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/Constants.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/DerivedTypes.h"
29 #include "llvm/IR/GlobalVariable.h"
30 #include "llvm/IR/IRBuilder.h"
31 #include "llvm/IR/Instructions.h"
32 #include "llvm/IR/IntrinsicInst.h"
33 #include "llvm/IR/LLVMContext.h"
34 #include "llvm/IR/MDBuilder.h"
35 #include "llvm/IR/Metadata.h"
36 #include "llvm/IR/Module.h"
37 #include "llvm/IR/Operator.h"
38 #include "llvm/IR/Type.h"
39 #include "llvm/Support/CFG.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/ConstantRange.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/NoFolder.h"
44 #include "llvm/Support/PatternMatch.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"));
66 HoistCondStores("simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
67 cl::desc("Hoist conditional stores if an unconditional store preceeds"));
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 TD;
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 *TD)
113 : TTI(TTI), TD(TD) {}
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 *TD) {
310 // Normal constant int.
311 ConstantInt *CI = dyn_cast<ConstantInt>(V);
312 if (CI || !TD || !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>(TD->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 *TD, 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), TD)) {
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, TD,
410 unsigned NumVals = Vals.size();
411 unsigned UsedICmpsBeforeRHS = UsedICmps;
412 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD,
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, TD,
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), TD))
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() == TD->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 /// Sees if any of the weights are too big for a uint32_t, and halves all the
752 /// weights if any are.
753 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)
768 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value
769 /// equality comparison instruction (either a switch or a branch on "X == c").
770 /// See if any of the predecessors of the terminator block are value comparisons
771 /// on the same value. If so, and if safe to do so, fold them together.
772 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
773 IRBuilder<> &Builder) {
774 BasicBlock *BB = TI->getParent();
775 Value *CV = isValueEqualityComparison(TI); // CondVal
776 assert(CV && "Not a comparison?");
777 bool Changed = false;
779 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
780 while (!Preds.empty()) {
781 BasicBlock *Pred = Preds.pop_back_val();
783 // See if the predecessor is a comparison with the same value.
784 TerminatorInst *PTI = Pred->getTerminator();
785 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
787 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
788 // Figure out which 'cases' to copy from SI to PSI.
789 std::vector<ValueEqualityComparisonCase> BBCases;
790 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
792 std::vector<ValueEqualityComparisonCase> PredCases;
793 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
795 // Based on whether the default edge from PTI goes to BB or not, fill in
796 // PredCases and PredDefault with the new switch cases we would like to
798 SmallVector<BasicBlock*, 8> NewSuccessors;
800 // Update the branch weight metadata along the way
801 SmallVector<uint64_t, 8> Weights;
802 bool PredHasWeights = HasBranchWeights(PTI);
803 bool SuccHasWeights = HasBranchWeights(TI);
805 if (PredHasWeights) {
806 GetBranchWeights(PTI, Weights);
807 // branch-weight metadata is inconsistent here.
808 if (Weights.size() != 1 + PredCases.size())
809 PredHasWeights = SuccHasWeights = false;
810 } else if (SuccHasWeights)
811 // If there are no predecessor weights but there are successor weights,
812 // populate Weights with 1, which will later be scaled to the sum of
813 // successor's weights
814 Weights.assign(1 + PredCases.size(), 1);
816 SmallVector<uint64_t, 8> SuccWeights;
817 if (SuccHasWeights) {
818 GetBranchWeights(TI, SuccWeights);
819 // branch-weight metadata is inconsistent here.
820 if (SuccWeights.size() != 1 + BBCases.size())
821 PredHasWeights = SuccHasWeights = false;
822 } else if (PredHasWeights)
823 SuccWeights.assign(1 + BBCases.size(), 1);
825 if (PredDefault == BB) {
826 // If this is the default destination from PTI, only the edges in TI
827 // that don't occur in PTI, or that branch to BB will be activated.
828 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
829 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
830 if (PredCases[i].Dest != BB)
831 PTIHandled.insert(PredCases[i].Value);
833 // The default destination is BB, we don't need explicit targets.
834 std::swap(PredCases[i], PredCases.back());
836 if (PredHasWeights || SuccHasWeights) {
837 // Increase weight for the default case.
838 Weights[0] += Weights[i+1];
839 std::swap(Weights[i+1], Weights.back());
843 PredCases.pop_back();
847 // Reconstruct the new switch statement we will be building.
848 if (PredDefault != BBDefault) {
849 PredDefault->removePredecessor(Pred);
850 PredDefault = BBDefault;
851 NewSuccessors.push_back(BBDefault);
854 unsigned CasesFromPred = Weights.size();
855 uint64_t ValidTotalSuccWeight = 0;
856 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
857 if (!PTIHandled.count(BBCases[i].Value) &&
858 BBCases[i].Dest != BBDefault) {
859 PredCases.push_back(BBCases[i]);
860 NewSuccessors.push_back(BBCases[i].Dest);
861 if (SuccHasWeights || PredHasWeights) {
862 // The default weight is at index 0, so weight for the ith case
863 // should be at index i+1. Scale the cases from successor by
864 // PredDefaultWeight (Weights[0]).
865 Weights.push_back(Weights[0] * SuccWeights[i+1]);
866 ValidTotalSuccWeight += SuccWeights[i+1];
870 if (SuccHasWeights || PredHasWeights) {
871 ValidTotalSuccWeight += SuccWeights[0];
872 // Scale the cases from predecessor by ValidTotalSuccWeight.
873 for (unsigned i = 1; i < CasesFromPred; ++i)
874 Weights[i] *= ValidTotalSuccWeight;
875 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
876 Weights[0] *= SuccWeights[0];
879 // If this is not the default destination from PSI, only the edges
880 // in SI that occur in PSI with a destination of BB will be
882 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
883 std::map<ConstantInt*, uint64_t> WeightsForHandled;
884 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
885 if (PredCases[i].Dest == BB) {
886 PTIHandled.insert(PredCases[i].Value);
888 if (PredHasWeights || SuccHasWeights) {
889 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
890 std::swap(Weights[i+1], Weights.back());
894 std::swap(PredCases[i], PredCases.back());
895 PredCases.pop_back();
899 // Okay, now we know which constants were sent to BB from the
900 // predecessor. Figure out where they will all go now.
901 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
902 if (PTIHandled.count(BBCases[i].Value)) {
903 // If this is one we are capable of getting...
904 if (PredHasWeights || SuccHasWeights)
905 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
906 PredCases.push_back(BBCases[i]);
907 NewSuccessors.push_back(BBCases[i].Dest);
908 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
911 // If there are any constants vectored to BB that TI doesn't handle,
912 // they must go to the default destination of TI.
913 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
915 E = PTIHandled.end(); I != E; ++I) {
916 if (PredHasWeights || SuccHasWeights)
917 Weights.push_back(WeightsForHandled[*I]);
918 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
919 NewSuccessors.push_back(BBDefault);
923 // Okay, at this point, we know which new successor Pred will get. Make
924 // sure we update the number of entries in the PHI nodes for these
926 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i)
927 AddPredecessorToBlock(NewSuccessors[i], Pred, BB);
929 Builder.SetInsertPoint(PTI);
930 // Convert pointer to int before we switch.
931 if (CV->getType()->isPointerTy()) {
932 assert(TD && "Cannot switch on pointer without DataLayout");
933 CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getType()),
937 // Now that the successors are updated, create the new Switch instruction.
938 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
940 NewSI->setDebugLoc(PTI->getDebugLoc());
941 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
942 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest);
944 if (PredHasWeights || SuccHasWeights) {
945 // Halve the weights if any of them cannot fit in an uint32_t
948 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
950 NewSI->setMetadata(LLVMContext::MD_prof,
951 MDBuilder(BB->getContext()).
952 createBranchWeights(MDWeights));
955 EraseTerminatorInstAndDCECond(PTI);
957 // Okay, last check. If BB is still a successor of PSI, then we must
958 // have an infinite loop case. If so, add an infinitely looping block
959 // to handle the case to preserve the behavior of the code.
960 BasicBlock *InfLoopBlock = 0;
961 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
962 if (NewSI->getSuccessor(i) == BB) {
963 if (InfLoopBlock == 0) {
964 // Insert it at the end of the function, because it's either code,
965 // or it won't matter if it's hot. :)
966 InfLoopBlock = BasicBlock::Create(BB->getContext(),
967 "infloop", BB->getParent());
968 BranchInst::Create(InfLoopBlock, InfLoopBlock);
970 NewSI->setSuccessor(i, InfLoopBlock);
979 // isSafeToHoistInvoke - If we would need to insert a select that uses the
980 // value of this invoke (comments in HoistThenElseCodeToIf explain why we
981 // would need to do this), we can't hoist the invoke, as there is nowhere
982 // to put the select in this case.
983 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
984 Instruction *I1, Instruction *I2) {
985 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
987 for (BasicBlock::iterator BBI = SI->begin();
988 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
989 Value *BB1V = PN->getIncomingValueForBlock(BB1);
990 Value *BB2V = PN->getIncomingValueForBlock(BB2);
991 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
999 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
1000 /// BB2, hoist any common code in the two blocks up into the branch block. The
1001 /// caller of this function guarantees that BI's block dominates BB1 and BB2.
1002 static bool HoistThenElseCodeToIf(BranchInst *BI) {
1003 // This does very trivial matching, with limited scanning, to find identical
1004 // instructions in the two blocks. In particular, we don't want to get into
1005 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1006 // such, we currently just scan for obviously identical instructions in an
1008 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1009 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1011 BasicBlock::iterator BB1_Itr = BB1->begin();
1012 BasicBlock::iterator BB2_Itr = BB2->begin();
1014 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++;
1015 // Skip debug info if it is not identical.
1016 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1017 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1018 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1019 while (isa<DbgInfoIntrinsic>(I1))
1021 while (isa<DbgInfoIntrinsic>(I2))
1024 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1025 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1028 BasicBlock *BIParent = BI->getParent();
1030 bool Changed = false;
1032 // If we are hoisting the terminator instruction, don't move one (making a
1033 // broken BB), instead clone it, and remove BI.
1034 if (isa<TerminatorInst>(I1))
1035 goto HoistTerminator;
1037 // For a normal instruction, we just move one to right before the branch,
1038 // then replace all uses of the other with the first. Finally, we remove
1039 // the now redundant second instruction.
