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(NumSinkCommons, "Number of common instructions sunk down to the end block");
72 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
75 /// ValueEqualityComparisonCase - Represents a case of a switch.
76 struct ValueEqualityComparisonCase {
80 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
81 : Value(Value), Dest(Dest) {}
83 bool operator<(ValueEqualityComparisonCase RHS) const {
84 // Comparing pointers is ok as we only rely on the order for uniquing.
85 return Value < RHS.Value;
88 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
91 class SimplifyCFGOpt {
92 const TargetTransformInfo &TTI;
93 const DataLayout *const DL;
94 Value *isValueEqualityComparison(TerminatorInst *TI);
95 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
96 std::vector<ValueEqualityComparisonCase> &Cases);
97 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
99 IRBuilder<> &Builder);
100 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
101 IRBuilder<> &Builder);
103 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
104 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
105 bool SimplifyUnreachable(UnreachableInst *UI);
106 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
107 bool SimplifyIndirectBr(IndirectBrInst *IBI);
108 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder);
109 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder);
112 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout *DL)
113 : TTI(TTI), DL(DL) {}
114 bool run(BasicBlock *BB);
118 /// SafeToMergeTerminators - Return true if it is safe to merge these two
119 /// terminator instructions together.
121 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
122 if (SI1 == SI2) return false; // Can't merge with self!
124 // It is not safe to merge these two switch instructions if they have a common
125 // successor, and if that successor has a PHI node, and if *that* PHI node has
126 // conflicting incoming values from the two switch blocks.
127 BasicBlock *SI1BB = SI1->getParent();
128 BasicBlock *SI2BB = SI2->getParent();
129 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
131 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
132 if (SI1Succs.count(*I))
133 for (BasicBlock::iterator BBI = (*I)->begin();
134 isa<PHINode>(BBI); ++BBI) {
135 PHINode *PN = cast<PHINode>(BBI);
136 if (PN->getIncomingValueForBlock(SI1BB) !=
137 PN->getIncomingValueForBlock(SI2BB))
144 /// isProfitableToFoldUnconditional - Return true if it is safe and profitable
145 /// to merge these two terminator instructions together, where SI1 is an
146 /// unconditional branch. PhiNodes will store all PHI nodes in common
149 static bool isProfitableToFoldUnconditional(BranchInst *SI1,
152 SmallVectorImpl<PHINode*> &PhiNodes) {
153 if (SI1 == SI2) return false; // Can't merge with self!
154 assert(SI1->isUnconditional() && SI2->isConditional());
156 // We fold the unconditional branch if we can easily update all PHI nodes in
157 // common successors:
158 // 1> We have a constant incoming value for the conditional branch;
159 // 2> We have "Cond" as the incoming value for the unconditional branch;
160 // 3> SI2->getCondition() and Cond have same operands.
161 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
162 if (!Ci2) return false;
163 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
164 Cond->getOperand(1) == Ci2->getOperand(1)) &&
165 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
166 Cond->getOperand(1) == Ci2->getOperand(0)))
169 BasicBlock *SI1BB = SI1->getParent();
170 BasicBlock *SI2BB = SI2->getParent();
171 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
172 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
173 if (SI1Succs.count(*I))
174 for (BasicBlock::iterator BBI = (*I)->begin();
175 isa<PHINode>(BBI); ++BBI) {
176 PHINode *PN = cast<PHINode>(BBI);
177 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
178 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
180 PhiNodes.push_back(PN);
185 /// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will
186 /// now be entries in it from the 'NewPred' block. The values that will be
187 /// flowing into the PHI nodes will be the same as those coming in from
188 /// ExistPred, an existing predecessor of Succ.
189 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
190 BasicBlock *ExistPred) {
191 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
194 for (BasicBlock::iterator I = Succ->begin();
195 (PN = dyn_cast<PHINode>(I)); ++I)
196 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
199 /// ComputeSpeculationCost - Compute an abstract "cost" of speculating the
200 /// given instruction, which is assumed to be safe to speculate. 1 means
201 /// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive.
202 static unsigned ComputeSpeculationCost(const User *I) {
203 assert(isSafeToSpeculativelyExecute(I) &&
204 "Instruction is not safe to speculatively execute!");
205 switch (Operator::getOpcode(I)) {
207 // In doubt, be conservative.
209 case Instruction::GetElementPtr:
210 // GEPs are cheap if all indices are constant.
211 if (!cast<GEPOperator>(I)->hasAllConstantIndices())
214 case Instruction::Load:
215 case Instruction::Add:
216 case Instruction::Sub:
217 case Instruction::And:
218 case Instruction::Or:
219 case Instruction::Xor:
220 case Instruction::Shl:
221 case Instruction::LShr:
222 case Instruction::AShr:
223 case Instruction::ICmp:
224 case Instruction::Trunc:
225 case Instruction::ZExt:
226 case Instruction::SExt:
227 return 1; // These are all cheap.
229 case Instruction::Call:
230 case Instruction::Select:
235 /// DominatesMergePoint - If we have a merge point of an "if condition" as
236 /// accepted above, return true if the specified value dominates the block. We
237 /// don't handle the true generality of domination here, just a special case
238 /// which works well enough for us.
240 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
241 /// see if V (which must be an instruction) and its recursive operands
242 /// that do not dominate BB have a combined cost lower than CostRemaining and
243 /// are non-trapping. If both are true, the instruction is inserted into the
244 /// set and true is returned.
246 /// The cost for most non-trapping instructions is defined as 1 except for
247 /// Select whose cost is 2.
249 /// After this function returns, CostRemaining is decreased by the cost of
250 /// V plus its non-dominating operands. If that cost is greater than
251 /// CostRemaining, false is returned and CostRemaining is undefined.
252 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
253 SmallPtrSet<Instruction*, 4> *AggressiveInsts,
254 unsigned &CostRemaining) {
255 Instruction *I = dyn_cast<Instruction>(V);
257 // Non-instructions all dominate instructions, but not all constantexprs
258 // can be executed unconditionally.
259 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
264 BasicBlock *PBB = I->getParent();
266 // We don't want to allow weird loops that might have the "if condition" in
267 // the bottom of this block.
268 if (PBB == BB) return false;
270 // If this instruction is defined in a block that contains an unconditional
271 // branch to BB, then it must be in the 'conditional' part of the "if
272 // statement". If not, it definitely dominates the region.
273 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
274 if (BI == 0 || BI->isConditional() || BI->getSuccessor(0) != BB)
277 // If we aren't allowing aggressive promotion anymore, then don't consider
278 // instructions in the 'if region'.
279 if (AggressiveInsts == 0) return false;
281 // If we have seen this instruction before, don't count it again.
282 if (AggressiveInsts->count(I)) return true;
284 // Okay, it looks like the instruction IS in the "condition". Check to
285 // see if it's a cheap instruction to unconditionally compute, and if it
286 // only uses stuff defined outside of the condition. If so, hoist it out.
287 if (!isSafeToSpeculativelyExecute(I))
290 unsigned Cost = ComputeSpeculationCost(I);
292 if (Cost > CostRemaining)
295 CostRemaining -= Cost;
297 // Okay, we can only really hoist these out if their operands do
298 // not take us over the cost threshold.
299 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
300 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining))
302 // Okay, it's safe to do this! Remember this instruction.
303 AggressiveInsts->insert(I);
307 /// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr
308 /// and PointerNullValue. Return NULL if value is not a constant int.
309 static ConstantInt *GetConstantInt(Value *V, const DataLayout *DL) {
310 // Normal constant int.
311 ConstantInt *CI = dyn_cast<ConstantInt>(V);
312 if (CI || !DL || !isa<Constant>(V) || !V->getType()->isPointerTy())
315 // This is some kind of pointer constant. Turn it into a pointer-sized
316 // ConstantInt if possible.
317 IntegerType *PtrTy = cast<IntegerType>(DL->getIntPtrType(V->getType()));
319 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
320 if (isa<ConstantPointerNull>(V))
321 return ConstantInt::get(PtrTy, 0);
323 // IntToPtr const int.
324 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
325 if (CE->getOpcode() == Instruction::IntToPtr)
326 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
327 // The constant is very likely to have the right type already.
328 if (CI->getType() == PtrTy)
331 return cast<ConstantInt>
332 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
337 /// GatherConstantCompares - Given a potentially 'or'd or 'and'd together
338 /// collection of icmp eq/ne instructions that compare a value against a
339 /// constant, return the value being compared, and stick the constant into the
342 GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra,
343 const DataLayout *DL, bool isEQ, unsigned &UsedICmps) {
344 Instruction *I = dyn_cast<Instruction>(V);
345 if (I == 0) return 0;
347 // If this is an icmp against a constant, handle this as one of the cases.
348 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) {
349 if (ConstantInt *C = GetConstantInt(I->getOperand(1), DL)) {
353 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
354 // (x & ~2^x) == y --> x == y || x == y|2^x
355 // This undoes a transformation done by instcombine to fuse 2 compares.
356 if (match(ICI->getOperand(0),
357 m_And(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
358 APInt Not = ~RHSC->getValue();
359 if (Not.isPowerOf2()) {
362 ConstantInt::get(C->getContext(), C->getValue() | Not));
370 return I->getOperand(0);
373 // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to
376 ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue());
378 // Shift the range if the compare is fed by an add. This is the range
379 // compare idiom as emitted by instcombine.
381 match(I->getOperand(0), m_Add(m_Value(RHSVal), m_ConstantInt(RHSC)));
383 Span = Span.subtract(RHSC->getValue());
385 // If this is an and/!= check then we want to optimize "x ugt 2" into
388 Span = Span.inverse();
390 // If there are a ton of values, we don't want to make a ginormous switch.
391 if (Span.getSetSize().ugt(8) || Span.isEmptySet())
394 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
395 Vals.push_back(ConstantInt::get(V->getContext(), Tmp));
397 return hasAdd ? RHSVal : I->getOperand(0);
402 // Otherwise, we can only handle an | or &, depending on isEQ.
403 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And))
406 unsigned NumValsBeforeLHS = Vals.size();
407 unsigned UsedICmpsBeforeLHS = UsedICmps;
408 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, DL,
410 unsigned NumVals = Vals.size();
411 unsigned UsedICmpsBeforeRHS = UsedICmps;
412 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, DL,
416 Vals.resize(NumVals);
417 UsedICmps = UsedICmpsBeforeRHS;
420 // The RHS of the or/and can't be folded in and we haven't used "Extra" yet,
421 // set it and return success.
422 if (Extra == 0 || Extra == I->getOperand(1)) {
423 Extra = I->getOperand(1);
427 Vals.resize(NumValsBeforeLHS);
428 UsedICmps = UsedICmpsBeforeLHS;
432 // If the LHS can't be folded in, but Extra is available and RHS can, try to
434 if (Extra == 0 || Extra == I->getOperand(0)) {
435 Value *OldExtra = Extra;
436 Extra = I->getOperand(0);
437 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, DL,
440 assert(Vals.size() == NumValsBeforeLHS);
447 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
448 Instruction *Cond = 0;
449 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
450 Cond = dyn_cast<Instruction>(SI->getCondition());
451 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
452 if (BI->isConditional())
453 Cond = dyn_cast<Instruction>(BI->getCondition());
454 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
455 Cond = dyn_cast<Instruction>(IBI->getAddress());
458 TI->eraseFromParent();
459 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
462 /// isValueEqualityComparison - Return true if the specified terminator checks
463 /// to see if a value is equal to constant integer value.
464 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
466 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
467 // Do not permit merging of large switch instructions into their
468 // predecessors unless there is only one predecessor.
469 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()),
470 pred_end(SI->getParent())) <= 128)
471 CV = SI->getCondition();
472 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
473 if (BI->isConditional() && BI->getCondition()->hasOneUse())
474 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition()))
475 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
476 CV = ICI->getOperand(0);
478 // Unwrap any lossless ptrtoint cast.
480 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
481 Value *Ptr = PTII->getPointerOperand();
482 if (PTII->getType() == DL->getIntPtrType(Ptr->getType()))
489 /// GetValueEqualityComparisonCases - Given a value comparison instruction,
490 /// decode all of the 'cases' that it represents and return the 'default' block.
491 BasicBlock *SimplifyCFGOpt::
492 GetValueEqualityComparisonCases(TerminatorInst *TI,
493 std::vector<ValueEqualityComparisonCase>
495 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
496 Cases.reserve(SI->getNumCases());
497 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
498 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
499 i.getCaseSuccessor()));
500 return SI->getDefaultDest();
503 BranchInst *BI = cast<BranchInst>(TI);
504 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
505 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
506 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
509 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
513 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries
514 /// in the list that match the specified block.
515 static void EliminateBlockCases(BasicBlock *BB,
516 std::vector<ValueEqualityComparisonCase> &Cases) {
517 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
520 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as
523 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
524 std::vector<ValueEqualityComparisonCase > &C2) {
525 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
527 // Make V1 be smaller than V2.
528 if (V1->size() > V2->size())
531 if (V1->size() == 0) return false;
532 if (V1->size() == 1) {
534 ConstantInt *TheVal = (*V1)[0].Value;
535 for (unsigned i = 0, e = V2->size(); i != e; ++i)
536 if (TheVal == (*V2)[i].Value)
540 // Otherwise, just sort both lists and compare element by element.
541 array_pod_sort(V1->begin(), V1->end());
542 array_pod_sort(V2->begin(), V2->end());
543 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
544 while (i1 != e1 && i2 != e2) {
545 if ((*V1)[i1].Value == (*V2)[i2].Value)
547 if ((*V1)[i1].Value < (*V2)[i2].Value)
555 /// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a
556 /// terminator instruction and its block is known to only have a single
557 /// predecessor block, check to see if that predecessor is also a value
558 /// comparison with the same value, and if that comparison determines the
559 /// outcome of this comparison. If so, simplify TI. This does a very limited
560 /// form of jump threading.
561 bool SimplifyCFGOpt::
562 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
564 IRBuilder<> &Builder) {
565 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
566 if (!PredVal) return false; // Not a value comparison in predecessor.
568 Value *ThisVal = isValueEqualityComparison(TI);
569 assert(ThisVal && "This isn't a value comparison!!");
570 if (ThisVal != PredVal) return false; // Different predicates.
572 // TODO: Preserve branch weight metadata, similarly to how
573 // FoldValueComparisonIntoPredecessors preserves it.
575 // Find out information about when control will move from Pred to TI's block.
576 std::vector<ValueEqualityComparisonCase> PredCases;
577 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
579 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
581 // Find information about how control leaves this block.
582 std::vector<ValueEqualityComparisonCase> ThisCases;
583 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
584 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
586 // If TI's block is the default block from Pred's comparison, potentially
587 // simplify TI based on this knowledge.
