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 #define DEBUG_TYPE "simplifycfg"
15 #include "llvm/Transforms/Utils/Local.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SetVector.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/Analysis/ConstantFolding.h"
23 #include "llvm/Analysis/InstructionSimplify.h"
24 #include "llvm/Analysis/TargetTransformInfo.h"
25 #include "llvm/Analysis/ValueTracking.h"
26 #include "llvm/IR/CFG.h"
27 #include "llvm/IR/ConstantRange.h"
28 #include "llvm/IR/Constants.h"
29 #include "llvm/IR/DataLayout.h"
30 #include "llvm/IR/DerivedTypes.h"
31 #include "llvm/IR/GlobalVariable.h"
32 #include "llvm/IR/IRBuilder.h"
33 #include "llvm/IR/Instructions.h"
34 #include "llvm/IR/IntrinsicInst.h"
35 #include "llvm/IR/LLVMContext.h"
36 #include "llvm/IR/MDBuilder.h"
37 #include "llvm/IR/Metadata.h"
38 #include "llvm/IR/Module.h"
39 #include "llvm/IR/NoFolder.h"
40 #include "llvm/IR/Operator.h"
41 #include "llvm/IR/PatternMatch.h"
42 #include "llvm/IR/Type.h"
43 #include "llvm/Support/CommandLine.h"
44 #include "llvm/Support/Debug.h"
45 #include "llvm/Support/raw_ostream.h"
46 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
51 using namespace PatternMatch;
53 static cl::opt<unsigned>
54 PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1),
55 cl::desc("Control the amount of phi node folding to perform (default = 1)"));
58 DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false),
59 cl::desc("Duplicate return instructions into unconditional branches"));
62 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
63 cl::desc("Sink common instructions down to the end block"));
65 static cl::opt<bool> HoistCondStores(
66 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
67 cl::desc("Hoist conditional stores if an unconditional store precedes"));
69 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
70 STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables");
71 STATISTIC(NumLookupTablesHoles, "Number of switch instructions turned into lookup tables (holes checked)");
72 STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block");
73 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
76 /// ValueEqualityComparisonCase - Represents a case of a switch.
77 struct ValueEqualityComparisonCase {
81 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
82 : Value(Value), Dest(Dest) {}
84 bool operator<(ValueEqualityComparisonCase RHS) const {
85 // Comparing pointers is ok as we only rely on the order for uniquing.
86 return Value < RHS.Value;
89 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
92 class SimplifyCFGOpt {
93 const TargetTransformInfo &TTI;
94 const DataLayout *const DL;
95 Value *isValueEqualityComparison(TerminatorInst *TI);
96 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
97 std::vector<ValueEqualityComparisonCase> &Cases);
98 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
100 IRBuilder<> &Builder);
101 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
102 IRBuilder<> &Builder);
104 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
105 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
106 bool SimplifyUnreachable(UnreachableInst *UI);
107 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
108 bool SimplifyIndirectBr(IndirectBrInst *IBI);
109 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder);
110 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder);
113 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout *DL)
114 : TTI(TTI), DL(DL) {}
115 bool run(BasicBlock *BB);
119 /// SafeToMergeTerminators - Return true if it is safe to merge these two
120 /// terminator instructions together.
122 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
123 if (SI1 == SI2) return false; // Can't merge with self!
125 // It is not safe to merge these two switch instructions if they have a common
126 // successor, and if that successor has a PHI node, and if *that* PHI node has
127 // conflicting incoming values from the two switch blocks.
128 BasicBlock *SI1BB = SI1->getParent();
129 BasicBlock *SI2BB = SI2->getParent();
130 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
132 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
133 if (SI1Succs.count(*I))
134 for (BasicBlock::iterator BBI = (*I)->begin();
135 isa<PHINode>(BBI); ++BBI) {
136 PHINode *PN = cast<PHINode>(BBI);
137 if (PN->getIncomingValueForBlock(SI1BB) !=
138 PN->getIncomingValueForBlock(SI2BB))
145 /// isProfitableToFoldUnconditional - Return true if it is safe and profitable
146 /// to merge these two terminator instructions together, where SI1 is an
147 /// unconditional branch. PhiNodes will store all PHI nodes in common
150 static bool isProfitableToFoldUnconditional(BranchInst *SI1,
153 SmallVectorImpl<PHINode*> &PhiNodes) {
154 if (SI1 == SI2) return false; // Can't merge with self!
155 assert(SI1->isUnconditional() && SI2->isConditional());
157 // We fold the unconditional branch if we can easily update all PHI nodes in
158 // common successors:
159 // 1> We have a constant incoming value for the conditional branch;
160 // 2> We have "Cond" as the incoming value for the unconditional branch;
161 // 3> SI2->getCondition() and Cond have same operands.
162 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
163 if (!Ci2) return false;
164 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
165 Cond->getOperand(1) == Ci2->getOperand(1)) &&
166 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
167 Cond->getOperand(1) == Ci2->getOperand(0)))
170 BasicBlock *SI1BB = SI1->getParent();
171 BasicBlock *SI2BB = SI2->getParent();
172 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
173 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
174 if (SI1Succs.count(*I))
175 for (BasicBlock::iterator BBI = (*I)->begin();
176 isa<PHINode>(BBI); ++BBI) {
177 PHINode *PN = cast<PHINode>(BBI);
178 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
179 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
181 PhiNodes.push_back(PN);
186 /// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will
187 /// now be entries in it from the 'NewPred' block. The values that will be
188 /// flowing into the PHI nodes will be the same as those coming in from
189 /// ExistPred, an existing predecessor of Succ.
190 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
191 BasicBlock *ExistPred) {
192 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
195 for (BasicBlock::iterator I = Succ->begin();
196 (PN = dyn_cast<PHINode>(I)); ++I)
197 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
200 /// ComputeSpeculationCost - Compute an abstract "cost" of speculating the
201 /// given instruction, which is assumed to be safe to speculate. 1 means
202 /// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive.
203 static unsigned ComputeSpeculationCost(const User *I) {
204 assert(isSafeToSpeculativelyExecute(I) &&
205 "Instruction is not safe to speculatively execute!");
206 switch (Operator::getOpcode(I)) {
208 // In doubt, be conservative.
210 case Instruction::GetElementPtr:
211 // GEPs are cheap if all indices are constant.
212 if (!cast<GEPOperator>(I)->hasAllConstantIndices())
215 case Instruction::Load:
216 case Instruction::Add:
217 case Instruction::Sub:
218 case Instruction::And:
219 case Instruction::Or:
220 case Instruction::Xor:
221 case Instruction::Shl:
222 case Instruction::LShr:
223 case Instruction::AShr:
224 case Instruction::ICmp:
225 case Instruction::Trunc:
226 case Instruction::ZExt:
227 case Instruction::SExt:
228 return 1; // These are all cheap.
230 case Instruction::Call:
231 case Instruction::Select:
236 /// DominatesMergePoint - If we have a merge point of an "if condition" as
237 /// accepted above, return true if the specified value dominates the block. We
238 /// don't handle the true generality of domination here, just a special case
239 /// which works well enough for us.
241 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
242 /// see if V (which must be an instruction) and its recursive operands
243 /// that do not dominate BB have a combined cost lower than CostRemaining and
244 /// are non-trapping. If both are true, the instruction is inserted into the
245 /// set and true is returned.
247 /// The cost for most non-trapping instructions is defined as 1 except for
248 /// Select whose cost is 2.
250 /// After this function returns, CostRemaining is decreased by the cost of
251 /// V plus its non-dominating operands. If that cost is greater than
252 /// CostRemaining, false is returned and CostRemaining is undefined.
253 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
254 SmallPtrSet<Instruction*, 4> *AggressiveInsts,
255 unsigned &CostRemaining) {
256 Instruction *I = dyn_cast<Instruction>(V);
258 // Non-instructions all dominate instructions, but not all constantexprs
259 // can be executed unconditionally.
260 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
265 BasicBlock *PBB = I->getParent();
267 // We don't want to allow weird loops that might have the "if condition" in
268 // the bottom of this block.
269 if (PBB == BB) return false;
271 // If this instruction is defined in a block that contains an unconditional
272 // branch to BB, then it must be in the 'conditional' part of the "if
273 // statement". If not, it definitely dominates the region.
274 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
275 if (BI == 0 || BI->isConditional() || BI->getSuccessor(0) != BB)
278 // If we aren't allowing aggressive promotion anymore, then don't consider
279 // instructions in the 'if region'.
280 if (AggressiveInsts == 0) return false;
282 // If we have seen this instruction before, don't count it again.
283 if (AggressiveInsts->count(I)) return true;
285 // Okay, it looks like the instruction IS in the "condition". Check to
286 // see if it's a cheap instruction to unconditionally compute, and if it
287 // only uses stuff defined outside of the condition. If so, hoist it out.
288 if (!isSafeToSpeculativelyExecute(I))
291 unsigned Cost = ComputeSpeculationCost(I);
293 if (Cost > CostRemaining)
296 CostRemaining -= Cost;
298 // Okay, we can only really hoist these out if their operands do
299 // not take us over the cost threshold.
300 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
301 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining))
303 // Okay, it's safe to do this! Remember this instruction.
304 AggressiveInsts->insert(I);
308 /// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr
309 /// and PointerNullValue. Return NULL if value is not a constant int.
310 static ConstantInt *GetConstantInt(Value *V, const DataLayout *DL) {
311 // Normal constant int.
312 ConstantInt *CI = dyn_cast<ConstantInt>(V);
313 if (CI || !DL || !isa<Constant>(V) || !V->getType()->isPointerTy())
316 // This is some kind of pointer constant. Turn it into a pointer-sized
317 // ConstantInt if possible.
318 IntegerType *PtrTy = cast<IntegerType>(DL->getIntPtrType(V->getType()));
320 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
321 if (isa<ConstantPointerNull>(V))
322 return ConstantInt::get(PtrTy, 0);
324 // IntToPtr const int.
325 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
326 if (CE->getOpcode() == Instruction::IntToPtr)
327 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
328 // The constant is very likely to have the right type already.
329 if (CI->getType() == PtrTy)
332 return cast<ConstantInt>
333 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
338 /// GatherConstantCompares - Given a potentially 'or'd or 'and'd together
339 /// collection of icmp eq/ne instructions that compare a value against a
340 /// constant, return the value being compared, and stick the constant into the
343 GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra,
344 const DataLayout *DL, bool isEQ, unsigned &UsedICmps) {
345 Instruction *I = dyn_cast<Instruction>(V);
346 if (I == 0) return 0;
348 // If this is an icmp against a constant, handle this as one of the cases.
349 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) {
350 if (ConstantInt *C = GetConstantInt(I->getOperand(1), DL)) {
354 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
355 // (x & ~2^x) == y --> x == y || x == y|2^x
356 // This undoes a transformation done by instcombine to fuse 2 compares.
357 if (match(ICI->getOperand(0),
358 m_And(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
359 APInt Not = ~RHSC->getValue();
360 if (Not.isPowerOf2()) {
363 ConstantInt::get(C->getContext(), C->getValue() | Not));
371 return I->getOperand(0);
374 // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to
377 ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue());
379 // Shift the range if the compare is fed by an add. This is the range
380 // compare idiom as emitted by instcombine.
382 match(I->getOperand(0), m_Add(m_Value(RHSVal), m_ConstantInt(RHSC)));
384 Span = Span.subtract(RHSC->getValue());
386 // If this is an and/!= check then we want to optimize "x ugt 2" into
389 Span = Span.inverse();
391 // If there are a ton of values, we don't want to make a ginormous switch.
392 if (Span.getSetSize().ugt(8) || Span.isEmptySet())
395 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
396 Vals.push_back(ConstantInt::get(V->getContext(), Tmp));
398 return hasAdd ? RHSVal : I->getOperand(0);
403 // Otherwise, we can only handle an | or &, depending on isEQ.
404 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And))
407 unsigned NumValsBeforeLHS = Vals.size();
408 unsigned UsedICmpsBeforeLHS = UsedICmps;
409 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, DL,
411 unsigned NumVals = Vals.size();
412 unsigned UsedICmpsBeforeRHS = UsedICmps;
413 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, DL,
417 Vals.resize(NumVals);
418 UsedICmps = UsedICmpsBeforeRHS;
421 // The RHS of the or/and can't be folded in and we haven't used "Extra" yet,
422 // set it and return success.
423 if (Extra == 0 || Extra == I->getOperand(1)) {
424 Extra = I->getOperand(1);
428 Vals.resize(NumValsBeforeLHS);
429 UsedICmps = UsedICmpsBeforeLHS;
433 // If the LHS can't be folded in, but Extra is available and RHS can, try to
435 if (Extra == 0 || Extra == I->getOperand(0)) {
436 Value *OldExtra = Extra;
437 Extra = I->getOperand(0);
438 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, DL,
441 assert(Vals.size() == NumValsBeforeLHS);
448 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
449 Instruction *Cond = 0;
450 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
451 Cond = dyn_cast<Instruction>(SI->getCondition());
452 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
453 if (BI->isConditional())
454 Cond = dyn_cast<Instruction>(BI->getCondition());
455 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
456 Cond = dyn_cast<Instruction>(IBI->getAddress());
459 TI->eraseFromParent();
460 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
463 /// isValueEqualityComparison - Return true if the specified terminator checks
464 /// to see if a value is equal to constant integer value.
465 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
467 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
468 // Do not permit merging of large switch instructions into their
469 // predecessors unless there is only one predecessor.
470 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()),
471 pred_end(SI->getParent())) <= 128)
472 CV = SI->getCondition();
473 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
474 if (BI->isConditional() && BI->getCondition()->hasOneUse())
475 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition()))
476 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
477 CV = ICI->getOperand(0);
479 // Unwrap any lossless ptrtoint cast.
481 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
482 Value *Ptr = PTII->getPointerOperand();
483 if (PTII->getType() == DL->getIntPtrType(Ptr->getType()))
490 /// GetValueEqualityComparisonCases - Given a value comparison instruction,
491 /// decode all of the 'cases' that it represents and return the 'default' block.
492 BasicBlock *SimplifyCFGOpt::
493 GetValueEqualityComparisonCases(TerminatorInst *TI,
494 std::vector<ValueEqualityComparisonCase>
496 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
497 Cases.reserve(SI->getNumCases());
498 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
499 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
500 i.getCaseSuccessor()));
501 return SI->getDefaultDest();
504 BranchInst *BI = cast<BranchInst>(TI);
505 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
506 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
507 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
510 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
514 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries
515 /// in the list that match the specified block.
516 static void EliminateBlockCases(BasicBlock *BB,
517 std::vector<ValueEqualityComparisonCase> &Cases) {
518 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
521 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as
524 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
525 std::vector<ValueEqualityComparisonCase > &C2) {
526 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
528 // Make V1 be smaller than V2.
529 if (V1->size() > V2->size())
532 if (V1->size() == 0) return false;
533 if (V1->size() == 1) {
535 ConstantInt *TheVal = (*V1)[0].Value;
536 for (unsigned i = 0, e = V2->size(); i != e; ++i)
537 if (TheVal == (*V2)[i].Value)
541 // Otherwise, just sort both lists and compare element by element.
