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 const DataLayout &DL;
114 unsigned BonusInstThreshold;
116 Value *isValueEqualityComparison(TerminatorInst *TI);
117 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
118 std::vector<ValueEqualityComparisonCase> &Cases);
119 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
121 IRBuilder<> &Builder);
122 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
123 IRBuilder<> &Builder);
125 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
126 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
127 bool 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, const DataLayout &DL,
135 unsigned BonusInstThreshold, AssumptionCache *AC)
136 : TTI(TTI), DL(DL), BonusInstThreshold(BonusInstThreshold), AC(AC) {}
137 bool run(BasicBlock *BB);
141 /// Return true if it is safe to merge these two
142 /// terminator instructions together.
143 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
144 if (SI1 == SI2) return false; // Can't merge with self!
146 // It is not safe to merge these two switch instructions if they have a common
147 // successor, and if that successor has a PHI node, and if *that* PHI node has
148 // conflicting incoming values from the two switch blocks.
149 BasicBlock *SI1BB = SI1->getParent();
150 BasicBlock *SI2BB = SI2->getParent();
151 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
153 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
154 if (SI1Succs.count(*I))
155 for (BasicBlock::iterator BBI = (*I)->begin();
156 isa<PHINode>(BBI); ++BBI) {
157 PHINode *PN = cast<PHINode>(BBI);
158 if (PN->getIncomingValueForBlock(SI1BB) !=
159 PN->getIncomingValueForBlock(SI2BB))
166 /// Return true if it is safe and profitable to merge these two terminator
167 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
168 /// store all PHI nodes in common successors.
169 static bool isProfitableToFoldUnconditional(BranchInst *SI1,
172 SmallVectorImpl<PHINode*> &PhiNodes) {
173 if (SI1 == SI2) return false; // Can't merge with self!
174 assert(SI1->isUnconditional() && SI2->isConditional());
176 // We fold the unconditional branch if we can easily update all PHI nodes in
177 // common successors:
178 // 1> We have a constant incoming value for the conditional branch;
179 // 2> We have "Cond" as the incoming value for the unconditional branch;
180 // 3> SI2->getCondition() and Cond have same operands.
181 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
182 if (!Ci2) return false;
183 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
184 Cond->getOperand(1) == Ci2->getOperand(1)) &&
185 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
186 Cond->getOperand(1) == Ci2->getOperand(0)))
189 BasicBlock *SI1BB = SI1->getParent();
190 BasicBlock *SI2BB = SI2->getParent();
191 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
192 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
193 if (SI1Succs.count(*I))
194 for (BasicBlock::iterator BBI = (*I)->begin();
195 isa<PHINode>(BBI); ++BBI) {
196 PHINode *PN = cast<PHINode>(BBI);
197 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
198 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
200 PhiNodes.push_back(PN);
205 /// Update PHI nodes in Succ to indicate that there will now be entries in it
206 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
207 /// will be the same as those coming in from ExistPred, an existing predecessor
209 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
210 BasicBlock *ExistPred) {
211 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
214 for (BasicBlock::iterator I = Succ->begin();
215 (PN = dyn_cast<PHINode>(I)); ++I)
216 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
219 /// Compute an abstract "cost" of speculating the given instruction,
220 /// which is assumed to be safe to speculate. TCC_Free means cheap,
221 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
223 static unsigned ComputeSpeculationCost(const User *I,
224 const TargetTransformInfo &TTI) {
225 assert(isSafeToSpeculativelyExecute(I) &&
226 "Instruction is not safe to speculatively execute!");
227 return TTI.getUserCost(I);
229 /// If we have a merge point of an "if condition" as accepted above,
230 /// return true if the specified value dominates the block. We
231 /// don't handle the true generality of domination here, just a special case
232 /// which works well enough for us.
234 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
235 /// see if V (which must be an instruction) and its recursive operands
236 /// that do not dominate BB have a combined cost lower than CostRemaining and
237 /// are non-trapping. If both are true, the instruction is inserted into the
238 /// set and true is returned.
240 /// The cost for most non-trapping instructions is defined as 1 except for
241 /// Select whose cost is 2.
243 /// After this function returns, CostRemaining is decreased by the cost of
244 /// V plus its non-dominating operands. If that cost is greater than
245 /// CostRemaining, false is returned and CostRemaining is undefined.
246 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
247 SmallPtrSetImpl<Instruction*> *AggressiveInsts,
248 unsigned &CostRemaining,
249 const TargetTransformInfo &TTI) {
250 Instruction *I = dyn_cast<Instruction>(V);
252 // Non-instructions all dominate instructions, but not all constantexprs
253 // can be executed unconditionally.
254 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
259 BasicBlock *PBB = I->getParent();
261 // We don't want to allow weird loops that might have the "if condition" in
262 // the bottom of this block.
263 if (PBB == BB) return false;
265 // If this instruction is defined in a block that contains an unconditional
266 // branch to BB, then it must be in the 'conditional' part of the "if
267 // statement". If not, it definitely dominates the region.
268 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
269 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
272 // If we aren't allowing aggressive promotion anymore, then don't consider
273 // instructions in the 'if region'.
274 if (!AggressiveInsts) return false;
276 // If we have seen this instruction before, don't count it again.
277 if (AggressiveInsts->count(I)) return true;
279 // Okay, it looks like the instruction IS in the "condition". Check to
280 // see if it's a cheap instruction to unconditionally compute, and if it
281 // only uses stuff defined outside of the condition. If so, hoist it out.
282 if (!isSafeToSpeculativelyExecute(I))
285 unsigned Cost = ComputeSpeculationCost(I, TTI);
287 if (Cost > CostRemaining)
290 CostRemaining -= Cost;
292 // Okay, we can only really hoist these out if their operands do
293 // not take us over the cost threshold.
294 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
295 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, TTI))
297 // Okay, it's safe to do this! Remember this instruction.
298 AggressiveInsts->insert(I);
302 /// Extract ConstantInt from value, looking through IntToPtr
303 /// and PointerNullValue. Return NULL if value is not a constant int.
304 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
305 // Normal constant int.
306 ConstantInt *CI = dyn_cast<ConstantInt>(V);
307 if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
310 // This is some kind of pointer constant. Turn it into a pointer-sized
311 // ConstantInt if possible.
312 IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
314 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
315 if (isa<ConstantPointerNull>(V))
316 return ConstantInt::get(PtrTy, 0);
318 // IntToPtr const int.
319 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
320 if (CE->getOpcode() == Instruction::IntToPtr)
321 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
322 // The constant is very likely to have the right type already.
323 if (CI->getType() == PtrTy)
326 return cast<ConstantInt>
327 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
334 /// Given a chain of or (||) or and (&&) comparison of a value against a
335 /// constant, this will try to recover the information required for a switch
337 /// It will depth-first traverse the chain of comparison, seeking for patterns
338 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
339 /// representing the different cases for the switch.
340 /// Note that if the chain is composed of '||' it will build the set of elements
341 /// that matches the comparisons (i.e. any of this value validate the chain)
342 /// while for a chain of '&&' it will build the set elements that make the test
344 struct ConstantComparesGatherer {
345 const DataLayout &DL;
346 Value *CompValue; /// Value found for the switch comparison
347 Value *Extra; /// Extra clause to be checked before the switch
348 SmallVector<ConstantInt *, 8> Vals; /// Set of integers to match in switch
349 unsigned UsedICmps; /// Number of comparisons matched in the and/or chain
351 /// Construct and compute the result for the comparison instruction Cond
352 ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL)
353 : DL(DL), CompValue(nullptr), Extra(nullptr), UsedICmps(0) {
358 ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
359 ConstantComparesGatherer &
360 operator=(const ConstantComparesGatherer &) = delete;
364 /// Try to set the current value used for the comparison, it succeeds only if
365 /// it wasn't set before or if the new value is the same as the old one
366 bool setValueOnce(Value *NewVal) {
367 if(CompValue && CompValue != NewVal) return false;
369 return (CompValue != nullptr);
372 /// Try to match Instruction "I" as a comparison against a constant and
373 /// populates the array Vals with the set of values that match (or do not
374 /// match depending on isEQ).
375 /// Return false on failure. On success, the Value the comparison matched
376 /// against is placed in CompValue.
377 /// If CompValue is already set, the function is expected to fail if a match
378 /// is found but the value compared to is different.
379 bool matchInstruction(Instruction *I, bool isEQ) {
380 // If this is an icmp against a constant, handle this as one of the cases.
383 if (!((ICI = dyn_cast<ICmpInst>(I)) &&
384 (C = GetConstantInt(I->getOperand(1), DL)))) {
391 // Pattern match a special case
392 // (x & ~2^x) == y --> x == y || x == y|2^x
393 // This undoes a transformation done by instcombine to fuse 2 compares.
394 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
395 if (match(ICI->getOperand(0),
396 m_And(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
397 APInt Not = ~RHSC->getValue();
398 if (Not.isPowerOf2()) {
399 // If we already have a value for the switch, it has to match!
400 if(!setValueOnce(RHSVal))
404 Vals.push_back(ConstantInt::get(C->getContext(),
405 C->getValue() | Not));
411 // If we already have a value for the switch, it has to match!
412 if(!setValueOnce(ICI->getOperand(0)))
417 return ICI->getOperand(0);
420 // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
421 ConstantRange Span = ConstantRange::makeAllowedICmpRegion(
422 ICI->getPredicate(), C->getValue());
424 // Shift the range if the compare is fed by an add. This is the range
425 // compare idiom as emitted by instcombine.
426 Value *CandidateVal = I->getOperand(0);
427 if(match(I->getOperand(0), m_Add(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
428 Span = Span.subtract(RHSC->getValue());
429 CandidateVal = RHSVal;
432 // If this is an and/!= check, then we are looking to build the set of
433 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
436 Span = Span.inverse();
438 // If there are a ton of values, we don't want to make a ginormous switch.
439 if (Span.getSetSize().ugt(8) || Span.isEmptySet()) {
443 // If we already have a value for the switch, it has to match!
444 if(!setValueOnce(CandidateVal))
447 // Add all values from the range to the set
448 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
449 Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
456 /// Given a potentially 'or'd or 'and'd together collection of icmp
457 /// eq/ne/lt/gt instructions that compare a value against a constant, extract
458 /// the value being compared, and stick the list constants into the Vals
460 /// One "Extra" case is allowed to differ from the other.
461 void gather(Value *V) {
462 Instruction *I = dyn_cast<Instruction>(V);
463 bool isEQ = (I->getOpcode() == Instruction::Or);
465 // Keep a stack (SmallVector for efficiency) for depth-first traversal
466 SmallVector<Value *, 8> DFT;
471 while(!DFT.empty()) {
472 V = DFT.pop_back_val();
474 if (Instruction *I = dyn_cast<Instruction>(V)) {
475 // If it is a || (or && depending on isEQ), process the operands.
476 if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
477 DFT.push_back(I->getOperand(1));
478 DFT.push_back(I->getOperand(0));
482 // Try to match the current instruction
483 if (matchInstruction(I, isEQ))
484 // Match succeed, continue the loop
488 // One element of the sequence of || (or &&) could not be match as a
489 // comparison against the same value as the others.
490 // We allow only one "Extra" case to be checked before the switch
495 // Failed to parse a proper sequence, abort now
504 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
505 Instruction *Cond = nullptr;
506 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
507 Cond = dyn_cast<Instruction>(SI->getCondition());
508 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
509 if (BI->isConditional())
510 Cond = dyn_cast<Instruction>(BI->getCondition());
511 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
512 Cond = dyn_cast<Instruction>(IBI->getAddress());
515 TI->eraseFromParent();
516 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
519 /// Return true if the specified terminator checks
520 /// to see if a value is equal to constant integer value.
521 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
523 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
524 // Do not permit merging of large switch instructions into their
525 // predecessors unless there is only one predecessor.
526 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()),
527 pred_end(SI->getParent())) <= 128)
528 CV = SI->getCondition();
529 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
530 if (BI->isConditional() && BI->getCondition()->hasOneUse())
531 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
532 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
533 CV = ICI->getOperand(0);
536 // Unwrap any lossless ptrtoint cast.
538 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
539 Value *Ptr = PTII->getPointerOperand();
540 if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
547 /// Given a value comparison instruction,
548 /// decode all of the 'cases' that it represents and return the 'default' block.
549 BasicBlock *SimplifyCFGOpt::
550 GetValueEqualityComparisonCases(TerminatorInst *TI,
551 std::vector<ValueEqualityComparisonCase>
553 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
554 Cases.reserve(SI->getNumCases());
555 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
556 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
557 i.getCaseSuccessor()));
558 return SI->getDefaultDest();
561 BranchInst *BI = cast<BranchInst>(TI);
562 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
563 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
564 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
567 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
571 /// Given a vector of bb/value pairs, remove any entries
572 /// in the list that match the specified block.
573 static void EliminateBlockCases(BasicBlock *BB,
574 std::vector<ValueEqualityComparisonCase> &Cases) {
575 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
578 /// Return true if there are any keys in C1 that exist in C2 as well.
580 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
581 std::vector<ValueEqualityComparisonCase > &C2) {
582 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
584 // Make V1 be smaller than V2.
585 if (V1->size() > V2->size())
588 if (V1->size() == 0) return false;
589 if (V1->size() == 1) {
591 ConstantInt *TheVal = (*V1)[0].Value;
592 for (unsigned i = 0, e = V2->size(); i != e; ++i)
593 if (TheVal == (*V2)[i].Value)
597 // Otherwise, just sort both lists and compare element by element.
598 array_pod_sort(V1->begin(), V1->end());
599 array_pod_sort(V2->begin(), V2->end());
600 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
601 while (i1 != e1 && i2 != e2) {
602 if ((*V1)[i1].Value == (*V2)[i2].Value)
604 if ((*V1)[i1].Value < (*V2)[i2].Value)
612 /// If TI is known to be a terminator instruction and its block is known to
613 /// only have a single predecessor block, check to see if that predecessor is
614 /// also a value comparison with the same value, and if that comparison
615 /// determines the outcome of this comparison. If so, simplify TI. This does a
616 /// very limited form of jump threading.
617 bool SimplifyCFGOpt::
618 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
620 IRBuilder<> &Builder) {
621 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
622 if (!PredVal) return false; // Not a value comparison in predecessor.
624 Value *ThisVal = isValueEqualityComparison(TI);
625 assert(ThisVal && "This isn't a value comparison!!");
626 if (ThisVal != PredVal) return false; // Different predicates.
628 // TODO: Preserve branch weight metadata, similarly to how
629 // FoldValueComparisonIntoPredecessors preserves it.
631 // Find out information about when control will move from Pred to TI's block.
632 std::vector<ValueEqualityComparisonCase> PredCases;
633 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
635 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
637 // Find information about how control leaves this block.
638 std::vector<ValueEqualityComparisonCase> ThisCases;
639 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
640 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
642 // If TI's block is the default block from Pred's comparison, potentially
643 // simplify TI based on this knowledge.
644 if (PredDef == TI->getParent()) {
645 // If we are here, we know that the value is none of those cases listed in
646 // PredCases. If there are any cases in ThisCases that are in PredCases, we
648 if (!ValuesOverlap(PredCases, ThisCases))
651 if (isa<BranchInst>(TI)) {
652 // Okay, one of the successors of this condbr is dead. Convert it to a
654 assert(ThisCases.size() == 1 && "Branch can only have one case!");
655 // Insert the new branch.
656 Instruction *NI = Builder.CreateBr(ThisDef);
659 // Remove PHI node entries for the dead edge.
660 ThisCases[0].Dest->removePredecessor(TI->getParent());
662 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
663 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
665 EraseTerminatorInstAndDCECond(TI);
669 SwitchInst *SI = cast<SwitchInst>(TI);
670 // Okay, TI has cases that are statically dead, prune them away.
671 SmallPtrSet<Constant*, 16> DeadCases;
672 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
673 DeadCases.insert(PredCases[i].Value);
675 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
676 << "Through successor TI: " << *TI);
678 // Collect branch weights into a vector.
679 SmallVector<uint32_t, 8> Weights;
680 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
681 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
683 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
685 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
686 Weights.push_back(CI->getValue().getZExtValue());
688 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
690 if (DeadCases.count(i.getCaseValue())) {
692 std::swap(Weights[i.getCaseIndex()+1], Weights.back());
695 i.getCaseSuccessor()->removePredecessor(TI->getParent());
699 if (HasWeight && Weights.size() >= 2)
700 SI->setMetadata(LLVMContext::MD_prof,
701 MDBuilder(SI->getParent()->getContext()).
702 createBranchWeights(Weights));
704 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
708 // Otherwise, TI's block must correspond to some matched value. Find out
709 // which value (or set of values) this is.
