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/SetOperations.h"
18 #include "llvm/ADT/SetVector.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/Analysis/ConstantFolding.h"
23 #include "llvm/Analysis/InstructionSimplify.h"
24 #include "llvm/Analysis/TargetTransformInfo.h"
25 #include "llvm/Analysis/ValueTracking.h"
26 #include "llvm/IR/CFG.h"
27 #include "llvm/IR/ConstantRange.h"
28 #include "llvm/IR/Constants.h"
29 #include "llvm/IR/DataLayout.h"
30 #include "llvm/IR/DerivedTypes.h"
31 #include "llvm/IR/GlobalVariable.h"
32 #include "llvm/IR/IRBuilder.h"
33 #include "llvm/IR/Instructions.h"
34 #include "llvm/IR/IntrinsicInst.h"
35 #include "llvm/IR/LLVMContext.h"
36 #include "llvm/IR/MDBuilder.h"
37 #include "llvm/IR/Metadata.h"
38 #include "llvm/IR/Module.h"
39 #include "llvm/IR/NoFolder.h"
40 #include "llvm/IR/Operator.h"
41 #include "llvm/IR/PatternMatch.h"
42 #include "llvm/IR/Type.h"
43 #include "llvm/Support/CommandLine.h"
44 #include "llvm/Support/Debug.h"
45 #include "llvm/Support/raw_ostream.h"
46 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
47 #include "llvm/Transforms/Utils/ValueMapper.h"
52 using namespace PatternMatch;
54 #define DEBUG_TYPE "simplifycfg"
56 // Chosen as 2 so as to be cheap, but still to have enough power to fold
57 // a select, so the "clamp" idiom (of a min followed by a max) will be caught.
58 // To catch this, we need to fold a compare and a select, hence '2' being the
59 // minimum reasonable default.
60 static cl::opt<unsigned>
61 PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(2),
62 cl::desc("Control the amount of phi node folding to perform (default = 2)"));
65 DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false),
66 cl::desc("Duplicate return instructions into unconditional branches"));
69 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
70 cl::desc("Sink common instructions down to the end block"));
72 static cl::opt<bool> HoistCondStores(
73 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
74 cl::desc("Hoist conditional stores if an unconditional store precedes"));
76 static cl::opt<bool> MergeCondStores(
77 "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
78 cl::desc("Hoist conditional stores even if an unconditional store does not "
79 "precede - hoist multiple conditional stores into a single "
82 static cl::opt<bool> MergeCondStoresAggressively(
83 "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
84 cl::desc("When merging conditional stores, do so even if the resultant "
85 "basic blocks are unlikely to be if-converted as a result"));
87 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
88 STATISTIC(NumLinearMaps, "Number of switch instructions turned into linear mapping");
89 STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables");
90 STATISTIC(NumLookupTablesHoles, "Number of switch instructions turned into lookup tables (holes checked)");
91 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
92 STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block");
93 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
96 // The first field contains the value that the switch produces when a certain
97 // case group is selected, and the second field is a vector containing the
98 // cases composing the case group.
99 typedef SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>
100 SwitchCaseResultVectorTy;
101 // The first field contains the phi node that generates a result of the switch
102 // and the second field contains the value generated for a certain case in the
103 // switch for that PHI.
104 typedef SmallVector<std::pair<PHINode *, Constant *>, 4> SwitchCaseResultsTy;
106 /// ValueEqualityComparisonCase - Represents a case of a switch.
107 struct ValueEqualityComparisonCase {
111 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
112 : Value(Value), Dest(Dest) {}
114 bool operator<(ValueEqualityComparisonCase RHS) const {
115 // Comparing pointers is ok as we only rely on the order for uniquing.
116 return Value < RHS.Value;
119 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
122 class SimplifyCFGOpt {
123 const TargetTransformInfo &TTI;
124 const DataLayout &DL;
125 unsigned BonusInstThreshold;
127 Value *isValueEqualityComparison(TerminatorInst *TI);
128 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
129 std::vector<ValueEqualityComparisonCase> &Cases);
130 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
132 IRBuilder<> &Builder);
133 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
134 IRBuilder<> &Builder);
136 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
137 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
138 bool SimplifyCleanupReturn(CleanupReturnInst *RI);
139 bool SimplifyUnreachable(UnreachableInst *UI);
140 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
141 bool SimplifyIndirectBr(IndirectBrInst *IBI);
142 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder);
143 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder);
146 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
147 unsigned BonusInstThreshold, AssumptionCache *AC)
148 : TTI(TTI), DL(DL), BonusInstThreshold(BonusInstThreshold), AC(AC) {}
149 bool run(BasicBlock *BB);
153 /// Return true if it is safe to merge these two
154 /// terminator instructions together.
155 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
156 if (SI1 == SI2) return false; // Can't merge with self!
158 // It is not safe to merge these two switch instructions if they have a common
159 // successor, and if that successor has a PHI node, and if *that* PHI node has
160 // conflicting incoming values from the two switch blocks.
161 BasicBlock *SI1BB = SI1->getParent();
162 BasicBlock *SI2BB = SI2->getParent();
163 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
165 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
166 if (SI1Succs.count(*I))
167 for (BasicBlock::iterator BBI = (*I)->begin();
168 isa<PHINode>(BBI); ++BBI) {
169 PHINode *PN = cast<PHINode>(BBI);
170 if (PN->getIncomingValueForBlock(SI1BB) !=
171 PN->getIncomingValueForBlock(SI2BB))
178 /// Return true if it is safe and profitable to merge these two terminator
179 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
180 /// store all PHI nodes in common successors.
181 static bool isProfitableToFoldUnconditional(BranchInst *SI1,
184 SmallVectorImpl<PHINode*> &PhiNodes) {
185 if (SI1 == SI2) return false; // Can't merge with self!
186 assert(SI1->isUnconditional() && SI2->isConditional());
188 // We fold the unconditional branch if we can easily update all PHI nodes in
189 // common successors:
190 // 1> We have a constant incoming value for the conditional branch;
191 // 2> We have "Cond" as the incoming value for the unconditional branch;
192 // 3> SI2->getCondition() and Cond have same operands.
193 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
194 if (!Ci2) return false;
195 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
196 Cond->getOperand(1) == Ci2->getOperand(1)) &&
197 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
198 Cond->getOperand(1) == Ci2->getOperand(0)))
201 BasicBlock *SI1BB = SI1->getParent();
202 BasicBlock *SI2BB = SI2->getParent();
203 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
204 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
205 if (SI1Succs.count(*I))
206 for (BasicBlock::iterator BBI = (*I)->begin();
207 isa<PHINode>(BBI); ++BBI) {
208 PHINode *PN = cast<PHINode>(BBI);
209 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
210 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
212 PhiNodes.push_back(PN);
217 /// Update PHI nodes in Succ to indicate that there will now be entries in it
218 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
219 /// will be the same as those coming in from ExistPred, an existing predecessor
221 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
222 BasicBlock *ExistPred) {
223 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
226 for (BasicBlock::iterator I = Succ->begin();
227 (PN = dyn_cast<PHINode>(I)); ++I)
228 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
231 /// Compute an abstract "cost" of speculating the given instruction,
232 /// which is assumed to be safe to speculate. TCC_Free means cheap,
233 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
235 static unsigned ComputeSpeculationCost(const User *I,
236 const TargetTransformInfo &TTI) {
237 assert(isSafeToSpeculativelyExecute(I) &&
238 "Instruction is not safe to speculatively execute!");
239 return TTI.getUserCost(I);
242 /// If we have a merge point of an "if condition" as accepted above,
243 /// return true if the specified value dominates the block. We
244 /// don't handle the true generality of domination here, just a special case
245 /// which works well enough for us.
247 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
248 /// see if V (which must be an instruction) and its recursive operands
249 /// that do not dominate BB have a combined cost lower than CostRemaining and
250 /// are non-trapping. If both are true, the instruction is inserted into the
251 /// set and true is returned.
253 /// The cost for most non-trapping instructions is defined as 1 except for
254 /// Select whose cost is 2.
256 /// After this function returns, CostRemaining is decreased by the cost of
257 /// V plus its non-dominating operands. If that cost is greater than
258 /// CostRemaining, false is returned and CostRemaining is undefined.
259 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
260 SmallPtrSetImpl<Instruction*> *AggressiveInsts,
261 unsigned &CostRemaining,
262 const TargetTransformInfo &TTI) {
263 Instruction *I = dyn_cast<Instruction>(V);
265 // Non-instructions all dominate instructions, but not all constantexprs
266 // can be executed unconditionally.
267 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
272 BasicBlock *PBB = I->getParent();
274 // We don't want to allow weird loops that might have the "if condition" in
275 // the bottom of this block.
276 if (PBB == BB) return false;
278 // If this instruction is defined in a block that contains an unconditional
279 // branch to BB, then it must be in the 'conditional' part of the "if
280 // statement". If not, it definitely dominates the region.
281 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
282 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
285 // If we aren't allowing aggressive promotion anymore, then don't consider
286 // instructions in the 'if region'.
287 if (!AggressiveInsts) return false;
289 // If we have seen this instruction before, don't count it again.
290 if (AggressiveInsts->count(I)) return true;
292 // Okay, it looks like the instruction IS in the "condition". Check to
293 // see if it's a cheap instruction to unconditionally compute, and if it
294 // only uses stuff defined outside of the condition. If so, hoist it out.
295 if (!isSafeToSpeculativelyExecute(I))
298 unsigned Cost = ComputeSpeculationCost(I, TTI);
300 if (Cost > CostRemaining)
303 CostRemaining -= Cost;
305 // Okay, we can only really hoist these out if their operands do
306 // not take us over the cost threshold.
307 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
308 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, TTI))
310 // Okay, it's safe to do this! Remember this instruction.
311 AggressiveInsts->insert(I);
315 /// Extract ConstantInt from value, looking through IntToPtr
316 /// and PointerNullValue. Return NULL if value is not a constant int.
317 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
318 // Normal constant int.
319 ConstantInt *CI = dyn_cast<ConstantInt>(V);
320 if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
323 // This is some kind of pointer constant. Turn it into a pointer-sized
324 // ConstantInt if possible.
325 IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
327 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
328 if (isa<ConstantPointerNull>(V))
329 return ConstantInt::get(PtrTy, 0);
331 // IntToPtr const int.
332 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
333 if (CE->getOpcode() == Instruction::IntToPtr)
334 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
335 // The constant is very likely to have the right type already.
336 if (CI->getType() == PtrTy)
339 return cast<ConstantInt>
340 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
347 /// Given a chain of or (||) or and (&&) comparison of a value against a
348 /// constant, this will try to recover the information required for a switch
350 /// It will depth-first traverse the chain of comparison, seeking for patterns
351 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
352 /// representing the different cases for the switch.
353 /// Note that if the chain is composed of '||' it will build the set of elements
354 /// that matches the comparisons (i.e. any of this value validate the chain)
355 /// while for a chain of '&&' it will build the set elements that make the test
357 struct ConstantComparesGatherer {
358 const DataLayout &DL;
359 Value *CompValue; /// Value found for the switch comparison
360 Value *Extra; /// Extra clause to be checked before the switch
361 SmallVector<ConstantInt *, 8> Vals; /// Set of integers to match in switch
362 unsigned UsedICmps; /// Number of comparisons matched in the and/or chain
364 /// Construct and compute the result for the comparison instruction Cond
365 ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL)
366 : DL(DL), CompValue(nullptr), Extra(nullptr), UsedICmps(0) {
371 ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
372 ConstantComparesGatherer &
373 operator=(const ConstantComparesGatherer &) = delete;
377 /// Try to set the current value used for the comparison, it succeeds only if
378 /// it wasn't set before or if the new value is the same as the old one
379 bool setValueOnce(Value *NewVal) {
380 if(CompValue && CompValue != NewVal) return false;
382 return (CompValue != nullptr);
385 /// Try to match Instruction "I" as a comparison against a constant and
386 /// populates the array Vals with the set of values that match (or do not
387 /// match depending on isEQ).
388 /// Return false on failure. On success, the Value the comparison matched
389 /// against is placed in CompValue.
390 /// If CompValue is already set, the function is expected to fail if a match
391 /// is found but the value compared to is different.
392 bool matchInstruction(Instruction *I, bool isEQ) {
393 // If this is an icmp against a constant, handle this as one of the cases.
396 if (!((ICI = dyn_cast<ICmpInst>(I)) &&
397 (C = GetConstantInt(I->getOperand(1), DL)))) {
404 // Pattern match a special case
405 // (x & ~2^x) == y --> x == y || x == y|2^x
406 // This undoes a transformation done by instcombine to fuse 2 compares.
407 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
408 if (match(ICI->getOperand(0),
409 m_And(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
410 APInt Not = ~RHSC->getValue();
411 if (Not.isPowerOf2()) {
412 // If we already have a value for the switch, it has to match!
413 if(!setValueOnce(RHSVal))
417 Vals.push_back(ConstantInt::get(C->getContext(),
418 C->getValue() | Not));
424 // If we already have a value for the switch, it has to match!
425 if(!setValueOnce(ICI->getOperand(0)))
430 return ICI->getOperand(0);
433 // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
434 ConstantRange Span = ConstantRange::makeAllowedICmpRegion(
435 ICI->getPredicate(), C->getValue());
437 // Shift the range if the compare is fed by an add. This is the range
438 // compare idiom as emitted by instcombine.
439 Value *CandidateVal = I->getOperand(0);
440 if(match(I->getOperand(0), m_Add(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
441 Span = Span.subtract(RHSC->getValue());
442 CandidateVal = RHSVal;
445 // If this is an and/!= check, then we are looking to build the set of
446 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
449 Span = Span.inverse();
451 // If there are a ton of values, we don't want to make a ginormous switch.
452 if (Span.getSetSize().ugt(8) || Span.isEmptySet()) {
456 // If we already have a value for the switch, it has to match!
457 if(!setValueOnce(CandidateVal))
460 // Add all values from the range to the set
461 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
462 Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
469 /// Given a potentially 'or'd or 'and'd together collection of icmp
470 /// eq/ne/lt/gt instructions that compare a value against a constant, extract
471 /// the value being compared, and stick the list constants into the Vals
473 /// One "Extra" case is allowed to differ from the other.
474 void gather(Value *V) {
475 Instruction *I = dyn_cast<Instruction>(V);
476 bool isEQ = (I->getOpcode() == Instruction::Or);
478 // Keep a stack (SmallVector for efficiency) for depth-first traversal
479 SmallVector<Value *, 8> DFT;
484 while(!DFT.empty()) {
485 V = DFT.pop_back_val();
487 if (Instruction *I = dyn_cast<Instruction>(V)) {
488 // If it is a || (or && depending on isEQ), process the operands.
489 if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
490 DFT.push_back(I->getOperand(1));
491 DFT.push_back(I->getOperand(0));
495 // Try to match the current instruction
496 if (matchInstruction(I, isEQ))
497 // Match succeed, continue the loop
501 // One element of the sequence of || (or &&) could not be match as a
502 // comparison against the same value as the others.
503 // We allow only one "Extra" case to be checked before the switch
508 // Failed to parse a proper sequence, abort now
517 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
518 Instruction *Cond = nullptr;
519 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
520 Cond = dyn_cast<Instruction>(SI->getCondition());
521 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
522 if (BI->isConditional())
523 Cond = dyn_cast<Instruction>(BI->getCondition());
524 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
525 Cond = dyn_cast<Instruction>(IBI->getAddress());
528 TI->eraseFromParent();
529 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
532 /// Return true if the specified terminator checks
533 /// to see if a value is equal to constant integer value.
534 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
536 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
537 // Do not permit merging of large switch instructions into their
538 // predecessors unless there is only one predecessor.
539 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()),
540 pred_end(SI->getParent())) <= 128)
541 CV = SI->getCondition();
542 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
543 if (BI->isConditional() && BI->getCondition()->hasOneUse())
544 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
545 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
546 CV = ICI->getOperand(0);
549 // Unwrap any lossless ptrtoint cast.
551 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
552 Value *Ptr = PTII->getPointerOperand();
553 if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
560 /// Given a value comparison instruction,
561 /// decode all of the 'cases' that it represents and return the 'default' block.
562 BasicBlock *SimplifyCFGOpt::
563 GetValueEqualityComparisonCases(TerminatorInst *TI,
564 std::vector<ValueEqualityComparisonCase>
566 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
567 Cases.reserve(SI->getNumCases());
568 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
569 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
570 i.getCaseSuccessor()));
571 return SI->getDefaultDest();
574 BranchInst *BI = cast<BranchInst>(TI);
575 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
576 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
577 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
580 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
584 /// Given a vector of bb/value pairs, remove any entries
585 /// in the list that match the specified block.
586 static void EliminateBlockCases(BasicBlock *BB,
587 std::vector<ValueEqualityComparisonCase> &Cases) {
588 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
591 /// Return true if there are any keys in C1 that exist in C2 as well.
593 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
594 std::vector<ValueEqualityComparisonCase > &C2) {
595 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
597 // Make V1 be smaller than V2.
598 if (V1->size() > V2->size())
601 if (V1->size() == 0) return false;
602 if (V1->size() == 1) {
604 ConstantInt *TheVal = (*V1)[0].Value;
605 for (unsigned i = 0, e = V2->size(); i != e; ++i)
606 if (TheVal == (*V2)[i].Value)
610 // Otherwise, just sort both lists and compare element by element.
611 array_pod_sort(V1->begin(), V1->end());
612 array_pod_sort(V2->begin(), V2->end());
613 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
614 while (i1 != e1 && i2 != e2) {
615 if ((*V1)[i1].Value == (*V2)[i2].Value)
617 if ((*V1)[i1].Value < (*V2)[i2].Value)
625 /// If TI is known to be a terminator instruction and its block is known to
626 /// only have a single predecessor block, check to see if that predecessor is
627 /// also a value comparison with the same value, and if that comparison
628 /// determines the outcome of this comparison. If so, simplify TI. This does a
629 /// very limited form of jump threading.
630 bool SimplifyCFGOpt::
631 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
633 IRBuilder<> &Builder) {
634 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
635 if (!PredVal) return false; // Not a value comparison in predecessor.
637 Value *ThisVal = isValueEqualityComparison(TI);
638 assert(ThisVal && "This isn't a value comparison!!");
639 if (ThisVal != PredVal) return false; // Different predicates.
