1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Peephole optimize the CFG.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Transforms/Utils/Local.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/ADT/SetVector.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/ConstantFolding.h"
22 #include "llvm/Analysis/InstructionSimplify.h"
23 #include "llvm/Analysis/TargetTransformInfo.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/IR/CFG.h"
26 #include "llvm/IR/ConstantRange.h"
27 #include "llvm/IR/Constants.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/DerivedTypes.h"
30 #include "llvm/IR/GlobalVariable.h"
31 #include "llvm/IR/IRBuilder.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/LLVMContext.h"
35 #include "llvm/IR/MDBuilder.h"
36 #include "llvm/IR/Metadata.h"
37 #include "llvm/IR/Module.h"
38 #include "llvm/IR/NoFolder.h"
39 #include "llvm/IR/Operator.h"
40 #include "llvm/IR/PatternMatch.h"
41 #include "llvm/IR/Type.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/Debug.h"
44 #include "llvm/Support/raw_ostream.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Transforms/Utils/ValueMapper.h"
52 using namespace PatternMatch;
54 #define DEBUG_TYPE "simplifycfg"
56 static cl::opt<unsigned>
57 PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1),
58 cl::desc("Control the amount of phi node folding to perform (default = 1)"));
61 DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false),
62 cl::desc("Duplicate return instructions into unconditional branches"));
65 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
66 cl::desc("Sink common instructions down to the end block"));
68 static cl::opt<bool> HoistCondStores(
69 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
70 cl::desc("Hoist conditional stores if an unconditional store precedes"));
72 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
73 STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables");
74 STATISTIC(NumLookupTablesHoles, "Number of switch instructions turned into lookup tables (holes checked)");
75 STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block");
76 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
79 /// ValueEqualityComparisonCase - Represents a case of a switch.
80 struct ValueEqualityComparisonCase {
84 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
85 : Value(Value), Dest(Dest) {}
87 bool operator<(ValueEqualityComparisonCase RHS) const {
88 // Comparing pointers is ok as we only rely on the order for uniquing.
89 return Value < RHS.Value;
92 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
95 class SimplifyCFGOpt {
96 const TargetTransformInfo &TTI;
97 unsigned BonusInstThreshold;
98 const DataLayout *const DL;
99 AssumptionTracker *AT;
100 Value *isValueEqualityComparison(TerminatorInst *TI);
101 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
102 std::vector<ValueEqualityComparisonCase> &Cases);
103 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
105 IRBuilder<> &Builder);
106 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
107 IRBuilder<> &Builder);
109 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
110 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
111 bool SimplifyUnreachable(UnreachableInst *UI);
112 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
113 bool SimplifyIndirectBr(IndirectBrInst *IBI);
114 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder);
115 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder);
118 SimplifyCFGOpt(const TargetTransformInfo &TTI, unsigned BonusInstThreshold,
119 const DataLayout *DL, AssumptionTracker *AT)
120 : TTI(TTI), BonusInstThreshold(BonusInstThreshold), DL(DL), AT(AT) {}
121 bool run(BasicBlock *BB);
125 /// SafeToMergeTerminators - Return true if it is safe to merge these two
126 /// terminator instructions together.
128 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
129 if (SI1 == SI2) return false; // Can't merge with self!
131 // It is not safe to merge these two switch instructions if they have a common
132 // successor, and if that successor has a PHI node, and if *that* PHI node has
133 // conflicting incoming values from the two switch blocks.
134 BasicBlock *SI1BB = SI1->getParent();
135 BasicBlock *SI2BB = SI2->getParent();
136 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
138 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
139 if (SI1Succs.count(*I))
140 for (BasicBlock::iterator BBI = (*I)->begin();
141 isa<PHINode>(BBI); ++BBI) {
142 PHINode *PN = cast<PHINode>(BBI);
143 if (PN->getIncomingValueForBlock(SI1BB) !=
144 PN->getIncomingValueForBlock(SI2BB))
151 /// isProfitableToFoldUnconditional - Return true if it is safe and profitable
152 /// to merge these two terminator instructions together, where SI1 is an
153 /// unconditional branch. PhiNodes will store all PHI nodes in common
156 static bool isProfitableToFoldUnconditional(BranchInst *SI1,
159 SmallVectorImpl<PHINode*> &PhiNodes) {
160 if (SI1 == SI2) return false; // Can't merge with self!
161 assert(SI1->isUnconditional() && SI2->isConditional());
163 // We fold the unconditional branch if we can easily update all PHI nodes in
164 // common successors:
165 // 1> We have a constant incoming value for the conditional branch;
166 // 2> We have "Cond" as the incoming value for the unconditional branch;
167 // 3> SI2->getCondition() and Cond have same operands.
168 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
169 if (!Ci2) return false;
170 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
171 Cond->getOperand(1) == Ci2->getOperand(1)) &&
172 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
173 Cond->getOperand(1) == Ci2->getOperand(0)))
176 BasicBlock *SI1BB = SI1->getParent();
177 BasicBlock *SI2BB = SI2->getParent();
178 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
179 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
180 if (SI1Succs.count(*I))
181 for (BasicBlock::iterator BBI = (*I)->begin();
182 isa<PHINode>(BBI); ++BBI) {
183 PHINode *PN = cast<PHINode>(BBI);
184 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
185 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
187 PhiNodes.push_back(PN);
192 /// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will
193 /// now be entries in it from the 'NewPred' block. The values that will be
194 /// flowing into the PHI nodes will be the same as those coming in from
195 /// ExistPred, an existing predecessor of Succ.
196 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
197 BasicBlock *ExistPred) {
198 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
201 for (BasicBlock::iterator I = Succ->begin();
202 (PN = dyn_cast<PHINode>(I)); ++I)
203 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
206 /// ComputeSpeculationCost - Compute an abstract "cost" of speculating the
207 /// given instruction, which is assumed to be safe to speculate. 1 means
208 /// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive.
209 static unsigned ComputeSpeculationCost(const User *I, const DataLayout *DL) {
210 assert(isSafeToSpeculativelyExecute(I, DL) &&
211 "Instruction is not safe to speculatively execute!");
212 switch (Operator::getOpcode(I)) {
214 // In doubt, be conservative.
216 case Instruction::GetElementPtr:
217 // GEPs are cheap if all indices are constant.
218 if (!cast<GEPOperator>(I)->hasAllConstantIndices())
221 case Instruction::ExtractValue:
222 case Instruction::Load:
223 case Instruction::Add:
224 case Instruction::Sub:
225 case Instruction::And:
226 case Instruction::Or:
227 case Instruction::Xor:
228 case Instruction::Shl:
229 case Instruction::LShr:
230 case Instruction::AShr:
231 case Instruction::ICmp:
232 case Instruction::Trunc:
233 case Instruction::ZExt:
234 case Instruction::SExt:
235 case Instruction::BitCast:
236 case Instruction::ExtractElement:
237 case Instruction::InsertElement:
238 return 1; // These are all cheap.
240 case Instruction::Call:
241 case Instruction::Select:
246 /// DominatesMergePoint - If we have a merge point of an "if condition" as
247 /// accepted above, return true if the specified value dominates the block. We
248 /// don't handle the true generality of domination here, just a special case
249 /// which works well enough for us.
251 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
252 /// see if V (which must be an instruction) and its recursive operands
253 /// that do not dominate BB have a combined cost lower than CostRemaining and
254 /// are non-trapping. If both are true, the instruction is inserted into the
255 /// set and true is returned.
257 /// The cost for most non-trapping instructions is defined as 1 except for
258 /// Select whose cost is 2.
260 /// After this function returns, CostRemaining is decreased by the cost of
261 /// V plus its non-dominating operands. If that cost is greater than
262 /// CostRemaining, false is returned and CostRemaining is undefined.
263 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
264 SmallPtrSetImpl<Instruction*> *AggressiveInsts,
265 unsigned &CostRemaining,
266 const DataLayout *DL) {
267 Instruction *I = dyn_cast<Instruction>(V);
269 // Non-instructions all dominate instructions, but not all constantexprs
270 // can be executed unconditionally.
271 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
276 BasicBlock *PBB = I->getParent();
278 // We don't want to allow weird loops that might have the "if condition" in
279 // the bottom of this block.
280 if (PBB == BB) return false;
282 // If this instruction is defined in a block that contains an unconditional
283 // branch to BB, then it must be in the 'conditional' part of the "if
284 // statement". If not, it definitely dominates the region.
285 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
286 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
289 // If we aren't allowing aggressive promotion anymore, then don't consider
290 // instructions in the 'if region'.
291 if (!AggressiveInsts) return false;
293 // If we have seen this instruction before, don't count it again.
294 if (AggressiveInsts->count(I)) return true;
296 // Okay, it looks like the instruction IS in the "condition". Check to
297 // see if it's a cheap instruction to unconditionally compute, and if it
298 // only uses stuff defined outside of the condition. If so, hoist it out.
299 if (!isSafeToSpeculativelyExecute(I, DL))
302 unsigned Cost = ComputeSpeculationCost(I, DL);
304 if (Cost > CostRemaining)
307 CostRemaining -= Cost;
309 // Okay, we can only really hoist these out if their operands do
310 // not take us over the cost threshold.
311 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
312 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, DL))
314 // Okay, it's safe to do this! Remember this instruction.
315 AggressiveInsts->insert(I);
319 /// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr
320 /// and PointerNullValue. Return NULL if value is not a constant int.
321 static ConstantInt *GetConstantInt(Value *V, const DataLayout *DL) {
322 // Normal constant int.
323 ConstantInt *CI = dyn_cast<ConstantInt>(V);
324 if (CI || !DL || !isa<Constant>(V) || !V->getType()->isPointerTy())
327 // This is some kind of pointer constant. Turn it into a pointer-sized
328 // ConstantInt if possible.
329 IntegerType *PtrTy = cast<IntegerType>(DL->getIntPtrType(V->getType()));
331 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
332 if (isa<ConstantPointerNull>(V))
333 return ConstantInt::get(PtrTy, 0);
335 // IntToPtr const int.
336 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
337 if (CE->getOpcode() == Instruction::IntToPtr)
338 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
339 // The constant is very likely to have the right type already.
340 if (CI->getType() == PtrTy)
343 return cast<ConstantInt>
344 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
349 /// GatherConstantCompares - Given a potentially 'or'd or 'and'd together
350 /// collection of icmp eq/ne instructions that compare a value against a
351 /// constant, return the value being compared, and stick the constant into the
354 GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra,
355 const DataLayout *DL, bool isEQ, unsigned &UsedICmps) {
356 Instruction *I = dyn_cast<Instruction>(V);
357 if (!I) return nullptr;
359 // If this is an icmp against a constant, handle this as one of the cases.
360 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) {
361 if (ConstantInt *C = GetConstantInt(I->getOperand(1), DL)) {
365 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
366 // (x & ~2^x) == y --> x == y || x == y|2^x
367 // This undoes a transformation done by instcombine to fuse 2 compares.
368 if (match(ICI->getOperand(0),
369 m_And(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
370 APInt Not = ~RHSC->getValue();
371 if (Not.isPowerOf2()) {
374 ConstantInt::get(C->getContext(), C->getValue() | Not));
382 return I->getOperand(0);
385 // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to
388 ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue());
390 // Shift the range if the compare is fed by an add. This is the range
391 // compare idiom as emitted by instcombine.
393 match(I->getOperand(0), m_Add(m_Value(RHSVal), m_ConstantInt(RHSC)));
395 Span = Span.subtract(RHSC->getValue());
397 // If this is an and/!= check then we want to optimize "x ugt 2" into
400 Span = Span.inverse();
402 // If there are a ton of values, we don't want to make a ginormous switch.
403 if (Span.getSetSize().ugt(8) || Span.isEmptySet())
406 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
407 Vals.push_back(ConstantInt::get(V->getContext(), Tmp));
409 return hasAdd ? RHSVal : I->getOperand(0);
414 // Otherwise, we can only handle an | or &, depending on isEQ.
415 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And))
418 unsigned NumValsBeforeLHS = Vals.size();
419 unsigned UsedICmpsBeforeLHS = UsedICmps;
420 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, DL,
422 unsigned NumVals = Vals.size();
423 unsigned UsedICmpsBeforeRHS = UsedICmps;
424 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, DL,
428 Vals.resize(NumVals);
429 UsedICmps = UsedICmpsBeforeRHS;
432 // The RHS of the or/and can't be folded in and we haven't used "Extra" yet,
433 // set it and return success.
434 if (Extra == nullptr || Extra == I->getOperand(1)) {
435 Extra = I->getOperand(1);
439 Vals.resize(NumValsBeforeLHS);
440 UsedICmps = UsedICmpsBeforeLHS;
444 // If the LHS can't be folded in, but Extra is available and RHS can, try to
446 if (Extra == nullptr || Extra == I->getOperand(0)) {
447 Value *OldExtra = Extra;
448 Extra = I->getOperand(0);
449 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, DL,
452 assert(Vals.size() == NumValsBeforeLHS);
459 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
460 Instruction *Cond = nullptr;
461 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
462 Cond = dyn_cast<Instruction>(SI->getCondition());
463 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
464 if (BI->isConditional())
465 Cond = dyn_cast<Instruction>(BI->getCondition());
466 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
467 Cond = dyn_cast<Instruction>(IBI->getAddress());
470 TI->eraseFromParent();
471 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
474 /// isValueEqualityComparison - Return true if the specified terminator checks
475 /// to see if a value is equal to constant integer value.
476 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
478 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
479 // Do not permit merging of large switch instructions into their
480 // predecessors unless there is only one predecessor.
481 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()),
482 pred_end(SI->getParent())) <= 128)
483 CV = SI->getCondition();
484 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
485 if (BI->isConditional() && BI->getCondition()->hasOneUse())
486 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition()))
487 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
488 CV = ICI->getOperand(0);
490 // Unwrap any lossless ptrtoint cast.
492 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
493 Value *Ptr = PTII->getPointerOperand();
494 if (PTII->getType() == DL->getIntPtrType(Ptr->getType()))
501 /// GetValueEqualityComparisonCases - Given a value comparison instruction,
502 /// decode all of the 'cases' that it represents and return the 'default' block.
503 BasicBlock *SimplifyCFGOpt::
504 GetValueEqualityComparisonCases(TerminatorInst *TI,
505 std::vector<ValueEqualityComparisonCase>
507 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
508 Cases.reserve(SI->getNumCases());
509 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
510 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
511 i.getCaseSuccessor()));
512 return SI->getDefaultDest();
515 BranchInst *BI = cast<BranchInst>(TI);
516 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
517 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
518 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
521 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
525 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries
526 /// in the list that match the specified block.
527 static void EliminateBlockCases(BasicBlock *BB,
528 std::vector<ValueEqualityComparisonCase> &Cases) {
529 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
532 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as
535 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
536 std::vector<ValueEqualityComparisonCase > &C2) {
537 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
539 // Make V1 be smaller than V2.
540 if (V1->size() > V2->size())
543 if (V1->size() == 0) return false;
544 if (V1->size() == 1) {
546 ConstantInt *TheVal = (*V1)[0].Value;
547 for (unsigned i = 0, e = V2->size(); i != e; ++i)
548 if (TheVal == (*V2)[i].Value)
552 // Otherwise, just sort both lists and compare element by element.
