value at phi nodes for this case value.
3539 SwitchCaseResultsTy Results;
3540 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
3544 // Only one value per case is permitted
3545 if (Results.size() > 1)
3547 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
3549 // Check the PHI consistency.
3551 PHI = Results[0].first;
3552 else if (PHI != Results[0].first)
3555 // Find the default result value.
3556 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
3557 BasicBlock *DefaultDest = SI->getDefaultDest();
3558 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
3560 // If the default value is not found abort unless the default destination
3563 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
3564 if ((!DefaultResult &&
3565 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
3571 // ConvertTwoCaseSwitch - Helper function that checks if it is possible to
3572 // transform a switch with only two cases (or two cases + default)
3573 // that produces a result into a value select.
3576 // case 10: %0 = icmp eq i32 %a, 10
3577 // return 10; %1 = select i1 %0, i32 10, i32 4
3578 // case 20: ----> %2 = icmp eq i32 %a, 20
3579 // return 2; %3 = select i1 %2, i32 2, i32 %1
3584 ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
3585 Constant *DefaultResult, Value *Condition,
3586 IRBuilder<> &Builder) {
3587 assert(ResultVector.size() == 2 &&
3588 "We should have exactly two unique results at this point");
3589 // If we are selecting between only two cases transform into a simple
3590 // select or a two-way select if default is possible.
3591 if (ResultVector[0].second.size() == 1 &&
3592 ResultVector[1].second.size() == 1) {
3593 ConstantInt *const FirstCase = ResultVector[0].second[0];
3594 ConstantInt *const SecondCase = ResultVector[1].second[0];
3596 bool DefaultCanTrigger = DefaultResult;
3597 Value *SelectValue = ResultVector[1].first;
3598 if (DefaultCanTrigger) {
3599 Value *const ValueCompare =
3600 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
3601 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
3602 DefaultResult, "switch.select");
3604 Value *const ValueCompare =
3605 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
3606 return Builder.CreateSelect(ValueCompare, ResultVector[0].first, SelectValue,
3613 // RemoveSwitchAfterSelectConversion - Helper function to cleanup a switch
3614 // instruction that has been converted into a select, fixing up PHI nodes and
3616 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
3618 IRBuilder<> &Builder) {
3619 BasicBlock *SelectBB = SI->getParent();
3620 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
3621 PHI->removeIncomingValue(SelectBB);
3622 PHI->addIncoming(SelectValue, SelectBB);
3624 Builder.CreateBr(PHI->getParent());
3626 // Remove the switch.
3627 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
3628 BasicBlock *Succ = SI->getSuccessor(i);
3630 if (Succ == PHI->getParent())
3632 Succ->removePredecessor(SelectBB);
3634 SI->eraseFromParent();
3637 /// SwitchToSelect - If the switch is only used to initialize one or more
3638 /// phi nodes in a common successor block with only two different
3639 /// constant values, replace the switch with select.
3640 static bool SwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
3641 const DataLayout *DL, AssumptionCache *AC) {
3642 Value *const Cond = SI->getCondition();
3643 PHINode *PHI = nullptr;
3644 BasicBlock *CommonDest = nullptr;
3645 Constant *DefaultResult;
3646 SwitchCaseResultVectorTy UniqueResults;
3647 // Collect all the cases that will deliver the same value from the switch.
3648 if (!InitializeUniqueCases(SI, DL, PHI, CommonDest, UniqueResults,
3651 // Selects choose between maximum two values.
3652 if (UniqueResults.size() != 2)
3654 assert(PHI != nullptr && "PHI for value select not found");
3656 Builder.SetInsertPoint(SI);
3657 Value *SelectValue = ConvertTwoCaseSwitch(
3659 DefaultResult, Cond, Builder);
3661 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
3664 // The switch couldn't be converted into a select.
3669 /// SwitchLookupTable - This class represents a lookup table that can be used
3670 /// to replace a switch.
