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 (User *U : BBI->users()) {
1594 Instruction *UI = cast<Instruction>(U);
1595 if (UI->getParent() != BB || isa<PHINode>(UI)) return false;
1598 // Looks ok, continue checking.
1604 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1605 /// that is defined in the same block as the branch and if any PHI entries are
1606 /// constants, thread edges corresponding to that entry to be branches to their
1607 /// ultimate destination.
1608 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *DL) {
1609 BasicBlock *BB = BI->getParent();
1610 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1611 // NOTE: we currently cannot transform this case if the PHI node is used
1612 // outside of the block.
1613 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1616 // Degenerate case of a single entry PHI.
1617 if (PN->getNumIncomingValues() == 1) {
1618 FoldSingleEntryPHINodes(PN->getParent());
1622 // Now we know that this block has multiple preds and two succs.
1623 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1625 if (HasNoDuplicateCall(BB)) return false;
1627 // Okay, this is a simple enough basic block. See if any phi values are
1629 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1630 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1631 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue;
1633 // Okay, we now know that all edges from PredBB should be revectored to
1634 // branch to RealDest.
1635 BasicBlock *PredBB = PN->getIncomingBlock(i);
1636 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1638 if (RealDest == BB) continue; // Skip self loops.
1639 // Skip if the predecessor's terminator is an indirect branch.
1640 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1642 // The dest block might have PHI nodes, other predecessors and other
1643 // difficult cases. Instead of being smart about this, just insert a new
1644 // block that jumps to the destination block, effectively splitting
1645 // the edge we are about to create.
1646 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1647 RealDest->getName()+".critedge",
1648 RealDest->getParent(), RealDest);
1649 BranchInst::Create(RealDest, EdgeBB);
1651 // Update PHI nodes.
1652 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1654 // BB may have instructions that are being threaded over. Clone these
1655 // instructions into EdgeBB. We know that there will be no uses of the
1656 // cloned instructions outside of EdgeBB.
1657 BasicBlock::iterator InsertPt = EdgeBB->begin();
1658 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1659 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1660 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1661 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1664 // Clone the instruction.
1665 Instruction *N = BBI->clone();
1666 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1668 // Update operands due to translation.
1669 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1671 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1672 if (PI != TranslateMap.end())
1676 // Check for trivial simplification.
1677 if (Value *V = SimplifyInstruction(N, DL)) {
1678 TranslateMap[BBI] = V;
1679 delete N; // Instruction folded away, don't need actual inst
1681 // Insert the new instruction into its new home.
1682 EdgeBB->getInstList().insert(InsertPt, N);
1683 if (!BBI->use_empty())
1684 TranslateMap[BBI] = N;
1688 // Loop over all of the edges from PredBB to BB, changing them to branch
1689 // to EdgeBB instead.
1690 TerminatorInst *PredBBTI = PredBB->getTerminator();
1691 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1692 if (PredBBTI->getSuccessor(i) == BB) {
1693 BB->removePredecessor(PredBB);
1694 PredBBTI->setSuccessor(i, EdgeBB);
1697 // Recurse, simplifying any other constants.
1698 return FoldCondBranchOnPHI(BI, DL) | true;
1704 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1705 /// PHI node, see if we can eliminate it.
1706 static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *DL) {
1707 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1708 // statement", which has a very simple dominance structure. Basically, we
1709 // are trying to find the condition that is being branched on, which
1710 // subsequently causes this merge to happen. We really want control
1711 // dependence information for this check, but simplifycfg can't keep it up
1712 // to date, and this catches most of the cases we care about anyway.
1713 BasicBlock *BB = PN->getParent();
1714 BasicBlock *IfTrue, *IfFalse;
1715 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1717 // Don't bother if the branch will be constant folded trivially.
1718 isa<ConstantInt>(IfCond))
1721 // Okay, we found that we can merge this two-entry phi node into a select.
1722 // Doing so would require us to fold *all* two entry phi nodes in this block.
1723 // At some point this becomes non-profitable (particularly if the target
1724 // doesn't support cmov's). Only do this transformation if there are two or
1725 // fewer PHI nodes in this block.
1726 unsigned NumPhis = 0;
1727 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1731 // Loop over the PHI's seeing if we can promote them all to select
1732 // instructions. While we are at it, keep track of the instructions
1733 // that need to be moved to the dominating block.
1734 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1735 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1736 MaxCostVal1 = PHINodeFoldingThreshold;
1738 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1739 PHINode *PN = cast<PHINode>(II++);
1740 if (Value *V = SimplifyInstruction(PN, DL)) {
1741 PN->replaceAllUsesWith(V);
1742 PN->eraseFromParent();
1746 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1748 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1753 // If we folded the first phi, PN dangles at this point. Refresh it. If
1754 // we ran out of PHIs then we simplified them all.
1755 PN = dyn_cast<PHINode>(BB->begin());
1756 if (PN == 0) return true;
1758 // Don't fold i1 branches on PHIs which contain binary operators. These can
1759 // often be turned into switches and other things.
1760 if (PN->getType()->isIntegerTy(1) &&
1761 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1762 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1763 isa<BinaryOperator>(IfCond)))
1766 // If we all PHI nodes are promotable, check to make sure that all
1767 // instructions in the predecessor blocks can be promoted as well. If
1768 // not, we won't be able to get rid of the control flow, so it's not
1769 // worth promoting to select instructions.
1770 BasicBlock *DomBlock = 0;
1771 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1772 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1773 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1776 DomBlock = *pred_begin(IfBlock1);
1777 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1778 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1779 // This is not an aggressive instruction that we can promote.
1780 // Because of this, we won't be able to get rid of the control
1781 // flow, so the xform is not worth it.
1786 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1789 DomBlock = *pred_begin(IfBlock2);
1790 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1791 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1792 // This is not an aggressive instruction that we can promote.
1793 // Because of this, we won't be able to get rid of the control
1794 // flow, so the xform is not worth it.
1799 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1800 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1802 // If we can still promote the PHI nodes after this gauntlet of tests,
1803 // do all of the PHI's now.
1804 Instruction *InsertPt = DomBlock->getTerminator();
1805 IRBuilder<true, NoFolder> Builder(InsertPt);
1807 // Move all 'aggressive' instructions, which are defined in the
1808 // conditional parts of the if's up to the dominating block.
1810 DomBlock->getInstList().splice(InsertPt,
1811 IfBlock1->getInstList(), IfBlock1->begin(),
1812 IfBlock1->getTerminator());
1814 DomBlock->getInstList().splice(InsertPt,
1815 IfBlock2->getInstList(), IfBlock2->begin(),
1816 IfBlock2->getTerminator());
1818 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1819 // Change the PHI node into a select instruction.
1820 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1821 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1824 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1825 PN->replaceAllUsesWith(NV);
1827 PN->eraseFromParent();
1830 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1831 // has been flattened. Change DomBlock to jump directly to our new block to
1832 // avoid other simplifycfg's kicking in on the diamond.
1833 TerminatorInst *OldTI = DomBlock->getTerminator();
1834 Builder.SetInsertPoint(OldTI);
1835 Builder.CreateBr(BB);
1836 OldTI->eraseFromParent();
1840 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1841 /// to two returning blocks, try to merge them together into one return,
1842 /// introducing a select if the return values disagree.
1843 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1844 IRBuilder<> &Builder) {
1845 assert(BI->isConditional() && "Must be a conditional branch");
1846 BasicBlock *TrueSucc = BI->getSuccessor(0);
1847 BasicBlock *FalseSucc = BI->getSuccessor(1);
1848 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1849 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1851 // Check to ensure both blocks are empty (just a return) or optionally empty
1852 // with PHI nodes. If there are other instructions, merging would cause extra
1853 // computation on one path or the other.
1854 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1856 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1859 Builder.SetInsertPoint(BI);
1860 // Okay, we found a branch that is going to two return nodes. If
1861 // there is no return value for this function, just change the
1862 // branch into a return.
1863 if (FalseRet->getNumOperands() == 0) {
1864 TrueSucc->removePredecessor(BI->getParent());
1865 FalseSucc->removePredecessor(BI->getParent());
1866 Builder.CreateRetVoid();
1867 EraseTerminatorInstAndDCECond(BI);
1871 // Otherwise, figure out what the true and false return values are
1872 // so we can insert a new select instruction.
1873 Value *TrueValue = TrueRet->getReturnValue();
1874 Value *FalseValue = FalseRet->getReturnValue();
1876 // Unwrap any PHI nodes in the return blocks.
1877 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1878 if (TVPN->getParent() == TrueSucc)
1879 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1880 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1881 if (FVPN->getParent() == FalseSucc)
1882 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1884 // In order for this transformation to be safe, we must be able to
1885 // unconditionally execute both operands to the return. This is
1886 // normally the case, but we could have a potentially-trapping
1887 // constant expression that prevents this transformation from being
1889 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1892 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1896 // Okay, we collected all the mapped values and checked them for sanity, and
1897 // defined to really do this transformation. First, update the CFG.
1898 TrueSucc->removePredecessor(BI->getParent());
1899 FalseSucc->removePredecessor(BI->getParent());
1901 // Insert select instructions where needed.
1902 Value *BrCond = BI->getCondition();
1904 // Insert a select if the results differ.
