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 #include "llvm/Transforms/Utils/Local.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/ADT/SetVector.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/ConstantFolding.h"
22 #include "llvm/Analysis/InstructionSimplify.h"
23 #include "llvm/Analysis/TargetTransformInfo.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/IR/CFG.h"
26 #include "llvm/IR/ConstantRange.h"
27 #include "llvm/IR/Constants.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/DerivedTypes.h"
30 #include "llvm/IR/GlobalVariable.h"
31 #include "llvm/IR/IRBuilder.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/LLVMContext.h"
35 #include "llvm/IR/MDBuilder.h"
36 #include "llvm/IR/Metadata.h"
37 #include "llvm/IR/Module.h"
38 #include "llvm/IR/NoFolder.h"
39 #include "llvm/IR/Operator.h"
40 #include "llvm/IR/PatternMatch.h"
41 #include "llvm/IR/Type.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/Debug.h"
44 #include "llvm/Support/raw_ostream.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Transforms/Utils/ValueMapper.h"
52 using namespace PatternMatch;
54 #define DEBUG_TYPE "simplifycfg"
56 static cl::opt<unsigned>
57 PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1),
58 cl::desc("Control the amount of phi node folding to perform (default = 1)"));
61 DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false),
62 cl::desc("Duplicate return instructions into unconditional branches"));
65 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
66 cl::desc("Sink common instructions down to the end block"));
68 static cl::opt<bool> HoistCondStores(
69 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
70 cl::desc("Hoist conditional stores if an unconditional store precedes"));
72 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
73 STATISTIC(NumLinearMaps, "Number of switch instructions turned into linear mapping");
74 STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables");
75 STATISTIC(NumLookupTablesHoles, "Number of switch instructions turned into lookup tables (holes checked)");
76 STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block");
77 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
80 // The first field contains the value that the switch produces when a certain
81 // case group is selected, and the second field is a vector containing the cases
82 // composing the case group.
83 typedef SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>
84 SwitchCaseResultVectorTy;
85 // The first field contains the phi node that generates a result of the switch
86 // and the second field contains the value generated for a certain case in the switch
88 typedef SmallVector<std::pair<PHINode *, Constant *>, 4> SwitchCaseResultsTy;
90 /// ValueEqualityComparisonCase - Represents a case of a switch.
91 struct ValueEqualityComparisonCase {
95 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
96 : Value(Value), Dest(Dest) {}
98 bool operator<(ValueEqualityComparisonCase RHS) const {
99 // Comparing pointers is ok as we only rely on the order for uniquing.
100 return Value < RHS.Value;
103 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
106 class SimplifyCFGOpt {
107 const TargetTransformInfo &TTI;
108 unsigned BonusInstThreshold;
109 const DataLayout *const DL;
110 AssumptionTracker *AT;
111 Value *isValueEqualityComparison(TerminatorInst *TI);
112 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
113 std::vector<ValueEqualityComparisonCase> &Cases);
114 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
116 IRBuilder<> &Builder);
117 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
118 IRBuilder<> &Builder);
120 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
121 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
122 bool SimplifyUnreachable(UnreachableInst *UI);
123 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
124 bool SimplifyIndirectBr(IndirectBrInst *IBI);
125 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder);
126 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder);
129 SimplifyCFGOpt(const TargetTransformInfo &TTI, unsigned BonusInstThreshold,
130 const DataLayout *DL, AssumptionTracker *AT)
131 : TTI(TTI), BonusInstThreshold(BonusInstThreshold), DL(DL), AT(AT) {}
132 bool run(BasicBlock *BB);
136 /// SafeToMergeTerminators - Return true if it is safe to merge these two
137 /// terminator instructions together.
139 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
140 if (SI1 == SI2) return false; // Can't merge with self!
142 // It is not safe to merge these two switch instructions if they have a common
143 // successor, and if that successor has a PHI node, and if *that* PHI node has
144 // conflicting incoming values from the two switch blocks.
145 BasicBlock *SI1BB = SI1->getParent();
146 BasicBlock *SI2BB = SI2->getParent();
147 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
149 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
150 if (SI1Succs.count(*I))
151 for (BasicBlock::iterator BBI = (*I)->begin();
152 isa<PHINode>(BBI); ++BBI) {
153 PHINode *PN = cast<PHINode>(BBI);
154 if (PN->getIncomingValueForBlock(SI1BB) !=
155 PN->getIncomingValueForBlock(SI2BB))
162 /// isProfitableToFoldUnconditional - Return true if it is safe and profitable
163 /// to merge these two terminator instructions together, where SI1 is an
164 /// unconditional branch. PhiNodes will store all PHI nodes in common
167 static bool isProfitableToFoldUnconditional(BranchInst *SI1,
170 SmallVectorImpl<PHINode*> &PhiNodes) {
171 if (SI1 == SI2) return false; // Can't merge with self!
172 assert(SI1->isUnconditional() && SI2->isConditional());
174 // We fold the unconditional branch if we can easily update all PHI nodes in
175 // common successors:
176 // 1> We have a constant incoming value for the conditional branch;
177 // 2> We have "Cond" as the incoming value for the unconditional branch;
178 // 3> SI2->getCondition() and Cond have same operands.
179 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
180 if (!Ci2) return false;
181 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
182 Cond->getOperand(1) == Ci2->getOperand(1)) &&
183 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
184 Cond->getOperand(1) == Ci2->getOperand(0)))
187 BasicBlock *SI1BB = SI1->getParent();
188 BasicBlock *SI2BB = SI2->getParent();
189 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
190 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
191 if (SI1Succs.count(*I))
192 for (BasicBlock::iterator BBI = (*I)->begin();
193 isa<PHINode>(BBI); ++BBI) {
194 PHINode *PN = cast<PHINode>(BBI);
195 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
196 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
198 PhiNodes.push_back(PN);
203 /// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will
204 /// now be entries in it from the 'NewPred' block. The values that will be
205 /// flowing into the PHI nodes will be the same as those coming in from
206 /// ExistPred, an existing predecessor of Succ.
207 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
208 BasicBlock *ExistPred) {
209 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
212 for (BasicBlock::iterator I = Succ->begin();
213 (PN = dyn_cast<PHINode>(I)); ++I)
214 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
217 /// ComputeSpeculationCost - Compute an abstract "cost" of speculating the
218 /// given instruction, which is assumed to be safe to speculate. 1 means
219 /// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive.
220 static unsigned ComputeSpeculationCost(const User *I, const DataLayout *DL) {
221 assert(isSafeToSpeculativelyExecute(I, DL) &&
222 "Instruction is not safe to speculatively execute!");
223 switch (Operator::getOpcode(I)) {
225 // In doubt, be conservative.
227 case Instruction::GetElementPtr:
228 // GEPs are cheap if all indices are constant.
229 if (!cast<GEPOperator>(I)->hasAllConstantIndices())
232 case Instruction::ExtractValue:
233 case Instruction::Load:
234 case Instruction::Add:
235 case Instruction::Sub:
236 case Instruction::And:
237 case Instruction::Or:
238 case Instruction::Xor:
239 case Instruction::Shl:
240 case Instruction::LShr:
241 case Instruction::AShr:
242 case Instruction::ICmp:
243 case Instruction::Trunc:
244 case Instruction::ZExt:
245 case Instruction::SExt:
246 case Instruction::BitCast:
247 case Instruction::ExtractElement:
248 case Instruction::InsertElement:
249 return 1; // These are all cheap.
251 case Instruction::Call:
252 case Instruction::Select:
257 /// DominatesMergePoint - If we have a merge point of an "if condition" as
258 /// accepted above, return true if the specified value dominates the block. We
259 /// don't handle the true generality of domination here, just a special case
260 /// which works well enough for us.
262 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
263 /// see if V (which must be an instruction) and its recursive operands
264 /// that do not dominate BB have a combined cost lower than CostRemaining and
265 /// are non-trapping. If both are true, the instruction is inserted into the
266 /// set and true is returned.
268 /// The cost for most non-trapping instructions is defined as 1 except for
269 /// Select whose cost is 2.
271 /// After this function returns, CostRemaining is decreased by the cost of
272 /// V plus its non-dominating operands. If that cost is greater than
273 /// CostRemaining, false is returned and CostRemaining is undefined.
274 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
275 SmallPtrSetImpl<Instruction*> *AggressiveInsts,
276 unsigned &CostRemaining,
277 const DataLayout *DL) {
278 Instruction *I = dyn_cast<Instruction>(V);
280 // Non-instructions all dominate instructions, but not all constantexprs
281 // can be executed unconditionally.
282 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
287 BasicBlock *PBB = I->getParent();
289 // We don't want to allow weird loops that might have the "if condition" in
290 // the bottom of this block.
291 if (PBB == BB) return false;
293 // If this instruction is defined in a block that contains an unconditional
294 // branch to BB, then it must be in the 'conditional' part of the "if
295 // statement". If not, it definitely dominates the region.
296 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
297 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
300 // If we aren't allowing aggressive promotion anymore, then don't consider
301 // instructions in the 'if region'.
302 if (!AggressiveInsts) return false;
304 // If we have seen this instruction before, don't count it again.
305 if (AggressiveInsts->count(I)) return true;
307 // Okay, it looks like the instruction IS in the "condition". Check to
308 // see if it's a cheap instruction to unconditionally compute, and if it
309 // only uses stuff defined outside of the condition. If so, hoist it out.
310 if (!isSafeToSpeculativelyExecute(I, DL))
313 unsigned Cost = ComputeSpeculationCost(I, DL);
315 if (Cost > CostRemaining)
318 CostRemaining -= Cost;
320 // Okay, we can only really hoist these out if their operands do
321 // not take us over the cost threshold.
322 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
323 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, DL))
325 // Okay, it's safe to do this! Remember this instruction.
326 AggressiveInsts->insert(I);
330 /// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr
331 /// and PointerNullValue. Return NULL if value is not a constant int.
332 static ConstantInt *GetConstantInt(Value *V, const DataLayout *DL) {
333 // Normal constant int.
334 ConstantInt *CI = dyn_cast<ConstantInt>(V);
335 if (CI || !DL || !isa<Constant>(V) || !V->getType()->isPointerTy())
338 // This is some kind of pointer constant. Turn it into a pointer-sized
339 // ConstantInt if possible.
340 IntegerType *PtrTy = cast<IntegerType>(DL->getIntPtrType(V->getType()));
342 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
343 if (isa<ConstantPointerNull>(V))
344 return ConstantInt::get(PtrTy, 0);
346 // IntToPtr const int.
347 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
348 if (CE->getOpcode() == Instruction::IntToPtr)
349 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
350 // The constant is very likely to have the right type already.
351 if (CI->getType() == PtrTy)
354 return cast<ConstantInt>
355 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
360 /// Given a chain of or (||) or and (&&) comparison of a value against a
361 /// constant, this will try to recover the information required for a switch
363 /// It will depth-first traverse the chain of comparison, seeking for patterns
364 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
365 /// representing the different cases for the switch.
366 /// Note that if the chain is composed of '||' it will build the set of elements
367 /// that matches the comparisons (i.e. any of this value validate the chain)
368 /// while for a chain of '&&' it will build the set elements that make the test
370 struct ConstantComparesGatherer {
372 Value *CompValue; /// Value found for the switch comparison
373 Value *Extra; /// Extra clause to be checked before the switch
374 SmallVector<ConstantInt *, 8> Vals; /// Set of integers to match in switch
375 unsigned UsedICmps; /// Number of comparisons matched in the and/or chain
377 /// Construct and compute the result for the comparison instruction Cond
378 ConstantComparesGatherer(Instruction *Cond, const DataLayout *DL)
379 : CompValue(nullptr), Extra(nullptr), UsedICmps(0) {
384 ConstantComparesGatherer(const ConstantComparesGatherer &)
385 LLVM_DELETED_FUNCTION;
386 ConstantComparesGatherer &
387 operator=(const ConstantComparesGatherer &) LLVM_DELETED_FUNCTION;
391 /// Try to set the current value used for the comparison, it succeeds only if
392 /// it wasn't set before or if the new value is the same as the old one
393 bool setValueOnce(Value *NewVal) {
394 if(CompValue && CompValue != NewVal) return false;
395 return CompValue == NewVal;
398 /// Try to match Instruction "I" as a comparison against a constant and
399 /// populates the array Vals with the set of values that match (or do not
400 /// match depending on isEQ).
401 /// Return false on failure. On success, the Value the comparison matched
402 /// against is placed in CompValue.
403 /// If CompValue is already set, the function is expected to fail if a match
404 /// is found but the value compared to is different.
405 bool matchInstruction(Instruction *I, const DataLayout *DL, bool isEQ) {
406 // If this is an icmp against a constant, handle this as one of the cases.
409 if (!((ICI = dyn_cast<ICmpInst>(I)) &&
410 (C = GetConstantInt(I->getOperand(1), DL)))) {
417 // Pattern match a special case
418 // (x & ~2^x) == y --> x == y || x == y|2^x
419 // This undoes a transformation done by instcombine to fuse 2 compares.
420 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
421 if (match(ICI->getOperand(0),
422 m_And(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
423 APInt Not = ~RHSC->getValue();
424 if (Not.isPowerOf2()) {
425 // If we already have a value for the switch, it has to match!
426 if(!setValueOnce(RHSVal))
430 Vals.push_back(ConstantInt::get(C->getContext(),
431 C->getValue() | Not));
437 // If we already have a value for the switch, it has to match!
438 if(!setValueOnce(ICI->getOperand(0)))
443 return ICI->getOperand(0);
446 // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
447 ConstantRange Span = ConstantRange::makeICmpRegion(ICI->getPredicate(),
450 // Shift the range if the compare is fed by an add. This is the range
451 // compare idiom as emitted by instcombine.
452 Value *CandidateVal = I->getOperand(0);
453 if(match(I->getOperand(0), m_Add(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
454 Span = Span.subtract(RHSC->getValue());
455 CandidateVal = RHSVal;
458 // If this is an and/!= check, then we are looking to build the set of
459 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
462 Span = Span.inverse();
464 // If there are a ton of values, we don't want to make a ginormous switch.
465 if (Span.getSetSize().ugt(8) || Span.isEmptySet()) {
469 // If we already have a value for the switch, it has to match!
470 if(!setValueOnce(CandidateVal))
473 // Add all values from the range to the set
474 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
475 Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
482 /// gather - Given a potentially 'or'd or 'and'd together collection of icmp
483 /// eq/ne/lt/gt instructions that compare a value against a constant, extract
484 /// the value being compared, and stick the list constants into the Vals
486 /// One "Extra" case is allowed to differ from the other.
487 void gather(Value *V, const DataLayout *DL) {
488 Instruction *I = dyn_cast<Instruction>(V);
489 bool isEQ = (I->getOpcode() == Instruction::Or);
491 // Keep a stack (SmallVector for efficiency) for depth-first traversal
492 SmallVector<Value *, 8> DFT;
497 while(!DFT.empty()) {
498 V = DFT.pop_back_val();
500 if (Instruction *I = dyn_cast<Instruction>(V)) {
501 // If it is a || (or && depending on isEQ), process the operands.
502 if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
503 DFT.push_back(I->getOperand(1));
504 DFT.push_back(I->getOperand(0));
508 // Try to match the current instruction
509 if (matchInstruction(I, DL, isEQ))
510 // Match succeed, continue the loop
514 // One element of the sequence of || (or &&) could not be match as a
515 // comparison against the same value as the others.
516 // We allow only one "Extra" case to be checked before the switch
521 // Failed to parse a proper sequence, abort now
528 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
529 Instruction *Cond = nullptr;
530 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
531 Cond = dyn_cast<Instruction>(SI->getCondition());
532 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
533 if (BI->isConditional())
534 Cond = dyn_cast<Instruction>(BI->getCondition());
535 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
536 Cond = dyn_cast<Instruction>(IBI->getAddress());
539 TI->eraseFromParent();
540 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
543 /// isValueEqualityComparison - Return true if the specified terminator checks
544 /// to see if a value is equal to constant integer value.
545 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
547 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
548 // Do not permit merging of large switch instructions into their
549 // predecessors unless there is only one predecessor.
550 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()),
551 pred_end(SI->getParent())) <= 128)
552 CV = SI->getCondition();
553 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
554 if (BI->isConditional() && BI->getCondition()->hasOneUse())
555 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition()))
556 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
557 CV = ICI->getOperand(0);
559 // Unwrap any lossless ptrtoint cast.
