1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Peephole optimize the CFG.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Transforms/Utils/Local.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/ADT/SetVector.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/ConstantFolding.h"
22 #include "llvm/Analysis/InstructionSimplify.h"
23 #include "llvm/Analysis/TargetTransformInfo.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/IR/CFG.h"
26 #include "llvm/IR/ConstantRange.h"
27 #include "llvm/IR/Constants.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/DerivedTypes.h"
30 #include "llvm/IR/GlobalVariable.h"
31 #include "llvm/IR/IRBuilder.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/LLVMContext.h"
35 #include "llvm/IR/MDBuilder.h"
36 #include "llvm/IR/Metadata.h"
37 #include "llvm/IR/Module.h"
38 #include "llvm/IR/NoFolder.h"
39 #include "llvm/IR/Operator.h"
40 #include "llvm/IR/PatternMatch.h"
41 #include "llvm/IR/Type.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/Debug.h"
44 #include "llvm/Support/raw_ostream.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Transforms/Utils/ValueMapper.h"
52 using namespace PatternMatch;
54 #define DEBUG_TYPE "simplifycfg"
56 // Chosen as 2 so as to be cheap, but still to have enough power to fold
57 // a select, so the "clamp" idiom (of a min followed by a max) will be caught.
58 // To catch this, we need to fold a compare and a select, hence '2' being the
59 // minimum reasonable default.
60 static cl::opt<unsigned>
61 PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(2),
62 cl::desc("Control the amount of phi node folding to perform (default = 2)"));
65 DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false),
66 cl::desc("Duplicate return instructions into unconditional branches"));
69 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
70 cl::desc("Sink common instructions down to the end block"));
72 static cl::opt<bool> HoistCondStores(
73 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
74 cl::desc("Hoist conditional stores if an unconditional store precedes"));
76 static cl::opt<unsigned> SpeculativeFlattenBias(
77 "speculative-flatten-bias", cl::Hidden, cl::init(100),
78 cl::desc("Control how biased a branch needs to be to be considered rarely"
79 " failing for speculative flattening (default = 100)"));
81 static cl::opt<unsigned> SpeculativeFlattenThreshold(
82 "speculative-flatten-threshold", cl::Hidden, cl::init(10),
83 cl::desc("Control how much speculation happens due to speculative"
84 " flattening (default = 10)"));
87 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
88 STATISTIC(NumLinearMaps, "Number of switch instructions turned into linear mapping");
89 STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables");
90 STATISTIC(NumLookupTablesHoles, "Number of switch instructions turned into lookup tables (holes checked)");
91 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
92 STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block");
93 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
96 // The first field contains the value that the switch produces when a certain
97 // case group is selected, and the second field is a vector containing the
98 // cases composing the case group.
99 typedef SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>
100 SwitchCaseResultVectorTy;
101 // The first field contains the phi node that generates a result of the switch
102 // and the second field contains the value generated for a certain case in the
103 // switch for that PHI.
104 typedef SmallVector<std::pair<PHINode *, Constant *>, 4> SwitchCaseResultsTy;
106 /// ValueEqualityComparisonCase - Represents a case of a switch.
107 struct ValueEqualityComparisonCase {
111 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
112 : Value(Value), Dest(Dest) {}
114 bool operator<(ValueEqualityComparisonCase RHS) const {
115 // Comparing pointers is ok as we only rely on the order for uniquing.
116 return Value < RHS.Value;
119 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
122 class SimplifyCFGOpt {
123 const TargetTransformInfo &TTI;
124 const DataLayout &DL;
125 unsigned BonusInstThreshold;
127 Value *isValueEqualityComparison(TerminatorInst *TI);
128 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
129 std::vector<ValueEqualityComparisonCase> &Cases);
130 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
132 IRBuilder<> &Builder);
133 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
134 IRBuilder<> &Builder);
136 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
137 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
138 bool SimplifyCleanupReturn(CleanupReturnInst *RI);
139 bool SimplifyUnreachable(UnreachableInst *UI);
140 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
141 bool SimplifyIndirectBr(IndirectBrInst *IBI);
142 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder);
143 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder);
146 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
147 unsigned BonusInstThreshold, AssumptionCache *AC)
148 : TTI(TTI), DL(DL), BonusInstThreshold(BonusInstThreshold), AC(AC) {}
149 bool run(BasicBlock *BB);
153 /// Return true if it is safe to merge these two
154 /// terminator instructions together.
155 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
156 if (SI1 == SI2) return false; // Can't merge with self!
158 // It is not safe to merge these two switch instructions if they have a common
159 // successor, and if that successor has a PHI node, and if *that* PHI node has
160 // conflicting incoming values from the two switch blocks.
161 BasicBlock *SI1BB = SI1->getParent();
162 BasicBlock *SI2BB = SI2->getParent();
163 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
165 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
166 if (SI1Succs.count(*I))
167 for (BasicBlock::iterator BBI = (*I)->begin();
168 isa<PHINode>(BBI); ++BBI) {
169 PHINode *PN = cast<PHINode>(BBI);
170 if (PN->getIncomingValueForBlock(SI1BB) !=
171 PN->getIncomingValueForBlock(SI2BB))
178 /// Return true if it is safe and profitable to merge these two terminator
179 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
180 /// store all PHI nodes in common successors.
181 static bool isProfitableToFoldUnconditional(BranchInst *SI1,
184 SmallVectorImpl<PHINode*> &PhiNodes) {
185 if (SI1 == SI2) return false; // Can't merge with self!
186 assert(SI1->isUnconditional() && SI2->isConditional());
188 // We fold the unconditional branch if we can easily update all PHI nodes in
189 // common successors:
190 // 1> We have a constant incoming value for the conditional branch;
191 // 2> We have "Cond" as the incoming value for the unconditional branch;
192 // 3> SI2->getCondition() and Cond have same operands.
193 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
194 if (!Ci2) return false;
195 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
196 Cond->getOperand(1) == Ci2->getOperand(1)) &&
197 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
198 Cond->getOperand(1) == Ci2->getOperand(0)))
201 BasicBlock *SI1BB = SI1->getParent();
202 BasicBlock *SI2BB = SI2->getParent();
203 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
204 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
205 if (SI1Succs.count(*I))
206 for (BasicBlock::iterator BBI = (*I)->begin();
207 isa<PHINode>(BBI); ++BBI) {
208 PHINode *PN = cast<PHINode>(BBI);
209 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
210 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
212 PhiNodes.push_back(PN);
217 /// Update PHI nodes in Succ to indicate that there will now be entries in it
218 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
219 /// will be the same as those coming in from ExistPred, an existing predecessor
221 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
222 BasicBlock *ExistPred) {
223 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
226 for (BasicBlock::iterator I = Succ->begin();
227 (PN = dyn_cast<PHINode>(I)); ++I)
228 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
231 /// Compute an abstract "cost" of speculating the given instruction,
232 /// which is assumed to be safe to speculate. TCC_Free means cheap,
233 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
235 static unsigned ComputeSpeculationCost(const User *I,
236 const TargetTransformInfo &TTI) {
237 assert(isSafeToSpeculativelyExecute(I) &&
238 "Instruction is not safe to speculatively execute!");
239 return TTI.getUserCost(I);
242 /// If we have a merge point of an "if condition" as accepted above,
243 /// return true if the specified value dominates the block. We
244 /// don't handle the true generality of domination here, just a special case
245 /// which works well enough for us.
247 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
248 /// see if V (which must be an instruction) and its recursive operands
249 /// that do not dominate BB have a combined cost lower than CostRemaining and
250 /// are non-trapping. If both are true, the instruction is inserted into the
251 /// set and true is returned.
253 /// The cost for most non-trapping instructions is defined as 1 except for
254 /// Select whose cost is 2.
256 /// After this function returns, CostRemaining is decreased by the cost of
257 /// V plus its non-dominating operands. If that cost is greater than
258 /// CostRemaining, false is returned and CostRemaining is undefined.
259 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
260 SmallPtrSetImpl<Instruction*> *AggressiveInsts,
261 unsigned &CostRemaining,
262 const TargetTransformInfo &TTI) {
263 Instruction *I = dyn_cast<Instruction>(V);
265 // Non-instructions all dominate instructions, but not all constantexprs
266 // can be executed unconditionally.
267 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
272 BasicBlock *PBB = I->getParent();
274 // We don't want to allow weird loops that might have the "if condition" in
275 // the bottom of this block.
276 if (PBB == BB) return false;
278 // If this instruction is defined in a block that contains an unconditional
279 // branch to BB, then it must be in the 'conditional' part of the "if
280 // statement". If not, it definitely dominates the region.
281 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
282 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
285 // If we aren't allowing aggressive promotion anymore, then don't consider
286 // instructions in the 'if region'.
287 if (!AggressiveInsts) return false;
289 // If we have seen this instruction before, don't count it again.
290 if (AggressiveInsts->count(I)) return true;
292 // Okay, it looks like the instruction IS in the "condition". Check to
293 // see if it's a cheap instruction to unconditionally compute, and if it
294 // only uses stuff defined outside of the condition. If so, hoist it out.
295 if (!isSafeToSpeculativelyExecute(I))
298 unsigned Cost = ComputeSpeculationCost(I, TTI);
300 if (Cost > CostRemaining)
303 CostRemaining -= Cost;
305 // Okay, we can only really hoist these out if their operands do
306 // not take us over the cost threshold.
307 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
308 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, TTI))
310 // Okay, it's safe to do this! Remember this instruction.
311 AggressiveInsts->insert(I);
315 /// Extract ConstantInt from value, looking through IntToPtr
316 /// and PointerNullValue. Return NULL if value is not a constant int.
317 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
318 // Normal constant int.
319 ConstantInt *CI = dyn_cast<ConstantInt>(V);
320 if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
323 // This is some kind of pointer constant. Turn it into a pointer-sized
324 // ConstantInt if possible.
325 IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
327 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
328 if (isa<ConstantPointerNull>(V))
329 return ConstantInt::get(PtrTy, 0);
331 // IntToPtr const int.
332 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
333 if (CE->getOpcode() == Instruction::IntToPtr)
334 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
335 // The constant is very likely to have the right type already.
336 if (CI->getType() == PtrTy)
339 return cast<ConstantInt>
340 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
347 /// Given a chain of or (||) or and (&&) comparison of a value against a
348 /// constant, this will try to recover the information required for a switch
350 /// It will depth-first traverse the chain of comparison, seeking for patterns
351 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
352 /// representing the different cases for the switch.
353 /// Note that if the chain is composed of '||' it will build the set of elements
354 /// that matches the comparisons (i.e. any of this value validate the chain)
355 /// while for a chain of '&&' it will build the set elements that make the test
357 struct ConstantComparesGatherer {
358 const DataLayout &DL;
359 Value *CompValue; /// Value found for the switch comparison
360 Value *Extra; /// Extra clause to be checked before the switch
361 SmallVector<ConstantInt *, 8> Vals; /// Set of integers to match in switch
362 unsigned UsedICmps; /// Number of comparisons matched in the and/or chain
364 /// Construct and compute the result for the comparison instruction Cond
365 ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL)
366 : DL(DL), CompValue(nullptr), Extra(nullptr), UsedICmps(0) {
371 ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
372 ConstantComparesGatherer &
373 operator=(const ConstantComparesGatherer &) = delete;
377 /// Try to set the current value used for the comparison, it succeeds only if
378 /// it wasn't set before or if the new value is the same as the old one
379 bool setValueOnce(Value *NewVal) {
380 if(CompValue && CompValue != NewVal) return false;
382 return (CompValue != nullptr);
385 /// Try to match Instruction "I" as a comparison against a constant and
386 /// populates the array Vals with the set of values that match (or do not
387 /// match depending on isEQ).
388 /// Return false on failure. On success, the Value the comparison matched
389 /// against is placed in CompValue.
390 /// If CompValue is already set, the function is expected to fail if a match
391 /// is found but the value compared to is different.
392 bool matchInstruction(Instruction *I, bool isEQ) {
393 // If this is an icmp against a constant, handle this as one of the cases.
