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/InstructionSimplify.h"
23 #include "llvm/Analysis/TargetTransformInfo.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/GlobalVariable.h"
29 #include "llvm/IR/IRBuilder.h"
30 #include "llvm/IR/Instructions.h"
31 #include "llvm/IR/IntrinsicInst.h"
32 #include "llvm/IR/LLVMContext.h"
33 #include "llvm/IR/MDBuilder.h"
34 #include "llvm/IR/Metadata.h"
35 #include "llvm/IR/Module.h"
36 #include "llvm/IR/Operator.h"
37 #include "llvm/IR/Type.h"
38 #include "llvm/Support/CFG.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Support/ConstantRange.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/NoFolder.h"
43 #include "llvm/Support/PatternMatch.h"
44 #include "llvm/Support/raw_ostream.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
50 using namespace PatternMatch;
52 static cl::opt<unsigned>
53 PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1),
54 cl::desc("Control the amount of phi node folding to perform (default = 1)"));
57 DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false),
58 cl::desc("Duplicate return instructions into unconditional branches"));
61 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
62 cl::desc("Sink common instructions down to the end block"));
65 HoistCondStores("simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
66 cl::desc("Hoist conditional stores if an unconditional store preceeds"));
68 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
69 STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables");
70 STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block");
71 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
74 /// ValueEqualityComparisonCase - Represents a case of a switch.
75 struct ValueEqualityComparisonCase {
79 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
80 : Value(Value), Dest(Dest) {}
82 bool operator<(ValueEqualityComparisonCase RHS) const {
83 // Comparing pointers is ok as we only rely on the order for uniquing.
84 return Value < RHS.Value;
87 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
90 class SimplifyCFGOpt {
91 const TargetTransformInfo &TTI;
92 const DataLayout *const TD;
93 Value *isValueEqualityComparison(TerminatorInst *TI);
94 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
95 std::vector<ValueEqualityComparisonCase> &Cases);
96 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
98 IRBuilder<> &Builder);
99 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
100 IRBuilder<> &Builder);
102 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
103 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
104 bool SimplifyUnreachable(UnreachableInst *UI);
105 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
106 bool SimplifyIndirectBr(IndirectBrInst *IBI);
107 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder);
108 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder);
111 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout *TD)
112 : TTI(TTI), TD(TD) {}
113 bool run(BasicBlock *BB);
117 /// SafeToMergeTerminators - Return true if it is safe to merge these two
118 /// terminator instructions together.
120 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
121 if (SI1 == SI2) return false; // Can't merge with self!
123 // It is not safe to merge these two switch instructions if they have a common
124 // successor, and if that successor has a PHI node, and if *that* PHI node has
125 // conflicting incoming values from the two switch blocks.
126 BasicBlock *SI1BB = SI1->getParent();
127 BasicBlock *SI2BB = SI2->getParent();
128 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
130 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
131 if (SI1Succs.count(*I))
132 for (BasicBlock::iterator BBI = (*I)->begin();
133 isa<PHINode>(BBI); ++BBI) {
134 PHINode *PN = cast<PHINode>(BBI);
135 if (PN->getIncomingValueForBlock(SI1BB) !=
136 PN->getIncomingValueForBlock(SI2BB))
143 /// isProfitableToFoldUnconditional - Return true if it is safe and profitable
144 /// to merge these two terminator instructions together, where SI1 is an
145 /// unconditional branch. PhiNodes will store all PHI nodes in common
148 static bool isProfitableToFoldUnconditional(BranchInst *SI1,
151 SmallVectorImpl<PHINode*> &PhiNodes) {
152 if (SI1 == SI2) return false; // Can't merge with self!
153 assert(SI1->isUnconditional() && SI2->isConditional());
155 // We fold the unconditional branch if we can easily update all PHI nodes in
156 // common successors:
157 // 1> We have a constant incoming value for the conditional branch;
158 // 2> We have "Cond" as the incoming value for the unconditional branch;
159 // 3> SI2->getCondition() and Cond have same operands.
160 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
161 if (!Ci2) return false;
162 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
163 Cond->getOperand(1) == Ci2->getOperand(1)) &&
164 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
165 Cond->getOperand(1) == Ci2->getOperand(0)))
168 BasicBlock *SI1BB = SI1->getParent();
169 BasicBlock *SI2BB = SI2->getParent();
170 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
171 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
172 if (SI1Succs.count(*I))
173 for (BasicBlock::iterator BBI = (*I)->begin();
174 isa<PHINode>(BBI); ++BBI) {
175 PHINode *PN = cast<PHINode>(BBI);
176 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
177 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
179 PhiNodes.push_back(PN);
184 /// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will
185 /// now be entries in it from the 'NewPred' block. The values that will be
186 /// flowing into the PHI nodes will be the same as those coming in from
187 /// ExistPred, an existing predecessor of Succ.
188 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
189 BasicBlock *ExistPred) {
190 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
193 for (BasicBlock::iterator I = Succ->begin();
194 (PN = dyn_cast<PHINode>(I)); ++I)
195 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
198 /// ComputeSpeculationCost - Compute an abstract "cost" of speculating the
199 /// given instruction, which is assumed to be safe to speculate. 1 means
200 /// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive.
201 static unsigned ComputeSpeculationCost(const User *I) {
202 assert(isSafeToSpeculativelyExecute(I) &&
203 "Instruction is not safe to speculatively execute!");
204 switch (Operator::getOpcode(I)) {
206 // In doubt, be conservative.
208 case Instruction::GetElementPtr:
209 // GEPs are cheap if all indices are constant.
210 if (!cast<GEPOperator>(I)->hasAllConstantIndices())
213 case Instruction::Load:
214 case Instruction::Add:
215 case Instruction::Sub:
216 case Instruction::And:
217 case Instruction::Or:
218 case Instruction::Xor:
219 case Instruction::Shl:
220 case Instruction::LShr:
221 case Instruction::AShr:
222 case Instruction::ICmp:
223 case Instruction::Trunc:
224 case Instruction::ZExt:
225 case Instruction::SExt:
226 return 1; // These are all cheap.
228 case Instruction::Call:
229 case Instruction::Select:
234 /// DominatesMergePoint - If we have a merge point of an "if condition" as
235 /// accepted above, return true if the specified value dominates the block. We
236 /// don't handle the true generality of domination here, just a special case
237 /// which works well enough for us.
239 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
240 /// see if V (which must be an instruction) and its recursive operands
241 /// that do not dominate BB have a combined cost lower than CostRemaining and
242 /// are non-trapping. If both are true, the instruction is inserted into the
243 /// set and true is returned.
245 /// The cost for most non-trapping instructions is defined as 1 except for
246 /// Select whose cost is 2.
248 /// After this function returns, CostRemaining is decreased by the cost of
249 /// V plus its non-dominating operands. If that cost is greater than
250 /// CostRemaining, false is returned and CostRemaining is undefined.
251 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
252 SmallPtrSet<Instruction*, 4> *AggressiveInsts,
253 unsigned &CostRemaining) {
254 Instruction *I = dyn_cast<Instruction>(V);
256 // Non-instructions all dominate instructions, but not all constantexprs
257 // can be executed unconditionally.
258 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
263 BasicBlock *PBB = I->getParent();
265 // We don't want to allow weird loops that might have the "if condition" in
266 // the bottom of this block.
267 if (PBB == BB) return false;
269 // If this instruction is defined in a block that contains an unconditional
270 // branch to BB, then it must be in the 'conditional' part of the "if
271 // statement". If not, it definitely dominates the region.
272 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
273 if (BI == 0 || BI->isConditional() || BI->getSuccessor(0) != BB)
276 // If we aren't allowing aggressive promotion anymore, then don't consider
277 // instructions in the 'if region'.
278 if (AggressiveInsts == 0) return false;
280 // If we have seen this instruction before, don't count it again.
281 if (AggressiveInsts->count(I)) return true;
283 // Okay, it looks like the instruction IS in the "condition". Check to
284 // see if it's a cheap instruction to unconditionally compute, and if it
285 // only uses stuff defined outside of the condition. If so, hoist it out.
286 if (!isSafeToSpeculativelyExecute(I))
289 unsigned Cost = ComputeSpeculationCost(I);
291 if (Cost > CostRemaining)
294 CostRemaining -= Cost;
296 // Okay, we can only really hoist these out if their operands do
297 // not take us over the cost threshold.
298 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
299 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining))
301 // Okay, it's safe to do this! Remember this instruction.
302 AggressiveInsts->insert(I);
306 /// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr
307 /// and PointerNullValue. Return NULL if value is not a constant int.
308 static ConstantInt *GetConstantInt(Value *V, const DataLayout *TD) {
309 // Normal constant int.
310 ConstantInt *CI = dyn_cast<ConstantInt>(V);
311 if (CI || !TD || !isa<Constant>(V) || !V->getType()->isPointerTy())
314 // This is some kind of pointer constant. Turn it into a pointer-sized
315 // ConstantInt if possible.
316 IntegerType *PtrTy = cast<IntegerType>(TD->getIntPtrType(V->getType()));
318 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
319 if (isa<ConstantPointerNull>(V))
320 return ConstantInt::get(PtrTy, 0);
322 // IntToPtr const int.
323 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
324 if (CE->getOpcode() == Instruction::IntToPtr)
325 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
326 // The constant is very likely to have the right type already.
327 if (CI->getType() == PtrTy)
330 return cast<ConstantInt>
331 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
336 /// GatherConstantCompares - Given a potentially 'or'd or 'and'd together
337 /// collection of icmp eq/ne instructions that compare a value against a
338 /// constant, return the value being compared, and stick the constant into the
341 GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra,
342 const DataLayout *TD, bool isEQ, unsigned &UsedICmps) {
343 Instruction *I = dyn_cast<Instruction>(V);
344 if (I == 0) return 0;
346 // If this is an icmp against a constant, handle this as one of the cases.
347 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) {
348 if (ConstantInt *C = GetConstantInt(I->getOperand(1), TD)) {
352 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
353 // (x & ~2^x) == y --> x == y || x == y|2^x
354 // This undoes a transformation done by instcombine to fuse 2 compares.
355 if (match(ICI->getOperand(0),
356 m_And(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
357 APInt Not = ~RHSC->getValue();
358 if (Not.isPowerOf2()) {
361 ConstantInt::get(C->getContext(), C->getValue() | Not));
369 return I->getOperand(0);
372 // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to
375 ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue());
377 // Shift the range if the compare is fed by an add. This is the range
378 // compare idiom as emitted by instcombine.
380 match(I->getOperand(0), m_Add(m_Value(RHSVal), m_ConstantInt(RHSC)));
382 Span = Span.subtract(RHSC->getValue());
384 // If this is an and/!= check then we want to optimize "x ugt 2" into
387 Span = Span.inverse();
389 // If there are a ton of values, we don't want to make a ginormous switch.
390 if (Span.getSetSize().ugt(8) || Span.isEmptySet())
393 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
394 Vals.push_back(ConstantInt::get(V->getContext(), Tmp));
396 return hasAdd ? RHSVal : I->getOperand(0);
401 // Otherwise, we can only handle an | or &, depending on isEQ.
402 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And))
405 unsigned NumValsBeforeLHS = Vals.size();
406 unsigned UsedICmpsBeforeLHS = UsedICmps;
407 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, TD,
409 unsigned NumVals = Vals.size();
410 unsigned UsedICmpsBeforeRHS = UsedICmps;
411 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD,
415 Vals.resize(NumVals);
416 UsedICmps = UsedICmpsBeforeRHS;
419 // The RHS of the or/and can't be folded in and we haven't used "Extra" yet,
420 // set it and return success.
421 if (Extra == 0 || Extra == I->getOperand(1)) {
422 Extra = I->getOperand(1);
426 Vals.resize(NumValsBeforeLHS);
427 UsedICmps = UsedICmpsBeforeLHS;
431 // If the LHS can't be folded in, but Extra is available and RHS can, try to
433 if (Extra == 0 || Extra == I->getOperand(0)) {
434 Value *OldExtra = Extra;
435 Extra = I->getOperand(0);
436 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD,
439 assert(Vals.size() == NumValsBeforeLHS);
446 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
447 Instruction *Cond = 0;
448 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
449 Cond = dyn_cast<Instruction>(SI->getCondition());
450 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
451 if (BI->isConditional())
452 Cond = dyn_cast<Instruction>(BI->getCondition());
453 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
454 Cond = dyn_cast<Instruction>(IBI->getAddress());
457 TI->eraseFromParent();
458 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
461 /// isValueEqualityComparison - Return true if the specified terminator checks
462 /// to see if a value is equal to constant integer value.
463 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
465 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
466 // Do not permit merging of large switch instructions into their
467 // predecessors unless there is only one predecessor.
468 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()),
469 pred_end(SI->getParent())) <= 128)
470 CV = SI->getCondition();
471 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
472 if (BI->isConditional() && BI->getCondition()->hasOneUse())
473 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition()))
474 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), TD))
475 CV = ICI->getOperand(0);
477 // Unwrap any lossless ptrtoint cast.
479 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
480 Value *Ptr = PTII->getPointerOperand();
481 if (PTII->getType() == TD->getIntPtrType(Ptr->getType()))
488 /// GetValueEqualityComparisonCases - Given a value comparison instruction,
489 /// decode all of the 'cases' that it represents and return the 'default' block.
490 BasicBlock *SimplifyCFGOpt::
491 GetValueEqualityComparisonCases(TerminatorInst *TI,
492 std::vector<ValueEqualityComparisonCase>
494 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
495 Cases.reserve(SI->getNumCases());
496 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
497 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
498 i.getCaseSuccessor()));
499 return SI->getDefaultDest();
502 BranchInst *BI = cast<BranchInst>(TI);
503 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
504 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
505 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
508 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
512 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries
513 /// in the list that match the specified block.
514 static void EliminateBlockCases(BasicBlock *BB,
515 std::vector<ValueEqualityComparisonCase> &Cases) {
516 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
519 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as
522 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
523 std::vector<ValueEqualityComparisonCase > &C2) {
524 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
526 // Make V1 be smaller than V2.
527 if (V1->size() > V2->size())
530 if (V1->size() == 0) return false;
531 if (V1->size() == 1) {
533 ConstantInt *TheVal = (*V1)[0].Value;
534 for (unsigned i = 0, e = V2->size(); i != e; ++i)
535 if (TheVal == (*V2)[i].Value)
539 // Otherwise, just sort both lists and compare element by element.
540 array_pod_sort(V1->begin(), V1->end());
541 array_pod_sort(V2->begin(), V2->end());
542 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
543 while (i1 != e1 && i2 != e2) {
544 if ((*V1)[i1].Value == (*V2)[i2].Value)
546 if ((*V1)[i1].Value < (*V2)[i2].Value)
554 /// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a
555 /// terminator instruction and its block is known to only have a single
556 /// predecessor block, check to see if that predecessor is also a value
557 /// comparison with the same value, and if that comparison determines the
558 /// outcome of this comparison. If so, simplify TI. This does a very limited
559 /// form of jump threading.