641 // TODO: Preserve branch weight metadata, similarly to how
642 // FoldValueComparisonIntoPredecessors preserves it.
644 // Find out information about when control will move from Pred to TI's block.
645 std::vector<ValueEqualityComparisonCase> PredCases;
646 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
648 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
650 // Find information about how control leaves this block.
651 std::vector<ValueEqualityComparisonCase> ThisCases;
652 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
653 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
655 // If TI's block is the default block from Pred's comparison, potentially
656 // simplify TI based on this knowledge.
657 if (PredDef == TI->getParent()) {
658 // If we are here, we know that the value is none of those cases listed in
659 // PredCases. If there are any cases in ThisCases that are in PredCases, we
661 if (!ValuesOverlap(PredCases, ThisCases))
664 if (isa<BranchInst>(TI)) {
665 // Okay, one of the successors of this condbr is dead. Convert it to a
667 assert(ThisCases.size() == 1 && "Branch can only have one case!");
668 // Insert the new branch.
669 Instruction *NI = Builder.CreateBr(ThisDef);
672 // Remove PHI node entries for the dead edge.
673 ThisCases[0].Dest->removePredecessor(TI->getParent());
675 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
676 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
678 EraseTerminatorInstAndDCECond(TI);
682 SwitchInst *SI = cast<SwitchInst>(TI);
683 // Okay, TI has cases that are statically dead, prune them away.
684 SmallPtrSet<Constant*, 16> DeadCases;
685 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
686 DeadCases.insert(PredCases[i].Value);
688 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
689 << "Through successor TI: " << *TI);
691 // Collect branch weights into a vector.
692 SmallVector<uint32_t, 8> Weights;
693 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
694 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
696 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
698 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
699 Weights.push_back(CI->getValue().getZExtValue());
701 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
703 if (DeadCases.count(i.getCaseValue())) {
705 std::swap(Weights[i.getCaseIndex()+1], Weights.back());
708 i.getCaseSuccessor()->removePredecessor(TI->getParent());
712 if (HasWeight && Weights.size() >= 2)
713 SI->setMetadata(LLVMContext::MD_prof,
714 MDBuilder(SI->getParent()->getContext()).
715 createBranchWeights(Weights));
717 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
721 // Otherwise, TI's block must correspond to some matched value. Find out
722 // which value (or set of values) this is.
723 ConstantInt *TIV = nullptr;
724 BasicBlock *TIBB = TI->getParent();
725 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
726 if (PredCases[i].Dest == TIBB) {
728 return false; // Cannot handle multiple values coming to this block.
729 TIV = PredCases[i].Value;
731 assert(TIV && "No edge from pred to succ?");
733 // Okay, we found the one constant that our value can be if we get into TI's
734 // BB. Find out which successor will unconditionally be branched to.
735 BasicBlock *TheRealDest = nullptr;
736 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
737 if (ThisCases[i].Value == TIV) {
738 TheRealDest = ThisCases[i].Dest;
742 // If not handled by any explicit cases, it is handled by the default case.
743 if (!TheRealDest) TheRealDest = ThisDef;
745 // Remove PHI node entries for dead edges.
746 BasicBlock *CheckEdge = TheRealDest;
747 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
748 if (*SI != CheckEdge)
749 (*SI)->removePredecessor(TIBB);
753 // Insert the new branch.
754 Instruction *NI = Builder.CreateBr(TheRealDest);
757 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
758 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
760 EraseTerminatorInstAndDCECond(TI);
765 /// This class implements a stable ordering of constant
766 /// integers that does not depend on their address. This is important for
767 /// applications that sort ConstantInt's to ensure uniqueness.
768 struct ConstantIntOrdering {
769 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
770 return LHS->getValue().ult(RHS->getValue());
775 static int ConstantIntSortPredicate(ConstantInt *const *P1,
776 ConstantInt *const *P2) {
777 const ConstantInt *LHS = *P1;
778 const ConstantInt *RHS = *P2;
779 if (LHS->getValue().ult(RHS->getValue()))
781 if (LHS->getValue() == RHS->getValue())
786 static inline bool HasBranchWeights(const Instruction* I) {
787 MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
788 if (ProfMD && ProfMD->getOperand(0))
789 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
790 return MDS->getString().equals("branch_weights");
795 /// Get Weights of a given TerminatorInst, the default weight is at the front
796 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
798 static void GetBranchWeights(TerminatorInst *TI,
799 SmallVectorImpl<uint64_t> &Weights) {
800 MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
802 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
803 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
804 Weights.push_back(CI->getValue().getZExtValue());
807 // If TI is a conditional eq, the default case is the false case,
808 // and the corresponding branch-weight data is at index 2. We swap the
809 // default weight to be the first entry.
810 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
811 assert(Weights.size() == 2);
812 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
813 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
814 std::swap(Weights.front(), Weights.back());
818 /// Keep halving the weights until all can fit in uint32_t.
819 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
820 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
821 if (Max > UINT_MAX) {
822 unsigned Offset = 32 - countLeadingZeros(Max);
823 for (uint64_t &I : Weights)
828 /// The specified terminator is a value equality comparison instruction
829 /// (either a switch or a branch on "X == c").
830 /// See if any of the predecessors of the terminator block are value comparisons
831 /// on the same value. If so, and if safe to do so, fold them together.
832 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
833 IRBuilder<> &Builder) {
834 BasicBlock *BB = TI->getParent();
835 Value *CV = isValueEqualityComparison(TI); // CondVal
836 assert(CV && "Not a comparison?");
837 bool Changed = false;
839 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
840 while (!Preds.empty()) {
841 BasicBlock *Pred = Preds.pop_back_val();
843 // See if the predecessor is a comparison with the same value.
844 TerminatorInst *PTI = Pred->getTerminator();
845 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
847 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
848 // Figure out which 'cases' to copy from SI to PSI.
849 std::vector<ValueEqualityComparisonCase> BBCases;
850 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
852 std::vector<ValueEqualityComparisonCase> PredCases;
853 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
855 // Based on whether the default edge from PTI goes to BB or not, fill in
856 // PredCases and PredDefault with the new switch cases we would like to
858 SmallVector<BasicBlock*, 8> NewSuccessors;
860 // Update the branch weight metadata along the way
861 SmallVector<uint64_t, 8> Weights;
862 bool PredHasWeights = HasBranchWeights(PTI);
863 bool SuccHasWeights = HasBranchWeights(TI);
865 if (PredHasWeights) {
866 GetBranchWeights(PTI, Weights);
867 // branch-weight metadata is inconsistent here.
868 if (Weights.size() != 1 + PredCases.size())
869 PredHasWeights = SuccHasWeights = false;
870 } else if (SuccHasWeights)
871 // If there are no predecessor weights but there are successor weights,
872 // populate Weights with 1, which will later be scaled to the sum of
873 // successor's weights
874 Weights.assign(1 + PredCases.size(), 1);
876 SmallVector<uint64_t, 8> SuccWeights;
877 if (SuccHasWeights) {
878 GetBranchWeights(TI, SuccWeights);
879 // branch-weight metadata is inconsistent here.
880 if (SuccWeights.size() != 1 + BBCases.size())
881 PredHasWeights = SuccHasWeights = false;
882 } else if (PredHasWeights)
883 SuccWeights.assign(1 + BBCases.size(), 1);
885 if (PredDefault == BB) {
886 // If this is the default destination from PTI, only the edges in TI
887 // that don't occur in PTI, or that branch to BB will be activated.
888 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
889 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
890 if (PredCases[i].Dest != BB)
891 PTIHandled.insert(PredCases[i].Value);
893 // The default destination is BB, we don't need explicit targets.
894 std::swap(PredCases[i], PredCases.back());
896 if (PredHasWeights || SuccHasWeights) {
897 // Increase weight for the default case.
898 Weights[0] += Weights[i+1];
899 std::swap(Weights[i+1], Weights.back());
903 PredCases.pop_back();
907 // Reconstruct the new switch statement we will be building.
908 if (PredDefault != BBDefault) {
909 PredDefault->removePredecessor(Pred);
910 PredDefault = BBDefault;
911 NewSuccessors.push_back(BBDefault);
914 unsigned CasesFromPred = Weights.size();
915 uint64_t ValidTotalSuccWeight = 0;
916 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
917 if (!PTIHandled.count(BBCases[i].Value) &&
918 BBCases[i].Dest != BBDefault) {
919 PredCases.push_back(BBCases[i]);
920 NewSuccessors.push_back(BBCases[i].Dest);
921 if (SuccHasWeights || PredHasWeights) {
922 // The default weight is at index 0, so weight for the ith case
923 // should be at index i+1. Scale the cases from successor by
924 // PredDefaultWeight (Weights[0]).
925 Weights.push_back(Weights[0] * SuccWeights[i+1]);
926 ValidTotalSuccWeight += SuccWeights[i+1];
930 if (SuccHasWeights || PredHasWeights) {
931 ValidTotalSuccWeight += SuccWeights[0];
932 // Scale the cases from predecessor by ValidTotalSuccWeight.
933 for (unsigned i = 1; i < CasesFromPred; ++i)
934 Weights[i] *= ValidTotalSuccWeight;
935 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
936 Weights[0] *= SuccWeights[0];
939 // If this is not the default destination from PSI, only the edges
940 // in SI that occur in PSI with a destination of BB will be
942 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
943 std::map<ConstantInt*, uint64_t> WeightsForHandled;
944 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
945 if (PredCases[i].Dest == BB) {
946 PTIHandled.insert(PredCases[i].Value);
948 if (PredHasWeights || SuccHasWeights) {
949 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
950 std::swap(Weights[i+1], Weights.back());
954 std::swap(PredCases[i], PredCases.back());
955 PredCases.pop_back();
959 // Okay, now we know which constants were sent to BB from the
960 // predecessor. Figure out where they will all go now.
961 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
962 if (PTIHandled.count(BBCases[i].Value)) {
963 // If this is one we are capable of getting...
964 if (PredHasWeights || SuccHasWeights)
965 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
966 PredCases.push_back(BBCases[i]);
967 NewSuccessors.push_back(BBCases[i].Dest);
968 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
971 // If there are any constants vectored to BB that TI doesn't handle,
972 // they must go to the default destination of TI.
973 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
975 E = PTIHandled.end(); I != E; ++I) {
976 if (PredHasWeights || SuccHasWeights)
977 Weights.push_back(WeightsForHandled[*I]);
978 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
979 NewSuccessors.push_back(BBDefault);
983 // Okay, at this point, we know which new successor Pred will get. Make
984 // sure we update the number of entries in the PHI nodes for these
986 for (BasicBlock *NewSuccessor : NewSuccessors)
987 AddPredecessorToBlock(NewSuccessor, Pred, BB);
989 Builder.SetInsertPoint(PTI);
990 // Convert pointer to int before we switch.
991 if (CV->getType()->isPointerTy()) {
992 CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
996 // Now that the successors are updated, create the new Switch instruction.
997 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
999 NewSI->setDebugLoc(PTI->getDebugLoc());
1000 for (ValueEqualityComparisonCase &V : PredCases)
1001 NewSI->addCase(V.Value, V.Dest);
1003 if (PredHasWeights || SuccHasWeights) {
1004 // Halve the weights if any of them cannot fit in an uint32_t
1005 FitWeights(Weights);
1007 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1009 NewSI->setMetadata(LLVMContext::MD_prof,
1010 MDBuilder(BB->getContext()).
1011 createBranchWeights(MDWeights));
1014 EraseTerminatorInstAndDCECond(PTI);
1016 // Okay, last check. If BB is still a successor of PSI, then we must
1017 // have an infinite loop case. If so, add an infinitely looping block
1018 // to handle the case to preserve the behavior of the code.
1019 BasicBlock *InfLoopBlock = nullptr;
1020 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1021 if (NewSI->getSuccessor(i) == BB) {
1022 if (!InfLoopBlock) {
1023 // Insert it at the end of the function, because it's either code,
1024 // or it won't matter if it's hot. :)
1025 InfLoopBlock = BasicBlock::Create(BB->getContext(),
1026 "infloop", BB->getParent());
1027 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1029 NewSI->setSuccessor(i, InfLoopBlock);
1038 // If we would need to insert a select that uses the value of this invoke
1039 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1040 // can't hoist the invoke, as there is nowhere to put the select in this case.
1041 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1042 Instruction *I1, Instruction *I2) {
1043 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1045 for (BasicBlock::iterator BBI = SI->begin();
1046 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1047 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1048 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1049 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
1057 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I);
1059 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1060 /// in the two blocks up into the branch block. The caller of this function
1061 /// guarantees that BI's block dominates BB1 and BB2.
1062 static bool HoistThenElseCodeToIf(BranchInst *BI,
1063 const TargetTransformInfo &TTI) {
1064 // This does very trivial matching, with limited scanning, to find identical
1065 // instructions in the two blocks. In particular, we don't want to get into
1066 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1067 // such, we currently just scan for obviously identical instructions in an
1069 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1070 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1072 BasicBlock::iterator BB1_Itr = BB1->begin();
1073 BasicBlock::iterator BB2_Itr = BB2->begin();
1075 Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1076 // Skip debug info if it is not identical.
1077 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1078 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1079 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1080 while (isa<DbgInfoIntrinsic>(I1))
1082 while (isa<DbgInfoIntrinsic>(I2))
1085 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1086 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1089 BasicBlock *BIParent = BI->getParent();
1091 bool Changed = false;
1093 // If we are hoisting the terminator instruction, don't move one (making a
1094 // broken BB), instead clone it, and remove BI.
1095 if (isa<TerminatorInst>(I1))
1096 goto HoistTerminator;
1098 if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1101 // For a normal instruction, we just move one to right before the branch,
1102 // then replace all uses of the other with the first. Finally, we remove
1103 // the now redundant second instruction.
1104 BIParent->getInstList().splice(BI->getIterator(), BB1->getInstList(), I1);
1105 if (!I2->use_empty())
1106 I2->replaceAllUsesWith(I1);
1107 I1->intersectOptionalDataWith(I2);
1108 unsigned KnownIDs[] = {
1109 LLVMContext::MD_tbaa, LLVMContext::MD_range,
1110 LLVMContext::MD_fpmath, LLVMContext::MD_invariant_load,
1111 LLVMContext::MD_nonnull, LLVMContext::MD_invariant_group,
1112 LLVMContext::MD_align, LLVMContext::MD_dereferenceable,
1113 LLVMContext::MD_dereferenceable_or_null};
1114 combineMetadata(I1, I2, KnownIDs);
1115 I2->eraseFromParent();
1120 // Skip debug info if it is not identical.
1121 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1122 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1123 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1124 while (isa<DbgInfoIntrinsic>(I1))
1126 while (isa<DbgInfoIntrinsic>(I2))
1129 } while (I1->isIdenticalToWhenDefined(I2));
1134 // It may not be possible to hoist an invoke.
1135 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1138 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1140 for (BasicBlock::iterator BBI = SI->begin();
1141 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1142 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1143 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1147 // Check for passingValueIsAlwaysUndefined here because we would rather
1148 // eliminate undefined control flow then converting it to a select.
1149 if (passingValueIsAlwaysUndefined(BB1V, PN) ||
1150 passingValueIsAlwaysUndefined(BB2V, PN))
1153 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1155 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1160 // Okay, it is safe to hoist the terminator.
1161 Instruction *NT = I1->clone();
1162 BIParent->getInstList().insert(BI->getIterator(), NT);
1163 if (!NT->getType()->isVoidTy()) {
1164 I1->replaceAllUsesWith(NT);
1165 I2->replaceAllUsesWith(NT);
1169 IRBuilder<true, NoFolder> Builder(NT);
1170 // Hoisting one of the terminators from our successor is a great thing.
1171 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1172 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1173 // nodes, so we insert select instruction to compute the final result.
1174 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1175 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1177 for (BasicBlock::iterator BBI = SI->begin();
1178 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1179 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1180 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1181 if (BB1V == BB2V) continue;
1183 // These values do not agree. Insert a select instruction before NT
1184 // that determines the right value.
1185 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1187 SI = cast<SelectInst>
1188 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1189 BB1V->getName()+"."+BB2V->getName()));
1191 // Make the PHI node use the select for all incoming values for BB1/BB2
1192 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1193 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1194 PN->setIncomingValue(i, SI);
1198 // Update any PHI nodes in our new successors.
1199 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1200 AddPredecessorToBlock(*SI, BIParent, BB1);
1202 EraseTerminatorInstAndDCECond(BI);
1206 /// Given an unconditional branch that goes to BBEnd,
1207 /// check whether BBEnd has only two predecessors and the other predecessor
1208 /// ends with an unconditional branch. If it is true, sink any common code
1209 /// in the two predecessors to BBEnd.
1210 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1211 assert(BI1->isUnconditional());
1212 BasicBlock *BB1 = BI1->getParent();
1213 BasicBlock *BBEnd = BI1->getSuccessor(0);
1215 // Check that BBEnd has two predecessors and the other predecessor ends with
1216 // an unconditional branch.
1217 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1218 BasicBlock *Pred0 = *PI++;
1219 if (PI == PE) // Only one predecessor.
1221 BasicBlock *Pred1 = *PI++;
1222 if (PI != PE) // More than two predecessors.
1224 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1225 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1226 if (!BI2 || !BI2->isUnconditional())
1229 // Gather the PHI nodes in BBEnd.
1230 SmallDenseMap<std::pair<Value *, Value *>, PHINode *> JointValueMap;
1231 Instruction *FirstNonPhiInBBEnd = nullptr;
1232 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end(); I != E; ++I) {
1233 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1234 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1235 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1236 JointValueMap[std::make_pair(BB1V, BB2V)] = PN;
1238 FirstNonPhiInBBEnd = &*I;
1242 if (!FirstNonPhiInBBEnd)
1245 // This does very trivial matching, with limited scanning, to find identical
1246 // instructions in the two blocks. We scan backward for obviously identical
1247 // instructions in an identical order.
1248 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1249 RE1 = BB1->getInstList().rend(),
1250 RI2 = BB2->getInstList().rbegin(),
1251 RE2 = BB2->getInstList().rend();
1253 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1256 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1259 // Skip the unconditional branches.