396 if (!((ICI = dyn_cast<ICmpInst>(I)) &&
397 (C = GetConstantInt(I->getOperand(1), DL)))) {
404 // Pattern match a special case
405 // (x & ~2^x) == y --> x == y || x == y|2^x
406 // This undoes a transformation done by instcombine to fuse 2 compares.
407 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
408 if (match(ICI->getOperand(0),
409 m_And(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
410 APInt Not = ~RHSC->getValue();
411 if (Not.isPowerOf2()) {
412 // If we already have a value for the switch, it has to match!
413 if(!setValueOnce(RHSVal))
417 Vals.push_back(ConstantInt::get(C->getContext(),
418 C->getValue() | Not));
424 // If we already have a value for the switch, it has to match!
425 if(!setValueOnce(ICI->getOperand(0)))
430 return ICI->getOperand(0);
433 // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
434 ConstantRange Span = ConstantRange::makeAllowedICmpRegion(
435 ICI->getPredicate(), C->getValue());
437 // Shift the range if the compare is fed by an add. This is the range
438 // compare idiom as emitted by instcombine.
439 Value *CandidateVal = I->getOperand(0);
440 if(match(I->getOperand(0), m_Add(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
441 Span = Span.subtract(RHSC->getValue());
442 CandidateVal = RHSVal;
445 // If this is an and/!= check, then we are looking to build the set of
446 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
449 Span = Span.inverse();
451 // If there are a ton of values, we don't want to make a ginormous switch.
452 if (Span.getSetSize().ugt(8) || Span.isEmptySet()) {
456 // If we already have a value for the switch, it has to match!
457 if(!setValueOnce(CandidateVal))
460 // Add all values from the range to the set
461 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
462 Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
469 /// Given a potentially 'or'd or 'and'd together collection of icmp
470 /// eq/ne/lt/gt instructions that compare a value against a constant, extract
471 /// the value being compared, and stick the list constants into the Vals
473 /// One "Extra" case is allowed to differ from the other.
474 void gather(Value *V) {
475 Instruction *I = dyn_cast<Instruction>(V);
476 bool isEQ = (I->getOpcode() == Instruction::Or);
478 // Keep a stack (SmallVector for efficiency) for depth-first traversal
479 SmallVector<Value *, 8> DFT;
484 while(!DFT.empty()) {
485 V = DFT.pop_back_val();
487 if (Instruction *I = dyn_cast<Instruction>(V)) {
488 // If it is a || (or && depending on isEQ), process the operands.
489 if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
490 DFT.push_back(I->getOperand(1));
491 DFT.push_back(I->getOperand(0));
495 // Try to match the current instruction
496 if (matchInstruction(I, isEQ))
497 // Match succeed, continue the loop
501 // One element of the sequence of || (or &&) could not be match as a
502 // comparison against the same value as the others.
503 // We allow only one "Extra" case to be checked before the switch
508 // Failed to parse a proper sequence, abort now
517 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
518 Instruction *Cond = nullptr;
519 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
520 Cond = dyn_cast<Instruction>(SI->getCondition());
521 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
522 if (BI->isConditional())
523 Cond = dyn_cast<Instruction>(BI->getCondition());
524 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
525 Cond = dyn_cast<Instruction>(IBI->getAddress());
528 TI->eraseFromParent();
529 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
532 /// Return true if the specified terminator checks
533 /// to see if a value is equal to constant integer value.
534 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
536 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
537 // Do not permit merging of large switch instructions into their
538 // predecessors unless there is only one predecessor.
539 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()),
540 pred_end(SI->getParent())) <= 128)
541 CV = SI->getCondition();
542 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
543 if (BI->isConditional() && BI->getCondition()->hasOneUse())
544 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
545 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
546 CV = ICI->getOperand(0);
549 // Unwrap any lossless ptrtoint cast.
551 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
552 Value *Ptr = PTII->getPointerOperand();
553 if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
560 /// Given a value comparison instruction,
561 /// decode all of the 'cases' that it represents and return the 'default' block.
562 BasicBlock *SimplifyCFGOpt::
563 GetValueEqualityComparisonCases(TerminatorInst *TI,
564 std::vector<ValueEqualityComparisonCase>
566 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
567 Cases.reserve(SI->getNumCases());
568 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
569 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
570 i.getCaseSuccessor()));
571 return SI->getDefaultDest();
574 BranchInst *BI = cast<BranchInst>(TI);
575 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
576 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
577 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
580 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
584 /// Given a vector of bb/value pairs, remove any entries
585 /// in the list that match the specified block.
586 static void EliminateBlockCases(BasicBlock *BB,
587 std::vector<ValueEqualityComparisonCase> &Cases) {
588 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
591 /// Return true if there are any keys in C1 that exist in C2 as well.
593 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
594 std::vector<ValueEqualityComparisonCase > &C2) {
595 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
597 // Make V1 be smaller than V2.
598 if (V1->size() > V2->size())
601 if (V1->size() == 0) return false;
602 if (V1->size() == 1) {
604 ConstantInt *TheVal = (*V1)[0].Value;
605 for (unsigned i = 0, e = V2->size(); i != e; ++i)
606 if (TheVal == (*V2)[i].Value)
610 // Otherwise, just sort both lists and compare element by element.
611 array_pod_sort(V1->begin(), V1->end());
612 array_pod_sort(V2->begin(), V2->end());
613 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
614 while (i1 != e1 && i2 != e2) {
615 if ((*V1)[i1].Value == (*V2)[i2].Value)
617 if ((*V1)[i1].Value < (*V2)[i2].Value)
625 /// If TI is known to be a terminator instruction and its block is known to
626 /// only have a single predecessor block, check to see if that predecessor is
627 /// also a value comparison with the same value, and if that comparison
628 /// determines the outcome of this comparison. If so, simplify TI. This does a
629 /// very limited form of jump threading.
630 bool SimplifyCFGOpt::
631 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
633 IRBuilder<> &Builder) {
634 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
635 if (!PredVal) return false; // Not a value comparison in predecessor.
637 Value *ThisVal = isValueEqualityComparison(TI);
638 assert(ThisVal && "This isn't a value comparison!!");
639 if (ThisVal != PredVal) return false; // Different predicates.
641 // TODO: Preserve branch weight metadata, similarly to how
642 // FoldValueComparisonIntoPredecessors preserves it.
644 // Find out information about when control will move from Pred to TI's block.
645 std::vector<ValueEqualityComparisonCase> PredCases;
646 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
648 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
650 // Find information about how control leaves this block.
651 std::vector<ValueEqualityComparisonCase> ThisCases;
652 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
653 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
655 // If TI's block is the default block from Pred's comparison, potentially
656 // simplify TI based on this knowledge.
657 if (PredDef == TI->getParent()) {
658 // If we are here, we know that the value is none of those cases listed in
659 // PredCases. If there are any cases in ThisCases that are in PredCases, we
661 if (!ValuesOverlap(PredCases, ThisCases))
664 if (isa<BranchInst>(TI)) {
665 // Okay, one of the successors of this condbr is dead. Convert it to a
667 assert(ThisCases.size() == 1 && "Branch can only have one case!");
668 // Insert the new branch.
669 Instruction *NI = Builder.CreateBr(ThisDef);
672 // Remove PHI node entries for the dead edge.
673 ThisCases[0].Dest->removePredecessor(TI->getParent());
675 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
676 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
678 EraseTerminatorInstAndDCECond(TI);
682 SwitchInst *SI = cast<SwitchInst>(TI);
683 // Okay, TI has cases that are statically dead, prune them away.
684 SmallPtrSet<Constant*, 16> DeadCases;
685 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
686 DeadCases.insert(PredCases[i].Value);
688 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
689 << "Through successor TI: " << *TI);
691 // Collect branch weights into a vector.
692 SmallVector<uint32_t, 8> Weights;
693 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
694 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
696 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
698 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
699 Weights.push_back(CI->getValue().getZExtValue());
701 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
703 if (DeadCases.count(i.getCaseValue())) {
705 std::swap(Weights[i.getCaseIndex()+1], Weights.back());
708 i.getCaseSuccessor()->removePredecessor(TI->getParent());
712 if (HasWeight && Weights.size() >= 2)
713 SI->setMetadata(LLVMContext::MD_prof,
714 MDBuilder(SI->getParent()->getContext()).
715 createBranchWeights(Weights));
717 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
721 // Otherwise, TI's block must correspond to some matched value. Find out
722 // which value (or set of values) this is.
723 ConstantInt *TIV = nullptr;
724 BasicBlock *TIBB = TI->getParent();
725 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
726 if (PredCases[i].Dest == TIBB) {
728 return false; // Cannot handle multiple values coming to this block.
729 TIV = PredCases[i].Value;
731 assert(TIV && "No edge from pred to succ?");
733 // Okay, we found the one constant that our value can be if we get into TI's
734 // BB. Find out which successor will unconditionally be branched to.
735 BasicBlock *TheRealDest = nullptr;
736 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
737 if (ThisCases[i].Value == TIV) {
738 TheRealDest = ThisCases[i].Dest;
742 // If not handled by any explicit cases, it is handled by the default case.
743 if (!TheRealDest) TheRealDest = ThisDef;
745 // Remove PHI node entries for dead edges.
746 BasicBlock *CheckEdge = TheRealDest;
747 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
748 if (*SI != CheckEdge)
749 (*SI)->removePredecessor(TIBB);
753 // Insert the new branch.
754 Instruction *NI = Builder.CreateBr(TheRealDest);
757 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
758 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
760 EraseTerminatorInstAndDCECond(TI);
765 /// This class implements a stable ordering of constant
766 /// integers that does not depend on their address. This is important for
767 /// applications that sort ConstantInt's to ensure uniqueness.
768 struct ConstantIntOrdering {
769 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
770 return LHS->getValue().ult(RHS->getValue());
775 static int ConstantIntSortPredicate(ConstantInt *const *P1,
776 ConstantInt *const *P2) {
777 const ConstantInt *LHS = *P1;
778 const ConstantInt *RHS = *P2;
779 if (LHS->getValue().ult(RHS->getValue()))
781 if (LHS->getValue() == RHS->getValue())
786 static inline bool HasBranchWeights(const Instruction* I) {
787 MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
788 if (ProfMD && ProfMD->getOperand(0))
789 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
790 return MDS->getString().equals("branch_weights");
795 /// Get Weights of a given TerminatorInst, the default weight is at the front
796 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
798 static void GetBranchWeights(TerminatorInst *TI,
799 SmallVectorImpl<uint64_t> &Weights) {
800 MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
802 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
803 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
804 Weights.push_back(CI->getValue().getZExtValue());
807 // If TI is a conditional eq, the default case is the false case,
808 // and the corresponding branch-weight data is at index 2. We swap the
809 // default weight to be the first entry.
810 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
811 assert(Weights.size() == 2);
812 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
813 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
814 std::swap(Weights.front(), Weights.back());
818 /// Keep halving the weights until all can fit in uint32_t.
819 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
820 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
821 if (Max > UINT_MAX) {
822 unsigned Offset = 32 - countLeadingZeros(Max);
823 for (uint64_t &I : Weights)
828 /// The specified terminator is a value equality comparison instruction
829 /// (either a switch or a branch on "X == c").
830 /// See if any of the predecessors of the terminator block are value comparisons
831 /// on the same value. If so, and if safe to do so, fold them together.
832 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
833 IRBuilder<> &Builder) {
834 BasicBlock *BB = TI->getParent();
835 Value *CV = isValueEqualityComparison(TI); // CondVal
836 assert(CV && "Not a comparison?");
837 bool Changed = false;
839 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
840 while (!Preds.empty()) {
841 BasicBlock *Pred = Preds.pop_back_val();
843 // See if the predecessor is a comparison with the same value.
844 TerminatorInst *PTI = Pred->getTerminator();
845 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
847 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
848 // Figure out which 'cases' to copy from SI to PSI.