1905 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1906 } else if (isa<UndefValue>(TrueValue)) {
1907 TrueValue = FalseValue;
1909 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1910 FalseValue, "retval");
1914 Value *RI = !TrueValue ?
1915 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
1919 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
1920 << "\n " << *BI << "NewRet = " << *RI
1921 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
1923 EraseTerminatorInstAndDCECond(BI);
1928 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
1929 /// probabilities of the branch taking each edge. Fills in the two APInt
1930 /// parameters and return true, or returns false if no or invalid metadata was
1932 static bool ExtractBranchMetadata(BranchInst *BI,
1933 uint64_t &ProbTrue, uint64_t &ProbFalse) {
1934 assert(BI->isConditional() &&
1935 "Looking for probabilities on unconditional branch?");
1936 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
1937 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
1938 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
1939 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
1940 if (!CITrue || !CIFalse) return false;
1941 ProbTrue = CITrue->getValue().getZExtValue();
1942 ProbFalse = CIFalse->getValue().getZExtValue();
1946 /// checkCSEInPredecessor - Return true if the given instruction is available
1947 /// in its predecessor block. If yes, the instruction will be removed.
1949 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
1950 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
1952 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
1953 Instruction *PBI = &*I;
1954 // Check whether Inst and PBI generate the same value.
1955 if (Inst->isIdenticalTo(PBI)) {
1956 Inst->replaceAllUsesWith(PBI);
1957 Inst->eraseFromParent();
1964 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
1965 /// predecessor branches to us and one of our successors, fold the block into
1966 /// the predecessor and use logical operations to pick the right destination.
1967 bool llvm::FoldBranchToCommonDest(BranchInst *BI) {
1968 BasicBlock *BB = BI->getParent();
1970 Instruction *Cond = 0;
1971 if (BI->isConditional())
1972 Cond = dyn_cast<Instruction>(BI->getCondition());
1974 // For unconditional branch, check for a simple CFG pattern, where
1975 // BB has a single predecessor and BB's successor is also its predecessor's
1976 // successor. If such pattern exisits, check for CSE between BB and its
1978 if (BasicBlock *PB = BB->getSinglePredecessor())
1979 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
1980 if (PBI->isConditional() &&
1981 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
1982 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
1983 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
1985 Instruction *Curr = I++;
1986 if (isa<CmpInst>(Curr)) {
1990 // Quit if we can't remove this instruction.
1991 if (!checkCSEInPredecessor(Curr, PB))
2000 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2001 Cond->getParent() != BB || !Cond->hasOneUse())
2004 // Only allow this if the condition is a simple instruction that can be
2005 // executed unconditionally. It must be in the same block as the branch, and
2006 // must be at the front of the block.
2007 BasicBlock::iterator FrontIt = BB->front();
2009 // Ignore dbg intrinsics.
2010 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
2012 // Allow a single instruction to be hoisted in addition to the compare
2013 // that feeds the branch. We later ensure that any values that _it_ uses
2014 // were also live in the predecessor, so that we don't unnecessarily create
2015 // register pressure or inhibit out-of-order execution.
2016 Instruction *BonusInst = 0;
2017 if (&*FrontIt != Cond &&
2018 FrontIt->hasOneUse() && FrontIt->user_back() == Cond &&
2019 isSafeToSpeculativelyExecute(FrontIt)) {
2020 BonusInst = &*FrontIt;
2023 // Ignore dbg intrinsics.
2024 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
2027 // Only a single bonus inst is allowed.
2028 if (&*FrontIt != Cond)
2031 // Make sure the instruction after the condition is the cond branch.
2032 BasicBlock::iterator CondIt = Cond; ++CondIt;
2034 // Ingore dbg intrinsics.
2035 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
2040 // Cond is known to be a compare or binary operator. Check to make sure that
2041 // neither operand is a potentially-trapping constant expression.
2042 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2045 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2049 // Finally, don't infinitely unroll conditional loops.
2050 BasicBlock *TrueDest = BI->getSuccessor(0);
2051 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0;
2052 if (TrueDest == BB || FalseDest == BB)
2055 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2056 BasicBlock *PredBlock = *PI;
2057 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2059 // Check that we have two conditional branches. If there is a PHI node in
2060 // the common successor, verify that the same value flows in from both
2062 SmallVector<PHINode*, 4> PHIs;
2063 if (PBI == 0 || PBI->isUnconditional() ||
2064 (BI->isConditional() &&
2065 !SafeToMergeTerminators(BI, PBI)) ||
2066 (!BI->isConditional() &&
2067 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2070 // Determine if the two branches share a common destination.
2071 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2072 bool InvertPredCond = false;
2074 if (BI->isConditional()) {
2075 if (PBI->getSuccessor(0) == TrueDest)
2076 Opc = Instruction::Or;
2077 else if (PBI->getSuccessor(1) == FalseDest)
2078 Opc = Instruction::And;
2079 else if (PBI->getSuccessor(0) == FalseDest)
2080 Opc = Instruction::And, InvertPredCond = true;
2081 else if (PBI->getSuccessor(1) == TrueDest)
2082 Opc = Instruction::Or, InvertPredCond = true;
2086 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2090 // Ensure that any values used in the bonus instruction are also used
2091 // by the terminator of the predecessor. This means that those values
2092 // must already have been resolved, so we won't be inhibiting the
2093 // out-of-order core by speculating them earlier. We also allow
2094 // instructions that are used by the terminator's condition because it
2095 // exposes more merging opportunities.
2096 bool UsedByBranch = (BonusInst && BonusInst->hasOneUse() &&
2097 BonusInst->user_back() == Cond);
2099 if (BonusInst && !UsedByBranch) {
2100 // Collect the values used by the bonus inst
2101 SmallPtrSet<Value*, 4> UsedValues;
2102 for (Instruction::op_iterator OI = BonusInst->op_begin(),
2103 OE = BonusInst->op_end(); OI != OE; ++OI) {
2105 if (!isa<Constant>(V) && !isa<Argument>(V))
2106 UsedValues.insert(V);
2109 SmallVector<std::pair<Value*, unsigned>, 4> Worklist;
2110 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0));
2112 // Walk up to four levels back up the use-def chain of the predecessor's
2113 // terminator to see if all those values were used. The choice of four
2114 // levels is arbitrary, to provide a compile-time-cost bound.
2115 while (!Worklist.empty()) {
2116 std::pair<Value*, unsigned> Pair = Worklist.back();
2117 Worklist.pop_back();
2119 if (Pair.second >= 4) continue;
2120 UsedValues.erase(Pair.first);
2121 if (UsedValues.empty()) break;
2123 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) {
2124 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
2126 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1));
2130 if (!UsedValues.empty()) return false;
2133 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2134 IRBuilder<> Builder(PBI);
2136 // If we need to invert the condition in the pred block to match, do so now.
2137 if (InvertPredCond) {
2138 Value *NewCond = PBI->getCondition();
2140 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2141 CmpInst *CI = cast<CmpInst>(NewCond);
2142 CI->setPredicate(CI->getInversePredicate());
2144 NewCond = Builder.CreateNot(NewCond,
2145 PBI->getCondition()->getName()+".not");
2148 PBI->setCondition(NewCond);
2149 PBI->swapSuccessors();
2152 // If we have a bonus inst, clone it into the predecessor block.
2153 Instruction *NewBonus = 0;
2155 NewBonus = BonusInst->clone();
2157 // If we moved a load, we cannot any longer claim any knowledge about
2158 // its potential value. The previous information might have been valid
2159 // only given the branch precondition.
2160 // For an analogous reason, we must also drop all the metadata whose
2161 // semantics we don't understand.
2162 NewBonus->dropUnknownMetadata(LLVMContext::MD_dbg);
2164 PredBlock->getInstList().insert(PBI, NewBonus);
2165 NewBonus->takeName(BonusInst);
2166 BonusInst->setName(BonusInst->getName()+".old");
2169 // Clone Cond into the predecessor basic block, and or/and the
2170 // two conditions together.
2171 Instruction *New = Cond->clone();
2172 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus);
2173 PredBlock->getInstList().insert(PBI, New);
2174 New->takeName(Cond);
2175 Cond->setName(New->getName()+".old");
2177 if (BI->isConditional()) {
2178 Instruction *NewCond =
2179 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2181 PBI->setCondition(NewCond);
2183 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2184 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2186 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2188 SmallVector<uint64_t, 8> NewWeights;
2190 if (PBI->getSuccessor(0) == BB) {
2191 if (PredHasWeights && SuccHasWeights) {
2192 // PBI: br i1 %x, BB, FalseDest
2193 // BI: br i1 %y, TrueDest, FalseDest
2194 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2195 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2196 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2197 // TrueWeight for PBI * FalseWeight for BI.
2198 // We assume that total weights of a BranchInst can fit into 32 bits.
2199 // Therefore, we will not have overflow using 64-bit arithmetic.
2200 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2201 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2203 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2204 PBI->setSuccessor(0, TrueDest);
2206 if (PBI->getSuccessor(1) == BB) {
2207 if (PredHasWeights && SuccHasWeights) {
2208 // PBI: br i1 %x, TrueDest, BB
2209 // BI: br i1 %y, TrueDest, FalseDest
2210 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2211 // FalseWeight for PBI * TrueWeight for BI.