560 bool SimplifyCFGOpt::
561 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
563 IRBuilder<> &Builder) {
564 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
565 if (!PredVal) return false; // Not a value comparison in predecessor.
567 Value *ThisVal = isValueEqualityComparison(TI);
568 assert(ThisVal && "This isn't a value comparison!!");
569 if (ThisVal != PredVal) return false; // Different predicates.
571 // TODO: Preserve branch weight metadata, similarly to how
572 // FoldValueComparisonIntoPredecessors preserves it.
574 // Find out information about when control will move from Pred to TI's block.
575 std::vector<ValueEqualityComparisonCase> PredCases;
576 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
578 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
580 // Find information about how control leaves this block.
581 std::vector<ValueEqualityComparisonCase> ThisCases;
582 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
583 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
585 // If TI's block is the default block from Pred's comparison, potentially
586 // simplify TI based on this knowledge.
587 if (PredDef == TI->getParent()) {
588 // If we are here, we know that the value is none of those cases listed in
589 // PredCases. If there are any cases in ThisCases that are in PredCases, we
591 if (!ValuesOverlap(PredCases, ThisCases))
594 if (isa<BranchInst>(TI)) {
595 // Okay, one of the successors of this condbr is dead. Convert it to a
597 assert(ThisCases.size() == 1 && "Branch can only have one case!");
598 // Insert the new branch.
599 Instruction *NI = Builder.CreateBr(ThisDef);
602 // Remove PHI node entries for the dead edge.
603 ThisCases[0].Dest->removePredecessor(TI->getParent());
605 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
606 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
608 EraseTerminatorInstAndDCECond(TI);
612 SwitchInst *SI = cast<SwitchInst>(TI);
613 // Okay, TI has cases that are statically dead, prune them away.
614 SmallPtrSet<Constant*, 16> DeadCases;
615 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
616 DeadCases.insert(PredCases[i].Value);
618 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
619 << "Through successor TI: " << *TI);
621 // Collect branch weights into a vector.
622 SmallVector<uint32_t, 8> Weights;
623 MDNode* MD = SI->getMetadata(LLVMContext::MD_prof);
624 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
626 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
628 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i));
630 Weights.push_back(CI->getValue().getZExtValue());
632 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
634 if (DeadCases.count(i.getCaseValue())) {
636 std::swap(Weights[i.getCaseIndex()+1], Weights.back());
639 i.getCaseSuccessor()->removePredecessor(TI->getParent());
643 if (HasWeight && Weights.size() >= 2)
644 SI->setMetadata(LLVMContext::MD_prof,
645 MDBuilder(SI->getParent()->getContext()).
646 createBranchWeights(Weights));
648 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
652 // Otherwise, TI's block must correspond to some matched value. Find out
653 // which value (or set of values) this is.
654 ConstantInt *TIV = 0;
655 BasicBlock *TIBB = TI->getParent();
656 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
657 if (PredCases[i].Dest == TIBB) {
659 return false; // Cannot handle multiple values coming to this block.
660 TIV = PredCases[i].Value;
662 assert(TIV && "No edge from pred to succ?");
664 // Okay, we found the one constant that our value can be if we get into TI's
665 // BB. Find out which successor will unconditionally be branched to.
666 BasicBlock *TheRealDest = 0;
667 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
668 if (ThisCases[i].Value == TIV) {
669 TheRealDest = ThisCases[i].Dest;
673 // If not handled by any explicit cases, it is handled by the default case.
674 if (TheRealDest == 0) TheRealDest = ThisDef;
676 // Remove PHI node entries for dead edges.
677 BasicBlock *CheckEdge = TheRealDest;
678 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
679 if (*SI != CheckEdge)
680 (*SI)->removePredecessor(TIBB);
684 // Insert the new branch.
685 Instruction *NI = Builder.CreateBr(TheRealDest);
688 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
689 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
691 EraseTerminatorInstAndDCECond(TI);
696 /// ConstantIntOrdering - This class implements a stable ordering of constant
697 /// integers that does not depend on their address. This is important for
698 /// applications that sort ConstantInt's to ensure uniqueness.
699 struct ConstantIntOrdering {
700 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
701 return LHS->getValue().ult(RHS->getValue());
706 static int ConstantIntSortPredicate(ConstantInt *const *P1,
707 ConstantInt *const *P2) {
708 const ConstantInt *LHS = *P1;
709 const ConstantInt *RHS = *P2;
710 if (LHS->getValue().ult(RHS->getValue()))
712 if (LHS->getValue() == RHS->getValue())
717 static inline bool HasBranchWeights(const Instruction* I) {
718 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof);
719 if (ProfMD && ProfMD->getOperand(0))
720 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
721 return MDS->getString().equals("branch_weights");
726 /// Get Weights of a given TerminatorInst, the default weight is at the front
727 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
729 static void GetBranchWeights(TerminatorInst *TI,
730 SmallVectorImpl<uint64_t> &Weights) {
731 MDNode* MD = TI->getMetadata(LLVMContext::MD_prof);
733 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
734 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(i));
736 Weights.push_back(CI->getValue().getZExtValue());
739 // If TI is a conditional eq, the default case is the false case,
740 // and the corresponding branch-weight data is at index 2. We swap the
741 // default weight to be the first entry.
742 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
743 assert(Weights.size() == 2);
744 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
745 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
746 std::swap(Weights.front(), Weights.back());
750 /// Sees if any of the weights are too big for a uint32_t, and halves all the
751 /// weights if any are.
752 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
754 for (unsigned i = 0; i < Weights.size(); ++i)
755 if (Weights[i] > UINT_MAX) {
763 for (unsigned i = 0; i < Weights.size(); ++i)
767 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value
768 /// equality comparison instruction (either a switch or a branch on "X == c").
769 /// See if any of the predecessors of the terminator block are value comparisons
770 /// on the same value. If so, and if safe to do so, fold them together.
771 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
772 IRBuilder<> &Builder) {
773 BasicBlock *BB = TI->getParent();
774 Value *CV = isValueEqualityComparison(TI); // CondVal
775 assert(CV && "Not a comparison?");
776 bool Changed = false;
778 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
779 while (!Preds.empty()) {
780 BasicBlock *Pred = Preds.pop_back_val();
782 // See if the predecessor is a comparison with the same value.
783 TerminatorInst *PTI = Pred->getTerminator();
784 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
786 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
787 // Figure out which 'cases' to copy from SI to PSI.
788 std::vector<ValueEqualityComparisonCase> BBCases;
789 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
791 std::vector<ValueEqualityComparisonCase> PredCases;
792 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
794 // Based on whether the default edge from PTI goes to BB or not, fill in
795 // PredCases and PredDefault with the new switch cases we would like to
797 SmallVector<BasicBlock*, 8> NewSuccessors;
799 // Update the branch weight metadata along the way
800 SmallVector<uint64_t, 8> Weights;
801 bool PredHasWeights = HasBranchWeights(PTI);
802 bool SuccHasWeights = HasBranchWeights(TI);
804 if (PredHasWeights) {
805 GetBranchWeights(PTI, Weights);
806 // branch-weight metadata is inconsistent here.
807 if (Weights.size() != 1 + PredCases.size())
808 PredHasWeights = SuccHasWeights = false;
809 } else if (SuccHasWeights)
810 // If there are no predecessor weights but there are successor weights,
811 // populate Weights with 1, which will later be scaled to the sum of
812 // successor's weights
813 Weights.assign(1 + PredCases.size(), 1);
815 SmallVector<uint64_t, 8> SuccWeights;
816 if (SuccHasWeights) {
817 GetBranchWeights(TI, SuccWeights);
818 // branch-weight metadata is inconsistent here.
819 if (SuccWeights.size() != 1 + BBCases.size())
820 PredHasWeights = SuccHasWeights = false;
821 } else if (PredHasWeights)
822 SuccWeights.assign(1 + BBCases.size(), 1);
824 if (PredDefault == BB) {
825 // If this is the default destination from PTI, only the edges in TI
826 // that don't occur in PTI, or that branch to BB will be activated.
827 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
828 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
829 if (PredCases[i].Dest != BB)
830 PTIHandled.insert(PredCases[i].Value);
832 // The default destination is BB, we don't need explicit targets.
833 std::swap(PredCases[i], PredCases.back());
835 if (PredHasWeights || SuccHasWeights) {
836 // Increase weight for the default case.
837 Weights[0] += Weights[i+1];
838 std::swap(Weights[i+1], Weights.back());
842 PredCases.pop_back();
846 // Reconstruct the new switch statement we will be building.
847 if (PredDefault != BBDefault) {
848 PredDefault->removePredecessor(Pred);
849 PredDefault = BBDefault;
850 NewSuccessors.push_back(BBDefault);
853 unsigned CasesFromPred = Weights.size();
854 uint64_t ValidTotalSuccWeight = 0;
855 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
856 if (!PTIHandled.count(BBCases[i].Value) &&
857 BBCases[i].Dest != BBDefault) {
858 PredCases.push_back(BBCases[i]);
859 NewSuccessors.push_back(BBCases[i].Dest);
860 if (SuccHasWeights || PredHasWeights) {
861 // The default weight is at index 0, so weight for the ith case
862 // should be at index i+1. Scale the cases from successor by
863 // PredDefaultWeight (Weights[0]).
864 Weights.push_back(Weights[0] * SuccWeights[i+1]);
865 ValidTotalSuccWeight += SuccWeights[i+1];
869 if (SuccHasWeights || PredHasWeights) {
870 ValidTotalSuccWeight += SuccWeights[0];
871 // Scale the cases from predecessor by ValidTotalSuccWeight.
872 for (unsigned i = 1; i < CasesFromPred; ++i)
873 Weights[i] *= ValidTotalSuccWeight;
874 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
875 Weights[0] *= SuccWeights[0];
878 // If this is not the default destination from PSI, only the edges
879 // in SI that occur in PSI with a destination of BB will be
881 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
882 std::map<ConstantInt*, uint64_t> WeightsForHandled;
883 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
884 if (PredCases[i].Dest == BB) {
885 PTIHandled.insert(PredCases[i].Value);
887 if (PredHasWeights || SuccHasWeights) {
888 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
889 std::swap(Weights[i+1], Weights.back());
893 std::swap(PredCases[i], PredCases.back());
894 PredCases.pop_back();
898 // Okay, now we know which constants were sent to BB from the
899 // predecessor. Figure out where they will all go now.
900 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
901 if (PTIHandled.count(BBCases[i].Value)) {
902 // If this is one we are capable of getting...
903 if (PredHasWeights || SuccHasWeights)
904 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
905 PredCases.push_back(BBCases[i]);
906 NewSuccessors.push_back(BBCases[i].Dest);
907 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
910 // If there are any constants vectored to BB that TI doesn't handle,
911 // they must go to the default destination of TI.
912 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
914 E = PTIHandled.end(); I != E; ++I) {
915 if (PredHasWeights || SuccHasWeights)
916 Weights.push_back(WeightsForHandled[*I]);
917 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
918 NewSuccessors.push_back(BBDefault);
922 // Okay, at this point, we know which new successor Pred will get. Make
923 // sure we update the number of entries in the PHI nodes for these
925 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i)
926 AddPredecessorToBlock(NewSuccessors[i], Pred, BB);
928 Builder.SetInsertPoint(PTI);
929 // Convert pointer to int before we switch.
930 if (CV->getType()->isPointerTy()) {
931 assert(TD && "Cannot switch on pointer without DataLayout");
932 CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getType()),
936 // Now that the successors are updated, create the new Switch instruction.
937 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
939 NewSI->setDebugLoc(PTI->getDebugLoc());
940 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
941 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest);
943 if (PredHasWeights || SuccHasWeights) {
944 // Halve the weights if any of them cannot fit in an uint32_t
947 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
949 NewSI->setMetadata(LLVMContext::MD_prof,
950 MDBuilder(BB->getContext()).
951 createBranchWeights(MDWeights));
954 EraseTerminatorInstAndDCECond(PTI);
956 // Okay, last check. If BB is still a successor of PSI, then we must
957 // have an infinite loop case. If so, add an infinitely looping block
958 // to handle the case to preserve the behavior of the code.
959 BasicBlock *InfLoopBlock = 0;
960 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
961 if (NewSI->getSuccessor(i) == BB) {
962 if (InfLoopBlock == 0) {
963 // Insert it at the end of the function, because it's either code,
964 // or it won't matter if it's hot. :)
965 InfLoopBlock = BasicBlock::Create(BB->getContext(),
966 "infloop", BB->getParent());
967 BranchInst::Create(InfLoopBlock, InfLoopBlock);
969 NewSI->setSuccessor(i, InfLoopBlock);
978 // isSafeToHoistInvoke - If we would need to insert a select that uses the
979 // value of this invoke (comments in HoistThenElseCodeToIf explain why we
980 // would need to do this), we can't hoist the invoke, as there is nowhere
981 // to put the select in this case.
982 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
983 Instruction *I1, Instruction *I2) {
984 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
986 for (BasicBlock::iterator BBI = SI->begin();
987 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
988 Value *BB1V = PN->getIncomingValueForBlock(BB1);
989 Value *BB2V = PN->getIncomingValueForBlock(BB2);
990 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
998 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
999 /// BB2, hoist any common code in the two blocks up into the branch block. The
1000 /// caller of this function guarantees that BI's block dominates BB1 and BB2.
1001 static bool HoistThenElseCodeToIf(BranchInst *BI) {
1002 // This does very trivial matching, with limited scanning, to find identical
1003 // instructions in the two blocks. In particular, we don't want to get into
1004 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1005 // such, we currently just scan for obviously identical instructions in an
1007 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1008 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1010 BasicBlock::iterator BB1_Itr = BB1->begin();
1011 BasicBlock::iterator BB2_Itr = BB2->begin();
1013 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++;
1014 // Skip debug info if it is not identical.
1015 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1016 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1017 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1018 while (isa<DbgInfoIntrinsic>(I1))
1020 while (isa<DbgInfoIntrinsic>(I2))
1023 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1024 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1027 BasicBlock *BIParent = BI->getParent();
1029 bool Changed = false;
1031 // If we are hoisting the terminator instruction, don't move one (making a
1032 // broken BB), instead clone it, and remove BI.
1033 if (isa<TerminatorInst>(I1))
1034 goto HoistTerminator;
1036 // For a normal instruction, we just move one to right before the branch,
1037 // then replace all uses of the other with the first. Finally, we remove
1038 // the now redundant second instruction.