542 array_pod_sort(V1->begin(), V1->end());
543 array_pod_sort(V2->begin(), V2->end());
544 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
545 while (i1 != e1 && i2 != e2) {
546 if ((*V1)[i1].Value == (*V2)[i2].Value)
548 if ((*V1)[i1].Value < (*V2)[i2].Value)
556 /// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a
557 /// terminator instruction and its block is known to only have a single
558 /// predecessor block, check to see if that predecessor is also a value
559 /// comparison with the same value, and if that comparison determines the
560 /// outcome of this comparison. If so, simplify TI. This does a very limited
561 /// form of jump threading.
562 bool SimplifyCFGOpt::
563 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
565 IRBuilder<> &Builder) {
566 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
567 if (!PredVal) return false; // Not a value comparison in predecessor.
569 Value *ThisVal = isValueEqualityComparison(TI);
570 assert(ThisVal && "This isn't a value comparison!!");
571 if (ThisVal != PredVal) return false; // Different predicates.
573 // TODO: Preserve branch weight metadata, similarly to how
574 // FoldValueComparisonIntoPredecessors preserves it.
576 // Find out information about when control will move from Pred to TI's block.
577 std::vector<ValueEqualityComparisonCase> PredCases;
578 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
580 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
582 // Find information about how control leaves this block.
583 std::vector<ValueEqualityComparisonCase> ThisCases;
584 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
585 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
587 // If TI's block is the default block from Pred's comparison, potentially
588 // simplify TI based on this knowledge.
589 if (PredDef == TI->getParent()) {
590 // If we are here, we know that the value is none of those cases listed in
591 // PredCases. If there are any cases in ThisCases that are in PredCases, we
593 if (!ValuesOverlap(PredCases, ThisCases))
596 if (isa<BranchInst>(TI)) {
597 // Okay, one of the successors of this condbr is dead. Convert it to a
599 assert(ThisCases.size() == 1 && "Branch can only have one case!");
600 // Insert the new branch.
601 Instruction *NI = Builder.CreateBr(ThisDef);
604 // Remove PHI node entries for the dead edge.
605 ThisCases[0].Dest->removePredecessor(TI->getParent());
607 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
608 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
610 EraseTerminatorInstAndDCECond(TI);
614 SwitchInst *SI = cast<SwitchInst>(TI);
615 // Okay, TI has cases that are statically dead, prune them away.
616 SmallPtrSet<Constant*, 16> DeadCases;
617 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
618 DeadCases.insert(PredCases[i].Value);
620 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
621 << "Through successor TI: " << *TI);
623 // Collect branch weights into a vector.
624 SmallVector<uint32_t, 8> Weights;
625 MDNode* MD = SI->getMetadata(LLVMContext::MD_prof);
626 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
628 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
630 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i));
632 Weights.push_back(CI->getValue().getZExtValue());
634 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
636 if (DeadCases.count(i.getCaseValue())) {
638 std::swap(Weights[i.getCaseIndex()+1], Weights.back());
641 i.getCaseSuccessor()->removePredecessor(TI->getParent());
645 if (HasWeight && Weights.size() >= 2)
646 SI->setMetadata(LLVMContext::MD_prof,
647 MDBuilder(SI->getParent()->getContext()).
648 createBranchWeights(Weights));
650 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
654 // Otherwise, TI's block must correspond to some matched value. Find out
655 // which value (or set of values) this is.
656 ConstantInt *TIV = 0;
657 BasicBlock *TIBB = TI->getParent();
658 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
659 if (PredCases[i].Dest == TIBB) {
661 return false; // Cannot handle multiple values coming to this block.
662 TIV = PredCases[i].Value;
664 assert(TIV && "No edge from pred to succ?");
666 // Okay, we found the one constant that our value can be if we get into TI's
667 // BB. Find out which successor will unconditionally be branched to.
668 BasicBlock *TheRealDest = 0;
669 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
670 if (ThisCases[i].Value == TIV) {
671 TheRealDest = ThisCases[i].Dest;
675 // If not handled by any explicit cases, it is handled by the default case.
676 if (TheRealDest == 0) TheRealDest = ThisDef;
678 // Remove PHI node entries for dead edges.
679 BasicBlock *CheckEdge = TheRealDest;
680 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
681 if (*SI != CheckEdge)
682 (*SI)->removePredecessor(TIBB);
686 // Insert the new branch.
687 Instruction *NI = Builder.CreateBr(TheRealDest);
690 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
691 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
693 EraseTerminatorInstAndDCECond(TI);
698 /// ConstantIntOrdering - This class implements a stable ordering of constant
699 /// integers that does not depend on their address. This is important for
700 /// applications that sort ConstantInt's to ensure uniqueness.
701 struct ConstantIntOrdering {
702 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
703 return LHS->getValue().ult(RHS->getValue());
708 static int ConstantIntSortPredicate(ConstantInt *const *P1,
709 ConstantInt *const *P2) {
710 const ConstantInt *LHS = *P1;
711 const ConstantInt *RHS = *P2;
712 if (LHS->getValue().ult(RHS->getValue()))
714 if (LHS->getValue() == RHS->getValue())
719 static inline bool HasBranchWeights(const Instruction* I) {
720 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof);
721 if (ProfMD && ProfMD->getOperand(0))
722 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
723 return MDS->getString().equals("branch_weights");
728 /// Get Weights of a given TerminatorInst, the default weight is at the front
729 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
731 static void GetBranchWeights(TerminatorInst *TI,
732 SmallVectorImpl<uint64_t> &Weights) {
733 MDNode* MD = TI->getMetadata(LLVMContext::MD_prof);
735 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
736 ConstantInt *CI = cast<ConstantInt>(MD->getOperand(i));
737 Weights.push_back(CI->getValue().getZExtValue());
740 // If TI is a conditional eq, the default case is the false case,
741 // and the corresponding branch-weight data is at index 2. We swap the
742 // default weight to be the first entry.
743 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
744 assert(Weights.size() == 2);
745 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
746 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
747 std::swap(Weights.front(), Weights.back());
751 /// Keep halving the weights until all can fit in uint32_t.
752 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
753 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
754 if (Max > UINT_MAX) {
755 unsigned Offset = 32 - countLeadingZeros(Max);
756 for (uint64_t &I : Weights)
761 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value
762 /// equality comparison instruction (either a switch or a branch on "X == c").
763 /// See if any of the predecessors of the terminator block are value comparisons
764 /// on the same value. If so, and if safe to do so, fold them together.
765 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
766 IRBuilder<> &Builder) {
767 BasicBlock *BB = TI->getParent();
768 Value *CV = isValueEqualityComparison(TI); // CondVal
769 assert(CV && "Not a comparison?");
770 bool Changed = false;
772 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
773 while (!Preds.empty()) {
774 BasicBlock *Pred = Preds.pop_back_val();
776 // See if the predecessor is a comparison with the same value.
777 TerminatorInst *PTI = Pred->getTerminator();
778 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
780 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
781 // Figure out which 'cases' to copy from SI to PSI.
782 std::vector<ValueEqualityComparisonCase> BBCases;
783 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
785 std::vector<ValueEqualityComparisonCase> PredCases;
786 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
788 // Based on whether the default edge from PTI goes to BB or not, fill in
789 // PredCases and PredDefault with the new switch cases we would like to
791 SmallVector<BasicBlock*, 8> NewSuccessors;
793 // Update the branch weight metadata along the way
794 SmallVector<uint64_t, 8> Weights;
795 bool PredHasWeights = HasBranchWeights(PTI);
796 bool SuccHasWeights = HasBranchWeights(TI);
798 if (PredHasWeights) {
799 GetBranchWeights(PTI, Weights);
800 // branch-weight metadata is inconsistent here.
801 if (Weights.size() != 1 + PredCases.size())
802 PredHasWeights = SuccHasWeights = false;
803 } else if (SuccHasWeights)
804 // If there are no predecessor weights but there are successor weights,
805 // populate Weights with 1, which will later be scaled to the sum of
806 // successor's weights
807 Weights.assign(1 + PredCases.size(), 1);
809 SmallVector<uint64_t, 8> SuccWeights;
810 if (SuccHasWeights) {
811 GetBranchWeights(TI, SuccWeights);
812 // branch-weight metadata is inconsistent here.
813 if (SuccWeights.size() != 1 + BBCases.size())
814 PredHasWeights = SuccHasWeights = false;
815 } else if (PredHasWeights)
816 SuccWeights.assign(1 + BBCases.size(), 1);
818 if (PredDefault == BB) {
819 // If this is the default destination from PTI, only the edges in TI
820 // that don't occur in PTI, or that branch to BB will be activated.
821 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
822 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
823 if (PredCases[i].Dest != BB)
824 PTIHandled.insert(PredCases[i].Value);
826 // The default destination is BB, we don't need explicit targets.
827 std::swap(PredCases[i], PredCases.back());
829 if (PredHasWeights || SuccHasWeights) {
830 // Increase weight for the default case.
831 Weights[0] += Weights[i+1];
832 std::swap(Weights[i+1], Weights.back());
836 PredCases.pop_back();
840 // Reconstruct the new switch statement we will be building.
841 if (PredDefault != BBDefault) {
842 PredDefault->removePredecessor(Pred);
843 PredDefault = BBDefault;
844 NewSuccessors.push_back(BBDefault);
847 unsigned CasesFromPred = Weights.size();
848 uint64_t ValidTotalSuccWeight = 0;
849 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
850 if (!PTIHandled.count(BBCases[i].Value) &&
851 BBCases[i].Dest != BBDefault) {
852 PredCases.push_back(BBCases[i]);
853 NewSuccessors.push_back(BBCases[i].Dest);
854 if (SuccHasWeights || PredHasWeights) {
855 // The default weight is at index 0, so weight for the ith case
856 // should be at index i+1. Scale the cases from successor by
857 // PredDefaultWeight (Weights[0]).
858 Weights.push_back(Weights[0] * SuccWeights[i+1]);
859 ValidTotalSuccWeight += SuccWeights[i+1];
863 if (SuccHasWeights || PredHasWeights) {
864 ValidTotalSuccWeight += SuccWeights[0];
865 // Scale the cases from predecessor by ValidTotalSuccWeight.
866 for (unsigned i = 1; i < CasesFromPred; ++i)
867 Weights[i] *= ValidTotalSuccWeight;
868 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
869 Weights[0] *= SuccWeights[0];
872 // If this is not the default destination from PSI, only the edges
873 // in SI that occur in PSI with a destination of BB will be
875 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
876 std::map<ConstantInt*, uint64_t> WeightsForHandled;
877 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
878 if (PredCases[i].Dest == BB) {
879 PTIHandled.insert(PredCases[i].Value);
881 if (PredHasWeights || SuccHasWeights) {
882 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
883 std::swap(Weights[i+1], Weights.back());
887 std::swap(PredCases[i], PredCases.back());
888 PredCases.pop_back();
892 // Okay, now we know which constants were sent to BB from the
893 // predecessor. Figure out where they will all go now.
894 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
895 if (PTIHandled.count(BBCases[i].Value)) {
896 // If this is one we are capable of getting...
897 if (PredHasWeights || SuccHasWeights)
898 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
899 PredCases.push_back(BBCases[i]);
900 NewSuccessors.push_back(BBCases[i].Dest);
901 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
904 // If there are any constants vectored to BB that TI doesn't handle,
905 // they must go to the default destination of TI.
906 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
908 E = PTIHandled.end(); I != E; ++I) {
909 if (PredHasWeights || SuccHasWeights)
910 Weights.push_back(WeightsForHandled[*I]);
911 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
912 NewSuccessors.push_back(BBDefault);
916 // Okay, at this point, we know which new successor Pred will get. Make
917 // sure we update the number of entries in the PHI nodes for these
919 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i)
920 AddPredecessorToBlock(NewSuccessors[i], Pred, BB);
922 Builder.SetInsertPoint(PTI);
923 // Convert pointer to int before we switch.
924 if (CV->getType()->isPointerTy()) {
925 assert(DL && "Cannot switch on pointer without DataLayout");
926 CV = Builder.CreatePtrToInt(CV, DL->getIntPtrType(CV->getType()),
930 // Now that the successors are updated, create the new Switch instruction.
931 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
933 NewSI->setDebugLoc(PTI->getDebugLoc());
934 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
935 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest);
937 if (PredHasWeights || SuccHasWeights) {
938 // Halve the weights if any of them cannot fit in an uint32_t
941 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
943 NewSI->setMetadata(LLVMContext::MD_prof,
944 MDBuilder(BB->getContext()).
945 createBranchWeights(MDWeights));
948 EraseTerminatorInstAndDCECond(PTI);
950 // Okay, last check. If BB is still a successor of PSI, then we must
951 // have an infinite loop case. If so, add an infinitely looping block
952 // to handle the case to preserve the behavior of the code.
953 BasicBlock *InfLoopBlock = 0;
954 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
955 if (NewSI->getSuccessor(i) == BB) {
956 if (InfLoopBlock == 0) {
957 // Insert it at the end of the function, because it's either code,
958 // or it won't matter if it's hot. :)
959 InfLoopBlock = BasicBlock::Create(BB->getContext(),
960 "infloop", BB->getParent());
961 BranchInst::Create(InfLoopBlock, InfLoopBlock);
963 NewSI->setSuccessor(i, InfLoopBlock);
972 // isSafeToHoistInvoke - If we would need to insert a select that uses the
973 // value of this invoke (comments in HoistThenElseCodeToIf explain why we
974 // would need to do this), we can't hoist the invoke, as there is nowhere
975 // to put the select in this case.
976 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
977 Instruction *I1, Instruction *I2) {
978 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
980 for (BasicBlock::iterator BBI = SI->begin();
981 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
982 Value *BB1V = PN->getIncomingValueForBlock(BB1);
983 Value *BB2V = PN->getIncomingValueForBlock(BB2);
984 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
992 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
993 /// BB2, hoist any common code in the two blocks up into the branch block. The
994 /// caller of this function guarantees that BI's block dominates BB1 and BB2.
995 static bool HoistThenElseCodeToIf(BranchInst *BI) {
996 // This does very trivial matching, with limited scanning, to find identical
997 // instructions in the two blocks. In particular, we don't want to get into
998 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
999 // such, we currently just scan for obviously identical instructions in an
1001 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1002 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1004 BasicBlock::iterator BB1_Itr = BB1->begin();
1005 BasicBlock::iterator BB2_Itr = BB2->begin();
1007 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++;
1008 // Skip debug info if it is not identical.
1009 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1010 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1011 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1012 while (isa<DbgInfoIntrinsic>(I1))
1014 while (isa<DbgInfoIntrinsic>(I2))
1017 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1018 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1021 BasicBlock *BIParent = BI->getParent();
1023 bool Changed = false;
1025 // If we are hoisting the terminator instruction, don't move one (making a
1026 // broken BB), instead clone it, and remove BI.
1027 if (isa<TerminatorInst>(I1))
1028 goto HoistTerminator;
1030 // For a normal instruction, we just move one to right before the branch,
1031 // then replace all uses of the other with the first. Finally, we remove
1032 // the now redundant second instruction.