561 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
562 Value *Ptr = PTII->getPointerOperand();
563 if (PTII->getType() == DL->getIntPtrType(Ptr->getType()))
570 /// GetValueEqualityComparisonCases - Given a value comparison instruction,
571 /// decode all of the 'cases' that it represents and return the 'default' block.
572 BasicBlock *SimplifyCFGOpt::
573 GetValueEqualityComparisonCases(TerminatorInst *TI,
574 std::vector<ValueEqualityComparisonCase>
576 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
577 Cases.reserve(SI->getNumCases());
578 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
579 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
580 i.getCaseSuccessor()));
581 return SI->getDefaultDest();
584 BranchInst *BI = cast<BranchInst>(TI);
585 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
586 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
587 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
590 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
594 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries
595 /// in the list that match the specified block.
596 static void EliminateBlockCases(BasicBlock *BB,
597 std::vector<ValueEqualityComparisonCase> &Cases) {
598 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
601 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as
604 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
605 std::vector<ValueEqualityComparisonCase > &C2) {
606 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
608 // Make V1 be smaller than V2.
609 if (V1->size() > V2->size())
612 if (V1->size() == 0) return false;
613 if (V1->size() == 1) {
615 ConstantInt *TheVal = (*V1)[0].Value;
616 for (unsigned i = 0, e = V2->size(); i != e; ++i)
617 if (TheVal == (*V2)[i].Value)
621 // Otherwise, just sort both lists and compare element by element.
622 array_pod_sort(V1->begin(), V1->end());
623 array_pod_sort(V2->begin(), V2->end());
624 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
625 while (i1 != e1 && i2 != e2) {
626 if ((*V1)[i1].Value == (*V2)[i2].Value)
628 if ((*V1)[i1].Value < (*V2)[i2].Value)
636 /// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a
637 /// terminator instruction and its block is known to only have a single
638 /// predecessor block, check to see if that predecessor is also a value
639 /// comparison with the same value, and if that comparison determines the
640 /// outcome of this comparison. If so, simplify TI. This does a very limited
641 /// form of jump threading.
642 bool SimplifyCFGOpt::
643 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
645 IRBuilder<> &Builder) {
646 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
647 if (!PredVal) return false; // Not a value comparison in predecessor.
649 Value *ThisVal = isValueEqualityComparison(TI);
650 assert(ThisVal && "This isn't a value comparison!!");
651 if (ThisVal != PredVal) return false; // Different predicates.
653 // TODO: Preserve branch weight metadata, similarly to how
654 // FoldValueComparisonIntoPredecessors preserves it.
656 // Find out information about when control will move from Pred to TI's block.
657 std::vector<ValueEqualityComparisonCase> PredCases;
658 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
660 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
662 // Find information about how control leaves this block.
663 std::vector<ValueEqualityComparisonCase> ThisCases;
664 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
665 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
667 // If TI's block is the default block from Pred's comparison, potentially
668 // simplify TI based on this knowledge.
669 if (PredDef == TI->getParent()) {
670 // If we are here, we know that the value is none of those cases listed in
671 // PredCases. If there are any cases in ThisCases that are in PredCases, we
673 if (!ValuesOverlap(PredCases, ThisCases))
676 if (isa<BranchInst>(TI)) {
677 // Okay, one of the successors of this condbr is dead. Convert it to a
679 assert(ThisCases.size() == 1 && "Branch can only have one case!");
680 // Insert the new branch.
681 Instruction *NI = Builder.CreateBr(ThisDef);
684 // Remove PHI node entries for the dead edge.
685 ThisCases[0].Dest->removePredecessor(TI->getParent());
687 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
688 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
690 EraseTerminatorInstAndDCECond(TI);
694 SwitchInst *SI = cast<SwitchInst>(TI);
695 // Okay, TI has cases that are statically dead, prune them away.
696 SmallPtrSet<Constant*, 16> DeadCases;
697 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
698 DeadCases.insert(PredCases[i].Value);
700 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
701 << "Through successor TI: " << *TI);
703 // Collect branch weights into a vector.
704 SmallVector<uint32_t, 8> Weights;
705 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
706 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
708 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
710 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i));
712 Weights.push_back(CI->getValue().getZExtValue());
714 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
716 if (DeadCases.count(i.getCaseValue())) {
718 std::swap(Weights[i.getCaseIndex()+1], Weights.back());
721 i.getCaseSuccessor()->removePredecessor(TI->getParent());
725 if (HasWeight && Weights.size() >= 2)
726 SI->setMetadata(LLVMContext::MD_prof,
727 MDBuilder(SI->getParent()->getContext()).
728 createBranchWeights(Weights));
730 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
734 // Otherwise, TI's block must correspond to some matched value. Find out
735 // which value (or set of values) this is.
736 ConstantInt *TIV = nullptr;
737 BasicBlock *TIBB = TI->getParent();
738 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
739 if (PredCases[i].Dest == TIBB) {
741 return false; // Cannot handle multiple values coming to this block.
742 TIV = PredCases[i].Value;
744 assert(TIV && "No edge from pred to succ?");
746 // Okay, we found the one constant that our value can be if we get into TI's
747 // BB. Find out which successor will unconditionally be branched to.
748 BasicBlock *TheRealDest = nullptr;
749 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
750 if (ThisCases[i].Value == TIV) {
751 TheRealDest = ThisCases[i].Dest;
755 // If not handled by any explicit cases, it is handled by the default case.
756 if (!TheRealDest) TheRealDest = ThisDef;
758 // Remove PHI node entries for dead edges.
759 BasicBlock *CheckEdge = TheRealDest;
760 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
761 if (*SI != CheckEdge)
762 (*SI)->removePredecessor(TIBB);
766 // Insert the new branch.
767 Instruction *NI = Builder.CreateBr(TheRealDest);
770 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
771 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
773 EraseTerminatorInstAndDCECond(TI);
778 /// ConstantIntOrdering - This class implements a stable ordering of constant
779 /// integers that does not depend on their address. This is important for
780 /// applications that sort ConstantInt's to ensure uniqueness.
781 struct ConstantIntOrdering {
782 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
783 return LHS->getValue().ult(RHS->getValue());
788 static int ConstantIntSortPredicate(ConstantInt *const *P1,
789 ConstantInt *const *P2) {
790 const ConstantInt *LHS = *P1;
791 const ConstantInt *RHS = *P2;
792 if (LHS->getValue().ult(RHS->getValue()))
794 if (LHS->getValue() == RHS->getValue())
799 static inline bool HasBranchWeights(const Instruction* I) {
800 MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
801 if (ProfMD && ProfMD->getOperand(0))
802 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
803 return MDS->getString().equals("branch_weights");
808 /// Get Weights of a given TerminatorInst, the default weight is at the front
809 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
811 static void GetBranchWeights(TerminatorInst *TI,
812 SmallVectorImpl<uint64_t> &Weights) {
813 MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
815 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
816 ConstantInt *CI = cast<ConstantInt>(MD->getOperand(i));
817 Weights.push_back(CI->getValue().getZExtValue());
820 // If TI is a conditional eq, the default case is the false case,
821 // and the corresponding branch-weight data is at index 2. We swap the
822 // default weight to be the first entry.
823 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
824 assert(Weights.size() == 2);
825 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
826 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
827 std::swap(Weights.front(), Weights.back());
831 /// Keep halving the weights until all can fit in uint32_t.
832 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
833 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
834 if (Max > UINT_MAX) {
835 unsigned Offset = 32 - countLeadingZeros(Max);
836 for (uint64_t &I : Weights)
841 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value
842 /// equality comparison instruction (either a switch or a branch on "X == c").
843 /// See if any of the predecessors of the terminator block are value comparisons
844 /// on the same value. If so, and if safe to do so, fold them together.
845 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
846 IRBuilder<> &Builder) {
847 BasicBlock *BB = TI->getParent();
848 Value *CV = isValueEqualityComparison(TI); // CondVal
849 assert(CV && "Not a comparison?");
850 bool Changed = false;
852 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
853 while (!Preds.empty()) {
854 BasicBlock *Pred = Preds.pop_back_val();
856 // See if the predecessor is a comparison with the same value.
857 TerminatorInst *PTI = Pred->getTerminator();
858 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
860 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
861 // Figure out which 'cases' to copy from SI to PSI.
862 std::vector<ValueEqualityComparisonCase> BBCases;
863 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
865 std::vector<ValueEqualityComparisonCase> PredCases;
866 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
868 // Based on whether the default edge from PTI goes to BB or not, fill in
869 // PredCases and PredDefault with the new switch cases we would like to
871 SmallVector<BasicBlock*, 8> NewSuccessors;
873 // Update the branch weight metadata along the way
874 SmallVector<uint64_t, 8> Weights;
875 bool PredHasWeights = HasBranchWeights(PTI);
876 bool SuccHasWeights = HasBranchWeights(TI);
878 if (PredHasWeights) {
879 GetBranchWeights(PTI, Weights);
880 // branch-weight metadata is inconsistent here.
881 if (Weights.size() != 1 + PredCases.size())
882 PredHasWeights = SuccHasWeights = false;
883 } else if (SuccHasWeights)
884 // If there are no predecessor weights but there are successor weights,
885 // populate Weights with 1, which will later be scaled to the sum of
886 // successor's weights
887 Weights.assign(1 + PredCases.size(), 1);
889 SmallVector<uint64_t, 8> SuccWeights;
890 if (SuccHasWeights) {
891 GetBranchWeights(TI, SuccWeights);
892 // branch-weight metadata is inconsistent here.
893 if (SuccWeights.size() != 1 + BBCases.size())
894 PredHasWeights = SuccHasWeights = false;
895 } else if (PredHasWeights)
896 SuccWeights.assign(1 + BBCases.size(), 1);
898 if (PredDefault == BB) {
899 // If this is the default destination from PTI, only the edges in TI
900 // that don't occur in PTI, or that branch to BB will be activated.
901 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
902 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
903 if (PredCases[i].Dest != BB)
904 PTIHandled.insert(PredCases[i].Value);
906 // The default destination is BB, we don't need explicit targets.
907 std::swap(PredCases[i], PredCases.back());
909 if (PredHasWeights || SuccHasWeights) {
910 // Increase weight for the default case.
911 Weights[0] += Weights[i+1];
912 std::swap(Weights[i+1], Weights.back());
916 PredCases.pop_back();
920 // Reconstruct the new switch statement we will be building.
921 if (PredDefault != BBDefault) {
922 PredDefault->removePredecessor(Pred);
923 PredDefault = BBDefault;
924 NewSuccessors.push_back(BBDefault);
927 unsigned CasesFromPred = Weights.size();
928 uint64_t ValidTotalSuccWeight = 0;
929 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
930 if (!PTIHandled.count(BBCases[i].Value) &&
931 BBCases[i].Dest != BBDefault) {
932 PredCases.push_back(BBCases[i]);
933 NewSuccessors.push_back(BBCases[i].Dest);
934 if (SuccHasWeights || PredHasWeights) {
935 // The default weight is at index 0, so weight for the ith case
936 // should be at index i+1. Scale the cases from successor by
937 // PredDefaultWeight (Weights[0]).
938 Weights.push_back(Weights[0] * SuccWeights[i+1]);
939 ValidTotalSuccWeight += SuccWeights[i+1];
943 if (SuccHasWeights || PredHasWeights) {
944 ValidTotalSuccWeight += SuccWeights[0];
945 // Scale the cases from predecessor by ValidTotalSuccWeight.
946 for (unsigned i = 1; i < CasesFromPred; ++i)
947 Weights[i] *= ValidTotalSuccWeight;
948 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
949 Weights[0] *= SuccWeights[0];
952 // If this is not the default destination from PSI, only the edges
953 // in SI that occur in PSI with a destination of BB will be
955 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
956 std::map<ConstantInt*, uint64_t> WeightsForHandled;
957 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
958 if (PredCases[i].Dest == BB) {
959 PTIHandled.insert(PredCases[i].Value);
961 if (PredHasWeights || SuccHasWeights) {
962 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
963 std::swap(Weights[i+1], Weights.back());
967 std::swap(PredCases[i], PredCases.back());
968 PredCases.pop_back();
972 // Okay, now we know which constants were sent to BB from the
973 // predecessor. Figure out where they will all go now.
974 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
975 if (PTIHandled.count(BBCases[i].Value)) {
976 // If this is one we are capable of getting...
977 if (PredHasWeights || SuccHasWeights)
978 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
979 PredCases.push_back(BBCases[i]);
980 NewSuccessors.push_back(BBCases[i].Dest);
981 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
984 // If there are any constants vectored to BB that TI doesn't handle,
985 // they must go to the default destination of TI.
986 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
988 E = PTIHandled.end(); I != E; ++I) {
989 if (PredHasWeights || SuccHasWeights)
990 Weights.push_back(WeightsForHandled[*I]);
991 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
992 NewSuccessors.push_back(BBDefault);
996 // Okay, at this point, we know which new successor Pred will get. Make
997 // sure we update the number of entries in the PHI nodes for these
999 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i)
1000 AddPredecessorToBlock(NewSuccessors[i], Pred, BB);
1002 Builder.SetInsertPoint(PTI);
1003 // Convert pointer to int before we switch.
1004 if (CV->getType()->isPointerTy()) {
1005 assert(DL && "Cannot switch on pointer without DataLayout");
1006 CV = Builder.CreatePtrToInt(CV, DL->getIntPtrType(CV->getType()),
1010 // Now that the successors are updated, create the new Switch instruction.
1011 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
1013 NewSI->setDebugLoc(PTI->getDebugLoc());
1014 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1015 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest);
1017 if (PredHasWeights || SuccHasWeights) {
1018 // Halve the weights if any of them cannot fit in an uint32_t
1019 FitWeights(Weights);
1021 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1023 NewSI->setMetadata(LLVMContext::MD_prof,
1024 MDBuilder(BB->getContext()).
1025 createBranchWeights(MDWeights));
1028 EraseTerminatorInstAndDCECond(PTI);
1030 // Okay, last check. If BB is still a successor of PSI, then we must
1031 // have an infinite loop case. If so, add an infinitely looping block
1032 // to handle the case to preserve the behavior of the code.
1033 BasicBlock *InfLoopBlock = nullptr;
1034 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1035 if (NewSI->getSuccessor(i) == BB) {
1036 if (!InfLoopBlock) {
1037 // Insert it at the end of the function, because it's either code,
1038 // or it won't matter if it's hot. :)
1039 InfLoopBlock = BasicBlock::Create(BB->getContext(),
1040 "infloop", BB->getParent());
1041 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1043 NewSI->setSuccessor(i, InfLoopBlock);
1052 // isSafeToHoistInvoke - If we would need to insert a select that uses the
1053 // value of this invoke (comments in HoistThenElseCodeToIf explain why we
1054 // would need to do this), we can't hoist the invoke, as there is nowhere
1055 // to put the select in this case.
1056 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1057 Instruction *I1, Instruction *I2) {
1058 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1060 for (BasicBlock::iterator BBI = SI->begin();
1061 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1062 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1063 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1064 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
1072 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I);
1074 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
1075 /// BB2, hoist any common code in the two blocks up into the branch block. The
1076 /// caller of this function guarantees that BI's block dominates BB1 and BB2.
1077 static bool HoistThenElseCodeToIf(BranchInst *BI, const DataLayout *DL) {
1078 // This does very trivial matching, with limited scanning, to find identical
1079 // instructions in the two blocks. In particular, we don't want to get into
1080 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1081 // such, we currently just scan for obviously identical instructions in an
1083 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1084 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1086 BasicBlock::iterator BB1_Itr = BB1->begin();
1087 BasicBlock::iterator BB2_Itr = BB2->begin();
1089 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++;
1090 // Skip debug info if it is not identical.
1091 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1092 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1093 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1094 while (isa<DbgInfoIntrinsic>(I1))
1096 while (isa<DbgInfoIntrinsic>(I2))
1099 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1100 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1103 BasicBlock *BIParent = BI->getParent();
1105 bool Changed = false;
1107 // If we are hoisting the terminator instruction, don't move one (making a
1108 // broken BB), instead clone it, and remove BI.
1109 if (isa<TerminatorInst>(I1))
1110 goto HoistTerminator;
1112 // For a normal instruction, we just move one to right before the branch,
1113 // then replace all uses of the other with the first. Finally, we remove
1114 // the now redundant second instruction.