3671 class SwitchLookupTable {
3673 /// SwitchLookupTable - Create a lookup table to use as a switch replacement
3674 /// with the contents of Values, using DefaultValue to fill any holes in the
3676 SwitchLookupTable(Module &M,
3678 ConstantInt *Offset,
3679 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3680 Constant *DefaultValue,
3681 const DataLayout *DL);
3683 /// BuildLookup - Build instructions with Builder to retrieve the value at
3684 /// the position given by Index in the lookup table.
3685 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3687 /// WouldFitInRegister - Return true if a table with TableSize elements of
3688 /// type ElementType would fit in a target-legal register.
3689 static bool WouldFitInRegister(const DataLayout *DL,
3691 const Type *ElementType);
3694 // Depending on the contents of the table, it can be represented in
3697 // For tables where each element contains the same value, we just have to
3698 // store that single value and return it for each lookup.
3701 // For tables where there is a linear relationship between table index
3702 // and values. We calculate the result with a simple multiplication
3703 // and addition instead of a table lookup.
3706 // For small tables with integer elements, we can pack them into a bitmap
3707 // that fits into a target-legal register. Values are retrieved by
3708 // shift and mask operations.
3711 // The table is stored as an array of values. Values are retrieved by load
3712 // instructions from the table.
3716 // For SingleValueKind, this is the single value.
3717 Constant *SingleValue;
3719 // For BitMapKind, this is the bitmap.
3720 ConstantInt *BitMap;
3721 IntegerType *BitMapElementTy;
3723 // For LinearMapKind, these are the constants used to derive the value.
3724 ConstantInt *LinearOffset;
3725 ConstantInt *LinearMultiplier;
3727 // For ArrayKind, this is the array.
3728 GlobalVariable *Array;
3732 SwitchLookupTable::SwitchLookupTable(Module &M,
3734 ConstantInt *Offset,
3735 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3736 Constant *DefaultValue,
3737 const DataLayout *DL)
3738 : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
3739 LinearOffset(nullptr), LinearMultiplier(nullptr), Array(nullptr) {
3740 assert(Values.size() && "Can't build lookup table without values!");
3741 assert(TableSize >= Values.size() && "Can't fit values in table!");
3743 // If all values in the table are equal, this is that value.
3744 SingleValue = Values.begin()->second;
3746 Type *ValueType = Values.begin()->second->getType();
3748 // Build up the table contents.
3749 SmallVector<Constant*, 64> TableContents(TableSize);
3750 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3751 ConstantInt *CaseVal = Values[I].first;
3752 Constant *CaseRes = Values[I].second;
3753 assert(CaseRes->getType() == ValueType);
3755 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3757 TableContents[Idx] = CaseRes;
3759 if (CaseRes != SingleValue)
3760 SingleValue = nullptr;
3763 // Fill in any holes in the table with the default result.
3764 if (Values.size() < TableSize) {
3765 assert(DefaultValue &&
3766 "Need a default value to fill the lookup table holes.");
3767 assert(DefaultValue->getType() == ValueType);
3768 for (uint64_t I = 0; I < TableSize; ++I) {
3769 if (!TableContents[I])
3770 TableContents[I] = DefaultValue;
3773 if (DefaultValue != SingleValue)
3774 SingleValue = nullptr;
3777 // If each element in the table contains the same value, we only need to store
3778 // that single value.
3780 Kind = SingleValueKind;
3784 // Check if we can derive the value with a linear transformation from the
3786 if (isa<IntegerType>(ValueType)) {
3787 bool LinearMappingPossible = true;
3790 assert(TableSize >= 2 && "Should be a SingleValue table.");
3791 // Check if there is the same distance between two consecutive values.
3792 for (uint64_t I = 0; I < TableSize; ++I) {
3793 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
3795 // This is an undef. We could deal with it, but undefs in lookup tables
3796 // are very seldom. It's probably not worth the additional complexity.
3797 LinearMappingPossible = false;
3800 APInt Val = ConstVal->getValue();
3802 APInt Dist = Val - PrevVal;
3805 } else if (Dist != DistToPrev) {
3806 LinearMappingPossible = false;
3812 if (LinearMappingPossible) {
3813 LinearOffset = cast<ConstantInt>(TableContents[0]);
3814 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
3815 Kind = LinearMapKind;
3821 // If the type is integer and the table fits in a register, build a bitmap.