1033 BIParent->getInstList().splice(BI, BB1->getInstList(), I1);
1034 if (!I2->use_empty())
1035 I2->replaceAllUsesWith(I1);
1036 I1->intersectOptionalDataWith(I2);
1037 I2->eraseFromParent();
1042 // Skip debug info if it is not identical.
1043 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1044 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1045 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1046 while (isa<DbgInfoIntrinsic>(I1))
1048 while (isa<DbgInfoIntrinsic>(I2))
1051 } while (I1->isIdenticalToWhenDefined(I2));
1056 // It may not be possible to hoist an invoke.
1057 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1060 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1062 for (BasicBlock::iterator BBI = SI->begin();
1063 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1064 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1065 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1069 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1071 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1076 // Okay, it is safe to hoist the terminator.
1077 Instruction *NT = I1->clone();
1078 BIParent->getInstList().insert(BI, NT);
1079 if (!NT->getType()->isVoidTy()) {
1080 I1->replaceAllUsesWith(NT);
1081 I2->replaceAllUsesWith(NT);
1085 IRBuilder<true, NoFolder> Builder(NT);
1086 // Hoisting one of the terminators from our successor is a great thing.
1087 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1088 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1089 // nodes, so we insert select instruction to compute the final result.
1090 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1091 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1093 for (BasicBlock::iterator BBI = SI->begin();
1094 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1095 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1096 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1097 if (BB1V == BB2V) continue;
1099 // These values do not agree. Insert a select instruction before NT
1100 // that determines the right value.
1101 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1103 SI = cast<SelectInst>
1104 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1105 BB1V->getName()+"."+BB2V->getName()));
1107 // Make the PHI node use the select for all incoming values for BB1/BB2
1108 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1109 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1110 PN->setIncomingValue(i, SI);
1114 // Update any PHI nodes in our new successors.
1115 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1116 AddPredecessorToBlock(*SI, BIParent, BB1);
1118 EraseTerminatorInstAndDCECond(BI);
1122 /// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd,
1123 /// check whether BBEnd has only two predecessors and the other predecessor
1124 /// ends with an unconditional branch. If it is true, sink any common code
1125 /// in the two predecessors to BBEnd.
1126 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1127 assert(BI1->isUnconditional());
1128 BasicBlock *BB1 = BI1->getParent();
1129 BasicBlock *BBEnd = BI1->getSuccessor(0);
1131 // Check that BBEnd has two predecessors and the other predecessor ends with
1132 // an unconditional branch.
1133 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1134 BasicBlock *Pred0 = *PI++;
1135 if (PI == PE) // Only one predecessor.
1137 BasicBlock *Pred1 = *PI++;
1138 if (PI != PE) // More than two predecessors.
1140 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1141 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1142 if (!BI2 || !BI2->isUnconditional())
1145 // Gather the PHI nodes in BBEnd.
1146 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2;
1147 Instruction *FirstNonPhiInBBEnd = 0;
1148 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end();
1150 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1151 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1152 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1153 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN);
1155 FirstNonPhiInBBEnd = &*I;
1159 if (!FirstNonPhiInBBEnd)
1163 // This does very trivial matching, with limited scanning, to find identical
1164 // instructions in the two blocks. We scan backward for obviously identical
1165 // instructions in an identical order.
1166 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1167 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(),
1168 RE2 = BB2->getInstList().rend();
1170 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1173 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1176 // Skip the unconditional branches.
1180 bool Changed = false;
1181 while (RI1 != RE1 && RI2 != RE2) {
1183 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1186 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1190 Instruction *I1 = &*RI1, *I2 = &*RI2;
1191 // I1 and I2 should have a single use in the same PHI node, and they
1192 // perform the same operation.
1193 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1194 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1195 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1196 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) ||
1197 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1198 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1199 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1200 !I1->hasOneUse() || !I2->hasOneUse() ||
1201 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() ||
1202 MapValueFromBB1ToBB2[I1].first != I2)
1205 // Check whether we should swap the operands of ICmpInst.
1206 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1207 bool SwapOpnds = false;
1208 if (ICmp1 && ICmp2 &&
1209 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1210 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1211 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1212 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1213 ICmp2->swapOperands();
1216 if (!I1->isSameOperationAs(I2)) {
1218 ICmp2->swapOperands();
1222 // The operands should be either the same or they need to be generated
1223 // with a PHI node after sinking. We only handle the case where there is
1224 // a single pair of different operands.
1225 Value *DifferentOp1 = 0, *DifferentOp2 = 0;
1226 unsigned Op1Idx = 0;
1227 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1228 if (I1->getOperand(I) == I2->getOperand(I))
1230 // Early exit if we have more-than one pair of different operands or
1231 // the different operand is already in MapValueFromBB1ToBB2.
1232 // Early exit if we need a PHI node to replace a constant.
1234 MapValueFromBB1ToBB2.find(I1->getOperand(I)) !=
1235 MapValueFromBB1ToBB2.end() ||
1236 isa<Constant>(I1->getOperand(I)) ||
1237 isa<Constant>(I2->getOperand(I))) {
1238 // If we can't sink the instructions, undo the swapping.
1240 ICmp2->swapOperands();
1243 DifferentOp1 = I1->getOperand(I);
1245 DifferentOp2 = I2->getOperand(I);
1248 // We insert the pair of different operands to MapValueFromBB1ToBB2 and
1249 // remove (I1, I2) from MapValueFromBB1ToBB2.
1251 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2,
1252 DifferentOp1->getName() + ".sink",
1254 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN);
1255 // I1 should use NewPN instead of DifferentOp1.
1256 I1->setOperand(Op1Idx, NewPN);
1257 NewPN->addIncoming(DifferentOp1, BB1);
1258 NewPN->addIncoming(DifferentOp2, BB2);
1259 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1261 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second;
1262 MapValueFromBB1ToBB2.erase(I1);
1264 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";);
1265 DEBUG(dbgs() << " " << *I2 << "\n";);
1266 // We need to update RE1 and RE2 if we are going to sink the first
1267 // instruction in the basic block down.
1268 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1269 // Sink the instruction.
1270 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1);
1271 if (!OldPN->use_empty())
1272 OldPN->replaceAllUsesWith(I1);
1273 OldPN->eraseFromParent();
1275 if (!I2->use_empty())
1276 I2->replaceAllUsesWith(I1);
1277 I1->intersectOptionalDataWith(I2);
1278 I2->eraseFromParent();
1281 RE1 = BB1->getInstList().rend();
1283 RE2 = BB2->getInstList().rend();
1284 FirstNonPhiInBBEnd = I1;
1291 /// \brief Determine if we can hoist sink a sole store instruction out of a
1292 /// conditional block.
1294 /// We are looking for code like the following:
1296 /// store i32 %add, i32* %arrayidx2
1297 /// ... // No other stores or function calls (we could be calling a memory
1298 /// ... // function).
1299 /// %cmp = icmp ult %x, %y
1300 /// br i1 %cmp, label %EndBB, label %ThenBB
1302 /// store i32 %add5, i32* %arrayidx2
1306 /// We are going to transform this into:
1308 /// store i32 %add, i32* %arrayidx2
1310 /// %cmp = icmp ult %x, %y
1311 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1312 /// store i32 %add.add5, i32* %arrayidx2
1315 /// \return The pointer to the value of the previous store if the store can be
1316 /// hoisted into the predecessor block. 0 otherwise.
1317 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1318 BasicBlock *StoreBB, BasicBlock *EndBB) {
1319 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1323 // Volatile or atomic.
1324 if (!StoreToHoist->isSimple())
1327 Value *StorePtr = StoreToHoist->getPointerOperand();
1329 // Look for a store to the same pointer in BrBB.
1330 unsigned MaxNumInstToLookAt = 10;
1331 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(),
1332 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) {
1333 Instruction *CurI = &*RI;
1335 // Could be calling an instruction that effects memory like free().
1336 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI))
1339 StoreInst *SI = dyn_cast<StoreInst>(CurI);
1340 // Found the previous store make sure it stores to the same location.
1341 if (SI && SI->getPointerOperand() == StorePtr)
1342 // Found the previous store, return its value operand.
1343 return SI->getValueOperand();
1345 return 0; // Unknown store.
1351 /// \brief Speculate a conditional basic block flattening the CFG.
1353 /// Note that this is a very risky transform currently. Speculating
1354 /// instructions like this is most often not desirable. Instead, there is an MI
1355 /// pass which can do it with full awareness of the resource constraints.
1356 /// However, some cases are "obvious" and we should do directly. An example of
1357 /// this is speculating a single, reasonably cheap instruction.
1359 /// There is only one distinct advantage to flattening the CFG at the IR level:
1360 /// it makes very common but simplistic optimizations such as are common in
1361 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1362 /// modeling their effects with easier to reason about SSA value graphs.
1365 /// An illustration of this transform is turning this IR:
1368 /// %cmp = icmp ult %x, %y
1369 /// br i1 %cmp, label %EndBB, label %ThenBB
1371 /// %sub = sub %x, %y
1374 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1381 /// %cmp = icmp ult %x, %y
1382 /// %sub = sub %x, %y
1383 /// %cond = select i1 %cmp, 0, %sub
1387 /// \returns true if the conditional block is removed.
1388 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB) {
1389 // Be conservative for now. FP select instruction can often be expensive.
1390 Value *BrCond = BI->getCondition();
1391 if (isa<FCmpInst>(BrCond))
1394 BasicBlock *BB = BI->getParent();
1395 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1397 // If ThenBB is actually on the false edge of the conditional branch, remember
1398 // to swap the select operands later.
1399 bool Invert = false;
1400 if (ThenBB != BI->getSuccessor(0)) {
1401 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1404 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1406 // Keep a count of how many times instructions are used within CondBB when
1407 // they are candidates for sinking into CondBB. Specifically:
1408 // - They are defined in BB, and
1409 // - They have no side effects, and
1410 // - All of their uses are in CondBB.
1411 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1413 unsigned SpeculationCost = 0;
1414 Value *SpeculatedStoreValue = 0;
1415 StoreInst *SpeculatedStore = 0;
1416 for (BasicBlock::iterator BBI = ThenBB->begin(),
1417 BBE = std::prev(ThenBB->end());
1418 BBI != BBE; ++BBI) {
1419 Instruction *I = BBI;
1421 if (isa<DbgInfoIntrinsic>(I))
1424 // Only speculatively execution a single instruction (not counting the
1425 // terminator) for now.
1427 if (SpeculationCost > 1)
1430 // Don't hoist the instruction if it's unsafe or expensive.
1431 if (!isSafeToSpeculativelyExecute(I) &&
1432 !(HoistCondStores &&
1433 (SpeculatedStoreValue = isSafeToSpeculateStore(I, BB, ThenBB,
1436 if (!SpeculatedStoreValue &&
1437 ComputeSpeculationCost(I) > PHINodeFoldingThreshold)
1440 // Store the store speculation candidate.
1441 if (SpeculatedStoreValue)
1442 SpeculatedStore = cast<StoreInst>(I);
1444 // Do not hoist the instruction if any of its operands are defined but not
1445 // used in BB. The transformation will prevent the operand from
1446 // being sunk into the use block.
1447 for (User::op_iterator i = I->op_begin(), e = I->op_end();
1449 Instruction *OpI = dyn_cast<Instruction>(*i);
1450 if (!OpI || OpI->getParent() != BB ||
1451 OpI->mayHaveSideEffects())
1452 continue; // Not a candidate for sinking.
1454 ++SinkCandidateUseCounts[OpI];
1458 // Consider any sink candidates which are only used in CondBB as costs for
1459 // speculation. Note, while we iterate over a DenseMap here, we are summing
1460 // and so iteration order isn't significant.
1461 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I =
1462 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end();
1464 if (I->first->getNumUses() == I->second) {
1466 if (SpeculationCost > 1)
1470 // Check that the PHI nodes can be converted to selects.
1471 bool HaveRewritablePHIs = false;
1472 for (BasicBlock::iterator I = EndBB->begin();
1473 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1474 Value *OrigV = PN->getIncomingValueForBlock(BB);
1475 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1477 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1478 // Skip PHIs which are trivial.
1482 HaveRewritablePHIs = true;
1483 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1484 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1485 if (!OrigCE && !ThenCE)
1486 continue; // Known safe and cheap.
1488 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
1489 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
1491 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE) : 0;
1492 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE) : 0;
1493 if (OrigCost + ThenCost > 2 * PHINodeFoldingThreshold)
1496 // Account for the cost of an unfolded ConstantExpr which could end up
1497 // getting expanded into Instructions.
1498 // FIXME: This doesn't account for how many operations are combined in the
1499 // constant expression.
1501 if (SpeculationCost > 1)
1505 // If there are no PHIs to process, bail early. This helps ensure idempotence
1507 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
1510 // If we get here, we can hoist the instruction and if-convert.
1511 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1513 // Insert a select of the value of the speculated store.
1514 if (SpeculatedStoreValue) {
1515 IRBuilder<true, NoFolder> Builder(BI);
1516 Value *TrueV = SpeculatedStore->getValueOperand();
1517 Value *FalseV = SpeculatedStoreValue;
1519 std::swap(TrueV, FalseV);
1520 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() +
1521 "." + FalseV->getName());
1522 SpeculatedStore->setOperand(0, S);
1525 // Hoist the instructions.
1526 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(),
1527 std::prev(ThenBB->end()));
1529 // Insert selects and rewrite the PHI operands.
1530 IRBuilder<true, NoFolder> Builder(BI);
1531 for (BasicBlock::iterator I = EndBB->begin();
1532 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1533 unsigned OrigI = PN->getBasicBlockIndex(BB);
1534 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1535 Value *OrigV = PN->getIncomingValue(OrigI);
1536 Value *ThenV = PN->getIncomingValue(ThenI);
1538 // Skip PHIs which are trivial.
1542 // Create a select whose true value is the speculatively executed value and
1543 // false value is the preexisting value. Swap them if the branch
1544 // destinations were inverted.
1545 Value *TrueV = ThenV, *FalseV = OrigV;
1547 std::swap(TrueV, FalseV);
1548 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV,
1549 TrueV->getName() + "." + FalseV->getName());
1550 PN->setIncomingValue(OrigI, V);
1551 PN->setIncomingValue(ThenI, V);
1558 /// \returns True if this block contains a CallInst with the NoDuplicate
1560 static bool HasNoDuplicateCall(const BasicBlock *BB) {
1561 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1562 const CallInst *CI = dyn_cast<CallInst>(I);
1565 if (CI->cannotDuplicate())
1571 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1572 /// across this block.
1573 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1574 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1577 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1578 if (isa<DbgInfoIntrinsic>(BBI))
1580 if (Size > 10) return false; // Don't clone large BB's.
1583 // We can only support instructions that do not define values that are
1584 // live outside of the current basic block.
1585 for (User *U : BBI->users()) {
1586 Instruction *UI = cast<Instruction>(U);
1587 if (UI->getParent() != BB || isa<PHINode>(UI)) return false;
1590 // Looks ok, continue checking.
1596 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1597 /// that is defined in the same block as the branch and if any PHI entries are
1598 /// constants, thread edges corresponding to that entry to be branches to their
1599 /// ultimate destination.