2212 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2213 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2214 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2215 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2217 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2218 PBI->setSuccessor(1, FalseDest);
2220 if (NewWeights.size() == 2) {
2221 // Halve the weights if any of them cannot fit in an uint32_t
2222 FitWeights(NewWeights);
2224 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2225 PBI->setMetadata(LLVMContext::MD_prof,
2226 MDBuilder(BI->getContext()).
2227 createBranchWeights(MDWeights));
2229 PBI->setMetadata(LLVMContext::MD_prof, NULL);
2231 // Update PHI nodes in the common successors.
2232 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2233 ConstantInt *PBI_C = cast<ConstantInt>(
2234 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2235 assert(PBI_C->getType()->isIntegerTy(1));
2236 Instruction *MergedCond = 0;
2237 if (PBI->getSuccessor(0) == TrueDest) {
2238 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2239 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2240 // is false: !PBI_Cond and BI_Value
2241 Instruction *NotCond =
2242 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2245 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2250 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2251 PBI->getCondition(), MergedCond,
2254 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2255 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2256 // is false: PBI_Cond and BI_Value
2258 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2259 PBI->getCondition(), New,
2261 if (PBI_C->isOne()) {
2262 Instruction *NotCond =
2263 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2266 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2267 NotCond, MergedCond,
2272 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2275 // Change PBI from Conditional to Unconditional.
2276 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2277 EraseTerminatorInstAndDCECond(PBI);
2281 // TODO: If BB is reachable from all paths through PredBlock, then we
2282 // could replace PBI's branch probabilities with BI's.
2284 // Copy any debug value intrinsics into the end of PredBlock.
2285 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2286 if (isa<DbgInfoIntrinsic>(*I))
2287 I->clone()->insertBefore(PBI);
2294 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2295 /// predecessor of another block, this function tries to simplify it. We know
2296 /// that PBI and BI are both conditional branches, and BI is in one of the
2297 /// successor blocks of PBI - PBI branches to BI.
2298 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2299 assert(PBI->isConditional() && BI->isConditional());
2300 BasicBlock *BB = BI->getParent();
2302 // If this block ends with a branch instruction, and if there is a
2303 // predecessor that ends on a branch of the same condition, make
2304 // this conditional branch redundant.
2305 if (PBI->getCondition() == BI->getCondition() &&
2306 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2307 // Okay, the outcome of this conditional branch is statically
2308 // knowable. If this block had a single pred, handle specially.
2309 if (BB->getSinglePredecessor()) {
2310 // Turn this into a branch on constant.
2311 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2312 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2314 return true; // Nuke the branch on constant.
2317 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2318 // in the constant and simplify the block result. Subsequent passes of
2319 // simplifycfg will thread the block.
2320 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2321 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2322 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2323 std::distance(PB, PE),
2324 BI->getCondition()->getName() + ".pr",
2326 // Okay, we're going to insert the PHI node. Since PBI is not the only
2327 // predecessor, compute the PHI'd conditional value for all of the preds.
2328 // Any predecessor where the condition is not computable we keep symbolic.
2329 for (pred_iterator PI = PB; PI != PE; ++PI) {
2330 BasicBlock *P = *PI;
2331 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2332 PBI != BI && PBI->isConditional() &&
2333 PBI->getCondition() == BI->getCondition() &&
2334 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2335 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2336 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2339 NewPN->addIncoming(BI->getCondition(), P);
2343 BI->setCondition(NewPN);
2348 // If this is a conditional branch in an empty block, and if any
2349 // predecessors is a conditional branch to one of our destinations,
2350 // fold the conditions into logical ops and one cond br.
2351 BasicBlock::iterator BBI = BB->begin();
2352 // Ignore dbg intrinsics.
2353 while (isa<DbgInfoIntrinsic>(BBI))
2359 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2364 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2366 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2367 PBIOp = 0, BIOp = 1;
2368 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2369 PBIOp = 1, BIOp = 0;
2370 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2375 // Check to make sure that the other destination of this branch
2376 // isn't BB itself. If so, this is an infinite loop that will
2377 // keep getting unwound.
2378 if (PBI->getSuccessor(PBIOp) == BB)
2381 // Do not perform this transformation if it would require
2382 // insertion of a large number of select instructions. For targets
2383 // without predication/cmovs, this is a big pessimization.
2384 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2386 unsigned NumPhis = 0;
2387 for (BasicBlock::iterator II = CommonDest->begin();
2388 isa<PHINode>(II); ++II, ++NumPhis)
2389 if (NumPhis > 2) // Disable this xform.
2392 // Finally, if everything is ok, fold the branches to logical ops.
2393 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2395 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2396 << "AND: " << *BI->getParent());
2399 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2400 // branch in it, where one edge (OtherDest) goes back to itself but the other
2401 // exits. We don't *know* that the program avoids the infinite loop
2402 // (even though that seems likely). If we do this xform naively, we'll end up
2403 // recursively unpeeling the loop. Since we know that (after the xform is
2404 // done) that the block *is* infinite if reached, we just make it an obviously
2405 // infinite loop with no cond branch.
2406 if (OtherDest == BB) {
2407 // Insert it at the end of the function, because it's either code,
2408 // or it won't matter if it's hot. :)
2409 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2410 "infloop", BB->getParent());
2411 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2412 OtherDest = InfLoopBlock;
2415 DEBUG(dbgs() << *PBI->getParent()->getParent());
2417 // BI may have other predecessors. Because of this, we leave
2418 // it alone, but modify PBI.
2420 // Make sure we get to CommonDest on True&True directions.
2421 Value *PBICond = PBI->getCondition();
2422 IRBuilder<true, NoFolder> Builder(PBI);
2424 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2426 Value *BICond = BI->getCondition();
2428 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2430 // Merge the conditions.
2431 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2433 // Modify PBI to branch on the new condition to the new dests.
2434 PBI->setCondition(Cond);
2435 PBI->setSuccessor(0, CommonDest);
2436 PBI->setSuccessor(1, OtherDest);
2438 // Update branch weight for PBI.
2439 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2440 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2442 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2444 if (PredHasWeights && SuccHasWeights) {
2445 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2446 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2447 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2448 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2449 // The weight to CommonDest should be PredCommon * SuccTotal +
2450 // PredOther * SuccCommon.
2451 // The weight to OtherDest should be PredOther * SuccOther.
2452 SmallVector<uint64_t, 2> NewWeights;
2453 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) +
2454 PredOther * SuccCommon);
2455 NewWeights.push_back(PredOther * SuccOther);
2456 // Halve the weights if any of them cannot fit in an uint32_t
2457 FitWeights(NewWeights);
2459 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end());
2460 PBI->setMetadata(LLVMContext::MD_prof,
2461 MDBuilder(BI->getContext()).
2462 createBranchWeights(MDWeights));
2465 // OtherDest may have phi nodes. If so, add an entry from PBI's
2466 // block that are identical to the entries for BI's block.
2467 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2469 // We know that the CommonDest already had an edge from PBI to
2470 // it. If it has PHIs though, the PHIs may have different
2471 // entries for BB and PBI's BB. If so, insert a select to make
2474 for (BasicBlock::iterator II = CommonDest->begin();
2475 (PN = dyn_cast<PHINode>(II)); ++II) {
2476 Value *BIV = PN->getIncomingValueForBlock(BB);
2477 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2478 Value *PBIV = PN->getIncomingValue(PBBIdx);
2480 // Insert a select in PBI to pick the right value.
2481 Value *NV = cast<SelectInst>
2482 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2483 PN->setIncomingValue(PBBIdx, NV);
2487 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2488 DEBUG(dbgs() << *PBI->getParent()->getParent());
2490 // This basic block is probably dead. We know it has at least
2491 // one fewer predecessor.
2495 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2496 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2497 // Takes care of updating the successors and removing the old terminator.
2498 // Also makes sure not to introduce new successors by assuming that edges to
2499 // non-successor TrueBBs and FalseBBs aren't reachable.
2500 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2501 BasicBlock *TrueBB, BasicBlock *FalseBB,
2502 uint32_t TrueWeight,
2503 uint32_t FalseWeight){
2504 // Remove any superfluous successor edges from the CFG.
2505 // First, figure out which successors to preserve.
2506 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2508 BasicBlock *KeepEdge1 = TrueBB;
2509 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0;
2511 // Then remove the rest.
2512 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2513 BasicBlock *Succ = OldTerm->getSuccessor(I);
2514 // Make sure only to keep exactly one copy of each edge.
2515 if (Succ == KeepEdge1)
2517 else if (Succ == KeepEdge2)
2520 Succ->removePredecessor(OldTerm->getParent());
2523 IRBuilder<> Builder(OldTerm);
2524 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2526 // Insert an appropriate new terminator.
2527 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) {
2528 if (TrueBB == FalseBB)
2529 // We were only looking for one successor, and it was present.
2530 // Create an unconditional branch to it.
2531 Builder.CreateBr(TrueBB);
2533 // We found both of the successors we were looking for.
2534 // Create a conditional branch sharing the condition of the select.
2535 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2536 if (TrueWeight != FalseWeight)
2537 NewBI->setMetadata(LLVMContext::MD_prof,
2538 MDBuilder(OldTerm->getContext()).
2539 createBranchWeights(TrueWeight, FalseWeight));
2541 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2542 // Neither of the selected blocks were successors, so this
2543 // terminator must be unreachable.
2544 new UnreachableInst(OldTerm->getContext(), OldTerm);
2546 // One of the selected values was a successor, but the other wasn't.
2547 // Insert an unconditional branch to the one that was found;
2548 // the edge to the one that wasn't must be unreachable.
2550 // Only TrueBB was found.