553 array_pod_sort(V1->begin(), V1->end());
554 array_pod_sort(V2->begin(), V2->end());
555 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
556 while (i1 != e1 && i2 != e2) {
557 if ((*V1)[i1].Value == (*V2)[i2].Value)
559 if ((*V1)[i1].Value < (*V2)[i2].Value)
567 /// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a
568 /// terminator instruction and its block is known to only have a single
569 /// predecessor block, check to see if that predecessor is also a value
570 /// comparison with the same value, and if that comparison determines the
571 /// outcome of this comparison. If so, simplify TI. This does a very limited
572 /// form of jump threading.
573 bool SimplifyCFGOpt::
574 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
576 IRBuilder<> &Builder) {
577 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
578 if (!PredVal) return false; // Not a value comparison in predecessor.
580 Value *ThisVal = isValueEqualityComparison(TI);
581 assert(ThisVal && "This isn't a value comparison!!");
582 if (ThisVal != PredVal) return false; // Different predicates.
584 // TODO: Preserve branch weight metadata, similarly to how
585 // FoldValueComparisonIntoPredecessors preserves it.
587 // Find out information about when control will move from Pred to TI's block.
588 std::vector<ValueEqualityComparisonCase> PredCases;
589 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
591 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
593 // Find information about how control leaves this block.
594 std::vector<ValueEqualityComparisonCase> ThisCases;
595 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
596 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
598 // If TI's block is the default block from Pred's comparison, potentially
599 // simplify TI based on this knowledge.
600 if (PredDef == TI->getParent()) {
601 // If we are here, we know that the value is none of those cases listed in
602 // PredCases. If there are any cases in ThisCases that are in PredCases, we
604 if (!ValuesOverlap(PredCases, ThisCases))
607 if (isa<BranchInst>(TI)) {
608 // Okay, one of the successors of this condbr is dead. Convert it to a
610 assert(ThisCases.size() == 1 && "Branch can only have one case!");
611 // Insert the new branch.
612 Instruction *NI = Builder.CreateBr(ThisDef);
615 // Remove PHI node entries for the dead edge.
616 ThisCases[0].Dest->removePredecessor(TI->getParent());
618 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
619 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
621 EraseTerminatorInstAndDCECond(TI);
625 SwitchInst *SI = cast<SwitchInst>(TI);
626 // Okay, TI has cases that are statically dead, prune them away.
627 SmallPtrSet<Constant*, 16> DeadCases;
628 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
629 DeadCases.insert(PredCases[i].Value);
631 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
632 << "Through successor TI: " << *TI);
634 // Collect branch weights into a vector.
635 SmallVector<uint32_t, 8> Weights;
636 MDNode* MD = SI->getMetadata(LLVMContext::MD_prof);
637 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
639 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
641 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i));
643 Weights.push_back(CI->getValue().getZExtValue());
645 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
647 if (DeadCases.count(i.getCaseValue())) {
649 std::swap(Weights[i.getCaseIndex()+1], Weights.back());
652 i.getCaseSuccessor()->removePredecessor(TI->getParent());
656 if (HasWeight && Weights.size() >= 2)
657 SI->setMetadata(LLVMContext::MD_prof,
658 MDBuilder(SI->getParent()->getContext()).
659 createBranchWeights(Weights));
661 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
665 // Otherwise, TI's block must correspond to some matched value. Find out
666 // which value (or set of values) this is.
667 ConstantInt *TIV = nullptr;
668 BasicBlock *TIBB = TI->getParent();
669 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
670 if (PredCases[i].Dest == TIBB) {
672 return false; // Cannot handle multiple values coming to this block.
673 TIV = PredCases[i].Value;
675 assert(TIV && "No edge from pred to succ?");
677 // Okay, we found the one constant that our value can be if we get into TI's
678 // BB. Find out which successor will unconditionally be branched to.
679 BasicBlock *TheRealDest = nullptr;
680 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
681 if (ThisCases[i].Value == TIV) {
682 TheRealDest = ThisCases[i].Dest;
686 // If not handled by any explicit cases, it is handled by the default case.
687 if (!TheRealDest) TheRealDest = ThisDef;
689 // Remove PHI node entries for dead edges.
690 BasicBlock *CheckEdge = TheRealDest;
691 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
692 if (*SI != CheckEdge)
693 (*SI)->removePredecessor(TIBB);
697 // Insert the new branch.
698 Instruction *NI = Builder.CreateBr(TheRealDest);
701 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
702 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
704 EraseTerminatorInstAndDCECond(TI);
709 /// ConstantIntOrdering - This class implements a stable ordering of constant
710 /// integers that does not depend on their address. This is important for
711 /// applications that sort ConstantInt's to ensure uniqueness.
712 struct ConstantIntOrdering {
713 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
714 return LHS->getValue().ult(RHS->getValue());
719 static int ConstantIntSortPredicate(ConstantInt *const *P1,
720 ConstantInt *const *P2) {
721 const ConstantInt *LHS = *P1;
722 const ConstantInt *RHS = *P2;
723 if (LHS->getValue().ult(RHS->getValue()))
725 if (LHS->getValue() == RHS->getValue())
730 static inline bool HasBranchWeights(const Instruction* I) {
731 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof);
732 if (ProfMD && ProfMD->getOperand(0))
733 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
734 return MDS->getString().equals("branch_weights");
739 /// Get Weights of a given TerminatorInst, the default weight is at the front
740 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
742 static void GetBranchWeights(TerminatorInst *TI,
743 SmallVectorImpl<uint64_t> &Weights) {
744 MDNode* MD = TI->getMetadata(LLVMContext::MD_prof);
746 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
747 ConstantInt *CI = cast<ConstantInt>(MD->getOperand(i));
748 Weights.push_back(CI->getValue().getZExtValue());
751 // If TI is a conditional eq, the default case is the false case,
752 // and the corresponding branch-weight data is at index 2. We swap the
753 // default weight to be the first entry.
754 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
755 assert(Weights.size() == 2);
756 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
757 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
758 std::swap(Weights.front(), Weights.back());
762 /// Keep halving the weights until all can fit in uint32_t.
763 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
764 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
765 if (Max > UINT_MAX) {
766 unsigned Offset = 32 - countLeadingZeros(Max);
767 for (uint64_t &I : Weights)
772 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value
773 /// equality comparison instruction (either a switch or a branch on "X == c").
774 /// See if any of the predecessors of the terminator block are value comparisons
775 /// on the same value. If so, and if safe to do so, fold them together.
776 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
777 IRBuilder<> &Builder) {
778 BasicBlock *BB = TI->getParent();
779 Value *CV = isValueEqualityComparison(TI); // CondVal
780 assert(CV && "Not a comparison?");
781 bool Changed = false;
783 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
784 while (!Preds.empty()) {
785 BasicBlock *Pred = Preds.pop_back_val();
787 // See if the predecessor is a comparison with the same value.
788 TerminatorInst *PTI = Pred->getTerminator();
789 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
791 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
792 // Figure out which 'cases' to copy from SI to PSI.
793 std::vector<ValueEqualityComparisonCase> BBCases;
794 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
796 std::vector<ValueEqualityComparisonCase> PredCases;
797 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
799 // Based on whether the default edge from PTI goes to BB or not, fill in
800 // PredCases and PredDefault with the new switch cases we would like to
802 SmallVector<BasicBlock*, 8> NewSuccessors;
804 // Update the branch weight metadata along the way
805 SmallVector<uint64_t, 8> Weights;
806 bool PredHasWeights = HasBranchWeights(PTI);
807 bool SuccHasWeights = HasBranchWeights(TI);
809 if (PredHasWeights) {
810 GetBranchWeights(PTI, Weights);
811 // branch-weight metadata is inconsistent here.
812 if (Weights.size() != 1 + PredCases.size())
813 PredHasWeights = SuccHasWeights = false;
814 } else if (SuccHasWeights)
815 // If there are no predecessor weights but there are successor weights,
816 // populate Weights with 1, which will later be scaled to the sum of
817 // successor's weights
818 Weights.assign(1 + PredCases.size(), 1);
820 SmallVector<uint64_t, 8> SuccWeights;
821 if (SuccHasWeights) {
822 GetBranchWeights(TI, SuccWeights);
823 // branch-weight metadata is inconsistent here.
824 if (SuccWeights.size() != 1 + BBCases.size())
825 PredHasWeights = SuccHasWeights = false;
826 } else if (PredHasWeights)
827 SuccWeights.assign(1 + BBCases.size(), 1);
829 if (PredDefault == BB) {
830 // If this is the default destination from PTI, only the edges in TI
831 // that don't occur in PTI, or that branch to BB will be activated.
832 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
833 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
834 if (PredCases[i].Dest != BB)
835 PTIHandled.insert(PredCases[i].Value);
837 // The default destination is BB, we don't need explicit targets.
838 std::swap(PredCases[i], PredCases.back());
840 if (PredHasWeights || SuccHasWeights) {
841 // Increase weight for the default case.
842 Weights[0] += Weights[i+1];
843 std::swap(Weights[i+1], Weights.back());
847 PredCases.pop_back();
851 // Reconstruct the new switch statement we will be building.
852 if (PredDefault != BBDefault) {
853 PredDefault->removePredecessor(Pred);
854 PredDefault = BBDefault;
855 NewSuccessors.push_back(BBDefault);
858 unsigned CasesFromPred = Weights.size();
859 uint64_t ValidTotalSuccWeight = 0;
860 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
861 if (!PTIHandled.count(BBCases[i].Value) &&
862 BBCases[i].Dest != BBDefault) {
863 PredCases.push_back(BBCases[i]);
864 NewSuccessors.push_back(BBCases[i].Dest);
865 if (SuccHasWeights || PredHasWeights) {
866 // The default weight is at index 0, so weight for the ith case
867 // should be at index i+1. Scale the cases from successor by
868 // PredDefaultWeight (Weights[0]).
869 Weights.push_back(Weights[0] * SuccWeights[i+1]);
870 ValidTotalSuccWeight += SuccWeights[i+1];
874 if (SuccHasWeights || PredHasWeights) {
875 ValidTotalSuccWeight += SuccWeights[0];
876 // Scale the cases from predecessor by ValidTotalSuccWeight.
877 for (unsigned i = 1; i < CasesFromPred; ++i)
878 Weights[i] *= ValidTotalSuccWeight;
879 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
880 Weights[0] *= SuccWeights[0];
883 // If this is not the default destination from PSI, only the edges
884 // in SI that occur in PSI with a destination of BB will be
886 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
887 std::map<ConstantInt*, uint64_t> WeightsForHandled;
888 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
889 if (PredCases[i].Dest == BB) {
890 PTIHandled.insert(PredCases[i].Value);
892 if (PredHasWeights || SuccHasWeights) {
893 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
894 std::swap(Weights[i+1], Weights.back());
898 std::swap(PredCases[i], PredCases.back());
899 PredCases.pop_back();
903 // Okay, now we know which constants were sent to BB from the
904 // predecessor. Figure out where they will all go now.
905 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
906 if (PTIHandled.count(BBCases[i].Value)) {
907 // If this is one we are capable of getting...
908 if (PredHasWeights || SuccHasWeights)
909 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
910 PredCases.push_back(BBCases[i]);
911 NewSuccessors.push_back(BBCases[i].Dest);
912 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
915 // If there are any constants vectored to BB that TI doesn't handle,
916 // they must go to the default destination of TI.
917 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
919 E = PTIHandled.end(); I != E; ++I) {
920 if (PredHasWeights || SuccHasWeights)
921 Weights.push_back(WeightsForHandled[*I]);
922 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
923 NewSuccessors.push_back(BBDefault);
927 // Okay, at this point, we know which new successor Pred will get. Make
928 // sure we update the number of entries in the PHI nodes for these
930 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i)
931 AddPredecessorToBlock(NewSuccessors[i], Pred, BB);
933 Builder.SetInsertPoint(PTI);
934 // Convert pointer to int before we switch.
935 if (CV->getType()->isPointerTy()) {
936 assert(DL && "Cannot switch on pointer without DataLayout");
937 CV = Builder.CreatePtrToInt(CV, DL->getIntPtrType(CV->getType()),
941 // Now that the successors are updated, create the new Switch instruction.
942 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
944 NewSI->setDebugLoc(PTI->getDebugLoc());
945 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
946 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest);
948 if (PredHasWeights || SuccHasWeights) {
949 // Halve the weights if any of them cannot fit in an uint32_t
952 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
954 NewSI->setMetadata(LLVMContext::MD_prof,
955 MDBuilder(BB->getContext()).
956 createBranchWeights(MDWeights));
959 EraseTerminatorInstAndDCECond(PTI);
961 // Okay, last check. If BB is still a successor of PSI, then we must
962 // have an infinite loop case. If so, add an infinitely looping block
963 // to handle the case to preserve the behavior of the code.
964 BasicBlock *InfLoopBlock = nullptr;
965 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
966 if (NewSI->getSuccessor(i) == BB) {
968 // Insert it at the end of the function, because it's either code,
969 // or it won't matter if it's hot. :)
970 InfLoopBlock = BasicBlock::Create(BB->getContext(),
971 "infloop", BB->getParent());
972 BranchInst::Create(InfLoopBlock, InfLoopBlock);
974 NewSI->setSuccessor(i, InfLoopBlock);
983 // isSafeToHoistInvoke - If we would need to insert a select that uses the
984 // value of this invoke (comments in HoistThenElseCodeToIf explain why we
985 // would need to do this), we can't hoist the invoke, as there is nowhere
986 // to put the select in this case.
987 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
988 Instruction *I1, Instruction *I2) {
989 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
991 for (BasicBlock::iterator BBI = SI->begin();
992 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
993 Value *BB1V = PN->getIncomingValueForBlock(BB1);
994 Value *BB2V = PN->getIncomingValueForBlock(BB2);
995 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
1003 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I);
1005 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
1006 /// BB2, hoist any common code in the two blocks up into the branch block. The
1007 /// caller of this function guarantees that BI's block dominates BB1 and BB2.
1008 static bool HoistThenElseCodeToIf(BranchInst *BI, const DataLayout *DL) {
1009 // This does very trivial matching, with limited scanning, to find identical
1010 // instructions in the two blocks. In particular, we don't want to get into
1011 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1012 // such, we currently just scan for obviously identical instructions in an
1014 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1015 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1017 BasicBlock::iterator BB1_Itr = BB1->begin();
1018 BasicBlock::iterator BB2_Itr = BB2->begin();
1020 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++;
1021 // Skip debug info if it is not identical.
1022 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1023 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1024 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1025 while (isa<DbgInfoIntrinsic>(I1))
1027 while (isa<DbgInfoIntrinsic>(I2))
1030 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1031 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1034 BasicBlock *BIParent = BI->getParent();
1036 bool Changed = false;
1038 // If we are hoisting the terminator instruction, don't move one (making a
1039 // broken BB), instead clone it, and remove BI.
1040 if (isa<TerminatorInst>(I1))
1041 goto HoistTerminator;
1043 // For a normal instruction, we just move one to right before the branch,
1044 // then replace all uses of the other with the first. Finally, we remove
1045 // the now redundant second instruction.