1039 BIParent->getInstList().splice(BI, BB1->getInstList(), I1);
1040 if (!I2->use_empty())
1041 I2->replaceAllUsesWith(I1);
1042 I1->intersectOptionalDataWith(I2);
1043 I2->eraseFromParent();
1048 // Skip debug info if it is not identical.
1049 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1050 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1051 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1052 while (isa<DbgInfoIntrinsic>(I1))
1054 while (isa<DbgInfoIntrinsic>(I2))
1057 } while (I1->isIdenticalToWhenDefined(I2));
1062 // It may not be possible to hoist an invoke.
1063 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1066 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1068 for (BasicBlock::iterator BBI = SI->begin();
1069 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1070 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1071 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1075 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1077 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1082 // Okay, it is safe to hoist the terminator.
1083 Instruction *NT = I1->clone();
1084 BIParent->getInstList().insert(BI, NT);
1085 if (!NT->getType()->isVoidTy()) {
1086 I1->replaceAllUsesWith(NT);
1087 I2->replaceAllUsesWith(NT);
1091 IRBuilder<true, NoFolder> Builder(NT);
1092 // Hoisting one of the terminators from our successor is a great thing.
1093 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1094 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1095 // nodes, so we insert select instruction to compute the final result.
1096 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1097 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1099 for (BasicBlock::iterator BBI = SI->begin();
1100 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1101 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1102 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1103 if (BB1V == BB2V) continue;
1105 // These values do not agree. Insert a select instruction before NT
1106 // that determines the right value.
1107 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1109 SI = cast<SelectInst>
1110 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1111 BB1V->getName()+"."+BB2V->getName()));
1113 // Make the PHI node use the select for all incoming values for BB1/BB2
1114 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1115 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1116 PN->setIncomingValue(i, SI);
1120 // Update any PHI nodes in our new successors.
1121 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1122 AddPredecessorToBlock(*SI, BIParent, BB1);
1124 EraseTerminatorInstAndDCECond(BI);
1128 /// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd,
1129 /// check whether BBEnd has only two predecessors and the other predecessor
1130 /// ends with an unconditional branch. If it is true, sink any common code
1131 /// in the two predecessors to BBEnd.
1132 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1133 assert(BI1->isUnconditional());
1134 BasicBlock *BB1 = BI1->getParent();
1135 BasicBlock *BBEnd = BI1->getSuccessor(0);
1137 // Check that BBEnd has two predecessors and the other predecessor ends with
1138 // an unconditional branch.
1139 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1140 BasicBlock *Pred0 = *PI++;
1141 if (PI == PE) // Only one predecessor.
1143 BasicBlock *Pred1 = *PI++;
1144 if (PI != PE) // More than two predecessors.
1146 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1147 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1148 if (!BI2 || !BI2->isUnconditional())
1151 // Gather the PHI nodes in BBEnd.
1152 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2;
1153 Instruction *FirstNonPhiInBBEnd = 0;
1154 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end();
1156 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1157 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1158 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1159 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN);
1161 FirstNonPhiInBBEnd = &*I;
1165 if (!FirstNonPhiInBBEnd)
1169 // This does very trivial matching, with limited scanning, to find identical
1170 // instructions in the two blocks. We scan backward for obviously identical
1171 // instructions in an identical order.
1172 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1173 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(),
1174 RE2 = BB2->getInstList().rend();
1176 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1179 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1182 // Skip the unconditional branches.
1186 bool Changed = false;
1187 while (RI1 != RE1 && RI2 != RE2) {
1189 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1192 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1196 Instruction *I1 = &*RI1, *I2 = &*RI2;
1197 // I1 and I2 should have a single use in the same PHI node, and they
1198 // perform the same operation.
1199 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1200 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1201 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1202 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) ||
1203 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1204 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1205 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1206 !I1->hasOneUse() || !I2->hasOneUse() ||
1207 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() ||
1208 MapValueFromBB1ToBB2[I1].first != I2)
1211 // Check whether we should swap the operands of ICmpInst.
1212 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1213 bool SwapOpnds = false;
1214 if (ICmp1 && ICmp2 &&
1215 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1216 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1217 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1218 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1219 ICmp2->swapOperands();
1222 if (!I1->isSameOperationAs(I2)) {
1224 ICmp2->swapOperands();
1228 // The operands should be either the same or they need to be generated
1229 // with a PHI node after sinking. We only handle the case where there is
1230 // a single pair of different operands.
1231 Value *DifferentOp1 = 0, *DifferentOp2 = 0;
1232 unsigned Op1Idx = 0;
1233 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1234 if (I1->getOperand(I) == I2->getOperand(I))
1236 // Early exit if we have more-than one pair of different operands or
1237 // the different operand is already in MapValueFromBB1ToBB2.
1238 // Early exit if we need a PHI node to replace a constant.
1240 MapValueFromBB1ToBB2.find(I1->getOperand(I)) !=
1241 MapValueFromBB1ToBB2.end() ||
1242 isa<Constant>(I1->getOperand(I)) ||
1243 isa<Constant>(I2->getOperand(I))) {
1244 // If we can't sink the instructions, undo the swapping.
1246 ICmp2->swapOperands();
1249 DifferentOp1 = I1->getOperand(I);
1251 DifferentOp2 = I2->getOperand(I);
1254 // We insert the pair of different operands to MapValueFromBB1ToBB2 and
1255 // remove (I1, I2) from MapValueFromBB1ToBB2.
1257 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2,
1258 DifferentOp1->getName() + ".sink",
1260 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN);
1261 // I1 should use NewPN instead of DifferentOp1.
1262 I1->setOperand(Op1Idx, NewPN);
1263 NewPN->addIncoming(DifferentOp1, BB1);
1264 NewPN->addIncoming(DifferentOp2, BB2);
1265 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1267 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second;
1268 MapValueFromBB1ToBB2.erase(I1);
1270 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";);
1271 DEBUG(dbgs() << " " << *I2 << "\n";);
1272 // We need to update RE1 and RE2 if we are going to sink the first
1273 // instruction in the basic block down.
1274 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1275 // Sink the instruction.
1276 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1);
1277 if (!OldPN->use_empty())
1278 OldPN->replaceAllUsesWith(I1);
1279 OldPN->eraseFromParent();
1281 if (!I2->use_empty())
1282 I2->replaceAllUsesWith(I1);
1283 I1->intersectOptionalDataWith(I2);
1284 I2->eraseFromParent();
1287 RE1 = BB1->getInstList().rend();
1289 RE2 = BB2->getInstList().rend();
1290 FirstNonPhiInBBEnd = I1;
1297 /// \brief Determine if we can hoist sink a sole store instruction out of a
1298 /// conditional block.
1300 /// We are looking for code like the following:
1302 /// store i32 %add, i32* %arrayidx2
1303 /// ... // No other stores or function calls (we could be calling a memory
1304 /// ... // function).
1305 /// %cmp = icmp ult %x, %y
1306 /// br i1 %cmp, label %EndBB, label %ThenBB
1308 /// store i32 %add5, i32* %arrayidx2
1312 /// We are going to transform this into:
1314 /// store i32 %add, i32* %arrayidx2
1316 /// %cmp = icmp ult %x, %y
1317 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1318 /// store i32 %add.add5, i32* %arrayidx2
1321 /// \return The pointer to the value of the previous store if the store can be
1322 /// hoisted into the predecessor block. 0 otherwise.
1323 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1324 BasicBlock *StoreBB, BasicBlock *EndBB) {
1325 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1329 // Volatile or atomic.
1330 if (!StoreToHoist->isSimple())
1333 Value *StorePtr = StoreToHoist->getPointerOperand();
1335 // Look for a store to the same pointer in BrBB.
1336 unsigned MaxNumInstToLookAt = 10;
1337 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(),
1338 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) {
1339 Instruction *CurI = &*RI;
1341 // Could be calling an instruction that effects memory like free().
1342 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI))
1345 StoreInst *SI = dyn_cast<StoreInst>(CurI);
1346 // Found the previous store make sure it stores to the same location.
1347 if (SI && SI->getPointerOperand() == StorePtr)
1348 // Found the previous store, return its value operand.
1349 return SI->getValueOperand();
1351 return 0; // Unknown store.
1357 /// \brief Speculate a conditional basic block flattening the CFG.
1359 /// Note that this is a very risky transform currently. Speculating
1360 /// instructions like this is most often not desirable. Instead, there is an MI
1361 /// pass which can do it with full awareness of the resource constraints.
1362 /// However, some cases are "obvious" and we should do directly. An example of
1363 /// this is speculating a single, reasonably cheap instruction.
1365 /// There is only one distinct advantage to flattening the CFG at the IR level:
1366 /// it makes very common but simplistic optimizations such as are common in
1367 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1368 /// modeling their effects with easier to reason about SSA value graphs.
1371 /// An illustration of this transform is turning this IR:
1374 /// %cmp = icmp ult %x, %y
1375 /// br i1 %cmp, label %EndBB, label %ThenBB
1377 /// %sub = sub %x, %y
1380 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1387 /// %cmp = icmp ult %x, %y
1388 /// %sub = sub %x, %y
1389 /// %cond = select i1 %cmp, 0, %sub
1393 /// \returns true if the conditional block is removed.
1394 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB) {
1395 // Be conservative for now. FP select instruction can often be expensive.
1396 Value *BrCond = BI->getCondition();
1397 if (isa<FCmpInst>(BrCond))
1400 BasicBlock *BB = BI->getParent();
1401 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1403 // If ThenBB is actually on the false edge of the conditional branch, remember
1404 // to swap the select operands later.
1405 bool Invert = false;
1406 if (ThenBB != BI->getSuccessor(0)) {
1407 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1410 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1412 // Keep a count of how many times instructions are used within CondBB when
1413 // they are candidates for sinking into CondBB. Specifically:
1414 // - They are defined in BB, and
1415 // - They have no side effects, and
1416 // - All of their uses are in CondBB.
1417 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1419 unsigned SpeculationCost = 0;
1420 Value *SpeculatedStoreValue = 0;
1421 StoreInst *SpeculatedStore = 0;
1422 for (BasicBlock::iterator BBI = ThenBB->begin(),
1423 BBE = llvm::prior(ThenBB->end());
1424 BBI != BBE; ++BBI) {
1425 Instruction *I = BBI;
1427 if (isa<DbgInfoIntrinsic>(I))
1430 // Only speculatively execution a single instruction (not counting the
1431 // terminator) for now.
1433 if (SpeculationCost > 1)
1436 // Don't hoist the instruction if it's unsafe or expensive.
1437 if (!isSafeToSpeculativelyExecute(I) &&
1438 !(HoistCondStores &&
1439 (SpeculatedStoreValue = isSafeToSpeculateStore(I, BB, ThenBB,
1442 if (!SpeculatedStoreValue &&
1443 ComputeSpeculationCost(I) > PHINodeFoldingThreshold)
1446 // Store the store speculation candidate.
1447 if (SpeculatedStoreValue)
1448 SpeculatedStore = cast<StoreInst>(I);
1450 // Do not hoist the instruction if any of its operands are defined but not
1451 // used in BB. The transformation will prevent the operand from
1452 // being sunk into the use block.
1453 for (User::op_iterator i = I->op_begin(), e = I->op_end();
1455 Instruction *OpI = dyn_cast<Instruction>(*i);
1456 if (!OpI || OpI->getParent() != BB ||
1457 OpI->mayHaveSideEffects())
1458 continue; // Not a candidate for sinking.
1460 ++SinkCandidateUseCounts[OpI];
1464 // Consider any sink candidates which are only used in CondBB as costs for
1465 // speculation. Note, while we iterate over a DenseMap here, we are summing
1466 // and so iteration order isn't significant.
1467 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I =
1468 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end();
1470 if (I->first->getNumUses() == I->second) {
1472 if (SpeculationCost > 1)
1476 // Check that the PHI nodes can be converted to selects.
1477 bool HaveRewritablePHIs = false;
1478 for (BasicBlock::iterator I = EndBB->begin();
1479 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1480 Value *OrigV = PN->getIncomingValueForBlock(BB);
1481 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1483 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1484 // Skip PHIs which are trivial.
1488 HaveRewritablePHIs = true;
1489 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1490 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1491 if (!OrigCE && !ThenCE)
1492 continue; // Known safe and cheap.
1494 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
1495 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
1497 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE) : 0;
1498 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE) : 0;
1499 if (OrigCost + ThenCost > 2 * PHINodeFoldingThreshold)
1502 // Account for the cost of an unfolded ConstantExpr which could end up
1503 // getting expanded into Instructions.
1504 // FIXME: This doesn't account for how many operations are combined in the
1505 // constant expression.
1507 if (SpeculationCost > 1)
1511 // If there are no PHIs to process, bail early. This helps ensure idempotence
1513 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
1516 // If we get here, we can hoist the instruction and if-convert.
1517 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1519 // Insert a select of the value of the speculated store.
1520 if (SpeculatedStoreValue) {
1521 IRBuilder<true, NoFolder> Builder(BI);
1522 Value *TrueV = SpeculatedStore->getValueOperand();
1523 Value *FalseV = SpeculatedStoreValue;
1525 std::swap(TrueV, FalseV);
1526 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() +
1527 "." + FalseV->getName());
1528 SpeculatedStore->setOperand(0, S);
1531 // Hoist the instructions.
1532 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(),
1533 llvm::prior(ThenBB->end()));
1535 // Insert selects and rewrite the PHI operands.
1536 IRBuilder<true, NoFolder> Builder(BI);
1537 for (BasicBlock::iterator I = EndBB->begin();
1538 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1539 unsigned OrigI = PN->getBasicBlockIndex(BB);
1540 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1541 Value *OrigV = PN->getIncomingValue(OrigI);
1542 Value *ThenV = PN->getIncomingValue(ThenI);
1544 // Skip PHIs which are trivial.
1548 // Create a select whose true value is the speculatively executed value and
1549 // false value is the preexisting value. Swap them if the branch
1550 // destinations were inverted.
1551 Value *TrueV = ThenV, *FalseV = OrigV;
1553 std::swap(TrueV, FalseV);
1554 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV,
1555 TrueV->getName() + "." + FalseV->getName());
1556 PN->setIncomingValue(OrigI, V);
1557 PN->setIncomingValue(ThenI, V);
1564 /// \returns True if this block contains a CallInst with the NoDuplicate
1566 static bool HasNoDuplicateCall(const BasicBlock *BB) {
1567 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1568 const CallInst *CI = dyn_cast<CallInst>(I);
1571 if (CI->cannotDuplicate())
1577 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1578 /// across this block.
1579 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1580 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1583 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1584 if (isa<DbgInfoIntrinsic>(BBI))
1586 if (Size > 10) return false; // Don't clone large BB's.
1589 // We can only support instructions that do not define values that are
1590 // live outside of the current basic block.
1591 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end();
1593 Instruction *U = cast<Instruction>(*UI);
1594 if (U->getParent() != BB || isa<PHINode>(U)) return false;
1597 // Looks ok, continue checking.