710 ConstantInt *TIV = nullptr;
711 BasicBlock *TIBB = TI->getParent();
712 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
713 if (PredCases[i].Dest == TIBB) {
715 return false; // Cannot handle multiple values coming to this block.
716 TIV = PredCases[i].Value;
718 assert(TIV && "No edge from pred to succ?");
720 // Okay, we found the one constant that our value can be if we get into TI's
721 // BB. Find out which successor will unconditionally be branched to.
722 BasicBlock *TheRealDest = nullptr;
723 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
724 if (ThisCases[i].Value == TIV) {
725 TheRealDest = ThisCases[i].Dest;
729 // If not handled by any explicit cases, it is handled by the default case.
730 if (!TheRealDest) TheRealDest = ThisDef;
732 // Remove PHI node entries for dead edges.
733 BasicBlock *CheckEdge = TheRealDest;
734 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
735 if (*SI != CheckEdge)
736 (*SI)->removePredecessor(TIBB);
740 // Insert the new branch.
741 Instruction *NI = Builder.CreateBr(TheRealDest);
744 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
745 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
747 EraseTerminatorInstAndDCECond(TI);
752 /// This class implements a stable ordering of constant
753 /// integers that does not depend on their address. This is important for
754 /// applications that sort ConstantInt's to ensure uniqueness.
755 struct ConstantIntOrdering {
756 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
757 return LHS->getValue().ult(RHS->getValue());
762 static int ConstantIntSortPredicate(ConstantInt *const *P1,
763 ConstantInt *const *P2) {
764 const ConstantInt *LHS = *P1;
765 const ConstantInt *RHS = *P2;
766 if (LHS->getValue().ult(RHS->getValue()))
768 if (LHS->getValue() == RHS->getValue())
773 static inline bool HasBranchWeights(const Instruction* I) {
774 MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
775 if (ProfMD && ProfMD->getOperand(0))
776 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
777 return MDS->getString().equals("branch_weights");
782 /// Get Weights of a given TerminatorInst, the default weight is at the front
783 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
785 static void GetBranchWeights(TerminatorInst *TI,
786 SmallVectorImpl<uint64_t> &Weights) {
787 MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
789 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
790 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
791 Weights.push_back(CI->getValue().getZExtValue());
794 // If TI is a conditional eq, the default case is the false case,
795 // and the corresponding branch-weight data is at index 2. We swap the
796 // default weight to be the first entry.
797 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
798 assert(Weights.size() == 2);
799 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
800 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
801 std::swap(Weights.front(), Weights.back());
805 /// Keep halving the weights until all can fit in uint32_t.
806 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
807 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
808 if (Max > UINT_MAX) {
809 unsigned Offset = 32 - countLeadingZeros(Max);
810 for (uint64_t &I : Weights)
815 /// The specified terminator is a value equality comparison instruction
816 /// (either a switch or a branch on "X == c").
817 /// See if any of the predecessors of the terminator block are value comparisons
818 /// on the same value. If so, and if safe to do so, fold them together.
819 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
820 IRBuilder<> &Builder) {
821 BasicBlock *BB = TI->getParent();
822 Value *CV = isValueEqualityComparison(TI); // CondVal
823 assert(CV && "Not a comparison?");
824 bool Changed = false;
826 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
827 while (!Preds.empty()) {
828 BasicBlock *Pred = Preds.pop_back_val();
830 // See if the predecessor is a comparison with the same value.
831 TerminatorInst *PTI = Pred->getTerminator();
832 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
834 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
835 // Figure out which 'cases' to copy from SI to PSI.
836 std::vector<ValueEqualityComparisonCase> BBCases;
837 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
839 std::vector<ValueEqualityComparisonCase> PredCases;
840 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
842 // Based on whether the default edge from PTI goes to BB or not, fill in
843 // PredCases and PredDefault with the new switch cases we would like to
845 SmallVector<BasicBlock*, 8> NewSuccessors;
847 // Update the branch weight metadata along the way
848 SmallVector<uint64_t, 8> Weights;
849 bool PredHasWeights = HasBranchWeights(PTI);
850 bool SuccHasWeights = HasBranchWeights(TI);
852 if (PredHasWeights) {
853 GetBranchWeights(PTI, Weights);
854 // branch-weight metadata is inconsistent here.
855 if (Weights.size() != 1 + PredCases.size())
856 PredHasWeights = SuccHasWeights = false;
857 } else if (SuccHasWeights)
858 // If there are no predecessor weights but there are successor weights,
859 // populate Weights with 1, which will later be scaled to the sum of
860 // successor's weights
861 Weights.assign(1 + PredCases.size(), 1);
863 SmallVector<uint64_t, 8> SuccWeights;
864 if (SuccHasWeights) {
865 GetBranchWeights(TI, SuccWeights);
866 // branch-weight metadata is inconsistent here.
867 if (SuccWeights.size() != 1 + BBCases.size())
868 PredHasWeights = SuccHasWeights = false;
869 } else if (PredHasWeights)
870 SuccWeights.assign(1 + BBCases.size(), 1);
872 if (PredDefault == BB) {
873 // If this is the default destination from PTI, only the edges in TI
874 // that don't occur in PTI, or that branch to BB will be activated.
875 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
876 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
877 if (PredCases[i].Dest != BB)
878 PTIHandled.insert(PredCases[i].Value);
880 // The default destination is BB, we don't need explicit targets.
881 std::swap(PredCases[i], PredCases.back());
883 if (PredHasWeights || SuccHasWeights) {
884 // Increase weight for the default case.
885 Weights[0] += Weights[i+1];
886 std::swap(Weights[i+1], Weights.back());
890 PredCases.pop_back();
894 // Reconstruct the new switch statement we will be building.
895 if (PredDefault != BBDefault) {
896 PredDefault->removePredecessor(Pred);
897 PredDefault = BBDefault;
898 NewSuccessors.push_back(BBDefault);
901 unsigned CasesFromPred = Weights.size();
902 uint64_t ValidTotalSuccWeight = 0;
903 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
904 if (!PTIHandled.count(BBCases[i].Value) &&
905 BBCases[i].Dest != BBDefault) {
906 PredCases.push_back(BBCases[i]);
907 NewSuccessors.push_back(BBCases[i].Dest);
908 if (SuccHasWeights || PredHasWeights) {
909 // The default weight is at index 0, so weight for the ith case
910 // should be at index i+1. Scale the cases from successor by
911 // PredDefaultWeight (Weights[0]).
912 Weights.push_back(Weights[0] * SuccWeights[i+1]);
913 ValidTotalSuccWeight += SuccWeights[i+1];
917 if (SuccHasWeights || PredHasWeights) {
918 ValidTotalSuccWeight += SuccWeights[0];
919 // Scale the cases from predecessor by ValidTotalSuccWeight.
920 for (unsigned i = 1; i < CasesFromPred; ++i)
921 Weights[i] *= ValidTotalSuccWeight;
922 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
923 Weights[0] *= SuccWeights[0];
926 // If this is not the default destination from PSI, only the edges
927 // in SI that occur in PSI with a destination of BB will be
929 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
930 std::map<ConstantInt*, uint64_t> WeightsForHandled;
931 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
932 if (PredCases[i].Dest == BB) {
933 PTIHandled.insert(PredCases[i].Value);
935 if (PredHasWeights || SuccHasWeights) {
936 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
937 std::swap(Weights[i+1], Weights.back());
941 std::swap(PredCases[i], PredCases.back());
942 PredCases.pop_back();
946 // Okay, now we know which constants were sent to BB from the
947 // predecessor. Figure out where they will all go now.
948 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
949 if (PTIHandled.count(BBCases[i].Value)) {
950 // If this is one we are capable of getting...
951 if (PredHasWeights || SuccHasWeights)
952 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
953 PredCases.push_back(BBCases[i]);
954 NewSuccessors.push_back(BBCases[i].Dest);
955 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
958 // If there are any constants vectored to BB that TI doesn't handle,
959 // they must go to the default destination of TI.
960 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
962 E = PTIHandled.end(); I != E; ++I) {
963 if (PredHasWeights || SuccHasWeights)
964 Weights.push_back(WeightsForHandled[*I]);
965 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
966 NewSuccessors.push_back(BBDefault);
970 // Okay, at this point, we know which new successor Pred will get. Make
971 // sure we update the number of entries in the PHI nodes for these
973 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i)
974 AddPredecessorToBlock(NewSuccessors[i], Pred, BB);
976 Builder.SetInsertPoint(PTI);
977 // Convert pointer to int before we switch.
978 if (CV->getType()->isPointerTy()) {
979 CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
983 // Now that the successors are updated, create the new Switch instruction.
984 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
986 NewSI->setDebugLoc(PTI->getDebugLoc());
987 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
988 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest);
990 if (PredHasWeights || SuccHasWeights) {
991 // Halve the weights if any of them cannot fit in an uint32_t
994 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
996 NewSI->setMetadata(LLVMContext::MD_prof,
997 MDBuilder(BB->getContext()).
998 createBranchWeights(MDWeights));
1001 EraseTerminatorInstAndDCECond(PTI);
1003 // Okay, last check. If BB is still a successor of PSI, then we must
1004 // have an infinite loop case. If so, add an infinitely looping block
1005 // to handle the case to preserve the behavior of the code.
1006 BasicBlock *InfLoopBlock = nullptr;
1007 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1008 if (NewSI->getSuccessor(i) == BB) {
1009 if (!InfLoopBlock) {
1010 // Insert it at the end of the function, because it's either code,
1011 // or it won't matter if it's hot. :)
1012 InfLoopBlock = BasicBlock::Create(BB->getContext(),
1013 "infloop", BB->getParent());
1014 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1016 NewSI->setSuccessor(i, InfLoopBlock);
1025 // If we would need to insert a select that uses the value of this invoke
1026 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1027 // can't hoist the invoke, as there is nowhere to put the select in this case.
1028 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1029 Instruction *I1, Instruction *I2) {
1030 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1032 for (BasicBlock::iterator BBI = SI->begin();
1033 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1034 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1035 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1036 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
1044 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I);
1046 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1047 /// in the two blocks up into the branch block. The caller of this function
1048 /// guarantees that BI's block dominates BB1 and BB2.
1049 static bool HoistThenElseCodeToIf(BranchInst *BI,
1050 const TargetTransformInfo &TTI) {
1051 // This does very trivial matching, with limited scanning, to find identical
1052 // instructions in the two blocks. In particular, we don't want to get into
1053 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1054 // such, we currently just scan for obviously identical instructions in an
1056 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1057 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1059 BasicBlock::iterator BB1_Itr = BB1->begin();
1060 BasicBlock::iterator BB2_Itr = BB2->begin();
1062 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++;
1063 // Skip debug info if it is not identical.
1064 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1065 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1066 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1067 while (isa<DbgInfoIntrinsic>(I1))
1069 while (isa<DbgInfoIntrinsic>(I2))
1072 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1073 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1076 BasicBlock *BIParent = BI->getParent();
1078 bool Changed = false;
1080 // If we are hoisting the terminator instruction, don't move one (making a
1081 // broken BB), instead clone it, and remove BI.
1082 if (isa<TerminatorInst>(I1))
1083 goto HoistTerminator;
1085 if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1088 // For a normal instruction, we just move one to right before the branch,
1089 // then replace all uses of the other with the first. Finally, we remove
1090 // the now redundant second instruction.
1091 BIParent->getInstList().splice(BI, BB1->getInstList(), I1);
1092 if (!I2->use_empty())
1093 I2->replaceAllUsesWith(I1);
1094 I1->intersectOptionalDataWith(I2);
1095 unsigned KnownIDs[] = {
1096 LLVMContext::MD_tbaa,
1097 LLVMContext::MD_range,
1098 LLVMContext::MD_fpmath,
1099 LLVMContext::MD_invariant_load,
1100 LLVMContext::MD_nonnull
1102 combineMetadata(I1, I2, KnownIDs);
1103 I2->eraseFromParent();
1108 // Skip debug info if it is not identical.
1109 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1110 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1111 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1112 while (isa<DbgInfoIntrinsic>(I1))
1114 while (isa<DbgInfoIntrinsic>(I2))
1117 } while (I1->isIdenticalToWhenDefined(I2));
1122 // It may not be possible to hoist an invoke.
1123 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1126 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1128 for (BasicBlock::iterator BBI = SI->begin();
1129 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1130 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1131 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1135 // Check for passingValueIsAlwaysUndefined here because we would rather
1136 // eliminate undefined control flow then converting it to a select.
1137 if (passingValueIsAlwaysUndefined(BB1V, PN) ||
1138 passingValueIsAlwaysUndefined(BB2V, PN))
1141 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1143 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1148 // Okay, it is safe to hoist the terminator.
1149 Instruction *NT = I1->clone();
1150 BIParent->getInstList().insert(BI, NT);
1151 if (!NT->getType()->isVoidTy()) {
1152 I1->replaceAllUsesWith(NT);
1153 I2->replaceAllUsesWith(NT);
1157 IRBuilder<true, NoFolder> Builder(NT);
1158 // Hoisting one of the terminators from our successor is a great thing.
1159 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1160 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1161 // nodes, so we insert select instruction to compute the final result.
1162 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1163 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1165 for (BasicBlock::iterator BBI = SI->begin();
1166 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1167 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1168 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1169 if (BB1V == BB2V) continue;
1171 // These values do not agree. Insert a select instruction before NT
1172 // that determines the right value.
1173 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1175 SI = cast<SelectInst>
1176 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1177 BB1V->getName()+"."+BB2V->getName()));
1179 // Make the PHI node use the select for all incoming values for BB1/BB2
1180 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1181 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1182 PN->setIncomingValue(i, SI);
1186 // Update any PHI nodes in our new successors.
1187 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1188 AddPredecessorToBlock(*SI, BIParent, BB1);
1190 EraseTerminatorInstAndDCECond(BI);
1194 /// Given an unconditional branch that goes to BBEnd,
1195 /// check whether BBEnd has only two predecessors and the other predecessor
1196 /// ends with an unconditional branch. If it is true, sink any common code
1197 /// in the two predecessors to BBEnd.
1198 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1199 assert(BI1->isUnconditional());
1200 BasicBlock *BB1 = BI1->getParent();
1201 BasicBlock *BBEnd = BI1->getSuccessor(0);
1203 // Check that BBEnd has two predecessors and the other predecessor ends with
1204 // an unconditional branch.
1205 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1206 BasicBlock *Pred0 = *PI++;
1207 if (PI == PE) // Only one predecessor.
1209 BasicBlock *Pred1 = *PI++;
1210 if (PI != PE) // More than two predecessors.
1212 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1213 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1214 if (!BI2 || !BI2->isUnconditional())
1217 // Gather the PHI nodes in BBEnd.
1218 SmallDenseMap<std::pair<Value *, Value *>, PHINode *> JointValueMap;
1219 Instruction *FirstNonPhiInBBEnd = nullptr;
1220 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end(); I != E; ++I) {
1221 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1222 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1223 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1224 JointValueMap[std::make_pair(BB1V, BB2V)] = PN;
1226 FirstNonPhiInBBEnd = &*I;
1230 if (!FirstNonPhiInBBEnd)
1233 // This does very trivial matching, with limited scanning, to find identical
1234 // instructions in the two blocks. We scan backward for obviously identical
1235 // instructions in an identical order.
1236 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1237 RE1 = BB1->getInstList().rend(),
1238 RI2 = BB2->getInstList().rbegin(),
1239 RE2 = BB2->getInstList().rend();
1241 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1244 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1247 // Skip the unconditional branches.
1251 bool Changed = false;
1252 while (RI1 != RE1 && RI2 != RE2) {
1254 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1257 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1261 Instruction *I1 = &*RI1, *I2 = &*RI2;
1262 auto InstPair = std::make_pair(I1, I2);
1263 // I1 and I2 should have a single use in the same PHI node, and they
1264 // perform the same operation.
1265 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1266 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1267 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1268 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) ||
1269 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1270 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1271 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1272 !I1->hasOneUse() || !I2->hasOneUse() ||
1273 !JointValueMap.count(InstPair))
1276 // Check whether we should swap the operands of ICmpInst.