1263 bool Changed = false;
1264 while (RI1 != RE1 && RI2 != RE2) {
1266 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1269 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1273 Instruction *I1 = &*RI1, *I2 = &*RI2;
1274 auto InstPair = std::make_pair(I1, I2);
1275 // I1 and I2 should have a single use in the same PHI node, and they
1276 // perform the same operation.
1277 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1278 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1279 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1280 I1->isEHPad() || I2->isEHPad() ||
1281 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1282 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1283 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1284 !I1->hasOneUse() || !I2->hasOneUse() ||
1285 !JointValueMap.count(InstPair))
1288 // Check whether we should swap the operands of ICmpInst.
1289 // TODO: Add support of communativity.
1290 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1291 bool SwapOpnds = false;
1292 if (ICmp1 && ICmp2 &&
1293 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1294 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1295 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1296 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1297 ICmp2->swapOperands();
1300 if (!I1->isSameOperationAs(I2)) {
1302 ICmp2->swapOperands();
1306 // The operands should be either the same or they need to be generated
1307 // with a PHI node after sinking. We only handle the case where there is
1308 // a single pair of different operands.
1309 Value *DifferentOp1 = nullptr, *DifferentOp2 = nullptr;
1310 unsigned Op1Idx = ~0U;
1311 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1312 if (I1->getOperand(I) == I2->getOperand(I))
1314 // Early exit if we have more-than one pair of different operands or if
1315 // we need a PHI node to replace a constant.
1316 if (Op1Idx != ~0U ||
1317 isa<Constant>(I1->getOperand(I)) ||
1318 isa<Constant>(I2->getOperand(I))) {
1319 // If we can't sink the instructions, undo the swapping.
1321 ICmp2->swapOperands();
1324 DifferentOp1 = I1->getOperand(I);
1326 DifferentOp2 = I2->getOperand(I);
1329 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n");
1330 DEBUG(dbgs() << " " << *I2 << "\n");
1332 // We insert the pair of different operands to JointValueMap and
1333 // remove (I1, I2) from JointValueMap.
1334 if (Op1Idx != ~0U) {
1335 auto &NewPN = JointValueMap[std::make_pair(DifferentOp1, DifferentOp2)];
1338 PHINode::Create(DifferentOp1->getType(), 2,
1339 DifferentOp1->getName() + ".sink", &BBEnd->front());
1340 NewPN->addIncoming(DifferentOp1, BB1);
1341 NewPN->addIncoming(DifferentOp2, BB2);
1342 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1344 // I1 should use NewPN instead of DifferentOp1.
1345 I1->setOperand(Op1Idx, NewPN);
1347 PHINode *OldPN = JointValueMap[InstPair];
1348 JointValueMap.erase(InstPair);
1350 // We need to update RE1 and RE2 if we are going to sink the first
1351 // instruction in the basic block down.
1352 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1353 // Sink the instruction.
1354 BBEnd->getInstList().splice(FirstNonPhiInBBEnd->getIterator(),
1355 BB1->getInstList(), I1);
1356 if (!OldPN->use_empty())
1357 OldPN->replaceAllUsesWith(I1);
1358 OldPN->eraseFromParent();
1360 if (!I2->use_empty())
1361 I2->replaceAllUsesWith(I1);
1362 I1->intersectOptionalDataWith(I2);
1363 // TODO: Use combineMetadata here to preserve what metadata we can
1364 // (analogous to the hoisting case above).
1365 I2->eraseFromParent();
1368 RE1 = BB1->getInstList().rend();
1370 RE2 = BB2->getInstList().rend();
1371 FirstNonPhiInBBEnd = &*I1;
1378 /// \brief Determine if we can hoist sink a sole store instruction out of a
1379 /// conditional block.
1381 /// We are looking for code like the following:
1383 /// store i32 %add, i32* %arrayidx2
1384 /// ... // No other stores or function calls (we could be calling a memory
1385 /// ... // function).
1386 /// %cmp = icmp ult %x, %y
1387 /// br i1 %cmp, label %EndBB, label %ThenBB
1389 /// store i32 %add5, i32* %arrayidx2
1393 /// We are going to transform this into:
1395 /// store i32 %add, i32* %arrayidx2
1397 /// %cmp = icmp ult %x, %y
1398 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1399 /// store i32 %add.add5, i32* %arrayidx2
1402 /// \return The pointer to the value of the previous store if the store can be
1403 /// hoisted into the predecessor block. 0 otherwise.
1404 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1405 BasicBlock *StoreBB, BasicBlock *EndBB) {
1406 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1410 // Volatile or atomic.
1411 if (!StoreToHoist->isSimple())
1414 Value *StorePtr = StoreToHoist->getPointerOperand();
1416 // Look for a store to the same pointer in BrBB.
1417 unsigned MaxNumInstToLookAt = 10;
1418 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(),
1419 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) {
1420 Instruction *CurI = &*RI;
1422 // Could be calling an instruction that effects memory like free().
1423 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI))
1426 StoreInst *SI = dyn_cast<StoreInst>(CurI);
1427 // Found the previous store make sure it stores to the same location.
1428 if (SI && SI->getPointerOperand() == StorePtr)
1429 // Found the previous store, return its value operand.
1430 return SI->getValueOperand();
1432 return nullptr; // Unknown store.
1438 /// \brief Speculate a conditional basic block flattening the CFG.
1440 /// Note that this is a very risky transform currently. Speculating
1441 /// instructions like this is most often not desirable. Instead, there is an MI
1442 /// pass which can do it with full awareness of the resource constraints.
1443 /// However, some cases are "obvious" and we should do directly. An example of
1444 /// this is speculating a single, reasonably cheap instruction.
1446 /// There is only one distinct advantage to flattening the CFG at the IR level:
1447 /// it makes very common but simplistic optimizations such as are common in
1448 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1449 /// modeling their effects with easier to reason about SSA value graphs.
1452 /// An illustration of this transform is turning this IR:
1455 /// %cmp = icmp ult %x, %y
1456 /// br i1 %cmp, label %EndBB, label %ThenBB
1458 /// %sub = sub %x, %y
1461 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1468 /// %cmp = icmp ult %x, %y
1469 /// %sub = sub %x, %y
1470 /// %cond = select i1 %cmp, 0, %sub
1474 /// \returns true if the conditional block is removed.
1475 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1476 const TargetTransformInfo &TTI) {
1477 // Be conservative for now. FP select instruction can often be expensive.
1478 Value *BrCond = BI->getCondition();
1479 if (isa<FCmpInst>(BrCond))
1482 BasicBlock *BB = BI->getParent();
1483 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1485 // If ThenBB is actually on the false edge of the conditional branch, remember
1486 // to swap the select operands later.
1487 bool Invert = false;
1488 if (ThenBB != BI->getSuccessor(0)) {
1489 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1492 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1494 // Keep a count of how many times instructions are used within CondBB when
1495 // they are candidates for sinking into CondBB. Specifically:
1496 // - They are defined in BB, and
1497 // - They have no side effects, and
1498 // - All of their uses are in CondBB.
1499 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1501 unsigned SpeculationCost = 0;
1502 Value *SpeculatedStoreValue = nullptr;
1503 StoreInst *SpeculatedStore = nullptr;
1504 for (BasicBlock::iterator BBI = ThenBB->begin(),
1505 BBE = std::prev(ThenBB->end());
1506 BBI != BBE; ++BBI) {
1507 Instruction *I = &*BBI;
1509 if (isa<DbgInfoIntrinsic>(I))
1512 // Only speculatively execute a single instruction (not counting the
1513 // terminator) for now.
1515 if (SpeculationCost > 1)
1518 // Don't hoist the instruction if it's unsafe or expensive.
1519 if (!isSafeToSpeculativelyExecute(I) &&
1520 !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1521 I, BB, ThenBB, EndBB))))
1523 if (!SpeculatedStoreValue &&
1524 ComputeSpeculationCost(I, TTI) >
1525 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
1528 // Store the store speculation candidate.
1529 if (SpeculatedStoreValue)
1530 SpeculatedStore = cast<StoreInst>(I);
1532 // Do not hoist the instruction if any of its operands are defined but not
1533 // used in BB. The transformation will prevent the operand from
1534 // being sunk into the use block.
1535 for (User::op_iterator i = I->op_begin(), e = I->op_end();
1537 Instruction *OpI = dyn_cast<Instruction>(*i);
1538 if (!OpI || OpI->getParent() != BB ||
1539 OpI->mayHaveSideEffects())
1540 continue; // Not a candidate for sinking.
1542 ++SinkCandidateUseCounts[OpI];
1546 // Consider any sink candidates which are only used in CondBB as costs for
1547 // speculation. Note, while we iterate over a DenseMap here, we are summing
1548 // and so iteration order isn't significant.
1549 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I =
1550 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end();
1552 if (I->first->getNumUses() == I->second) {
1554 if (SpeculationCost > 1)
1558 // Check that the PHI nodes can be converted to selects.
1559 bool HaveRewritablePHIs = false;
1560 for (BasicBlock::iterator I = EndBB->begin();
1561 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1562 Value *OrigV = PN->getIncomingValueForBlock(BB);
1563 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1565 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1566 // Skip PHIs which are trivial.
1570 // Don't convert to selects if we could remove undefined behavior instead.
1571 if (passingValueIsAlwaysUndefined(OrigV, PN) ||
1572 passingValueIsAlwaysUndefined(ThenV, PN))
1575 HaveRewritablePHIs = true;
1576 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1577 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1578 if (!OrigCE && !ThenCE)
1579 continue; // Known safe and cheap.
1581 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
1582 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
1584 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
1585 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
1586 unsigned MaxCost = 2 * PHINodeFoldingThreshold *
1587 TargetTransformInfo::TCC_Basic;
1588 if (OrigCost + ThenCost > MaxCost)
1591 // Account for the cost of an unfolded ConstantExpr which could end up
1592 // getting expanded into Instructions.
1593 // FIXME: This doesn't account for how many operations are combined in the
1594 // constant expression.
1596 if (SpeculationCost > 1)
1600 // If there are no PHIs to process, bail early. This helps ensure idempotence
1602 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
1605 // If we get here, we can hoist the instruction and if-convert.
1606 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1608 // Insert a select of the value of the speculated store.
1609 if (SpeculatedStoreValue) {
1610 IRBuilder<true, NoFolder> Builder(BI);
1611 Value *TrueV = SpeculatedStore->getValueOperand();
1612 Value *FalseV = SpeculatedStoreValue;
1614 std::swap(TrueV, FalseV);
1615 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() +
1616 "." + FalseV->getName());
1617 SpeculatedStore->setOperand(0, S);
1620 // Metadata can be dependent on the condition we are hoisting above.
1621 // Conservatively strip all metadata on the instruction.
1622 for (auto &I: *ThenBB)
1623 I.dropUnknownNonDebugMetadata();
1625 // Hoist the instructions.
1626 BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
1627 ThenBB->begin(), std::prev(ThenBB->end()));
1629 // Insert selects and rewrite the PHI operands.
1630 IRBuilder<true, NoFolder> Builder(BI);
1631 for (BasicBlock::iterator I = EndBB->begin();
1632 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1633 unsigned OrigI = PN->getBasicBlockIndex(BB);
1634 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1635 Value *OrigV = PN->getIncomingValue(OrigI);
1636 Value *ThenV = PN->getIncomingValue(ThenI);
1638 // Skip PHIs which are trivial.
1642 // Create a select whose true value is the speculatively executed value and
1643 // false value is the preexisting value. Swap them if the branch
1644 // destinations were inverted.
1645 Value *TrueV = ThenV, *FalseV = OrigV;
1647 std::swap(TrueV, FalseV);
1648 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV,
1649 TrueV->getName() + "." + FalseV->getName());
1650 PN->setIncomingValue(OrigI, V);
1651 PN->setIncomingValue(ThenI, V);
1658 /// \returns True if this block contains a CallInst with the NoDuplicate
1660 static bool HasNoDuplicateCall(const BasicBlock *BB) {
1661 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1662 const CallInst *CI = dyn_cast<CallInst>(I);
1665 if (CI->cannotDuplicate())
1671 /// Return true if we can thread a branch across this block.
1672 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1673 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1676 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1677 if (isa<DbgInfoIntrinsic>(BBI))
1679 if (Size > 10) return false; // Don't clone large BB's.
1682 // We can only support instructions that do not define values that are
1683 // live outside of the current basic block.
1684 for (User *U : BBI->users()) {
1685 Instruction *UI = cast<Instruction>(U);
1686 if (UI->getParent() != BB || isa<PHINode>(UI)) return false;
1689 // Looks ok, continue checking.
1695 /// If we have a conditional branch on a PHI node value that is defined in the
1696 /// same block as the branch and if any PHI entries are constants, thread edges
1697 /// corresponding to that entry to be branches to their ultimate destination.
1698 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL) {
1699 BasicBlock *BB = BI->getParent();
1700 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1701 // NOTE: we currently cannot transform this case if the PHI node is used
1702 // outside of the block.
1703 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1706 // Degenerate case of a single entry PHI.
1707 if (PN->getNumIncomingValues() == 1) {
1708 FoldSingleEntryPHINodes(PN->getParent());
1712 // Now we know that this block has multiple preds and two succs.
1713 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1715 if (HasNoDuplicateCall(BB)) return false;
1717 // Okay, this is a simple enough basic block. See if any phi values are
1719 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1720 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1721 if (!CB || !CB->getType()->isIntegerTy(1)) continue;
1723 // Okay, we now know that all edges from PredBB should be revectored to
1724 // branch to RealDest.
1725 BasicBlock *PredBB = PN->getIncomingBlock(i);
1726 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1728 if (RealDest == BB) continue; // Skip self loops.
1729 // Skip if the predecessor's terminator is an indirect branch.
1730 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1732 // The dest block might have PHI nodes, other predecessors and other
1733 // difficult cases. Instead of being smart about this, just insert a new
1734 // block that jumps to the destination block, effectively splitting
1735 // the edge we are about to create.
1736 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1737 RealDest->getName()+".critedge",
1738 RealDest->getParent(), RealDest);
1739 BranchInst::Create(RealDest, EdgeBB);
1741 // Update PHI nodes.
1742 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1744 // BB may have instructions that are being threaded over. Clone these
1745 // instructions into EdgeBB. We know that there will be no uses of the
1746 // cloned instructions outside of EdgeBB.
1747 BasicBlock::iterator InsertPt = EdgeBB->begin();
1748 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1749 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1750 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1751 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1754 // Clone the instruction.
1755 Instruction *N = BBI->clone();
1756 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1758 // Update operands due to translation.
1759 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1761 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1762 if (PI != TranslateMap.end())
1766 // Check for trivial simplification.
1767 if (Value *V = SimplifyInstruction(N, DL)) {
1768 TranslateMap[&*BBI] = V;
1769 delete N; // Instruction folded away, don't need actual inst
1771 // Insert the new instruction into its new home.
1772 EdgeBB->getInstList().insert(InsertPt, N);
1773 if (!BBI->use_empty())
1774 TranslateMap[&*BBI] = N;
1778 // Loop over all of the edges from PredBB to BB, changing them to branch
1779 // to EdgeBB instead.
1780 TerminatorInst *PredBBTI = PredBB->getTerminator();
1781 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1782 if (PredBBTI->getSuccessor(i) == BB) {
1783 BB->removePredecessor(PredBB);
1784 PredBBTI->setSuccessor(i, EdgeBB);
1787 // Recurse, simplifying any other constants.
1788 return FoldCondBranchOnPHI(BI, DL) | true;
1794 /// Given a BB that starts with the specified two-entry PHI node,
1795 /// see if we can eliminate it.
1796 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI,
1797 const DataLayout &DL) {
1798 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1799 // statement", which has a very simple dominance structure. Basically, we
1800 // are trying to find the condition that is being branched on, which
1801 // subsequently causes this merge to happen. We really want control
1802 // dependence information for this check, but simplifycfg can't keep it up
1803 // to date, and this catches most of the cases we care about anyway.
1804 BasicBlock *BB = PN->getParent();
1805 BasicBlock *IfTrue, *IfFalse;
1806 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1808 // Don't bother if the branch will be constant folded trivially.
1809 isa<ConstantInt>(IfCond))
1812 // Okay, we found that we can merge this two-entry phi node into a select.
1813 // Doing so would require us to fold *all* two entry phi nodes in this block.
1814 // At some point this becomes non-profitable (particularly if the target
1815 // doesn't support cmov's). Only do this transformation if there are two or
1816 // fewer PHI nodes in this block.
1817 unsigned NumPhis = 0;
1818 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1822 // Loop over the PHI's seeing if we can promote them all to select
1823 // instructions. While we are at it, keep track of the instructions
1824 // that need to be moved to the dominating block.
1825 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1826 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1827 MaxCostVal1 = PHINodeFoldingThreshold;
1828 MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
1829 MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
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,
1840 MaxCostVal0, TTI) ||
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->getIterator(),
1904 IfBlock1->getInstList(), IfBlock1->begin(),
1905 IfBlock1->getTerminator()->getIterator());
1907 DomBlock->getInstList().splice(InsertPt->getIterator(),
1908 IfBlock2->getInstList(), IfBlock2->begin(),
1909 IfBlock2->getTerminator()->getIterator());
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 /// If we found a conditional branch that goes to two returning blocks,
1934 /// 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 /// Given a conditional BranchInstruction, retrieve the probabilities of the
2022 /// branch taking each edge. Fills in the two APInt parameters and returns true,
2023 /// or returns false if no or invalid metadata was found.
2024 static bool ExtractBranchMetadata(BranchInst *BI,
2025 uint64_t &ProbTrue, uint64_t &ProbFalse) {
2026 assert(BI->isConditional() &&
2027 "Looking for probabilities on unconditional branch?");
2028 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
2029 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
2030 ConstantInt *CITrue =
2031 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
2032 ConstantInt *CIFalse =
2033 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
2034 if (!CITrue || !CIFalse) return false;
2035 ProbTrue = CITrue->getValue().getZExtValue();
2036 ProbFalse = CIFalse->getValue().getZExtValue();
2040 /// Return true if the given instruction is available
2041 /// 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 /// If this basic block is simple enough, and if a predecessor branches to us
2058 /// and one of our successors, fold the block into the predecessor and use
2059 /// logical operations to pick the right destination.