1600 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *DL) {
1601 BasicBlock *BB = BI->getParent();
1602 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1603 // NOTE: we currently cannot transform this case if the PHI node is used
1604 // outside of the block.
1605 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1608 // Degenerate case of a single entry PHI.
1609 if (PN->getNumIncomingValues() == 1) {
1610 FoldSingleEntryPHINodes(PN->getParent());
1614 // Now we know that this block has multiple preds and two succs.
1615 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1617 if (HasNoDuplicateCall(BB)) return false;
1619 // Okay, this is a simple enough basic block. See if any phi values are
1621 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1622 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1623 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue;
1625 // Okay, we now know that all edges from PredBB should be revectored to
1626 // branch to RealDest.
1627 BasicBlock *PredBB = PN->getIncomingBlock(i);
1628 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1630 if (RealDest == BB) continue; // Skip self loops.
1631 // Skip if the predecessor's terminator is an indirect branch.
1632 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1634 // The dest block might have PHI nodes, other predecessors and other
1635 // difficult cases. Instead of being smart about this, just insert a new
1636 // block that jumps to the destination block, effectively splitting
1637 // the edge we are about to create.
1638 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1639 RealDest->getName()+".critedge",
1640 RealDest->getParent(), RealDest);
1641 BranchInst::Create(RealDest, EdgeBB);
1643 // Update PHI nodes.
1644 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1646 // BB may have instructions that are being threaded over. Clone these
1647 // instructions into EdgeBB. We know that there will be no uses of the
1648 // cloned instructions outside of EdgeBB.
1649 BasicBlock::iterator InsertPt = EdgeBB->begin();
1650 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1651 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1652 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1653 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1656 // Clone the instruction.
1657 Instruction *N = BBI->clone();
1658 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1660 // Update operands due to translation.
1661 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1663 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1664 if (PI != TranslateMap.end())
1668 // Check for trivial simplification.
1669 if (Value *V = SimplifyInstruction(N, DL)) {
1670 TranslateMap[BBI] = V;
1671 delete N; // Instruction folded away, don't need actual inst
1673 // Insert the new instruction into its new home.
1674 EdgeBB->getInstList().insert(InsertPt, N);
1675 if (!BBI->use_empty())
1676 TranslateMap[BBI] = N;
1680 // Loop over all of the edges from PredBB to BB, changing them to branch
1681 // to EdgeBB instead.
1682 TerminatorInst *PredBBTI = PredBB->getTerminator();
1683 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1684 if (PredBBTI->getSuccessor(i) == BB) {
1685 BB->removePredecessor(PredBB);
1686 PredBBTI->setSuccessor(i, EdgeBB);
1689 // Recurse, simplifying any other constants.
1690 return FoldCondBranchOnPHI(BI, DL) | true;
1696 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1697 /// PHI node, see if we can eliminate it.
1698 static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *DL) {
1699 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1700 // statement", which has a very simple dominance structure. Basically, we
1701 // are trying to find the condition that is being branched on, which
1702 // subsequently causes this merge to happen. We really want control
1703 // dependence information for this check, but simplifycfg can't keep it up
1704 // to date, and this catches most of the cases we care about anyway.
1705 BasicBlock *BB = PN->getParent();
1706 BasicBlock *IfTrue, *IfFalse;
1707 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1709 // Don't bother if the branch will be constant folded trivially.
1710 isa<ConstantInt>(IfCond))
1713 // Okay, we found that we can merge this two-entry phi node into a select.
1714 // Doing so would require us to fold *all* two entry phi nodes in this block.
1715 // At some point this becomes non-profitable (particularly if the target
1716 // doesn't support cmov's). Only do this transformation if there are two or
1717 // fewer PHI nodes in this block.
1718 unsigned NumPhis = 0;
1719 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1723 // Loop over the PHI's seeing if we can promote them all to select
1724 // instructions. While we are at it, keep track of the instructions
1725 // that need to be moved to the dominating block.
1726 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1727 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1728 MaxCostVal1 = PHINodeFoldingThreshold;
1730 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1731 PHINode *PN = cast<PHINode>(II++);
1732 if (Value *V = SimplifyInstruction(PN, DL)) {
1733 PN->replaceAllUsesWith(V);
1734 PN->eraseFromParent();
1738 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1740 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1745 // If we folded the first phi, PN dangles at this point. Refresh it. If
1746 // we ran out of PHIs then we simplified them all.
1747 PN = dyn_cast<PHINode>(BB->begin());
1748 if (PN == 0) return true;
1750 // Don't fold i1 branches on PHIs which contain binary operators. These can
1751 // often be turned into switches and other things.
1752 if (PN->getType()->isIntegerTy(1) &&
1753 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1754 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1755 isa<BinaryOperator>(IfCond)))
1758 // If we all PHI nodes are promotable, check to make sure that all
1759 // instructions in the predecessor blocks can be promoted as well. If
1760 // not, we won't be able to get rid of the control flow, so it's not
1761 // worth promoting to select instructions.
1762 BasicBlock *DomBlock = 0;
1763 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1764 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1765 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1768 DomBlock = *pred_begin(IfBlock1);
1769 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1770 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1771 // This is not an aggressive instruction that we can promote.
1772 // Because of this, we won't be able to get rid of the control
1773 // flow, so the xform is not worth it.
1778 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1781 DomBlock = *pred_begin(IfBlock2);
1782 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1783 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1784 // This is not an aggressive instruction that we can promote.
1785 // Because of this, we won't be able to get rid of the control
1786 // flow, so the xform is not worth it.
1791 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1792 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1794 // If we can still promote the PHI nodes after this gauntlet of tests,
1795 // do all of the PHI's now.
1796 Instruction *InsertPt = DomBlock->getTerminator();
1797 IRBuilder<true, NoFolder> Builder(InsertPt);
1799 // Move all 'aggressive' instructions, which are defined in the
1800 // conditional parts of the if's up to the dominating block.
1802 DomBlock->getInstList().splice(InsertPt,
1803 IfBlock1->getInstList(), IfBlock1->begin(),
1804 IfBlock1->getTerminator());
1806 DomBlock->getInstList().splice(InsertPt,
1807 IfBlock2->getInstList(), IfBlock2->begin(),
1808 IfBlock2->getTerminator());
1810 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1811 // Change the PHI node into a select instruction.
1812 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1813 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1816 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1817 PN->replaceAllUsesWith(NV);
1819 PN->eraseFromParent();
1822 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1823 // has been flattened. Change DomBlock to jump directly to our new block to
1824 // avoid other simplifycfg's kicking in on the diamond.
1825 TerminatorInst *OldTI = DomBlock->getTerminator();
1826 Builder.SetInsertPoint(OldTI);
1827 Builder.CreateBr(BB);
1828 OldTI->eraseFromParent();
1832 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1833 /// to two returning blocks, try to merge them together into one return,
1834 /// introducing a select if the return values disagree.
1835 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1836 IRBuilder<> &Builder) {
1837 assert(BI->isConditional() && "Must be a conditional branch");
1838 BasicBlock *TrueSucc = BI->getSuccessor(0);
1839 BasicBlock *FalseSucc = BI->getSuccessor(1);
1840 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1841 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1843 // Check to ensure both blocks are empty (just a return) or optionally empty
1844 // with PHI nodes. If there are other instructions, merging would cause extra
1845 // computation on one path or the other.
1846 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1848 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1851 Builder.SetInsertPoint(BI);
1852 // Okay, we found a branch that is going to two return nodes. If
1853 // there is no return value for this function, just change the
1854 // branch into a return.
1855 if (FalseRet->getNumOperands() == 0) {
1856 TrueSucc->removePredecessor(BI->getParent());
1857 FalseSucc->removePredecessor(BI->getParent());
1858 Builder.CreateRetVoid();
1859 EraseTerminatorInstAndDCECond(BI);
1863 // Otherwise, figure out what the true and false return values are
1864 // so we can insert a new select instruction.
1865 Value *TrueValue = TrueRet->getReturnValue();
1866 Value *FalseValue = FalseRet->getReturnValue();
1868 // Unwrap any PHI nodes in the return blocks.
1869 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1870 if (TVPN->getParent() == TrueSucc)
1871 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1872 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1873 if (FVPN->getParent() == FalseSucc)
1874 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1876 // In order for this transformation to be safe, we must be able to
1877 // unconditionally execute both operands to the return. This is
1878 // normally the case, but we could have a potentially-trapping
1879 // constant expression that prevents this transformation from being
1881 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1884 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1888 // Okay, we collected all the mapped values and checked them for sanity, and
1889 // defined to really do this transformation. First, update the CFG.
1890 TrueSucc->removePredecessor(BI->getParent());
1891 FalseSucc->removePredecessor(BI->getParent());
1893 // Insert select instructions where needed.
1894 Value *BrCond = BI->getCondition();
1896 // Insert a select if the results differ.
1897 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1898 } else if (isa<UndefValue>(TrueValue)) {
1899 TrueValue = FalseValue;
1901 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1902 FalseValue, "retval");
1906 Value *RI = !TrueValue ?
1907 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
1911 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
1912 << "\n " << *BI << "NewRet = " << *RI
1913 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
1915 EraseTerminatorInstAndDCECond(BI);
1920 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
1921 /// probabilities of the branch taking each edge. Fills in the two APInt
1922 /// parameters and return true, or returns false if no or invalid metadata was
1924 static bool ExtractBranchMetadata(BranchInst *BI,
1925 uint64_t &ProbTrue, uint64_t &ProbFalse) {
1926 assert(BI->isConditional() &&
1927 "Looking for probabilities on unconditional branch?");
1928 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
1929 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
1930 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
1931 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
1932 if (!CITrue || !CIFalse) return false;
1933 ProbTrue = CITrue->getValue().getZExtValue();
1934 ProbFalse = CIFalse->getValue().getZExtValue();
1938 /// checkCSEInPredecessor - Return true if the given instruction is available
1939 /// in its predecessor block. If yes, the instruction will be removed.
1941 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
1942 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
1944 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
1945 Instruction *PBI = &*I;
1946 // Check whether Inst and PBI generate the same value.
1947 if (Inst->isIdenticalTo(PBI)) {
1948 Inst->replaceAllUsesWith(PBI);
1949 Inst->eraseFromParent();
1956 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
1957 /// predecessor branches to us and one of our successors, fold the block into
1958 /// the predecessor and use logical operations to pick the right destination.
1959 bool llvm::FoldBranchToCommonDest(BranchInst *BI) {
1960 BasicBlock *BB = BI->getParent();
1962 Instruction *Cond = 0;
1963 if (BI->isConditional())
1964 Cond = dyn_cast<Instruction>(BI->getCondition());
1966 // For unconditional branch, check for a simple CFG pattern, where
1967 // BB has a single predecessor and BB's successor is also its predecessor's
1968 // successor. If such pattern exisits, check for CSE between BB and its
1970 if (BasicBlock *PB = BB->getSinglePredecessor())
1971 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
1972 if (PBI->isConditional() &&
1973 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
1974 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
1975 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
1977 Instruction *Curr = I++;
1978 if (isa<CmpInst>(Curr)) {
1982 // Quit if we can't remove this instruction.
1983 if (!checkCSEInPredecessor(Curr, PB))
1992 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
1993 Cond->getParent() != BB || !Cond->hasOneUse())
1996 // Only allow this if the condition is a simple instruction that can be
1997 // executed unconditionally. It must be in the same block as the branch, and
1998 // must be at the front of the block.
1999 BasicBlock::iterator FrontIt = BB->front();
2001 // Ignore dbg intrinsics.
2002 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
2004 // Allow a single instruction to be hoisted in addition to the compare
2005 // that feeds the branch. We later ensure that any values that _it_ uses
2006 // were also live in the predecessor, so that we don't unnecessarily create
2007 // register pressure or inhibit out-of-order execution.
2008 Instruction *BonusInst = 0;
2009 if (&*FrontIt != Cond &&
2010 FrontIt->hasOneUse() && FrontIt->user_back() == Cond &&
2011 isSafeToSpeculativelyExecute(FrontIt)) {
2012 BonusInst = &*FrontIt;
2015 // Ignore dbg intrinsics.
2016 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
2019 // Only a single bonus inst is allowed.
2020 if (&*FrontIt != Cond)
2023 // Make sure the instruction after the condition is the cond branch.
2024 BasicBlock::iterator CondIt = Cond; ++CondIt;
2026 // Ingore dbg intrinsics.
2027 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
2032 // Cond is known to be a compare or binary operator. Check to make sure that
2033 // neither operand is a potentially-trapping constant expression.
2034 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2037 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2041 // Finally, don't infinitely unroll conditional loops.
2042 BasicBlock *TrueDest = BI->getSuccessor(0);
2043 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0;
2044 if (TrueDest == BB || FalseDest == BB)
2047 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2048 BasicBlock *PredBlock = *PI;
2049 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2051 // Check that we have two conditional branches. If there is a PHI node in
2052 // the common successor, verify that the same value flows in from both
2054 SmallVector<PHINode*, 4> PHIs;
2055 if (PBI == 0 || PBI->isUnconditional() ||
2056 (BI->isConditional() &&
2057 !SafeToMergeTerminators(BI, PBI)) ||
2058 (!BI->isConditional() &&
2059 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2062 // Determine if the two branches share a common destination.
2063 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2064 bool InvertPredCond = false;
2066 if (BI->isConditional()) {
2067 if (PBI->getSuccessor(0) == TrueDest)
2068 Opc = Instruction::Or;
2069 else if (PBI->getSuccessor(1) == FalseDest)
2070 Opc = Instruction::And;
2071 else if (PBI->getSuccessor(0) == FalseDest)
2072 Opc = Instruction::And, InvertPredCond = true;
2073 else if (PBI->getSuccessor(1) == TrueDest)
2074 Opc = Instruction::Or, InvertPredCond = true;
2078 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2082 // Ensure that any values used in the bonus instruction are also used
2083 // by the terminator of the predecessor. This means that those values
2084 // must already have been resolved, so we won't be inhibiting the
2085 // out-of-order core by speculating them earlier. We also allow
2086 // instructions that are used by the terminator's condition because it
2087 // exposes more merging opportunities.
2088 bool UsedByBranch = (BonusInst && BonusInst->hasOneUse() &&
2089 BonusInst->user_back() == Cond);
2091 if (BonusInst && !UsedByBranch) {
2092 // Collect the values used by the bonus inst
2093 SmallPtrSet<Value*, 4> UsedValues;
2094 for (Instruction::op_iterator OI = BonusInst->op_begin(),
2095 OE = BonusInst->op_end(); OI != OE; ++OI) {
2097 if (!isa<Constant>(V) && !isa<Argument>(V))
2098 UsedValues.insert(V);
2101 SmallVector<std::pair<Value*, unsigned>, 4> Worklist;
2102 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0));
2104 // Walk up to four levels back up the use-def chain of the predecessor's
2105 // terminator to see if all those values were used. The choice of four
2106 // levels is arbitrary, to provide a compile-time-cost bound.