1115 BIParent->getInstList().splice(BI, BB1->getInstList(), I1);
1116 if (!I2->use_empty())
1117 I2->replaceAllUsesWith(I1);
1118 I1->intersectOptionalDataWith(I2);
1119 unsigned KnownIDs[] = {
1120 LLVMContext::MD_tbaa,
1121 LLVMContext::MD_range,
1122 LLVMContext::MD_fpmath,
1123 LLVMContext::MD_invariant_load,
1124 LLVMContext::MD_nonnull
1126 combineMetadata(I1, I2, KnownIDs);
1127 I2->eraseFromParent();
1132 // Skip debug info if it is not identical.
1133 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1134 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1135 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1136 while (isa<DbgInfoIntrinsic>(I1))
1138 while (isa<DbgInfoIntrinsic>(I2))
1141 } while (I1->isIdenticalToWhenDefined(I2));
1146 // It may not be possible to hoist an invoke.
1147 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1150 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1152 for (BasicBlock::iterator BBI = SI->begin();
1153 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1154 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1155 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1159 // Check for passingValueIsAlwaysUndefined here because we would rather
1160 // eliminate undefined control flow then converting it to a select.
1161 if (passingValueIsAlwaysUndefined(BB1V, PN) ||
1162 passingValueIsAlwaysUndefined(BB2V, PN))
1165 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V, DL))
1167 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V, DL))
1172 // Okay, it is safe to hoist the terminator.
1173 Instruction *NT = I1->clone();
1174 BIParent->getInstList().insert(BI, NT);
1175 if (!NT->getType()->isVoidTy()) {
1176 I1->replaceAllUsesWith(NT);
1177 I2->replaceAllUsesWith(NT);
1181 IRBuilder<true, NoFolder> Builder(NT);
1182 // Hoisting one of the terminators from our successor is a great thing.
1183 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1184 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1185 // nodes, so we insert select instruction to compute the final result.
1186 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1187 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1189 for (BasicBlock::iterator BBI = SI->begin();
1190 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1191 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1192 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1193 if (BB1V == BB2V) continue;
1195 // These values do not agree. Insert a select instruction before NT
1196 // that determines the right value.
1197 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1199 SI = cast<SelectInst>
1200 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1201 BB1V->getName()+"."+BB2V->getName()));
1203 // Make the PHI node use the select for all incoming values for BB1/BB2
1204 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1205 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1206 PN->setIncomingValue(i, SI);
1210 // Update any PHI nodes in our new successors.
1211 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1212 AddPredecessorToBlock(*SI, BIParent, BB1);
1214 EraseTerminatorInstAndDCECond(BI);
1218 /// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd,
1219 /// check whether BBEnd has only two predecessors and the other predecessor
1220 /// ends with an unconditional branch. If it is true, sink any common code
1221 /// in the two predecessors to BBEnd.
1222 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1223 assert(BI1->isUnconditional());
1224 BasicBlock *BB1 = BI1->getParent();
1225 BasicBlock *BBEnd = BI1->getSuccessor(0);
1227 // Check that BBEnd has two predecessors and the other predecessor ends with
1228 // an unconditional branch.
1229 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1230 BasicBlock *Pred0 = *PI++;
1231 if (PI == PE) // Only one predecessor.
1233 BasicBlock *Pred1 = *PI++;
1234 if (PI != PE) // More than two predecessors.
1236 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1237 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1238 if (!BI2 || !BI2->isUnconditional())
1241 // Gather the PHI nodes in BBEnd.
1242 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2;
1243 Instruction *FirstNonPhiInBBEnd = nullptr;
1244 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end();
1246 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1247 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1248 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1249 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN);
1251 FirstNonPhiInBBEnd = &*I;
1255 if (!FirstNonPhiInBBEnd)
1259 // This does very trivial matching, with limited scanning, to find identical
1260 // instructions in the two blocks. We scan backward for obviously identical
1261 // instructions in an identical order.
1262 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1263 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(),
1264 RE2 = BB2->getInstList().rend();
1266 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1269 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1272 // Skip the unconditional branches.
1276 bool Changed = false;
1277 while (RI1 != RE1 && RI2 != RE2) {
1279 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1282 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1286 Instruction *I1 = &*RI1, *I2 = &*RI2;
1287 // I1 and I2 should have a single use in the same PHI node, and they
1288 // perform the same operation.
1289 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1290 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1291 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1292 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) ||
1293 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1294 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1295 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1296 !I1->hasOneUse() || !I2->hasOneUse() ||
1297 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() ||
1298 MapValueFromBB1ToBB2[I1].first != I2)
1301 // Check whether we should swap the operands of ICmpInst.
1302 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1303 bool SwapOpnds = false;
1304 if (ICmp1 && ICmp2 &&
1305 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1306 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1307 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1308 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1309 ICmp2->swapOperands();
1312 if (!I1->isSameOperationAs(I2)) {
1314 ICmp2->swapOperands();
1318 // The operands should be either the same or they need to be generated
1319 // with a PHI node after sinking. We only handle the case where there is
1320 // a single pair of different operands.
1321 Value *DifferentOp1 = nullptr, *DifferentOp2 = nullptr;
1322 unsigned Op1Idx = 0;
1323 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1324 if (I1->getOperand(I) == I2->getOperand(I))
1326 // Early exit if we have more-than one pair of different operands or
1327 // the different operand is already in MapValueFromBB1ToBB2.
1328 // Early exit if we need a PHI node to replace a constant.
1330 MapValueFromBB1ToBB2.find(I1->getOperand(I)) !=
1331 MapValueFromBB1ToBB2.end() ||
1332 isa<Constant>(I1->getOperand(I)) ||
1333 isa<Constant>(I2->getOperand(I))) {
1334 // If we can't sink the instructions, undo the swapping.
1336 ICmp2->swapOperands();
1339 DifferentOp1 = I1->getOperand(I);
1341 DifferentOp2 = I2->getOperand(I);
1344 // We insert the pair of different operands to MapValueFromBB1ToBB2 and
1345 // remove (I1, I2) from MapValueFromBB1ToBB2.
1347 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2,
1348 DifferentOp1->getName() + ".sink",
1350 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN);
1351 // I1 should use NewPN instead of DifferentOp1.
1352 I1->setOperand(Op1Idx, NewPN);
1353 NewPN->addIncoming(DifferentOp1, BB1);
1354 NewPN->addIncoming(DifferentOp2, BB2);
1355 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1357 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second;
1358 MapValueFromBB1ToBB2.erase(I1);
1360 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";);
1361 DEBUG(dbgs() << " " << *I2 << "\n";);
1362 // We need to update RE1 and RE2 if we are going to sink the first
1363 // instruction in the basic block down.
1364 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1365 // Sink the instruction.
1366 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1);
1367 if (!OldPN->use_empty())
1368 OldPN->replaceAllUsesWith(I1);
1369 OldPN->eraseFromParent();
1371 if (!I2->use_empty())
1372 I2->replaceAllUsesWith(I1);
1373 I1->intersectOptionalDataWith(I2);
1374 // TODO: Use combineMetadata here to preserve what metadata we can
1375 // (analogous to the hoisting case above).
1376 I2->eraseFromParent();
1379 RE1 = BB1->getInstList().rend();
1381 RE2 = BB2->getInstList().rend();
1382 FirstNonPhiInBBEnd = I1;
1389 /// \brief Determine if we can hoist sink a sole store instruction out of a
1390 /// conditional block.
1392 /// We are looking for code like the following:
1394 /// store i32 %add, i32* %arrayidx2
1395 /// ... // No other stores or function calls (we could be calling a memory
1396 /// ... // function).
1397 /// %cmp = icmp ult %x, %y
1398 /// br i1 %cmp, label %EndBB, label %ThenBB
1400 /// store i32 %add5, i32* %arrayidx2
1404 /// We are going to transform this into:
1406 /// store i32 %add, i32* %arrayidx2
1408 /// %cmp = icmp ult %x, %y
1409 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1410 /// store i32 %add.add5, i32* %arrayidx2
1413 /// \return The pointer to the value of the previous store if the store can be
1414 /// hoisted into the predecessor block. 0 otherwise.
1415 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1416 BasicBlock *StoreBB, BasicBlock *EndBB) {
1417 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1421 // Volatile or atomic.
1422 if (!StoreToHoist->isSimple())
1425 Value *StorePtr = StoreToHoist->getPointerOperand();
1427 // Look for a store to the same pointer in BrBB.
1428 unsigned MaxNumInstToLookAt = 10;
1429 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(),
1430 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) {
1431 Instruction *CurI = &*RI;
1433 // Could be calling an instruction that effects memory like free().
1434 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI))
1437 StoreInst *SI = dyn_cast<StoreInst>(CurI);
1438 // Found the previous store make sure it stores to the same location.
1439 if (SI && SI->getPointerOperand() == StorePtr)
1440 // Found the previous store, return its value operand.
1441 return SI->getValueOperand();
1443 return nullptr; // Unknown store.
1449 /// \brief Speculate a conditional basic block flattening the CFG.
1451 /// Note that this is a very risky transform currently. Speculating
1452 /// instructions like this is most often not desirable. Instead, there is an MI
1453 /// pass which can do it with full awareness of the resource constraints.
1454 /// However, some cases are "obvious" and we should do directly. An example of
1455 /// this is speculating a single, reasonably cheap instruction.
1457 /// There is only one distinct advantage to flattening the CFG at the IR level:
1458 /// it makes very common but simplistic optimizations such as are common in
1459 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1460 /// modeling their effects with easier to reason about SSA value graphs.
1463 /// An illustration of this transform is turning this IR:
1466 /// %cmp = icmp ult %x, %y
1467 /// br i1 %cmp, label %EndBB, label %ThenBB
1469 /// %sub = sub %x, %y
1472 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1479 /// %cmp = icmp ult %x, %y
1480 /// %sub = sub %x, %y
1481 /// %cond = select i1 %cmp, 0, %sub
1485 /// \returns true if the conditional block is removed.
1486 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1487 const DataLayout *DL) {
1488 // Be conservative for now. FP select instruction can often be expensive.
1489 Value *BrCond = BI->getCondition();
1490 if (isa<FCmpInst>(BrCond))
1493 BasicBlock *BB = BI->getParent();
1494 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1496 // If ThenBB is actually on the false edge of the conditional branch, remember
1497 // to swap the select operands later.
1498 bool Invert = false;
1499 if (ThenBB != BI->getSuccessor(0)) {
1500 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1503 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1505 // Keep a count of how many times instructions are used within CondBB when
1506 // they are candidates for sinking into CondBB. Specifically:
1507 // - They are defined in BB, and
1508 // - They have no side effects, and
1509 // - All of their uses are in CondBB.
1510 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1512 unsigned SpeculationCost = 0;
1513 Value *SpeculatedStoreValue = nullptr;
1514 StoreInst *SpeculatedStore = nullptr;
1515 for (BasicBlock::iterator BBI = ThenBB->begin(),
1516 BBE = std::prev(ThenBB->end());
1517 BBI != BBE; ++BBI) {
1518 Instruction *I = BBI;
1520 if (isa<DbgInfoIntrinsic>(I))
1523 // Only speculatively execution a single instruction (not counting the
1524 // terminator) for now.
1526 if (SpeculationCost > 1)
1529 // Don't hoist the instruction if it's unsafe or expensive.
1530 if (!isSafeToSpeculativelyExecute(I, DL) &&
1531 !(HoistCondStores &&
1532 (SpeculatedStoreValue = isSafeToSpeculateStore(I, BB, ThenBB,
1535 if (!SpeculatedStoreValue &&
1536 ComputeSpeculationCost(I, DL) > PHINodeFoldingThreshold)
1539 // Store the store speculation candidate.
1540 if (SpeculatedStoreValue)
1541 SpeculatedStore = cast<StoreInst>(I);
1543 // Do not hoist the instruction if any of its operands are defined but not
1544 // used in BB. The transformation will prevent the operand from
1545 // being sunk into the use block.
1546 for (User::op_iterator i = I->op_begin(), e = I->op_end();
1548 Instruction *OpI = dyn_cast<Instruction>(*i);
1549 if (!OpI || OpI->getParent() != BB ||
1550 OpI->mayHaveSideEffects())
1551 continue; // Not a candidate for sinking.
1553 ++SinkCandidateUseCounts[OpI];
1557 // Consider any sink candidates which are only used in CondBB as costs for
1558 // speculation. Note, while we iterate over a DenseMap here, we are summing
1559 // and so iteration order isn't significant.
1560 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I =
1561 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end();
1563 if (I->first->getNumUses() == I->second) {
1565 if (SpeculationCost > 1)
1569 // Check that the PHI nodes can be converted to selects.
1570 bool HaveRewritablePHIs = false;
1571 for (BasicBlock::iterator I = EndBB->begin();
1572 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1573 Value *OrigV = PN->getIncomingValueForBlock(BB);
1574 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1576 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1577 // Skip PHIs which are trivial.
1581 // Don't convert to selects if we could remove undefined behavior instead.
1582 if (passingValueIsAlwaysUndefined(OrigV, PN) ||
1583 passingValueIsAlwaysUndefined(ThenV, PN))
1586 HaveRewritablePHIs = true;
1587 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1588 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1589 if (!OrigCE && !ThenCE)
1590 continue; // Known safe and cheap.
1592 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE, DL)) ||
1593 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE, DL)))
1595 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, DL) : 0;
1596 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, DL) : 0;
1597 if (OrigCost + ThenCost > 2 * PHINodeFoldingThreshold)
1600 // Account for the cost of an unfolded ConstantExpr which could end up
1601 // getting expanded into Instructions.
1602 // FIXME: This doesn't account for how many operations are combined in the
1603 // constant expression.
1605 if (SpeculationCost > 1)
1609 // If there are no PHIs to process, bail early. This helps ensure idempotence
1611 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
1614 // If we get here, we can hoist the instruction and if-convert.
1615 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1617 // Insert a select of the value of the speculated store.
1618 if (SpeculatedStoreValue) {
1619 IRBuilder<true, NoFolder> Builder(BI);
1620 Value *TrueV = SpeculatedStore->getValueOperand();
1621 Value *FalseV = SpeculatedStoreValue;
1623 std::swap(TrueV, FalseV);
1624 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() +
1625 "." + FalseV->getName());
1626 SpeculatedStore->setOperand(0, S);
1629 // Hoist the instructions.
1630 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(),
1631 std::prev(ThenBB->end()));
1633 // Insert selects and rewrite the PHI operands.
1634 IRBuilder<true, NoFolder> Builder(BI);
1635 for (BasicBlock::iterator I = EndBB->begin();
1636 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1637 unsigned OrigI = PN->getBasicBlockIndex(BB);
1638 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1639 Value *OrigV = PN->getIncomingValue(OrigI);
1640 Value *ThenV = PN->getIncomingValue(ThenI);
1642 // Skip PHIs which are trivial.
1646 // Create a select whose true value is the speculatively executed value and
1647 // false value is the preexisting value. Swap them if the branch
1648 // destinations were inverted.
1649 Value *TrueV = ThenV, *FalseV = OrigV;
1651 std::swap(TrueV, FalseV);
1652 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV,
1653 TrueV->getName() + "." + FalseV->getName());
1654 PN->setIncomingValue(OrigI, V);
1655 PN->setIncomingValue(ThenI, V);
1662 /// \returns True if this block contains a CallInst with the NoDuplicate
1664 static bool HasNoDuplicateCall(const BasicBlock *BB) {
1665 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1666 const CallInst *CI = dyn_cast<CallInst>(I);
1669 if (CI->cannotDuplicate())
1675 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1676 /// across this block.
1677 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1678 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1681 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1682 if (isa<DbgInfoIntrinsic>(BBI))
1684 if (Size > 10) return false; // Don't clone large BB's.
1687 // We can only support instructions that do not define values that are
1688 // live outside of the current basic block.
1689 for (User *U : BBI->users()) {
1690 Instruction *UI = cast<Instruction>(U);
1691 if (UI->getParent() != BB || isa<PHINode>(UI)) return false;
1694 // Looks ok, continue checking.
1700 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1701 /// that is defined in the same block as the branch and if any PHI entries are
1702 /// constants, thread edges corresponding to that entry to be branches to their
1703 /// ultimate destination.
1704 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *DL) {
1705 BasicBlock *BB = BI->getParent();
1706 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1707 // NOTE: we currently cannot transform this case if the PHI node is used
1708 // outside of the block.