3822 if (WouldFitInRegister(DL, TableSize, ValueType)) {
3823 IntegerType *IT = cast<IntegerType>(ValueType);
3824 APInt TableInt(TableSize * IT->getBitWidth(), 0);
3825 for (uint64_t I = TableSize; I > 0; --I) {
3826 TableInt <<= IT->getBitWidth();
3827 // Insert values into the bitmap. Undef values are set to zero.
3828 if (!isa<UndefValue>(TableContents[I - 1])) {
3829 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
3830 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
3833 BitMap = ConstantInt::get(M.getContext(), TableInt);
3834 BitMapElementTy = IT;
3840 // Store the table in an array.
3841 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
3842 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3844 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3845 GlobalVariable::PrivateLinkage,
3848 Array->setUnnamedAddr(true);
3852 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
3854 case SingleValueKind:
3856 case LinearMapKind: {
3857 // Derive the result value from the input value.
3858 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
3859 false, "switch.idx.cast");
3860 if (!LinearMultiplier->isOne())
3861 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
3862 if (!LinearOffset->isZero())
3863 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
3867 // Type of the bitmap (e.g. i59).
3868 IntegerType *MapTy = BitMap->getType();
3870 // Cast Index to the same type as the bitmap.
3871 // Note: The Index is <= the number of elements in the table, so
3872 // truncating it to the width of the bitmask is safe.
3873 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
3875 // Multiply the shift amount by the element width.
3876 ShiftAmt = Builder.CreateMul(ShiftAmt,
3877 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
3881 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
3882 "switch.downshift");
3884 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
3888 // Make sure the table index will not overflow when treated as signed.
3889 IntegerType *IT = cast<IntegerType>(Index->getType());
3890 uint64_t TableSize = Array->getInitializer()->getType()
3891 ->getArrayNumElements();
3892 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
3893 Index = Builder.CreateZExt(Index,
3894 IntegerType::get(IT->getContext(),
3895 IT->getBitWidth() + 1),
3896 "switch.tableidx.zext");
3898 Value *GEPIndices[] = { Builder.getInt32(0), Index };
3899 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices,
3901 return Builder.CreateLoad(GEP, "switch.load");
3904 llvm_unreachable("Unknown lookup table kind!");
3907 bool SwitchLookupTable::WouldFitInRegister(const DataLayout *DL,
3909 const Type *ElementType) {
3912 const IntegerType *IT = dyn_cast<IntegerType>(ElementType);
3915 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
3916 // are <= 15, we could try to narrow the type.
3918 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
3919 if (TableSize >= UINT_MAX/IT->getBitWidth())
3921 return DL->fitsInLegalInteger(TableSize * IT->getBitWidth());
3924 /// ShouldBuildLookupTable - Determine whether a lookup table should be built
3925 /// for this switch, based on the number of cases, size of the table and the
3926 /// types of the results.
3927 static bool ShouldBuildLookupTable(SwitchInst *SI,
3929 const TargetTransformInfo &TTI,
3930 const DataLayout *DL,
3931 const SmallDenseMap<PHINode*, Type*>& ResultTypes) {
3932 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
3933 return false; // TableSize overflowed, or mul below might overflow.
3935 bool AllTablesFitInRegister = true;
3936 bool HasIllegalType = false;
3937 for (const auto &I : ResultTypes) {
3938 Type *Ty = I.second;
3940 // Saturate this flag to true.
3941 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
3943 // Saturate this flag to false.
3944 AllTablesFitInRegister = AllTablesFitInRegister &&
3945 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
3947 // If both flags saturate, we're done. NOTE: This *only* works with
3948 // saturating flags, and all flags have to saturate first due to the
3949 // non-deterministic behavior of iterating over a dense map.
3950 if (HasIllegalType && !AllTablesFitInRegister)
3954 // If each table would fit in a register, we should build it anyway.
3955 if (AllTablesFitInRegister)
3958 // Don't build a table that doesn't fit in-register if it has illegal types.
3962 // The table density should be at least 40%. This is the same criterion as for
3963 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3964 // FIXME: Find the best cut-off.
