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 = nullptr; /// Value found for the switch comparison
373 Value *Extra = nullptr; /// Extra clause to be checked before the switch
374 SmallVector<ConstantInt*, 8> Vals; /// Set of integers to match in switch
375 unsigned UsedICmps = 0; /// 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) {
383 ConstantComparesGatherer(const ConstantComparesGatherer&) = delete;
384 ConstantComparesGatherer &operator=(const ConstantComparesGatherer&) = delete;
388 /// Try to set the current value used for the comparison, it succeeds only if
389 /// it wasn't set before or if the new value is the same as the old one
390 bool setValueOnce(Value *NewVal) {
391 if(CompValue && CompValue != NewVal) return false;
392 return CompValue = NewVal;
395 /// Try to match Instruction "I" as a comparison against a constant and
396 /// populates the array Vals with the set of values that match (or do not
397 /// match depending on isEQ).
398 /// Return false on failure. On success, the Value the comparison matched
399 /// against is placed in CompValue.
400 /// If CompValue is already set, the function is expected to fail if a match
401 /// is found but the value compared to is different.
402 bool matchInstruction(Instruction *I, const DataLayout *DL, bool isEQ) {
403 // If this is an icmp against a constant, handle this as one of the cases.
406 if (!((ICI = dyn_cast<ICmpInst>(I)) &&
407 (C = GetConstantInt(I->getOperand(1), DL)))) {
414 // Pattern match a special case
415 // (x & ~2^x) == y --> x == y || x == y|2^x
416 // This undoes a transformation done by instcombine to fuse 2 compares.
417 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
418 if (match(ICI->getOperand(0),
419 m_And(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
420 APInt Not = ~RHSC->getValue();
421 if (Not.isPowerOf2()) {
422 // If we already have a value for the switch, it has to match!
423 if(!setValueOnce(RHSVal))
427 Vals.push_back(ConstantInt::get(C->getContext(),
428 C->getValue() | Not));
434 // If we already have a value for the switch, it has to match!
435 if(!setValueOnce(ICI->getOperand(0)))
440 return ICI->getOperand(0);
443 // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
444 ConstantRange Span = ConstantRange::makeICmpRegion(ICI->getPredicate(),
447 // Shift the range if the compare is fed by an add. This is the range
448 // compare idiom as emitted by instcombine.
449 Value *CandidateVal = I->getOperand(0);
450 if(match(I->getOperand(0), m_Add(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
451 Span = Span.subtract(RHSC->getValue());
452 CandidateVal = RHSVal;
455 // If this is an and/!= check, then we are looking to build the set of
456 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
459 Span = Span.inverse();
461 // If there are a ton of values, we don't want to make a ginormous switch.
462 if (Span.getSetSize().ugt(8) || Span.isEmptySet()) {
466 // If we already have a value for the switch, it has to match!
467 if(!setValueOnce(CandidateVal))
470 // Add all values from the range to the set
471 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
472 Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
479 /// gather - Given a potentially 'or'd or 'and'd together collection of icmp
480 /// eq/ne/lt/gt instructions that compare a value against a constant, extract
481 /// the value being compared, and stick the list constants into the Vals
483 /// One "Extra" case is allowed to differ from the other.
484 void gather(Value *V, const DataLayout *DL) {
485 Instruction *I = dyn_cast<Instruction>(V);
486 bool isEQ = (I->getOpcode() == Instruction::Or);
488 // Keep a stack (SmallVector for efficiency) for depth-first traversal
489 SmallVector<Value *, 8> DFT;
494 while(!DFT.empty()) {
495 V = DFT.pop_back_val();
497 if (Instruction *I = dyn_cast<Instruction>(V)) {
498 // If it is a || (or && depending on isEQ), process the operands.
499 if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
500 DFT.push_back(I->getOperand(1));
501 DFT.push_back(I->getOperand(0));
505 // Try to match the current instruction
506 if (matchInstruction(I, DL, isEQ))
507 // Match succeed, continue the loop
511 // One element of the sequence of || (or &&) could not be match as a
512 // comparison against the same value as the others.
513 // We allow only one "Extra" case to be checked before the switch
518 // Failed to parse a proper sequence, abort now
525 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
526 Instruction *Cond = nullptr;
527 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
528 Cond = dyn_cast<Instruction>(SI->getCondition());
529 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
530 if (BI->isConditional())
531 Cond = dyn_cast<Instruction>(BI->getCondition());
532 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
533 Cond = dyn_cast<Instruction>(IBI->getAddress());
536 TI->eraseFromParent();
537 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
540 /// isValueEqualityComparison - Return true if the specified terminator checks
541 /// to see if a value is equal to constant integer value.
542 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
544 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
545 // Do not permit merging of large switch instructions into their
546 // predecessors unless there is only one predecessor.
547 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()),
548 pred_end(SI->getParent())) <= 128)
549 CV = SI->getCondition();
550 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
551 if (BI->isConditional() && BI->getCondition()->hasOneUse())
552 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition()))
553 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
554 CV = ICI->getOperand(0);
556 // Unwrap any lossless ptrtoint cast.
558 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
559 Value *Ptr = PTII->getPointerOperand();
560 if (PTII->getType() == DL->getIntPtrType(Ptr->getType()))
567 /// GetValueEqualityComparisonCases - Given a value comparison instruction,
568 /// decode all of the 'cases' that it represents and return the 'default' block.
569 BasicBlock *SimplifyCFGOpt::
570 GetValueEqualityComparisonCases(TerminatorInst *TI,
571 std::vector<ValueEqualityComparisonCase>
573 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
574 Cases.reserve(SI->getNumCases());
575 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
576 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
577 i.getCaseSuccessor()));
578 return SI->getDefaultDest();
581 BranchInst *BI = cast<BranchInst>(TI);
582 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
583 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
584 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
587 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
591 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries
592 /// in the list that match the specified block.
593 static void EliminateBlockCases(BasicBlock *BB,
594 std::vector<ValueEqualityComparisonCase> &Cases) {
595 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
598 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as
601 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
602 std::vector<ValueEqualityComparisonCase > &C2) {
603 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
605 // Make V1 be smaller than V2.
606 if (V1->size() > V2->size())
609 if (V1->size() == 0) return false;
610 if (V1->size() == 1) {
612 ConstantInt *TheVal = (*V1)[0].Value;
613 for (unsigned i = 0, e = V2->size(); i != e; ++i)
614 if (TheVal == (*V2)[i].Value)
618 // Otherwise, just sort both lists and compare element by element.
619 array_pod_sort(V1->begin(), V1->end());
620 array_pod_sort(V2->begin(), V2->end());
621 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
622 while (i1 != e1 && i2 != e2) {
623 if ((*V1)[i1].Value == (*V2)[i2].Value)
625 if ((*V1)[i1].Value < (*V2)[i2].Value)
633 /// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a
634 /// terminator instruction and its block is known to only have a single
635 /// predecessor block, check to see if that predecessor is also a value
636 /// comparison with the same value, and if that comparison determines the
637 /// outcome of this comparison. If so, simplify TI. This does a very limited
638 /// form of jump threading.
639 bool SimplifyCFGOpt::
640 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
642 IRBuilder<> &Builder) {
643 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
644 if (!PredVal) return false; // Not a value comparison in predecessor.
646 Value *ThisVal = isValueEqualityComparison(TI);
647 assert(ThisVal && "This isn't a value comparison!!");
648 if (ThisVal != PredVal) return false; // Different predicates.
650 // TODO: Preserve branch weight metadata, similarly to how
651 // FoldValueComparisonIntoPredecessors preserves it.
653 // Find out information about when control will move from Pred to TI's block.
654 std::vector<ValueEqualityComparisonCase> PredCases;
655 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
657 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
659 // Find information about how control leaves this block.
660 std::vector<ValueEqualityComparisonCase> ThisCases;
661 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
662 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
664 // If TI's block is the default block from Pred's comparison, potentially
665 // simplify TI based on this knowledge.
666 if (PredDef == TI->getParent()) {
667 // If we are here, we know that the value is none of those cases listed in
668 // PredCases. If there are any cases in ThisCases that are in PredCases, we
670 if (!ValuesOverlap(PredCases, ThisCases))
673 if (isa<BranchInst>(TI)) {
674 // Okay, one of the successors of this condbr is dead. Convert it to a
676 assert(ThisCases.size() == 1 && "Branch can only have one case!");
677 // Insert the new branch.
678 Instruction *NI = Builder.CreateBr(ThisDef);
681 // Remove PHI node entries for the dead edge.
682 ThisCases[0].Dest->removePredecessor(TI->getParent());
684 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
685 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
687 EraseTerminatorInstAndDCECond(TI);
691 SwitchInst *SI = cast<SwitchInst>(TI);
692 // Okay, TI has cases that are statically dead, prune them away.
693 SmallPtrSet<Constant*, 16> DeadCases;
694 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
695 DeadCases.insert(PredCases[i].Value);
697 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
698 << "Through successor TI: " << *TI);
700 // Collect branch weights into a vector.
701 SmallVector<uint32_t, 8> Weights;
702 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
703 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
705 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
707 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i));
709 Weights.push_back(CI->getValue().getZExtValue());
711 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
713 if (DeadCases.count(i.getCaseValue())) {
715 std::swap(Weights[i.getCaseIndex()+1], Weights.back());
718 i.getCaseSuccessor()->removePredecessor(TI->getParent());
722 if (HasWeight && Weights.size() >= 2)
723 SI->setMetadata(LLVMContext::MD_prof,
724 MDBuilder(SI->getParent()->getContext()).
725 createBranchWeights(Weights));
727 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
731 // Otherwise, TI's block must correspond to some matched value. Find out
732 // which value (or set of values) this is.
733 ConstantInt *TIV = nullptr;
734 BasicBlock *TIBB = TI->getParent();
735 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
736 if (PredCases[i].Dest == TIBB) {
738 return false; // Cannot handle multiple values coming to this block.
739 TIV = PredCases[i].Value;
741 assert(TIV && "No edge from pred to succ?");
743 // Okay, we found the one constant that our value can be if we get into TI's
744 // BB. Find out which successor will unconditionally be branched to.
745 BasicBlock *TheRealDest = nullptr;
746 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
747 if (ThisCases[i].Value == TIV) {
748 TheRealDest = ThisCases[i].Dest;
752 // If not handled by any explicit cases, it is handled by the default case.
753 if (!TheRealDest) TheRealDest = ThisDef;
755 // Remove PHI node entries for dead edges.
756 BasicBlock *CheckEdge = TheRealDest;
757 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
758 if (*SI != CheckEdge)
759 (*SI)->removePredecessor(TIBB);
763 // Insert the new branch.
764 Instruction *NI = Builder.CreateBr(TheRealDest);
767 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
768 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
770 EraseTerminatorInstAndDCECond(TI);
775 /// ConstantIntOrdering - This class implements a stable ordering of constant
776 /// integers that does not depend on their address. This is important for
777 /// applications that sort ConstantInt's to ensure uniqueness.
778 struct ConstantIntOrdering {
779 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
780 return LHS->getValue().ult(RHS->getValue());
785 static int ConstantIntSortPredicate(ConstantInt *const *P1,
786 ConstantInt *const *P2) {
787 const ConstantInt *LHS = *P1;
788 const ConstantInt *RHS = *P2;
789 if (LHS->getValue().ult(RHS->getValue()))
791 if (LHS->getValue() == RHS->getValue())
796 static inline bool HasBranchWeights(const Instruction* I) {
797 MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
798 if (ProfMD && ProfMD->getOperand(0))
799 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
800 return MDS->getString().equals("branch_weights");
805 /// Get Weights of a given TerminatorInst, the default weight is at the front
806 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
808 static void GetBranchWeights(TerminatorInst *TI,
809 SmallVectorImpl<uint64_t> &Weights) {
810 MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
812 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
813 ConstantInt *CI = cast<ConstantInt>(MD->getOperand(i));
814 Weights.push_back(CI->getValue().getZExtValue());
817 // If TI is a conditional eq, the default case is the false case,
818 // and the corresponding branch-weight data is at index 2. We swap the
819 // default weight to be the first entry.
820 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
821 assert(Weights.size() == 2);
822 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
823 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
824 std::swap(Weights.front(), Weights.back());
828 /// Keep halving the weights until all can fit in uint32_t.
829 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
830 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
831 if (Max > UINT_MAX) {
832 unsigned Offset = 32 - countLeadingZeros(Max);
833 for (uint64_t &I : Weights)
838 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value
839 /// equality comparison instruction (either a switch or a branch on "X == c").
840 /// See if any of the predecessors of the terminator block are value comparisons
841 /// on the same value. If so, and if safe to do so, fold them together.
842 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
843 IRBuilder<> &Builder) {
844 BasicBlock *BB = TI->getParent();
845 Value *CV = isValueEqualityComparison(TI); // CondVal
846 assert(CV && "Not a comparison?");
847 bool Changed = false;
849 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
850 while (!Preds.empty()) {
851 BasicBlock *Pred = Preds.pop_back_val();
853 // See if the predecessor is a comparison with the same value.
854 TerminatorInst *PTI = Pred->getTerminator();
855 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
857 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
858 // Figure out which 'cases' to copy from SI to PSI.
859 std::vector<ValueEqualityComparisonCase> BBCases;
860 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
862 std::vector<ValueEqualityComparisonCase> PredCases;
863 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
865 // Based on whether the default edge from PTI goes to BB or not, fill in
866 // PredCases and PredDefault with the new switch cases we would like to
868 SmallVector<BasicBlock*, 8> NewSuccessors;
870 // Update the branch weight metadata along the way
871 SmallVector<uint64_t, 8> Weights;
872 bool PredHasWeights = HasBranchWeights(PTI);
873 bool SuccHasWeights = HasBranchWeights(TI);
875 if (PredHasWeights) {
876 GetBranchWeights(PTI, Weights);
877 // branch-weight metadata is inconsistent here.
878 if (Weights.size() != 1 + PredCases.size())
879 PredHasWeights = SuccHasWeights = false;
880 } else if (SuccHasWeights)
881 // If there are no predecessor weights but there are successor weights,
882 // populate Weights with 1, which will later be scaled to the sum of
883 // successor's weights
884 Weights.assign(1 + PredCases.size(), 1);
886 SmallVector<uint64_t, 8> SuccWeights;
887 if (SuccHasWeights) {
888 GetBranchWeights(TI, SuccWeights);
889 // branch-weight metadata is inconsistent here.
890 if (SuccWeights.size() != 1 + BBCases.size())
891 PredHasWeights = SuccHasWeights = false;
892 } else if (PredHasWeights)
893 SuccWeights.assign(1 + BBCases.size(), 1);
895 if (PredDefault == BB) {
896 // If this is the default destination from PTI, only the edges in TI
897 // that don't occur in PTI, or that branch to BB will be activated.
898 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
899 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
900 if (PredCases[i].Dest != BB)
901 PTIHandled.insert(PredCases[i].Value);
903 // The default destination is BB, we don't need explicit targets.