1709 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1712 // Degenerate case of a single entry PHI.
1713 if (PN->getNumIncomingValues() == 1) {
1714 FoldSingleEntryPHINodes(PN->getParent());
1718 // Now we know that this block has multiple preds and two succs.
1719 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1721 if (HasNoDuplicateCall(BB)) return false;
1723 // Okay, this is a simple enough basic block. See if any phi values are
1725 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1726 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1727 if (!CB || !CB->getType()->isIntegerTy(1)) continue;
1729 // Okay, we now know that all edges from PredBB should be revectored to
1730 // branch to RealDest.
1731 BasicBlock *PredBB = PN->getIncomingBlock(i);
1732 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1734 if (RealDest == BB) continue; // Skip self loops.
1735 // Skip if the predecessor's terminator is an indirect branch.
1736 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1738 // The dest block might have PHI nodes, other predecessors and other
1739 // difficult cases. Instead of being smart about this, just insert a new
1740 // block that jumps to the destination block, effectively splitting
1741 // the edge we are about to create.
1742 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1743 RealDest->getName()+".critedge",
1744 RealDest->getParent(), RealDest);
1745 BranchInst::Create(RealDest, EdgeBB);
1747 // Update PHI nodes.
1748 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1750 // BB may have instructions that are being threaded over. Clone these
1751 // instructions into EdgeBB. We know that there will be no uses of the
1752 // cloned instructions outside of EdgeBB.
1753 BasicBlock::iterator InsertPt = EdgeBB->begin();
1754 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1755 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1756 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1757 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1760 // Clone the instruction.
1761 Instruction *N = BBI->clone();
1762 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1764 // Update operands due to translation.
1765 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1767 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1768 if (PI != TranslateMap.end())
1772 // Check for trivial simplification.
1773 if (Value *V = SimplifyInstruction(N, DL)) {
1774 TranslateMap[BBI] = V;
1775 delete N; // Instruction folded away, don't need actual inst
1777 // Insert the new instruction into its new home.
1778 EdgeBB->getInstList().insert(InsertPt, N);
1779 if (!BBI->use_empty())
1780 TranslateMap[BBI] = N;
1784 // Loop over all of the edges from PredBB to BB, changing them to branch
1785 // to EdgeBB instead.
1786 TerminatorInst *PredBBTI = PredBB->getTerminator();
1787 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1788 if (PredBBTI->getSuccessor(i) == BB) {
1789 BB->removePredecessor(PredBB);
1790 PredBBTI->setSuccessor(i, EdgeBB);
1793 // Recurse, simplifying any other constants.
1794 return FoldCondBranchOnPHI(BI, DL) | true;
1800 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1801 /// PHI node, see if we can eliminate it.
1802 static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *DL) {
1803 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1804 // statement", which has a very simple dominance structure. Basically, we
1805 // are trying to find the condition that is being branched on, which
1806 // subsequently causes this merge to happen. We really want control
1807 // dependence information for this check, but simplifycfg can't keep it up
1808 // to date, and this catches most of the cases we care about anyway.
1809 BasicBlock *BB = PN->getParent();
1810 BasicBlock *IfTrue, *IfFalse;
1811 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1813 // Don't bother if the branch will be constant folded trivially.
1814 isa<ConstantInt>(IfCond))
1817 // Okay, we found that we can merge this two-entry phi node into a select.
1818 // Doing so would require us to fold *all* two entry phi nodes in this block.
1819 // At some point this becomes non-profitable (particularly if the target
1820 // doesn't support cmov's). Only do this transformation if there are two or
1821 // fewer PHI nodes in this block.
1822 unsigned NumPhis = 0;
1823 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1827 // Loop over the PHI's seeing if we can promote them all to select
1828 // instructions. While we are at it, keep track of the instructions
1829 // that need to be moved to the dominating block.
1830 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1831 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1832 MaxCostVal1 = PHINodeFoldingThreshold;
1834 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1835 PHINode *PN = cast<PHINode>(II++);
1836 if (Value *V = SimplifyInstruction(PN, DL)) {
1837 PN->replaceAllUsesWith(V);
1838 PN->eraseFromParent();
1842 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1844 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1849 // If we folded the first phi, PN dangles at this point. Refresh it. If
1850 // we ran out of PHIs then we simplified them all.
1851 PN = dyn_cast<PHINode>(BB->begin());
1852 if (!PN) return true;
1854 // Don't fold i1 branches on PHIs which contain binary operators. These can
1855 // often be turned into switches and other things.
1856 if (PN->getType()->isIntegerTy(1) &&
1857 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1858 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1859 isa<BinaryOperator>(IfCond)))
1862 // If we all PHI nodes are promotable, check to make sure that all
1863 // instructions in the predecessor blocks can be promoted as well. If
1864 // not, we won't be able to get rid of the control flow, so it's not
1865 // worth promoting to select instructions.
1866 BasicBlock *DomBlock = nullptr;
1867 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1868 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1869 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1872 DomBlock = *pred_begin(IfBlock1);
1873 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1874 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1875 // This is not an aggressive instruction that we can promote.
1876 // Because of this, we won't be able to get rid of the control
1877 // flow, so the xform is not worth it.
1882 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1885 DomBlock = *pred_begin(IfBlock2);
1886 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1887 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1888 // This is not an aggressive instruction that we can promote.
1889 // Because of this, we won't be able to get rid of the control
1890 // flow, so the xform is not worth it.
1895 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1896 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1898 // If we can still promote the PHI nodes after this gauntlet of tests,
1899 // do all of the PHI's now.
1900 Instruction *InsertPt = DomBlock->getTerminator();
1901 IRBuilder<true, NoFolder> Builder(InsertPt);
1903 // Move all 'aggressive' instructions, which are defined in the
1904 // conditional parts of the if's up to the dominating block.
1906 DomBlock->getInstList().splice(InsertPt,
1907 IfBlock1->getInstList(), IfBlock1->begin(),
1908 IfBlock1->getTerminator());
1910 DomBlock->getInstList().splice(InsertPt,
1911 IfBlock2->getInstList(), IfBlock2->begin(),
1912 IfBlock2->getTerminator());
1914 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1915 // Change the PHI node into a select instruction.
1916 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1917 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1920 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1921 PN->replaceAllUsesWith(NV);
1923 PN->eraseFromParent();
1926 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1927 // has been flattened. Change DomBlock to jump directly to our new block to
1928 // avoid other simplifycfg's kicking in on the diamond.
1929 TerminatorInst *OldTI = DomBlock->getTerminator();
1930 Builder.SetInsertPoint(OldTI);
1931 Builder.CreateBr(BB);
1932 OldTI->eraseFromParent();
1936 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1937 /// to two returning blocks, try to merge them together into one return,
1938 /// introducing a select if the return values disagree.
1939 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1940 IRBuilder<> &Builder) {
1941 assert(BI->isConditional() && "Must be a conditional branch");
1942 BasicBlock *TrueSucc = BI->getSuccessor(0);
1943 BasicBlock *FalseSucc = BI->getSuccessor(1);
1944 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1945 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1947 // Check to ensure both blocks are empty (just a return) or optionally empty
1948 // with PHI nodes. If there are other instructions, merging would cause extra
1949 // computation on one path or the other.
1950 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1952 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1955 Builder.SetInsertPoint(BI);
1956 // Okay, we found a branch that is going to two return nodes. If
1957 // there is no return value for this function, just change the
1958 // branch into a return.
1959 if (FalseRet->getNumOperands() == 0) {
1960 TrueSucc->removePredecessor(BI->getParent());
1961 FalseSucc->removePredecessor(BI->getParent());
1962 Builder.CreateRetVoid();
1963 EraseTerminatorInstAndDCECond(BI);
1967 // Otherwise, figure out what the true and false return values are
1968 // so we can insert a new select instruction.
1969 Value *TrueValue = TrueRet->getReturnValue();
1970 Value *FalseValue = FalseRet->getReturnValue();
1972 // Unwrap any PHI nodes in the return blocks.
1973 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1974 if (TVPN->getParent() == TrueSucc)
1975 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1976 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1977 if (FVPN->getParent() == FalseSucc)
1978 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1980 // In order for this transformation to be safe, we must be able to
1981 // unconditionally execute both operands to the return. This is
1982 // normally the case, but we could have a potentially-trapping
1983 // constant expression that prevents this transformation from being
1985 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1988 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1992 // Okay, we collected all the mapped values and checked them for sanity, and
1993 // defined to really do this transformation. First, update the CFG.
1994 TrueSucc->removePredecessor(BI->getParent());
1995 FalseSucc->removePredecessor(BI->getParent());
1997 // Insert select instructions where needed.
1998 Value *BrCond = BI->getCondition();
2000 // Insert a select if the results differ.
2001 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2002 } else if (isa<UndefValue>(TrueValue)) {
2003 TrueValue = FalseValue;
2005 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
2006 FalseValue, "retval");
2010 Value *RI = !TrueValue ?
2011 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2015 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2016 << "\n " << *BI << "NewRet = " << *RI
2017 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
2019 EraseTerminatorInstAndDCECond(BI);
2024 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
2025 /// probabilities of the branch taking each edge. Fills in the two APInt
2026 /// parameters and return true, or returns false if no or invalid metadata was
2028 static bool ExtractBranchMetadata(BranchInst *BI,
2029 uint64_t &ProbTrue, uint64_t &ProbFalse) {
2030 assert(BI->isConditional() &&
2031 "Looking for probabilities on unconditional branch?");
2032 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
2033 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
2034 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
2035 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
2036 if (!CITrue || !CIFalse) return false;
2037 ProbTrue = CITrue->getValue().getZExtValue();
2038 ProbFalse = CIFalse->getValue().getZExtValue();
2042 /// checkCSEInPredecessor - Return true if the given instruction is available
2043 /// in its predecessor block. If yes, the instruction will be removed.
2045 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
2046 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2048 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
2049 Instruction *PBI = &*I;
2050 // Check whether Inst and PBI generate the same value.
2051 if (Inst->isIdenticalTo(PBI)) {
2052 Inst->replaceAllUsesWith(PBI);
2053 Inst->eraseFromParent();
2060 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
2061 /// predecessor branches to us and one of our successors, fold the block into
2062 /// the predecessor and use logical operations to pick the right destination.
2063 bool llvm::FoldBranchToCommonDest(BranchInst *BI, const DataLayout *DL,
2064 unsigned BonusInstThreshold) {
2065 BasicBlock *BB = BI->getParent();
2067 Instruction *Cond = nullptr;
2068 if (BI->isConditional())
2069 Cond = dyn_cast<Instruction>(BI->getCondition());
2071 // For unconditional branch, check for a simple CFG pattern, where
2072 // BB has a single predecessor and BB's successor is also its predecessor's
2073 // successor. If such pattern exisits, check for CSE between BB and its
2075 if (BasicBlock *PB = BB->getSinglePredecessor())
2076 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2077 if (PBI->isConditional() &&
2078 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2079 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2080 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
2082 Instruction *Curr = I++;
2083 if (isa<CmpInst>(Curr)) {
2087 // Quit if we can't remove this instruction.
2088 if (!checkCSEInPredecessor(Curr, PB))
2097 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2098 Cond->getParent() != BB || !Cond->hasOneUse())
2101 // Make sure the instruction after the condition is the cond branch.
2102 BasicBlock::iterator CondIt = Cond; ++CondIt;
2104 // Ignore dbg intrinsics.
2105 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
2110 // Only allow this transformation if computing the condition doesn't involve
2111 // too many instructions and these involved instructions can be executed
2112 // unconditionally. We denote all involved instructions except the condition
2113 // as "bonus instructions", and only allow this transformation when the
2114 // number of the bonus instructions does not exceed a certain threshold.
2115 unsigned NumBonusInsts = 0;
2116 for (auto I = BB->begin(); Cond != I; ++I) {
2117 // Ignore dbg intrinsics.
2118 if (isa<DbgInfoIntrinsic>(I))
2120 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(I, DL))
2122 // I has only one use and can be executed unconditionally.
2123 Instruction *User = dyn_cast<Instruction>(I->user_back());
2124 if (User == nullptr || User->getParent() != BB)
2126 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2127 // to use any other instruction, User must be an instruction between next(I)
2130 // Early exits once we reach the limit.
2131 if (NumBonusInsts > BonusInstThreshold)
2135 // Cond is known to be a compare or binary operator. Check to make sure that
2136 // neither operand is a potentially-trapping constant expression.
2137 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2140 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2144 // Finally, don't infinitely unroll conditional loops.
2145 BasicBlock *TrueDest = BI->getSuccessor(0);
2146 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2147 if (TrueDest == BB || FalseDest == BB)
2150 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2151 BasicBlock *PredBlock = *PI;
2152 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2154 // Check that we have two conditional branches. If there is a PHI node in
2155 // the common successor, verify that the same value flows in from both
2157 SmallVector<PHINode*, 4> PHIs;
2158 if (!PBI || PBI->isUnconditional() ||
2159 (BI->isConditional() &&
2160 !SafeToMergeTerminators(BI, PBI)) ||
2161 (!BI->isConditional() &&
2162 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2165 // Determine if the two branches share a common destination.
2166 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2167 bool InvertPredCond = false;
2169 if (BI->isConditional()) {
2170 if (PBI->getSuccessor(0) == TrueDest)
2171 Opc = Instruction::Or;
2172 else if (PBI->getSuccessor(1) == FalseDest)
2173 Opc = Instruction::And;
2174 else if (PBI->getSuccessor(0) == FalseDest)
2175 Opc = Instruction::And, InvertPredCond = true;
2176 else if (PBI->getSuccessor(1) == TrueDest)
2177 Opc = Instruction::Or, InvertPredCond = true;
2181 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2185 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2186 IRBuilder<> Builder(PBI);
2188 // If we need to invert the condition in the pred block to match, do so now.
2189 if (InvertPredCond) {
2190 Value *NewCond = PBI->getCondition();
2192 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2193 CmpInst *CI = cast<CmpInst>(NewCond);
2194 CI->setPredicate(CI->getInversePredicate());
2196 NewCond = Builder.CreateNot(NewCond,
2197 PBI->getCondition()->getName()+".not");
2200 PBI->setCondition(NewCond);
2201 PBI->swapSuccessors();
2204 // If we have bonus instructions, clone them into the predecessor block.
2205 // Note that there may be mutliple predecessor blocks, so we cannot move
2206 // bonus instructions to a predecessor block.
2207 ValueToValueMapTy VMap; // maps original values to cloned values
2208 // We already make sure Cond is the last instruction before BI. Therefore,
2209 // every instructions before Cond other than DbgInfoIntrinsic are bonus
2211 for (auto BonusInst = BB->begin(); Cond != BonusInst; ++BonusInst) {
2212 if (isa<DbgInfoIntrinsic>(BonusInst))
2214 Instruction *NewBonusInst = BonusInst->clone();
2215 RemapInstruction(NewBonusInst, VMap,
2216 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2217 VMap[BonusInst] = NewBonusInst;
2219 // If we moved a load, we cannot any longer claim any knowledge about
2220 // its potential value. The previous information might have been valid
2221 // only given the branch precondition.
2222 // For an analogous reason, we must also drop all the metadata whose
2223 // semantics we don't understand.
2224 NewBonusInst->dropUnknownMetadata(LLVMContext::MD_dbg);
2226 PredBlock->getInstList().insert(PBI, NewBonusInst);
2227 NewBonusInst->takeName(BonusInst);
2228 BonusInst->setName(BonusInst->getName() + ".old");
2231 // Clone Cond into the predecessor basic block, and or/and the
2232 // two conditions together.
2233 Instruction *New = Cond->clone();
2234 RemapInstruction(New, VMap,
2235 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2236 PredBlock->getInstList().insert(PBI, New);
2237 New->takeName(Cond);
2238 Cond->setName(New->getName() + ".old");
2240 if (BI->isConditional()) {
2241 Instruction *NewCond =
2242 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2244 PBI->setCondition(NewCond);
2246 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2247 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2249 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2251 SmallVector<uint64_t, 8> NewWeights;
2253 if (PBI->getSuccessor(0) == BB) {
2254 if (PredHasWeights && SuccHasWeights) {
2255 // PBI: br i1 %x, BB, FalseDest
2256 // BI: br i1 %y, TrueDest, FalseDest
2257 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2258 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2259 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2260 // TrueWeight for PBI * FalseWeight for BI.