3965 return SI->getNumCases() * 10 >= TableSize * 4;
3968 /// Try to reuse the switch table index compare. Following pattern:
3970 /// if (idx < tablesize)
3971 /// r = table[idx]; // table does not contain default_value
3973 /// r = default_value;
3974 /// if (r != default_value)
3977 /// Is optimized to:
3979 /// cond = idx < tablesize;
3983 /// r = default_value;
3987 /// Jump threading will then eliminate the second if(cond).
3988 static void reuseTableCompare(User *PhiUser, BasicBlock *PhiBlock,
3989 BranchInst *RangeCheckBranch, Constant *DefaultValue,
3990 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values) {
3992 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
3996 // We require that the compare is in the same block as the phi so that jump
3997 // threading can do its work afterwards.
3998 if (CmpInst->getParent() != PhiBlock)
4001 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
4005 Value *RangeCmp = RangeCheckBranch->getCondition();
4006 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
4007 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
4009 // Check if the compare with the default value is constant true or false.
4010 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
4011 DefaultValue, CmpOp1, true);
4012 if (DefaultConst != TrueConst && DefaultConst != FalseConst)
4015 // Check if the compare with the case values is distinct from the default
4017 for (auto ValuePair : Values) {
4018 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
4019 ValuePair.second, CmpOp1, true);
4020 if (!CaseConst || CaseConst == DefaultConst)
4022 assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
4023 "Expect true or false as compare result.");
4026 // Check if the branch instruction dominates the phi node. It's a simple
4027 // dominance check, but sufficient for our needs.
4028 // Although this check is invariant in the calling loops, it's better to do it
4029 // at this late stage. Practically we do it at most once for a switch.
4030 BasicBlock *BranchBlock = RangeCheckBranch->getParent();
4031 for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
4032 BasicBlock *Pred = *PI;
4033 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
4037 if (DefaultConst == FalseConst) {
4038 // The compare yields the same result. We can replace it.
4039 CmpInst->replaceAllUsesWith(RangeCmp);
4040 ++NumTableCmpReuses;
4042 // The compare yields the same result, just inverted. We can replace it.
4043 Value *InvertedTableCmp = BinaryOperator::CreateXor(RangeCmp,
4044 ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
4046 CmpInst->replaceAllUsesWith(InvertedTableCmp);
4047 ++NumTableCmpReuses;
4051 /// SwitchToLookupTable - If the switch is only used to initialize one or more
4052 /// phi nodes in a common successor block with different constant values,
4053 /// replace the switch with lookup tables.
4054 static bool SwitchToLookupTable(SwitchInst *SI,
4055 IRBuilder<> &Builder,
4056 const TargetTransformInfo &TTI,
4057 const DataLayout* DL) {
4058 assert(SI->getNumCases() > 1 && "Degenerate switch?");
4060 // Only build lookup table when we have a target that supports it.
4061 if (!TTI.shouldBuildLookupTables())
4064 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
4065 // split off a dense part and build a lookup table for that.
4067 // FIXME: This creates arrays of GEPs to constant strings, which means each
4068 // GEP needs a runtime relocation in PIC code. We should just build one big
4069 // string and lookup indices into that.
4071 // Ignore switches with less than three cases. Lookup tables will not make them
4072 // faster, so we don't analyze them.
4073 if (SI->getNumCases() < 3)
4076 // Figure out the corresponding result for each case value and phi node in the
4077 // common destination, as well as the the min and max case values.
4078 assert(SI->case_begin() != SI->case_end());
4079 SwitchInst::CaseIt CI = SI->case_begin();
4080 ConstantInt *MinCaseVal = CI.getCaseValue();
4081 ConstantInt *MaxCaseVal = CI.getCaseValue();
4083 BasicBlock *CommonDest = nullptr;
4084 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
4085 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
4086 SmallDenseMap<PHINode*, Constant*> DefaultResults;
4087 SmallDenseMap<PHINode*, Type*> ResultTypes;
4088 SmallVector<PHINode*, 4> PHIs;
4090 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
4091 ConstantInt *CaseVal = CI.getCaseValue();
4092 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
4093 MinCaseVal = CaseVal;
4094 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
4095 MaxCaseVal = CaseVal;
4097 // Resulting value at phi nodes for this case value.