1603 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1604 /// that is defined in the same block as the branch and if any PHI entries are
1605 /// constants, thread edges corresponding to that entry to be branches to their
1606 /// ultimate destination.
1607 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *TD) {
1608 BasicBlock *BB = BI->getParent();
1609 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1610 // NOTE: we currently cannot transform this case if the PHI node is used
1611 // outside of the block.
1612 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1615 // Degenerate case of a single entry PHI.
1616 if (PN->getNumIncomingValues() == 1) {
1617 FoldSingleEntryPHINodes(PN->getParent());
1621 // Now we know that this block has multiple preds and two succs.
1622 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1624 if (HasNoDuplicateCall(BB)) return false;
1626 // Okay, this is a simple enough basic block. See if any phi values are
1628 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1629 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1630 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue;
1632 // Okay, we now know that all edges from PredBB should be revectored to
1633 // branch to RealDest.
1634 BasicBlock *PredBB = PN->getIncomingBlock(i);
1635 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1637 if (RealDest == BB) continue; // Skip self loops.
1638 // Skip if the predecessor's terminator is an indirect branch.
1639 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1641 // The dest block might have PHI nodes, other predecessors and other
1642 // difficult cases. Instead of being smart about this, just insert a new
1643 // block that jumps to the destination block, effectively splitting
1644 // the edge we are about to create.
1645 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1646 RealDest->getName()+".critedge",
1647 RealDest->getParent(), RealDest);
1648 BranchInst::Create(RealDest, EdgeBB);
1650 // Update PHI nodes.
1651 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1653 // BB may have instructions that are being threaded over. Clone these
1654 // instructions into EdgeBB. We know that there will be no uses of the
1655 // cloned instructions outside of EdgeBB.
1656 BasicBlock::iterator InsertPt = EdgeBB->begin();
1657 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1658 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1659 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1660 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1663 // Clone the instruction.
1664 Instruction *N = BBI->clone();
1665 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1667 // Update operands due to translation.
1668 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1670 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1671 if (PI != TranslateMap.end())
1675 // Check for trivial simplification.
1676 if (Value *V = SimplifyInstruction(N, TD)) {
1677 TranslateMap[BBI] = V;
1678 delete N; // Instruction folded away, don't need actual inst
1680 // Insert the new instruction into its new home.
1681 EdgeBB->getInstList().insert(InsertPt, N);
1682 if (!BBI->use_empty())
1683 TranslateMap[BBI] = N;
1687 // Loop over all of the edges from PredBB to BB, changing them to branch
1688 // to EdgeBB instead.
1689 TerminatorInst *PredBBTI = PredBB->getTerminator();
1690 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1691 if (PredBBTI->getSuccessor(i) == BB) {
1692 BB->removePredecessor(PredBB);
1693 PredBBTI->setSuccessor(i, EdgeBB);
1696 // Recurse, simplifying any other constants.
1697 return FoldCondBranchOnPHI(BI, TD) | true;
1703 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1704 /// PHI node, see if we can eliminate it.
1705 static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *TD) {
1706 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1707 // statement", which has a very simple dominance structure. Basically, we
1708 // are trying to find the condition that is being branched on, which
1709 // subsequently causes this merge to happen. We really want control
1710 // dependence information for this check, but simplifycfg can't keep it up
1711 // to date, and this catches most of the cases we care about anyway.
1712 BasicBlock *BB = PN->getParent();
1713 BasicBlock *IfTrue, *IfFalse;
1714 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1716 // Don't bother if the branch will be constant folded trivially.
1717 isa<ConstantInt>(IfCond))
1720 // Okay, we found that we can merge this two-entry phi node into a select.
1721 // Doing so would require us to fold *all* two entry phi nodes in this block.
1722 // At some point this becomes non-profitable (particularly if the target
1723 // doesn't support cmov's). Only do this transformation if there are two or
1724 // fewer PHI nodes in this block.
1725 unsigned NumPhis = 0;
1726 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1730 // Loop over the PHI's seeing if we can promote them all to select
1731 // instructions. While we are at it, keep track of the instructions
1732 // that need to be moved to the dominating block.
1733 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1734 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1735 MaxCostVal1 = PHINodeFoldingThreshold;
1737 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1738 PHINode *PN = cast<PHINode>(II++);
1739 if (Value *V = SimplifyInstruction(PN, TD)) {
1740 PN->replaceAllUsesWith(V);
1741 PN->eraseFromParent();
1745 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1747 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1752 // If we folded the first phi, PN dangles at this point. Refresh it. If
1753 // we ran out of PHIs then we simplified them all.
1754 PN = dyn_cast<PHINode>(BB->begin());
1755 if (PN == 0) return true;
1757 // Don't fold i1 branches on PHIs which contain binary operators. These can
1758 // often be turned into switches and other things.
1759 if (PN->getType()->isIntegerTy(1) &&
1760 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1761 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1762 isa<BinaryOperator>(IfCond)))
1765 // If we all PHI nodes are promotable, check to make sure that all
1766 // instructions in the predecessor blocks can be promoted as well. If
1767 // not, we won't be able to get rid of the control flow, so it's not
1768 // worth promoting to select instructions.
1769 BasicBlock *DomBlock = 0;
1770 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1771 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1772 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1775 DomBlock = *pred_begin(IfBlock1);
1776 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1777 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1778 // This is not an aggressive instruction that we can promote.
1779 // Because of this, we won't be able to get rid of the control
1780 // flow, so the xform is not worth it.
1785 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1788 DomBlock = *pred_begin(IfBlock2);
1789 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1790 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1791 // This is not an aggressive instruction that we can promote.
1792 // Because of this, we won't be able to get rid of the control
1793 // flow, so the xform is not worth it.
1798 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1799 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1801 // If we can still promote the PHI nodes after this gauntlet of tests,
1802 // do all of the PHI's now.
1803 Instruction *InsertPt = DomBlock->getTerminator();
1804 IRBuilder<true, NoFolder> Builder(InsertPt);
1806 // Move all 'aggressive' instructions, which are defined in the
1807 // conditional parts of the if's up to the dominating block.
1809 DomBlock->getInstList().splice(InsertPt,
1810 IfBlock1->getInstList(), IfBlock1->begin(),
1811 IfBlock1->getTerminator());
1813 DomBlock->getInstList().splice(InsertPt,
1814 IfBlock2->getInstList(), IfBlock2->begin(),
1815 IfBlock2->getTerminator());
1817 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1818 // Change the PHI node into a select instruction.
1819 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1820 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1823 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1824 PN->replaceAllUsesWith(NV);
1826 PN->eraseFromParent();
1829 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1830 // has been flattened. Change DomBlock to jump directly to our new block to
1831 // avoid other simplifycfg's kicking in on the diamond.
1832 TerminatorInst *OldTI = DomBlock->getTerminator();
1833 Builder.SetInsertPoint(OldTI);
1834 Builder.CreateBr(BB);
1835 OldTI->eraseFromParent();
1839 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1840 /// to two returning blocks, try to merge them together into one return,
1841 /// introducing a select if the return values disagree.
1842 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1843 IRBuilder<> &Builder) {
1844 assert(BI->isConditional() && "Must be a conditional branch");
1845 BasicBlock *TrueSucc = BI->getSuccessor(0);
1846 BasicBlock *FalseSucc = BI->getSuccessor(1);
1847 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1848 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1850 // Check to ensure both blocks are empty (just a return) or optionally empty
1851 // with PHI nodes. If there are other instructions, merging would cause extra
1852 // computation on one path or the other.
1853 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1855 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1858 Builder.SetInsertPoint(BI);
1859 // Okay, we found a branch that is going to two return nodes. If
1860 // there is no return value for this function, just change the
1861 // branch into a return.
1862 if (FalseRet->getNumOperands() == 0) {
1863 TrueSucc->removePredecessor(BI->getParent());
1864 FalseSucc->removePredecessor(BI->getParent());
1865 Builder.CreateRetVoid();
1866 EraseTerminatorInstAndDCECond(BI);
1870 // Otherwise, figure out what the true and false return values are
1871 // so we can insert a new select instruction.
1872 Value *TrueValue = TrueRet->getReturnValue();
1873 Value *FalseValue = FalseRet->getReturnValue();
1875 // Unwrap any PHI nodes in the return blocks.
1876 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1877 if (TVPN->getParent() == TrueSucc)
1878 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1879 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1880 if (FVPN->getParent() == FalseSucc)
1881 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1883 // In order for this transformation to be safe, we must be able to
1884 // unconditionally execute both operands to the return. This is
1885 // normally the case, but we could have a potentially-trapping
1886 // constant expression that prevents this transformation from being
1888 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1891 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1895 // Okay, we collected all the mapped values and checked them for sanity, and
1896 // defined to really do this transformation. First, update the CFG.
1897 TrueSucc->removePredecessor(BI->getParent());
1898 FalseSucc->removePredecessor(BI->getParent());
1900 // Insert select instructions where needed.
1901 Value *BrCond = BI->getCondition();
1903 // Insert a select if the results differ.
1904 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1905 } else if (isa<UndefValue>(TrueValue)) {
1906 TrueValue = FalseValue;
1908 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1909 FalseValue, "retval");
1913 Value *RI = !TrueValue ?
1914 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
1918 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
1919 << "\n " << *BI << "NewRet = " << *RI
1920 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
1922 EraseTerminatorInstAndDCECond(BI);
1927 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
1928 /// probabilities of the branch taking each edge. Fills in the two APInt
1929 /// parameters and return true, or returns false if no or invalid metadata was
1931 static bool ExtractBranchMetadata(BranchInst *BI,
1932 uint64_t &ProbTrue, uint64_t &ProbFalse) {
1933 assert(BI->isConditional() &&
1934 "Looking for probabilities on unconditional branch?");
1935 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
1936 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
1937 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
1938 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
1939 if (!CITrue || !CIFalse) return false;
1940 ProbTrue = CITrue->getValue().getZExtValue();
1941 ProbFalse = CIFalse->getValue().getZExtValue();
1945 /// checkCSEInPredecessor - Return true if the given instruction is available
1946 /// in its predecessor block. If yes, the instruction will be removed.
1948 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
1949 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
1951 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
1952 Instruction *PBI = &*I;
1953 // Check whether Inst and PBI generate the same value.
1954 if (Inst->isIdenticalTo(PBI)) {
1955 Inst->replaceAllUsesWith(PBI);
1956 Inst->eraseFromParent();
1963 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
1964 /// predecessor branches to us and one of our successors, fold the block into
1965 /// the predecessor and use logical operations to pick the right destination.
1966 bool llvm::FoldBranchToCommonDest(BranchInst *BI) {
1967 BasicBlock *BB = BI->getParent();
1969 Instruction *Cond = 0;
1970 if (BI->isConditional())
1971 Cond = dyn_cast<Instruction>(BI->getCondition());
1973 // For unconditional branch, check for a simple CFG pattern, where
1974 // BB has a single predecessor and BB's successor is also its predecessor's
1975 // successor. If such pattern exisits, check for CSE between BB and its
1977 if (BasicBlock *PB = BB->getSinglePredecessor())
1978 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
1979 if (PBI->isConditional() &&
1980 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
1981 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
1982 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
1984 Instruction *Curr = I++;
1985 if (isa<CmpInst>(Curr)) {
1989 // Quit if we can't remove this instruction.
1990 if (!checkCSEInPredecessor(Curr, PB))
1999 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2000 Cond->getParent() != BB || !Cond->hasOneUse())
2003 // Only allow this if the condition is a simple instruction that can be
2004 // executed unconditionally. It must be in the same block as the branch, and
2005 // must be at the front of the block.
2006 BasicBlock::iterator FrontIt = BB->front();
2008 // Ignore dbg intrinsics.
2009 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
2011 // Allow a single instruction to be hoisted in addition to the compare
2012 // that feeds the branch. We later ensure that any values that _it_ uses
2013 // were also live in the predecessor, so that we don't unnecessarily create
2014 // register pressure or inhibit out-of-order execution.
2015 Instruction *BonusInst = 0;
2016 if (&*FrontIt != Cond &&
2017 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond &&
2018 isSafeToSpeculativelyExecute(FrontIt)) {
2019 BonusInst = &*FrontIt;
2022 // Ignore dbg intrinsics.
2023 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
2026 // Only a single bonus inst is allowed.
2027 if (&*FrontIt != Cond)
2030 // Make sure the instruction after the condition is the cond branch.
2031 BasicBlock::iterator CondIt = Cond; ++CondIt;
2033 // Ingore dbg intrinsics.
2034 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
2039 // Cond is known to be a compare or binary operator. Check to make sure that
2040 // neither operand is a potentially-trapping constant expression.
2041 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2044 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2048 // Finally, don't infinitely unroll conditional loops.
2049 BasicBlock *TrueDest = BI->getSuccessor(0);
2050 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0;
2051 if (TrueDest == BB || FalseDest == BB)
2054 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2055 BasicBlock *PredBlock = *PI;
2056 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2058 // Check that we have two conditional branches. If there is a PHI node in
2059 // the common successor, verify that the same value flows in from both
2061 SmallVector<PHINode*, 4> PHIs;
2062 if (PBI == 0 || PBI->isUnconditional() ||
2063 (BI->isConditional() &&
2064 !SafeToMergeTerminators(BI, PBI)) ||
2065 (!BI->isConditional() &&
2066 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2069 // Determine if the two branches share a common destination.
2070 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2071 bool InvertPredCond = false;
2073 if (BI->isConditional()) {
2074 if (PBI->getSuccessor(0) == TrueDest)
2075 Opc = Instruction::Or;
2076 else if (PBI->getSuccessor(1) == FalseDest)
2077 Opc = Instruction::And;
2078 else if (PBI->getSuccessor(0) == FalseDest)
2079 Opc = Instruction::And, InvertPredCond = true;
2080 else if (PBI->getSuccessor(1) == TrueDest)
2081 Opc = Instruction::Or, InvertPredCond = true;
2085 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2089 // Ensure that any values used in the bonus instruction are also used
2090 // by the terminator of the predecessor. This means that those values
2091 // must already have been resolved, so we won't be inhibiting the
2092 // out-of-order core by speculating them earlier.
2094 // Collect the values used by the bonus inst
2095 SmallPtrSet<Value*, 4> UsedValues;
2096 for (Instruction::op_iterator OI = BonusInst->op_begin(),
2097 OE = BonusInst->op_end(); OI != OE; ++OI) {
2099 if (!isa<Constant>(V))
2100 UsedValues.insert(V);
2103 SmallVector<std::pair<Value*, unsigned>, 4> Worklist;
2104 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0));
2106 // Walk up to four levels back up the use-def chain of the predecessor's
2107 // terminator to see if all those values were used. The choice of four
2108 // levels is arbitrary, to provide a compile-time-cost bound.