1040 BIParent->getInstList().splice(BI, BB1->getInstList(), I1);
1041 if (!I2->use_empty())
1042 I2->replaceAllUsesWith(I1);
1043 I1->intersectOptionalDataWith(I2);
1044 I2->eraseFromParent();
1049 // Skip debug info if it is not identical.
1050 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1051 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1052 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1053 while (isa<DbgInfoIntrinsic>(I1))
1055 while (isa<DbgInfoIntrinsic>(I2))
1058 } while (I1->isIdenticalToWhenDefined(I2));
1063 // It may not be possible to hoist an invoke.
1064 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1067 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1069 for (BasicBlock::iterator BBI = SI->begin();
1070 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1071 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1072 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1076 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1078 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1083 // Okay, it is safe to hoist the terminator.
1084 Instruction *NT = I1->clone();
1085 BIParent->getInstList().insert(BI, NT);
1086 if (!NT->getType()->isVoidTy()) {
1087 I1->replaceAllUsesWith(NT);
1088 I2->replaceAllUsesWith(NT);
1092 IRBuilder<true, NoFolder> Builder(NT);
1093 // Hoisting one of the terminators from our successor is a great thing.
1094 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1095 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1096 // nodes, so we insert select instruction to compute the final result.
1097 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1098 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1100 for (BasicBlock::iterator BBI = SI->begin();
1101 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1102 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1103 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1104 if (BB1V == BB2V) continue;
1106 // These values do not agree. Insert a select instruction before NT
1107 // that determines the right value.
1108 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1110 SI = cast<SelectInst>
1111 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1112 BB1V->getName()+"."+BB2V->getName()));
1114 // Make the PHI node use the select for all incoming values for BB1/BB2
1115 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1116 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1117 PN->setIncomingValue(i, SI);
1121 // Update any PHI nodes in our new successors.
1122 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1123 AddPredecessorToBlock(*SI, BIParent, BB1);
1125 EraseTerminatorInstAndDCECond(BI);
1129 /// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd,
1130 /// check whether BBEnd has only two predecessors and the other predecessor
1131 /// ends with an unconditional branch. If it is true, sink any common code
1132 /// in the two predecessors to BBEnd.
1133 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1134 assert(BI1->isUnconditional());
1135 BasicBlock *BB1 = BI1->getParent();
1136 BasicBlock *BBEnd = BI1->getSuccessor(0);
1138 // Check that BBEnd has two predecessors and the other predecessor ends with
1139 // an unconditional branch.
1140 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1141 BasicBlock *Pred0 = *PI++;
1142 if (PI == PE) // Only one predecessor.
1144 BasicBlock *Pred1 = *PI++;
1145 if (PI != PE) // More than two predecessors.
1147 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1148 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1149 if (!BI2 || !BI2->isUnconditional())
1152 // Gather the PHI nodes in BBEnd.
1153 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2;
1154 Instruction *FirstNonPhiInBBEnd = 0;
1155 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end();
1157 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1158 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1159 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1160 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN);
1162 FirstNonPhiInBBEnd = &*I;
1166 if (!FirstNonPhiInBBEnd)
1170 // This does very trivial matching, with limited scanning, to find identical
1171 // instructions in the two blocks. We scan backward for obviously identical
1172 // instructions in an identical order.
1173 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1174 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(),
1175 RE2 = BB2->getInstList().rend();
1177 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1180 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1183 // Skip the unconditional branches.
1187 bool Changed = false;
1188 while (RI1 != RE1 && RI2 != RE2) {
1190 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1193 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1197 Instruction *I1 = &*RI1, *I2 = &*RI2;
1198 // I1 and I2 should have a single use in the same PHI node, and they
1199 // perform the same operation.
1200 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1201 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1202 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1203 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) ||
1204 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1205 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1206 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1207 !I1->hasOneUse() || !I2->hasOneUse() ||
1208 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() ||
1209 MapValueFromBB1ToBB2[I1].first != I2)
1212 // Check whether we should swap the operands of ICmpInst.
1213 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1214 bool SwapOpnds = false;
1215 if (ICmp1 && ICmp2 &&
1216 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1217 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1218 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1219 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1220 ICmp2->swapOperands();
1223 if (!I1->isSameOperationAs(I2)) {
1225 ICmp2->swapOperands();
1229 // The operands should be either the same or they need to be generated
1230 // with a PHI node after sinking. We only handle the case where there is
1231 // a single pair of different operands.
1232 Value *DifferentOp1 = 0, *DifferentOp2 = 0;
1233 unsigned Op1Idx = 0;
1234 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1235 if (I1->getOperand(I) == I2->getOperand(I))
1237 // Early exit if we have more-than one pair of different operands or
1238 // the different operand is already in MapValueFromBB1ToBB2.
1239 // Early exit if we need a PHI node to replace a constant.
1241 MapValueFromBB1ToBB2.find(I1->getOperand(I)) !=
1242 MapValueFromBB1ToBB2.end() ||
1243 isa<Constant>(I1->getOperand(I)) ||
1244 isa<Constant>(I2->getOperand(I))) {
1245 // If we can't sink the instructions, undo the swapping.
1247 ICmp2->swapOperands();
1250 DifferentOp1 = I1->getOperand(I);
1252 DifferentOp2 = I2->getOperand(I);
1255 // We insert the pair of different operands to MapValueFromBB1ToBB2 and
1256 // remove (I1, I2) from MapValueFromBB1ToBB2.
1258 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2,
1259 DifferentOp1->getName() + ".sink",
1261 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN);
1262 // I1 should use NewPN instead of DifferentOp1.
1263 I1->setOperand(Op1Idx, NewPN);
1264 NewPN->addIncoming(DifferentOp1, BB1);
1265 NewPN->addIncoming(DifferentOp2, BB2);
1266 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1268 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second;
1269 MapValueFromBB1ToBB2.erase(I1);
1271 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";);
1272 DEBUG(dbgs() << " " << *I2 << "\n";);
1273 // We need to update RE1 and RE2 if we are going to sink the first
1274 // instruction in the basic block down.
1275 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1276 // Sink the instruction.
1277 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1);
1278 if (!OldPN->use_empty())
1279 OldPN->replaceAllUsesWith(I1);
1280 OldPN->eraseFromParent();
1282 if (!I2->use_empty())
1283 I2->replaceAllUsesWith(I1);
1284 I1->intersectOptionalDataWith(I2);
1285 I2->eraseFromParent();
1288 RE1 = BB1->getInstList().rend();
1290 RE2 = BB2->getInstList().rend();
1291 FirstNonPhiInBBEnd = I1;
1298 /// \brief Determine if we can hoist sink a sole store instruction out of a
1299 /// conditional block.
1301 /// We are looking for code like the following:
1303 /// store i32 %add, i32* %arrayidx2
1304 /// ... // No other stores or function calls (we could be calling a memory
1305 /// ... // function).
1306 /// %cmp = icmp ult %x, %y
1307 /// br i1 %cmp, label %EndBB, label %ThenBB
1309 /// store i32 %add5, i32* %arrayidx2
1313 /// We are going to transform this into:
1315 /// store i32 %add, i32* %arrayidx2
1317 /// %cmp = icmp ult %x, %y
1318 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1319 /// store i32 %add.add5, i32* %arrayidx2
1322 /// \return The pointer to the value of the previous store if the store can be
1323 /// hoisted into the predecessor block. 0 otherwise.
1324 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1325 BasicBlock *StoreBB, BasicBlock *EndBB) {
1326 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1330 // Volatile or atomic.
1331 if (!StoreToHoist->isSimple())
1334 Value *StorePtr = StoreToHoist->getPointerOperand();
1336 // Look for a store to the same pointer in BrBB.
1337 unsigned MaxNumInstToLookAt = 10;
1338 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(),
1339 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) {
1340 Instruction *CurI = &*RI;
1342 // Could be calling an instruction that effects memory like free().
1343 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI))
1346 StoreInst *SI = dyn_cast<StoreInst>(CurI);
1347 // Found the previous store make sure it stores to the same location.
1348 if (SI && SI->getPointerOperand() == StorePtr)
1349 // Found the previous store, return its value operand.
1350 return SI->getValueOperand();
1352 return 0; // Unknown store.
1358 /// \brief Speculate a conditional basic block flattening the CFG.
1360 /// Note that this is a very risky transform currently. Speculating
1361 /// instructions like this is most often not desirable. Instead, there is an MI
1362 /// pass which can do it with full awareness of the resource constraints.
1363 /// However, some cases are "obvious" and we should do directly. An example of
1364 /// this is speculating a single, reasonably cheap instruction.
1366 /// There is only one distinct advantage to flattening the CFG at the IR level:
1367 /// it makes very common but simplistic optimizations such as are common in
1368 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1369 /// modeling their effects with easier to reason about SSA value graphs.
1372 /// An illustration of this transform is turning this IR:
1375 /// %cmp = icmp ult %x, %y
1376 /// br i1 %cmp, label %EndBB, label %ThenBB
1378 /// %sub = sub %x, %y
1381 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1388 /// %cmp = icmp ult %x, %y
1389 /// %sub = sub %x, %y
1390 /// %cond = select i1 %cmp, 0, %sub
1394 /// \returns true if the conditional block is removed.
1395 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB) {
1396 // Be conservative for now. FP select instruction can often be expensive.
1397 Value *BrCond = BI->getCondition();
1398 if (isa<FCmpInst>(BrCond))
1401 BasicBlock *BB = BI->getParent();
1402 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1404 // If ThenBB is actually on the false edge of the conditional branch, remember
1405 // to swap the select operands later.
1406 bool Invert = false;
1407 if (ThenBB != BI->getSuccessor(0)) {
1408 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1411 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1413 // Keep a count of how many times instructions are used within CondBB when
1414 // they are candidates for sinking into CondBB. Specifically:
1415 // - They are defined in BB, and
1416 // - They have no side effects, and
1417 // - All of their uses are in CondBB.
1418 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1420 unsigned SpeculationCost = 0;
1421 Value *SpeculatedStoreValue = 0;
1422 StoreInst *SpeculatedStore = 0;
1423 for (BasicBlock::iterator BBI = ThenBB->begin(),
1424 BBE = llvm::prior(ThenBB->end());
1425 BBI != BBE; ++BBI) {
1426 Instruction *I = BBI;
1428 if (isa<DbgInfoIntrinsic>(I))
1431 // Only speculatively execution a single instruction (not counting the
1432 // terminator) for now.
1434 if (SpeculationCost > 1)
1437 // Don't hoist the instruction if it's unsafe or expensive.
