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/Local.h"
48 #include "llvm/Transforms/Utils/ValueMapper.h"
53 using namespace PatternMatch;
55 #define DEBUG_TYPE "simplifycfg"
57 // Chosen as 2 so as to be cheap, but still to have enough power to fold
58 // a select, so the "clamp" idiom (of a min followed by a max) will be caught.
59 // To catch this, we need to fold a compare and a select, hence '2' being the
60 // minimum reasonable default.
61 static cl::opt<unsigned>
62 PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(2),
63 cl::desc("Control the amount of phi node folding to perform (default = 2)"));
66 DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false),
67 cl::desc("Duplicate return instructions into unconditional branches"));
70 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
71 cl::desc("Sink common instructions down to the end block"));
73 static cl::opt<bool> HoistCondStores(
74 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
75 cl::desc("Hoist conditional stores if an unconditional store precedes"));
77 static cl::opt<bool> MergeCondStores(
78 "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
79 cl::desc("Hoist conditional stores even if an unconditional store does not "
80 "precede - hoist multiple conditional stores into a single "
83 static cl::opt<bool> MergeCondStoresAggressively(
84 "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
85 cl::desc("When merging conditional stores, do so even if the resultant "
86 "basic blocks are unlikely to be if-converted as a result"));
88 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
89 STATISTIC(NumLinearMaps, "Number of switch instructions turned into linear mapping");
90 STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables");
91 STATISTIC(NumLookupTablesHoles, "Number of switch instructions turned into lookup tables (holes checked)");
92 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
93 STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block");
94 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
97 // The first field contains the value that the switch produces when a certain
98 // case group is selected, and the second field is a vector containing the
99 // cases composing the case group.
100 typedef SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>
101 SwitchCaseResultVectorTy;
102 // The first field contains the phi node that generates a result of the switch
103 // and the second field contains the value generated for a certain case in the
104 // switch for that PHI.
105 typedef SmallVector<std::pair<PHINode *, Constant *>, 4> SwitchCaseResultsTy;
107 /// ValueEqualityComparisonCase - Represents a case of a switch.
108 struct ValueEqualityComparisonCase {
112 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
113 : Value(Value), Dest(Dest) {}
115 bool operator<(ValueEqualityComparisonCase RHS) const {
116 // Comparing pointers is ok as we only rely on the order for uniquing.
117 return Value < RHS.Value;
120 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
123 class SimplifyCFGOpt {
124 const TargetTransformInfo &TTI;
125 const DataLayout &DL;
126 unsigned BonusInstThreshold;
128 Value *isValueEqualityComparison(TerminatorInst *TI);
129 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
130 std::vector<ValueEqualityComparisonCase> &Cases);
131 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
133 IRBuilder<> &Builder);
134 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
135 IRBuilder<> &Builder);
137 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
138 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
139 bool SimplifyCleanupReturn(CleanupReturnInst *RI);
140 bool SimplifyUnreachable(UnreachableInst *UI);
141 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
142 bool SimplifyIndirectBr(IndirectBrInst *IBI);
143 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder);
144 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder);
147 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
148 unsigned BonusInstThreshold, AssumptionCache *AC)
149 : TTI(TTI), DL(DL), BonusInstThreshold(BonusInstThreshold), AC(AC) {}
150 bool run(BasicBlock *BB);
154 /// Return true if it is safe to merge these two
155 /// terminator instructions together.
156 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
157 if (SI1 == SI2) return false; // Can't merge with self!
159 // It is not safe to merge these two switch instructions if they have a common
160 // successor, and if that successor has a PHI node, and if *that* PHI node has
161 // conflicting incoming values from the two switch blocks.
162 BasicBlock *SI1BB = SI1->getParent();
163 BasicBlock *SI2BB = SI2->getParent();
164 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
166 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
167 if (SI1Succs.count(*I))
168 for (BasicBlock::iterator BBI = (*I)->begin();
169 isa<PHINode>(BBI); ++BBI) {
170 PHINode *PN = cast<PHINode>(BBI);
171 if (PN->getIncomingValueForBlock(SI1BB) !=
172 PN->getIncomingValueForBlock(SI2BB))
179 /// Return true if it is safe and profitable to merge these two terminator
180 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
181 /// store all PHI nodes in common successors.
182 static bool isProfitableToFoldUnconditional(BranchInst *SI1,
185 SmallVectorImpl<PHINode*> &PhiNodes) {
186 if (SI1 == SI2) return false; // Can't merge with self!
187 assert(SI1->isUnconditional() && SI2->isConditional());
189 // We fold the unconditional branch if we can easily update all PHI nodes in
190 // common successors:
191 // 1> We have a constant incoming value for the conditional branch;
192 // 2> We have "Cond" as the incoming value for the unconditional branch;
193 // 3> SI2->getCondition() and Cond have same operands.
194 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
195 if (!Ci2) return false;
196 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
197 Cond->getOperand(1) == Ci2->getOperand(1)) &&
198 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
199 Cond->getOperand(1) == Ci2->getOperand(0)))
202 BasicBlock *SI1BB = SI1->getParent();
203 BasicBlock *SI2BB = SI2->getParent();
204 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
205 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
206 if (SI1Succs.count(*I))
207 for (BasicBlock::iterator BBI = (*I)->begin();
208 isa<PHINode>(BBI); ++BBI) {
209 PHINode *PN = cast<PHINode>(BBI);
210 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
211 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
213 PhiNodes.push_back(PN);
218 /// Update PHI nodes in Succ to indicate that there will now be entries in it
219 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
220 /// will be the same as those coming in from ExistPred, an existing predecessor
222 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
223 BasicBlock *ExistPred) {
224 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
227 for (BasicBlock::iterator I = Succ->begin();
228 (PN = dyn_cast<PHINode>(I)); ++I)
229 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
232 /// Compute an abstract "cost" of speculating the given instruction,
233 /// which is assumed to be safe to speculate. TCC_Free means cheap,
234 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
236 static unsigned ComputeSpeculationCost(const User *I,
237 const TargetTransformInfo &TTI) {
238 assert(isSafeToSpeculativelyExecute(I) &&
239 "Instruction is not safe to speculatively execute!");
240 return TTI.getUserCost(I);
243 /// If we have a merge point of an "if condition" as accepted above,
244 /// return true if the specified value dominates the block. We
245 /// don't handle the true generality of domination here, just a special case
246 /// which works well enough for us.
248 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
249 /// see if V (which must be an instruction) and its recursive operands
250 /// that do not dominate BB have a combined cost lower than CostRemaining and
251 /// are non-trapping. If both are true, the instruction is inserted into the
252 /// set and true is returned.
254 /// The cost for most non-trapping instructions is defined as 1 except for
255 /// Select whose cost is 2.
257 /// After this function returns, CostRemaining is decreased by the cost of
258 /// V plus its non-dominating operands. If that cost is greater than
259 /// CostRemaining, false is returned and CostRemaining is undefined.
260 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
261 SmallPtrSetImpl<Instruction*> *AggressiveInsts,
262 unsigned &CostRemaining,
263 const TargetTransformInfo &TTI) {
264 Instruction *I = dyn_cast<Instruction>(V);
266 // Non-instructions all dominate instructions, but not all constantexprs
267 // can be executed unconditionally.
268 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
273 BasicBlock *PBB = I->getParent();
275 // We don't want to allow weird loops that might have the "if condition" in
276 // the bottom of this block.
277 if (PBB == BB) return false;
279 // If this instruction is defined in a block that contains an unconditional
280 // branch to BB, then it must be in the 'conditional' part of the "if
281 // statement". If not, it definitely dominates the region.
282 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
283 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
286 // If we aren't allowing aggressive promotion anymore, then don't consider
287 // instructions in the 'if region'.
288 if (!AggressiveInsts) return false;
290 // If we have seen this instruction before, don't count it again.
291 if (AggressiveInsts->count(I)) return true;
293 // Okay, it looks like the instruction IS in the "condition". Check to
294 // see if it's a cheap instruction to unconditionally compute, and if it
295 // only uses stuff defined outside of the condition. If so, hoist it out.
296 if (!isSafeToSpeculativelyExecute(I))
299 unsigned Cost = ComputeSpeculationCost(I, TTI);
301 if (Cost > CostRemaining)
304 CostRemaining -= Cost;
306 // Okay, we can only really hoist these out if their operands do
307 // not take us over the cost threshold.
308 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
309 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, TTI))
311 // Okay, it's safe to do this! Remember this instruction.
312 AggressiveInsts->insert(I);
316 /// Extract ConstantInt from value, looking through IntToPtr
317 /// and PointerNullValue. Return NULL if value is not a constant int.
318 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
319 // Normal constant int.
320 ConstantInt *CI = dyn_cast<ConstantInt>(V);
321 if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
324 // This is some kind of pointer constant. Turn it into a pointer-sized
325 // ConstantInt if possible.
326 IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
328 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
329 if (isa<ConstantPointerNull>(V))
330 return ConstantInt::get(PtrTy, 0);
332 // IntToPtr const int.
333 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
334 if (CE->getOpcode() == Instruction::IntToPtr)
335 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
336 // The constant is very likely to have the right type already.
337 if (CI->getType() == PtrTy)
340 return cast<ConstantInt>
341 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
348 /// Given a chain of or (||) or and (&&) comparison of a value against a
349 /// constant, this will try to recover the information required for a switch
351 /// It will depth-first traverse the chain of comparison, seeking for patterns
352 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
353 /// representing the different cases for the switch.
354 /// Note that if the chain is composed of '||' it will build the set of elements
355 /// that matches the comparisons (i.e. any of this value validate the chain)
356 /// while for a chain of '&&' it will build the set elements that make the test
358 struct ConstantComparesGatherer {
359 const DataLayout &DL;
360 Value *CompValue; /// Value found for the switch comparison
361 Value *Extra; /// Extra clause to be checked before the switch
362 SmallVector<ConstantInt *, 8> Vals; /// Set of integers to match in switch
363 unsigned UsedICmps; /// Number of comparisons matched in the and/or chain
365 /// Construct and compute the result for the comparison instruction Cond
366 ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL)
367 : DL(DL), CompValue(nullptr), Extra(nullptr), UsedICmps(0) {
372 ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
373 ConstantComparesGatherer &
374 operator=(const ConstantComparesGatherer &) = delete;
378 /// Try to set the current value used for the comparison, it succeeds only if
379 /// it wasn't set before or if the new value is the same as the old one
380 bool setValueOnce(Value *NewVal) {
381 if(CompValue && CompValue != NewVal) return false;
383 return (CompValue != nullptr);
386 /// Try to match Instruction "I" as a comparison against a constant and
387 /// populates the array Vals with the set of values that match (or do not
388 /// match depending on isEQ).
389 /// Return false on failure. On success, the Value the comparison matched
390 /// against is placed in CompValue.
391 /// If CompValue is already set, the function is expected to fail if a match
392 /// is found but the value compared to is different.
393 bool matchInstruction(Instruction *I, bool isEQ) {
394 // If this is an icmp against a constant, handle this as one of the cases.
397 if (!((ICI = dyn_cast<ICmpInst>(I)) &&
398 (C = GetConstantInt(I->getOperand(1), DL)))) {
405 // Pattern match a special case
406 // (x & ~2^x) == y --> x == y || x == y|2^x
407 // This undoes a transformation done by instcombine to fuse 2 compares.
408 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
409 if (match(ICI->getOperand(0),
410 m_And(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
411 APInt Not = ~RHSC->getValue();
412 if (Not.isPowerOf2()) {
413 // If we already have a value for the switch, it has to match!
414 if(!setValueOnce(RHSVal))
418 Vals.push_back(ConstantInt::get(C->getContext(),
419 C->getValue() | Not));
425 // If we already have a value for the switch, it has to match!
426 if(!setValueOnce(ICI->getOperand(0)))
431 return ICI->getOperand(0);
434 // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
435 ConstantRange Span = ConstantRange::makeAllowedICmpRegion(
436 ICI->getPredicate(), C->getValue());
438 // Shift the range if the compare is fed by an add. This is the range
439 // compare idiom as emitted by instcombine.
440 Value *CandidateVal = I->getOperand(0);
441 if(match(I->getOperand(0), m_Add(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
442 Span = Span.subtract(RHSC->getValue());
443 CandidateVal = RHSVal;
446 // If this is an and/!= check, then we are looking to build the set of
447 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
450 Span = Span.inverse();
452 // If there are a ton of values, we don't want to make a ginormous switch.
453 if (Span.getSetSize().ugt(8) || Span.isEmptySet()) {
457 // If we already have a value for the switch, it has to match!
458 if(!setValueOnce(CandidateVal))
461 // Add all values from the range to the set
462 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
463 Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
470 /// Given a potentially 'or'd or 'and'd together collection of icmp
471 /// eq/ne/lt/gt instructions that compare a value against a constant, extract
472 /// the value being compared, and stick the list constants into the Vals
474 /// One "Extra" case is allowed to differ from the other.
475 void gather(Value *V) {
476 Instruction *I = dyn_cast<Instruction>(V);
477 bool isEQ = (I->getOpcode() == Instruction::Or);
479 // Keep a stack (SmallVector for efficiency) for depth-first traversal
480 SmallVector<Value *, 8> DFT;
485 while(!DFT.empty()) {
486 V = DFT.pop_back_val();
488 if (Instruction *I = dyn_cast<Instruction>(V)) {
489 // If it is a || (or && depending on isEQ), process the operands.
490 if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
491 DFT.push_back(I->getOperand(1));
492 DFT.push_back(I->getOperand(0));
496 // Try to match the current instruction
497 if (matchInstruction(I, isEQ))
498 // Match succeed, continue the loop
502 // One element of the sequence of || (or &&) could not be match as a
503 // comparison against the same value as the others.
504 // We allow only one "Extra" case to be checked before the switch
509 // Failed to parse a proper sequence, abort now
518 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
519 Instruction *Cond = nullptr;
520 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
521 Cond = dyn_cast<Instruction>(SI->getCondition());
522 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
523 if (BI->isConditional())
524 Cond = dyn_cast<Instruction>(BI->getCondition());
525 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
526 Cond = dyn_cast<Instruction>(IBI->getAddress());
529 TI->eraseFromParent();
530 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
533 /// Return true if the specified terminator checks
534 /// to see if a value is equal to constant integer value.
535 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
537 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
538 // Do not permit merging of large switch instructions into their
539 // predecessors unless there is only one predecessor.
540 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()),
541 pred_end(SI->getParent())) <= 128)
542 CV = SI->getCondition();
543 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
544 if (BI->isConditional() && BI->getCondition()->hasOneUse())
545 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
546 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
547 CV = ICI->getOperand(0);
550 // Unwrap any lossless ptrtoint cast.
552 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
553 Value *Ptr = PTII->getPointerOperand();
554 if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
561 /// Given a value comparison instruction,
562 /// decode all of the 'cases' that it represents and return the 'default' block.
563 BasicBlock *SimplifyCFGOpt::
564 GetValueEqualityComparisonCases(TerminatorInst *TI,
565 std::vector<ValueEqualityComparisonCase>
567 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
568 Cases.reserve(SI->getNumCases());
569 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
570 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
571 i.getCaseSuccessor()));
572 return SI->getDefaultDest();
575 BranchInst *BI = cast<BranchInst>(TI);
576 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
577 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
578 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
581 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
585 /// Given a vector of bb/value pairs, remove any entries
586 /// in the list that match the specified block.
587 static void EliminateBlockCases(BasicBlock *BB,
588 std::vector<ValueEqualityComparisonCase> &Cases) {
589 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
592 /// Return true if there are any keys in C1 that exist in C2 as well.
594 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
595 std::vector<ValueEqualityComparisonCase > &C2) {
596 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
598 // Make V1 be smaller than V2.
599 if (V1->size() > V2->size())
602 if (V1->size() == 0) return false;
603 if (V1->size() == 1) {
605 ConstantInt *TheVal = (*V1)[0].Value;
606 for (unsigned i = 0, e = V2->size(); i != e; ++i)
607 if (TheVal == (*V2)[i].Value)
611 // Otherwise, just sort both lists and compare element by element.
612 array_pod_sort(V1->begin(), V1->end());
613 array_pod_sort(V2->begin(), V2->end());
614 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
615 while (i1 != e1 && i2 != e2) {
616 if ((*V1)[i1].Value == (*V2)[i2].Value)
618 if ((*V1)[i1].Value < (*V2)[i2].Value)
626 /// If TI is known to be a terminator instruction and its block is known to
627 /// only have a single predecessor block, check to see if that predecessor is
628 /// also a value comparison with the same value, and if that comparison
629 /// determines the outcome of this comparison. If so, simplify TI. This does a
630 /// very limited form of jump threading.
631 bool SimplifyCFGOpt::
632 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
634 IRBuilder<> &Builder) {
635 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
636 if (!PredVal) return false; // Not a value comparison in predecessor.
638 Value *ThisVal = isValueEqualityComparison(TI);
639 assert(ThisVal && "This isn't a value comparison!!");
640 if (ThisVal != PredVal) return false; // Different predicates.
642 // TODO: Preserve branch weight metadata, similarly to how
643 // FoldValueComparisonIntoPredecessors preserves it.
645 // Find out information about when control will move from Pred to TI's block.
646 std::vector<ValueEqualityComparisonCase> PredCases;
647 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
649 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
651 // Find information about how control leaves this block.
652 std::vector<ValueEqualityComparisonCase> ThisCases;
653 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
654 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
656 // If TI's block is the default block from Pred's comparison, potentially
657 // simplify TI based on this knowledge.
658 if (PredDef == TI->getParent()) {
659 // If we are here, we know that the value is none of those cases listed in
660 // PredCases. If there are any cases in ThisCases that are in PredCases, we
662 if (!ValuesOverlap(PredCases, ThisCases))
665 if (isa<BranchInst>(TI)) {
666 // Okay, one of the successors of this condbr is dead. Convert it to a
668 assert(ThisCases.size() == 1 && "Branch can only have one case!");
669 // Insert the new branch.
670 Instruction *NI = Builder.CreateBr(ThisDef);
673 // Remove PHI node entries for the dead edge.
674 ThisCases[0].Dest->removePredecessor(TI->getParent());
676 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
677 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
679 EraseTerminatorInstAndDCECond(TI);
683 SwitchInst *SI = cast<SwitchInst>(TI);
684 // Okay, TI has cases that are statically dead, prune them away.
685 SmallPtrSet<Constant*, 16> DeadCases;
686 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
687 DeadCases.insert(PredCases[i].Value);
689 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
690 << "Through successor TI: " << *TI);
692 // Collect branch weights into a vector.