849 std::vector<ValueEqualityComparisonCase> BBCases;
850 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
852 std::vector<ValueEqualityComparisonCase> PredCases;
853 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
855 // Based on whether the default edge from PTI goes to BB or not, fill in
856 // PredCases and PredDefault with the new switch cases we would like to
858 SmallVector<BasicBlock*, 8> NewSuccessors;
860 // Update the branch weight metadata along the way
861 SmallVector<uint64_t, 8> Weights;
862 bool PredHasWeights = HasBranchWeights(PTI);
863 bool SuccHasWeights = HasBranchWeights(TI);
865 if (PredHasWeights) {
866 GetBranchWeights(PTI, Weights);
867 // branch-weight metadata is inconsistent here.
868 if (Weights.size() != 1 + PredCases.size())
869 PredHasWeights = SuccHasWeights = false;
870 } else if (SuccHasWeights)
871 // If there are no predecessor weights but there are successor weights,
872 // populate Weights with 1, which will later be scaled to the sum of
873 // successor's weights
874 Weights.assign(1 + PredCases.size(), 1);
876 SmallVector<uint64_t, 8> SuccWeights;
877 if (SuccHasWeights) {
878 GetBranchWeights(TI, SuccWeights);
879 // branch-weight metadata is inconsistent here.
880 if (SuccWeights.size() != 1 + BBCases.size())
881 PredHasWeights = SuccHasWeights = false;
882 } else if (PredHasWeights)
883 SuccWeights.assign(1 + BBCases.size(), 1);
885 if (PredDefault == BB) {
886 // If this is the default destination from PTI, only the edges in TI
887 // that don't occur in PTI, or that branch to BB will be activated.
888 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
889 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
890 if (PredCases[i].Dest != BB)
891 PTIHandled.insert(PredCases[i].Value);
893 // The default destination is BB, we don't need explicit targets.
894 std::swap(PredCases[i], PredCases.back());
896 if (PredHasWeights || SuccHasWeights) {
897 // Increase weight for the default case.
898 Weights[0] += Weights[i+1];
899 std::swap(Weights[i+1], Weights.back());
903 PredCases.pop_back();
907 // Reconstruct the new switch statement we will be building.
908 if (PredDefault != BBDefault) {
909 PredDefault->removePredecessor(Pred);
910 PredDefault = BBDefault;
911 NewSuccessors.push_back(BBDefault);
914 unsigned CasesFromPred = Weights.size();
915 uint64_t ValidTotalSuccWeight = 0;
916 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
917 if (!PTIHandled.count(BBCases[i].Value) &&
918 BBCases[i].Dest != BBDefault) {
919 PredCases.push_back(BBCases[i]);
920 NewSuccessors.push_back(BBCases[i].Dest);
921 if (SuccHasWeights || PredHasWeights) {
922 // The default weight is at index 0, so weight for the ith case
923 // should be at index i+1. Scale the cases from successor by
924 // PredDefaultWeight (Weights[0]).
925 Weights.push_back(Weights[0] * SuccWeights[i+1]);
926 ValidTotalSuccWeight += SuccWeights[i+1];
930 if (SuccHasWeights || PredHasWeights) {
931 ValidTotalSuccWeight += SuccWeights[0];
932 // Scale the cases from predecessor by ValidTotalSuccWeight.
933 for (unsigned i = 1; i < CasesFromPred; ++i)
934 Weights[i] *= ValidTotalSuccWeight;
935 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
936 Weights[0] *= SuccWeights[0];
939 // If this is not the default destination from PSI, only the edges
940 // in SI that occur in PSI with a destination of BB will be
942 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
943 std::map<ConstantInt*, uint64_t> WeightsForHandled;
944 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
945 if (PredCases[i].Dest == BB) {
946 PTIHandled.insert(PredCases[i].Value);
948 if (PredHasWeights || SuccHasWeights) {
949 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
950 std::swap(Weights[i+1], Weights.back());
954 std::swap(PredCases[i], PredCases.back());
955 PredCases.pop_back();
959 // Okay, now we know which constants were sent to BB from the
960 // predecessor. Figure out where they will all go now.
961 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
962 if (PTIHandled.count(BBCases[i].Value)) {
963 // If this is one we are capable of getting...
964 if (PredHasWeights || SuccHasWeights)
965 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
966 PredCases.push_back(BBCases[i]);
967 NewSuccessors.push_back(BBCases[i].Dest);
968 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
971 // If there are any constants vectored to BB that TI doesn't handle,
972 // they must go to the default destination of TI.
973 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
975 E = PTIHandled.end(); I != E; ++I) {
976 if (PredHasWeights || SuccHasWeights)
977 Weights.push_back(WeightsForHandled[*I]);
978 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
979 NewSuccessors.push_back(BBDefault);
983 // Okay, at this point, we know which new successor Pred will get. Make
984 // sure we update the number of entries in the PHI nodes for these
986 for (BasicBlock *NewSuccessor : NewSuccessors)
987 AddPredecessorToBlock(NewSuccessor, Pred, BB);
989 Builder.SetInsertPoint(PTI);
990 // Convert pointer to int before we switch.
991 if (CV->getType()->isPointerTy()) {
992 CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
996 // Now that the successors are updated, create the new Switch instruction.
997 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
999 NewSI->setDebugLoc(PTI->getDebugLoc());
1000 for (ValueEqualityComparisonCase &V : PredCases)
1001 NewSI->addCase(V.Value, V.Dest);
1003 if (PredHasWeights || SuccHasWeights) {
1004 // Halve the weights if any of them cannot fit in an uint32_t
1005 FitWeights(Weights);
1007 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1009 NewSI->setMetadata(LLVMContext::MD_prof,
1010 MDBuilder(BB->getContext()).
1011 createBranchWeights(MDWeights));
1014 EraseTerminatorInstAndDCECond(PTI);
1016 // Okay, last check. If BB is still a successor of PSI, then we must
1017 // have an infinite loop case. If so, add an infinitely looping block
1018 // to handle the case to preserve the behavior of the code.
1019 BasicBlock *InfLoopBlock = nullptr;
1020 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1021 if (NewSI->getSuccessor(i) == BB) {
1022 if (!InfLoopBlock) {
1023 // Insert it at the end of the function, because it's either code,
1024 // or it won't matter if it's hot. :)
1025 InfLoopBlock = BasicBlock::Create(BB->getContext(),
1026 "infloop", BB->getParent());
1027 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1029 NewSI->setSuccessor(i, InfLoopBlock);
1038 // If we would need to insert a select that uses the value of this invoke
1039 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1040 // can't hoist the invoke, as there is nowhere to put the select in this case.
1041 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1042 Instruction *I1, Instruction *I2) {
1043 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1045 for (BasicBlock::iterator BBI = SI->begin();
1046 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1047 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1048 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1049 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
1057 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I);
1059 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1060 /// in the two blocks up into the branch block. The caller of this function
1061 /// guarantees that BI's block dominates BB1 and BB2.
1062 static bool HoistThenElseCodeToIf(BranchInst *BI,
1063 const TargetTransformInfo &TTI) {
1064 // This does very trivial matching, with limited scanning, to find identical
1065 // instructions in the two blocks. In particular, we don't want to get into
1066 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1067 // such, we currently just scan for obviously identical instructions in an
1069 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1070 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1072 BasicBlock::iterator BB1_Itr = BB1->begin();
1073 BasicBlock::iterator BB2_Itr = BB2->begin();
1075 Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1076 // Skip debug info if it is not identical.
1077 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1078 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1079 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1080 while (isa<DbgInfoIntrinsic>(I1))
1082 while (isa<DbgInfoIntrinsic>(I2))
1085 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1086 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1089 BasicBlock *BIParent = BI->getParent();
1091 bool Changed = false;
1093 // If we are hoisting the terminator instruction, don't move one (making a
1094 // broken BB), instead clone it, and remove BI.
1095 if (isa<TerminatorInst>(I1))
1096 goto HoistTerminator;
1098 if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1101 // For a normal instruction, we just move one to right before the branch,
1102 // then replace all uses of the other with the first. Finally, we remove
1103 // the now redundant second instruction.
1104 BIParent->getInstList().splice(BI->getIterator(), BB1->getInstList(), I1);
1105 if (!I2->use_empty())
1106 I2->replaceAllUsesWith(I1);
1107 I1->intersectOptionalDataWith(I2);
1108 unsigned KnownIDs[] = {
1109 LLVMContext::MD_tbaa, LLVMContext::MD_range,
1110 LLVMContext::MD_fpmath, LLVMContext::MD_invariant_load,
1111 LLVMContext::MD_nonnull, LLVMContext::MD_invariant_group};
1112 combineMetadata(I1, I2, KnownIDs);
1113 I2->eraseFromParent();
1118 // Skip debug info if it is not identical.
1119 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1120 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1121 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1122 while (isa<DbgInfoIntrinsic>(I1))
1124 while (isa<DbgInfoIntrinsic>(I2))
1127 } while (I1->isIdenticalToWhenDefined(I2));
1132 // It may not be possible to hoist an invoke.
1133 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1136 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1138 for (BasicBlock::iterator BBI = SI->begin();
1139 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1140 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1141 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1145 // Check for passingValueIsAlwaysUndefined here because we would rather
1146 // eliminate undefined control flow then converting it to a select.
1147 if (passingValueIsAlwaysUndefined(BB1V, PN) ||
1148 passingValueIsAlwaysUndefined(BB2V, PN))
1151 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1153 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1158 // Okay, it is safe to hoist the terminator.
1159 Instruction *NT = I1->clone();
1160 BIParent->getInstList().insert(BI->getIterator(), NT);
1161 if (!NT->getType()->isVoidTy()) {
1162 I1->replaceAllUsesWith(NT);
1163 I2->replaceAllUsesWith(NT);
1167 IRBuilder<true, NoFolder> Builder(NT);
1168 // Hoisting one of the terminators from our successor is a great thing.
1169 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1170 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1171 // nodes, so we insert select instruction to compute the final result.
1172 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1173 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1175 for (BasicBlock::iterator BBI = SI->begin();
1176 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1177 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1178 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1179 if (BB1V == BB2V) continue;
1181 // These values do not agree. Insert a select instruction before NT
1182 // that determines the right value.
1183 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1185 SI = cast<SelectInst>
1186 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1187 BB1V->getName()+"."+BB2V->getName()));
1189 // Make the PHI node use the select for all incoming values for BB1/BB2
1190 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1191 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1192 PN->setIncomingValue(i, SI);
1196 // Update any PHI nodes in our new successors.
1197 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1198 AddPredecessorToBlock(*SI, BIParent, BB1);
1200 EraseTerminatorInstAndDCECond(BI);
1204 /// Given an unconditional branch that goes to BBEnd,
1205 /// check whether BBEnd has only two predecessors and the other predecessor
1206 /// ends with an unconditional branch. If it is true, sink any common code
1207 /// in the two predecessors to BBEnd.
1208 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1209 assert(BI1->isUnconditional());
1210 BasicBlock *BB1 = BI1->getParent();
1211 BasicBlock *BBEnd = BI1->getSuccessor(0);
1213 // Check that BBEnd has two predecessors and the other predecessor ends with
1214 // an unconditional branch.
1215 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1216 BasicBlock *Pred0 = *PI++;
1217 if (PI == PE) // Only one predecessor.
1219 BasicBlock *Pred1 = *PI++;
1220 if (PI != PE) // More than two predecessors.
1222 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1223 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1224 if (!BI2 || !BI2->isUnconditional())
1227 // Gather the PHI nodes in BBEnd.
1228 SmallDenseMap<std::pair<Value *, Value *>, PHINode *> JointValueMap;
1229 Instruction *FirstNonPhiInBBEnd = nullptr;
1230 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end(); I != E; ++I) {
1231 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1232 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1233 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1234 JointValueMap[std::make_pair(BB1V, BB2V)] = PN;
1236 FirstNonPhiInBBEnd = &*I;
1240 if (!FirstNonPhiInBBEnd)
1243 // This does very trivial matching, with limited scanning, to find identical
1244 // instructions in the two blocks. We scan backward for obviously identical
1245 // instructions in an identical order.