904 std::swap(PredCases[i], PredCases.back());
906 if (PredHasWeights || SuccHasWeights) {
907 // Increase weight for the default case.
908 Weights[0] += Weights[i+1];
909 std::swap(Weights[i+1], Weights.back());
913 PredCases.pop_back();
917 // Reconstruct the new switch statement we will be building.
918 if (PredDefault != BBDefault) {
919 PredDefault->removePredecessor(Pred);
920 PredDefault = BBDefault;
921 NewSuccessors.push_back(BBDefault);
924 unsigned CasesFromPred = Weights.size();
925 uint64_t ValidTotalSuccWeight = 0;
926 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
927 if (!PTIHandled.count(BBCases[i].Value) &&
928 BBCases[i].Dest != BBDefault) {
929 PredCases.push_back(BBCases[i]);
930 NewSuccessors.push_back(BBCases[i].Dest);
931 if (SuccHasWeights || PredHasWeights) {
932 // The default weight is at index 0, so weight for the ith case
933 // should be at index i+1. Scale the cases from successor by
934 // PredDefaultWeight (Weights[0]).
935 Weights.push_back(Weights[0] * SuccWeights[i+1]);
936 ValidTotalSuccWeight += SuccWeights[i+1];
940 if (SuccHasWeights || PredHasWeights) {
941 ValidTotalSuccWeight += SuccWeights[0];
942 // Scale the cases from predecessor by ValidTotalSuccWeight.
943 for (unsigned i = 1; i < CasesFromPred; ++i)
944 Weights[i] *= ValidTotalSuccWeight;
945 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
946 Weights[0] *= SuccWeights[0];
949 // If this is not the default destination from PSI, only the edges
950 // in SI that occur in PSI with a destination of BB will be
952 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
953 std::map<ConstantInt*, uint64_t> WeightsForHandled;
954 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
955 if (PredCases[i].Dest == BB) {
956 PTIHandled.insert(PredCases[i].Value);
958 if (PredHasWeights || SuccHasWeights) {
959 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
960 std::swap(Weights[i+1], Weights.back());
964 std::swap(PredCases[i], PredCases.back());
965 PredCases.pop_back();
969 // Okay, now we know which constants were sent to BB from the
970 // predecessor. Figure out where they will all go now.
971 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
972 if (PTIHandled.count(BBCases[i].Value)) {
973 // If this is one we are capable of getting...
974 if (PredHasWeights || SuccHasWeights)
975 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
976 PredCases.push_back(BBCases[i]);
977 NewSuccessors.push_back(BBCases[i].Dest);
978 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
981 // If there are any constants vectored to BB that TI doesn't handle,
982 // they must go to the default destination of TI.
983 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
985 E = PTIHandled.end(); I != E; ++I) {
986 if (PredHasWeights || SuccHasWeights)
987 Weights.push_back(WeightsForHandled[*I]);
988 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
989 NewSuccessors.push_back(BBDefault);
993 // Okay, at this point, we know which new successor Pred will get. Make
994 // sure we update the number of entries in the PHI nodes for these
996 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i)
997 AddPredecessorToBlock(NewSuccessors[i], Pred, BB);
999 Builder.SetInsertPoint(PTI);
1000 // Convert pointer to int before we switch.
1001 if (CV->getType()->isPointerTy()) {
1002 assert(DL && "Cannot switch on pointer without DataLayout");
1003 CV = Builder.CreatePtrToInt(CV, DL->getIntPtrType(CV->getType()),
1007 // Now that the successors are updated, create the new Switch instruction.
1008 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
1010 NewSI->setDebugLoc(PTI->getDebugLoc());
1011 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1012 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest);
1014 if (PredHasWeights || SuccHasWeights) {
1015 // Halve the weights if any of them cannot fit in an uint32_t
1016 FitWeights(Weights);
1018 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1020 NewSI->setMetadata(LLVMContext::MD_prof,
1021 MDBuilder(BB->getContext()).
1022 createBranchWeights(MDWeights));
1025 EraseTerminatorInstAndDCECond(PTI);
1027 // Okay, last check. If BB is still a successor of PSI, then we must
1028 // have an infinite loop case. If so, add an infinitely looping block
1029 // to handle the case to preserve the behavior of the code.
1030 BasicBlock *InfLoopBlock = nullptr;
1031 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1032 if (NewSI->getSuccessor(i) == BB) {
1033 if (!InfLoopBlock) {
1034 // Insert it at the end of the function, because it's either code,
1035 // or it won't matter if it's hot. :)
1036 InfLoopBlock = BasicBlock::Create(BB->getContext(),
1037 "infloop", BB->getParent());
1038 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1040 NewSI->setSuccessor(i, InfLoopBlock);
1049 // isSafeToHoistInvoke - If we would need to insert a select that uses the
1050 // value of this invoke (comments in HoistThenElseCodeToIf explain why we
1051 // would need to do this), we can't hoist the invoke, as there is nowhere
1052 // to put the select in this case.
1053 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1054 Instruction *I1, Instruction *I2) {
1055 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1057 for (BasicBlock::iterator BBI = SI->begin();
1058 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1059 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1060 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1061 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
1069 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I);
1071 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
1072 /// BB2, hoist any common code in the two blocks up into the branch block. The
1073 /// caller of this function guarantees that BI's block dominates BB1 and BB2.
1074 static bool HoistThenElseCodeToIf(BranchInst *BI, const DataLayout *DL) {
1075 // This does very trivial matching, with limited scanning, to find identical
1076 // instructions in the two blocks. In particular, we don't want to get into
1077 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1078 // such, we currently just scan for obviously identical instructions in an
1080 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1081 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1083 BasicBlock::iterator BB1_Itr = BB1->begin();
1084 BasicBlock::iterator BB2_Itr = BB2->begin();
1086 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++;
1087 // Skip debug info if it is not identical.
1088 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1089 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1090 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1091 while (isa<DbgInfoIntrinsic>(I1))
1093 while (isa<DbgInfoIntrinsic>(I2))
1096 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1097 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1100 BasicBlock *BIParent = BI->getParent();
1102 bool Changed = false;
1104 // If we are hoisting the terminator instruction, don't move one (making a
1105 // broken BB), instead clone it, and remove BI.
1106 if (isa<TerminatorInst>(I1))
1107 goto HoistTerminator;
1109 // For a normal instruction, we just move one to right before the branch,
1110 // then replace all uses of the other with the first. Finally, we remove
1111 // the now redundant second instruction.
1112 BIParent->getInstList().splice(BI, BB1->getInstList(), I1);
1113 if (!I2->use_empty())
1114 I2->replaceAllUsesWith(I1);
1115 I1->intersectOptionalDataWith(I2);
1116 unsigned KnownIDs[] = {
1117 LLVMContext::MD_tbaa,
1118 LLVMContext::MD_range,
1119 LLVMContext::MD_fpmath,
1120 LLVMContext::MD_invariant_load,
1121 LLVMContext::MD_nonnull
1123 combineMetadata(I1, I2, KnownIDs);
1124 I2->eraseFromParent();
1129 // Skip debug info if it is not identical.
1130 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1131 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1132 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1133 while (isa<DbgInfoIntrinsic>(I1))
1135 while (isa<DbgInfoIntrinsic>(I2))
1138 } while (I1->isIdenticalToWhenDefined(I2));
1143 // It may not be possible to hoist an invoke.
1144 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1147 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1149 for (BasicBlock::iterator BBI = SI->begin();
1150 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1151 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1152 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1156 // Check for passingValueIsAlwaysUndefined here because we would rather
1157 // eliminate undefined control flow then converting it to a select.
1158 if (passingValueIsAlwaysUndefined(BB1V, PN) ||
1159 passingValueIsAlwaysUndefined(BB2V, PN))
1162 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V, DL))
1164 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V, DL))
1169 // Okay, it is safe to hoist the terminator.
1170 Instruction *NT = I1->clone();
1171 BIParent->getInstList().insert(BI, NT);
1172 if (!NT->getType()->isVoidTy()) {
1173 I1->replaceAllUsesWith(NT);
1174 I2->replaceAllUsesWith(NT);
1178 IRBuilder<true, NoFolder> Builder(NT);
1179 // Hoisting one of the terminators from our successor is a great thing.
1180 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1181 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1182 // nodes, so we insert select instruction to compute the final result.
1183 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1184 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1186 for (BasicBlock::iterator BBI = SI->begin();
1187 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1188 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1189 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1190 if (BB1V == BB2V) continue;
1192 // These values do not agree. Insert a select instruction before NT
1193 // that determines the right value.
1194 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1196 SI = cast<SelectInst>
1197 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1198 BB1V->getName()+"."+BB2V->getName()));
1200 // Make the PHI node use the select for all incoming values for BB1/BB2
1201 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1202 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1203 PN->setIncomingValue(i, SI);
1207 // Update any PHI nodes in our new successors.
1208 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1209 AddPredecessorToBlock(*SI, BIParent, BB1);
1211 EraseTerminatorInstAndDCECond(BI);
1215 /// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd,
1216 /// check whether BBEnd has only two predecessors and the other predecessor
1217 /// ends with an unconditional branch. If it is true, sink any common code
1218 /// in the two predecessors to BBEnd.
1219 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1220 assert(BI1->isUnconditional());
1221 BasicBlock *BB1 = BI1->getParent();
1222 BasicBlock *BBEnd = BI1->getSuccessor(0);
1224 // Check that BBEnd has two predecessors and the other predecessor ends with
1225 // an unconditional branch.
1226 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1227 BasicBlock *Pred0 = *PI++;
1228 if (PI == PE) // Only one predecessor.
1230 BasicBlock *Pred1 = *PI++;
1231 if (PI != PE) // More than two predecessors.
1233 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1234 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1235 if (!BI2 || !BI2->isUnconditional())
1238 // Gather the PHI nodes in BBEnd.
1239 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2;
1240 Instruction *FirstNonPhiInBBEnd = nullptr;
1241 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end();
1243 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1244 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1245 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1246 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN);
1248 FirstNonPhiInBBEnd = &*I;
1252 if (!FirstNonPhiInBBEnd)
1256 // This does very trivial matching, with limited scanning, to find identical
1257 // instructions in the two blocks. We scan backward for obviously identical
1258 // instructions in an identical order.
1259 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1260 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(),
1261 RE2 = BB2->getInstList().rend();
1263 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1266 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1269 // Skip the unconditional branches.
1273 bool Changed = false;
1274 while (RI1 != RE1 && RI2 != RE2) {
1276 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1279 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1283 Instruction *I1 = &*RI1, *I2 = &*RI2;
1284 // I1 and I2 should have a single use in the same PHI node, and they
1285 // perform the same operation.
1286 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1287 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1288 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1289 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) ||
1290 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1291 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1292 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1293 !I1->hasOneUse() || !I2->hasOneUse() ||
1294 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() ||
1295 MapValueFromBB1ToBB2[I1].first != I2)
1298 // Check whether we should swap the operands of ICmpInst.
1299 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1300 bool SwapOpnds = false;
1301 if (ICmp1 && ICmp2 &&
1302 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1303 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1304 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1305 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1306 ICmp2->swapOperands();
1309 if (!I1->isSameOperationAs(I2)) {
1311 ICmp2->swapOperands();
1315 // The operands should be either the same or they need to be generated
1316 // with a PHI node after sinking. We only handle the case where there is
1317 // a single pair of different operands.
1318 Value *DifferentOp1 = nullptr, *DifferentOp2 = nullptr;
1319 unsigned Op1Idx = 0;
1320 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1321 if (I1->getOperand(I) == I2->getOperand(I))
1323 // Early exit if we have more-than one pair of different operands or
1324 // the different operand is already in MapValueFromBB1ToBB2.
1325 // Early exit if we need a PHI node to replace a constant.
1327 MapValueFromBB1ToBB2.find(I1->getOperand(I)) !=
1328 MapValueFromBB1ToBB2.end() ||
1329 isa<Constant>(I1->getOperand(I)) ||
1330 isa<Constant>(I2->getOperand(I))) {
1331 // If we can't sink the instructions, undo the swapping.
1333 ICmp2->swapOperands();
1336 DifferentOp1 = I1->getOperand(I);
1338 DifferentOp2 = I2->getOperand(I);
1341 // We insert the pair of different operands to MapValueFromBB1ToBB2 and
1342 // remove (I1, I2) from MapValueFromBB1ToBB2.
1344 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2,
1345 DifferentOp1->getName() + ".sink",
1347 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN);
1348 // I1 should use NewPN instead of DifferentOp1.
1349 I1->setOperand(Op1Idx, NewPN);
1350 NewPN->addIncoming(DifferentOp1, BB1);
1351 NewPN->addIncoming(DifferentOp2, BB2);
1352 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1354 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second;
1355 MapValueFromBB1ToBB2.erase(I1);
1357 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";);
1358 DEBUG(dbgs() << " " << *I2 << "\n";);
1359 // We need to update RE1 and RE2 if we are going to sink the first
1360 // instruction in the basic block down.
1361 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1362 // Sink the instruction.
1363 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1);
1364 if (!OldPN->use_empty())
1365 OldPN->replaceAllUsesWith(I1);
1366 OldPN->eraseFromParent();
1368 if (!I2->use_empty())
1369 I2->replaceAllUsesWith(I1);
1370 I1->intersectOptionalDataWith(I2);
1371 // TODO: Use combineMetadata here to preserve what metadata we can
1372 // (analogous to the hoisting case above).
1373 I2->eraseFromParent();
1376 RE1 = BB1->getInstList().rend();
1378 RE2 = BB2->getInstList().rend();
1379 FirstNonPhiInBBEnd = I1;
1386 /// \brief Determine if we can hoist sink a sole store instruction out of a
1387 /// conditional block.
1389 /// We are looking for code like the following:
1391 /// store i32 %add, i32* %arrayidx2
1392 /// ... // No other stores or function calls (we could be calling a memory
1393 /// ... // function).
1394 /// %cmp = icmp ult %x, %y
1395 /// br i1 %cmp, label %EndBB, label %ThenBB
1397 /// store i32 %add5, i32* %arrayidx2
1401 /// We are going to transform this into:
1403 /// store i32 %add, i32* %arrayidx2
1405 /// %cmp = icmp ult %x, %y
1406 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1407 /// store i32 %add.add5, i32* %arrayidx2
1410 /// \return The pointer to the value of the previous store if the store can be
1411 /// hoisted into the predecessor block. 0 otherwise.
1412 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1413 BasicBlock *StoreBB, BasicBlock *EndBB) {
1414 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1418 // Volatile or atomic.
1419 if (!StoreToHoist->isSimple())
1422 Value *StorePtr = StoreToHoist->getPointerOperand();
1424 // Look for a store to the same pointer in BrBB.
1425 unsigned MaxNumInstToLookAt = 10;
1426 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(),
1427 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) {
1428 Instruction *CurI = &*RI;
1430 // Could be calling an instruction that effects memory like free().
1431 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI))
1434 StoreInst *SI = dyn_cast<StoreInst>(CurI);
1435 // Found the previous store make sure it stores to the same location.