2060 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2061 BasicBlock *BB = BI->getParent();
2063 Instruction *Cond = nullptr;
2064 if (BI->isConditional())
2065 Cond = dyn_cast<Instruction>(BI->getCondition());
2067 // For unconditional branch, check for a simple CFG pattern, where
2068 // BB has a single predecessor and BB's successor is also its predecessor's
2069 // successor. If such pattern exisits, check for CSE between BB and its
2071 if (BasicBlock *PB = BB->getSinglePredecessor())
2072 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2073 if (PBI->isConditional() &&
2074 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2075 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2076 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
2078 Instruction *Curr = &*I++;
2079 if (isa<CmpInst>(Curr)) {
2083 // Quit if we can't remove this instruction.
2084 if (!checkCSEInPredecessor(Curr, PB))
2093 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2094 Cond->getParent() != BB || !Cond->hasOneUse())
2097 // Make sure the instruction after the condition is the cond branch.
2098 BasicBlock::iterator CondIt = ++Cond->getIterator();
2100 // Ignore dbg intrinsics.
2101 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
2106 // Only allow this transformation if computing the condition doesn't involve
2107 // too many instructions and these involved instructions can be executed
2108 // unconditionally. We denote all involved instructions except the condition
2109 // as "bonus instructions", and only allow this transformation when the
2110 // number of the bonus instructions does not exceed a certain threshold.
2111 unsigned NumBonusInsts = 0;
2112 for (auto I = BB->begin(); Cond != I; ++I) {
2113 // Ignore dbg intrinsics.
2114 if (isa<DbgInfoIntrinsic>(I))
2116 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2118 // I has only one use and can be executed unconditionally.
2119 Instruction *User = dyn_cast<Instruction>(I->user_back());
2120 if (User == nullptr || User->getParent() != BB)
2122 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2123 // to use any other instruction, User must be an instruction between next(I)
2126 // Early exits once we reach the limit.
2127 if (NumBonusInsts > BonusInstThreshold)
2131 // Cond is known to be a compare or binary operator. Check to make sure that
2132 // neither operand is a potentially-trapping constant expression.
2133 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2136 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2140 // Finally, don't infinitely unroll conditional loops.
2141 BasicBlock *TrueDest = BI->getSuccessor(0);
2142 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2143 if (TrueDest == BB || FalseDest == BB)
2146 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2147 BasicBlock *PredBlock = *PI;
2148 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2150 // Check that we have two conditional branches. If there is a PHI node in
2151 // the common successor, verify that the same value flows in from both
2153 SmallVector<PHINode*, 4> PHIs;
2154 if (!PBI || PBI->isUnconditional() ||
2155 (BI->isConditional() &&
2156 !SafeToMergeTerminators(BI, PBI)) ||
2157 (!BI->isConditional() &&
2158 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2161 // Determine if the two branches share a common destination.
2162 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2163 bool InvertPredCond = false;
2165 if (BI->isConditional()) {
2166 if (PBI->getSuccessor(0) == TrueDest)
2167 Opc = Instruction::Or;
2168 else if (PBI->getSuccessor(1) == FalseDest)
2169 Opc = Instruction::And;
2170 else if (PBI->getSuccessor(0) == FalseDest)
2171 Opc = Instruction::And, InvertPredCond = true;
2172 else if (PBI->getSuccessor(1) == TrueDest)
2173 Opc = Instruction::Or, InvertPredCond = true;
2177 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2181 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2182 IRBuilder<> Builder(PBI);
2184 // If we need to invert the condition in the pred block to match, do so now.
2185 if (InvertPredCond) {
2186 Value *NewCond = PBI->getCondition();
2188 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2189 CmpInst *CI = cast<CmpInst>(NewCond);
2190 CI->setPredicate(CI->getInversePredicate());
2192 NewCond = Builder.CreateNot(NewCond,
2193 PBI->getCondition()->getName()+".not");
2196 PBI->setCondition(NewCond);
2197 PBI->swapSuccessors();
2200 // If we have bonus instructions, clone them into the predecessor block.
2201 // Note that there may be multiple predecessor blocks, so we cannot move
2202 // bonus instructions to a predecessor block.
2203 ValueToValueMapTy VMap; // maps original values to cloned values
2204 // We already make sure Cond is the last instruction before BI. Therefore,
2205 // all instructions before Cond other than DbgInfoIntrinsic are bonus
2207 for (auto BonusInst = BB->begin(); Cond != BonusInst; ++BonusInst) {
2208 if (isa<DbgInfoIntrinsic>(BonusInst))
2210 Instruction *NewBonusInst = BonusInst->clone();
2211 RemapInstruction(NewBonusInst, VMap,
2212 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2213 VMap[&*BonusInst] = NewBonusInst;
2215 // If we moved a load, we cannot any longer claim any knowledge about
2216 // its potential value. The previous information might have been valid
2217 // only given the branch precondition.
2218 // For an analogous reason, we must also drop all the metadata whose
2219 // semantics we don't understand.
2220 NewBonusInst->dropUnknownNonDebugMetadata();
2222 PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2223 NewBonusInst->takeName(&*BonusInst);
2224 BonusInst->setName(BonusInst->getName() + ".old");
2227 // Clone Cond into the predecessor basic block, and or/and the
2228 // two conditions together.
2229 Instruction *New = Cond->clone();
2230 RemapInstruction(New, VMap,
2231 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2232 PredBlock->getInstList().insert(PBI->getIterator(), New);
2233 New->takeName(Cond);
2234 Cond->setName(New->getName() + ".old");
2236 if (BI->isConditional()) {
2237 Instruction *NewCond =
2238 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2240 PBI->setCondition(NewCond);
2242 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2243 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2245 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2247 SmallVector<uint64_t, 8> NewWeights;
2249 if (PBI->getSuccessor(0) == BB) {
2250 if (PredHasWeights && SuccHasWeights) {
2251 // PBI: br i1 %x, BB, FalseDest
2252 // BI: br i1 %y, TrueDest, FalseDest
2253 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2254 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2255 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2256 // TrueWeight for PBI * FalseWeight for BI.
2257 // We assume that total weights of a BranchInst can fit into 32 bits.
2258 // Therefore, we will not have overflow using 64-bit arithmetic.
2259 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2260 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2262 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2263 PBI->setSuccessor(0, TrueDest);
2265 if (PBI->getSuccessor(1) == BB) {
2266 if (PredHasWeights && SuccHasWeights) {
2267 // PBI: br i1 %x, TrueDest, BB
2268 // BI: br i1 %y, TrueDest, FalseDest
2269 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2270 // FalseWeight for PBI * TrueWeight for BI.
2271 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2272 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2273 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2274 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2276 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2277 PBI->setSuccessor(1, FalseDest);
2279 if (NewWeights.size() == 2) {
2280 // Halve the weights if any of them cannot fit in an uint32_t
2281 FitWeights(NewWeights);
2283 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2284 PBI->setMetadata(LLVMContext::MD_prof,
2285 MDBuilder(BI->getContext()).
2286 createBranchWeights(MDWeights));
2288 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2290 // Update PHI nodes in the common successors.
2291 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2292 ConstantInt *PBI_C = cast<ConstantInt>(
2293 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2294 assert(PBI_C->getType()->isIntegerTy(1));
2295 Instruction *MergedCond = nullptr;
2296 if (PBI->getSuccessor(0) == TrueDest) {
2297 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2298 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2299 // is false: !PBI_Cond and BI_Value
2300 Instruction *NotCond =
2301 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2304 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2309 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2310 PBI->getCondition(), MergedCond,
2313 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2314 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2315 // is false: PBI_Cond and BI_Value
2317 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2318 PBI->getCondition(), New,
2320 if (PBI_C->isOne()) {
2321 Instruction *NotCond =
2322 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2325 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2326 NotCond, MergedCond,
2331 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2334 // Change PBI from Conditional to Unconditional.
2335 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2336 EraseTerminatorInstAndDCECond(PBI);
2340 // TODO: If BB is reachable from all paths through PredBlock, then we
2341 // could replace PBI's branch probabilities with BI's.
2343 // Copy any debug value intrinsics into the end of PredBlock.
2344 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2345 if (isa<DbgInfoIntrinsic>(*I))
2346 I->clone()->insertBefore(PBI);
2353 // If there is only one store in BB1 and BB2, return it, otherwise return
2355 static StoreInst *findUniqueStoreInBlocks(BasicBlock *BB1, BasicBlock *BB2) {
2356 StoreInst *S = nullptr;
2357 for (auto *BB : {BB1, BB2}) {
2361 if (auto *SI = dyn_cast<StoreInst>(&I)) {
2363 // Multiple stores seen.
2372 static Value *ensureValueAvailableInSuccessor(Value *V, BasicBlock *BB,
2373 Value *AlternativeV = nullptr) {
2374 // PHI is going to be a PHI node that allows the value V that is defined in
2375 // BB to be referenced in BB's only successor.
2377 // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2378 // doesn't matter to us what the other operand is (it'll never get used). We
2379 // could just create a new PHI with an undef incoming value, but that could
2380 // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2381 // other PHI. So here we directly look for some PHI in BB's successor with V
2382 // as an incoming operand. If we find one, we use it, else we create a new
2385 // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2386 // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2387 // where OtherBB is the single other predecessor of BB's only successor.
2388 PHINode *PHI = nullptr;
2389 BasicBlock *Succ = BB->getSingleSuccessor();
2391 for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2392 if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2393 PHI = cast<PHINode>(I);
2397 assert(std::distance(pred_begin(Succ), pred_end(Succ)) == 2);
2398 auto PredI = pred_begin(Succ);
2399 BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2400 if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2407 PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2408 PHI->addIncoming(V, BB);
2409 for (BasicBlock *PredBB : predecessors(Succ))
2411 PHI->addIncoming(AlternativeV ? AlternativeV : UndefValue::get(V->getType()),
2416 static bool mergeConditionalStoreToAddress(BasicBlock *PTB, BasicBlock *PFB,
2417 BasicBlock *QTB, BasicBlock *QFB,
2418 BasicBlock *PostBB, Value *Address,
2419 bool InvertPCond, bool InvertQCond) {
2420 auto IsaBitcastOfPointerType = [](const Instruction &I) {
2421 return Operator::getOpcode(&I) == Instruction::BitCast &&
2422 I.getType()->isPointerTy();
2425 // If we're not in aggressive mode, we only optimize if we have some
2426 // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2427 auto IsWorthwhile = [&](BasicBlock *BB) {
2430 // Heuristic: if the block can be if-converted/phi-folded and the
2431 // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2432 // thread this store.
2434 for (auto &I : *BB) {
2435 // Cheap instructions viable for folding.
2436 if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2439 // Free instructions.
2440 else if (isa<TerminatorInst>(I) || isa<DbgInfoIntrinsic>(I) ||
2441 IsaBitcastOfPointerType(I))
2446 return N <= PHINodeFoldingThreshold;
2449 if (!MergeCondStoresAggressively && (!IsWorthwhile(PTB) ||
2450 !IsWorthwhile(PFB) ||
2451 !IsWorthwhile(QTB) ||
2452 !IsWorthwhile(QFB)))
2455 // For every pointer, there must be exactly two stores, one coming from
2456 // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2457 // store (to any address) in PTB,PFB or QTB,QFB.
2458 // FIXME: We could relax this restriction with a bit more work and performance
2460 StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2461 StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2462 if (!PStore || !QStore)
2465 // Now check the stores are compatible.
2466 if (!QStore->isUnordered() || !PStore->isUnordered())
2469 // Check that sinking the store won't cause program behavior changes. Sinking
2470 // the store out of the Q blocks won't change any behavior as we're sinking
2471 // from a block to its unconditional successor. But we're moving a store from
2472 // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2473 // So we need to check that there are no aliasing loads or stores in
2474 // QBI, QTB and QFB. We also need to check there are no conflicting memory
2475 // operations between PStore and the end of its parent block.
2477 // The ideal way to do this is to query AliasAnalysis, but we don't
2478 // preserve AA currently so that is dangerous. Be super safe and just
2479 // check there are no other memory operations at all.
2480 for (auto &I : *QFB->getSinglePredecessor())
2481 if (I.mayReadOrWriteMemory())
2483 for (auto &I : *QFB)
2484 if (&I != QStore && I.mayReadOrWriteMemory())
2487 for (auto &I : *QTB)
2488 if (&I != QStore && I.mayReadOrWriteMemory())
2490 for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2492 if (&*I != PStore && I->mayReadOrWriteMemory())
2495 // OK, we're going to sink the stores to PostBB. The store has to be
2496 // conditional though, so first create the predicate.
2497 Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
2499 Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
2502 Value *PPHI = ensureValueAvailableInSuccessor(PStore->getValueOperand(),
2503 PStore->getParent());
2504 Value *QPHI = ensureValueAvailableInSuccessor(QStore->getValueOperand(),
2505 QStore->getParent(), PPHI);
2507 IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
2509 Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
2510 Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
2513 PPred = QB.CreateNot(PPred);
2515 QPred = QB.CreateNot(QPred);
2516 Value *CombinedPred = QB.CreateOr(PPred, QPred);
2519 SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
2520 QB.SetInsertPoint(T);
2521 StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
2523 PStore->getAAMetadata(AAMD, /*Merge=*/false);
2524 PStore->getAAMetadata(AAMD, /*Merge=*/true);
2525 SI->setAAMetadata(AAMD);
2527 QStore->eraseFromParent();
2528 PStore->eraseFromParent();
2533 static bool mergeConditionalStores(BranchInst *PBI, BranchInst *QBI) {
2534 // The intention here is to find diamonds or triangles (see below) where each
2535 // conditional block contains a store to the same address. Both of these
2536 // stores are conditional, so they can't be unconditionally sunk. But it may
2537 // be profitable to speculatively sink the stores into one merged store at the
2538 // end, and predicate the merged store on the union of the two conditions of
2541 // This can reduce the number of stores executed if both of the conditions are
2542 // true, and can allow the blocks to become small enough to be if-converted.
2543 // This optimization will also chain, so that ladders of test-and-set
2544 // sequences can be if-converted away.
2546 // We only deal with simple diamonds or triangles:
2548 // PBI or PBI or a combination of the two
2558 // We model triangles as a type of diamond with a nullptr "true" block.
2559 // Triangles are canonicalized so that the fallthrough edge is represented by
2560 // a true condition, as in the diagram above.
2562 BasicBlock *PTB = PBI->getSuccessor(0);
2563 BasicBlock *PFB = PBI->getSuccessor(1);
2564 BasicBlock *QTB = QBI->getSuccessor(0);
2565 BasicBlock *QFB = QBI->getSuccessor(1);
2566 BasicBlock *PostBB = QFB->getSingleSuccessor();
2568 bool InvertPCond = false, InvertQCond = false;
2569 // Canonicalize fallthroughs to the true branches.
2570 if (PFB == QBI->getParent()) {
2571 std::swap(PFB, PTB);
2574 if (QFB == PostBB) {
2575 std::swap(QFB, QTB);
2579 // From this point on we can assume PTB or QTB may be fallthroughs but PFB
2580 // and QFB may not. Model fallthroughs as a nullptr block.
2581 if (PTB == QBI->getParent())
2586 // Legality bailouts. We must have at least the non-fallthrough blocks and
2587 // the post-dominating block, and the non-fallthroughs must only have one
2589 auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
2590 return BB->getSinglePredecessor() == P &&
2591 BB->getSingleSuccessor() == S;
2594 !HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
2595 !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
2597 if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
2598 (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
2600 if (PostBB->getNumUses() != 2 || QBI->getParent()->getNumUses() != 2)
2603 // OK, this is a sequence of two diamonds or triangles.
2604 // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
2605 SmallPtrSet<Value *,4> PStoreAddresses, QStoreAddresses;
2606 for (auto *BB : {PTB, PFB}) {
2610 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
2611 PStoreAddresses.insert(SI->getPointerOperand());
2613 for (auto *BB : {QTB, QFB}) {
2617 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
2618 QStoreAddresses.insert(SI->getPointerOperand());
2621 set_intersect(PStoreAddresses, QStoreAddresses);
2622 // set_intersect mutates PStoreAddresses in place. Rename it here to make it
2623 // clear what it contains.
2624 auto &CommonAddresses = PStoreAddresses;
2626 bool Changed = false;
2627 for (auto *Address : CommonAddresses)
2628 Changed |= mergeConditionalStoreToAddress(
2629 PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond);
2633 /// If we have a conditional branch as a predecessor of another block,
2634 /// this function tries to simplify it. We know
2635 /// that PBI and BI are both conditional branches, and BI is in one of the
2636 /// successor blocks of PBI - PBI branches to BI.
2637 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
2638 const DataLayout &DL) {
2639 assert(PBI->isConditional() && BI->isConditional());
2640 BasicBlock *BB = BI->getParent();
2642 // If this block ends with a branch instruction, and if there is a
2643 // predecessor that ends on a branch of the same condition, make
2644 // this conditional branch redundant.
2645 if (PBI->getCondition() == BI->getCondition() &&
2646 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2647 // Okay, the outcome of this conditional branch is statically
2648 // knowable. If this block had a single pred, handle specially.
2649 if (BB->getSinglePredecessor()) {
2650 // Turn this into a branch on constant.
2651 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2652 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2654 return true; // Nuke the branch on constant.
2657 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2658 // in the constant and simplify the block result. Subsequent passes of
2659 // simplifycfg will thread the block.
2660 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2661 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2662 PHINode *NewPN = PHINode::Create(
2663 Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
2664 BI->getCondition()->getName() + ".pr", &BB->front());
2665 // Okay, we're going to insert the PHI node. Since PBI is not the only
2666 // predecessor, compute the PHI'd conditional value for all of the preds.
2667 // Any predecessor where the condition is not computable we keep symbolic.
2668 for (pred_iterator PI = PB; PI != PE; ++PI) {
2669 BasicBlock *P = *PI;
2670 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2671 PBI != BI && PBI->isConditional() &&
2672 PBI->getCondition() == BI->getCondition() &&
2673 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2674 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2675 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2678 NewPN->addIncoming(BI->getCondition(), P);
2682 BI->setCondition(NewPN);
2687 if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2691 // If BI is reached from the true path of PBI and PBI's condition implies
2692 // BI's condition, we know the direction of the BI branch.
2693 if (PBI->getSuccessor(0) == BI->getParent() &&
2694 isImpliedCondition(PBI->getCondition(), BI->getCondition(), DL) &&
2695 PBI->getSuccessor(0) != PBI->getSuccessor(1) &&
2696 BB->getSinglePredecessor()) {
2697 // Turn this into a branch on constant.