1046 BIParent->getInstList().splice(BI, BB1->getInstList(), I1);
1047 if (!I2->use_empty())
1048 I2->replaceAllUsesWith(I1);
1049 I1->intersectOptionalDataWith(I2);
1050 unsigned KnownIDs[] = {
1051 LLVMContext::MD_tbaa,
1052 LLVMContext::MD_range,
1053 LLVMContext::MD_fpmath,
1054 LLVMContext::MD_invariant_load
1056 combineMetadata(I1, I2, KnownIDs);
1057 I2->eraseFromParent();
1062 // Skip debug info if it is not identical.
1063 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1064 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1065 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1066 while (isa<DbgInfoIntrinsic>(I1))
1068 while (isa<DbgInfoIntrinsic>(I2))
1071 } while (I1->isIdenticalToWhenDefined(I2));
1076 // It may not be possible to hoist an invoke.
1077 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1080 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1082 for (BasicBlock::iterator BBI = SI->begin();
1083 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1084 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1085 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1089 // Check for passingValueIsAlwaysUndefined here because we would rather
1090 // eliminate undefined control flow then converting it to a select.
1091 if (passingValueIsAlwaysUndefined(BB1V, PN) ||
1092 passingValueIsAlwaysUndefined(BB2V, PN))
1095 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V, DL))
1097 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V, DL))
1102 // Okay, it is safe to hoist the terminator.
1103 Instruction *NT = I1->clone();
1104 BIParent->getInstList().insert(BI, NT);
1105 if (!NT->getType()->isVoidTy()) {
1106 I1->replaceAllUsesWith(NT);
1107 I2->replaceAllUsesWith(NT);
1111 IRBuilder<true, NoFolder> Builder(NT);
1112 // Hoisting one of the terminators from our successor is a great thing.
1113 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1114 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1115 // nodes, so we insert select instruction to compute the final result.
1116 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1117 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1119 for (BasicBlock::iterator BBI = SI->begin();
1120 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1121 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1122 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1123 if (BB1V == BB2V) continue;
1125 // These values do not agree. Insert a select instruction before NT
1126 // that determines the right value.
1127 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1129 SI = cast<SelectInst>
1130 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1131 BB1V->getName()+"."+BB2V->getName()));
1133 // Make the PHI node use the select for all incoming values for BB1/BB2
1134 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1135 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1136 PN->setIncomingValue(i, SI);
1140 // Update any PHI nodes in our new successors.
1141 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1142 AddPredecessorToBlock(*SI, BIParent, BB1);
1144 EraseTerminatorInstAndDCECond(BI);
1148 /// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd,
1149 /// check whether BBEnd has only two predecessors and the other predecessor
1150 /// ends with an unconditional branch. If it is true, sink any common code
1151 /// in the two predecessors to BBEnd.
1152 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1153 assert(BI1->isUnconditional());
1154 BasicBlock *BB1 = BI1->getParent();
1155 BasicBlock *BBEnd = BI1->getSuccessor(0);
1157 // Check that BBEnd has two predecessors and the other predecessor ends with
1158 // an unconditional branch.
1159 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1160 BasicBlock *Pred0 = *PI++;
1161 if (PI == PE) // Only one predecessor.
1163 BasicBlock *Pred1 = *PI++;
1164 if (PI != PE) // More than two predecessors.
1166 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1167 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1168 if (!BI2 || !BI2->isUnconditional())
1171 // Gather the PHI nodes in BBEnd.
1172 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2;
1173 Instruction *FirstNonPhiInBBEnd = nullptr;
1174 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end();
1176 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1177 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1178 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1179 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN);
1181 FirstNonPhiInBBEnd = &*I;
1185 if (!FirstNonPhiInBBEnd)
1189 // This does very trivial matching, with limited scanning, to find identical
1190 // instructions in the two blocks. We scan backward for obviously identical
1191 // instructions in an identical order.
1192 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1193 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(),
1194 RE2 = BB2->getInstList().rend();
1196 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1199 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1202 // Skip the unconditional branches.
1206 bool Changed = false;
1207 while (RI1 != RE1 && RI2 != RE2) {
1209 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1212 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1216 Instruction *I1 = &*RI1, *I2 = &*RI2;
1217 // I1 and I2 should have a single use in the same PHI node, and they
1218 // perform the same operation.
1219 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1220 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1221 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1222 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) ||
1223 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1224 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1225 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1226 !I1->hasOneUse() || !I2->hasOneUse() ||
1227 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() ||
1228 MapValueFromBB1ToBB2[I1].first != I2)
1231 // Check whether we should swap the operands of ICmpInst.
1232 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1233 bool SwapOpnds = false;
1234 if (ICmp1 && ICmp2 &&
1235 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1236 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1237 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1238 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1239 ICmp2->swapOperands();
1242 if (!I1->isSameOperationAs(I2)) {
1244 ICmp2->swapOperands();
1248 // The operands should be either the same or they need to be generated
1249 // with a PHI node after sinking. We only handle the case where there is
1250 // a single pair of different operands.
1251 Value *DifferentOp1 = nullptr, *DifferentOp2 = nullptr;
1252 unsigned Op1Idx = 0;
1253 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1254 if (I1->getOperand(I) == I2->getOperand(I))
1256 // Early exit if we have more-than one pair of different operands or
1257 // the different operand is already in MapValueFromBB1ToBB2.
1258 // Early exit if we need a PHI node to replace a constant.
1260 MapValueFromBB1ToBB2.find(I1->getOperand(I)) !=
1261 MapValueFromBB1ToBB2.end() ||
1262 isa<Constant>(I1->getOperand(I)) ||
1263 isa<Constant>(I2->getOperand(I))) {
1264 // If we can't sink the instructions, undo the swapping.
1266 ICmp2->swapOperands();
1269 DifferentOp1 = I1->getOperand(I);
1271 DifferentOp2 = I2->getOperand(I);
1274 // We insert the pair of different operands to MapValueFromBB1ToBB2 and
1275 // remove (I1, I2) from MapValueFromBB1ToBB2.
1277 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2,
1278 DifferentOp1->getName() + ".sink",
1280 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN);
1281 // I1 should use NewPN instead of DifferentOp1.
1282 I1->setOperand(Op1Idx, NewPN);
1283 NewPN->addIncoming(DifferentOp1, BB1);
1284 NewPN->addIncoming(DifferentOp2, BB2);
1285 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1287 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second;
1288 MapValueFromBB1ToBB2.erase(I1);
1290 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";);
1291 DEBUG(dbgs() << " " << *I2 << "\n";);
1292 // We need to update RE1 and RE2 if we are going to sink the first
1293 // instruction in the basic block down.
1294 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1295 // Sink the instruction.
1296 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1);
1297 if (!OldPN->use_empty())
1298 OldPN->replaceAllUsesWith(I1);
1299 OldPN->eraseFromParent();
1301 if (!I2->use_empty())
1302 I2->replaceAllUsesWith(I1);
1303 I1->intersectOptionalDataWith(I2);
1304 I2->eraseFromParent();
1307 RE1 = BB1->getInstList().rend();
1309 RE2 = BB2->getInstList().rend();
1310 FirstNonPhiInBBEnd = I1;
1317 /// \brief Determine if we can hoist sink a sole store instruction out of a
1318 /// conditional block.
1320 /// We are looking for code like the following:
1322 /// store i32 %add, i32* %arrayidx2
1323 /// ... // No other stores or function calls (we could be calling a memory
1324 /// ... // function).
1325 /// %cmp = icmp ult %x, %y
1326 /// br i1 %cmp, label %EndBB, label %ThenBB
1328 /// store i32 %add5, i32* %arrayidx2
1332 /// We are going to transform this into:
1334 /// store i32 %add, i32* %arrayidx2
1336 /// %cmp = icmp ult %x, %y
1337 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1338 /// store i32 %add.add5, i32* %arrayidx2
1341 /// \return The pointer to the value of the previous store if the store can be
1342 /// hoisted into the predecessor block. 0 otherwise.
1343 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1344 BasicBlock *StoreBB, BasicBlock *EndBB) {
1345 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1349 // Volatile or atomic.
1350 if (!StoreToHoist->isSimple())
1353 Value *StorePtr = StoreToHoist->getPointerOperand();
1355 // Look for a store to the same pointer in BrBB.
1356 unsigned MaxNumInstToLookAt = 10;
1357 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(),
1358 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) {
1359 Instruction *CurI = &*RI;
1361 // Could be calling an instruction that effects memory like free().
1362 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI))
1365 StoreInst *SI = dyn_cast<StoreInst>(CurI);
1366 // Found the previous store make sure it stores to the same location.
1367 if (SI && SI->getPointerOperand() == StorePtr)
1368 // Found the previous store, return its value operand.
1369 return SI->getValueOperand();
1371 return nullptr; // Unknown store.
1377 /// \brief Speculate a conditional basic block flattening the CFG.
1379 /// Note that this is a very risky transform currently. Speculating
1380 /// instructions like this is most often not desirable. Instead, there is an MI
1381 /// pass which can do it with full awareness of the resource constraints.
1382 /// However, some cases are "obvious" and we should do directly. An example of
1383 /// this is speculating a single, reasonably cheap instruction.
1385 /// There is only one distinct advantage to flattening the CFG at the IR level:
1386 /// it makes very common but simplistic optimizations such as are common in
1387 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1388 /// modeling their effects with easier to reason about SSA value graphs.
1391 /// An illustration of this transform is turning this IR:
1394 /// %cmp = icmp ult %x, %y
1395 /// br i1 %cmp, label %EndBB, label %ThenBB
1397 /// %sub = sub %x, %y
1400 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1407 /// %cmp = icmp ult %x, %y
1408 /// %sub = sub %x, %y
1409 /// %cond = select i1 %cmp, 0, %sub
1413 /// \returns true if the conditional block is removed.
1414 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1415 const DataLayout *DL) {
1416 // Be conservative for now. FP select instruction can often be expensive.
1417 Value *BrCond = BI->getCondition();
1418 if (isa<FCmpInst>(BrCond))
1421 BasicBlock *BB = BI->getParent();
1422 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1424 // If ThenBB is actually on the false edge of the conditional branch, remember
1425 // to swap the select operands later.
1426 bool Invert = false;
1427 if (ThenBB != BI->getSuccessor(0)) {
1428 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1431 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1433 // Keep a count of how many times instructions are used within CondBB when
1434 // they are candidates for sinking into CondBB. Specifically:
1435 // - They are defined in BB, and
1436 // - They have no side effects, and
1437 // - All of their uses are in CondBB.
1438 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1440 unsigned SpeculationCost = 0;
1441 Value *SpeculatedStoreValue = nullptr;
1442 StoreInst *SpeculatedStore = nullptr;
1443 for (BasicBlock::iterator BBI = ThenBB->begin(),
1444 BBE = std::prev(ThenBB->end());
1445 BBI != BBE; ++BBI) {
1446 Instruction *I = BBI;
1448 if (isa<DbgInfoIntrinsic>(I))
1451 // Only speculatively execution a single instruction (not counting the
1452 // terminator) for now.
1454 if (SpeculationCost > 1)
1457 // Don't hoist the instruction if it's unsafe or expensive.
1458 if (!isSafeToSpeculativelyExecute(I, DL) &&
1459 !(HoistCondStores &&
1460 (SpeculatedStoreValue = isSafeToSpeculateStore(I, BB, ThenBB,
1463 if (!SpeculatedStoreValue &&
1464 ComputeSpeculationCost(I, DL) > PHINodeFoldingThreshold)
1467 // Store the store speculation candidate.
1468 if (SpeculatedStoreValue)
1469 SpeculatedStore = cast<StoreInst>(I);
1471 // Do not hoist the instruction if any of its operands are defined but not
1472 // used in BB. The transformation will prevent the operand from
1473 // being sunk into the use block.
1474 for (User::op_iterator i = I->op_begin(), e = I->op_end();
1476 Instruction *OpI = dyn_cast<Instruction>(*i);
1477 if (!OpI || OpI->getParent() != BB ||
1478 OpI->mayHaveSideEffects())
1479 continue; // Not a candidate for sinking.
1481 ++SinkCandidateUseCounts[OpI];
1485 // Consider any sink candidates which are only used in CondBB as costs for
1486 // speculation. Note, while we iterate over a DenseMap here, we are summing
1487 // and so iteration order isn't significant.
1488 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I =
1489 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end();
1491 if (I->first->getNumUses() == I->second) {
1493 if (SpeculationCost > 1)
1497 // Check that the PHI nodes can be converted to selects.
1498 bool HaveRewritablePHIs = false;
1499 for (BasicBlock::iterator I = EndBB->begin();
1500 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1501 Value *OrigV = PN->getIncomingValueForBlock(BB);
1502 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1504 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1505 // Skip PHIs which are trivial.
1509 // Don't convert to selects if we could remove undefined behavior instead.
1510 if (passingValueIsAlwaysUndefined(OrigV, PN) ||
1511 passingValueIsAlwaysUndefined(ThenV, PN))
1514 HaveRewritablePHIs = true;
1515 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1516 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1517 if (!OrigCE && !ThenCE)
1518 continue; // Known safe and cheap.
1520 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE, DL)) ||
1521 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE, DL)))
1523 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, DL) : 0;
1524 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, DL) : 0;
1525 if (OrigCost + ThenCost > 2 * PHINodeFoldingThreshold)
1528 // Account for the cost of an unfolded ConstantExpr which could end up
1529 // getting expanded into Instructions.
1530 // FIXME: This doesn't account for how many operations are combined in the
1531 // constant expression.
1533 if (SpeculationCost > 1)
1537 // If there are no PHIs to process, bail early. This helps ensure idempotence
1539 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
1542 // If we get here, we can hoist the instruction and if-convert.
1543 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1545 // Insert a select of the value of the speculated store.
1546 if (SpeculatedStoreValue) {
1547 IRBuilder<true, NoFolder> Builder(BI);
1548 Value *TrueV = SpeculatedStore->getValueOperand();
1549 Value *FalseV = SpeculatedStoreValue;
1551 std::swap(TrueV, FalseV);
1552 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() +
1553 "." + FalseV->getName());
1554 SpeculatedStore->setOperand(0, S);
1557 // Hoist the instructions.
1558 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(),
1559 std::prev(ThenBB->end()));
1561 // Insert selects and rewrite the PHI operands.
1562 IRBuilder<true, NoFolder> Builder(BI);
1563 for (BasicBlock::iterator I = EndBB->begin();
1564 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1565 unsigned OrigI = PN->getBasicBlockIndex(BB);
1566 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1567 Value *OrigV = PN->getIncomingValue(OrigI);
1568 Value *ThenV = PN->getIncomingValue(ThenI);
1570 // Skip PHIs which are trivial.
1574 // Create a select whose true value is the speculatively executed value and
1575 // false value is the preexisting value. Swap them if the branch
1576 // destinations were inverted.
1577 Value *TrueV = ThenV, *FalseV = OrigV;
1579 std::swap(TrueV, FalseV);
1580 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV,
1581 TrueV->getName() + "." + FalseV->getName());
1582 PN->setIncomingValue(OrigI, V);
1583 PN->setIncomingValue(ThenI, V);
1590 /// \returns True if this block contains a CallInst with the NoDuplicate
1592 static bool HasNoDuplicateCall(const BasicBlock *BB) {
1593 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1594 const CallInst *CI = dyn_cast<CallInst>(I);
1597 if (CI->cannotDuplicate())
1603 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1604 /// across this block.