1436 if (SI && SI->getPointerOperand() == StorePtr)
1437 // Found the previous store, return its value operand.
1438 return SI->getValueOperand();
1440 return nullptr; // Unknown store.
1446 /// \brief Speculate a conditional basic block flattening the CFG.
1448 /// Note that this is a very risky transform currently. Speculating
1449 /// instructions like this is most often not desirable. Instead, there is an MI
1450 /// pass which can do it with full awareness of the resource constraints.
1451 /// However, some cases are "obvious" and we should do directly. An example of
1452 /// this is speculating a single, reasonably cheap instruction.
1454 /// There is only one distinct advantage to flattening the CFG at the IR level:
1455 /// it makes very common but simplistic optimizations such as are common in
1456 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1457 /// modeling their effects with easier to reason about SSA value graphs.
1460 /// An illustration of this transform is turning this IR:
1463 /// %cmp = icmp ult %x, %y
1464 /// br i1 %cmp, label %EndBB, label %ThenBB
1466 /// %sub = sub %x, %y
1469 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1476 /// %cmp = icmp ult %x, %y
1477 /// %sub = sub %x, %y
1478 /// %cond = select i1 %cmp, 0, %sub
1482 /// \returns true if the conditional block is removed.
1483 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1484 const DataLayout *DL) {
1485 // Be conservative for now. FP select instruction can often be expensive.
1486 Value *BrCond = BI->getCondition();
1487 if (isa<FCmpInst>(BrCond))
1490 BasicBlock *BB = BI->getParent();
1491 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1493 // If ThenBB is actually on the false edge of the conditional branch, remember
1494 // to swap the select operands later.
1495 bool Invert = false;
1496 if (ThenBB != BI->getSuccessor(0)) {
1497 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1500 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1502 // Keep a count of how many times instructions are used within CondBB when
1503 // they are candidates for sinking into CondBB. Specifically:
1504 // - They are defined in BB, and
1505 // - They have no side effects, and
1506 // - All of their uses are in CondBB.
1507 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1509 unsigned SpeculationCost = 0;
1510 Value *SpeculatedStoreValue = nullptr;
1511 StoreInst *SpeculatedStore = nullptr;
1512 for (BasicBlock::iterator BBI = ThenBB->begin(),
1513 BBE = std::prev(ThenBB->end());
1514 BBI != BBE; ++BBI) {
1515 Instruction *I = BBI;
1517 if (isa<DbgInfoIntrinsic>(I))
1520 // Only speculatively execution a single instruction (not counting the
1521 // terminator) for now.
1523 if (SpeculationCost > 1)
1526 // Don't hoist the instruction if it's unsafe or expensive.
1527 if (!isSafeToSpeculativelyExecute(I, DL) &&
1528 !(HoistCondStores &&
1529 (SpeculatedStoreValue = isSafeToSpeculateStore(I, BB, ThenBB,
1532 if (!SpeculatedStoreValue &&
1533 ComputeSpeculationCost(I, DL) > PHINodeFoldingThreshold)
1536 // Store the store speculation candidate.
1537 if (SpeculatedStoreValue)
1538 SpeculatedStore = cast<StoreInst>(I);
1540 // Do not hoist the instruction if any of its operands are defined but not
1541 // used in BB. The transformation will prevent the operand from
1542 // being sunk into the use block.
1543 for (User::op_iterator i = I->op_begin(), e = I->op_end();
1545 Instruction *OpI = dyn_cast<Instruction>(*i);
1546 if (!OpI || OpI->getParent() != BB ||
1547 OpI->mayHaveSideEffects())
1548 continue; // Not a candidate for sinking.
1550 ++SinkCandidateUseCounts[OpI];
1554 // Consider any sink candidates which are only used in CondBB as costs for
1555 // speculation. Note, while we iterate over a DenseMap here, we are summing
1556 // and so iteration order isn't significant.
1557 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I =
1558 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end();
1560 if (I->first->getNumUses() == I->second) {
1562 if (SpeculationCost > 1)
1566 // Check that the PHI nodes can be converted to selects.
1567 bool HaveRewritablePHIs = false;
1568 for (BasicBlock::iterator I = EndBB->begin();
1569 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1570 Value *OrigV = PN->getIncomingValueForBlock(BB);
1571 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1573 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1574 // Skip PHIs which are trivial.
1578 // Don't convert to selects if we could remove undefined behavior instead.
1579 if (passingValueIsAlwaysUndefined(OrigV, PN) ||
1580 passingValueIsAlwaysUndefined(ThenV, PN))
1583 HaveRewritablePHIs = true;
1584 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1585 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1586 if (!OrigCE && !ThenCE)
1587 continue; // Known safe and cheap.
1589 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE, DL)) ||
1590 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE, DL)))
1592 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, DL) : 0;
1593 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, DL) : 0;
1594 if (OrigCost + ThenCost > 2 * PHINodeFoldingThreshold)
1597 // Account for the cost of an unfolded ConstantExpr which could end up
1598 // getting expanded into Instructions.
1599 // FIXME: This doesn't account for how many operations are combined in the
1600 // constant expression.
1602 if (SpeculationCost > 1)
1606 // If there are no PHIs to process, bail early. This helps ensure idempotence
1608 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
1611 // If we get here, we can hoist the instruction and if-convert.
1612 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1614 // Insert a select of the value of the speculated store.
1615 if (SpeculatedStoreValue) {
1616 IRBuilder<true, NoFolder> Builder(BI);
1617 Value *TrueV = SpeculatedStore->getValueOperand();
1618 Value *FalseV = SpeculatedStoreValue;
1620 std::swap(TrueV, FalseV);
1621 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() +
1622 "." + FalseV->getName());
1623 SpeculatedStore->setOperand(0, S);
1626 // Hoist the instructions.
1627 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(),
1628 std::prev(ThenBB->end()));
1630 // Insert selects and rewrite the PHI operands.
1631 IRBuilder<true, NoFolder> Builder(BI);
1632 for (BasicBlock::iterator I = EndBB->begin();
1633 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1634 unsigned OrigI = PN->getBasicBlockIndex(BB);
1635 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1636 Value *OrigV = PN->getIncomingValue(OrigI);
1637 Value *ThenV = PN->getIncomingValue(ThenI);
1639 // Skip PHIs which are trivial.
1643 // Create a select whose true value is the speculatively executed value and
1644 // false value is the preexisting value. Swap them if the branch
1645 // destinations were inverted.
1646 Value *TrueV = ThenV, *FalseV = OrigV;
1648 std::swap(TrueV, FalseV);
1649 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV,
1650 TrueV->getName() + "." + FalseV->getName());
1651 PN->setIncomingValue(OrigI, V);
1652 PN->setIncomingValue(ThenI, V);
1659 /// \returns True if this block contains a CallInst with the NoDuplicate
1661 static bool HasNoDuplicateCall(const BasicBlock *BB) {
1662 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1663 const CallInst *CI = dyn_cast<CallInst>(I);
1666 if (CI->cannotDuplicate())
1672 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1673 /// across this block.
1674 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1675 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1678 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1679 if (isa<DbgInfoIntrinsic>(BBI))
1681 if (Size > 10) return false; // Don't clone large BB's.
1684 // We can only support instructions that do not define values that are
1685 // live outside of the current basic block.
1686 for (User *U : BBI->users()) {
1687 Instruction *UI = cast<Instruction>(U);
1688 if (UI->getParent() != BB || isa<PHINode>(UI)) return false;
1691 // Looks ok, continue checking.
1697 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1698 /// that is defined in the same block as the branch and if any PHI entries are
1699 /// constants, thread edges corresponding to that entry to be branches to their
1700 /// ultimate destination.
1701 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *DL) {
1702 BasicBlock *BB = BI->getParent();
1703 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1704 // NOTE: we currently cannot transform this case if the PHI node is used
1705 // outside of the block.
1706 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1709 // Degenerate case of a single entry PHI.
1710 if (PN->getNumIncomingValues() == 1) {
1711 FoldSingleEntryPHINodes(PN->getParent());
1715 // Now we know that this block has multiple preds and two succs.
1716 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1718 if (HasNoDuplicateCall(BB)) return false;
1720 // Okay, this is a simple enough basic block. See if any phi values are
1722 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1723 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1724 if (!CB || !CB->getType()->isIntegerTy(1)) continue;
1726 // Okay, we now know that all edges from PredBB should be revectored to
1727 // branch to RealDest.
1728 BasicBlock *PredBB = PN->getIncomingBlock(i);
1729 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1731 if (RealDest == BB) continue; // Skip self loops.
1732 // Skip if the predecessor's terminator is an indirect branch.
1733 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1735 // The dest block might have PHI nodes, other predecessors and other
1736 // difficult cases. Instead of being smart about this, just insert a new
1737 // block that jumps to the destination block, effectively splitting
1738 // the edge we are about to create.
1739 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1740 RealDest->getName()+".critedge",
1741 RealDest->getParent(), RealDest);
1742 BranchInst::Create(RealDest, EdgeBB);
1744 // Update PHI nodes.
1745 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1747 // BB may have instructions that are being threaded over. Clone these
1748 // instructions into EdgeBB. We know that there will be no uses of the
1749 // cloned instructions outside of EdgeBB.
1750 BasicBlock::iterator InsertPt = EdgeBB->begin();
1751 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1752 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1753 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1754 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1757 // Clone the instruction.
1758 Instruction *N = BBI->clone();
1759 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1761 // Update operands due to translation.
1762 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1764 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1765 if (PI != TranslateMap.end())
1769 // Check for trivial simplification.
1770 if (Value *V = SimplifyInstruction(N, DL)) {
1771 TranslateMap[BBI] = V;
1772 delete N; // Instruction folded away, don't need actual inst
1774 // Insert the new instruction into its new home.
1775 EdgeBB->getInstList().insert(InsertPt, N);
1776 if (!BBI->use_empty())
1777 TranslateMap[BBI] = N;
1781 // Loop over all of the edges from PredBB to BB, changing them to branch
1782 // to EdgeBB instead.
1783 TerminatorInst *PredBBTI = PredBB->getTerminator();
1784 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1785 if (PredBBTI->getSuccessor(i) == BB) {
1786 BB->removePredecessor(PredBB);
1787 PredBBTI->setSuccessor(i, EdgeBB);
1790 // Recurse, simplifying any other constants.
1791 return FoldCondBranchOnPHI(BI, DL) | true;
1797 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1798 /// PHI node, see if we can eliminate it.
1799 static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *DL) {
1800 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1801 // statement", which has a very simple dominance structure. Basically, we
1802 // are trying to find the condition that is being branched on, which
1803 // subsequently causes this merge to happen. We really want control
1804 // dependence information for this check, but simplifycfg can't keep it up
1805 // to date, and this catches most of the cases we care about anyway.
1806 BasicBlock *BB = PN->getParent();
1807 BasicBlock *IfTrue, *IfFalse;
1808 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1810 // Don't bother if the branch will be constant folded trivially.
1811 isa<ConstantInt>(IfCond))
1814 // Okay, we found that we can merge this two-entry phi node into a select.
1815 // Doing so would require us to fold *all* two entry phi nodes in this block.
1816 // At some point this becomes non-profitable (particularly if the target
1817 // doesn't support cmov's). Only do this transformation if there are two or
1818 // fewer PHI nodes in this block.
1819 unsigned NumPhis = 0;
1820 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1824 // Loop over the PHI's seeing if we can promote them all to select
1825 // instructions. While we are at it, keep track of the instructions
1826 // that need to be moved to the dominating block.
1827 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1828 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1829 MaxCostVal1 = PHINodeFoldingThreshold;
1831 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1832 PHINode *PN = cast<PHINode>(II++);
1833 if (Value *V = SimplifyInstruction(PN, DL)) {
1834 PN->replaceAllUsesWith(V);
1835 PN->eraseFromParent();
1839 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1841 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1846 // If we folded the first phi, PN dangles at this point. Refresh it. If
1847 // we ran out of PHIs then we simplified them all.
1848 PN = dyn_cast<PHINode>(BB->begin());
1849 if (!PN) return true;
1851 // Don't fold i1 branches on PHIs which contain binary operators. These can
1852 // often be turned into switches and other things.
1853 if (PN->getType()->isIntegerTy(1) &&
1854 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1855 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1856 isa<BinaryOperator>(IfCond)))
1859 // If we all PHI nodes are promotable, check to make sure that all
1860 // instructions in the predecessor blocks can be promoted as well. If
1861 // not, we won't be able to get rid of the control flow, so it's not
1862 // worth promoting to select instructions.
1863 BasicBlock *DomBlock = nullptr;
1864 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1865 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1866 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1869 DomBlock = *pred_begin(IfBlock1);
1870 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1871 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1872 // This is not an aggressive instruction that we can promote.
1873 // Because of this, we won't be able to get rid of the control
1874 // flow, so the xform is not worth it.
1879 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1882 DomBlock = *pred_begin(IfBlock2);
1883 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1884 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1885 // This is not an aggressive instruction that we can promote.
1886 // Because of this, we won't be able to get rid of the control
1887 // flow, so the xform is not worth it.
1892 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1893 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1895 // If we can still promote the PHI nodes after this gauntlet of tests,
1896 // do all of the PHI's now.
1897 Instruction *InsertPt = DomBlock->getTerminator();
1898 IRBuilder<true, NoFolder> Builder(InsertPt);
1900 // Move all 'aggressive' instructions, which are defined in the
1901 // conditional parts of the if's up to the dominating block.
1903 DomBlock->getInstList().splice(InsertPt,
1904 IfBlock1->getInstList(), IfBlock1->begin(),
1905 IfBlock1->getTerminator());
1907 DomBlock->getInstList().splice(InsertPt,
1908 IfBlock2->getInstList(), IfBlock2->begin(),
1909 IfBlock2->getTerminator());
1911 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1912 // Change the PHI node into a select instruction.
1913 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1914 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1917 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1918 PN->replaceAllUsesWith(NV);
1920 PN->eraseFromParent();
1923 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1924 // has been flattened. Change DomBlock to jump directly to our new block to
1925 // avoid other simplifycfg's kicking in on the diamond.
1926 TerminatorInst *OldTI = DomBlock->getTerminator();
1927 Builder.SetInsertPoint(OldTI);
1928 Builder.CreateBr(BB);
1929 OldTI->eraseFromParent();
1933 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1934 /// to two returning blocks, try to merge them together into one return,
1935 /// introducing a select if the return values disagree.
1936 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1937 IRBuilder<> &Builder) {
1938 assert(BI->isConditional() && "Must be a conditional branch");
1939 BasicBlock *TrueSucc = BI->getSuccessor(0);
1940 BasicBlock *FalseSucc = BI->getSuccessor(1);
1941 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1942 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1944 // Check to ensure both blocks are empty (just a return) or optionally empty
1945 // with PHI nodes. If there are other instructions, merging would cause extra
1946 // computation on one path or the other.