1438 if (!isSafeToSpeculativelyExecute(I) &&
1439 !(HoistCondStores &&
1440 (SpeculatedStoreValue = isSafeToSpeculateStore(I, BB, ThenBB,
1443 if (!SpeculatedStoreValue &&
1444 ComputeSpeculationCost(I) > PHINodeFoldingThreshold)
1447 // Store the store speculation candidate.
1448 if (SpeculatedStoreValue)
1449 SpeculatedStore = cast<StoreInst>(I);
1451 // Do not hoist the instruction if any of its operands are defined but not
1452 // used in BB. The transformation will prevent the operand from
1453 // being sunk into the use block.
1454 for (User::op_iterator i = I->op_begin(), e = I->op_end();
1456 Instruction *OpI = dyn_cast<Instruction>(*i);
1457 if (!OpI || OpI->getParent() != BB ||
1458 OpI->mayHaveSideEffects())
1459 continue; // Not a candidate for sinking.
1461 ++SinkCandidateUseCounts[OpI];
1465 // Consider any sink candidates which are only used in CondBB as costs for
1466 // speculation. Note, while we iterate over a DenseMap here, we are summing
1467 // and so iteration order isn't significant.
1468 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I =
1469 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end();
1471 if (I->first->getNumUses() == I->second) {
1473 if (SpeculationCost > 1)
1477 // Check that the PHI nodes can be converted to selects.
1478 bool HaveRewritablePHIs = false;
1479 for (BasicBlock::iterator I = EndBB->begin();
1480 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1481 Value *OrigV = PN->getIncomingValueForBlock(BB);
1482 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1484 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1485 // Skip PHIs which are trivial.
1489 HaveRewritablePHIs = true;
1490 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1491 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1492 if (!OrigCE && !ThenCE)
1493 continue; // Known safe and cheap.
1495 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
1496 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
1498 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE) : 0;
1499 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE) : 0;
1500 if (OrigCost + ThenCost > 2 * PHINodeFoldingThreshold)
1503 // Account for the cost of an unfolded ConstantExpr which could end up
1504 // getting expanded into Instructions.
1505 // FIXME: This doesn't account for how many operations are combined in the
1506 // constant expression.
1508 if (SpeculationCost > 1)
1512 // If there are no PHIs to process, bail early. This helps ensure idempotence
1514 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
1517 // If we get here, we can hoist the instruction and if-convert.
1518 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1520 // Insert a select of the value of the speculated store.
1521 if (SpeculatedStoreValue) {
1522 IRBuilder<true, NoFolder> Builder(BI);
1523 Value *TrueV = SpeculatedStore->getValueOperand();
1524 Value *FalseV = SpeculatedStoreValue;
1526 std::swap(TrueV, FalseV);
1527 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() +
1528 "." + FalseV->getName());
1529 SpeculatedStore->setOperand(0, S);
1532 // Hoist the instructions.
1533 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(),
1534 llvm::prior(ThenBB->end()));
1536 // Insert selects and rewrite the PHI operands.
1537 IRBuilder<true, NoFolder> Builder(BI);
1538 for (BasicBlock::iterator I = EndBB->begin();
1539 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1540 unsigned OrigI = PN->getBasicBlockIndex(BB);
1541 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1542 Value *OrigV = PN->getIncomingValue(OrigI);
1543 Value *ThenV = PN->getIncomingValue(ThenI);
1545 // Skip PHIs which are trivial.
1549 // Create a select whose true value is the speculatively executed value and
1550 // false value is the preexisting value. Swap them if the branch
1551 // destinations were inverted.
1552 Value *TrueV = ThenV, *FalseV = OrigV;
1554 std::swap(TrueV, FalseV);
1555 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV,
1556 TrueV->getName() + "." + FalseV->getName());
1557 PN->setIncomingValue(OrigI, V);
1558 PN->setIncomingValue(ThenI, V);
1565 /// \returns True if this block contains a CallInst with the NoDuplicate
1567 static bool HasNoDuplicateCall(const BasicBlock *BB) {
1568 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1569 const CallInst *CI = dyn_cast<CallInst>(I);
1572 if (CI->cannotDuplicate())
1578 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1579 /// across this block.
1580 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1581 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1584 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1585 if (isa<DbgInfoIntrinsic>(BBI))
1587 if (Size > 10) return false; // Don't clone large BB's.
1590 // We can only support instructions that do not define values that are
1591 // live outside of the current basic block.
1592 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end();
1594 Instruction *U = cast<Instruction>(*UI);
1595 if (U->getParent() != BB || isa<PHINode>(U)) 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 *TD) {
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, TD)) {
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, TD) | 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 *TD) {
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, TD)) {
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->use_begin() == 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->use_begin() == 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();
2156 PredBlock->getInstList().insert(PBI, NewBonus);
2157 NewBonus->takeName(BonusInst);
2158 BonusInst->setName(BonusInst->getName()+".old");
2161 // Clone Cond into the predecessor basic block, and or/and the
2162 // two conditions together.
2163 Instruction *New = Cond->clone();
2164 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus);
2165 PredBlock->getInstList().insert(PBI, New);
2166 New->takeName(Cond);
2167 Cond->setName(New->getName()+".old");
2169 if (BI->isConditional()) {
2170 Instruction *NewCond =
2171 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2173 PBI->setCondition(NewCond);
2175 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2176 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2178 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2180 SmallVector<uint64_t, 8> NewWeights;
2182 if (PBI->getSuccessor(0) == BB) {
2183 if (PredHasWeights && SuccHasWeights) {
2184 // PBI: br i1 %x, BB, FalseDest
2185 // BI: br i1 %y, TrueDest, FalseDest
2186 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2187 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2188 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2189 // TrueWeight for PBI * FalseWeight for BI.
2190 // We assume that total weights of a BranchInst can fit into 32 bits.
2191 // Therefore, we will not have overflow using 64-bit arithmetic.
2192 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2193 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2195 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2196 PBI->setSuccessor(0, TrueDest);
2198 if (PBI->getSuccessor(1) == BB) {
2199 if (PredHasWeights && SuccHasWeights) {
2200 // PBI: br i1 %x, TrueDest, BB
2201 // BI: br i1 %y, TrueDest, FalseDest
2202 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2203 // FalseWeight for PBI * TrueWeight for BI.
2204 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2205 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2206 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2207 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2209 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2210 PBI->setSuccessor(1, FalseDest);
2212 if (NewWeights.size() == 2) {
2213 // Halve the weights if any of them cannot fit in an uint32_t
2214 FitWeights(NewWeights);
2216 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2217 PBI->setMetadata(LLVMContext::MD_prof,
2218 MDBuilder(BI->getContext()).
2219 createBranchWeights(MDWeights));
2221 PBI->setMetadata(LLVMContext::MD_prof, NULL);
2223 // Update PHI nodes in the common successors.
2224 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2225 ConstantInt *PBI_C = cast<ConstantInt>(
2226 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2227 assert(PBI_C->getType()->isIntegerTy(1));
2228 Instruction *MergedCond = 0;
2229 if (PBI->getSuccessor(0) == TrueDest) {
2230 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2231 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2232 // is false: !PBI_Cond and BI_Value
2233 Instruction *NotCond =
2234 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2237 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2242 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2243 PBI->getCondition(), MergedCond,
2246 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2247 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2248 // is false: PBI_Cond and BI_Value
2250 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2251 PBI->getCondition(), New,
2253 if (PBI_C->isOne()) {
2254 Instruction *NotCond =
2255 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2258 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2259 NotCond, MergedCond,
2264 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2267 // Change PBI from Conditional to Unconditional.
2268 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2269 EraseTerminatorInstAndDCECond(PBI);
2273 // TODO: If BB is reachable from all paths through PredBlock, then we
2274 // could replace PBI's branch probabilities with BI's.
2276 // Copy any debug value intrinsics into the end of PredBlock.
2277 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2278 if (isa<DbgInfoIntrinsic>(*I))
2279 I->clone()->insertBefore(PBI);
2286 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2287 /// predecessor of another block, this function tries to simplify it. We know
2288 /// that PBI and BI are both conditional branches, and BI is in one of the
2289 /// successor blocks of PBI - PBI branches to BI.
2290 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2291 assert(PBI->isConditional() && BI->isConditional());
2292 BasicBlock *BB = BI->getParent();
2294 // If this block ends with a branch instruction, and if there is a
2295 // predecessor that ends on a branch of the same condition, make
2296 // this conditional branch redundant.
2297 if (PBI->getCondition() == BI->getCondition() &&
2298 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2299 // Okay, the outcome of this conditional branch is statically
2300 // knowable. If this block had a single pred, handle specially.
2301 if (BB->getSinglePredecessor()) {
2302 // Turn this into a branch on constant.
2303 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2304 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2306 return true; // Nuke the branch on constant.
2309 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2310 // in the constant and simplify the block result. Subsequent passes of
2311 // simplifycfg will thread the block.
2312 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2313 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2314 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2315 std::distance(PB, PE),
2316 BI->getCondition()->getName() + ".pr",
2318 // Okay, we're going to insert the PHI node. Since PBI is not the only
2319 // predecessor, compute the PHI'd conditional value for all of the preds.
2320 // Any predecessor where the condition is not computable we keep symbolic.
2321 for (pred_iterator PI = PB; PI != PE; ++PI) {
2322 BasicBlock *P = *PI;
2323 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2324 PBI != BI && PBI->isConditional() &&
2325 PBI->getCondition() == BI->getCondition() &&
2326 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2327 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2328 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2331 NewPN->addIncoming(BI->getCondition(), P);
2335 BI->setCondition(NewPN);
2340 // If this is a conditional branch in an empty block, and if any
2341 // predecessors is a conditional branch to one of our destinations,
2342 // fold the conditions into logical ops and one cond br.
2343 BasicBlock::iterator BBI = BB->begin();
2344 // Ignore dbg intrinsics.
2345 while (isa<DbgInfoIntrinsic>(BBI))
2351 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2356 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2358 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2359 PBIOp = 0, BIOp = 1;
2360 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2361 PBIOp = 1, BIOp = 0;
2362 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2367 // Check to make sure that the other destination of this branch
2368 // isn't BB itself. If so, this is an infinite loop that will
2369 // keep getting unwound.
2370 if (PBI->getSuccessor(PBIOp) == BB)
2373 // Do not perform this transformation if it would require
2374 // insertion of a large number of select instructions. For targets
2375 // without predication/cmovs, this is a big pessimization.