693 SmallVector<uint32_t, 8> Weights;
694 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
695 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
697 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
699 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
700 Weights.push_back(CI->getValue().getZExtValue());
702 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
704 if (DeadCases.count(i.getCaseValue())) {
706 std::swap(Weights[i.getCaseIndex()+1], Weights.back());
709 i.getCaseSuccessor()->removePredecessor(TI->getParent());
713 if (HasWeight && Weights.size() >= 2)
714 SI->setMetadata(LLVMContext::MD_prof,
715 MDBuilder(SI->getParent()->getContext()).
716 createBranchWeights(Weights));
718 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
722 // Otherwise, TI's block must correspond to some matched value. Find out
723 // which value (or set of values) this is.
724 ConstantInt *TIV = nullptr;
725 BasicBlock *TIBB = TI->getParent();
726 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
727 if (PredCases[i].Dest == TIBB) {
729 return false; // Cannot handle multiple values coming to this block.
730 TIV = PredCases[i].Value;
732 assert(TIV && "No edge from pred to succ?");
734 // Okay, we found the one constant that our value can be if we get into TI's
735 // BB. Find out which successor will unconditionally be branched to.
736 BasicBlock *TheRealDest = nullptr;
737 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
738 if (ThisCases[i].Value == TIV) {
739 TheRealDest = ThisCases[i].Dest;
743 // If not handled by any explicit cases, it is handled by the default case.
744 if (!TheRealDest) TheRealDest = ThisDef;
746 // Remove PHI node entries for dead edges.
747 BasicBlock *CheckEdge = TheRealDest;
748 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
749 if (*SI != CheckEdge)
750 (*SI)->removePredecessor(TIBB);
754 // Insert the new branch.
755 Instruction *NI = Builder.CreateBr(TheRealDest);
758 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
759 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
761 EraseTerminatorInstAndDCECond(TI);
766 /// This class implements a stable ordering of constant
767 /// integers that does not depend on their address. This is important for
768 /// applications that sort ConstantInt's to ensure uniqueness.
769 struct ConstantIntOrdering {
770 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
771 return LHS->getValue().ult(RHS->getValue());
776 static int ConstantIntSortPredicate(ConstantInt *const *P1,
777 ConstantInt *const *P2) {
778 const ConstantInt *LHS = *P1;
779 const ConstantInt *RHS = *P2;
780 if (LHS->getValue().ult(RHS->getValue()))
782 if (LHS->getValue() == RHS->getValue())
787 static inline bool HasBranchWeights(const Instruction* I) {
788 MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
789 if (ProfMD && ProfMD->getOperand(0))
790 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
791 return MDS->getString().equals("branch_weights");
796 /// Get Weights of a given TerminatorInst, the default weight is at the front
797 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
799 static void GetBranchWeights(TerminatorInst *TI,
800 SmallVectorImpl<uint64_t> &Weights) {
801 MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
803 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
804 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
805 Weights.push_back(CI->getValue().getZExtValue());
808 // If TI is a conditional eq, the default case is the false case,
809 // and the corresponding branch-weight data is at index 2. We swap the
810 // default weight to be the first entry.
811 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
812 assert(Weights.size() == 2);
813 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
814 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
815 std::swap(Weights.front(), Weights.back());
819 /// Keep halving the weights until all can fit in uint32_t.
820 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
821 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
822 if (Max > UINT_MAX) {
823 unsigned Offset = 32 - countLeadingZeros(Max);
824 for (uint64_t &I : Weights)
829 /// The specified terminator is a value equality comparison instruction
830 /// (either a switch or a branch on "X == c").
831 /// See if any of the predecessors of the terminator block are value comparisons
832 /// on the same value. If so, and if safe to do so, fold them together.
833 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
834 IRBuilder<> &Builder) {
835 BasicBlock *BB = TI->getParent();
836 Value *CV = isValueEqualityComparison(TI); // CondVal
837 assert(CV && "Not a comparison?");
838 bool Changed = false;
840 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
841 while (!Preds.empty()) {
842 BasicBlock *Pred = Preds.pop_back_val();
844 // See if the predecessor is a comparison with the same value.
845 TerminatorInst *PTI = Pred->getTerminator();
846 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
848 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
849 // Figure out which 'cases' to copy from SI to PSI.
850 std::vector<ValueEqualityComparisonCase> BBCases;
851 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
853 std::vector<ValueEqualityComparisonCase> PredCases;
854 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
856 // Based on whether the default edge from PTI goes to BB or not, fill in
857 // PredCases and PredDefault with the new switch cases we would like to
859 SmallVector<BasicBlock*, 8> NewSuccessors;
861 // Update the branch weight metadata along the way
862 SmallVector<uint64_t, 8> Weights;
863 bool PredHasWeights = HasBranchWeights(PTI);
864 bool SuccHasWeights = HasBranchWeights(TI);
866 if (PredHasWeights) {
867 GetBranchWeights(PTI, Weights);
868 // branch-weight metadata is inconsistent here.
869 if (Weights.size() != 1 + PredCases.size())
870 PredHasWeights = SuccHasWeights = false;
871 } else if (SuccHasWeights)
872 // If there are no predecessor weights but there are successor weights,
873 // populate Weights with 1, which will later be scaled to the sum of
874 // successor's weights
875 Weights.assign(1 + PredCases.size(), 1);
877 SmallVector<uint64_t, 8> SuccWeights;
878 if (SuccHasWeights) {
879 GetBranchWeights(TI, SuccWeights);
880 // branch-weight metadata is inconsistent here.
881 if (SuccWeights.size() != 1 + BBCases.size())
882 PredHasWeights = SuccHasWeights = false;
883 } else if (PredHasWeights)
884 SuccWeights.assign(1 + BBCases.size(), 1);
886 if (PredDefault == BB) {
887 // If this is the default destination from PTI, only the edges in TI
888 // that don't occur in PTI, or that branch to BB will be activated.
889 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
890 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
891 if (PredCases[i].Dest != BB)
892 PTIHandled.insert(PredCases[i].Value);
894 // The default destination is BB, we don't need explicit targets.
895 std::swap(PredCases[i], PredCases.back());
897 if (PredHasWeights || SuccHasWeights) {
898 // Increase weight for the default case.
899 Weights[0] += Weights[i+1];
900 std::swap(Weights[i+1], Weights.back());
904 PredCases.pop_back();
908 // Reconstruct the new switch statement we will be building.
909 if (PredDefault != BBDefault) {
910 PredDefault->removePredecessor(Pred);
911 PredDefault = BBDefault;
912 NewSuccessors.push_back(BBDefault);
915 unsigned CasesFromPred = Weights.size();
916 uint64_t ValidTotalSuccWeight = 0;
917 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
918 if (!PTIHandled.count(BBCases[i].Value) &&
919 BBCases[i].Dest != BBDefault) {
920 PredCases.push_back(BBCases[i]);
921 NewSuccessors.push_back(BBCases[i].Dest);
922 if (SuccHasWeights || PredHasWeights) {
923 // The default weight is at index 0, so weight for the ith case
924 // should be at index i+1. Scale the cases from successor by
925 // PredDefaultWeight (Weights[0]).
926 Weights.push_back(Weights[0] * SuccWeights[i+1]);
927 ValidTotalSuccWeight += SuccWeights[i+1];
931 if (SuccHasWeights || PredHasWeights) {
932 ValidTotalSuccWeight += SuccWeights[0];
933 // Scale the cases from predecessor by ValidTotalSuccWeight.
934 for (unsigned i = 1; i < CasesFromPred; ++i)
935 Weights[i] *= ValidTotalSuccWeight;
936 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
937 Weights[0] *= SuccWeights[0];
940 // If this is not the default destination from PSI, only the edges
941 // in SI that occur in PSI with a destination of BB will be
943 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
944 std::map<ConstantInt*, uint64_t> WeightsForHandled;
945 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
946 if (PredCases[i].Dest == BB) {
947 PTIHandled.insert(PredCases[i].Value);
949 if (PredHasWeights || SuccHasWeights) {
950 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
951 std::swap(Weights[i+1], Weights.back());
955 std::swap(PredCases[i], PredCases.back());
956 PredCases.pop_back();
960 // Okay, now we know which constants were sent to BB from the
961 // predecessor. Figure out where they will all go now.
962 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
963 if (PTIHandled.count(BBCases[i].Value)) {
964 // If this is one we are capable of getting...
965 if (PredHasWeights || SuccHasWeights)
966 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
967 PredCases.push_back(BBCases[i]);
968 NewSuccessors.push_back(BBCases[i].Dest);
969 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
972 // If there are any constants vectored to BB that TI doesn't handle,
973 // they must go to the default destination of TI.
974 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
976 E = PTIHandled.end(); I != E; ++I) {
977 if (PredHasWeights || SuccHasWeights)
978 Weights.push_back(WeightsForHandled[*I]);
979 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
980 NewSuccessors.push_back(BBDefault);
984 // Okay, at this point, we know which new successor Pred will get. Make
985 // sure we update the number of entries in the PHI nodes for these
987 for (BasicBlock *NewSuccessor : NewSuccessors)
988 AddPredecessorToBlock(NewSuccessor, Pred, BB);
990 Builder.SetInsertPoint(PTI);
991 // Convert pointer to int before we switch.
992 if (CV->getType()->isPointerTy()) {
993 CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
997 // Now that the successors are updated, create the new Switch instruction.
998 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
1000 NewSI->setDebugLoc(PTI->getDebugLoc());
1001 for (ValueEqualityComparisonCase &V : PredCases)
1002 NewSI->addCase(V.Value, V.Dest);
1004 if (PredHasWeights || SuccHasWeights) {
1005 // Halve the weights if any of them cannot fit in an uint32_t
1006 FitWeights(Weights);
1008 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1010 NewSI->setMetadata(LLVMContext::MD_prof,
1011 MDBuilder(BB->getContext()).
1012 createBranchWeights(MDWeights));
1015 EraseTerminatorInstAndDCECond(PTI);
1017 // Okay, last check. If BB is still a successor of PSI, then we must
1018 // have an infinite loop case. If so, add an infinitely looping block
1019 // to handle the case to preserve the behavior of the code.
1020 BasicBlock *InfLoopBlock = nullptr;
1021 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1022 if (NewSI->getSuccessor(i) == BB) {
1023 if (!InfLoopBlock) {
1024 // Insert it at the end of the function, because it's either code,
1025 // or it won't matter if it's hot. :)
1026 InfLoopBlock = BasicBlock::Create(BB->getContext(),
1027 "infloop", BB->getParent());
1028 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1030 NewSI->setSuccessor(i, InfLoopBlock);
1039 // If we would need to insert a select that uses the value of this invoke
1040 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1041 // can't hoist the invoke, as there is nowhere to put the select in this case.
1042 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1043 Instruction *I1, Instruction *I2) {
1044 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1046 for (BasicBlock::iterator BBI = SI->begin();
1047 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1048 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1049 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1050 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
1058 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I);
1060 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1061 /// in the two blocks up into the branch block. The caller of this function
1062 /// guarantees that BI's block dominates BB1 and BB2.
1063 static bool HoistThenElseCodeToIf(BranchInst *BI,
1064 const TargetTransformInfo &TTI) {
1065 // This does very trivial matching, with limited scanning, to find identical
1066 // instructions in the two blocks. In particular, we don't want to get into
1067 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1068 // such, we currently just scan for obviously identical instructions in an
1070 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1071 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1073 BasicBlock::iterator BB1_Itr = BB1->begin();
1074 BasicBlock::iterator BB2_Itr = BB2->begin();
1076 Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1077 // Skip debug info if it is not identical.
1078 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1079 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1080 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1081 while (isa<DbgInfoIntrinsic>(I1))
1083 while (isa<DbgInfoIntrinsic>(I2))
1086 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1087 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1090 BasicBlock *BIParent = BI->getParent();
1092 bool Changed = false;
1094 // If we are hoisting the terminator instruction, don't move one (making a
1095 // broken BB), instead clone it, and remove BI.
1096 if (isa<TerminatorInst>(I1))
1097 goto HoistTerminator;
1099 if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1102 // For a normal instruction, we just move one to right before the branch,
1103 // then replace all uses of the other with the first. Finally, we remove
1104 // the now redundant second instruction.
1105 BIParent->getInstList().splice(BI->getIterator(), BB1->getInstList(), I1);
1106 if (!I2->use_empty())
1107 I2->replaceAllUsesWith(I1);
1108 I1->intersectOptionalDataWith(I2);
1109 unsigned KnownIDs[] = {
1110 LLVMContext::MD_tbaa, LLVMContext::MD_range,
1111 LLVMContext::MD_fpmath, LLVMContext::MD_invariant_load,
1112 LLVMContext::MD_nonnull, LLVMContext::MD_invariant_group,
1113 LLVMContext::MD_align, LLVMContext::MD_dereferenceable,
1114 LLVMContext::MD_dereferenceable_or_null};
1115 combineMetadata(I1, I2, KnownIDs);
1116 I2->eraseFromParent();
1121 // Skip debug info if it is not identical.
1122 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1123 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1124 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1125 while (isa<DbgInfoIntrinsic>(I1))
1127 while (isa<DbgInfoIntrinsic>(I2))
1130 } while (I1->isIdenticalToWhenDefined(I2));
1135 // It may not be possible to hoist an invoke.
1136 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1139 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1141 for (BasicBlock::iterator BBI = SI->begin();
1142 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1143 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1144 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1148 // Check for passingValueIsAlwaysUndefined here because we would rather
1149 // eliminate undefined control flow then converting it to a select.
1150 if (passingValueIsAlwaysUndefined(BB1V, PN) ||
1151 passingValueIsAlwaysUndefined(BB2V, PN))
1154 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1156 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1161 // Okay, it is safe to hoist the terminator.
1162 Instruction *NT = I1->clone();
1163 BIParent->getInstList().insert(BI->getIterator(), NT);
1164 if (!NT->getType()->isVoidTy()) {
1165 I1->replaceAllUsesWith(NT);
1166 I2->replaceAllUsesWith(NT);
1170 IRBuilder<true, NoFolder> Builder(NT);
1171 // Hoisting one of the terminators from our successor is a great thing.
1172 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1173 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1174 // nodes, so we insert select instruction to compute the final result.
1175 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1176 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1178 for (BasicBlock::iterator BBI = SI->begin();
1179 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1180 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1181 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1182 if (BB1V == BB2V) continue;
1184 // These values do not agree. Insert a select instruction before NT
1185 // that determines the right value.
1186 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1188 SI = cast<SelectInst>
1189 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1190 BB1V->getName()+"."+BB2V->getName()));
1192 // Make the PHI node use the select for all incoming values for BB1/BB2
1193 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1194 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1195 PN->setIncomingValue(i, SI);
1199 // Update any PHI nodes in our new successors.
1200 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1201 AddPredecessorToBlock(*SI, BIParent, BB1);
1203 EraseTerminatorInstAndDCECond(BI);
1207 /// Given an unconditional branch that goes to BBEnd,
1208 /// check whether BBEnd has only two predecessors and the other predecessor
1209 /// ends with an unconditional branch. If it is true, sink any common code
1210 /// in the two predecessors to BBEnd.
1211 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1212 assert(BI1->isUnconditional());
1213 BasicBlock *BB1 = BI1->getParent();
1214 BasicBlock *BBEnd = BI1->getSuccessor(0);
1216 // Check that BBEnd has two predecessors and the other predecessor ends with
1217 // an unconditional branch.
1218 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1219 BasicBlock *Pred0 = *PI++;
1220 if (PI == PE) // Only one predecessor.
1222 BasicBlock *Pred1 = *PI++;
1223 if (PI != PE) // More than two predecessors.
1225 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1226 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1227 if (!BI2 || !BI2->isUnconditional())
1230 // Gather the PHI nodes in BBEnd.
1231 SmallDenseMap<std::pair<Value *, Value *>, PHINode *> JointValueMap;
1232 Instruction *FirstNonPhiInBBEnd = nullptr;
1233 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end(); I != E; ++I) {
1234 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1235 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1236 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1237 JointValueMap[std::make_pair(BB1V, BB2V)] = PN;
1239 FirstNonPhiInBBEnd = &*I;
1243 if (!FirstNonPhiInBBEnd)
1246 // This does very trivial matching, with limited scanning, to find identical
1247 // instructions in the two blocks. We scan backward for obviously identical
1248 // instructions in an identical order.
1249 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1250 RE1 = BB1->getInstList().rend(),
1251 RI2 = BB2->getInstList().rbegin(),
1252 RE2 = BB2->getInstList().rend();
1254 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1257 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1260 // Skip the unconditional branches.
1264 bool Changed = false;
1265 while (RI1 != RE1 && RI2 != RE2) {
1267 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1270 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1274 Instruction *I1 = &*RI1, *I2 = &*RI2;
1275 auto InstPair = std::make_pair(I1, I2);
1276 // I1 and I2 should have a single use in the same PHI node, and they
1277 // perform the same operation.
1278 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1279 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1280 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1281 I1->isEHPad() || I2->isEHPad() ||
1282 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1283 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1284 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1285 !I1->hasOneUse() || !I2->hasOneUse() ||
1286 !JointValueMap.count(InstPair))
1289 // Check whether we should swap the operands of ICmpInst.
1290 // TODO: Add support of communativity.
1291 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1292 bool SwapOpnds = false;
1293 if (ICmp1 && ICmp2 &&
1294 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1295 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1296 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1297 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1298 ICmp2->swapOperands();
1301 if (!I1->isSameOperationAs(I2)) {
1303 ICmp2->swapOperands();
1307 // The operands should be either the same or they need to be generated
1308 // with a PHI node after sinking. We only handle the case where there is
1309 // a single pair of different operands.
1310 Value *DifferentOp1 = nullptr, *DifferentOp2 = nullptr;
1311 unsigned Op1Idx = ~0U;
1312 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1313 if (I1->getOperand(I) == I2->getOperand(I))
1315 // Early exit if we have more-than one pair of different operands or if
1316 // we need a PHI node to replace a constant.
1317 if (Op1Idx != ~0U ||
1318 isa<Constant>(I1->getOperand(I)) ||
1319 isa<Constant>(I2->getOperand(I))) {
1320 // If we can't sink the instructions, undo the swapping.
1322 ICmp2->swapOperands();
1325 DifferentOp1 = I1->getOperand(I);
1327 DifferentOp2 = I2->getOperand(I);
1330 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n");
1331 DEBUG(dbgs() << " " << *I2 << "\n");
1333 // We insert the pair of different operands to JointValueMap and
1334 // remove (I1, I2) from JointValueMap.
1335 if (Op1Idx != ~0U) {
1336 auto &NewPN = JointValueMap[std::make_pair(DifferentOp1, DifferentOp2)];
1339 PHINode::Create(DifferentOp1->getType(), 2,
1340 DifferentOp1->getName() + ".sink", &BBEnd->front());
1341 NewPN->addIncoming(DifferentOp1, BB1);
1342 NewPN->addIncoming(DifferentOp2, BB2);
1343 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1345 // I1 should use NewPN instead of DifferentOp1.