588 if (PredDef == TI->getParent()) {
589 // If we are here, we know that the value is none of those cases listed in
590 // PredCases. If there are any cases in ThisCases that are in PredCases, we
592 if (!ValuesOverlap(PredCases, ThisCases))
595 if (isa<BranchInst>(TI)) {
596 // Okay, one of the successors of this condbr is dead. Convert it to a
598 assert(ThisCases.size() == 1 && "Branch can only have one case!");
599 // Insert the new branch.
600 Instruction *NI = Builder.CreateBr(ThisDef);
603 // Remove PHI node entries for the dead edge.
604 ThisCases[0].Dest->removePredecessor(TI->getParent());
606 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
607 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
609 EraseTerminatorInstAndDCECond(TI);
613 SwitchInst *SI = cast<SwitchInst>(TI);
614 // Okay, TI has cases that are statically dead, prune them away.
615 SmallPtrSet<Constant*, 16> DeadCases;
616 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
617 DeadCases.insert(PredCases[i].Value);
619 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
620 << "Through successor TI: " << *TI);
622 // Collect branch weights into a vector.
623 SmallVector<uint32_t, 8> Weights;
624 MDNode* MD = SI->getMetadata(LLVMContext::MD_prof);
625 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
627 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
629 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i));
631 Weights.push_back(CI->getValue().getZExtValue());
633 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
635 if (DeadCases.count(i.getCaseValue())) {
637 std::swap(Weights[i.getCaseIndex()+1], Weights.back());
640 i.getCaseSuccessor()->removePredecessor(TI->getParent());
644 if (HasWeight && Weights.size() >= 2)
645 SI->setMetadata(LLVMContext::MD_prof,
646 MDBuilder(SI->getParent()->getContext()).
647 createBranchWeights(Weights));
649 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
653 // Otherwise, TI's block must correspond to some matched value. Find out
654 // which value (or set of values) this is.
655 ConstantInt *TIV = 0;
656 BasicBlock *TIBB = TI->getParent();
657 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
658 if (PredCases[i].Dest == TIBB) {
660 return false; // Cannot handle multiple values coming to this block.
661 TIV = PredCases[i].Value;
663 assert(TIV && "No edge from pred to succ?");
665 // Okay, we found the one constant that our value can be if we get into TI's
666 // BB. Find out which successor will unconditionally be branched to.
667 BasicBlock *TheRealDest = 0;
668 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
669 if (ThisCases[i].Value == TIV) {
670 TheRealDest = ThisCases[i].Dest;
674 // If not handled by any explicit cases, it is handled by the default case.
675 if (TheRealDest == 0) TheRealDest = ThisDef;
677 // Remove PHI node entries for dead edges.
678 BasicBlock *CheckEdge = TheRealDest;
679 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
680 if (*SI != CheckEdge)
681 (*SI)->removePredecessor(TIBB);
685 // Insert the new branch.
686 Instruction *NI = Builder.CreateBr(TheRealDest);
689 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
690 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
692 EraseTerminatorInstAndDCECond(TI);
697 /// ConstantIntOrdering - This class implements a stable ordering of constant
698 /// integers that does not depend on their address. This is important for
699 /// applications that sort ConstantInt's to ensure uniqueness.
700 struct ConstantIntOrdering {
701 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
702 return LHS->getValue().ult(RHS->getValue());
707 static int ConstantIntSortPredicate(ConstantInt *const *P1,
708 ConstantInt *const *P2) {
709 const ConstantInt *LHS = *P1;
710 const ConstantInt *RHS = *P2;
711 if (LHS->getValue().ult(RHS->getValue()))
713 if (LHS->getValue() == RHS->getValue())
718 static inline bool HasBranchWeights(const Instruction* I) {
719 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof);
720 if (ProfMD && ProfMD->getOperand(0))
721 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
722 return MDS->getString().equals("branch_weights");
727 /// Get Weights of a given TerminatorInst, the default weight is at the front
728 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
730 static void GetBranchWeights(TerminatorInst *TI,
731 SmallVectorImpl<uint64_t> &Weights) {
732 MDNode* MD = TI->getMetadata(LLVMContext::MD_prof);
734 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
735 ConstantInt* CI = dyn_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) {
755 for (unsigned i = 0; i < Weights.size(); ++i)
756 if (Weights[i] > UINT_MAX) {
764 for (unsigned i = 0; i < Weights.size(); ++i)
769 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value
770 /// equality comparison instruction (either a switch or a branch on "X == c").
771 /// See if any of the predecessors of the terminator block are value comparisons
772 /// on the same value. If so, and if safe to do so, fold them together.
773 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
774 IRBuilder<> &Builder) {
775 BasicBlock *BB = TI->getParent();
776 Value *CV = isValueEqualityComparison(TI); // CondVal
777 assert(CV && "Not a comparison?");
778 bool Changed = false;
780 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
781 while (!Preds.empty()) {
782 BasicBlock *Pred = Preds.pop_back_val();
784 // See if the predecessor is a comparison with the same value.
785 TerminatorInst *PTI = Pred->getTerminator();
786 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
788 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
789 // Figure out which 'cases' to copy from SI to PSI.
790 std::vector<ValueEqualityComparisonCase> BBCases;
791 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
793 std::vector<ValueEqualityComparisonCase> PredCases;
794 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
796 // Based on whether the default edge from PTI goes to BB or not, fill in
797 // PredCases and PredDefault with the new switch cases we would like to
799 SmallVector<BasicBlock*, 8> NewSuccessors;
801 // Update the branch weight metadata along the way
802 SmallVector<uint64_t, 8> Weights;
803 bool PredHasWeights = HasBranchWeights(PTI);
804 bool SuccHasWeights = HasBranchWeights(TI);
806 if (PredHasWeights) {
807 GetBranchWeights(PTI, Weights);
808 // branch-weight metadata is inconsistent here.
809 if (Weights.size() != 1 + PredCases.size())
810 PredHasWeights = SuccHasWeights = false;
811 } else if (SuccHasWeights)
812 // If there are no predecessor weights but there are successor weights,
813 // populate Weights with 1, which will later be scaled to the sum of
814 // successor's weights
815 Weights.assign(1 + PredCases.size(), 1);
817 SmallVector<uint64_t, 8> SuccWeights;
818 if (SuccHasWeights) {
819 GetBranchWeights(TI, SuccWeights);
820 // branch-weight metadata is inconsistent here.
821 if (SuccWeights.size() != 1 + BBCases.size())
822 PredHasWeights = SuccHasWeights = false;
823 } else if (PredHasWeights)
824 SuccWeights.assign(1 + BBCases.size(), 1);
826 if (PredDefault == BB) {
827 // If this is the default destination from PTI, only the edges in TI
828 // that don't occur in PTI, or that branch to BB will be activated.
829 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
830 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
831 if (PredCases[i].Dest != BB)
832 PTIHandled.insert(PredCases[i].Value);
834 // The default destination is BB, we don't need explicit targets.
835 std::swap(PredCases[i], PredCases.back());
837 if (PredHasWeights || SuccHasWeights) {
838 // Increase weight for the default case.
839 Weights[0] += Weights[i+1];
840 std::swap(Weights[i+1], Weights.back());
844 PredCases.pop_back();
848 // Reconstruct the new switch statement we will be building.
849 if (PredDefault != BBDefault) {
850 PredDefault->removePredecessor(Pred);
851 PredDefault = BBDefault;
852 NewSuccessors.push_back(BBDefault);
855 unsigned CasesFromPred = Weights.size();
856 uint64_t ValidTotalSuccWeight = 0;
857 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
858 if (!PTIHandled.count(BBCases[i].Value) &&
859 BBCases[i].Dest != BBDefault) {
860 PredCases.push_back(BBCases[i]);
861 NewSuccessors.push_back(BBCases[i].Dest);
862 if (SuccHasWeights || PredHasWeights) {
863 // The default weight is at index 0, so weight for the ith case
864 // should be at index i+1. Scale the cases from successor by
865 // PredDefaultWeight (Weights[0]).
866 Weights.push_back(Weights[0] * SuccWeights[i+1]);
867 ValidTotalSuccWeight += SuccWeights[i+1];
871 if (SuccHasWeights || PredHasWeights) {
872 ValidTotalSuccWeight += SuccWeights[0];
873 // Scale the cases from predecessor by ValidTotalSuccWeight.
874 for (unsigned i = 1; i < CasesFromPred; ++i)
875 Weights[i] *= ValidTotalSuccWeight;
876 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
877 Weights[0] *= SuccWeights[0];
880 // If this is not the default destination from PSI, only the edges
881 // in SI that occur in PSI with a destination of BB will be
883 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
884 std::map<ConstantInt*, uint64_t> WeightsForHandled;
885 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
886 if (PredCases[i].Dest == BB) {
887 PTIHandled.insert(PredCases[i].Value);
889 if (PredHasWeights || SuccHasWeights) {
890 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
891 std::swap(Weights[i+1], Weights.back());
895 std::swap(PredCases[i], PredCases.back());
896 PredCases.pop_back();
900 // Okay, now we know which constants were sent to BB from the
901 // predecessor. Figure out where they will all go now.
902 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
903 if (PTIHandled.count(BBCases[i].Value)) {
904 // If this is one we are capable of getting...
905 if (PredHasWeights || SuccHasWeights)
906 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
907 PredCases.push_back(BBCases[i]);
908 NewSuccessors.push_back(BBCases[i].Dest);
909 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
912 // If there are any constants vectored to BB that TI doesn't handle,
913 // they must go to the default destination of TI.
914 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
916 E = PTIHandled.end(); I != E; ++I) {
917 if (PredHasWeights || SuccHasWeights)
918 Weights.push_back(WeightsForHandled[*I]);
919 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
920 NewSuccessors.push_back(BBDefault);
924 // Okay, at this point, we know which new successor Pred will get. Make
925 // sure we update the number of entries in the PHI nodes for these
927 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i)
928 AddPredecessorToBlock(NewSuccessors[i], Pred, BB);
930 Builder.SetInsertPoint(PTI);
931 // Convert pointer to int before we switch.
932 if (CV->getType()->isPointerTy()) {
933 assert(DL && "Cannot switch on pointer without DataLayout");
934 CV = Builder.CreatePtrToInt(CV, DL->getIntPtrType(CV->getType()),
938 // Now that the successors are updated, create the new Switch instruction.
939 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
941 NewSI->setDebugLoc(PTI->getDebugLoc());
942 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
943 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest);
945 if (PredHasWeights || SuccHasWeights) {
946 // Halve the weights if any of them cannot fit in an uint32_t
949 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
951 NewSI->setMetadata(LLVMContext::MD_prof,
952 MDBuilder(BB->getContext()).
953 createBranchWeights(MDWeights));
956 EraseTerminatorInstAndDCECond(PTI);
958 // Okay, last check. If BB is still a successor of PSI, then we must
959 // have an infinite loop case. If so, add an infinitely looping block
960 // to handle the case to preserve the behavior of the code.
961 BasicBlock *InfLoopBlock = 0;
962 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
963 if (NewSI->getSuccessor(i) == BB) {
964 if (InfLoopBlock == 0) {
965 // Insert it at the end of the function, because it's either code,
966 // or it won't matter if it's hot. :)
967 InfLoopBlock = BasicBlock::Create(BB->getContext(),
968 "infloop", BB->getParent());
969 BranchInst::Create(InfLoopBlock, InfLoopBlock);
971 NewSI->setSuccessor(i, InfLoopBlock);
980 // isSafeToHoistInvoke - If we would need to insert a select that uses the
981 // value of this invoke (comments in HoistThenElseCodeToIf explain why we
982 // would need to do this), we can't hoist the invoke, as there is nowhere
983 // to put the select in this case.
984 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
985 Instruction *I1, Instruction *I2) {
986 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
988 for (BasicBlock::iterator BBI = SI->begin();
989 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
990 Value *BB1V = PN->getIncomingValueForBlock(BB1);
991 Value *BB2V = PN->getIncomingValueForBlock(BB2);
992 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
1000 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
1001 /// BB2, hoist any common code in the two blocks up into the branch block. The
1002 /// caller of this function guarantees that BI's block dominates BB1 and BB2.
1003 static bool HoistThenElseCodeToIf(BranchInst *BI) {
1004 // This does very trivial matching, with limited scanning, to find identical
1005 // instructions in the two blocks. In particular, we don't want to get into
1006 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1007 // such, we currently just scan for obviously identical instructions in an
1009 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1010 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1012 BasicBlock::iterator BB1_Itr = BB1->begin();
1013 BasicBlock::iterator BB2_Itr = BB2->begin();
1015 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++;
1016 // Skip debug info if it is not identical.
1017 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1018 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1019 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1020 while (isa<DbgInfoIntrinsic>(I1))
1022 while (isa<DbgInfoIntrinsic>(I2))
1025 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1026 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1029 BasicBlock *BIParent = BI->getParent();
1031 bool Changed = false;
1033 // If we are hoisting the terminator instruction, don't move one (making a
1034 // broken BB), instead clone it, and remove BI.
1035 if (isa<TerminatorInst>(I1))
1036 goto HoistTerminator;
1038 // For a normal instruction, we just move one to right before the branch,
1039 // then replace all uses of the other with the first. Finally, we remove
1040 // the now redundant second instruction.
1041 BIParent->getInstList().splice(BI, BB1->getInstList(), I1);
1042 if (!I2->use_empty())
1043 I2->replaceAllUsesWith(I1);
1044 I1->intersectOptionalDataWith(I2);
1045 I2->eraseFromParent();
1050 // Skip debug info if it is not identical.
1051 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1052 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1053 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1054 while (isa<DbgInfoIntrinsic>(I1))
1056 while (isa<DbgInfoIntrinsic>(I2))
1059 } while (I1->isIdenticalToWhenDefined(I2));
1064 // It may not be possible to hoist an invoke.
1065 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1068 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1070 for (BasicBlock::iterator BBI = SI->begin();
1071 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1072 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1073 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1077 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1079 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1084 // Okay, it is safe to hoist the terminator.
1085 Instruction *NT = I1->clone();
1086 BIParent->getInstList().insert(BI, NT);
1087 if (!NT->getType()->isVoidTy()) {
1088 I1->replaceAllUsesWith(NT);
1089 I2->replaceAllUsesWith(NT);
1093 IRBuilder<true, NoFolder> Builder(NT);
1094 // Hoisting one of the terminators from our successor is a great thing.
1095 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1096 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1097 // nodes, so we insert select instruction to compute the final result.