1947 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1949 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1952 Builder.SetInsertPoint(BI);
1953 // Okay, we found a branch that is going to two return nodes. If
1954 // there is no return value for this function, just change the
1955 // branch into a return.
1956 if (FalseRet->getNumOperands() == 0) {
1957 TrueSucc->removePredecessor(BI->getParent());
1958 FalseSucc->removePredecessor(BI->getParent());
1959 Builder.CreateRetVoid();
1960 EraseTerminatorInstAndDCECond(BI);
1964 // Otherwise, figure out what the true and false return values are
1965 // so we can insert a new select instruction.
1966 Value *TrueValue = TrueRet->getReturnValue();
1967 Value *FalseValue = FalseRet->getReturnValue();
1969 // Unwrap any PHI nodes in the return blocks.
1970 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1971 if (TVPN->getParent() == TrueSucc)
1972 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1973 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1974 if (FVPN->getParent() == FalseSucc)
1975 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1977 // In order for this transformation to be safe, we must be able to
1978 // unconditionally execute both operands to the return. This is
1979 // normally the case, but we could have a potentially-trapping
1980 // constant expression that prevents this transformation from being
1982 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1985 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1989 // Okay, we collected all the mapped values and checked them for sanity, and
1990 // defined to really do this transformation. First, update the CFG.
1991 TrueSucc->removePredecessor(BI->getParent());
1992 FalseSucc->removePredecessor(BI->getParent());
1994 // Insert select instructions where needed.
1995 Value *BrCond = BI->getCondition();
1997 // Insert a select if the results differ.
1998 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1999 } else if (isa<UndefValue>(TrueValue)) {
2000 TrueValue = FalseValue;
2002 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
2003 FalseValue, "retval");
2007 Value *RI = !TrueValue ?
2008 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2012 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2013 << "\n " << *BI << "NewRet = " << *RI
2014 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
2016 EraseTerminatorInstAndDCECond(BI);
2021 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
2022 /// probabilities of the branch taking each edge. Fills in the two APInt
2023 /// parameters and return true, or returns false if no or invalid metadata was
2025 static bool ExtractBranchMetadata(BranchInst *BI,
2026 uint64_t &ProbTrue, uint64_t &ProbFalse) {
2027 assert(BI->isConditional() &&
2028 "Looking for probabilities on unconditional branch?");
2029 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
2030 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
2031 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
2032 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
2033 if (!CITrue || !CIFalse) return false;
2034 ProbTrue = CITrue->getValue().getZExtValue();
2035 ProbFalse = CIFalse->getValue().getZExtValue();
2039 /// checkCSEInPredecessor - Return true if the given instruction is available
2040 /// in its predecessor block. If yes, the instruction will be removed.
2042 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
2043 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2045 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
2046 Instruction *PBI = &*I;
2047 // Check whether Inst and PBI generate the same value.
2048 if (Inst->isIdenticalTo(PBI)) {
2049 Inst->replaceAllUsesWith(PBI);
2050 Inst->eraseFromParent();
2057 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
2058 /// predecessor branches to us and one of our successors, fold the block into
2059 /// the predecessor and use logical operations to pick the right destination.
2060 bool llvm::FoldBranchToCommonDest(BranchInst *BI, const DataLayout *DL,
2061 unsigned BonusInstThreshold) {
2062 BasicBlock *BB = BI->getParent();
2064 Instruction *Cond = nullptr;
2065 if (BI->isConditional())
2066 Cond = dyn_cast<Instruction>(BI->getCondition());
2068 // For unconditional branch, check for a simple CFG pattern, where
2069 // BB has a single predecessor and BB's successor is also its predecessor's
2070 // successor. If such pattern exisits, check for CSE between BB and its
2072 if (BasicBlock *PB = BB->getSinglePredecessor())
2073 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2074 if (PBI->isConditional() &&
2075 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2076 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2077 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
2079 Instruction *Curr = I++;
2080 if (isa<CmpInst>(Curr)) {
2084 // Quit if we can't remove this instruction.
2085 if (!checkCSEInPredecessor(Curr, PB))
2094 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2095 Cond->getParent() != BB || !Cond->hasOneUse())
2098 // Make sure the instruction after the condition is the cond branch.
2099 BasicBlock::iterator CondIt = Cond; ++CondIt;
2101 // Ignore dbg intrinsics.
2102 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
2107 // Only allow this transformation if computing the condition doesn't involve
2108 // too many instructions and these involved instructions can be executed
2109 // unconditionally. We denote all involved instructions except the condition
2110 // as "bonus instructions", and only allow this transformation when the
2111 // number of the bonus instructions does not exceed a certain threshold.
2112 unsigned NumBonusInsts = 0;
2113 for (auto I = BB->begin(); Cond != I; ++I) {
2114 // Ignore dbg intrinsics.
2115 if (isa<DbgInfoIntrinsic>(I))
2117 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(I, DL))
2119 // I has only one use and can be executed unconditionally.
2120 Instruction *User = dyn_cast<Instruction>(I->user_back());
2121 if (User == nullptr || User->getParent() != BB)
2123 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2124 // to use any other instruction, User must be an instruction between next(I)
2127 // Early exits once we reach the limit.
2128 if (NumBonusInsts > BonusInstThreshold)
2132 // Cond is known to be a compare or binary operator. Check to make sure that
2133 // neither operand is a potentially-trapping constant expression.
2134 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2137 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2141 // Finally, don't infinitely unroll conditional loops.
2142 BasicBlock *TrueDest = BI->getSuccessor(0);
2143 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2144 if (TrueDest == BB || FalseDest == BB)
2147 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2148 BasicBlock *PredBlock = *PI;
2149 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2151 // Check that we have two conditional branches. If there is a PHI node in
2152 // the common successor, verify that the same value flows in from both
2154 SmallVector<PHINode*, 4> PHIs;
2155 if (!PBI || PBI->isUnconditional() ||
2156 (BI->isConditional() &&
2157 !SafeToMergeTerminators(BI, PBI)) ||
2158 (!BI->isConditional() &&
2159 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2162 // Determine if the two branches share a common destination.
2163 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2164 bool InvertPredCond = false;
2166 if (BI->isConditional()) {
2167 if (PBI->getSuccessor(0) == TrueDest)
2168 Opc = Instruction::Or;
2169 else if (PBI->getSuccessor(1) == FalseDest)
2170 Opc = Instruction::And;
2171 else if (PBI->getSuccessor(0) == FalseDest)
2172 Opc = Instruction::And, InvertPredCond = true;
2173 else if (PBI->getSuccessor(1) == TrueDest)
2174 Opc = Instruction::Or, InvertPredCond = true;
2178 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2182 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2183 IRBuilder<> Builder(PBI);
2185 // If we need to invert the condition in the pred block to match, do so now.
2186 if (InvertPredCond) {
2187 Value *NewCond = PBI->getCondition();
2189 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2190 CmpInst *CI = cast<CmpInst>(NewCond);
2191 CI->setPredicate(CI->getInversePredicate());
2193 NewCond = Builder.CreateNot(NewCond,
2194 PBI->getCondition()->getName()+".not");
2197 PBI->setCondition(NewCond);
2198 PBI->swapSuccessors();
2201 // If we have bonus instructions, clone them into the predecessor block.
2202 // Note that there may be mutliple predecessor blocks, so we cannot move
2203 // bonus instructions to a predecessor block.
2204 ValueToValueMapTy VMap; // maps original values to cloned values
2205 // We already make sure Cond is the last instruction before BI. Therefore,
2206 // every instructions before Cond other than DbgInfoIntrinsic are bonus
2208 for (auto BonusInst = BB->begin(); Cond != BonusInst; ++BonusInst) {
2209 if (isa<DbgInfoIntrinsic>(BonusInst))
2211 Instruction *NewBonusInst = BonusInst->clone();
2212 RemapInstruction(NewBonusInst, VMap,
2213 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2214 VMap[BonusInst] = NewBonusInst;
2216 // If we moved a load, we cannot any longer claim any knowledge about
2217 // its potential value. The previous information might have been valid
2218 // only given the branch precondition.
2219 // For an analogous reason, we must also drop all the metadata whose
2220 // semantics we don't understand.
2221 NewBonusInst->dropUnknownMetadata(LLVMContext::MD_dbg);
2223 PredBlock->getInstList().insert(PBI, NewBonusInst);
2224 NewBonusInst->takeName(BonusInst);
2225 BonusInst->setName(BonusInst->getName() + ".old");
2228 // Clone Cond into the predecessor basic block, and or/and the
2229 // two conditions together.
2230 Instruction *New = Cond->clone();
2231 RemapInstruction(New, VMap,
2232 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2233 PredBlock->getInstList().insert(PBI, New);
2234 New->takeName(Cond);
2235 Cond->setName(New->getName() + ".old");
2237 if (BI->isConditional()) {
2238 Instruction *NewCond =
2239 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2241 PBI->setCondition(NewCond);
2243 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2244 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2246 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2248 SmallVector<uint64_t, 8> NewWeights;
2250 if (PBI->getSuccessor(0) == BB) {
2251 if (PredHasWeights && SuccHasWeights) {
2252 // PBI: br i1 %x, BB, FalseDest
2253 // BI: br i1 %y, TrueDest, FalseDest
2254 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2255 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2256 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2257 // TrueWeight for PBI * FalseWeight for BI.
2258 // We assume that total weights of a BranchInst can fit into 32 bits.
2259 // Therefore, we will not have overflow using 64-bit arithmetic.
2260 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2261 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2263 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2264 PBI->setSuccessor(0, TrueDest);
2266 if (PBI->getSuccessor(1) == BB) {
2267 if (PredHasWeights && SuccHasWeights) {
2268 // PBI: br i1 %x, TrueDest, BB
2269 // BI: br i1 %y, TrueDest, FalseDest
2270 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2271 // FalseWeight for PBI * TrueWeight for BI.
2272 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2273 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2274 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2275 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2277 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2278 PBI->setSuccessor(1, FalseDest);
2280 if (NewWeights.size() == 2) {
2281 // Halve the weights if any of them cannot fit in an uint32_t
2282 FitWeights(NewWeights);
2284 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2285 PBI->setMetadata(LLVMContext::MD_prof,
2286 MDBuilder(BI->getContext()).
2287 createBranchWeights(MDWeights));
2289 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2291 // Update PHI nodes in the common successors.
2292 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2293 ConstantInt *PBI_C = cast<ConstantInt>(
2294 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2295 assert(PBI_C->getType()->isIntegerTy(1));
2296 Instruction *MergedCond = nullptr;
2297 if (PBI->getSuccessor(0) == TrueDest) {
2298 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2299 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2300 // is false: !PBI_Cond and BI_Value
2301 Instruction *NotCond =
2302 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2305 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2310 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2311 PBI->getCondition(), MergedCond,
2314 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2315 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2316 // is false: PBI_Cond and BI_Value
2318 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2319 PBI->getCondition(), New,
2321 if (PBI_C->isOne()) {
2322 Instruction *NotCond =
2323 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2326 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2327 NotCond, MergedCond,
2332 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2335 // Change PBI from Conditional to Unconditional.
2336 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2337 EraseTerminatorInstAndDCECond(PBI);
2341 // TODO: If BB is reachable from all paths through PredBlock, then we
2342 // could replace PBI's branch probabilities with BI's.
2344 // Copy any debug value intrinsics into the end of PredBlock.
2345 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2346 if (isa<DbgInfoIntrinsic>(*I))
2347 I->clone()->insertBefore(PBI);
2354 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2355 /// predecessor of another block, this function tries to simplify it. We know
2356 /// that PBI and BI are both conditional branches, and BI is in one of the
2357 /// successor blocks of PBI - PBI branches to BI.
2358 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2359 assert(PBI->isConditional() && BI->isConditional());
2360 BasicBlock *BB = BI->getParent();
2362 // If this block ends with a branch instruction, and if there is a
2363 // predecessor that ends on a branch of the same condition, make
2364 // this conditional branch redundant.
2365 if (PBI->getCondition() == BI->getCondition() &&
2366 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2367 // Okay, the outcome of this conditional branch is statically
2368 // knowable. If this block had a single pred, handle specially.
2369 if (BB->getSinglePredecessor()) {
2370 // Turn this into a branch on constant.
2371 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2372 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2374 return true; // Nuke the branch on constant.
2377 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2378 // in the constant and simplify the block result. Subsequent passes of
2379 // simplifycfg will thread the block.
2380 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2381 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2382 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2383 std::distance(PB, PE),
2384 BI->getCondition()->getName() + ".pr",
2386 // Okay, we're going to insert the PHI node. Since PBI is not the only
2387 // predecessor, compute the PHI'd conditional value for all of the preds.
2388 // Any predecessor where the condition is not computable we keep symbolic.
2389 for (pred_iterator PI = PB; PI != PE; ++PI) {
2390 BasicBlock *P = *PI;
2391 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2392 PBI != BI && PBI->isConditional() &&
2393 PBI->getCondition() == BI->getCondition() &&
2394 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2395 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2396 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2399 NewPN->addIncoming(BI->getCondition(), P);
2403 BI->setCondition(NewPN);
2408 // If this is a conditional branch in an empty block, and if any
2409 // predecessors are a conditional branch to one of our destinations,
2410 // fold the conditions into logical ops and one cond br.
2411 BasicBlock::iterator BBI = BB->begin();
2412 // Ignore dbg intrinsics.
2413 while (isa<DbgInfoIntrinsic>(BBI))
2419 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2424 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2426 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2427 PBIOp = 0, BIOp = 1;
2428 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2429 PBIOp = 1, BIOp = 0;
2430 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2435 // Check to make sure that the other destination of this branch
2436 // isn't BB itself. If so, this is an infinite loop that will
2437 // keep getting unwound.
2438 if (PBI->getSuccessor(PBIOp) == BB)
2441 // Do not perform this transformation if it would require
2442 // insertion of a large number of select instructions. For targets
2443 // without predication/cmovs, this is a big pessimization.
2445 // Also do not perform this transformation if any phi node in the common
2446 // destination block can trap when reached by BB or PBB (PR17073). In that
2447 // case, it would be unsafe to hoist the operation into a select instruction.
2449 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2450 unsigned NumPhis = 0;
2451 for (BasicBlock::iterator II = CommonDest->begin();
2452 isa<PHINode>(II); ++II, ++NumPhis) {
2453 if (NumPhis > 2) // Disable this xform.
2456 PHINode *PN = cast<PHINode>(II);
2457 Value *BIV = PN->getIncomingValueForBlock(BB);
2458 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
2462 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2463 Value *PBIV = PN->getIncomingValue(PBBIdx);
2464 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
2469 // Finally, if everything is ok, fold the branches to logical ops.
2470 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2472 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2473 << "AND: " << *BI->getParent());
2476 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2477 // branch in it, where one edge (OtherDest) goes back to itself but the other
2478 // exits. We don't *know* that the program avoids the infinite loop
2479 // (even though that seems likely). If we do this xform naively, we'll end up
2480 // recursively unpeeling the loop. Since we know that (after the xform is
2481 // done) that the block *is* infinite if reached, we just make it an obviously
2482 // infinite loop with no cond branch.