1605 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1606 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1609 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1610 if (isa<DbgInfoIntrinsic>(BBI))
1612 if (Size > 10) return false; // Don't clone large BB's.
1615 // We can only support instructions that do not define values that are
1616 // live outside of the current basic block.
1617 for (User *U : BBI->users()) {
1618 Instruction *UI = cast<Instruction>(U);
1619 if (UI->getParent() != BB || isa<PHINode>(UI)) return false;
1622 // Looks ok, continue checking.
1628 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1629 /// that is defined in the same block as the branch and if any PHI entries are
1630 /// constants, thread edges corresponding to that entry to be branches to their
1631 /// ultimate destination.
1632 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *DL) {
1633 BasicBlock *BB = BI->getParent();
1634 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1635 // NOTE: we currently cannot transform this case if the PHI node is used
1636 // outside of the block.
1637 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1640 // Degenerate case of a single entry PHI.
1641 if (PN->getNumIncomingValues() == 1) {
1642 FoldSingleEntryPHINodes(PN->getParent());
1646 // Now we know that this block has multiple preds and two succs.
1647 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1649 if (HasNoDuplicateCall(BB)) return false;
1651 // Okay, this is a simple enough basic block. See if any phi values are
1653 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1654 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1655 if (!CB || !CB->getType()->isIntegerTy(1)) continue;
1657 // Okay, we now know that all edges from PredBB should be revectored to
1658 // branch to RealDest.
1659 BasicBlock *PredBB = PN->getIncomingBlock(i);
1660 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1662 if (RealDest == BB) continue; // Skip self loops.
1663 // Skip if the predecessor's terminator is an indirect branch.
1664 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1666 // The dest block might have PHI nodes, other predecessors and other
1667 // difficult cases. Instead of being smart about this, just insert a new
1668 // block that jumps to the destination block, effectively splitting
1669 // the edge we are about to create.
1670 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1671 RealDest->getName()+".critedge",
1672 RealDest->getParent(), RealDest);
1673 BranchInst::Create(RealDest, EdgeBB);
1675 // Update PHI nodes.
1676 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1678 // BB may have instructions that are being threaded over. Clone these
1679 // instructions into EdgeBB. We know that there will be no uses of the
1680 // cloned instructions outside of EdgeBB.
1681 BasicBlock::iterator InsertPt = EdgeBB->begin();
1682 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1683 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1684 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1685 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1688 // Clone the instruction.
1689 Instruction *N = BBI->clone();
1690 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1692 // Update operands due to translation.
1693 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1695 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1696 if (PI != TranslateMap.end())
1700 // Check for trivial simplification.
1701 if (Value *V = SimplifyInstruction(N, DL)) {
1702 TranslateMap[BBI] = V;
1703 delete N; // Instruction folded away, don't need actual inst
1705 // Insert the new instruction into its new home.
1706 EdgeBB->getInstList().insert(InsertPt, N);
1707 if (!BBI->use_empty())
1708 TranslateMap[BBI] = N;
1712 // Loop over all of the edges from PredBB to BB, changing them to branch
1713 // to EdgeBB instead.
1714 TerminatorInst *PredBBTI = PredBB->getTerminator();
1715 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1716 if (PredBBTI->getSuccessor(i) == BB) {
1717 BB->removePredecessor(PredBB);
1718 PredBBTI->setSuccessor(i, EdgeBB);
1721 // Recurse, simplifying any other constants.
1722 return FoldCondBranchOnPHI(BI, DL) | true;
1728 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1729 /// PHI node, see if we can eliminate it.
1730 static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *DL) {
1731 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1732 // statement", which has a very simple dominance structure. Basically, we
1733 // are trying to find the condition that is being branched on, which
1734 // subsequently causes this merge to happen. We really want control
1735 // dependence information for this check, but simplifycfg can't keep it up
1736 // to date, and this catches most of the cases we care about anyway.
1737 BasicBlock *BB = PN->getParent();
1738 BasicBlock *IfTrue, *IfFalse;
1739 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1741 // Don't bother if the branch will be constant folded trivially.
1742 isa<ConstantInt>(IfCond))
1745 // Okay, we found that we can merge this two-entry phi node into a select.
1746 // Doing so would require us to fold *all* two entry phi nodes in this block.
1747 // At some point this becomes non-profitable (particularly if the target
1748 // doesn't support cmov's). Only do this transformation if there are two or
1749 // fewer PHI nodes in this block.
1750 unsigned NumPhis = 0;
1751 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1755 // Loop over the PHI's seeing if we can promote them all to select
1756 // instructions. While we are at it, keep track of the instructions
1757 // that need to be moved to the dominating block.
1758 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1759 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1760 MaxCostVal1 = PHINodeFoldingThreshold;
1762 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1763 PHINode *PN = cast<PHINode>(II++);
1764 if (Value *V = SimplifyInstruction(PN, DL)) {
1765 PN->replaceAllUsesWith(V);
1766 PN->eraseFromParent();
1770 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1772 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1777 // If we folded the first phi, PN dangles at this point. Refresh it. If
1778 // we ran out of PHIs then we simplified them all.
1779 PN = dyn_cast<PHINode>(BB->begin());
1780 if (!PN) return true;
1782 // Don't fold i1 branches on PHIs which contain binary operators. These can
1783 // often be turned into switches and other things.
1784 if (PN->getType()->isIntegerTy(1) &&
1785 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1786 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1787 isa<BinaryOperator>(IfCond)))
1790 // If we all PHI nodes are promotable, check to make sure that all
1791 // instructions in the predecessor blocks can be promoted as well. If
1792 // not, we won't be able to get rid of the control flow, so it's not
1793 // worth promoting to select instructions.
1794 BasicBlock *DomBlock = nullptr;
1795 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1796 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1797 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1800 DomBlock = *pred_begin(IfBlock1);
1801 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1802 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1803 // This is not an aggressive instruction that we can promote.
1804 // Because of this, we won't be able to get rid of the control
1805 // flow, so the xform is not worth it.
1810 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1813 DomBlock = *pred_begin(IfBlock2);
1814 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1815 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1816 // This is not an aggressive instruction that we can promote.
1817 // Because of this, we won't be able to get rid of the control
1818 // flow, so the xform is not worth it.
1823 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1824 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1826 // If we can still promote the PHI nodes after this gauntlet of tests,
1827 // do all of the PHI's now.
1828 Instruction *InsertPt = DomBlock->getTerminator();
1829 IRBuilder<true, NoFolder> Builder(InsertPt);
1831 // Move all 'aggressive' instructions, which are defined in the
1832 // conditional parts of the if's up to the dominating block.
1834 DomBlock->getInstList().splice(InsertPt,
1835 IfBlock1->getInstList(), IfBlock1->begin(),
1836 IfBlock1->getTerminator());
1838 DomBlock->getInstList().splice(InsertPt,
1839 IfBlock2->getInstList(), IfBlock2->begin(),
1840 IfBlock2->getTerminator());
1842 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1843 // Change the PHI node into a select instruction.
1844 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1845 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1848 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1849 PN->replaceAllUsesWith(NV);
1851 PN->eraseFromParent();
1854 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1855 // has been flattened. Change DomBlock to jump directly to our new block to
1856 // avoid other simplifycfg's kicking in on the diamond.
1857 TerminatorInst *OldTI = DomBlock->getTerminator();
1858 Builder.SetInsertPoint(OldTI);
1859 Builder.CreateBr(BB);
1860 OldTI->eraseFromParent();
1864 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1865 /// to two returning blocks, try to merge them together into one return,
1866 /// introducing a select if the return values disagree.
1867 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1868 IRBuilder<> &Builder) {
1869 assert(BI->isConditional() && "Must be a conditional branch");
1870 BasicBlock *TrueSucc = BI->getSuccessor(0);
1871 BasicBlock *FalseSucc = BI->getSuccessor(1);
1872 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1873 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1875 // Check to ensure both blocks are empty (just a return) or optionally empty
1876 // with PHI nodes. If there are other instructions, merging would cause extra
1877 // computation on one path or the other.
1878 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1880 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1883 Builder.SetInsertPoint(BI);
1884 // Okay, we found a branch that is going to two return nodes. If
1885 // there is no return value for this function, just change the
1886 // branch into a return.
1887 if (FalseRet->getNumOperands() == 0) {
1888 TrueSucc->removePredecessor(BI->getParent());
1889 FalseSucc->removePredecessor(BI->getParent());
1890 Builder.CreateRetVoid();
1891 EraseTerminatorInstAndDCECond(BI);
1895 // Otherwise, figure out what the true and false return values are
1896 // so we can insert a new select instruction.
1897 Value *TrueValue = TrueRet->getReturnValue();
1898 Value *FalseValue = FalseRet->getReturnValue();
1900 // Unwrap any PHI nodes in the return blocks.
1901 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1902 if (TVPN->getParent() == TrueSucc)
1903 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1904 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1905 if (FVPN->getParent() == FalseSucc)
1906 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1908 // In order for this transformation to be safe, we must be able to
1909 // unconditionally execute both operands to the return. This is
1910 // normally the case, but we could have a potentially-trapping
1911 // constant expression that prevents this transformation from being
1913 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1916 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1920 // Okay, we collected all the mapped values and checked them for sanity, and
1921 // defined to really do this transformation. First, update the CFG.
1922 TrueSucc->removePredecessor(BI->getParent());
1923 FalseSucc->removePredecessor(BI->getParent());
1925 // Insert select instructions where needed.
1926 Value *BrCond = BI->getCondition();
1928 // Insert a select if the results differ.
1929 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1930 } else if (isa<UndefValue>(TrueValue)) {
1931 TrueValue = FalseValue;
1933 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1934 FalseValue, "retval");
1938 Value *RI = !TrueValue ?
1939 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
1943 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
1944 << "\n " << *BI << "NewRet = " << *RI
1945 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
1947 EraseTerminatorInstAndDCECond(BI);
1952 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
1953 /// probabilities of the branch taking each edge. Fills in the two APInt
1954 /// parameters and return true, or returns false if no or invalid metadata was
1956 static bool ExtractBranchMetadata(BranchInst *BI,
1957 uint64_t &ProbTrue, uint64_t &ProbFalse) {
1958 assert(BI->isConditional() &&
1959 "Looking for probabilities on unconditional branch?");
1960 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
1961 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
1962 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
1963 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
1964 if (!CITrue || !CIFalse) return false;
1965 ProbTrue = CITrue->getValue().getZExtValue();
1966 ProbFalse = CIFalse->getValue().getZExtValue();
1970 /// checkCSEInPredecessor - Return true if the given instruction is available
1971 /// in its predecessor block. If yes, the instruction will be removed.
1973 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
1974 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
1976 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
1977 Instruction *PBI = &*I;
1978 // Check whether Inst and PBI generate the same value.
1979 if (Inst->isIdenticalTo(PBI)) {
1980 Inst->replaceAllUsesWith(PBI);
1981 Inst->eraseFromParent();
1988 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
1989 /// predecessor branches to us and one of our successors, fold the block into
1990 /// the predecessor and use logical operations to pick the right destination.
1991 bool llvm::FoldBranchToCommonDest(BranchInst *BI, const DataLayout *DL,
1992 unsigned BonusInstThreshold) {
1993 BasicBlock *BB = BI->getParent();
1995 Instruction *Cond = nullptr;
1996 if (BI->isConditional())
1997 Cond = dyn_cast<Instruction>(BI->getCondition());
1999 // For unconditional branch, check for a simple CFG pattern, where
2000 // BB has a single predecessor and BB's successor is also its predecessor's
2001 // successor. If such pattern exisits, check for CSE between BB and its
2003 if (BasicBlock *PB = BB->getSinglePredecessor())
2004 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2005 if (PBI->isConditional() &&
2006 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2007 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2008 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
2010 Instruction *Curr = I++;
2011 if (isa<CmpInst>(Curr)) {
2015 // Quit if we can't remove this instruction.
2016 if (!checkCSEInPredecessor(Curr, PB))
2025 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2026 Cond->getParent() != BB || !Cond->hasOneUse())
2029 // Make sure the instruction after the condition is the cond branch.
2030 BasicBlock::iterator CondIt = Cond; ++CondIt;
2032 // Ignore dbg intrinsics.
2033 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
2038 // Only allow this transformation if computing the condition doesn't involve
2039 // too many instructions and these involved instructions can be executed
2040 // unconditionally. We denote all involved instructions except the condition
2041 // as "bonus instructions", and only allow this transformation when the
2042 // number of the bonus instructions does not exceed a certain threshold.
2043 unsigned NumBonusInsts = 0;
2044 for (auto I = BB->begin(); Cond != I; ++I) {
2045 // Ignore dbg intrinsics.
2046 if (isa<DbgInfoIntrinsic>(I))
2048 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(I, DL))
2050 // I has only one use and can be executed unconditionally.
2051 Instruction *User = dyn_cast<Instruction>(I->user_back());
2052 if (User == nullptr || User->getParent() != BB)
2054 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2055 // to use any other instruction, User must be an instruction between next(I)
2058 // Early exits once we reach the limit.
2059 if (NumBonusInsts > BonusInstThreshold)
2063 // Cond is known to be a compare or binary operator. Check to make sure that
2064 // neither operand is a potentially-trapping constant expression.
2065 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2068 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2072 // Finally, don't infinitely unroll conditional loops.
2073 BasicBlock *TrueDest = BI->getSuccessor(0);
2074 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2075 if (TrueDest == BB || FalseDest == BB)
2078 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2079 BasicBlock *PredBlock = *PI;
2080 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2082 // Check that we have two conditional branches. If there is a PHI node in
2083 // the common successor, verify that the same value flows in from both
2085 SmallVector<PHINode*, 4> PHIs;
2086 if (!PBI || PBI->isUnconditional() ||
2087 (BI->isConditional() &&
2088 !SafeToMergeTerminators(BI, PBI)) ||
2089 (!BI->isConditional() &&
2090 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2093 // Determine if the two branches share a common destination.
2094 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2095 bool InvertPredCond = false;
2097 if (BI->isConditional()) {
2098 if (PBI->getSuccessor(0) == TrueDest)
2099 Opc = Instruction::Or;
2100 else if (PBI->getSuccessor(1) == FalseDest)
2101 Opc = Instruction::And;
2102 else if (PBI->getSuccessor(0) == FalseDest)
2103 Opc = Instruction::And, InvertPredCond = true;
2104 else if (PBI->getSuccessor(1) == TrueDest)
2105 Opc = Instruction::Or, InvertPredCond = true;
2109 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2113 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2114 IRBuilder<> Builder(PBI);
2116 // If we need to invert the condition in the pred block to match, do so now.
2117 if (InvertPredCond) {
2118 Value *NewCond = PBI->getCondition();
2120 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2121 CmpInst *CI = cast<CmpInst>(NewCond);
2122 CI->setPredicate(CI->getInversePredicate());
2124 NewCond = Builder.CreateNot(NewCond,
2125 PBI->getCondition()->getName()+".not");
2128 PBI->setCondition(NewCond);
2129 PBI->swapSuccessors();
2132 // If we have bonus instructions, clone them into the predecessor block.
2133 // Note that there may be mutliple predecessor blocks, so we cannot move
2134 // bonus instructions to a predecessor block.