1346 I1->setOperand(Op1Idx, NewPN);
1348 PHINode *OldPN = JointValueMap[InstPair];
1349 JointValueMap.erase(InstPair);
1351 // We need to update RE1 and RE2 if we are going to sink the first
1352 // instruction in the basic block down.
1353 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1354 // Sink the instruction.
1355 BBEnd->getInstList().splice(FirstNonPhiInBBEnd->getIterator(),
1356 BB1->getInstList(), I1);
1357 if (!OldPN->use_empty())
1358 OldPN->replaceAllUsesWith(I1);
1359 OldPN->eraseFromParent();
1361 if (!I2->use_empty())
1362 I2->replaceAllUsesWith(I1);
1363 I1->intersectOptionalDataWith(I2);
1364 // TODO: Use combineMetadata here to preserve what metadata we can
1365 // (analogous to the hoisting case above).
1366 I2->eraseFromParent();
1369 RE1 = BB1->getInstList().rend();
1371 RE2 = BB2->getInstList().rend();
1372 FirstNonPhiInBBEnd = &*I1;
1379 /// \brief Determine if we can hoist sink a sole store instruction out of a
1380 /// conditional block.
1382 /// We are looking for code like the following:
1384 /// store i32 %add, i32* %arrayidx2
1385 /// ... // No other stores or function calls (we could be calling a memory
1386 /// ... // function).
1387 /// %cmp = icmp ult %x, %y
1388 /// br i1 %cmp, label %EndBB, label %ThenBB
1390 /// store i32 %add5, i32* %arrayidx2
1394 /// We are going to transform this into:
1396 /// store i32 %add, i32* %arrayidx2
1398 /// %cmp = icmp ult %x, %y
1399 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1400 /// store i32 %add.add5, i32* %arrayidx2
1403 /// \return The pointer to the value of the previous store if the store can be
1404 /// hoisted into the predecessor block. 0 otherwise.
1405 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1406 BasicBlock *StoreBB, BasicBlock *EndBB) {
1407 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1411 // Volatile or atomic.
1412 if (!StoreToHoist->isSimple())
1415 Value *StorePtr = StoreToHoist->getPointerOperand();
1417 // Look for a store to the same pointer in BrBB.
1418 unsigned MaxNumInstToLookAt = 10;
1419 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(),
1420 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) {
1421 Instruction *CurI = &*RI;
1423 // Could be calling an instruction that effects memory like free().
1424 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI))
1427 StoreInst *SI = dyn_cast<StoreInst>(CurI);
1428 // Found the previous store make sure it stores to the same location.
1429 if (SI && SI->getPointerOperand() == StorePtr)
1430 // Found the previous store, return its value operand.
1431 return SI->getValueOperand();
1433 return nullptr; // Unknown store.
1439 /// \brief Speculate a conditional basic block flattening the CFG.
1441 /// Note that this is a very risky transform currently. Speculating
1442 /// instructions like this is most often not desirable. Instead, there is an MI
1443 /// pass which can do it with full awareness of the resource constraints.
1444 /// However, some cases are "obvious" and we should do directly. An example of
1445 /// this is speculating a single, reasonably cheap instruction.
1447 /// There is only one distinct advantage to flattening the CFG at the IR level:
1448 /// it makes very common but simplistic optimizations such as are common in
1449 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1450 /// modeling their effects with easier to reason about SSA value graphs.
1453 /// An illustration of this transform is turning this IR:
1456 /// %cmp = icmp ult %x, %y
1457 /// br i1 %cmp, label %EndBB, label %ThenBB
1459 /// %sub = sub %x, %y
1462 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1469 /// %cmp = icmp ult %x, %y
1470 /// %sub = sub %x, %y
1471 /// %cond = select i1 %cmp, 0, %sub
1475 /// \returns true if the conditional block is removed.
1476 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1477 const TargetTransformInfo &TTI) {
1478 // Be conservative for now. FP select instruction can often be expensive.
1479 Value *BrCond = BI->getCondition();
1480 if (isa<FCmpInst>(BrCond))
1483 BasicBlock *BB = BI->getParent();
1484 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1486 // If ThenBB is actually on the false edge of the conditional branch, remember
1487 // to swap the select operands later.
1488 bool Invert = false;
1489 if (ThenBB != BI->getSuccessor(0)) {
1490 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1493 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1495 // Keep a count of how many times instructions are used within CondBB when
1496 // they are candidates for sinking into CondBB. Specifically:
1497 // - They are defined in BB, and
1498 // - They have no side effects, and
1499 // - All of their uses are in CondBB.
1500 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1502 unsigned SpeculationCost = 0;
1503 Value *SpeculatedStoreValue = nullptr;
1504 StoreInst *SpeculatedStore = nullptr;
1505 for (BasicBlock::iterator BBI = ThenBB->begin(),
1506 BBE = std::prev(ThenBB->end());
1507 BBI != BBE; ++BBI) {
1508 Instruction *I = &*BBI;
1510 if (isa<DbgInfoIntrinsic>(I))
1513 // Only speculatively execute a single instruction (not counting the
1514 // terminator) for now.
1516 if (SpeculationCost > 1)
1519 // Don't hoist the instruction if it's unsafe or expensive.
1520 if (!isSafeToSpeculativelyExecute(I) &&
1521 !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1522 I, BB, ThenBB, EndBB))))
1524 if (!SpeculatedStoreValue &&
1525 ComputeSpeculationCost(I, TTI) >
1526 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
1529 // Store the store speculation candidate.
1530 if (SpeculatedStoreValue)
1531 SpeculatedStore = cast<StoreInst>(I);
1533 // Do not hoist the instruction if any of its operands are defined but not
1534 // used in BB. The transformation will prevent the operand from
1535 // being sunk into the use block.
1536 for (User::op_iterator i = I->op_begin(), e = I->op_end();
1538 Instruction *OpI = dyn_cast<Instruction>(*i);
1539 if (!OpI || OpI->getParent() != BB ||
1540 OpI->mayHaveSideEffects())
1541 continue; // Not a candidate for sinking.
1543 ++SinkCandidateUseCounts[OpI];
1547 // Consider any sink candidates which are only used in CondBB as costs for
1548 // speculation. Note, while we iterate over a DenseMap here, we are summing
1549 // and so iteration order isn't significant.
1550 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I =
1551 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end();
1553 if (I->first->getNumUses() == I->second) {
1555 if (SpeculationCost > 1)
1559 // Check that the PHI nodes can be converted to selects.
1560 bool HaveRewritablePHIs = false;
1561 for (BasicBlock::iterator I = EndBB->begin();
1562 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1563 Value *OrigV = PN->getIncomingValueForBlock(BB);
1564 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1566 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1567 // Skip PHIs which are trivial.
1571 // Don't convert to selects if we could remove undefined behavior instead.
1572 if (passingValueIsAlwaysUndefined(OrigV, PN) ||
1573 passingValueIsAlwaysUndefined(ThenV, PN))
1576 HaveRewritablePHIs = true;
1577 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1578 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1579 if (!OrigCE && !ThenCE)
1580 continue; // Known safe and cheap.
1582 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
1583 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
1585 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
1586 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
1587 unsigned MaxCost = 2 * PHINodeFoldingThreshold *
1588 TargetTransformInfo::TCC_Basic;
1589 if (OrigCost + ThenCost > MaxCost)
1592 // Account for the cost of an unfolded ConstantExpr which could end up
1593 // getting expanded into Instructions.
1594 // FIXME: This doesn't account for how many operations are combined in the
1595 // constant expression.
1597 if (SpeculationCost > 1)
1601 // If there are no PHIs to process, bail early. This helps ensure idempotence
1603 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
1606 // If we get here, we can hoist the instruction and if-convert.
1607 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1609 // Insert a select of the value of the speculated store.
1610 if (SpeculatedStoreValue) {
1611 IRBuilder<true, NoFolder> Builder(BI);
1612 Value *TrueV = SpeculatedStore->getValueOperand();
1613 Value *FalseV = SpeculatedStoreValue;
1615 std::swap(TrueV, FalseV);
1616 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() +
1617 "." + FalseV->getName());
1618 SpeculatedStore->setOperand(0, S);
1621 // Metadata can be dependent on the condition we are hoisting above.
1622 // Conservatively strip all metadata on the instruction.
1623 for (auto &I: *ThenBB)
1624 I.dropUnknownNonDebugMetadata();
1626 // Hoist the instructions.
1627 BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
1628 ThenBB->begin(), std::prev(ThenBB->end()));
1630 // Insert selects and rewrite the PHI operands.
1631 IRBuilder<true, NoFolder> Builder(BI);
1632 for (BasicBlock::iterator I = EndBB->begin();
1633 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1634 unsigned OrigI = PN->getBasicBlockIndex(BB);
1635 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1636 Value *OrigV = PN->getIncomingValue(OrigI);
1637 Value *ThenV = PN->getIncomingValue(ThenI);
1639 // Skip PHIs which are trivial.
1643 // Create a select whose true value is the speculatively executed value and
1644 // false value is the preexisting value. Swap them if the branch
1645 // destinations were inverted.
1646 Value *TrueV = ThenV, *FalseV = OrigV;
1648 std::swap(TrueV, FalseV);
1649 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV,
1650 TrueV->getName() + "." + FalseV->getName());
1651 PN->setIncomingValue(OrigI, V);
1652 PN->setIncomingValue(ThenI, V);
1659 /// \returns True if this block contains a CallInst with the NoDuplicate
1661 static bool HasNoDuplicateCall(const BasicBlock *BB) {
1662 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1663 const CallInst *CI = dyn_cast<CallInst>(I);
1666 if (CI->cannotDuplicate())
1672 /// Return true if we can thread a branch across this block.
1673 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1674 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1677 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1678 if (isa<DbgInfoIntrinsic>(BBI))
1680 if (Size > 10) return false; // Don't clone large BB's.
1683 // We can only support instructions that do not define values that are
1684 // live outside of the current basic block.
1685 for (User *U : BBI->users()) {
1686 Instruction *UI = cast<Instruction>(U);
1687 if (UI->getParent() != BB || isa<PHINode>(UI)) return false;
1690 // Looks ok, continue checking.
1696 /// If we have a conditional branch on a PHI node value that is defined in the
1697 /// same block as the branch and if any PHI entries are constants, thread edges
1698 /// corresponding to that entry to be branches to their ultimate destination.
1699 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL) {
1700 BasicBlock *BB = BI->getParent();
1701 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1702 // NOTE: we currently cannot transform this case if the PHI node is used
1703 // outside of the block.
1704 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1707 // Degenerate case of a single entry PHI.
1708 if (PN->getNumIncomingValues() == 1) {
1709 FoldSingleEntryPHINodes(PN->getParent());
1713 // Now we know that this block has multiple preds and two succs.
1714 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1716 if (HasNoDuplicateCall(BB)) return false;
1718 // Okay, this is a simple enough basic block. See if any phi values are
1720 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1721 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1722 if (!CB || !CB->getType()->isIntegerTy(1)) continue;
1724 // Okay, we now know that all edges from PredBB should be revectored to
1725 // branch to RealDest.
1726 BasicBlock *PredBB = PN->getIncomingBlock(i);
1727 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1729 if (RealDest == BB) continue; // Skip self loops.
1730 // Skip if the predecessor's terminator is an indirect branch.
1731 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1733 // The dest block might have PHI nodes, other predecessors and other
1734 // difficult cases. Instead of being smart about this, just insert a new
1735 // block that jumps to the destination block, effectively splitting
1736 // the edge we are about to create.
1737 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1738 RealDest->getName()+".critedge",
1739 RealDest->getParent(), RealDest);
1740 BranchInst::Create(RealDest, EdgeBB);
1742 // Update PHI nodes.
1743 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1745 // BB may have instructions that are being threaded over. Clone these
1746 // instructions into EdgeBB. We know that there will be no uses of the
1747 // cloned instructions outside of EdgeBB.
1748 BasicBlock::iterator InsertPt = EdgeBB->begin();
1749 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1750 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1751 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1752 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1755 // Clone the instruction.
1756 Instruction *N = BBI->clone();
1757 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1759 // Update operands due to translation.
1760 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1762 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1763 if (PI != TranslateMap.end())
1767 // Check for trivial simplification.
1768 if (Value *V = SimplifyInstruction(N, DL)) {
1769 TranslateMap[&*BBI] = V;
1770 delete N; // Instruction folded away, don't need actual inst
1772 // Insert the new instruction into its new home.
1773 EdgeBB->getInstList().insert(InsertPt, N);
1774 if (!BBI->use_empty())
1775 TranslateMap[&*BBI] = N;
1779 // Loop over all of the edges from PredBB to BB, changing them to branch
1780 // to EdgeBB instead.
1781 TerminatorInst *PredBBTI = PredBB->getTerminator();
1782 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1783 if (PredBBTI->getSuccessor(i) == BB) {
1784 BB->removePredecessor(PredBB);
1785 PredBBTI->setSuccessor(i, EdgeBB);
1788 // Recurse, simplifying any other constants.
1789 return FoldCondBranchOnPHI(BI, DL) | true;
1795 /// Given a BB that starts with the specified two-entry PHI node,
1796 /// see if we can eliminate it.
1797 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI,
1798 const DataLayout &DL) {
1799 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1800 // statement", which has a very simple dominance structure. Basically, we
1801 // are trying to find the condition that is being branched on, which
1802 // subsequently causes this merge to happen. We really want control
1803 // dependence information for this check, but simplifycfg can't keep it up
1804 // to date, and this catches most of the cases we care about anyway.
1805 BasicBlock *BB = PN->getParent();
1806 BasicBlock *IfTrue, *IfFalse;
1807 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1809 // Don't bother if the branch will be constant folded trivially.
1810 isa<ConstantInt>(IfCond))
1813 // Okay, we found that we can merge this two-entry phi node into a select.
1814 // Doing so would require us to fold *all* two entry phi nodes in this block.
1815 // At some point this becomes non-profitable (particularly if the target
1816 // doesn't support cmov's). Only do this transformation if there are two or
1817 // fewer PHI nodes in this block.
1818 unsigned NumPhis = 0;
1819 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1823 // Loop over the PHI's seeing if we can promote them all to select
1824 // instructions. While we are at it, keep track of the instructions
1825 // that need to be moved to the dominating block.
1826 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1827 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1828 MaxCostVal1 = PHINodeFoldingThreshold;
1829 MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
1830 MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
1832 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1833 PHINode *PN = cast<PHINode>(II++);
1834 if (Value *V = SimplifyInstruction(PN, DL)) {
1835 PN->replaceAllUsesWith(V);
1836 PN->eraseFromParent();
1840 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1841 MaxCostVal0, TTI) ||
1842 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1847 // If we folded the first phi, PN dangles at this point. Refresh it. If
1848 // we ran out of PHIs then we simplified them all.
1849 PN = dyn_cast<PHINode>(BB->begin());
1850 if (!PN) return true;
1852 // Don't fold i1 branches on PHIs which contain binary operators. These can
1853 // often be turned into switches and other things.
1854 if (PN->getType()->isIntegerTy(1) &&
1855 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1856 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1857 isa<BinaryOperator>(IfCond)))
1860 // If we all PHI nodes are promotable, check to make sure that all
1861 // instructions in the predecessor blocks can be promoted as well. If
1862 // not, we won't be able to get rid of the control flow, so it's not
1863 // worth promoting to select instructions.
1864 BasicBlock *DomBlock = nullptr;
1865 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1866 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1867 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1870 DomBlock = *pred_begin(IfBlock1);
1871 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1872 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
1873 // This is not an aggressive instruction that we can promote.
1874 // Because of this, we won't be able to get rid of the control
1875 // flow, so the xform is not worth it.
1880 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1883 DomBlock = *pred_begin(IfBlock2);
1884 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1885 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
1886 // This is not an aggressive instruction that we can promote.
1887 // Because of this, we won't be able to get rid of the control
1888 // flow, so the xform is not worth it.
1893 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1894 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1896 // If we can still promote the PHI nodes after this gauntlet of tests,
1897 // do all of the PHI's now.
1898 Instruction *InsertPt = DomBlock->getTerminator();
1899 IRBuilder<true, NoFolder> Builder(InsertPt);
1901 // Move all 'aggressive' instructions, which are defined in the
1902 // conditional parts of the if's up to the dominating block.
1904 DomBlock->getInstList().splice(InsertPt->getIterator(),
1905 IfBlock1->getInstList(), IfBlock1->begin(),
1906 IfBlock1->getTerminator()->getIterator());
1908 DomBlock->getInstList().splice(InsertPt->getIterator(),
1909 IfBlock2->getInstList(), IfBlock2->begin(),
1910 IfBlock2->getTerminator()->getIterator());
1912 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1913 // Change the PHI node into a select instruction.
1914 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1915 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1918 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1919 PN->replaceAllUsesWith(NV);
1921 PN->eraseFromParent();
1924 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1925 // has been flattened. Change DomBlock to jump directly to our new block to
1926 // avoid other simplifycfg's kicking in on the diamond.
1927 TerminatorInst *OldTI = DomBlock->getTerminator();
1928 Builder.SetInsertPoint(OldTI);
1929 Builder.CreateBr(BB);
1930 OldTI->eraseFromParent();
1934 /// If we found a conditional branch that goes to two returning blocks,
1935 /// try to merge them together into one return,
1936 /// introducing a select if the return values disagree.
1937 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1938 IRBuilder<> &Builder) {
1939 assert(BI->isConditional() && "Must be a conditional branch");
1940 BasicBlock *TrueSucc = BI->getSuccessor(0);
1941 BasicBlock *FalseSucc = BI->getSuccessor(1);
1942 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1943 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1945 // Check to ensure both blocks are empty (just a return) or optionally empty
1946 // with PHI nodes. If there are other instructions, merging would cause extra
1947 // computation on one path or the other.
1948 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1950 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1953 Builder.SetInsertPoint(BI);
1954 // Okay, we found a branch that is going to two return nodes. If
1955 // there is no return value for this function, just change the
1956 // branch into a return.
1957 if (FalseRet->getNumOperands() == 0) {
1958 TrueSucc->removePredecessor(BI->getParent());
1959 FalseSucc->removePredecessor(BI->getParent());
1960 Builder.CreateRetVoid();
1961 EraseTerminatorInstAndDCECond(BI);
1965 // Otherwise, figure out what the true and false return values are
1966 // so we can insert a new select instruction.
1967 Value *TrueValue = TrueRet->getReturnValue();
1968 Value *FalseValue = FalseRet->getReturnValue();
1970 // Unwrap any PHI nodes in the return blocks.