2698 auto *OldCond = BI->getCondition();
2699 BI->setCondition(ConstantInt::getTrue(BB->getContext()));
2700 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
2701 return true; // Nuke the branch on constant.
2704 // If both branches are conditional and both contain stores to the same
2705 // address, remove the stores from the conditionals and create a conditional
2706 // merged store at the end.
2707 if (MergeCondStores && mergeConditionalStores(PBI, BI))
2710 // If this is a conditional branch in an empty block, and if any
2711 // predecessors are a conditional branch to one of our destinations,
2712 // fold the conditions into logical ops and one cond br.
2713 BasicBlock::iterator BBI = BB->begin();
2714 // Ignore dbg intrinsics.
2715 while (isa<DbgInfoIntrinsic>(BBI))
2721 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2723 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2724 PBIOp = 0, BIOp = 1;
2725 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2726 PBIOp = 1, BIOp = 0;
2727 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2732 // Check to make sure that the other destination of this branch
2733 // isn't BB itself. If so, this is an infinite loop that will
2734 // keep getting unwound.
2735 if (PBI->getSuccessor(PBIOp) == BB)
2738 // Do not perform this transformation if it would require
2739 // insertion of a large number of select instructions. For targets
2740 // without predication/cmovs, this is a big pessimization.
2742 // Also do not perform this transformation if any phi node in the common
2743 // destination block can trap when reached by BB or PBB (PR17073). In that
2744 // case, it would be unsafe to hoist the operation into a select instruction.
2746 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2747 unsigned NumPhis = 0;
2748 for (BasicBlock::iterator II = CommonDest->begin();
2749 isa<PHINode>(II); ++II, ++NumPhis) {
2750 if (NumPhis > 2) // Disable this xform.
2753 PHINode *PN = cast<PHINode>(II);
2754 Value *BIV = PN->getIncomingValueForBlock(BB);
2755 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
2759 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2760 Value *PBIV = PN->getIncomingValue(PBBIdx);
2761 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
2766 // Finally, if everything is ok, fold the branches to logical ops.
2767 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2769 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2770 << "AND: " << *BI->getParent());
2773 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2774 // branch in it, where one edge (OtherDest) goes back to itself but the other
2775 // exits. We don't *know* that the program avoids the infinite loop
2776 // (even though that seems likely). If we do this xform naively, we'll end up
2777 // recursively unpeeling the loop. Since we know that (after the xform is
2778 // done) that the block *is* infinite if reached, we just make it an obviously
2779 // infinite loop with no cond branch.
2780 if (OtherDest == BB) {
2781 // Insert it at the end of the function, because it's either code,
2782 // or it won't matter if it's hot. :)
2783 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2784 "infloop", BB->getParent());
2785 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2786 OtherDest = InfLoopBlock;
2789 DEBUG(dbgs() << *PBI->getParent()->getParent());
2791 // BI may have other predecessors. Because of this, we leave
2792 // it alone, but modify PBI.
2794 // Make sure we get to CommonDest on True&True directions.
2795 Value *PBICond = PBI->getCondition();
2796 IRBuilder<true, NoFolder> Builder(PBI);
2798 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2800 Value *BICond = BI->getCondition();
2802 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2804 // Merge the conditions.
2805 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2807 // Modify PBI to branch on the new condition to the new dests.
2808 PBI->setCondition(Cond);
2809 PBI->setSuccessor(0, CommonDest);
2810 PBI->setSuccessor(1, OtherDest);
2812 // Update branch weight for PBI.
2813 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2814 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2816 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2818 if (PredHasWeights && SuccHasWeights) {
2819 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2820 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2821 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2822 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2823 // The weight to CommonDest should be PredCommon * SuccTotal +
2824 // PredOther * SuccCommon.
2825 // The weight to OtherDest should be PredOther * SuccOther.
2826 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
2827 PredOther * SuccCommon,
2828 PredOther * SuccOther};
2829 // Halve the weights if any of them cannot fit in an uint32_t
2830 FitWeights(NewWeights);
2832 PBI->setMetadata(LLVMContext::MD_prof,
2833 MDBuilder(BI->getContext())
2834 .createBranchWeights(NewWeights[0], NewWeights[1]));
2837 // OtherDest may have phi nodes. If so, add an entry from PBI's
2838 // block that are identical to the entries for BI's block.
2839 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2841 // We know that the CommonDest already had an edge from PBI to
2842 // it. If it has PHIs though, the PHIs may have different
2843 // entries for BB and PBI's BB. If so, insert a select to make
2846 for (BasicBlock::iterator II = CommonDest->begin();
2847 (PN = dyn_cast<PHINode>(II)); ++II) {
2848 Value *BIV = PN->getIncomingValueForBlock(BB);
2849 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2850 Value *PBIV = PN->getIncomingValue(PBBIdx);
2852 // Insert a select in PBI to pick the right value.
2853 Value *NV = cast<SelectInst>
2854 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2855 PN->setIncomingValue(PBBIdx, NV);
2859 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2860 DEBUG(dbgs() << *PBI->getParent()->getParent());
2862 // This basic block is probably dead. We know it has at least
2863 // one fewer predecessor.
2867 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
2868 // true or to FalseBB if Cond is false.
2869 // Takes care of updating the successors and removing the old terminator.
2870 // Also makes sure not to introduce new successors by assuming that edges to
2871 // non-successor TrueBBs and FalseBBs aren't reachable.
2872 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2873 BasicBlock *TrueBB, BasicBlock *FalseBB,
2874 uint32_t TrueWeight,
2875 uint32_t FalseWeight){
2876 // Remove any superfluous successor edges from the CFG.
2877 // First, figure out which successors to preserve.
2878 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2880 BasicBlock *KeepEdge1 = TrueBB;
2881 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
2883 // Then remove the rest.
2884 for (BasicBlock *Succ : OldTerm->successors()) {
2885 // Make sure only to keep exactly one copy of each edge.
2886 if (Succ == KeepEdge1)
2887 KeepEdge1 = nullptr;
2888 else if (Succ == KeepEdge2)
2889 KeepEdge2 = nullptr;
2891 Succ->removePredecessor(OldTerm->getParent(),
2892 /*DontDeleteUselessPHIs=*/true);
2895 IRBuilder<> Builder(OldTerm);
2896 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2898 // Insert an appropriate new terminator.
2899 if (!KeepEdge1 && !KeepEdge2) {
2900 if (TrueBB == FalseBB)
2901 // We were only looking for one successor, and it was present.
2902 // Create an unconditional branch to it.
2903 Builder.CreateBr(TrueBB);
2905 // We found both of the successors we were looking for.
2906 // Create a conditional branch sharing the condition of the select.
2907 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2908 if (TrueWeight != FalseWeight)
2909 NewBI->setMetadata(LLVMContext::MD_prof,
2910 MDBuilder(OldTerm->getContext()).
2911 createBranchWeights(TrueWeight, FalseWeight));
2913 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2914 // Neither of the selected blocks were successors, so this
2915 // terminator must be unreachable.
2916 new UnreachableInst(OldTerm->getContext(), OldTerm);
2918 // One of the selected values was a successor, but the other wasn't.
2919 // Insert an unconditional branch to the one that was found;
2920 // the edge to the one that wasn't must be unreachable.
2922 // Only TrueBB was found.
2923 Builder.CreateBr(TrueBB);
2925 // Only FalseBB was found.
2926 Builder.CreateBr(FalseBB);
2929 EraseTerminatorInstAndDCECond(OldTerm);
2934 // (switch (select cond, X, Y)) on constant X, Y
2935 // with a branch - conditional if X and Y lead to distinct BBs,
2936 // unconditional otherwise.
2937 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2938 // Check for constant integer values in the select.
2939 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2940 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2941 if (!TrueVal || !FalseVal)
2944 // Find the relevant condition and destinations.
2945 Value *Condition = Select->getCondition();
2946 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2947 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2949 // Get weight for TrueBB and FalseBB.
2950 uint32_t TrueWeight = 0, FalseWeight = 0;
2951 SmallVector<uint64_t, 8> Weights;
2952 bool HasWeights = HasBranchWeights(SI);
2954 GetBranchWeights(SI, Weights);
2955 if (Weights.size() == 1 + SI->getNumCases()) {
2956 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2957 getSuccessorIndex()];
2958 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2959 getSuccessorIndex()];
2963 // Perform the actual simplification.
2964 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2965 TrueWeight, FalseWeight);
2969 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2970 // blockaddress(@fn, BlockB)))
2972 // (br cond, BlockA, BlockB).
2973 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2974 // Check that both operands of the select are block addresses.
2975 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2976 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2980 // Extract the actual blocks.
2981 BasicBlock *TrueBB = TBA->getBasicBlock();
2982 BasicBlock *FalseBB = FBA->getBasicBlock();
2984 // Perform the actual simplification.
2985 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2989 /// This is called when we find an icmp instruction
2990 /// (a seteq/setne with a constant) as the only instruction in a
2991 /// block that ends with an uncond branch. We are looking for a very specific
2992 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2993 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2994 /// default value goes to an uncond block with a seteq in it, we get something
2997 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2999 /// %tmp = icmp eq i8 %A, 92
3002 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3004 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3005 /// the PHI, merging the third icmp into the switch.
3006 static bool TryToSimplifyUncondBranchWithICmpInIt(
3007 ICmpInst *ICI, IRBuilder<> &Builder, const DataLayout &DL,
3008 const TargetTransformInfo &TTI, unsigned BonusInstThreshold,
3009 AssumptionCache *AC) {
3010 BasicBlock *BB = ICI->getParent();
3012 // If the block has any PHIs in it or the icmp has multiple uses, it is too
3014 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
3016 Value *V = ICI->getOperand(0);
3017 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3019 // The pattern we're looking for is where our only predecessor is a switch on
3020 // 'V' and this block is the default case for the switch. In this case we can
3021 // fold the compared value into the switch to simplify things.
3022 BasicBlock *Pred = BB->getSinglePredecessor();
3023 if (!Pred || !isa<SwitchInst>(Pred->getTerminator())) return false;
3025 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3026 if (SI->getCondition() != V)
3029 // If BB is reachable on a non-default case, then we simply know the value of
3030 // V in this block. Substitute it and constant fold the icmp instruction
3032 if (SI->getDefaultDest() != BB) {
3033 ConstantInt *VVal = SI->findCaseDest(BB);
3034 assert(VVal && "Should have a unique destination value");
3035 ICI->setOperand(0, VVal);
3037 if (Value *V = SimplifyInstruction(ICI, DL)) {
3038 ICI->replaceAllUsesWith(V);
3039 ICI->eraseFromParent();
3041 // BB is now empty, so it is likely to simplify away.
3042 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3045 // Ok, the block is reachable from the default dest. If the constant we're
3046 // comparing exists in one of the other edges, then we can constant fold ICI
3048 if (SI->findCaseValue(Cst) != SI->case_default()) {
3050 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3051 V = ConstantInt::getFalse(BB->getContext());
3053 V = ConstantInt::getTrue(BB->getContext());
3055 ICI->replaceAllUsesWith(V);
3056 ICI->eraseFromParent();
3057 // BB is now empty, so it is likely to simplify away.
3058 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3061 // The use of the icmp has to be in the 'end' block, by the only PHI node in
3063 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3064 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3065 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3066 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3069 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3071 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3072 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3074 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3075 std::swap(DefaultCst, NewCst);
3077 // Replace ICI (which is used by the PHI for the default value) with true or
3078 // false depending on if it is EQ or NE.
3079 ICI->replaceAllUsesWith(DefaultCst);
3080 ICI->eraseFromParent();
3082 // Okay, the switch goes to this block on a default value. Add an edge from
3083 // the switch to the merge point on the compared value.
3084 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
3085 BB->getParent(), BB);
3086 SmallVector<uint64_t, 8> Weights;
3087 bool HasWeights = HasBranchWeights(SI);
3089 GetBranchWeights(SI, Weights);
3090 if (Weights.size() == 1 + SI->getNumCases()) {
3091 // Split weight for default case to case for "Cst".
3092 Weights[0] = (Weights[0]+1) >> 1;
3093 Weights.push_back(Weights[0]);
3095 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3096 SI->setMetadata(LLVMContext::MD_prof,
3097 MDBuilder(SI->getContext()).
3098 createBranchWeights(MDWeights));
3101 SI->addCase(Cst, NewBB);
3103 // NewBB branches to the phi block, add the uncond branch and the phi entry.
3104 Builder.SetInsertPoint(NewBB);
3105 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3106 Builder.CreateBr(SuccBlock);
3107 PHIUse->addIncoming(NewCst, NewBB);
3111 /// The specified branch is a conditional branch.
3112 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3113 /// fold it into a switch instruction if so.
3114 static bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
3115 const DataLayout &DL) {
3116 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3117 if (!Cond) return false;
3119 // Change br (X == 0 | X == 1), T, F into a switch instruction.
3120 // If this is a bunch of seteq's or'd together, or if it's a bunch of
3121 // 'setne's and'ed together, collect them.
3123 // Try to gather values from a chain of and/or to be turned into a switch
3124 ConstantComparesGatherer ConstantCompare(Cond, DL);
3125 // Unpack the result
3126 SmallVectorImpl<ConstantInt*> &Values = ConstantCompare.Vals;
3127 Value *CompVal = ConstantCompare.CompValue;
3128 unsigned UsedICmps = ConstantCompare.UsedICmps;
3129 Value *ExtraCase = ConstantCompare.Extra;
3131 // If we didn't have a multiply compared value, fail.
3132 if (!CompVal) return false;
3134 // Avoid turning single icmps into a switch.
3138 bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3140 // There might be duplicate constants in the list, which the switch
3141 // instruction can't handle, remove them now.
3142 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3143 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3145 // If Extra was used, we require at least two switch values to do the
3146 // transformation. A switch with one value is just a conditional branch.
3147 if (ExtraCase && Values.size() < 2) return false;
3149 // TODO: Preserve branch weight metadata, similarly to how
3150 // FoldValueComparisonIntoPredecessors preserves it.
3152 // Figure out which block is which destination.
3153 BasicBlock *DefaultBB = BI->getSuccessor(1);
3154 BasicBlock *EdgeBB = BI->getSuccessor(0);
3155 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
3157 BasicBlock *BB = BI->getParent();
3159 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3160 << " cases into SWITCH. BB is:\n" << *BB);
3162 // If there are any extra values that couldn't be folded into the switch
3163 // then we evaluate them with an explicit branch first. Split the block
3164 // right before the condbr to handle it.
3167 BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3168 // Remove the uncond branch added to the old block.
3169 TerminatorInst *OldTI = BB->getTerminator();
3170 Builder.SetInsertPoint(OldTI);
3173 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3175 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3177 OldTI->eraseFromParent();
3179 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3180 // for the edge we just added.
3181 AddPredecessorToBlock(EdgeBB, BB, NewBB);
3183 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3184 << "\nEXTRABB = " << *BB);
3188 Builder.SetInsertPoint(BI);
3189 // Convert pointer to int before we switch.
3190 if (CompVal->getType()->isPointerTy()) {
3191 CompVal = Builder.CreatePtrToInt(
3192 CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3195 // Create the new switch instruction now.
3196 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3198 // Add all of the 'cases' to the switch instruction.
3199 for (unsigned i = 0, e = Values.size(); i != e; ++i)
3200 New->addCase(Values[i], EdgeBB);
3202 // We added edges from PI to the EdgeBB. As such, if there were any
3203 // PHI nodes in EdgeBB, they need entries to be added corresponding to
3204 // the number of edges added.
3205 for (BasicBlock::iterator BBI = EdgeBB->begin();
3206 isa<PHINode>(BBI); ++BBI) {
3207 PHINode *PN = cast<PHINode>(BBI);
3208 Value *InVal = PN->getIncomingValueForBlock(BB);
3209 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
3210 PN->addIncoming(InVal, BB);
3213 // Erase the old branch instruction.
3214 EraseTerminatorInstAndDCECond(BI);
3216 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3220 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3221 // If this is a trivial landing pad that just continues unwinding the caught
3222 // exception then zap the landing pad, turning its invokes into calls.
3223 BasicBlock *BB = RI->getParent();
3224 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
3225 if (RI->getValue() != LPInst)
3226 // Not a landing pad, or the resume is not unwinding the exception that
3227 // caused control to branch here.
3230 // Check that there are no other instructions except for debug intrinsics.
3231 BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3233 if (!isa<DbgInfoIntrinsic>(I))
3236 // Turn all invokes that unwind here into calls and delete the basic block.
3237 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3238 BasicBlock *Pred = *PI++;
3239 removeUnwindEdge(Pred);
3242 // The landingpad is now unreachable. Zap it.
3243 BB->eraseFromParent();
3247 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
3248 // If this is a trivial cleanup pad that executes no instructions, it can be
3249 // eliminated. If the cleanup pad continues to the caller, any predecessor
3250 // that is an EH pad will be updated to continue to the caller and any
3251 // predecessor that terminates with an invoke instruction will have its invoke
3252 // instruction converted to a call instruction. If the cleanup pad being
3253 // simplified does not continue to the caller, each predecessor will be
3254 // updated to continue to the unwind destination of the cleanup pad being
3256 BasicBlock *BB = RI->getParent();
3257 Instruction *CPInst = dyn_cast<CleanupPadInst>(BB->getFirstNonPHI());
3259 // This isn't an empty cleanup.
3262 // Check that there are no other instructions except for debug intrinsics.
3263 BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3265 if (!isa<DbgInfoIntrinsic>(I))
3268 // If the cleanup return we are simplifying unwinds to the caller, this
3269 // will set UnwindDest to nullptr.
3270 BasicBlock *UnwindDest = RI->getUnwindDest();
3272 // We're about to remove BB from the control flow. Before we do, sink any
3273 // PHINodes into the unwind destination. Doing this before changing the
3274 // control flow avoids some potentially slow checks, since we can currently
3275 // be certain that UnwindDest and BB have no common predecessors (since they
3276 // are both EH pads).