2261 // We assume that total weights of a BranchInst can fit into 32 bits.
2262 // Therefore, we will not have overflow using 64-bit arithmetic.
2263 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2264 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2266 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2267 PBI->setSuccessor(0, TrueDest);
2269 if (PBI->getSuccessor(1) == BB) {
2270 if (PredHasWeights && SuccHasWeights) {
2271 // PBI: br i1 %x, TrueDest, BB
2272 // BI: br i1 %y, TrueDest, FalseDest
2273 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2274 // FalseWeight for PBI * TrueWeight for BI.
2275 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2276 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2277 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2278 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2280 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2281 PBI->setSuccessor(1, FalseDest);
2283 if (NewWeights.size() == 2) {
2284 // Halve the weights if any of them cannot fit in an uint32_t
2285 FitWeights(NewWeights);
2287 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2288 PBI->setMetadata(LLVMContext::MD_prof,
2289 MDBuilder(BI->getContext()).
2290 createBranchWeights(MDWeights));
2292 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2294 // Update PHI nodes in the common successors.
2295 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2296 ConstantInt *PBI_C = cast<ConstantInt>(
2297 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2298 assert(PBI_C->getType()->isIntegerTy(1));
2299 Instruction *MergedCond = nullptr;
2300 if (PBI->getSuccessor(0) == TrueDest) {
2301 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2302 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2303 // is false: !PBI_Cond and BI_Value
2304 Instruction *NotCond =
2305 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2308 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2313 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2314 PBI->getCondition(), MergedCond,
2317 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2318 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2319 // is false: PBI_Cond and BI_Value
2321 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2322 PBI->getCondition(), New,
2324 if (PBI_C->isOne()) {
2325 Instruction *NotCond =
2326 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2329 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2330 NotCond, MergedCond,
2335 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2338 // Change PBI from Conditional to Unconditional.
2339 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2340 EraseTerminatorInstAndDCECond(PBI);
2344 // TODO: If BB is reachable from all paths through PredBlock, then we
2345 // could replace PBI's branch probabilities with BI's.
2347 // Copy any debug value intrinsics into the end of PredBlock.
2348 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2349 if (isa<DbgInfoIntrinsic>(*I))
2350 I->clone()->insertBefore(PBI);
2357 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2358 /// predecessor of another block, this function tries to simplify it. We know
2359 /// that PBI and BI are both conditional branches, and BI is in one of the
2360 /// successor blocks of PBI - PBI branches to BI.
2361 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2362 assert(PBI->isConditional() && BI->isConditional());
2363 BasicBlock *BB = BI->getParent();
2365 // If this block ends with a branch instruction, and if there is a
2366 // predecessor that ends on a branch of the same condition, make
2367 // this conditional branch redundant.
2368 if (PBI->getCondition() == BI->getCondition() &&
2369 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2370 // Okay, the outcome of this conditional branch is statically
2371 // knowable. If this block had a single pred, handle specially.
2372 if (BB->getSinglePredecessor()) {
2373 // Turn this into a branch on constant.
2374 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2375 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2377 return true; // Nuke the branch on constant.
2380 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2381 // in the constant and simplify the block result. Subsequent passes of
2382 // simplifycfg will thread the block.
2383 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2384 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2385 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2386 std::distance(PB, PE),
2387 BI->getCondition()->getName() + ".pr",
2389 // Okay, we're going to insert the PHI node. Since PBI is not the only
2390 // predecessor, compute the PHI'd conditional value for all of the preds.
2391 // Any predecessor where the condition is not computable we keep symbolic.
2392 for (pred_iterator PI = PB; PI != PE; ++PI) {
2393 BasicBlock *P = *PI;
2394 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2395 PBI != BI && PBI->isConditional() &&
2396 PBI->getCondition() == BI->getCondition() &&
2397 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2398 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2399 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2402 NewPN->addIncoming(BI->getCondition(), P);
2406 BI->setCondition(NewPN);
2411 // If this is a conditional branch in an empty block, and if any
2412 // predecessors are a conditional branch to one of our destinations,
2413 // fold the conditions into logical ops and one cond br.
2414 BasicBlock::iterator BBI = BB->begin();
2415 // Ignore dbg intrinsics.
2416 while (isa<DbgInfoIntrinsic>(BBI))
2422 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2427 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2429 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2430 PBIOp = 0, BIOp = 1;
2431 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2432 PBIOp = 1, BIOp = 0;
2433 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2438 // Check to make sure that the other destination of this branch
2439 // isn't BB itself. If so, this is an infinite loop that will
2440 // keep getting unwound.
2441 if (PBI->getSuccessor(PBIOp) == BB)
2444 // Do not perform this transformation if it would require
2445 // insertion of a large number of select instructions. For targets
2446 // without predication/cmovs, this is a big pessimization.
2448 // Also do not perform this transformation if any phi node in the common
2449 // destination block can trap when reached by BB or PBB (PR17073). In that
2450 // case, it would be unsafe to hoist the operation into a select instruction.
2452 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2453 unsigned NumPhis = 0;
2454 for (BasicBlock::iterator II = CommonDest->begin();
2455 isa<PHINode>(II); ++II, ++NumPhis) {
2456 if (NumPhis > 2) // Disable this xform.
2459 PHINode *PN = cast<PHINode>(II);
2460 Value *BIV = PN->getIncomingValueForBlock(BB);
2461 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
2465 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2466 Value *PBIV = PN->getIncomingValue(PBBIdx);
2467 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
2472 // Finally, if everything is ok, fold the branches to logical ops.
2473 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2475 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2476 << "AND: " << *BI->getParent());
2479 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2480 // branch in it, where one edge (OtherDest) goes back to itself but the other
2481 // exits. We don't *know* that the program avoids the infinite loop
2482 // (even though that seems likely). If we do this xform naively, we'll end up
2483 // recursively unpeeling the loop. Since we know that (after the xform is
2484 // done) that the block *is* infinite if reached, we just make it an obviously
2485 // infinite loop with no cond branch.
2486 if (OtherDest == BB) {
2487 // Insert it at the end of the function, because it's either code,
2488 // or it won't matter if it's hot. :)
2489 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2490 "infloop", BB->getParent());
2491 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2492 OtherDest = InfLoopBlock;
2495 DEBUG(dbgs() << *PBI->getParent()->getParent());
2497 // BI may have other predecessors. Because of this, we leave
2498 // it alone, but modify PBI.
2500 // Make sure we get to CommonDest on True&True directions.
2501 Value *PBICond = PBI->getCondition();
2502 IRBuilder<true, NoFolder> Builder(PBI);
2504 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2506 Value *BICond = BI->getCondition();
2508 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2510 // Merge the conditions.
2511 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2513 // Modify PBI to branch on the new condition to the new dests.
2514 PBI->setCondition(Cond);
2515 PBI->setSuccessor(0, CommonDest);
2516 PBI->setSuccessor(1, OtherDest);
2518 // Update branch weight for PBI.
2519 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2520 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2522 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2524 if (PredHasWeights && SuccHasWeights) {
2525 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2526 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2527 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2528 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2529 // The weight to CommonDest should be PredCommon * SuccTotal +
2530 // PredOther * SuccCommon.
2531 // The weight to OtherDest should be PredOther * SuccOther.
2532 SmallVector<uint64_t, 2> NewWeights;
2533 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) +
2534 PredOther * SuccCommon);
2535 NewWeights.push_back(PredOther * SuccOther);
2536 // Halve the weights if any of them cannot fit in an uint32_t
2537 FitWeights(NewWeights);
2539 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end());
2540 PBI->setMetadata(LLVMContext::MD_prof,
2541 MDBuilder(BI->getContext()).
2542 createBranchWeights(MDWeights));
2545 // OtherDest may have phi nodes. If so, add an entry from PBI's
2546 // block that are identical to the entries for BI's block.
2547 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2549 // We know that the CommonDest already had an edge from PBI to
2550 // it. If it has PHIs though, the PHIs may have different
2551 // entries for BB and PBI's BB. If so, insert a select to make
2554 for (BasicBlock::iterator II = CommonDest->begin();
2555 (PN = dyn_cast<PHINode>(II)); ++II) {
2556 Value *BIV = PN->getIncomingValueForBlock(BB);
2557 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2558 Value *PBIV = PN->getIncomingValue(PBBIdx);
2560 // Insert a select in PBI to pick the right value.
2561 Value *NV = cast<SelectInst>
2562 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2563 PN->setIncomingValue(PBBIdx, NV);
2567 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2568 DEBUG(dbgs() << *PBI->getParent()->getParent());
2570 // This basic block is probably dead. We know it has at least
2571 // one fewer predecessor.
2575 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2576 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2577 // Takes care of updating the successors and removing the old terminator.
2578 // Also makes sure not to introduce new successors by assuming that edges to
2579 // non-successor TrueBBs and FalseBBs aren't reachable.
2580 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2581 BasicBlock *TrueBB, BasicBlock *FalseBB,
2582 uint32_t TrueWeight,
2583 uint32_t FalseWeight){
2584 // Remove any superfluous successor edges from the CFG.
2585 // First, figure out which successors to preserve.
2586 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2588 BasicBlock *KeepEdge1 = TrueBB;
2589 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
2591 // Then remove the rest.
2592 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2593 BasicBlock *Succ = OldTerm->getSuccessor(I);
2594 // Make sure only to keep exactly one copy of each edge.
2595 if (Succ == KeepEdge1)
2596 KeepEdge1 = nullptr;
2597 else if (Succ == KeepEdge2)
2598 KeepEdge2 = nullptr;
2600 Succ->removePredecessor(OldTerm->getParent());
2603 IRBuilder<> Builder(OldTerm);
2604 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2606 // Insert an appropriate new terminator.
2607 if (!KeepEdge1 && !KeepEdge2) {
2608 if (TrueBB == FalseBB)
2609 // We were only looking for one successor, and it was present.
2610 // Create an unconditional branch to it.
2611 Builder.CreateBr(TrueBB);
2613 // We found both of the successors we were looking for.
2614 // Create a conditional branch sharing the condition of the select.
2615 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2616 if (TrueWeight != FalseWeight)
2617 NewBI->setMetadata(LLVMContext::MD_prof,
2618 MDBuilder(OldTerm->getContext()).
2619 createBranchWeights(TrueWeight, FalseWeight));
2621 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2622 // Neither of the selected blocks were successors, so this
2623 // terminator must be unreachable.
2624 new UnreachableInst(OldTerm->getContext(), OldTerm);
2626 // One of the selected values was a successor, but the other wasn't.
2627 // Insert an unconditional branch to the one that was found;
2628 // the edge to the one that wasn't must be unreachable.
2630 // Only TrueBB was found.
2631 Builder.CreateBr(TrueBB);
2633 // Only FalseBB was found.
2634 Builder.CreateBr(FalseBB);
2637 EraseTerminatorInstAndDCECond(OldTerm);
2641 // SimplifySwitchOnSelect - Replaces
2642 // (switch (select cond, X, Y)) on constant X, Y
2643 // with a branch - conditional if X and Y lead to distinct BBs,
2644 // unconditional otherwise.
2645 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2646 // Check for constant integer values in the select.
2647 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2648 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2649 if (!TrueVal || !FalseVal)
2652 // Find the relevant condition and destinations.
2653 Value *Condition = Select->getCondition();
2654 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2655 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2657 // Get weight for TrueBB and FalseBB.
2658 uint32_t TrueWeight = 0, FalseWeight = 0;
2659 SmallVector<uint64_t, 8> Weights;
2660 bool HasWeights = HasBranchWeights(SI);
2662 GetBranchWeights(SI, Weights);
2663 if (Weights.size() == 1 + SI->getNumCases()) {
2664 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2665 getSuccessorIndex()];
2666 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2667 getSuccessorIndex()];
2671 // Perform the actual simplification.
2672 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2673 TrueWeight, FalseWeight);
2676 // SimplifyIndirectBrOnSelect - Replaces
2677 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2678 // blockaddress(@fn, BlockB)))
2680 // (br cond, BlockA, BlockB).
2681 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2682 // Check that both operands of the select are block addresses.
2683 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2684 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2688 // Extract the actual blocks.
2689 BasicBlock *TrueBB = TBA->getBasicBlock();
2690 BasicBlock *FalseBB = FBA->getBasicBlock();
2692 // Perform the actual simplification.
2693 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2697 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2698 /// instruction (a seteq/setne with a constant) as the only instruction in a
2699 /// block that ends with an uncond branch. We are looking for a very specific
2700 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2701 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2702 /// default value goes to an uncond block with a seteq in it, we get something
2705 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2707 /// %tmp = icmp eq i8 %A, 92
2710 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2712 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2713 /// the PHI, merging the third icmp into the switch.
2714 static bool TryToSimplifyUncondBranchWithICmpInIt(
2715 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI,
2716 unsigned BonusInstThreshold, const DataLayout *DL, AssumptionTracker *AT) {
2717 BasicBlock *BB = ICI->getParent();
2719 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2721 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2723 Value *V = ICI->getOperand(0);
2724 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2726 // The pattern we're looking for is where our only predecessor is a switch on
2727 // 'V' and this block is the default case for the switch. In this case we can
2728 // fold the compared value into the switch to simplify things.
2729 BasicBlock *Pred = BB->getSinglePredecessor();
2730 if (!Pred || !isa<SwitchInst>(Pred->getTerminator())) return false;
2732 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2733 if (SI->getCondition() != V)
2736 // If BB is reachable on a non-default case, then we simply know the value of
2737 // V in this block. Substitute it and constant fold the icmp instruction
2739 if (SI->getDefaultDest() != BB) {
2740 ConstantInt *VVal = SI->findCaseDest(BB);
2741 assert(VVal && "Should have a unique destination value");
2742 ICI->setOperand(0, VVal);
2744 if (Value *V = SimplifyInstruction(ICI, DL)) {
2745 ICI->replaceAllUsesWith(V);
2746 ICI->eraseFromParent();
2748 // BB is now empty, so it is likely to simplify away.
2749 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
2752 // Ok, the block is reachable from the default dest. If the constant we're
2753 // comparing exists in one of the other edges, then we can constant fold ICI
2755 if (SI->findCaseValue(Cst) != SI->case_default()) {
2757 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2758 V = ConstantInt::getFalse(BB->getContext());
2760 V = ConstantInt::getTrue(BB->getContext());
2762 ICI->replaceAllUsesWith(V);
2763 ICI->eraseFromParent();
2764 // BB is now empty, so it is likely to simplify away.
2765 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
2768 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2770 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2771 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
2772 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
2773 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2776 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2778 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2779 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2781 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2782 std::swap(DefaultCst, NewCst);
2784 // Replace ICI (which is used by the PHI for the default value) with true or
2785 // false depending on if it is EQ or NE.
2786 ICI->replaceAllUsesWith(DefaultCst);
2787 ICI->eraseFromParent();
2789 // Okay, the switch goes to this block on a default value. Add an edge from
2790 // the switch to the merge point on the compared value.
2791 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2792 BB->getParent(), BB);
2793 SmallVector<uint64_t, 8> Weights;
2794 bool HasWeights = HasBranchWeights(SI);
2796 GetBranchWeights(SI, Weights);
2797 if (Weights.size() == 1 + SI->getNumCases()) {
2798 // Split weight for default case to case for "Cst".
2799 Weights[0] = (Weights[0]+1) >> 1;
2800 Weights.push_back(Weights[0]);
2802 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2803 SI->setMetadata(LLVMContext::MD_prof,
2804 MDBuilder(SI->getContext()).
2805 createBranchWeights(MDWeights));
2808 SI->addCase(Cst, NewBB);
2810 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2811 Builder.SetInsertPoint(NewBB);
2812 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2813 Builder.CreateBr(SuccBlock);
2814 PHIUse->addIncoming(NewCst, NewBB);
2818 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2819 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2820 /// fold it into a switch instruction if so.
2821 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *DL,
2822 IRBuilder<> &Builder) {
2823 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2824 if (!Cond) return false;
2826 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2827 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2828 // 'setne's and'ed together, collect them.
2830 // Try to gather values from a chain of and/or to be turned into a switch
2831 ConstantComparesGatherer ConstantCompare(Cond, DL);
2832 // Unpack the result
2833 SmallVectorImpl<ConstantInt*> &Values = ConstantCompare.Vals;
2834 Value *CompVal = ConstantCompare.CompValue;
2835 unsigned UsedICmps = ConstantCompare.UsedICmps;
2836 Value *ExtraCase = ConstantCompare.Extra;
2838 // If we didn't have a multiply compared value, fail.