1971 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1972 if (TVPN->getParent() == TrueSucc)
1973 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1974 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1975 if (FVPN->getParent() == FalseSucc)
1976 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1978 // In order for this transformation to be safe, we must be able to
1979 // unconditionally execute both operands to the return. This is
1980 // normally the case, but we could have a potentially-trapping
1981 // constant expression that prevents this transformation from being
1983 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1986 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1990 // Okay, we collected all the mapped values and checked them for sanity, and
1991 // defined to really do this transformation. First, update the CFG.
1992 TrueSucc->removePredecessor(BI->getParent());
1993 FalseSucc->removePredecessor(BI->getParent());
1995 // Insert select instructions where needed.
1996 Value *BrCond = BI->getCondition();
1998 // Insert a select if the results differ.
1999 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2000 } else if (isa<UndefValue>(TrueValue)) {
2001 TrueValue = FalseValue;
2003 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
2004 FalseValue, "retval");
2008 Value *RI = !TrueValue ?
2009 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2013 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2014 << "\n " << *BI << "NewRet = " << *RI
2015 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
2017 EraseTerminatorInstAndDCECond(BI);
2022 /// Given a conditional BranchInstruction, retrieve the probabilities of the
2023 /// branch taking each edge. Fills in the two APInt parameters and returns true,
2024 /// or returns false if no or invalid metadata was found.
2025 static bool ExtractBranchMetadata(BranchInst *BI,
2026 uint64_t &ProbTrue, uint64_t &ProbFalse) {
2027 assert(BI->isConditional() &&
2028 "Looking for probabilities on unconditional branch?");
2029 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
2030 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
2031 ConstantInt *CITrue =
2032 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
2033 ConstantInt *CIFalse =
2034 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
2035 if (!CITrue || !CIFalse) return false;
2036 ProbTrue = CITrue->getValue().getZExtValue();
2037 ProbFalse = CIFalse->getValue().getZExtValue();
2041 /// Return true if the given instruction is available
2042 /// in its predecessor block. If yes, the instruction will be removed.
2043 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
2044 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2046 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
2047 Instruction *PBI = &*I;
2048 // Check whether Inst and PBI generate the same value.
2049 if (Inst->isIdenticalTo(PBI)) {
2050 Inst->replaceAllUsesWith(PBI);
2051 Inst->eraseFromParent();
2058 /// If this basic block is simple enough, and if a predecessor branches to us
2059 /// and one of our successors, fold the block into the predecessor and use
2060 /// logical operations to pick the right destination.
2061 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2062 BasicBlock *BB = BI->getParent();
2064 Instruction *Cond = nullptr;
2065 if (BI->isConditional())
2066 Cond = dyn_cast<Instruction>(BI->getCondition());
2068 // For unconditional branch, check for a simple CFG pattern, where
2069 // BB has a single predecessor and BB's successor is also its predecessor's
2070 // successor. If such pattern exisits, check for CSE between BB and its
2072 if (BasicBlock *PB = BB->getSinglePredecessor())
2073 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2074 if (PBI->isConditional() &&
2075 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2076 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2077 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
2079 Instruction *Curr = &*I++;
2080 if (isa<CmpInst>(Curr)) {
2084 // Quit if we can't remove this instruction.
2085 if (!checkCSEInPredecessor(Curr, PB))
2094 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2095 Cond->getParent() != BB || !Cond->hasOneUse())
2098 // Make sure the instruction after the condition is the cond branch.
2099 BasicBlock::iterator CondIt = ++Cond->getIterator();
2101 // Ignore dbg intrinsics.
2102 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
2107 // Only allow this transformation if computing the condition doesn't involve
2108 // too many instructions and these involved instructions can be executed
2109 // unconditionally. We denote all involved instructions except the condition
2110 // as "bonus instructions", and only allow this transformation when the
2111 // number of the bonus instructions does not exceed a certain threshold.
2112 unsigned NumBonusInsts = 0;
2113 for (auto I = BB->begin(); Cond != I; ++I) {
2114 // Ignore dbg intrinsics.
2115 if (isa<DbgInfoIntrinsic>(I))
2117 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2119 // I has only one use and can be executed unconditionally.
2120 Instruction *User = dyn_cast<Instruction>(I->user_back());
2121 if (User == nullptr || User->getParent() != BB)
2123 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2124 // to use any other instruction, User must be an instruction between next(I)
2127 // Early exits once we reach the limit.
2128 if (NumBonusInsts > BonusInstThreshold)
2132 // Cond is known to be a compare or binary operator. Check to make sure that
2133 // neither operand is a potentially-trapping constant expression.
2134 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2137 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2141 // Finally, don't infinitely unroll conditional loops.
2142 BasicBlock *TrueDest = BI->getSuccessor(0);
2143 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2144 if (TrueDest == BB || FalseDest == BB)
2147 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2148 BasicBlock *PredBlock = *PI;
2149 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2151 // Check that we have two conditional branches. If there is a PHI node in
2152 // the common successor, verify that the same value flows in from both
2154 SmallVector<PHINode*, 4> PHIs;
2155 if (!PBI || PBI->isUnconditional() ||
2156 (BI->isConditional() &&
2157 !SafeToMergeTerminators(BI, PBI)) ||
2158 (!BI->isConditional() &&
2159 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2162 // Determine if the two branches share a common destination.
2163 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2164 bool InvertPredCond = false;
2166 if (BI->isConditional()) {
2167 if (PBI->getSuccessor(0) == TrueDest)
2168 Opc = Instruction::Or;
2169 else if (PBI->getSuccessor(1) == FalseDest)
2170 Opc = Instruction::And;
2171 else if (PBI->getSuccessor(0) == FalseDest)
2172 Opc = Instruction::And, InvertPredCond = true;
2173 else if (PBI->getSuccessor(1) == TrueDest)
2174 Opc = Instruction::Or, InvertPredCond = true;
2178 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2182 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2183 IRBuilder<> Builder(PBI);
2185 // If we need to invert the condition in the pred block to match, do so now.
2186 if (InvertPredCond) {
2187 Value *NewCond = PBI->getCondition();
2189 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2190 CmpInst *CI = cast<CmpInst>(NewCond);
2191 CI->setPredicate(CI->getInversePredicate());
2193 NewCond = Builder.CreateNot(NewCond,
2194 PBI->getCondition()->getName()+".not");
2197 PBI->setCondition(NewCond);
2198 PBI->swapSuccessors();
2201 // If we have bonus instructions, clone them into the predecessor block.
2202 // Note that there may be multiple predecessor blocks, so we cannot move
2203 // bonus instructions to a predecessor block.
2204 ValueToValueMapTy VMap; // maps original values to cloned values
2205 // We already make sure Cond is the last instruction before BI. Therefore,
2206 // all instructions before Cond other than DbgInfoIntrinsic are bonus
2208 for (auto BonusInst = BB->begin(); Cond != BonusInst; ++BonusInst) {
2209 if (isa<DbgInfoIntrinsic>(BonusInst))
2211 Instruction *NewBonusInst = BonusInst->clone();
2212 RemapInstruction(NewBonusInst, VMap,
2213 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2214 VMap[&*BonusInst] = NewBonusInst;
2216 // If we moved a load, we cannot any longer claim any knowledge about
2217 // its potential value. The previous information might have been valid
2218 // only given the branch precondition.
2219 // For an analogous reason, we must also drop all the metadata whose
2220 // semantics we don't understand.
2221 NewBonusInst->dropUnknownNonDebugMetadata();
2223 PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2224 NewBonusInst->takeName(&*BonusInst);
2225 BonusInst->setName(BonusInst->getName() + ".old");
2228 // Clone Cond into the predecessor basic block, and or/and the
2229 // two conditions together.
2230 Instruction *New = Cond->clone();
2231 RemapInstruction(New, VMap,
2232 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2233 PredBlock->getInstList().insert(PBI->getIterator(), New);
2234 New->takeName(Cond);
2235 Cond->setName(New->getName() + ".old");
2237 if (BI->isConditional()) {
2238 Instruction *NewCond =
2239 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2241 PBI->setCondition(NewCond);
2243 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2244 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2246 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2248 SmallVector<uint64_t, 8> NewWeights;
2250 if (PBI->getSuccessor(0) == BB) {
2251 if (PredHasWeights && SuccHasWeights) {
2252 // PBI: br i1 %x, BB, FalseDest
2253 // BI: br i1 %y, TrueDest, FalseDest
2254 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2255 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2256 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2257 // TrueWeight for PBI * FalseWeight for BI.
2258 // We assume that total weights of a BranchInst can fit into 32 bits.
2259 // Therefore, we will not have overflow using 64-bit arithmetic.
2260 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2261 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2263 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2264 PBI->setSuccessor(0, TrueDest);
2266 if (PBI->getSuccessor(1) == BB) {
2267 if (PredHasWeights && SuccHasWeights) {
2268 // PBI: br i1 %x, TrueDest, BB
2269 // BI: br i1 %y, TrueDest, FalseDest
2270 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2271 // FalseWeight for PBI * TrueWeight for BI.
2272 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2273 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2274 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2275 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2277 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2278 PBI->setSuccessor(1, FalseDest);
2280 if (NewWeights.size() == 2) {
2281 // Halve the weights if any of them cannot fit in an uint32_t
2282 FitWeights(NewWeights);
2284 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2285 PBI->setMetadata(LLVMContext::MD_prof,
2286 MDBuilder(BI->getContext()).
2287 createBranchWeights(MDWeights));
2289 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2291 // Update PHI nodes in the common successors.
2292 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2293 ConstantInt *PBI_C = cast<ConstantInt>(
2294 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2295 assert(PBI_C->getType()->isIntegerTy(1));
2296 Instruction *MergedCond = nullptr;
2297 if (PBI->getSuccessor(0) == TrueDest) {
2298 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2299 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2300 // is false: !PBI_Cond and BI_Value
2301 Instruction *NotCond =
2302 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2305 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2310 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2311 PBI->getCondition(), MergedCond,
2314 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2315 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2316 // is false: PBI_Cond and BI_Value
2318 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2319 PBI->getCondition(), New,
2321 if (PBI_C->isOne()) {
2322 Instruction *NotCond =
2323 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2326 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2327 NotCond, MergedCond,
2332 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2335 // Change PBI from Conditional to Unconditional.
2336 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2337 EraseTerminatorInstAndDCECond(PBI);
2341 // TODO: If BB is reachable from all paths through PredBlock, then we
2342 // could replace PBI's branch probabilities with BI's.
2344 // Copy any debug value intrinsics into the end of PredBlock.
2345 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2346 if (isa<DbgInfoIntrinsic>(*I))
2347 I->clone()->insertBefore(PBI);
2354 // If there is only one store in BB1 and BB2, return it, otherwise return
2356 static StoreInst *findUniqueStoreInBlocks(BasicBlock *BB1, BasicBlock *BB2) {
2357 StoreInst *S = nullptr;
2358 for (auto *BB : {BB1, BB2}) {
2362 if (auto *SI = dyn_cast<StoreInst>(&I)) {
2364 // Multiple stores seen.
2373 static Value *ensureValueAvailableInSuccessor(Value *V, BasicBlock *BB,
2374 Value *AlternativeV = nullptr) {
2375 // PHI is going to be a PHI node that allows the value V that is defined in
2376 // BB to be referenced in BB's only successor.
2378 // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2379 // doesn't matter to us what the other operand is (it'll never get used). We
2380 // could just create a new PHI with an undef incoming value, but that could
2381 // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2382 // other PHI. So here we directly look for some PHI in BB's successor with V
2383 // as an incoming operand. If we find one, we use it, else we create a new
2386 // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2387 // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2388 // where OtherBB is the single other predecessor of BB's only successor.
2389 PHINode *PHI = nullptr;
2390 BasicBlock *Succ = BB->getSingleSuccessor();
2392 for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2393 if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2394 PHI = cast<PHINode>(I);
2398 assert(std::distance(pred_begin(Succ), pred_end(Succ)) == 2);
2399 auto PredI = pred_begin(Succ);
2400 BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2401 if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2408 PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2409 PHI->addIncoming(V, BB);
2410 for (BasicBlock *PredBB : predecessors(Succ))
2412 PHI->addIncoming(AlternativeV ? AlternativeV : UndefValue::get(V->getType()),
2417 static bool mergeConditionalStoreToAddress(BasicBlock *PTB, BasicBlock *PFB,
2418 BasicBlock *QTB, BasicBlock *QFB,
2419 BasicBlock *PostBB, Value *Address,
2420 bool InvertPCond, bool InvertQCond) {
2421 auto IsaBitcastOfPointerType = [](const Instruction &I) {
2422 return Operator::getOpcode(&I) == Instruction::BitCast &&
2423 I.getType()->isPointerTy();
2426 // If we're not in aggressive mode, we only optimize if we have some
2427 // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2428 auto IsWorthwhile = [&](BasicBlock *BB) {
2431 // Heuristic: if the block can be if-converted/phi-folded and the
2432 // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2433 // thread this store.
2435 for (auto &I : *BB) {
2436 // Cheap instructions viable for folding.
2437 if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2440 // Free instructions.
2441 else if (isa<TerminatorInst>(I) || isa<DbgInfoIntrinsic>(I) ||
2442 IsaBitcastOfPointerType(I))
2447 return N <= PHINodeFoldingThreshold;
2450 if (!MergeCondStoresAggressively && (!IsWorthwhile(PTB) ||
2451 !IsWorthwhile(PFB) ||
2452 !IsWorthwhile(QTB) ||
2453 !IsWorthwhile(QFB)))
2456 // For every pointer, there must be exactly two stores, one coming from
2457 // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2458 // store (to any address) in PTB,PFB or QTB,QFB.
2459 // FIXME: We could relax this restriction with a bit more work and performance
2461 StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2462 StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2463 if (!PStore || !QStore)
2466 // Now check the stores are compatible.
2467 if (!QStore->isUnordered() || !PStore->isUnordered())
2470 // Check that sinking the store won't cause program behavior changes. Sinking
2471 // the store out of the Q blocks won't change any behavior as we're sinking
2472 // from a block to its unconditional successor. But we're moving a store from
2473 // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2474 // So we need to check that there are no aliasing loads or stores in
2475 // QBI, QTB and QFB. We also need to check there are no conflicting memory
2476 // operations between PStore and the end of its parent block.
2478 // The ideal way to do this is to query AliasAnalysis, but we don't
2479 // preserve AA currently so that is dangerous. Be super safe and just
2480 // check there are no other memory operations at all.
2481 for (auto &I : *QFB->getSinglePredecessor())
2482 if (I.mayReadOrWriteMemory())
2484 for (auto &I : *QFB)
2485 if (&I != QStore && I.mayReadOrWriteMemory())
2488 for (auto &I : *QTB)
2489 if (&I != QStore && I.mayReadOrWriteMemory())
2491 for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2493 if (&*I != PStore && I->mayReadOrWriteMemory())
2496 // OK, we're going to sink the stores to PostBB. The store has to be
2497 // conditional though, so first create the predicate.
2498 Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
2500 Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
2503 Value *PPHI = ensureValueAvailableInSuccessor(PStore->getValueOperand(),
2504 PStore->getParent());
2505 Value *QPHI = ensureValueAvailableInSuccessor(QStore->getValueOperand(),
2506 QStore->getParent(), PPHI);
2508 IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
2510 Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
2511 Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
2514 PPred = QB.CreateNot(PPred);
2516 QPred = QB.CreateNot(QPred);
2517 Value *CombinedPred = QB.CreateOr(PPred, QPred);
2520 SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
2521 QB.SetInsertPoint(T);
2522 StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
2524 PStore->getAAMetadata(AAMD, /*Merge=*/false);
2525 PStore->getAAMetadata(AAMD, /*Merge=*/true);
2526 SI->setAAMetadata(AAMD);
2528 QStore->eraseFromParent();
2529 PStore->eraseFromParent();
2534 static bool mergeConditionalStores(BranchInst *PBI, BranchInst *QBI) {
2535 // The intention here is to find diamonds or triangles (see below) where each
2536 // conditional block contains a store to the same address. Both of these
2537 // stores are conditional, so they can't be unconditionally sunk. But it may
2538 // be profitable to speculatively sink the stores into one merged store at the
2539 // end, and predicate the merged store on the union of the two conditions of
2542 // This can reduce the number of stores executed if both of the conditions are
2543 // true, and can allow the blocks to become small enough to be if-converted.
2544 // This optimization will also chain, so that ladders of test-and-set
2545 // sequences can be if-converted away.
2547 // We only deal with simple diamonds or triangles:
2549 // PBI or PBI or a combination of the two
2559 // We model triangles as a type of diamond with a nullptr "true" block.
2560 // Triangles are canonicalized so that the fallthrough edge is represented by
2561 // a true condition, as in the diagram above.
2563 BasicBlock *PTB = PBI->getSuccessor(0);
2564 BasicBlock *PFB = PBI->getSuccessor(1);
2565 BasicBlock *QTB = QBI->getSuccessor(0);
2566 BasicBlock *QFB = QBI->getSuccessor(1);
2567 BasicBlock *PostBB = QFB->getSingleSuccessor();
2569 bool InvertPCond = false, InvertQCond = false;
2570 // Canonicalize fallthroughs to the true branches.
2571 if (PFB == QBI->getParent()) {
2572 std::swap(PFB, PTB);
2575 if (QFB == PostBB) {
2576 std::swap(QFB, QTB);
2580 // From this point on we can assume PTB or QTB may be fallthroughs but PFB
2581 // and QFB may not. Model fallthroughs as a nullptr block.
2582 if (PTB == QBI->getParent())
2587 // Legality bailouts. We must have at least the non-fallthrough blocks and
2588 // the post-dominating block, and the non-fallthroughs must only have one
2590 auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
2591 return BB->getSinglePredecessor() == P &&
2592 BB->getSingleSuccessor() == S;
2595 !HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
2596 !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
2598 if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
2599 (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
2601 if (PostBB->getNumUses() != 2 || QBI->getParent()->getNumUses() != 2)
2604 // OK, this is a sequence of two diamonds or triangles.
2605 // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
2606 SmallPtrSet<Value *,4> PStoreAddresses, QStoreAddresses;
2607 for (auto *BB : {PTB, PFB}) {
2611 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
2612 PStoreAddresses.insert(SI->getPointerOperand());
2614 for (auto *BB : {QTB, QFB}) {
2618 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
2619 QStoreAddresses.insert(SI->getPointerOperand());
2622 set_intersect(PStoreAddresses, QStoreAddresses);
2623 // set_intersect mutates PStoreAddresses in place. Rename it here to make it
2624 // clear what it contains.
2625 auto &CommonAddresses = PStoreAddresses;
2627 bool Changed = false;
2628 for (auto *Address : CommonAddresses)
2629 Changed |= mergeConditionalStoreToAddress(
2630 PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond);
2634 /// If we have a conditional branch as a predecessor of another block,
2635 /// this function tries to simplify it. We know
2636 /// that PBI and BI are both conditional branches, and BI is in one of the
2637 /// successor blocks of PBI - PBI branches to BI.