2483 if (OtherDest == BB) {
2484 // Insert it at the end of the function, because it's either code,
2485 // or it won't matter if it's hot. :)
2486 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2487 "infloop", BB->getParent());
2488 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2489 OtherDest = InfLoopBlock;
2492 DEBUG(dbgs() << *PBI->getParent()->getParent());
2494 // BI may have other predecessors. Because of this, we leave
2495 // it alone, but modify PBI.
2497 // Make sure we get to CommonDest on True&True directions.
2498 Value *PBICond = PBI->getCondition();
2499 IRBuilder<true, NoFolder> Builder(PBI);
2501 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2503 Value *BICond = BI->getCondition();
2505 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2507 // Merge the conditions.
2508 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2510 // Modify PBI to branch on the new condition to the new dests.
2511 PBI->setCondition(Cond);
2512 PBI->setSuccessor(0, CommonDest);
2513 PBI->setSuccessor(1, OtherDest);
2515 // Update branch weight for PBI.
2516 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2517 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2519 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2521 if (PredHasWeights && SuccHasWeights) {
2522 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2523 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2524 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2525 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2526 // The weight to CommonDest should be PredCommon * SuccTotal +
2527 // PredOther * SuccCommon.
2528 // The weight to OtherDest should be PredOther * SuccOther.
2529 SmallVector<uint64_t, 2> NewWeights;
2530 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) +
2531 PredOther * SuccCommon);
2532 NewWeights.push_back(PredOther * SuccOther);
2533 // Halve the weights if any of them cannot fit in an uint32_t
2534 FitWeights(NewWeights);
2536 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end());
2537 PBI->setMetadata(LLVMContext::MD_prof,
2538 MDBuilder(BI->getContext()).
2539 createBranchWeights(MDWeights));
2542 // OtherDest may have phi nodes. If so, add an entry from PBI's
2543 // block that are identical to the entries for BI's block.
2544 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2546 // We know that the CommonDest already had an edge from PBI to
2547 // it. If it has PHIs though, the PHIs may have different
2548 // entries for BB and PBI's BB. If so, insert a select to make
2551 for (BasicBlock::iterator II = CommonDest->begin();
2552 (PN = dyn_cast<PHINode>(II)); ++II) {
2553 Value *BIV = PN->getIncomingValueForBlock(BB);
2554 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2555 Value *PBIV = PN->getIncomingValue(PBBIdx);
2557 // Insert a select in PBI to pick the right value.
2558 Value *NV = cast<SelectInst>
2559 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2560 PN->setIncomingValue(PBBIdx, NV);
2564 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2565 DEBUG(dbgs() << *PBI->getParent()->getParent());
2567 // This basic block is probably dead. We know it has at least
2568 // one fewer predecessor.
2572 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2573 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2574 // Takes care of updating the successors and removing the old terminator.
2575 // Also makes sure not to introduce new successors by assuming that edges to
2576 // non-successor TrueBBs and FalseBBs aren't reachable.
2577 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2578 BasicBlock *TrueBB, BasicBlock *FalseBB,
2579 uint32_t TrueWeight,
2580 uint32_t FalseWeight){
2581 // Remove any superfluous successor edges from the CFG.
2582 // First, figure out which successors to preserve.
2583 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2585 BasicBlock *KeepEdge1 = TrueBB;
2586 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
2588 // Then remove the rest.
2589 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2590 BasicBlock *Succ = OldTerm->getSuccessor(I);
2591 // Make sure only to keep exactly one copy of each edge.
2592 if (Succ == KeepEdge1)
2593 KeepEdge1 = nullptr;
2594 else if (Succ == KeepEdge2)
2595 KeepEdge2 = nullptr;
2597 Succ->removePredecessor(OldTerm->getParent());
2600 IRBuilder<> Builder(OldTerm);
2601 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2603 // Insert an appropriate new terminator.
2604 if (!KeepEdge1 && !KeepEdge2) {
2605 if (TrueBB == FalseBB)
2606 // We were only looking for one successor, and it was present.
2607 // Create an unconditional branch to it.
2608 Builder.CreateBr(TrueBB);
2610 // We found both of the successors we were looking for.
2611 // Create a conditional branch sharing the condition of the select.
2612 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2613 if (TrueWeight != FalseWeight)
2614 NewBI->setMetadata(LLVMContext::MD_prof,
2615 MDBuilder(OldTerm->getContext()).
2616 createBranchWeights(TrueWeight, FalseWeight));
2618 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2619 // Neither of the selected blocks were successors, so this
2620 // terminator must be unreachable.
2621 new UnreachableInst(OldTerm->getContext(), OldTerm);
2623 // One of the selected values was a successor, but the other wasn't.
2624 // Insert an unconditional branch to the one that was found;
2625 // the edge to the one that wasn't must be unreachable.
2627 // Only TrueBB was found.
2628 Builder.CreateBr(TrueBB);
2630 // Only FalseBB was found.
2631 Builder.CreateBr(FalseBB);
2634 EraseTerminatorInstAndDCECond(OldTerm);
2638 // SimplifySwitchOnSelect - Replaces
2639 // (switch (select cond, X, Y)) on constant X, Y
2640 // with a branch - conditional if X and Y lead to distinct BBs,
2641 // unconditional otherwise.
2642 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2643 // Check for constant integer values in the select.
2644 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2645 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2646 if (!TrueVal || !FalseVal)
2649 // Find the relevant condition and destinations.
2650 Value *Condition = Select->getCondition();
2651 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2652 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2654 // Get weight for TrueBB and FalseBB.
2655 uint32_t TrueWeight = 0, FalseWeight = 0;
2656 SmallVector<uint64_t, 8> Weights;
2657 bool HasWeights = HasBranchWeights(SI);
2659 GetBranchWeights(SI, Weights);
2660 if (Weights.size() == 1 + SI->getNumCases()) {
2661 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2662 getSuccessorIndex()];
2663 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2664 getSuccessorIndex()];
2668 // Perform the actual simplification.
2669 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2670 TrueWeight, FalseWeight);
2673 // SimplifyIndirectBrOnSelect - Replaces
2674 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2675 // blockaddress(@fn, BlockB)))
2677 // (br cond, BlockA, BlockB).
2678 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2679 // Check that both operands of the select are block addresses.
2680 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2681 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2685 // Extract the actual blocks.
2686 BasicBlock *TrueBB = TBA->getBasicBlock();
2687 BasicBlock *FalseBB = FBA->getBasicBlock();
2689 // Perform the actual simplification.
2690 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2694 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2695 /// instruction (a seteq/setne with a constant) as the only instruction in a
2696 /// block that ends with an uncond branch. We are looking for a very specific
2697 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2698 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2699 /// default value goes to an uncond block with a seteq in it, we get something
2702 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2704 /// %tmp = icmp eq i8 %A, 92
2707 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2709 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2710 /// the PHI, merging the third icmp into the switch.
2711 static bool TryToSimplifyUncondBranchWithICmpInIt(
2712 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI,
2713 unsigned BonusInstThreshold, const DataLayout *DL, AssumptionTracker *AT) {
2714 BasicBlock *BB = ICI->getParent();
2716 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2718 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2720 Value *V = ICI->getOperand(0);
2721 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2723 // The pattern we're looking for is where our only predecessor is a switch on
2724 // 'V' and this block is the default case for the switch. In this case we can
2725 // fold the compared value into the switch to simplify things.
2726 BasicBlock *Pred = BB->getSinglePredecessor();
2727 if (!Pred || !isa<SwitchInst>(Pred->getTerminator())) return false;
2729 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2730 if (SI->getCondition() != V)
2733 // If BB is reachable on a non-default case, then we simply know the value of
2734 // V in this block. Substitute it and constant fold the icmp instruction
2736 if (SI->getDefaultDest() != BB) {
2737 ConstantInt *VVal = SI->findCaseDest(BB);
2738 assert(VVal && "Should have a unique destination value");
2739 ICI->setOperand(0, VVal);
2741 if (Value *V = SimplifyInstruction(ICI, DL)) {
2742 ICI->replaceAllUsesWith(V);
2743 ICI->eraseFromParent();
2745 // BB is now empty, so it is likely to simplify away.
2746 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
2749 // Ok, the block is reachable from the default dest. If the constant we're
2750 // comparing exists in one of the other edges, then we can constant fold ICI
2752 if (SI->findCaseValue(Cst) != SI->case_default()) {
2754 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2755 V = ConstantInt::getFalse(BB->getContext());
2757 V = ConstantInt::getTrue(BB->getContext());
2759 ICI->replaceAllUsesWith(V);
2760 ICI->eraseFromParent();
2761 // BB is now empty, so it is likely to simplify away.
2762 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
2765 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2767 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2768 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
2769 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
2770 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2773 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2775 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2776 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2778 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2779 std::swap(DefaultCst, NewCst);
2781 // Replace ICI (which is used by the PHI for the default value) with true or
2782 // false depending on if it is EQ or NE.
2783 ICI->replaceAllUsesWith(DefaultCst);
2784 ICI->eraseFromParent();
2786 // Okay, the switch goes to this block on a default value. Add an edge from
2787 // the switch to the merge point on the compared value.
2788 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2789 BB->getParent(), BB);
2790 SmallVector<uint64_t, 8> Weights;
2791 bool HasWeights = HasBranchWeights(SI);
2793 GetBranchWeights(SI, Weights);
2794 if (Weights.size() == 1 + SI->getNumCases()) {
2795 // Split weight for default case to case for "Cst".
2796 Weights[0] = (Weights[0]+1) >> 1;
2797 Weights.push_back(Weights[0]);
2799 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2800 SI->setMetadata(LLVMContext::MD_prof,
2801 MDBuilder(SI->getContext()).
2802 createBranchWeights(MDWeights));
2805 SI->addCase(Cst, NewBB);
2807 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2808 Builder.SetInsertPoint(NewBB);
2809 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2810 Builder.CreateBr(SuccBlock);
2811 PHIUse->addIncoming(NewCst, NewBB);
2815 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2816 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2817 /// fold it into a switch instruction if so.
2818 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *DL,
2819 IRBuilder<> &Builder) {
2820 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2821 if (!Cond) return false;
2823 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2824 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2825 // 'setne's and'ed together, collect them.
2827 // Try to gather values from a chain of and/or to be turned into a switch
2828 ConstantComparesGatherer ConstantCompare{Cond, DL};
2829 // Unpack the result
2830 SmallVectorImpl<ConstantInt*> &Values = ConstantCompare.Vals;
2831 Value *CompVal = ConstantCompare.CompValue;
2832 unsigned UsedICmps = ConstantCompare.UsedICmps;
2833 Value *ExtraCase = ConstantCompare.Extra;
2835 // If we didn't have a multiply compared value, fail.
2836 if (!CompVal) return false;
2838 // Avoid turning single icmps into a switch.
2842 bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
2844 // There might be duplicate constants in the list, which the switch
2845 // instruction can't handle, remove them now.
2846 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2847 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2849 // If Extra was used, we require at least two switch values to do the
2850 // transformation. A switch with one value is just an cond branch.
2851 if (ExtraCase && Values.size() < 2) return false;
2853 // TODO: Preserve branch weight metadata, similarly to how
2854 // FoldValueComparisonIntoPredecessors preserves it.
2856 // Figure out which block is which destination.
2857 BasicBlock *DefaultBB = BI->getSuccessor(1);
2858 BasicBlock *EdgeBB = BI->getSuccessor(0);
2859 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2861 BasicBlock *BB = BI->getParent();
2863 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2864 << " cases into SWITCH. BB is:\n" << *BB);
2866 // If there are any extra values that couldn't be folded into the switch
2867 // then we evaluate them with an explicit branch first. Split the block
2868 // right before the condbr to handle it.
2870 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2871 // Remove the uncond branch added to the old block.
2872 TerminatorInst *OldTI = BB->getTerminator();
2873 Builder.SetInsertPoint(OldTI);
2876 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2878 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2880 OldTI->eraseFromParent();
2882 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2883 // for the edge we just added.
2884 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2886 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2887 << "\nEXTRABB = " << *BB);
2891 Builder.SetInsertPoint(BI);
2892 // Convert pointer to int before we switch.
2893 if (CompVal->getType()->isPointerTy()) {
2894 assert(DL && "Cannot switch on pointer without DataLayout");
2895 CompVal = Builder.CreatePtrToInt(CompVal,
2896 DL->getIntPtrType(CompVal->getType()),
2900 // Create the new switch instruction now.
2901 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2903 // Add all of the 'cases' to the switch instruction.
2904 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2905 New->addCase(Values[i], EdgeBB);
2907 // We added edges from PI to the EdgeBB. As such, if there were any
2908 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2909 // the number of edges added.
2910 for (BasicBlock::iterator BBI = EdgeBB->begin();
2911 isa<PHINode>(BBI); ++BBI) {
2912 PHINode *PN = cast<PHINode>(BBI);
2913 Value *InVal = PN->getIncomingValueForBlock(BB);
2914 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2915 PN->addIncoming(InVal, BB);
2918 // Erase the old branch instruction.
2919 EraseTerminatorInstAndDCECond(BI);
2921 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2925 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2926 // If this is a trivial landing pad that just continues unwinding the caught
2927 // exception then zap the landing pad, turning its invokes into calls.
2928 BasicBlock *BB = RI->getParent();
2929 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2930 if (RI->getValue() != LPInst)
2931 // Not a landing pad, or the resume is not unwinding the exception that
2932 // caused control to branch here.
2935 // Check that there are no other instructions except for debug intrinsics.
2936 BasicBlock::iterator I = LPInst, E = RI;
2938 if (!isa<DbgInfoIntrinsic>(I))
2941 // Turn all invokes that unwind here into calls and delete the basic block.
2942 bool InvokeRequiresTableEntry = false;
2943 bool Changed = false;
2944 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2945 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2947 if (II->hasFnAttr(Attribute::UWTable)) {
2948 // Don't remove an `invoke' instruction if the ABI requires an entry into
2950 InvokeRequiresTableEntry = true;
2954 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2956 // Insert a call instruction before the invoke.
2957 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2959 Call->setCallingConv(II->getCallingConv());
2960 Call->setAttributes(II->getAttributes());
2961 Call->setDebugLoc(II->getDebugLoc());
2963 // Anything that used the value produced by the invoke instruction now uses
2964 // the value produced by the call instruction. Note that we do this even
2965 // for void functions and calls with no uses so that the callgraph edge is
2967 II->replaceAllUsesWith(Call);
2968 BB->removePredecessor(II->getParent());
2970 // Insert a branch to the normal destination right before the invoke.
2971 BranchInst::Create(II->getNormalDest(), II);
2973 // Finally, delete the invoke instruction!
2974 II->eraseFromParent();
2978 if (!InvokeRequiresTableEntry)
2979 // The landingpad is now unreachable. Zap it.
2980 BB->eraseFromParent();
2985 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2986 BasicBlock *BB = RI->getParent();
2987 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2989 // Find predecessors that end with branches.