4098 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
4100 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
4104 // Append the result from this case to the list for each phi.
4105 for (const auto &I : Results) {
4106 PHINode *PHI = I.first;
4107 Constant *Value = I.second;
4108 if (!ResultLists.count(PHI))
4109 PHIs.push_back(PHI);
4110 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
4114 // Keep track of the result types.
4115 for (PHINode *PHI : PHIs) {
4116 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
4119 uint64_t NumResults = ResultLists[PHIs[0]].size();
4120 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
4121 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
4122 bool TableHasHoles = (NumResults < TableSize);
4124 // If the table has holes, we need a constant result for the default case
4125 // or a bitmask that fits in a register.
4126 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
4127 bool HasDefaultResults = GetCaseResults(SI, nullptr, SI->getDefaultDest(),
4128 &CommonDest, DefaultResultsList, DL);
4130 bool NeedMask = (TableHasHoles && !HasDefaultResults);
4132 // As an extra penalty for the validity test we require more cases.
4133 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
4135 if (!(DL && DL->fitsInLegalInteger(TableSize)))
4139 for (const auto &I : DefaultResultsList) {
4140 PHINode *PHI = I.first;
4141 Constant *Result = I.second;
4142 DefaultResults[PHI] = Result;
4145 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
4148 // Create the BB that does the lookups.
4149 Module &Mod = *CommonDest->getParent()->getParent();
4150 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
4152 CommonDest->getParent(),
4155 // Compute the table index value.
4156 Builder.SetInsertPoint(SI);
4157 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
4160 // Compute the maximum table size representable by the integer type we are
4162 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
4163 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
4164 assert(MaxTableSize >= TableSize &&
4165 "It is impossible for a switch to have more entries than the max "
4166 "representable value of its input integer type's size.");
4168 // If the default destination is unreachable, or if the lookup table covers
4169 // all values of the conditional variable, branch directly to the lookup table
4170 // BB. Otherwise, check that the condition is within the case range.
4171 const bool DefaultIsReachable =
4172 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4173 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
4174 BranchInst *RangeCheckBranch = nullptr;
4176 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
4177 Builder.CreateBr(LookupBB);
4178 // We cached PHINodes in PHIs, to avoid accessing deleted PHINodes later,
4179 // do not delete PHINodes here.
4180 SI->getDefaultDest()->removePredecessor(SI->getParent(),
4181 /*DontDeleteUselessPHIs=*/true);
4183 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
4184 MinCaseVal->getType(), TableSize));
4185 RangeCheckBranch = Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
4188 // Populate the BB that does the lookups.
4189 Builder.SetInsertPoint(LookupBB);
4192 // Before doing the lookup we do the hole check.
4193 // The LookupBB is therefore re-purposed to do the hole check
4194 // and we create a new LookupBB.
4195 BasicBlock *MaskBB = LookupBB;
4196 MaskBB->setName("switch.hole_check");
4197 LookupBB = BasicBlock::Create(Mod.getContext(),
4199 CommonDest->getParent(),
4202 // Make the mask's bitwidth at least 8bit and a power-of-2 to avoid
4203 // unnecessary illegal types.
4204 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
4205 APInt MaskInt(TableSizePowOf2, 0);
4206 APInt One(TableSizePowOf2, 1);
4207 // Build bitmask; fill in a 1 bit for every case.
4208 const ResultListTy &ResultList = ResultLists[PHIs[0]];
4209 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
4210 uint64_t Idx = (ResultList[I].first->getValue() -
4211 MinCaseVal->getValue()).getLimitedValue();
4212 MaskInt |= One << Idx;
4214 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
4216 // Get the TableIndex'th bit of the bitmask.
4217 // If this bit is 0 (meaning hole) jump to the default destination,
4218 // else continue with table lookup.