2109 while (!Worklist.empty()) {
2110 std::pair<Value*, unsigned> Pair = Worklist.back();
2111 Worklist.pop_back();
2113 if (Pair.second >= 4) continue;
2114 UsedValues.erase(Pair.first);
2115 if (UsedValues.empty()) break;
2117 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) {
2118 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
2120 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1));
2124 if (!UsedValues.empty()) return false;
2127 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2128 IRBuilder<> Builder(PBI);
2130 // If we need to invert the condition in the pred block to match, do so now.
2131 if (InvertPredCond) {
2132 Value *NewCond = PBI->getCondition();
2134 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2135 CmpInst *CI = cast<CmpInst>(NewCond);
2136 CI->setPredicate(CI->getInversePredicate());
2138 NewCond = Builder.CreateNot(NewCond,
2139 PBI->getCondition()->getName()+".not");
2142 PBI->setCondition(NewCond);
2143 PBI->swapSuccessors();
2146 // If we have a bonus inst, clone it into the predecessor block.
2147 Instruction *NewBonus = 0;
2149 NewBonus = BonusInst->clone();
2150 PredBlock->getInstList().insert(PBI, NewBonus);
2151 NewBonus->takeName(BonusInst);
2152 BonusInst->setName(BonusInst->getName()+".old");
2155 // Clone Cond into the predecessor basic block, and or/and the
2156 // two conditions together.
2157 Instruction *New = Cond->clone();
2158 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus);
2159 PredBlock->getInstList().insert(PBI, New);
2160 New->takeName(Cond);
2161 Cond->setName(New->getName()+".old");
2163 if (BI->isConditional()) {
2164 Instruction *NewCond =
2165 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2167 PBI->setCondition(NewCond);
2169 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2170 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2172 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2174 SmallVector<uint64_t, 8> NewWeights;
2176 if (PBI->getSuccessor(0) == BB) {
2177 if (PredHasWeights && SuccHasWeights) {
2178 // PBI: br i1 %x, BB, FalseDest
2179 // BI: br i1 %y, TrueDest, FalseDest
2180 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2181 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2182 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2183 // TrueWeight for PBI * FalseWeight for BI.
2184 // We assume that total weights of a BranchInst can fit into 32 bits.
2185 // Therefore, we will not have overflow using 64-bit arithmetic.
2186 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2187 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2189 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2190 PBI->setSuccessor(0, TrueDest);
2192 if (PBI->getSuccessor(1) == BB) {
2193 if (PredHasWeights && SuccHasWeights) {
2194 // PBI: br i1 %x, TrueDest, BB
2195 // BI: br i1 %y, TrueDest, FalseDest
2196 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2197 // FalseWeight for PBI * TrueWeight for BI.
2198 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2199 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2200 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2201 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2203 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2204 PBI->setSuccessor(1, FalseDest);
2206 if (NewWeights.size() == 2) {
2207 // Halve the weights if any of them cannot fit in an uint32_t
2208 FitWeights(NewWeights);
2210 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2211 PBI->setMetadata(LLVMContext::MD_prof,
2212 MDBuilder(BI->getContext()).
2213 createBranchWeights(MDWeights));
2215 PBI->setMetadata(LLVMContext::MD_prof, NULL);
2217 // Update PHI nodes in the common successors.
2218 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2219 ConstantInt *PBI_C = cast<ConstantInt>(
2220 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2221 assert(PBI_C->getType()->isIntegerTy(1));
2222 Instruction *MergedCond = 0;
2223 if (PBI->getSuccessor(0) == TrueDest) {
2224 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2225 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2226 // is false: !PBI_Cond and BI_Value
2227 Instruction *NotCond =
2228 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2231 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2236 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2237 PBI->getCondition(), MergedCond,
2240 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2241 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2242 // is false: PBI_Cond and BI_Value
2244 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2245 PBI->getCondition(), New,
2247 if (PBI_C->isOne()) {
2248 Instruction *NotCond =
2249 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2252 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2253 NotCond, MergedCond,
2258 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2261 // Change PBI from Conditional to Unconditional.
2262 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2263 EraseTerminatorInstAndDCECond(PBI);
2267 // TODO: If BB is reachable from all paths through PredBlock, then we
2268 // could replace PBI's branch probabilities with BI's.
2270 // Copy any debug value intrinsics into the end of PredBlock.
2271 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2272 if (isa<DbgInfoIntrinsic>(*I))
2273 I->clone()->insertBefore(PBI);
2280 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2281 /// predecessor of another block, this function tries to simplify it. We know
2282 /// that PBI and BI are both conditional branches, and BI is in one of the
2283 /// successor blocks of PBI - PBI branches to BI.
2284 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2285 assert(PBI->isConditional() && BI->isConditional());
2286 BasicBlock *BB = BI->getParent();
2288 // If this block ends with a branch instruction, and if there is a
2289 // predecessor that ends on a branch of the same condition, make
2290 // this conditional branch redundant.
2291 if (PBI->getCondition() == BI->getCondition() &&
2292 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2293 // Okay, the outcome of this conditional branch is statically
2294 // knowable. If this block had a single pred, handle specially.
2295 if (BB->getSinglePredecessor()) {
2296 // Turn this into a branch on constant.
2297 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2298 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2300 return true; // Nuke the branch on constant.
2303 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2304 // in the constant and simplify the block result. Subsequent passes of
2305 // simplifycfg will thread the block.
2306 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2307 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2308 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2309 std::distance(PB, PE),
2310 BI->getCondition()->getName() + ".pr",
2312 // Okay, we're going to insert the PHI node. Since PBI is not the only
2313 // predecessor, compute the PHI'd conditional value for all of the preds.
2314 // Any predecessor where the condition is not computable we keep symbolic.
2315 for (pred_iterator PI = PB; PI != PE; ++PI) {
2316 BasicBlock *P = *PI;
2317 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2318 PBI != BI && PBI->isConditional() &&
2319 PBI->getCondition() == BI->getCondition() &&
2320 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2321 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2322 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2325 NewPN->addIncoming(BI->getCondition(), P);
2329 BI->setCondition(NewPN);
2334 // If this is a conditional branch in an empty block, and if any
2335 // predecessors is a conditional branch to one of our destinations,
2336 // fold the conditions into logical ops and one cond br.
2337 BasicBlock::iterator BBI = BB->begin();
2338 // Ignore dbg intrinsics.
2339 while (isa<DbgInfoIntrinsic>(BBI))
2345 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2350 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2352 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2353 PBIOp = 0, BIOp = 1;
2354 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2355 PBIOp = 1, BIOp = 0;
2356 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2361 // Check to make sure that the other destination of this branch
2362 // isn't BB itself. If so, this is an infinite loop that will
2363 // keep getting unwound.
2364 if (PBI->getSuccessor(PBIOp) == BB)
2367 // Do not perform this transformation if it would require
2368 // insertion of a large number of select instructions. For targets
2369 // without predication/cmovs, this is a big pessimization.
2370 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2372 unsigned NumPhis = 0;
2373 for (BasicBlock::iterator II = CommonDest->begin();
2374 isa<PHINode>(II); ++II, ++NumPhis)
2375 if (NumPhis > 2) // Disable this xform.
2378 // Finally, if everything is ok, fold the branches to logical ops.
2379 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2381 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2382 << "AND: " << *BI->getParent());
2385 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2386 // branch in it, where one edge (OtherDest) goes back to itself but the other
2387 // exits. We don't *know* that the program avoids the infinite loop
2388 // (even though that seems likely). If we do this xform naively, we'll end up
2389 // recursively unpeeling the loop. Since we know that (after the xform is
2390 // done) that the block *is* infinite if reached, we just make it an obviously
2391 // infinite loop with no cond branch.
2392 if (OtherDest == BB) {
2393 // Insert it at the end of the function, because it's either code,
2394 // or it won't matter if it's hot. :)
2395 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2396 "infloop", BB->getParent());
2397 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2398 OtherDest = InfLoopBlock;
2401 DEBUG(dbgs() << *PBI->getParent()->getParent());
2403 // BI may have other predecessors. Because of this, we leave
2404 // it alone, but modify PBI.
2406 // Make sure we get to CommonDest on True&True directions.
2407 Value *PBICond = PBI->getCondition();
2408 IRBuilder<true, NoFolder> Builder(PBI);
2410 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2412 Value *BICond = BI->getCondition();
2414 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2416 // Merge the conditions.
2417 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2419 // Modify PBI to branch on the new condition to the new dests.
2420 PBI->setCondition(Cond);
2421 PBI->setSuccessor(0, CommonDest);
2422 PBI->setSuccessor(1, OtherDest);
2424 // Update branch weight for PBI.
2425 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2426 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2428 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2430 if (PredHasWeights && SuccHasWeights) {
2431 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2432 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2433 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2434 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2435 // The weight to CommonDest should be PredCommon * SuccTotal +
2436 // PredOther * SuccCommon.
2437 // The weight to OtherDest should be PredOther * SuccOther.
2438 SmallVector<uint64_t, 2> NewWeights;
2439 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) +
2440 PredOther * SuccCommon);
2441 NewWeights.push_back(PredOther * SuccOther);
2442 // Halve the weights if any of them cannot fit in an uint32_t
2443 FitWeights(NewWeights);
2445 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end());
2446 PBI->setMetadata(LLVMContext::MD_prof,
2447 MDBuilder(BI->getContext()).
2448 createBranchWeights(MDWeights));
2451 // OtherDest may have phi nodes. If so, add an entry from PBI's
2452 // block that are identical to the entries for BI's block.
2453 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2455 // We know that the CommonDest already had an edge from PBI to
2456 // it. If it has PHIs though, the PHIs may have different
2457 // entries for BB and PBI's BB. If so, insert a select to make
2460 for (BasicBlock::iterator II = CommonDest->begin();
2461 (PN = dyn_cast<PHINode>(II)); ++II) {
2462 Value *BIV = PN->getIncomingValueForBlock(BB);
2463 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2464 Value *PBIV = PN->getIncomingValue(PBBIdx);
2466 // Insert a select in PBI to pick the right value.
2467 Value *NV = cast<SelectInst>
2468 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2469 PN->setIncomingValue(PBBIdx, NV);
2473 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2474 DEBUG(dbgs() << *PBI->getParent()->getParent());
2476 // This basic block is probably dead. We know it has at least
2477 // one fewer predecessor.
2481 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2482 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2483 // Takes care of updating the successors and removing the old terminator.
2484 // Also makes sure not to introduce new successors by assuming that edges to
2485 // non-successor TrueBBs and FalseBBs aren't reachable.
2486 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2487 BasicBlock *TrueBB, BasicBlock *FalseBB,
2488 uint32_t TrueWeight,
2489 uint32_t FalseWeight){
2490 // Remove any superfluous successor edges from the CFG.
2491 // First, figure out which successors to preserve.
2492 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2494 BasicBlock *KeepEdge1 = TrueBB;
2495 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0;
2497 // Then remove the rest.
2498 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2499 BasicBlock *Succ = OldTerm->getSuccessor(I);
2500 // Make sure only to keep exactly one copy of each edge.
2501 if (Succ == KeepEdge1)
2503 else if (Succ == KeepEdge2)
2506 Succ->removePredecessor(OldTerm->getParent());
2509 IRBuilder<> Builder(OldTerm);
2510 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2512 // Insert an appropriate new terminator.
2513 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) {
2514 if (TrueBB == FalseBB)
2515 // We were only looking for one successor, and it was present.
2516 // Create an unconditional branch to it.
2517 Builder.CreateBr(TrueBB);
2519 // We found both of the successors we were looking for.
2520 // Create a conditional branch sharing the condition of the select.
2521 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2522 if (TrueWeight != FalseWeight)
2523 NewBI->setMetadata(LLVMContext::MD_prof,
2524 MDBuilder(OldTerm->getContext()).
2525 createBranchWeights(TrueWeight, FalseWeight));
2527 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2528 // Neither of the selected blocks were successors, so this
2529 // terminator must be unreachable.
2530 new UnreachableInst(OldTerm->getContext(), OldTerm);
2532 // One of the selected values was a successor, but the other wasn't.
2533 // Insert an unconditional branch to the one that was found;
2534 // the edge to the one that wasn't must be unreachable.
2536 // Only TrueBB was found.
2537 Builder.CreateBr(TrueBB);
2539 // Only FalseBB was found.
2540 Builder.CreateBr(FalseBB);
2543 EraseTerminatorInstAndDCECond(OldTerm);
2547 // SimplifySwitchOnSelect - Replaces
2548 // (switch (select cond, X, Y)) on constant X, Y
2549 // with a branch - conditional if X and Y lead to distinct BBs,
2550 // unconditional otherwise.
2551 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2552 // Check for constant integer values in the select.
2553 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2554 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2555 if (!TrueVal || !FalseVal)
2558 // Find the relevant condition and destinations.
2559 Value *Condition = Select->getCondition();
2560 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2561 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2563 // Get weight for TrueBB and FalseBB.
2564 uint32_t TrueWeight = 0, FalseWeight = 0;
2565 SmallVector<uint64_t, 8> Weights;
2566 bool HasWeights = HasBranchWeights(SI);
2568 GetBranchWeights(SI, Weights);
2569 if (Weights.size() == 1 + SI->getNumCases()) {
2570 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2571 getSuccessorIndex()];
2572 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2573 getSuccessorIndex()];
2577 // Perform the actual simplification.
2578 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2579 TrueWeight, FalseWeight);
2582 // SimplifyIndirectBrOnSelect - Replaces
2583 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2584 // blockaddress(@fn, BlockB)))
2586 // (br cond, BlockA, BlockB).
2587 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2588 // Check that both operands of the select are block addresses.
2589 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2590 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2594 // Extract the actual blocks.
2595 BasicBlock *TrueBB = TBA->getBasicBlock();
2596 BasicBlock *FalseBB = FBA->getBasicBlock();
2598 // Perform the actual simplification.
2599 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2603 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2604 /// instruction (a seteq/setne with a constant) as the only instruction in a
2605 /// block that ends with an uncond branch. We are looking for a very specific
2606 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2607 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2608 /// default value goes to an uncond block with a seteq in it, we get something
2611 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2613 /// %tmp = icmp eq i8 %A, 92
2616 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2618 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2619 /// the PHI, merging the third icmp into the switch.
2620 static bool TryToSimplifyUncondBranchWithICmpInIt(
2621 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI,
2622 const DataLayout *TD) {
2623 BasicBlock *BB = ICI->getParent();
2625 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2627 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2629 Value *V = ICI->getOperand(0);
2630 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2632 // The pattern we're looking for is where our only predecessor is a switch on
2633 // 'V' and this block is the default case for the switch. In this case we can
2634 // fold the compared value into the switch to simplify things.