2135 ValueToValueMapTy VMap; // maps original values to cloned values
2136 // We already make sure Cond is the last instruction before BI. Therefore,
2137 // every instructions before Cond other than DbgInfoIntrinsic are bonus
2139 for (auto BonusInst = BB->begin(); Cond != BonusInst; ++BonusInst) {
2140 if (isa<DbgInfoIntrinsic>(BonusInst))
2142 Instruction *NewBonusInst = BonusInst->clone();
2143 RemapInstruction(NewBonusInst, VMap,
2144 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2145 VMap[BonusInst] = NewBonusInst;
2147 // If we moved a load, we cannot any longer claim any knowledge about
2148 // its potential value. The previous information might have been valid
2149 // only given the branch precondition.
2150 // For an analogous reason, we must also drop all the metadata whose
2151 // semantics we don't understand.
2152 NewBonusInst->dropUnknownMetadata(LLVMContext::MD_dbg);
2154 PredBlock->getInstList().insert(PBI, NewBonusInst);
2155 NewBonusInst->takeName(BonusInst);
2156 BonusInst->setName(BonusInst->getName() + ".old");
2159 // Clone Cond into the predecessor basic block, and or/and the
2160 // two conditions together.
2161 Instruction *New = Cond->clone();
2162 RemapInstruction(New, VMap,
2163 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2164 PredBlock->getInstList().insert(PBI, New);
2165 New->takeName(Cond);
2166 Cond->setName(New->getName() + ".old");
2168 if (BI->isConditional()) {
2169 Instruction *NewCond =
2170 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2172 PBI->setCondition(NewCond);
2174 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2175 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2177 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2179 SmallVector<uint64_t, 8> NewWeights;
2181 if (PBI->getSuccessor(0) == BB) {
2182 if (PredHasWeights && SuccHasWeights) {
2183 // PBI: br i1 %x, BB, FalseDest
2184 // BI: br i1 %y, TrueDest, FalseDest
2185 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2186 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2187 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2188 // TrueWeight for PBI * FalseWeight for BI.
2189 // We assume that total weights of a BranchInst can fit into 32 bits.
2190 // Therefore, we will not have overflow using 64-bit arithmetic.
2191 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2192 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2194 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2195 PBI->setSuccessor(0, TrueDest);
2197 if (PBI->getSuccessor(1) == BB) {
2198 if (PredHasWeights && SuccHasWeights) {
2199 // PBI: br i1 %x, TrueDest, BB
2200 // BI: br i1 %y, TrueDest, FalseDest
2201 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2202 // FalseWeight for PBI * TrueWeight for BI.
2203 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2204 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2205 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2206 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2208 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2209 PBI->setSuccessor(1, FalseDest);
2211 if (NewWeights.size() == 2) {
2212 // Halve the weights if any of them cannot fit in an uint32_t
2213 FitWeights(NewWeights);
2215 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2216 PBI->setMetadata(LLVMContext::MD_prof,
2217 MDBuilder(BI->getContext()).
2218 createBranchWeights(MDWeights));
2220 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2222 // Update PHI nodes in the common successors.
2223 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2224 ConstantInt *PBI_C = cast<ConstantInt>(
2225 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2226 assert(PBI_C->getType()->isIntegerTy(1));
2227 Instruction *MergedCond = nullptr;
2228 if (PBI->getSuccessor(0) == TrueDest) {
2229 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2230 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2231 // is false: !PBI_Cond and BI_Value
2232 Instruction *NotCond =
2233 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2236 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2241 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2242 PBI->getCondition(), MergedCond,
2245 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2246 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2247 // is false: PBI_Cond and BI_Value
2249 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2250 PBI->getCondition(), New,
2252 if (PBI_C->isOne()) {
2253 Instruction *NotCond =
2254 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2257 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2258 NotCond, MergedCond,
2263 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2266 // Change PBI from Conditional to Unconditional.
2267 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2268 EraseTerminatorInstAndDCECond(PBI);
2272 // TODO: If BB is reachable from all paths through PredBlock, then we
2273 // could replace PBI's branch probabilities with BI's.
2275 // Copy any debug value intrinsics into the end of PredBlock.
2276 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2277 if (isa<DbgInfoIntrinsic>(*I))
2278 I->clone()->insertBefore(PBI);
2285 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2286 /// predecessor of another block, this function tries to simplify it. We know
2287 /// that PBI and BI are both conditional branches, and BI is in one of the
2288 /// successor blocks of PBI - PBI branches to BI.
2289 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2290 assert(PBI->isConditional() && BI->isConditional());
2291 BasicBlock *BB = BI->getParent();
2293 // If this block ends with a branch instruction, and if there is a
2294 // predecessor that ends on a branch of the same condition, make
2295 // this conditional branch redundant.
2296 if (PBI->getCondition() == BI->getCondition() &&
2297 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2298 // Okay, the outcome of this conditional branch is statically
2299 // knowable. If this block had a single pred, handle specially.
2300 if (BB->getSinglePredecessor()) {
2301 // Turn this into a branch on constant.
2302 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2303 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2305 return true; // Nuke the branch on constant.
2308 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2309 // in the constant and simplify the block result. Subsequent passes of
2310 // simplifycfg will thread the block.
2311 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2312 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2313 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2314 std::distance(PB, PE),
2315 BI->getCondition()->getName() + ".pr",
2317 // Okay, we're going to insert the PHI node. Since PBI is not the only
2318 // predecessor, compute the PHI'd conditional value for all of the preds.
2319 // Any predecessor where the condition is not computable we keep symbolic.
2320 for (pred_iterator PI = PB; PI != PE; ++PI) {
2321 BasicBlock *P = *PI;
2322 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2323 PBI != BI && PBI->isConditional() &&
2324 PBI->getCondition() == BI->getCondition() &&
2325 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2326 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2327 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2330 NewPN->addIncoming(BI->getCondition(), P);
2334 BI->setCondition(NewPN);
2339 // If this is a conditional branch in an empty block, and if any
2340 // predecessors are a conditional branch to one of our destinations,
2341 // fold the conditions into logical ops and one cond br.
2342 BasicBlock::iterator BBI = BB->begin();
2343 // Ignore dbg intrinsics.
2344 while (isa<DbgInfoIntrinsic>(BBI))
2350 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2355 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2357 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2358 PBIOp = 0, BIOp = 1;
2359 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2360 PBIOp = 1, BIOp = 0;
2361 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2366 // Check to make sure that the other destination of this branch
2367 // isn't BB itself. If so, this is an infinite loop that will
2368 // keep getting unwound.
2369 if (PBI->getSuccessor(PBIOp) == BB)
2372 // Do not perform this transformation if it would require
2373 // insertion of a large number of select instructions. For targets
2374 // without predication/cmovs, this is a big pessimization.
2376 // Also do not perform this transformation if any phi node in the common
2377 // destination block can trap when reached by BB or PBB (PR17073). In that
2378 // case, it would be unsafe to hoist the operation into a select instruction.
2380 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2381 unsigned NumPhis = 0;
2382 for (BasicBlock::iterator II = CommonDest->begin();
2383 isa<PHINode>(II); ++II, ++NumPhis) {
2384 if (NumPhis > 2) // Disable this xform.
2387 PHINode *PN = cast<PHINode>(II);
2388 Value *BIV = PN->getIncomingValueForBlock(BB);
2389 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
2393 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2394 Value *PBIV = PN->getIncomingValue(PBBIdx);
2395 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
2400 // Finally, if everything is ok, fold the branches to logical ops.
2401 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2403 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2404 << "AND: " << *BI->getParent());
2407 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2408 // branch in it, where one edge (OtherDest) goes back to itself but the other
2409 // exits. We don't *know* that the program avoids the infinite loop
2410 // (even though that seems likely). If we do this xform naively, we'll end up
2411 // recursively unpeeling the loop. Since we know that (after the xform is
2412 // done) that the block *is* infinite if reached, we just make it an obviously
2413 // infinite loop with no cond branch.
2414 if (OtherDest == BB) {
2415 // Insert it at the end of the function, because it's either code,
2416 // or it won't matter if it's hot. :)
2417 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2418 "infloop", BB->getParent());
2419 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2420 OtherDest = InfLoopBlock;
2423 DEBUG(dbgs() << *PBI->getParent()->getParent());
2425 // BI may have other predecessors. Because of this, we leave
2426 // it alone, but modify PBI.
2428 // Make sure we get to CommonDest on True&True directions.
2429 Value *PBICond = PBI->getCondition();
2430 IRBuilder<true, NoFolder> Builder(PBI);
2432 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2434 Value *BICond = BI->getCondition();
2436 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2438 // Merge the conditions.
2439 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2441 // Modify PBI to branch on the new condition to the new dests.
2442 PBI->setCondition(Cond);
2443 PBI->setSuccessor(0, CommonDest);
2444 PBI->setSuccessor(1, OtherDest);
2446 // Update branch weight for PBI.
2447 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2448 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2450 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2452 if (PredHasWeights && SuccHasWeights) {
2453 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2454 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2455 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2456 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2457 // The weight to CommonDest should be PredCommon * SuccTotal +
2458 // PredOther * SuccCommon.
2459 // The weight to OtherDest should be PredOther * SuccOther.
2460 SmallVector<uint64_t, 2> NewWeights;
2461 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) +
2462 PredOther * SuccCommon);
2463 NewWeights.push_back(PredOther * SuccOther);
2464 // Halve the weights if any of them cannot fit in an uint32_t
2465 FitWeights(NewWeights);
2467 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end());
2468 PBI->setMetadata(LLVMContext::MD_prof,
2469 MDBuilder(BI->getContext()).
2470 createBranchWeights(MDWeights));
2473 // OtherDest may have phi nodes. If so, add an entry from PBI's
2474 // block that are identical to the entries for BI's block.
2475 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2477 // We know that the CommonDest already had an edge from PBI to
2478 // it. If it has PHIs though, the PHIs may have different
2479 // entries for BB and PBI's BB. If so, insert a select to make
2482 for (BasicBlock::iterator II = CommonDest->begin();
2483 (PN = dyn_cast<PHINode>(II)); ++II) {
2484 Value *BIV = PN->getIncomingValueForBlock(BB);
2485 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2486 Value *PBIV = PN->getIncomingValue(PBBIdx);
2488 // Insert a select in PBI to pick the right value.
2489 Value *NV = cast<SelectInst>
2490 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2491 PN->setIncomingValue(PBBIdx, NV);
2495 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2496 DEBUG(dbgs() << *PBI->getParent()->getParent());
2498 // This basic block is probably dead. We know it has at least
2499 // one fewer predecessor.
2503 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2504 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2505 // Takes care of updating the successors and removing the old terminator.
2506 // Also makes sure not to introduce new successors by assuming that edges to
2507 // non-successor TrueBBs and FalseBBs aren't reachable.
2508 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2509 BasicBlock *TrueBB, BasicBlock *FalseBB,
2510 uint32_t TrueWeight,
2511 uint32_t FalseWeight){
2512 // Remove any superfluous successor edges from the CFG.
2513 // First, figure out which successors to preserve.
2514 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2516 BasicBlock *KeepEdge1 = TrueBB;
2517 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
2519 // Then remove the rest.
2520 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2521 BasicBlock *Succ = OldTerm->getSuccessor(I);
2522 // Make sure only to keep exactly one copy of each edge.
2523 if (Succ == KeepEdge1)
2524 KeepEdge1 = nullptr;
2525 else if (Succ == KeepEdge2)
2526 KeepEdge2 = nullptr;
2528 Succ->removePredecessor(OldTerm->getParent());
2531 IRBuilder<> Builder(OldTerm);
2532 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2534 // Insert an appropriate new terminator.
2535 if (!KeepEdge1 && !KeepEdge2) {
2536 if (TrueBB == FalseBB)
2537 // We were only looking for one successor, and it was present.
2538 // Create an unconditional branch to it.
2539 Builder.CreateBr(TrueBB);
2541 // We found both of the successors we were looking for.
2542 // Create a conditional branch sharing the condition of the select.
2543 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2544 if (TrueWeight != FalseWeight)
2545 NewBI->setMetadata(LLVMContext::MD_prof,
2546 MDBuilder(OldTerm->getContext()).
2547 createBranchWeights(TrueWeight, FalseWeight));
2549 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2550 // Neither of the selected blocks were successors, so this
2551 // terminator must be unreachable.
2552 new UnreachableInst(OldTerm->getContext(), OldTerm);
2554 // One of the selected values was a successor, but the other wasn't.
2555 // Insert an unconditional branch to the one that was found;
2556 // the edge to the one that wasn't must be unreachable.
2558 // Only TrueBB was found.
2559 Builder.CreateBr(TrueBB);
2561 // Only FalseBB was found.
2562 Builder.CreateBr(FalseBB);
2565 EraseTerminatorInstAndDCECond(OldTerm);
2569 // SimplifySwitchOnSelect - Replaces
2570 // (switch (select cond, X, Y)) on constant X, Y
2571 // with a branch - conditional if X and Y lead to distinct BBs,
2572 // unconditional otherwise.
2573 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2574 // Check for constant integer values in the select.
2575 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2576 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2577 if (!TrueVal || !FalseVal)
2580 // Find the relevant condition and destinations.
2581 Value *Condition = Select->getCondition();
2582 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2583 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2585 // Get weight for TrueBB and FalseBB.
2586 uint32_t TrueWeight = 0, FalseWeight = 0;
2587 SmallVector<uint64_t, 8> Weights;
2588 bool HasWeights = HasBranchWeights(SI);
2590 GetBranchWeights(SI, Weights);
2591 if (Weights.size() == 1 + SI->getNumCases()) {
2592 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2593 getSuccessorIndex()];
2594 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2595 getSuccessorIndex()];
2599 // Perform the actual simplification.
2600 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2601 TrueWeight, FalseWeight);
2604 // SimplifyIndirectBrOnSelect - Replaces
2605 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2606 // blockaddress(@fn, BlockB)))
2608 // (br cond, BlockA, BlockB).
2609 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2610 // Check that both operands of the select are block addresses.
2611 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2612 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2616 // Extract the actual blocks.
2617 BasicBlock *TrueBB = TBA->getBasicBlock();
2618 BasicBlock *FalseBB = FBA->getBasicBlock();
2620 // Perform the actual simplification.
2621 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2625 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2626 /// instruction (a seteq/setne with a constant) as the only instruction in a
2627 /// block that ends with an uncond branch. We are looking for a very specific
2628 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2629 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2630 /// default value goes to an uncond block with a seteq in it, we get something
2633 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2635 /// %tmp = icmp eq i8 %A, 92
2638 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2640 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2641 /// the PHI, merging the third icmp into the switch.
2642 static bool TryToSimplifyUncondBranchWithICmpInIt(
2643 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI,
2644 unsigned BonusInstThreshold, const DataLayout *DL, AssumptionTracker *AT) {
2645 BasicBlock *BB = ICI->getParent();
2647 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2649 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2651 Value *V = ICI->getOperand(0);
2652 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2654 // The pattern we're looking for is where our only predecessor is a switch on
2655 // 'V' and this block is the default case for the switch. In this case we can
2656 // fold the compared value into the switch to simplify things.