2638 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
2639 const DataLayout &DL) {
2640 assert(PBI->isConditional() && BI->isConditional());
2641 BasicBlock *BB = BI->getParent();
2643 // If this block ends with a branch instruction, and if there is a
2644 // predecessor that ends on a branch of the same condition, make
2645 // this conditional branch redundant.
2646 if (PBI->getCondition() == BI->getCondition() &&
2647 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2648 // Okay, the outcome of this conditional branch is statically
2649 // knowable. If this block had a single pred, handle specially.
2650 if (BB->getSinglePredecessor()) {
2651 // Turn this into a branch on constant.
2652 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2653 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2655 return true; // Nuke the branch on constant.
2658 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2659 // in the constant and simplify the block result. Subsequent passes of
2660 // simplifycfg will thread the block.
2661 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2662 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2663 PHINode *NewPN = PHINode::Create(
2664 Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
2665 BI->getCondition()->getName() + ".pr", &BB->front());
2666 // Okay, we're going to insert the PHI node. Since PBI is not the only
2667 // predecessor, compute the PHI'd conditional value for all of the preds.
2668 // Any predecessor where the condition is not computable we keep symbolic.
2669 for (pred_iterator PI = PB; PI != PE; ++PI) {
2670 BasicBlock *P = *PI;
2671 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2672 PBI != BI && PBI->isConditional() &&
2673 PBI->getCondition() == BI->getCondition() &&
2674 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2675 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2676 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2679 NewPN->addIncoming(BI->getCondition(), P);
2683 BI->setCondition(NewPN);
2688 if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2692 // If BI is reached from the true path of PBI and PBI's condition implies
2693 // BI's condition, we know the direction of the BI branch.
2694 if (PBI->getSuccessor(0) == BI->getParent() &&
2695 isImpliedCondition(PBI->getCondition(), BI->getCondition(), DL) &&
2696 PBI->getSuccessor(0) != PBI->getSuccessor(1) &&
2697 BB->getSinglePredecessor()) {
2698 // Turn this into a branch on constant.
2699 auto *OldCond = BI->getCondition();
2700 BI->setCondition(ConstantInt::getTrue(BB->getContext()));
2701 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
2702 return true; // Nuke the branch on constant.
2705 // If both branches are conditional and both contain stores to the same
2706 // address, remove the stores from the conditionals and create a conditional
2707 // merged store at the end.
2708 if (MergeCondStores && mergeConditionalStores(PBI, BI))
2711 // If this is a conditional branch in an empty block, and if any
2712 // predecessors are a conditional branch to one of our destinations,
2713 // fold the conditions into logical ops and one cond br.
2714 BasicBlock::iterator BBI = BB->begin();
2715 // Ignore dbg intrinsics.
2716 while (isa<DbgInfoIntrinsic>(BBI))
2722 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2724 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2725 PBIOp = 0, BIOp = 1;
2726 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2727 PBIOp = 1, BIOp = 0;
2728 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2733 // Check to make sure that the other destination of this branch
2734 // isn't BB itself. If so, this is an infinite loop that will
2735 // keep getting unwound.
2736 if (PBI->getSuccessor(PBIOp) == BB)
2739 // Do not perform this transformation if it would require
2740 // insertion of a large number of select instructions. For targets
2741 // without predication/cmovs, this is a big pessimization.
2743 // Also do not perform this transformation if any phi node in the common
2744 // destination block can trap when reached by BB or PBB (PR17073). In that
2745 // case, it would be unsafe to hoist the operation into a select instruction.
2747 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2748 unsigned NumPhis = 0;
2749 for (BasicBlock::iterator II = CommonDest->begin();
2750 isa<PHINode>(II); ++II, ++NumPhis) {
2751 if (NumPhis > 2) // Disable this xform.
2754 PHINode *PN = cast<PHINode>(II);
2755 Value *BIV = PN->getIncomingValueForBlock(BB);
2756 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
2760 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2761 Value *PBIV = PN->getIncomingValue(PBBIdx);
2762 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
2767 // Finally, if everything is ok, fold the branches to logical ops.
2768 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2770 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2771 << "AND: " << *BI->getParent());
2774 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2775 // branch in it, where one edge (OtherDest) goes back to itself but the other
2776 // exits. We don't *know* that the program avoids the infinite loop
2777 // (even though that seems likely). If we do this xform naively, we'll end up
2778 // recursively unpeeling the loop. Since we know that (after the xform is
2779 // done) that the block *is* infinite if reached, we just make it an obviously
2780 // infinite loop with no cond branch.
2781 if (OtherDest == BB) {
2782 // Insert it at the end of the function, because it's either code,
2783 // or it won't matter if it's hot. :)
2784 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2785 "infloop", BB->getParent());
2786 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2787 OtherDest = InfLoopBlock;
2790 DEBUG(dbgs() << *PBI->getParent()->getParent());
2792 // BI may have other predecessors. Because of this, we leave
2793 // it alone, but modify PBI.
2795 // Make sure we get to CommonDest on True&True directions.
2796 Value *PBICond = PBI->getCondition();
2797 IRBuilder<true, NoFolder> Builder(PBI);
2799 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2801 Value *BICond = BI->getCondition();
2803 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2805 // Merge the conditions.
2806 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2808 // Modify PBI to branch on the new condition to the new dests.
2809 PBI->setCondition(Cond);
2810 PBI->setSuccessor(0, CommonDest);
2811 PBI->setSuccessor(1, OtherDest);
2813 // Update branch weight for PBI.
2814 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2815 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2817 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2819 if (PredHasWeights && SuccHasWeights) {
2820 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2821 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2822 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2823 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2824 // The weight to CommonDest should be PredCommon * SuccTotal +
2825 // PredOther * SuccCommon.
2826 // The weight to OtherDest should be PredOther * SuccOther.
2827 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
2828 PredOther * SuccCommon,
2829 PredOther * SuccOther};
2830 // Halve the weights if any of them cannot fit in an uint32_t
2831 FitWeights(NewWeights);
2833 PBI->setMetadata(LLVMContext::MD_prof,
2834 MDBuilder(BI->getContext())
2835 .createBranchWeights(NewWeights[0], NewWeights[1]));
2838 // OtherDest may have phi nodes. If so, add an entry from PBI's
2839 // block that are identical to the entries for BI's block.
2840 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2842 // We know that the CommonDest already had an edge from PBI to
2843 // it. If it has PHIs though, the PHIs may have different
2844 // entries for BB and PBI's BB. If so, insert a select to make
2847 for (BasicBlock::iterator II = CommonDest->begin();
2848 (PN = dyn_cast<PHINode>(II)); ++II) {
2849 Value *BIV = PN->getIncomingValueForBlock(BB);
2850 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2851 Value *PBIV = PN->getIncomingValue(PBBIdx);
2853 // Insert a select in PBI to pick the right value.
2854 Value *NV = cast<SelectInst>
2855 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2856 PN->setIncomingValue(PBBIdx, NV);
2860 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2861 DEBUG(dbgs() << *PBI->getParent()->getParent());
2863 // This basic block is probably dead. We know it has at least
2864 // one fewer predecessor.
2868 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
2869 // true or to FalseBB if Cond is false.
2870 // Takes care of updating the successors and removing the old terminator.
2871 // Also makes sure not to introduce new successors by assuming that edges to
2872 // non-successor TrueBBs and FalseBBs aren't reachable.
2873 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2874 BasicBlock *TrueBB, BasicBlock *FalseBB,
2875 uint32_t TrueWeight,
2876 uint32_t FalseWeight){
2877 // Remove any superfluous successor edges from the CFG.
2878 // First, figure out which successors to preserve.
2879 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2881 BasicBlock *KeepEdge1 = TrueBB;
2882 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
2884 // Then remove the rest.
2885 for (BasicBlock *Succ : OldTerm->successors()) {
2886 // Make sure only to keep exactly one copy of each edge.
2887 if (Succ == KeepEdge1)
2888 KeepEdge1 = nullptr;
2889 else if (Succ == KeepEdge2)
2890 KeepEdge2 = nullptr;
2892 Succ->removePredecessor(OldTerm->getParent(),
2893 /*DontDeleteUselessPHIs=*/true);
2896 IRBuilder<> Builder(OldTerm);
2897 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2899 // Insert an appropriate new terminator.
2900 if (!KeepEdge1 && !KeepEdge2) {
2901 if (TrueBB == FalseBB)
2902 // We were only looking for one successor, and it was present.
2903 // Create an unconditional branch to it.
2904 Builder.CreateBr(TrueBB);
2906 // We found both of the successors we were looking for.
2907 // Create a conditional branch sharing the condition of the select.
2908 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2909 if (TrueWeight != FalseWeight)
2910 NewBI->setMetadata(LLVMContext::MD_prof,
2911 MDBuilder(OldTerm->getContext()).
2912 createBranchWeights(TrueWeight, FalseWeight));
2914 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2915 // Neither of the selected blocks were successors, so this
2916 // terminator must be unreachable.
2917 new UnreachableInst(OldTerm->getContext(), OldTerm);
2919 // One of the selected values was a successor, but the other wasn't.
2920 // Insert an unconditional branch to the one that was found;
2921 // the edge to the one that wasn't must be unreachable.
2923 // Only TrueBB was found.
2924 Builder.CreateBr(TrueBB);
2926 // Only FalseBB was found.
2927 Builder.CreateBr(FalseBB);
2930 EraseTerminatorInstAndDCECond(OldTerm);
2935 // (switch (select cond, X, Y)) on constant X, Y
2936 // with a branch - conditional if X and Y lead to distinct BBs,
2937 // unconditional otherwise.
2938 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2939 // Check for constant integer values in the select.
2940 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2941 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2942 if (!TrueVal || !FalseVal)
2945 // Find the relevant condition and destinations.
2946 Value *Condition = Select->getCondition();
2947 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2948 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2950 // Get weight for TrueBB and FalseBB.
2951 uint32_t TrueWeight = 0, FalseWeight = 0;
2952 SmallVector<uint64_t, 8> Weights;
2953 bool HasWeights = HasBranchWeights(SI);
2955 GetBranchWeights(SI, Weights);
2956 if (Weights.size() == 1 + SI->getNumCases()) {
2957 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2958 getSuccessorIndex()];
2959 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2960 getSuccessorIndex()];
2964 // Perform the actual simplification.
2965 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2966 TrueWeight, FalseWeight);
2970 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2971 // blockaddress(@fn, BlockB)))
2973 // (br cond, BlockA, BlockB).
2974 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2975 // Check that both operands of the select are block addresses.
2976 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2977 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2981 // Extract the actual blocks.
2982 BasicBlock *TrueBB = TBA->getBasicBlock();
2983 BasicBlock *FalseBB = FBA->getBasicBlock();
2985 // Perform the actual simplification.
2986 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2990 /// This is called when we find an icmp instruction
2991 /// (a seteq/setne with a constant) as the only instruction in a
2992 /// block that ends with an uncond branch. We are looking for a very specific
2993 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2994 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2995 /// default value goes to an uncond block with a seteq in it, we get something
2998 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3000 /// %tmp = icmp eq i8 %A, 92
3003 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3005 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3006 /// the PHI, merging the third icmp into the switch.
3007 static bool TryToSimplifyUncondBranchWithICmpInIt(
3008 ICmpInst *ICI, IRBuilder<> &Builder, const DataLayout &DL,
3009 const TargetTransformInfo &TTI, unsigned BonusInstThreshold,
3010 AssumptionCache *AC) {
3011 BasicBlock *BB = ICI->getParent();
3013 // If the block has any PHIs in it or the icmp has multiple uses, it is too
3015 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
3017 Value *V = ICI->getOperand(0);
3018 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3020 // The pattern we're looking for is where our only predecessor is a switch on
3021 // 'V' and this block is the default case for the switch. In this case we can
3022 // fold the compared value into the switch to simplify things.
3023 BasicBlock *Pred = BB->getSinglePredecessor();
3024 if (!Pred || !isa<SwitchInst>(Pred->getTerminator())) return false;
3026 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3027 if (SI->getCondition() != V)
3030 // If BB is reachable on a non-default case, then we simply know the value of
3031 // V in this block. Substitute it and constant fold the icmp instruction
3033 if (SI->getDefaultDest() != BB) {
3034 ConstantInt *VVal = SI->findCaseDest(BB);
3035 assert(VVal && "Should have a unique destination value");
3036 ICI->setOperand(0, VVal);
3038 if (Value *V = SimplifyInstruction(ICI, DL)) {
3039 ICI->replaceAllUsesWith(V);
3040 ICI->eraseFromParent();
3042 // BB is now empty, so it is likely to simplify away.
3043 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3046 // Ok, the block is reachable from the default dest. If the constant we're
3047 // comparing exists in one of the other edges, then we can constant fold ICI
3049 if (SI->findCaseValue(Cst) != SI->case_default()) {
3051 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3052 V = ConstantInt::getFalse(BB->getContext());
3054 V = ConstantInt::getTrue(BB->getContext());
3056 ICI->replaceAllUsesWith(V);
3057 ICI->eraseFromParent();
3058 // BB is now empty, so it is likely to simplify away.
3059 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3062 // The use of the icmp has to be in the 'end' block, by the only PHI node in
3064 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3065 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3066 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3067 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3070 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3072 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3073 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3075 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3076 std::swap(DefaultCst, NewCst);
3078 // Replace ICI (which is used by the PHI for the default value) with true or
3079 // false depending on if it is EQ or NE.
3080 ICI->replaceAllUsesWith(DefaultCst);
3081 ICI->eraseFromParent();
3083 // Okay, the switch goes to this block on a default value. Add an edge from
3084 // the switch to the merge point on the compared value.
3085 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
3086 BB->getParent(), BB);
3087 SmallVector<uint64_t, 8> Weights;
3088 bool HasWeights = HasBranchWeights(SI);
3090 GetBranchWeights(SI, Weights);
3091 if (Weights.size() == 1 + SI->getNumCases()) {
3092 // Split weight for default case to case for "Cst".
3093 Weights[0] = (Weights[0]+1) >> 1;
3094 Weights.push_back(Weights[0]);
3096 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3097 SI->setMetadata(LLVMContext::MD_prof,
3098 MDBuilder(SI->getContext()).
3099 createBranchWeights(MDWeights));
3102 SI->addCase(Cst, NewBB);
3104 // NewBB branches to the phi block, add the uncond branch and the phi entry.
3105 Builder.SetInsertPoint(NewBB);
3106 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3107 Builder.CreateBr(SuccBlock);
3108 PHIUse->addIncoming(NewCst, NewBB);
3112 /// The specified branch is a conditional branch.
3113 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3114 /// fold it into a switch instruction if so.
3115 static bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
3116 const DataLayout &DL) {
3117 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3118 if (!Cond) return false;
3120 // Change br (X == 0 | X == 1), T, F into a switch instruction.
3121 // If this is a bunch of seteq's or'd together, or if it's a bunch of
3122 // 'setne's and'ed together, collect them.
3124 // Try to gather values from a chain of and/or to be turned into a switch
3125 ConstantComparesGatherer ConstantCompare(Cond, DL);
3126 // Unpack the result
3127 SmallVectorImpl<ConstantInt*> &Values = ConstantCompare.Vals;
3128 Value *CompVal = ConstantCompare.CompValue;
3129 unsigned UsedICmps = ConstantCompare.UsedICmps;
3130 Value *ExtraCase = ConstantCompare.Extra;
3132 // If we didn't have a multiply compared value, fail.
3133 if (!CompVal) return false;
3135 // Avoid turning single icmps into a switch.
3139 bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3141 // There might be duplicate constants in the list, which the switch
3142 // instruction can't handle, remove them now.
3143 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3144 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3146 // If Extra was used, we require at least two switch values to do the
3147 // transformation. A switch with one value is just a conditional branch.
3148 if (ExtraCase && Values.size() < 2) return false;
3150 // TODO: Preserve branch weight metadata, similarly to how
3151 // FoldValueComparisonIntoPredecessors preserves it.
3153 // Figure out which block is which destination.
3154 BasicBlock *DefaultBB = BI->getSuccessor(1);
3155 BasicBlock *EdgeBB = BI->getSuccessor(0);
3156 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
3158 BasicBlock *BB = BI->getParent();
3160 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3161 << " cases into SWITCH. BB is:\n" << *BB);
3163 // If there are any extra values that couldn't be folded into the switch
3164 // then we evaluate them with an explicit branch first. Split the block
3165 // right before the condbr to handle it.
3168 BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3169 // Remove the uncond branch added to the old block.
3170 TerminatorInst *OldTI = BB->getTerminator();
3171 Builder.SetInsertPoint(OldTI);
3174 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3176 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3178 OldTI->eraseFromParent();
3180 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3181 // for the edge we just added.
3182 AddPredecessorToBlock(EdgeBB, BB, NewBB);
3184 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3185 << "\nEXTRABB = " << *BB);
3189 Builder.SetInsertPoint(BI);
3190 // Convert pointer to int before we switch.
3191 if (CompVal->getType()->isPointerTy()) {
3192 CompVal = Builder.CreatePtrToInt(
3193 CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3196 // Create the new switch instruction now.
3197 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3199 // Add all of the 'cases' to the switch instruction.
3200 for (unsigned i = 0, e = Values.size(); i != e; ++i)
3201 New->addCase(Values[i], EdgeBB);
3203 // We added edges from PI to the EdgeBB. As such, if there were any
3204 // PHI nodes in EdgeBB, they need entries to be added corresponding to
3205 // the number of edges added.
3206 for (BasicBlock::iterator BBI = EdgeBB->begin();
3207 isa<PHINode>(BBI); ++BBI) {
3208 PHINode *PN = cast<PHINode>(BBI);
3209 Value *InVal = PN->getIncomingValueForBlock(BB);
3210 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
3211 PN->addIncoming(InVal, BB);
3214 // Erase the old branch instruction.
3215 EraseTerminatorInstAndDCECond(BI);
3217 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3221 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3222 // If this is a trivial landing pad that just continues unwinding the caught
3223 // exception then zap the landing pad, turning its invokes into calls.
3224 BasicBlock *BB = RI->getParent();
3225 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
3226 if (RI->getValue() != LPInst)
3227 // Not a landing pad, or the resume is not unwinding the exception that
3228 // caused control to branch here.
3231 // Check that there are no other instructions except for debug intrinsics.
3232 BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3234 if (!isa<DbgInfoIntrinsic>(I))
3237 // Turn all invokes that unwind here into calls and delete the basic block.
3238 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3239 BasicBlock *Pred = *PI++;
3240 removeUnwindEdge(Pred);
3243 // The landingpad is now unreachable. Zap it.