1277 // TODO: Add support of communativity.
1278 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1279 bool SwapOpnds = false;
1280 if (ICmp1 && ICmp2 &&
1281 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1282 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1283 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1284 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1285 ICmp2->swapOperands();
1288 if (!I1->isSameOperationAs(I2)) {
1290 ICmp2->swapOperands();
1294 // The operands should be either the same or they need to be generated
1295 // with a PHI node after sinking. We only handle the case where there is
1296 // a single pair of different operands.
1297 Value *DifferentOp1 = nullptr, *DifferentOp2 = nullptr;
1298 unsigned Op1Idx = ~0U;
1299 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1300 if (I1->getOperand(I) == I2->getOperand(I))
1302 // Early exit if we have more-than one pair of different operands or if
1303 // we need a PHI node to replace a constant.
1304 if (Op1Idx != ~0U ||
1305 isa<Constant>(I1->getOperand(I)) ||
1306 isa<Constant>(I2->getOperand(I))) {
1307 // If we can't sink the instructions, undo the swapping.
1309 ICmp2->swapOperands();
1312 DifferentOp1 = I1->getOperand(I);
1314 DifferentOp2 = I2->getOperand(I);
1317 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n");
1318 DEBUG(dbgs() << " " << *I2 << "\n");
1320 // We insert the pair of different operands to JointValueMap and
1321 // remove (I1, I2) from JointValueMap.
1322 if (Op1Idx != ~0U) {
1323 auto &NewPN = JointValueMap[std::make_pair(DifferentOp1, DifferentOp2)];
1326 PHINode::Create(DifferentOp1->getType(), 2,
1327 DifferentOp1->getName() + ".sink", BBEnd->begin());
1328 NewPN->addIncoming(DifferentOp1, BB1);
1329 NewPN->addIncoming(DifferentOp2, BB2);
1330 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1332 // I1 should use NewPN instead of DifferentOp1.
1333 I1->setOperand(Op1Idx, NewPN);
1335 PHINode *OldPN = JointValueMap[InstPair];
1336 JointValueMap.erase(InstPair);
1338 // We need to update RE1 and RE2 if we are going to sink the first
1339 // instruction in the basic block down.
1340 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1341 // Sink the instruction.
1342 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1);
1343 if (!OldPN->use_empty())
1344 OldPN->replaceAllUsesWith(I1);
1345 OldPN->eraseFromParent();
1347 if (!I2->use_empty())
1348 I2->replaceAllUsesWith(I1);
1349 I1->intersectOptionalDataWith(I2);
1350 // TODO: Use combineMetadata here to preserve what metadata we can
1351 // (analogous to the hoisting case above).
1352 I2->eraseFromParent();
1355 RE1 = BB1->getInstList().rend();
1357 RE2 = BB2->getInstList().rend();
1358 FirstNonPhiInBBEnd = I1;
1365 /// \brief Determine if we can hoist sink a sole store instruction out of a
1366 /// conditional block.
1368 /// We are looking for code like the following:
1370 /// store i32 %add, i32* %arrayidx2
1371 /// ... // No other stores or function calls (we could be calling a memory
1372 /// ... // function).
1373 /// %cmp = icmp ult %x, %y
1374 /// br i1 %cmp, label %EndBB, label %ThenBB
1376 /// store i32 %add5, i32* %arrayidx2
1380 /// We are going to transform this into:
1382 /// store i32 %add, i32* %arrayidx2
1384 /// %cmp = icmp ult %x, %y
1385 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1386 /// store i32 %add.add5, i32* %arrayidx2
1389 /// \return The pointer to the value of the previous store if the store can be
1390 /// hoisted into the predecessor block. 0 otherwise.
1391 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1392 BasicBlock *StoreBB, BasicBlock *EndBB) {
1393 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1397 // Volatile or atomic.
1398 if (!StoreToHoist->isSimple())
1401 Value *StorePtr = StoreToHoist->getPointerOperand();
1403 // Look for a store to the same pointer in BrBB.
1404 unsigned MaxNumInstToLookAt = 10;
1405 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(),
1406 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) {
1407 Instruction *CurI = &*RI;
1409 // Could be calling an instruction that effects memory like free().
1410 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI))
1413 StoreInst *SI = dyn_cast<StoreInst>(CurI);
1414 // Found the previous store make sure it stores to the same location.
1415 if (SI && SI->getPointerOperand() == StorePtr)
1416 // Found the previous store, return its value operand.
1417 return SI->getValueOperand();
1419 return nullptr; // Unknown store.
1425 /// \brief Speculate a conditional basic block flattening the CFG.
1427 /// Note that this is a very risky transform currently. Speculating
1428 /// instructions like this is most often not desirable. Instead, there is an MI
1429 /// pass which can do it with full awareness of the resource constraints.
1430 /// However, some cases are "obvious" and we should do directly. An example of
1431 /// this is speculating a single, reasonably cheap instruction.
1433 /// There is only one distinct advantage to flattening the CFG at the IR level:
1434 /// it makes very common but simplistic optimizations such as are common in
1435 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1436 /// modeling their effects with easier to reason about SSA value graphs.
1439 /// An illustration of this transform is turning this IR:
1442 /// %cmp = icmp ult %x, %y
1443 /// br i1 %cmp, label %EndBB, label %ThenBB
1445 /// %sub = sub %x, %y
1448 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1455 /// %cmp = icmp ult %x, %y
1456 /// %sub = sub %x, %y
1457 /// %cond = select i1 %cmp, 0, %sub
1461 /// \returns true if the conditional block is removed.
1462 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1463 const TargetTransformInfo &TTI) {
1464 // Be conservative for now. FP select instruction can often be expensive.
1465 Value *BrCond = BI->getCondition();
1466 if (isa<FCmpInst>(BrCond))
1469 BasicBlock *BB = BI->getParent();
1470 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1472 // If ThenBB is actually on the false edge of the conditional branch, remember
1473 // to swap the select operands later.
1474 bool Invert = false;
1475 if (ThenBB != BI->getSuccessor(0)) {
1476 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1479 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1481 // Keep a count of how many times instructions are used within CondBB when
1482 // they are candidates for sinking into CondBB. Specifically:
1483 // - They are defined in BB, and
1484 // - They have no side effects, and
1485 // - All of their uses are in CondBB.
1486 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1488 unsigned SpeculationCost = 0;
1489 Value *SpeculatedStoreValue = nullptr;
1490 StoreInst *SpeculatedStore = nullptr;
1491 for (BasicBlock::iterator BBI = ThenBB->begin(),
1492 BBE = std::prev(ThenBB->end());
1493 BBI != BBE; ++BBI) {
1494 Instruction *I = BBI;
1496 if (isa<DbgInfoIntrinsic>(I))
1499 // Only speculatively execute a single instruction (not counting the
1500 // terminator) for now.
1502 if (SpeculationCost > 1)
1505 // Don't hoist the instruction if it's unsafe or expensive.
1506 if (!isSafeToSpeculativelyExecute(I) &&
1507 !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1508 I, BB, ThenBB, EndBB))))
1510 if (!SpeculatedStoreValue &&
1511 ComputeSpeculationCost(I, TTI) >
1512 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
1515 // Store the store speculation candidate.
1516 if (SpeculatedStoreValue)
1517 SpeculatedStore = cast<StoreInst>(I);
1519 // Do not hoist the instruction if any of its operands are defined but not
1520 // used in BB. The transformation will prevent the operand from
1521 // being sunk into the use block.
1522 for (User::op_iterator i = I->op_begin(), e = I->op_end();
1524 Instruction *OpI = dyn_cast<Instruction>(*i);
1525 if (!OpI || OpI->getParent() != BB ||
1526 OpI->mayHaveSideEffects())
1527 continue; // Not a candidate for sinking.
1529 ++SinkCandidateUseCounts[OpI];
1533 // Consider any sink candidates which are only used in CondBB as costs for
1534 // speculation. Note, while we iterate over a DenseMap here, we are summing
1535 // and so iteration order isn't significant.
1536 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I =
1537 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end();
1539 if (I->first->getNumUses() == I->second) {
1541 if (SpeculationCost > 1)
1545 // Check that the PHI nodes can be converted to selects.
1546 bool HaveRewritablePHIs = false;
1547 for (BasicBlock::iterator I = EndBB->begin();
1548 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1549 Value *OrigV = PN->getIncomingValueForBlock(BB);
1550 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1552 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1553 // Skip PHIs which are trivial.
1557 // Don't convert to selects if we could remove undefined behavior instead.
1558 if (passingValueIsAlwaysUndefined(OrigV, PN) ||
1559 passingValueIsAlwaysUndefined(ThenV, PN))
1562 HaveRewritablePHIs = true;
1563 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1564 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1565 if (!OrigCE && !ThenCE)
1566 continue; // Known safe and cheap.
1568 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
1569 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
1571 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
1572 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
1573 unsigned MaxCost = 2 * PHINodeFoldingThreshold *
1574 TargetTransformInfo::TCC_Basic;
1575 if (OrigCost + ThenCost > MaxCost)
1578 // Account for the cost of an unfolded ConstantExpr which could end up
1579 // getting expanded into Instructions.
1580 // FIXME: This doesn't account for how many operations are combined in the
1581 // constant expression.
1583 if (SpeculationCost > 1)
1587 // If there are no PHIs to process, bail early. This helps ensure idempotence
1589 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
1592 // If we get here, we can hoist the instruction and if-convert.
1593 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1595 // Insert a select of the value of the speculated store.
1596 if (SpeculatedStoreValue) {
1597 IRBuilder<true, NoFolder> Builder(BI);
1598 Value *TrueV = SpeculatedStore->getValueOperand();
1599 Value *FalseV = SpeculatedStoreValue;
1601 std::swap(TrueV, FalseV);
1602 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() +
1603 "." + FalseV->getName());
1604 SpeculatedStore->setOperand(0, S);
1607 // Hoist the instructions.
1608 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(),
1609 std::prev(ThenBB->end()));
1611 // Insert selects and rewrite the PHI operands.
1612 IRBuilder<true, NoFolder> Builder(BI);
1613 for (BasicBlock::iterator I = EndBB->begin();
1614 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1615 unsigned OrigI = PN->getBasicBlockIndex(BB);
1616 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1617 Value *OrigV = PN->getIncomingValue(OrigI);
1618 Value *ThenV = PN->getIncomingValue(ThenI);
1620 // Skip PHIs which are trivial.
1624 // Create a select whose true value is the speculatively executed value and
1625 // false value is the preexisting value. Swap them if the branch
1626 // destinations were inverted.
1627 Value *TrueV = ThenV, *FalseV = OrigV;
1629 std::swap(TrueV, FalseV);
1630 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV,
1631 TrueV->getName() + "." + FalseV->getName());
1632 PN->setIncomingValue(OrigI, V);
1633 PN->setIncomingValue(ThenI, V);
1640 /// \returns True if this block contains a CallInst with the NoDuplicate
1642 static bool HasNoDuplicateCall(const BasicBlock *BB) {
1643 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1644 const CallInst *CI = dyn_cast<CallInst>(I);
1647 if (CI->cannotDuplicate())
1653 /// Return true if we can thread a branch across this block.
1654 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1655 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1658 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1659 if (isa<DbgInfoIntrinsic>(BBI))
1661 if (Size > 10) return false; // Don't clone large BB's.
1664 // We can only support instructions that do not define values that are
1665 // live outside of the current basic block.
1666 for (User *U : BBI->users()) {
1667 Instruction *UI = cast<Instruction>(U);
1668 if (UI->getParent() != BB || isa<PHINode>(UI)) return false;
1671 // Looks ok, continue checking.
1677 /// If we have a conditional branch on a PHI node value that is defined in the
1678 /// same block as the branch and if any PHI entries are constants, thread edges
1679 /// corresponding to that entry to be branches to their ultimate destination.
1680 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL) {
1681 BasicBlock *BB = BI->getParent();
1682 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1683 // NOTE: we currently cannot transform this case if the PHI node is used
1684 // outside of the block.
1685 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1688 // Degenerate case of a single entry PHI.
1689 if (PN->getNumIncomingValues() == 1) {
1690 FoldSingleEntryPHINodes(PN->getParent());
1694 // Now we know that this block has multiple preds and two succs.
1695 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1697 if (HasNoDuplicateCall(BB)) return false;
1699 // Okay, this is a simple enough basic block. See if any phi values are
1701 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1702 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1703 if (!CB || !CB->getType()->isIntegerTy(1)) continue;
1705 // Okay, we now know that all edges from PredBB should be revectored to
1706 // branch to RealDest.
1707 BasicBlock *PredBB = PN->getIncomingBlock(i);
1708 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1710 if (RealDest == BB) continue; // Skip self loops.
1711 // Skip if the predecessor's terminator is an indirect branch.
1712 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1714 // The dest block might have PHI nodes, other predecessors and other
1715 // difficult cases. Instead of being smart about this, just insert a new
1716 // block that jumps to the destination block, effectively splitting
1717 // the edge we are about to create.
1718 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1719 RealDest->getName()+".critedge",
1720 RealDest->getParent(), RealDest);
1721 BranchInst::Create(RealDest, EdgeBB);
1723 // Update PHI nodes.
1724 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1726 // BB may have instructions that are being threaded over. Clone these
1727 // instructions into EdgeBB. We know that there will be no uses of the
1728 // cloned instructions outside of EdgeBB.
1729 BasicBlock::iterator InsertPt = EdgeBB->begin();
1730 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1731 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1732 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1733 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1736 // Clone the instruction.
1737 Instruction *N = BBI->clone();
1738 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1740 // Update operands due to translation.
1741 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1743 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1744 if (PI != TranslateMap.end())
1748 // Check for trivial simplification.
1749 if (Value *V = SimplifyInstruction(N, DL)) {
1750 TranslateMap[BBI] = V;
1751 delete N; // Instruction folded away, don't need actual inst
1753 // Insert the new instruction into its new home.
1754 EdgeBB->getInstList().insert(InsertPt, N);
1755 if (!BBI->use_empty())
1756 TranslateMap[BBI] = N;
1760 // Loop over all of the edges from PredBB to BB, changing them to branch
1761 // to EdgeBB instead.
1762 TerminatorInst *PredBBTI = PredBB->getTerminator();
1763 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1764 if (PredBBTI->getSuccessor(i) == BB) {
1765 BB->removePredecessor(PredBB);
1766 PredBBTI->setSuccessor(i, EdgeBB);
1769 // Recurse, simplifying any other constants.
1770 return FoldCondBranchOnPHI(BI, DL) | true;
1776 /// Given a BB that starts with the specified two-entry PHI node,
1777 /// see if we can eliminate it.
1778 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI,
1779 const DataLayout &DL) {
1780 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1781 // statement", which has a very simple dominance structure. Basically, we
1782 // are trying to find the condition that is being branched on, which
1783 // subsequently causes this merge to happen. We really want control
1784 // dependence information for this check, but simplifycfg can't keep it up
1785 // to date, and this catches most of the cases we care about anyway.
1786 BasicBlock *BB = PN->getParent();
1787 BasicBlock *IfTrue, *IfFalse;
1788 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1790 // Don't bother if the branch will be constant folded trivially.
1791 isa<ConstantInt>(IfCond))
1794 // Okay, we found that we can merge this two-entry phi node into a select.
1795 // Doing so would require us to fold *all* two entry phi nodes in this block.
1796 // At some point this becomes non-profitable (particularly if the target
1797 // doesn't support cmov's). Only do this transformation if there are two or
1798 // fewer PHI nodes in this block.
1799 unsigned NumPhis = 0;
1800 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1804 // Loop over the PHI's seeing if we can promote them all to select
1805 // instructions. While we are at it, keep track of the instructions
1806 // that need to be moved to the dominating block.
1807 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1808 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1809 MaxCostVal1 = PHINodeFoldingThreshold;
1810 MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
1811 MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
1813 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1814 PHINode *PN = cast<PHINode>(II++);
1815 if (Value *V = SimplifyInstruction(PN, DL)) {
1816 PN->replaceAllUsesWith(V);
1817 PN->eraseFromParent();
1821 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1822 MaxCostVal0, TTI) ||
1823 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1828 // If we folded the first phi, PN dangles at this point. Refresh it. If
1829 // we ran out of PHIs then we simplified them all.
1830 PN = dyn_cast<PHINode>(BB->begin());
1831 if (!PN) return true;
1833 // Don't fold i1 branches on PHIs which contain binary operators. These can
1834 // often be turned into switches and other things.