2376 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2378 unsigned NumPhis = 0;
2379 for (BasicBlock::iterator II = CommonDest->begin();
2380 isa<PHINode>(II); ++II, ++NumPhis)
2381 if (NumPhis > 2) // Disable this xform.
2384 // Finally, if everything is ok, fold the branches to logical ops.
2385 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2387 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2388 << "AND: " << *BI->getParent());
2391 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2392 // branch in it, where one edge (OtherDest) goes back to itself but the other
2393 // exits. We don't *know* that the program avoids the infinite loop
2394 // (even though that seems likely). If we do this xform naively, we'll end up
2395 // recursively unpeeling the loop. Since we know that (after the xform is
2396 // done) that the block *is* infinite if reached, we just make it an obviously
2397 // infinite loop with no cond branch.
2398 if (OtherDest == BB) {
2399 // Insert it at the end of the function, because it's either code,
2400 // or it won't matter if it's hot. :)
2401 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2402 "infloop", BB->getParent());
2403 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2404 OtherDest = InfLoopBlock;
2407 DEBUG(dbgs() << *PBI->getParent()->getParent());
2409 // BI may have other predecessors. Because of this, we leave
2410 // it alone, but modify PBI.
2412 // Make sure we get to CommonDest on True&True directions.
2413 Value *PBICond = PBI->getCondition();
2414 IRBuilder<true, NoFolder> Builder(PBI);
2416 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2418 Value *BICond = BI->getCondition();
2420 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2422 // Merge the conditions.
2423 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2425 // Modify PBI to branch on the new condition to the new dests.
2426 PBI->setCondition(Cond);
2427 PBI->setSuccessor(0, CommonDest);
2428 PBI->setSuccessor(1, OtherDest);
2430 // Update branch weight for PBI.
2431 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2432 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2434 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2436 if (PredHasWeights && SuccHasWeights) {
2437 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2438 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2439 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2440 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2441 // The weight to CommonDest should be PredCommon * SuccTotal +
2442 // PredOther * SuccCommon.
2443 // The weight to OtherDest should be PredOther * SuccOther.
2444 SmallVector<uint64_t, 2> NewWeights;
2445 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) +
2446 PredOther * SuccCommon);
2447 NewWeights.push_back(PredOther * SuccOther);
2448 // Halve the weights if any of them cannot fit in an uint32_t
2449 FitWeights(NewWeights);
2451 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end());
2452 PBI->setMetadata(LLVMContext::MD_prof,
2453 MDBuilder(BI->getContext()).
2454 createBranchWeights(MDWeights));
2457 // OtherDest may have phi nodes. If so, add an entry from PBI's
2458 // block that are identical to the entries for BI's block.
2459 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2461 // We know that the CommonDest already had an edge from PBI to
2462 // it. If it has PHIs though, the PHIs may have different
2463 // entries for BB and PBI's BB. If so, insert a select to make
2466 for (BasicBlock::iterator II = CommonDest->begin();
2467 (PN = dyn_cast<PHINode>(II)); ++II) {
2468 Value *BIV = PN->getIncomingValueForBlock(BB);
2469 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2470 Value *PBIV = PN->getIncomingValue(PBBIdx);
2472 // Insert a select in PBI to pick the right value.
2473 Value *NV = cast<SelectInst>
2474 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2475 PN->setIncomingValue(PBBIdx, NV);
2479 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2480 DEBUG(dbgs() << *PBI->getParent()->getParent());
2482 // This basic block is probably dead. We know it has at least
2483 // one fewer predecessor.
2487 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2488 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2489 // Takes care of updating the successors and removing the old terminator.
2490 // Also makes sure not to introduce new successors by assuming that edges to
2491 // non-successor TrueBBs and FalseBBs aren't reachable.
2492 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2493 BasicBlock *TrueBB, BasicBlock *FalseBB,
2494 uint32_t TrueWeight,
2495 uint32_t FalseWeight){
2496 // Remove any superfluous successor edges from the CFG.
2497 // First, figure out which successors to preserve.
2498 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2500 BasicBlock *KeepEdge1 = TrueBB;
2501 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0;
2503 // Then remove the rest.
2504 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2505 BasicBlock *Succ = OldTerm->getSuccessor(I);
2506 // Make sure only to keep exactly one copy of each edge.
2507 if (Succ == KeepEdge1)
2509 else if (Succ == KeepEdge2)
2512 Succ->removePredecessor(OldTerm->getParent());
2515 IRBuilder<> Builder(OldTerm);
2516 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2518 // Insert an appropriate new terminator.
2519 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) {
2520 if (TrueBB == FalseBB)
2521 // We were only looking for one successor, and it was present.
2522 // Create an unconditional branch to it.
2523 Builder.CreateBr(TrueBB);
2525 // We found both of the successors we were looking for.
2526 // Create a conditional branch sharing the condition of the select.
2527 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2528 if (TrueWeight != FalseWeight)
2529 NewBI->setMetadata(LLVMContext::MD_prof,
2530 MDBuilder(OldTerm->getContext()).
2531 createBranchWeights(TrueWeight, FalseWeight));
2533 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2534 // Neither of the selected blocks were successors, so this
2535 // terminator must be unreachable.
2536 new UnreachableInst(OldTerm->getContext(), OldTerm);
2538 // One of the selected values was a successor, but the other wasn't.
2539 // Insert an unconditional branch to the one that was found;
2540 // the edge to the one that wasn't must be unreachable.
2542 // Only TrueBB was found.
2543 Builder.CreateBr(TrueBB);
2545 // Only FalseBB was found.
2546 Builder.CreateBr(FalseBB);
2549 EraseTerminatorInstAndDCECond(OldTerm);
2553 // SimplifySwitchOnSelect - Replaces
2554 // (switch (select cond, X, Y)) on constant X, Y
2555 // with a branch - conditional if X and Y lead to distinct BBs,
2556 // unconditional otherwise.
2557 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2558 // Check for constant integer values in the select.
2559 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2560 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2561 if (!TrueVal || !FalseVal)
2564 // Find the relevant condition and destinations.
2565 Value *Condition = Select->getCondition();
2566 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2567 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2569 // Get weight for TrueBB and FalseBB.
2570 uint32_t TrueWeight = 0, FalseWeight = 0;
2571 SmallVector<uint64_t, 8> Weights;
2572 bool HasWeights = HasBranchWeights(SI);
2574 GetBranchWeights(SI, Weights);
2575 if (Weights.size() == 1 + SI->getNumCases()) {
2576 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2577 getSuccessorIndex()];
2578 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2579 getSuccessorIndex()];
2583 // Perform the actual simplification.
2584 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2585 TrueWeight, FalseWeight);
2588 // SimplifyIndirectBrOnSelect - Replaces
2589 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2590 // blockaddress(@fn, BlockB)))
2592 // (br cond, BlockA, BlockB).
2593 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2594 // Check that both operands of the select are block addresses.
2595 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2596 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2600 // Extract the actual blocks.
2601 BasicBlock *TrueBB = TBA->getBasicBlock();
2602 BasicBlock *FalseBB = FBA->getBasicBlock();
2604 // Perform the actual simplification.
2605 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2609 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2610 /// instruction (a seteq/setne with a constant) as the only instruction in a
2611 /// block that ends with an uncond branch. We are looking for a very specific
2612 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2613 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2614 /// default value goes to an uncond block with a seteq in it, we get something
2617 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2619 /// %tmp = icmp eq i8 %A, 92
2622 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2624 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2625 /// the PHI, merging the third icmp into the switch.
2626 static bool TryToSimplifyUncondBranchWithICmpInIt(
2627 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI,
2628 const DataLayout *TD) {
2629 BasicBlock *BB = ICI->getParent();
2631 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2633 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2635 Value *V = ICI->getOperand(0);
2636 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2638 // The pattern we're looking for is where our only predecessor is a switch on
2639 // 'V' and this block is the default case for the switch. In this case we can
2640 // fold the compared value into the switch to simplify things.
2641 BasicBlock *Pred = BB->getSinglePredecessor();
2642 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false;
2644 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2645 if (SI->getCondition() != V)
2648 // If BB is reachable on a non-default case, then we simply know the value of
2649 // V in this block. Substitute it and constant fold the icmp instruction
2651 if (SI->getDefaultDest() != BB) {
2652 ConstantInt *VVal = SI->findCaseDest(BB);
2653 assert(VVal && "Should have a unique destination value");
2654 ICI->setOperand(0, VVal);
2656 if (Value *V = SimplifyInstruction(ICI, TD)) {
2657 ICI->replaceAllUsesWith(V);
2658 ICI->eraseFromParent();
2660 // BB is now empty, so it is likely to simplify away.
2661 return SimplifyCFG(BB, TTI, TD) | true;
2664 // Ok, the block is reachable from the default dest. If the constant we're
2665 // comparing exists in one of the other edges, then we can constant fold ICI
2667 if (SI->findCaseValue(Cst) != SI->case_default()) {
2669 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2670 V = ConstantInt::getFalse(BB->getContext());
2672 V = ConstantInt::getTrue(BB->getContext());
2674 ICI->replaceAllUsesWith(V);
2675 ICI->eraseFromParent();
2676 // BB is now empty, so it is likely to simplify away.
2677 return SimplifyCFG(BB, TTI, TD) | true;
2680 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2682 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2683 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back());
2684 if (PHIUse == 0 || PHIUse != &SuccBlock->front() ||
2685 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2688 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2690 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2691 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2693 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2694 std::swap(DefaultCst, NewCst);
2696 // Replace ICI (which is used by the PHI for the default value) with true or
2697 // false depending on if it is EQ or NE.
2698 ICI->replaceAllUsesWith(DefaultCst);
2699 ICI->eraseFromParent();
2701 // Okay, the switch goes to this block on a default value. Add an edge from
2702 // the switch to the merge point on the compared value.
2703 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2704 BB->getParent(), BB);
2705 SmallVector<uint64_t, 8> Weights;
2706 bool HasWeights = HasBranchWeights(SI);
2708 GetBranchWeights(SI, Weights);
2709 if (Weights.size() == 1 + SI->getNumCases()) {
2710 // Split weight for default case to case for "Cst".
2711 Weights[0] = (Weights[0]+1) >> 1;
2712 Weights.push_back(Weights[0]);
2714 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2715 SI->setMetadata(LLVMContext::MD_prof,
2716 MDBuilder(SI->getContext()).
2717 createBranchWeights(MDWeights));
2720 SI->addCase(Cst, NewBB);
2722 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2723 Builder.SetInsertPoint(NewBB);
2724 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2725 Builder.CreateBr(SuccBlock);
2726 PHIUse->addIncoming(NewCst, NewBB);
2730 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2731 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2732 /// fold it into a switch instruction if so.