2657 BasicBlock *Pred = BB->getSinglePredecessor();
2658 if (!Pred || !isa<SwitchInst>(Pred->getTerminator())) return false;
2660 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2661 if (SI->getCondition() != V)
2664 // If BB is reachable on a non-default case, then we simply know the value of
2665 // V in this block. Substitute it and constant fold the icmp instruction
2667 if (SI->getDefaultDest() != BB) {
2668 ConstantInt *VVal = SI->findCaseDest(BB);
2669 assert(VVal && "Should have a unique destination value");
2670 ICI->setOperand(0, VVal);
2672 if (Value *V = SimplifyInstruction(ICI, DL)) {
2673 ICI->replaceAllUsesWith(V);
2674 ICI->eraseFromParent();
2676 // BB is now empty, so it is likely to simplify away.
2677 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
2680 // Ok, the block is reachable from the default dest. If the constant we're
2681 // comparing exists in one of the other edges, then we can constant fold ICI
2683 if (SI->findCaseValue(Cst) != SI->case_default()) {
2685 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2686 V = ConstantInt::getFalse(BB->getContext());
2688 V = ConstantInt::getTrue(BB->getContext());
2690 ICI->replaceAllUsesWith(V);
2691 ICI->eraseFromParent();
2692 // BB is now empty, so it is likely to simplify away.
2693 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
2696 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2698 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2699 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
2700 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
2701 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2704 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2706 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2707 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2709 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2710 std::swap(DefaultCst, NewCst);
2712 // Replace ICI (which is used by the PHI for the default value) with true or
2713 // false depending on if it is EQ or NE.
2714 ICI->replaceAllUsesWith(DefaultCst);
2715 ICI->eraseFromParent();
2717 // Okay, the switch goes to this block on a default value. Add an edge from
2718 // the switch to the merge point on the compared value.
2719 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2720 BB->getParent(), BB);
2721 SmallVector<uint64_t, 8> Weights;
2722 bool HasWeights = HasBranchWeights(SI);
2724 GetBranchWeights(SI, Weights);
2725 if (Weights.size() == 1 + SI->getNumCases()) {
2726 // Split weight for default case to case for "Cst".
2727 Weights[0] = (Weights[0]+1) >> 1;
2728 Weights.push_back(Weights[0]);
2730 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2731 SI->setMetadata(LLVMContext::MD_prof,
2732 MDBuilder(SI->getContext()).
2733 createBranchWeights(MDWeights));
2736 SI->addCase(Cst, NewBB);
2738 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2739 Builder.SetInsertPoint(NewBB);
2740 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2741 Builder.CreateBr(SuccBlock);
2742 PHIUse->addIncoming(NewCst, NewBB);
2746 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2747 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2748 /// fold it into a switch instruction if so.
2749 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *DL,
2750 IRBuilder<> &Builder) {
2751 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2752 if (!Cond) return false;
2755 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2756 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2757 // 'setne's and'ed together, collect them.
2758 Value *CompVal = nullptr;
2759 std::vector<ConstantInt*> Values;
2760 bool TrueWhenEqual = true;
2761 Value *ExtraCase = nullptr;
2762 unsigned UsedICmps = 0;
2764 if (Cond->getOpcode() == Instruction::Or) {
2765 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, DL, true,
2767 } else if (Cond->getOpcode() == Instruction::And) {
2768 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, DL, false,
2770 TrueWhenEqual = false;
2773 // If we didn't have a multiply compared value, fail.
2774 if (!CompVal) return false;
2776 // Avoid turning single icmps into a switch.
2780 // There might be duplicate constants in the list, which the switch
2781 // instruction can't handle, remove them now.
2782 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2783 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2785 // If Extra was used, we require at least two switch values to do the
2786 // transformation. A switch with one value is just an cond branch.
2787 if (ExtraCase && Values.size() < 2) return false;
2789 // TODO: Preserve branch weight metadata, similarly to how
2790 // FoldValueComparisonIntoPredecessors preserves it.
2792 // Figure out which block is which destination.
2793 BasicBlock *DefaultBB = BI->getSuccessor(1);
2794 BasicBlock *EdgeBB = BI->getSuccessor(0);
2795 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2797 BasicBlock *BB = BI->getParent();
2799 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2800 << " cases into SWITCH. BB is:\n" << *BB);
2802 // If there are any extra values that couldn't be folded into the switch
2803 // then we evaluate them with an explicit branch first. Split the block
2804 // right before the condbr to handle it.
2806 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2807 // Remove the uncond branch added to the old block.
2808 TerminatorInst *OldTI = BB->getTerminator();
2809 Builder.SetInsertPoint(OldTI);
2812 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2814 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2816 OldTI->eraseFromParent();
2818 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2819 // for the edge we just added.
2820 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2822 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2823 << "\nEXTRABB = " << *BB);
2827 Builder.SetInsertPoint(BI);
2828 // Convert pointer to int before we switch.
2829 if (CompVal->getType()->isPointerTy()) {
2830 assert(DL && "Cannot switch on pointer without DataLayout");
2831 CompVal = Builder.CreatePtrToInt(CompVal,
2832 DL->getIntPtrType(CompVal->getType()),
2836 // Create the new switch instruction now.
2837 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2839 // Add all of the 'cases' to the switch instruction.
2840 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2841 New->addCase(Values[i], EdgeBB);
2843 // We added edges from PI to the EdgeBB. As such, if there were any
2844 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2845 // the number of edges added.
2846 for (BasicBlock::iterator BBI = EdgeBB->begin();
2847 isa<PHINode>(BBI); ++BBI) {
2848 PHINode *PN = cast<PHINode>(BBI);
2849 Value *InVal = PN->getIncomingValueForBlock(BB);
2850 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2851 PN->addIncoming(InVal, BB);
2854 // Erase the old branch instruction.
2855 EraseTerminatorInstAndDCECond(BI);
2857 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2861 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2862 // If this is a trivial landing pad that just continues unwinding the caught
2863 // exception then zap the landing pad, turning its invokes into calls.
2864 BasicBlock *BB = RI->getParent();
2865 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2866 if (RI->getValue() != LPInst)
2867 // Not a landing pad, or the resume is not unwinding the exception that
2868 // caused control to branch here.
2871 // Check that there are no other instructions except for debug intrinsics.
2872 BasicBlock::iterator I = LPInst, E = RI;
2874 if (!isa<DbgInfoIntrinsic>(I))
2877 // Turn all invokes that unwind here into calls and delete the basic block.
2878 bool InvokeRequiresTableEntry = false;
2879 bool Changed = false;
2880 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2881 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2883 if (II->hasFnAttr(Attribute::UWTable)) {
2884 // Don't remove an `invoke' instruction if the ABI requires an entry into
2886 InvokeRequiresTableEntry = true;
2890 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2892 // Insert a call instruction before the invoke.
2893 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2895 Call->setCallingConv(II->getCallingConv());
2896 Call->setAttributes(II->getAttributes());
2897 Call->setDebugLoc(II->getDebugLoc());
2899 // Anything that used the value produced by the invoke instruction now uses
2900 // the value produced by the call instruction. Note that we do this even
2901 // for void functions and calls with no uses so that the callgraph edge is
2903 II->replaceAllUsesWith(Call);
2904 BB->removePredecessor(II->getParent());
2906 // Insert a branch to the normal destination right before the invoke.
2907 BranchInst::Create(II->getNormalDest(), II);
2909 // Finally, delete the invoke instruction!
2910 II->eraseFromParent();
2914 if (!InvokeRequiresTableEntry)
2915 // The landingpad is now unreachable. Zap it.
2916 BB->eraseFromParent();
2921 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2922 BasicBlock *BB = RI->getParent();
2923 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2925 // Find predecessors that end with branches.
2926 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2927 SmallVector<BranchInst*, 8> CondBranchPreds;
2928 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2929 BasicBlock *P = *PI;
2930 TerminatorInst *PTI = P->getTerminator();
2931 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2932 if (BI->isUnconditional())
2933 UncondBranchPreds.push_back(P);
2935 CondBranchPreds.push_back(BI);
2939 // If we found some, do the transformation!
2940 if (!UncondBranchPreds.empty() && DupRet) {
2941 while (!UncondBranchPreds.empty()) {
2942 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
2943 DEBUG(dbgs() << "FOLDING: " << *BB
2944 << "INTO UNCOND BRANCH PRED: " << *Pred);
2945 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
2948 // If we eliminated all predecessors of the block, delete the block now.
2949 if (pred_begin(BB) == pred_end(BB))
2950 // We know there are no successors, so just nuke the block.
2951 BB->eraseFromParent();
2956 // Check out all of the conditional branches going to this return
2957 // instruction. If any of them just select between returns, change the
2958 // branch itself into a select/return pair.
2959 while (!CondBranchPreds.empty()) {
2960 BranchInst *BI = CondBranchPreds.pop_back_val();
2962 // Check to see if the non-BB successor is also a return block.
2963 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
2964 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
2965 SimplifyCondBranchToTwoReturns(BI, Builder))
2971 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
2972 BasicBlock *BB = UI->getParent();
2974 bool Changed = false;
2976 // If there are any instructions immediately before the unreachable that can
2977 // be removed, do so.
2978 while (UI != BB->begin()) {
2979 BasicBlock::iterator BBI = UI;
2981 // Do not delete instructions that can have side effects which might cause
2982 // the unreachable to not be reachable; specifically, calls and volatile
2983 // operations may have this effect.
2984 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
2986 if (BBI->mayHaveSideEffects()) {
2987 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
2988 if (SI->isVolatile())
2990 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
2991 if (LI->isVolatile())
2993 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
2994 if (RMWI->isVolatile())
2996 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
2997 if (CXI->isVolatile())
2999 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
3000 !isa<LandingPadInst>(BBI)) {
3003 // Note that deleting LandingPad's here is in fact okay, although it
3004 // involves a bit of subtle reasoning. If this inst is a LandingPad,
3005 // all the predecessors of this block will be the unwind edges of Invokes,
3006 // and we can therefore guarantee this block will be erased.
3009 // Delete this instruction (any uses are guaranteed to be dead)
3010 if (!BBI->use_empty())
3011 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
3012 BBI->eraseFromParent();
3016 // If the unreachable instruction is the first in the block, take a gander
3017 // at all of the predecessors of this instruction, and simplify them.
3018 if (&BB->front() != UI) return Changed;
3020 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
3021 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
3022 TerminatorInst *TI = Preds[i]->getTerminator();
3023 IRBuilder<> Builder(TI);
3024 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3025 if (BI->isUnconditional()) {
3026 if (BI->getSuccessor(0) == BB) {
3027 new UnreachableInst(TI->getContext(), TI);
3028 TI->eraseFromParent();
3032 if (BI->getSuccessor(0) == BB) {
3033 Builder.CreateBr(BI->getSuccessor(1));
3034 EraseTerminatorInstAndDCECond(BI);
3035 } else if (BI->getSuccessor(1) == BB) {
3036 Builder.CreateBr(BI->getSuccessor(0));
3037 EraseTerminatorInstAndDCECond(BI);
3041 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3042 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3044 if (i.getCaseSuccessor() == BB) {
3045 BB->removePredecessor(SI->getParent());
3050 // If the default value is unreachable, figure out the most popular
3051 // destination and make it the default.
3052 if (SI->getDefaultDest() == BB) {
3053 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
3054 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3056 std::pair<unsigned, unsigned> &entry =
3057 Popularity[i.getCaseSuccessor()];
3058 if (entry.first == 0) {
3060 entry.second = i.getCaseIndex();
3066 // Find the most popular block.
3067 unsigned MaxPop = 0;
3068 unsigned MaxIndex = 0;
3069 BasicBlock *MaxBlock = nullptr;
3070 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
3071 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
3072 if (I->second.first > MaxPop ||
3073 (I->second.first == MaxPop && MaxIndex > I->second.second)) {
3074 MaxPop = I->second.first;
3075 MaxIndex = I->second.second;
3076 MaxBlock = I->first;
3080 // Make this the new default, allowing us to delete any explicit
3082 SI->setDefaultDest(MaxBlock);
3085 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
3087 if (isa<PHINode>(MaxBlock->begin()))
3088 for (unsigned i = 0; i != MaxPop-1; ++i)
3089 MaxBlock->removePredecessor(SI->getParent());
3091 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3093 if (i.getCaseSuccessor() == MaxBlock) {
3099 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
3100 if (II->getUnwindDest() == BB) {
3101 // Convert the invoke to a call instruction. This would be a good
3102 // place to note that the call does not throw though.
3103 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
3104 II->removeFromParent(); // Take out of symbol table
3106 // Insert the call now...
3107 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
3108 Builder.SetInsertPoint(BI);
3109 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
3110 Args, II->getName());
3111 CI->setCallingConv(II->getCallingConv());
3112 CI->setAttributes(II->getAttributes());
3113 // If the invoke produced a value, the call does now instead.
3114 II->replaceAllUsesWith(CI);
3121 // If this block is now dead, remove it.
3122 if (pred_begin(BB) == pred_end(BB) &&
3123 BB != &BB->getParent()->getEntryBlock()) {
3124 // We know there are no successors, so just nuke the block.
3125 BB->eraseFromParent();
3132 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
3133 /// integer range comparison into a sub, an icmp and a branch.
3134 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3135 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3137 // Make sure all cases point to the same destination and gather the values.
3138 SmallVector<ConstantInt *, 16> Cases;
3139 SwitchInst::CaseIt I = SI->case_begin();
3140 Cases.push_back(I.getCaseValue());
3141 SwitchInst::CaseIt PrevI = I++;
3142 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
3143 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
3145 Cases.push_back(I.getCaseValue());
3147 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
3149 // Sort the case values, then check if they form a range we can transform.
3150 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3151 for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
3152 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
3156 Constant *Offset = ConstantExpr::getNeg(Cases.back());
3157 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
3159 Value *Sub = SI->getCondition();
3160 if (!Offset->isNullValue())
3161 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
3163 // If NumCases overflowed, then all possible values jump to the successor.
3164 if (NumCases->isNullValue() && SI->getNumCases() != 0)
3165 Cmp = ConstantInt::getTrue(SI->getContext());
3167 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3168 BranchInst *NewBI = Builder.CreateCondBr(
3169 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
3171 // Update weight for the newly-created conditional branch.
3172 SmallVector<uint64_t, 8> Weights;
3173 bool HasWeights = HasBranchWeights(SI);
3175 GetBranchWeights(SI, Weights);
3176 if (Weights.size() == 1 + SI->getNumCases()) {
3177 // Combine all weights for the cases to be the true weight of NewBI.
3178 // We assume that the sum of all weights for a Terminator can fit into 32
3180 uint32_t NewTrueWeight = 0;
3181 for (unsigned I = 1, E = Weights.size(); I != E; ++I)
3182 NewTrueWeight += (uint32_t)Weights[I];
3183 NewBI->setMetadata(LLVMContext::MD_prof,
3184 MDBuilder(SI->getContext()).
3185 createBranchWeights(NewTrueWeight,
3186 (uint32_t)Weights[0]));
3190 // Prune obsolete incoming values off the successor's PHI nodes.
3191 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
3192 isa<PHINode>(BBI); ++BBI) {
3193 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
3194 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3196 SI->eraseFromParent();
3201 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
3202 /// and use it to remove dead cases.