4219 IntegerType *MapTy = TableMask->getType();
4220 Value *MaskIndex = Builder.CreateZExtOrTrunc(TableIndex, MapTy,
4221 "switch.maskindex");
4222 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex,
4224 Value *LoBit = Builder.CreateTrunc(Shifted,
4225 Type::getInt1Ty(Mod.getContext()),
4227 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
4229 Builder.SetInsertPoint(LookupBB);
4230 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
4233 bool ReturnedEarly = false;
4234 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
4235 PHINode *PHI = PHIs[I];
4236 const ResultListTy &ResultList = ResultLists[PHI];
4238 // If using a bitmask, use any value to fill the lookup table holes.
4239 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
4240 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL);
4242 Value *Result = Table.BuildLookup(TableIndex, Builder);
4244 // If the result is used to return immediately from the function, we want to
4245 // do that right here.
4246 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
4247 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
4248 Builder.CreateRet(Result);
4249 ReturnedEarly = true;
4253 // Do a small peephole optimization: re-use the switch table compare if
4255 if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
4256 BasicBlock *PhiBlock = PHI->getParent();
4257 // Search for compare instructions which use the phi.
4258 for (auto *User : PHI->users()) {
4259 reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
4263 PHI->addIncoming(Result, LookupBB);
4267 Builder.CreateBr(CommonDest);
4269 // Remove the switch.
4270 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4271 BasicBlock *Succ = SI->getSuccessor(i);
4273 if (Succ == SI->getDefaultDest())
4275 Succ->removePredecessor(SI->getParent());
4277 SI->eraseFromParent();
4281 ++NumLookupTablesHoles;
4285 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
4286 BasicBlock *BB = SI->getParent();
4288 if (isValueEqualityComparison(SI)) {
4289 // If we only have one predecessor, and if it is a branch on this value,
4290 // see if that predecessor totally determines the outcome of this switch.
4291 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4292 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
4293 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AC) | true;
4295 Value *Cond = SI->getCondition();
4296 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
4297 if (SimplifySwitchOnSelect(SI, Select))
4298 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AC) | true;
4300 // If the block only contains the switch, see if we can fold the block
4301 // away into any preds.
4302 BasicBlock::iterator BBI = BB->begin();
4303 // Ignore dbg intrinsics.
4304 while (isa<DbgInfoIntrinsic>(BBI))
4307 if (FoldValueComparisonIntoPredecessors(SI, Builder))
4308 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AC) | true;
4311 // Try to transform the switch into an icmp and a branch.
4312 if (TurnSwitchRangeIntoICmp(SI, Builder))
4313 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AC) | true;
4315 // Remove unreachable cases.
4316 if (EliminateDeadSwitchCases(SI, DL, AC))
4317 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AC) | true;
4319 if (SwitchToSelect(SI, Builder, DL, AC))
4320 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AC) | true;
4322 if (ForwardSwitchConditionToPHI(SI))
4323 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AC) | true;
4325 if (SwitchToLookupTable(SI, Builder, TTI, DL))
4326 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AC) | true;
4331 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
4332 BasicBlock *BB = IBI->getParent();
4333 bool Changed = false;
4335 // Eliminate redundant destinations.
4336 SmallPtrSet<Value *, 8> Succs;
4337 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
4338 BasicBlock *Dest = IBI->getDestination(i);
4339 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
4340 Dest->removePredecessor(BB);
4341 IBI->removeDestination(i);
4347 if (IBI->getNumDestinations() == 0) {
4348 // If the indirectbr has no successors, change it to unreachable.
4349 new UnreachableInst(IBI->getContext(), IBI);
4350 EraseTerminatorInstAndDCECond(IBI);
4354 if (IBI->getNumDestinations() == 1) {
4355 // If the indirectbr has one successor, change it to a direct branch.
4356 BranchInst::Create(IBI->getDestination(0), IBI);
4357 EraseTerminatorInstAndDCECond(IBI);
4361 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
4362 if (SimplifyIndirectBrOnSelect(IBI, SI))
4363 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AC) | true;
4368 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
4369 BasicBlock *BB = BI->getParent();
4371 if (SinkCommon && SinkThenElseCodeToEnd(BI))
4374 // If the Terminator is the only non-phi instruction, simplify the block.