2551 Builder.CreateBr(TrueBB);
2553 // Only FalseBB was found.
2554 Builder.CreateBr(FalseBB);
2557 EraseTerminatorInstAndDCECond(OldTerm);
2561 // SimplifySwitchOnSelect - Replaces
2562 // (switch (select cond, X, Y)) on constant X, Y
2563 // with a branch - conditional if X and Y lead to distinct BBs,
2564 // unconditional otherwise.
2565 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2566 // Check for constant integer values in the select.
2567 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2568 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2569 if (!TrueVal || !FalseVal)
2572 // Find the relevant condition and destinations.
2573 Value *Condition = Select->getCondition();
2574 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2575 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2577 // Get weight for TrueBB and FalseBB.
2578 uint32_t TrueWeight = 0, FalseWeight = 0;
2579 SmallVector<uint64_t, 8> Weights;
2580 bool HasWeights = HasBranchWeights(SI);
2582 GetBranchWeights(SI, Weights);
2583 if (Weights.size() == 1 + SI->getNumCases()) {
2584 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2585 getSuccessorIndex()];
2586 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2587 getSuccessorIndex()];
2591 // Perform the actual simplification.
2592 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2593 TrueWeight, FalseWeight);
2596 // SimplifyIndirectBrOnSelect - Replaces
2597 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2598 // blockaddress(@fn, BlockB)))
2600 // (br cond, BlockA, BlockB).
2601 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2602 // Check that both operands of the select are block addresses.
2603 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2604 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2608 // Extract the actual blocks.
2609 BasicBlock *TrueBB = TBA->getBasicBlock();
2610 BasicBlock *FalseBB = FBA->getBasicBlock();
2612 // Perform the actual simplification.
2613 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2617 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2618 /// instruction (a seteq/setne with a constant) as the only instruction in a
2619 /// block that ends with an uncond branch. We are looking for a very specific
2620 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2621 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2622 /// default value goes to an uncond block with a seteq in it, we get something
2625 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2627 /// %tmp = icmp eq i8 %A, 92
2630 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2632 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2633 /// the PHI, merging the third icmp into the switch.
2634 static bool TryToSimplifyUncondBranchWithICmpInIt(
2635 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI,
2636 const DataLayout *DL) {
2637 BasicBlock *BB = ICI->getParent();
2639 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2641 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2643 Value *V = ICI->getOperand(0);
2644 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2646 // The pattern we're looking for is where our only predecessor is a switch on
2647 // 'V' and this block is the default case for the switch. In this case we can
2648 // fold the compared value into the switch to simplify things.
2649 BasicBlock *Pred = BB->getSinglePredecessor();
2650 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false;
2652 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2653 if (SI->getCondition() != V)
2656 // If BB is reachable on a non-default case, then we simply know the value of
2657 // V in this block. Substitute it and constant fold the icmp instruction
2659 if (SI->getDefaultDest() != BB) {
2660 ConstantInt *VVal = SI->findCaseDest(BB);
2661 assert(VVal && "Should have a unique destination value");
2662 ICI->setOperand(0, VVal);
2664 if (Value *V = SimplifyInstruction(ICI, DL)) {
2665 ICI->replaceAllUsesWith(V);
2666 ICI->eraseFromParent();
2668 // BB is now empty, so it is likely to simplify away.
2669 return SimplifyCFG(BB, TTI, DL) | true;
2672 // Ok, the block is reachable from the default dest. If the constant we're
2673 // comparing exists in one of the other edges, then we can constant fold ICI
2675 if (SI->findCaseValue(Cst) != SI->case_default()) {
2677 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2678 V = ConstantInt::getFalse(BB->getContext());
2680 V = ConstantInt::getTrue(BB->getContext());
2682 ICI->replaceAllUsesWith(V);
2683 ICI->eraseFromParent();
2684 // BB is now empty, so it is likely to simplify away.
2685 return SimplifyCFG(BB, TTI, DL) | true;
2688 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2690 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2691 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
2692 if (PHIUse == 0 || PHIUse != &SuccBlock->front() ||
2693 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2696 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2698 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2699 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2701 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2702 std::swap(DefaultCst, NewCst);
2704 // Replace ICI (which is used by the PHI for the default value) with true or
2705 // false depending on if it is EQ or NE.
2706 ICI->replaceAllUsesWith(DefaultCst);
2707 ICI->eraseFromParent();
2709 // Okay, the switch goes to this block on a default value. Add an edge from
2710 // the switch to the merge point on the compared value.
2711 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2712 BB->getParent(), BB);
2713 SmallVector<uint64_t, 8> Weights;
2714 bool HasWeights = HasBranchWeights(SI);
2716 GetBranchWeights(SI, Weights);
2717 if (Weights.size() == 1 + SI->getNumCases()) {
2718 // Split weight for default case to case for "Cst".
2719 Weights[0] = (Weights[0]+1) >> 1;
2720 Weights.push_back(Weights[0]);
2722 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2723 SI->setMetadata(LLVMContext::MD_prof,
2724 MDBuilder(SI->getContext()).
2725 createBranchWeights(MDWeights));
2728 SI->addCase(Cst, NewBB);
2730 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2731 Builder.SetInsertPoint(NewBB);
2732 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2733 Builder.CreateBr(SuccBlock);
2734 PHIUse->addIncoming(NewCst, NewBB);
2738 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2739 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2740 /// fold it into a switch instruction if so.
2741 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *DL,
2742 IRBuilder<> &Builder) {
2743 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2744 if (Cond == 0) return false;
2747 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2748 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2749 // 'setne's and'ed together, collect them.
2751 std::vector<ConstantInt*> Values;
2752 bool TrueWhenEqual = true;
2753 Value *ExtraCase = 0;
2754 unsigned UsedICmps = 0;
2756 if (Cond->getOpcode() == Instruction::Or) {
2757 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, DL, true,
2759 } else if (Cond->getOpcode() == Instruction::And) {
2760 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, DL, false,
2762 TrueWhenEqual = false;
2765 // If we didn't have a multiply compared value, fail.
2766 if (CompVal == 0) return false;
2768 // Avoid turning single icmps into a switch.
2772 // There might be duplicate constants in the list, which the switch
2773 // instruction can't handle, remove them now.
2774 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2775 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2777 // If Extra was used, we require at least two switch values to do the
2778 // transformation. A switch with one value is just an cond branch.
2779 if (ExtraCase && Values.size() < 2) return false;
2781 // TODO: Preserve branch weight metadata, similarly to how
2782 // FoldValueComparisonIntoPredecessors preserves it.
2784 // Figure out which block is which destination.
2785 BasicBlock *DefaultBB = BI->getSuccessor(1);
2786 BasicBlock *EdgeBB = BI->getSuccessor(0);
2787 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2789 BasicBlock *BB = BI->getParent();
2791 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2792 << " cases into SWITCH. BB is:\n" << *BB);
2794 // If there are any extra values that couldn't be folded into the switch
2795 // then we evaluate them with an explicit branch first. Split the block
2796 // right before the condbr to handle it.
2798 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2799 // Remove the uncond branch added to the old block.
2800 TerminatorInst *OldTI = BB->getTerminator();
2801 Builder.SetInsertPoint(OldTI);
2804 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2806 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2808 OldTI->eraseFromParent();
2810 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2811 // for the edge we just added.
2812 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2814 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2815 << "\nEXTRABB = " << *BB);
2819 Builder.SetInsertPoint(BI);
2820 // Convert pointer to int before we switch.
2821 if (CompVal->getType()->isPointerTy()) {
2822 assert(DL && "Cannot switch on pointer without DataLayout");
2823 CompVal = Builder.CreatePtrToInt(CompVal,
2824 DL->getIntPtrType(CompVal->getType()),
2828 // Create the new switch instruction now.
2829 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2831 // Add all of the 'cases' to the switch instruction.
2832 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2833 New->addCase(Values[i], EdgeBB);
2835 // We added edges from PI to the EdgeBB. As such, if there were any
2836 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2837 // the number of edges added.
2838 for (BasicBlock::iterator BBI = EdgeBB->begin();
2839 isa<PHINode>(BBI); ++BBI) {
2840 PHINode *PN = cast<PHINode>(BBI);
2841 Value *InVal = PN->getIncomingValueForBlock(BB);
2842 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2843 PN->addIncoming(InVal, BB);
2846 // Erase the old branch instruction.
2847 EraseTerminatorInstAndDCECond(BI);
2849 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2853 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2854 // If this is a trivial landing pad that just continues unwinding the caught
2855 // exception then zap the landing pad, turning its invokes into calls.
2856 BasicBlock *BB = RI->getParent();
2857 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2858 if (RI->getValue() != LPInst)
2859 // Not a landing pad, or the resume is not unwinding the exception that
2860 // caused control to branch here.
2863 // Check that there are no other instructions except for debug intrinsics.
2864 BasicBlock::iterator I = LPInst, E = RI;
2866 if (!isa<DbgInfoIntrinsic>(I))
2869 // Turn all invokes that unwind here into calls and delete the basic block.
2870 bool InvokeRequiresTableEntry = false;
2871 bool Changed = false;
2872 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2873 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2875 if (II->hasFnAttr(Attribute::UWTable)) {
2876 // Don't remove an `invoke' instruction if the ABI requires an entry into
2878 InvokeRequiresTableEntry = true;
2882 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2884 // Insert a call instruction before the invoke.