1098 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1099 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1101 for (BasicBlock::iterator BBI = SI->begin();
1102 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1103 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1104 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1105 if (BB1V == BB2V) continue;
1107 // These values do not agree. Insert a select instruction before NT
1108 // that determines the right value.
1109 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1111 SI = cast<SelectInst>
1112 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1113 BB1V->getName()+"."+BB2V->getName()));
1115 // Make the PHI node use the select for all incoming values for BB1/BB2
1116 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1117 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1118 PN->setIncomingValue(i, SI);
1122 // Update any PHI nodes in our new successors.
1123 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1124 AddPredecessorToBlock(*SI, BIParent, BB1);
1126 EraseTerminatorInstAndDCECond(BI);
1130 /// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd,
1131 /// check whether BBEnd has only two predecessors and the other predecessor
1132 /// ends with an unconditional branch. If it is true, sink any common code
1133 /// in the two predecessors to BBEnd.
1134 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1135 assert(BI1->isUnconditional());
1136 BasicBlock *BB1 = BI1->getParent();
1137 BasicBlock *BBEnd = BI1->getSuccessor(0);
1139 // Check that BBEnd has two predecessors and the other predecessor ends with
1140 // an unconditional branch.
1141 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1142 BasicBlock *Pred0 = *PI++;
1143 if (PI == PE) // Only one predecessor.
1145 BasicBlock *Pred1 = *PI++;
1146 if (PI != PE) // More than two predecessors.
1148 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1149 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1150 if (!BI2 || !BI2->isUnconditional())
1153 // Gather the PHI nodes in BBEnd.
1154 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2;
1155 Instruction *FirstNonPhiInBBEnd = 0;
1156 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end();
1158 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1159 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1160 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1161 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN);
1163 FirstNonPhiInBBEnd = &*I;
1167 if (!FirstNonPhiInBBEnd)
1171 // This does very trivial matching, with limited scanning, to find identical
1172 // instructions in the two blocks. We scan backward for obviously identical
1173 // instructions in an identical order.
1174 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1175 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(),
1176 RE2 = BB2->getInstList().rend();
1178 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1181 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1184 // Skip the unconditional branches.
1188 bool Changed = false;
1189 while (RI1 != RE1 && RI2 != RE2) {
1191 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1194 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1198 Instruction *I1 = &*RI1, *I2 = &*RI2;
1199 // I1 and I2 should have a single use in the same PHI node, and they
1200 // perform the same operation.
1201 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1202 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1203 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1204 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) ||
1205 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1206 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1207 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1208 !I1->hasOneUse() || !I2->hasOneUse() ||
1209 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() ||
1210 MapValueFromBB1ToBB2[I1].first != I2)
1213 // Check whether we should swap the operands of ICmpInst.
1214 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1215 bool SwapOpnds = false;
1216 if (ICmp1 && ICmp2 &&
1217 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1218 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1219 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1220 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1221 ICmp2->swapOperands();
1224 if (!I1->isSameOperationAs(I2)) {
1226 ICmp2->swapOperands();
1230 // The operands should be either the same or they need to be generated
1231 // with a PHI node after sinking. We only handle the case where there is
1232 // a single pair of different operands.
1233 Value *DifferentOp1 = 0, *DifferentOp2 = 0;
1234 unsigned Op1Idx = 0;
1235 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1236 if (I1->getOperand(I) == I2->getOperand(I))
1238 // Early exit if we have more-than one pair of different operands or
1239 // the different operand is already in MapValueFromBB1ToBB2.
1240 // Early exit if we need a PHI node to replace a constant.
1242 MapValueFromBB1ToBB2.find(I1->getOperand(I)) !=
1243 MapValueFromBB1ToBB2.end() ||
1244 isa<Constant>(I1->getOperand(I)) ||
1245 isa<Constant>(I2->getOperand(I))) {
1246 // If we can't sink the instructions, undo the swapping.
1248 ICmp2->swapOperands();
1251 DifferentOp1 = I1->getOperand(I);
1253 DifferentOp2 = I2->getOperand(I);
1256 // We insert the pair of different operands to MapValueFromBB1ToBB2 and
1257 // remove (I1, I2) from MapValueFromBB1ToBB2.
1259 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2,
1260 DifferentOp1->getName() + ".sink",
1262 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN);
1263 // I1 should use NewPN instead of DifferentOp1.
1264 I1->setOperand(Op1Idx, NewPN);
1265 NewPN->addIncoming(DifferentOp1, BB1);
1266 NewPN->addIncoming(DifferentOp2, BB2);
1267 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1269 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second;
1270 MapValueFromBB1ToBB2.erase(I1);
1272 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";);
1273 DEBUG(dbgs() << " " << *I2 << "\n";);
1274 // We need to update RE1 and RE2 if we are going to sink the first
1275 // instruction in the basic block down.
1276 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1277 // Sink the instruction.
1278 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1);
1279 if (!OldPN->use_empty())
1280 OldPN->replaceAllUsesWith(I1);
1281 OldPN->eraseFromParent();
1283 if (!I2->use_empty())
1284 I2->replaceAllUsesWith(I1);
1285 I1->intersectOptionalDataWith(I2);
1286 I2->eraseFromParent();
1289 RE1 = BB1->getInstList().rend();
1291 RE2 = BB2->getInstList().rend();
1292 FirstNonPhiInBBEnd = I1;
1299 /// \brief Determine if we can hoist sink a sole store instruction out of a
1300 /// conditional block.
1302 /// We are looking for code like the following:
1304 /// store i32 %add, i32* %arrayidx2
1305 /// ... // No other stores or function calls (we could be calling a memory
1306 /// ... // function).
1307 /// %cmp = icmp ult %x, %y
1308 /// br i1 %cmp, label %EndBB, label %ThenBB
1310 /// store i32 %add5, i32* %arrayidx2
1314 /// We are going to transform this into:
1316 /// store i32 %add, i32* %arrayidx2
1318 /// %cmp = icmp ult %x, %y
1319 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1320 /// store i32 %add.add5, i32* %arrayidx2
1323 /// \return The pointer to the value of the previous store if the store can be
1324 /// hoisted into the predecessor block. 0 otherwise.
1325 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1326 BasicBlock *StoreBB, BasicBlock *EndBB) {
1327 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1331 // Volatile or atomic.
1332 if (!StoreToHoist->isSimple())
1335 Value *StorePtr = StoreToHoist->getPointerOperand();
1337 // Look for a store to the same pointer in BrBB.
1338 unsigned MaxNumInstToLookAt = 10;
1339 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(),
1340 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) {
1341 Instruction *CurI = &*RI;
1343 // Could be calling an instruction that effects memory like free().
1344 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI))
1347 StoreInst *SI = dyn_cast<StoreInst>(CurI);
1348 // Found the previous store make sure it stores to the same location.
1349 if (SI && SI->getPointerOperand() == StorePtr)
1350 // Found the previous store, return its value operand.
1351 return SI->getValueOperand();
1353 return 0; // Unknown store.
1359 /// \brief Speculate a conditional basic block flattening the CFG.
1361 /// Note that this is a very risky transform currently. Speculating
1362 /// instructions like this is most often not desirable. Instead, there is an MI
1363 /// pass which can do it with full awareness of the resource constraints.
1364 /// However, some cases are "obvious" and we should do directly. An example of
1365 /// this is speculating a single, reasonably cheap instruction.
1367 /// There is only one distinct advantage to flattening the CFG at the IR level:
1368 /// it makes very common but simplistic optimizations such as are common in
1369 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1370 /// modeling their effects with easier to reason about SSA value graphs.
1373 /// An illustration of this transform is turning this IR:
1376 /// %cmp = icmp ult %x, %y
1377 /// br i1 %cmp, label %EndBB, label %ThenBB
1379 /// %sub = sub %x, %y
1382 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1389 /// %cmp = icmp ult %x, %y
1390 /// %sub = sub %x, %y
1391 /// %cond = select i1 %cmp, 0, %sub
1395 /// \returns true if the conditional block is removed.
1396 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB) {
1397 // Be conservative for now. FP select instruction can often be expensive.
1398 Value *BrCond = BI->getCondition();
1399 if (isa<FCmpInst>(BrCond))
1402 BasicBlock *BB = BI->getParent();
1403 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1405 // If ThenBB is actually on the false edge of the conditional branch, remember
1406 // to swap the select operands later.
1407 bool Invert = false;
1408 if (ThenBB != BI->getSuccessor(0)) {
1409 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1412 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1414 // Keep a count of how many times instructions are used within CondBB when
1415 // they are candidates for sinking into CondBB. Specifically:
1416 // - They are defined in BB, and
1417 // - They have no side effects, and
1418 // - All of their uses are in CondBB.
1419 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1421 unsigned SpeculationCost = 0;
1422 Value *SpeculatedStoreValue = 0;
1423 StoreInst *SpeculatedStore = 0;
1424 for (BasicBlock::iterator BBI = ThenBB->begin(),
1425 BBE = std::prev(ThenBB->end());
1426 BBI != BBE; ++BBI) {
1427 Instruction *I = BBI;
1429 if (isa<DbgInfoIntrinsic>(I))
1432 // Only speculatively execution a single instruction (not counting the
1433 // terminator) for now.
1435 if (SpeculationCost > 1)
1438 // Don't hoist the instruction if it's unsafe or expensive.
1439 if (!isSafeToSpeculativelyExecute(I) &&
1440 !(HoistCondStores &&
1441 (SpeculatedStoreValue = isSafeToSpeculateStore(I, BB, ThenBB,
1444 if (!SpeculatedStoreValue &&
1445 ComputeSpeculationCost(I) > PHINodeFoldingThreshold)
1448 // Store the store speculation candidate.
1449 if (SpeculatedStoreValue)
1450 SpeculatedStore = cast<StoreInst>(I);
1452 // Do not hoist the instruction if any of its operands are defined but not
1453 // used in BB. The transformation will prevent the operand from
1454 // being sunk into the use block.
1455 for (User::op_iterator i = I->op_begin(), e = I->op_end();
1457 Instruction *OpI = dyn_cast<Instruction>(*i);
1458 if (!OpI || OpI->getParent() != BB ||
1459 OpI->mayHaveSideEffects())
1460 continue; // Not a candidate for sinking.
1462 ++SinkCandidateUseCounts[OpI];
1466 // Consider any sink candidates which are only used in CondBB as costs for
1467 // speculation. Note, while we iterate over a DenseMap here, we are summing
1468 // and so iteration order isn't significant.
1469 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I =
1470 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end();
1472 if (I->first->getNumUses() == I->second) {
1474 if (SpeculationCost > 1)
1478 // Check that the PHI nodes can be converted to selects.
1479 bool HaveRewritablePHIs = false;
1480 for (BasicBlock::iterator I = EndBB->begin();
1481 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1482 Value *OrigV = PN->getIncomingValueForBlock(BB);
1483 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1485 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1486 // Skip PHIs which are trivial.
1490 HaveRewritablePHIs = true;
1491 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1492 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1493 if (!OrigCE && !ThenCE)
1494 continue; // Known safe and cheap.
1496 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
1497 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
1499 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE) : 0;
1500 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE) : 0;
1501 if (OrigCost + ThenCost > 2 * PHINodeFoldingThreshold)
1504 // Account for the cost of an unfolded ConstantExpr which could end up
1505 // getting expanded into Instructions.
1506 // FIXME: This doesn't account for how many operations are combined in the
1507 // constant expression.
1509 if (SpeculationCost > 1)
1513 // If there are no PHIs to process, bail early. This helps ensure idempotence
1515 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
1518 // If we get here, we can hoist the instruction and if-convert.
1519 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1521 // Insert a select of the value of the speculated store.
1522 if (SpeculatedStoreValue) {
1523 IRBuilder<true, NoFolder> Builder(BI);
1524 Value *TrueV = SpeculatedStore->getValueOperand();
1525 Value *FalseV = SpeculatedStoreValue;
1527 std::swap(TrueV, FalseV);
1528 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() +
1529 "." + FalseV->getName());
1530 SpeculatedStore->setOperand(0, S);
1533 // Hoist the instructions.
1534 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(),
1535 std::prev(ThenBB->end()));
1537 // Insert selects and rewrite the PHI operands.
1538 IRBuilder<true, NoFolder> Builder(BI);
1539 for (BasicBlock::iterator I = EndBB->begin();
1540 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1541 unsigned OrigI = PN->getBasicBlockIndex(BB);
1542 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1543 Value *OrigV = PN->getIncomingValue(OrigI);
1544 Value *ThenV = PN->getIncomingValue(ThenI);
1546 // Skip PHIs which are trivial.
1550 // Create a select whose true value is the speculatively executed value and
1551 // false value is the preexisting value. Swap them if the branch
1552 // destinations were inverted.
1553 Value *TrueV = ThenV, *FalseV = OrigV;
1555 std::swap(TrueV, FalseV);
1556 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV,
1557 TrueV->getName() + "." + FalseV->getName());
1558 PN->setIncomingValue(OrigI, V);
1559 PN->setIncomingValue(ThenI, V);
1566 /// \returns True if this block contains a CallInst with the NoDuplicate
1568 static bool HasNoDuplicateCall(const BasicBlock *BB) {
1569 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1570 const CallInst *CI = dyn_cast<CallInst>(I);
1573 if (CI->cannotDuplicate())
1579 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1580 /// across this block.
1581 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1582 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1585 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1586 if (isa<DbgInfoIntrinsic>(BBI))
1588 if (Size > 10) return false; // Don't clone large BB's.
1591 // We can only support instructions that do not define values that are
1592 // live outside of the current basic block.
1593 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end();
1595 Instruction *U = cast<Instruction>(*UI);
1596 if (U->getParent() != BB || isa<PHINode>(U)) return false;
1599 // Looks ok, continue checking.
1605 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1606 /// that is defined in the same block as the branch and if any PHI entries are
1607 /// constants, thread edges corresponding to that entry to be branches to their
1608 /// ultimate destination.
1609 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *DL) {
1610 BasicBlock *BB = BI->getParent();
1611 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1612 // NOTE: we currently cannot transform this case if the PHI node is used
1613 // outside of the block.
1614 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1617 // Degenerate case of a single entry PHI.
1618 if (PN->getNumIncomingValues() == 1) {
1619 FoldSingleEntryPHINodes(PN->getParent());
1623 // Now we know that this block has multiple preds and two succs.
1624 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1626 if (HasNoDuplicateCall(BB)) return false;
1628 // Okay, this is a simple enough basic block. See if any phi values are
1630 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1631 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1632 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue;
1634 // Okay, we now know that all edges from PredBB should be revectored to
1635 // branch to RealDest.
1636 BasicBlock *PredBB = PN->getIncomingBlock(i);
1637 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1639 if (RealDest == BB) continue; // Skip self loops.