2990 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2991 SmallVector<BranchInst*, 8> CondBranchPreds;
2992 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2993 BasicBlock *P = *PI;
2994 TerminatorInst *PTI = P->getTerminator();
2995 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2996 if (BI->isUnconditional())
2997 UncondBranchPreds.push_back(P);
2999 CondBranchPreds.push_back(BI);
3003 // If we found some, do the transformation!
3004 if (!UncondBranchPreds.empty() && DupRet) {
3005 while (!UncondBranchPreds.empty()) {
3006 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
3007 DEBUG(dbgs() << "FOLDING: " << *BB
3008 << "INTO UNCOND BRANCH PRED: " << *Pred);
3009 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
3012 // If we eliminated all predecessors of the block, delete the block now.
3013 if (pred_begin(BB) == pred_end(BB))
3014 // We know there are no successors, so just nuke the block.
3015 BB->eraseFromParent();
3020 // Check out all of the conditional branches going to this return
3021 // instruction. If any of them just select between returns, change the
3022 // branch itself into a select/return pair.
3023 while (!CondBranchPreds.empty()) {
3024 BranchInst *BI = CondBranchPreds.pop_back_val();
3026 // Check to see if the non-BB successor is also a return block.
3027 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
3028 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
3029 SimplifyCondBranchToTwoReturns(BI, Builder))
3035 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
3036 BasicBlock *BB = UI->getParent();
3038 bool Changed = false;
3040 // If there are any instructions immediately before the unreachable that can
3041 // be removed, do so.
3042 while (UI != BB->begin()) {
3043 BasicBlock::iterator BBI = UI;
3045 // Do not delete instructions that can have side effects which might cause
3046 // the unreachable to not be reachable; specifically, calls and volatile
3047 // operations may have this effect.
3048 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
3050 if (BBI->mayHaveSideEffects()) {
3051 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
3052 if (SI->isVolatile())
3054 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
3055 if (LI->isVolatile())
3057 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
3058 if (RMWI->isVolatile())
3060 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
3061 if (CXI->isVolatile())
3063 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
3064 !isa<LandingPadInst>(BBI)) {
3067 // Note that deleting LandingPad's here is in fact okay, although it
3068 // involves a bit of subtle reasoning. If this inst is a LandingPad,
3069 // all the predecessors of this block will be the unwind edges of Invokes,
3070 // and we can therefore guarantee this block will be erased.
3073 // Delete this instruction (any uses are guaranteed to be dead)
3074 if (!BBI->use_empty())
3075 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
3076 BBI->eraseFromParent();
3080 // If the unreachable instruction is the first in the block, take a gander
3081 // at all of the predecessors of this instruction, and simplify them.
3082 if (&BB->front() != UI) return Changed;
3084 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
3085 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
3086 TerminatorInst *TI = Preds[i]->getTerminator();
3087 IRBuilder<> Builder(TI);
3088 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3089 if (BI->isUnconditional()) {
3090 if (BI->getSuccessor(0) == BB) {
3091 new UnreachableInst(TI->getContext(), TI);
3092 TI->eraseFromParent();
3096 if (BI->getSuccessor(0) == BB) {
3097 Builder.CreateBr(BI->getSuccessor(1));
3098 EraseTerminatorInstAndDCECond(BI);
3099 } else if (BI->getSuccessor(1) == BB) {
3100 Builder.CreateBr(BI->getSuccessor(0));
3101 EraseTerminatorInstAndDCECond(BI);
3105 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3106 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3108 if (i.getCaseSuccessor() == BB) {
3109 BB->removePredecessor(SI->getParent());
3114 // If the default value is unreachable, figure out the most popular
3115 // destination and make it the default.
3116 if (SI->getDefaultDest() == BB) {
3117 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
3118 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3120 std::pair<unsigned, unsigned> &entry =
3121 Popularity[i.getCaseSuccessor()];
3122 if (entry.first == 0) {
3124 entry.second = i.getCaseIndex();
3130 // Find the most popular block.
3131 unsigned MaxPop = 0;
3132 unsigned MaxIndex = 0;
3133 BasicBlock *MaxBlock = nullptr;
3134 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
3135 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
3136 if (I->second.first > MaxPop ||
3137 (I->second.first == MaxPop && MaxIndex > I->second.second)) {
3138 MaxPop = I->second.first;
3139 MaxIndex = I->second.second;
3140 MaxBlock = I->first;
3144 // Make this the new default, allowing us to delete any explicit
3146 SI->setDefaultDest(MaxBlock);
3149 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
3151 if (isa<PHINode>(MaxBlock->begin()))
3152 for (unsigned i = 0; i != MaxPop-1; ++i)
3153 MaxBlock->removePredecessor(SI->getParent());
3155 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3157 if (i.getCaseSuccessor() == MaxBlock) {
3163 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
3164 if (II->getUnwindDest() == BB) {
3165 // Convert the invoke to a call instruction. This would be a good
3166 // place to note that the call does not throw though.
3167 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
3168 II->removeFromParent(); // Take out of symbol table
3170 // Insert the call now...
3171 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
3172 Builder.SetInsertPoint(BI);
3173 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
3174 Args, II->getName());
3175 CI->setCallingConv(II->getCallingConv());
3176 CI->setAttributes(II->getAttributes());
3177 // If the invoke produced a value, the call does now instead.
3178 II->replaceAllUsesWith(CI);
3185 // If this block is now dead, remove it.
3186 if (pred_begin(BB) == pred_end(BB) &&
3187 BB != &BB->getParent()->getEntryBlock()) {
3188 // We know there are no successors, so just nuke the block.
3189 BB->eraseFromParent();
3196 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
3197 /// integer range comparison into a sub, an icmp and a branch.
3198 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3199 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3201 // Make sure all cases point to the same destination and gather the values.
3202 SmallVector<ConstantInt *, 16> Cases;
3203 SwitchInst::CaseIt I = SI->case_begin();
3204 Cases.push_back(I.getCaseValue());
3205 SwitchInst::CaseIt PrevI = I++;
3206 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
3207 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
3209 Cases.push_back(I.getCaseValue());
3211 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
3213 // Sort the case values, then check if they form a range we can transform.
3214 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3215 for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
3216 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
3220 Constant *Offset = ConstantExpr::getNeg(Cases.back());
3221 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
3223 Value *Sub = SI->getCondition();
3224 if (!Offset->isNullValue())
3225 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
3227 // If NumCases overflowed, then all possible values jump to the successor.
3228 if (NumCases->isNullValue() && SI->getNumCases() != 0)
3229 Cmp = ConstantInt::getTrue(SI->getContext());
3231 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3232 BranchInst *NewBI = Builder.CreateCondBr(
3233 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
3235 // Update weight for the newly-created conditional branch.
3236 SmallVector<uint64_t, 8> Weights;
3237 bool HasWeights = HasBranchWeights(SI);
3239 GetBranchWeights(SI, Weights);
3240 if (Weights.size() == 1 + SI->getNumCases()) {
3241 // Combine all weights for the cases to be the true weight of NewBI.
3242 // We assume that the sum of all weights for a Terminator can fit into 32
3244 uint32_t NewTrueWeight = 0;
3245 for (unsigned I = 1, E = Weights.size(); I != E; ++I)
3246 NewTrueWeight += (uint32_t)Weights[I];
3247 NewBI->setMetadata(LLVMContext::MD_prof,
3248 MDBuilder(SI->getContext()).
3249 createBranchWeights(NewTrueWeight,
3250 (uint32_t)Weights[0]));
3254 // Prune obsolete incoming values off the successor's PHI nodes.
3255 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
3256 isa<PHINode>(BBI); ++BBI) {
3257 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
3258 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3260 SI->eraseFromParent();
3265 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
3266 /// and use it to remove dead cases.
3267 static bool EliminateDeadSwitchCases(SwitchInst *SI, const DataLayout *DL,
3268 AssumptionTracker *AT) {
3269 Value *Cond = SI->getCondition();
3270 unsigned Bits = Cond->getType()->getIntegerBitWidth();
3271 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3272 computeKnownBits(Cond, KnownZero, KnownOne, DL, 0, AT, SI);
3274 // Gather dead cases.
3275 SmallVector<ConstantInt*, 8> DeadCases;
3276 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3277 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3278 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3279 DeadCases.push_back(I.getCaseValue());
3280 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3281 << I.getCaseValue() << "' is dead.\n");
3285 SmallVector<uint64_t, 8> Weights;
3286 bool HasWeight = HasBranchWeights(SI);
3288 GetBranchWeights(SI, Weights);
3289 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3292 // Remove dead cases from the switch.
3293 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3294 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3295 assert(Case != SI->case_default() &&
3296 "Case was not found. Probably mistake in DeadCases forming.");
3298 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3302 // Prune unused values from PHI nodes.
3303 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3304 SI->removeCase(Case);
3306 if (HasWeight && Weights.size() >= 2) {
3307 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3308 SI->setMetadata(LLVMContext::MD_prof,
3309 MDBuilder(SI->getParent()->getContext()).
3310 createBranchWeights(MDWeights));
3313 return !DeadCases.empty();
3316 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
3317 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3318 /// by an unconditional branch), look at the phi node for BB in the successor
3319 /// block and see if the incoming value is equal to CaseValue. If so, return
3320 /// the phi node, and set PhiIndex to BB's index in the phi node.
3321 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3324 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3325 return nullptr; // BB must be empty to be a candidate for simplification.
3326 if (!BB->getSinglePredecessor())
3327 return nullptr; // BB must be dominated by the switch.
3329 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3330 if (!Branch || !Branch->isUnconditional())
3331 return nullptr; // Terminator must be unconditional branch.
3333 BasicBlock *Succ = Branch->getSuccessor(0);
3335 BasicBlock::iterator I = Succ->begin();
3336 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3337 int Idx = PHI->getBasicBlockIndex(BB);
3338 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3340 Value *InValue = PHI->getIncomingValue(Idx);
3341 if (InValue != CaseValue) continue;
3350 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
3351 /// instruction to a phi node dominated by the switch, if that would mean that
3352 /// some of the destination blocks of the switch can be folded away.
3353 /// Returns true if a change is made.
3354 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3355 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3356 ForwardingNodesMap ForwardingNodes;
3358 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3359 ConstantInt *CaseValue = I.getCaseValue();
3360 BasicBlock *CaseDest = I.getCaseSuccessor();
3363 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3367 ForwardingNodes[PHI].push_back(PhiIndex);
3370 bool Changed = false;
3372 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3373 E = ForwardingNodes.end(); I != E; ++I) {
3374 PHINode *Phi = I->first;
3375 SmallVectorImpl<int> &Indexes = I->second;
3377 if (Indexes.size() < 2) continue;
3379 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3380 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3387 /// ValidLookupTableConstant - Return true if the backend will be able to handle
3388 /// initializing an array of constants like C.
3389 static bool ValidLookupTableConstant(Constant *C) {
3390 if (C->isThreadDependent())
3392 if (C->isDLLImportDependent())
3395 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3396 return CE->isGEPWithNoNotionalOverIndexing();
3398 return isa<ConstantFP>(C) ||
3399 isa<ConstantInt>(C) ||
3400 isa<ConstantPointerNull>(C) ||
3401 isa<GlobalValue>(C) ||
3405 /// LookupConstant - If V is a Constant, return it. Otherwise, try to look up
3406 /// its constant value in ConstantPool, returning 0 if it's not there.
3407 static Constant *LookupConstant(Value *V,
3408 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3409 if (Constant *C = dyn_cast<Constant>(V))
3411 return ConstantPool.lookup(V);
3414 /// ConstantFold - Try to fold instruction I into a constant. This works for
3415 /// simple instructions such as binary operations where both operands are
3416 /// constant or can be replaced by constants from the ConstantPool. Returns the
3417 /// resulting constant on success, 0 otherwise.
3419 ConstantFold(Instruction *I,
3420 const SmallDenseMap<Value *, Constant *> &ConstantPool,
3421 const DataLayout *DL) {
3422 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3423 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3426 if (A->isAllOnesValue())
3427 return LookupConstant(Select->getTrueValue(), ConstantPool);
3428 if (A->isNullValue())
3429 return LookupConstant(Select->getFalseValue(), ConstantPool);
3433 SmallVector<Constant *, 4> COps;
3434 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
3435 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
3441 if (CmpInst *Cmp = dyn_cast<CmpInst>(I))
3442 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
3445 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), COps, DL);
3448 /// GetCaseResults - Try to determine the resulting constant values in phi nodes
3449 /// at the common destination basic block, *CommonDest, for one of the case
3450 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3451 /// case), of a switch instruction SI.
3453 GetCaseResults(SwitchInst *SI,
3454 ConstantInt *CaseVal,
3455 BasicBlock *CaseDest,
3456 BasicBlock **CommonDest,
3457 SmallVectorImpl<std::pair<PHINode *, Constant *> > &Res,
3458 const DataLayout *DL) {
3459 // The block from which we enter the common destination.
3460 BasicBlock *Pred = SI->getParent();
3462 // If CaseDest is empty except for some side-effect free instructions through
3463 // which we can constant-propagate the CaseVal, continue to its successor.
3464 SmallDenseMap<Value*, Constant*> ConstantPool;
3465 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3466 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3468 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3469 // If the terminator is a simple branch, continue to the next block.
3470 if (T->getNumSuccessors() != 1)
3473 CaseDest = T->getSuccessor(0);
3474 } else if (isa<DbgInfoIntrinsic>(I)) {
3475 // Skip debug intrinsic.
3477 } else if (Constant *C = ConstantFold(I, ConstantPool, DL)) {
3478 // Instruction is side-effect free and constant.
3479 ConstantPool.insert(std::make_pair(I, C));
3485 // If we did not have a CommonDest before, use the current one.
3487 *CommonDest = CaseDest;
3488 // If the destination isn't the common one, abort.
3489 if (CaseDest != *CommonDest)
3492 // Get the values for this case from phi nodes in the destination block.
3493 BasicBlock::iterator I = (*CommonDest)->begin();
3494 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3495 int Idx = PHI->getBasicBlockIndex(Pred);
3499 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3504 // Note: If the constant comes from constant-propagating the case value
3505 // through the CaseDest basic block, it will be safe to remove the
3506 // instructions in that block. They cannot be used (except in the phi nodes
3507 // we visit) outside CaseDest, because that block does not dominate its
3508 // successor. If it did, we would not be in this phi node.
3510 // Be conservative about which kinds of constants we support.
3511 if (!ValidLookupTableConstant(ConstVal))
3514 Res.push_back(std::make_pair(PHI, ConstVal));
3517 return Res.size() > 0;
3520 // MapCaseToResult - Helper function used to
3521 // add CaseVal to the list of cases that generate Result.