3244 BB->eraseFromParent();
3248 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
3249 // If this is a trivial cleanup pad that executes no instructions, it can be
3250 // eliminated. If the cleanup pad continues to the caller, any predecessor
3251 // that is an EH pad will be updated to continue to the caller and any
3252 // predecessor that terminates with an invoke instruction will have its invoke
3253 // instruction converted to a call instruction. If the cleanup pad being
3254 // simplified does not continue to the caller, each predecessor will be
3255 // updated to continue to the unwind destination of the cleanup pad being
3257 BasicBlock *BB = RI->getParent();
3258 Instruction *CPInst = dyn_cast<CleanupPadInst>(BB->getFirstNonPHI());
3260 // This isn't an empty cleanup.
3263 // Check that there are no other instructions except for debug intrinsics.
3264 BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3266 if (!isa<DbgInfoIntrinsic>(I))
3269 // If the cleanup return we are simplifying unwinds to the caller, this
3270 // will set UnwindDest to nullptr.
3271 BasicBlock *UnwindDest = RI->getUnwindDest();
3273 // We're about to remove BB from the control flow. Before we do, sink any
3274 // PHINodes into the unwind destination. Doing this before changing the
3275 // control flow avoids some potentially slow checks, since we can currently
3276 // be certain that UnwindDest and BB have no common predecessors (since they
3277 // are both EH pads).
3279 // First, go through the PHI nodes in UnwindDest and update any nodes that
3280 // reference the block we are removing
3281 for (BasicBlock::iterator I = UnwindDest->begin(),
3282 IE = UnwindDest->getFirstNonPHI()->getIterator();
3284 PHINode *DestPN = cast<PHINode>(I);
3286 int Idx = DestPN->getBasicBlockIndex(BB);
3287 // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3289 // This PHI node has an incoming value that corresponds to a control
3290 // path through the cleanup pad we are removing. If the incoming
3291 // value is in the cleanup pad, it must be a PHINode (because we
3292 // verified above that the block is otherwise empty). Otherwise, the
3293 // value is either a constant or a value that dominates the cleanup
3294 // pad being removed.
3296 // Because BB and UnwindDest are both EH pads, all of their
3297 // predecessors must unwind to these blocks, and since no instruction
3298 // can have multiple unwind destinations, there will be no overlap in
3299 // incoming blocks between SrcPN and DestPN.
3300 Value *SrcVal = DestPN->getIncomingValue(Idx);
3301 PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3303 // Remove the entry for the block we are deleting.
3304 DestPN->removeIncomingValue(Idx, false);
3306 if (SrcPN && SrcPN->getParent() == BB) {
3307 // If the incoming value was a PHI node in the cleanup pad we are
3308 // removing, we need to merge that PHI node's incoming values into
3310 for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3311 SrcIdx != SrcE; ++SrcIdx) {
3312 DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3313 SrcPN->getIncomingBlock(SrcIdx));
3316 // Otherwise, the incoming value came from above BB and
3317 // so we can just reuse it. We must associate all of BB's
3318 // predecessors with this value.
3319 for (auto *pred : predecessors(BB)) {
3320 DestPN->addIncoming(SrcVal, pred);
3325 // Sink any remaining PHI nodes directly into UnwindDest.
3326 Instruction *InsertPt = UnwindDest->getFirstNonPHI();
3327 for (BasicBlock::iterator I = BB->begin(),
3328 IE = BB->getFirstNonPHI()->getIterator();
3330 // The iterator must be incremented here because the instructions are
3331 // being moved to another block.
3332 PHINode *PN = cast<PHINode>(I++);
3333 if (PN->use_empty())
3334 // If the PHI node has no uses, just leave it. It will be erased
3335 // when we erase BB below.
3338 // Otherwise, sink this PHI node into UnwindDest.
3339 // Any predecessors to UnwindDest which are not already represented
3340 // must be back edges which inherit the value from the path through
3341 // BB. In this case, the PHI value must reference itself.
3342 for (auto *pred : predecessors(UnwindDest))
3344 PN->addIncoming(PN, pred);
3345 PN->moveBefore(InsertPt);
3349 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3350 // The iterator must be updated here because we are removing this pred.
3351 BasicBlock *PredBB = *PI++;
3352 if (UnwindDest == nullptr) {
3353 removeUnwindEdge(PredBB);
3355 TerminatorInst *TI = PredBB->getTerminator();
3356 TI->replaceUsesOfWith(BB, UnwindDest);
3360 // The cleanup pad is now unreachable. Zap it.
3361 BB->eraseFromParent();
3365 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
3366 BasicBlock *BB = RI->getParent();
3367 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
3369 // Find predecessors that end with branches.
3370 SmallVector<BasicBlock*, 8> UncondBranchPreds;
3371 SmallVector<BranchInst*, 8> CondBranchPreds;
3372 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
3373 BasicBlock *P = *PI;
3374 TerminatorInst *PTI = P->getTerminator();
3375 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
3376 if (BI->isUnconditional())
3377 UncondBranchPreds.push_back(P);
3379 CondBranchPreds.push_back(BI);
3383 // If we found some, do the transformation!
3384 if (!UncondBranchPreds.empty() && DupRet) {
3385 while (!UncondBranchPreds.empty()) {
3386 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
3387 DEBUG(dbgs() << "FOLDING: " << *BB
3388 << "INTO UNCOND BRANCH PRED: " << *Pred);
3389 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
3392 // If we eliminated all predecessors of the block, delete the block now.
3394 // We know there are no successors, so just nuke the block.
3395 BB->eraseFromParent();
3400 // Check out all of the conditional branches going to this return
3401 // instruction. If any of them just select between returns, change the
3402 // branch itself into a select/return pair.
3403 while (!CondBranchPreds.empty()) {
3404 BranchInst *BI = CondBranchPreds.pop_back_val();
3406 // Check to see if the non-BB successor is also a return block.
3407 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
3408 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
3409 SimplifyCondBranchToTwoReturns(BI, Builder))
3415 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
3416 BasicBlock *BB = UI->getParent();
3418 bool Changed = false;
3420 // If there are any instructions immediately before the unreachable that can
3421 // be removed, do so.
3422 while (UI->getIterator() != BB->begin()) {
3423 BasicBlock::iterator BBI = UI->getIterator();
3425 // Do not delete instructions that can have side effects which might cause
3426 // the unreachable to not be reachable; specifically, calls and volatile
3427 // operations may have this effect.
3428 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
3430 if (BBI->mayHaveSideEffects()) {
3431 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
3432 if (SI->isVolatile())
3434 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
3435 if (LI->isVolatile())
3437 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
3438 if (RMWI->isVolatile())
3440 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
3441 if (CXI->isVolatile())
3443 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
3444 !isa<LandingPadInst>(BBI)) {
3447 // Note that deleting LandingPad's here is in fact okay, although it
3448 // involves a bit of subtle reasoning. If this inst is a LandingPad,
3449 // all the predecessors of this block will be the unwind edges of Invokes,
3450 // and we can therefore guarantee this block will be erased.
3453 // Delete this instruction (any uses are guaranteed to be dead)
3454 if (!BBI->use_empty())
3455 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
3456 BBI->eraseFromParent();
3460 // If the unreachable instruction is the first in the block, take a gander
3461 // at all of the predecessors of this instruction, and simplify them.
3462 if (&BB->front() != UI) return Changed;
3464 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
3465 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
3466 TerminatorInst *TI = Preds[i]->getTerminator();
3467 IRBuilder<> Builder(TI);
3468 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3469 if (BI->isUnconditional()) {
3470 if (BI->getSuccessor(0) == BB) {
3471 new UnreachableInst(TI->getContext(), TI);
3472 TI->eraseFromParent();
3476 if (BI->getSuccessor(0) == BB) {
3477 Builder.CreateBr(BI->getSuccessor(1));
3478 EraseTerminatorInstAndDCECond(BI);
3479 } else if (BI->getSuccessor(1) == BB) {
3480 Builder.CreateBr(BI->getSuccessor(0));
3481 EraseTerminatorInstAndDCECond(BI);
3485 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3486 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3488 if (i.getCaseSuccessor() == BB) {
3489 BB->removePredecessor(SI->getParent());
3494 } else if ((isa<InvokeInst>(TI) &&
3495 cast<InvokeInst>(TI)->getUnwindDest() == BB) ||
3496 isa<CatchEndPadInst>(TI) || isa<TerminatePadInst>(TI)) {
3497 removeUnwindEdge(TI->getParent());
3499 } else if (isa<CleanupReturnInst>(TI) || isa<CleanupEndPadInst>(TI)) {
3500 new UnreachableInst(TI->getContext(), TI);
3501 TI->eraseFromParent();
3504 // TODO: If TI is a CatchPadInst, then (BB must be its normal dest and)
3505 // we can eliminate it, redirecting its preds to its unwind successor,
3506 // or to the next outer handler if the removed catch is the last for its
3510 // If this block is now dead, remove it.
3511 if (pred_empty(BB) &&
3512 BB != &BB->getParent()->getEntryBlock()) {
3513 // We know there are no successors, so just nuke the block.
3514 BB->eraseFromParent();
3521 static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) {
3522 assert(Cases.size() >= 1);
3524 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3525 for (size_t I = 1, E = Cases.size(); I != E; ++I) {
3526 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
3532 /// Turn a switch with two reachable destinations into an integer range
3533 /// comparison and branch.
3534 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3535 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3538 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
3540 // Partition the cases into two sets with different destinations.
3541 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
3542 BasicBlock *DestB = nullptr;
3543 SmallVector <ConstantInt *, 16> CasesA;
3544 SmallVector <ConstantInt *, 16> CasesB;
3546 for (SwitchInst::CaseIt I : SI->cases()) {
3547 BasicBlock *Dest = I.getCaseSuccessor();
3548 if (!DestA) DestA = Dest;
3549 if (Dest == DestA) {
3550 CasesA.push_back(I.getCaseValue());
3553 if (!DestB) DestB = Dest;
3554 if (Dest == DestB) {
3555 CasesB.push_back(I.getCaseValue());
3558 return false; // More than two destinations.
3561 assert(DestA && DestB && "Single-destination switch should have been folded.");
3562 assert(DestA != DestB);
3563 assert(DestB != SI->getDefaultDest());
3564 assert(!CasesB.empty() && "There must be non-default cases.");
3565 assert(!CasesA.empty() || HasDefault);
3567 // Figure out if one of the sets of cases form a contiguous range.
3568 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
3569 BasicBlock *ContiguousDest = nullptr;
3570 BasicBlock *OtherDest = nullptr;
3571 if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
3572 ContiguousCases = &CasesA;
3573 ContiguousDest = DestA;
3575 } else if (CasesAreContiguous(CasesB)) {
3576 ContiguousCases = &CasesB;
3577 ContiguousDest = DestB;
3582 // Start building the compare and branch.
3584 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
3585 Constant *NumCases = ConstantInt::get(Offset->getType(), ContiguousCases->size());
3587 Value *Sub = SI->getCondition();
3588 if (!Offset->isNullValue())
3589 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
3592 // If NumCases overflowed, then all possible values jump to the successor.
3593 if (NumCases->isNullValue() && !ContiguousCases->empty())
3594 Cmp = ConstantInt::getTrue(SI->getContext());
3596 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3597 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
3599 // Update weight for the newly-created conditional branch.
3600 if (HasBranchWeights(SI)) {
3601 SmallVector<uint64_t, 8> Weights;
3602 GetBranchWeights(SI, Weights);
3603 if (Weights.size() == 1 + SI->getNumCases()) {
3604 uint64_t TrueWeight = 0;
3605 uint64_t FalseWeight = 0;
3606 for (size_t I = 0, E = Weights.size(); I != E; ++I) {
3607 if (SI->getSuccessor(I) == ContiguousDest)
3608 TrueWeight += Weights[I];
3610 FalseWeight += Weights[I];
3612 while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
3616 NewBI->setMetadata(LLVMContext::MD_prof,
3617 MDBuilder(SI->getContext()).createBranchWeights(
3618 (uint32_t)TrueWeight, (uint32_t)FalseWeight));
3622 // Prune obsolete incoming values off the successors' PHI nodes.
3623 for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
3624 unsigned PreviousEdges = ContiguousCases->size();
3625 if (ContiguousDest == SI->getDefaultDest()) ++PreviousEdges;
3626 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
3627 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3629 for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
3630 unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
3631 if (OtherDest == SI->getDefaultDest()) ++PreviousEdges;
3632 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
3633 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3637 SI->eraseFromParent();
3642 /// Compute masked bits for the condition of a switch
3643 /// and use it to remove dead cases.
3644 static bool EliminateDeadSwitchCases(SwitchInst *SI, AssumptionCache *AC,
3645 const DataLayout &DL) {
3646 Value *Cond = SI->getCondition();
3647 unsigned Bits = Cond->getType()->getIntegerBitWidth();
3648 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3649 computeKnownBits(Cond, KnownZero, KnownOne, DL, 0, AC, SI);
3651 // Gather dead cases.
3652 SmallVector<ConstantInt*, 8> DeadCases;
3653 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3654 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3655 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3656 DeadCases.push_back(I.getCaseValue());
3657 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3658 << I.getCaseValue() << "' is dead.\n");
3662 // If we can prove that the cases must cover all possible values, the
3663 // default destination becomes dead and we can remove it. If we know some
3664 // of the bits in the value, we can use that to more precisely compute the
3665 // number of possible unique case values.
3667 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
3668 const unsigned NumUnknownBits = Bits -
3669 (KnownZero.Or(KnownOne)).countPopulation();
3670 assert(NumUnknownBits <= Bits);
3671 if (HasDefault && DeadCases.empty() &&
3672 NumUnknownBits < 64 /* avoid overflow */ &&
3673 SI->getNumCases() == (1ULL << NumUnknownBits)) {
3674 DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
3675 BasicBlock *NewDefault = SplitBlockPredecessors(SI->getDefaultDest(),
3676 SI->getParent(), "");
3677 SI->setDefaultDest(&*NewDefault);
3678 SplitBlock(&*NewDefault, &NewDefault->front());
3679 auto *OldTI = NewDefault->getTerminator();
3680 new UnreachableInst(SI->getContext(), OldTI);
3681 EraseTerminatorInstAndDCECond(OldTI);
3685 SmallVector<uint64_t, 8> Weights;
3686 bool HasWeight = HasBranchWeights(SI);
3688 GetBranchWeights(SI, Weights);
3689 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3692 // Remove dead cases from the switch.
3693 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3694 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3695 assert(Case != SI->case_default() &&
3696 "Case was not found. Probably mistake in DeadCases forming.");
3698 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3702 // Prune unused values from PHI nodes.
3703 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3704 SI->removeCase(Case);
3706 if (HasWeight && Weights.size() >= 2) {
3707 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3708 SI->setMetadata(LLVMContext::MD_prof,
3709 MDBuilder(SI->getParent()->getContext()).
3710 createBranchWeights(MDWeights));
3713 return !DeadCases.empty();
3716 /// If BB would be eligible for simplification by
3717 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3718 /// by an unconditional branch), look at the phi node for BB in the successor
3719 /// block and see if the incoming value is equal to CaseValue. If so, return
3720 /// the phi node, and set PhiIndex to BB's index in the phi node.
3721 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3724 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3725 return nullptr; // BB must be empty to be a candidate for simplification.
3726 if (!BB->getSinglePredecessor())
3727 return nullptr; // BB must be dominated by the switch.
3729 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3730 if (!Branch || !Branch->isUnconditional())
3731 return nullptr; // Terminator must be unconditional branch.
3733 BasicBlock *Succ = Branch->getSuccessor(0);
3735 BasicBlock::iterator I = Succ->begin();
3736 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3737 int Idx = PHI->getBasicBlockIndex(BB);
3738 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3740 Value *InValue = PHI->getIncomingValue(Idx);
3741 if (InValue != CaseValue) continue;
3750 /// Try to forward the condition of a switch instruction to a phi node
3751 /// dominated by the switch, if that would mean that some of the destination
3752 /// blocks of the switch can be folded away.
3753 /// Returns true if a change is made.
3754 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3755 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3756 ForwardingNodesMap ForwardingNodes;
3758 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3759 ConstantInt *CaseValue = I.getCaseValue();
3760 BasicBlock *CaseDest = I.getCaseSuccessor();
3763 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3767 ForwardingNodes[PHI].push_back(PhiIndex);
3770 bool Changed = false;
3772 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3773 E = ForwardingNodes.end(); I != E; ++I) {
3774 PHINode *Phi = I->first;
3775 SmallVectorImpl<int> &Indexes = I->second;
3777 if (Indexes.size() < 2) continue;
3779 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3780 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3787 /// Return true if the backend will be able to handle
3788 /// initializing an array of constants like C.
3789 static bool ValidLookupTableConstant(Constant *C) {
3790 if (C->isThreadDependent())
3792 if (C->isDLLImportDependent())
3795 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3796 return CE->isGEPWithNoNotionalOverIndexing();
3798 return isa<ConstantFP>(C) ||
3799 isa<ConstantInt>(C) ||
3800 isa<ConstantPointerNull>(C) ||
3801 isa<GlobalValue>(C) ||
3805 /// If V is a Constant, return it. Otherwise, try to look up
3806 /// its constant value in ConstantPool, returning 0 if it's not there.
3807 static Constant *LookupConstant(Value *V,
3808 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3809 if (Constant *C = dyn_cast<Constant>(V))
3811 return ConstantPool.lookup(V);
3814 /// Try to fold instruction I into a constant. This works for
3815 /// simple instructions such as binary operations where both operands are
3816 /// constant or can be replaced by constants from the ConstantPool. Returns the
3817 /// resulting constant on success, 0 otherwise.
3819 ConstantFold(Instruction *I, const DataLayout &DL,
3820 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
3821 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3822 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3825 if (A->isAllOnesValue())
3826 return LookupConstant(Select->getTrueValue(), ConstantPool);
3827 if (A->isNullValue())
3828 return LookupConstant(Select->getFalseValue(), ConstantPool);
3832 SmallVector<Constant *, 4> COps;
3833 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
3834 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
3840 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
3841 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
3845 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), COps, DL);
3848 /// Try to determine the resulting constant values in phi nodes
3849 /// at the common destination basic block, *CommonDest, for one of the case
3850 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3851 /// case), of a switch instruction SI.
3853 GetCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest,
3854 BasicBlock **CommonDest,
3855 SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
3856 const DataLayout &DL) {
3857 // The block from which we enter the common destination.
3858 BasicBlock *Pred = SI->getParent();
3860 // If CaseDest is empty except for some side-effect free instructions through
3861 // which we can constant-propagate the CaseVal, continue to its successor.
3862 SmallDenseMap<Value*, Constant*> ConstantPool;
3863 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3864 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3866 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3867 // If the terminator is a simple branch, continue to the next block.