1640 // Skip if the predecessor's terminator is an indirect branch.
1641 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1643 // The dest block might have PHI nodes, other predecessors and other
1644 // difficult cases. Instead of being smart about this, just insert a new
1645 // block that jumps to the destination block, effectively splitting
1646 // the edge we are about to create.
1647 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1648 RealDest->getName()+".critedge",
1649 RealDest->getParent(), RealDest);
1650 BranchInst::Create(RealDest, EdgeBB);
1652 // Update PHI nodes.
1653 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1655 // BB may have instructions that are being threaded over. Clone these
1656 // instructions into EdgeBB. We know that there will be no uses of the
1657 // cloned instructions outside of EdgeBB.
1658 BasicBlock::iterator InsertPt = EdgeBB->begin();
1659 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1660 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1661 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1662 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1665 // Clone the instruction.
1666 Instruction *N = BBI->clone();
1667 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1669 // Update operands due to translation.
1670 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1672 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1673 if (PI != TranslateMap.end())
1677 // Check for trivial simplification.
1678 if (Value *V = SimplifyInstruction(N, DL)) {
1679 TranslateMap[BBI] = V;
1680 delete N; // Instruction folded away, don't need actual inst
1682 // Insert the new instruction into its new home.
1683 EdgeBB->getInstList().insert(InsertPt, N);
1684 if (!BBI->use_empty())
1685 TranslateMap[BBI] = N;
1689 // Loop over all of the edges from PredBB to BB, changing them to branch
1690 // to EdgeBB instead.
1691 TerminatorInst *PredBBTI = PredBB->getTerminator();
1692 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1693 if (PredBBTI->getSuccessor(i) == BB) {
1694 BB->removePredecessor(PredBB);
1695 PredBBTI->setSuccessor(i, EdgeBB);
1698 // Recurse, simplifying any other constants.
1699 return FoldCondBranchOnPHI(BI, DL) | true;
1705 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1706 /// PHI node, see if we can eliminate it.
1707 static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *DL) {
1708 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1709 // statement", which has a very simple dominance structure. Basically, we
1710 // are trying to find the condition that is being branched on, which
1711 // subsequently causes this merge to happen. We really want control
1712 // dependence information for this check, but simplifycfg can't keep it up
1713 // to date, and this catches most of the cases we care about anyway.
1714 BasicBlock *BB = PN->getParent();
1715 BasicBlock *IfTrue, *IfFalse;
1716 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1718 // Don't bother if the branch will be constant folded trivially.
1719 isa<ConstantInt>(IfCond))
1722 // Okay, we found that we can merge this two-entry phi node into a select.
1723 // Doing so would require us to fold *all* two entry phi nodes in this block.
1724 // At some point this becomes non-profitable (particularly if the target
1725 // doesn't support cmov's). Only do this transformation if there are two or
1726 // fewer PHI nodes in this block.
1727 unsigned NumPhis = 0;
1728 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1732 // Loop over the PHI's seeing if we can promote them all to select
1733 // instructions. While we are at it, keep track of the instructions
1734 // that need to be moved to the dominating block.
1735 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1736 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1737 MaxCostVal1 = PHINodeFoldingThreshold;
1739 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1740 PHINode *PN = cast<PHINode>(II++);
1741 if (Value *V = SimplifyInstruction(PN, DL)) {
1742 PN->replaceAllUsesWith(V);
1743 PN->eraseFromParent();
1747 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1749 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1754 // If we folded the first phi, PN dangles at this point. Refresh it. If
1755 // we ran out of PHIs then we simplified them all.
1756 PN = dyn_cast<PHINode>(BB->begin());
1757 if (PN == 0) return true;
1759 // Don't fold i1 branches on PHIs which contain binary operators. These can
1760 // often be turned into switches and other things.
1761 if (PN->getType()->isIntegerTy(1) &&
1762 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1763 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1764 isa<BinaryOperator>(IfCond)))
1767 // If we all PHI nodes are promotable, check to make sure that all
1768 // instructions in the predecessor blocks can be promoted as well. If
1769 // not, we won't be able to get rid of the control flow, so it's not
1770 // worth promoting to select instructions.
1771 BasicBlock *DomBlock = 0;
1772 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1773 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1774 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1777 DomBlock = *pred_begin(IfBlock1);
1778 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1779 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1780 // This is not an aggressive instruction that we can promote.
1781 // Because of this, we won't be able to get rid of the control
1782 // flow, so the xform is not worth it.
1787 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1790 DomBlock = *pred_begin(IfBlock2);
1791 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1792 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1793 // This is not an aggressive instruction that we can promote.
1794 // Because of this, we won't be able to get rid of the control
1795 // flow, so the xform is not worth it.
1800 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1801 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1803 // If we can still promote the PHI nodes after this gauntlet of tests,
1804 // do all of the PHI's now.
1805 Instruction *InsertPt = DomBlock->getTerminator();
1806 IRBuilder<true, NoFolder> Builder(InsertPt);
1808 // Move all 'aggressive' instructions, which are defined in the
1809 // conditional parts of the if's up to the dominating block.
1811 DomBlock->getInstList().splice(InsertPt,
1812 IfBlock1->getInstList(), IfBlock1->begin(),
1813 IfBlock1->getTerminator());
1815 DomBlock->getInstList().splice(InsertPt,
1816 IfBlock2->getInstList(), IfBlock2->begin(),
1817 IfBlock2->getTerminator());
1819 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1820 // Change the PHI node into a select instruction.
1821 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1822 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1825 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1826 PN->replaceAllUsesWith(NV);
1828 PN->eraseFromParent();
1831 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1832 // has been flattened. Change DomBlock to jump directly to our new block to
1833 // avoid other simplifycfg's kicking in on the diamond.
1834 TerminatorInst *OldTI = DomBlock->getTerminator();
1835 Builder.SetInsertPoint(OldTI);
1836 Builder.CreateBr(BB);
1837 OldTI->eraseFromParent();
1841 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1842 /// to two returning blocks, try to merge them together into one return,
1843 /// introducing a select if the return values disagree.
1844 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1845 IRBuilder<> &Builder) {
1846 assert(BI->isConditional() && "Must be a conditional branch");
1847 BasicBlock *TrueSucc = BI->getSuccessor(0);
1848 BasicBlock *FalseSucc = BI->getSuccessor(1);
1849 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1850 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1852 // Check to ensure both blocks are empty (just a return) or optionally empty
1853 // with PHI nodes. If there are other instructions, merging would cause extra
1854 // computation on one path or the other.
1855 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1857 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1860 Builder.SetInsertPoint(BI);
1861 // Okay, we found a branch that is going to two return nodes. If
1862 // there is no return value for this function, just change the
1863 // branch into a return.
1864 if (FalseRet->getNumOperands() == 0) {
1865 TrueSucc->removePredecessor(BI->getParent());
1866 FalseSucc->removePredecessor(BI->getParent());
1867 Builder.CreateRetVoid();
1868 EraseTerminatorInstAndDCECond(BI);
1872 // Otherwise, figure out what the true and false return values are
1873 // so we can insert a new select instruction.
1874 Value *TrueValue = TrueRet->getReturnValue();
1875 Value *FalseValue = FalseRet->getReturnValue();
1877 // Unwrap any PHI nodes in the return blocks.
1878 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1879 if (TVPN->getParent() == TrueSucc)
1880 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1881 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1882 if (FVPN->getParent() == FalseSucc)
1883 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1885 // In order for this transformation to be safe, we must be able to
1886 // unconditionally execute both operands to the return. This is
1887 // normally the case, but we could have a potentially-trapping
1888 // constant expression that prevents this transformation from being
1890 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1893 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1897 // Okay, we collected all the mapped values and checked them for sanity, and
1898 // defined to really do this transformation. First, update the CFG.
1899 TrueSucc->removePredecessor(BI->getParent());
1900 FalseSucc->removePredecessor(BI->getParent());
1902 // Insert select instructions where needed.
1903 Value *BrCond = BI->getCondition();
1905 // Insert a select if the results differ.
1906 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1907 } else if (isa<UndefValue>(TrueValue)) {
1908 TrueValue = FalseValue;
1910 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1911 FalseValue, "retval");
1915 Value *RI = !TrueValue ?
1916 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
1920 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
1921 << "\n " << *BI << "NewRet = " << *RI
1922 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
1924 EraseTerminatorInstAndDCECond(BI);
1929 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
1930 /// probabilities of the branch taking each edge. Fills in the two APInt
1931 /// parameters and return true, or returns false if no or invalid metadata was
1933 static bool ExtractBranchMetadata(BranchInst *BI,
1934 uint64_t &ProbTrue, uint64_t &ProbFalse) {
1935 assert(BI->isConditional() &&
1936 "Looking for probabilities on unconditional branch?");
1937 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
1938 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
1939 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
1940 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
1941 if (!CITrue || !CIFalse) return false;
1942 ProbTrue = CITrue->getValue().getZExtValue();
1943 ProbFalse = CIFalse->getValue().getZExtValue();
1947 /// checkCSEInPredecessor - Return true if the given instruction is available
1948 /// in its predecessor block. If yes, the instruction will be removed.
1950 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
1951 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
1953 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
1954 Instruction *PBI = &*I;
1955 // Check whether Inst and PBI generate the same value.
1956 if (Inst->isIdenticalTo(PBI)) {
1957 Inst->replaceAllUsesWith(PBI);
1958 Inst->eraseFromParent();
1965 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
1966 /// predecessor branches to us and one of our successors, fold the block into
1967 /// the predecessor and use logical operations to pick the right destination.
1968 bool llvm::FoldBranchToCommonDest(BranchInst *BI) {
1969 BasicBlock *BB = BI->getParent();
1971 Instruction *Cond = 0;
1972 if (BI->isConditional())
1973 Cond = dyn_cast<Instruction>(BI->getCondition());
1975 // For unconditional branch, check for a simple CFG pattern, where
1976 // BB has a single predecessor and BB's successor is also its predecessor's
1977 // successor. If such pattern exisits, check for CSE between BB and its
1979 if (BasicBlock *PB = BB->getSinglePredecessor())
1980 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
1981 if (PBI->isConditional() &&
1982 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
1983 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
1984 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
1986 Instruction *Curr = I++;
1987 if (isa<CmpInst>(Curr)) {
1991 // Quit if we can't remove this instruction.
1992 if (!checkCSEInPredecessor(Curr, PB))
2001 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2002 Cond->getParent() != BB || !Cond->hasOneUse())
2005 // Only allow this if the condition is a simple instruction that can be
2006 // executed unconditionally. It must be in the same block as the branch, and
2007 // must be at the front of the block.
2008 BasicBlock::iterator FrontIt = BB->front();
2010 // Ignore dbg intrinsics.
2011 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
2013 // Allow a single instruction to be hoisted in addition to the compare
2014 // that feeds the branch. We later ensure that any values that _it_ uses
2015 // were also live in the predecessor, so that we don't unnecessarily create
2016 // register pressure or inhibit out-of-order execution.
2017 Instruction *BonusInst = 0;
2018 if (&*FrontIt != Cond &&
2019 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond &&
2020 isSafeToSpeculativelyExecute(FrontIt)) {
2021 BonusInst = &*FrontIt;
2024 // Ignore dbg intrinsics.
2025 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
2028 // Only a single bonus inst is allowed.
2029 if (&*FrontIt != Cond)
2032 // Make sure the instruction after the condition is the cond branch.
2033 BasicBlock::iterator CondIt = Cond; ++CondIt;
2035 // Ingore dbg intrinsics.
2036 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
2041 // Cond is known to be a compare or binary operator. Check to make sure that
2042 // neither operand is a potentially-trapping constant expression.
2043 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2046 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2050 // Finally, don't infinitely unroll conditional loops.
2051 BasicBlock *TrueDest = BI->getSuccessor(0);
2052 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0;
2053 if (TrueDest == BB || FalseDest == BB)
2056 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2057 BasicBlock *PredBlock = *PI;
2058 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2060 // Check that we have two conditional branches. If there is a PHI node in
2061 // the common successor, verify that the same value flows in from both
2063 SmallVector<PHINode*, 4> PHIs;
2064 if (PBI == 0 || PBI->isUnconditional() ||
2065 (BI->isConditional() &&
2066 !SafeToMergeTerminators(BI, PBI)) ||
2067 (!BI->isConditional() &&
2068 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2071 // Determine if the two branches share a common destination.
2072 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2073 bool InvertPredCond = false;
2075 if (BI->isConditional()) {
2076 if (PBI->getSuccessor(0) == TrueDest)
2077 Opc = Instruction::Or;
2078 else if (PBI->getSuccessor(1) == FalseDest)
2079 Opc = Instruction::And;
2080 else if (PBI->getSuccessor(0) == FalseDest)
2081 Opc = Instruction::And, InvertPredCond = true;
2082 else if (PBI->getSuccessor(1) == TrueDest)
2083 Opc = Instruction::Or, InvertPredCond = true;
2087 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2091 // Ensure that any values used in the bonus instruction are also used
2092 // by the terminator of the predecessor. This means that those values
2093 // must already have been resolved, so we won't be inhibiting the
2094 // out-of-order core by speculating them earlier. We also allow
2095 // instructions that are used by the terminator's condition because it
2096 // exposes more merging opportunities.
2097 bool UsedByBranch = (BonusInst && BonusInst->hasOneUse() &&
2098 *BonusInst->use_begin() == Cond);
2100 if (BonusInst && !UsedByBranch) {
2101 // Collect the values used by the bonus inst
2102 SmallPtrSet<Value*, 4> UsedValues;
2103 for (Instruction::op_iterator OI = BonusInst->op_begin(),
2104 OE = BonusInst->op_end(); OI != OE; ++OI) {
2106 if (!isa<Constant>(V) && !isa<Argument>(V))
2107 UsedValues.insert(V);
2110 SmallVector<std::pair<Value*, unsigned>, 4> Worklist;
2111 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0));
2113 // Walk up to four levels back up the use-def chain of the predecessor's
2114 // terminator to see if all those values were used. The choice of four
2115 // levels is arbitrary, to provide a compile-time-cost bound.
2116 while (!Worklist.empty()) {
2117 std::pair<Value*, unsigned> Pair = Worklist.back();
2118 Worklist.pop_back();
2120 if (Pair.second >= 4) continue;
2121 UsedValues.erase(Pair.first);
2122 if (UsedValues.empty()) break;
2124 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) {
2125 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
2127 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1));
2131 if (!UsedValues.empty()) return false;
2134 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2135 IRBuilder<> Builder(PBI);
2137 // If we need to invert the condition in the pred block to match, do so now.