2733 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *TD,
2734 IRBuilder<> &Builder) {
2735 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2736 if (Cond == 0) return false;
2739 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2740 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2741 // 'setne's and'ed together, collect them.
2743 std::vector<ConstantInt*> Values;
2744 bool TrueWhenEqual = true;
2745 Value *ExtraCase = 0;
2746 unsigned UsedICmps = 0;
2748 if (Cond->getOpcode() == Instruction::Or) {
2749 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true,
2751 } else if (Cond->getOpcode() == Instruction::And) {
2752 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false,
2754 TrueWhenEqual = false;
2757 // If we didn't have a multiply compared value, fail.
2758 if (CompVal == 0) return false;
2760 // Avoid turning single icmps into a switch.
2764 // There might be duplicate constants in the list, which the switch
2765 // instruction can't handle, remove them now.
2766 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2767 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2769 // If Extra was used, we require at least two switch values to do the
2770 // transformation. A switch with one value is just an cond branch.
2771 if (ExtraCase && Values.size() < 2) return false;
2773 // TODO: Preserve branch weight metadata, similarly to how
2774 // FoldValueComparisonIntoPredecessors preserves it.
2776 // Figure out which block is which destination.
2777 BasicBlock *DefaultBB = BI->getSuccessor(1);
2778 BasicBlock *EdgeBB = BI->getSuccessor(0);
2779 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2781 BasicBlock *BB = BI->getParent();
2783 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2784 << " cases into SWITCH. BB is:\n" << *BB);
2786 // If there are any extra values that couldn't be folded into the switch
2787 // then we evaluate them with an explicit branch first. Split the block
2788 // right before the condbr to handle it.
2790 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2791 // Remove the uncond branch added to the old block.
2792 TerminatorInst *OldTI = BB->getTerminator();
2793 Builder.SetInsertPoint(OldTI);
2796 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2798 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2800 OldTI->eraseFromParent();
2802 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2803 // for the edge we just added.
2804 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2806 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2807 << "\nEXTRABB = " << *BB);
2811 Builder.SetInsertPoint(BI);
2812 // Convert pointer to int before we switch.
2813 if (CompVal->getType()->isPointerTy()) {
2814 assert(TD && "Cannot switch on pointer without DataLayout");
2815 CompVal = Builder.CreatePtrToInt(CompVal,
2816 TD->getIntPtrType(CompVal->getType()),
2820 // Create the new switch instruction now.
2821 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2823 // Add all of the 'cases' to the switch instruction.
2824 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2825 New->addCase(Values[i], EdgeBB);
2827 // We added edges from PI to the EdgeBB. As such, if there were any
2828 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2829 // the number of edges added.
2830 for (BasicBlock::iterator BBI = EdgeBB->begin();
2831 isa<PHINode>(BBI); ++BBI) {
2832 PHINode *PN = cast<PHINode>(BBI);
2833 Value *InVal = PN->getIncomingValueForBlock(BB);
2834 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2835 PN->addIncoming(InVal, BB);
2838 // Erase the old branch instruction.
2839 EraseTerminatorInstAndDCECond(BI);
2841 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2845 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2846 // If this is a trivial landing pad that just continues unwinding the caught
2847 // exception then zap the landing pad, turning its invokes into calls.
2848 BasicBlock *BB = RI->getParent();
2849 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2850 if (RI->getValue() != LPInst)
2851 // Not a landing pad, or the resume is not unwinding the exception that
2852 // caused control to branch here.
2855 // Check that there are no other instructions except for debug intrinsics.
2856 BasicBlock::iterator I = LPInst, E = RI;
2858 if (!isa<DbgInfoIntrinsic>(I))
2861 // Turn all invokes that unwind here into calls and delete the basic block.
2862 bool InvokeRequiresTableEntry = false;
2863 bool Changed = false;
2864 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2865 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2867 if (II->hasFnAttr(Attribute::UWTable)) {
2868 // Don't remove an `invoke' instruction if the ABI requires an entry into
2870 InvokeRequiresTableEntry = true;
2874 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2876 // Insert a call instruction before the invoke.
2877 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2879 Call->setCallingConv(II->getCallingConv());
2880 Call->setAttributes(II->getAttributes());
2881 Call->setDebugLoc(II->getDebugLoc());
2883 // Anything that used the value produced by the invoke instruction now uses
2884 // the value produced by the call instruction. Note that we do this even
2885 // for void functions and calls with no uses so that the callgraph edge is
2887 II->replaceAllUsesWith(Call);
2888 BB->removePredecessor(II->getParent());
2890 // Insert a branch to the normal destination right before the invoke.
2891 BranchInst::Create(II->getNormalDest(), II);
2893 // Finally, delete the invoke instruction!
2894 II->eraseFromParent();
2898 if (!InvokeRequiresTableEntry)
2899 // The landingpad is now unreachable. Zap it.
2900 BB->eraseFromParent();
2905 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2906 BasicBlock *BB = RI->getParent();
2907 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2909 // Find predecessors that end with branches.
2910 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2911 SmallVector<BranchInst*, 8> CondBranchPreds;
2912 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2913 BasicBlock *P = *PI;
2914 TerminatorInst *PTI = P->getTerminator();
2915 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2916 if (BI->isUnconditional())
2917 UncondBranchPreds.push_back(P);
2919 CondBranchPreds.push_back(BI);
2923 // If we found some, do the transformation!
2924 if (!UncondBranchPreds.empty() && DupRet) {
2925 while (!UncondBranchPreds.empty()) {
2926 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
2927 DEBUG(dbgs() << "FOLDING: " << *BB
2928 << "INTO UNCOND BRANCH PRED: " << *Pred);
2929 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
2932 // If we eliminated all predecessors of the block, delete the block now.
2933 if (pred_begin(BB) == pred_end(BB))
2934 // We know there are no successors, so just nuke the block.
2935 BB->eraseFromParent();
2940 // Check out all of the conditional branches going to this return
2941 // instruction. If any of them just select between returns, change the
2942 // branch itself into a select/return pair.
2943 while (!CondBranchPreds.empty()) {
2944 BranchInst *BI = CondBranchPreds.pop_back_val();
2946 // Check to see if the non-BB successor is also a return block.
2947 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
2948 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
2949 SimplifyCondBranchToTwoReturns(BI, Builder))
2955 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
2956 BasicBlock *BB = UI->getParent();
2958 bool Changed = false;
2960 // If there are any instructions immediately before the unreachable that can
2961 // be removed, do so.
2962 while (UI != BB->begin()) {
2963 BasicBlock::iterator BBI = UI;
2965 // Do not delete instructions that can have side effects which might cause
2966 // the unreachable to not be reachable; specifically, calls and volatile
2967 // operations may have this effect.
2968 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
2970 if (BBI->mayHaveSideEffects()) {
2971 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
2972 if (SI->isVolatile())
2974 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
2975 if (LI->isVolatile())
2977 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
2978 if (RMWI->isVolatile())
2980 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
2981 if (CXI->isVolatile())
2983 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
2984 !isa<LandingPadInst>(BBI)) {
2987 // Note that deleting LandingPad's here is in fact okay, although it
2988 // involves a bit of subtle reasoning. If this inst is a LandingPad,
2989 // all the predecessors of this block will be the unwind edges of Invokes,
2990 // and we can therefore guarantee this block will be erased.
2993 // Delete this instruction (any uses are guaranteed to be dead)
2994 if (!BBI->use_empty())
2995 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
2996 BBI->eraseFromParent();
3000 // If the unreachable instruction is the first in the block, take a gander
3001 // at all of the predecessors of this instruction, and simplify them.
3002 if (&BB->front() != UI) return Changed;
3004 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
3005 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
3006 TerminatorInst *TI = Preds[i]->getTerminator();
3007 IRBuilder<> Builder(TI);
3008 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3009 if (BI->isUnconditional()) {
3010 if (BI->getSuccessor(0) == BB) {
3011 new UnreachableInst(TI->getContext(), TI);
3012 TI->eraseFromParent();
3016 if (BI->getSuccessor(0) == BB) {
3017 Builder.CreateBr(BI->getSuccessor(1));
3018 EraseTerminatorInstAndDCECond(BI);
3019 } else if (BI->getSuccessor(1) == BB) {
3020 Builder.CreateBr(BI->getSuccessor(0));
3021 EraseTerminatorInstAndDCECond(BI);
3025 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3026 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3028 if (i.getCaseSuccessor() == BB) {
3029 BB->removePredecessor(SI->getParent());
3034 // If the default value is unreachable, figure out the most popular
3035 // destination and make it the default.
3036 if (SI->getDefaultDest() == BB) {
3037 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
3038 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3040 std::pair<unsigned, unsigned> &entry =
3041 Popularity[i.getCaseSuccessor()];
3042 if (entry.first == 0) {
3044 entry.second = i.getCaseIndex();
3050 // Find the most popular block.
3051 unsigned MaxPop = 0;
3052 unsigned MaxIndex = 0;
3053 BasicBlock *MaxBlock = 0;
3054 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
3055 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
3056 if (I->second.first > MaxPop ||
3057 (I->second.first == MaxPop && MaxIndex > I->second.second)) {
3058 MaxPop = I->second.first;
3059 MaxIndex = I->second.second;
3060 MaxBlock = I->first;
3064 // Make this the new default, allowing us to delete any explicit
3066 SI->setDefaultDest(MaxBlock);
3069 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
3071 if (isa<PHINode>(MaxBlock->begin()))
3072 for (unsigned i = 0; i != MaxPop-1; ++i)
3073 MaxBlock->removePredecessor(SI->getParent());
3075 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3077 if (i.getCaseSuccessor() == MaxBlock) {
3083 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
3084 if (II->getUnwindDest() == BB) {
3085 // Convert the invoke to a call instruction. This would be a good
3086 // place to note that the call does not throw though.
3087 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
3088 II->removeFromParent(); // Take out of symbol table
3090 // Insert the call now...
3091 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
3092 Builder.SetInsertPoint(BI);
3093 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
3094 Args, II->getName());
3095 CI->setCallingConv(II->getCallingConv());
3096 CI->setAttributes(II->getAttributes());
3097 // If the invoke produced a value, the call does now instead.