2839 if (!CompVal) return false;
2841 // Avoid turning single icmps into a switch.
2845 bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
2847 // There might be duplicate constants in the list, which the switch
2848 // instruction can't handle, remove them now.
2849 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2850 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2852 // If Extra was used, we require at least two switch values to do the
2853 // transformation. A switch with one value is just an cond branch.
2854 if (ExtraCase && Values.size() < 2) return false;
2856 // TODO: Preserve branch weight metadata, similarly to how
2857 // FoldValueComparisonIntoPredecessors preserves it.
2859 // Figure out which block is which destination.
2860 BasicBlock *DefaultBB = BI->getSuccessor(1);
2861 BasicBlock *EdgeBB = BI->getSuccessor(0);
2862 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2864 BasicBlock *BB = BI->getParent();
2866 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2867 << " cases into SWITCH. BB is:\n" << *BB);
2869 // If there are any extra values that couldn't be folded into the switch
2870 // then we evaluate them with an explicit branch first. Split the block
2871 // right before the condbr to handle it.
2873 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2874 // Remove the uncond branch added to the old block.
2875 TerminatorInst *OldTI = BB->getTerminator();
2876 Builder.SetInsertPoint(OldTI);
2879 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2881 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2883 OldTI->eraseFromParent();
2885 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2886 // for the edge we just added.
2887 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2889 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2890 << "\nEXTRABB = " << *BB);
2894 Builder.SetInsertPoint(BI);
2895 // Convert pointer to int before we switch.
2896 if (CompVal->getType()->isPointerTy()) {
2897 assert(DL && "Cannot switch on pointer without DataLayout");
2898 CompVal = Builder.CreatePtrToInt(CompVal,
2899 DL->getIntPtrType(CompVal->getType()),
2903 // Create the new switch instruction now.
2904 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2906 // Add all of the 'cases' to the switch instruction.
2907 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2908 New->addCase(Values[i], EdgeBB);
2910 // We added edges from PI to the EdgeBB. As such, if there were any
2911 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2912 // the number of edges added.
2913 for (BasicBlock::iterator BBI = EdgeBB->begin();
2914 isa<PHINode>(BBI); ++BBI) {
2915 PHINode *PN = cast<PHINode>(BBI);
2916 Value *InVal = PN->getIncomingValueForBlock(BB);
2917 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2918 PN->addIncoming(InVal, BB);
2921 // Erase the old branch instruction.
2922 EraseTerminatorInstAndDCECond(BI);
2924 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2928 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2929 // If this is a trivial landing pad that just continues unwinding the caught
2930 // exception then zap the landing pad, turning its invokes into calls.
2931 BasicBlock *BB = RI->getParent();
2932 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2933 if (RI->getValue() != LPInst)
2934 // Not a landing pad, or the resume is not unwinding the exception that
2935 // caused control to branch here.
2938 // Check that there are no other instructions except for debug intrinsics.
2939 BasicBlock::iterator I = LPInst, E = RI;
2941 if (!isa<DbgInfoIntrinsic>(I))
2944 // Turn all invokes that unwind here into calls and delete the basic block.
2945 bool InvokeRequiresTableEntry = false;
2946 bool Changed = false;
2947 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2948 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2950 if (II->hasFnAttr(Attribute::UWTable)) {
2951 // Don't remove an `invoke' instruction if the ABI requires an entry into
2953 InvokeRequiresTableEntry = true;
2957 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2959 // Insert a call instruction before the invoke.
2960 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2962 Call->setCallingConv(II->getCallingConv());
2963 Call->setAttributes(II->getAttributes());
2964 Call->setDebugLoc(II->getDebugLoc());
2966 // Anything that used the value produced by the invoke instruction now uses
2967 // the value produced by the call instruction. Note that we do this even
2968 // for void functions and calls with no uses so that the callgraph edge is
2970 II->replaceAllUsesWith(Call);
2971 BB->removePredecessor(II->getParent());
2973 // Insert a branch to the normal destination right before the invoke.
2974 BranchInst::Create(II->getNormalDest(), II);
2976 // Finally, delete the invoke instruction!
2977 II->eraseFromParent();
2981 if (!InvokeRequiresTableEntry)
2982 // The landingpad is now unreachable. Zap it.
2983 BB->eraseFromParent();
2988 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2989 BasicBlock *BB = RI->getParent();
2990 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2992 // Find predecessors that end with branches.
2993 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2994 SmallVector<BranchInst*, 8> CondBranchPreds;
2995 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2996 BasicBlock *P = *PI;
2997 TerminatorInst *PTI = P->getTerminator();
2998 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2999 if (BI->isUnconditional())
3000 UncondBranchPreds.push_back(P);
3002 CondBranchPreds.push_back(BI);
3006 // If we found some, do the transformation!
3007 if (!UncondBranchPreds.empty() && DupRet) {
3008 while (!UncondBranchPreds.empty()) {
3009 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
3010 DEBUG(dbgs() << "FOLDING: " << *BB
3011 << "INTO UNCOND BRANCH PRED: " << *Pred);
3012 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
3015 // If we eliminated all predecessors of the block, delete the block now.
3016 if (pred_begin(BB) == pred_end(BB))
3017 // We know there are no successors, so just nuke the block.
3018 BB->eraseFromParent();
3023 // Check out all of the conditional branches going to this return
3024 // instruction. If any of them just select between returns, change the
3025 // branch itself into a select/return pair.
3026 while (!CondBranchPreds.empty()) {
3027 BranchInst *BI = CondBranchPreds.pop_back_val();
3029 // Check to see if the non-BB successor is also a return block.
3030 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
3031 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
3032 SimplifyCondBranchToTwoReturns(BI, Builder))
3038 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
3039 BasicBlock *BB = UI->getParent();
3041 bool Changed = false;
3043 // If there are any instructions immediately before the unreachable that can
3044 // be removed, do so.
3045 while (UI != BB->begin()) {
3046 BasicBlock::iterator BBI = UI;
3048 // Do not delete instructions that can have side effects which might cause
3049 // the unreachable to not be reachable; specifically, calls and volatile
3050 // operations may have this effect.
3051 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
3053 if (BBI->mayHaveSideEffects()) {
3054 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
3055 if (SI->isVolatile())
3057 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
3058 if (LI->isVolatile())
3060 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
3061 if (RMWI->isVolatile())
3063 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
3064 if (CXI->isVolatile())
3066 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
3067 !isa<LandingPadInst>(BBI)) {
3070 // Note that deleting LandingPad's here is in fact okay, although it
3071 // involves a bit of subtle reasoning. If this inst is a LandingPad,
3072 // all the predecessors of this block will be the unwind edges of Invokes,
3073 // and we can therefore guarantee this block will be erased.
3076 // Delete this instruction (any uses are guaranteed to be dead)
3077 if (!BBI->use_empty())
3078 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
3079 BBI->eraseFromParent();
3083 // If the unreachable instruction is the first in the block, take a gander
3084 // at all of the predecessors of this instruction, and simplify them.
3085 if (&BB->front() != UI) return Changed;
3087 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
3088 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
3089 TerminatorInst *TI = Preds[i]->getTerminator();
3090 IRBuilder<> Builder(TI);
3091 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3092 if (BI->isUnconditional()) {
3093 if (BI->getSuccessor(0) == BB) {
3094 new UnreachableInst(TI->getContext(), TI);
3095 TI->eraseFromParent();
3099 if (BI->getSuccessor(0) == BB) {
3100 Builder.CreateBr(BI->getSuccessor(1));
3101 EraseTerminatorInstAndDCECond(BI);
3102 } else if (BI->getSuccessor(1) == BB) {
3103 Builder.CreateBr(BI->getSuccessor(0));
3104 EraseTerminatorInstAndDCECond(BI);
3108 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3109 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3111 if (i.getCaseSuccessor() == BB) {
3112 BB->removePredecessor(SI->getParent());
3117 // If the default value is unreachable, figure out the most popular
3118 // destination and make it the default.
3119 if (SI->getDefaultDest() == BB) {
3120 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
3121 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3123 std::pair<unsigned, unsigned> &entry =
3124 Popularity[i.getCaseSuccessor()];
3125 if (entry.first == 0) {
3127 entry.second = i.getCaseIndex();
3133 // Find the most popular block.
3134 unsigned MaxPop = 0;
3135 unsigned MaxIndex = 0;
3136 BasicBlock *MaxBlock = nullptr;
3137 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
3138 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
3139 if (I->second.first > MaxPop ||
3140 (I->second.first == MaxPop && MaxIndex > I->second.second)) {
3141 MaxPop = I->second.first;
3142 MaxIndex = I->second.second;
3143 MaxBlock = I->first;
3147 // Make this the new default, allowing us to delete any explicit
3149 SI->setDefaultDest(MaxBlock);
3152 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
3154 if (isa<PHINode>(MaxBlock->begin()))
3155 for (unsigned i = 0; i != MaxPop-1; ++i)
3156 MaxBlock->removePredecessor(SI->getParent());
3158 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3160 if (i.getCaseSuccessor() == MaxBlock) {
3166 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
3167 if (II->getUnwindDest() == BB) {
3168 // Convert the invoke to a call instruction. This would be a good
3169 // place to note that the call does not throw though.
3170 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
3171 II->removeFromParent(); // Take out of symbol table
3173 // Insert the call now...
3174 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
3175 Builder.SetInsertPoint(BI);
3176 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
3177 Args, II->getName());
3178 CI->setCallingConv(II->getCallingConv());
3179 CI->setAttributes(II->getAttributes());
3180 // If the invoke produced a value, the call does now instead.
3181 II->replaceAllUsesWith(CI);
3188 // If this block is now dead, remove it.
3189 if (pred_begin(BB) == pred_end(BB) &&
3190 BB != &BB->getParent()->getEntryBlock()) {
3191 // We know there are no successors, so just nuke the block.
3192 BB->eraseFromParent();
3199 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
3200 /// integer range comparison into a sub, an icmp and a branch.
3201 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3202 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3204 // Make sure all cases point to the same destination and gather the values.
3205 SmallVector<ConstantInt *, 16> Cases;
3206 SwitchInst::CaseIt I = SI->case_begin();
3207 Cases.push_back(I.getCaseValue());
3208 SwitchInst::CaseIt PrevI = I++;
3209 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
3210 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
3212 Cases.push_back(I.getCaseValue());
3214 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
3216 // Sort the case values, then check if they form a range we can transform.
3217 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3218 for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
3219 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
3223 Constant *Offset = ConstantExpr::getNeg(Cases.back());
3224 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
3226 Value *Sub = SI->getCondition();
3227 if (!Offset->isNullValue())
3228 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
3230 // If NumCases overflowed, then all possible values jump to the successor.
3231 if (NumCases->isNullValue() && SI->getNumCases() != 0)
3232 Cmp = ConstantInt::getTrue(SI->getContext());
3234 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3235 BranchInst *NewBI = Builder.CreateCondBr(
3236 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
3238 // Update weight for the newly-created conditional branch.
3239 SmallVector<uint64_t, 8> Weights;
3240 bool HasWeights = HasBranchWeights(SI);
3242 GetBranchWeights(SI, Weights);
3243 if (Weights.size() == 1 + SI->getNumCases()) {
3244 // Combine all weights for the cases to be the true weight of NewBI.
3245 // We assume that the sum of all weights for a Terminator can fit into 32
3247 uint32_t NewTrueWeight = 0;
3248 for (unsigned I = 1, E = Weights.size(); I != E; ++I)
3249 NewTrueWeight += (uint32_t)Weights[I];
3250 NewBI->setMetadata(LLVMContext::MD_prof,
3251 MDBuilder(SI->getContext()).
3252 createBranchWeights(NewTrueWeight,
3253 (uint32_t)Weights[0]));
3257 // Prune obsolete incoming values off the successor's PHI nodes.
3258 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
3259 isa<PHINode>(BBI); ++BBI) {
3260 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
3261 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3263 SI->eraseFromParent();
3268 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
3269 /// and use it to remove dead cases.
3270 static bool EliminateDeadSwitchCases(SwitchInst *SI, const DataLayout *DL,
3271 AssumptionTracker *AT) {
3272 Value *Cond = SI->getCondition();
3273 unsigned Bits = Cond->getType()->getIntegerBitWidth();
3274 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3275 computeKnownBits(Cond, KnownZero, KnownOne, DL, 0, AT, SI);
3277 // Gather dead cases.
3278 SmallVector<ConstantInt*, 8> DeadCases;
3279 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3280 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3281 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3282 DeadCases.push_back(I.getCaseValue());
3283 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3284 << I.getCaseValue() << "' is dead.\n");
3288 SmallVector<uint64_t, 8> Weights;
3289 bool HasWeight = HasBranchWeights(SI);
3291 GetBranchWeights(SI, Weights);
3292 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3295 // Remove dead cases from the switch.
3296 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3297 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3298 assert(Case != SI->case_default() &&
3299 "Case was not found. Probably mistake in DeadCases forming.");
3301 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3305 // Prune unused values from PHI nodes.
3306 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3307 SI->removeCase(Case);
3309 if (HasWeight && Weights.size() >= 2) {
3310 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3311 SI->setMetadata(LLVMContext::MD_prof,
3312 MDBuilder(SI->getParent()->getContext()).
3313 createBranchWeights(MDWeights));
3316 return !DeadCases.empty();
3319 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
3320 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3321 /// by an unconditional branch), look at the phi node for BB in the successor
3322 /// block and see if the incoming value is equal to CaseValue. If so, return
3323 /// the phi node, and set PhiIndex to BB's index in the phi node.
3324 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3327 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3328 return nullptr; // BB must be empty to be a candidate for simplification.
3329 if (!BB->getSinglePredecessor())
3330 return nullptr; // BB must be dominated by the switch.
3332 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3333 if (!Branch || !Branch->isUnconditional())
3334 return nullptr; // Terminator must be unconditional branch.
3336 BasicBlock *Succ = Branch->getSuccessor(0);
3338 BasicBlock::iterator I = Succ->begin();
3339 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3340 int Idx = PHI->getBasicBlockIndex(BB);
3341 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3343 Value *InValue = PHI->getIncomingValue(Idx);
3344 if (InValue != CaseValue) continue;
3353 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
3354 /// instruction to a phi node dominated by the switch, if that would mean that
3355 /// some of the destination blocks of the switch can be folded away.
3356 /// Returns true if a change is made.
3357 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3358 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3359 ForwardingNodesMap ForwardingNodes;
3361 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3362 ConstantInt *CaseValue = I.getCaseValue();
3363 BasicBlock *CaseDest = I.getCaseSuccessor();
3366 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3370 ForwardingNodes[PHI].push_back(PhiIndex);
3373 bool Changed = false;
3375 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3376 E = ForwardingNodes.end(); I != E; ++I) {
3377 PHINode *Phi = I->first;
3378 SmallVectorImpl<int> &Indexes = I->second;
3380 if (Indexes.size() < 2) continue;
3382 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3383 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3390 /// ValidLookupTableConstant - Return true if the backend will be able to handle
3391 /// initializing an array of constants like C.
3392 static bool ValidLookupTableConstant(Constant *C) {
3393 if (C->isThreadDependent())
3395 if (C->isDLLImportDependent())
3398 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3399 return CE->isGEPWithNoNotionalOverIndexing();
3401 return isa<ConstantFP>(C) ||
3402 isa<ConstantInt>(C) ||
3403 isa<ConstantPointerNull>(C) ||
3404 isa<GlobalValue>(C) ||
3408 /// LookupConstant - If V is a Constant, return it. Otherwise, try to look up
3409 /// its constant value in ConstantPool, returning 0 if it's not there.
3410 static Constant *LookupConstant(Value *V,
3411 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3412 if (Constant *C = dyn_cast<Constant>(V))
3414 return ConstantPool.lookup(V);
3417 /// ConstantFold - Try to fold instruction I into a constant. This works for
3418 /// simple instructions such as binary operations where both operands are
3419 /// constant or can be replaced by constants from the ConstantPool. Returns the
3420 /// resulting constant on success, 0 otherwise.