1835 if (PN->getType()->isIntegerTy(1) &&
1836 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1837 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1838 isa<BinaryOperator>(IfCond)))
1841 // If we all PHI nodes are promotable, check to make sure that all
1842 // instructions in the predecessor blocks can be promoted as well. If
1843 // not, we won't be able to get rid of the control flow, so it's not
1844 // worth promoting to select instructions.
1845 BasicBlock *DomBlock = nullptr;
1846 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1847 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1848 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1851 DomBlock = *pred_begin(IfBlock1);
1852 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1853 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1854 // This is not an aggressive instruction that we can promote.
1855 // Because of this, we won't be able to get rid of the control
1856 // flow, so the xform is not worth it.
1861 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1864 DomBlock = *pred_begin(IfBlock2);
1865 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1866 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1867 // This is not an aggressive instruction that we can promote.
1868 // Because of this, we won't be able to get rid of the control
1869 // flow, so the xform is not worth it.
1874 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1875 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1877 // If we can still promote the PHI nodes after this gauntlet of tests,
1878 // do all of the PHI's now.
1879 Instruction *InsertPt = DomBlock->getTerminator();
1880 IRBuilder<true, NoFolder> Builder(InsertPt);
1882 // Move all 'aggressive' instructions, which are defined in the
1883 // conditional parts of the if's up to the dominating block.
1885 DomBlock->getInstList().splice(InsertPt,
1886 IfBlock1->getInstList(), IfBlock1->begin(),
1887 IfBlock1->getTerminator());
1889 DomBlock->getInstList().splice(InsertPt,
1890 IfBlock2->getInstList(), IfBlock2->begin(),
1891 IfBlock2->getTerminator());
1893 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1894 // Change the PHI node into a select instruction.
1895 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1896 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1899 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1900 PN->replaceAllUsesWith(NV);
1902 PN->eraseFromParent();
1905 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1906 // has been flattened. Change DomBlock to jump directly to our new block to
1907 // avoid other simplifycfg's kicking in on the diamond.
1908 TerminatorInst *OldTI = DomBlock->getTerminator();
1909 Builder.SetInsertPoint(OldTI);
1910 Builder.CreateBr(BB);
1911 OldTI->eraseFromParent();
1915 /// If we found a conditional branch that goes to two returning blocks,
1916 /// try to merge them together into one return,
1917 /// introducing a select if the return values disagree.
1918 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1919 IRBuilder<> &Builder) {
1920 assert(BI->isConditional() && "Must be a conditional branch");
1921 BasicBlock *TrueSucc = BI->getSuccessor(0);
1922 BasicBlock *FalseSucc = BI->getSuccessor(1);
1923 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1924 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1926 // Check to ensure both blocks are empty (just a return) or optionally empty
1927 // with PHI nodes. If there are other instructions, merging would cause extra
1928 // computation on one path or the other.
1929 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1931 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1934 Builder.SetInsertPoint(BI);
1935 // Okay, we found a branch that is going to two return nodes. If
1936 // there is no return value for this function, just change the
1937 // branch into a return.
1938 if (FalseRet->getNumOperands() == 0) {
1939 TrueSucc->removePredecessor(BI->getParent());
1940 FalseSucc->removePredecessor(BI->getParent());
1941 Builder.CreateRetVoid();
1942 EraseTerminatorInstAndDCECond(BI);
1946 // Otherwise, figure out what the true and false return values are
1947 // so we can insert a new select instruction.
1948 Value *TrueValue = TrueRet->getReturnValue();
1949 Value *FalseValue = FalseRet->getReturnValue();
1951 // Unwrap any PHI nodes in the return blocks.
1952 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1953 if (TVPN->getParent() == TrueSucc)
1954 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1955 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1956 if (FVPN->getParent() == FalseSucc)
1957 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1959 // In order for this transformation to be safe, we must be able to
1960 // unconditionally execute both operands to the return. This is
1961 // normally the case, but we could have a potentially-trapping
1962 // constant expression that prevents this transformation from being
1964 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1967 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1971 // Okay, we collected all the mapped values and checked them for sanity, and
1972 // defined to really do this transformation. First, update the CFG.
1973 TrueSucc->removePredecessor(BI->getParent());
1974 FalseSucc->removePredecessor(BI->getParent());
1976 // Insert select instructions where needed.
1977 Value *BrCond = BI->getCondition();
1979 // Insert a select if the results differ.
1980 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1981 } else if (isa<UndefValue>(TrueValue)) {
1982 TrueValue = FalseValue;
1984 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1985 FalseValue, "retval");
1989 Value *RI = !TrueValue ?
1990 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
1994 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
1995 << "\n " << *BI << "NewRet = " << *RI
1996 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
1998 EraseTerminatorInstAndDCECond(BI);
2003 /// Given a conditional BranchInstruction, retrieve the probabilities of the
2004 /// branch taking each edge. Fills in the two APInt parameters and returns true,
2005 /// or returns false if no or invalid metadata was found.
2006 static bool ExtractBranchMetadata(BranchInst *BI,
2007 uint64_t &ProbTrue, uint64_t &ProbFalse) {
2008 assert(BI->isConditional() &&
2009 "Looking for probabilities on unconditional branch?");
2010 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
2011 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
2012 ConstantInt *CITrue =
2013 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
2014 ConstantInt *CIFalse =
2015 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
2016 if (!CITrue || !CIFalse) return false;
2017 ProbTrue = CITrue->getValue().getZExtValue();
2018 ProbFalse = CIFalse->getValue().getZExtValue();
2022 /// Return true if the given instruction is available
2023 /// in its predecessor block. If yes, the instruction will be removed.
2024 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
2025 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2027 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
2028 Instruction *PBI = &*I;
2029 // Check whether Inst and PBI generate the same value.
2030 if (Inst->isIdenticalTo(PBI)) {
2031 Inst->replaceAllUsesWith(PBI);
2032 Inst->eraseFromParent();
2039 /// If this basic block is simple enough, and if a predecessor branches to us
2040 /// and one of our successors, fold the block into the predecessor and use
2041 /// logical operations to pick the right destination.
2042 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2043 BasicBlock *BB = BI->getParent();
2045 Instruction *Cond = nullptr;
2046 if (BI->isConditional())
2047 Cond = dyn_cast<Instruction>(BI->getCondition());
2049 // For unconditional branch, check for a simple CFG pattern, where
2050 // BB has a single predecessor and BB's successor is also its predecessor's
2051 // successor. If such pattern exisits, check for CSE between BB and its
2053 if (BasicBlock *PB = BB->getSinglePredecessor())
2054 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2055 if (PBI->isConditional() &&
2056 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2057 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2058 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
2060 Instruction *Curr = I++;
2061 if (isa<CmpInst>(Curr)) {
2065 // Quit if we can't remove this instruction.
2066 if (!checkCSEInPredecessor(Curr, PB))
2075 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2076 Cond->getParent() != BB || !Cond->hasOneUse())
2079 // Make sure the instruction after the condition is the cond branch.
2080 BasicBlock::iterator CondIt = Cond; ++CondIt;
2082 // Ignore dbg intrinsics.
2083 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
2088 // Only allow this transformation if computing the condition doesn't involve
2089 // too many instructions and these involved instructions can be executed
2090 // unconditionally. We denote all involved instructions except the condition
2091 // as "bonus instructions", and only allow this transformation when the
2092 // number of the bonus instructions does not exceed a certain threshold.
2093 unsigned NumBonusInsts = 0;
2094 for (auto I = BB->begin(); Cond != I; ++I) {
2095 // Ignore dbg intrinsics.
2096 if (isa<DbgInfoIntrinsic>(I))
2098 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(I))
2100 // I has only one use and can be executed unconditionally.
2101 Instruction *User = dyn_cast<Instruction>(I->user_back());
2102 if (User == nullptr || User->getParent() != BB)
2104 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2105 // to use any other instruction, User must be an instruction between next(I)
2108 // Early exits once we reach the limit.
2109 if (NumBonusInsts > BonusInstThreshold)
2113 // Cond is known to be a compare or binary operator. Check to make sure that
2114 // neither operand is a potentially-trapping constant expression.
2115 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2118 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2122 // Finally, don't infinitely unroll conditional loops.
2123 BasicBlock *TrueDest = BI->getSuccessor(0);
2124 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2125 if (TrueDest == BB || FalseDest == BB)
2128 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2129 BasicBlock *PredBlock = *PI;
2130 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2132 // Check that we have two conditional branches. If there is a PHI node in
2133 // the common successor, verify that the same value flows in from both
2135 SmallVector<PHINode*, 4> PHIs;
2136 if (!PBI || PBI->isUnconditional() ||
2137 (BI->isConditional() &&
2138 !SafeToMergeTerminators(BI, PBI)) ||
2139 (!BI->isConditional() &&
2140 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2143 // Determine if the two branches share a common destination.
2144 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2145 bool InvertPredCond = false;
2147 if (BI->isConditional()) {
2148 if (PBI->getSuccessor(0) == TrueDest)
2149 Opc = Instruction::Or;
2150 else if (PBI->getSuccessor(1) == FalseDest)
2151 Opc = Instruction::And;
2152 else if (PBI->getSuccessor(0) == FalseDest)
2153 Opc = Instruction::And, InvertPredCond = true;
2154 else if (PBI->getSuccessor(1) == TrueDest)
2155 Opc = Instruction::Or, InvertPredCond = true;
2159 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2163 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2164 IRBuilder<> Builder(PBI);
2166 // If we need to invert the condition in the pred block to match, do so now.
2167 if (InvertPredCond) {
2168 Value *NewCond = PBI->getCondition();
2170 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2171 CmpInst *CI = cast<CmpInst>(NewCond);
2172 CI->setPredicate(CI->getInversePredicate());
2174 NewCond = Builder.CreateNot(NewCond,
2175 PBI->getCondition()->getName()+".not");
2178 PBI->setCondition(NewCond);
2179 PBI->swapSuccessors();
2182 // If we have bonus instructions, clone them into the predecessor block.
2183 // Note that there may be multiple predecessor blocks, so we cannot move
2184 // bonus instructions to a predecessor block.
2185 ValueToValueMapTy VMap; // maps original values to cloned values
2186 // We already make sure Cond is the last instruction before BI. Therefore,
2187 // all instructions before Cond other than DbgInfoIntrinsic are bonus
2189 for (auto BonusInst = BB->begin(); Cond != BonusInst; ++BonusInst) {
2190 if (isa<DbgInfoIntrinsic>(BonusInst))
2192 Instruction *NewBonusInst = BonusInst->clone();
2193 RemapInstruction(NewBonusInst, VMap,
2194 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2195 VMap[BonusInst] = NewBonusInst;
2197 // If we moved a load, we cannot any longer claim any knowledge about
2198 // its potential value. The previous information might have been valid
2199 // only given the branch precondition.
2200 // For an analogous reason, we must also drop all the metadata whose
2201 // semantics we don't understand.
2202 NewBonusInst->dropUnknownMetadata(LLVMContext::MD_dbg);
2204 PredBlock->getInstList().insert(PBI, NewBonusInst);
2205 NewBonusInst->takeName(BonusInst);
2206 BonusInst->setName(BonusInst->getName() + ".old");
2209 // Clone Cond into the predecessor basic block, and or/and the
2210 // two conditions together.
2211 Instruction *New = Cond->clone();
2212 RemapInstruction(New, VMap,
2213 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2214 PredBlock->getInstList().insert(PBI, New);
2215 New->takeName(Cond);
2216 Cond->setName(New->getName() + ".old");
2218 if (BI->isConditional()) {
2219 Instruction *NewCond =
2220 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2222 PBI->setCondition(NewCond);
2224 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2225 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2227 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2229 SmallVector<uint64_t, 8> NewWeights;
2231 if (PBI->getSuccessor(0) == BB) {
2232 if (PredHasWeights && SuccHasWeights) {
2233 // PBI: br i1 %x, BB, FalseDest
2234 // BI: br i1 %y, TrueDest, FalseDest
2235 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2236 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2237 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2238 // TrueWeight for PBI * FalseWeight for BI.
2239 // We assume that total weights of a BranchInst can fit into 32 bits.
2240 // Therefore, we will not have overflow using 64-bit arithmetic.
2241 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2242 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2244 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2245 PBI->setSuccessor(0, TrueDest);
2247 if (PBI->getSuccessor(1) == BB) {
2248 if (PredHasWeights && SuccHasWeights) {
2249 // PBI: br i1 %x, TrueDest, BB
2250 // BI: br i1 %y, TrueDest, FalseDest
2251 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2252 // FalseWeight for PBI * TrueWeight for BI.
2253 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2254 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2255 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2256 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2258 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2259 PBI->setSuccessor(1, FalseDest);
2261 if (NewWeights.size() == 2) {
2262 // Halve the weights if any of them cannot fit in an uint32_t
2263 FitWeights(NewWeights);
2265 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2266 PBI->setMetadata(LLVMContext::MD_prof,
2267 MDBuilder(BI->getContext()).
2268 createBranchWeights(MDWeights));
2270 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2272 // Update PHI nodes in the common successors.
2273 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2274 ConstantInt *PBI_C = cast<ConstantInt>(
2275 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2276 assert(PBI_C->getType()->isIntegerTy(1));
2277 Instruction *MergedCond = nullptr;
2278 if (PBI->getSuccessor(0) == TrueDest) {
2279 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2280 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2281 // is false: !PBI_Cond and BI_Value
2282 Instruction *NotCond =
2283 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2286 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2291 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2292 PBI->getCondition(), MergedCond,
2295 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2296 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2297 // is false: PBI_Cond and BI_Value
2299 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2300 PBI->getCondition(), New,
2302 if (PBI_C->isOne()) {
2303 Instruction *NotCond =
2304 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2307 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2308 NotCond, MergedCond,
2313 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2316 // Change PBI from Conditional to Unconditional.
2317 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2318 EraseTerminatorInstAndDCECond(PBI);
2322 // TODO: If BB is reachable from all paths through PredBlock, then we
2323 // could replace PBI's branch probabilities with BI's.
2325 // Copy any debug value intrinsics into the end of PredBlock.
2326 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2327 if (isa<DbgInfoIntrinsic>(*I))
2328 I->clone()->insertBefore(PBI);
2335 /// If we have a conditional branch as a predecessor of another block,
2336 /// this function tries to simplify it. We know
2337 /// that PBI and BI are both conditional branches, and BI is in one of the
2338 /// successor blocks of PBI - PBI branches to BI.
2339 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2340 assert(PBI->isConditional() && BI->isConditional());
2341 BasicBlock *BB = BI->getParent();
2343 // If this block ends with a branch instruction, and if there is a
2344 // predecessor that ends on a branch of the same condition, make
2345 // this conditional branch redundant.
2346 if (PBI->getCondition() == BI->getCondition() &&
2347 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2348 // Okay, the outcome of this conditional branch is statically
2349 // knowable. If this block had a single pred, handle specially.
2350 if (BB->getSinglePredecessor()) {
2351 // Turn this into a branch on constant.
2352 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2353 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2355 return true; // Nuke the branch on constant.
2358 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2359 // in the constant and simplify the block result. Subsequent passes of
2360 // simplifycfg will thread the block.
2361 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2362 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2363 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2364 std::distance(PB, PE),
2365 BI->getCondition()->getName() + ".pr",
2367 // Okay, we're going to insert the PHI node. Since PBI is not the only
2368 // predecessor, compute the PHI'd conditional value for all of the preds.
2369 // Any predecessor where the condition is not computable we keep symbolic.
2370 for (pred_iterator PI = PB; PI != PE; ++PI) {
2371 BasicBlock *P = *PI;
2372 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2373 PBI != BI && PBI->isConditional() &&
2374 PBI->getCondition() == BI->getCondition() &&
2375 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2376 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2377 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2380 NewPN->addIncoming(BI->getCondition(), P);
2384 BI->setCondition(NewPN);
2389 // If this is a conditional branch in an empty block, and if any
2390 // predecessors are a conditional branch to one of our destinations,
2391 // fold the conditions into logical ops and one cond br.
2392 BasicBlock::iterator BBI = BB->begin();
2393 // Ignore dbg intrinsics.
2394 while (isa<DbgInfoIntrinsic>(BBI))
2400 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2405 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2407 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2408 PBIOp = 0, BIOp = 1;
2409 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2410 PBIOp = 1, BIOp = 0;
2411 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2416 // Check to make sure that the other destination of this branch
2417 // isn't BB itself. If so, this is an infinite loop that will
2418 // keep getting unwound.