3278 // First, go through the PHI nodes in UnwindDest and update any nodes that
3279 // reference the block we are removing
3280 for (BasicBlock::iterator I = UnwindDest->begin(),
3281 IE = UnwindDest->getFirstNonPHI()->getIterator();
3283 PHINode *DestPN = cast<PHINode>(I);
3285 int Idx = DestPN->getBasicBlockIndex(BB);
3286 // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3288 // This PHI node has an incoming value that corresponds to a control
3289 // path through the cleanup pad we are removing. If the incoming
3290 // value is in the cleanup pad, it must be a PHINode (because we
3291 // verified above that the block is otherwise empty). Otherwise, the
3292 // value is either a constant or a value that dominates the cleanup
3293 // pad being removed.
3295 // Because BB and UnwindDest are both EH pads, all of their
3296 // predecessors must unwind to these blocks, and since no instruction
3297 // can have multiple unwind destinations, there will be no overlap in
3298 // incoming blocks between SrcPN and DestPN.
3299 Value *SrcVal = DestPN->getIncomingValue(Idx);
3300 PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3302 // Remove the entry for the block we are deleting.
3303 DestPN->removeIncomingValue(Idx, false);
3305 if (SrcPN && SrcPN->getParent() == BB) {
3306 // If the incoming value was a PHI node in the cleanup pad we are
3307 // removing, we need to merge that PHI node's incoming values into
3309 for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3310 SrcIdx != SrcE; ++SrcIdx) {
3311 DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3312 SrcPN->getIncomingBlock(SrcIdx));
3315 // Otherwise, the incoming value came from above BB and
3316 // so we can just reuse it. We must associate all of BB's
3317 // predecessors with this value.
3318 for (auto *pred : predecessors(BB)) {
3319 DestPN->addIncoming(SrcVal, pred);
3324 // Sink any remaining PHI nodes directly into UnwindDest.
3325 Instruction *InsertPt = UnwindDest->getFirstNonPHI();
3326 for (BasicBlock::iterator I = BB->begin(),
3327 IE = BB->getFirstNonPHI()->getIterator();
3329 // The iterator must be incremented here because the instructions are
3330 // being moved to another block.
3331 PHINode *PN = cast<PHINode>(I++);
3332 if (PN->use_empty())
3333 // If the PHI node has no uses, just leave it. It will be erased
3334 // when we erase BB below.
3337 // Otherwise, sink this PHI node into UnwindDest.
3338 // Any predecessors to UnwindDest which are not already represented
3339 // must be back edges which inherit the value from the path through
3340 // BB. In this case, the PHI value must reference itself.
3341 for (auto *pred : predecessors(UnwindDest))
3343 PN->addIncoming(PN, pred);
3344 PN->moveBefore(InsertPt);
3348 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3349 // The iterator must be updated here because we are removing this pred.
3350 BasicBlock *PredBB = *PI++;
3351 if (UnwindDest == nullptr) {
3352 removeUnwindEdge(PredBB);
3354 TerminatorInst *TI = PredBB->getTerminator();
3355 TI->replaceUsesOfWith(BB, UnwindDest);
3359 // The cleanup pad is now unreachable. Zap it.
3360 BB->eraseFromParent();
3364 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
3365 BasicBlock *BB = RI->getParent();
3366 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
3368 // Find predecessors that end with branches.
3369 SmallVector<BasicBlock*, 8> UncondBranchPreds;
3370 SmallVector<BranchInst*, 8> CondBranchPreds;
3371 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
3372 BasicBlock *P = *PI;
3373 TerminatorInst *PTI = P->getTerminator();
3374 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
3375 if (BI->isUnconditional())
3376 UncondBranchPreds.push_back(P);
3378 CondBranchPreds.push_back(BI);
3382 // If we found some, do the transformation!
3383 if (!UncondBranchPreds.empty() && DupRet) {
3384 while (!UncondBranchPreds.empty()) {
3385 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
3386 DEBUG(dbgs() << "FOLDING: " << *BB
3387 << "INTO UNCOND BRANCH PRED: " << *Pred);
3388 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
3391 // If we eliminated all predecessors of the block, delete the block now.
3393 // We know there are no successors, so just nuke the block.
3394 BB->eraseFromParent();
3399 // Check out all of the conditional branches going to this return
3400 // instruction. If any of them just select between returns, change the
3401 // branch itself into a select/return pair.
3402 while (!CondBranchPreds.empty()) {
3403 BranchInst *BI = CondBranchPreds.pop_back_val();
3405 // Check to see if the non-BB successor is also a return block.
3406 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
3407 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
3408 SimplifyCondBranchToTwoReturns(BI, Builder))
3414 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
3415 BasicBlock *BB = UI->getParent();
3417 bool Changed = false;
3419 // If there are any instructions immediately before the unreachable that can
3420 // be removed, do so.
3421 while (UI->getIterator() != BB->begin()) {
3422 BasicBlock::iterator BBI = UI->getIterator();
3424 // Do not delete instructions that can have side effects which might cause
3425 // the unreachable to not be reachable; specifically, calls and volatile
3426 // operations may have this effect.
3427 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
3429 if (BBI->mayHaveSideEffects()) {
3430 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
3431 if (SI->isVolatile())
3433 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
3434 if (LI->isVolatile())
3436 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
3437 if (RMWI->isVolatile())
3439 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
3440 if (CXI->isVolatile())
3442 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
3443 !isa<LandingPadInst>(BBI)) {
3446 // Note that deleting LandingPad's here is in fact okay, although it
3447 // involves a bit of subtle reasoning. If this inst is a LandingPad,
3448 // all the predecessors of this block will be the unwind edges of Invokes,
3449 // and we can therefore guarantee this block will be erased.
3452 // Delete this instruction (any uses are guaranteed to be dead)
3453 if (!BBI->use_empty())
3454 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
3455 BBI->eraseFromParent();
3459 // If the unreachable instruction is the first in the block, take a gander
3460 // at all of the predecessors of this instruction, and simplify them.
3461 if (&BB->front() != UI) return Changed;
3463 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
3464 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
3465 TerminatorInst *TI = Preds[i]->getTerminator();
3466 IRBuilder<> Builder(TI);
3467 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3468 if (BI->isUnconditional()) {
3469 if (BI->getSuccessor(0) == BB) {
3470 new UnreachableInst(TI->getContext(), TI);
3471 TI->eraseFromParent();
3475 if (BI->getSuccessor(0) == BB) {
3476 Builder.CreateBr(BI->getSuccessor(1));
3477 EraseTerminatorInstAndDCECond(BI);
3478 } else if (BI->getSuccessor(1) == BB) {
3479 Builder.CreateBr(BI->getSuccessor(0));
3480 EraseTerminatorInstAndDCECond(BI);
3484 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3485 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3487 if (i.getCaseSuccessor() == BB) {
3488 BB->removePredecessor(SI->getParent());
3493 } else if ((isa<InvokeInst>(TI) &&
3494 cast<InvokeInst>(TI)->getUnwindDest() == BB) ||
3495 isa<CatchEndPadInst>(TI) || isa<TerminatePadInst>(TI)) {
3496 removeUnwindEdge(TI->getParent());
3498 } else if (isa<CleanupReturnInst>(TI) || isa<CleanupEndPadInst>(TI)) {
3499 new UnreachableInst(TI->getContext(), TI);
3500 TI->eraseFromParent();
3503 // TODO: If TI is a CatchPadInst, then (BB must be its normal dest and)
3504 // we can eliminate it, redirecting its preds to its unwind successor,
3505 // or to the next outer handler if the removed catch is the last for its
3509 // If this block is now dead, remove it.
3510 if (pred_empty(BB) &&
3511 BB != &BB->getParent()->getEntryBlock()) {
3512 // We know there are no successors, so just nuke the block.
3513 BB->eraseFromParent();
3520 static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) {
3521 assert(Cases.size() >= 1);
3523 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3524 for (size_t I = 1, E = Cases.size(); I != E; ++I) {
3525 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
3531 /// Turn a switch with two reachable destinations into an integer range
3532 /// comparison and branch.
3533 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3534 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3537 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
3539 // Partition the cases into two sets with different destinations.
3540 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
3541 BasicBlock *DestB = nullptr;
3542 SmallVector <ConstantInt *, 16> CasesA;
3543 SmallVector <ConstantInt *, 16> CasesB;
3545 for (SwitchInst::CaseIt I : SI->cases()) {
3546 BasicBlock *Dest = I.getCaseSuccessor();
3547 if (!DestA) DestA = Dest;
3548 if (Dest == DestA) {
3549 CasesA.push_back(I.getCaseValue());
3552 if (!DestB) DestB = Dest;
3553 if (Dest == DestB) {
3554 CasesB.push_back(I.getCaseValue());
3557 return false; // More than two destinations.
3560 assert(DestA && DestB && "Single-destination switch should have been folded.");
3561 assert(DestA != DestB);
3562 assert(DestB != SI->getDefaultDest());
3563 assert(!CasesB.empty() && "There must be non-default cases.");
3564 assert(!CasesA.empty() || HasDefault);
3566 // Figure out if one of the sets of cases form a contiguous range.
3567 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
3568 BasicBlock *ContiguousDest = nullptr;
3569 BasicBlock *OtherDest = nullptr;
3570 if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
3571 ContiguousCases = &CasesA;
3572 ContiguousDest = DestA;
3574 } else if (CasesAreContiguous(CasesB)) {
3575 ContiguousCases = &CasesB;
3576 ContiguousDest = DestB;
3581 // Start building the compare and branch.
3583 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
3584 Constant *NumCases = ConstantInt::get(Offset->getType(), ContiguousCases->size());
3586 Value *Sub = SI->getCondition();
3587 if (!Offset->isNullValue())
3588 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
3591 // If NumCases overflowed, then all possible values jump to the successor.
3592 if (NumCases->isNullValue() && !ContiguousCases->empty())
3593 Cmp = ConstantInt::getTrue(SI->getContext());
3595 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3596 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
3598 // Update weight for the newly-created conditional branch.
3599 if (HasBranchWeights(SI)) {
3600 SmallVector<uint64_t, 8> Weights;
3601 GetBranchWeights(SI, Weights);
3602 if (Weights.size() == 1 + SI->getNumCases()) {
3603 uint64_t TrueWeight = 0;
3604 uint64_t FalseWeight = 0;
3605 for (size_t I = 0, E = Weights.size(); I != E; ++I) {
3606 if (SI->getSuccessor(I) == ContiguousDest)
3607 TrueWeight += Weights[I];
3609 FalseWeight += Weights[I];
3611 while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
3615 NewBI->setMetadata(LLVMContext::MD_prof,
3616 MDBuilder(SI->getContext()).createBranchWeights(
3617 (uint32_t)TrueWeight, (uint32_t)FalseWeight));
3621 // Prune obsolete incoming values off the successors' PHI nodes.
3622 for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
3623 unsigned PreviousEdges = ContiguousCases->size();
3624 if (ContiguousDest == SI->getDefaultDest()) ++PreviousEdges;
3625 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
3626 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3628 for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
3629 unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
3630 if (OtherDest == SI->getDefaultDest()) ++PreviousEdges;
3631 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
3632 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3636 SI->eraseFromParent();
3641 /// Compute masked bits for the condition of a switch
3642 /// and use it to remove dead cases.
3643 static bool EliminateDeadSwitchCases(SwitchInst *SI, AssumptionCache *AC,
3644 const DataLayout &DL) {
3645 Value *Cond = SI->getCondition();
3646 unsigned Bits = Cond->getType()->getIntegerBitWidth();
3647 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3648 computeKnownBits(Cond, KnownZero, KnownOne, DL, 0, AC, SI);
3650 // Gather dead cases.
3651 SmallVector<ConstantInt*, 8> DeadCases;
3652 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3653 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3654 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3655 DeadCases.push_back(I.getCaseValue());
3656 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3657 << I.getCaseValue() << "' is dead.\n");
3661 // If we can prove that the cases must cover all possible values, the
3662 // default destination becomes dead and we can remove it. If we know some
3663 // of the bits in the value, we can use that to more precisely compute the
3664 // number of possible unique case values.
3666 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
3667 const unsigned NumUnknownBits = Bits -
3668 (KnownZero.Or(KnownOne)).countPopulation();
3669 assert(NumUnknownBits <= Bits);
3670 if (HasDefault && DeadCases.empty() &&
3671 NumUnknownBits < 64 /* avoid overflow */ &&
3672 SI->getNumCases() == (1ULL << NumUnknownBits)) {
3673 DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
3674 BasicBlock *NewDefault = SplitBlockPredecessors(SI->getDefaultDest(),
3675 SI->getParent(), "");
3676 SI->setDefaultDest(&*NewDefault);
3677 SplitBlock(&*NewDefault, &NewDefault->front());
3678 auto *OldTI = NewDefault->getTerminator();
3679 new UnreachableInst(SI->getContext(), OldTI);
3680 EraseTerminatorInstAndDCECond(OldTI);
3684 SmallVector<uint64_t, 8> Weights;
3685 bool HasWeight = HasBranchWeights(SI);
3687 GetBranchWeights(SI, Weights);
3688 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3691 // Remove dead cases from the switch.
3692 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3693 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3694 assert(Case != SI->case_default() &&
3695 "Case was not found. Probably mistake in DeadCases forming.");
3697 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3701 // Prune unused values from PHI nodes.
3702 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3703 SI->removeCase(Case);
3705 if (HasWeight && Weights.size() >= 2) {
3706 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3707 SI->setMetadata(LLVMContext::MD_prof,
3708 MDBuilder(SI->getParent()->getContext()).
3709 createBranchWeights(MDWeights));
3712 return !DeadCases.empty();
3715 /// If BB would be eligible for simplification by
3716 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3717 /// by an unconditional branch), look at the phi node for BB in the successor
3718 /// block and see if the incoming value is equal to CaseValue. If so, return
3719 /// the phi node, and set PhiIndex to BB's index in the phi node.
3720 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3723 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3724 return nullptr; // BB must be empty to be a candidate for simplification.
3725 if (!BB->getSinglePredecessor())
3726 return nullptr; // BB must be dominated by the switch.
3728 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3729 if (!Branch || !Branch->isUnconditional())
3730 return nullptr; // Terminator must be unconditional branch.
3732 BasicBlock *Succ = Branch->getSuccessor(0);
3734 BasicBlock::iterator I = Succ->begin();
3735 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3736 int Idx = PHI->getBasicBlockIndex(BB);
3737 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3739 Value *InValue = PHI->getIncomingValue(Idx);
3740 if (InValue != CaseValue) continue;
3749 /// Try to forward the condition of a switch instruction to a phi node
3750 /// dominated by the switch, if that would mean that some of the destination
3751 /// blocks of the switch can be folded away.
3752 /// Returns true if a change is made.
3753 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3754 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3755 ForwardingNodesMap ForwardingNodes;
3757 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3758 ConstantInt *CaseValue = I.getCaseValue();
3759 BasicBlock *CaseDest = I.getCaseSuccessor();
3762 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3766 ForwardingNodes[PHI].push_back(PhiIndex);
3769 bool Changed = false;
3771 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3772 E = ForwardingNodes.end(); I != E; ++I) {
3773 PHINode *Phi = I->first;
3774 SmallVectorImpl<int> &Indexes = I->second;
3776 if (Indexes.size() < 2) continue;
3778 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3779 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3786 /// Return true if the backend will be able to handle
3787 /// initializing an array of constants like C.
3788 static bool ValidLookupTableConstant(Constant *C) {
3789 if (C->isThreadDependent())
3791 if (C->isDLLImportDependent())
3794 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3795 return CE->isGEPWithNoNotionalOverIndexing();
3797 return isa<ConstantFP>(C) ||
3798 isa<ConstantInt>(C) ||
3799 isa<ConstantPointerNull>(C) ||
3800 isa<GlobalValue>(C) ||
3804 /// If V is a Constant, return it. Otherwise, try to look up
3805 /// its constant value in ConstantPool, returning 0 if it's not there.
3806 static Constant *LookupConstant(Value *V,
3807 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3808 if (Constant *C = dyn_cast<Constant>(V))
3810 return ConstantPool.lookup(V);
3813 /// Try to fold instruction I into a constant. This works for
3814 /// simple instructions such as binary operations where both operands are
3815 /// constant or can be replaced by constants from the ConstantPool. Returns the
3816 /// resulting constant on success, 0 otherwise.
3818 ConstantFold(Instruction *I, const DataLayout &DL,
3819 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
3820 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3821 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3824 if (A->isAllOnesValue())
3825 return LookupConstant(Select->getTrueValue(), ConstantPool);
3826 if (A->isNullValue())
3827 return LookupConstant(Select->getFalseValue(), ConstantPool);
3831 SmallVector<Constant *, 4> COps;
3832 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
3833 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
3839 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
3840 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
3844 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), COps, DL);
3847 /// Try to determine the resulting constant values in phi nodes
3848 /// at the common destination basic block, *CommonDest, for one of the case
3849 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3850 /// case), of a switch instruction SI.
3852 GetCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest,
3853 BasicBlock **CommonDest,
3854 SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
3855 const DataLayout &DL) {
3856 // The block from which we enter the common destination.
3857 BasicBlock *Pred = SI->getParent();
3859 // If CaseDest is empty except for some side-effect free instructions through
3860 // which we can constant-propagate the CaseVal, continue to its successor.
3861 SmallDenseMap<Value*, Constant*> ConstantPool;
3862 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3863 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3865 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3866 // If the terminator is a simple branch, continue to the next block.
3867 if (T->getNumSuccessors() != 1)
3870 CaseDest = T->getSuccessor(0);
3871 } else if (isa<DbgInfoIntrinsic>(I)) {
3872 // Skip debug intrinsic.
3874 } else if (Constant *C = ConstantFold(&*I, DL, ConstantPool)) {
3875 // Instruction is side-effect free and constant.
3877 // If the instruction has uses outside this block or a phi node slot for
3878 // the block, it is not safe to bypass the instruction since it would then
3879 // no longer dominate all its uses.
3880 for (auto &Use : I->uses()) {
3881 User *User = Use.getUser();
3882 if (Instruction *I = dyn_cast<Instruction>(User))
3883 if (I->getParent() == CaseDest)
3885 if (PHINode *Phi = dyn_cast<PHINode>(User))
3886 if (Phi->getIncomingBlock(Use) == CaseDest)
3891 ConstantPool.insert(std::make_pair(&*I, C));
3897 // If we did not have a CommonDest before, use the current one.