2885 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2887 Call->setCallingConv(II->getCallingConv());
2888 Call->setAttributes(II->getAttributes());
2889 Call->setDebugLoc(II->getDebugLoc());
2891 // Anything that used the value produced by the invoke instruction now uses
2892 // the value produced by the call instruction. Note that we do this even
2893 // for void functions and calls with no uses so that the callgraph edge is
2895 II->replaceAllUsesWith(Call);
2896 BB->removePredecessor(II->getParent());
2898 // Insert a branch to the normal destination right before the invoke.
2899 BranchInst::Create(II->getNormalDest(), II);
2901 // Finally, delete the invoke instruction!
2902 II->eraseFromParent();
2906 if (!InvokeRequiresTableEntry)
2907 // The landingpad is now unreachable. Zap it.
2908 BB->eraseFromParent();
2913 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2914 BasicBlock *BB = RI->getParent();
2915 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2917 // Find predecessors that end with branches.
2918 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2919 SmallVector<BranchInst*, 8> CondBranchPreds;
2920 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2921 BasicBlock *P = *PI;
2922 TerminatorInst *PTI = P->getTerminator();
2923 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2924 if (BI->isUnconditional())
2925 UncondBranchPreds.push_back(P);
2927 CondBranchPreds.push_back(BI);
2931 // If we found some, do the transformation!
2932 if (!UncondBranchPreds.empty() && DupRet) {
2933 while (!UncondBranchPreds.empty()) {
2934 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
2935 DEBUG(dbgs() << "FOLDING: " << *BB
2936 << "INTO UNCOND BRANCH PRED: " << *Pred);
2937 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
2940 // If we eliminated all predecessors of the block, delete the block now.
2941 if (pred_begin(BB) == pred_end(BB))
2942 // We know there are no successors, so just nuke the block.
2943 BB->eraseFromParent();
2948 // Check out all of the conditional branches going to this return
2949 // instruction. If any of them just select between returns, change the
2950 // branch itself into a select/return pair.
2951 while (!CondBranchPreds.empty()) {
2952 BranchInst *BI = CondBranchPreds.pop_back_val();
2954 // Check to see if the non-BB successor is also a return block.
2955 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
2956 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
2957 SimplifyCondBranchToTwoReturns(BI, Builder))
2963 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
2964 BasicBlock *BB = UI->getParent();
2966 bool Changed = false;
2968 // If there are any instructions immediately before the unreachable that can
2969 // be removed, do so.
2970 while (UI != BB->begin()) {
2971 BasicBlock::iterator BBI = UI;
2973 // Do not delete instructions that can have side effects which might cause
2974 // the unreachable to not be reachable; specifically, calls and volatile
2975 // operations may have this effect.
2976 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
2978 if (BBI->mayHaveSideEffects()) {
2979 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
2980 if (SI->isVolatile())
2982 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
2983 if (LI->isVolatile())
2985 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
2986 if (RMWI->isVolatile())
2988 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
2989 if (CXI->isVolatile())
2991 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
2992 !isa<LandingPadInst>(BBI)) {
2995 // Note that deleting LandingPad's here is in fact okay, although it
2996 // involves a bit of subtle reasoning. If this inst is a LandingPad,
2997 // all the predecessors of this block will be the unwind edges of Invokes,
2998 // and we can therefore guarantee this block will be erased.
3001 // Delete this instruction (any uses are guaranteed to be dead)
3002 if (!BBI->use_empty())
3003 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
3004 BBI->eraseFromParent();
3008 // If the unreachable instruction is the first in the block, take a gander
3009 // at all of the predecessors of this instruction, and simplify them.
3010 if (&BB->front() != UI) return Changed;
3012 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
3013 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
3014 TerminatorInst *TI = Preds[i]->getTerminator();
3015 IRBuilder<> Builder(TI);
3016 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3017 if (BI->isUnconditional()) {
3018 if (BI->getSuccessor(0) == BB) {
3019 new UnreachableInst(TI->getContext(), TI);
3020 TI->eraseFromParent();
3024 if (BI->getSuccessor(0) == BB) {
3025 Builder.CreateBr(BI->getSuccessor(1));
3026 EraseTerminatorInstAndDCECond(BI);
3027 } else if (BI->getSuccessor(1) == BB) {
3028 Builder.CreateBr(BI->getSuccessor(0));
3029 EraseTerminatorInstAndDCECond(BI);
3033 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3034 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3036 if (i.getCaseSuccessor() == BB) {
3037 BB->removePredecessor(SI->getParent());
3042 // If the default value is unreachable, figure out the most popular
3043 // destination and make it the default.
3044 if (SI->getDefaultDest() == BB) {
3045 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
3046 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3048 std::pair<unsigned, unsigned> &entry =
3049 Popularity[i.getCaseSuccessor()];
3050 if (entry.first == 0) {
3052 entry.second = i.getCaseIndex();
3058 // Find the most popular block.
3059 unsigned MaxPop = 0;
3060 unsigned MaxIndex = 0;
3061 BasicBlock *MaxBlock = 0;
3062 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
3063 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
3064 if (I->second.first > MaxPop ||
3065 (I->second.first == MaxPop && MaxIndex > I->second.second)) {
3066 MaxPop = I->second.first;
3067 MaxIndex = I->second.second;
3068 MaxBlock = I->first;
3072 // Make this the new default, allowing us to delete any explicit
3074 SI->setDefaultDest(MaxBlock);
3077 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
3079 if (isa<PHINode>(MaxBlock->begin()))
3080 for (unsigned i = 0; i != MaxPop-1; ++i)
3081 MaxBlock->removePredecessor(SI->getParent());
3083 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3085 if (i.getCaseSuccessor() == MaxBlock) {
3091 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
3092 if (II->getUnwindDest() == BB) {
3093 // Convert the invoke to a call instruction. This would be a good
3094 // place to note that the call does not throw though.
3095 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
3096 II->removeFromParent(); // Take out of symbol table
3098 // Insert the call now...
3099 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
3100 Builder.SetInsertPoint(BI);
3101 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
3102 Args, II->getName());
3103 CI->setCallingConv(II->getCallingConv());
3104 CI->setAttributes(II->getAttributes());
3105 // If the invoke produced a value, the call does now instead.
3106 II->replaceAllUsesWith(CI);
3113 // If this block is now dead, remove it.
3114 if (pred_begin(BB) == pred_end(BB) &&
3115 BB != &BB->getParent()->getEntryBlock()) {
3116 // We know there are no successors, so just nuke the block.
3117 BB->eraseFromParent();
3124 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
3125 /// integer range comparison into a sub, an icmp and a branch.
3126 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3127 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3129 // Make sure all cases point to the same destination and gather the values.
3130 SmallVector<ConstantInt *, 16> Cases;
3131 SwitchInst::CaseIt I = SI->case_begin();
3132 Cases.push_back(I.getCaseValue());
3133 SwitchInst::CaseIt PrevI = I++;
3134 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
3135 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
3137 Cases.push_back(I.getCaseValue());
3139 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
3141 // Sort the case values, then check if they form a range we can transform.
3142 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3143 for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
3144 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
3148 Constant *Offset = ConstantExpr::getNeg(Cases.back());
3149 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
3151 Value *Sub = SI->getCondition();
3152 if (!Offset->isNullValue())
3153 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
3155 // If NumCases overflowed, then all possible values jump to the successor.
3156 if (NumCases->isNullValue() && SI->getNumCases() != 0)
3157 Cmp = ConstantInt::getTrue(SI->getContext());
3159 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3160 BranchInst *NewBI = Builder.CreateCondBr(
3161 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
3163 // Update weight for the newly-created conditional branch.
3164 SmallVector<uint64_t, 8> Weights;
3165 bool HasWeights = HasBranchWeights(SI);
3167 GetBranchWeights(SI, Weights);
3168 if (Weights.size() == 1 + SI->getNumCases()) {
3169 // Combine all weights for the cases to be the true weight of NewBI.
3170 // We assume that the sum of all weights for a Terminator can fit into 32
3172 uint32_t NewTrueWeight = 0;
3173 for (unsigned I = 1, E = Weights.size(); I != E; ++I)
3174 NewTrueWeight += (uint32_t)Weights[I];
3175 NewBI->setMetadata(LLVMContext::MD_prof,
3176 MDBuilder(SI->getContext()).
3177 createBranchWeights(NewTrueWeight,
3178 (uint32_t)Weights[0]));
3182 // Prune obsolete incoming values off the successor's PHI nodes.
3183 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
3184 isa<PHINode>(BBI); ++BBI) {
3185 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
3186 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3188 SI->eraseFromParent();
3193 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
3194 /// and use it to remove dead cases.
3195 static bool EliminateDeadSwitchCases(SwitchInst *SI) {
3196 Value *Cond = SI->getCondition();
3197 unsigned Bits = Cond->getType()->getIntegerBitWidth();
3198 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3199 ComputeMaskedBits(Cond, KnownZero, KnownOne);
3201 // Gather dead cases.
3202 SmallVector<ConstantInt*, 8> DeadCases;
3203 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3204 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3205 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3206 DeadCases.push_back(I.getCaseValue());
3207 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3208 << I.getCaseValue() << "' is dead.\n");
3212 SmallVector<uint64_t, 8> Weights;
3213 bool HasWeight = HasBranchWeights(SI);
3215 GetBranchWeights(SI, Weights);
3216 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3219 // Remove dead cases from the switch.