1246 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1247 RE1 = BB1->getInstList().rend(),
1248 RI2 = BB2->getInstList().rbegin(),
1249 RE2 = BB2->getInstList().rend();
1251 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1254 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1257 // Skip the unconditional branches.
1261 bool Changed = false;
1262 while (RI1 != RE1 && RI2 != RE2) {
1264 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1267 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1271 Instruction *I1 = &*RI1, *I2 = &*RI2;
1272 auto InstPair = std::make_pair(I1, I2);
1273 // I1 and I2 should have a single use in the same PHI node, and they
1274 // perform the same operation.
1275 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1276 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1277 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1278 I1->isEHPad() || I2->isEHPad() ||
1279 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1280 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1281 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1282 !I1->hasOneUse() || !I2->hasOneUse() ||
1283 !JointValueMap.count(InstPair))
1286 // Check whether we should swap the operands of ICmpInst.
1287 // TODO: Add support of communativity.
1288 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1289 bool SwapOpnds = false;
1290 if (ICmp1 && ICmp2 &&
1291 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1292 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1293 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1294 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1295 ICmp2->swapOperands();
1298 if (!I1->isSameOperationAs(I2)) {
1300 ICmp2->swapOperands();
1304 // The operands should be either the same or they need to be generated
1305 // with a PHI node after sinking. We only handle the case where there is
1306 // a single pair of different operands.
1307 Value *DifferentOp1 = nullptr, *DifferentOp2 = nullptr;
1308 unsigned Op1Idx = ~0U;
1309 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1310 if (I1->getOperand(I) == I2->getOperand(I))
1312 // Early exit if we have more-than one pair of different operands or if
1313 // we need a PHI node to replace a constant.
1314 if (Op1Idx != ~0U ||
1315 isa<Constant>(I1->getOperand(I)) ||
1316 isa<Constant>(I2->getOperand(I))) {
1317 // If we can't sink the instructions, undo the swapping.
1319 ICmp2->swapOperands();
1322 DifferentOp1 = I1->getOperand(I);
1324 DifferentOp2 = I2->getOperand(I);
1327 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n");
1328 DEBUG(dbgs() << " " << *I2 << "\n");
1330 // We insert the pair of different operands to JointValueMap and
1331 // remove (I1, I2) from JointValueMap.
1332 if (Op1Idx != ~0U) {
1333 auto &NewPN = JointValueMap[std::make_pair(DifferentOp1, DifferentOp2)];
1336 PHINode::Create(DifferentOp1->getType(), 2,
1337 DifferentOp1->getName() + ".sink", &BBEnd->front());
1338 NewPN->addIncoming(DifferentOp1, BB1);
1339 NewPN->addIncoming(DifferentOp2, BB2);
1340 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1342 // I1 should use NewPN instead of DifferentOp1.
1343 I1->setOperand(Op1Idx, NewPN);
1345 PHINode *OldPN = JointValueMap[InstPair];
1346 JointValueMap.erase(InstPair);
1348 // We need to update RE1 and RE2 if we are going to sink the first
1349 // instruction in the basic block down.
1350 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1351 // Sink the instruction.
1352 BBEnd->getInstList().splice(FirstNonPhiInBBEnd->getIterator(),
1353 BB1->getInstList(), I1);
1354 if (!OldPN->use_empty())
1355 OldPN->replaceAllUsesWith(I1);
1356 OldPN->eraseFromParent();
1358 if (!I2->use_empty())
1359 I2->replaceAllUsesWith(I1);
1360 I1->intersectOptionalDataWith(I2);
1361 // TODO: Use combineMetadata here to preserve what metadata we can
1362 // (analogous to the hoisting case above).
1363 I2->eraseFromParent();
1366 RE1 = BB1->getInstList().rend();
1368 RE2 = BB2->getInstList().rend();
1369 FirstNonPhiInBBEnd = &*I1;
1376 /// \brief Determine if we can hoist sink a sole store instruction out of a
1377 /// conditional block.
1379 /// We are looking for code like the following:
1381 /// store i32 %add, i32* %arrayidx2
1382 /// ... // No other stores or function calls (we could be calling a memory
1383 /// ... // function).
1384 /// %cmp = icmp ult %x, %y
1385 /// br i1 %cmp, label %EndBB, label %ThenBB
1387 /// store i32 %add5, i32* %arrayidx2
1391 /// We are going to transform this into:
1393 /// store i32 %add, i32* %arrayidx2
1395 /// %cmp = icmp ult %x, %y
1396 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1397 /// store i32 %add.add5, i32* %arrayidx2
1400 /// \return The pointer to the value of the previous store if the store can be
1401 /// hoisted into the predecessor block. 0 otherwise.
1402 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1403 BasicBlock *StoreBB, BasicBlock *EndBB) {
1404 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1408 // Volatile or atomic.
1409 if (!StoreToHoist->isSimple())
1412 Value *StorePtr = StoreToHoist->getPointerOperand();
1414 // Look for a store to the same pointer in BrBB.
1415 unsigned MaxNumInstToLookAt = 10;
1416 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(),
1417 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) {
1418 Instruction *CurI = &*RI;
1420 // Could be calling an instruction that effects memory like free().
1421 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI))
1424 StoreInst *SI = dyn_cast<StoreInst>(CurI);
1425 // Found the previous store make sure it stores to the same location.
1426 if (SI && SI->getPointerOperand() == StorePtr)
1427 // Found the previous store, return its value operand.
1428 return SI->getValueOperand();
1430 return nullptr; // Unknown store.
1436 /// \brief Speculate a conditional basic block flattening the CFG.
1438 /// Note that this is a very risky transform currently. Speculating
1439 /// instructions like this is most often not desirable. Instead, there is an MI
1440 /// pass which can do it with full awareness of the resource constraints.
1441 /// However, some cases are "obvious" and we should do directly. An example of
1442 /// this is speculating a single, reasonably cheap instruction.
1444 /// There is only one distinct advantage to flattening the CFG at the IR level:
1445 /// it makes very common but simplistic optimizations such as are common in
1446 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1447 /// modeling their effects with easier to reason about SSA value graphs.
1450 /// An illustration of this transform is turning this IR:
1453 /// %cmp = icmp ult %x, %y
1454 /// br i1 %cmp, label %EndBB, label %ThenBB
1456 /// %sub = sub %x, %y
1459 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1466 /// %cmp = icmp ult %x, %y
1467 /// %sub = sub %x, %y
1468 /// %cond = select i1 %cmp, 0, %sub
1472 /// \returns true if the conditional block is removed.
1473 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1474 const TargetTransformInfo &TTI) {
1475 // Be conservative for now. FP select instruction can often be expensive.
1476 Value *BrCond = BI->getCondition();
1477 if (isa<FCmpInst>(BrCond))
1480 BasicBlock *BB = BI->getParent();
1481 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1483 // If ThenBB is actually on the false edge of the conditional branch, remember
1484 // to swap the select operands later.
1485 bool Invert = false;
1486 if (ThenBB != BI->getSuccessor(0)) {
1487 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1490 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1492 // Keep a count of how many times instructions are used within CondBB when
1493 // they are candidates for sinking into CondBB. Specifically:
1494 // - They are defined in BB, and
1495 // - They have no side effects, and
1496 // - All of their uses are in CondBB.
1497 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1499 unsigned SpeculationCost = 0;
1500 Value *SpeculatedStoreValue = nullptr;
1501 StoreInst *SpeculatedStore = nullptr;
1502 for (BasicBlock::iterator BBI = ThenBB->begin(),
1503 BBE = std::prev(ThenBB->end());
1504 BBI != BBE; ++BBI) {
1505 Instruction *I = &*BBI;
1507 if (isa<DbgInfoIntrinsic>(I))
1510 // Only speculatively execute a single instruction (not counting the
1511 // terminator) for now.
1513 if (SpeculationCost > 1)
1516 // Don't hoist the instruction if it's unsafe or expensive.
1517 if (!isSafeToSpeculativelyExecute(I) &&
1518 !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1519 I, BB, ThenBB, EndBB))))
1521 if (!SpeculatedStoreValue &&
1522 ComputeSpeculationCost(I, TTI) >
1523 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
1526 // Store the store speculation candidate.
1527 if (SpeculatedStoreValue)
1528 SpeculatedStore = cast<StoreInst>(I);
1530 // Do not hoist the instruction if any of its operands are defined but not
1531 // used in BB. The transformation will prevent the operand from
1532 // being sunk into the use block.
1533 for (User::op_iterator i = I->op_begin(), e = I->op_end();
1535 Instruction *OpI = dyn_cast<Instruction>(*i);
1536 if (!OpI || OpI->getParent() != BB ||
1537 OpI->mayHaveSideEffects())
1538 continue; // Not a candidate for sinking.
1540 ++SinkCandidateUseCounts[OpI];
1544 // Consider any sink candidates which are only used in CondBB as costs for
1545 // speculation. Note, while we iterate over a DenseMap here, we are summing
1546 // and so iteration order isn't significant.
1547 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I =
1548 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end();
1550 if (I->first->getNumUses() == I->second) {
1552 if (SpeculationCost > 1)
1556 // Check that the PHI nodes can be converted to selects.
1557 bool HaveRewritablePHIs = false;
1558 for (BasicBlock::iterator I = EndBB->begin();
1559 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1560 Value *OrigV = PN->getIncomingValueForBlock(BB);
1561 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1563 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1564 // Skip PHIs which are trivial.
1568 // Don't convert to selects if we could remove undefined behavior instead.
1569 if (passingValueIsAlwaysUndefined(OrigV, PN) ||
1570 passingValueIsAlwaysUndefined(ThenV, PN))
1573 HaveRewritablePHIs = true;
1574 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1575 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1576 if (!OrigCE && !ThenCE)
1577 continue; // Known safe and cheap.
1579 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
1580 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
1582 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
1583 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
1584 unsigned MaxCost = 2 * PHINodeFoldingThreshold *
1585 TargetTransformInfo::TCC_Basic;
1586 if (OrigCost + ThenCost > MaxCost)
1589 // Account for the cost of an unfolded ConstantExpr which could end up
1590 // getting expanded into Instructions.
1591 // FIXME: This doesn't account for how many operations are combined in the
1592 // constant expression.
1594 if (SpeculationCost > 1)
1598 // If there are no PHIs to process, bail early. This helps ensure idempotence
1600 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
1603 // If we get here, we can hoist the instruction and if-convert.
1604 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1606 // Insert a select of the value of the speculated store.
1607 if (SpeculatedStoreValue) {
1608 IRBuilder<true, NoFolder> Builder(BI);
1609 Value *TrueV = SpeculatedStore->getValueOperand();
1610 Value *FalseV = SpeculatedStoreValue;
1612 std::swap(TrueV, FalseV);
1613 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() +
1614 "." + FalseV->getName());
1615 SpeculatedStore->setOperand(0, S);
1618 // Hoist the instructions.
1619 BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
1620 ThenBB->begin(), std::prev(ThenBB->end()));
1622 // Insert selects and rewrite the PHI operands.
1623 IRBuilder<true, NoFolder> Builder(BI);
1624 for (BasicBlock::iterator I = EndBB->begin();
1625 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1626 unsigned OrigI = PN->getBasicBlockIndex(BB);
1627 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1628 Value *OrigV = PN->getIncomingValue(OrigI);
1629 Value *ThenV = PN->getIncomingValue(ThenI);
1631 // Skip PHIs which are trivial.
1635 // Create a select whose true value is the speculatively executed value and
1636 // false value is the preexisting value. Swap them if the branch
1637 // destinations were inverted.
1638 Value *TrueV = ThenV, *FalseV = OrigV;
1640 std::swap(TrueV, FalseV);
1641 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV,
1642 TrueV->getName() + "." + FalseV->getName());
1643 PN->setIncomingValue(OrigI, V);
1644 PN->setIncomingValue(ThenI, V);
1651 /// \returns True if this block contains a CallInst with the NoDuplicate
1653 static bool HasNoDuplicateCall(const BasicBlock *BB) {
1654 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1655 const CallInst *CI = dyn_cast<CallInst>(I);
1658 if (CI->cannotDuplicate())
1664 /// Return true if we can thread a branch across this block.