4375 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg();
4376 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
4377 TryToSimplifyUncondBranchFromEmptyBlock(BB))
4380 // If the only instruction in the block is a seteq/setne comparison
4381 // against a constant, try to simplify the block.
4382 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
4383 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
4384 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
4386 if (I->isTerminator() &&
4387 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI,
4388 BonusInstThreshold, DL, AC))
4392 // If this basic block is ONLY a compare and a branch, and if a predecessor
4393 // branches to us and our successor, fold the comparison into the
4394 // predecessor and use logical operations to update the incoming value
4395 // for PHI nodes in common successor.
4396 if (FoldBranchToCommonDest(BI, DL, BonusInstThreshold))
4397 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AC) | true;
4402 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
4403 BasicBlock *BB = BI->getParent();
4405 // Conditional branch
4406 if (isValueEqualityComparison(BI)) {
4407 // If we only have one predecessor, and if it is a branch on this value,
4408 // see if that predecessor totally determines the outcome of this
4410 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4411 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
4412 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AC) | true;
4414 // This block must be empty, except for the setcond inst, if it exists.
4415 // Ignore dbg intrinsics.
4416 BasicBlock::iterator I = BB->begin();
4417 // Ignore dbg intrinsics.
4418 while (isa<DbgInfoIntrinsic>(I))
4421 if (FoldValueComparisonIntoPredecessors(BI, Builder))
4422 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AC) | true;
4423 } else if (&*I == cast<Instruction>(BI->getCondition())){
4425 // Ignore dbg intrinsics.
4426 while (isa<DbgInfoIntrinsic>(I))
4428 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
4429 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AC) | true;
4433 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
4434 if (SimplifyBranchOnICmpChain(BI, DL, Builder))
4437 // If this basic block is ONLY a compare and a branch, and if a predecessor
4438 // branches to us and one of our successors, fold the comparison into the
4439 // predecessor and use logical operations to pick the right destination.
4440 if (FoldBranchToCommonDest(BI, DL, BonusInstThreshold))
4441 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AC) | true;
4443 // We have a conditional branch to two blocks that are only reachable
4444 // from BI. We know that the condbr dominates the two blocks, so see if
4445 // there is any identical code in the "then" and "else" blocks. If so, we
4446 // can hoist it up to the branching block.
4447 if (BI->getSuccessor(0)->getSinglePredecessor()) {
4448 if (BI->getSuccessor(1)->getSinglePredecessor()) {
4449 if (HoistThenElseCodeToIf(BI, DL, TTI))
4450 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AC) | true;
4452 // If Successor #1 has multiple preds, we may be able to conditionally
4453 // execute Successor #0 if it branches to Successor #1.
4454 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
4455 if (Succ0TI->getNumSuccessors() == 1 &&
4456 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
4457 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), DL, TTI))
4458 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AC) | true;
4460 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
4461 // If Successor #0 has multiple preds, we may be able to conditionally
4462 // execute Successor #1 if it branches to Successor #0.
4463 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
4464 if (Succ1TI->getNumSuccessors() == 1 &&
4465 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
4466 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), DL, TTI))
4467 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AC) | true;
4470 // If this is a branch on a phi node in the current block, thread control
4471 // through this block if any PHI node entries are constants.
4472 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
4473 if (PN->getParent() == BI->getParent())
4474 if (FoldCondBranchOnPHI(BI, DL))
4475 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AC) | true;
4477 // Scan predecessor blocks for conditional branches.
4478 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
4479 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
4480 if (PBI != BI && PBI->isConditional())
4481 if (SimplifyCondBranchToCondBranch(PBI, BI))
4482 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AC) | true;
4487 /// Check if passing a value to an instruction will cause undefined behavior.
4488 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
4489 Constant *C = dyn_cast<Constant>(V);
4496 if (C->isNullValue()) {
4497 // Only look at the first use, avoid hurting compile time with long uselists
4498 User *Use = *I->user_begin();
4500 // Now make sure that there are no instructions in between that can alter
4501 // control flow (eg. calls)
4502 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
4503 if (i == I->getParent()->end() || i->mayHaveSideEffects())
4506 // Look through GEPs. A load from a GEP derived from NULL is still undefined
4507 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
4508 if (GEP->getPointerOperand() == I)
4509 return passingValueIsAlwaysUndefined(V, GEP);
4511 // Look through bitcasts.