3220 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3221 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3222 assert(Case != SI->case_default() &&
3223 "Case was not found. Probably mistake in DeadCases forming.");
3225 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3229 // Prune unused values from PHI nodes.
3230 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3231 SI->removeCase(Case);
3233 if (HasWeight && Weights.size() >= 2) {
3234 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3235 SI->setMetadata(LLVMContext::MD_prof,
3236 MDBuilder(SI->getParent()->getContext()).
3237 createBranchWeights(MDWeights));
3240 return !DeadCases.empty();
3243 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
3244 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3245 /// by an unconditional branch), look at the phi node for BB in the successor
3246 /// block and see if the incoming value is equal to CaseValue. If so, return
3247 /// the phi node, and set PhiIndex to BB's index in the phi node.
3248 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3251 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3252 return NULL; // BB must be empty to be a candidate for simplification.
3253 if (!BB->getSinglePredecessor())
3254 return NULL; // BB must be dominated by the switch.
3256 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3257 if (!Branch || !Branch->isUnconditional())
3258 return NULL; // Terminator must be unconditional branch.
3260 BasicBlock *Succ = Branch->getSuccessor(0);
3262 BasicBlock::iterator I = Succ->begin();
3263 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3264 int Idx = PHI->getBasicBlockIndex(BB);
3265 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3267 Value *InValue = PHI->getIncomingValue(Idx);
3268 if (InValue != CaseValue) continue;
3277 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
3278 /// instruction to a phi node dominated by the switch, if that would mean that
3279 /// some of the destination blocks of the switch can be folded away.
3280 /// Returns true if a change is made.
3281 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3282 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3283 ForwardingNodesMap ForwardingNodes;
3285 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3286 ConstantInt *CaseValue = I.getCaseValue();
3287 BasicBlock *CaseDest = I.getCaseSuccessor();
3290 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3294 ForwardingNodes[PHI].push_back(PhiIndex);
3297 bool Changed = false;
3299 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3300 E = ForwardingNodes.end(); I != E; ++I) {
3301 PHINode *Phi = I->first;
3302 SmallVectorImpl<int> &Indexes = I->second;
3304 if (Indexes.size() < 2) continue;
3306 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3307 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3314 /// ValidLookupTableConstant - Return true if the backend will be able to handle
3315 /// initializing an array of constants like C.
3316 static bool ValidLookupTableConstant(Constant *C) {
3317 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3318 return CE->isGEPWithNoNotionalOverIndexing();
3320 return isa<ConstantFP>(C) ||
3321 isa<ConstantInt>(C) ||
3322 isa<ConstantPointerNull>(C) ||
3323 isa<GlobalValue>(C) ||
3327 /// LookupConstant - If V is a Constant, return it. Otherwise, try to look up
3328 /// its constant value in ConstantPool, returning 0 if it's not there.
3329 static Constant *LookupConstant(Value *V,
3330 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3331 if (Constant *C = dyn_cast<Constant>(V))
3333 return ConstantPool.lookup(V);
3336 /// ConstantFold - Try to fold instruction I into a constant. This works for
3337 /// simple instructions such as binary operations where both operands are
3338 /// constant or can be replaced by constants from the ConstantPool. Returns the
3339 /// resulting constant on success, 0 otherwise.
3341 ConstantFold(Instruction *I,
3342 const SmallDenseMap<Value *, Constant *> &ConstantPool,
3343 const DataLayout *DL) {
3344 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3345 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3348 if (A->isAllOnesValue())
3349 return LookupConstant(Select->getTrueValue(), ConstantPool);
3350 if (A->isNullValue())
3351 return LookupConstant(Select->getFalseValue(), ConstantPool);
3355 SmallVector<Constant *, 4> COps;
3356 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
3357 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
3363 if (CmpInst *Cmp = dyn_cast<CmpInst>(I))
3364 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
3367 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), COps, DL);
3370 /// GetCaseResults - Try to determine the resulting constant values in phi nodes
3371 /// at the common destination basic block, *CommonDest, for one of the case
3372 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3373 /// case), of a switch instruction SI.
3375 GetCaseResults(SwitchInst *SI,
3376 ConstantInt *CaseVal,
3377 BasicBlock *CaseDest,
3378 BasicBlock **CommonDest,
3379 SmallVectorImpl<std::pair<PHINode *, Constant *> > &Res,
3380 const DataLayout *DL) {
3381 // The block from which we enter the common destination.
3382 BasicBlock *Pred = SI->getParent();
3384 // If CaseDest is empty except for some side-effect free instructions through
3385 // which we can constant-propagate the CaseVal, continue to its successor.
3386 SmallDenseMap<Value*, Constant*> ConstantPool;
3387 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3388 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3390 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3391 // If the terminator is a simple branch, continue to the next block.
3392 if (T->getNumSuccessors() != 1)
3395 CaseDest = T->getSuccessor(0);
3396 } else if (isa<DbgInfoIntrinsic>(I)) {
3397 // Skip debug intrinsic.
3399 } else if (Constant *C = ConstantFold(I, ConstantPool, DL)) {
3400 // Instruction is side-effect free and constant.
3401 ConstantPool.insert(std::make_pair(I, C));
3407 // If we did not have a CommonDest before, use the current one.
3409 *CommonDest = CaseDest;
3410 // If the destination isn't the common one, abort.
3411 if (CaseDest != *CommonDest)
3414 // Get the values for this case from phi nodes in the destination block.
3415 BasicBlock::iterator I = (*CommonDest)->begin();
3416 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3417 int Idx = PHI->getBasicBlockIndex(Pred);
3421 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3426 // Note: If the constant comes from constant-propagating the case value
3427 // through the CaseDest basic block, it will be safe to remove the
3428 // instructions in that block. They cannot be used (except in the phi nodes
3429 // we visit) outside CaseDest, because that block does not dominate its
3430 // successor. If it did, we would not be in this phi node.
3432 // Be conservative about which kinds of constants we support.
3433 if (!ValidLookupTableConstant(ConstVal))
3436 Res.push_back(std::make_pair(PHI, ConstVal));
3439 return Res.size() > 0;
3443 /// SwitchLookupTable - This class represents a lookup table that can be used
3444 /// to replace a switch.
3445 class SwitchLookupTable {
3447 /// SwitchLookupTable - Create a lookup table to use as a switch replacement
3448 /// with the contents of Values, using DefaultValue to fill any holes in the
3450 SwitchLookupTable(Module &M,
3452 ConstantInt *Offset,
3453 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3454 Constant *DefaultValue,
3455 const DataLayout *DL);
3457 /// BuildLookup - Build instructions with Builder to retrieve the value at
3458 /// the position given by Index in the lookup table.
3459 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3461 /// WouldFitInRegister - Return true if a table with TableSize elements of
3462 /// type ElementType would fit in a target-legal register.
3463 static bool WouldFitInRegister(const DataLayout *DL,
3465 const Type *ElementType);
3468 // Depending on the contents of the table, it can be represented in
3471 // For tables where each element contains the same value, we just have to
3472 // store that single value and return it for each lookup.
3475 // For small tables with integer elements, we can pack them into a bitmap
3476 // that fits into a target-legal register. Values are retrieved by
3477 // shift and mask operations.
3480 // The table is stored as an array of values. Values are retrieved by load
3481 // instructions from the table.
3485 // For SingleValueKind, this is the single value.
3486 Constant *SingleValue;
3488 // For BitMapKind, this is the bitmap.
3489 ConstantInt *BitMap;
3490 IntegerType *BitMapElementTy;
3492 // For ArrayKind, this is the array.
3493 GlobalVariable *Array;
3497 SwitchLookupTable::SwitchLookupTable(Module &M,
3499 ConstantInt *Offset,
3500 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3501 Constant *DefaultValue,
3502 const DataLayout *DL)
3503 : SingleValue(0), BitMap(0), BitMapElementTy(0), Array(0) {
3504 assert(Values.size() && "Can't build lookup table without values!");
3505 assert(TableSize >= Values.size() && "Can't fit values in table!");
3507 // If all values in the table are equal, this is that value.
3508 SingleValue = Values.begin()->second;
3510 Type *ValueType = Values.begin()->second->getType();
3512 // Build up the table contents.
3513 SmallVector<Constant*, 64> TableContents(TableSize);
3514 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3515 ConstantInt *CaseVal = Values[I].first;
3516 Constant *CaseRes = Values[I].second;
3517 assert(CaseRes->getType() == ValueType);
3519 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3521 TableContents[Idx] = CaseRes;
3523 if (CaseRes != SingleValue)
3527 // Fill in any holes in the table with the default result.
3528 if (Values.size() < TableSize) {
3529 assert(DefaultValue && "Need a default value to fill the lookup table holes.");
3530 assert(DefaultValue->getType() == ValueType);
3531 for (uint64_t I = 0; I < TableSize; ++I) {
3532 if (!TableContents[I])
3533 TableContents[I] = DefaultValue;
3536 if (DefaultValue != SingleValue)
3540 // If each element in the table contains the same value, we only need to store
3541 // that single value.
3543 Kind = SingleValueKind;
3547 // If the type is integer and the table fits in a register, build a bitmap.
3548 if (WouldFitInRegister(DL, TableSize, ValueType)) {
3549 IntegerType *IT = cast<IntegerType>(ValueType);
3550 APInt TableInt(TableSize * IT->getBitWidth(), 0);
3551 for (uint64_t I = TableSize; I > 0; --I) {
3552 TableInt <<= IT->getBitWidth();
3553 // Insert values into the bitmap. Undef values are set to zero.