2419 if (PBI->getSuccessor(PBIOp) == BB)
2422 // Do not perform this transformation if it would require
2423 // insertion of a large number of select instructions. For targets
2424 // without predication/cmovs, this is a big pessimization.
2426 // Also do not perform this transformation if any phi node in the common
2427 // destination block can trap when reached by BB or PBB (PR17073). In that
2428 // case, it would be unsafe to hoist the operation into a select instruction.
2430 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2431 unsigned NumPhis = 0;
2432 for (BasicBlock::iterator II = CommonDest->begin();
2433 isa<PHINode>(II); ++II, ++NumPhis) {
2434 if (NumPhis > 2) // Disable this xform.
2437 PHINode *PN = cast<PHINode>(II);
2438 Value *BIV = PN->getIncomingValueForBlock(BB);
2439 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
2443 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2444 Value *PBIV = PN->getIncomingValue(PBBIdx);
2445 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
2450 // Finally, if everything is ok, fold the branches to logical ops.
2451 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2453 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2454 << "AND: " << *BI->getParent());
2457 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2458 // branch in it, where one edge (OtherDest) goes back to itself but the other
2459 // exits. We don't *know* that the program avoids the infinite loop
2460 // (even though that seems likely). If we do this xform naively, we'll end up
2461 // recursively unpeeling the loop. Since we know that (after the xform is
2462 // done) that the block *is* infinite if reached, we just make it an obviously
2463 // infinite loop with no cond branch.
2464 if (OtherDest == BB) {
2465 // Insert it at the end of the function, because it's either code,
2466 // or it won't matter if it's hot. :)
2467 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2468 "infloop", BB->getParent());
2469 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2470 OtherDest = InfLoopBlock;
2473 DEBUG(dbgs() << *PBI->getParent()->getParent());
2475 // BI may have other predecessors. Because of this, we leave
2476 // it alone, but modify PBI.
2478 // Make sure we get to CommonDest on True&True directions.
2479 Value *PBICond = PBI->getCondition();
2480 IRBuilder<true, NoFolder> Builder(PBI);
2482 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2484 Value *BICond = BI->getCondition();
2486 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2488 // Merge the conditions.
2489 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2491 // Modify PBI to branch on the new condition to the new dests.
2492 PBI->setCondition(Cond);
2493 PBI->setSuccessor(0, CommonDest);
2494 PBI->setSuccessor(1, OtherDest);
2496 // Update branch weight for PBI.
2497 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2498 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2500 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2502 if (PredHasWeights && SuccHasWeights) {
2503 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2504 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2505 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2506 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2507 // The weight to CommonDest should be PredCommon * SuccTotal +
2508 // PredOther * SuccCommon.
2509 // The weight to OtherDest should be PredOther * SuccOther.
2510 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
2511 PredOther * SuccCommon,
2512 PredOther * SuccOther};
2513 // Halve the weights if any of them cannot fit in an uint32_t
2514 FitWeights(NewWeights);
2516 PBI->setMetadata(LLVMContext::MD_prof,
2517 MDBuilder(BI->getContext())
2518 .createBranchWeights(NewWeights[0], NewWeights[1]));
2521 // OtherDest may have phi nodes. If so, add an entry from PBI's
2522 // block that are identical to the entries for BI's block.
2523 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2525 // We know that the CommonDest already had an edge from PBI to
2526 // it. If it has PHIs though, the PHIs may have different
2527 // entries for BB and PBI's BB. If so, insert a select to make
2530 for (BasicBlock::iterator II = CommonDest->begin();
2531 (PN = dyn_cast<PHINode>(II)); ++II) {
2532 Value *BIV = PN->getIncomingValueForBlock(BB);
2533 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2534 Value *PBIV = PN->getIncomingValue(PBBIdx);
2536 // Insert a select in PBI to pick the right value.
2537 Value *NV = cast<SelectInst>
2538 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2539 PN->setIncomingValue(PBBIdx, NV);
2543 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2544 DEBUG(dbgs() << *PBI->getParent()->getParent());
2546 // This basic block is probably dead. We know it has at least
2547 // one fewer predecessor.
2551 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
2552 // true or to FalseBB if Cond is false.
2553 // Takes care of updating the successors and removing the old terminator.
2554 // Also makes sure not to introduce new successors by assuming that edges to
2555 // non-successor TrueBBs and FalseBBs aren't reachable.
2556 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2557 BasicBlock *TrueBB, BasicBlock *FalseBB,
2558 uint32_t TrueWeight,
2559 uint32_t FalseWeight){
2560 // Remove any superfluous successor edges from the CFG.
2561 // First, figure out which successors to preserve.
2562 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2564 BasicBlock *KeepEdge1 = TrueBB;
2565 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
2567 // Then remove the rest.
2568 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2569 BasicBlock *Succ = OldTerm->getSuccessor(I);
2570 // Make sure only to keep exactly one copy of each edge.
2571 if (Succ == KeepEdge1)
2572 KeepEdge1 = nullptr;
2573 else if (Succ == KeepEdge2)
2574 KeepEdge2 = nullptr;
2576 Succ->removePredecessor(OldTerm->getParent());
2579 IRBuilder<> Builder(OldTerm);
2580 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2582 // Insert an appropriate new terminator.
2583 if (!KeepEdge1 && !KeepEdge2) {
2584 if (TrueBB == FalseBB)
2585 // We were only looking for one successor, and it was present.
2586 // Create an unconditional branch to it.
2587 Builder.CreateBr(TrueBB);
2589 // We found both of the successors we were looking for.
2590 // Create a conditional branch sharing the condition of the select.
2591 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2592 if (TrueWeight != FalseWeight)
2593 NewBI->setMetadata(LLVMContext::MD_prof,
2594 MDBuilder(OldTerm->getContext()).
2595 createBranchWeights(TrueWeight, FalseWeight));
2597 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2598 // Neither of the selected blocks were successors, so this
2599 // terminator must be unreachable.
2600 new UnreachableInst(OldTerm->getContext(), OldTerm);
2602 // One of the selected values was a successor, but the other wasn't.
2603 // Insert an unconditional branch to the one that was found;
2604 // the edge to the one that wasn't must be unreachable.
2606 // Only TrueBB was found.
2607 Builder.CreateBr(TrueBB);
2609 // Only FalseBB was found.
2610 Builder.CreateBr(FalseBB);
2613 EraseTerminatorInstAndDCECond(OldTerm);
2618 // (switch (select cond, X, Y)) on constant X, Y
2619 // with a branch - conditional if X and Y lead to distinct BBs,
2620 // unconditional otherwise.
2621 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2622 // Check for constant integer values in the select.
2623 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2624 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2625 if (!TrueVal || !FalseVal)
2628 // Find the relevant condition and destinations.
2629 Value *Condition = Select->getCondition();
2630 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2631 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2633 // Get weight for TrueBB and FalseBB.
2634 uint32_t TrueWeight = 0, FalseWeight = 0;
2635 SmallVector<uint64_t, 8> Weights;
2636 bool HasWeights = HasBranchWeights(SI);
2638 GetBranchWeights(SI, Weights);
2639 if (Weights.size() == 1 + SI->getNumCases()) {
2640 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2641 getSuccessorIndex()];
2642 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2643 getSuccessorIndex()];
2647 // Perform the actual simplification.
2648 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2649 TrueWeight, FalseWeight);
2653 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2654 // blockaddress(@fn, BlockB)))
2656 // (br cond, BlockA, BlockB).
2657 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2658 // Check that both operands of the select are block addresses.
2659 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2660 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2664 // Extract the actual blocks.
2665 BasicBlock *TrueBB = TBA->getBasicBlock();
2666 BasicBlock *FalseBB = FBA->getBasicBlock();
2668 // Perform the actual simplification.
2669 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2673 /// This is called when we find an icmp instruction
2674 /// (a seteq/setne with a constant) as the only instruction in a
2675 /// block that ends with an uncond branch. We are looking for a very specific
2676 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2677 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2678 /// default value goes to an uncond block with a seteq in it, we get something
2681 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2683 /// %tmp = icmp eq i8 %A, 92
2686 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2688 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2689 /// the PHI, merging the third icmp into the switch.
2690 static bool TryToSimplifyUncondBranchWithICmpInIt(
2691 ICmpInst *ICI, IRBuilder<> &Builder, const DataLayout &DL,
2692 const TargetTransformInfo &TTI, unsigned BonusInstThreshold,
2693 AssumptionCache *AC) {
2694 BasicBlock *BB = ICI->getParent();
2696 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2698 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2700 Value *V = ICI->getOperand(0);
2701 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2703 // The pattern we're looking for is where our only predecessor is a switch on
2704 // 'V' and this block is the default case for the switch. In this case we can
2705 // fold the compared value into the switch to simplify things.
2706 BasicBlock *Pred = BB->getSinglePredecessor();
2707 if (!Pred || !isa<SwitchInst>(Pred->getTerminator())) return false;
2709 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2710 if (SI->getCondition() != V)
2713 // If BB is reachable on a non-default case, then we simply know the value of
2714 // V in this block. Substitute it and constant fold the icmp instruction
2716 if (SI->getDefaultDest() != BB) {
2717 ConstantInt *VVal = SI->findCaseDest(BB);
2718 assert(VVal && "Should have a unique destination value");
2719 ICI->setOperand(0, VVal);
2721 if (Value *V = SimplifyInstruction(ICI, DL)) {
2722 ICI->replaceAllUsesWith(V);
2723 ICI->eraseFromParent();
2725 // BB is now empty, so it is likely to simplify away.
2726 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
2729 // Ok, the block is reachable from the default dest. If the constant we're
2730 // comparing exists in one of the other edges, then we can constant fold ICI
2732 if (SI->findCaseValue(Cst) != SI->case_default()) {
2734 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2735 V = ConstantInt::getFalse(BB->getContext());
2737 V = ConstantInt::getTrue(BB->getContext());
2739 ICI->replaceAllUsesWith(V);
2740 ICI->eraseFromParent();
2741 // BB is now empty, so it is likely to simplify away.
2742 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
2745 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2747 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2748 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
2749 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
2750 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2753 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2755 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2756 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2758 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2759 std::swap(DefaultCst, NewCst);
2761 // Replace ICI (which is used by the PHI for the default value) with true or
2762 // false depending on if it is EQ or NE.
2763 ICI->replaceAllUsesWith(DefaultCst);
2764 ICI->eraseFromParent();
2766 // Okay, the switch goes to this block on a default value. Add an edge from
2767 // the switch to the merge point on the compared value.
2768 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2769 BB->getParent(), BB);
2770 SmallVector<uint64_t, 8> Weights;
2771 bool HasWeights = HasBranchWeights(SI);
2773 GetBranchWeights(SI, Weights);
2774 if (Weights.size() == 1 + SI->getNumCases()) {
2775 // Split weight for default case to case for "Cst".
2776 Weights[0] = (Weights[0]+1) >> 1;
2777 Weights.push_back(Weights[0]);
2779 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2780 SI->setMetadata(LLVMContext::MD_prof,
2781 MDBuilder(SI->getContext()).
2782 createBranchWeights(MDWeights));
2785 SI->addCase(Cst, NewBB);
2787 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2788 Builder.SetInsertPoint(NewBB);
2789 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2790 Builder.CreateBr(SuccBlock);
2791 PHIUse->addIncoming(NewCst, NewBB);
2795 /// The specified branch is a conditional branch.
2796 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2797 /// fold it into a switch instruction if so.
2798 static bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
2799 const DataLayout &DL) {
2800 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2801 if (!Cond) return false;
2803 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2804 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2805 // 'setne's and'ed together, collect them.
2807 // Try to gather values from a chain of and/or to be turned into a switch
2808 ConstantComparesGatherer ConstantCompare(Cond, DL);
2809 // Unpack the result
2810 SmallVectorImpl<ConstantInt*> &Values = ConstantCompare.Vals;
2811 Value *CompVal = ConstantCompare.CompValue;
2812 unsigned UsedICmps = ConstantCompare.UsedICmps;
2813 Value *ExtraCase = ConstantCompare.Extra;
2815 // If we didn't have a multiply compared value, fail.
2816 if (!CompVal) return false;
2818 // Avoid turning single icmps into a switch.
2822 bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
2824 // There might be duplicate constants in the list, which the switch
2825 // instruction can't handle, remove them now.
2826 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2827 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2829 // If Extra was used, we require at least two switch values to do the
2830 // transformation. A switch with one value is just an cond branch.
2831 if (ExtraCase && Values.size() < 2) return false;
2833 // TODO: Preserve branch weight metadata, similarly to how
2834 // FoldValueComparisonIntoPredecessors preserves it.
2836 // Figure out which block is which destination.
2837 BasicBlock *DefaultBB = BI->getSuccessor(1);
2838 BasicBlock *EdgeBB = BI->getSuccessor(0);
2839 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2841 BasicBlock *BB = BI->getParent();
2843 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2844 << " cases into SWITCH. BB is:\n" << *BB);
2846 // If there are any extra values that couldn't be folded into the switch
2847 // then we evaluate them with an explicit branch first. Split the block
2848 // right before the condbr to handle it.
2850 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2851 // Remove the uncond branch added to the old block.
2852 TerminatorInst *OldTI = BB->getTerminator();
2853 Builder.SetInsertPoint(OldTI);
2856 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2858 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2860 OldTI->eraseFromParent();
2862 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2863 // for the edge we just added.
2864 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2866 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2867 << "\nEXTRABB = " << *BB);
2871 Builder.SetInsertPoint(BI);
2872 // Convert pointer to int before we switch.
2873 if (CompVal->getType()->isPointerTy()) {
2874 CompVal = Builder.CreatePtrToInt(
2875 CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
2878 // Create the new switch instruction now.
2879 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2881 // Add all of the 'cases' to the switch instruction.
2882 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2883 New->addCase(Values[i], EdgeBB);
2885 // We added edges from PI to the EdgeBB. As such, if there were any
2886 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2887 // the number of edges added.
2888 for (BasicBlock::iterator BBI = EdgeBB->begin();
2889 isa<PHINode>(BBI); ++BBI) {
2890 PHINode *PN = cast<PHINode>(BBI);
2891 Value *InVal = PN->getIncomingValueForBlock(BB);
2892 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2893 PN->addIncoming(InVal, BB);
2896 // Erase the old branch instruction.
2897 EraseTerminatorInstAndDCECond(BI);
2899 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2903 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2904 // If this is a trivial landing pad that just continues unwinding the caught
2905 // exception then zap the landing pad, turning its invokes into calls.
2906 BasicBlock *BB = RI->getParent();
2907 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2908 if (RI->getValue() != LPInst)
2909 // Not a landing pad, or the resume is not unwinding the exception that
2910 // caused control to branch here.
2913 // Check that there are no other instructions except for debug intrinsics.
2914 BasicBlock::iterator I = LPInst, E = RI;
2916 if (!isa<DbgInfoIntrinsic>(I))
2919 // Turn all invokes that unwind here into calls and delete the basic block.
2920 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2921 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2922 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2923 // Insert a call instruction before the invoke.
2924 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2926 Call->setCallingConv(II->getCallingConv());
2927 Call->setAttributes(II->getAttributes());
2928 Call->setDebugLoc(II->getDebugLoc());
2930 // Anything that used the value produced by the invoke instruction now uses
2931 // the value produced by the call instruction. Note that we do this even
2932 // for void functions and calls with no uses so that the callgraph edge is
2934 II->replaceAllUsesWith(Call);
2935 BB->removePredecessor(II->getParent());
2937 // Insert a branch to the normal destination right before the invoke.
2938 BranchInst::Create(II->getNormalDest(), II);
2940 // Finally, delete the invoke instruction!
2941 II->eraseFromParent();
2944 // The landingpad is now unreachable. Zap it.
2945 BB->eraseFromParent();
2949 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2950 BasicBlock *BB = RI->getParent();
2951 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2953 // Find predecessors that end with branches.
2954 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2955 SmallVector<BranchInst*, 8> CondBranchPreds;
2956 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2957 BasicBlock *P = *PI;
2958 TerminatorInst *PTI = P->getTerminator();
2959 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2960 if (BI->isUnconditional())
2961 UncondBranchPreds.push_back(P);
2963 CondBranchPreds.push_back(BI);
2967 // If we found some, do the transformation!