3098 II->replaceAllUsesWith(CI);
3105 // If this block is now dead, remove it.
3106 if (pred_begin(BB) == pred_end(BB) &&
3107 BB != &BB->getParent()->getEntryBlock()) {
3108 // We know there are no successors, so just nuke the block.
3109 BB->eraseFromParent();
3116 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
3117 /// integer range comparison into a sub, an icmp and a branch.
3118 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3119 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3121 // Make sure all cases point to the same destination and gather the values.
3122 SmallVector<ConstantInt *, 16> Cases;
3123 SwitchInst::CaseIt I = SI->case_begin();
3124 Cases.push_back(I.getCaseValue());
3125 SwitchInst::CaseIt PrevI = I++;
3126 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
3127 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
3129 Cases.push_back(I.getCaseValue());
3131 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
3133 // Sort the case values, then check if they form a range we can transform.
3134 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3135 for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
3136 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
3140 Constant *Offset = ConstantExpr::getNeg(Cases.back());
3141 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
3143 Value *Sub = SI->getCondition();
3144 if (!Offset->isNullValue())
3145 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
3147 // If NumCases overflowed, then all possible values jump to the successor.
3148 if (NumCases->isNullValue() && SI->getNumCases() != 0)
3149 Cmp = ConstantInt::getTrue(SI->getContext());
3151 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3152 BranchInst *NewBI = Builder.CreateCondBr(
3153 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
3155 // Update weight for the newly-created conditional branch.
3156 SmallVector<uint64_t, 8> Weights;
3157 bool HasWeights = HasBranchWeights(SI);
3159 GetBranchWeights(SI, Weights);
3160 if (Weights.size() == 1 + SI->getNumCases()) {
3161 // Combine all weights for the cases to be the true weight of NewBI.
3162 // We assume that the sum of all weights for a Terminator can fit into 32
3164 uint32_t NewTrueWeight = 0;
3165 for (unsigned I = 1, E = Weights.size(); I != E; ++I)
3166 NewTrueWeight += (uint32_t)Weights[I];
3167 NewBI->setMetadata(LLVMContext::MD_prof,
3168 MDBuilder(SI->getContext()).
3169 createBranchWeights(NewTrueWeight,
3170 (uint32_t)Weights[0]));
3174 // Prune obsolete incoming values off the successor's PHI nodes.
3175 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
3176 isa<PHINode>(BBI); ++BBI) {
3177 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
3178 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3180 SI->eraseFromParent();
3185 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
3186 /// and use it to remove dead cases.
3187 static bool EliminateDeadSwitchCases(SwitchInst *SI) {
3188 Value *Cond = SI->getCondition();
3189 unsigned Bits = Cond->getType()->getIntegerBitWidth();
3190 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3191 ComputeMaskedBits(Cond, KnownZero, KnownOne);
3193 // Gather dead cases.
3194 SmallVector<ConstantInt*, 8> DeadCases;
3195 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3196 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3197 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3198 DeadCases.push_back(I.getCaseValue());
3199 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3200 << I.getCaseValue() << "' is dead.\n");
3204 SmallVector<uint64_t, 8> Weights;
3205 bool HasWeight = HasBranchWeights(SI);
3207 GetBranchWeights(SI, Weights);
3208 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3211 // Remove dead cases from the switch.
3212 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3213 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3214 assert(Case != SI->case_default() &&
3215 "Case was not found. Probably mistake in DeadCases forming.");
3217 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3221 // Prune unused values from PHI nodes.
3222 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3223 SI->removeCase(Case);
3226 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3227 SI->setMetadata(LLVMContext::MD_prof,
3228 MDBuilder(SI->getParent()->getContext()).
3229 createBranchWeights(MDWeights));
3232 return !DeadCases.empty();
3235 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
3236 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3237 /// by an unconditional branch), look at the phi node for BB in the successor
3238 /// block and see if the incoming value is equal to CaseValue. If so, return
3239 /// the phi node, and set PhiIndex to BB's index in the phi node.
3240 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3243 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3244 return NULL; // BB must be empty to be a candidate for simplification.
3245 if (!BB->getSinglePredecessor())
3246 return NULL; // BB must be dominated by the switch.
3248 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3249 if (!Branch || !Branch->isUnconditional())
3250 return NULL; // Terminator must be unconditional branch.
3252 BasicBlock *Succ = Branch->getSuccessor(0);
3254 BasicBlock::iterator I = Succ->begin();
3255 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3256 int Idx = PHI->getBasicBlockIndex(BB);
3257 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3259 Value *InValue = PHI->getIncomingValue(Idx);
3260 if (InValue != CaseValue) continue;
3269 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
3270 /// instruction to a phi node dominated by the switch, if that would mean that
3271 /// some of the destination blocks of the switch can be folded away.
3272 /// Returns true if a change is made.
3273 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3274 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3275 ForwardingNodesMap ForwardingNodes;
3277 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3278 ConstantInt *CaseValue = I.getCaseValue();
3279 BasicBlock *CaseDest = I.getCaseSuccessor();
3282 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3286 ForwardingNodes[PHI].push_back(PhiIndex);
3289 bool Changed = false;
3291 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3292 E = ForwardingNodes.end(); I != E; ++I) {
3293 PHINode *Phi = I->first;
3294 SmallVectorImpl<int> &Indexes = I->second;
3296 if (Indexes.size() < 2) continue;
3298 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3299 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3306 /// ValidLookupTableConstant - Return true if the backend will be able to handle
3307 /// initializing an array of constants like C.
3308 static bool ValidLookupTableConstant(Constant *C) {
3309 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3310 return CE->isGEPWithNoNotionalOverIndexing();
3312 return isa<ConstantFP>(C) ||
3313 isa<ConstantInt>(C) ||
3314 isa<ConstantPointerNull>(C) ||
3315 isa<GlobalValue>(C) ||
3319 /// LookupConstant - If V is a Constant, return it. Otherwise, try to look up
3320 /// its constant value in ConstantPool, returning 0 if it's not there.
3321 static Constant *LookupConstant(Value *V,
3322 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3323 if (Constant *C = dyn_cast<Constant>(V))
3325 return ConstantPool.lookup(V);
3328 /// ConstantFold - Try to fold instruction I into a constant. This works for
3329 /// simple instructions such as binary operations where both operands are
3330 /// constant or can be replaced by constants from the ConstantPool. Returns the
3331 /// resulting constant on success, 0 otherwise.
3333 ConstantFold(Instruction *I,
3334 const SmallDenseMap<Value *, Constant *> &ConstantPool,
3335 const DataLayout *DL) {
3336 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3337 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3340 if (A->isAllOnesValue())
3341 return LookupConstant(Select->getTrueValue(), ConstantPool);
3342 if (A->isNullValue())
3343 return LookupConstant(Select->getFalseValue(), ConstantPool);
3347 SmallVector<Constant *, 4> COps;
3348 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
3349 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
3355 if (CmpInst *Cmp = dyn_cast<CmpInst>(I))
3356 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
3359 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), COps, DL);
3362 /// GetCaseResults - Try to determine the resulting constant values in phi nodes
3363 /// at the common destination basic block, *CommonDest, for one of the case
3364 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3365 /// case), of a switch instruction SI.
3367 GetCaseResults(SwitchInst *SI,
3368 ConstantInt *CaseVal,
3369 BasicBlock *CaseDest,
3370 BasicBlock **CommonDest,
3371 SmallVectorImpl<std::pair<PHINode *, Constant *> > &Res,
3372 const DataLayout *DL) {
3373 // The block from which we enter the common destination.
3374 BasicBlock *Pred = SI->getParent();
3376 // If CaseDest is empty except for some side-effect free instructions through
3377 // which we can constant-propagate the CaseVal, continue to its successor.
3378 SmallDenseMap<Value*, Constant*> ConstantPool;
3379 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3380 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3382 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3383 // If the terminator is a simple branch, continue to the next block.
3384 if (T->getNumSuccessors() != 1)
3387 CaseDest = T->getSuccessor(0);
3388 } else if (isa<DbgInfoIntrinsic>(I)) {
3389 // Skip debug intrinsic.
3391 } else if (Constant *C = ConstantFold(I, ConstantPool, DL)) {
3392 // Instruction is side-effect free and constant.
3393 ConstantPool.insert(std::make_pair(I, C));
3399 // If we did not have a CommonDest before, use the current one.
3401 *CommonDest = CaseDest;
3402 // If the destination isn't the common one, abort.
3403 if (CaseDest != *CommonDest)
3406 // Get the values for this case from phi nodes in the destination block.
3407 BasicBlock::iterator I = (*CommonDest)->begin();
3408 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3409 int Idx = PHI->getBasicBlockIndex(Pred);
3413 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3418 // Note: If the constant comes from constant-propagating the case value
3419 // through the CaseDest basic block, it will be safe to remove the
3420 // instructions in that block. They cannot be used (except in the phi nodes
3421 // we visit) outside CaseDest, because that block does not dominate its
3422 // successor. If it did, we would not be in this phi node.
3424 // Be conservative about which kinds of constants we support.
3425 if (!ValidLookupTableConstant(ConstVal))
3428 Res.push_back(std::make_pair(PHI, ConstVal));
3435 /// SwitchLookupTable - This class represents a lookup table that can be used
3436 /// to replace a switch.
3437 class SwitchLookupTable {
3439 /// SwitchLookupTable - Create a lookup table to use as a switch replacement
3440 /// with the contents of Values, using DefaultValue to fill any holes in the
3442 SwitchLookupTable(Module &M,
3444 ConstantInt *Offset,
3445 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3446 Constant *DefaultValue,
3447 const DataLayout *TD);
3449 /// BuildLookup - Build instructions with Builder to retrieve the value at
3450 /// the position given by Index in the lookup table.
3451 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3453 /// WouldFitInRegister - Return true if a table with TableSize elements of
3454 /// type ElementType would fit in a target-legal register.
3455 static bool WouldFitInRegister(const DataLayout *TD,
3457 const Type *ElementType);
3460 // Depending on the contents of the table, it can be represented in
3463 // For tables where each element contains the same value, we just have to
3464 // store that single value and return it for each lookup.
3467 // For small tables with integer elements, we can pack them into a bitmap
3468 // that fits into a target-legal register. Values are retrieved by
3469 // shift and mask operations.
3472 // The table is stored as an array of values. Values are retrieved by load
3473 // instructions from the table.
3477 // For SingleValueKind, this is the single value.