2138 if (InvertPredCond) {
2139 Value *NewCond = PBI->getCondition();
2141 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2142 CmpInst *CI = cast<CmpInst>(NewCond);
2143 CI->setPredicate(CI->getInversePredicate());
2145 NewCond = Builder.CreateNot(NewCond,
2146 PBI->getCondition()->getName()+".not");
2149 PBI->setCondition(NewCond);
2150 PBI->swapSuccessors();
2153 // If we have a bonus inst, clone it into the predecessor block.
2154 Instruction *NewBonus = 0;
2156 NewBonus = BonusInst->clone();
2158 // If we moved a load, we cannot any longer claim any knowledge about
2159 // its potential value. The previous information might have been valid
2160 // only given the branch precondition.
2161 // For an analogous reason, we must also drop all the metadata whose
2162 // semantics we don't understand.
2163 NewBonus->dropUnknownMetadata(LLVMContext::MD_dbg);
2165 PredBlock->getInstList().insert(PBI, NewBonus);
2166 NewBonus->takeName(BonusInst);
2167 BonusInst->setName(BonusInst->getName()+".old");
2170 // Clone Cond into the predecessor basic block, and or/and the
2171 // two conditions together.
2172 Instruction *New = Cond->clone();
2173 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus);
2174 PredBlock->getInstList().insert(PBI, New);
2175 New->takeName(Cond);
2176 Cond->setName(New->getName()+".old");
2178 if (BI->isConditional()) {
2179 Instruction *NewCond =
2180 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2182 PBI->setCondition(NewCond);
2184 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2185 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2187 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2189 SmallVector<uint64_t, 8> NewWeights;
2191 if (PBI->getSuccessor(0) == BB) {
2192 if (PredHasWeights && SuccHasWeights) {
2193 // PBI: br i1 %x, BB, FalseDest
2194 // BI: br i1 %y, TrueDest, FalseDest
2195 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2196 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2197 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2198 // TrueWeight for PBI * FalseWeight for BI.
2199 // We assume that total weights of a BranchInst can fit into 32 bits.
2200 // Therefore, we will not have overflow using 64-bit arithmetic.
2201 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2202 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2204 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2205 PBI->setSuccessor(0, TrueDest);
2207 if (PBI->getSuccessor(1) == BB) {
2208 if (PredHasWeights && SuccHasWeights) {
2209 // PBI: br i1 %x, TrueDest, BB
2210 // BI: br i1 %y, TrueDest, FalseDest
2211 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2212 // FalseWeight for PBI * TrueWeight for BI.
2213 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2214 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2215 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2216 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2218 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2219 PBI->setSuccessor(1, FalseDest);
2221 if (NewWeights.size() == 2) {
2222 // Halve the weights if any of them cannot fit in an uint32_t
2223 FitWeights(NewWeights);
2225 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2226 PBI->setMetadata(LLVMContext::MD_prof,
2227 MDBuilder(BI->getContext()).
2228 createBranchWeights(MDWeights));
2230 PBI->setMetadata(LLVMContext::MD_prof, NULL);
2232 // Update PHI nodes in the common successors.
2233 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2234 ConstantInt *PBI_C = cast<ConstantInt>(
2235 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2236 assert(PBI_C->getType()->isIntegerTy(1));
2237 Instruction *MergedCond = 0;
2238 if (PBI->getSuccessor(0) == TrueDest) {
2239 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2240 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2241 // is false: !PBI_Cond and BI_Value
2242 Instruction *NotCond =
2243 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2246 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2251 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2252 PBI->getCondition(), MergedCond,
2255 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2256 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2257 // is false: PBI_Cond and BI_Value
2259 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2260 PBI->getCondition(), New,
2262 if (PBI_C->isOne()) {
2263 Instruction *NotCond =
2264 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2267 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2268 NotCond, MergedCond,
2273 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2276 // Change PBI from Conditional to Unconditional.
2277 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2278 EraseTerminatorInstAndDCECond(PBI);
2282 // TODO: If BB is reachable from all paths through PredBlock, then we
2283 // could replace PBI's branch probabilities with BI's.
2285 // Copy any debug value intrinsics into the end of PredBlock.
2286 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2287 if (isa<DbgInfoIntrinsic>(*I))
2288 I->clone()->insertBefore(PBI);
2295 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2296 /// predecessor of another block, this function tries to simplify it. We know
2297 /// that PBI and BI are both conditional branches, and BI is in one of the
2298 /// successor blocks of PBI - PBI branches to BI.
2299 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2300 assert(PBI->isConditional() && BI->isConditional());
2301 BasicBlock *BB = BI->getParent();
2303 // If this block ends with a branch instruction, and if there is a
2304 // predecessor that ends on a branch of the same condition, make
2305 // this conditional branch redundant.
2306 if (PBI->getCondition() == BI->getCondition() &&
2307 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2308 // Okay, the outcome of this conditional branch is statically
2309 // knowable. If this block had a single pred, handle specially.
2310 if (BB->getSinglePredecessor()) {
2311 // Turn this into a branch on constant.
2312 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2313 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2315 return true; // Nuke the branch on constant.
2318 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2319 // in the constant and simplify the block result. Subsequent passes of
2320 // simplifycfg will thread the block.
2321 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2322 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2323 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2324 std::distance(PB, PE),
2325 BI->getCondition()->getName() + ".pr",
2327 // Okay, we're going to insert the PHI node. Since PBI is not the only
2328 // predecessor, compute the PHI'd conditional value for all of the preds.
2329 // Any predecessor where the condition is not computable we keep symbolic.
2330 for (pred_iterator PI = PB; PI != PE; ++PI) {
2331 BasicBlock *P = *PI;
2332 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2333 PBI != BI && PBI->isConditional() &&
2334 PBI->getCondition() == BI->getCondition() &&
2335 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2336 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2337 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2340 NewPN->addIncoming(BI->getCondition(), P);
2344 BI->setCondition(NewPN);
2349 // If this is a conditional branch in an empty block, and if any
2350 // predecessors is a conditional branch to one of our destinations,
2351 // fold the conditions into logical ops and one cond br.
2352 BasicBlock::iterator BBI = BB->begin();
2353 // Ignore dbg intrinsics.
2354 while (isa<DbgInfoIntrinsic>(BBI))
2360 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2365 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2367 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2368 PBIOp = 0, BIOp = 1;
2369 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2370 PBIOp = 1, BIOp = 0;
2371 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2376 // Check to make sure that the other destination of this branch
2377 // isn't BB itself. If so, this is an infinite loop that will
2378 // keep getting unwound.
2379 if (PBI->getSuccessor(PBIOp) == BB)
2382 // Do not perform this transformation if it would require
2383 // insertion of a large number of select instructions. For targets
2384 // without predication/cmovs, this is a big pessimization.
2385 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2387 unsigned NumPhis = 0;
2388 for (BasicBlock::iterator II = CommonDest->begin();
2389 isa<PHINode>(II); ++II, ++NumPhis)
2390 if (NumPhis > 2) // Disable this xform.
2393 // Finally, if everything is ok, fold the branches to logical ops.
2394 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2396 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2397 << "AND: " << *BI->getParent());
2400 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2401 // branch in it, where one edge (OtherDest) goes back to itself but the other
2402 // exits. We don't *know* that the program avoids the infinite loop
2403 // (even though that seems likely). If we do this xform naively, we'll end up
2404 // recursively unpeeling the loop. Since we know that (after the xform is
2405 // done) that the block *is* infinite if reached, we just make it an obviously
2406 // infinite loop with no cond branch.
2407 if (OtherDest == BB) {
2408 // Insert it at the end of the function, because it's either code,
2409 // or it won't matter if it's hot. :)
2410 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2411 "infloop", BB->getParent());
2412 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2413 OtherDest = InfLoopBlock;
2416 DEBUG(dbgs() << *PBI->getParent()->getParent());
2418 // BI may have other predecessors. Because of this, we leave
2419 // it alone, but modify PBI.
2421 // Make sure we get to CommonDest on True&True directions.
2422 Value *PBICond = PBI->getCondition();
2423 IRBuilder<true, NoFolder> Builder(PBI);
2425 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2427 Value *BICond = BI->getCondition();
2429 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2431 // Merge the conditions.
2432 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2434 // Modify PBI to branch on the new condition to the new dests.
2435 PBI->setCondition(Cond);
2436 PBI->setSuccessor(0, CommonDest);
2437 PBI->setSuccessor(1, OtherDest);
2439 // Update branch weight for PBI.
2440 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2441 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2443 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2445 if (PredHasWeights && SuccHasWeights) {
2446 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2447 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2448 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2449 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2450 // The weight to CommonDest should be PredCommon * SuccTotal +
2451 // PredOther * SuccCommon.
2452 // The weight to OtherDest should be PredOther * SuccOther.
2453 SmallVector<uint64_t, 2> NewWeights;
2454 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) +
2455 PredOther * SuccCommon);
2456 NewWeights.push_back(PredOther * SuccOther);
2457 // Halve the weights if any of them cannot fit in an uint32_t
2458 FitWeights(NewWeights);
2460 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end());
2461 PBI->setMetadata(LLVMContext::MD_prof,
2462 MDBuilder(BI->getContext()).
2463 createBranchWeights(MDWeights));
2466 // OtherDest may have phi nodes. If so, add an entry from PBI's
2467 // block that are identical to the entries for BI's block.
2468 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2470 // We know that the CommonDest already had an edge from PBI to
2471 // it. If it has PHIs though, the PHIs may have different
2472 // entries for BB and PBI's BB. If so, insert a select to make
2475 for (BasicBlock::iterator II = CommonDest->begin();
2476 (PN = dyn_cast<PHINode>(II)); ++II) {
2477 Value *BIV = PN->getIncomingValueForBlock(BB);
2478 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2479 Value *PBIV = PN->getIncomingValue(PBBIdx);
2481 // Insert a select in PBI to pick the right value.
2482 Value *NV = cast<SelectInst>
2483 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2484 PN->setIncomingValue(PBBIdx, NV);
2488 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2489 DEBUG(dbgs() << *PBI->getParent()->getParent());
2491 // This basic block is probably dead. We know it has at least
2492 // one fewer predecessor.
2496 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2497 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2498 // Takes care of updating the successors and removing the old terminator.
2499 // Also makes sure not to introduce new successors by assuming that edges to
2500 // non-successor TrueBBs and FalseBBs aren't reachable.
2501 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2502 BasicBlock *TrueBB, BasicBlock *FalseBB,
2503 uint32_t TrueWeight,
2504 uint32_t FalseWeight){
2505 // Remove any superfluous successor edges from the CFG.
2506 // First, figure out which successors to preserve.
2507 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2509 BasicBlock *KeepEdge1 = TrueBB;
2510 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0;
2512 // Then remove the rest.
2513 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2514 BasicBlock *Succ = OldTerm->getSuccessor(I);
2515 // Make sure only to keep exactly one copy of each edge.
2516 if (Succ == KeepEdge1)
2518 else if (Succ == KeepEdge2)
2521 Succ->removePredecessor(OldTerm->getParent());
2524 IRBuilder<> Builder(OldTerm);
2525 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2527 // Insert an appropriate new terminator.
2528 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) {
2529 if (TrueBB == FalseBB)
2530 // We were only looking for one successor, and it was present.
2531 // Create an unconditional branch to it.
2532 Builder.CreateBr(TrueBB);
2534 // We found both of the successors we were looking for.
2535 // Create a conditional branch sharing the condition of the select.
2536 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2537 if (TrueWeight != FalseWeight)
2538 NewBI->setMetadata(LLVMContext::MD_prof,
2539 MDBuilder(OldTerm->getContext()).
2540 createBranchWeights(TrueWeight, FalseWeight));
2542 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2543 // Neither of the selected blocks were successors, so this
2544 // terminator must be unreachable.
2545 new UnreachableInst(OldTerm->getContext(), OldTerm);
2547 // One of the selected values was a successor, but the other wasn't.
2548 // Insert an unconditional branch to the one that was found;
2549 // the edge to the one that wasn't must be unreachable.
2551 // Only TrueBB was found.
2552 Builder.CreateBr(TrueBB);
2554 // Only FalseBB was found.
2555 Builder.CreateBr(FalseBB);
2558 EraseTerminatorInstAndDCECond(OldTerm);
2562 // SimplifySwitchOnSelect - Replaces
2563 // (switch (select cond, X, Y)) on constant X, Y
2564 // with a branch - conditional if X and Y lead to distinct BBs,
2565 // unconditional otherwise.
2566 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2567 // Check for constant integer values in the select.
2568 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2569 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2570 if (!TrueVal || !FalseVal)
2573 // Find the relevant condition and destinations.
2574 Value *Condition = Select->getCondition();
2575 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2576 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2578 // Get weight for TrueBB and FalseBB.
2579 uint32_t TrueWeight = 0, FalseWeight = 0;
2580 SmallVector<uint64_t, 8> Weights;
2581 bool HasWeights = HasBranchWeights(SI);
2583 GetBranchWeights(SI, Weights);
2584 if (Weights.size() == 1 + SI->getNumCases()) {
2585 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2586 getSuccessorIndex()];
2587 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2588 getSuccessorIndex()];
2592 // Perform the actual simplification.
2593 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2594 TrueWeight, FalseWeight);
2597 // SimplifyIndirectBrOnSelect - Replaces
2598 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2599 // blockaddress(@fn, BlockB)))
2601 // (br cond, BlockA, BlockB).
2602 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2603 // Check that both operands of the select are block addresses.
2604 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2605 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2609 // Extract the actual blocks.
2610 BasicBlock *TrueBB = TBA->getBasicBlock();
2611 BasicBlock *FalseBB = FBA->getBasicBlock();
2613 // Perform the actual simplification.
2614 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2618 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2619 /// instruction (a seteq/setne with a constant) as the only instruction in a
2620 /// block that ends with an uncond branch. We are looking for a very specific
2621 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2622 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2623 /// default value goes to an uncond block with a seteq in it, we get something
2626 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2628 /// %tmp = icmp eq i8 %A, 92
2631 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2633 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2634 /// the PHI, merging the third icmp into the switch.
2635 static bool TryToSimplifyUncondBranchWithICmpInIt(
2636 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI,
2637 const DataLayout *DL) {
2638 BasicBlock *BB = ICI->getParent();
2640 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2642 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2644 Value *V = ICI->getOperand(0);
2645 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2647 // The pattern we're looking for is where our only predecessor is a switch on
2648 // 'V' and this block is the default case for the switch. In this case we can
2649 // fold the compared value into the switch to simplify things.