2107 while (!Worklist.empty()) {
2108 std::pair<Value*, unsigned> Pair = Worklist.back();
2109 Worklist.pop_back();
2111 if (Pair.second >= 4) continue;
2112 UsedValues.erase(Pair.first);
2113 if (UsedValues.empty()) break;
2115 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) {
2116 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
2118 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1));
2122 if (!UsedValues.empty()) return false;
2125 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2126 IRBuilder<> Builder(PBI);
2128 // If we need to invert the condition in the pred block to match, do so now.
2129 if (InvertPredCond) {
2130 Value *NewCond = PBI->getCondition();
2132 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2133 CmpInst *CI = cast<CmpInst>(NewCond);
2134 CI->setPredicate(CI->getInversePredicate());
2136 NewCond = Builder.CreateNot(NewCond,
2137 PBI->getCondition()->getName()+".not");
2140 PBI->setCondition(NewCond);
2141 PBI->swapSuccessors();
2144 // If we have a bonus inst, clone it into the predecessor block.
2145 Instruction *NewBonus = 0;
2147 NewBonus = BonusInst->clone();
2149 // If we moved a load, we cannot any longer claim any knowledge about
2150 // its potential value. The previous information might have been valid
2151 // only given the branch precondition.
2152 // For an analogous reason, we must also drop all the metadata whose
2153 // semantics we don't understand.
2154 NewBonus->dropUnknownMetadata(LLVMContext::MD_dbg);
2156 PredBlock->getInstList().insert(PBI, NewBonus);
2157 NewBonus->takeName(BonusInst);
2158 BonusInst->setName(BonusInst->getName()+".old");
2161 // Clone Cond into the predecessor basic block, and or/and the
2162 // two conditions together.
2163 Instruction *New = Cond->clone();
2164 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus);
2165 PredBlock->getInstList().insert(PBI, New);
2166 New->takeName(Cond);
2167 Cond->setName(New->getName()+".old");
2169 if (BI->isConditional()) {
2170 Instruction *NewCond =
2171 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2173 PBI->setCondition(NewCond);
2175 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2176 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2178 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2180 SmallVector<uint64_t, 8> NewWeights;
2182 if (PBI->getSuccessor(0) == BB) {
2183 if (PredHasWeights && SuccHasWeights) {
2184 // PBI: br i1 %x, BB, FalseDest
2185 // BI: br i1 %y, TrueDest, FalseDest
2186 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2187 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2188 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2189 // TrueWeight for PBI * FalseWeight for BI.
2190 // We assume that total weights of a BranchInst can fit into 32 bits.
2191 // Therefore, we will not have overflow using 64-bit arithmetic.
2192 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2193 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2195 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2196 PBI->setSuccessor(0, TrueDest);
2198 if (PBI->getSuccessor(1) == BB) {
2199 if (PredHasWeights && SuccHasWeights) {
2200 // PBI: br i1 %x, TrueDest, BB
2201 // BI: br i1 %y, TrueDest, FalseDest
2202 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2203 // FalseWeight for PBI * TrueWeight for BI.
2204 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2205 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2206 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2207 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2209 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2210 PBI->setSuccessor(1, FalseDest);
2212 if (NewWeights.size() == 2) {
2213 // Halve the weights if any of them cannot fit in an uint32_t
2214 FitWeights(NewWeights);
2216 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2217 PBI->setMetadata(LLVMContext::MD_prof,
2218 MDBuilder(BI->getContext()).
2219 createBranchWeights(MDWeights));
2221 PBI->setMetadata(LLVMContext::MD_prof, NULL);
2223 // Update PHI nodes in the common successors.
2224 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2225 ConstantInt *PBI_C = cast<ConstantInt>(
2226 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2227 assert(PBI_C->getType()->isIntegerTy(1));
2228 Instruction *MergedCond = 0;
2229 if (PBI->getSuccessor(0) == TrueDest) {
2230 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2231 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2232 // is false: !PBI_Cond and BI_Value
2233 Instruction *NotCond =
2234 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2237 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2242 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2243 PBI->getCondition(), MergedCond,
2246 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2247 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2248 // is false: PBI_Cond and BI_Value
2250 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2251 PBI->getCondition(), New,
2253 if (PBI_C->isOne()) {
2254 Instruction *NotCond =
2255 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2258 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2259 NotCond, MergedCond,
2264 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2267 // Change PBI from Conditional to Unconditional.
2268 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2269 EraseTerminatorInstAndDCECond(PBI);
2273 // TODO: If BB is reachable from all paths through PredBlock, then we
2274 // could replace PBI's branch probabilities with BI's.
2276 // Copy any debug value intrinsics into the end of PredBlock.
2277 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2278 if (isa<DbgInfoIntrinsic>(*I))
2279 I->clone()->insertBefore(PBI);
2286 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2287 /// predecessor of another block, this function tries to simplify it. We know
2288 /// that PBI and BI are both conditional branches, and BI is in one of the
2289 /// successor blocks of PBI - PBI branches to BI.
2290 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2291 assert(PBI->isConditional() && BI->isConditional());
2292 BasicBlock *BB = BI->getParent();
2294 // If this block ends with a branch instruction, and if there is a
2295 // predecessor that ends on a branch of the same condition, make
2296 // this conditional branch redundant.
2297 if (PBI->getCondition() == BI->getCondition() &&
2298 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2299 // Okay, the outcome of this conditional branch is statically
2300 // knowable. If this block had a single pred, handle specially.
2301 if (BB->getSinglePredecessor()) {
2302 // Turn this into a branch on constant.
2303 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2304 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2306 return true; // Nuke the branch on constant.
2309 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2310 // in the constant and simplify the block result. Subsequent passes of
2311 // simplifycfg will thread the block.
2312 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2313 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2314 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2315 std::distance(PB, PE),
2316 BI->getCondition()->getName() + ".pr",
2318 // Okay, we're going to insert the PHI node. Since PBI is not the only
2319 // predecessor, compute the PHI'd conditional value for all of the preds.
2320 // Any predecessor where the condition is not computable we keep symbolic.
2321 for (pred_iterator PI = PB; PI != PE; ++PI) {
2322 BasicBlock *P = *PI;
2323 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2324 PBI != BI && PBI->isConditional() &&
2325 PBI->getCondition() == BI->getCondition() &&
2326 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2327 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2328 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2331 NewPN->addIncoming(BI->getCondition(), P);
2335 BI->setCondition(NewPN);
2340 // If this is a conditional branch in an empty block, and if any
2341 // predecessors is a conditional branch to one of our destinations,
2342 // fold the conditions into logical ops and one cond br.
2343 BasicBlock::iterator BBI = BB->begin();
2344 // Ignore dbg intrinsics.
2345 while (isa<DbgInfoIntrinsic>(BBI))
2351 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2356 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2358 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2359 PBIOp = 0, BIOp = 1;
2360 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2361 PBIOp = 1, BIOp = 0;
2362 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2367 // Check to make sure that the other destination of this branch
2368 // isn't BB itself. If so, this is an infinite loop that will
2369 // keep getting unwound.
2370 if (PBI->getSuccessor(PBIOp) == BB)
2373 // Do not perform this transformation if it would require
2374 // insertion of a large number of select instructions. For targets
2375 // without predication/cmovs, this is a big pessimization.
2376 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2378 unsigned NumPhis = 0;
2379 for (BasicBlock::iterator II = CommonDest->begin();
2380 isa<PHINode>(II); ++II, ++NumPhis)
2381 if (NumPhis > 2) // Disable this xform.
2384 // Finally, if everything is ok, fold the branches to logical ops.
2385 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2387 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2388 << "AND: " << *BI->getParent());
2391 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2392 // branch in it, where one edge (OtherDest) goes back to itself but the other
2393 // exits. We don't *know* that the program avoids the infinite loop
2394 // (even though that seems likely). If we do this xform naively, we'll end up
2395 // recursively unpeeling the loop. Since we know that (after the xform is
2396 // done) that the block *is* infinite if reached, we just make it an obviously
2397 // infinite loop with no cond branch.
2398 if (OtherDest == BB) {
2399 // Insert it at the end of the function, because it's either code,
2400 // or it won't matter if it's hot. :)
2401 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2402 "infloop", BB->getParent());
2403 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2404 OtherDest = InfLoopBlock;
2407 DEBUG(dbgs() << *PBI->getParent()->getParent());
2409 // BI may have other predecessors. Because of this, we leave
2410 // it alone, but modify PBI.
2412 // Make sure we get to CommonDest on True&True directions.
2413 Value *PBICond = PBI->getCondition();
2414 IRBuilder<true, NoFolder> Builder(PBI);
2416 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2418 Value *BICond = BI->getCondition();
2420 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2422 // Merge the conditions.
2423 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2425 // Modify PBI to branch on the new condition to the new dests.
2426 PBI->setCondition(Cond);
2427 PBI->setSuccessor(0, CommonDest);
2428 PBI->setSuccessor(1, OtherDest);
2430 // Update branch weight for PBI.
2431 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2432 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2434 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2436 if (PredHasWeights && SuccHasWeights) {
2437 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2438 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2439 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2440 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2441 // The weight to CommonDest should be PredCommon * SuccTotal +
2442 // PredOther * SuccCommon.
2443 // The weight to OtherDest should be PredOther * SuccOther.
2444 SmallVector<uint64_t, 2> NewWeights;
2445 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) +
2446 PredOther * SuccCommon);
2447 NewWeights.push_back(PredOther * SuccOther);
2448 // Halve the weights if any of them cannot fit in an uint32_t
2449 FitWeights(NewWeights);
2451 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end());
2452 PBI->setMetadata(LLVMContext::MD_prof,
2453 MDBuilder(BI->getContext()).
2454 createBranchWeights(MDWeights));
2457 // OtherDest may have phi nodes. If so, add an entry from PBI's
2458 // block that are identical to the entries for BI's block.
2459 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2461 // We know that the CommonDest already had an edge from PBI to
2462 // it. If it has PHIs though, the PHIs may have different
2463 // entries for BB and PBI's BB. If so, insert a select to make
2466 for (BasicBlock::iterator II = CommonDest->begin();
2467 (PN = dyn_cast<PHINode>(II)); ++II) {
2468 Value *BIV = PN->getIncomingValueForBlock(BB);
2469 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2470 Value *PBIV = PN->getIncomingValue(PBBIdx);
2472 // Insert a select in PBI to pick the right value.
2473 Value *NV = cast<SelectInst>
2474 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2475 PN->setIncomingValue(PBBIdx, NV);
2479 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2480 DEBUG(dbgs() << *PBI->getParent()->getParent());
2482 // This basic block is probably dead. We know it has at least
2483 // one fewer predecessor.
2487 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2488 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2489 // Takes care of updating the successors and removing the old terminator.
2490 // Also makes sure not to introduce new successors by assuming that edges to
2491 // non-successor TrueBBs and FalseBBs aren't reachable.
2492 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2493 BasicBlock *TrueBB, BasicBlock *FalseBB,
2494 uint32_t TrueWeight,
2495 uint32_t FalseWeight){
2496 // Remove any superfluous successor edges from the CFG.
2497 // First, figure out which successors to preserve.
2498 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2500 BasicBlock *KeepEdge1 = TrueBB;
2501 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0;
2503 // Then remove the rest.
2504 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2505 BasicBlock *Succ = OldTerm->getSuccessor(I);
2506 // Make sure only to keep exactly one copy of each edge.
2507 if (Succ == KeepEdge1)
2509 else if (Succ == KeepEdge2)
2512 Succ->removePredecessor(OldTerm->getParent());
2515 IRBuilder<> Builder(OldTerm);
2516 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2518 // Insert an appropriate new terminator.
2519 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) {
2520 if (TrueBB == FalseBB)
2521 // We were only looking for one successor, and it was present.
2522 // Create an unconditional branch to it.
2523 Builder.CreateBr(TrueBB);
2525 // We found both of the successors we were looking for.
2526 // Create a conditional branch sharing the condition of the select.
2527 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2528 if (TrueWeight != FalseWeight)
2529 NewBI->setMetadata(LLVMContext::MD_prof,
2530 MDBuilder(OldTerm->getContext()).
2531 createBranchWeights(TrueWeight, FalseWeight));
2533 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2534 // Neither of the selected blocks were successors, so this
2535 // terminator must be unreachable.
2536 new UnreachableInst(OldTerm->getContext(), OldTerm);
2538 // One of the selected values was a successor, but the other wasn't.
2539 // Insert an unconditional branch to the one that was found;
2540 // the edge to the one that wasn't must be unreachable.
2542 // Only TrueBB was found.
2543 Builder.CreateBr(TrueBB);
2545 // Only FalseBB was found.
2546 Builder.CreateBr(FalseBB);
2549 EraseTerminatorInstAndDCECond(OldTerm);
2553 // SimplifySwitchOnSelect - Replaces
2554 // (switch (select cond, X, Y)) on constant X, Y
2555 // with a branch - conditional if X and Y lead to distinct BBs,
2556 // unconditional otherwise.
2557 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2558 // Check for constant integer values in the select.
2559 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2560 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2561 if (!TrueVal || !FalseVal)
2564 // Find the relevant condition and destinations.
2565 Value *Condition = Select->getCondition();
2566 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2567 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2569 // Get weight for TrueBB and FalseBB.
2570 uint32_t TrueWeight = 0, FalseWeight = 0;
2571 SmallVector<uint64_t, 8> Weights;
2572 bool HasWeights = HasBranchWeights(SI);
2574 GetBranchWeights(SI, Weights);
2575 if (Weights.size() == 1 + SI->getNumCases()) {
2576 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2577 getSuccessorIndex()];
2578 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2579 getSuccessorIndex()];
2583 // Perform the actual simplification.
2584 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2585 TrueWeight, FalseWeight);
2588 // SimplifyIndirectBrOnSelect - Replaces
2589 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2590 // blockaddress(@fn, BlockB)))
2592 // (br cond, BlockA, BlockB).
2593 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2594 // Check that both operands of the select are block addresses.
2595 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2596 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2600 // Extract the actual blocks.
2601 BasicBlock *TrueBB = TBA->getBasicBlock();
2602 BasicBlock *FalseBB = FBA->getBasicBlock();
2604 // Perform the actual simplification.
2605 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2609 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2610 /// instruction (a seteq/setne with a constant) as the only instruction in a
2611 /// block that ends with an uncond branch. We are looking for a very specific
2612 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2613 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2614 /// default value goes to an uncond block with a seteq in it, we get something
2617 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2619 /// %tmp = icmp eq i8 %A, 92
2622 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2624 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2625 /// the PHI, merging the third icmp into the switch.
2626 static bool TryToSimplifyUncondBranchWithICmpInIt(
2627 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI,
2628 const DataLayout *DL) {
2629 BasicBlock *BB = ICI->getParent();
2631 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2633 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2635 Value *V = ICI->getOperand(0);
2636 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2638 // The pattern we're looking for is where our only predecessor is a switch on
2639 // 'V' and this block is the default case for the switch. In this case we can
2640 // fold the compared value into the switch to simplify things.