3422 ConstantFold(Instruction *I,
3423 const SmallDenseMap<Value *, Constant *> &ConstantPool,
3424 const DataLayout *DL) {
3425 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3426 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3429 if (A->isAllOnesValue())
3430 return LookupConstant(Select->getTrueValue(), ConstantPool);
3431 if (A->isNullValue())
3432 return LookupConstant(Select->getFalseValue(), ConstantPool);
3436 SmallVector<Constant *, 4> COps;
3437 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
3438 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
3444 if (CmpInst *Cmp = dyn_cast<CmpInst>(I))
3445 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
3448 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), COps, DL);
3451 /// GetCaseResults - Try to determine the resulting constant values in phi nodes
3452 /// at the common destination basic block, *CommonDest, for one of the case
3453 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3454 /// case), of a switch instruction SI.
3456 GetCaseResults(SwitchInst *SI,
3457 ConstantInt *CaseVal,
3458 BasicBlock *CaseDest,
3459 BasicBlock **CommonDest,
3460 SmallVectorImpl<std::pair<PHINode *, Constant *> > &Res,
3461 const DataLayout *DL) {
3462 // The block from which we enter the common destination.
3463 BasicBlock *Pred = SI->getParent();
3465 // If CaseDest is empty except for some side-effect free instructions through
3466 // which we can constant-propagate the CaseVal, continue to its successor.
3467 SmallDenseMap<Value*, Constant*> ConstantPool;
3468 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3469 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3471 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3472 // If the terminator is a simple branch, continue to the next block.
3473 if (T->getNumSuccessors() != 1)
3476 CaseDest = T->getSuccessor(0);
3477 } else if (isa<DbgInfoIntrinsic>(I)) {
3478 // Skip debug intrinsic.
3480 } else if (Constant *C = ConstantFold(I, ConstantPool, DL)) {
3481 // Instruction is side-effect free and constant.
3482 ConstantPool.insert(std::make_pair(I, C));
3488 // If we did not have a CommonDest before, use the current one.
3490 *CommonDest = CaseDest;
3491 // If the destination isn't the common one, abort.
3492 if (CaseDest != *CommonDest)
3495 // Get the values for this case from phi nodes in the destination block.
3496 BasicBlock::iterator I = (*CommonDest)->begin();
3497 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3498 int Idx = PHI->getBasicBlockIndex(Pred);
3502 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3507 // Note: If the constant comes from constant-propagating the case value
3508 // through the CaseDest basic block, it will be safe to remove the
3509 // instructions in that block. They cannot be used (except in the phi nodes
3510 // we visit) outside CaseDest, because that block does not dominate its
3511 // successor. If it did, we would not be in this phi node.
3513 // Be conservative about which kinds of constants we support.
3514 if (!ValidLookupTableConstant(ConstVal))
3517 Res.push_back(std::make_pair(PHI, ConstVal));
3520 return Res.size() > 0;
3523 // MapCaseToResult - Helper function used to
3524 // add CaseVal to the list of cases that generate Result.
3525 static void MapCaseToResult(ConstantInt *CaseVal,
3526 SwitchCaseResultVectorTy &UniqueResults,
3528 for (auto &I : UniqueResults) {
3529 if (I.first == Result) {
3530 I.second.push_back(CaseVal);
3534 UniqueResults.push_back(std::make_pair(Result,
3535 SmallVector<ConstantInt*, 4>(1, CaseVal)));
3538 // InitializeUniqueCases - Helper function that initializes a map containing
3539 // results for the PHI node of the common destination block for a switch
3540 // instruction. Returns false if multiple PHI nodes have been found or if
3541 // there is not a common destination block for the switch.
3542 static bool InitializeUniqueCases(
3543 SwitchInst *SI, const DataLayout *DL, PHINode *&PHI,
3544 BasicBlock *&CommonDest,
3545 SwitchCaseResultVectorTy &UniqueResults,
3546 Constant *&DefaultResult) {
3547 for (auto &I : SI->cases()) {
3548 ConstantInt *CaseVal = I.getCaseValue();
3550 // Resulting value at phi nodes for this case value.
3551 SwitchCaseResultsTy Results;
3552 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
3556 // Only one value per case is permitted
3557 if (Results.size() > 1)
3559 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
3561 // Check the PHI consistency.
3563 PHI = Results[0].first;
3564 else if (PHI != Results[0].first)
3567 // Find the default result value.
3568 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
3569 BasicBlock *DefaultDest = SI->getDefaultDest();
3570 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
3572 // If the default value is not found abort unless the default destination
3575 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
3576 if ((!DefaultResult &&
3577 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
3583 // ConvertTwoCaseSwitch - Helper function that checks if it is possible to
3584 // transform a switch with only two cases (or two cases + default)
3585 // that produces a result into a value select.
3588 // case 10: %0 = icmp eq i32 %a, 10
3589 // return 10; %1 = select i1 %0, i32 10, i32 4
3590 // case 20: ----> %2 = icmp eq i32 %a, 20
3591 // return 2; %3 = select i1 %2, i32 2, i32 %1
3596 ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
3597 Constant *DefaultResult, Value *Condition,
3598 IRBuilder<> &Builder) {
3599 assert(ResultVector.size() == 2 &&
3600 "We should have exactly two unique results at this point");
3601 // If we are selecting between only two cases transform into a simple
3602 // select or a two-way select if default is possible.
3603 if (ResultVector[0].second.size() == 1 &&
3604 ResultVector[1].second.size() == 1) {
3605 ConstantInt *const FirstCase = ResultVector[0].second[0];
3606 ConstantInt *const SecondCase = ResultVector[1].second[0];
3608 bool DefaultCanTrigger = DefaultResult;
3609 Value *SelectValue = ResultVector[1].first;
3610 if (DefaultCanTrigger) {
3611 Value *const ValueCompare =
3612 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
3613 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
3614 DefaultResult, "switch.select");
3616 Value *const ValueCompare =
3617 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
3618 return Builder.CreateSelect(ValueCompare, ResultVector[0].first, SelectValue,
3625 // RemoveSwitchAfterSelectConversion - Helper function to cleanup a switch
3626 // instruction that has been converted into a select, fixing up PHI nodes and
3628 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
3630 IRBuilder<> &Builder) {
3631 BasicBlock *SelectBB = SI->getParent();
3632 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
3633 PHI->removeIncomingValue(SelectBB);
3634 PHI->addIncoming(SelectValue, SelectBB);
3636 Builder.CreateBr(PHI->getParent());
3638 // Remove the switch.
3639 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
3640 BasicBlock *Succ = SI->getSuccessor(i);
3642 if (Succ == PHI->getParent())
3644 Succ->removePredecessor(SelectBB);
3646 SI->eraseFromParent();
3649 /// SwitchToSelect - If the switch is only used to initialize one or more
3650 /// phi nodes in a common successor block with only two different
3651 /// constant values, replace the switch with select.
3652 static bool SwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
3653 const DataLayout *DL, AssumptionTracker *AT) {
3654 Value *const Cond = SI->getCondition();
3655 PHINode *PHI = nullptr;
3656 BasicBlock *CommonDest = nullptr;
3657 Constant *DefaultResult;
3658 SwitchCaseResultVectorTy UniqueResults;
3659 // Collect all the cases that will deliver the same value from the switch.
3660 if (!InitializeUniqueCases(SI, DL, PHI, CommonDest, UniqueResults,
3663 // Selects choose between maximum two values.
3664 if (UniqueResults.size() != 2)
3666 assert(PHI != nullptr && "PHI for value select not found");
3668 Builder.SetInsertPoint(SI);
3669 Value *SelectValue = ConvertTwoCaseSwitch(
3671 DefaultResult, Cond, Builder);
3673 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
3676 // The switch couldn't be converted into a select.
3681 /// SwitchLookupTable - This class represents a lookup table that can be used
3682 /// to replace a switch.
3683 class SwitchLookupTable {
3685 /// SwitchLookupTable - Create a lookup table to use as a switch replacement
3686 /// with the contents of Values, using DefaultValue to fill any holes in the
3688 SwitchLookupTable(Module &M,
3690 ConstantInt *Offset,
3691 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3692 Constant *DefaultValue,
3693 const DataLayout *DL);
3695 /// BuildLookup - Build instructions with Builder to retrieve the value at
3696 /// the position given by Index in the lookup table.
3697 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3699 /// WouldFitInRegister - Return true if a table with TableSize elements of
3700 /// type ElementType would fit in a target-legal register.
3701 static bool WouldFitInRegister(const DataLayout *DL,
3703 const Type *ElementType);
3706 // Depending on the contents of the table, it can be represented in
3709 // For tables where each element contains the same value, we just have to
3710 // store that single value and return it for each lookup.
3713 // For tables where there is a linear relationship between table index
3714 // and values. We calculate the result with a simple multiplication
3715 // and addition instead of a table lookup.
3718 // For small tables with integer elements, we can pack them into a bitmap
3719 // that fits into a target-legal register. Values are retrieved by
3720 // shift and mask operations.
3723 // The table is stored as an array of values. Values are retrieved by load
3724 // instructions from the table.
3728 // For SingleValueKind, this is the single value.
3729 Constant *SingleValue;
3731 // For BitMapKind, this is the bitmap.
3732 ConstantInt *BitMap;
3733 IntegerType *BitMapElementTy;
3735 // For LinearMapKind, these are the constants used to derive the value.
3736 ConstantInt *LinearOffset;
3737 ConstantInt *LinearMultiplier;
3739 // For ArrayKind, this is the array.
3740 GlobalVariable *Array;
3744 SwitchLookupTable::SwitchLookupTable(Module &M,
3746 ConstantInt *Offset,
3747 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3748 Constant *DefaultValue,
3749 const DataLayout *DL)
3750 : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
3751 LinearOffset(nullptr), LinearMultiplier(nullptr), Array(nullptr) {
3752 assert(Values.size() && "Can't build lookup table without values!");
3753 assert(TableSize >= Values.size() && "Can't fit values in table!");
3755 // If all values in the table are equal, this is that value.
3756 SingleValue = Values.begin()->second;
3758 Type *ValueType = Values.begin()->second->getType();
3760 // Build up the table contents.
3761 SmallVector<Constant*, 64> TableContents(TableSize);
3762 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3763 ConstantInt *CaseVal = Values[I].first;
3764 Constant *CaseRes = Values[I].second;
3765 assert(CaseRes->getType() == ValueType);
3767 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3769 TableContents[Idx] = CaseRes;
3771 if (CaseRes != SingleValue)
3772 SingleValue = nullptr;
3775 // Fill in any holes in the table with the default result.
3776 if (Values.size() < TableSize) {
3777 assert(DefaultValue &&
3778 "Need a default value to fill the lookup table holes.");
3779 assert(DefaultValue->getType() == ValueType);
3780 for (uint64_t I = 0; I < TableSize; ++I) {
3781 if (!TableContents[I])
3782 TableContents[I] = DefaultValue;
3785 if (DefaultValue != SingleValue)
3786 SingleValue = nullptr;
3789 // If each element in the table contains the same value, we only need to store
3790 // that single value.
3792 Kind = SingleValueKind;
3796 // Check if we can derive the value with a linear transformation from the
3798 if (isa<IntegerType>(ValueType)) {
3799 bool LinearMappingPossible = true;
3802 assert(TableSize >= 2 && "Should be a SingleValue table.");
3803 // Check if there is the same distance between two consecutive values.
3804 for (uint64_t I = 0; I < TableSize; ++I) {
3805 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
3807 // This is an undef. We could deal with it, but undefs in lookup tables
3808 // are very seldom. It's probably not worth the additional complexity.
3809 LinearMappingPossible = false;
3812 APInt Val = ConstVal->getValue();
3814 APInt Dist = Val - PrevVal;
3817 } else if (Dist != DistToPrev) {
3818 LinearMappingPossible = false;
3824 if (LinearMappingPossible) {
3825 LinearOffset = cast<ConstantInt>(TableContents[0]);
3826 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
3827 Kind = LinearMapKind;
3833 // If the type is integer and the table fits in a register, build a bitmap.
3834 if (WouldFitInRegister(DL, TableSize, ValueType)) {
3835 IntegerType *IT = cast<IntegerType>(ValueType);
3836 APInt TableInt(TableSize * IT->getBitWidth(), 0);
3837 for (uint64_t I = TableSize; I > 0; --I) {
3838 TableInt <<= IT->getBitWidth();
3839 // Insert values into the bitmap. Undef values are set to zero.
3840 if (!isa<UndefValue>(TableContents[I - 1])) {
3841 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
3842 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
3845 BitMap = ConstantInt::get(M.getContext(), TableInt);
3846 BitMapElementTy = IT;
3852 // Store the table in an array.
3853 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
3854 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3856 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3857 GlobalVariable::PrivateLinkage,
3860 Array->setUnnamedAddr(true);
3864 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
3866 case SingleValueKind:
3868 case LinearMapKind: {
3869 // Derive the result value from the input value.
3870 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
3871 false, "switch.idx.cast");
3872 if (!LinearMultiplier->isOne())
3873 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
3874 if (!LinearOffset->isZero())
3875 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
3879 // Type of the bitmap (e.g. i59).
3880 IntegerType *MapTy = BitMap->getType();
3882 // Cast Index to the same type as the bitmap.
3883 // Note: The Index is <= the number of elements in the table, so
3884 // truncating it to the width of the bitmask is safe.
3885 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
3887 // Multiply the shift amount by the element width.
3888 ShiftAmt = Builder.CreateMul(ShiftAmt,
3889 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
3893 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
3894 "switch.downshift");
3896 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
3900 // Make sure the table index will not overflow when treated as signed.
3901 IntegerType *IT = cast<IntegerType>(Index->getType());
3902 uint64_t TableSize = Array->getInitializer()->getType()
3903 ->getArrayNumElements();
3904 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
3905 Index = Builder.CreateZExt(Index,
3906 IntegerType::get(IT->getContext(),
3907 IT->getBitWidth() + 1),
3908 "switch.tableidx.zext");
3910 Value *GEPIndices[] = { Builder.getInt32(0), Index };
3911 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices,
3913 return Builder.CreateLoad(GEP, "switch.load");
3916 llvm_unreachable("Unknown lookup table kind!");
3919 bool SwitchLookupTable::WouldFitInRegister(const DataLayout *DL,
3921 const Type *ElementType) {
3924 const IntegerType *IT = dyn_cast<IntegerType>(ElementType);
3927 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
3928 // are <= 15, we could try to narrow the type.
3930 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
3931 if (TableSize >= UINT_MAX/IT->getBitWidth())
3933 return DL->fitsInLegalInteger(TableSize * IT->getBitWidth());
3936 /// ShouldBuildLookupTable - Determine whether a lookup table should be built
3937 /// for this switch, based on the number of cases, size of the table and the
3938 /// types of the results.
3939 static bool ShouldBuildLookupTable(SwitchInst *SI,
3941 const TargetTransformInfo &TTI,
3942 const DataLayout *DL,
3943 const SmallDenseMap<PHINode*, Type*>& ResultTypes) {
3944 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
3945 return false; // TableSize overflowed, or mul below might overflow.
3947 bool AllTablesFitInRegister = true;
3948 bool HasIllegalType = false;
3949 for (const auto &I : ResultTypes) {
3950 Type *Ty = I.second;
3952 // Saturate this flag to true.
3953 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
3955 // Saturate this flag to false.
3956 AllTablesFitInRegister = AllTablesFitInRegister &&
3957 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
3959 // If both flags saturate, we're done. NOTE: This *only* works with
3960 // saturating flags, and all flags have to saturate first due to the
3961 // non-deterministic behavior of iterating over a dense map.
3962 if (HasIllegalType && !AllTablesFitInRegister)
3966 // If each table would fit in a register, we should build it anyway.
3967 if (AllTablesFitInRegister)
3970 // Don't build a table that doesn't fit in-register if it has illegal types.
3974 // The table density should be at least 40%. This is the same criterion as for
3975 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3976 // FIXME: Find the best cut-off.
3977 return SI->getNumCases() * 10 >= TableSize * 4;
3980 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3981 /// phi nodes in a common successor block with different constant values,
3982 /// replace the switch with lookup tables.
3983 static bool SwitchToLookupTable(SwitchInst *SI,
3984 IRBuilder<> &Builder,
3985 const TargetTransformInfo &TTI,
3986 const DataLayout* DL) {
3987 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3989 // Only build lookup table when we have a target that supports it.
3990 if (!TTI.shouldBuildLookupTables())
3993 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3994 // split off a dense part and build a lookup table for that.
3996 // FIXME: This creates arrays of GEPs to constant strings, which means each
3997 // GEP needs a runtime relocation in PIC code. We should just build one big
3998 // string and lookup indices into that.