2650 BasicBlock *Pred = BB->getSinglePredecessor();
2651 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false;
2653 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2654 if (SI->getCondition() != V)
2657 // If BB is reachable on a non-default case, then we simply know the value of
2658 // V in this block. Substitute it and constant fold the icmp instruction
2660 if (SI->getDefaultDest() != BB) {
2661 ConstantInt *VVal = SI->findCaseDest(BB);
2662 assert(VVal && "Should have a unique destination value");
2663 ICI->setOperand(0, VVal);
2665 if (Value *V = SimplifyInstruction(ICI, DL)) {
2666 ICI->replaceAllUsesWith(V);
2667 ICI->eraseFromParent();
2669 // BB is now empty, so it is likely to simplify away.
2670 return SimplifyCFG(BB, TTI, DL) | true;
2673 // Ok, the block is reachable from the default dest. If the constant we're
2674 // comparing exists in one of the other edges, then we can constant fold ICI
2676 if (SI->findCaseValue(Cst) != SI->case_default()) {
2678 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2679 V = ConstantInt::getFalse(BB->getContext());
2681 V = ConstantInt::getTrue(BB->getContext());
2683 ICI->replaceAllUsesWith(V);
2684 ICI->eraseFromParent();
2685 // BB is now empty, so it is likely to simplify away.
2686 return SimplifyCFG(BB, TTI, DL) | true;
2689 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2691 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2692 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back());
2693 if (PHIUse == 0 || PHIUse != &SuccBlock->front() ||
2694 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2697 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2699 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2700 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2702 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2703 std::swap(DefaultCst, NewCst);
2705 // Replace ICI (which is used by the PHI for the default value) with true or
2706 // false depending on if it is EQ or NE.
2707 ICI->replaceAllUsesWith(DefaultCst);
2708 ICI->eraseFromParent();
2710 // Okay, the switch goes to this block on a default value. Add an edge from
2711 // the switch to the merge point on the compared value.
2712 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2713 BB->getParent(), BB);
2714 SmallVector<uint64_t, 8> Weights;
2715 bool HasWeights = HasBranchWeights(SI);
2717 GetBranchWeights(SI, Weights);
2718 if (Weights.size() == 1 + SI->getNumCases()) {
2719 // Split weight for default case to case for "Cst".
2720 Weights[0] = (Weights[0]+1) >> 1;
2721 Weights.push_back(Weights[0]);
2723 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2724 SI->setMetadata(LLVMContext::MD_prof,
2725 MDBuilder(SI->getContext()).
2726 createBranchWeights(MDWeights));
2729 SI->addCase(Cst, NewBB);
2731 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2732 Builder.SetInsertPoint(NewBB);
2733 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2734 Builder.CreateBr(SuccBlock);
2735 PHIUse->addIncoming(NewCst, NewBB);
2739 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2740 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2741 /// fold it into a switch instruction if so.
2742 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *DL,
2743 IRBuilder<> &Builder) {
2744 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2745 if (Cond == 0) return false;
2748 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2749 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2750 // 'setne's and'ed together, collect them.
2752 std::vector<ConstantInt*> Values;
2753 bool TrueWhenEqual = true;
2754 Value *ExtraCase = 0;
2755 unsigned UsedICmps = 0;
2757 if (Cond->getOpcode() == Instruction::Or) {
2758 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, DL, true,
2760 } else if (Cond->getOpcode() == Instruction::And) {
2761 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, DL, false,
2763 TrueWhenEqual = false;
2766 // If we didn't have a multiply compared value, fail.
2767 if (CompVal == 0) return false;
2769 // Avoid turning single icmps into a switch.
2773 // There might be duplicate constants in the list, which the switch
2774 // instruction can't handle, remove them now.
2775 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2776 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2778 // If Extra was used, we require at least two switch values to do the
2779 // transformation. A switch with one value is just an cond branch.
2780 if (ExtraCase && Values.size() < 2) return false;
2782 // TODO: Preserve branch weight metadata, similarly to how
2783 // FoldValueComparisonIntoPredecessors preserves it.
2785 // Figure out which block is which destination.
2786 BasicBlock *DefaultBB = BI->getSuccessor(1);
2787 BasicBlock *EdgeBB = BI->getSuccessor(0);
2788 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2790 BasicBlock *BB = BI->getParent();
2792 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2793 << " cases into SWITCH. BB is:\n" << *BB);
2795 // If there are any extra values that couldn't be folded into the switch
2796 // then we evaluate them with an explicit branch first. Split the block
2797 // right before the condbr to handle it.
2799 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2800 // Remove the uncond branch added to the old block.
2801 TerminatorInst *OldTI = BB->getTerminator();
2802 Builder.SetInsertPoint(OldTI);
2805 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2807 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2809 OldTI->eraseFromParent();
2811 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2812 // for the edge we just added.
2813 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2815 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2816 << "\nEXTRABB = " << *BB);
2820 Builder.SetInsertPoint(BI);
2821 // Convert pointer to int before we switch.
2822 if (CompVal->getType()->isPointerTy()) {
2823 assert(DL && "Cannot switch on pointer without DataLayout");
2824 CompVal = Builder.CreatePtrToInt(CompVal,
2825 DL->getIntPtrType(CompVal->getType()),
2829 // Create the new switch instruction now.
2830 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2832 // Add all of the 'cases' to the switch instruction.
2833 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2834 New->addCase(Values[i], EdgeBB);
2836 // We added edges from PI to the EdgeBB. As such, if there were any
2837 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2838 // the number of edges added.
2839 for (BasicBlock::iterator BBI = EdgeBB->begin();
2840 isa<PHINode>(BBI); ++BBI) {
2841 PHINode *PN = cast<PHINode>(BBI);
2842 Value *InVal = PN->getIncomingValueForBlock(BB);
2843 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2844 PN->addIncoming(InVal, BB);
2847 // Erase the old branch instruction.
2848 EraseTerminatorInstAndDCECond(BI);
2850 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2854 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2855 // If this is a trivial landing pad that just continues unwinding the caught
2856 // exception then zap the landing pad, turning its invokes into calls.
2857 BasicBlock *BB = RI->getParent();
2858 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2859 if (RI->getValue() != LPInst)
2860 // Not a landing pad, or the resume is not unwinding the exception that
2861 // caused control to branch here.
2864 // Check that there are no other instructions except for debug intrinsics.
2865 BasicBlock::iterator I = LPInst, E = RI;
2867 if (!isa<DbgInfoIntrinsic>(I))
2870 // Turn all invokes that unwind here into calls and delete the basic block.
2871 bool InvokeRequiresTableEntry = false;
2872 bool Changed = false;
2873 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2874 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2876 if (II->hasFnAttr(Attribute::UWTable)) {
2877 // Don't remove an `invoke' instruction if the ABI requires an entry into
2879 InvokeRequiresTableEntry = true;
2883 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2885 // Insert a call instruction before the invoke.
2886 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2888 Call->setCallingConv(II->getCallingConv());
2889 Call->setAttributes(II->getAttributes());
2890 Call->setDebugLoc(II->getDebugLoc());
2892 // Anything that used the value produced by the invoke instruction now uses
2893 // the value produced by the call instruction. Note that we do this even
2894 // for void functions and calls with no uses so that the callgraph edge is
2896 II->replaceAllUsesWith(Call);
2897 BB->removePredecessor(II->getParent());
2899 // Insert a branch to the normal destination right before the invoke.
2900 BranchInst::Create(II->getNormalDest(), II);
2902 // Finally, delete the invoke instruction!
2903 II->eraseFromParent();
2907 if (!InvokeRequiresTableEntry)
2908 // The landingpad is now unreachable. Zap it.
2909 BB->eraseFromParent();
2914 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2915 BasicBlock *BB = RI->getParent();
2916 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2918 // Find predecessors that end with branches.
2919 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2920 SmallVector<BranchInst*, 8> CondBranchPreds;
2921 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2922 BasicBlock *P = *PI;
2923 TerminatorInst *PTI = P->getTerminator();
2924 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2925 if (BI->isUnconditional())
2926 UncondBranchPreds.push_back(P);
2928 CondBranchPreds.push_back(BI);
2932 // If we found some, do the transformation!
2933 if (!UncondBranchPreds.empty() && DupRet) {
2934 while (!UncondBranchPreds.empty()) {
2935 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
2936 DEBUG(dbgs() << "FOLDING: " << *BB
2937 << "INTO UNCOND BRANCH PRED: " << *Pred);
2938 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
2941 // If we eliminated all predecessors of the block, delete the block now.
2942 if (pred_begin(BB) == pred_end(BB))
2943 // We know there are no successors, so just nuke the block.
2944 BB->eraseFromParent();
2949 // Check out all of the conditional branches going to this return
2950 // instruction. If any of them just select between returns, change the
2951 // branch itself into a select/return pair.
2952 while (!CondBranchPreds.empty()) {
2953 BranchInst *BI = CondBranchPreds.pop_back_val();
2955 // Check to see if the non-BB successor is also a return block.
2956 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
2957 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
2958 SimplifyCondBranchToTwoReturns(BI, Builder))
2964 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
2965 BasicBlock *BB = UI->getParent();
2967 bool Changed = false;
2969 // If there are any instructions immediately before the unreachable that can
2970 // be removed, do so.
2971 while (UI != BB->begin()) {
2972 BasicBlock::iterator BBI = UI;
2974 // Do not delete instructions that can have side effects which might cause
2975 // the unreachable to not be reachable; specifically, calls and volatile
2976 // operations may have this effect.
2977 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
2979 if (BBI->mayHaveSideEffects()) {
2980 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
2981 if (SI->isVolatile())
2983 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
2984 if (LI->isVolatile())
2986 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
2987 if (RMWI->isVolatile())
2989 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
2990 if (CXI->isVolatile())
2992 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
2993 !isa<LandingPadInst>(BBI)) {
2996 // Note that deleting LandingPad's here is in fact okay, although it
2997 // involves a bit of subtle reasoning. If this inst is a LandingPad,
2998 // all the predecessors of this block will be the unwind edges of Invokes,
2999 // and we can therefore guarantee this block will be erased.
3002 // Delete this instruction (any uses are guaranteed to be dead)
3003 if (!BBI->use_empty())
3004 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
3005 BBI->eraseFromParent();
3009 // If the unreachable instruction is the first in the block, take a gander
3010 // at all of the predecessors of this instruction, and simplify them.
3011 if (&BB->front() != UI) return Changed;
3013 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
3014 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
3015 TerminatorInst *TI = Preds[i]->getTerminator();
3016 IRBuilder<> Builder(TI);
3017 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3018 if (BI->isUnconditional()) {
3019 if (BI->getSuccessor(0) == BB) {
3020 new UnreachableInst(TI->getContext(), TI);
3021 TI->eraseFromParent();
3025 if (BI->getSuccessor(0) == BB) {
3026 Builder.CreateBr(BI->getSuccessor(1));
3027 EraseTerminatorInstAndDCECond(BI);
3028 } else if (BI->getSuccessor(1) == BB) {
3029 Builder.CreateBr(BI->getSuccessor(0));
3030 EraseTerminatorInstAndDCECond(BI);
3034 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3035 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3037 if (i.getCaseSuccessor() == BB) {
3038 BB->removePredecessor(SI->getParent());
3043 // If the default value is unreachable, figure out the most popular
3044 // destination and make it the default.
3045 if (SI->getDefaultDest() == BB) {
3046 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
3047 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3049 std::pair<unsigned, unsigned> &entry =
3050 Popularity[i.getCaseSuccessor()];
3051 if (entry.first == 0) {
3053 entry.second = i.getCaseIndex();
3059 // Find the most popular block.
3060 unsigned MaxPop = 0;
3061 unsigned MaxIndex = 0;
3062 BasicBlock *MaxBlock = 0;
3063 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
3064 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
3065 if (I->second.first > MaxPop ||
3066 (I->second.first == MaxPop && MaxIndex > I->second.second)) {
3067 MaxPop = I->second.first;
3068 MaxIndex = I->second.second;
3069 MaxBlock = I->first;
3073 // Make this the new default, allowing us to delete any explicit
3075 SI->setDefaultDest(MaxBlock);
3078 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
3080 if (isa<PHINode>(MaxBlock->begin()))
3081 for (unsigned i = 0; i != MaxPop-1; ++i)
3082 MaxBlock->removePredecessor(SI->getParent());
3084 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3086 if (i.getCaseSuccessor() == MaxBlock) {
3092 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
3093 if (II->getUnwindDest() == BB) {
3094 // Convert the invoke to a call instruction. This would be a good
3095 // place to note that the call does not throw though.
3096 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
3097 II->removeFromParent(); // Take out of symbol table
3099 // Insert the call now...
3100 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
3101 Builder.SetInsertPoint(BI);
3102 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
3103 Args, II->getName());
3104 CI->setCallingConv(II->getCallingConv());
3105 CI->setAttributes(II->getAttributes());
3106 // If the invoke produced a value, the call does now instead.
3107 II->replaceAllUsesWith(CI);
3114 // If this block is now dead, remove it.
3115 if (pred_begin(BB) == pred_end(BB) &&
3116 BB != &BB->getParent()->getEntryBlock()) {
3117 // We know there are no successors, so just nuke the block.
3118 BB->eraseFromParent();
3125 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
3126 /// integer range comparison into a sub, an icmp and a branch.
3127 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3128 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3130 // Make sure all cases point to the same destination and gather the values.
3131 SmallVector<ConstantInt *, 16> Cases;
3132 SwitchInst::CaseIt I = SI->case_begin();
3133 Cases.push_back(I.getCaseValue());
3134 SwitchInst::CaseIt PrevI = I++;
3135 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
3136 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
3138 Cases.push_back(I.getCaseValue());
3140 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
3142 // Sort the case values, then check if they form a range we can transform.
3143 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3144 for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
3145 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
3149 Constant *Offset = ConstantExpr::getNeg(Cases.back());
3150 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
3152 Value *Sub = SI->getCondition();
3153 if (!Offset->isNullValue())
3154 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
3156 // If NumCases overflowed, then all possible values jump to the successor.
3157 if (NumCases->isNullValue() && SI->getNumCases() != 0)
3158 Cmp = ConstantInt::getTrue(SI->getContext());
3160 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3161 BranchInst *NewBI = Builder.CreateCondBr(
3162 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
3164 // Update weight for the newly-created conditional branch.
3165 SmallVector<uint64_t, 8> Weights;
3166 bool HasWeights = HasBranchWeights(SI);
3168 GetBranchWeights(SI, Weights);
3169 if (Weights.size() == 1 + SI->getNumCases()) {
3170 // Combine all weights for the cases to be the true weight of NewBI.