2635 BasicBlock *Pred = BB->getSinglePredecessor();
2636 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false;
2638 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2639 if (SI->getCondition() != V)
2642 // If BB is reachable on a non-default case, then we simply know the value of
2643 // V in this block. Substitute it and constant fold the icmp instruction
2645 if (SI->getDefaultDest() != BB) {
2646 ConstantInt *VVal = SI->findCaseDest(BB);
2647 assert(VVal && "Should have a unique destination value");
2648 ICI->setOperand(0, VVal);
2650 if (Value *V = SimplifyInstruction(ICI, TD)) {
2651 ICI->replaceAllUsesWith(V);
2652 ICI->eraseFromParent();
2654 // BB is now empty, so it is likely to simplify away.
2655 return SimplifyCFG(BB, TTI, TD) | true;
2658 // Ok, the block is reachable from the default dest. If the constant we're
2659 // comparing exists in one of the other edges, then we can constant fold ICI
2661 if (SI->findCaseValue(Cst) != SI->case_default()) {
2663 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2664 V = ConstantInt::getFalse(BB->getContext());
2666 V = ConstantInt::getTrue(BB->getContext());
2668 ICI->replaceAllUsesWith(V);
2669 ICI->eraseFromParent();
2670 // BB is now empty, so it is likely to simplify away.
2671 return SimplifyCFG(BB, TTI, TD) | true;
2674 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2676 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2677 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back());
2678 if (PHIUse == 0 || PHIUse != &SuccBlock->front() ||
2679 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2682 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2684 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2685 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2687 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2688 std::swap(DefaultCst, NewCst);
2690 // Replace ICI (which is used by the PHI for the default value) with true or
2691 // false depending on if it is EQ or NE.
2692 ICI->replaceAllUsesWith(DefaultCst);
2693 ICI->eraseFromParent();
2695 // Okay, the switch goes to this block on a default value. Add an edge from
2696 // the switch to the merge point on the compared value.
2697 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2698 BB->getParent(), BB);
2699 SmallVector<uint64_t, 8> Weights;
2700 bool HasWeights = HasBranchWeights(SI);
2702 GetBranchWeights(SI, Weights);
2703 if (Weights.size() == 1 + SI->getNumCases()) {
2704 // Split weight for default case to case for "Cst".
2705 Weights[0] = (Weights[0]+1) >> 1;
2706 Weights.push_back(Weights[0]);
2708 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2709 SI->setMetadata(LLVMContext::MD_prof,
2710 MDBuilder(SI->getContext()).
2711 createBranchWeights(MDWeights));
2714 SI->addCase(Cst, NewBB);
2716 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2717 Builder.SetInsertPoint(NewBB);
2718 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2719 Builder.CreateBr(SuccBlock);
2720 PHIUse->addIncoming(NewCst, NewBB);
2724 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2725 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2726 /// fold it into a switch instruction if so.
2727 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *TD,
2728 IRBuilder<> &Builder) {
2729 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2730 if (Cond == 0) return false;
2733 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2734 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2735 // 'setne's and'ed together, collect them.
2737 std::vector<ConstantInt*> Values;
2738 bool TrueWhenEqual = true;
2739 Value *ExtraCase = 0;
2740 unsigned UsedICmps = 0;
2742 if (Cond->getOpcode() == Instruction::Or) {
2743 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true,
2745 } else if (Cond->getOpcode() == Instruction::And) {
2746 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false,
2748 TrueWhenEqual = false;
2751 // If we didn't have a multiply compared value, fail.
2752 if (CompVal == 0) return false;
2754 // Avoid turning single icmps into a switch.
2758 // There might be duplicate constants in the list, which the switch
2759 // instruction can't handle, remove them now.
2760 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2761 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2763 // If Extra was used, we require at least two switch values to do the
2764 // transformation. A switch with one value is just an cond branch.
2765 if (ExtraCase && Values.size() < 2) return false;
2767 // TODO: Preserve branch weight metadata, similarly to how
2768 // FoldValueComparisonIntoPredecessors preserves it.
2770 // Figure out which block is which destination.
2771 BasicBlock *DefaultBB = BI->getSuccessor(1);
2772 BasicBlock *EdgeBB = BI->getSuccessor(0);
2773 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2775 BasicBlock *BB = BI->getParent();
2777 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2778 << " cases into SWITCH. BB is:\n" << *BB);
2780 // If there are any extra values that couldn't be folded into the switch
2781 // then we evaluate them with an explicit branch first. Split the block
2782 // right before the condbr to handle it.
2784 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2785 // Remove the uncond branch added to the old block.
2786 TerminatorInst *OldTI = BB->getTerminator();
2787 Builder.SetInsertPoint(OldTI);
2790 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2792 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2794 OldTI->eraseFromParent();
2796 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2797 // for the edge we just added.
2798 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2800 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2801 << "\nEXTRABB = " << *BB);
2805 Builder.SetInsertPoint(BI);
2806 // Convert pointer to int before we switch.
2807 if (CompVal->getType()->isPointerTy()) {
2808 assert(TD && "Cannot switch on pointer without DataLayout");
2809 CompVal = Builder.CreatePtrToInt(CompVal,
2810 TD->getIntPtrType(CompVal->getType()),
2814 // Create the new switch instruction now.
2815 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2817 // Add all of the 'cases' to the switch instruction.
2818 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2819 New->addCase(Values[i], EdgeBB);
2821 // We added edges from PI to the EdgeBB. As such, if there were any
2822 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2823 // the number of edges added.
2824 for (BasicBlock::iterator BBI = EdgeBB->begin();
2825 isa<PHINode>(BBI); ++BBI) {
2826 PHINode *PN = cast<PHINode>(BBI);
2827 Value *InVal = PN->getIncomingValueForBlock(BB);
2828 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2829 PN->addIncoming(InVal, BB);
2832 // Erase the old branch instruction.
2833 EraseTerminatorInstAndDCECond(BI);
2835 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2839 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2840 // If this is a trivial landing pad that just continues unwinding the caught
2841 // exception then zap the landing pad, turning its invokes into calls.
2842 BasicBlock *BB = RI->getParent();
2843 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2844 if (RI->getValue() != LPInst)
2845 // Not a landing pad, or the resume is not unwinding the exception that
2846 // caused control to branch here.
2849 // Check that there are no other instructions except for debug intrinsics.
2850 BasicBlock::iterator I = LPInst, E = RI;
2852 if (!isa<DbgInfoIntrinsic>(I))
2855 // Turn all invokes that unwind here into calls and delete the basic block.
2856 bool InvokeRequiresTableEntry = false;
2857 bool Changed = false;
2858 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2859 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2861 if (II->hasFnAttr(Attribute::UWTable)) {
2862 // Don't remove an `invoke' instruction if the ABI requires an entry into
2864 InvokeRequiresTableEntry = true;
2868 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2870 // Insert a call instruction before the invoke.
2871 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2873 Call->setCallingConv(II->getCallingConv());
2874 Call->setAttributes(II->getAttributes());
2875 Call->setDebugLoc(II->getDebugLoc());
2877 // Anything that used the value produced by the invoke instruction now uses
2878 // the value produced by the call instruction. Note that we do this even
2879 // for void functions and calls with no uses so that the callgraph edge is
2881 II->replaceAllUsesWith(Call);
2882 BB->removePredecessor(II->getParent());
2884 // Insert a branch to the normal destination right before the invoke.
2885 BranchInst::Create(II->getNormalDest(), II);
2887 // Finally, delete the invoke instruction!
2888 II->eraseFromParent();
2892 if (!InvokeRequiresTableEntry)
2893 // The landingpad is now unreachable. Zap it.
2894 BB->eraseFromParent();
2899 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2900 BasicBlock *BB = RI->getParent();
2901 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2903 // Find predecessors that end with branches.
2904 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2905 SmallVector<BranchInst*, 8> CondBranchPreds;
2906 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2907 BasicBlock *P = *PI;
2908 TerminatorInst *PTI = P->getTerminator();
2909 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2910 if (BI->isUnconditional())
2911 UncondBranchPreds.push_back(P);
2913 CondBranchPreds.push_back(BI);
2917 // If we found some, do the transformation!
2918 if (!UncondBranchPreds.empty() && DupRet) {
2919 while (!UncondBranchPreds.empty()) {
2920 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
2921 DEBUG(dbgs() << "FOLDING: " << *BB
2922 << "INTO UNCOND BRANCH PRED: " << *Pred);
2923 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
2926 // If we eliminated all predecessors of the block, delete the block now.
2927 if (pred_begin(BB) == pred_end(BB))
2928 // We know there are no successors, so just nuke the block.
2929 BB->eraseFromParent();
2934 // Check out all of the conditional branches going to this return
2935 // instruction. If any of them just select between returns, change the
2936 // branch itself into a select/return pair.
2937 while (!CondBranchPreds.empty()) {
2938 BranchInst *BI = CondBranchPreds.pop_back_val();
2940 // Check to see if the non-BB successor is also a return block.
2941 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
2942 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
2943 SimplifyCondBranchToTwoReturns(BI, Builder))
2949 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
2950 BasicBlock *BB = UI->getParent();
2952 bool Changed = false;
2954 // If there are any instructions immediately before the unreachable that can
2955 // be removed, do so.
2956 while (UI != BB->begin()) {
2957 BasicBlock::iterator BBI = UI;
2959 // Do not delete instructions that can have side effects which might cause
2960 // the unreachable to not be reachable; specifically, calls and volatile
2961 // operations may have this effect.
2962 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
2964 if (BBI->mayHaveSideEffects()) {
2965 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
2966 if (SI->isVolatile())
2968 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
2969 if (LI->isVolatile())
2971 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
2972 if (RMWI->isVolatile())
2974 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
2975 if (CXI->isVolatile())
2977 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
2978 !isa<LandingPadInst>(BBI)) {
2981 // Note that deleting LandingPad's here is in fact okay, although it
2982 // involves a bit of subtle reasoning. If this inst is a LandingPad,
2983 // all the predecessors of this block will be the unwind edges of Invokes,
2984 // and we can therefore guarantee this block will be erased.
2987 // Delete this instruction (any uses are guaranteed to be dead)
2988 if (!BBI->use_empty())
2989 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
2990 BBI->eraseFromParent();
2994 // If the unreachable instruction is the first in the block, take a gander
2995 // at all of the predecessors of this instruction, and simplify them.
2996 if (&BB->front() != UI) return Changed;
2998 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
2999 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
3000 TerminatorInst *TI = Preds[i]->getTerminator();
3001 IRBuilder<> Builder(TI);
3002 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3003 if (BI->isUnconditional()) {
3004 if (BI->getSuccessor(0) == BB) {
3005 new UnreachableInst(TI->getContext(), TI);
3006 TI->eraseFromParent();
3010 if (BI->getSuccessor(0) == BB) {
3011 Builder.CreateBr(BI->getSuccessor(1));
3012 EraseTerminatorInstAndDCECond(BI);
3013 } else if (BI->getSuccessor(1) == BB) {
3014 Builder.CreateBr(BI->getSuccessor(0));
3015 EraseTerminatorInstAndDCECond(BI);
3019 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3020 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3022 if (i.getCaseSuccessor() == BB) {
3023 BB->removePredecessor(SI->getParent());
3028 // If the default value is unreachable, figure out the most popular
3029 // destination and make it the default.
3030 if (SI->getDefaultDest() == BB) {
3031 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
3032 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3034 std::pair<unsigned, unsigned> &entry =
3035 Popularity[i.getCaseSuccessor()];
3036 if (entry.first == 0) {
3038 entry.second = i.getCaseIndex();
3044 // Find the most popular block.
3045 unsigned MaxPop = 0;
3046 unsigned MaxIndex = 0;
3047 BasicBlock *MaxBlock = 0;
3048 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
3049 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
3050 if (I->second.first > MaxPop ||
3051 (I->second.first == MaxPop && MaxIndex > I->second.second)) {
3052 MaxPop = I->second.first;
3053 MaxIndex = I->second.second;
3054 MaxBlock = I->first;
3058 // Make this the new default, allowing us to delete any explicit
3060 SI->setDefaultDest(MaxBlock);
3063 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
3065 if (isa<PHINode>(MaxBlock->begin()))
3066 for (unsigned i = 0; i != MaxPop-1; ++i)
3067 MaxBlock->removePredecessor(SI->getParent());
3069 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3071 if (i.getCaseSuccessor() == MaxBlock) {
3077 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
3078 if (II->getUnwindDest() == BB) {
3079 // Convert the invoke to a call instruction. This would be a good
3080 // place to note that the call does not throw though.
3081 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
3082 II->removeFromParent(); // Take out of symbol table
3084 // Insert the call now...
3085 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
3086 Builder.SetInsertPoint(BI);
3087 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
3088 Args, II->getName());
3089 CI->setCallingConv(II->getCallingConv());
3090 CI->setAttributes(II->getAttributes());
3091 // If the invoke produced a value, the call does now instead.
3092 II->replaceAllUsesWith(CI);
3099 // If this block is now dead, remove it.
3100 if (pred_begin(BB) == pred_end(BB) &&
3101 BB != &BB->getParent()->getEntryBlock()) {
3102 // We know there are no successors, so just nuke the block.
3103 BB->eraseFromParent();
3110 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
3111 /// integer range comparison into a sub, an icmp and a branch.
3112 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3113 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3115 // Make sure all cases point to the same destination and gather the values.
3116 SmallVector<ConstantInt *, 16> Cases;
3117 SwitchInst::CaseIt I = SI->case_begin();
3118 Cases.push_back(I.getCaseValue());
3119 SwitchInst::CaseIt PrevI = I++;
3120 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
3121 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
3123 Cases.push_back(I.getCaseValue());
3125 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
3127 // Sort the case values, then check if they form a range we can transform.
3128 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3129 for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
3130 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
3134 Constant *Offset = ConstantExpr::getNeg(Cases.back());
3135 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
3137 Value *Sub = SI->getCondition();
3138 if (!Offset->isNullValue())
3139 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
3141 // If NumCases overflowed, then all possible values jump to the successor.
3142 if (NumCases->isNullValue() && SI->getNumCases() != 0)
3143 Cmp = ConstantInt::getTrue(SI->getContext());
3145 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3146 BranchInst *NewBI = Builder.CreateCondBr(
3147 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
3149 // Update weight for the newly-created conditional branch.
3150 SmallVector<uint64_t, 8> Weights;
3151 bool HasWeights = HasBranchWeights(SI);
3153 GetBranchWeights(SI, Weights);
3154 if (Weights.size() == 1 + SI->getNumCases()) {
3155 // Combine all weights for the cases to be the true weight of NewBI.