2641 BasicBlock *Pred = BB->getSinglePredecessor();
2642 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false;
2644 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2645 if (SI->getCondition() != V)
2648 // If BB is reachable on a non-default case, then we simply know the value of
2649 // V in this block. Substitute it and constant fold the icmp instruction
2651 if (SI->getDefaultDest() != BB) {
2652 ConstantInt *VVal = SI->findCaseDest(BB);
2653 assert(VVal && "Should have a unique destination value");
2654 ICI->setOperand(0, VVal);
2656 if (Value *V = SimplifyInstruction(ICI, DL)) {
2657 ICI->replaceAllUsesWith(V);
2658 ICI->eraseFromParent();
2660 // BB is now empty, so it is likely to simplify away.
2661 return SimplifyCFG(BB, TTI, DL) | true;
2664 // Ok, the block is reachable from the default dest. If the constant we're
2665 // comparing exists in one of the other edges, then we can constant fold ICI
2667 if (SI->findCaseValue(Cst) != SI->case_default()) {
2669 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2670 V = ConstantInt::getFalse(BB->getContext());
2672 V = ConstantInt::getTrue(BB->getContext());
2674 ICI->replaceAllUsesWith(V);
2675 ICI->eraseFromParent();
2676 // BB is now empty, so it is likely to simplify away.
2677 return SimplifyCFG(BB, TTI, DL) | true;
2680 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2682 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2683 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
2684 if (PHIUse == 0 || PHIUse != &SuccBlock->front() ||
2685 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2688 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2690 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2691 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2693 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2694 std::swap(DefaultCst, NewCst);
2696 // Replace ICI (which is used by the PHI for the default value) with true or
2697 // false depending on if it is EQ or NE.
2698 ICI->replaceAllUsesWith(DefaultCst);
2699 ICI->eraseFromParent();
2701 // Okay, the switch goes to this block on a default value. Add an edge from
2702 // the switch to the merge point on the compared value.
2703 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2704 BB->getParent(), BB);
2705 SmallVector<uint64_t, 8> Weights;
2706 bool HasWeights = HasBranchWeights(SI);
2708 GetBranchWeights(SI, Weights);
2709 if (Weights.size() == 1 + SI->getNumCases()) {
2710 // Split weight for default case to case for "Cst".
2711 Weights[0] = (Weights[0]+1) >> 1;
2712 Weights.push_back(Weights[0]);
2714 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2715 SI->setMetadata(LLVMContext::MD_prof,
2716 MDBuilder(SI->getContext()).
2717 createBranchWeights(MDWeights));
2720 SI->addCase(Cst, NewBB);
2722 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2723 Builder.SetInsertPoint(NewBB);
2724 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2725 Builder.CreateBr(SuccBlock);
2726 PHIUse->addIncoming(NewCst, NewBB);
2730 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2731 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2732 /// fold it into a switch instruction if so.
2733 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *DL,
2734 IRBuilder<> &Builder) {
2735 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2736 if (Cond == 0) return false;
2739 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2740 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2741 // 'setne's and'ed together, collect them.
2743 std::vector<ConstantInt*> Values;
2744 bool TrueWhenEqual = true;
2745 Value *ExtraCase = 0;
2746 unsigned UsedICmps = 0;
2748 if (Cond->getOpcode() == Instruction::Or) {
2749 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, DL, true,
2751 } else if (Cond->getOpcode() == Instruction::And) {
2752 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, DL, false,
2754 TrueWhenEqual = false;
2757 // If we didn't have a multiply compared value, fail.
2758 if (CompVal == 0) return false;
2760 // Avoid turning single icmps into a switch.
2764 // There might be duplicate constants in the list, which the switch
2765 // instruction can't handle, remove them now.
2766 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2767 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2769 // If Extra was used, we require at least two switch values to do the
2770 // transformation. A switch with one value is just an cond branch.
2771 if (ExtraCase && Values.size() < 2) return false;
2773 // TODO: Preserve branch weight metadata, similarly to how
2774 // FoldValueComparisonIntoPredecessors preserves it.
2776 // Figure out which block is which destination.
2777 BasicBlock *DefaultBB = BI->getSuccessor(1);
2778 BasicBlock *EdgeBB = BI->getSuccessor(0);
2779 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2781 BasicBlock *BB = BI->getParent();
2783 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2784 << " cases into SWITCH. BB is:\n" << *BB);
2786 // If there are any extra values that couldn't be folded into the switch
2787 // then we evaluate them with an explicit branch first. Split the block
2788 // right before the condbr to handle it.
2790 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2791 // Remove the uncond branch added to the old block.
2792 TerminatorInst *OldTI = BB->getTerminator();
2793 Builder.SetInsertPoint(OldTI);
2796 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2798 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2800 OldTI->eraseFromParent();
2802 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2803 // for the edge we just added.
2804 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2806 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2807 << "\nEXTRABB = " << *BB);
2811 Builder.SetInsertPoint(BI);
2812 // Convert pointer to int before we switch.
2813 if (CompVal->getType()->isPointerTy()) {
2814 assert(DL && "Cannot switch on pointer without DataLayout");
2815 CompVal = Builder.CreatePtrToInt(CompVal,
2816 DL->getIntPtrType(CompVal->getType()),
2820 // Create the new switch instruction now.
2821 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2823 // Add all of the 'cases' to the switch instruction.
2824 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2825 New->addCase(Values[i], EdgeBB);
2827 // We added edges from PI to the EdgeBB. As such, if there were any
2828 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2829 // the number of edges added.
2830 for (BasicBlock::iterator BBI = EdgeBB->begin();
2831 isa<PHINode>(BBI); ++BBI) {
2832 PHINode *PN = cast<PHINode>(BBI);
2833 Value *InVal = PN->getIncomingValueForBlock(BB);
2834 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2835 PN->addIncoming(InVal, BB);
2838 // Erase the old branch instruction.
2839 EraseTerminatorInstAndDCECond(BI);
2841 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2845 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2846 // If this is a trivial landing pad that just continues unwinding the caught
2847 // exception then zap the landing pad, turning its invokes into calls.
2848 BasicBlock *BB = RI->getParent();
2849 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2850 if (RI->getValue() != LPInst)
2851 // Not a landing pad, or the resume is not unwinding the exception that
2852 // caused control to branch here.
2855 // Check that there are no other instructions except for debug intrinsics.
2856 BasicBlock::iterator I = LPInst, E = RI;
2858 if (!isa<DbgInfoIntrinsic>(I))
2861 // Turn all invokes that unwind here into calls and delete the basic block.
2862 bool InvokeRequiresTableEntry = false;
2863 bool Changed = false;
2864 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2865 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2867 if (II->hasFnAttr(Attribute::UWTable)) {
2868 // Don't remove an `invoke' instruction if the ABI requires an entry into
2870 InvokeRequiresTableEntry = true;
2874 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2876 // Insert a call instruction before the invoke.
2877 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2879 Call->setCallingConv(II->getCallingConv());
2880 Call->setAttributes(II->getAttributes());
2881 Call->setDebugLoc(II->getDebugLoc());
2883 // Anything that used the value produced by the invoke instruction now uses
2884 // the value produced by the call instruction. Note that we do this even
2885 // for void functions and calls with no uses so that the callgraph edge is
2887 II->replaceAllUsesWith(Call);
2888 BB->removePredecessor(II->getParent());
2890 // Insert a branch to the normal destination right before the invoke.
2891 BranchInst::Create(II->getNormalDest(), II);
2893 // Finally, delete the invoke instruction!
2894 II->eraseFromParent();
2898 if (!InvokeRequiresTableEntry)
2899 // The landingpad is now unreachable. Zap it.
2900 BB->eraseFromParent();
2905 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2906 BasicBlock *BB = RI->getParent();
2907 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2909 // Find predecessors that end with branches.
2910 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2911 SmallVector<BranchInst*, 8> CondBranchPreds;
2912 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2913 BasicBlock *P = *PI;
2914 TerminatorInst *PTI = P->getTerminator();
2915 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2916 if (BI->isUnconditional())
2917 UncondBranchPreds.push_back(P);
2919 CondBranchPreds.push_back(BI);
2923 // If we found some, do the transformation!
2924 if (!UncondBranchPreds.empty() && DupRet) {
2925 while (!UncondBranchPreds.empty()) {
2926 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
2927 DEBUG(dbgs() << "FOLDING: " << *BB
2928 << "INTO UNCOND BRANCH PRED: " << *Pred);
2929 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
2932 // If we eliminated all predecessors of the block, delete the block now.
2933 if (pred_begin(BB) == pred_end(BB))
2934 // We know there are no successors, so just nuke the block.
2935 BB->eraseFromParent();
2940 // Check out all of the conditional branches going to this return
2941 // instruction. If any of them just select between returns, change the
2942 // branch itself into a select/return pair.
2943 while (!CondBranchPreds.empty()) {
2944 BranchInst *BI = CondBranchPreds.pop_back_val();
2946 // Check to see if the non-BB successor is also a return block.
2947 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
2948 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
2949 SimplifyCondBranchToTwoReturns(BI, Builder))
2955 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
2956 BasicBlock *BB = UI->getParent();
2958 bool Changed = false;
2960 // If there are any instructions immediately before the unreachable that can
2961 // be removed, do so.
2962 while (UI != BB->begin()) {
2963 BasicBlock::iterator BBI = UI;
2965 // Do not delete instructions that can have side effects which might cause
2966 // the unreachable to not be reachable; specifically, calls and volatile
2967 // operations may have this effect.
2968 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
2970 if (BBI->mayHaveSideEffects()) {
2971 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
2972 if (SI->isVolatile())
2974 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
2975 if (LI->isVolatile())
2977 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
2978 if (RMWI->isVolatile())
2980 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
2981 if (CXI->isVolatile())
2983 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
2984 !isa<LandingPadInst>(BBI)) {
2987 // Note that deleting LandingPad's here is in fact okay, although it
2988 // involves a bit of subtle reasoning. If this inst is a LandingPad,
2989 // all the predecessors of this block will be the unwind edges of Invokes,
2990 // and we can therefore guarantee this block will be erased.
2993 // Delete this instruction (any uses are guaranteed to be dead)
2994 if (!BBI->use_empty())
2995 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
2996 BBI->eraseFromParent();
3000 // If the unreachable instruction is the first in the block, take a gander
3001 // at all of the predecessors of this instruction, and simplify them.
3002 if (&BB->front() != UI) return Changed;
3004 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
3005 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
3006 TerminatorInst *TI = Preds[i]->getTerminator();
3007 IRBuilder<> Builder(TI);
3008 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3009 if (BI->isUnconditional()) {
3010 if (BI->getSuccessor(0) == BB) {
3011 new UnreachableInst(TI->getContext(), TI);
3012 TI->eraseFromParent();
3016 if (BI->getSuccessor(0) == BB) {
3017 Builder.CreateBr(BI->getSuccessor(1));
3018 EraseTerminatorInstAndDCECond(BI);
3019 } else if (BI->getSuccessor(1) == BB) {
3020 Builder.CreateBr(BI->getSuccessor(0));
3021 EraseTerminatorInstAndDCECond(BI);
3025 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3026 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3028 if (i.getCaseSuccessor() == BB) {
3029 BB->removePredecessor(SI->getParent());
3034 // If the default value is unreachable, figure out the most popular
3035 // destination and make it the default.
3036 if (SI->getDefaultDest() == BB) {
3037 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
3038 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3040 std::pair<unsigned, unsigned> &entry =
3041 Popularity[i.getCaseSuccessor()];
3042 if (entry.first == 0) {
3044 entry.second = i.getCaseIndex();
3050 // Find the most popular block.
3051 unsigned MaxPop = 0;
3052 unsigned MaxIndex = 0;
3053 BasicBlock *MaxBlock = 0;
3054 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
3055 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
3056 if (I->second.first > MaxPop ||
3057 (I->second.first == MaxPop && MaxIndex > I->second.second)) {
3058 MaxPop = I->second.first;
3059 MaxIndex = I->second.second;
3060 MaxBlock = I->first;
3064 // Make this the new default, allowing us to delete any explicit
3066 SI->setDefaultDest(MaxBlock);
3069 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
3071 if (isa<PHINode>(MaxBlock->begin()))
3072 for (unsigned i = 0; i != MaxPop-1; ++i)
3073 MaxBlock->removePredecessor(SI->getParent());
3075 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3077 if (i.getCaseSuccessor() == MaxBlock) {
3083 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
3084 if (II->getUnwindDest() == BB) {
3085 // Convert the invoke to a call instruction. This would be a good
3086 // place to note that the call does not throw though.
3087 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
3088 II->removeFromParent(); // Take out of symbol table
3090 // Insert the call now...
3091 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
3092 Builder.SetInsertPoint(BI);
3093 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
3094 Args, II->getName());
3095 CI->setCallingConv(II->getCallingConv());
3096 CI->setAttributes(II->getAttributes());
3097 // If the invoke produced a value, the call does now instead.
3098 II->replaceAllUsesWith(CI);
3105 // If this block is now dead, remove it.
3106 if (pred_begin(BB) == pred_end(BB) &&
3107 BB != &BB->getParent()->getEntryBlock()) {
3108 // We know there are no successors, so just nuke the block.
3109 BB->eraseFromParent();
3116 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
3117 /// integer range comparison into a sub, an icmp and a branch.
3118 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3119 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3121 // Make sure all cases point to the same destination and gather the values.
3122 SmallVector<ConstantInt *, 16> Cases;
3123 SwitchInst::CaseIt I = SI->case_begin();
3124 Cases.push_back(I.getCaseValue());
3125 SwitchInst::CaseIt PrevI = I++;
3126 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
3127 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
3129 Cases.push_back(I.getCaseValue());
3131 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
3133 // Sort the case values, then check if they form a range we can transform.
3134 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3135 for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
3136 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
3140 Constant *Offset = ConstantExpr::getNeg(Cases.back());
3141 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
3143 Value *Sub = SI->getCondition();
3144 if (!Offset->isNullValue())
3145 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
3147 // If NumCases overflowed, then all possible values jump to the successor.
3148 if (NumCases->isNullValue() && SI->getNumCases() != 0)
3149 Cmp = ConstantInt::getTrue(SI->getContext());
3151 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3152 BranchInst *NewBI = Builder.CreateCondBr(
3153 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
3155 // Update weight for the newly-created conditional branch.
3156 SmallVector<uint64_t, 8> Weights;
3157 bool HasWeights = HasBranchWeights(SI);
3159 GetBranchWeights(SI, Weights);
3160 if (Weights.size() == 1 + SI->getNumCases()) {
3161 // Combine all weights for the cases to be the true weight of NewBI.
3162 // We assume that the sum of all weights for a Terminator can fit into 32
3164 uint32_t NewTrueWeight = 0;
3165 for (unsigned I = 1, E = Weights.size(); I != E; ++I)
3166 NewTrueWeight += (uint32_t)Weights[I];
3167 NewBI->setMetadata(LLVMContext::MD_prof,
3168 MDBuilder(SI->getContext()).
3169 createBranchWeights(NewTrueWeight,
3170 (uint32_t)Weights[0]));
3174 // Prune obsolete incoming values off the successor's PHI nodes.
3175 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
3176 isa<PHINode>(BBI); ++BBI) {
3177 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
3178 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3180 SI->eraseFromParent();
3185 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
3186 /// and use it to remove dead cases.