2968 if (!UncondBranchPreds.empty() && DupRet) {
2969 while (!UncondBranchPreds.empty()) {
2970 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
2971 DEBUG(dbgs() << "FOLDING: " << *BB
2972 << "INTO UNCOND BRANCH PRED: " << *Pred);
2973 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
2976 // If we eliminated all predecessors of the block, delete the block now.
2978 // We know there are no successors, so just nuke the block.
2979 BB->eraseFromParent();
2984 // Check out all of the conditional branches going to this return
2985 // instruction. If any of them just select between returns, change the
2986 // branch itself into a select/return pair.
2987 while (!CondBranchPreds.empty()) {
2988 BranchInst *BI = CondBranchPreds.pop_back_val();
2990 // Check to see if the non-BB successor is also a return block.
2991 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
2992 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
2993 SimplifyCondBranchToTwoReturns(BI, Builder))
2999 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
3000 BasicBlock *BB = UI->getParent();
3002 bool Changed = false;
3004 // If there are any instructions immediately before the unreachable that can
3005 // be removed, do so.
3006 while (UI != BB->begin()) {
3007 BasicBlock::iterator BBI = UI;
3009 // Do not delete instructions that can have side effects which might cause
3010 // the unreachable to not be reachable; specifically, calls and volatile
3011 // operations may have this effect.
3012 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
3014 if (BBI->mayHaveSideEffects()) {
3015 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
3016 if (SI->isVolatile())
3018 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
3019 if (LI->isVolatile())
3021 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
3022 if (RMWI->isVolatile())
3024 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
3025 if (CXI->isVolatile())
3027 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
3028 !isa<LandingPadInst>(BBI)) {
3031 // Note that deleting LandingPad's here is in fact okay, although it
3032 // involves a bit of subtle reasoning. If this inst is a LandingPad,
3033 // all the predecessors of this block will be the unwind edges of Invokes,
3034 // and we can therefore guarantee this block will be erased.
3037 // Delete this instruction (any uses are guaranteed to be dead)
3038 if (!BBI->use_empty())
3039 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
3040 BBI->eraseFromParent();
3044 // If the unreachable instruction is the first in the block, take a gander
3045 // at all of the predecessors of this instruction, and simplify them.
3046 if (&BB->front() != UI) return Changed;
3048 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
3049 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
3050 TerminatorInst *TI = Preds[i]->getTerminator();
3051 IRBuilder<> Builder(TI);
3052 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3053 if (BI->isUnconditional()) {
3054 if (BI->getSuccessor(0) == BB) {
3055 new UnreachableInst(TI->getContext(), TI);
3056 TI->eraseFromParent();
3060 if (BI->getSuccessor(0) == BB) {
3061 Builder.CreateBr(BI->getSuccessor(1));
3062 EraseTerminatorInstAndDCECond(BI);
3063 } else if (BI->getSuccessor(1) == BB) {
3064 Builder.CreateBr(BI->getSuccessor(0));
3065 EraseTerminatorInstAndDCECond(BI);
3069 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3070 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3072 if (i.getCaseSuccessor() == BB) {
3073 BB->removePredecessor(SI->getParent());
3078 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
3079 if (II->getUnwindDest() == BB) {
3080 // Convert the invoke to a call instruction. This would be a good
3081 // place to note that the call does not throw though.
3082 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
3083 II->removeFromParent(); // Take out of symbol table
3085 // Insert the call now...
3086 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
3087 Builder.SetInsertPoint(BI);
3088 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
3089 Args, II->getName());
3090 CI->setCallingConv(II->getCallingConv());
3091 CI->setAttributes(II->getAttributes());
3092 // If the invoke produced a value, the call does now instead.
3093 II->replaceAllUsesWith(CI);
3100 // If this block is now dead, remove it.
3101 if (pred_empty(BB) &&
3102 BB != &BB->getParent()->getEntryBlock()) {
3103 // We know there are no successors, so just nuke the block.
3104 BB->eraseFromParent();
3111 static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) {
3112 assert(Cases.size() >= 1);
3114 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3115 for (size_t I = 1, E = Cases.size(); I != E; ++I) {
3116 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
3122 /// Turn a switch with two reachable destinations into an integer range
3123 /// comparison and branch.
3124 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3125 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3128 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
3130 // Partition the cases into two sets with different destinations.
3131 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
3132 BasicBlock *DestB = nullptr;
3133 SmallVector <ConstantInt *, 16> CasesA;
3134 SmallVector <ConstantInt *, 16> CasesB;
3136 for (SwitchInst::CaseIt I : SI->cases()) {
3137 BasicBlock *Dest = I.getCaseSuccessor();
3138 if (!DestA) DestA = Dest;
3139 if (Dest == DestA) {
3140 CasesA.push_back(I.getCaseValue());
3143 if (!DestB) DestB = Dest;
3144 if (Dest == DestB) {
3145 CasesB.push_back(I.getCaseValue());
3148 return false; // More than two destinations.
3151 assert(DestA && DestB && "Single-destination switch should have been folded.");
3152 assert(DestA != DestB);
3153 assert(DestB != SI->getDefaultDest());
3154 assert(!CasesB.empty() && "There must be non-default cases.");
3155 assert(!CasesA.empty() || HasDefault);
3157 // Figure out if one of the sets of cases form a contiguous range.
3158 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
3159 BasicBlock *ContiguousDest = nullptr;
3160 BasicBlock *OtherDest = nullptr;
3161 if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
3162 ContiguousCases = &CasesA;
3163 ContiguousDest = DestA;
3165 } else if (CasesAreContiguous(CasesB)) {
3166 ContiguousCases = &CasesB;
3167 ContiguousDest = DestB;
3172 // Start building the compare and branch.
3174 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
3175 Constant *NumCases = ConstantInt::get(Offset->getType(), ContiguousCases->size());
3177 Value *Sub = SI->getCondition();
3178 if (!Offset->isNullValue())
3179 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
3182 // If NumCases overflowed, then all possible values jump to the successor.
3183 if (NumCases->isNullValue() && !ContiguousCases->empty())
3184 Cmp = ConstantInt::getTrue(SI->getContext());
3186 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3187 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
3189 // Update weight for the newly-created conditional branch.
3190 if (HasBranchWeights(SI)) {
3191 SmallVector<uint64_t, 8> Weights;
3192 GetBranchWeights(SI, Weights);
3193 if (Weights.size() == 1 + SI->getNumCases()) {
3194 uint64_t TrueWeight = 0;
3195 uint64_t FalseWeight = 0;
3196 for (size_t I = 0, E = Weights.size(); I != E; ++I) {
3197 if (SI->getSuccessor(I) == ContiguousDest)
3198 TrueWeight += Weights[I];
3200 FalseWeight += Weights[I];
3202 while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
3206 NewBI->setMetadata(LLVMContext::MD_prof,
3207 MDBuilder(SI->getContext()).createBranchWeights(
3208 (uint32_t)TrueWeight, (uint32_t)FalseWeight));
3212 // Prune obsolete incoming values off the successors' PHI nodes.
3213 for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
3214 unsigned PreviousEdges = ContiguousCases->size();
3215 if (ContiguousDest == SI->getDefaultDest()) ++PreviousEdges;
3216 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
3217 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3219 for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
3220 unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
3221 if (OtherDest == SI->getDefaultDest()) ++PreviousEdges;
3222 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
3223 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3227 SI->eraseFromParent();
3232 /// Compute masked bits for the condition of a switch
3233 /// and use it to remove dead cases.
3234 static bool EliminateDeadSwitchCases(SwitchInst *SI, AssumptionCache *AC,
3235 const DataLayout &DL) {
3236 Value *Cond = SI->getCondition();
3237 unsigned Bits = Cond->getType()->getIntegerBitWidth();
3238 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3239 computeKnownBits(Cond, KnownZero, KnownOne, DL, 0, AC, SI);
3241 // Gather dead cases.
3242 SmallVector<ConstantInt*, 8> DeadCases;
3243 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3244 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3245 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3246 DeadCases.push_back(I.getCaseValue());
3247 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3248 << I.getCaseValue() << "' is dead.\n");
3252 SmallVector<uint64_t, 8> Weights;
3253 bool HasWeight = HasBranchWeights(SI);
3255 GetBranchWeights(SI, Weights);
3256 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3259 // Remove dead cases from the switch.
3260 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3261 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3262 assert(Case != SI->case_default() &&
3263 "Case was not found. Probably mistake in DeadCases forming.");
3265 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3269 // Prune unused values from PHI nodes.
3270 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3271 SI->removeCase(Case);
3273 if (HasWeight && Weights.size() >= 2) {
3274 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3275 SI->setMetadata(LLVMContext::MD_prof,
3276 MDBuilder(SI->getParent()->getContext()).
3277 createBranchWeights(MDWeights));
3280 return !DeadCases.empty();
3283 /// If BB would be eligible for simplification by
3284 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3285 /// by an unconditional branch), look at the phi node for BB in the successor
3286 /// block and see if the incoming value is equal to CaseValue. If so, return
3287 /// the phi node, and set PhiIndex to BB's index in the phi node.
3288 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3291 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3292 return nullptr; // BB must be empty to be a candidate for simplification.
3293 if (!BB->getSinglePredecessor())
3294 return nullptr; // BB must be dominated by the switch.
3296 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3297 if (!Branch || !Branch->isUnconditional())
3298 return nullptr; // Terminator must be unconditional branch.
3300 BasicBlock *Succ = Branch->getSuccessor(0);
3302 BasicBlock::iterator I = Succ->begin();
3303 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3304 int Idx = PHI->getBasicBlockIndex(BB);
3305 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3307 Value *InValue = PHI->getIncomingValue(Idx);
3308 if (InValue != CaseValue) continue;
3317 /// Try to forward the condition of a switch instruction to a phi node
3318 /// dominated by the switch, if that would mean that some of the destination
3319 /// blocks of the switch can be folded away.
3320 /// Returns true if a change is made.
3321 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3322 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3323 ForwardingNodesMap ForwardingNodes;
3325 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3326 ConstantInt *CaseValue = I.getCaseValue();
3327 BasicBlock *CaseDest = I.getCaseSuccessor();
3330 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3334 ForwardingNodes[PHI].push_back(PhiIndex);
3337 bool Changed = false;
3339 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3340 E = ForwardingNodes.end(); I != E; ++I) {
3341 PHINode *Phi = I->first;
3342 SmallVectorImpl<int> &Indexes = I->second;
3344 if (Indexes.size() < 2) continue;
3346 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3347 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3354 /// Return true if the backend will be able to handle
3355 /// initializing an array of constants like C.
3356 static bool ValidLookupTableConstant(Constant *C) {
3357 if (C->isThreadDependent())
3359 if (C->isDLLImportDependent())
3362 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3363 return CE->isGEPWithNoNotionalOverIndexing();
3365 return isa<ConstantFP>(C) ||
3366 isa<ConstantInt>(C) ||
3367 isa<ConstantPointerNull>(C) ||
3368 isa<GlobalValue>(C) ||
3372 /// If V is a Constant, return it. Otherwise, try to look up
3373 /// its constant value in ConstantPool, returning 0 if it's not there.
3374 static Constant *LookupConstant(Value *V,
3375 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3376 if (Constant *C = dyn_cast<Constant>(V))
3378 return ConstantPool.lookup(V);
3381 /// Try to fold instruction I into a constant. This works for
3382 /// simple instructions such as binary operations where both operands are
3383 /// constant or can be replaced by constants from the ConstantPool. Returns the
3384 /// resulting constant on success, 0 otherwise.
3386 ConstantFold(Instruction *I, const DataLayout &DL,
3387 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
3388 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3389 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3392 if (A->isAllOnesValue())
3393 return LookupConstant(Select->getTrueValue(), ConstantPool);
3394 if (A->isNullValue())
3395 return LookupConstant(Select->getFalseValue(), ConstantPool);
3399 SmallVector<Constant *, 4> COps;
3400 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
3401 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
3407 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
3408 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
3412 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), COps, DL);
3415 /// Try to determine the resulting constant values in phi nodes
3416 /// at the common destination basic block, *CommonDest, for one of the case
3417 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3418 /// case), of a switch instruction SI.
3420 GetCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest,
3421 BasicBlock **CommonDest,
3422 SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
3423 const DataLayout &DL) {
3424 // The block from which we enter the common destination.
3425 BasicBlock *Pred = SI->getParent();
3427 // If CaseDest is empty except for some side-effect free instructions through
3428 // which we can constant-propagate the CaseVal, continue to its successor.
3429 SmallDenseMap<Value*, Constant*> ConstantPool;
3430 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3431 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3433 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3434 // If the terminator is a simple branch, continue to the next block.
3435 if (T->getNumSuccessors() != 1)
3438 CaseDest = T->getSuccessor(0);
3439 } else if (isa<DbgInfoIntrinsic>(I)) {
3440 // Skip debug intrinsic.
3442 } else if (Constant *C = ConstantFold(I, DL, ConstantPool)) {
3443 // Instruction is side-effect free and constant.
3445 // If the instruction has uses outside this block or a phi node slot for
3446 // the block, it is not safe to bypass the instruction since it would then
3447 // no longer dominate all its uses.
3448 for (auto &Use : I->uses()) {
3449 User *User = Use.getUser();
3450 if (Instruction *I = dyn_cast<Instruction>(User))
3451 if (I->getParent() == CaseDest)
3453 if (PHINode *Phi = dyn_cast<PHINode>(User))
3454 if (Phi->getIncomingBlock(Use) == CaseDest)
3459 ConstantPool.insert(std::make_pair(I, C));
3465 // If we did not have a CommonDest before, use the current one.
3467 *CommonDest = CaseDest;
3468 // If the destination isn't the common one, abort.
3469 if (CaseDest != *CommonDest)
3472 // Get the values for this case from phi nodes in the destination block.
3473 BasicBlock::iterator I = (*CommonDest)->begin();
3474 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3475 int Idx = PHI->getBasicBlockIndex(Pred);
3479 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3484 // Be conservative about which kinds of constants we support.
3485 if (!ValidLookupTableConstant(ConstVal))
3488 Res.push_back(std::make_pair(PHI, ConstVal));
3491 return Res.size() > 0;
3494 // Helper function used to add CaseVal to the list of cases that generate
3496 static void MapCaseToResult(ConstantInt *CaseVal,
3497 SwitchCaseResultVectorTy &UniqueResults,
3499 for (auto &I : UniqueResults) {
3500 if (I.first == Result) {
3501 I.second.push_back(CaseVal);
3505 UniqueResults.push_back(std::make_pair(Result,
3506 SmallVector<ConstantInt*, 4>(1, CaseVal)));
3509 // Helper function that initializes a map containing
3510 // results for the PHI node of the common destination block for a switch
3511 // instruction. Returns false if multiple PHI nodes have been found or if
3512 // there is not a common destination block for the switch.
3513 static bool InitializeUniqueCases(SwitchInst *SI, PHINode *&PHI,
3514 BasicBlock *&CommonDest,
3515 SwitchCaseResultVectorTy &UniqueResults,
3516 Constant *&DefaultResult,
3517 const DataLayout &DL) {
3518 for (auto &I : SI->cases()) {
3519 ConstantInt *CaseVal = I.getCaseValue();
3521 // Resulting value at phi nodes for this case value.
3522 SwitchCaseResultsTy Results;
3523 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
3527 // Only one value per case is permitted
3528 if (Results.size() > 1)
3530 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
3532 // Check the PHI consistency.
3534 PHI = Results[0].first;
3535 else if (PHI != Results[0].first)
3538 // Find the default result value.
3539 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
3540 BasicBlock *DefaultDest = SI->getDefaultDest();
3541 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
3543 // If the default value is not found abort unless the default destination
3546 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
3547 if ((!DefaultResult &&
3548 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
3554 // Helper function that checks if it is possible to transform a switch with only
3555 // two cases (or two cases + default) that produces a result into a select.
3558 // case 10: %0 = icmp eq i32 %a, 10
3559 // return 10; %1 = select i1 %0, i32 10, i32 4
3560 // case 20: ----> %2 = icmp eq i32 %a, 20
3561 // return 2; %3 = select i1 %2, i32 2, i32 %1
3566 ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
3567 Constant *DefaultResult, Value *Condition,
3568 IRBuilder<> &Builder) {
3569 assert(ResultVector.size() == 2 &&
3570 "We should have exactly two unique results at this point");
3571 // If we are selecting between only two cases transform into a simple
3572 // select or a two-way select if default is possible.