3156 // We assume that the sum of all weights for a Terminator can fit into 32
3158 uint32_t NewTrueWeight = 0;
3159 for (unsigned I = 1, E = Weights.size(); I != E; ++I)
3160 NewTrueWeight += (uint32_t)Weights[I];
3161 NewBI->setMetadata(LLVMContext::MD_prof,
3162 MDBuilder(SI->getContext()).
3163 createBranchWeights(NewTrueWeight,
3164 (uint32_t)Weights[0]));
3168 // Prune obsolete incoming values off the successor's PHI nodes.
3169 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
3170 isa<PHINode>(BBI); ++BBI) {
3171 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
3172 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3174 SI->eraseFromParent();
3179 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
3180 /// and use it to remove dead cases.
3181 static bool EliminateDeadSwitchCases(SwitchInst *SI) {
3182 Value *Cond = SI->getCondition();
3183 unsigned Bits = Cond->getType()->getIntegerBitWidth();
3184 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3185 ComputeMaskedBits(Cond, KnownZero, KnownOne);
3187 // Gather dead cases.
3188 SmallVector<ConstantInt*, 8> DeadCases;
3189 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3190 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3191 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3192 DeadCases.push_back(I.getCaseValue());
3193 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3194 << I.getCaseValue() << "' is dead.\n");
3198 SmallVector<uint64_t, 8> Weights;
3199 bool HasWeight = HasBranchWeights(SI);
3201 GetBranchWeights(SI, Weights);
3202 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3205 // Remove dead cases from the switch.
3206 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3207 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3208 assert(Case != SI->case_default() &&
3209 "Case was not found. Probably mistake in DeadCases forming.");
3211 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3215 // Prune unused values from PHI nodes.
3216 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3217 SI->removeCase(Case);
3220 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3221 SI->setMetadata(LLVMContext::MD_prof,
3222 MDBuilder(SI->getParent()->getContext()).
3223 createBranchWeights(MDWeights));
3226 return !DeadCases.empty();
3229 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
3230 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3231 /// by an unconditional branch), look at the phi node for BB in the successor
3232 /// block and see if the incoming value is equal to CaseValue. If so, return
3233 /// the phi node, and set PhiIndex to BB's index in the phi node.
3234 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3237 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3238 return NULL; // BB must be empty to be a candidate for simplification.
3239 if (!BB->getSinglePredecessor())
3240 return NULL; // BB must be dominated by the switch.
3242 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3243 if (!Branch || !Branch->isUnconditional())
3244 return NULL; // Terminator must be unconditional branch.
3246 BasicBlock *Succ = Branch->getSuccessor(0);
3248 BasicBlock::iterator I = Succ->begin();
3249 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3250 int Idx = PHI->getBasicBlockIndex(BB);
3251 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3253 Value *InValue = PHI->getIncomingValue(Idx);
3254 if (InValue != CaseValue) continue;
3263 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
3264 /// instruction to a phi node dominated by the switch, if that would mean that
3265 /// some of the destination blocks of the switch can be folded away.
3266 /// Returns true if a change is made.
3267 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3268 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3269 ForwardingNodesMap ForwardingNodes;
3271 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3272 ConstantInt *CaseValue = I.getCaseValue();
3273 BasicBlock *CaseDest = I.getCaseSuccessor();
3276 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3280 ForwardingNodes[PHI].push_back(PhiIndex);
3283 bool Changed = false;
3285 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3286 E = ForwardingNodes.end(); I != E; ++I) {
3287 PHINode *Phi = I->first;
3288 SmallVectorImpl<int> &Indexes = I->second;
3290 if (Indexes.size() < 2) continue;
3292 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3293 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3300 /// ValidLookupTableConstant - Return true if the backend will be able to handle
3301 /// initializing an array of constants like C.
3302 static bool ValidLookupTableConstant(Constant *C) {
3303 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3304 return CE->isGEPWithNoNotionalOverIndexing();
3306 return isa<ConstantFP>(C) ||
3307 isa<ConstantInt>(C) ||
3308 isa<ConstantPointerNull>(C) ||
3309 isa<GlobalValue>(C) ||
3313 /// LookupConstant - If V is a Constant, return it. Otherwise, try to look up
3314 /// its constant value in ConstantPool, returning 0 if it's not there.
3315 static Constant *LookupConstant(Value *V,
3316 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3317 if (Constant *C = dyn_cast<Constant>(V))
3319 return ConstantPool.lookup(V);
3322 /// ConstantFold - Try to fold instruction I into a constant. This works for
3323 /// simple instructions such as binary operations where both operands are
3324 /// constant or can be replaced by constants from the ConstantPool. Returns the
3325 /// resulting constant on success, 0 otherwise.
3326 static Constant *ConstantFold(Instruction *I,
3327 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3328 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
3329 Constant *A = LookupConstant(BO->getOperand(0), ConstantPool);
3332 Constant *B = LookupConstant(BO->getOperand(1), ConstantPool);
3335 return ConstantExpr::get(BO->getOpcode(), A, B);
3338 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
3339 Constant *A = LookupConstant(I->getOperand(0), ConstantPool);
3342 Constant *B = LookupConstant(I->getOperand(1), ConstantPool);
3345 return ConstantExpr::getCompare(Cmp->getPredicate(), A, B);
3348 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3349 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3352 if (A->isAllOnesValue())
3353 return LookupConstant(Select->getTrueValue(), ConstantPool);
3354 if (A->isNullValue())
3355 return LookupConstant(Select->getFalseValue(), ConstantPool);
3359 if (CastInst *Cast = dyn_cast<CastInst>(I)) {
3360 Constant *A = LookupConstant(I->getOperand(0), ConstantPool);
3363 return ConstantExpr::getCast(Cast->getOpcode(), A, Cast->getDestTy());
3369 /// GetCaseResults - Try to determine the resulting constant values in phi nodes
3370 /// at the common destination basic block, *CommonDest, for one of the case
3371 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3372 /// case), of a switch instruction SI.
3374 GetCaseResults(SwitchInst *SI,
3375 ConstantInt *CaseVal,
3376 BasicBlock *CaseDest,
3377 BasicBlock **CommonDest,
3378 SmallVectorImpl<std::pair<PHINode*,Constant*> > &Res) {
3379 // The block from which we enter the common destination.
3380 BasicBlock *Pred = SI->getParent();
3382 // If CaseDest is empty except for some side-effect free instructions through
3383 // which we can constant-propagate the CaseVal, continue to its successor.
3384 SmallDenseMap<Value*, Constant*> ConstantPool;
3385 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3386 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3388 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3389 // If the terminator is a simple branch, continue to the next block.
3390 if (T->getNumSuccessors() != 1)
3393 CaseDest = T->getSuccessor(0);
3394 } else if (isa<DbgInfoIntrinsic>(I)) {
3395 // Skip debug intrinsic.
3397 } else if (Constant *C = ConstantFold(I, ConstantPool)) {
3398 // Instruction is side-effect free and constant.
3399 ConstantPool.insert(std::make_pair(I, C));
3405 // If we did not have a CommonDest before, use the current one.
3407 *CommonDest = CaseDest;
3408 // If the destination isn't the common one, abort.
3409 if (CaseDest != *CommonDest)
3412 // Get the values for this case from phi nodes in the destination block.
3413 BasicBlock::iterator I = (*CommonDest)->begin();
3414 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3415 int Idx = PHI->getBasicBlockIndex(Pred);
3419 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3424 // Note: If the constant comes from constant-propagating the case value
3425 // through the CaseDest basic block, it will be safe to remove the
3426 // instructions in that block. They cannot be used (except in the phi nodes
3427 // we visit) outside CaseDest, because that block does not dominate its
3428 // successor. If it did, we would not be in this phi node.
3430 // Be conservative about which kinds of constants we support.
3431 if (!ValidLookupTableConstant(ConstVal))
3434 Res.push_back(std::make_pair(PHI, ConstVal));
3441 /// SwitchLookupTable - This class represents a lookup table that can be used
3442 /// to replace a switch.
3443 class SwitchLookupTable {
3445 /// SwitchLookupTable - Create a lookup table to use as a switch replacement
3446 /// with the contents of Values, using DefaultValue to fill any holes in the
3448 SwitchLookupTable(Module &M,
3450 ConstantInt *Offset,
3451 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3452 Constant *DefaultValue,
3453 const DataLayout *TD);
3455 /// BuildLookup - Build instructions with Builder to retrieve the value at
3456 /// the position given by Index in the lookup table.
3457 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3459 /// WouldFitInRegister - Return true if a table with TableSize elements of
3460 /// type ElementType would fit in a target-legal register.
3461 static bool WouldFitInRegister(const DataLayout *TD,
3463 const Type *ElementType);
3466 // Depending on the contents of the table, it can be represented in
3469 // For tables where each element contains the same value, we just have to
3470 // store that single value and return it for each lookup.
3473 // For small tables with integer elements, we can pack them into a bitmap
3474 // that fits into a target-legal register. Values are retrieved by
3475 // shift and mask operations.
3478 // The table is stored as an array of values. Values are retrieved by load
3479 // instructions from the table.
3483 // For SingleValueKind, this is the single value.
3484 Constant *SingleValue;
3486 // For BitMapKind, this is the bitmap.
3487 ConstantInt *BitMap;
3488 IntegerType *BitMapElementTy;
3490 // For ArrayKind, this is the array.
3491 GlobalVariable *Array;
3495 SwitchLookupTable::SwitchLookupTable(Module &M,
3497 ConstantInt *Offset,
3498 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3499 Constant *DefaultValue,
3500 const DataLayout *TD)
3501 : SingleValue(0), BitMap(0), BitMapElementTy(0), Array(0) {
3502 assert(Values.size() && "Can't build lookup table without values!");
3503 assert(TableSize >= Values.size() && "Can't fit values in table!");
3505 // If all values in the table are equal, this is that value.
3506 SingleValue = Values.begin()->second;
3508 // Build up the table contents.
3509 SmallVector<Constant*, 64> TableContents(TableSize);
3510 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3511 ConstantInt *CaseVal = Values[I].first;
3512 Constant *CaseRes = Values[I].second;
3513 assert(CaseRes->getType() == DefaultValue->getType());
3515 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3517 TableContents[Idx] = CaseRes;
3519 if (CaseRes != SingleValue)
3523 // Fill in any holes in the table with the default result.
3524 if (Values.size() < TableSize) {
3525 for (uint64_t I = 0; I < TableSize; ++I) {
3526 if (!TableContents[I])
3527 TableContents[I] = DefaultValue;
3530 if (DefaultValue != SingleValue)
3534 // If each element in the table contains the same value, we only need to store
3535 // that single value.
3537 Kind = SingleValueKind;
3541 // If the type is integer and the table fits in a register, build a bitmap.
3542 if (WouldFitInRegister(TD, TableSize, DefaultValue->getType())) {
3543 IntegerType *IT = cast<IntegerType>(DefaultValue->getType());
3544 APInt TableInt(TableSize * IT->getBitWidth(), 0);
3545 for (uint64_t I = TableSize; I > 0; --I) {
3546 TableInt <<= IT->getBitWidth();
3547 // Insert values into the bitmap. Undef values are set to zero.
3548 if (!isa<UndefValue>(TableContents[I - 1])) {
3549 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
3550 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
3553 BitMap = ConstantInt::get(M.getContext(), TableInt);
3554 BitMapElementTy = IT;
3560 // Store the table in an array.
3561 ArrayType *ArrayTy = ArrayType::get(DefaultValue->getType(), TableSize);
3562 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3564 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3565 GlobalVariable::PrivateLinkage,
3568 Array->setUnnamedAddr(true);
3572 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
3574 case SingleValueKind:
3577 // Type of the bitmap (e.g. i59).
3578 IntegerType *MapTy = BitMap->getType();
3580 // Cast Index to the same type as the bitmap.
3581 // Note: The Index is <= the number of elements in the table, so
3582 // truncating it to the width of the bitmask is safe.
3583 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
3585 // Multiply the shift amount by the element width.
3586 ShiftAmt = Builder.CreateMul(ShiftAmt,
3587 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
3591 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
3592 "switch.downshift");
3594 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
3598 Value *GEPIndices[] = { Builder.getInt32(0), Index };
3599 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices,
3601 return Builder.CreateLoad(GEP, "switch.load");
3604 llvm_unreachable("Unknown lookup table kind!");
3607 bool SwitchLookupTable::WouldFitInRegister(const DataLayout *TD,
3609 const Type *ElementType) {
3612 const IntegerType *IT = dyn_cast<IntegerType>(ElementType);
3615 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
3616 // are <= 15, we could try to narrow the type.
3618 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
3619 if (TableSize >= UINT_MAX/IT->getBitWidth())
3621 return TD->fitsInLegalInteger(TableSize * IT->getBitWidth());
3624 /// ShouldBuildLookupTable - Determine whether a lookup table should be built
3625 /// for this switch, based on the number of cases, size of the table and the
3626 /// types of the results.
3627 static bool ShouldBuildLookupTable(SwitchInst *SI,
3629 const TargetTransformInfo &TTI,
3630 const DataLayout *TD,
3631 const SmallDenseMap<PHINode*, Type*>& ResultTypes) {
3632 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
3633 return false; // TableSize overflowed, or mul below might overflow.
3635 bool AllTablesFitInRegister = true;
3636 bool HasIllegalType = false;
3637 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(),
3638 E = ResultTypes.end(); I != E; ++I) {
3639 Type *Ty = I->second;
3641 // Saturate this flag to true.
3642 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
3644 // Saturate this flag to false.
3645 AllTablesFitInRegister = AllTablesFitInRegister &&
3646 SwitchLookupTable::WouldFitInRegister(TD, TableSize, Ty);
3648 // If both flags saturate, we're done. NOTE: This *only* works with
3649 // saturating flags, and all flags have to saturate first due to the
3650 // non-deterministic behavior of iterating over a dense map.
3651 if (HasIllegalType && !AllTablesFitInRegister)
3655 // If each table would fit in a register, we should build it anyway.
3656 if (AllTablesFitInRegister)
3659 // Don't build a table that doesn't fit in-register if it has illegal types.
3663 // The table density should be at least 40%. This is the same criterion as for
3664 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3665 // FIXME: Find the best cut-off.
3666 return SI->getNumCases() * 10 >= TableSize * 4;
3669 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3670 /// phi nodes in a common successor block with different constant values,
3671 /// replace the switch with lookup tables.
3672 static bool SwitchToLookupTable(SwitchInst *SI,
3673 IRBuilder<> &Builder,
3674 const TargetTransformInfo &TTI,
3675 const DataLayout* TD) {
3676 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3678 // Only build lookup table when we have a target that supports it.
3679 if (!TTI.shouldBuildLookupTables())
3682 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3683 // split off a dense part and build a lookup table for that.
3685 // FIXME: This creates arrays of GEPs to constant strings, which means each
3686 // GEP needs a runtime relocation in PIC code. We should just build one big
3687 // string and lookup indices into that.
3689 // Ignore the switch if the number of cases is too small.