3522 static void MapCaseToResult(ConstantInt *CaseVal,
3523 SwitchCaseResultVectorTy &UniqueResults,
3525 for (auto &I : UniqueResults) {
3526 if (I.first == Result) {
3527 I.second.push_back(CaseVal);
3531 UniqueResults.push_back(std::make_pair(Result,
3532 SmallVector<ConstantInt*, 4>(1, CaseVal)));
3535 // InitializeUniqueCases - Helper function that initializes a map containing
3536 // results for the PHI node of the common destination block for a switch
3537 // instruction. Returns false if multiple PHI nodes have been found or if
3538 // there is not a common destination block for the switch.
3539 static bool InitializeUniqueCases(
3540 SwitchInst *SI, const DataLayout *DL, PHINode *&PHI,
3541 BasicBlock *&CommonDest,
3542 SwitchCaseResultVectorTy &UniqueResults,
3543 Constant *&DefaultResult) {
3544 for (auto &I : SI->cases()) {
3545 ConstantInt *CaseVal = I.getCaseValue();
3547 // Resulting value at phi nodes for this case value.
3548 SwitchCaseResultsTy Results;
3549 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
3553 // Only one value per case is permitted
3554 if (Results.size() > 1)
3556 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
3558 // Check the PHI consistency.
3560 PHI = Results[0].first;
3561 else if (PHI != Results[0].first)
3564 // Find the default result value.
3565 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
3566 BasicBlock *DefaultDest = SI->getDefaultDest();
3567 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
3569 // If the default value is not found abort unless the default destination
3572 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
3573 if ((!DefaultResult &&
3574 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
3580 // ConvertTwoCaseSwitch - Helper function that checks if it is possible to
3581 // transform a switch with only two cases (or two cases + default)
3582 // that produces a result into a value select.
3585 // case 10: %0 = icmp eq i32 %a, 10
3586 // return 10; %1 = select i1 %0, i32 10, i32 4
3587 // case 20: ----> %2 = icmp eq i32 %a, 20
3588 // return 2; %3 = select i1 %2, i32 2, i32 %1
3593 ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
3594 Constant *DefaultResult, Value *Condition,
3595 IRBuilder<> &Builder) {
3596 assert(ResultVector.size() == 2 &&
3597 "We should have exactly two unique results at this point");
3598 // If we are selecting between only two cases transform into a simple
3599 // select or a two-way select if default is possible.
3600 if (ResultVector[0].second.size() == 1 &&
3601 ResultVector[1].second.size() == 1) {
3602 ConstantInt *const FirstCase = ResultVector[0].second[0];
3603 ConstantInt *const SecondCase = ResultVector[1].second[0];
3605 bool DefaultCanTrigger = DefaultResult;
3606 Value *SelectValue = ResultVector[1].first;
3607 if (DefaultCanTrigger) {
3608 Value *const ValueCompare =
3609 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
3610 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
3611 DefaultResult, "switch.select");
3613 Value *const ValueCompare =
3614 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
3615 return Builder.CreateSelect(ValueCompare, ResultVector[0].first, SelectValue,
3622 // RemoveSwitchAfterSelectConversion - Helper function to cleanup a switch
3623 // instruction that has been converted into a select, fixing up PHI nodes and
3625 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
3627 IRBuilder<> &Builder) {
3628 BasicBlock *SelectBB = SI->getParent();
3629 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
3630 PHI->removeIncomingValue(SelectBB);
3631 PHI->addIncoming(SelectValue, SelectBB);
3633 Builder.CreateBr(PHI->getParent());
3635 // Remove the switch.
3636 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
3637 BasicBlock *Succ = SI->getSuccessor(i);
3639 if (Succ == PHI->getParent())
3641 Succ->removePredecessor(SelectBB);
3643 SI->eraseFromParent();
3646 /// SwitchToSelect - If the switch is only used to initialize one or more
3647 /// phi nodes in a common successor block with only two different
3648 /// constant values, replace the switch with select.
3649 static bool SwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
3650 const DataLayout *DL, AssumptionTracker *AT) {
3651 Value *const Cond = SI->getCondition();
3652 PHINode *PHI = nullptr;
3653 BasicBlock *CommonDest = nullptr;
3654 Constant *DefaultResult;
3655 SwitchCaseResultVectorTy UniqueResults;
3656 // Collect all the cases that will deliver the same value from the switch.
3657 if (!InitializeUniqueCases(SI, DL, PHI, CommonDest, UniqueResults,
3660 // Selects choose between maximum two values.
3661 if (UniqueResults.size() != 2)
3663 assert(PHI != nullptr && "PHI for value select not found");
3665 Builder.SetInsertPoint(SI);
3666 Value *SelectValue = ConvertTwoCaseSwitch(
3668 DefaultResult, Cond, Builder);
3670 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
3673 // The switch couldn't be converted into a select.
3678 /// SwitchLookupTable - This class represents a lookup table that can be used
3679 /// to replace a switch.
3680 class SwitchLookupTable {
3682 /// SwitchLookupTable - Create a lookup table to use as a switch replacement
3683 /// with the contents of Values, using DefaultValue to fill any holes in the
3685 SwitchLookupTable(Module &M,
3687 ConstantInt *Offset,
3688 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3689 Constant *DefaultValue,
3690 const DataLayout *DL);
3692 /// BuildLookup - Build instructions with Builder to retrieve the value at
3693 /// the position given by Index in the lookup table.
3694 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3696 /// WouldFitInRegister - Return true if a table with TableSize elements of
3697 /// type ElementType would fit in a target-legal register.
3698 static bool WouldFitInRegister(const DataLayout *DL,
3700 const Type *ElementType);
3703 // Depending on the contents of the table, it can be represented in
3706 // For tables where each element contains the same value, we just have to
3707 // store that single value and return it for each lookup.
3710 // For tables where there is a linear relationship between table index
3711 // and values. We calculate the result with a simple multiplication
3712 // and addition instead of a table lookup.
3715 // For small tables with integer elements, we can pack them into a bitmap
3716 // that fits into a target-legal register. Values are retrieved by
3717 // shift and mask operations.
3720 // The table is stored as an array of values. Values are retrieved by load
3721 // instructions from the table.
3725 // For SingleValueKind, this is the single value.
3726 Constant *SingleValue;
3728 // For BitMapKind, this is the bitmap.
3729 ConstantInt *BitMap;
3730 IntegerType *BitMapElementTy;
3732 // For LinearMapKind, these are the constants used to derive the value.
3733 ConstantInt *LinearOffset;
3734 ConstantInt *LinearMultiplier;
3736 // For ArrayKind, this is the array.
3737 GlobalVariable *Array;
3741 SwitchLookupTable::SwitchLookupTable(Module &M,
3743 ConstantInt *Offset,
3744 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3745 Constant *DefaultValue,
3746 const DataLayout *DL)
3747 : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
3748 LinearOffset(nullptr), LinearMultiplier(nullptr), Array(nullptr) {
3749 assert(Values.size() && "Can't build lookup table without values!");
3750 assert(TableSize >= Values.size() && "Can't fit values in table!");
3752 // If all values in the table are equal, this is that value.
3753 SingleValue = Values.begin()->second;
3755 Type *ValueType = Values.begin()->second->getType();
3757 // Build up the table contents.
3758 SmallVector<Constant*, 64> TableContents(TableSize);
3759 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3760 ConstantInt *CaseVal = Values[I].first;
3761 Constant *CaseRes = Values[I].second;
3762 assert(CaseRes->getType() == ValueType);
3764 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3766 TableContents[Idx] = CaseRes;
3768 if (CaseRes != SingleValue)
3769 SingleValue = nullptr;
3772 // Fill in any holes in the table with the default result.
3773 if (Values.size() < TableSize) {
3774 assert(DefaultValue &&
3775 "Need a default value to fill the lookup table holes.");
3776 assert(DefaultValue->getType() == ValueType);
3777 for (uint64_t I = 0; I < TableSize; ++I) {
3778 if (!TableContents[I])
3779 TableContents[I] = DefaultValue;
3782 if (DefaultValue != SingleValue)
3783 SingleValue = nullptr;
3786 // If each element in the table contains the same value, we only need to store
3787 // that single value.
3789 Kind = SingleValueKind;
3793 // Check if we can derive the value with a linear transformation from the
3795 if (isa<IntegerType>(ValueType)) {
3796 bool LinearMappingPossible = true;
3799 assert(TableSize >= 2 && "Should be a SingleValue table.");
3800 // Check if there is the same distance between two consecutive values.
3801 for (uint64_t I = 0; I < TableSize; ++I) {
3802 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
3804 // This is an undef. We could deal with it, but undefs in lookup tables
3805 // are very seldom. It's probably not worth the additional complexity.
3806 LinearMappingPossible = false;
3809 APInt Val = ConstVal->getValue();
3811 APInt Dist = Val - PrevVal;
3814 } else if (Dist != DistToPrev) {
3815 LinearMappingPossible = false;
3821 if (LinearMappingPossible) {
3822 LinearOffset = cast<ConstantInt>(TableContents[0]);
3823 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
3824 Kind = LinearMapKind;
3830 // If the type is integer and the table fits in a register, build a bitmap.
3831 if (WouldFitInRegister(DL, TableSize, ValueType)) {
3832 IntegerType *IT = cast<IntegerType>(ValueType);
3833 APInt TableInt(TableSize * IT->getBitWidth(), 0);
3834 for (uint64_t I = TableSize; I > 0; --I) {
3835 TableInt <<= IT->getBitWidth();
3836 // Insert values into the bitmap. Undef values are set to zero.
3837 if (!isa<UndefValue>(TableContents[I - 1])) {
3838 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
3839 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
3842 BitMap = ConstantInt::get(M.getContext(), TableInt);
3843 BitMapElementTy = IT;
3849 // Store the table in an array.
3850 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
3851 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3853 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3854 GlobalVariable::PrivateLinkage,
3857 Array->setUnnamedAddr(true);
3861 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
3863 case SingleValueKind:
3865 case LinearMapKind: {
3866 // Derive the result value from the input value.
3867 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
3868 false, "switch.idx.cast");
3869 if (!LinearMultiplier->isOne())
3870 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
3871 if (!LinearOffset->isZero())
3872 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
3876 // Type of the bitmap (e.g. i59).
3877 IntegerType *MapTy = BitMap->getType();
3879 // Cast Index to the same type as the bitmap.
3880 // Note: The Index is <= the number of elements in the table, so
3881 // truncating it to the width of the bitmask is safe.
3882 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
3884 // Multiply the shift amount by the element width.
3885 ShiftAmt = Builder.CreateMul(ShiftAmt,
3886 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
3890 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
3891 "switch.downshift");
3893 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
3897 // Make sure the table index will not overflow when treated as signed.
3898 IntegerType *IT = cast<IntegerType>(Index->getType());
3899 uint64_t TableSize = Array->getInitializer()->getType()
3900 ->getArrayNumElements();
3901 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
3902 Index = Builder.CreateZExt(Index,
3903 IntegerType::get(IT->getContext(),
3904 IT->getBitWidth() + 1),
3905 "switch.tableidx.zext");
3907 Value *GEPIndices[] = { Builder.getInt32(0), Index };
3908 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices,
3910 return Builder.CreateLoad(GEP, "switch.load");
3913 llvm_unreachable("Unknown lookup table kind!");
3916 bool SwitchLookupTable::WouldFitInRegister(const DataLayout *DL,
3918 const Type *ElementType) {
3921 const IntegerType *IT = dyn_cast<IntegerType>(ElementType);
3924 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
3925 // are <= 15, we could try to narrow the type.
3927 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
3928 if (TableSize >= UINT_MAX/IT->getBitWidth())
3930 return DL->fitsInLegalInteger(TableSize * IT->getBitWidth());
3933 /// ShouldBuildLookupTable - Determine whether a lookup table should be built
3934 /// for this switch, based on the number of cases, size of the table and the
3935 /// types of the results.
3936 static bool ShouldBuildLookupTable(SwitchInst *SI,
3938 const TargetTransformInfo &TTI,
3939 const DataLayout *DL,
3940 const SmallDenseMap<PHINode*, Type*>& ResultTypes) {
3941 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
3942 return false; // TableSize overflowed, or mul below might overflow.
3944 bool AllTablesFitInRegister = true;
3945 bool HasIllegalType = false;
3946 for (const auto &I : ResultTypes) {
3947 Type *Ty = I.second;
3949 // Saturate this flag to true.
3950 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
3952 // Saturate this flag to false.
3953 AllTablesFitInRegister = AllTablesFitInRegister &&
3954 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
3956 // If both flags saturate, we're done. NOTE: This *only* works with
3957 // saturating flags, and all flags have to saturate first due to the
3958 // non-deterministic behavior of iterating over a dense map.
3959 if (HasIllegalType && !AllTablesFitInRegister)
3963 // If each table would fit in a register, we should build it anyway.
3964 if (AllTablesFitInRegister)
3967 // Don't build a table that doesn't fit in-register if it has illegal types.
3971 // The table density should be at least 40%. This is the same criterion as for
3972 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3973 // FIXME: Find the best cut-off.
3974 return SI->getNumCases() * 10 >= TableSize * 4;
3977 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3978 /// phi nodes in a common successor block with different constant values,
3979 /// replace the switch with lookup tables.
3980 static bool SwitchToLookupTable(SwitchInst *SI,
3981 IRBuilder<> &Builder,
3982 const TargetTransformInfo &TTI,
3983 const DataLayout* DL) {
3984 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3986 // Only build lookup table when we have a target that supports it.
3987 if (!TTI.shouldBuildLookupTables())
3990 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3991 // split off a dense part and build a lookup table for that.
3993 // FIXME: This creates arrays of GEPs to constant strings, which means each
3994 // GEP needs a runtime relocation in PIC code. We should just build one big
3995 // string and lookup indices into that.
3997 // Ignore switches with less than three cases. Lookup tables will not make them
3998 // faster, so we don't analyze them.
3999 if (SI->getNumCases() < 3)
4002 // Figure out the corresponding result for each case value and phi node in the
4003 // common destination, as well as the the min and max case values.
4004 assert(SI->case_begin() != SI->case_end());
4005 SwitchInst::CaseIt CI = SI->case_begin();
4006 ConstantInt *MinCaseVal = CI.getCaseValue();
4007 ConstantInt *MaxCaseVal = CI.getCaseValue();
4009 BasicBlock *CommonDest = nullptr;
4010 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
4011 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
4012 SmallDenseMap<PHINode*, Constant*> DefaultResults;
4013 SmallDenseMap<PHINode*, Type*> ResultTypes;
4014 SmallVector<PHINode*, 4> PHIs;
4016 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
4017 ConstantInt *CaseVal = CI.getCaseValue();
4018 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
4019 MinCaseVal = CaseVal;
4020 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
4021 MaxCaseVal = CaseVal;
4023 // Resulting value at phi nodes for this case value.
4024 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
4026 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
4030 // Append the result from this case to the list for each phi.
4031 for (const auto &I : Results) {
4032 PHINode *PHI = I.first;
4033 Constant *Value = I.second;
4034 if (!ResultLists.count(PHI))
4035 PHIs.push_back(PHI);
4036 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
4040 // Keep track of the result types.