3554 if (!isa<UndefValue>(TableContents[I - 1])) {
3555 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
3556 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
3559 BitMap = ConstantInt::get(M.getContext(), TableInt);
3560 BitMapElementTy = IT;
3566 // Store the table in an array.
3567 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
3568 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3570 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3571 GlobalVariable::PrivateLinkage,
3574 Array->setUnnamedAddr(true);
3578 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
3580 case SingleValueKind:
3583 // Type of the bitmap (e.g. i59).
3584 IntegerType *MapTy = BitMap->getType();
3586 // Cast Index to the same type as the bitmap.
3587 // Note: The Index is <= the number of elements in the table, so
3588 // truncating it to the width of the bitmask is safe.
3589 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
3591 // Multiply the shift amount by the element width.
3592 ShiftAmt = Builder.CreateMul(ShiftAmt,
3593 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
3597 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
3598 "switch.downshift");
3600 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
3604 Value *GEPIndices[] = { Builder.getInt32(0), Index };
3605 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices,
3607 return Builder.CreateLoad(GEP, "switch.load");
3610 llvm_unreachable("Unknown lookup table kind!");
3613 bool SwitchLookupTable::WouldFitInRegister(const DataLayout *DL,
3615 const Type *ElementType) {
3618 const IntegerType *IT = dyn_cast<IntegerType>(ElementType);
3621 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
3622 // are <= 15, we could try to narrow the type.
3624 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
3625 if (TableSize >= UINT_MAX/IT->getBitWidth())
3627 return DL->fitsInLegalInteger(TableSize * IT->getBitWidth());
3630 /// ShouldBuildLookupTable - Determine whether a lookup table should be built
3631 /// for this switch, based on the number of cases, size of the table and the
3632 /// types of the results.
3633 static bool ShouldBuildLookupTable(SwitchInst *SI,
3635 const TargetTransformInfo &TTI,
3636 const DataLayout *DL,
3637 const SmallDenseMap<PHINode*, Type*>& ResultTypes) {
3638 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
3639 return false; // TableSize overflowed, or mul below might overflow.
3641 bool AllTablesFitInRegister = true;
3642 bool HasIllegalType = false;
3643 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(),
3644 E = ResultTypes.end(); I != E; ++I) {
3645 Type *Ty = I->second;
3647 // Saturate this flag to true.
3648 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
3650 // Saturate this flag to false.
3651 AllTablesFitInRegister = AllTablesFitInRegister &&
3652 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
3654 // If both flags saturate, we're done. NOTE: This *only* works with
3655 // saturating flags, and all flags have to saturate first due to the
3656 // non-deterministic behavior of iterating over a dense map.
3657 if (HasIllegalType && !AllTablesFitInRegister)
3661 // If each table would fit in a register, we should build it anyway.
3662 if (AllTablesFitInRegister)
3665 // Don't build a table that doesn't fit in-register if it has illegal types.
3669 // The table density should be at least 40%. This is the same criterion as for
3670 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3671 // FIXME: Find the best cut-off.
3672 return SI->getNumCases() * 10 >= TableSize * 4;
3675 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3676 /// phi nodes in a common successor block with different constant values,
3677 /// replace the switch with lookup tables.
3678 static bool SwitchToLookupTable(SwitchInst *SI,
3679 IRBuilder<> &Builder,
3680 const TargetTransformInfo &TTI,
3681 const DataLayout* DL) {
3682 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3684 // Only build lookup table when we have a target that supports it.
3685 if (!TTI.shouldBuildLookupTables())
3688 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3689 // split off a dense part and build a lookup table for that.
3691 // FIXME: This creates arrays of GEPs to constant strings, which means each
3692 // GEP needs a runtime relocation in PIC code. We should just build one big
3693 // string and lookup indices into that.
3695 // Ignore switches with less than three cases. Lookup tables will not make them
3696 // faster, so we don't analyze them.
3697 if (SI->getNumCases() < 3)
3700 // Figure out the corresponding result for each case value and phi node in the
3701 // common destination, as well as the the min and max case values.
3702 assert(SI->case_begin() != SI->case_end());
3703 SwitchInst::CaseIt CI = SI->case_begin();
3704 ConstantInt *MinCaseVal = CI.getCaseValue();
3705 ConstantInt *MaxCaseVal = CI.getCaseValue();
3707 BasicBlock *CommonDest = 0;
3708 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
3709 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
3710 SmallDenseMap<PHINode*, Constant*> DefaultResults;
3711 SmallDenseMap<PHINode*, Type*> ResultTypes;
3712 SmallVector<PHINode*, 4> PHIs;
3714 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
3715 ConstantInt *CaseVal = CI.getCaseValue();
3716 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
3717 MinCaseVal = CaseVal;
3718 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
3719 MaxCaseVal = CaseVal;
3721 // Resulting value at phi nodes for this case value.
3722 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
3724 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
3728 // Append the result from this case to the list for each phi.
3729 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) {
3730 if (!ResultLists.count(I->first))
3731 PHIs.push_back(I->first);
3732 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second));
3736 // Keep track of the result types.
3737 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
3738 PHINode *PHI = PHIs[I];
3739 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
3742 uint64_t NumResults = ResultLists[PHIs[0]].size();
3743 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
3744 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
3745 bool TableHasHoles = (NumResults < TableSize);
3747 // If the table has holes, we need a constant result for the default case.
3748 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
3749 if (TableHasHoles && !GetCaseResults(SI, 0, SI->getDefaultDest(), &CommonDest,
3750 DefaultResultsList, DL))
3753 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) {
3754 PHINode *PHI = DefaultResultsList[I].first;
3755 Constant *Result = DefaultResultsList[I].second;
3756 DefaultResults[PHI] = Result;
3759 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
3762 // Create the BB that does the lookups.
3763 Module &Mod = *CommonDest->getParent()->getParent();
3764 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
3766 CommonDest->getParent(),
3769 // Compute the table index value.
3770 Builder.SetInsertPoint(SI);
3771 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
3774 // Compute the maximum table size representable by the integer type we are
3776 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
3777 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
3778 assert(MaxTableSize >= TableSize &&
3779 "It is impossible for a switch to have more entries than the max "
3780 "representable value of its input integer type's size.");
3782 // If we have a fully covered lookup table, unconditionally branch to the
3783 // lookup table BB. Otherwise, check if the condition value is within the case
3784 // range. If it is so, branch to the new BB. Otherwise branch to SI's default
3786 const bool GeneratingCoveredLookupTable = MaxTableSize == TableSize;
3787 if (GeneratingCoveredLookupTable) {
3788 Builder.CreateBr(LookupBB);
3789 SI->getDefaultDest()->removePredecessor(SI->getParent());
3791 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
3792 MinCaseVal->getType(), TableSize));
3793 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
3796 // Populate the BB that does the lookups.
3797 Builder.SetInsertPoint(LookupBB);
3798 bool ReturnedEarly = false;
3799 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
3800 PHINode *PHI = PHIs[I];
3802 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI],
3803 DefaultResults[PHI], DL);
3805 Value *Result = Table.BuildLookup(TableIndex, Builder);
3807 // If the result is used to return immediately from the function, we want to
3808 // do that right here.
3809 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
3810 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
3811 Builder.CreateRet(Result);
3812 ReturnedEarly = true;
3816 PHI->addIncoming(Result, LookupBB);
3820 Builder.CreateBr(CommonDest);
3822 // Remove the switch.
3823 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
3824 BasicBlock *Succ = SI->getSuccessor(i);
3826 if (Succ == SI->getDefaultDest())
3828 Succ->removePredecessor(SI->getParent());
3830 SI->eraseFromParent();
3836 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
3837 BasicBlock *BB = SI->getParent();
3839 if (isValueEqualityComparison(SI)) {
3840 // If we only have one predecessor, and if it is a branch on this value,
3841 // see if that predecessor totally determines the outcome of this switch.
3842 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3843 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
3844 return SimplifyCFG(BB, TTI, DL) | true;
3846 Value *Cond = SI->getCondition();
3847 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
3848 if (SimplifySwitchOnSelect(SI, Select))
3849 return SimplifyCFG(BB, TTI, DL) | true;
3851 // If the block only contains the switch, see if we can fold the block
3852 // away into any preds.
3853 BasicBlock::iterator BBI = BB->begin();
3854 // Ignore dbg intrinsics.
3855 while (isa<DbgInfoIntrinsic>(BBI))
3858 if (FoldValueComparisonIntoPredecessors(SI, Builder))
3859 return SimplifyCFG(BB, TTI, DL) | true;
3862 // Try to transform the switch into an icmp and a branch.
3863 if (TurnSwitchRangeIntoICmp(SI, Builder))
3864 return SimplifyCFG(BB, TTI, DL) | true;
3866 // Remove unreachable cases.
3867 if (EliminateDeadSwitchCases(SI))
3868 return SimplifyCFG(BB, TTI, DL) | true;
3870 if (ForwardSwitchConditionToPHI(SI))
3871 return SimplifyCFG(BB, TTI, DL) | true;
3873 if (SwitchToLookupTable(SI, Builder, TTI, DL))
3874 return SimplifyCFG(BB, TTI, DL) | true;
3879 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
3880 BasicBlock *BB = IBI->getParent();
3881 bool Changed = false;
3883 // Eliminate redundant destinations.