3690 // This is similar to the check when building jump tables in
3691 // SelectionDAGBuilder::handleJTSwitchCase.
3692 // FIXME: Determine the best cut-off.
3693 if (SI->getNumCases() < 4)
3696 // Figure out the corresponding result for each case value and phi node in the
3697 // common destination, as well as the the min and max case values.
3698 assert(SI->case_begin() != SI->case_end());
3699 SwitchInst::CaseIt CI = SI->case_begin();
3700 ConstantInt *MinCaseVal = CI.getCaseValue();
3701 ConstantInt *MaxCaseVal = CI.getCaseValue();
3703 BasicBlock *CommonDest = 0;
3704 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
3705 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
3706 SmallDenseMap<PHINode*, Constant*> DefaultResults;
3707 SmallDenseMap<PHINode*, Type*> ResultTypes;
3708 SmallVector<PHINode*, 4> PHIs;
3710 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
3711 ConstantInt *CaseVal = CI.getCaseValue();
3712 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
3713 MinCaseVal = CaseVal;
3714 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
3715 MaxCaseVal = CaseVal;
3717 // Resulting value at phi nodes for this case value.
3718 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
3720 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
3724 // Append the result from this case to the list for each phi.
3725 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) {
3726 if (!ResultLists.count(I->first))
3727 PHIs.push_back(I->first);
3728 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second));
3732 // Get the resulting values for the default case.
3733 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
3734 if (!GetCaseResults(SI, 0, SI->getDefaultDest(), &CommonDest,
3735 DefaultResultsList))
3737 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) {
3738 PHINode *PHI = DefaultResultsList[I].first;
3739 Constant *Result = DefaultResultsList[I].second;
3740 DefaultResults[PHI] = Result;
3741 ResultTypes[PHI] = Result->getType();
3744 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
3745 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
3746 if (!ShouldBuildLookupTable(SI, TableSize, TTI, TD, ResultTypes))
3749 // Create the BB that does the lookups.
3750 Module &Mod = *CommonDest->getParent()->getParent();
3751 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
3753 CommonDest->getParent(),
3756 // Compute the table index value.
3757 Builder.SetInsertPoint(SI);
3758 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
3761 // Compute the maximum table size representable by the integer type we are
3763 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
3764 uint64_t MaxTableSize = CaseSize > 63? UINT64_MAX : 1ULL << CaseSize;
3765 assert(MaxTableSize >= TableSize &&
3766 "It is impossible for a switch to have more entries than the max "
3767 "representable value of its input integer type's size.");
3769 // If we have a fully covered lookup table, unconditionally branch to the
3770 // lookup table BB. Otherwise, check if the condition value is within the case
3771 // range. If it is so, branch to the new BB. Otherwise branch to SI's default
3773 const bool GeneratingCoveredLookupTable = MaxTableSize == TableSize;
3774 if (GeneratingCoveredLookupTable) {
3775 Builder.CreateBr(LookupBB);
3776 SI->getDefaultDest()->removePredecessor(SI->getParent());
3778 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
3779 MinCaseVal->getType(), TableSize));
3780 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
3783 // Populate the BB that does the lookups.
3784 Builder.SetInsertPoint(LookupBB);
3785 bool ReturnedEarly = false;
3786 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
3787 PHINode *PHI = PHIs[I];
3789 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI],
3790 DefaultResults[PHI], TD);
3792 Value *Result = Table.BuildLookup(TableIndex, Builder);
3794 // If the result is used to return immediately from the function, we want to
3795 // do that right here.
3796 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin()) &&
3797 *PHI->use_begin() == CommonDest->getFirstNonPHIOrDbg()) {
3798 Builder.CreateRet(Result);
3799 ReturnedEarly = true;
3803 PHI->addIncoming(Result, LookupBB);
3807 Builder.CreateBr(CommonDest);
3809 // Remove the switch.
3810 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
3811 BasicBlock *Succ = SI->getSuccessor(i);
3813 if (Succ == SI->getDefaultDest())
3815 Succ->removePredecessor(SI->getParent());
3817 SI->eraseFromParent();
3823 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
3824 BasicBlock *BB = SI->getParent();
3826 if (isValueEqualityComparison(SI)) {
3827 // If we only have one predecessor, and if it is a branch on this value,
3828 // see if that predecessor totally determines the outcome of this switch.
3829 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3830 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
3831 return SimplifyCFG(BB, TTI, TD) | true;
3833 Value *Cond = SI->getCondition();
3834 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
3835 if (SimplifySwitchOnSelect(SI, Select))
3836 return SimplifyCFG(BB, TTI, TD) | true;
3838 // If the block only contains the switch, see if we can fold the block
3839 // away into any preds.
3840 BasicBlock::iterator BBI = BB->begin();
3841 // Ignore dbg intrinsics.
3842 while (isa<DbgInfoIntrinsic>(BBI))
3845 if (FoldValueComparisonIntoPredecessors(SI, Builder))
3846 return SimplifyCFG(BB, TTI, TD) | true;
3849 // Try to transform the switch into an icmp and a branch.
3850 if (TurnSwitchRangeIntoICmp(SI, Builder))
3851 return SimplifyCFG(BB, TTI, TD) | true;
3853 // Remove unreachable cases.
3854 if (EliminateDeadSwitchCases(SI))
3855 return SimplifyCFG(BB, TTI, TD) | true;
3857 if (ForwardSwitchConditionToPHI(SI))
3858 return SimplifyCFG(BB, TTI, TD) | true;
3860 if (SwitchToLookupTable(SI, Builder, TTI, TD))
3861 return SimplifyCFG(BB, TTI, TD) | true;
3866 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
3867 BasicBlock *BB = IBI->getParent();
3868 bool Changed = false;
3870 // Eliminate redundant destinations.
3871 SmallPtrSet<Value *, 8> Succs;
3872 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
3873 BasicBlock *Dest = IBI->getDestination(i);
3874 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) {
3875 Dest->removePredecessor(BB);
3876 IBI->removeDestination(i);
3882 if (IBI->getNumDestinations() == 0) {
3883 // If the indirectbr has no successors, change it to unreachable.
3884 new UnreachableInst(IBI->getContext(), IBI);
3885 EraseTerminatorInstAndDCECond(IBI);
3889 if (IBI->getNumDestinations() == 1) {
3890 // If the indirectbr has one successor, change it to a direct branch.
3891 BranchInst::Create(IBI->getDestination(0), IBI);
3892 EraseTerminatorInstAndDCECond(IBI);
3896 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
3897 if (SimplifyIndirectBrOnSelect(IBI, SI))
3898 return SimplifyCFG(BB, TTI, TD) | true;
3903 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
3904 BasicBlock *BB = BI->getParent();
3906 if (SinkCommon && SinkThenElseCodeToEnd(BI))
3909 // If the Terminator is the only non-phi instruction, simplify the block.
3910 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime();
3911 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
3912 TryToSimplifyUncondBranchFromEmptyBlock(BB))
3915 // If the only instruction in the block is a seteq/setne comparison
3916 // against a constant, try to simplify the block.
3917 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
3918 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
3919 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
3921 if (I->isTerminator() &&
3922 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI, TD))
3926 // If this basic block is ONLY a compare and a branch, and if a predecessor
3927 // branches to us and our successor, fold the comparison into the
3928 // predecessor and use logical operations to update the incoming value
3929 // for PHI nodes in common successor.
3930 if (FoldBranchToCommonDest(BI))
3931 return SimplifyCFG(BB, TTI, TD) | true;
3936 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
3937 BasicBlock *BB = BI->getParent();
3939 // Conditional branch
3940 if (isValueEqualityComparison(BI)) {
3941 // If we only have one predecessor, and if it is a branch on this value,
3942 // see if that predecessor totally determines the outcome of this
3944 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3945 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
3946 return SimplifyCFG(BB, TTI, TD) | true;
3948 // This block must be empty, except for the setcond inst, if it exists.
3949 // Ignore dbg intrinsics.
3950 BasicBlock::iterator I = BB->begin();
3951 // Ignore dbg intrinsics.
3952 while (isa<DbgInfoIntrinsic>(I))
3955 if (FoldValueComparisonIntoPredecessors(BI, Builder))
3956 return SimplifyCFG(BB, TTI, TD) | true;
3957 } else if (&*I == cast<Instruction>(BI->getCondition())){
3959 // Ignore dbg intrinsics.
3960 while (isa<DbgInfoIntrinsic>(I))
3962 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
3963 return SimplifyCFG(BB, TTI, TD) | true;
3967 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
3968 if (SimplifyBranchOnICmpChain(BI, TD, Builder))
3971 // If this basic block is ONLY a compare and a branch, and if a predecessor
3972 // branches to us and one of our successors, fold the comparison into the
3973 // predecessor and use logical operations to pick the right destination.
3974 if (FoldBranchToCommonDest(BI))
3975 return SimplifyCFG(BB, TTI, TD) | true;
3977 // We have a conditional branch to two blocks that are only reachable
3978 // from BI. We know that the condbr dominates the two blocks, so see if
3979 // there is any identical code in the "then" and "else" blocks. If so, we
3980 // can hoist it up to the branching block.
3981 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) {
3982 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3983 if (HoistThenElseCodeToIf(BI))
3984 return SimplifyCFG(BB, TTI, TD) | true;
3986 // If Successor #1 has multiple preds, we may be able to conditionally
3987 // execute Successor #0 if it branches to successor #1.
3988 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
3989 if (Succ0TI->getNumSuccessors() == 1 &&
3990 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
3991 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0)))
3992 return SimplifyCFG(BB, TTI, TD) | true;
3994 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3995 // If Successor #0 has multiple preds, we may be able to conditionally
3996 // execute Successor #1 if it branches to successor #0.
3997 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
3998 if (Succ1TI->getNumSuccessors() == 1 &&
3999 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
4000 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1)))
4001 return SimplifyCFG(BB, TTI, TD) | true;
4004 // If this is a branch on a phi node in the current block, thread control
4005 // through this block if any PHI node entries are constants.
4006 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
4007 if (PN->getParent() == BI->getParent())
4008 if (FoldCondBranchOnPHI(BI, TD))
4009 return SimplifyCFG(BB, TTI, TD) | true;
4011 // Scan predecessor blocks for conditional branches.
4012 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
4013 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
4014 if (PBI != BI && PBI->isConditional())
4015 if (SimplifyCondBranchToCondBranch(PBI, BI))
4016 return SimplifyCFG(BB, TTI, TD) | true;
4021 /// Check if passing a value to an instruction will cause undefined behavior.
4022 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
4023 Constant *C = dyn_cast<Constant>(V);
4030 if (C->isNullValue()) {
4031 // Only look at the first use, avoid hurting compile time with long uselists
4032 User *Use = *I->use_begin();
4034 // Now make sure that there are no instructions in between that can alter
4035 // control flow (eg. calls)
4036 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
4037 if (i == I->getParent()->end() || i->mayHaveSideEffects())
4040 // Look through GEPs. A load from a GEP derived from NULL is still undefined
4041 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
4042 if (GEP->getPointerOperand() == I)
4043 return passingValueIsAlwaysUndefined(V, GEP);
4045 // Look through bitcasts.
4046 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
4047 return passingValueIsAlwaysUndefined(V, BC);
4049 // Load from null is undefined.
4050 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
4051 if (!LI->isVolatile())
4052 return LI->getPointerAddressSpace() == 0;
4054 // Store to null is undefined.
4055 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
4056 if (!SI->isVolatile())
4057 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
4062 /// If BB has an incoming value that will always trigger undefined behavior
4063 /// (eg. null pointer dereference), remove the branch leading here.
4064 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
4065 for (BasicBlock::iterator i = BB->begin();
4066 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
4067 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
4068 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
4069 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
4070 IRBuilder<> Builder(T);
4071 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
4072 BB->removePredecessor(PHI->getIncomingBlock(i));
4073 // Turn uncoditional branches into unreachables and remove the dead
4074 // destination from conditional branches.
4075 if (BI->isUnconditional())
4076 Builder.CreateUnreachable();
4078 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
4079 BI->getSuccessor(0));
4080 BI->eraseFromParent();
4083 // TODO: SwitchInst.
4089 bool SimplifyCFGOpt::run(BasicBlock *BB) {
4090 bool Changed = false;
4092 assert(BB && BB->getParent() && "Block not embedded in function!");
4093 assert(BB->getTerminator() && "Degenerate basic block encountered!");
4095 // Remove basic blocks that have no predecessors (except the entry block)...
4096 // or that just have themself as a predecessor. These are unreachable.
4097 if ((pred_begin(BB) == pred_end(BB) &&
4098 BB != &BB->getParent()->getEntryBlock()) ||
4099 BB->getSinglePredecessor() == BB) {
4100 DEBUG(dbgs() << "Removing BB: \n" << *BB);
4101 DeleteDeadBlock(BB);
4105 // Check to see if we can constant propagate this terminator instruction
4107 Changed |= ConstantFoldTerminator(BB, true);
4109 // Check for and eliminate duplicate PHI nodes in this block.
4110 Changed |= EliminateDuplicatePHINodes(BB);
4112 // Check for and remove branches that will always cause undefined behavior.
4113 Changed |= removeUndefIntroducingPredecessor(BB);
4115 // Merge basic blocks into their predecessor if there is only one distinct
4116 // pred, and if there is only one distinct successor of the predecessor, and
4117 // if there are no PHI nodes.
4119 if (MergeBlockIntoPredecessor(BB))
4122 IRBuilder<> Builder(BB);
4124 // If there is a trivial two-entry PHI node in this basic block, and we can
4125 // eliminate it, do so now.
4126 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
4127 if (PN->getNumIncomingValues() == 2)
4128 Changed |= FoldTwoEntryPHINode(PN, TD);
4130 Builder.SetInsertPoint(BB->getTerminator());
4131 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
4132 if (BI->isUnconditional()) {
4133 if (SimplifyUncondBranch(BI, Builder)) return true;
4135 if (SimplifyCondBranch(BI, Builder)) return true;
4137 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
4138 if (SimplifyReturn(RI, Builder)) return true;
4139 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
4140 if (SimplifyResume(RI, Builder)) return true;
4141 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
4142 if (SimplifySwitch(SI, Builder)) return true;
4143 } else if (UnreachableInst *UI =
4144 dyn_cast<UnreachableInst>(BB->getTerminator())) {
4145 if (SimplifyUnreachable(UI)) return true;
4146 } else if (IndirectBrInst *IBI =
4147 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
4148 if (SimplifyIndirectBr(IBI)) return true;
4154 /// SimplifyCFG - This function is used to do simplification of a CFG. For
4155 /// example, it adjusts branches to branches to eliminate the extra hop, it
4156 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4157 /// of the CFG. It returns true if a modification was made.
4159 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
4160 const DataLayout *TD) {
4161 return SimplifyCFGOpt(TTI, TD).run(BB);