3478 Constant *SingleValue;
3480 // For BitMapKind, this is the bitmap.
3481 ConstantInt *BitMap;
3482 IntegerType *BitMapElementTy;
3484 // For ArrayKind, this is the array.
3485 GlobalVariable *Array;
3489 SwitchLookupTable::SwitchLookupTable(Module &M,
3491 ConstantInt *Offset,
3492 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3493 Constant *DefaultValue,
3494 const DataLayout *TD)
3495 : SingleValue(0), BitMap(0), BitMapElementTy(0), Array(0) {
3496 assert(Values.size() && "Can't build lookup table without values!");
3497 assert(TableSize >= Values.size() && "Can't fit values in table!");
3499 // If all values in the table are equal, this is that value.
3500 SingleValue = Values.begin()->second;
3502 // Build up the table contents.
3503 SmallVector<Constant*, 64> TableContents(TableSize);
3504 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3505 ConstantInt *CaseVal = Values[I].first;
3506 Constant *CaseRes = Values[I].second;
3507 assert(CaseRes->getType() == DefaultValue->getType());
3509 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3511 TableContents[Idx] = CaseRes;
3513 if (CaseRes != SingleValue)
3517 // Fill in any holes in the table with the default result.
3518 if (Values.size() < TableSize) {
3519 for (uint64_t I = 0; I < TableSize; ++I) {
3520 if (!TableContents[I])
3521 TableContents[I] = DefaultValue;
3524 if (DefaultValue != SingleValue)
3528 // If each element in the table contains the same value, we only need to store
3529 // that single value.
3531 Kind = SingleValueKind;
3535 // If the type is integer and the table fits in a register, build a bitmap.
3536 if (WouldFitInRegister(TD, TableSize, DefaultValue->getType())) {
3537 IntegerType *IT = cast<IntegerType>(DefaultValue->getType());
3538 APInt TableInt(TableSize * IT->getBitWidth(), 0);
3539 for (uint64_t I = TableSize; I > 0; --I) {
3540 TableInt <<= IT->getBitWidth();
3541 // Insert values into the bitmap. Undef values are set to zero.
3542 if (!isa<UndefValue>(TableContents[I - 1])) {
3543 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
3544 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
3547 BitMap = ConstantInt::get(M.getContext(), TableInt);
3548 BitMapElementTy = IT;
3554 // Store the table in an array.
3555 ArrayType *ArrayTy = ArrayType::get(DefaultValue->getType(), TableSize);
3556 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3558 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3559 GlobalVariable::PrivateLinkage,
3562 Array->setUnnamedAddr(true);
3566 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
3568 case SingleValueKind:
3571 // Type of the bitmap (e.g. i59).
3572 IntegerType *MapTy = BitMap->getType();
3574 // Cast Index to the same type as the bitmap.
3575 // Note: The Index is <= the number of elements in the table, so
3576 // truncating it to the width of the bitmask is safe.
3577 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
3579 // Multiply the shift amount by the element width.
3580 ShiftAmt = Builder.CreateMul(ShiftAmt,
3581 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
3585 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
3586 "switch.downshift");
3588 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
3592 Value *GEPIndices[] = { Builder.getInt32(0), Index };
3593 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices,
3595 return Builder.CreateLoad(GEP, "switch.load");
3598 llvm_unreachable("Unknown lookup table kind!");
3601 bool SwitchLookupTable::WouldFitInRegister(const DataLayout *TD,
3603 const Type *ElementType) {
3606 const IntegerType *IT = dyn_cast<IntegerType>(ElementType);
3609 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
3610 // are <= 15, we could try to narrow the type.
3612 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
3613 if (TableSize >= UINT_MAX/IT->getBitWidth())
3615 return TD->fitsInLegalInteger(TableSize * IT->getBitWidth());
3618 /// ShouldBuildLookupTable - Determine whether a lookup table should be built
3619 /// for this switch, based on the number of cases, size of the table and the
3620 /// types of the results.
3621 static bool ShouldBuildLookupTable(SwitchInst *SI,
3623 const TargetTransformInfo &TTI,
3624 const DataLayout *TD,
3625 const SmallDenseMap<PHINode*, Type*>& ResultTypes) {
3626 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
3627 return false; // TableSize overflowed, or mul below might overflow.
3629 bool AllTablesFitInRegister = true;
3630 bool HasIllegalType = false;
3631 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(),
3632 E = ResultTypes.end(); I != E; ++I) {
3633 Type *Ty = I->second;
3635 // Saturate this flag to true.
3636 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
3638 // Saturate this flag to false.
3639 AllTablesFitInRegister = AllTablesFitInRegister &&
3640 SwitchLookupTable::WouldFitInRegister(TD, TableSize, Ty);
3642 // If both flags saturate, we're done. NOTE: This *only* works with
3643 // saturating flags, and all flags have to saturate first due to the
3644 // non-deterministic behavior of iterating over a dense map.
3645 if (HasIllegalType && !AllTablesFitInRegister)
3649 // If each table would fit in a register, we should build it anyway.
3650 if (AllTablesFitInRegister)
3653 // Don't build a table that doesn't fit in-register if it has illegal types.
3657 // The table density should be at least 40%. This is the same criterion as for
3658 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3659 // FIXME: Find the best cut-off.
3660 return SI->getNumCases() * 10 >= TableSize * 4;
3663 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3664 /// phi nodes in a common successor block with different constant values,
3665 /// replace the switch with lookup tables.
3666 static bool SwitchToLookupTable(SwitchInst *SI,
3667 IRBuilder<> &Builder,
3668 const TargetTransformInfo &TTI,
3669 const DataLayout* TD) {
3670 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3672 // Only build lookup table when we have a target that supports it.
3673 if (!TTI.shouldBuildLookupTables())
3676 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3677 // split off a dense part and build a lookup table for that.
3679 // FIXME: This creates arrays of GEPs to constant strings, which means each
3680 // GEP needs a runtime relocation in PIC code. We should just build one big
3681 // string and lookup indices into that.
3683 // Ignore the switch if the number of cases is too small.
3684 // This is similar to the check when building jump tables in
3685 // SelectionDAGBuilder::handleJTSwitchCase.
3686 // FIXME: Determine the best cut-off.
3687 if (SI->getNumCases() < 4)
3690 // Figure out the corresponding result for each case value and phi node in the
3691 // common destination, as well as the the min and max case values.
3692 assert(SI->case_begin() != SI->case_end());
3693 SwitchInst::CaseIt CI = SI->case_begin();
3694 ConstantInt *MinCaseVal = CI.getCaseValue();
3695 ConstantInt *MaxCaseVal = CI.getCaseValue();
3697 BasicBlock *CommonDest = 0;
3698 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
3699 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
3700 SmallDenseMap<PHINode*, Constant*> DefaultResults;
3701 SmallDenseMap<PHINode*, Type*> ResultTypes;
3702 SmallVector<PHINode*, 4> PHIs;
3704 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
3705 ConstantInt *CaseVal = CI.getCaseValue();
3706 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
3707 MinCaseVal = CaseVal;
3708 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
3709 MaxCaseVal = CaseVal;
3711 // Resulting value at phi nodes for this case value.
3712 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
3714 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
3718 // Append the result from this case to the list for each phi.
3719 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) {
3720 if (!ResultLists.count(I->first))
3721 PHIs.push_back(I->first);
3722 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second));
3726 // Get the resulting values for the default case.
3727 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
3728 if (!GetCaseResults(SI, 0, SI->getDefaultDest(), &CommonDest,
3729 DefaultResultsList, TD))
3731 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) {
3732 PHINode *PHI = DefaultResultsList[I].first;
3733 Constant *Result = DefaultResultsList[I].second;
3734 DefaultResults[PHI] = Result;
3735 ResultTypes[PHI] = Result->getType();
3738 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
3739 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
3740 if (!ShouldBuildLookupTable(SI, TableSize, TTI, TD, ResultTypes))
3743 // Create the BB that does the lookups.
3744 Module &Mod = *CommonDest->getParent()->getParent();
3745 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
3747 CommonDest->getParent(),
3750 // Compute the table index value.
3751 Builder.SetInsertPoint(SI);
3752 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
3755 // Compute the maximum table size representable by the integer type we are
3757 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
3758 uint64_t MaxTableSize = CaseSize > 63? UINT64_MAX : 1ULL << CaseSize;
3759 assert(MaxTableSize >= TableSize &&
3760 "It is impossible for a switch to have more entries than the max "
3761 "representable value of its input integer type's size.");
3763 // If we have a fully covered lookup table, unconditionally branch to the
3764 // lookup table BB. Otherwise, check if the condition value is within the case
3765 // range. If it is so, branch to the new BB. Otherwise branch to SI's default
3767 const bool GeneratingCoveredLookupTable = MaxTableSize == TableSize;
3768 if (GeneratingCoveredLookupTable) {
3769 Builder.CreateBr(LookupBB);
3770 SI->getDefaultDest()->removePredecessor(SI->getParent());
3772 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
3773 MinCaseVal->getType(), TableSize));
3774 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
3777 // Populate the BB that does the lookups.
3778 Builder.SetInsertPoint(LookupBB);
3779 bool ReturnedEarly = false;
3780 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
3781 PHINode *PHI = PHIs[I];
3783 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI],
3784 DefaultResults[PHI], TD);
3786 Value *Result = Table.BuildLookup(TableIndex, Builder);
3788 // If the result is used to return immediately from the function, we want to
3789 // do that right here.
3790 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin()) &&
3791 *PHI->use_begin() == CommonDest->getFirstNonPHIOrDbg()) {
3792 Builder.CreateRet(Result);
3793 ReturnedEarly = true;
3797 PHI->addIncoming(Result, LookupBB);
3801 Builder.CreateBr(CommonDest);
3803 // Remove the switch.
3804 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
3805 BasicBlock *Succ = SI->getSuccessor(i);
3807 if (Succ == SI->getDefaultDest())
3809 Succ->removePredecessor(SI->getParent());
3811 SI->eraseFromParent();
3817 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
3818 BasicBlock *BB = SI->getParent();
3820 if (isValueEqualityComparison(SI)) {
3821 // If we only have one predecessor, and if it is a branch on this value,
3822 // see if that predecessor totally determines the outcome of this switch.
3823 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3824 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
3825 return SimplifyCFG(BB, TTI, TD) | true;
3827 Value *Cond = SI->getCondition();
3828 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
3829 if (SimplifySwitchOnSelect(SI, Select))
3830 return SimplifyCFG(BB, TTI, TD) | true;
3832 // If the block only contains the switch, see if we can fold the block
3833 // away into any preds.