4000 // Ignore switches with less than three cases. Lookup tables will not make them
4001 // faster, so we don't analyze them.
4002 if (SI->getNumCases() < 3)
4005 // Figure out the corresponding result for each case value and phi node in the
4006 // common destination, as well as the the min and max case values.
4007 assert(SI->case_begin() != SI->case_end());
4008 SwitchInst::CaseIt CI = SI->case_begin();
4009 ConstantInt *MinCaseVal = CI.getCaseValue();
4010 ConstantInt *MaxCaseVal = CI.getCaseValue();
4012 BasicBlock *CommonDest = nullptr;
4013 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
4014 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
4015 SmallDenseMap<PHINode*, Constant*> DefaultResults;
4016 SmallDenseMap<PHINode*, Type*> ResultTypes;
4017 SmallVector<PHINode*, 4> PHIs;
4019 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
4020 ConstantInt *CaseVal = CI.getCaseValue();
4021 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
4022 MinCaseVal = CaseVal;
4023 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
4024 MaxCaseVal = CaseVal;
4026 // Resulting value at phi nodes for this case value.
4027 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
4029 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
4033 // Append the result from this case to the list for each phi.
4034 for (const auto &I : Results) {
4035 PHINode *PHI = I.first;
4036 Constant *Value = I.second;
4037 if (!ResultLists.count(PHI))
4038 PHIs.push_back(PHI);
4039 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
4043 // Keep track of the result types.
4044 for (PHINode *PHI : PHIs) {
4045 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
4048 uint64_t NumResults = ResultLists[PHIs[0]].size();
4049 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
4050 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
4051 bool TableHasHoles = (NumResults < TableSize);
4053 // If the table has holes, we need a constant result for the default case
4054 // or a bitmask that fits in a register.
4055 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
4056 bool HasDefaultResults = false;
4057 if (TableHasHoles) {
4058 HasDefaultResults = GetCaseResults(SI, nullptr, SI->getDefaultDest(),
4059 &CommonDest, DefaultResultsList, DL);
4062 bool NeedMask = (TableHasHoles && !HasDefaultResults);
4064 // As an extra penalty for the validity test we require more cases.
4065 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
4067 if (!(DL && DL->fitsInLegalInteger(TableSize)))
4071 for (const auto &I : DefaultResultsList) {
4072 PHINode *PHI = I.first;
4073 Constant *Result = I.second;
4074 DefaultResults[PHI] = Result;
4077 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
4080 // Create the BB that does the lookups.
4081 Module &Mod = *CommonDest->getParent()->getParent();
4082 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
4084 CommonDest->getParent(),
4087 // Compute the table index value.
4088 Builder.SetInsertPoint(SI);
4089 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
4092 // Compute the maximum table size representable by the integer type we are
4094 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
4095 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
4096 assert(MaxTableSize >= TableSize &&
4097 "It is impossible for a switch to have more entries than the max "
4098 "representable value of its input integer type's size.");
4100 // If we have a fully covered lookup table, unconditionally branch to the
4101 // lookup table BB. Otherwise, check if the condition value is within the case
4102 // range. If it is so, branch to the new BB. Otherwise branch to SI's default
4104 const bool GeneratingCoveredLookupTable = MaxTableSize == TableSize;
4105 if (GeneratingCoveredLookupTable) {
4106 Builder.CreateBr(LookupBB);
4107 // We cached PHINodes in PHIs, to avoid accessing deleted PHINodes later,
4108 // do not delete PHINodes here.
4109 SI->getDefaultDest()->removePredecessor(SI->getParent(),
4110 true/*DontDeleteUselessPHIs*/);
4112 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
4113 MinCaseVal->getType(), TableSize));
4114 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
4117 // Populate the BB that does the lookups.
4118 Builder.SetInsertPoint(LookupBB);
4121 // Before doing the lookup we do the hole check.
4122 // The LookupBB is therefore re-purposed to do the hole check
4123 // and we create a new LookupBB.
4124 BasicBlock *MaskBB = LookupBB;
4125 MaskBB->setName("switch.hole_check");
4126 LookupBB = BasicBlock::Create(Mod.getContext(),
4128 CommonDest->getParent(),
4131 // Make the mask's bitwidth at least 8bit and a power-of-2 to avoid
4132 // unnecessary illegal types.
4133 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
4134 APInt MaskInt(TableSizePowOf2, 0);
4135 APInt One(TableSizePowOf2, 1);
4136 // Build bitmask; fill in a 1 bit for every case.
4137 const ResultListTy &ResultList = ResultLists[PHIs[0]];
4138 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
4139 uint64_t Idx = (ResultList[I].first->getValue() -
4140 MinCaseVal->getValue()).getLimitedValue();
4141 MaskInt |= One << Idx;
4143 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
4145 // Get the TableIndex'th bit of the bitmask.
4146 // If this bit is 0 (meaning hole) jump to the default destination,
4147 // else continue with table lookup.
4148 IntegerType *MapTy = TableMask->getType();
4149 Value *MaskIndex = Builder.CreateZExtOrTrunc(TableIndex, MapTy,
4150 "switch.maskindex");
4151 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex,
4153 Value *LoBit = Builder.CreateTrunc(Shifted,
4154 Type::getInt1Ty(Mod.getContext()),
4156 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
4158 Builder.SetInsertPoint(LookupBB);
4159 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
4162 bool ReturnedEarly = false;
4163 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
4164 PHINode *PHI = PHIs[I];
4166 // If using a bitmask, use any value to fill the lookup table holes.
4167 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
4168 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI],
4171 Value *Result = Table.BuildLookup(TableIndex, Builder);
4173 // If the result is used to return immediately from the function, we want to
4174 // do that right here.
4175 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
4176 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
4177 Builder.CreateRet(Result);
4178 ReturnedEarly = true;
4182 PHI->addIncoming(Result, LookupBB);
4186 Builder.CreateBr(CommonDest);
4188 // Remove the switch.
4189 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4190 BasicBlock *Succ = SI->getSuccessor(i);
4192 if (Succ == SI->getDefaultDest())
4194 Succ->removePredecessor(SI->getParent());
4196 SI->eraseFromParent();
4200 ++NumLookupTablesHoles;
4204 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
4205 BasicBlock *BB = SI->getParent();
4207 if (isValueEqualityComparison(SI)) {
4208 // If we only have one predecessor, and if it is a branch on this value,
4209 // see if that predecessor totally determines the outcome of this switch.
4210 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4211 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
4212 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4214 Value *Cond = SI->getCondition();
4215 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
4216 if (SimplifySwitchOnSelect(SI, Select))
4217 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4219 // If the block only contains the switch, see if we can fold the block
4220 // away into any preds.
4221 BasicBlock::iterator BBI = BB->begin();
4222 // Ignore dbg intrinsics.
4223 while (isa<DbgInfoIntrinsic>(BBI))
4226 if (FoldValueComparisonIntoPredecessors(SI, Builder))
4227 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4230 // Try to transform the switch into an icmp and a branch.
4231 if (TurnSwitchRangeIntoICmp(SI, Builder))
4232 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4234 // Remove unreachable cases.
4235 if (EliminateDeadSwitchCases(SI, DL, AT))
4236 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4238 if (SwitchToSelect(SI, Builder, DL, AT))
4239 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4241 if (ForwardSwitchConditionToPHI(SI))
4242 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4244 if (SwitchToLookupTable(SI, Builder, TTI, DL))
4245 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4250 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
4251 BasicBlock *BB = IBI->getParent();
4252 bool Changed = false;
4254 // Eliminate redundant destinations.
4255 SmallPtrSet<Value *, 8> Succs;
4256 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
4257 BasicBlock *Dest = IBI->getDestination(i);
4258 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
4259 Dest->removePredecessor(BB);
4260 IBI->removeDestination(i);
4266 if (IBI->getNumDestinations() == 0) {
4267 // If the indirectbr has no successors, change it to unreachable.
4268 new UnreachableInst(IBI->getContext(), IBI);
4269 EraseTerminatorInstAndDCECond(IBI);
4273 if (IBI->getNumDestinations() == 1) {
4274 // If the indirectbr has one successor, change it to a direct branch.
4275 BranchInst::Create(IBI->getDestination(0), IBI);
4276 EraseTerminatorInstAndDCECond(IBI);
4280 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
4281 if (SimplifyIndirectBrOnSelect(IBI, SI))
4282 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4287 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
4288 BasicBlock *BB = BI->getParent();
4290 if (SinkCommon && SinkThenElseCodeToEnd(BI))
4293 // If the Terminator is the only non-phi instruction, simplify the block.
4294 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg();
4295 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
4296 TryToSimplifyUncondBranchFromEmptyBlock(BB))
4299 // If the only instruction in the block is a seteq/setne comparison
4300 // against a constant, try to simplify the block.
4301 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
4302 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
4303 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
4305 if (I->isTerminator() &&
4306 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI,
4307 BonusInstThreshold, DL, AT))
4311 // If this basic block is ONLY a compare and a branch, and if a predecessor
4312 // branches to us and our successor, fold the comparison into the
4313 // predecessor and use logical operations to update the incoming value
4314 // for PHI nodes in common successor.
4315 if (FoldBranchToCommonDest(BI, DL, BonusInstThreshold))
4316 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4321 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
4322 BasicBlock *BB = BI->getParent();
4324 // Conditional branch
4325 if (isValueEqualityComparison(BI)) {
4326 // If we only have one predecessor, and if it is a branch on this value,
4327 // see if that predecessor totally determines the outcome of this
4329 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4330 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
4331 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4333 // This block must be empty, except for the setcond inst, if it exists.
4334 // Ignore dbg intrinsics.
4335 BasicBlock::iterator I = BB->begin();
4336 // Ignore dbg intrinsics.
4337 while (isa<DbgInfoIntrinsic>(I))
4340 if (FoldValueComparisonIntoPredecessors(BI, Builder))
4341 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4342 } else if (&*I == cast<Instruction>(BI->getCondition())){
4344 // Ignore dbg intrinsics.
4345 while (isa<DbgInfoIntrinsic>(I))
4347 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
4348 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4352 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
4353 if (SimplifyBranchOnICmpChain(BI, DL, Builder))
4356 // If this basic block is ONLY a compare and a branch, and if a predecessor
4357 // branches to us and one of our successors, fold the comparison into the
4358 // predecessor and use logical operations to pick the right destination.
4359 if (FoldBranchToCommonDest(BI, DL, BonusInstThreshold))
4360 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4362 // We have a conditional branch to two blocks that are only reachable
4363 // from BI. We know that the condbr dominates the two blocks, so see if
4364 // there is any identical code in the "then" and "else" blocks. If so, we
4365 // can hoist it up to the branching block.
4366 if (BI->getSuccessor(0)->getSinglePredecessor()) {
4367 if (BI->getSuccessor(1)->getSinglePredecessor()) {
4368 if (HoistThenElseCodeToIf(BI, DL))
4369 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4371 // If Successor #1 has multiple preds, we may be able to conditionally
4372 // execute Successor #0 if it branches to Successor #1.
4373 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
4374 if (Succ0TI->getNumSuccessors() == 1 &&
4375 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
4376 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), DL))
4377 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4379 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
4380 // If Successor #0 has multiple preds, we may be able to conditionally
4381 // execute Successor #1 if it branches to Successor #0.
4382 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
4383 if (Succ1TI->getNumSuccessors() == 1 &&
4384 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
4385 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), DL))
4386 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4389 // If this is a branch on a phi node in the current block, thread control
4390 // through this block if any PHI node entries are constants.
4391 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
4392 if (PN->getParent() == BI->getParent())
4393 if (FoldCondBranchOnPHI(BI, DL))
4394 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4396 // Scan predecessor blocks for conditional branches.
4397 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
4398 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
4399 if (PBI != BI && PBI->isConditional())
4400 if (SimplifyCondBranchToCondBranch(PBI, BI))
4401 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4406 /// Check if passing a value to an instruction will cause undefined behavior.
4407 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
4408 Constant *C = dyn_cast<Constant>(V);
4415 if (C->isNullValue()) {
4416 // Only look at the first use, avoid hurting compile time with long uselists
4417 User *Use = *I->user_begin();
4419 // Now make sure that there are no instructions in between that can alter
4420 // control flow (eg. calls)
4421 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
4422 if (i == I->getParent()->end() || i->mayHaveSideEffects())
4425 // Look through GEPs. A load from a GEP derived from NULL is still undefined
4426 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
4427 if (GEP->getPointerOperand() == I)
4428 return passingValueIsAlwaysUndefined(V, GEP);
4430 // Look through bitcasts.
4431 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
4432 return passingValueIsAlwaysUndefined(V, BC);
4434 // Load from null is undefined.
4435 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
4436 if (!LI->isVolatile())
4437 return LI->getPointerAddressSpace() == 0;
4439 // Store to null is undefined.
4440 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
4441 if (!SI->isVolatile())
4442 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
4447 /// If BB has an incoming value that will always trigger undefined behavior
4448 /// (eg. null pointer dereference), remove the branch leading here.
4449 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
4450 for (BasicBlock::iterator i = BB->begin();
4451 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
4452 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
4453 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
4454 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
4455 IRBuilder<> Builder(T);
4456 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
4457 BB->removePredecessor(PHI->getIncomingBlock(i));
4458 // Turn uncoditional branches into unreachables and remove the dead
4459 // destination from conditional branches.
4460 if (BI->isUnconditional())
4461 Builder.CreateUnreachable();
4463 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
4464 BI->getSuccessor(0));
4465 BI->eraseFromParent();
4468 // TODO: SwitchInst.
4474 bool SimplifyCFGOpt::run(BasicBlock *BB) {
4475 bool Changed = false;
4477 assert(BB && BB->getParent() && "Block not embedded in function!");
4478 assert(BB->getTerminator() && "Degenerate basic block encountered!");
4480 // Remove basic blocks that have no predecessors (except the entry block)...
4481 // or that just have themself as a predecessor. These are unreachable.
4482 if ((pred_begin(BB) == pred_end(BB) &&
4483 BB != &BB->getParent()->getEntryBlock()) ||
4484 BB->getSinglePredecessor() == BB) {
4485 DEBUG(dbgs() << "Removing BB: \n" << *BB);
4486 DeleteDeadBlock(BB);
4490 // Check to see if we can constant propagate this terminator instruction
4492 Changed |= ConstantFoldTerminator(BB, true);
4494 // Check for and eliminate duplicate PHI nodes in this block.
4495 Changed |= EliminateDuplicatePHINodes(BB);
4497 // Check for and remove branches that will always cause undefined behavior.
4498 Changed |= removeUndefIntroducingPredecessor(BB);
4500 // Merge basic blocks into their predecessor if there is only one distinct
4501 // pred, and if there is only one distinct successor of the predecessor, and
4502 // if there are no PHI nodes.
4504 if (MergeBlockIntoPredecessor(BB))
4507 IRBuilder<> Builder(BB);
4509 // If there is a trivial two-entry PHI node in this basic block, and we can
4510 // eliminate it, do so now.
4511 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
4512 if (PN->getNumIncomingValues() == 2)
4513 Changed |= FoldTwoEntryPHINode(PN, DL);
4515 Builder.SetInsertPoint(BB->getTerminator());
4516 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
4517 if (BI->isUnconditional()) {
4518 if (SimplifyUncondBranch(BI, Builder)) return true;
4520 if (SimplifyCondBranch(BI, Builder)) return true;
4522 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
4523 if (SimplifyReturn(RI, Builder)) return true;
4524 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
4525 if (SimplifyResume(RI, Builder)) return true;
4526 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
4527 if (SimplifySwitch(SI, Builder)) return true;
4528 } else if (UnreachableInst *UI =
4529 dyn_cast<UnreachableInst>(BB->getTerminator())) {
4530 if (SimplifyUnreachable(UI)) return true;
4531 } else if (IndirectBrInst *IBI =
4532 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
4533 if (SimplifyIndirectBr(IBI)) return true;
4539 /// SimplifyCFG - This function is used to do simplification of a CFG. For
4540 /// example, it adjusts branches to branches to eliminate the extra hop, it
4541 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4542 /// of the CFG. It returns true if a modification was made.
4544 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
4545 unsigned BonusInstThreshold,
4546 const DataLayout *DL, AssumptionTracker *AT) {
4547 return SimplifyCFGOpt(TTI, BonusInstThreshold, DL, AT).run(BB);