3171 // We assume that the sum of all weights for a Terminator can fit into 32
3173 uint32_t NewTrueWeight = 0;
3174 for (unsigned I = 1, E = Weights.size(); I != E; ++I)
3175 NewTrueWeight += (uint32_t)Weights[I];
3176 NewBI->setMetadata(LLVMContext::MD_prof,
3177 MDBuilder(SI->getContext()).
3178 createBranchWeights(NewTrueWeight,
3179 (uint32_t)Weights[0]));
3183 // Prune obsolete incoming values off the successor's PHI nodes.
3184 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
3185 isa<PHINode>(BBI); ++BBI) {
3186 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
3187 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3189 SI->eraseFromParent();
3194 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
3195 /// and use it to remove dead cases.
3196 static bool EliminateDeadSwitchCases(SwitchInst *SI) {
3197 Value *Cond = SI->getCondition();
3198 unsigned Bits = Cond->getType()->getIntegerBitWidth();
3199 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3200 ComputeMaskedBits(Cond, KnownZero, KnownOne);
3202 // Gather dead cases.
3203 SmallVector<ConstantInt*, 8> DeadCases;
3204 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3205 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3206 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3207 DeadCases.push_back(I.getCaseValue());
3208 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3209 << I.getCaseValue() << "' is dead.\n");
3213 SmallVector<uint64_t, 8> Weights;
3214 bool HasWeight = HasBranchWeights(SI);
3216 GetBranchWeights(SI, Weights);
3217 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3220 // Remove dead cases from the switch.
3221 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3222 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3223 assert(Case != SI->case_default() &&
3224 "Case was not found. Probably mistake in DeadCases forming.");
3226 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3230 // Prune unused values from PHI nodes.
3231 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3232 SI->removeCase(Case);
3234 if (HasWeight && Weights.size() >= 2) {
3235 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3236 SI->setMetadata(LLVMContext::MD_prof,
3237 MDBuilder(SI->getParent()->getContext()).
3238 createBranchWeights(MDWeights));
3241 return !DeadCases.empty();
3244 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
3245 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3246 /// by an unconditional branch), look at the phi node for BB in the successor
3247 /// block and see if the incoming value is equal to CaseValue. If so, return
3248 /// the phi node, and set PhiIndex to BB's index in the phi node.
3249 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3252 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3253 return NULL; // BB must be empty to be a candidate for simplification.
3254 if (!BB->getSinglePredecessor())
3255 return NULL; // BB must be dominated by the switch.
3257 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3258 if (!Branch || !Branch->isUnconditional())
3259 return NULL; // Terminator must be unconditional branch.
3261 BasicBlock *Succ = Branch->getSuccessor(0);
3263 BasicBlock::iterator I = Succ->begin();
3264 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3265 int Idx = PHI->getBasicBlockIndex(BB);
3266 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3268 Value *InValue = PHI->getIncomingValue(Idx);
3269 if (InValue != CaseValue) continue;
3278 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
3279 /// instruction to a phi node dominated by the switch, if that would mean that
3280 /// some of the destination blocks of the switch can be folded away.
3281 /// Returns true if a change is made.
3282 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3283 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3284 ForwardingNodesMap ForwardingNodes;
3286 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3287 ConstantInt *CaseValue = I.getCaseValue();
3288 BasicBlock *CaseDest = I.getCaseSuccessor();
3291 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3295 ForwardingNodes[PHI].push_back(PhiIndex);
3298 bool Changed = false;
3300 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3301 E = ForwardingNodes.end(); I != E; ++I) {
3302 PHINode *Phi = I->first;
3303 SmallVectorImpl<int> &Indexes = I->second;
3305 if (Indexes.size() < 2) continue;
3307 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3308 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3315 /// ValidLookupTableConstant - Return true if the backend will be able to handle
3316 /// initializing an array of constants like C.
3317 static bool ValidLookupTableConstant(Constant *C) {
3318 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3319 return CE->isGEPWithNoNotionalOverIndexing();
3321 return isa<ConstantFP>(C) ||
3322 isa<ConstantInt>(C) ||
3323 isa<ConstantPointerNull>(C) ||
3324 isa<GlobalValue>(C) ||
3328 /// LookupConstant - If V is a Constant, return it. Otherwise, try to look up
3329 /// its constant value in ConstantPool, returning 0 if it's not there.
3330 static Constant *LookupConstant(Value *V,
3331 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3332 if (Constant *C = dyn_cast<Constant>(V))
3334 return ConstantPool.lookup(V);
3337 /// ConstantFold - Try to fold instruction I into a constant. This works for
3338 /// simple instructions such as binary operations where both operands are
3339 /// constant or can be replaced by constants from the ConstantPool. Returns the
3340 /// resulting constant on success, 0 otherwise.
3342 ConstantFold(Instruction *I,
3343 const SmallDenseMap<Value *, Constant *> &ConstantPool,
3344 const DataLayout *DL) {
3345 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3346 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3349 if (A->isAllOnesValue())
3350 return LookupConstant(Select->getTrueValue(), ConstantPool);
3351 if (A->isNullValue())
3352 return LookupConstant(Select->getFalseValue(), ConstantPool);
3356 SmallVector<Constant *, 4> COps;
3357 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
3358 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
3364 if (CmpInst *Cmp = dyn_cast<CmpInst>(I))
3365 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
3368 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), COps, DL);
3371 /// GetCaseResults - Try to determine the resulting constant values in phi nodes
3372 /// at the common destination basic block, *CommonDest, for one of the case
3373 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3374 /// case), of a switch instruction SI.
3376 GetCaseResults(SwitchInst *SI,
3377 ConstantInt *CaseVal,
3378 BasicBlock *CaseDest,
3379 BasicBlock **CommonDest,
3380 SmallVectorImpl<std::pair<PHINode *, Constant *> > &Res,
3381 const DataLayout *DL) {
3382 // The block from which we enter the common destination.
3383 BasicBlock *Pred = SI->getParent();
3385 // If CaseDest is empty except for some side-effect free instructions through
3386 // which we can constant-propagate the CaseVal, continue to its successor.
3387 SmallDenseMap<Value*, Constant*> ConstantPool;
3388 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3389 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3391 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3392 // If the terminator is a simple branch, continue to the next block.
3393 if (T->getNumSuccessors() != 1)
3396 CaseDest = T->getSuccessor(0);
3397 } else if (isa<DbgInfoIntrinsic>(I)) {
3398 // Skip debug intrinsic.
3400 } else if (Constant *C = ConstantFold(I, ConstantPool, DL)) {
3401 // Instruction is side-effect free and constant.
3402 ConstantPool.insert(std::make_pair(I, C));
3408 // If we did not have a CommonDest before, use the current one.
3410 *CommonDest = CaseDest;
3411 // If the destination isn't the common one, abort.
3412 if (CaseDest != *CommonDest)
3415 // Get the values for this case from phi nodes in the destination block.
3416 BasicBlock::iterator I = (*CommonDest)->begin();
3417 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3418 int Idx = PHI->getBasicBlockIndex(Pred);
3422 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3427 // Note: If the constant comes from constant-propagating the case value
3428 // through the CaseDest basic block, it will be safe to remove the
3429 // instructions in that block. They cannot be used (except in the phi nodes
3430 // we visit) outside CaseDest, because that block does not dominate its
3431 // successor. If it did, we would not be in this phi node.
3433 // Be conservative about which kinds of constants we support.
3434 if (!ValidLookupTableConstant(ConstVal))
3437 Res.push_back(std::make_pair(PHI, ConstVal));
3440 return Res.size() > 0;
3444 /// SwitchLookupTable - This class represents a lookup table that can be used
3445 /// to replace a switch.
3446 class SwitchLookupTable {
3448 /// SwitchLookupTable - Create a lookup table to use as a switch replacement
3449 /// with the contents of Values, using DefaultValue to fill any holes in the
3451 SwitchLookupTable(Module &M,
3453 ConstantInt *Offset,
3454 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3455 Constant *DefaultValue,
3456 const DataLayout *DL);
3458 /// BuildLookup - Build instructions with Builder to retrieve the value at
3459 /// the position given by Index in the lookup table.
3460 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3462 /// WouldFitInRegister - Return true if a table with TableSize elements of
3463 /// type ElementType would fit in a target-legal register.
3464 static bool WouldFitInRegister(const DataLayout *DL,
3466 const Type *ElementType);
3469 // Depending on the contents of the table, it can be represented in
3472 // For tables where each element contains the same value, we just have to
3473 // store that single value and return it for each lookup.
3476 // For small tables with integer elements, we can pack them into a bitmap
3477 // that fits into a target-legal register. Values are retrieved by
3478 // shift and mask operations.
3481 // The table is stored as an array of values. Values are retrieved by load
3482 // instructions from the table.
3486 // For SingleValueKind, this is the single value.
3487 Constant *SingleValue;
3489 // For BitMapKind, this is the bitmap.
3490 ConstantInt *BitMap;
3491 IntegerType *BitMapElementTy;
3493 // For ArrayKind, this is the array.
3494 GlobalVariable *Array;
3498 SwitchLookupTable::SwitchLookupTable(Module &M,
3500 ConstantInt *Offset,
3501 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3502 Constant *DefaultValue,
3503 const DataLayout *DL)
3504 : SingleValue(0), BitMap(0), BitMapElementTy(0), Array(0) {
3505 assert(Values.size() && "Can't build lookup table without values!");
3506 assert(TableSize >= Values.size() && "Can't fit values in table!");
3508 // If all values in the table are equal, this is that value.
3509 SingleValue = Values.begin()->second;
3511 Type *ValueType = Values.begin()->second->getType();
3513 // Build up the table contents.
3514 SmallVector<Constant*, 64> TableContents(TableSize);
3515 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3516 ConstantInt *CaseVal = Values[I].first;
3517 Constant *CaseRes = Values[I].second;
3518 assert(CaseRes->getType() == ValueType);
3520 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3522 TableContents[Idx] = CaseRes;
3524 if (CaseRes != SingleValue)
3528 // Fill in any holes in the table with the default result.
3529 if (Values.size() < TableSize) {
3530 assert(DefaultValue && "Need a default value to fill the lookup table holes.");
3531 assert(DefaultValue->getType() == ValueType);
3532 for (uint64_t I = 0; I < TableSize; ++I) {
3533 if (!TableContents[I])
3534 TableContents[I] = DefaultValue;
3537 if (DefaultValue != SingleValue)
3541 // If each element in the table contains the same value, we only need to store
3542 // that single value.
3544 Kind = SingleValueKind;
3548 // If the type is integer and the table fits in a register, build a bitmap.
3549 if (WouldFitInRegister(DL, TableSize, ValueType)) {
3550 IntegerType *IT = cast<IntegerType>(ValueType);
3551 APInt TableInt(TableSize * IT->getBitWidth(), 0);
3552 for (uint64_t I = TableSize; I > 0; --I) {
3553 TableInt <<= IT->getBitWidth();
3554 // Insert values into the bitmap. Undef values are set to zero.
3555 if (!isa<UndefValue>(TableContents[I - 1])) {
3556 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
3557 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
3560 BitMap = ConstantInt::get(M.getContext(), TableInt);
3561 BitMapElementTy = IT;
3567 // Store the table in an array.
3568 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
3569 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3571 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3572 GlobalVariable::PrivateLinkage,
3575 Array->setUnnamedAddr(true);
3579 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
3581 case SingleValueKind:
3584 // Type of the bitmap (e.g. i59).
3585 IntegerType *MapTy = BitMap->getType();
3587 // Cast Index to the same type as the bitmap.
3588 // Note: The Index is <= the number of elements in the table, so
3589 // truncating it to the width of the bitmask is safe.
3590 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
3592 // Multiply the shift amount by the element width.
3593 ShiftAmt = Builder.CreateMul(ShiftAmt,
3594 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
3598 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
3599 "switch.downshift");
3601 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
3605 Value *GEPIndices[] = { Builder.getInt32(0), Index };
3606 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices,
3608 return Builder.CreateLoad(GEP, "switch.load");
3611 llvm_unreachable("Unknown lookup table kind!");
3614 bool SwitchLookupTable::WouldFitInRegister(const DataLayout *DL,
3616 const Type *ElementType) {
3619 const IntegerType *IT = dyn_cast<IntegerType>(ElementType);
3622 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
3623 // are <= 15, we could try to narrow the type.
3625 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
3626 if (TableSize >= UINT_MAX/IT->getBitWidth())
3628 return DL->fitsInLegalInteger(TableSize * IT->getBitWidth());
3631 /// ShouldBuildLookupTable - Determine whether a lookup table should be built
3632 /// for this switch, based on the number of cases, size of the table and the
3633 /// types of the results.
3634 static bool ShouldBuildLookupTable(SwitchInst *SI,
3636 const TargetTransformInfo &TTI,
3637 const DataLayout *DL,
3638 const SmallDenseMap<PHINode*, Type*>& ResultTypes) {
3639 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
3640 return false; // TableSize overflowed, or mul below might overflow.
3642 bool AllTablesFitInRegister = true;
3643 bool HasIllegalType = false;
3644 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(),
3645 E = ResultTypes.end(); I != E; ++I) {
3646 Type *Ty = I->second;
3648 // Saturate this flag to true.
3649 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
3651 // Saturate this flag to false.
3652 AllTablesFitInRegister = AllTablesFitInRegister &&
3653 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
3655 // If both flags saturate, we're done. NOTE: This *only* works with
3656 // saturating flags, and all flags have to saturate first due to the
3657 // non-deterministic behavior of iterating over a dense map.
3658 if (HasIllegalType && !AllTablesFitInRegister)
3662 // If each table would fit in a register, we should build it anyway.
3663 if (AllTablesFitInRegister)
3666 // Don't build a table that doesn't fit in-register if it has illegal types.
3670 // The table density should be at least 40%. This is the same criterion as for
3671 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3672 // FIXME: Find the best cut-off.
3673 return SI->getNumCases() * 10 >= TableSize * 4;
3676 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3677 /// phi nodes in a common successor block with different constant values,
3678 /// replace the switch with lookup tables.
3679 static bool SwitchToLookupTable(SwitchInst *SI,
3680 IRBuilder<> &Builder,
3681 const TargetTransformInfo &TTI,
3682 const DataLayout* DL) {
3683 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3685 // Only build lookup table when we have a target that supports it.
3686 if (!TTI.shouldBuildLookupTables())
3689 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3690 // split off a dense part and build a lookup table for that.
3692 // FIXME: This creates arrays of GEPs to constant strings, which means each
3693 // GEP needs a runtime relocation in PIC code. We should just build one big
3694 // string and lookup indices into that.
3696 // Ignore switches with less than three cases. Lookup tables will not make them
3697 // faster, so we don't analyze them.