1665 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1666 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1669 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1670 if (isa<DbgInfoIntrinsic>(BBI))
1672 if (Size > 10) return false; // Don't clone large BB's.
1675 // We can only support instructions that do not define values that are
1676 // live outside of the current basic block.
1677 for (User *U : BBI->users()) {
1678 Instruction *UI = cast<Instruction>(U);
1679 if (UI->getParent() != BB || isa<PHINode>(UI)) return false;
1682 // Looks ok, continue checking.
1688 /// If we have a conditional branch on a PHI node value that is defined in the
1689 /// same block as the branch and if any PHI entries are constants, thread edges
1690 /// corresponding to that entry to be branches to their ultimate destination.
1691 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL) {
1692 BasicBlock *BB = BI->getParent();
1693 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1694 // NOTE: we currently cannot transform this case if the PHI node is used
1695 // outside of the block.
1696 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1699 // Degenerate case of a single entry PHI.
1700 if (PN->getNumIncomingValues() == 1) {
1701 FoldSingleEntryPHINodes(PN->getParent());
1705 // Now we know that this block has multiple preds and two succs.
1706 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1708 if (HasNoDuplicateCall(BB)) return false;
1710 // Okay, this is a simple enough basic block. See if any phi values are
1712 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1713 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1714 if (!CB || !CB->getType()->isIntegerTy(1)) continue;
1716 // Okay, we now know that all edges from PredBB should be revectored to
1717 // branch to RealDest.
1718 BasicBlock *PredBB = PN->getIncomingBlock(i);
1719 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1721 if (RealDest == BB) continue; // Skip self loops.
1722 // Skip if the predecessor's terminator is an indirect branch.
1723 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1725 // The dest block might have PHI nodes, other predecessors and other
1726 // difficult cases. Instead of being smart about this, just insert a new
1727 // block that jumps to the destination block, effectively splitting
1728 // the edge we are about to create.
1729 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1730 RealDest->getName()+".critedge",
1731 RealDest->getParent(), RealDest);
1732 BranchInst::Create(RealDest, EdgeBB);
1734 // Update PHI nodes.
1735 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1737 // BB may have instructions that are being threaded over. Clone these
1738 // instructions into EdgeBB. We know that there will be no uses of the
1739 // cloned instructions outside of EdgeBB.
1740 BasicBlock::iterator InsertPt = EdgeBB->begin();
1741 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1742 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1743 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1744 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1747 // Clone the instruction.
1748 Instruction *N = BBI->clone();
1749 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1751 // Update operands due to translation.
1752 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1754 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1755 if (PI != TranslateMap.end())
1759 // Check for trivial simplification.
1760 if (Value *V = SimplifyInstruction(N, DL)) {
1761 TranslateMap[&*BBI] = V;
1762 delete N; // Instruction folded away, don't need actual inst
1764 // Insert the new instruction into its new home.
1765 EdgeBB->getInstList().insert(InsertPt, N);
1766 if (!BBI->use_empty())
1767 TranslateMap[&*BBI] = N;
1771 // Loop over all of the edges from PredBB to BB, changing them to branch
1772 // to EdgeBB instead.
1773 TerminatorInst *PredBBTI = PredBB->getTerminator();
1774 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1775 if (PredBBTI->getSuccessor(i) == BB) {
1776 BB->removePredecessor(PredBB);
1777 PredBBTI->setSuccessor(i, EdgeBB);
1780 // Recurse, simplifying any other constants.
1781 return FoldCondBranchOnPHI(BI, DL) | true;
1787 /// Given a BB that starts with the specified two-entry PHI node,
1788 /// see if we can eliminate it.
1789 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI,
1790 const DataLayout &DL) {
1791 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1792 // statement", which has a very simple dominance structure. Basically, we
1793 // are trying to find the condition that is being branched on, which
1794 // subsequently causes this merge to happen. We really want control
1795 // dependence information for this check, but simplifycfg can't keep it up
1796 // to date, and this catches most of the cases we care about anyway.
1797 BasicBlock *BB = PN->getParent();
1798 BasicBlock *IfTrue, *IfFalse;
1799 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1801 // Don't bother if the branch will be constant folded trivially.
1802 isa<ConstantInt>(IfCond))
1805 // Okay, we found that we can merge this two-entry phi node into a select.
1806 // Doing so would require us to fold *all* two entry phi nodes in this block.
1807 // At some point this becomes non-profitable (particularly if the target
1808 // doesn't support cmov's). Only do this transformation if there are two or
1809 // fewer PHI nodes in this block.
1810 unsigned NumPhis = 0;
1811 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1815 // Loop over the PHI's seeing if we can promote them all to select
1816 // instructions. While we are at it, keep track of the instructions
1817 // that need to be moved to the dominating block.
1818 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1819 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1820 MaxCostVal1 = PHINodeFoldingThreshold;
1821 MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
1822 MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
1824 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1825 PHINode *PN = cast<PHINode>(II++);
1826 if (Value *V = SimplifyInstruction(PN, DL)) {
1827 PN->replaceAllUsesWith(V);
1828 PN->eraseFromParent();
1832 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1833 MaxCostVal0, TTI) ||
1834 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1839 // If we folded the first phi, PN dangles at this point. Refresh it. If
1840 // we ran out of PHIs then we simplified them all.
1841 PN = dyn_cast<PHINode>(BB->begin());
1842 if (!PN) return true;
1844 // Don't fold i1 branches on PHIs which contain binary operators. These can
1845 // often be turned into switches and other things.
1846 if (PN->getType()->isIntegerTy(1) &&
1847 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1848 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1849 isa<BinaryOperator>(IfCond)))
1852 // If we all PHI nodes are promotable, check to make sure that all
1853 // instructions in the predecessor blocks can be promoted as well. If
1854 // not, we won't be able to get rid of the control flow, so it's not
1855 // worth promoting to select instructions.
1856 BasicBlock *DomBlock = nullptr;
1857 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1858 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1859 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1862 DomBlock = *pred_begin(IfBlock1);
1863 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1864 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
1865 // This is not an aggressive instruction that we can promote.
1866 // Because of this, we won't be able to get rid of the control
1867 // flow, so the xform is not worth it.
1872 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1875 DomBlock = *pred_begin(IfBlock2);
1876 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1877 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
1878 // This is not an aggressive instruction that we can promote.
1879 // Because of this, we won't be able to get rid of the control
1880 // flow, so the xform is not worth it.
1885 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1886 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1888 // If we can still promote the PHI nodes after this gauntlet of tests,
1889 // do all of the PHI's now.
1890 Instruction *InsertPt = DomBlock->getTerminator();
1891 IRBuilder<true, NoFolder> Builder(InsertPt);
1893 // Move all 'aggressive' instructions, which are defined in the
1894 // conditional parts of the if's up to the dominating block.
1896 DomBlock->getInstList().splice(InsertPt->getIterator(),
1897 IfBlock1->getInstList(), IfBlock1->begin(),
1898 IfBlock1->getTerminator()->getIterator());
1900 DomBlock->getInstList().splice(InsertPt->getIterator(),
1901 IfBlock2->getInstList(), IfBlock2->begin(),
1902 IfBlock2->getTerminator()->getIterator());
1904 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1905 // Change the PHI node into a select instruction.
1906 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1907 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1910 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1911 PN->replaceAllUsesWith(NV);
1913 PN->eraseFromParent();
1916 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1917 // has been flattened. Change DomBlock to jump directly to our new block to
1918 // avoid other simplifycfg's kicking in on the diamond.
1919 TerminatorInst *OldTI = DomBlock->getTerminator();
1920 Builder.SetInsertPoint(OldTI);
1921 Builder.CreateBr(BB);
1922 OldTI->eraseFromParent();
1926 /// If we found a conditional branch that goes to two returning blocks,
1927 /// try to merge them together into one return,
1928 /// introducing a select if the return values disagree.
1929 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1930 IRBuilder<> &Builder) {
1931 assert(BI->isConditional() && "Must be a conditional branch");
1932 BasicBlock *TrueSucc = BI->getSuccessor(0);
1933 BasicBlock *FalseSucc = BI->getSuccessor(1);
1934 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1935 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1937 // Check to ensure both blocks are empty (just a return) or optionally empty
1938 // with PHI nodes. If there are other instructions, merging would cause extra
1939 // computation on one path or the other.
1940 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1942 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1945 Builder.SetInsertPoint(BI);
1946 // Okay, we found a branch that is going to two return nodes. If
1947 // there is no return value for this function, just change the
1948 // branch into a return.
1949 if (FalseRet->getNumOperands() == 0) {
1950 TrueSucc->removePredecessor(BI->getParent());
1951 FalseSucc->removePredecessor(BI->getParent());
1952 Builder.CreateRetVoid();
1953 EraseTerminatorInstAndDCECond(BI);
1957 // Otherwise, figure out what the true and false return values are
1958 // so we can insert a new select instruction.
1959 Value *TrueValue = TrueRet->getReturnValue();
1960 Value *FalseValue = FalseRet->getReturnValue();
1962 // Unwrap any PHI nodes in the return blocks.
1963 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1964 if (TVPN->getParent() == TrueSucc)
1965 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1966 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1967 if (FVPN->getParent() == FalseSucc)
1968 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1970 // In order for this transformation to be safe, we must be able to
1971 // unconditionally execute both operands to the return. This is
1972 // normally the case, but we could have a potentially-trapping
1973 // constant expression that prevents this transformation from being
1975 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1978 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1982 // Okay, we collected all the mapped values and checked them for sanity, and
1983 // defined to really do this transformation. First, update the CFG.
1984 TrueSucc->removePredecessor(BI->getParent());
1985 FalseSucc->removePredecessor(BI->getParent());
1987 // Insert select instructions where needed.
1988 Value *BrCond = BI->getCondition();
1990 // Insert a select if the results differ.
1991 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1992 } else if (isa<UndefValue>(TrueValue)) {
1993 TrueValue = FalseValue;
1995 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1996 FalseValue, "retval");
2000 Value *RI = !TrueValue ?
2001 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2005 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2006 << "\n " << *BI << "NewRet = " << *RI
2007 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
2009 EraseTerminatorInstAndDCECond(BI);
2014 /// Given a conditional BranchInstruction, retrieve the probabilities of the
2015 /// branch taking each edge. Fills in the two APInt parameters and returns true,
2016 /// or returns false if no or invalid metadata was found.
2017 static bool ExtractBranchMetadata(BranchInst *BI,
2018 uint64_t &ProbTrue, uint64_t &ProbFalse) {
2019 assert(BI->isConditional() &&
2020 "Looking for probabilities on unconditional branch?");
2021 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
2022 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
2023 ConstantInt *CITrue =
2024 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
2025 ConstantInt *CIFalse =
2026 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
2027 if (!CITrue || !CIFalse) return false;
2028 ProbTrue = CITrue->getValue().getZExtValue();
2029 ProbFalse = CIFalse->getValue().getZExtValue();
2033 /// Return true if the given instruction is available
2034 /// in its predecessor block. If yes, the instruction will be removed.
2035 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
2036 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2038 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
2039 Instruction *PBI = &*I;
2040 // Check whether Inst and PBI generate the same value.
2041 if (Inst->isIdenticalTo(PBI)) {
2042 Inst->replaceAllUsesWith(PBI);
2043 Inst->eraseFromParent();
2050 /// If this basic block is simple enough, and if a predecessor branches to us
2051 /// and one of our successors, fold the block into the predecessor and use
2052 /// logical operations to pick the right destination.