3203 static bool EliminateDeadSwitchCases(SwitchInst *SI, const DataLayout *DL,
3204 AssumptionTracker *AT) {
3205 Value *Cond = SI->getCondition();
3206 unsigned Bits = Cond->getType()->getIntegerBitWidth();
3207 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3208 computeKnownBits(Cond, KnownZero, KnownOne, DL, 0, AT, SI);
3210 // Gather dead cases.
3211 SmallVector<ConstantInt*, 8> DeadCases;
3212 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3213 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3214 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3215 DeadCases.push_back(I.getCaseValue());
3216 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3217 << I.getCaseValue() << "' is dead.\n");
3221 SmallVector<uint64_t, 8> Weights;
3222 bool HasWeight = HasBranchWeights(SI);
3224 GetBranchWeights(SI, Weights);
3225 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3228 // Remove dead cases from the switch.
3229 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3230 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3231 assert(Case != SI->case_default() &&
3232 "Case was not found. Probably mistake in DeadCases forming.");
3234 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3238 // Prune unused values from PHI nodes.
3239 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3240 SI->removeCase(Case);
3242 if (HasWeight && Weights.size() >= 2) {
3243 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3244 SI->setMetadata(LLVMContext::MD_prof,
3245 MDBuilder(SI->getParent()->getContext()).
3246 createBranchWeights(MDWeights));
3249 return !DeadCases.empty();
3252 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
3253 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3254 /// by an unconditional branch), look at the phi node for BB in the successor
3255 /// block and see if the incoming value is equal to CaseValue. If so, return
3256 /// the phi node, and set PhiIndex to BB's index in the phi node.
3257 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3260 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3261 return nullptr; // BB must be empty to be a candidate for simplification.
3262 if (!BB->getSinglePredecessor())
3263 return nullptr; // BB must be dominated by the switch.
3265 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3266 if (!Branch || !Branch->isUnconditional())
3267 return nullptr; // Terminator must be unconditional branch.
3269 BasicBlock *Succ = Branch->getSuccessor(0);
3271 BasicBlock::iterator I = Succ->begin();
3272 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3273 int Idx = PHI->getBasicBlockIndex(BB);
3274 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3276 Value *InValue = PHI->getIncomingValue(Idx);
3277 if (InValue != CaseValue) continue;
3286 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
3287 /// instruction to a phi node dominated by the switch, if that would mean that
3288 /// some of the destination blocks of the switch can be folded away.
3289 /// Returns true if a change is made.
3290 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3291 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3292 ForwardingNodesMap ForwardingNodes;
3294 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3295 ConstantInt *CaseValue = I.getCaseValue();
3296 BasicBlock *CaseDest = I.getCaseSuccessor();
3299 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3303 ForwardingNodes[PHI].push_back(PhiIndex);
3306 bool Changed = false;
3308 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3309 E = ForwardingNodes.end(); I != E; ++I) {
3310 PHINode *Phi = I->first;
3311 SmallVectorImpl<int> &Indexes = I->second;
3313 if (Indexes.size() < 2) continue;
3315 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3316 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3323 /// ValidLookupTableConstant - Return true if the backend will be able to handle
3324 /// initializing an array of constants like C.
3325 static bool ValidLookupTableConstant(Constant *C) {
3326 if (C->isThreadDependent())
3328 if (C->isDLLImportDependent())
3331 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3332 return CE->isGEPWithNoNotionalOverIndexing();
3334 return isa<ConstantFP>(C) ||
3335 isa<ConstantInt>(C) ||
3336 isa<ConstantPointerNull>(C) ||
3337 isa<GlobalValue>(C) ||
3341 /// LookupConstant - If V is a Constant, return it. Otherwise, try to look up
3342 /// its constant value in ConstantPool, returning 0 if it's not there.
3343 static Constant *LookupConstant(Value *V,
3344 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3345 if (Constant *C = dyn_cast<Constant>(V))
3347 return ConstantPool.lookup(V);
3350 /// ConstantFold - Try to fold instruction I into a constant. This works for
3351 /// simple instructions such as binary operations where both operands are
3352 /// constant or can be replaced by constants from the ConstantPool. Returns the
3353 /// resulting constant on success, 0 otherwise.
3355 ConstantFold(Instruction *I,
3356 const SmallDenseMap<Value *, Constant *> &ConstantPool,
3357 const DataLayout *DL) {
3358 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3359 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3362 if (A->isAllOnesValue())
3363 return LookupConstant(Select->getTrueValue(), ConstantPool);
3364 if (A->isNullValue())
3365 return LookupConstant(Select->getFalseValue(), ConstantPool);
3369 SmallVector<Constant *, 4> COps;
3370 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
3371 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
3377 if (CmpInst *Cmp = dyn_cast<CmpInst>(I))
3378 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
3381 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), COps, DL);
3384 /// GetCaseResults - Try to determine the resulting constant values in phi nodes
3385 /// at the common destination basic block, *CommonDest, for one of the case
3386 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3387 /// case), of a switch instruction SI.
3389 GetCaseResults(SwitchInst *SI,
3390 ConstantInt *CaseVal,
3391 BasicBlock *CaseDest,
3392 BasicBlock **CommonDest,
3393 SmallVectorImpl<std::pair<PHINode *, Constant *> > &Res,
3394 const DataLayout *DL) {
3395 // The block from which we enter the common destination.
3396 BasicBlock *Pred = SI->getParent();
3398 // If CaseDest is empty except for some side-effect free instructions through
3399 // which we can constant-propagate the CaseVal, continue to its successor.
3400 SmallDenseMap<Value*, Constant*> ConstantPool;
3401 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3402 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3404 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3405 // If the terminator is a simple branch, continue to the next block.
3406 if (T->getNumSuccessors() != 1)
3409 CaseDest = T->getSuccessor(0);
3410 } else if (isa<DbgInfoIntrinsic>(I)) {
3411 // Skip debug intrinsic.
3413 } else if (Constant *C = ConstantFold(I, ConstantPool, DL)) {
3414 // Instruction is side-effect free and constant.
3415 ConstantPool.insert(std::make_pair(I, C));
3421 // If we did not have a CommonDest before, use the current one.
3423 *CommonDest = CaseDest;
3424 // If the destination isn't the common one, abort.
3425 if (CaseDest != *CommonDest)
3428 // Get the values for this case from phi nodes in the destination block.
3429 BasicBlock::iterator I = (*CommonDest)->begin();
3430 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3431 int Idx = PHI->getBasicBlockIndex(Pred);
3435 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3440 // Note: If the constant comes from constant-propagating the case value
3441 // through the CaseDest basic block, it will be safe to remove the
3442 // instructions in that block. They cannot be used (except in the phi nodes
3443 // we visit) outside CaseDest, because that block does not dominate its
3444 // successor. If it did, we would not be in this phi node.
3446 // Be conservative about which kinds of constants we support.
3447 if (!ValidLookupTableConstant(ConstVal))
3450 Res.push_back(std::make_pair(PHI, ConstVal));
3453 return Res.size() > 0;
3457 /// SwitchLookupTable - This class represents a lookup table that can be used
3458 /// to replace a switch.
3459 class SwitchLookupTable {
3461 /// SwitchLookupTable - Create a lookup table to use as a switch replacement
3462 /// with the contents of Values, using DefaultValue to fill any holes in the
3464 SwitchLookupTable(Module &M,
3466 ConstantInt *Offset,
3467 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3468 Constant *DefaultValue,
3469 const DataLayout *DL);
3471 /// BuildLookup - Build instructions with Builder to retrieve the value at
3472 /// the position given by Index in the lookup table.
3473 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3475 /// WouldFitInRegister - Return true if a table with TableSize elements of
3476 /// type ElementType would fit in a target-legal register.
3477 static bool WouldFitInRegister(const DataLayout *DL,
3479 const Type *ElementType);
3482 // Depending on the contents of the table, it can be represented in
3485 // For tables where each element contains the same value, we just have to
3486 // store that single value and return it for each lookup.
3489 // For small tables with integer elements, we can pack them into a bitmap
3490 // that fits into a target-legal register. Values are retrieved by
3491 // shift and mask operations.
3494 // The table is stored as an array of values. Values are retrieved by load
3495 // instructions from the table.
3499 // For SingleValueKind, this is the single value.
3500 Constant *SingleValue;
3502 // For BitMapKind, this is the bitmap.
3503 ConstantInt *BitMap;
3504 IntegerType *BitMapElementTy;
3506 // For ArrayKind, this is the array.
3507 GlobalVariable *Array;
3511 SwitchLookupTable::SwitchLookupTable(Module &M,
3513 ConstantInt *Offset,
3514 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3515 Constant *DefaultValue,
3516 const DataLayout *DL)
3517 : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
3519 assert(Values.size() && "Can't build lookup table without values!");
3520 assert(TableSize >= Values.size() && "Can't fit values in table!");
3522 // If all values in the table are equal, this is that value.
3523 SingleValue = Values.begin()->second;
3525 Type *ValueType = Values.begin()->second->getType();
3527 // Build up the table contents.
3528 SmallVector<Constant*, 64> TableContents(TableSize);
3529 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3530 ConstantInt *CaseVal = Values[I].first;
3531 Constant *CaseRes = Values[I].second;
3532 assert(CaseRes->getType() == ValueType);
3534 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3536 TableContents[Idx] = CaseRes;
3538 if (CaseRes != SingleValue)
3539 SingleValue = nullptr;
3542 // Fill in any holes in the table with the default result.
3543 if (Values.size() < TableSize) {
3544 assert(DefaultValue &&
3545 "Need a default value to fill the lookup table holes.");
3546 assert(DefaultValue->getType() == ValueType);
3547 for (uint64_t I = 0; I < TableSize; ++I) {
3548 if (!TableContents[I])
3549 TableContents[I] = DefaultValue;
3552 if (DefaultValue != SingleValue)
3553 SingleValue = nullptr;
3556 // If each element in the table contains the same value, we only need to store
3557 // that single value.
3559 Kind = SingleValueKind;
3563 // If the type is integer and the table fits in a register, build a bitmap.
3564 if (WouldFitInRegister(DL, TableSize, ValueType)) {
3565 IntegerType *IT = cast<IntegerType>(ValueType);
3566 APInt TableInt(TableSize * IT->getBitWidth(), 0);
3567 for (uint64_t I = TableSize; I > 0; --I) {
3568 TableInt <<= IT->getBitWidth();
3569 // Insert values into the bitmap. Undef values are set to zero.
3570 if (!isa<UndefValue>(TableContents[I - 1])) {
3571 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
3572 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
3575 BitMap = ConstantInt::get(M.getContext(), TableInt);
3576 BitMapElementTy = IT;
3582 // Store the table in an array.
3583 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
3584 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3586 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3587 GlobalVariable::PrivateLinkage,
3590 Array->setUnnamedAddr(true);
3594 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
3596 case SingleValueKind:
3599 // Type of the bitmap (e.g. i59).
3600 IntegerType *MapTy = BitMap->getType();
3602 // Cast Index to the same type as the bitmap.
3603 // Note: The Index is <= the number of elements in the table, so
3604 // truncating it to the width of the bitmask is safe.
3605 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
3607 // Multiply the shift amount by the element width.
3608 ShiftAmt = Builder.CreateMul(ShiftAmt,
3609 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
3613 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
3614 "switch.downshift");
3616 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
3620 // Make sure the table index will not overflow when treated as signed.
3621 IntegerType *IT = cast<IntegerType>(Index->getType());
3622 uint64_t TableSize = Array->getInitializer()->getType()
3623 ->getArrayNumElements();
3624 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
3625 Index = Builder.CreateZExt(Index,
3626 IntegerType::get(IT->getContext(),
3627 IT->getBitWidth() + 1),
3628 "switch.tableidx.zext");
3630 Value *GEPIndices[] = { Builder.getInt32(0), Index };
3631 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices,
3633 return Builder.CreateLoad(GEP, "switch.load");
3636 llvm_unreachable("Unknown lookup table kind!");
3639 bool SwitchLookupTable::WouldFitInRegister(const DataLayout *DL,
3641 const Type *ElementType) {
3644 const IntegerType *IT = dyn_cast<IntegerType>(ElementType);
3647 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
3648 // are <= 15, we could try to narrow the type.
3650 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
3651 if (TableSize >= UINT_MAX/IT->getBitWidth())
3653 return DL->fitsInLegalInteger(TableSize * IT->getBitWidth());
3656 /// ShouldBuildLookupTable - Determine whether a lookup table should be built
3657 /// for this switch, based on the number of cases, size of the table and the
3658 /// types of the results.
3659 static bool ShouldBuildLookupTable(SwitchInst *SI,
3661 const TargetTransformInfo &TTI,
3662 const DataLayout *DL,
3663 const SmallDenseMap<PHINode*, Type*>& ResultTypes) {
3664 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
3665 return false; // TableSize overflowed, or mul below might overflow.
3667 bool AllTablesFitInRegister = true;
3668 bool HasIllegalType = false;
3669 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(),
3670 E = ResultTypes.end(); I != E; ++I) {
3671 Type *Ty = I->second;
3673 // Saturate this flag to true.
3674 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
3676 // Saturate this flag to false.
3677 AllTablesFitInRegister = AllTablesFitInRegister &&
3678 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
3680 // If both flags saturate, we're done. NOTE: This *only* works with
3681 // saturating flags, and all flags have to saturate first due to the
3682 // non-deterministic behavior of iterating over a dense map.
3683 if (HasIllegalType && !AllTablesFitInRegister)
3687 // If each table would fit in a register, we should build it anyway.
3688 if (AllTablesFitInRegister)
3691 // Don't build a table that doesn't fit in-register if it has illegal types.
3695 // The table density should be at least 40%. This is the same criterion as for
3696 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3697 // FIXME: Find the best cut-off.
3698 return SI->getNumCases() * 10 >= TableSize * 4;
3701 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3702 /// phi nodes in a common successor block with different constant values,
3703 /// replace the switch with lookup tables.
3704 static bool SwitchToLookupTable(SwitchInst *SI,
3705 IRBuilder<> &Builder,
3706 const TargetTransformInfo &TTI,
3707 const DataLayout* DL) {
3708 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3710 // Only build lookup table when we have a target that supports it.
3711 if (!TTI.shouldBuildLookupTables())
3714 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3715 // split off a dense part and build a lookup table for that.
3717 // FIXME: This creates arrays of GEPs to constant strings, which means each
3718 // GEP needs a runtime relocation in PIC code. We should just build one big
3719 // string and lookup indices into that.
3721 // Ignore switches with less than three cases. Lookup tables will not make them
3722 // faster, so we don't analyze them.
3723 if (SI->getNumCases() < 3)
3726 // Figure out the corresponding result for each case value and phi node in the
3727 // common destination, as well as the the min and max case values.
3728 assert(SI->case_begin() != SI->case_end());
3729 SwitchInst::CaseIt CI = SI->case_begin();
3730 ConstantInt *MinCaseVal = CI.getCaseValue();
3731 ConstantInt *MaxCaseVal = CI.getCaseValue();
3733 BasicBlock *CommonDest = nullptr;
3734 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
3735 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
3736 SmallDenseMap<PHINode*, Constant*> DefaultResults;
3737 SmallDenseMap<PHINode*, Type*> ResultTypes;
3738 SmallVector<PHINode*, 4> PHIs;
3740 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
3741 ConstantInt *CaseVal = CI.getCaseValue();
3742 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
3743 MinCaseVal = CaseVal;
3744 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
3745 MaxCaseVal = CaseVal;
3747 // Resulting value at phi nodes for this case value.