3573 if (ResultVector[0].second.size() == 1 &&
3574 ResultVector[1].second.size() == 1) {
3575 ConstantInt *const FirstCase = ResultVector[0].second[0];
3576 ConstantInt *const SecondCase = ResultVector[1].second[0];
3578 bool DefaultCanTrigger = DefaultResult;
3579 Value *SelectValue = ResultVector[1].first;
3580 if (DefaultCanTrigger) {
3581 Value *const ValueCompare =
3582 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
3583 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
3584 DefaultResult, "switch.select");
3586 Value *const ValueCompare =
3587 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
3588 return Builder.CreateSelect(ValueCompare, ResultVector[0].first, SelectValue,
3595 // Helper function to cleanup a switch instruction that has been converted into
3596 // a select, fixing up PHI nodes and basic blocks.
3597 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
3599 IRBuilder<> &Builder) {
3600 BasicBlock *SelectBB = SI->getParent();
3601 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
3602 PHI->removeIncomingValue(SelectBB);
3603 PHI->addIncoming(SelectValue, SelectBB);
3605 Builder.CreateBr(PHI->getParent());
3607 // Remove the switch.
3608 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
3609 BasicBlock *Succ = SI->getSuccessor(i);
3611 if (Succ == PHI->getParent())
3613 Succ->removePredecessor(SelectBB);
3615 SI->eraseFromParent();
3618 /// If the switch is only used to initialize one or more
3619 /// phi nodes in a common successor block with only two different
3620 /// constant values, replace the switch with select.
3621 static bool SwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
3622 AssumptionCache *AC, const DataLayout &DL) {
3623 Value *const Cond = SI->getCondition();
3624 PHINode *PHI = nullptr;
3625 BasicBlock *CommonDest = nullptr;
3626 Constant *DefaultResult;
3627 SwitchCaseResultVectorTy UniqueResults;
3628 // Collect all the cases that will deliver the same value from the switch.
3629 if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
3632 // Selects choose between maximum two values.
3633 if (UniqueResults.size() != 2)
3635 assert(PHI != nullptr && "PHI for value select not found");
3637 Builder.SetInsertPoint(SI);
3638 Value *SelectValue = ConvertTwoCaseSwitch(
3640 DefaultResult, Cond, Builder);
3642 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
3645 // The switch couldn't be converted into a select.
3650 /// This class represents a lookup table that can be used to replace a switch.
3651 class SwitchLookupTable {
3653 /// Create a lookup table to use as a switch replacement with the contents
3654 /// of Values, using DefaultValue to fill any holes in the table.
3656 Module &M, uint64_t TableSize, ConstantInt *Offset,
3657 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
3658 Constant *DefaultValue, const DataLayout &DL);
3660 /// Build instructions with Builder to retrieve the value at
3661 /// the position given by Index in the lookup table.
3662 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3664 /// Return true if a table with TableSize elements of
3665 /// type ElementType would fit in a target-legal register.
3666 static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
3670 // Depending on the contents of the table, it can be represented in
3673 // For tables where each element contains the same value, we just have to
3674 // store that single value and return it for each lookup.
3677 // For tables where there is a linear relationship between table index
3678 // and values. We calculate the result with a simple multiplication
3679 // and addition instead of a table lookup.
3682 // For small tables with integer elements, we can pack them into a bitmap
3683 // that fits into a target-legal register. Values are retrieved by
3684 // shift and mask operations.
3687 // The table is stored as an array of values. Values are retrieved by load
3688 // instructions from the table.
3692 // For SingleValueKind, this is the single value.
3693 Constant *SingleValue;
3695 // For BitMapKind, this is the bitmap.
3696 ConstantInt *BitMap;
3697 IntegerType *BitMapElementTy;
3699 // For LinearMapKind, these are the constants used to derive the value.
3700 ConstantInt *LinearOffset;
3701 ConstantInt *LinearMultiplier;
3703 // For ArrayKind, this is the array.
3704 GlobalVariable *Array;
3708 SwitchLookupTable::SwitchLookupTable(
3709 Module &M, uint64_t TableSize, ConstantInt *Offset,
3710 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
3711 Constant *DefaultValue, const DataLayout &DL)
3712 : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
3713 LinearOffset(nullptr), LinearMultiplier(nullptr), Array(nullptr) {
3714 assert(Values.size() && "Can't build lookup table without values!");
3715 assert(TableSize >= Values.size() && "Can't fit values in table!");
3717 // If all values in the table are equal, this is that value.
3718 SingleValue = Values.begin()->second;
3720 Type *ValueType = Values.begin()->second->getType();
3722 // Build up the table contents.
3723 SmallVector<Constant*, 64> TableContents(TableSize);
3724 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3725 ConstantInt *CaseVal = Values[I].first;
3726 Constant *CaseRes = Values[I].second;
3727 assert(CaseRes->getType() == ValueType);
3729 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3731 TableContents[Idx] = CaseRes;
3733 if (CaseRes != SingleValue)
3734 SingleValue = nullptr;
3737 // Fill in any holes in the table with the default result.
3738 if (Values.size() < TableSize) {
3739 assert(DefaultValue &&
3740 "Need a default value to fill the lookup table holes.");
3741 assert(DefaultValue->getType() == ValueType);
3742 for (uint64_t I = 0; I < TableSize; ++I) {
3743 if (!TableContents[I])
3744 TableContents[I] = DefaultValue;
3747 if (DefaultValue != SingleValue)
3748 SingleValue = nullptr;
3751 // If each element in the table contains the same value, we only need to store
3752 // that single value.
3754 Kind = SingleValueKind;
3758 // Check if we can derive the value with a linear transformation from the
3760 if (isa<IntegerType>(ValueType)) {
3761 bool LinearMappingPossible = true;
3764 assert(TableSize >= 2 && "Should be a SingleValue table.");
3765 // Check if there is the same distance between two consecutive values.
3766 for (uint64_t I = 0; I < TableSize; ++I) {
3767 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
3769 // This is an undef. We could deal with it, but undefs in lookup tables
3770 // are very seldom. It's probably not worth the additional complexity.
3771 LinearMappingPossible = false;
3774 APInt Val = ConstVal->getValue();
3776 APInt Dist = Val - PrevVal;
3779 } else if (Dist != DistToPrev) {
3780 LinearMappingPossible = false;
3786 if (LinearMappingPossible) {
3787 LinearOffset = cast<ConstantInt>(TableContents[0]);
3788 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
3789 Kind = LinearMapKind;
3795 // If the type is integer and the table fits in a register, build a bitmap.
3796 if (WouldFitInRegister(DL, TableSize, ValueType)) {
3797 IntegerType *IT = cast<IntegerType>(ValueType);
3798 APInt TableInt(TableSize * IT->getBitWidth(), 0);
3799 for (uint64_t I = TableSize; I > 0; --I) {
3800 TableInt <<= IT->getBitWidth();
3801 // Insert values into the bitmap. Undef values are set to zero.
3802 if (!isa<UndefValue>(TableContents[I - 1])) {
3803 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
3804 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
3807 BitMap = ConstantInt::get(M.getContext(), TableInt);
3808 BitMapElementTy = IT;
3814 // Store the table in an array.
3815 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
3816 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3818 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3819 GlobalVariable::PrivateLinkage,
3822 Array->setUnnamedAddr(true);
3826 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
3828 case SingleValueKind:
3830 case LinearMapKind: {
3831 // Derive the result value from the input value.
3832 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
3833 false, "switch.idx.cast");
3834 if (!LinearMultiplier->isOne())
3835 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
3836 if (!LinearOffset->isZero())
3837 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
3841 // Type of the bitmap (e.g. i59).
3842 IntegerType *MapTy = BitMap->getType();
3844 // Cast Index to the same type as the bitmap.
3845 // Note: The Index is <= the number of elements in the table, so
3846 // truncating it to the width of the bitmask is safe.
3847 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
3849 // Multiply the shift amount by the element width.
3850 ShiftAmt = Builder.CreateMul(ShiftAmt,
3851 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
3855 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
3856 "switch.downshift");
3858 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
3862 // Make sure the table index will not overflow when treated as signed.
3863 IntegerType *IT = cast<IntegerType>(Index->getType());
3864 uint64_t TableSize = Array->getInitializer()->getType()
3865 ->getArrayNumElements();
3866 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
3867 Index = Builder.CreateZExt(Index,
3868 IntegerType::get(IT->getContext(),
3869 IT->getBitWidth() + 1),
3870 "switch.tableidx.zext");
3872 Value *GEPIndices[] = { Builder.getInt32(0), Index };
3873 Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
3874 GEPIndices, "switch.gep");
3875 return Builder.CreateLoad(GEP, "switch.load");
3878 llvm_unreachable("Unknown lookup table kind!");
3881 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
3883 Type *ElementType) {
3884 auto *IT = dyn_cast<IntegerType>(ElementType);
3887 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
3888 // are <= 15, we could try to narrow the type.
3890 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
3891 if (TableSize >= UINT_MAX/IT->getBitWidth())
3893 return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
3896 /// Determine whether a lookup table should be built for this switch, based on
3897 /// the number of cases, size of the table, and the types of the results.
3899 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
3900 const TargetTransformInfo &TTI, const DataLayout &DL,
3901 const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
3902 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
3903 return false; // TableSize overflowed, or mul below might overflow.
3905 bool AllTablesFitInRegister = true;
3906 bool HasIllegalType = false;
3907 for (const auto &I : ResultTypes) {
3908 Type *Ty = I.second;
3910 // Saturate this flag to true.
3911 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
3913 // Saturate this flag to false.
3914 AllTablesFitInRegister = AllTablesFitInRegister &&
3915 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
3917 // If both flags saturate, we're done. NOTE: This *only* works with
3918 // saturating flags, and all flags have to saturate first due to the
3919 // non-deterministic behavior of iterating over a dense map.
3920 if (HasIllegalType && !AllTablesFitInRegister)
3924 // If each table would fit in a register, we should build it anyway.
3925 if (AllTablesFitInRegister)
3928 // Don't build a table that doesn't fit in-register if it has illegal types.
3932 // The table density should be at least 40%. This is the same criterion as for
3933 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3934 // FIXME: Find the best cut-off.
3935 return SI->getNumCases() * 10 >= TableSize * 4;
3938 /// Try to reuse the switch table index compare. Following pattern:
3940 /// if (idx < tablesize)
3941 /// r = table[idx]; // table does not contain default_value
3943 /// r = default_value;
3944 /// if (r != default_value)
3947 /// Is optimized to:
3949 /// cond = idx < tablesize;
3953 /// r = default_value;
3957 /// Jump threading will then eliminate the second if(cond).
3958 static void reuseTableCompare(User *PhiUser, BasicBlock *PhiBlock,
3959 BranchInst *RangeCheckBranch, Constant *DefaultValue,
3960 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values) {
3962 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
3966 // We require that the compare is in the same block as the phi so that jump
3967 // threading can do its work afterwards.
3968 if (CmpInst->getParent() != PhiBlock)
3971 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
3975 Value *RangeCmp = RangeCheckBranch->getCondition();
3976 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
3977 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
3979 // Check if the compare with the default value is constant true or false.
3980 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
3981 DefaultValue, CmpOp1, true);
3982 if (DefaultConst != TrueConst && DefaultConst != FalseConst)
3985 // Check if the compare with the case values is distinct from the default
3987 for (auto ValuePair : Values) {
3988 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
3989 ValuePair.second, CmpOp1, true);
3990 if (!CaseConst || CaseConst == DefaultConst)
3992 assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
3993 "Expect true or false as compare result.");
3996 // Check if the branch instruction dominates the phi node. It's a simple
3997 // dominance check, but sufficient for our needs.
3998 // Although this check is invariant in the calling loops, it's better to do it
3999 // at this late stage. Practically we do it at most once for a switch.
4000 BasicBlock *BranchBlock = RangeCheckBranch->getParent();
4001 for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
4002 BasicBlock *Pred = *PI;
4003 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
4007 if (DefaultConst == FalseConst) {
4008 // The compare yields the same result. We can replace it.
4009 CmpInst->replaceAllUsesWith(RangeCmp);
4010 ++NumTableCmpReuses;
4012 // The compare yields the same result, just inverted. We can replace it.
4013 Value *InvertedTableCmp = BinaryOperator::CreateXor(RangeCmp,
4014 ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
4016 CmpInst->replaceAllUsesWith(InvertedTableCmp);
4017 ++NumTableCmpReuses;
4021 /// If the switch is only used to initialize one or more phi nodes in a common
4022 /// successor block with different constant values, replace the switch with
4024 static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
4025 const DataLayout &DL,
4026 const TargetTransformInfo &TTI) {
4027 assert(SI->getNumCases() > 1 && "Degenerate switch?");
4029 // Only build lookup table when we have a target that supports it.
4030 if (!TTI.shouldBuildLookupTables())
4033 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
4034 // split off a dense part and build a lookup table for that.
4036 // FIXME: This creates arrays of GEPs to constant strings, which means each
4037 // GEP needs a runtime relocation in PIC code. We should just build one big
4038 // string and lookup indices into that.
4040 // Ignore switches with less than three cases. Lookup tables will not make them
4041 // faster, so we don't analyze them.
4042 if (SI->getNumCases() < 3)
4045 // Figure out the corresponding result for each case value and phi node in the
4046 // common destination, as well as the min and max case values.
4047 assert(SI->case_begin() != SI->case_end());
4048 SwitchInst::CaseIt CI = SI->case_begin();
4049 ConstantInt *MinCaseVal = CI.getCaseValue();
4050 ConstantInt *MaxCaseVal = CI.getCaseValue();
4052 BasicBlock *CommonDest = nullptr;
4053 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
4054 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
4055 SmallDenseMap<PHINode*, Constant*> DefaultResults;
4056 SmallDenseMap<PHINode*, Type*> ResultTypes;
4057 SmallVector<PHINode*, 4> PHIs;
4059 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
4060 ConstantInt *CaseVal = CI.getCaseValue();
4061 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
4062 MinCaseVal = CaseVal;
4063 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
4064 MaxCaseVal = CaseVal;
4066 // Resulting value at phi nodes for this case value.
4067 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
4069 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
4073 // Append the result from this case to the list for each phi.
4074 for (const auto &I : Results) {
4075 PHINode *PHI = I.first;
4076 Constant *Value = I.second;
4077 if (!ResultLists.count(PHI))
4078 PHIs.push_back(PHI);
4079 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
4083 // Keep track of the result types.
4084 for (PHINode *PHI : PHIs) {
4085 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
4088 uint64_t NumResults = ResultLists[PHIs[0]].size();
4089 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
4090 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
4091 bool TableHasHoles = (NumResults < TableSize);
4093 // If the table has holes, we need a constant result for the default case
4094 // or a bitmask that fits in a register.
4095 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
4096 bool HasDefaultResults = GetCaseResults(SI, nullptr, SI->getDefaultDest(),
4097 &CommonDest, DefaultResultsList, DL);
4099 bool NeedMask = (TableHasHoles && !HasDefaultResults);
4101 // As an extra penalty for the validity test we require more cases.
4102 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
4104 if (!DL.fitsInLegalInteger(TableSize))
4108 for (const auto &I : DefaultResultsList) {
4109 PHINode *PHI = I.first;
4110 Constant *Result = I.second;
4111 DefaultResults[PHI] = Result;
4114 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
4117 // Create the BB that does the lookups.
4118 Module &Mod = *CommonDest->getParent()->getParent();
4119 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
4121 CommonDest->getParent(),
4124 // Compute the table index value.
4125 Builder.SetInsertPoint(SI);
4126 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
4129 // Compute the maximum table size representable by the integer type we are
4131 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
4132 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
4133 assert(MaxTableSize >= TableSize &&
4134 "It is impossible for a switch to have more entries than the max "
4135 "representable value of its input integer type's size.");
4137 // If the default destination is unreachable, or if the lookup table covers
4138 // all values of the conditional variable, branch directly to the lookup table
4139 // BB. Otherwise, check that the condition is within the case range.
4140 const bool DefaultIsReachable =
4141 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4142 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
4143 BranchInst *RangeCheckBranch = nullptr;
4145 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
4146 Builder.CreateBr(LookupBB);
4147 // Note: We call removeProdecessor later since we need to be able to get the
4148 // PHI value for the default case in case we're using a bit mask.
4150 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
4151 MinCaseVal->getType(), TableSize));
4152 RangeCheckBranch = Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
4155 // Populate the BB that does the lookups.