2053 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2054 BasicBlock *BB = BI->getParent();
2056 Instruction *Cond = nullptr;
2057 if (BI->isConditional())
2058 Cond = dyn_cast<Instruction>(BI->getCondition());
2060 // For unconditional branch, check for a simple CFG pattern, where
2061 // BB has a single predecessor and BB's successor is also its predecessor's
2062 // successor. If such pattern exisits, check for CSE between BB and its
2064 if (BasicBlock *PB = BB->getSinglePredecessor())
2065 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2066 if (PBI->isConditional() &&
2067 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2068 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2069 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
2071 Instruction *Curr = &*I++;
2072 if (isa<CmpInst>(Curr)) {
2076 // Quit if we can't remove this instruction.
2077 if (!checkCSEInPredecessor(Curr, PB))
2086 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2087 Cond->getParent() != BB || !Cond->hasOneUse())
2090 // Make sure the instruction after the condition is the cond branch.
2091 BasicBlock::iterator CondIt = ++Cond->getIterator();
2093 // Ignore dbg intrinsics.
2094 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
2099 // Only allow this transformation if computing the condition doesn't involve
2100 // too many instructions and these involved instructions can be executed
2101 // unconditionally. We denote all involved instructions except the condition
2102 // as "bonus instructions", and only allow this transformation when the
2103 // number of the bonus instructions does not exceed a certain threshold.
2104 unsigned NumBonusInsts = 0;
2105 for (auto I = BB->begin(); Cond != I; ++I) {
2106 // Ignore dbg intrinsics.
2107 if (isa<DbgInfoIntrinsic>(I))
2109 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2111 // I has only one use and can be executed unconditionally.
2112 Instruction *User = dyn_cast<Instruction>(I->user_back());
2113 if (User == nullptr || User->getParent() != BB)
2115 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2116 // to use any other instruction, User must be an instruction between next(I)
2119 // Early exits once we reach the limit.
2120 if (NumBonusInsts > BonusInstThreshold)
2124 // Cond is known to be a compare or binary operator. Check to make sure that
2125 // neither operand is a potentially-trapping constant expression.
2126 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2129 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2133 // Finally, don't infinitely unroll conditional loops.
2134 BasicBlock *TrueDest = BI->getSuccessor(0);
2135 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2136 if (TrueDest == BB || FalseDest == BB)
2139 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2140 BasicBlock *PredBlock = *PI;
2141 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2143 // Check that we have two conditional branches. If there is a PHI node in
2144 // the common successor, verify that the same value flows in from both
2146 SmallVector<PHINode*, 4> PHIs;
2147 if (!PBI || PBI->isUnconditional() ||
2148 (BI->isConditional() &&
2149 !SafeToMergeTerminators(BI, PBI)) ||
2150 (!BI->isConditional() &&
2151 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2154 // Determine if the two branches share a common destination.
2155 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2156 bool InvertPredCond = false;
2158 if (BI->isConditional()) {
2159 if (PBI->getSuccessor(0) == TrueDest)
2160 Opc = Instruction::Or;
2161 else if (PBI->getSuccessor(1) == FalseDest)
2162 Opc = Instruction::And;
2163 else if (PBI->getSuccessor(0) == FalseDest)
2164 Opc = Instruction::And, InvertPredCond = true;
2165 else if (PBI->getSuccessor(1) == TrueDest)
2166 Opc = Instruction::Or, InvertPredCond = true;
2170 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2174 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2175 IRBuilder<> Builder(PBI);
2177 // If we need to invert the condition in the pred block to match, do so now.
2178 if (InvertPredCond) {
2179 Value *NewCond = PBI->getCondition();
2181 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2182 CmpInst *CI = cast<CmpInst>(NewCond);
2183 CI->setPredicate(CI->getInversePredicate());
2185 NewCond = Builder.CreateNot(NewCond,
2186 PBI->getCondition()->getName()+".not");
2189 PBI->setCondition(NewCond);
2190 PBI->swapSuccessors();
2193 // If we have bonus instructions, clone them into the predecessor block.
2194 // Note that there may be multiple predecessor blocks, so we cannot move
2195 // bonus instructions to a predecessor block.
2196 ValueToValueMapTy VMap; // maps original values to cloned values
2197 // We already make sure Cond is the last instruction before BI. Therefore,
2198 // all instructions before Cond other than DbgInfoIntrinsic are bonus
2200 for (auto BonusInst = BB->begin(); Cond != BonusInst; ++BonusInst) {
2201 if (isa<DbgInfoIntrinsic>(BonusInst))
2203 Instruction *NewBonusInst = BonusInst->clone();
2204 RemapInstruction(NewBonusInst, VMap,
2205 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2206 VMap[&*BonusInst] = NewBonusInst;
2208 // If we moved a load, we cannot any longer claim any knowledge about
2209 // its potential value. The previous information might have been valid
2210 // only given the branch precondition.
2211 // For an analogous reason, we must also drop all the metadata whose
2212 // semantics we don't understand.
2213 NewBonusInst->dropUnknownNonDebugMetadata();
2215 PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2216 NewBonusInst->takeName(&*BonusInst);
2217 BonusInst->setName(BonusInst->getName() + ".old");
2220 // Clone Cond into the predecessor basic block, and or/and the
2221 // two conditions together.
2222 Instruction *New = Cond->clone();
2223 RemapInstruction(New, VMap,
2224 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2225 PredBlock->getInstList().insert(PBI->getIterator(), New);
2226 New->takeName(Cond);
2227 Cond->setName(New->getName() + ".old");
2229 if (BI->isConditional()) {
2230 Instruction *NewCond =
2231 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2233 PBI->setCondition(NewCond);
2235 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2236 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2238 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2240 SmallVector<uint64_t, 8> NewWeights;
2242 if (PBI->getSuccessor(0) == BB) {
2243 if (PredHasWeights && SuccHasWeights) {
2244 // PBI: br i1 %x, BB, FalseDest
2245 // BI: br i1 %y, TrueDest, FalseDest
2246 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2247 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2248 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2249 // TrueWeight for PBI * FalseWeight for BI.
2250 // We assume that total weights of a BranchInst can fit into 32 bits.
2251 // Therefore, we will not have overflow using 64-bit arithmetic.
2252 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2253 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2255 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2256 PBI->setSuccessor(0, TrueDest);
2258 if (PBI->getSuccessor(1) == BB) {
2259 if (PredHasWeights && SuccHasWeights) {
2260 // PBI: br i1 %x, TrueDest, BB
2261 // BI: br i1 %y, TrueDest, FalseDest
2262 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2263 // FalseWeight for PBI * TrueWeight for BI.
2264 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2265 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2266 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2267 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2269 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2270 PBI->setSuccessor(1, FalseDest);
2272 if (NewWeights.size() == 2) {
2273 // Halve the weights if any of them cannot fit in an uint32_t
2274 FitWeights(NewWeights);
2276 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2277 PBI->setMetadata(LLVMContext::MD_prof,
2278 MDBuilder(BI->getContext()).
2279 createBranchWeights(MDWeights));
2281 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2283 // Update PHI nodes in the common successors.
2284 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2285 ConstantInt *PBI_C = cast<ConstantInt>(
2286 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2287 assert(PBI_C->getType()->isIntegerTy(1));
2288 Instruction *MergedCond = nullptr;
2289 if (PBI->getSuccessor(0) == TrueDest) {
2290 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2291 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2292 // is false: !PBI_Cond and BI_Value
2293 Instruction *NotCond =
2294 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2297 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2302 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2303 PBI->getCondition(), MergedCond,
2306 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2307 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2308 // is false: PBI_Cond and BI_Value
2310 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2311 PBI->getCondition(), New,
2313 if (PBI_C->isOne()) {
2314 Instruction *NotCond =
2315 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2318 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2319 NotCond, MergedCond,
2324 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2327 // Change PBI from Conditional to Unconditional.
2328 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2329 EraseTerminatorInstAndDCECond(PBI);
2333 // TODO: If BB is reachable from all paths through PredBlock, then we
2334 // could replace PBI's branch probabilities with BI's.
2336 // Copy any debug value intrinsics into the end of PredBlock.
2337 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2338 if (isa<DbgInfoIntrinsic>(*I))
2339 I->clone()->insertBefore(PBI);
2347 /// Return true if B is known to be implied by A. A & B must be i1 (boolean)
2348 static bool implies(Value *A, Value *B, const DataLayout &DL) {
2349 assert(A->getType()->isIntegerTy(1) && B->getType()->isIntegerTy(1));
2350 // Note that the truth table for implication is the same as <=u on i1
2352 Value *Simplified = SimplifyICmpInst(ICmpInst::ICMP_ULE, A, B, DL);
2353 Constant *Con = dyn_cast_or_null<Constant>(Simplified);
2354 return Con && Con->isOneValue();
2357 /// If we have a series of tests leading to a frequently executed fallthrough
2358 /// path and we can test all the conditions at once, we can execute a single
2359 /// test on the fast path and figure out which condition failed on the slow
2360 /// path. This transformation considers a pair of branches at a time since
2361 /// recursively considering branches pairwise will cause an entire chain to
2362 /// collapse. This transformation is code size neutral, but makes it
2363 /// dynamically more expensive to fail either check. As such, we only want to
2364 /// do this if both checks are expected to essentially never fail.
2365 /// The key motivating examples are cases like:
2366 /// br (i < Length), in_bounds, fail1
2368 /// br (i+1 < Length), in_bounds2, fail2
2372 /// We can rewrite this as:
2373 /// br (i+1 < Length), in_bounds2, dispatch
2377 /// br (i < Length), fail2, fail1
2379 /// TODO: we could consider duplicating some (non-speculatable) instructions
2380 /// from BI->getParent() down both paths
2381 static bool SpeculativelyFlattenCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
2382 const DataLayout &DL) {
2383 auto *PredBB = PBI->getParent();
2384 auto *BB = BI->getParent();
2386 /// Is the failing path of this branch taken rarely if at all?
2387 auto isRarelyUntaken = [](BranchInst *BI) {
2390 if (!ExtractBranchMetadata(BI, ProbTrue, ProbFalse))
2392 return ProbTrue > ProbFalse * SpeculativeFlattenBias;
2395 if (PBI->getSuccessor(0) != BB ||
2396 !isRarelyUntaken(PBI) || !isRarelyUntaken(BI) ||
2397 !implies(BI->getCondition(), PBI->getCondition(), DL))
2400 // TODO: The following code performs a similiar, but slightly distinct
2401 // transformation to that done by SpeculativelyExecuteBB. We should consider
2402 // combining them at some point.
2404 // Can we speculate everything in the given block (except for the terminator
2405 // instruction) at the instruction boundary before 'At'?
2406 unsigned SpeculationCost = 0;
2407 for (Instruction &I : *BB) {
2408 if (isa<TerminatorInst>(I)) break;
2409 if (!isSafeToSpeculativelyExecute(&I, PBI))
2412 // Only flatten relatively small BBs to avoid making the bad case terrible
2413 if (SpeculationCost > SpeculativeFlattenThreshold || isa<CallInst>(I))
2417 DEBUG(dbgs() << "Outlining slow path comparison: "
2418 << *PBI->getCondition() << " implied by "
2419 << *BI->getCondition() << "\n");
2420 // See the example in the function comment.
2421 Value *WhichCond = PBI->getCondition();
2422 auto *Success = BI->getSuccessor(0);
2423 auto *FailPBI = PBI->getSuccessor(1);
2424 auto *FailBI = BI->getSuccessor(1);
2425 // Have PBI branch directly to the fast path using BI's condition, branch
2426 // to this BI's block for the slow path dispatch
2427 PBI->setSuccessor(0, Success);
2428 PBI->setSuccessor(1, BB);
2429 PBI->setCondition(BI->getCondition());
2430 // Rewrite BI to distinguish between the two failing cases
2431 BI->setSuccessor(0, FailBI);
2432 BI->setSuccessor(1, FailPBI);
2433 BI->setCondition(WhichCond);
2434 // Move all of the instructions from BI->getParent other than
2435 // the terminator into the fastpath. This requires speculating them past
2436 // the original PBI branch, but we checked for that legality above.
2437 // TODO: This doesn't handle dependent loads. We could duplicate those
2438 // down both paths, but that involves further code growth. We need to
2439 // figure out a good cost model here.