3698 if (SI->getNumCases() < 3)
3701 // Figure out the corresponding result for each case value and phi node in the
3702 // common destination, as well as the the min and max case values.
3703 assert(SI->case_begin() != SI->case_end());
3704 SwitchInst::CaseIt CI = SI->case_begin();
3705 ConstantInt *MinCaseVal = CI.getCaseValue();
3706 ConstantInt *MaxCaseVal = CI.getCaseValue();
3708 BasicBlock *CommonDest = 0;
3709 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
3710 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
3711 SmallDenseMap<PHINode*, Constant*> DefaultResults;
3712 SmallDenseMap<PHINode*, Type*> ResultTypes;
3713 SmallVector<PHINode*, 4> PHIs;
3715 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
3716 ConstantInt *CaseVal = CI.getCaseValue();
3717 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
3718 MinCaseVal = CaseVal;
3719 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
3720 MaxCaseVal = CaseVal;
3722 // Resulting value at phi nodes for this case value.
3723 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
3725 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
3729 // Append the result from this case to the list for each phi.
3730 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) {
3731 if (!ResultLists.count(I->first))
3732 PHIs.push_back(I->first);
3733 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second));
3737 // Keep track of the result types.
3738 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
3739 PHINode *PHI = PHIs[I];
3740 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
3743 uint64_t NumResults = ResultLists[PHIs[0]].size();
3744 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
3745 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
3746 bool TableHasHoles = (NumResults < TableSize);
3748 // If the table has holes, we need a constant result for the default case.
3749 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
3750 if (TableHasHoles && !GetCaseResults(SI, 0, SI->getDefaultDest(), &CommonDest,
3751 DefaultResultsList, DL))
3754 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) {
3755 PHINode *PHI = DefaultResultsList[I].first;
3756 Constant *Result = DefaultResultsList[I].second;
3757 DefaultResults[PHI] = Result;
3760 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
3763 // Create the BB that does the lookups.
3764 Module &Mod = *CommonDest->getParent()->getParent();
3765 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
3767 CommonDest->getParent(),
3770 // Compute the table index value.
3771 Builder.SetInsertPoint(SI);
3772 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
3775 // Compute the maximum table size representable by the integer type we are
3777 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
3778 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
3779 assert(MaxTableSize >= TableSize &&
3780 "It is impossible for a switch to have more entries than the max "
3781 "representable value of its input integer type's size.");
3783 // If we have a fully covered lookup table, unconditionally branch to the
3784 // lookup table BB. Otherwise, check if the condition value is within the case
3785 // range. If it is so, branch to the new BB. Otherwise branch to SI's default
3787 const bool GeneratingCoveredLookupTable = MaxTableSize == TableSize;
3788 if (GeneratingCoveredLookupTable) {
3789 Builder.CreateBr(LookupBB);
3790 SI->getDefaultDest()->removePredecessor(SI->getParent());
3792 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
3793 MinCaseVal->getType(), TableSize));
3794 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
3797 // Populate the BB that does the lookups.
3798 Builder.SetInsertPoint(LookupBB);
3799 bool ReturnedEarly = false;
3800 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
3801 PHINode *PHI = PHIs[I];
3803 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI],
3804 DefaultResults[PHI], DL);
3806 Value *Result = Table.BuildLookup(TableIndex, Builder);
3808 // If the result is used to return immediately from the function, we want to
3809 // do that right here.
3810 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin()) &&
3811 *PHI->use_begin() == CommonDest->getFirstNonPHIOrDbg()) {
3812 Builder.CreateRet(Result);
3813 ReturnedEarly = true;
3817 PHI->addIncoming(Result, LookupBB);
3821 Builder.CreateBr(CommonDest);
3823 // Remove the switch.
3824 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
3825 BasicBlock *Succ = SI->getSuccessor(i);
3827 if (Succ == SI->getDefaultDest())
3829 Succ->removePredecessor(SI->getParent());
3831 SI->eraseFromParent();
3837 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
3838 BasicBlock *BB = SI->getParent();
3840 if (isValueEqualityComparison(SI)) {
3841 // If we only have one predecessor, and if it is a branch on this value,
3842 // see if that predecessor totally determines the outcome of this switch.
3843 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3844 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
3845 return SimplifyCFG(BB, TTI, DL) | true;
3847 Value *Cond = SI->getCondition();
3848 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
3849 if (SimplifySwitchOnSelect(SI, Select))
3850 return SimplifyCFG(BB, TTI, DL) | true;
3852 // If the block only contains the switch, see if we can fold the block
3853 // away into any preds.
3854 BasicBlock::iterator BBI = BB->begin();
3855 // Ignore dbg intrinsics.
3856 while (isa<DbgInfoIntrinsic>(BBI))
3859 if (FoldValueComparisonIntoPredecessors(SI, Builder))
3860 return SimplifyCFG(BB, TTI, DL) | true;
3863 // Try to transform the switch into an icmp and a branch.
3864 if (TurnSwitchRangeIntoICmp(SI, Builder))
3865 return SimplifyCFG(BB, TTI, DL) | true;
3867 // Remove unreachable cases.
3868 if (EliminateDeadSwitchCases(SI))
3869 return SimplifyCFG(BB, TTI, DL) | true;
3871 if (ForwardSwitchConditionToPHI(SI))
3872 return SimplifyCFG(BB, TTI, DL) | true;
3874 if (SwitchToLookupTable(SI, Builder, TTI, DL))
3875 return SimplifyCFG(BB, TTI, DL) | true;
3880 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
3881 BasicBlock *BB = IBI->getParent();
3882 bool Changed = false;
3884 // Eliminate redundant destinations.
3885 SmallPtrSet<Value *, 8> Succs;
3886 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
3887 BasicBlock *Dest = IBI->getDestination(i);
3888 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) {
3889 Dest->removePredecessor(BB);
3890 IBI->removeDestination(i);
3896 if (IBI->getNumDestinations() == 0) {
3897 // If the indirectbr has no successors, change it to unreachable.
3898 new UnreachableInst(IBI->getContext(), IBI);
3899 EraseTerminatorInstAndDCECond(IBI);
3903 if (IBI->getNumDestinations() == 1) {
3904 // If the indirectbr has one successor, change it to a direct branch.
3905 BranchInst::Create(IBI->getDestination(0), IBI);
3906 EraseTerminatorInstAndDCECond(IBI);
3910 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
3911 if (SimplifyIndirectBrOnSelect(IBI, SI))
3912 return SimplifyCFG(BB, TTI, DL) | true;
3917 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
3918 BasicBlock *BB = BI->getParent();
3920 if (SinkCommon && SinkThenElseCodeToEnd(BI))
3923 // If the Terminator is the only non-phi instruction, simplify the block.
3924 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime();
3925 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
3926 TryToSimplifyUncondBranchFromEmptyBlock(BB))
3929 // If the only instruction in the block is a seteq/setne comparison
3930 // against a constant, try to simplify the block.
3931 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
3932 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
3933 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
3935 if (I->isTerminator() &&
3936 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI, DL))
3940 // If this basic block is ONLY a compare and a branch, and if a predecessor
3941 // branches to us and our successor, fold the comparison into the
3942 // predecessor and use logical operations to update the incoming value
3943 // for PHI nodes in common successor.
3944 if (FoldBranchToCommonDest(BI))
3945 return SimplifyCFG(BB, TTI, DL) | true;
3950 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
3951 BasicBlock *BB = BI->getParent();
3953 // Conditional branch
3954 if (isValueEqualityComparison(BI)) {
3955 // If we only have one predecessor, and if it is a branch on this value,
3956 // see if that predecessor totally determines the outcome of this
3958 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3959 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
3960 return SimplifyCFG(BB, TTI, DL) | true;
3962 // This block must be empty, except for the setcond inst, if it exists.
3963 // Ignore dbg intrinsics.
3964 BasicBlock::iterator I = BB->begin();
3965 // Ignore dbg intrinsics.
3966 while (isa<DbgInfoIntrinsic>(I))
3969 if (FoldValueComparisonIntoPredecessors(BI, Builder))
3970 return SimplifyCFG(BB, TTI, DL) | true;
3971 } else if (&*I == cast<Instruction>(BI->getCondition())){
3973 // Ignore dbg intrinsics.
3974 while (isa<DbgInfoIntrinsic>(I))
3976 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
3977 return SimplifyCFG(BB, TTI, DL) | true;
3981 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
3982 if (SimplifyBranchOnICmpChain(BI, DL, Builder))
3985 // If this basic block is ONLY a compare and a branch, and if a predecessor
3986 // branches to us and one of our successors, fold the comparison into the
3987 // predecessor and use logical operations to pick the right destination.
3988 if (FoldBranchToCommonDest(BI))
3989 return SimplifyCFG(BB, TTI, DL) | true;
3991 // We have a conditional branch to two blocks that are only reachable
3992 // from BI. We know that the condbr dominates the two blocks, so see if
3993 // there is any identical code in the "then" and "else" blocks. If so, we
3994 // can hoist it up to the branching block.
3995 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) {
3996 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3997 if (HoistThenElseCodeToIf(BI))
3998 return SimplifyCFG(BB, TTI, DL) | true;
4000 // If Successor #1 has multiple preds, we may be able to conditionally
4001 // execute Successor #0 if it branches to successor #1.
4002 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
4003 if (Succ0TI->getNumSuccessors() == 1 &&
4004 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
4005 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0)))
4006 return SimplifyCFG(BB, TTI, DL) | true;
4008 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
4009 // If Successor #0 has multiple preds, we may be able to conditionally
4010 // execute Successor #1 if it branches to successor #0.
4011 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
4012 if (Succ1TI->getNumSuccessors() == 1 &&
4013 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
4014 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1)))
4015 return SimplifyCFG(BB, TTI, DL) | true;
4018 // If this is a branch on a phi node in the current block, thread control
4019 // through this block if any PHI node entries are constants.
4020 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
4021 if (PN->getParent() == BI->getParent())
4022 if (FoldCondBranchOnPHI(BI, DL))
4023 return SimplifyCFG(BB, TTI, DL) | true;
4025 // Scan predecessor blocks for conditional branches.
4026 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
4027 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
4028 if (PBI != BI && PBI->isConditional())
4029 if (SimplifyCondBranchToCondBranch(PBI, BI))
4030 return SimplifyCFG(BB, TTI, DL) | true;
4035 /// Check if passing a value to an instruction will cause undefined behavior.
4036 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
4037 Constant *C = dyn_cast<Constant>(V);
4044 if (C->isNullValue()) {
4045 // Only look at the first use, avoid hurting compile time with long uselists
4046 User *Use = *I->use_begin();
4048 // Now make sure that there are no instructions in between that can alter
4049 // control flow (eg. calls)
4050 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
4051 if (i == I->getParent()->end() || i->mayHaveSideEffects())
4054 // Look through GEPs. A load from a GEP derived from NULL is still undefined
4055 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
4056 if (GEP->getPointerOperand() == I)
4057 return passingValueIsAlwaysUndefined(V, GEP);
4059 // Look through bitcasts.
4060 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
4061 return passingValueIsAlwaysUndefined(V, BC);
4063 // Load from null is undefined.
4064 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
4065 if (!LI->isVolatile())
4066 return LI->getPointerAddressSpace() == 0;
4068 // Store to null is undefined.
4069 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
4070 if (!SI->isVolatile())
4071 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
4076 /// If BB has an incoming value that will always trigger undefined behavior
4077 /// (eg. null pointer dereference), remove the branch leading here.
4078 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
4079 for (BasicBlock::iterator i = BB->begin();
4080 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
4081 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
4082 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
4083 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
4084 IRBuilder<> Builder(T);
4085 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
4086 BB->removePredecessor(PHI->getIncomingBlock(i));
4087 // Turn uncoditional branches into unreachables and remove the dead
4088 // destination from conditional branches.
4089 if (BI->isUnconditional())
4090 Builder.CreateUnreachable();
4092 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
4093 BI->getSuccessor(0));
4094 BI->eraseFromParent();
4097 // TODO: SwitchInst.
4103 bool SimplifyCFGOpt::run(BasicBlock *BB) {
4104 bool Changed = false;
4106 assert(BB && BB->getParent() && "Block not embedded in function!");
4107 assert(BB->getTerminator() && "Degenerate basic block encountered!");
4109 // Remove basic blocks that have no predecessors (except the entry block)...
4110 // or that just have themself as a predecessor. These are unreachable.
4111 if ((pred_begin(BB) == pred_end(BB) &&
4112 BB != &BB->getParent()->getEntryBlock()) ||
4113 BB->getSinglePredecessor() == BB) {
4114 DEBUG(dbgs() << "Removing BB: \n" << *BB);
4115 DeleteDeadBlock(BB);
4119 // Check to see if we can constant propagate this terminator instruction
4121 Changed |= ConstantFoldTerminator(BB, true);
4123 // Check for and eliminate duplicate PHI nodes in this block.
4124 Changed |= EliminateDuplicatePHINodes(BB);
4126 // Check for and remove branches that will always cause undefined behavior.
4127 Changed |= removeUndefIntroducingPredecessor(BB);
4129 // Merge basic blocks into their predecessor if there is only one distinct
4130 // pred, and if there is only one distinct successor of the predecessor, and
4131 // if there are no PHI nodes.
4133 if (MergeBlockIntoPredecessor(BB))
4136 IRBuilder<> Builder(BB);
4138 // If there is a trivial two-entry PHI node in this basic block, and we can
4139 // eliminate it, do so now.
4140 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
4141 if (PN->getNumIncomingValues() == 2)
4142 Changed |= FoldTwoEntryPHINode(PN, DL);
4144 Builder.SetInsertPoint(BB->getTerminator());
4145 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
4146 if (BI->isUnconditional()) {
4147 if (SimplifyUncondBranch(BI, Builder)) return true;
4149 if (SimplifyCondBranch(BI, Builder)) return true;
4151 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
4152 if (SimplifyReturn(RI, Builder)) return true;
4153 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
4154 if (SimplifyResume(RI, Builder)) return true;
4155 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
4156 if (SimplifySwitch(SI, Builder)) return true;
4157 } else if (UnreachableInst *UI =
4158 dyn_cast<UnreachableInst>(BB->getTerminator())) {
4159 if (SimplifyUnreachable(UI)) return true;
4160 } else if (IndirectBrInst *IBI =
4161 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
4162 if (SimplifyIndirectBr(IBI)) return true;
4168 /// SimplifyCFG - This function is used to do simplification of a CFG. For
4169 /// example, it adjusts branches to branches to eliminate the extra hop, it
4170 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4171 /// of the CFG. It returns true if a modification was made.
4173 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
4174 const DataLayout *DL) {
4175 return SimplifyCFGOpt(TTI, DL).run(BB);