3899 *CommonDest = CaseDest;
3900 // If the destination isn't the common one, abort.
3901 if (CaseDest != *CommonDest)
3904 // Get the values for this case from phi nodes in the destination block.
3905 BasicBlock::iterator I = (*CommonDest)->begin();
3906 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3907 int Idx = PHI->getBasicBlockIndex(Pred);
3911 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3916 // Be conservative about which kinds of constants we support.
3917 if (!ValidLookupTableConstant(ConstVal))
3920 Res.push_back(std::make_pair(PHI, ConstVal));
3923 return Res.size() > 0;
3926 // Helper function used to add CaseVal to the list of cases that generate
3928 static void MapCaseToResult(ConstantInt *CaseVal,
3929 SwitchCaseResultVectorTy &UniqueResults,
3931 for (auto &I : UniqueResults) {
3932 if (I.first == Result) {
3933 I.second.push_back(CaseVal);
3937 UniqueResults.push_back(std::make_pair(Result,
3938 SmallVector<ConstantInt*, 4>(1, CaseVal)));
3941 // Helper function that initializes a map containing
3942 // results for the PHI node of the common destination block for a switch
3943 // instruction. Returns false if multiple PHI nodes have been found or if
3944 // there is not a common destination block for the switch.
3945 static bool InitializeUniqueCases(SwitchInst *SI, PHINode *&PHI,
3946 BasicBlock *&CommonDest,
3947 SwitchCaseResultVectorTy &UniqueResults,
3948 Constant *&DefaultResult,
3949 const DataLayout &DL) {
3950 for (auto &I : SI->cases()) {
3951 ConstantInt *CaseVal = I.getCaseValue();
3953 // Resulting value at phi nodes for this case value.
3954 SwitchCaseResultsTy Results;
3955 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
3959 // Only one value per case is permitted
3960 if (Results.size() > 1)
3962 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
3964 // Check the PHI consistency.
3966 PHI = Results[0].first;
3967 else if (PHI != Results[0].first)
3970 // Find the default result value.
3971 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
3972 BasicBlock *DefaultDest = SI->getDefaultDest();
3973 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
3975 // If the default value is not found abort unless the default destination
3978 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
3979 if ((!DefaultResult &&
3980 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
3986 // Helper function that checks if it is possible to transform a switch with only
3987 // two cases (or two cases + default) that produces a result into a select.
3990 // case 10: %0 = icmp eq i32 %a, 10
3991 // return 10; %1 = select i1 %0, i32 10, i32 4
3992 // case 20: ----> %2 = icmp eq i32 %a, 20
3993 // return 2; %3 = select i1 %2, i32 2, i32 %1
3998 ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
3999 Constant *DefaultResult, Value *Condition,
4000 IRBuilder<> &Builder) {
4001 assert(ResultVector.size() == 2 &&
4002 "We should have exactly two unique results at this point");
4003 // If we are selecting between only two cases transform into a simple
4004 // select or a two-way select if default is possible.
4005 if (ResultVector[0].second.size() == 1 &&
4006 ResultVector[1].second.size() == 1) {
4007 ConstantInt *const FirstCase = ResultVector[0].second[0];
4008 ConstantInt *const SecondCase = ResultVector[1].second[0];
4010 bool DefaultCanTrigger = DefaultResult;
4011 Value *SelectValue = ResultVector[1].first;
4012 if (DefaultCanTrigger) {
4013 Value *const ValueCompare =
4014 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4015 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4016 DefaultResult, "switch.select");
4018 Value *const ValueCompare =
4019 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4020 return Builder.CreateSelect(ValueCompare, ResultVector[0].first, SelectValue,
4027 // Helper function to cleanup a switch instruction that has been converted into
4028 // a select, fixing up PHI nodes and basic blocks.
4029 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
4031 IRBuilder<> &Builder) {
4032 BasicBlock *SelectBB = SI->getParent();
4033 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4034 PHI->removeIncomingValue(SelectBB);
4035 PHI->addIncoming(SelectValue, SelectBB);
4037 Builder.CreateBr(PHI->getParent());
4039 // Remove the switch.
4040 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4041 BasicBlock *Succ = SI->getSuccessor(i);
4043 if (Succ == PHI->getParent())
4045 Succ->removePredecessor(SelectBB);
4047 SI->eraseFromParent();
4050 /// If the switch is only used to initialize one or more
4051 /// phi nodes in a common successor block with only two different
4052 /// constant values, replace the switch with select.
4053 static bool SwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4054 AssumptionCache *AC, const DataLayout &DL) {
4055 Value *const Cond = SI->getCondition();
4056 PHINode *PHI = nullptr;
4057 BasicBlock *CommonDest = nullptr;
4058 Constant *DefaultResult;
4059 SwitchCaseResultVectorTy UniqueResults;
4060 // Collect all the cases that will deliver the same value from the switch.
4061 if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4064 // Selects choose between maximum two values.
4065 if (UniqueResults.size() != 2)
4067 assert(PHI != nullptr && "PHI for value select not found");
4069 Builder.SetInsertPoint(SI);
4070 Value *SelectValue = ConvertTwoCaseSwitch(
4072 DefaultResult, Cond, Builder);
4074 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4077 // The switch couldn't be converted into a select.
4082 /// This class represents a lookup table that can be used to replace a switch.
4083 class SwitchLookupTable {
4085 /// Create a lookup table to use as a switch replacement with the contents
4086 /// of Values, using DefaultValue to fill any holes in the table.
4088 Module &M, uint64_t TableSize, ConstantInt *Offset,
4089 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4090 Constant *DefaultValue, const DataLayout &DL);
4092 /// Build instructions with Builder to retrieve the value at
4093 /// the position given by Index in the lookup table.
4094 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4096 /// Return true if a table with TableSize elements of
4097 /// type ElementType would fit in a target-legal register.
4098 static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4102 // Depending on the contents of the table, it can be represented in
4105 // For tables where each element contains the same value, we just have to
4106 // store that single value and return it for each lookup.
4109 // For tables where there is a linear relationship between table index
4110 // and values. We calculate the result with a simple multiplication
4111 // and addition instead of a table lookup.
4114 // For small tables with integer elements, we can pack them into a bitmap
4115 // that fits into a target-legal register. Values are retrieved by
4116 // shift and mask operations.
4119 // The table is stored as an array of values. Values are retrieved by load
4120 // instructions from the table.
4124 // For SingleValueKind, this is the single value.
4125 Constant *SingleValue;
4127 // For BitMapKind, this is the bitmap.
4128 ConstantInt *BitMap;
4129 IntegerType *BitMapElementTy;
4131 // For LinearMapKind, these are the constants used to derive the value.
4132 ConstantInt *LinearOffset;
4133 ConstantInt *LinearMultiplier;
4135 // For ArrayKind, this is the array.
4136 GlobalVariable *Array;
4140 SwitchLookupTable::SwitchLookupTable(
4141 Module &M, uint64_t TableSize, ConstantInt *Offset,
4142 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4143 Constant *DefaultValue, const DataLayout &DL)
4144 : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
4145 LinearOffset(nullptr), LinearMultiplier(nullptr), Array(nullptr) {
4146 assert(Values.size() && "Can't build lookup table without values!");
4147 assert(TableSize >= Values.size() && "Can't fit values in table!");
4149 // If all values in the table are equal, this is that value.
4150 SingleValue = Values.begin()->second;
4152 Type *ValueType = Values.begin()->second->getType();
4154 // Build up the table contents.
4155 SmallVector<Constant*, 64> TableContents(TableSize);
4156 for (size_t I = 0, E = Values.size(); I != E; ++I) {
4157 ConstantInt *CaseVal = Values[I].first;
4158 Constant *CaseRes = Values[I].second;
4159 assert(CaseRes->getType() == ValueType);
4161 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
4163 TableContents[Idx] = CaseRes;
4165 if (CaseRes != SingleValue)
4166 SingleValue = nullptr;
4169 // Fill in any holes in the table with the default result.
4170 if (Values.size() < TableSize) {
4171 assert(DefaultValue &&
4172 "Need a default value to fill the lookup table holes.");
4173 assert(DefaultValue->getType() == ValueType);
4174 for (uint64_t I = 0; I < TableSize; ++I) {
4175 if (!TableContents[I])
4176 TableContents[I] = DefaultValue;
4179 if (DefaultValue != SingleValue)
4180 SingleValue = nullptr;
4183 // If each element in the table contains the same value, we only need to store
4184 // that single value.
4186 Kind = SingleValueKind;
4190 // Check if we can derive the value with a linear transformation from the
4192 if (isa<IntegerType>(ValueType)) {
4193 bool LinearMappingPossible = true;
4196 assert(TableSize >= 2 && "Should be a SingleValue table.");
4197 // Check if there is the same distance between two consecutive values.
4198 for (uint64_t I = 0; I < TableSize; ++I) {
4199 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4201 // This is an undef. We could deal with it, but undefs in lookup tables
4202 // are very seldom. It's probably not worth the additional complexity.
4203 LinearMappingPossible = false;
4206 APInt Val = ConstVal->getValue();
4208 APInt Dist = Val - PrevVal;
4211 } else if (Dist != DistToPrev) {
4212 LinearMappingPossible = false;
4218 if (LinearMappingPossible) {
4219 LinearOffset = cast<ConstantInt>(TableContents[0]);
4220 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
4221 Kind = LinearMapKind;
4227 // If the type is integer and the table fits in a register, build a bitmap.
4228 if (WouldFitInRegister(DL, TableSize, ValueType)) {
4229 IntegerType *IT = cast<IntegerType>(ValueType);
4230 APInt TableInt(TableSize * IT->getBitWidth(), 0);
4231 for (uint64_t I = TableSize; I > 0; --I) {
4232 TableInt <<= IT->getBitWidth();
4233 // Insert values into the bitmap. Undef values are set to zero.
4234 if (!isa<UndefValue>(TableContents[I - 1])) {
4235 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
4236 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
4239 BitMap = ConstantInt::get(M.getContext(), TableInt);
4240 BitMapElementTy = IT;
4246 // Store the table in an array.
4247 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
4248 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
4250 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
4251 GlobalVariable::PrivateLinkage,
4254 Array->setUnnamedAddr(true);
4258 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
4260 case SingleValueKind:
4262 case LinearMapKind: {
4263 // Derive the result value from the input value.
4264 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
4265 false, "switch.idx.cast");
4266 if (!LinearMultiplier->isOne())
4267 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
4268 if (!LinearOffset->isZero())
4269 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
4273 // Type of the bitmap (e.g. i59).
4274 IntegerType *MapTy = BitMap->getType();
4276 // Cast Index to the same type as the bitmap.
4277 // Note: The Index is <= the number of elements in the table, so
4278 // truncating it to the width of the bitmask is safe.
4279 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
4281 // Multiply the shift amount by the element width.
4282 ShiftAmt = Builder.CreateMul(ShiftAmt,
4283 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
4287 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
4288 "switch.downshift");
4290 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
4294 // Make sure the table index will not overflow when treated as signed.
4295 IntegerType *IT = cast<IntegerType>(Index->getType());
4296 uint64_t TableSize = Array->getInitializer()->getType()
4297 ->getArrayNumElements();
4298 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
4299 Index = Builder.CreateZExt(Index,
4300 IntegerType::get(IT->getContext(),
4301 IT->getBitWidth() + 1),
4302 "switch.tableidx.zext");
4304 Value *GEPIndices[] = { Builder.getInt32(0), Index };
4305 Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
4306 GEPIndices, "switch.gep");
4307 return Builder.CreateLoad(GEP, "switch.load");
4310 llvm_unreachable("Unknown lookup table kind!");
4313 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
4315 Type *ElementType) {
4316 auto *IT = dyn_cast<IntegerType>(ElementType);
4319 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
4320 // are <= 15, we could try to narrow the type.
4322 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
4323 if (TableSize >= UINT_MAX/IT->getBitWidth())
4325 return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
4328 /// Determine whether a lookup table should be built for this switch, based on
4329 /// the number of cases, size of the table, and the types of the results.
4331 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
4332 const TargetTransformInfo &TTI, const DataLayout &DL,
4333 const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
4334 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
4335 return false; // TableSize overflowed, or mul below might overflow.
4337 bool AllTablesFitInRegister = true;
4338 bool HasIllegalType = false;
4339 for (const auto &I : ResultTypes) {
4340 Type *Ty = I.second;
4342 // Saturate this flag to true.
4343 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
4345 // Saturate this flag to false.
4346 AllTablesFitInRegister = AllTablesFitInRegister &&
4347 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
4349 // If both flags saturate, we're done. NOTE: This *only* works with
4350 // saturating flags, and all flags have to saturate first due to the
4351 // non-deterministic behavior of iterating over a dense map.
4352 if (HasIllegalType && !AllTablesFitInRegister)
4356 // If each table would fit in a register, we should build it anyway.
4357 if (AllTablesFitInRegister)
4360 // Don't build a table that doesn't fit in-register if it has illegal types.
4364 // The table density should be at least 40%. This is the same criterion as for
4365 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
4366 // FIXME: Find the best cut-off.
4367 return SI->getNumCases() * 10 >= TableSize * 4;
4370 /// Try to reuse the switch table index compare. Following pattern:
4372 /// if (idx < tablesize)
4373 /// r = table[idx]; // table does not contain default_value
4375 /// r = default_value;
4376 /// if (r != default_value)
4379 /// Is optimized to:
4381 /// cond = idx < tablesize;
4385 /// r = default_value;
4389 /// Jump threading will then eliminate the second if(cond).
4390 static void reuseTableCompare(User *PhiUser, BasicBlock *PhiBlock,
4391 BranchInst *RangeCheckBranch, Constant *DefaultValue,
4392 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values) {
4394 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
4398 // We require that the compare is in the same block as the phi so that jump
4399 // threading can do its work afterwards.
4400 if (CmpInst->getParent() != PhiBlock)
4403 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
4407 Value *RangeCmp = RangeCheckBranch->getCondition();
4408 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
4409 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
4411 // Check if the compare with the default value is constant true or false.
4412 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
4413 DefaultValue, CmpOp1, true);
4414 if (DefaultConst != TrueConst && DefaultConst != FalseConst)
4417 // Check if the compare with the case values is distinct from the default
4419 for (auto ValuePair : Values) {
4420 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
4421 ValuePair.second, CmpOp1, true);
4422 if (!CaseConst || CaseConst == DefaultConst)
4424 assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
4425 "Expect true or false as compare result.");
4428 // Check if the branch instruction dominates the phi node. It's a simple
4429 // dominance check, but sufficient for our needs.
4430 // Although this check is invariant in the calling loops, it's better to do it
4431 // at this late stage. Practically we do it at most once for a switch.
4432 BasicBlock *BranchBlock = RangeCheckBranch->getParent();
4433 for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
4434 BasicBlock *Pred = *PI;
4435 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
4439 if (DefaultConst == FalseConst) {
4440 // The compare yields the same result. We can replace it.
4441 CmpInst->replaceAllUsesWith(RangeCmp);
4442 ++NumTableCmpReuses;
4444 // The compare yields the same result, just inverted. We can replace it.
4445 Value *InvertedTableCmp = BinaryOperator::CreateXor(RangeCmp,
4446 ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
4448 CmpInst->replaceAllUsesWith(InvertedTableCmp);
4449 ++NumTableCmpReuses;
4453 /// If the switch is only used to initialize one or more phi nodes in a common
4454 /// successor block with different constant values, replace the switch with
4456 static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
4457 const DataLayout &DL,
4458 const TargetTransformInfo &TTI) {
4459 assert(SI->getNumCases() > 1 && "Degenerate switch?");
4461 // Only build lookup table when we have a target that supports it.
4462 if (!TTI.shouldBuildLookupTables())
4465 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
4466 // split off a dense part and build a lookup table for that.
4468 // FIXME: This creates arrays of GEPs to constant strings, which means each
4469 // GEP needs a runtime relocation in PIC code. We should just build one big
4470 // string and lookup indices into that.
4472 // Ignore switches with less than three cases. Lookup tables will not make them
4473 // faster, so we don't analyze them.
4474 if (SI->getNumCases() < 3)
4477 // Figure out the corresponding result for each case value and phi node in the
4478 // common destination, as well as the min and max case values.
4479 assert(SI->case_begin() != SI->case_end());
4480 SwitchInst::CaseIt CI = SI->case_begin();
4481 ConstantInt *MinCaseVal = CI.getCaseValue();
4482 ConstantInt *MaxCaseVal = CI.getCaseValue();
4484 BasicBlock *CommonDest = nullptr;
4485 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
4486 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
4487 SmallDenseMap<PHINode*, Constant*> DefaultResults;
4488 SmallDenseMap<PHINode*, Type*> ResultTypes;
4489 SmallVector<PHINode*, 4> PHIs;
4491 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
4492 ConstantInt *CaseVal = CI.getCaseValue();
4493 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
4494 MinCaseVal = CaseVal;
4495 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
4496 MaxCaseVal = CaseVal;
4498 // Resulting value at phi nodes for this case value.
4499 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
4501 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
4505 // Append the result from this case to the list for each phi.
4506 for (const auto &I : Results) {
4507 PHINode *PHI = I.first;
4508 Constant *Value = I.second;
4509 if (!ResultLists.count(PHI))
4510 PHIs.push_back(PHI);
4511 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
4515 // Keep track of the result types.
4516 for (PHINode *PHI : PHIs) {
4517 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
4520 uint64_t NumResults = ResultLists[PHIs[0]].size();
4521 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
4522 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
4523 bool TableHasHoles = (NumResults < TableSize);
4525 // If the table has holes, we need a constant result for the default case
4526 // or a bitmask that fits in a register.
4527 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
4528 bool HasDefaultResults = GetCaseResults(SI, nullptr, SI->getDefaultDest(),
4529 &CommonDest, DefaultResultsList, DL);
4531 bool NeedMask = (TableHasHoles && !HasDefaultResults);
4533 // As an extra penalty for the validity test we require more cases.
4534 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
4536 if (!DL.fitsInLegalInteger(TableSize))
4540 for (const auto &I : DefaultResultsList) {
4541 PHINode *PHI = I.first;
4542 Constant *Result = I.second;
4543 DefaultResults[PHI] = Result;
4546 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
4549 // Create the BB that does the lookups.
4550 Module &Mod = *CommonDest->getParent()->getParent();
4551 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
4553 CommonDest->getParent(),
4556 // Compute the table index value.