2440 PredBB->getInstList().splice(PBI, BB->getInstList(),
2441 BB->begin(), std::prev(BB->end()));
2443 // To be conservatively correct, drop all metadata on the rewritten
2444 // branches. TODO: update metadata
2445 PBI->dropUnknownNonDebugMetadata();
2446 BI->dropUnknownNonDebugMetadata();
2449 /// If we have a conditional branch as a predecessor of another block,
2450 /// this function tries to simplify it. We know
2451 /// that PBI and BI are both conditional branches, and BI is in one of the
2452 /// successor blocks of PBI - PBI branches to BI.
2453 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
2454 const DataLayout &DL) {
2455 assert(PBI->isConditional() && BI->isConditional());
2456 BasicBlock *BB = BI->getParent();
2458 // If this block ends with a branch instruction, and if there is a
2459 // predecessor that ends on a branch of the same condition, make
2460 // this conditional branch redundant.
2461 if (PBI->getCondition() == BI->getCondition() &&
2462 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2463 // Okay, the outcome of this conditional branch is statically
2464 // knowable. If this block had a single pred, handle specially.
2465 if (BB->getSinglePredecessor()) {
2466 // Turn this into a branch on constant.
2467 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2468 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2470 return true; // Nuke the branch on constant.
2473 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2474 // in the constant and simplify the block result. Subsequent passes of
2475 // simplifycfg will thread the block.
2476 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2477 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2478 PHINode *NewPN = PHINode::Create(
2479 Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
2480 BI->getCondition()->getName() + ".pr", &BB->front());
2481 // Okay, we're going to insert the PHI node. Since PBI is not the only
2482 // predecessor, compute the PHI'd conditional value for all of the preds.
2483 // Any predecessor where the condition is not computable we keep symbolic.
2484 for (pred_iterator PI = PB; PI != PE; ++PI) {
2485 BasicBlock *P = *PI;
2486 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2487 PBI != BI && PBI->isConditional() &&
2488 PBI->getCondition() == BI->getCondition() &&
2489 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2490 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2491 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2494 NewPN->addIncoming(BI->getCondition(), P);
2498 BI->setCondition(NewPN);
2503 if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2507 if (SpeculativelyFlattenCondBranchToCondBranch(PBI, BI, DL))
2510 // If this is a conditional branch in an empty block, and if any
2511 // predecessors are a conditional branch to one of our destinations,
2512 // fold the conditions into logical ops and one cond br.
2513 BasicBlock::iterator BBI = BB->begin();
2514 // Ignore dbg intrinsics.
2515 while (isa<DbgInfoIntrinsic>(BBI))
2521 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2523 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2524 PBIOp = 0, BIOp = 1;
2525 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2526 PBIOp = 1, BIOp = 0;
2527 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2532 // Check to make sure that the other destination of this branch
2533 // isn't BB itself. If so, this is an infinite loop that will
2534 // keep getting unwound.
2535 if (PBI->getSuccessor(PBIOp) == BB)
2538 // Do not perform this transformation if it would require
2539 // insertion of a large number of select instructions. For targets
2540 // without predication/cmovs, this is a big pessimization.
2542 // Also do not perform this transformation if any phi node in the common
2543 // destination block can trap when reached by BB or PBB (PR17073). In that
2544 // case, it would be unsafe to hoist the operation into a select instruction.
2546 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2547 unsigned NumPhis = 0;
2548 for (BasicBlock::iterator II = CommonDest->begin();
2549 isa<PHINode>(II); ++II, ++NumPhis) {
2550 if (NumPhis > 2) // Disable this xform.
2553 PHINode *PN = cast<PHINode>(II);
2554 Value *BIV = PN->getIncomingValueForBlock(BB);
2555 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
2559 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2560 Value *PBIV = PN->getIncomingValue(PBBIdx);
2561 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
2566 // Finally, if everything is ok, fold the branches to logical ops.
2567 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2569 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2570 << "AND: " << *BI->getParent());
2573 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2574 // branch in it, where one edge (OtherDest) goes back to itself but the other
2575 // exits. We don't *know* that the program avoids the infinite loop
2576 // (even though that seems likely). If we do this xform naively, we'll end up
2577 // recursively unpeeling the loop. Since we know that (after the xform is
2578 // done) that the block *is* infinite if reached, we just make it an obviously
2579 // infinite loop with no cond branch.
2580 if (OtherDest == BB) {
2581 // Insert it at the end of the function, because it's either code,
2582 // or it won't matter if it's hot. :)
2583 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2584 "infloop", BB->getParent());
2585 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2586 OtherDest = InfLoopBlock;
2589 DEBUG(dbgs() << *PBI->getParent()->getParent());
2591 // BI may have other predecessors. Because of this, we leave
2592 // it alone, but modify PBI.
2594 // Make sure we get to CommonDest on True&True directions.
2595 Value *PBICond = PBI->getCondition();
2596 IRBuilder<true, NoFolder> Builder(PBI);
2598 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2600 Value *BICond = BI->getCondition();
2602 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2604 // Merge the conditions.
2605 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2607 // Modify PBI to branch on the new condition to the new dests.
2608 PBI->setCondition(Cond);
2609 PBI->setSuccessor(0, CommonDest);
2610 PBI->setSuccessor(1, OtherDest);
2612 // Update branch weight for PBI.
2613 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2614 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2616 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2618 if (PredHasWeights && SuccHasWeights) {
2619 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2620 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2621 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2622 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2623 // The weight to CommonDest should be PredCommon * SuccTotal +
2624 // PredOther * SuccCommon.
2625 // The weight to OtherDest should be PredOther * SuccOther.
2626 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
2627 PredOther * SuccCommon,
2628 PredOther * SuccOther};
2629 // Halve the weights if any of them cannot fit in an uint32_t
2630 FitWeights(NewWeights);
2632 PBI->setMetadata(LLVMContext::MD_prof,
2633 MDBuilder(BI->getContext())
2634 .createBranchWeights(NewWeights[0], NewWeights[1]));
2637 // OtherDest may have phi nodes. If so, add an entry from PBI's
2638 // block that are identical to the entries for BI's block.
2639 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2641 // We know that the CommonDest already had an edge from PBI to
2642 // it. If it has PHIs though, the PHIs may have different
2643 // entries for BB and PBI's BB. If so, insert a select to make
2646 for (BasicBlock::iterator II = CommonDest->begin();
2647 (PN = dyn_cast<PHINode>(II)); ++II) {
2648 Value *BIV = PN->getIncomingValueForBlock(BB);
2649 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2650 Value *PBIV = PN->getIncomingValue(PBBIdx);
2652 // Insert a select in PBI to pick the right value.
2653 Value *NV = cast<SelectInst>
2654 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2655 PN->setIncomingValue(PBBIdx, NV);
2659 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2660 DEBUG(dbgs() << *PBI->getParent()->getParent());
2662 // This basic block is probably dead. We know it has at least
2663 // one fewer predecessor.
2667 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
2668 // true or to FalseBB if Cond is false.
2669 // Takes care of updating the successors and removing the old terminator.
2670 // Also makes sure not to introduce new successors by assuming that edges to
2671 // non-successor TrueBBs and FalseBBs aren't reachable.
2672 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2673 BasicBlock *TrueBB, BasicBlock *FalseBB,
2674 uint32_t TrueWeight,
2675 uint32_t FalseWeight){
2676 // Remove any superfluous successor edges from the CFG.
2677 // First, figure out which successors to preserve.
2678 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2680 BasicBlock *KeepEdge1 = TrueBB;
2681 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
2683 // Then remove the rest.
2684 for (BasicBlock *Succ : OldTerm->successors()) {
2685 // Make sure only to keep exactly one copy of each edge.
2686 if (Succ == KeepEdge1)
2687 KeepEdge1 = nullptr;
2688 else if (Succ == KeepEdge2)
2689 KeepEdge2 = nullptr;
2691 Succ->removePredecessor(OldTerm->getParent());
2694 IRBuilder<> Builder(OldTerm);
2695 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2697 // Insert an appropriate new terminator.
2698 if (!KeepEdge1 && !KeepEdge2) {
2699 if (TrueBB == FalseBB)
2700 // We were only looking for one successor, and it was present.
2701 // Create an unconditional branch to it.
2702 Builder.CreateBr(TrueBB);
2704 // We found both of the successors we were looking for.
2705 // Create a conditional branch sharing the condition of the select.
2706 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2707 if (TrueWeight != FalseWeight)
2708 NewBI->setMetadata(LLVMContext::MD_prof,
2709 MDBuilder(OldTerm->getContext()).
2710 createBranchWeights(TrueWeight, FalseWeight));
2712 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2713 // Neither of the selected blocks were successors, so this
2714 // terminator must be unreachable.
2715 new UnreachableInst(OldTerm->getContext(), OldTerm);
2717 // One of the selected values was a successor, but the other wasn't.
2718 // Insert an unconditional branch to the one that was found;
2719 // the edge to the one that wasn't must be unreachable.
2721 // Only TrueBB was found.
2722 Builder.CreateBr(TrueBB);
2724 // Only FalseBB was found.
2725 Builder.CreateBr(FalseBB);
2728 EraseTerminatorInstAndDCECond(OldTerm);
2733 // (switch (select cond, X, Y)) on constant X, Y
2734 // with a branch - conditional if X and Y lead to distinct BBs,
2735 // unconditional otherwise.
2736 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2737 // Check for constant integer values in the select.
2738 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2739 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2740 if (!TrueVal || !FalseVal)
2743 // Find the relevant condition and destinations.
2744 Value *Condition = Select->getCondition();
2745 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2746 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2748 // Get weight for TrueBB and FalseBB.
2749 uint32_t TrueWeight = 0, FalseWeight = 0;
2750 SmallVector<uint64_t, 8> Weights;
2751 bool HasWeights = HasBranchWeights(SI);
2753 GetBranchWeights(SI, Weights);
2754 if (Weights.size() == 1 + SI->getNumCases()) {
2755 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2756 getSuccessorIndex()];
2757 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2758 getSuccessorIndex()];
2762 // Perform the actual simplification.
2763 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2764 TrueWeight, FalseWeight);
2768 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2769 // blockaddress(@fn, BlockB)))
2771 // (br cond, BlockA, BlockB).
2772 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2773 // Check that both operands of the select are block addresses.
2774 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2775 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2779 // Extract the actual blocks.
2780 BasicBlock *TrueBB = TBA->getBasicBlock();
2781 BasicBlock *FalseBB = FBA->getBasicBlock();
2783 // Perform the actual simplification.
2784 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2788 /// This is called when we find an icmp instruction
2789 /// (a seteq/setne with a constant) as the only instruction in a
2790 /// block that ends with an uncond branch. We are looking for a very specific
2791 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2792 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2793 /// default value goes to an uncond block with a seteq in it, we get something
2796 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2798 /// %tmp = icmp eq i8 %A, 92
2801 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2803 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2804 /// the PHI, merging the third icmp into the switch.
2805 static bool TryToSimplifyUncondBranchWithICmpInIt(
2806 ICmpInst *ICI, IRBuilder<> &Builder, const DataLayout &DL,
2807 const TargetTransformInfo &TTI, unsigned BonusInstThreshold,
2808 AssumptionCache *AC) {
2809 BasicBlock *BB = ICI->getParent();
2811 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2813 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2815 Value *V = ICI->getOperand(0);
2816 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2818 // The pattern we're looking for is where our only predecessor is a switch on
2819 // 'V' and this block is the default case for the switch. In this case we can
2820 // fold the compared value into the switch to simplify things.
2821 BasicBlock *Pred = BB->getSinglePredecessor();
2822 if (!Pred || !isa<SwitchInst>(Pred->getTerminator())) return false;
2824 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2825 if (SI->getCondition() != V)
2828 // If BB is reachable on a non-default case, then we simply know the value of
2829 // V in this block. Substitute it and constant fold the icmp instruction
2831 if (SI->getDefaultDest() != BB) {
2832 ConstantInt *VVal = SI->findCaseDest(BB);
2833 assert(VVal && "Should have a unique destination value");
2834 ICI->setOperand(0, VVal);
2836 if (Value *V = SimplifyInstruction(ICI, DL)) {
2837 ICI->replaceAllUsesWith(V);
2838 ICI->eraseFromParent();
2840 // BB is now empty, so it is likely to simplify away.