3868 if (T->getNumSuccessors() != 1)
3871 CaseDest = T->getSuccessor(0);
3872 } else if (isa<DbgInfoIntrinsic>(I)) {
3873 // Skip debug intrinsic.
3875 } else if (Constant *C = ConstantFold(&*I, DL, ConstantPool)) {
3876 // Instruction is side-effect free and constant.
3878 // If the instruction has uses outside this block or a phi node slot for
3879 // the block, it is not safe to bypass the instruction since it would then
3880 // no longer dominate all its uses.
3881 for (auto &Use : I->uses()) {
3882 User *User = Use.getUser();
3883 if (Instruction *I = dyn_cast<Instruction>(User))
3884 if (I->getParent() == CaseDest)
3886 if (PHINode *Phi = dyn_cast<PHINode>(User))
3887 if (Phi->getIncomingBlock(Use) == CaseDest)
3892 ConstantPool.insert(std::make_pair(&*I, C));
3898 // If we did not have a CommonDest before, use the current one.
3900 *CommonDest = CaseDest;
3901 // If the destination isn't the common one, abort.
3902 if (CaseDest != *CommonDest)
3905 // Get the values for this case from phi nodes in the destination block.
3906 BasicBlock::iterator I = (*CommonDest)->begin();
3907 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3908 int Idx = PHI->getBasicBlockIndex(Pred);
3912 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3917 // Be conservative about which kinds of constants we support.
3918 if (!ValidLookupTableConstant(ConstVal))
3921 Res.push_back(std::make_pair(PHI, ConstVal));
3924 return Res.size() > 0;
3927 // Helper function used to add CaseVal to the list of cases that generate
3929 static void MapCaseToResult(ConstantInt *CaseVal,
3930 SwitchCaseResultVectorTy &UniqueResults,
3932 for (auto &I : UniqueResults) {
3933 if (I.first == Result) {
3934 I.second.push_back(CaseVal);
3938 UniqueResults.push_back(std::make_pair(Result,
3939 SmallVector<ConstantInt*, 4>(1, CaseVal)));
3942 // Helper function that initializes a map containing
3943 // results for the PHI node of the common destination block for a switch
3944 // instruction. Returns false if multiple PHI nodes have been found or if
3945 // there is not a common destination block for the switch.
3946 static bool InitializeUniqueCases(SwitchInst *SI, PHINode *&PHI,
3947 BasicBlock *&CommonDest,
3948 SwitchCaseResultVectorTy &UniqueResults,
3949 Constant *&DefaultResult,
3950 const DataLayout &DL) {
3951 for (auto &I : SI->cases()) {
3952 ConstantInt *CaseVal = I.getCaseValue();
3954 // Resulting value at phi nodes for this case value.
3955 SwitchCaseResultsTy Results;
3956 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
3960 // Only one value per case is permitted
3961 if (Results.size() > 1)
3963 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
3965 // Check the PHI consistency.
3967 PHI = Results[0].first;
3968 else if (PHI != Results[0].first)
3971 // Find the default result value.
3972 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
3973 BasicBlock *DefaultDest = SI->getDefaultDest();
3974 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
3976 // If the default value is not found abort unless the default destination
3979 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
3980 if ((!DefaultResult &&
3981 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
3987 // Helper function that checks if it is possible to transform a switch with only
3988 // two cases (or two cases + default) that produces a result into a select.
3991 // case 10: %0 = icmp eq i32 %a, 10
3992 // return 10; %1 = select i1 %0, i32 10, i32 4
3993 // case 20: ----> %2 = icmp eq i32 %a, 20
3994 // return 2; %3 = select i1 %2, i32 2, i32 %1
3999 ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4000 Constant *DefaultResult, Value *Condition,
4001 IRBuilder<> &Builder) {
4002 assert(ResultVector.size() == 2 &&
4003 "We should have exactly two unique results at this point");
4004 // If we are selecting between only two cases transform into a simple
4005 // select or a two-way select if default is possible.
4006 if (ResultVector[0].second.size() == 1 &&
4007 ResultVector[1].second.size() == 1) {
4008 ConstantInt *const FirstCase = ResultVector[0].second[0];
4009 ConstantInt *const SecondCase = ResultVector[1].second[0];
4011 bool DefaultCanTrigger = DefaultResult;
4012 Value *SelectValue = ResultVector[1].first;
4013 if (DefaultCanTrigger) {
4014 Value *const ValueCompare =
4015 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4016 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4017 DefaultResult, "switch.select");
4019 Value *const ValueCompare =
4020 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4021 return Builder.CreateSelect(ValueCompare, ResultVector[0].first, SelectValue,
4028 // Helper function to cleanup a switch instruction that has been converted into
4029 // a select, fixing up PHI nodes and basic blocks.
4030 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
4032 IRBuilder<> &Builder) {
4033 BasicBlock *SelectBB = SI->getParent();
4034 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4035 PHI->removeIncomingValue(SelectBB);
4036 PHI->addIncoming(SelectValue, SelectBB);
4038 Builder.CreateBr(PHI->getParent());
4040 // Remove the switch.
4041 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4042 BasicBlock *Succ = SI->getSuccessor(i);
4044 if (Succ == PHI->getParent())
4046 Succ->removePredecessor(SelectBB);
4048 SI->eraseFromParent();
4051 /// If the switch is only used to initialize one or more
4052 /// phi nodes in a common successor block with only two different
4053 /// constant values, replace the switch with select.
4054 static bool SwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4055 AssumptionCache *AC, const DataLayout &DL) {
4056 Value *const Cond = SI->getCondition();
4057 PHINode *PHI = nullptr;
4058 BasicBlock *CommonDest = nullptr;
4059 Constant *DefaultResult;
4060 SwitchCaseResultVectorTy UniqueResults;
4061 // Collect all the cases that will deliver the same value from the switch.
4062 if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4065 // Selects choose between maximum two values.
4066 if (UniqueResults.size() != 2)
4068 assert(PHI != nullptr && "PHI for value select not found");
4070 Builder.SetInsertPoint(SI);
4071 Value *SelectValue = ConvertTwoCaseSwitch(
4073 DefaultResult, Cond, Builder);
4075 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4078 // The switch couldn't be converted into a select.
4083 /// This class represents a lookup table that can be used to replace a switch.
4084 class SwitchLookupTable {
4086 /// Create a lookup table to use as a switch replacement with the contents
4087 /// of Values, using DefaultValue to fill any holes in the table.
4089 Module &M, uint64_t TableSize, ConstantInt *Offset,
4090 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4091 Constant *DefaultValue, const DataLayout &DL);
4093 /// Build instructions with Builder to retrieve the value at
4094 /// the position given by Index in the lookup table.
4095 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4097 /// Return true if a table with TableSize elements of
4098 /// type ElementType would fit in a target-legal register.
4099 static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4103 // Depending on the contents of the table, it can be represented in
4106 // For tables where each element contains the same value, we just have to
4107 // store that single value and return it for each lookup.
4110 // For tables where there is a linear relationship between table index
4111 // and values. We calculate the result with a simple multiplication
4112 // and addition instead of a table lookup.
4115 // For small tables with integer elements, we can pack them into a bitmap
4116 // that fits into a target-legal register. Values are retrieved by
4117 // shift and mask operations.
4120 // The table is stored as an array of values. Values are retrieved by load
4121 // instructions from the table.
4125 // For SingleValueKind, this is the single value.
4126 Constant *SingleValue;
4128 // For BitMapKind, this is the bitmap.
4129 ConstantInt *BitMap;
4130 IntegerType *BitMapElementTy;
4132 // For LinearMapKind, these are the constants used to derive the value.
4133 ConstantInt *LinearOffset;
4134 ConstantInt *LinearMultiplier;
4136 // For ArrayKind, this is the array.
4137 GlobalVariable *Array;
4141 SwitchLookupTable::SwitchLookupTable(
4142 Module &M, uint64_t TableSize, ConstantInt *Offset,
4143 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4144 Constant *DefaultValue, const DataLayout &DL)
4145 : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
4146 LinearOffset(nullptr), LinearMultiplier(nullptr), Array(nullptr) {
4147 assert(Values.size() && "Can't build lookup table without values!");
4148 assert(TableSize >= Values.size() && "Can't fit values in table!");
4150 // If all values in the table are equal, this is that value.
4151 SingleValue = Values.begin()->second;
4153 Type *ValueType = Values.begin()->second->getType();
4155 // Build up the table contents.
4156 SmallVector<Constant*, 64> TableContents(TableSize);
4157 for (size_t I = 0, E = Values.size(); I != E; ++I) {
4158 ConstantInt *CaseVal = Values[I].first;
4159 Constant *CaseRes = Values[I].second;
4160 assert(CaseRes->getType() == ValueType);
4162 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
4164 TableContents[Idx] = CaseRes;
4166 if (CaseRes != SingleValue)
4167 SingleValue = nullptr;
4170 // Fill in any holes in the table with the default result.
4171 if (Values.size() < TableSize) {
4172 assert(DefaultValue &&
4173 "Need a default value to fill the lookup table holes.");
4174 assert(DefaultValue->getType() == ValueType);
4175 for (uint64_t I = 0; I < TableSize; ++I) {
4176 if (!TableContents[I])
4177 TableContents[I] = DefaultValue;
4180 if (DefaultValue != SingleValue)
4181 SingleValue = nullptr;
4184 // If each element in the table contains the same value, we only need to store
4185 // that single value.
4187 Kind = SingleValueKind;
4191 // Check if we can derive the value with a linear transformation from the
4193 if (isa<IntegerType>(ValueType)) {
4194 bool LinearMappingPossible = true;
4197 assert(TableSize >= 2 && "Should be a SingleValue table.");
4198 // Check if there is the same distance between two consecutive values.
4199 for (uint64_t I = 0; I < TableSize; ++I) {
4200 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4202 // This is an undef. We could deal with it, but undefs in lookup tables
4203 // are very seldom. It's probably not worth the additional complexity.
4204 LinearMappingPossible = false;
4207 APInt Val = ConstVal->getValue();
4209 APInt Dist = Val - PrevVal;
4212 } else if (Dist != DistToPrev) {
4213 LinearMappingPossible = false;
4219 if (LinearMappingPossible) {
4220 LinearOffset = cast<ConstantInt>(TableContents[0]);
4221 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
4222 Kind = LinearMapKind;
4228 // If the type is integer and the table fits in a register, build a bitmap.
4229 if (WouldFitInRegister(DL, TableSize, ValueType)) {
4230 IntegerType *IT = cast<IntegerType>(ValueType);
4231 APInt TableInt(TableSize * IT->getBitWidth(), 0);
4232 for (uint64_t I = TableSize; I > 0; --I) {
4233 TableInt <<= IT->getBitWidth();
4234 // Insert values into the bitmap. Undef values are set to zero.
4235 if (!isa<UndefValue>(TableContents[I - 1])) {
4236 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
4237 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
4240 BitMap = ConstantInt::get(M.getContext(), TableInt);
4241 BitMapElementTy = IT;
4247 // Store the table in an array.
4248 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
4249 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
4251 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
4252 GlobalVariable::PrivateLinkage,
4255 Array->setUnnamedAddr(true);
4259 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
4261 case SingleValueKind:
4263 case LinearMapKind: {
4264 // Derive the result value from the input value.
4265 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
4266 false, "switch.idx.cast");
4267 if (!LinearMultiplier->isOne())
4268 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
4269 if (!LinearOffset->isZero())
4270 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
4274 // Type of the bitmap (e.g. i59).
4275 IntegerType *MapTy = BitMap->getType();
4277 // Cast Index to the same type as the bitmap.
4278 // Note: The Index is <= the number of elements in the table, so
4279 // truncating it to the width of the bitmask is safe.
4280 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
4282 // Multiply the shift amount by the element width.
4283 ShiftAmt = Builder.CreateMul(ShiftAmt,
4284 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
4288 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
4289 "switch.downshift");
4291 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
4295 // Make sure the table index will not overflow when treated as signed.
4296 IntegerType *IT = cast<IntegerType>(Index->getType());
4297 uint64_t TableSize = Array->getInitializer()->getType()
4298 ->getArrayNumElements();
4299 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
4300 Index = Builder.CreateZExt(Index,
4301 IntegerType::get(IT->getContext(),
4302 IT->getBitWidth() + 1),
4303 "switch.tableidx.zext");
4305 Value *GEPIndices[] = { Builder.getInt32(0), Index };
4306 Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
4307 GEPIndices, "switch.gep");
4308 return Builder.CreateLoad(GEP, "switch.load");
4311 llvm_unreachable("Unknown lookup table kind!");
4314 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
4316 Type *ElementType) {
4317 auto *IT = dyn_cast<IntegerType>(ElementType);
4320 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
4321 // are <= 15, we could try to narrow the type.
4323 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
4324 if (TableSize >= UINT_MAX/IT->getBitWidth())
4326 return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
4329 /// Determine whether a lookup table should be built for this switch, based on
4330 /// the number of cases, size of the table, and the types of the results.
4332 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
4333 const TargetTransformInfo &TTI, const DataLayout &DL,
4334 const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
4335 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
4336 return false; // TableSize overflowed, or mul below might overflow.
4338 bool AllTablesFitInRegister = true;
4339 bool HasIllegalType = false;
4340 for (const auto &I : ResultTypes) {
4341 Type *Ty = I.second;
4343 // Saturate this flag to true.
4344 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
4346 // Saturate this flag to false.
4347 AllTablesFitInRegister = AllTablesFitInRegister &&
4348 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
4350 // If both flags saturate, we're done. NOTE: This *only* works with
4351 // saturating flags, and all flags have to saturate first due to the
4352 // non-deterministic behavior of iterating over a dense map.
4353 if (HasIllegalType && !AllTablesFitInRegister)
4357 // If each table would fit in a register, we should build it anyway.
4358 if (AllTablesFitInRegister)
4361 // Don't build a table that doesn't fit in-register if it has illegal types.
4365 // The table density should be at least 40%. This is the same criterion as for
4366 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
4367 // FIXME: Find the best cut-off.
4368 return SI->getNumCases() * 10 >= TableSize * 4;
4371 /// Try to reuse the switch table index compare. Following pattern:
4373 /// if (idx < tablesize)
4374 /// r = table[idx]; // table does not contain default_value
4376 /// r = default_value;
4377 /// if (r != default_value)
4380 /// Is optimized to:
4382 /// cond = idx < tablesize;
4386 /// r = default_value;
4390 /// Jump threading will then eliminate the second if(cond).
4391 static void reuseTableCompare(User *PhiUser, BasicBlock *PhiBlock,
4392 BranchInst *RangeCheckBranch, Constant *DefaultValue,
4393 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values) {
4395 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
4399 // We require that the compare is in the same block as the phi so that jump
4400 // threading can do its work afterwards.
4401 if (CmpInst->getParent() != PhiBlock)
4404 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
4408 Value *RangeCmp = RangeCheckBranch->getCondition();
4409 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
4410 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
4412 // Check if the compare with the default value is constant true or false.
4413 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
4414 DefaultValue, CmpOp1, true);
4415 if (DefaultConst != TrueConst && DefaultConst != FalseConst)
4418 // Check if the compare with the case values is distinct from the default
4420 for (auto ValuePair : Values) {
4421 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
4422 ValuePair.second, CmpOp1, true);
4423 if (!CaseConst || CaseConst == DefaultConst)
4425 assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
4426 "Expect true or false as compare result.");
4429 // Check if the branch instruction dominates the phi node. It's a simple
4430 // dominance check, but sufficient for our needs.
4431 // Although this check is invariant in the calling loops, it's better to do it
4432 // at this late stage. Practically we do it at most once for a switch.
4433 BasicBlock *BranchBlock = RangeCheckBranch->getParent();
4434 for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
4435 BasicBlock *Pred = *PI;
4436 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
4440 if (DefaultConst == FalseConst) {
4441 // The compare yields the same result. We can replace it.
4442 CmpInst->replaceAllUsesWith(RangeCmp);
4443 ++NumTableCmpReuses;
4445 // The compare yields the same result, just inverted. We can replace it.
4446 Value *InvertedTableCmp = BinaryOperator::CreateXor(RangeCmp,
4447 ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
4449 CmpInst->replaceAllUsesWith(InvertedTableCmp);
4450 ++NumTableCmpReuses;
4454 /// If the switch is only used to initialize one or more phi nodes in a common
4455 /// successor block with different constant values, replace the switch with
4457 static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
4458 const DataLayout &DL,
4459 const TargetTransformInfo &TTI) {
4460 assert(SI->getNumCases() > 1 && "Degenerate switch?");
4462 // Only build lookup table when we have a target that supports it.
4463 if (!TTI.shouldBuildLookupTables())
4466 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
4467 // split off a dense part and build a lookup table for that.
4469 // FIXME: This creates arrays of GEPs to constant strings, which means each
4470 // GEP needs a runtime relocation in PIC code. We should just build one big
4471 // string and lookup indices into that.
4473 // Ignore switches with less than three cases. Lookup tables will not make them
4474 // faster, so we don't analyze them.
4475 if (SI->getNumCases() < 3)
4478 // Figure out the corresponding result for each case value and phi node in the
4479 // common destination, as well as the min and max case values.
4480 assert(SI->case_begin() != SI->case_end());
4481 SwitchInst::CaseIt CI = SI->case_begin();
4482 ConstantInt *MinCaseVal = CI.getCaseValue();
4483 ConstantInt *MaxCaseVal = CI.getCaseValue();
4485 BasicBlock *CommonDest = nullptr;
4486 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
4487 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
4488 SmallDenseMap<PHINode*, Constant*> DefaultResults;
4489 SmallDenseMap<PHINode*, Type*> ResultTypes;
4490 SmallVector<PHINode*, 4> PHIs;
4492 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
4493 ConstantInt *CaseVal = CI.getCaseValue();
4494 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
4495 MinCaseVal = CaseVal;
4496 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
4497 MaxCaseVal = CaseVal;
4499 // Resulting value at phi nodes for this case value.
4500 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
4502 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
4506 // Append the result from this case to the list for each phi.
4507 for (const auto &I : Results) {
4508 PHINode *PHI = I.first;
4509 Constant *Value = I.second;
4510 if (!ResultLists.count(PHI))
4511 PHIs.push_back(PHI);
4512 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
4516 // Keep track of the result types.
4517 for (PHINode *PHI : PHIs) {
4518 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
4521 uint64_t NumResults = ResultLists[PHIs[0]].size();
4522 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
4523 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
4524 bool TableHasHoles = (NumResults < TableSize);
4526 // If the table has holes, we need a constant result for the default case
4527 // or a bitmask that fits in a register.
4528 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
4529 bool HasDefaultResults = GetCaseResults(SI, nullptr, SI->getDefaultDest(),
4530 &CommonDest, DefaultResultsList, DL);
4532 bool NeedMask = (TableHasHoles && !HasDefaultResults);
4534 // As an extra penalty for the validity test we require more cases.