3834 BasicBlock::iterator BBI = BB->begin();
3835 // Ignore dbg intrinsics.
3836 while (isa<DbgInfoIntrinsic>(BBI))
3839 if (FoldValueComparisonIntoPredecessors(SI, Builder))
3840 return SimplifyCFG(BB, TTI, TD) | true;
3843 // Try to transform the switch into an icmp and a branch.
3844 if (TurnSwitchRangeIntoICmp(SI, Builder))
3845 return SimplifyCFG(BB, TTI, TD) | true;
3847 // Remove unreachable cases.
3848 if (EliminateDeadSwitchCases(SI))
3849 return SimplifyCFG(BB, TTI, TD) | true;
3851 if (ForwardSwitchConditionToPHI(SI))
3852 return SimplifyCFG(BB, TTI, TD) | true;
3854 if (SwitchToLookupTable(SI, Builder, TTI, TD))
3855 return SimplifyCFG(BB, TTI, TD) | true;
3860 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
3861 BasicBlock *BB = IBI->getParent();
3862 bool Changed = false;
3864 // Eliminate redundant destinations.
3865 SmallPtrSet<Value *, 8> Succs;
3866 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
3867 BasicBlock *Dest = IBI->getDestination(i);
3868 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) {
3869 Dest->removePredecessor(BB);
3870 IBI->removeDestination(i);
3876 if (IBI->getNumDestinations() == 0) {
3877 // If the indirectbr has no successors, change it to unreachable.
3878 new UnreachableInst(IBI->getContext(), IBI);
3879 EraseTerminatorInstAndDCECond(IBI);
3883 if (IBI->getNumDestinations() == 1) {
3884 // If the indirectbr has one successor, change it to a direct branch.
3885 BranchInst::Create(IBI->getDestination(0), IBI);
3886 EraseTerminatorInstAndDCECond(IBI);
3890 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
3891 if (SimplifyIndirectBrOnSelect(IBI, SI))
3892 return SimplifyCFG(BB, TTI, TD) | true;
3897 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
3898 BasicBlock *BB = BI->getParent();
3900 if (SinkCommon && SinkThenElseCodeToEnd(BI))
3903 // If the Terminator is the only non-phi instruction, simplify the block.
3904 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime();
3905 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
3906 TryToSimplifyUncondBranchFromEmptyBlock(BB))
3909 // If the only instruction in the block is a seteq/setne comparison
3910 // against a constant, try to simplify the block.
3911 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
3912 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
3913 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
3915 if (I->isTerminator() &&
3916 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI, TD))
3920 // If this basic block is ONLY a compare and a branch, and if a predecessor
3921 // branches to us and our successor, fold the comparison into the
3922 // predecessor and use logical operations to update the incoming value
3923 // for PHI nodes in common successor.
3924 if (FoldBranchToCommonDest(BI))
3925 return SimplifyCFG(BB, TTI, TD) | true;
3930 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
3931 BasicBlock *BB = BI->getParent();
3933 // Conditional branch
3934 if (isValueEqualityComparison(BI)) {
3935 // If we only have one predecessor, and if it is a branch on this value,
3936 // see if that predecessor totally determines the outcome of this
3938 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3939 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
3940 return SimplifyCFG(BB, TTI, TD) | true;
3942 // This block must be empty, except for the setcond inst, if it exists.
3943 // Ignore dbg intrinsics.
3944 BasicBlock::iterator I = BB->begin();
3945 // Ignore dbg intrinsics.
3946 while (isa<DbgInfoIntrinsic>(I))
3949 if (FoldValueComparisonIntoPredecessors(BI, Builder))
3950 return SimplifyCFG(BB, TTI, TD) | true;
3951 } else if (&*I == cast<Instruction>(BI->getCondition())){
3953 // Ignore dbg intrinsics.
3954 while (isa<DbgInfoIntrinsic>(I))
3956 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
3957 return SimplifyCFG(BB, TTI, TD) | true;
3961 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
3962 if (SimplifyBranchOnICmpChain(BI, TD, Builder))
3965 // If this basic block is ONLY a compare and a branch, and if a predecessor
3966 // branches to us and one of our successors, fold the comparison into the
3967 // predecessor and use logical operations to pick the right destination.
3968 if (FoldBranchToCommonDest(BI))
3969 return SimplifyCFG(BB, TTI, TD) | true;
3971 // We have a conditional branch to two blocks that are only reachable
3972 // from BI. We know that the condbr dominates the two blocks, so see if
3973 // there is any identical code in the "then" and "else" blocks. If so, we
3974 // can hoist it up to the branching block.
3975 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) {
3976 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3977 if (HoistThenElseCodeToIf(BI))
3978 return SimplifyCFG(BB, TTI, TD) | true;
3980 // If Successor #1 has multiple preds, we may be able to conditionally
3981 // execute Successor #0 if it branches to successor #1.
3982 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
3983 if (Succ0TI->getNumSuccessors() == 1 &&
3984 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
3985 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0)))
3986 return SimplifyCFG(BB, TTI, TD) | true;
3988 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3989 // If Successor #0 has multiple preds, we may be able to conditionally
3990 // execute Successor #1 if it branches to successor #0.
3991 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
3992 if (Succ1TI->getNumSuccessors() == 1 &&
3993 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
3994 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1)))
3995 return SimplifyCFG(BB, TTI, TD) | true;
3998 // If this is a branch on a phi node in the current block, thread control
3999 // through this block if any PHI node entries are constants.
4000 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
4001 if (PN->getParent() == BI->getParent())
4002 if (FoldCondBranchOnPHI(BI, TD))
4003 return SimplifyCFG(BB, TTI, TD) | true;
4005 // Scan predecessor blocks for conditional branches.
4006 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
4007 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
4008 if (PBI != BI && PBI->isConditional())
4009 if (SimplifyCondBranchToCondBranch(PBI, BI))
4010 return SimplifyCFG(BB, TTI, TD) | true;
4015 /// Check if passing a value to an instruction will cause undefined behavior.
4016 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
4017 Constant *C = dyn_cast<Constant>(V);
4024 if (C->isNullValue()) {
4025 // Only look at the first use, avoid hurting compile time with long uselists
4026 User *Use = *I->use_begin();
4028 // Now make sure that there are no instructions in between that can alter
4029 // control flow (eg. calls)
4030 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
4031 if (i == I->getParent()->end() || i->mayHaveSideEffects())
4034 // Look through GEPs. A load from a GEP derived from NULL is still undefined
4035 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
4036 if (GEP->getPointerOperand() == I)
4037 return passingValueIsAlwaysUndefined(V, GEP);
4039 // Look through bitcasts.
4040 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
4041 return passingValueIsAlwaysUndefined(V, BC);
4043 // Load from null is undefined.
4044 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
4045 if (!LI->isVolatile())
4046 return LI->getPointerAddressSpace() == 0;
4048 // Store to null is undefined.
4049 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
4050 if (!SI->isVolatile())
4051 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
4056 /// If BB has an incoming value that will always trigger undefined behavior
4057 /// (eg. null pointer dereference), remove the branch leading here.
4058 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
4059 for (BasicBlock::iterator i = BB->begin();
4060 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
4061 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
4062 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
4063 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
4064 IRBuilder<> Builder(T);
4065 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
4066 BB->removePredecessor(PHI->getIncomingBlock(i));
4067 // Turn uncoditional branches into unreachables and remove the dead
4068 // destination from conditional branches.
4069 if (BI->isUnconditional())
4070 Builder.CreateUnreachable();
4072 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
4073 BI->getSuccessor(0));
4074 BI->eraseFromParent();
4077 // TODO: SwitchInst.
4083 bool SimplifyCFGOpt::run(BasicBlock *BB) {
4084 bool Changed = false;
4086 assert(BB && BB->getParent() && "Block not embedded in function!");
4087 assert(BB->getTerminator() && "Degenerate basic block encountered!");
4089 // Remove basic blocks that have no predecessors (except the entry block)...
4090 // or that just have themself as a predecessor. These are unreachable.
4091 if ((pred_begin(BB) == pred_end(BB) &&
4092 BB != &BB->getParent()->getEntryBlock()) ||
4093 BB->getSinglePredecessor() == BB) {
4094 DEBUG(dbgs() << "Removing BB: \n" << *BB);
4095 DeleteDeadBlock(BB);
4099 // Check to see if we can constant propagate this terminator instruction
4101 Changed |= ConstantFoldTerminator(BB, true);
4103 // Check for and eliminate duplicate PHI nodes in this block.
4104 Changed |= EliminateDuplicatePHINodes(BB);
4106 // Check for and remove branches that will always cause undefined behavior.
4107 Changed |= removeUndefIntroducingPredecessor(BB);
4109 // Merge basic blocks into their predecessor if there is only one distinct
4110 // pred, and if there is only one distinct successor of the predecessor, and
4111 // if there are no PHI nodes.
4113 if (MergeBlockIntoPredecessor(BB))
4116 IRBuilder<> Builder(BB);
4118 // If there is a trivial two-entry PHI node in this basic block, and we can
4119 // eliminate it, do so now.
4120 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
4121 if (PN->getNumIncomingValues() == 2)
4122 Changed |= FoldTwoEntryPHINode(PN, TD);
4124 Builder.SetInsertPoint(BB->getTerminator());
4125 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
4126 if (BI->isUnconditional()) {
4127 if (SimplifyUncondBranch(BI, Builder)) return true;
4129 if (SimplifyCondBranch(BI, Builder)) return true;
4131 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
4132 if (SimplifyReturn(RI, Builder)) return true;
4133 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
4134 if (SimplifyResume(RI, Builder)) return true;
4135 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
4136 if (SimplifySwitch(SI, Builder)) return true;
4137 } else if (UnreachableInst *UI =
4138 dyn_cast<UnreachableInst>(BB->getTerminator())) {
4139 if (SimplifyUnreachable(UI)) return true;
4140 } else if (IndirectBrInst *IBI =
4141 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
4142 if (SimplifyIndirectBr(IBI)) return true;
4148 /// SimplifyCFG - This function is used to do simplification of a CFG. For
4149 /// example, it adjusts branches to branches to eliminate the extra hop, it
4150 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4151 /// of the CFG. It returns true if a modification was made.
4153 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
4154 const DataLayout *TD) {
4155 return SimplifyCFGOpt(TTI, TD).run(BB);