4041 for (PHINode *PHI : PHIs) {
4042 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
4045 uint64_t NumResults = ResultLists[PHIs[0]].size();
4046 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
4047 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
4048 bool TableHasHoles = (NumResults < TableSize);
4050 // If the table has holes, we need a constant result for the default case
4051 // or a bitmask that fits in a register.
4052 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
4053 bool HasDefaultResults = false;
4054 if (TableHasHoles) {
4055 HasDefaultResults = GetCaseResults(SI, nullptr, SI->getDefaultDest(),
4056 &CommonDest, DefaultResultsList, DL);
4059 bool NeedMask = (TableHasHoles && !HasDefaultResults);
4061 // As an extra penalty for the validity test we require more cases.
4062 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
4064 if (!(DL && DL->fitsInLegalInteger(TableSize)))
4068 for (const auto &I : DefaultResultsList) {
4069 PHINode *PHI = I.first;
4070 Constant *Result = I.second;
4071 DefaultResults[PHI] = Result;
4074 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
4077 // Create the BB that does the lookups.
4078 Module &Mod = *CommonDest->getParent()->getParent();
4079 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
4081 CommonDest->getParent(),
4084 // Compute the table index value.
4085 Builder.SetInsertPoint(SI);
4086 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
4089 // Compute the maximum table size representable by the integer type we are
4091 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
4092 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
4093 assert(MaxTableSize >= TableSize &&
4094 "It is impossible for a switch to have more entries than the max "
4095 "representable value of its input integer type's size.");
4097 // If we have a fully covered lookup table, unconditionally branch to the
4098 // lookup table BB. Otherwise, check if the condition value is within the case
4099 // range. If it is so, branch to the new BB. Otherwise branch to SI's default
4101 const bool GeneratingCoveredLookupTable = MaxTableSize == TableSize;
4102 if (GeneratingCoveredLookupTable) {
4103 Builder.CreateBr(LookupBB);
4104 // We cached PHINodes in PHIs, to avoid accessing deleted PHINodes later,
4105 // do not delete PHINodes here.
4106 SI->getDefaultDest()->removePredecessor(SI->getParent(),
4107 true/*DontDeleteUselessPHIs*/);
4109 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
4110 MinCaseVal->getType(), TableSize));
4111 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
4114 // Populate the BB that does the lookups.
4115 Builder.SetInsertPoint(LookupBB);
4118 // Before doing the lookup we do the hole check.
4119 // The LookupBB is therefore re-purposed to do the hole check
4120 // and we create a new LookupBB.
4121 BasicBlock *MaskBB = LookupBB;
4122 MaskBB->setName("switch.hole_check");
4123 LookupBB = BasicBlock::Create(Mod.getContext(),
4125 CommonDest->getParent(),
4128 // Make the mask's bitwidth at least 8bit and a power-of-2 to avoid
4129 // unnecessary illegal types.
4130 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
4131 APInt MaskInt(TableSizePowOf2, 0);
4132 APInt One(TableSizePowOf2, 1);
4133 // Build bitmask; fill in a 1 bit for every case.
4134 const ResultListTy &ResultList = ResultLists[PHIs[0]];
4135 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
4136 uint64_t Idx = (ResultList[I].first->getValue() -
4137 MinCaseVal->getValue()).getLimitedValue();
4138 MaskInt |= One << Idx;
4140 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
4142 // Get the TableIndex'th bit of the bitmask.
4143 // If this bit is 0 (meaning hole) jump to the default destination,
4144 // else continue with table lookup.
4145 IntegerType *MapTy = TableMask->getType();
4146 Value *MaskIndex = Builder.CreateZExtOrTrunc(TableIndex, MapTy,
4147 "switch.maskindex");
4148 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex,
4150 Value *LoBit = Builder.CreateTrunc(Shifted,
4151 Type::getInt1Ty(Mod.getContext()),
4153 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
4155 Builder.SetInsertPoint(LookupBB);
4156 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
4159 bool ReturnedEarly = false;
4160 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
4161 PHINode *PHI = PHIs[I];
4163 // If using a bitmask, use any value to fill the lookup table holes.
4164 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
4165 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI],
4168 Value *Result = Table.BuildLookup(TableIndex, Builder);
4170 // If the result is used to return immediately from the function, we want to
4171 // do that right here.
4172 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
4173 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
4174 Builder.CreateRet(Result);
4175 ReturnedEarly = true;
4179 PHI->addIncoming(Result, LookupBB);
4183 Builder.CreateBr(CommonDest);
4185 // Remove the switch.
4186 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4187 BasicBlock *Succ = SI->getSuccessor(i);
4189 if (Succ == SI->getDefaultDest())
4191 Succ->removePredecessor(SI->getParent());
4193 SI->eraseFromParent();
4197 ++NumLookupTablesHoles;
4201 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
4202 BasicBlock *BB = SI->getParent();
4204 if (isValueEqualityComparison(SI)) {
4205 // If we only have one predecessor, and if it is a branch on this value,
4206 // see if that predecessor totally determines the outcome of this switch.
4207 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4208 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
4209 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4211 Value *Cond = SI->getCondition();
4212 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
4213 if (SimplifySwitchOnSelect(SI, Select))
4214 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4216 // If the block only contains the switch, see if we can fold the block
4217 // away into any preds.
4218 BasicBlock::iterator BBI = BB->begin();
4219 // Ignore dbg intrinsics.
4220 while (isa<DbgInfoIntrinsic>(BBI))
4223 if (FoldValueComparisonIntoPredecessors(SI, Builder))
4224 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4227 // Try to transform the switch into an icmp and a branch.
4228 if (TurnSwitchRangeIntoICmp(SI, Builder))
4229 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4231 // Remove unreachable cases.
4232 if (EliminateDeadSwitchCases(SI, DL, AT))
4233 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4235 if (SwitchToSelect(SI, Builder, DL, AT))
4236 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4238 if (ForwardSwitchConditionToPHI(SI))
4239 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4241 if (SwitchToLookupTable(SI, Builder, TTI, DL))
4242 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4247 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
4248 BasicBlock *BB = IBI->getParent();
4249 bool Changed = false;
4251 // Eliminate redundant destinations.
4252 SmallPtrSet<Value *, 8> Succs;
4253 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
4254 BasicBlock *Dest = IBI->getDestination(i);
4255 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
4256 Dest->removePredecessor(BB);
4257 IBI->removeDestination(i);
4263 if (IBI->getNumDestinations() == 0) {
4264 // If the indirectbr has no successors, change it to unreachable.
4265 new UnreachableInst(IBI->getContext(), IBI);
4266 EraseTerminatorInstAndDCECond(IBI);
4270 if (IBI->getNumDestinations() == 1) {
4271 // If the indirectbr has one successor, change it to a direct branch.
4272 BranchInst::Create(IBI->getDestination(0), IBI);
4273 EraseTerminatorInstAndDCECond(IBI);
4277 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
4278 if (SimplifyIndirectBrOnSelect(IBI, SI))
4279 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4284 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
4285 BasicBlock *BB = BI->getParent();
4287 if (SinkCommon && SinkThenElseCodeToEnd(BI))
4290 // If the Terminator is the only non-phi instruction, simplify the block.
4291 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg();
4292 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
4293 TryToSimplifyUncondBranchFromEmptyBlock(BB))
4296 // If the only instruction in the block is a seteq/setne comparison
4297 // against a constant, try to simplify the block.
4298 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
4299 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
4300 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
4302 if (I->isTerminator() &&
4303 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI,
4304 BonusInstThreshold, DL, AT))
4308 // If this basic block is ONLY a compare and a branch, and if a predecessor
4309 // branches to us and our successor, fold the comparison into the
4310 // predecessor and use logical operations to update the incoming value
4311 // for PHI nodes in common successor.
4312 if (FoldBranchToCommonDest(BI, DL, BonusInstThreshold))
4313 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4318 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
4319 BasicBlock *BB = BI->getParent();
4321 // Conditional branch
4322 if (isValueEqualityComparison(BI)) {
4323 // If we only have one predecessor, and if it is a branch on this value,
4324 // see if that predecessor totally determines the outcome of this
4326 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4327 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
4328 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4330 // This block must be empty, except for the setcond inst, if it exists.
4331 // Ignore dbg intrinsics.
4332 BasicBlock::iterator I = BB->begin();
4333 // Ignore dbg intrinsics.
4334 while (isa<DbgInfoIntrinsic>(I))
4337 if (FoldValueComparisonIntoPredecessors(BI, Builder))
4338 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4339 } else if (&*I == cast<Instruction>(BI->getCondition())){
4341 // Ignore dbg intrinsics.
4342 while (isa<DbgInfoIntrinsic>(I))
4344 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
4345 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4349 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
4350 if (SimplifyBranchOnICmpChain(BI, DL, Builder))
4353 // If this basic block is ONLY a compare and a branch, and if a predecessor
4354 // branches to us and one of our successors, fold the comparison into the
4355 // predecessor and use logical operations to pick the right destination.
4356 if (FoldBranchToCommonDest(BI, DL, BonusInstThreshold))
4357 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4359 // We have a conditional branch to two blocks that are only reachable
4360 // from BI. We know that the condbr dominates the two blocks, so see if
4361 // there is any identical code in the "then" and "else" blocks. If so, we
4362 // can hoist it up to the branching block.
4363 if (BI->getSuccessor(0)->getSinglePredecessor()) {
4364 if (BI->getSuccessor(1)->getSinglePredecessor()) {
4365 if (HoistThenElseCodeToIf(BI, DL))
4366 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4368 // If Successor #1 has multiple preds, we may be able to conditionally
4369 // execute Successor #0 if it branches to Successor #1.
4370 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
4371 if (Succ0TI->getNumSuccessors() == 1 &&
4372 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
4373 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), DL))
4374 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4376 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
4377 // If Successor #0 has multiple preds, we may be able to conditionally
4378 // execute Successor #1 if it branches to Successor #0.
4379 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
4380 if (Succ1TI->getNumSuccessors() == 1 &&
4381 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
4382 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), DL))
4383 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4386 // If this is a branch on a phi node in the current block, thread control
4387 // through this block if any PHI node entries are constants.
4388 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
4389 if (PN->getParent() == BI->getParent())
4390 if (FoldCondBranchOnPHI(BI, DL))
4391 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4393 // Scan predecessor blocks for conditional branches.
4394 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
4395 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
4396 if (PBI != BI && PBI->isConditional())
4397 if (SimplifyCondBranchToCondBranch(PBI, BI))
4398 return SimplifyCFG(BB, TTI, BonusInstThreshold, DL, AT) | true;
4403 /// Check if passing a value to an instruction will cause undefined behavior.
4404 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
4405 Constant *C = dyn_cast<Constant>(V);
4412 if (C->isNullValue()) {
4413 // Only look at the first use, avoid hurting compile time with long uselists
4414 User *Use = *I->user_begin();
4416 // Now make sure that there are no instructions in between that can alter
4417 // control flow (eg. calls)
4418 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
4419 if (i == I->getParent()->end() || i->mayHaveSideEffects())
4422 // Look through GEPs. A load from a GEP derived from NULL is still undefined
4423 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
4424 if (GEP->getPointerOperand() == I)
4425 return passingValueIsAlwaysUndefined(V, GEP);
4427 // Look through bitcasts.
4428 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
4429 return passingValueIsAlwaysUndefined(V, BC);
4431 // Load from null is undefined.
4432 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
4433 if (!LI->isVolatile())
4434 return LI->getPointerAddressSpace() == 0;
4436 // Store to null is undefined.
4437 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
4438 if (!SI->isVolatile())
4439 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
4444 /// If BB has an incoming value that will always trigger undefined behavior
4445 /// (eg. null pointer dereference), remove the branch leading here.
4446 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
4447 for (BasicBlock::iterator i = BB->begin();
4448 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
4449 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
4450 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
4451 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
4452 IRBuilder<> Builder(T);
4453 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
4454 BB->removePredecessor(PHI->getIncomingBlock(i));
4455 // Turn uncoditional branches into unreachables and remove the dead
4456 // destination from conditional branches.
4457 if (BI->isUnconditional())
4458 Builder.CreateUnreachable();
4460 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
4461 BI->getSuccessor(0));
4462 BI->eraseFromParent();
4465 // TODO: SwitchInst.
4471 bool SimplifyCFGOpt::run(BasicBlock *BB) {
4472 bool Changed = false;
4474 assert(BB && BB->getParent() && "Block not embedded in function!");
4475 assert(BB->getTerminator() && "Degenerate basic block encountered!");
4477 // Remove basic blocks that have no predecessors (except the entry block)...
4478 // or that just have themself as a predecessor. These are unreachable.
4479 if ((pred_begin(BB) == pred_end(BB) &&
4480 BB != &BB->getParent()->getEntryBlock()) ||
4481 BB->getSinglePredecessor() == BB) {
4482 DEBUG(dbgs() << "Removing BB: \n" << *BB);
4483 DeleteDeadBlock(BB);
4487 // Check to see if we can constant propagate this terminator instruction
4489 Changed |= ConstantFoldTerminator(BB, true);
4491 // Check for and eliminate duplicate PHI nodes in this block.
4492 Changed |= EliminateDuplicatePHINodes(BB);
4494 // Check for and remove branches that will always cause undefined behavior.
4495 Changed |= removeUndefIntroducingPredecessor(BB);
4497 // Merge basic blocks into their predecessor if there is only one distinct
4498 // pred, and if there is only one distinct successor of the predecessor, and
4499 // if there are no PHI nodes.
4501 if (MergeBlockIntoPredecessor(BB))
4504 IRBuilder<> Builder(BB);
4506 // If there is a trivial two-entry PHI node in this basic block, and we can
4507 // eliminate it, do so now.
4508 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
4509 if (PN->getNumIncomingValues() == 2)
4510 Changed |= FoldTwoEntryPHINode(PN, DL);
4512 Builder.SetInsertPoint(BB->getTerminator());
4513 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
4514 if (BI->isUnconditional()) {
4515 if (SimplifyUncondBranch(BI, Builder)) return true;
4517 if (SimplifyCondBranch(BI, Builder)) return true;
4519 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
4520 if (SimplifyReturn(RI, Builder)) return true;
4521 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
4522 if (SimplifyResume(RI, Builder)) return true;
4523 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
4524 if (SimplifySwitch(SI, Builder)) return true;
4525 } else if (UnreachableInst *UI =
4526 dyn_cast<UnreachableInst>(BB->getTerminator())) {
4527 if (SimplifyUnreachable(UI)) return true;
4528 } else if (IndirectBrInst *IBI =
4529 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
4530 if (SimplifyIndirectBr(IBI)) return true;
4536 /// SimplifyCFG - This function is used to do simplification of a CFG. For
4537 /// example, it adjusts branches to branches to eliminate the extra hop, it
4538 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4539 /// of the CFG. It returns true if a modification was made.
4541 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
4542 unsigned BonusInstThreshold,
4543 const DataLayout *DL, AssumptionTracker *AT) {
4544 return SimplifyCFGOpt(TTI, BonusInstThreshold, DL, AT).run(BB);