4512 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
4513 return passingValueIsAlwaysUndefined(V, BC);
4515 // Load from null is undefined.
4516 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
4517 if (!LI->isVolatile())
4518 return LI->getPointerAddressSpace() == 0;
4520 // Store to null is undefined.
4521 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
4522 if (!SI->isVolatile())
4523 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
4528 /// If BB has an incoming value that will always trigger undefined behavior
4529 /// (eg. null pointer dereference), remove the branch leading here.
4530 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
4531 for (BasicBlock::iterator i = BB->begin();
4532 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
4533 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
4534 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
4535 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
4536 IRBuilder<> Builder(T);
4537 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
4538 BB->removePredecessor(PHI->getIncomingBlock(i));
4539 // Turn uncoditional branches into unreachables and remove the dead
4540 // destination from conditional branches.
4541 if (BI->isUnconditional())
4542 Builder.CreateUnreachable();
4544 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
4545 BI->getSuccessor(0));
4546 BI->eraseFromParent();
4549 // TODO: SwitchInst.
4555 bool SimplifyCFGOpt::run(BasicBlock *BB) {
4556 bool Changed = false;
4558 assert(BB && BB->getParent() && "Block not embedded in function!");
4559 assert(BB->getTerminator() && "Degenerate basic block encountered!");
4561 // Remove basic blocks that have no predecessors (except the entry block)...
4562 // or that just have themself as a predecessor. These are unreachable.
4563 if ((pred_empty(BB) &&
4564 BB != &BB->getParent()->getEntryBlock()) ||
4565 BB->getSinglePredecessor() == BB) {
4566 DEBUG(dbgs() << "Removing BB: \n" << *BB);
4567 DeleteDeadBlock(BB);
4571 // Check to see if we can constant propagate this terminator instruction
4573 Changed |= ConstantFoldTerminator(BB, true);
4575 // Check for and eliminate duplicate PHI nodes in this block.
4576 Changed |= EliminateDuplicatePHINodes(BB);
4578 // Check for and remove branches that will always cause undefined behavior.
4579 Changed |= removeUndefIntroducingPredecessor(BB);
4581 // Merge basic blocks into their predecessor if there is only one distinct
4582 // pred, and if there is only one distinct successor of the predecessor, and
4583 // if there are no PHI nodes.
4585 if (MergeBlockIntoPredecessor(BB))
4588 IRBuilder<> Builder(BB);
4590 // If there is a trivial two-entry PHI node in this basic block, and we can
4591 // eliminate it, do so now.
4592 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
4593 if (PN->getNumIncomingValues() == 2)
4594 Changed |= FoldTwoEntryPHINode(PN, DL, TTI);
4596 Builder.SetInsertPoint(BB->getTerminator());
4597 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
4598 if (BI->isUnconditional()) {
4599 if (SimplifyUncondBranch(BI, Builder)) return true;
4601 if (SimplifyCondBranch(BI, Builder)) return true;
4603 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
4604 if (SimplifyReturn(RI, Builder)) return true;
4605 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
4606 if (SimplifyResume(RI, Builder)) return true;
4607 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
4608 if (SimplifySwitch(SI, Builder)) return true;
4609 } else if (UnreachableInst *UI =
4610 dyn_cast<UnreachableInst>(BB->getTerminator())) {
4611 if (SimplifyUnreachable(UI)) return true;
4612 } else if (IndirectBrInst *IBI =
4613 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
4614 if (SimplifyIndirectBr(IBI)) return true;
4620 /// SimplifyCFG - This function is used to do simplification of a CFG. For
4621 /// example, it adjusts branches to branches to eliminate the extra hop, it
4622 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4623 /// of the CFG. It returns true if a modification was made.
4625 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
4626 unsigned BonusInstThreshold, const DataLayout *DL,
4627 AssumptionCache *AC) {
4628 return SimplifyCFGOpt(TTI, BonusInstThreshold, DL, AC).run(BB);