3884 SmallPtrSet<Value *, 8> Succs;
3885 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
3886 BasicBlock *Dest = IBI->getDestination(i);
3887 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) {
3888 Dest->removePredecessor(BB);
3889 IBI->removeDestination(i);
3895 if (IBI->getNumDestinations() == 0) {
3896 // If the indirectbr has no successors, change it to unreachable.
3897 new UnreachableInst(IBI->getContext(), IBI);
3898 EraseTerminatorInstAndDCECond(IBI);
3902 if (IBI->getNumDestinations() == 1) {
3903 // If the indirectbr has one successor, change it to a direct branch.
3904 BranchInst::Create(IBI->getDestination(0), IBI);
3905 EraseTerminatorInstAndDCECond(IBI);
3909 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
3910 if (SimplifyIndirectBrOnSelect(IBI, SI))
3911 return SimplifyCFG(BB, TTI, DL) | true;
3916 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
3917 BasicBlock *BB = BI->getParent();
3919 if (SinkCommon && SinkThenElseCodeToEnd(BI))
3922 // If the Terminator is the only non-phi instruction, simplify the block.
3923 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime();
3924 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
3925 TryToSimplifyUncondBranchFromEmptyBlock(BB))
3928 // If the only instruction in the block is a seteq/setne comparison
3929 // against a constant, try to simplify the block.
3930 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
3931 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
3932 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
3934 if (I->isTerminator() &&
3935 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI, DL))
3939 // If this basic block is ONLY a compare and a branch, and if a predecessor
3940 // branches to us and our successor, fold the comparison into the
3941 // predecessor and use logical operations to update the incoming value
3942 // for PHI nodes in common successor.
3943 if (FoldBranchToCommonDest(BI))
3944 return SimplifyCFG(BB, TTI, DL) | true;
3949 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
3950 BasicBlock *BB = BI->getParent();
3952 // Conditional branch
3953 if (isValueEqualityComparison(BI)) {
3954 // If we only have one predecessor, and if it is a branch on this value,
3955 // see if that predecessor totally determines the outcome of this
3957 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3958 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
3959 return SimplifyCFG(BB, TTI, DL) | true;
3961 // This block must be empty, except for the setcond inst, if it exists.
3962 // Ignore dbg intrinsics.
3963 BasicBlock::iterator I = BB->begin();
3964 // Ignore dbg intrinsics.
3965 while (isa<DbgInfoIntrinsic>(I))
3968 if (FoldValueComparisonIntoPredecessors(BI, Builder))
3969 return SimplifyCFG(BB, TTI, DL) | true;
3970 } else if (&*I == cast<Instruction>(BI->getCondition())){
3972 // Ignore dbg intrinsics.
3973 while (isa<DbgInfoIntrinsic>(I))
3975 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
3976 return SimplifyCFG(BB, TTI, DL) | true;
3980 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
3981 if (SimplifyBranchOnICmpChain(BI, DL, Builder))
3984 // If this basic block is ONLY a compare and a branch, and if a predecessor
3985 // branches to us and one of our successors, fold the comparison into the
3986 // predecessor and use logical operations to pick the right destination.
3987 if (FoldBranchToCommonDest(BI))
3988 return SimplifyCFG(BB, TTI, DL) | true;
3990 // We have a conditional branch to two blocks that are only reachable
3991 // from BI. We know that the condbr dominates the two blocks, so see if
3992 // there is any identical code in the "then" and "else" blocks. If so, we
3993 // can hoist it up to the branching block.
3994 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) {
3995 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3996 if (HoistThenElseCodeToIf(BI))
3997 return SimplifyCFG(BB, TTI, DL) | true;
3999 // If Successor #1 has multiple preds, we may be able to conditionally
4000 // execute Successor #0 if it branches to successor #1.
4001 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
4002 if (Succ0TI->getNumSuccessors() == 1 &&
4003 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
4004 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0)))
4005 return SimplifyCFG(BB, TTI, DL) | true;
4007 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
4008 // If Successor #0 has multiple preds, we may be able to conditionally
4009 // execute Successor #1 if it branches to successor #0.
4010 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
4011 if (Succ1TI->getNumSuccessors() == 1 &&
4012 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
4013 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1)))
4014 return SimplifyCFG(BB, TTI, DL) | true;
4017 // If this is a branch on a phi node in the current block, thread control
4018 // through this block if any PHI node entries are constants.
4019 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
4020 if (PN->getParent() == BI->getParent())
4021 if (FoldCondBranchOnPHI(BI, DL))
4022 return SimplifyCFG(BB, TTI, DL) | true;
4024 // Scan predecessor blocks for conditional branches.
4025 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
4026 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
4027 if (PBI != BI && PBI->isConditional())
4028 if (SimplifyCondBranchToCondBranch(PBI, BI))
4029 return SimplifyCFG(BB, TTI, DL) | true;
4034 /// Check if passing a value to an instruction will cause undefined behavior.
4035 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
4036 Constant *C = dyn_cast<Constant>(V);
4043 if (C->isNullValue()) {
4044 // Only look at the first use, avoid hurting compile time with long uselists
4045 User *Use = *I->user_begin();
4047 // Now make sure that there are no instructions in between that can alter
4048 // control flow (eg. calls)
4049 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
4050 if (i == I->getParent()->end() || i->mayHaveSideEffects())
4053 // Look through GEPs. A load from a GEP derived from NULL is still undefined
4054 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
4055 if (GEP->getPointerOperand() == I)
4056 return passingValueIsAlwaysUndefined(V, GEP);
4058 // Look through bitcasts.
4059 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
4060 return passingValueIsAlwaysUndefined(V, BC);
4062 // Load from null is undefined.
4063 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
4064 if (!LI->isVolatile())
4065 return LI->getPointerAddressSpace() == 0;
4067 // Store to null is undefined.
4068 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
4069 if (!SI->isVolatile())
4070 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
4075 /// If BB has an incoming value that will always trigger undefined behavior
4076 /// (eg. null pointer dereference), remove the branch leading here.
4077 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
4078 for (BasicBlock::iterator i = BB->begin();
4079 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
4080 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
4081 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
4082 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
4083 IRBuilder<> Builder(T);
4084 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
4085 BB->removePredecessor(PHI->getIncomingBlock(i));
4086 // Turn uncoditional branches into unreachables and remove the dead
4087 // destination from conditional branches.
4088 if (BI->isUnconditional())
4089 Builder.CreateUnreachable();
4091 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
4092 BI->getSuccessor(0));
4093 BI->eraseFromParent();
4096 // TODO: SwitchInst.
4102 bool SimplifyCFGOpt::run(BasicBlock *BB) {
4103 bool Changed = false;
4105 assert(BB && BB->getParent() && "Block not embedded in function!");
4106 assert(BB->getTerminator() && "Degenerate basic block encountered!");
4108 // Remove basic blocks that have no predecessors (except the entry block)...
4109 // or that just have themself as a predecessor. These are unreachable.
4110 if ((pred_begin(BB) == pred_end(BB) &&
4111 BB != &BB->getParent()->getEntryBlock()) ||
4112 BB->getSinglePredecessor() == BB) {
4113 DEBUG(dbgs() << "Removing BB: \n" << *BB);
4114 DeleteDeadBlock(BB);
4118 // Check to see if we can constant propagate this terminator instruction
4120 Changed |= ConstantFoldTerminator(BB, true);
4122 // Check for and eliminate duplicate PHI nodes in this block.
4123 Changed |= EliminateDuplicatePHINodes(BB);
4125 // Check for and remove branches that will always cause undefined behavior.
4126 Changed |= removeUndefIntroducingPredecessor(BB);
4128 // Merge basic blocks into their predecessor if there is only one distinct
4129 // pred, and if there is only one distinct successor of the predecessor, and
4130 // if there are no PHI nodes.
4132 if (MergeBlockIntoPredecessor(BB))
4135 IRBuilder<> Builder(BB);
4137 // If there is a trivial two-entry PHI node in this basic block, and we can
4138 // eliminate it, do so now.
4139 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
4140 if (PN->getNumIncomingValues() == 2)
4141 Changed |= FoldTwoEntryPHINode(PN, DL);
4143 Builder.SetInsertPoint(BB->getTerminator());
4144 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
4145 if (BI->isUnconditional()) {
4146 if (SimplifyUncondBranch(BI, Builder)) return true;
4148 if (SimplifyCondBranch(BI, Builder)) return true;
4150 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
4151 if (SimplifyReturn(RI, Builder)) return true;
4152 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
4153 if (SimplifyResume(RI, Builder)) return true;
4154 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
4155 if (SimplifySwitch(SI, Builder)) return true;
4156 } else if (UnreachableInst *UI =
4157 dyn_cast<UnreachableInst>(BB->getTerminator())) {
4158 if (SimplifyUnreachable(UI)) return true;
4159 } else if (IndirectBrInst *IBI =
4160 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
4161 if (SimplifyIndirectBr(IBI)) return true;
4167 /// SimplifyCFG - This function is used to do simplification of a CFG. For
4168 /// example, it adjusts branches to branches to eliminate the extra hop, it
4169 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4170 /// of the CFG. It returns true if a modification was made.
4172 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
4173 const DataLayout *DL) {
4174 return SimplifyCFGOpt(TTI, DL).run(BB);