3187 static bool EliminateDeadSwitchCases(SwitchInst *SI) {
3188 Value *Cond = SI->getCondition();
3189 unsigned Bits = Cond->getType()->getIntegerBitWidth();
3190 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3191 ComputeMaskedBits(Cond, KnownZero, KnownOne);
3193 // Gather dead cases.
3194 SmallVector<ConstantInt*, 8> DeadCases;
3195 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3196 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3197 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3198 DeadCases.push_back(I.getCaseValue());
3199 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3200 << I.getCaseValue() << "' is dead.\n");
3204 SmallVector<uint64_t, 8> Weights;
3205 bool HasWeight = HasBranchWeights(SI);
3207 GetBranchWeights(SI, Weights);
3208 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3211 // Remove dead cases from the switch.
3212 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3213 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3214 assert(Case != SI->case_default() &&
3215 "Case was not found. Probably mistake in DeadCases forming.");
3217 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3221 // Prune unused values from PHI nodes.
3222 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3223 SI->removeCase(Case);
3225 if (HasWeight && Weights.size() >= 2) {
3226 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3227 SI->setMetadata(LLVMContext::MD_prof,
3228 MDBuilder(SI->getParent()->getContext()).
3229 createBranchWeights(MDWeights));
3232 return !DeadCases.empty();
3235 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
3236 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3237 /// by an unconditional branch), look at the phi node for BB in the successor
3238 /// block and see if the incoming value is equal to CaseValue. If so, return
3239 /// the phi node, and set PhiIndex to BB's index in the phi node.
3240 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3243 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3244 return NULL; // BB must be empty to be a candidate for simplification.
3245 if (!BB->getSinglePredecessor())
3246 return NULL; // BB must be dominated by the switch.
3248 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3249 if (!Branch || !Branch->isUnconditional())
3250 return NULL; // Terminator must be unconditional branch.
3252 BasicBlock *Succ = Branch->getSuccessor(0);
3254 BasicBlock::iterator I = Succ->begin();
3255 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3256 int Idx = PHI->getBasicBlockIndex(BB);
3257 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3259 Value *InValue = PHI->getIncomingValue(Idx);
3260 if (InValue != CaseValue) continue;
3269 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
3270 /// instruction to a phi node dominated by the switch, if that would mean that
3271 /// some of the destination blocks of the switch can be folded away.
3272 /// Returns true if a change is made.
3273 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3274 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3275 ForwardingNodesMap ForwardingNodes;
3277 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3278 ConstantInt *CaseValue = I.getCaseValue();
3279 BasicBlock *CaseDest = I.getCaseSuccessor();
3282 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3286 ForwardingNodes[PHI].push_back(PhiIndex);
3289 bool Changed = false;
3291 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3292 E = ForwardingNodes.end(); I != E; ++I) {
3293 PHINode *Phi = I->first;
3294 SmallVectorImpl<int> &Indexes = I->second;
3296 if (Indexes.size() < 2) continue;
3298 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3299 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3306 /// ValidLookupTableConstant - Return true if the backend will be able to handle
3307 /// initializing an array of constants like C.
3308 static bool ValidLookupTableConstant(Constant *C) {
3309 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3310 return CE->isGEPWithNoNotionalOverIndexing();
3312 return isa<ConstantFP>(C) ||
3313 isa<ConstantInt>(C) ||
3314 isa<ConstantPointerNull>(C) ||
3315 isa<GlobalValue>(C) ||
3319 /// LookupConstant - If V is a Constant, return it. Otherwise, try to look up
3320 /// its constant value in ConstantPool, returning 0 if it's not there.
3321 static Constant *LookupConstant(Value *V,
3322 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3323 if (Constant *C = dyn_cast<Constant>(V))
3325 return ConstantPool.lookup(V);
3328 /// ConstantFold - Try to fold instruction I into a constant. This works for
3329 /// simple instructions such as binary operations where both operands are
3330 /// constant or can be replaced by constants from the ConstantPool. Returns the
3331 /// resulting constant on success, 0 otherwise.
3333 ConstantFold(Instruction *I,
3334 const SmallDenseMap<Value *, Constant *> &ConstantPool,
3335 const DataLayout *DL) {
3336 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3337 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3340 if (A->isAllOnesValue())
3341 return LookupConstant(Select->getTrueValue(), ConstantPool);
3342 if (A->isNullValue())
3343 return LookupConstant(Select->getFalseValue(), ConstantPool);
3347 SmallVector<Constant *, 4> COps;
3348 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
3349 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
3355 if (CmpInst *Cmp = dyn_cast<CmpInst>(I))
3356 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
3359 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), COps, DL);
3362 /// GetCaseResults - Try to determine the resulting constant values in phi nodes
3363 /// at the common destination basic block, *CommonDest, for one of the case
3364 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3365 /// case), of a switch instruction SI.
3367 GetCaseResults(SwitchInst *SI,
3368 ConstantInt *CaseVal,
3369 BasicBlock *CaseDest,
3370 BasicBlock **CommonDest,
3371 SmallVectorImpl<std::pair<PHINode *, Constant *> > &Res,
3372 const DataLayout *DL) {
3373 // The block from which we enter the common destination.
3374 BasicBlock *Pred = SI->getParent();
3376 // If CaseDest is empty except for some side-effect free instructions through
3377 // which we can constant-propagate the CaseVal, continue to its successor.
3378 SmallDenseMap<Value*, Constant*> ConstantPool;
3379 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3380 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3382 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3383 // If the terminator is a simple branch, continue to the next block.
3384 if (T->getNumSuccessors() != 1)
3387 CaseDest = T->getSuccessor(0);
3388 } else if (isa<DbgInfoIntrinsic>(I)) {
3389 // Skip debug intrinsic.
3391 } else if (Constant *C = ConstantFold(I, ConstantPool, DL)) {
3392 // Instruction is side-effect free and constant.
3393 ConstantPool.insert(std::make_pair(I, C));
3399 // If we did not have a CommonDest before, use the current one.
3401 *CommonDest = CaseDest;
3402 // If the destination isn't the common one, abort.
3403 if (CaseDest != *CommonDest)
3406 // Get the values for this case from phi nodes in the destination block.
3407 BasicBlock::iterator I = (*CommonDest)->begin();
3408 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3409 int Idx = PHI->getBasicBlockIndex(Pred);
3413 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3418 // Note: If the constant comes from constant-propagating the case value
3419 // through the CaseDest basic block, it will be safe to remove the
3420 // instructions in that block. They cannot be used (except in the phi nodes
3421 // we visit) outside CaseDest, because that block does not dominate its
3422 // successor. If it did, we would not be in this phi node.
3424 // Be conservative about which kinds of constants we support.
3425 if (!ValidLookupTableConstant(ConstVal))
3428 Res.push_back(std::make_pair(PHI, ConstVal));
3431 return Res.size() > 0;
3435 /// SwitchLookupTable - This class represents a lookup table that can be used
3436 /// to replace a switch.
3437 class SwitchLookupTable {
3439 /// SwitchLookupTable - Create a lookup table to use as a switch replacement
3440 /// with the contents of Values, using DefaultValue to fill any holes in the
3442 SwitchLookupTable(Module &M,
3444 ConstantInt *Offset,
3445 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3446 Constant *DefaultValue,
3447 const DataLayout *DL);
3449 /// BuildLookup - Build instructions with Builder to retrieve the value at
3450 /// the position given by Index in the lookup table.
3451 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3453 /// WouldFitInRegister - Return true if a table with TableSize elements of
3454 /// type ElementType would fit in a target-legal register.
3455 static bool WouldFitInRegister(const DataLayout *DL,
3457 const Type *ElementType);
3460 // Depending on the contents of the table, it can be represented in
3463 // For tables where each element contains the same value, we just have to
3464 // store that single value and return it for each lookup.
3467 // For small tables with integer elements, we can pack them into a bitmap
3468 // that fits into a target-legal register. Values are retrieved by
3469 // shift and mask operations.
3472 // The table is stored as an array of values. Values are retrieved by load
3473 // instructions from the table.
3477 // For SingleValueKind, this is the single value.
3478 Constant *SingleValue;
3480 // For BitMapKind, this is the bitmap.
3481 ConstantInt *BitMap;
3482 IntegerType *BitMapElementTy;
3484 // For ArrayKind, this is the array.
3485 GlobalVariable *Array;
3489 SwitchLookupTable::SwitchLookupTable(Module &M,
3491 ConstantInt *Offset,
3492 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3493 Constant *DefaultValue,
3494 const DataLayout *DL)
3495 : SingleValue(0), BitMap(0), BitMapElementTy(0), Array(0) {
3496 assert(Values.size() && "Can't build lookup table without values!");
3497 assert(TableSize >= Values.size() && "Can't fit values in table!");
3499 // If all values in the table are equal, this is that value.
3500 SingleValue = Values.begin()->second;
3502 Type *ValueType = Values.begin()->second->getType();
3504 // Build up the table contents.
3505 SmallVector<Constant*, 64> TableContents(TableSize);
3506 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3507 ConstantInt *CaseVal = Values[I].first;
3508 Constant *CaseRes = Values[I].second;
3509 assert(CaseRes->getType() == ValueType);
3511 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3513 TableContents[Idx] = CaseRes;
3515 if (CaseRes != SingleValue)
3519 // Fill in any holes in the table with the default result.
3520 if (Values.size() < TableSize) {
3521 assert(DefaultValue && "Need a default value to fill the lookup table holes.");
3522 assert(DefaultValue->getType() == ValueType);
3523 for (uint64_t I = 0; I < TableSize; ++I) {
3524 if (!TableContents[I])
3525 TableContents[I] = DefaultValue;
3528 if (DefaultValue != SingleValue)
3532 // If each element in the table contains the same value, we only need to store
3533 // that single value.
3535 Kind = SingleValueKind;
3539 // If the type is integer and the table fits in a register, build a bitmap.
3540 if (WouldFitInRegister(DL, TableSize, ValueType)) {
3541 IntegerType *IT = cast<IntegerType>(ValueType);
3542 APInt TableInt(TableSize * IT->getBitWidth(), 0);
3543 for (uint64_t I = TableSize; I > 0; --I) {
3544 TableInt <<= IT->getBitWidth();
3545 // Insert values into the bitmap. Undef values are set to zero.
3546 if (!isa<UndefValue>(TableContents[I - 1])) {
3547 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
3548 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
3551 BitMap = ConstantInt::get(M.getContext(), TableInt);
3552 BitMapElementTy = IT;
3558 // Store the table in an array.
3559 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
3560 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3562 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3563 GlobalVariable::PrivateLinkage,
3566 Array->setUnnamedAddr(true);
3570 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
3572 case SingleValueKind:
3575 // Type of the bitmap (e.g. i59).
3576 IntegerType *MapTy = BitMap->getType();
3578 // Cast Index to the same type as the bitmap.
3579 // Note: The Index is <= the number of elements in the table, so
3580 // truncating it to the width of the bitmask is safe.
3581 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
3583 // Multiply the shift amount by the element width.
3584 ShiftAmt = Builder.CreateMul(ShiftAmt,
3585 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
3589 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
3590 "switch.downshift");
3592 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
3596 Value *GEPIndices[] = { Builder.getInt32(0), Index };
3597 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices,
3599 return Builder.CreateLoad(GEP, "switch.load");
3602 llvm_unreachable("Unknown lookup table kind!");
3605 bool SwitchLookupTable::WouldFitInRegister(const DataLayout *DL,
3607 const Type *ElementType) {
3610 const IntegerType *IT = dyn_cast<IntegerType>(ElementType);
3613 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
3614 // are <= 15, we could try to narrow the type.
3616 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
3617 if (TableSize >= UINT_MAX/IT->getBitWidth())
3619 return DL->fitsInLegalInteger(TableSize * IT->getBitWidth());
3622 /// ShouldBuildLookupTable - Determine whether a lookup table should be built
3623 /// for this switch, based on the number of cases, size of the table and the
3624 /// types of the results.
3625 static bool ShouldBuildLookupTable(SwitchInst *SI,
3627 const TargetTransformInfo &TTI,
3628 const DataLayout *DL,
3629 const SmallDenseMap<PHINode*, Type*>& ResultTypes) {
3630 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
3631 return false; // TableSize overflowed, or mul below might overflow.
3633 bool AllTablesFitInRegister = true;
3634 bool HasIllegalType = false;
3635 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(),
3636 E = ResultTypes.end(); I != E; ++I) {
3637 Type *Ty = I->second;
3639 // Saturate this flag to true.
3640 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
3642 // Saturate this flag to false.
3643 AllTablesFitInRegister = AllTablesFitInRegister &&
3644 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
3646 // If both flags saturate, we're done. NOTE: This *only* works with
3647 // saturating flags, and all flags have to saturate first due to the
3648 // non-deterministic behavior of iterating over a dense map.
3649 if (HasIllegalType && !AllTablesFitInRegister)
3653 // If each table would fit in a register, we should build it anyway.
3654 if (AllTablesFitInRegister)
3657 // Don't build a table that doesn't fit in-register if it has illegal types.
3661 // The table density should be at least 40%. This is the same criterion as for
3662 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3663 // FIXME: Find the best cut-off.
3664 return SI->getNumCases() * 10 >= TableSize * 4;
3667 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3668 /// phi nodes in a common successor block with different constant values,
3669 /// replace the switch with lookup tables.
3670 static bool SwitchToLookupTable(SwitchInst *SI,
3671 IRBuilder<> &Builder,
3672 const TargetTransformInfo &TTI,
3673 const DataLayout* DL) {
3674 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3676 // Only build lookup table when we have a target that supports it.
3677 if (!TTI.shouldBuildLookupTables())
3680 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3681 // split off a dense part and build a lookup table for that.
3683 // FIXME: This creates arrays of GEPs to constant strings, which means each
3684 // GEP needs a runtime relocation in PIC code. We should just build one big
3685 // string and lookup indices into that.
3687 // Ignore switches with less than three cases. Lookup tables will not make them
3688 // faster, so we don't analyze them.
3689 if (SI->getNumCases() < 3)
3692 // Figure out the corresponding result for each case value and phi node in the
3693 // common destination, as well as the the min and max case values.
3694 assert(SI->case_begin() != SI->case_end());
3695 SwitchInst::CaseIt CI = SI->case_begin();
3696 ConstantInt *MinCaseVal = CI.getCaseValue();
3697 ConstantInt *MaxCaseVal = CI.getCaseValue();
3699 BasicBlock *CommonDest = 0;
3700 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
3701 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
3702 SmallDenseMap<PHINode*, Constant*> DefaultResults;
3703 SmallDenseMap<PHINode*, Type*> ResultTypes;
3704 SmallVector<PHINode*, 4> PHIs;
3706 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
3707 ConstantInt *CaseVal = CI.getCaseValue();
3708 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
3709 MinCaseVal = CaseVal;
3710 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
3711 MaxCaseVal = CaseVal;
3713 // Resulting value at phi nodes for this case value.
3714 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
3716 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
3720 // Append the result from this case to the list for each phi.
3721 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) {
3722 if (!ResultLists.count(I->first))
3723 PHIs.push_back(I->first);
3724 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second));
3728 // Keep track of the result types.