4156 Builder.SetInsertPoint(LookupBB);
4159 // Before doing the lookup we do the hole check.
4160 // The LookupBB is therefore re-purposed to do the hole check
4161 // and we create a new LookupBB.
4162 BasicBlock *MaskBB = LookupBB;
4163 MaskBB->setName("switch.hole_check");
4164 LookupBB = BasicBlock::Create(Mod.getContext(),
4166 CommonDest->getParent(),
4169 // Make the mask's bitwidth at least 8bit and a power-of-2 to avoid
4170 // unnecessary illegal types.
4171 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
4172 APInt MaskInt(TableSizePowOf2, 0);
4173 APInt One(TableSizePowOf2, 1);
4174 // Build bitmask; fill in a 1 bit for every case.
4175 const ResultListTy &ResultList = ResultLists[PHIs[0]];
4176 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
4177 uint64_t Idx = (ResultList[I].first->getValue() -
4178 MinCaseVal->getValue()).getLimitedValue();
4179 MaskInt |= One << Idx;
4181 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
4183 // Get the TableIndex'th bit of the bitmask.
4184 // If this bit is 0 (meaning hole) jump to the default destination,
4185 // else continue with table lookup.
4186 IntegerType *MapTy = TableMask->getType();
4187 Value *MaskIndex = Builder.CreateZExtOrTrunc(TableIndex, MapTy,
4188 "switch.maskindex");
4189 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex,
4191 Value *LoBit = Builder.CreateTrunc(Shifted,
4192 Type::getInt1Ty(Mod.getContext()),
4194 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
4196 Builder.SetInsertPoint(LookupBB);
4197 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
4200 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
4201 // We cached PHINodes in PHIs, to avoid accessing deleted PHINodes later,
4202 // do not delete PHINodes here.
4203 SI->getDefaultDest()->removePredecessor(SI->getParent(),
4204 /*DontDeleteUselessPHIs=*/true);
4207 bool ReturnedEarly = false;
4208 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
4209 PHINode *PHI = PHIs[I];
4210 const ResultListTy &ResultList = ResultLists[PHI];
4212 // If using a bitmask, use any value to fill the lookup table holes.
4213 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
4214 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL);
4216 Value *Result = Table.BuildLookup(TableIndex, Builder);
4218 // If the result is used to return immediately from the function, we want to
4219 // do that right here.
4220 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
4221 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
4222 Builder.CreateRet(Result);
4223 ReturnedEarly = true;
4227 // Do a small peephole optimization: re-use the switch table compare if
4229 if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
4230 BasicBlock *PhiBlock = PHI->getParent();
4231 // Search for compare instructions which use the phi.
4232 for (auto *User : PHI->users()) {
4233 reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
4237 PHI->addIncoming(Result, LookupBB);
4241 Builder.CreateBr(CommonDest);
4243 // Remove the switch.
4244 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4245 BasicBlock *Succ = SI->getSuccessor(i);
4247 if (Succ == SI->getDefaultDest())
4249 Succ->removePredecessor(SI->getParent());
4251 SI->eraseFromParent();
4255 ++NumLookupTablesHoles;
4259 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
4260 BasicBlock *BB = SI->getParent();
4262 if (isValueEqualityComparison(SI)) {
4263 // If we only have one predecessor, and if it is a branch on this value,
4264 // see if that predecessor totally determines the outcome of this switch.
4265 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4266 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
4267 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4269 Value *Cond = SI->getCondition();
4270 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
4271 if (SimplifySwitchOnSelect(SI, Select))
4272 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4274 // If the block only contains the switch, see if we can fold the block
4275 // away into any preds.
4276 BasicBlock::iterator BBI = BB->begin();
4277 // Ignore dbg intrinsics.
4278 while (isa<DbgInfoIntrinsic>(BBI))
4281 if (FoldValueComparisonIntoPredecessors(SI, Builder))
4282 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4285 // Try to transform the switch into an icmp and a branch.
4286 if (TurnSwitchRangeIntoICmp(SI, Builder))
4287 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4289 // Remove unreachable cases.
4290 if (EliminateDeadSwitchCases(SI, AC, DL))
4291 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4293 if (SwitchToSelect(SI, Builder, AC, DL))
4294 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4296 if (ForwardSwitchConditionToPHI(SI))
4297 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4299 if (SwitchToLookupTable(SI, Builder, DL, TTI))
4300 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4305 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
4306 BasicBlock *BB = IBI->getParent();
4307 bool Changed = false;
4309 // Eliminate redundant destinations.
4310 SmallPtrSet<Value *, 8> Succs;
4311 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
4312 BasicBlock *Dest = IBI->getDestination(i);
4313 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
4314 Dest->removePredecessor(BB);
4315 IBI->removeDestination(i);
4321 if (IBI->getNumDestinations() == 0) {
4322 // If the indirectbr has no successors, change it to unreachable.
4323 new UnreachableInst(IBI->getContext(), IBI);
4324 EraseTerminatorInstAndDCECond(IBI);
4328 if (IBI->getNumDestinations() == 1) {
4329 // If the indirectbr has one successor, change it to a direct branch.
4330 BranchInst::Create(IBI->getDestination(0), IBI);
4331 EraseTerminatorInstAndDCECond(IBI);
4335 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
4336 if (SimplifyIndirectBrOnSelect(IBI, SI))
4337 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4342 /// Given an block with only a single landing pad and a unconditional branch
4343 /// try to find another basic block which this one can be merged with. This
4344 /// handles cases where we have multiple invokes with unique landing pads, but
4345 /// a shared handler.
4347 /// We specifically choose to not worry about merging non-empty blocks
4348 /// here. That is a PRE/scheduling problem and is best solved elsewhere. In
4349 /// practice, the optimizer produces empty landing pad blocks quite frequently
4350 /// when dealing with exception dense code. (see: instcombine, gvn, if-else
4351 /// sinking in this file)
4353 /// This is primarily a code size optimization. We need to avoid performing
4354 /// any transform which might inhibit optimization (such as our ability to
4355 /// specialize a particular handler via tail commoning). We do this by not
4356 /// merging any blocks which require us to introduce a phi. Since the same
4357 /// values are flowing through both blocks, we don't loose any ability to
4358 /// specialize. If anything, we make such specialization more likely.
4360 /// TODO - This transformation could remove entries from a phi in the target
4361 /// block when the inputs in the phi are the same for the two blocks being
4362 /// merged. In some cases, this could result in removal of the PHI entirely.
4363 static bool TryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI,
4365 auto Succ = BB->getUniqueSuccessor();
4367 // If there's a phi in the successor block, we'd likely have to introduce
4368 // a phi into the merged landing pad block.
4369 if (isa<PHINode>(*Succ->begin()))
4372 for (BasicBlock *OtherPred : predecessors(Succ)) {
4373 if (BB == OtherPred)
4375 BasicBlock::iterator I = OtherPred->begin();
4376 LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(I);
4377 if (!LPad2 || !LPad2->isIdenticalTo(LPad))
4379 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {}
4380 BranchInst *BI2 = dyn_cast<BranchInst>(I);
4381 if (!BI2 || !BI2->isIdenticalTo(BI))
4384 // We've found an identical block. Update our predeccessors to take that
4385 // path instead and make ourselves dead.
4386 SmallSet<BasicBlock *, 16> Preds;
4387 Preds.insert(pred_begin(BB), pred_end(BB));
4388 for (BasicBlock *Pred : Preds) {
4389 InvokeInst *II = cast<InvokeInst>(Pred->getTerminator());
4390 assert(II->getNormalDest() != BB &&
4391 II->getUnwindDest() == BB && "unexpected successor");
4392 II->setUnwindDest(OtherPred);
4395 // The debug info in OtherPred doesn't cover the merged control flow that
4396 // used to go through BB. We need to delete it or update it.
4397 for (auto I = OtherPred->begin(), E = OtherPred->end();
4399 Instruction &Inst = *I; I++;
4400 if (isa<DbgInfoIntrinsic>(Inst))
4401 Inst.eraseFromParent();
4404 SmallSet<BasicBlock *, 16> Succs;
4405 Succs.insert(succ_begin(BB), succ_end(BB));
4406 for (BasicBlock *Succ : Succs) {
4407 Succ->removePredecessor(BB);
4410 IRBuilder<> Builder(BI);
4411 Builder.CreateUnreachable();
4412 BI->eraseFromParent();
4418 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
4419 BasicBlock *BB = BI->getParent();
4421 if (SinkCommon && SinkThenElseCodeToEnd(BI))
4424 // If the Terminator is the only non-phi instruction, simplify the block.
4425 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg();
4426 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
4427 TryToSimplifyUncondBranchFromEmptyBlock(BB))
4430 // If the only instruction in the block is a seteq/setne comparison
4431 // against a constant, try to simplify the block.
4432 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
4433 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
4434 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
4436 if (I->isTerminator() &&
4437 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, DL, TTI,
4438 BonusInstThreshold, AC))
4442 // See if we can merge an empty landing pad block with another which is
4444 if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(I)) {
4445 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {}
4446 if (I->isTerminator() &&
4447 TryToMergeLandingPad(LPad, BI, BB))
4451 // If this basic block is ONLY a compare and a branch, and if a predecessor
4452 // branches to us and our successor, fold the comparison into the
4453 // predecessor and use logical operations to update the incoming value
4454 // for PHI nodes in common successor.
4455 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
4456 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4461 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
4462 BasicBlock *BB = BI->getParent();
4464 // Conditional branch
4465 if (isValueEqualityComparison(BI)) {
4466 // If we only have one predecessor, and if it is a branch on this value,
4467 // see if that predecessor totally determines the outcome of this
4469 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4470 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
4471 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4473 // This block must be empty, except for the setcond inst, if it exists.
4474 // Ignore dbg intrinsics.
4475 BasicBlock::iterator I = BB->begin();
4476 // Ignore dbg intrinsics.
4477 while (isa<DbgInfoIntrinsic>(I))
4480 if (FoldValueComparisonIntoPredecessors(BI, Builder))
4481 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4482 } else if (&*I == cast<Instruction>(BI->getCondition())){
4484 // Ignore dbg intrinsics.
4485 while (isa<DbgInfoIntrinsic>(I))
4487 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
4488 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4492 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
4493 if (SimplifyBranchOnICmpChain(BI, Builder, DL))
4496 // If this basic block is ONLY a compare and a branch, and if a predecessor
4497 // branches to us and one of our successors, fold the comparison into the
4498 // predecessor and use logical operations to pick the right destination.
4499 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
4500 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4502 // We have a conditional branch to two blocks that are only reachable
4503 // from BI. We know that the condbr dominates the two blocks, so see if
4504 // there is any identical code in the "then" and "else" blocks. If so, we
4505 // can hoist it up to the branching block.
4506 if (BI->getSuccessor(0)->getSinglePredecessor()) {
4507 if (BI->getSuccessor(1)->getSinglePredecessor()) {
4508 if (HoistThenElseCodeToIf(BI, TTI))
4509 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4511 // If Successor #1 has multiple preds, we may be able to conditionally
4512 // execute Successor #0 if it branches to Successor #1.
4513 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
4514 if (Succ0TI->getNumSuccessors() == 1 &&
4515 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
4516 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), TTI))
4517 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4519 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
4520 // If Successor #0 has multiple preds, we may be able to conditionally
4521 // execute Successor #1 if it branches to Successor #0.
4522 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
4523 if (Succ1TI->getNumSuccessors() == 1 &&
4524 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
4525 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), TTI))
4526 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4529 // If this is a branch on a phi node in the current block, thread control
4530 // through this block if any PHI node entries are constants.
4531 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
4532 if (PN->getParent() == BI->getParent())
4533 if (FoldCondBranchOnPHI(BI, DL))
4534 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4536 // Scan predecessor blocks for conditional branches.
4537 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
4538 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
4539 if (PBI != BI && PBI->isConditional())
4540 if (SimplifyCondBranchToCondBranch(PBI, BI))
4541 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4546 /// Check if passing a value to an instruction will cause undefined behavior.
4547 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
4548 Constant *C = dyn_cast<Constant>(V);
4555 if (C->isNullValue()) {
4556 // Only look at the first use, avoid hurting compile time with long uselists
4557 User *Use = *I->user_begin();
4559 // Now make sure that there are no instructions in between that can alter
4560 // control flow (eg. calls)
4561 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
4562 if (i == I->getParent()->end() || i->mayHaveSideEffects())
4565 // Look through GEPs. A load from a GEP derived from NULL is still undefined
4566 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
4567 if (GEP->getPointerOperand() == I)
4568 return passingValueIsAlwaysUndefined(V, GEP);
4570 // Look through bitcasts.
4571 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
4572 return passingValueIsAlwaysUndefined(V, BC);
4574 // Load from null is undefined.
4575 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
4576 if (!LI->isVolatile())
4577 return LI->getPointerAddressSpace() == 0;
4579 // Store to null is undefined.
4580 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
4581 if (!SI->isVolatile())
4582 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
4587 /// If BB has an incoming value that will always trigger undefined behavior
4588 /// (eg. null pointer dereference), remove the branch leading here.
4589 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
4590 for (BasicBlock::iterator i = BB->begin();
4591 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
4592 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
4593 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
4594 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
4595 IRBuilder<> Builder(T);
4596 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
4597 BB->removePredecessor(PHI->getIncomingBlock(i));
4598 // Turn uncoditional branches into unreachables and remove the dead
4599 // destination from conditional branches.
4600 if (BI->isUnconditional())
4601 Builder.CreateUnreachable();
4603 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
4604 BI->getSuccessor(0));
4605 BI->eraseFromParent();
4608 // TODO: SwitchInst.
4614 bool SimplifyCFGOpt::run(BasicBlock *BB) {
4615 bool Changed = false;
4617 assert(BB && BB->getParent() && "Block not embedded in function!");
4618 assert(BB->getTerminator() && "Degenerate basic block encountered!");
4620 // Remove basic blocks that have no predecessors (except the entry block)...
4621 // or that just have themself as a predecessor. These are unreachable.
4622 if ((pred_empty(BB) &&
4623 BB != &BB->getParent()->getEntryBlock()) ||
4624 BB->getSinglePredecessor() == BB) {
4625 DEBUG(dbgs() << "Removing BB: \n" << *BB);
4626 DeleteDeadBlock(BB);
4630 // Check to see if we can constant propagate this terminator instruction
4632 Changed |= ConstantFoldTerminator(BB, true);
4634 // Check for and eliminate duplicate PHI nodes in this block.
4635 Changed |= EliminateDuplicatePHINodes(BB);
4637 // Check for and remove branches that will always cause undefined behavior.
4638 Changed |= removeUndefIntroducingPredecessor(BB);
4640 // Merge basic blocks into their predecessor if there is only one distinct
4641 // pred, and if there is only one distinct successor of the predecessor, and
4642 // if there are no PHI nodes.
4644 if (MergeBlockIntoPredecessor(BB))
4647 IRBuilder<> Builder(BB);
4649 // If there is a trivial two-entry PHI node in this basic block, and we can
4650 // eliminate it, do so now.
4651 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
4652 if (PN->getNumIncomingValues() == 2)
4653 Changed |= FoldTwoEntryPHINode(PN, TTI, DL);
4655 Builder.SetInsertPoint(BB->getTerminator());
4656 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
4657 if (BI->isUnconditional()) {
4658 if (SimplifyUncondBranch(BI, Builder)) return true;
4660 if (SimplifyCondBranch(BI, Builder)) return true;
4662 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
4663 if (SimplifyReturn(RI, Builder)) return true;
4664 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
4665 if (SimplifyResume(RI, Builder)) return true;
4666 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
4667 if (SimplifySwitch(SI, Builder)) return true;
4668 } else if (UnreachableInst *UI =
4669 dyn_cast<UnreachableInst>(BB->getTerminator())) {
4670 if (SimplifyUnreachable(UI)) return true;
4671 } else if (IndirectBrInst *IBI =
4672 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
4673 if (SimplifyIndirectBr(IBI)) return true;
4679 /// This function is used to do simplification of a CFG.
4680 /// For example, it adjusts branches to branches to eliminate the extra hop,
4681 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4682 /// of the CFG. It returns true if a modification was made.
4684 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
4685 unsigned BonusInstThreshold, AssumptionCache *AC) {
4686 return SimplifyCFGOpt(TTI, BB->getModule()->getDataLayout(),
4687 BonusInstThreshold, AC).run(BB);