4557 Builder.SetInsertPoint(SI);
4558 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
4561 // Compute the maximum table size representable by the integer type we are
4563 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
4564 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
4565 assert(MaxTableSize >= TableSize &&
4566 "It is impossible for a switch to have more entries than the max "
4567 "representable value of its input integer type's size.");
4569 // If the default destination is unreachable, or if the lookup table covers
4570 // all values of the conditional variable, branch directly to the lookup table
4571 // BB. Otherwise, check that the condition is within the case range.
4572 const bool DefaultIsReachable =
4573 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4574 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
4575 BranchInst *RangeCheckBranch = nullptr;
4577 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
4578 Builder.CreateBr(LookupBB);
4579 // Note: We call removeProdecessor later since we need to be able to get the
4580 // PHI value for the default case in case we're using a bit mask.
4582 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
4583 MinCaseVal->getType(), TableSize));
4584 RangeCheckBranch = Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
4587 // Populate the BB that does the lookups.
4588 Builder.SetInsertPoint(LookupBB);
4591 // Before doing the lookup we do the hole check.
4592 // The LookupBB is therefore re-purposed to do the hole check
4593 // and we create a new LookupBB.
4594 BasicBlock *MaskBB = LookupBB;
4595 MaskBB->setName("switch.hole_check");
4596 LookupBB = BasicBlock::Create(Mod.getContext(),
4598 CommonDest->getParent(),
4601 // Make the mask's bitwidth at least 8bit and a power-of-2 to avoid
4602 // unnecessary illegal types.
4603 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
4604 APInt MaskInt(TableSizePowOf2, 0);
4605 APInt One(TableSizePowOf2, 1);
4606 // Build bitmask; fill in a 1 bit for every case.
4607 const ResultListTy &ResultList = ResultLists[PHIs[0]];
4608 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
4609 uint64_t Idx = (ResultList[I].first->getValue() -
4610 MinCaseVal->getValue()).getLimitedValue();
4611 MaskInt |= One << Idx;
4613 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
4615 // Get the TableIndex'th bit of the bitmask.
4616 // If this bit is 0 (meaning hole) jump to the default destination,
4617 // else continue with table lookup.
4618 IntegerType *MapTy = TableMask->getType();
4619 Value *MaskIndex = Builder.CreateZExtOrTrunc(TableIndex, MapTy,
4620 "switch.maskindex");
4621 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex,
4623 Value *LoBit = Builder.CreateTrunc(Shifted,
4624 Type::getInt1Ty(Mod.getContext()),
4626 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
4628 Builder.SetInsertPoint(LookupBB);
4629 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
4632 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
4633 // We cached PHINodes in PHIs, to avoid accessing deleted PHINodes later,
4634 // do not delete PHINodes here.
4635 SI->getDefaultDest()->removePredecessor(SI->getParent(),
4636 /*DontDeleteUselessPHIs=*/true);
4639 bool ReturnedEarly = false;
4640 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
4641 PHINode *PHI = PHIs[I];
4642 const ResultListTy &ResultList = ResultLists[PHI];
4644 // If using a bitmask, use any value to fill the lookup table holes.
4645 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
4646 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL);
4648 Value *Result = Table.BuildLookup(TableIndex, Builder);
4650 // If the result is used to return immediately from the function, we want to
4651 // do that right here.
4652 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
4653 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
4654 Builder.CreateRet(Result);
4655 ReturnedEarly = true;
4659 // Do a small peephole optimization: re-use the switch table compare if
4661 if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
4662 BasicBlock *PhiBlock = PHI->getParent();
4663 // Search for compare instructions which use the phi.
4664 for (auto *User : PHI->users()) {
4665 reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
4669 PHI->addIncoming(Result, LookupBB);
4673 Builder.CreateBr(CommonDest);
4675 // Remove the switch.
4676 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4677 BasicBlock *Succ = SI->getSuccessor(i);
4679 if (Succ == SI->getDefaultDest())
4681 Succ->removePredecessor(SI->getParent());
4683 SI->eraseFromParent();
4687 ++NumLookupTablesHoles;
4691 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
4692 BasicBlock *BB = SI->getParent();
4694 if (isValueEqualityComparison(SI)) {
4695 // If we only have one predecessor, and if it is a branch on this value,
4696 // see if that predecessor totally determines the outcome of this switch.
4697 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4698 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
4699 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4701 Value *Cond = SI->getCondition();
4702 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
4703 if (SimplifySwitchOnSelect(SI, Select))
4704 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4706 // If the block only contains the switch, see if we can fold the block
4707 // away into any preds.
4708 BasicBlock::iterator BBI = BB->begin();
4709 // Ignore dbg intrinsics.
4710 while (isa<DbgInfoIntrinsic>(BBI))
4713 if (FoldValueComparisonIntoPredecessors(SI, Builder))
4714 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4717 // Try to transform the switch into an icmp and a branch.
4718 if (TurnSwitchRangeIntoICmp(SI, Builder))
4719 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4721 // Remove unreachable cases.
4722 if (EliminateDeadSwitchCases(SI, AC, DL))
4723 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4725 if (SwitchToSelect(SI, Builder, AC, DL))
4726 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4728 if (ForwardSwitchConditionToPHI(SI))
4729 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4731 if (SwitchToLookupTable(SI, Builder, DL, TTI))
4732 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4737 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
4738 BasicBlock *BB = IBI->getParent();
4739 bool Changed = false;
4741 // Eliminate redundant destinations.
4742 SmallPtrSet<Value *, 8> Succs;
4743 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
4744 BasicBlock *Dest = IBI->getDestination(i);
4745 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
4746 Dest->removePredecessor(BB);
4747 IBI->removeDestination(i);
4753 if (IBI->getNumDestinations() == 0) {
4754 // If the indirectbr has no successors, change it to unreachable.
4755 new UnreachableInst(IBI->getContext(), IBI);
4756 EraseTerminatorInstAndDCECond(IBI);
4760 if (IBI->getNumDestinations() == 1) {
4761 // If the indirectbr has one successor, change it to a direct branch.
4762 BranchInst::Create(IBI->getDestination(0), IBI);
4763 EraseTerminatorInstAndDCECond(IBI);
4767 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
4768 if (SimplifyIndirectBrOnSelect(IBI, SI))
4769 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4774 /// Given an block with only a single landing pad and a unconditional branch
4775 /// try to find another basic block which this one can be merged with. This
4776 /// handles cases where we have multiple invokes with unique landing pads, but
4777 /// a shared handler.
4779 /// We specifically choose to not worry about merging non-empty blocks
4780 /// here. That is a PRE/scheduling problem and is best solved elsewhere. In
4781 /// practice, the optimizer produces empty landing pad blocks quite frequently
4782 /// when dealing with exception dense code. (see: instcombine, gvn, if-else
4783 /// sinking in this file)
4785 /// This is primarily a code size optimization. We need to avoid performing
4786 /// any transform which might inhibit optimization (such as our ability to
4787 /// specialize a particular handler via tail commoning). We do this by not
4788 /// merging any blocks which require us to introduce a phi. Since the same
4789 /// values are flowing through both blocks, we don't loose any ability to
4790 /// specialize. If anything, we make such specialization more likely.
4792 /// TODO - This transformation could remove entries from a phi in the target
4793 /// block when the inputs in the phi are the same for the two blocks being
4794 /// merged. In some cases, this could result in removal of the PHI entirely.
4795 static bool TryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI,
4797 auto Succ = BB->getUniqueSuccessor();
4799 // If there's a phi in the successor block, we'd likely have to introduce
4800 // a phi into the merged landing pad block.
4801 if (isa<PHINode>(*Succ->begin()))
4804 for (BasicBlock *OtherPred : predecessors(Succ)) {
4805 if (BB == OtherPred)
4807 BasicBlock::iterator I = OtherPred->begin();
4808 LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(I);
4809 if (!LPad2 || !LPad2->isIdenticalTo(LPad))
4811 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {}
4812 BranchInst *BI2 = dyn_cast<BranchInst>(I);
4813 if (!BI2 || !BI2->isIdenticalTo(BI))
4816 // We've found an identical block. Update our predeccessors to take that
4817 // path instead and make ourselves dead.
4818 SmallSet<BasicBlock *, 16> Preds;
4819 Preds.insert(pred_begin(BB), pred_end(BB));
4820 for (BasicBlock *Pred : Preds) {
4821 InvokeInst *II = cast<InvokeInst>(Pred->getTerminator());
4822 assert(II->getNormalDest() != BB &&
4823 II->getUnwindDest() == BB && "unexpected successor");
4824 II->setUnwindDest(OtherPred);
4827 // The debug info in OtherPred doesn't cover the merged control flow that
4828 // used to go through BB. We need to delete it or update it.
4829 for (auto I = OtherPred->begin(), E = OtherPred->end();
4831 Instruction &Inst = *I; I++;
4832 if (isa<DbgInfoIntrinsic>(Inst))
4833 Inst.eraseFromParent();
4836 SmallSet<BasicBlock *, 16> Succs;
4837 Succs.insert(succ_begin(BB), succ_end(BB));
4838 for (BasicBlock *Succ : Succs) {
4839 Succ->removePredecessor(BB);
4842 IRBuilder<> Builder(BI);
4843 Builder.CreateUnreachable();
4844 BI->eraseFromParent();
4850 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
4851 BasicBlock *BB = BI->getParent();
4853 if (SinkCommon && SinkThenElseCodeToEnd(BI))
4856 // If the Terminator is the only non-phi instruction, simplify the block.
4857 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg()->getIterator();
4858 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
4859 TryToSimplifyUncondBranchFromEmptyBlock(BB))
4862 // If the only instruction in the block is a seteq/setne comparison
4863 // against a constant, try to simplify the block.
4864 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
4865 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
4866 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
4868 if (I->isTerminator() &&
4869 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, DL, TTI,
4870 BonusInstThreshold, AC))
4874 // See if we can merge an empty landing pad block with another which is
4876 if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(I)) {
4877 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {}
4878 if (I->isTerminator() &&
4879 TryToMergeLandingPad(LPad, BI, BB))
4883 // If this basic block is ONLY a compare and a branch, and if a predecessor
4884 // branches to us and our successor, fold the comparison into the
4885 // predecessor and use logical operations to update the incoming value
4886 // for PHI nodes in common successor.
4887 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
4888 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4892 static BasicBlock *allPredecessorsComeFromSameSource(BasicBlock *BB) {
4893 BasicBlock *PredPred = nullptr;
4894 for (auto *P : predecessors(BB)) {
4895 BasicBlock *PPred = P->getSinglePredecessor();
4896 if (!PPred || (PredPred && PredPred != PPred))
4903 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
4904 BasicBlock *BB = BI->getParent();
4906 // Conditional branch
4907 if (isValueEqualityComparison(BI)) {
4908 // If we only have one predecessor, and if it is a branch on this value,
4909 // see if that predecessor totally determines the outcome of this
4911 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4912 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
4913 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4915 // This block must be empty, except for the setcond inst, if it exists.
4916 // Ignore dbg intrinsics.
4917 BasicBlock::iterator I = BB->begin();
4918 // Ignore dbg intrinsics.
4919 while (isa<DbgInfoIntrinsic>(I))
4922 if (FoldValueComparisonIntoPredecessors(BI, Builder))
4923 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4924 } else if (&*I == cast<Instruction>(BI->getCondition())){
4926 // Ignore dbg intrinsics.
4927 while (isa<DbgInfoIntrinsic>(I))
4929 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
4930 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4934 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
4935 if (SimplifyBranchOnICmpChain(BI, Builder, DL))
4938 // If this basic block is ONLY a compare and a branch, and if a predecessor
4939 // branches to us and one of our successors, fold the comparison into the
4940 // predecessor and use logical operations to pick the right destination.
4941 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
4942 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4944 // We have a conditional branch to two blocks that are only reachable
4945 // from BI. We know that the condbr dominates the two blocks, so see if
4946 // there is any identical code in the "then" and "else" blocks. If so, we
4947 // can hoist it up to the branching block.
4948 if (BI->getSuccessor(0)->getSinglePredecessor()) {
4949 if (BI->getSuccessor(1)->getSinglePredecessor()) {
4950 if (HoistThenElseCodeToIf(BI, TTI))
4951 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4953 // If Successor #1 has multiple preds, we may be able to conditionally
4954 // execute Successor #0 if it branches to Successor #1.
4955 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
4956 if (Succ0TI->getNumSuccessors() == 1 &&
4957 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
4958 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), TTI))
4959 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4961 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
4962 // If Successor #0 has multiple preds, we may be able to conditionally
4963 // execute Successor #1 if it branches to Successor #0.
4964 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
4965 if (Succ1TI->getNumSuccessors() == 1 &&
4966 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
4967 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), TTI))
4968 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4971 // If this is a branch on a phi node in the current block, thread control
4972 // through this block if any PHI node entries are constants.
4973 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
4974 if (PN->getParent() == BI->getParent())
4975 if (FoldCondBranchOnPHI(BI, DL))
4976 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4978 // Scan predecessor blocks for conditional branches.
4979 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
4980 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
4981 if (PBI != BI && PBI->isConditional())
4982 if (SimplifyCondBranchToCondBranch(PBI, BI, DL))
4983 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4985 // Look for diamond patterns.
4986 if (MergeCondStores)
4987 if (BasicBlock *PrevBB = allPredecessorsComeFromSameSource(BB))
4988 if (BranchInst *PBI = dyn_cast<BranchInst>(PrevBB->getTerminator()))
4989 if (PBI != BI && PBI->isConditional())
4990 if (mergeConditionalStores(PBI, BI))
4991 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4996 /// Check if passing a value to an instruction will cause undefined behavior.
4997 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
4998 Constant *C = dyn_cast<Constant>(V);
5005 if (C->isNullValue()) {
5006 // Only look at the first use, avoid hurting compile time with long uselists
5007 User *Use = *I->user_begin();
5009 // Now make sure that there are no instructions in between that can alter
5010 // control flow (eg. calls)
5011 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
5012 if (i == I->getParent()->end() || i->mayHaveSideEffects())
5015 // Look through GEPs. A load from a GEP derived from NULL is still undefined
5016 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
5017 if (GEP->getPointerOperand() == I)
5018 return passingValueIsAlwaysUndefined(V, GEP);
5020 // Look through bitcasts.
5021 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
5022 return passingValueIsAlwaysUndefined(V, BC);
5024 // Load from null is undefined.
5025 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
5026 if (!LI->isVolatile())
5027 return LI->getPointerAddressSpace() == 0;
5029 // Store to null is undefined.
5030 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
5031 if (!SI->isVolatile())
5032 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
5037 /// If BB has an incoming value that will always trigger undefined behavior
5038 /// (eg. null pointer dereference), remove the branch leading here.
5039 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
5040 for (BasicBlock::iterator i = BB->begin();
5041 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
5042 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
5043 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
5044 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
5045 IRBuilder<> Builder(T);
5046 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
5047 BB->removePredecessor(PHI->getIncomingBlock(i));
5048 // Turn uncoditional branches into unreachables and remove the dead
5049 // destination from conditional branches.
5050 if (BI->isUnconditional())
5051 Builder.CreateUnreachable();
5053 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
5054 BI->getSuccessor(0));
5055 BI->eraseFromParent();
5058 // TODO: SwitchInst.
5064 bool SimplifyCFGOpt::run(BasicBlock *BB) {
5065 bool Changed = false;
5067 assert(BB && BB->getParent() && "Block not embedded in function!");
5068 assert(BB->getTerminator() && "Degenerate basic block encountered!");
5070 // Remove basic blocks that have no predecessors (except the entry block)...
5071 // or that just have themself as a predecessor. These are unreachable.
5072 if ((pred_empty(BB) &&
5073 BB != &BB->getParent()->getEntryBlock()) ||
5074 BB->getSinglePredecessor() == BB) {
5075 DEBUG(dbgs() << "Removing BB: \n" << *BB);
5076 DeleteDeadBlock(BB);
5080 // Check to see if we can constant propagate this terminator instruction
5082 Changed |= ConstantFoldTerminator(BB, true);
5084 // Check for and eliminate duplicate PHI nodes in this block.
5085 Changed |= EliminateDuplicatePHINodes(BB);
5087 // Check for and remove branches that will always cause undefined behavior.
5088 Changed |= removeUndefIntroducingPredecessor(BB);
5090 // Merge basic blocks into their predecessor if there is only one distinct
5091 // pred, and if there is only one distinct successor of the predecessor, and
5092 // if there are no PHI nodes.
5094 if (MergeBlockIntoPredecessor(BB))
5097 IRBuilder<> Builder(BB);
5099 // If there is a trivial two-entry PHI node in this basic block, and we can
5100 // eliminate it, do so now.
5101 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
5102 if (PN->getNumIncomingValues() == 2)
5103 Changed |= FoldTwoEntryPHINode(PN, TTI, DL);
5105 Builder.SetInsertPoint(BB->getTerminator());
5106 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
5107 if (BI->isUnconditional()) {
5108 if (SimplifyUncondBranch(BI, Builder)) return true;
5110 if (SimplifyCondBranch(BI, Builder)) return true;
5112 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
5113 if (SimplifyReturn(RI, Builder)) return true;
5114 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
5115 if (SimplifyResume(RI, Builder)) return true;
5116 } else if (CleanupReturnInst *RI =
5117 dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
5118 if (SimplifyCleanupReturn(RI)) return true;
5119 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
5120 if (SimplifySwitch(SI, Builder)) return true;
5121 } else if (UnreachableInst *UI =
5122 dyn_cast<UnreachableInst>(BB->getTerminator())) {
5123 if (SimplifyUnreachable(UI)) return true;
5124 } else if (IndirectBrInst *IBI =
5125 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
5126 if (SimplifyIndirectBr(IBI)) return true;
5132 /// This function is used to do simplification of a CFG.
5133 /// For example, it adjusts branches to branches to eliminate the extra hop,
5134 /// eliminates unreachable basic blocks, and does other "peephole" optimization
5135 /// of the CFG. It returns true if a modification was made.
5137 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
5138 unsigned BonusInstThreshold, AssumptionCache *AC) {
5139 return SimplifyCFGOpt(TTI, BB->getModule()->getDataLayout(),
5140 BonusInstThreshold, AC).run(BB);