3729 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
3730 PHINode *PHI = PHIs[I];
3731 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
3734 uint64_t NumResults = ResultLists[PHIs[0]].size();
3735 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
3736 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
3737 bool TableHasHoles = (NumResults < TableSize);
3739 // If the table has holes, we need a constant result for the default case
3740 // or a bitmask that fits in a register.
3741 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
3742 bool HasDefaultResults = false;
3743 if (TableHasHoles) {
3744 HasDefaultResults = GetCaseResults(SI, 0, SI->getDefaultDest(), &CommonDest,
3745 DefaultResultsList, DL);
3747 bool NeedMask = (TableHasHoles && !HasDefaultResults);
3749 // As an extra penalty for the validity test we require more cases.
3750 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
3752 if (!(DL && DL->fitsInLegalInteger(TableSize)))
3756 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) {
3757 PHINode *PHI = DefaultResultsList[I].first;
3758 Constant *Result = DefaultResultsList[I].second;
3759 DefaultResults[PHI] = Result;
3762 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
3765 // Create the BB that does the lookups.
3766 Module &Mod = *CommonDest->getParent()->getParent();
3767 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
3769 CommonDest->getParent(),
3772 // Compute the table index value.
3773 Builder.SetInsertPoint(SI);
3774 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
3777 // Compute the maximum table size representable by the integer type we are
3779 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
3780 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
3781 assert(MaxTableSize >= TableSize &&
3782 "It is impossible for a switch to have more entries than the max "
3783 "representable value of its input integer type's size.");
3785 // If we have a fully covered lookup table, unconditionally branch to the
3786 // lookup table BB. Otherwise, check if the condition value is within the case
3787 // range. If it is so, branch to the new BB. Otherwise branch to SI's default
3789 const bool GeneratingCoveredLookupTable = MaxTableSize == TableSize;
3790 if (GeneratingCoveredLookupTable) {
3791 Builder.CreateBr(LookupBB);
3792 SI->getDefaultDest()->removePredecessor(SI->getParent());
3794 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
3795 MinCaseVal->getType(), TableSize));
3796 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
3799 // Populate the BB that does the lookups.
3800 Builder.SetInsertPoint(LookupBB);
3803 // Before doing the lookup we do the hole check.
3804 // The LookupBB is therefore re-purposed to do the hole check
3805 // and we create a new LookupBB.
3806 BasicBlock *MaskBB = LookupBB;
3807 MaskBB->setName("switch.hole_check");
3808 LookupBB = BasicBlock::Create(Mod.getContext(),
3810 CommonDest->getParent(),
3813 // Build bitmask; fill in a 1 bit for every case.
3814 APInt MaskInt(TableSize, 0);
3815 APInt One(TableSize, 1);
3816 const ResultListTy &ResultList = ResultLists[PHIs[0]];
3817 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
3818 uint64_t Idx = (ResultList[I].first->getValue() -
3819 MinCaseVal->getValue()).getLimitedValue();
3820 MaskInt |= One << Idx;
3822 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
3824 // Get the TableIndex'th bit of the bitmask.
3825 // If this bit is 0 (meaning hole) jump to the default destination,
3826 // else continue with table lookup.
3827 IntegerType *MapTy = TableMask->getType();
3828 Value *MaskIndex = Builder.CreateZExtOrTrunc(TableIndex, MapTy,
3829 "switch.maskindex");
3830 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex,
3832 Value *LoBit = Builder.CreateTrunc(Shifted,
3833 Type::getInt1Ty(Mod.getContext()),
3835 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
3837 Builder.SetInsertPoint(LookupBB);
3838 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
3841 bool ReturnedEarly = false;
3842 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
3843 PHINode *PHI = PHIs[I];
3845 // If using a bitmask, use any value to fill the lookup table holes.
3846 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
3847 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI],
3850 Value *Result = Table.BuildLookup(TableIndex, Builder);
3852 // If the result is used to return immediately from the function, we want to
3853 // do that right here.
3854 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
3855 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
3856 Builder.CreateRet(Result);
3857 ReturnedEarly = true;
3861 PHI->addIncoming(Result, LookupBB);
3865 Builder.CreateBr(CommonDest);
3867 // Remove the switch.
3868 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
3869 BasicBlock *Succ = SI->getSuccessor(i);
3871 if (Succ == SI->getDefaultDest())
3873 Succ->removePredecessor(SI->getParent());
3875 SI->eraseFromParent();
3879 ++NumLookupTablesHoles;
3883 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
3884 BasicBlock *BB = SI->getParent();
3886 if (isValueEqualityComparison(SI)) {
3887 // If we only have one predecessor, and if it is a branch on this value,
3888 // see if that predecessor totally determines the outcome of this switch.
3889 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3890 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
3891 return SimplifyCFG(BB, TTI, DL) | true;
3893 Value *Cond = SI->getCondition();
3894 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
3895 if (SimplifySwitchOnSelect(SI, Select))
3896 return SimplifyCFG(BB, TTI, DL) | true;
3898 // If the block only contains the switch, see if we can fold the block
3899 // away into any preds.
3900 BasicBlock::iterator BBI = BB->begin();
3901 // Ignore dbg intrinsics.
3902 while (isa<DbgInfoIntrinsic>(BBI))
3905 if (FoldValueComparisonIntoPredecessors(SI, Builder))
3906 return SimplifyCFG(BB, TTI, DL) | true;
3909 // Try to transform the switch into an icmp and a branch.
3910 if (TurnSwitchRangeIntoICmp(SI, Builder))
3911 return SimplifyCFG(BB, TTI, DL) | true;
3913 // Remove unreachable cases.
3914 if (EliminateDeadSwitchCases(SI))
3915 return SimplifyCFG(BB, TTI, DL) | true;
3917 if (ForwardSwitchConditionToPHI(SI))
3918 return SimplifyCFG(BB, TTI, DL) | true;
3920 if (SwitchToLookupTable(SI, Builder, TTI, DL))
3921 return SimplifyCFG(BB, TTI, DL) | true;
3926 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
3927 BasicBlock *BB = IBI->getParent();
3928 bool Changed = false;
3930 // Eliminate redundant destinations.
3931 SmallPtrSet<Value *, 8> Succs;
3932 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
3933 BasicBlock *Dest = IBI->getDestination(i);
3934 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) {
3935 Dest->removePredecessor(BB);
3936 IBI->removeDestination(i);
3942 if (IBI->getNumDestinations() == 0) {
3943 // If the indirectbr has no successors, change it to unreachable.
3944 new UnreachableInst(IBI->getContext(), IBI);
3945 EraseTerminatorInstAndDCECond(IBI);
3949 if (IBI->getNumDestinations() == 1) {
3950 // If the indirectbr has one successor, change it to a direct branch.
3951 BranchInst::Create(IBI->getDestination(0), IBI);
3952 EraseTerminatorInstAndDCECond(IBI);
3956 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
3957 if (SimplifyIndirectBrOnSelect(IBI, SI))
3958 return SimplifyCFG(BB, TTI, DL) | true;
3963 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
3964 BasicBlock *BB = BI->getParent();
3966 if (SinkCommon && SinkThenElseCodeToEnd(BI))
3969 // If the Terminator is the only non-phi instruction, simplify the block.
3970 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime();
3971 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
3972 TryToSimplifyUncondBranchFromEmptyBlock(BB))
3975 // If the only instruction in the block is a seteq/setne comparison
3976 // against a constant, try to simplify the block.
3977 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
3978 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
3979 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
3981 if (I->isTerminator() &&
3982 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI, DL))
3986 // If this basic block is ONLY a compare and a branch, and if a predecessor
3987 // branches to us and our successor, fold the comparison into the
3988 // predecessor and use logical operations to update the incoming value
3989 // for PHI nodes in common successor.
3990 if (FoldBranchToCommonDest(BI))
3991 return SimplifyCFG(BB, TTI, DL) | true;
3996 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
3997 BasicBlock *BB = BI->getParent();
3999 // Conditional branch
4000 if (isValueEqualityComparison(BI)) {
4001 // If we only have one predecessor, and if it is a branch on this value,
4002 // see if that predecessor totally determines the outcome of this
4004 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4005 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
4006 return SimplifyCFG(BB, TTI, DL) | true;
4008 // This block must be empty, except for the setcond inst, if it exists.
4009 // Ignore dbg intrinsics.
4010 BasicBlock::iterator I = BB->begin();
4011 // Ignore dbg intrinsics.
4012 while (isa<DbgInfoIntrinsic>(I))
4015 if (FoldValueComparisonIntoPredecessors(BI, Builder))
4016 return SimplifyCFG(BB, TTI, DL) | true;
4017 } else if (&*I == cast<Instruction>(BI->getCondition())){
4019 // Ignore dbg intrinsics.
4020 while (isa<DbgInfoIntrinsic>(I))
4022 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
4023 return SimplifyCFG(BB, TTI, DL) | true;
4027 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
4028 if (SimplifyBranchOnICmpChain(BI, DL, Builder))
4031 // If this basic block is ONLY a compare and a branch, and if a predecessor
4032 // branches to us and one of our successors, fold the comparison into the
4033 // predecessor and use logical operations to pick the right destination.
4034 if (FoldBranchToCommonDest(BI))
4035 return SimplifyCFG(BB, TTI, DL) | true;
4037 // We have a conditional branch to two blocks that are only reachable
4038 // from BI. We know that the condbr dominates the two blocks, so see if
4039 // there is any identical code in the "then" and "else" blocks. If so, we
4040 // can hoist it up to the branching block.
4041 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) {
4042 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
4043 if (HoistThenElseCodeToIf(BI))
4044 return SimplifyCFG(BB, TTI, DL) | true;
4046 // If Successor #1 has multiple preds, we may be able to conditionally
4047 // execute Successor #0 if it branches to successor #1.
4048 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
4049 if (Succ0TI->getNumSuccessors() == 1 &&
4050 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
4051 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0)))
4052 return SimplifyCFG(BB, TTI, DL) | true;
4054 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
4055 // If Successor #0 has multiple preds, we may be able to conditionally
4056 // execute Successor #1 if it branches to successor #0.
4057 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
4058 if (Succ1TI->getNumSuccessors() == 1 &&
4059 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
4060 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1)))
4061 return SimplifyCFG(BB, TTI, DL) | true;
4064 // If this is a branch on a phi node in the current block, thread control
4065 // through this block if any PHI node entries are constants.
4066 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
4067 if (PN->getParent() == BI->getParent())
4068 if (FoldCondBranchOnPHI(BI, DL))
4069 return SimplifyCFG(BB, TTI, DL) | true;
4071 // Scan predecessor blocks for conditional branches.
4072 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
4073 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
4074 if (PBI != BI && PBI->isConditional())
4075 if (SimplifyCondBranchToCondBranch(PBI, BI))
4076 return SimplifyCFG(BB, TTI, DL) | true;
4081 /// Check if passing a value to an instruction will cause undefined behavior.
4082 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
4083 Constant *C = dyn_cast<Constant>(V);
4090 if (C->isNullValue()) {
4091 // Only look at the first use, avoid hurting compile time with long uselists
4092 User *Use = *I->user_begin();
4094 // Now make sure that there are no instructions in between that can alter
4095 // control flow (eg. calls)
4096 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
4097 if (i == I->getParent()->end() || i->mayHaveSideEffects())
4100 // Look through GEPs. A load from a GEP derived from NULL is still undefined
4101 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
4102 if (GEP->getPointerOperand() == I)
4103 return passingValueIsAlwaysUndefined(V, GEP);
4105 // Look through bitcasts.
4106 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
4107 return passingValueIsAlwaysUndefined(V, BC);
4109 // Load from null is undefined.
4110 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
4111 if (!LI->isVolatile())
4112 return LI->getPointerAddressSpace() == 0;
4114 // Store to null is undefined.
4115 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
4116 if (!SI->isVolatile())
4117 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
4122 /// If BB has an incoming value that will always trigger undefined behavior
4123 /// (eg. null pointer dereference), remove the branch leading here.
4124 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
4125 for (BasicBlock::iterator i = BB->begin();
4126 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
4127 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
4128 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
4129 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
4130 IRBuilder<> Builder(T);
4131 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
4132 BB->removePredecessor(PHI->getIncomingBlock(i));
4133 // Turn uncoditional branches into unreachables and remove the dead
4134 // destination from conditional branches.
4135 if (BI->isUnconditional())
4136 Builder.CreateUnreachable();
4138 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
4139 BI->getSuccessor(0));
4140 BI->eraseFromParent();
4143 // TODO: SwitchInst.
4149 bool SimplifyCFGOpt::run(BasicBlock *BB) {
4150 bool Changed = false;
4152 assert(BB && BB->getParent() && "Block not embedded in function!");
4153 assert(BB->getTerminator() && "Degenerate basic block encountered!");
4155 // Remove basic blocks that have no predecessors (except the entry block)...
4156 // or that just have themself as a predecessor. These are unreachable.
4157 if ((pred_begin(BB) == pred_end(BB) &&
4158 BB != &BB->getParent()->getEntryBlock()) ||
4159 BB->getSinglePredecessor() == BB) {
4160 DEBUG(dbgs() << "Removing BB: \n" << *BB);
4161 DeleteDeadBlock(BB);
4165 // Check to see if we can constant propagate this terminator instruction
4167 Changed |= ConstantFoldTerminator(BB, true);
4169 // Check for and eliminate duplicate PHI nodes in this block.
4170 Changed |= EliminateDuplicatePHINodes(BB);
4172 // Check for and remove branches that will always cause undefined behavior.
4173 Changed |= removeUndefIntroducingPredecessor(BB);
4175 // Merge basic blocks into their predecessor if there is only one distinct
4176 // pred, and if there is only one distinct successor of the predecessor, and
4177 // if there are no PHI nodes.
4179 if (MergeBlockIntoPredecessor(BB))
4182 IRBuilder<> Builder(BB);
4184 // If there is a trivial two-entry PHI node in this basic block, and we can
4185 // eliminate it, do so now.
4186 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
4187 if (PN->getNumIncomingValues() == 2)
4188 Changed |= FoldTwoEntryPHINode(PN, DL);
4190 Builder.SetInsertPoint(BB->getTerminator());
4191 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
4192 if (BI->isUnconditional()) {
4193 if (SimplifyUncondBranch(BI, Builder)) return true;
4195 if (SimplifyCondBranch(BI, Builder)) return true;
4197 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
4198 if (SimplifyReturn(RI, Builder)) return true;
4199 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
4200 if (SimplifyResume(RI, Builder)) return true;
4201 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
4202 if (SimplifySwitch(SI, Builder)) return true;
4203 } else if (UnreachableInst *UI =
4204 dyn_cast<UnreachableInst>(BB->getTerminator())) {
4205 if (SimplifyUnreachable(UI)) return true;
4206 } else if (IndirectBrInst *IBI =
4207 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
4208 if (SimplifyIndirectBr(IBI)) return true;
4214 /// SimplifyCFG - This function is used to do simplification of a CFG. For
4215 /// example, it adjusts branches to branches to eliminate the extra hop, it
4216 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4217 /// of the CFG. It returns true if a modification was made.
4219 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
4220 const DataLayout *DL) {
4221 return SimplifyCFGOpt(TTI, DL).run(BB);