4535 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
4537 if (!DL.fitsInLegalInteger(TableSize))
4541 for (const auto &I : DefaultResultsList) {
4542 PHINode *PHI = I.first;
4543 Constant *Result = I.second;
4544 DefaultResults[PHI] = Result;
4547 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
4550 // Create the BB that does the lookups.
4551 Module &Mod = *CommonDest->getParent()->getParent();
4552 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
4554 CommonDest->getParent(),
4557 // Compute the table index value.
4558 Builder.SetInsertPoint(SI);
4559 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
4562 // Compute the maximum table size representable by the integer type we are
4564 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
4565 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
4566 assert(MaxTableSize >= TableSize &&
4567 "It is impossible for a switch to have more entries than the max "
4568 "representable value of its input integer type's size.");
4570 // If the default destination is unreachable, or if the lookup table covers
4571 // all values of the conditional variable, branch directly to the lookup table
4572 // BB. Otherwise, check that the condition is within the case range.
4573 const bool DefaultIsReachable =
4574 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4575 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
4576 BranchInst *RangeCheckBranch = nullptr;
4578 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
4579 Builder.CreateBr(LookupBB);
4580 // Note: We call removeProdecessor later since we need to be able to get the
4581 // PHI value for the default case in case we're using a bit mask.
4583 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
4584 MinCaseVal->getType(), TableSize));
4585 RangeCheckBranch = Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
4588 // Populate the BB that does the lookups.
4589 Builder.SetInsertPoint(LookupBB);
4592 // Before doing the lookup we do the hole check.
4593 // The LookupBB is therefore re-purposed to do the hole check
4594 // and we create a new LookupBB.
4595 BasicBlock *MaskBB = LookupBB;
4596 MaskBB->setName("switch.hole_check");
4597 LookupBB = BasicBlock::Create(Mod.getContext(),
4599 CommonDest->getParent(),
4602 // Make the mask's bitwidth at least 8bit and a power-of-2 to avoid
4603 // unnecessary illegal types.
4604 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
4605 APInt MaskInt(TableSizePowOf2, 0);
4606 APInt One(TableSizePowOf2, 1);
4607 // Build bitmask; fill in a 1 bit for every case.
4608 const ResultListTy &ResultList = ResultLists[PHIs[0]];
4609 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
4610 uint64_t Idx = (ResultList[I].first->getValue() -
4611 MinCaseVal->getValue()).getLimitedValue();
4612 MaskInt |= One << Idx;
4614 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
4616 // Get the TableIndex'th bit of the bitmask.
4617 // If this bit is 0 (meaning hole) jump to the default destination,
4618 // else continue with table lookup.
4619 IntegerType *MapTy = TableMask->getType();
4620 Value *MaskIndex = Builder.CreateZExtOrTrunc(TableIndex, MapTy,
4621 "switch.maskindex");
4622 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex,
4624 Value *LoBit = Builder.CreateTrunc(Shifted,
4625 Type::getInt1Ty(Mod.getContext()),
4627 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
4629 Builder.SetInsertPoint(LookupBB);
4630 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
4633 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
4634 // We cached PHINodes in PHIs, to avoid accessing deleted PHINodes later,
4635 // do not delete PHINodes here.
4636 SI->getDefaultDest()->removePredecessor(SI->getParent(),
4637 /*DontDeleteUselessPHIs=*/true);
4640 bool ReturnedEarly = false;
4641 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
4642 PHINode *PHI = PHIs[I];
4643 const ResultListTy &ResultList = ResultLists[PHI];
4645 // If using a bitmask, use any value to fill the lookup table holes.
4646 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
4647 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL);
4649 Value *Result = Table.BuildLookup(TableIndex, Builder);
4651 // If the result is used to return immediately from the function, we want to
4652 // do that right here.
4653 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
4654 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
4655 Builder.CreateRet(Result);
4656 ReturnedEarly = true;
4660 // Do a small peephole optimization: re-use the switch table compare if
4662 if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
4663 BasicBlock *PhiBlock = PHI->getParent();
4664 // Search for compare instructions which use the phi.
4665 for (auto *User : PHI->users()) {
4666 reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
4670 PHI->addIncoming(Result, LookupBB);
4674 Builder.CreateBr(CommonDest);
4676 // Remove the switch.
4677 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4678 BasicBlock *Succ = SI->getSuccessor(i);
4680 if (Succ == SI->getDefaultDest())
4682 Succ->removePredecessor(SI->getParent());
4684 SI->eraseFromParent();
4688 ++NumLookupTablesHoles;
4692 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
4693 BasicBlock *BB = SI->getParent();
4695 if (isValueEqualityComparison(SI)) {
4696 // If we only have one predecessor, and if it is a branch on this value,
4697 // see if that predecessor totally determines the outcome of this switch.
4698 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4699 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
4700 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4702 Value *Cond = SI->getCondition();
4703 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
4704 if (SimplifySwitchOnSelect(SI, Select))
4705 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4707 // If the block only contains the switch, see if we can fold the block
4708 // away into any preds.
4709 BasicBlock::iterator BBI = BB->begin();
4710 // Ignore dbg intrinsics.
4711 while (isa<DbgInfoIntrinsic>(BBI))
4714 if (FoldValueComparisonIntoPredecessors(SI, Builder))
4715 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4718 // Try to transform the switch into an icmp and a branch.
4719 if (TurnSwitchRangeIntoICmp(SI, Builder))
4720 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4722 // Remove unreachable cases.
4723 if (EliminateDeadSwitchCases(SI, AC, DL))
4724 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4726 if (SwitchToSelect(SI, Builder, AC, DL))
4727 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4729 if (ForwardSwitchConditionToPHI(SI))
4730 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4732 if (SwitchToLookupTable(SI, Builder, DL, TTI))
4733 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4738 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
4739 BasicBlock *BB = IBI->getParent();
4740 bool Changed = false;
4742 // Eliminate redundant destinations.
4743 SmallPtrSet<Value *, 8> Succs;
4744 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
4745 BasicBlock *Dest = IBI->getDestination(i);
4746 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
4747 Dest->removePredecessor(BB);
4748 IBI->removeDestination(i);
4754 if (IBI->getNumDestinations() == 0) {
4755 // If the indirectbr has no successors, change it to unreachable.
4756 new UnreachableInst(IBI->getContext(), IBI);
4757 EraseTerminatorInstAndDCECond(IBI);
4761 if (IBI->getNumDestinations() == 1) {
4762 // If the indirectbr has one successor, change it to a direct branch.
4763 BranchInst::Create(IBI->getDestination(0), IBI);
4764 EraseTerminatorInstAndDCECond(IBI);
4768 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
4769 if (SimplifyIndirectBrOnSelect(IBI, SI))
4770 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4775 /// Given an block with only a single landing pad and a unconditional branch
4776 /// try to find another basic block which this one can be merged with. This
4777 /// handles cases where we have multiple invokes with unique landing pads, but
4778 /// a shared handler.
4780 /// We specifically choose to not worry about merging non-empty blocks
4781 /// here. That is a PRE/scheduling problem and is best solved elsewhere. In
4782 /// practice, the optimizer produces empty landing pad blocks quite frequently
4783 /// when dealing with exception dense code. (see: instcombine, gvn, if-else
4784 /// sinking in this file)
4786 /// This is primarily a code size optimization. We need to avoid performing
4787 /// any transform which might inhibit optimization (such as our ability to
4788 /// specialize a particular handler via tail commoning). We do this by not
4789 /// merging any blocks which require us to introduce a phi. Since the same
4790 /// values are flowing through both blocks, we don't loose any ability to
4791 /// specialize. If anything, we make such specialization more likely.
4793 /// TODO - This transformation could remove entries from a phi in the target
4794 /// block when the inputs in the phi are the same for the two blocks being
4795 /// merged. In some cases, this could result in removal of the PHI entirely.
4796 static bool TryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI,
4798 auto Succ = BB->getUniqueSuccessor();
4800 // If there's a phi in the successor block, we'd likely have to introduce
4801 // a phi into the merged landing pad block.
4802 if (isa<PHINode>(*Succ->begin()))
4805 for (BasicBlock *OtherPred : predecessors(Succ)) {
4806 if (BB == OtherPred)
4808 BasicBlock::iterator I = OtherPred->begin();
4809 LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(I);
4810 if (!LPad2 || !LPad2->isIdenticalTo(LPad))
4812 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {}
4813 BranchInst *BI2 = dyn_cast<BranchInst>(I);
4814 if (!BI2 || !BI2->isIdenticalTo(BI))
4817 // We've found an identical block. Update our predeccessors to take that
4818 // path instead and make ourselves dead.
4819 SmallSet<BasicBlock *, 16> Preds;
4820 Preds.insert(pred_begin(BB), pred_end(BB));
4821 for (BasicBlock *Pred : Preds) {
4822 InvokeInst *II = cast<InvokeInst>(Pred->getTerminator());
4823 assert(II->getNormalDest() != BB &&
4824 II->getUnwindDest() == BB && "unexpected successor");
4825 II->setUnwindDest(OtherPred);
4828 // The debug info in OtherPred doesn't cover the merged control flow that
4829 // used to go through BB. We need to delete it or update it.
4830 for (auto I = OtherPred->begin(), E = OtherPred->end();
4832 Instruction &Inst = *I; I++;
4833 if (isa<DbgInfoIntrinsic>(Inst))
4834 Inst.eraseFromParent();
4837 SmallSet<BasicBlock *, 16> Succs;
4838 Succs.insert(succ_begin(BB), succ_end(BB));
4839 for (BasicBlock *Succ : Succs) {
4840 Succ->removePredecessor(BB);
4843 IRBuilder<> Builder(BI);
4844 Builder.CreateUnreachable();
4845 BI->eraseFromParent();
4851 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
4852 BasicBlock *BB = BI->getParent();
4854 if (SinkCommon && SinkThenElseCodeToEnd(BI))
4857 // If the Terminator is the only non-phi instruction, simplify the block.
4858 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg()->getIterator();
4859 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
4860 TryToSimplifyUncondBranchFromEmptyBlock(BB))
4863 // If the only instruction in the block is a seteq/setne comparison
4864 // against a constant, try to simplify the block.
4865 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
4866 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
4867 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
4869 if (I->isTerminator() &&
4870 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, DL, TTI,
4871 BonusInstThreshold, AC))
4875 // See if we can merge an empty landing pad block with another which is
4877 if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(I)) {
4878 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {}
4879 if (I->isTerminator() &&
4880 TryToMergeLandingPad(LPad, BI, BB))
4884 // If this basic block is ONLY a compare and a branch, and if a predecessor
4885 // branches to us and our successor, fold the comparison into the
4886 // predecessor and use logical operations to update the incoming value
4887 // for PHI nodes in common successor.
4888 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
4889 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4893 static BasicBlock *allPredecessorsComeFromSameSource(BasicBlock *BB) {
4894 BasicBlock *PredPred = nullptr;
4895 for (auto *P : predecessors(BB)) {
4896 BasicBlock *PPred = P->getSinglePredecessor();
4897 if (!PPred || (PredPred && PredPred != PPred))
4904 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
4905 BasicBlock *BB = BI->getParent();
4907 // Conditional branch
4908 if (isValueEqualityComparison(BI)) {
4909 // If we only have one predecessor, and if it is a branch on this value,
4910 // see if that predecessor totally determines the outcome of this
4912 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4913 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
4914 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4916 // This block must be empty, except for the setcond inst, if it exists.
4917 // Ignore dbg intrinsics.
4918 BasicBlock::iterator I = BB->begin();
4919 // Ignore dbg intrinsics.
4920 while (isa<DbgInfoIntrinsic>(I))
4923 if (FoldValueComparisonIntoPredecessors(BI, Builder))
4924 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4925 } else if (&*I == cast<Instruction>(BI->getCondition())){
4927 // Ignore dbg intrinsics.
4928 while (isa<DbgInfoIntrinsic>(I))
4930 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
4931 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4935 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
4936 if (SimplifyBranchOnICmpChain(BI, Builder, DL))
4939 // If this basic block is ONLY a compare and a branch, and if a predecessor
4940 // branches to us and one of our successors, fold the comparison into the
4941 // predecessor and use logical operations to pick the right destination.
4942 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
4943 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4945 // We have a conditional branch to two blocks that are only reachable
4946 // from BI. We know that the condbr dominates the two blocks, so see if
4947 // there is any identical code in the "then" and "else" blocks. If so, we
4948 // can hoist it up to the branching block.
4949 if (BI->getSuccessor(0)->getSinglePredecessor()) {
4950 if (BI->getSuccessor(1)->getSinglePredecessor()) {
4951 if (HoistThenElseCodeToIf(BI, TTI))
4952 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4954 // If Successor #1 has multiple preds, we may be able to conditionally
4955 // execute Successor #0 if it branches to Successor #1.
4956 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
4957 if (Succ0TI->getNumSuccessors() == 1 &&
4958 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
4959 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), TTI))
4960 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4962 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
4963 // If Successor #0 has multiple preds, we may be able to conditionally
4964 // execute Successor #1 if it branches to Successor #0.
4965 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
4966 if (Succ1TI->getNumSuccessors() == 1 &&
4967 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
4968 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), TTI))
4969 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4972 // If this is a branch on a phi node in the current block, thread control
4973 // through this block if any PHI node entries are constants.
4974 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
4975 if (PN->getParent() == BI->getParent())
4976 if (FoldCondBranchOnPHI(BI, DL))
4977 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4979 // Scan predecessor blocks for conditional branches.
4980 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
4981 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
4982 if (PBI != BI && PBI->isConditional())
4983 if (SimplifyCondBranchToCondBranch(PBI, BI, DL))
4984 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4986 // Look for diamond patterns.
4987 if (MergeCondStores)
4988 if (BasicBlock *PrevBB = allPredecessorsComeFromSameSource(BB))
4989 if (BranchInst *PBI = dyn_cast<BranchInst>(PrevBB->getTerminator()))
4990 if (PBI != BI && PBI->isConditional())
4991 if (mergeConditionalStores(PBI, BI))
4992 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4997 /// Check if passing a value to an instruction will cause undefined behavior.
4998 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
4999 Constant *C = dyn_cast<Constant>(V);
5006 if (C->isNullValue()) {
5007 // Only look at the first use, avoid hurting compile time with long uselists
5008 User *Use = *I->user_begin();
5010 // Now make sure that there are no instructions in between that can alter
5011 // control flow (eg. calls)
5012 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
5013 if (i == I->getParent()->end() || i->mayHaveSideEffects())
5016 // Look through GEPs. A load from a GEP derived from NULL is still undefined
5017 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
5018 if (GEP->getPointerOperand() == I)
5019 return passingValueIsAlwaysUndefined(V, GEP);
5021 // Look through bitcasts.
5022 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
5023 return passingValueIsAlwaysUndefined(V, BC);
5025 // Load from null is undefined.
5026 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
5027 if (!LI->isVolatile())
5028 return LI->getPointerAddressSpace() == 0;
5030 // Store to null is undefined.
5031 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
5032 if (!SI->isVolatile())
5033 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
5038 /// If BB has an incoming value that will always trigger undefined behavior
5039 /// (eg. null pointer dereference), remove the branch leading here.
5040 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
5041 for (BasicBlock::iterator i = BB->begin();
5042 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
5043 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
5044 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
5045 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
5046 IRBuilder<> Builder(T);
5047 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
5048 BB->removePredecessor(PHI->getIncomingBlock(i));
5049 // Turn uncoditional branches into unreachables and remove the dead
5050 // destination from conditional branches.
5051 if (BI->isUnconditional())
5052 Builder.CreateUnreachable();
5054 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
5055 BI->getSuccessor(0));
5056 BI->eraseFromParent();
5059 // TODO: SwitchInst.
5065 bool SimplifyCFGOpt::run(BasicBlock *BB) {
5066 bool Changed = false;
5068 assert(BB && BB->getParent() && "Block not embedded in function!");
5069 assert(BB->getTerminator() && "Degenerate basic block encountered!");
5071 // Remove basic blocks that have no predecessors (except the entry block)...
5072 // or that just have themself as a predecessor. These are unreachable.
5073 if ((pred_empty(BB) &&
5074 BB != &BB->getParent()->getEntryBlock()) ||
5075 BB->getSinglePredecessor() == BB) {
5076 DEBUG(dbgs() << "Removing BB: \n" << *BB);
5077 DeleteDeadBlock(BB);
5081 // Check to see if we can constant propagate this terminator instruction
5083 Changed |= ConstantFoldTerminator(BB, true);
5085 // Check for and eliminate duplicate PHI nodes in this block.
5086 Changed |= EliminateDuplicatePHINodes(BB);
5088 // Check for and remove branches that will always cause undefined behavior.
5089 Changed |= removeUndefIntroducingPredecessor(BB);
5091 // Merge basic blocks into their predecessor if there is only one distinct
5092 // pred, and if there is only one distinct successor of the predecessor, and
5093 // if there are no PHI nodes.
5095 if (MergeBlockIntoPredecessor(BB))
5098 IRBuilder<> Builder(BB);
5100 // If there is a trivial two-entry PHI node in this basic block, and we can
5101 // eliminate it, do so now.
5102 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
5103 if (PN->getNumIncomingValues() == 2)
5104 Changed |= FoldTwoEntryPHINode(PN, TTI, DL);
5106 Builder.SetInsertPoint(BB->getTerminator());
5107 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
5108 if (BI->isUnconditional()) {
5109 if (SimplifyUncondBranch(BI, Builder)) return true;
5111 if (SimplifyCondBranch(BI, Builder)) return true;
5113 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
5114 if (SimplifyReturn(RI, Builder)) return true;
5115 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
5116 if (SimplifyResume(RI, Builder)) return true;
5117 } else if (CleanupReturnInst *RI =
5118 dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
5119 if (SimplifyCleanupReturn(RI)) return true;
5120 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
5121 if (SimplifySwitch(SI, Builder)) return true;
5122 } else if (UnreachableInst *UI =
5123 dyn_cast<UnreachableInst>(BB->getTerminator())) {
5124 if (SimplifyUnreachable(UI)) return true;
5125 } else if (IndirectBrInst *IBI =
5126 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
5127 if (SimplifyIndirectBr(IBI)) return true;
5133 /// This function is used to do simplification of a CFG.
5134 /// For example, it adjusts branches to branches to eliminate the extra hop,
5135 /// eliminates unreachable basic blocks, and does other "peephole" optimization
5136 /// of the CFG. It returns true if a modification was made.
5138 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
5139 unsigned BonusInstThreshold, AssumptionCache *AC) {
5140 return SimplifyCFGOpt(TTI, BB->getModule()->getDataLayout(),
5141 BonusInstThreshold, AC).run(BB);