3748 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
3750 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
3754 // Append the result from this case to the list for each phi.
3755 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) {
3756 if (!ResultLists.count(I->first))
3757 PHIs.push_back(I->first);
3758 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second));
3762 // Keep track of the result types.
3763 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
3764 PHINode *PHI = PHIs[I];
3765 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
3768 uint64_t NumResults = ResultLists[PHIs[0]].size();
3769 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
3770 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
3771 bool TableHasHoles = (NumResults < TableSize);
3773 // If the table has holes, we need a constant result for the default case
3774 // or a bitmask that fits in a register.
3775 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
3776 bool HasDefaultResults = false;
3777 if (TableHasHoles) {
3778 HasDefaultResults = GetCaseResults(SI, nullptr, SI->getDefaultDest(),
3779 &CommonDest, DefaultResultsList, DL);
3781 bool NeedMask = (TableHasHoles && !HasDefaultResults);
3783 // As an extra penalty for the validity test we require more cases.
3784 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
3786 if (!(DL && DL->fitsInLegalInteger(TableSize)))
3790 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) {
3791 PHINode *PHI = DefaultResultsList[I].first;
3792 Constant *Result = DefaultResultsList[I].second;
3793 DefaultResults[PHI] = Result;
3796 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
3799 // Create the BB that does the lookups.
3800 Module &Mod = *CommonDest->getParent()->getParent();
3801 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
3803 CommonDest->getParent(),
3806 // Compute the table index value.
3807 Builder.SetInsertPoint(SI);
3808 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
3811 // Compute the maximum table size representable by the integer type we are
3813 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
3814 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
3815 assert(MaxTableSize >= TableSize &&
3816 "It is impossible for a switch to have more entries than the max "
3817 "representable value of its input integer type's size.");
3819 // If we have a fully covered lookup table, unconditionally branch to the
3820 // lookup table BB. Otherwise, check if the condition value is within the case
3821 // range. If it is so, branch to the new BB. Otherwise branch to SI's default
3823 const bool GeneratingCoveredLookupTable = MaxTableSize == TableSize;
3824 if (GeneratingCoveredLookupTable) {
3825 Builder.CreateBr(LookupBB);
3826 // We cached PHINodes in PHIs, to avoid accessing deleted PHINodes later,
3827 // do not delete PHINodes here.
3828 SI->getDefaultDest()->removePredecessor(SI->getParent(),
3829 true/*DontDeleteUselessPHIs*/);
3831 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
3832 MinCaseVal->getType(), TableSize));
3833 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
3836 // Populate the BB that does the lookups.
3837 Builder.SetInsertPoint(LookupBB);
3840 // Before doing the lookup we do the hole check.
3841 // The LookupBB is therefore re-purposed to do the hole check
3842 // and we create a new LookupBB.
3843 BasicBlock *MaskBB = LookupBB;
3844 MaskBB->setName("switch.hole_check");
3845 LookupBB = BasicBlock::Create(Mod.getContext(),
3847 CommonDest->getParent(),
3850 // Build bitmask; fill in a 1 bit for every case.
3851 APInt MaskInt(TableSize, 0);
3852 APInt One(TableSize, 1);
3853 const ResultListTy &ResultList = ResultLists[PHIs[0]];
3854 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
3855 uint64_t Idx = (ResultList[I].first->getValue() -
3856 MinCaseVal->getValue()).getLimitedValue();
3857 MaskInt |= One << Idx;
3859 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
3861 // Get the TableIndex'th bit of the bitmask.
3862 // If this bit is 0 (meaning hole) jump to the default destination,
3863 // else continue with table lookup.
3864 IntegerType *MapTy = TableMask->getType();
3865 Value *MaskIndex = Builder.CreateZExtOrTrunc(TableIndex, MapTy,
3866 "switch.maskindex");
3867 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex,
3869 Value *LoBit = Builder.CreateTrunc(Shifted,
3870 Type::getInt1Ty(Mod.getContext()),
3872 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
3874 Builder.SetInsertPoint(LookupBB);
3875 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
3878 bool ReturnedEarly = false;
3879 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
3880 PHINode *PHI = PHIs[I];
3882 // If using a bitmask, use any value to fill the lookup table holes.
3883 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
3884 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI],
3887 Value *Result = Table.BuildLookup(TableIndex, Builder);
3889 // If the result is used to return immediately from the function, we want to
3890 // do that right here.
3891 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
3892 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
3893 Builder.CreateRet(Result);
3894 ReturnedEarly = true;
3898 PHI->addIncoming(Result, LookupBB);
3902 Builder.CreateBr(CommonDest);
3904 // Remove the switch.
3905 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
3906 BasicBlock *Succ = SI->getSuccessor(i);
3908 if (Succ == SI->getDefaultDest())
3910 Succ->removePredecessor(SI->getParent());
3912 SI->eraseFromParent();
3916 ++NumLookupTablesHoles;
3920 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
3921 BasicBlock *BB = SI->getParent();
3923 if (isValueEqualityComparison(SI)) {
3924 // If we only have one predecessor, and if it is a branch on this value,
3925 // see if that predecessor totally determines the outcome of this switch.
3926 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3927 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
3928 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
3930 Value *Cond = SI->getCondition();
3931 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
3932 if (SimplifySwitchOnSelect(SI, Select))
3933 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
3935 // If the block only contains the switch, see if we can fold the block
3936 // away into any preds.
3937 BasicBlock::iterator BBI = BB->begin();
3938 // Ignore dbg intrinsics.
3939 while (isa<DbgInfoIntrinsic>(BBI))
3942 if (FoldValueComparisonIntoPredecessors(SI, Builder))
3943 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
3946 // Try to transform the switch into an icmp and a branch.
3947 if (TurnSwitchRangeIntoICmp(SI, Builder))
3948 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
3950 // Remove unreachable cases.
3951 if (EliminateDeadSwitchCases(SI, DL, AT))
3952 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
3954 if (ForwardSwitchConditionToPHI(SI))
3955 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
3957 if (SwitchToLookupTable(SI, Builder, TTI, DL))
3958 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
3963 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
3964 BasicBlock *BB = IBI->getParent();
3965 bool Changed = false;
3967 // Eliminate redundant destinations.
3968 SmallPtrSet<Value *, 8> Succs;
3969 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
3970 BasicBlock *Dest = IBI->getDestination(i);
3971 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) {
3972 Dest->removePredecessor(BB);
3973 IBI->removeDestination(i);
3979 if (IBI->getNumDestinations() == 0) {
3980 // If the indirectbr has no successors, change it to unreachable.
3981 new UnreachableInst(IBI->getContext(), IBI);
3982 EraseTerminatorInstAndDCECond(IBI);
3986 if (IBI->getNumDestinations() == 1) {
3987 // If the indirectbr has one successor, change it to a direct branch.
3988 BranchInst::Create(IBI->getDestination(0), IBI);
3989 EraseTerminatorInstAndDCECond(IBI);
3993 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
3994 if (SimplifyIndirectBrOnSelect(IBI, SI))
3995 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4000 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
4001 BasicBlock *BB = BI->getParent();
4003 if (SinkCommon && SinkThenElseCodeToEnd(BI))
4006 // If the Terminator is the only non-phi instruction, simplify the block.
4007 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg();
4008 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
4009 TryToSimplifyUncondBranchFromEmptyBlock(BB))
4012 // If the only instruction in the block is a seteq/setne comparison
4013 // against a constant, try to simplify the block.
4014 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
4015 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
4016 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
4018 if (I->isTerminator() &&
4019 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI,
4020 BonusInstThreshold, DL, AT))
4024 // If this basic block is ONLY a compare and a branch, and if a predecessor
4025 // branches to us and our successor, fold the comparison into the
4026 // predecessor and use logical operations to update the incoming value
4027 // for PHI nodes in common successor.
4028 if (FoldBranchToCommonDest(BI, DL, BonusInstThreshold))
4029 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4034 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
4035 BasicBlock *BB = BI->getParent();
4037 // Conditional branch
4038 if (isValueEqualityComparison(BI)) {
4039 // If we only have one predecessor, and if it is a branch on this value,
4040 // see if that predecessor totally determines the outcome of this
4042 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4043 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
4044 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4046 // This block must be empty, except for the setcond inst, if it exists.
4047 // Ignore dbg intrinsics.
4048 BasicBlock::iterator I = BB->begin();
4049 // Ignore dbg intrinsics.
4050 while (isa<DbgInfoIntrinsic>(I))
4053 if (FoldValueComparisonIntoPredecessors(BI, Builder))
4054 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4055 } else if (&*I == cast<Instruction>(BI->getCondition())){
4057 // Ignore dbg intrinsics.
4058 while (isa<DbgInfoIntrinsic>(I))
4060 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
4061 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4065 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
4066 if (SimplifyBranchOnICmpChain(BI, DL, Builder))
4069 // If this basic block is ONLY a compare and a branch, and if a predecessor
4070 // branches to us and one of our successors, fold the comparison into the
4071 // predecessor and use logical operations to pick the right destination.
4072 if (FoldBranchToCommonDest(BI, DL, BonusInstThreshold))
4073 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4075 // We have a conditional branch to two blocks that are only reachable
4076 // from BI. We know that the condbr dominates the two blocks, so see if
4077 // there is any identical code in the "then" and "else" blocks. If so, we
4078 // can hoist it up to the branching block.
4079 if (BI->getSuccessor(0)->getSinglePredecessor()) {
4080 if (BI->getSuccessor(1)->getSinglePredecessor()) {
4081 if (HoistThenElseCodeToIf(BI, DL))
4082 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4084 // If Successor #1 has multiple preds, we may be able to conditionally
4085 // execute Successor #0 if it branches to Successor #1.
4086 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
4087 if (Succ0TI->getNumSuccessors() == 1 &&
4088 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
4089 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), DL))
4090 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4092 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
4093 // If Successor #0 has multiple preds, we may be able to conditionally
4094 // execute Successor #1 if it branches to Successor #0.
4095 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
4096 if (Succ1TI->getNumSuccessors() == 1 &&
4097 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
4098 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), DL))
4099 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4102 // If this is a branch on a phi node in the current block, thread control
4103 // through this block if any PHI node entries are constants.
4104 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
4105 if (PN->getParent() == BI->getParent())
4106 if (FoldCondBranchOnPHI(BI, DL))
4107 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4109 // Scan predecessor blocks for conditional branches.
4110 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
4111 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
4112 if (PBI != BI && PBI->isConditional())
4113 if (SimplifyCondBranchToCondBranch(PBI, BI))
4114 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4119 /// Check if passing a value to an instruction will cause undefined behavior.
4120 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
4121 Constant *C = dyn_cast<Constant>(V);
4128 if (C->isNullValue()) {
4129 // Only look at the first use, avoid hurting compile time with long uselists
4130 User *Use = *I->user_begin();
4132 // Now make sure that there are no instructions in between that can alter
4133 // control flow (eg. calls)
4134 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
4135 if (i == I->getParent()->end() || i->mayHaveSideEffects())
4138 // Look through GEPs. A load from a GEP derived from NULL is still undefined
4139 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
4140 if (GEP->getPointerOperand() == I)
4141 return passingValueIsAlwaysUndefined(V, GEP);
4143 // Look through bitcasts.
4144 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
4145 return passingValueIsAlwaysUndefined(V, BC);
4147 // Load from null is undefined.
4148 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
4149 if (!LI->isVolatile())
4150 return LI->getPointerAddressSpace() == 0;
4152 // Store to null is undefined.
4153 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
4154 if (!SI->isVolatile())
4155 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
4160 /// If BB has an incoming value that will always trigger undefined behavior
4161 /// (eg. null pointer dereference), remove the branch leading here.
4162 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
4163 for (BasicBlock::iterator i = BB->begin();
4164 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
4165 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
4166 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
4167 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
4168 IRBuilder<> Builder(T);
4169 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
4170 BB->removePredecessor(PHI->getIncomingBlock(i));
4171 // Turn uncoditional branches into unreachables and remove the dead
4172 // destination from conditional branches.
4173 if (BI->isUnconditional())
4174 Builder.CreateUnreachable();
4176 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
4177 BI->getSuccessor(0));
4178 BI->eraseFromParent();
4181 // TODO: SwitchInst.
4187 bool SimplifyCFGOpt::run(BasicBlock *BB) {
4188 bool Changed = false;
4190 assert(BB && BB->getParent() && "Block not embedded in function!");
4191 assert(BB->getTerminator() && "Degenerate basic block encountered!");
4193 // Remove basic blocks that have no predecessors (except the entry block)...
4194 // or that just have themself as a predecessor. These are unreachable.
4195 if ((pred_begin(BB) == pred_end(BB) &&
4196 BB != &BB->getParent()->getEntryBlock()) ||
4197 BB->getSinglePredecessor() == BB) {
4198 DEBUG(dbgs() << "Removing BB: \n" << *BB);
4199 DeleteDeadBlock(BB);
4203 // Check to see if we can constant propagate this terminator instruction
4205 Changed |= ConstantFoldTerminator(BB, true);
4207 // Check for and eliminate duplicate PHI nodes in this block.
4208 Changed |= EliminateDuplicatePHINodes(BB);
4210 // Check for and remove branches that will always cause undefined behavior.
4211 Changed |= removeUndefIntroducingPredecessor(BB);
4213 // Merge basic blocks into their predecessor if there is only one distinct
4214 // pred, and if there is only one distinct successor of the predecessor, and
4215 // if there are no PHI nodes.
4217 if (MergeBlockIntoPredecessor(BB))
4220 IRBuilder<> Builder(BB);
4222 // If there is a trivial two-entry PHI node in this basic block, and we can
4223 // eliminate it, do so now.
4224 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
4225 if (PN->getNumIncomingValues() == 2)
4226 Changed |= FoldTwoEntryPHINode(PN, DL);
4228 Builder.SetInsertPoint(BB->getTerminator());
4229 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
4230 if (BI->isUnconditional()) {
4231 if (SimplifyUncondBranch(BI, Builder)) return true;
4233 if (SimplifyCondBranch(BI, Builder)) return true;
4235 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
4236 if (SimplifyReturn(RI, Builder)) return true;
4237 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
4238 if (SimplifyResume(RI, Builder)) return true;
4239 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
4240 if (SimplifySwitch(SI, Builder)) return true;
4241 } else if (UnreachableInst *UI =
4242 dyn_cast<UnreachableInst>(BB->getTerminator())) {
4243 if (SimplifyUnreachable(UI)) return true;
4244 } else if (IndirectBrInst *IBI =
4245 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
4246 if (SimplifyIndirectBr(IBI)) return true;
4252 /// SimplifyCFG - This function is used to do simplification of a CFG. For
4253 /// example, it adjusts branches to branches to eliminate the extra hop, it
4254 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4255 /// of the CFG. It returns true if a modification was made.
4257 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
4258 unsigned BonusInstThreshold,
4259 const DataLayout *DL, AssumptionTracker *AT) {
4260 return SimplifyCFGOpt(TTI, BonusInstThreshold, DL, AT).run(BB);