2841 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
2844 // Ok, the block is reachable from the default dest. If the constant we're
2845 // comparing exists in one of the other edges, then we can constant fold ICI
2847 if (SI->findCaseValue(Cst) != SI->case_default()) {
2849 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2850 V = ConstantInt::getFalse(BB->getContext());
2852 V = ConstantInt::getTrue(BB->getContext());
2854 ICI->replaceAllUsesWith(V);
2855 ICI->eraseFromParent();
2856 // BB is now empty, so it is likely to simplify away.
2857 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
2860 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2862 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2863 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
2864 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
2865 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2868 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2870 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2871 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2873 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2874 std::swap(DefaultCst, NewCst);
2876 // Replace ICI (which is used by the PHI for the default value) with true or
2877 // false depending on if it is EQ or NE.
2878 ICI->replaceAllUsesWith(DefaultCst);
2879 ICI->eraseFromParent();
2881 // Okay, the switch goes to this block on a default value. Add an edge from
2882 // the switch to the merge point on the compared value.
2883 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2884 BB->getParent(), BB);
2885 SmallVector<uint64_t, 8> Weights;
2886 bool HasWeights = HasBranchWeights(SI);
2888 GetBranchWeights(SI, Weights);
2889 if (Weights.size() == 1 + SI->getNumCases()) {
2890 // Split weight for default case to case for "Cst".
2891 Weights[0] = (Weights[0]+1) >> 1;
2892 Weights.push_back(Weights[0]);
2894 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2895 SI->setMetadata(LLVMContext::MD_prof,
2896 MDBuilder(SI->getContext()).
2897 createBranchWeights(MDWeights));
2900 SI->addCase(Cst, NewBB);
2902 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2903 Builder.SetInsertPoint(NewBB);
2904 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2905 Builder.CreateBr(SuccBlock);
2906 PHIUse->addIncoming(NewCst, NewBB);
2910 /// The specified branch is a conditional branch.
2911 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2912 /// fold it into a switch instruction if so.
2913 static bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
2914 const DataLayout &DL) {
2915 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2916 if (!Cond) return false;
2918 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2919 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2920 // 'setne's and'ed together, collect them.
2922 // Try to gather values from a chain of and/or to be turned into a switch
2923 ConstantComparesGatherer ConstantCompare(Cond, DL);
2924 // Unpack the result
2925 SmallVectorImpl<ConstantInt*> &Values = ConstantCompare.Vals;
2926 Value *CompVal = ConstantCompare.CompValue;
2927 unsigned UsedICmps = ConstantCompare.UsedICmps;
2928 Value *ExtraCase = ConstantCompare.Extra;
2930 // If we didn't have a multiply compared value, fail.
2931 if (!CompVal) return false;
2933 // Avoid turning single icmps into a switch.
2937 bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
2939 // There might be duplicate constants in the list, which the switch
2940 // instruction can't handle, remove them now.
2941 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2942 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2944 // If Extra was used, we require at least two switch values to do the
2945 // transformation. A switch with one value is just a conditional branch.
2946 if (ExtraCase && Values.size() < 2) return false;
2948 // TODO: Preserve branch weight metadata, similarly to how
2949 // FoldValueComparisonIntoPredecessors preserves it.
2951 // Figure out which block is which destination.
2952 BasicBlock *DefaultBB = BI->getSuccessor(1);
2953 BasicBlock *EdgeBB = BI->getSuccessor(0);
2954 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2956 BasicBlock *BB = BI->getParent();
2958 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2959 << " cases into SWITCH. BB is:\n" << *BB);
2961 // If there are any extra values that couldn't be folded into the switch
2962 // then we evaluate them with an explicit branch first. Split the block
2963 // right before the condbr to handle it.
2966 BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
2967 // Remove the uncond branch added to the old block.
2968 TerminatorInst *OldTI = BB->getTerminator();
2969 Builder.SetInsertPoint(OldTI);
2972 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2974 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2976 OldTI->eraseFromParent();
2978 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2979 // for the edge we just added.
2980 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2982 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2983 << "\nEXTRABB = " << *BB);
2987 Builder.SetInsertPoint(BI);
2988 // Convert pointer to int before we switch.
2989 if (CompVal->getType()->isPointerTy()) {
2990 CompVal = Builder.CreatePtrToInt(
2991 CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
2994 // Create the new switch instruction now.
2995 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2997 // Add all of the 'cases' to the switch instruction.
2998 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2999 New->addCase(Values[i], EdgeBB);
3001 // We added edges from PI to the EdgeBB. As such, if there were any
3002 // PHI nodes in EdgeBB, they need entries to be added corresponding to
3003 // the number of edges added.
3004 for (BasicBlock::iterator BBI = EdgeBB->begin();
3005 isa<PHINode>(BBI); ++BBI) {
3006 PHINode *PN = cast<PHINode>(BBI);
3007 Value *InVal = PN->getIncomingValueForBlock(BB);
3008 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
3009 PN->addIncoming(InVal, BB);
3012 // Erase the old branch instruction.
3013 EraseTerminatorInstAndDCECond(BI);
3015 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3019 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3020 // If this is a trivial landing pad that just continues unwinding the caught
3021 // exception then zap the landing pad, turning its invokes into calls.
3022 BasicBlock *BB = RI->getParent();
3023 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
3024 if (RI->getValue() != LPInst)
3025 // Not a landing pad, or the resume is not unwinding the exception that
3026 // caused control to branch here.
3029 // Check that there are no other instructions except for debug intrinsics.
3030 BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3032 if (!isa<DbgInfoIntrinsic>(I))
3035 // Turn all invokes that unwind here into calls and delete the basic block.
3036 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3037 BasicBlock *Pred = *PI++;
3038 removeUnwindEdge(Pred);
3041 // The landingpad is now unreachable. Zap it.
3042 BB->eraseFromParent();
3046 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
3047 // If this is a trivial cleanup pad that executes no instructions, it can be
3048 // eliminated. If the cleanup pad continues to the caller, any predecessor
3049 // that is an EH pad will be updated to continue to the caller and any
3050 // predecessor that terminates with an invoke instruction will have its invoke
3051 // instruction converted to a call instruction. If the cleanup pad being
3052 // simplified does not continue to the caller, each predecessor will be
3053 // updated to continue to the unwind destination of the cleanup pad being
3055 BasicBlock *BB = RI->getParent();
3056 Instruction *CPInst = dyn_cast<CleanupPadInst>(BB->getFirstNonPHI());
3058 // This isn't an empty cleanup.
3061 // Check that there are no other instructions except for debug intrinsics.
3062 BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3064 if (!isa<DbgInfoIntrinsic>(I))
3067 // If the cleanup return we are simplifying unwinds to the caller, this
3068 // will set UnwindDest to nullptr.
3069 BasicBlock *UnwindDest = RI->getUnwindDest();
3071 // We're about to remove BB from the control flow. Before we do, sink any
3072 // PHINodes into the unwind destination. Doing this before changing the
3073 // control flow avoids some potentially slow checks, since we can currently
3074 // be certain that UnwindDest and BB have no common predecessors (since they
3075 // are both EH pads).
3077 // First, go through the PHI nodes in UnwindDest and update any nodes that
3078 // reference the block we are removing
3079 for (BasicBlock::iterator I = UnwindDest->begin(),
3080 IE = UnwindDest->getFirstNonPHI()->getIterator();
3082 PHINode *DestPN = cast<PHINode>(I);
3084 int Idx = DestPN->getBasicBlockIndex(BB);
3085 // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3087 // This PHI node has an incoming value that corresponds to a control
3088 // path through the cleanup pad we are removing. If the incoming
3089 // value is in the cleanup pad, it must be a PHINode (because we
3090 // verified above that the block is otherwise empty). Otherwise, the
3091 // value is either a constant or a value that dominates the cleanup
3092 // pad being removed.
3094 // Because BB and UnwindDest are both EH pads, all of their
3095 // predecessors must unwind to these blocks, and since no instruction
3096 // can have multiple unwind destinations, there will be no overlap in
3097 // incoming blocks between SrcPN and DestPN.
3098 Value *SrcVal = DestPN->getIncomingValue(Idx);
3099 PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3101 // Remove the entry for the block we are deleting.
3102 DestPN->removeIncomingValue(Idx, false);
3104 if (SrcPN && SrcPN->getParent() == BB) {
3105 // If the incoming value was a PHI node in the cleanup pad we are
3106 // removing, we need to merge that PHI node's incoming values into
3108 for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3109 SrcIdx != SrcE; ++SrcIdx) {
3110 DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3111 SrcPN->getIncomingBlock(SrcIdx));
3114 // Otherwise, the incoming value came from above BB and
3115 // so we can just reuse it. We must associate all of BB's
3116 // predecessors with this value.
3117 for (auto *pred : predecessors(BB)) {
3118 DestPN->addIncoming(SrcVal, pred);
3123 // Sink any remaining PHI nodes directly into UnwindDest.
3124 Instruction *InsertPt = UnwindDest->getFirstNonPHI();
3125 for (BasicBlock::iterator I = BB->begin(),
3126 IE = BB->getFirstNonPHI()->getIterator();
3128 // The iterator must be incremented here because the instructions are
3129 // being moved to another block.
3130 PHINode *PN = cast<PHINode>(I++);
3131 if (PN->use_empty())
3132 // If the PHI node has no uses, just leave it. It will be erased
3133 // when we erase BB below.
3136 // Otherwise, sink this PHI node into UnwindDest.
3137 // Any predecessors to UnwindDest which are not already represented
3138 // must be back edges which inherit the value from the path through
3139 // BB. In this case, the PHI value must reference itself.
3140 for (auto *pred : predecessors(UnwindDest))
3142 PN->addIncoming(PN, pred);
3143 PN->moveBefore(InsertPt);
3147 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3148 // The iterator must be updated here because we are removing this pred.
3149 BasicBlock *PredBB = *PI++;
3150 if (UnwindDest == nullptr) {
3151 removeUnwindEdge(PredBB);
3153 TerminatorInst *TI = PredBB->getTerminator();
3154 TI->replaceUsesOfWith(BB, UnwindDest);
3158 // The cleanup pad is now unreachable. Zap it.
3159 BB->eraseFromParent();
3163 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
3164 BasicBlock *BB = RI->getParent();
3165 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
3167 // Find predecessors that end with branches.
3168 SmallVector<BasicBlock*, 8> UncondBranchPreds;
3169 SmallVector<BranchInst*, 8> CondBranchPreds;
3170 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
3171 BasicBlock *P = *PI;
3172 TerminatorInst *PTI = P->getTerminator();
3173 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
3174 if (BI->isUnconditional())
3175 UncondBranchPreds.push_back(P);
3177 CondBranchPreds.push_back(BI);
3181 // If we found some, do the transformation!
3182 if (!UncondBranchPreds.empty() && DupRet) {
3183 while (!UncondBranchPreds.empty()) {
3184 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
3185 DEBUG(dbgs() << "FOLDING: " << *BB
3186 << "INTO UNCOND BRANCH PRED: " << *Pred);
3187 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
3190 // If we eliminated all predecessors of the block, delete the block now.
3192 // We know there are no successors, so just nuke the block.
3193 BB->eraseFromParent();
3198 // Check out all of the conditional branches going to this return
3199 // instruction. If any of them just select between returns, change the
3200 // branch itself into a select/return pair.
3201 while (!CondBranchPreds.empty()) {
3202 BranchInst *BI = CondBranchPreds.pop_back_val();
3204 // Check to see if the non-BB successor is also a return block.
3205 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
3206 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
3207 SimplifyCondBranchToTwoReturns(BI, Builder))
3213 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
3214 BasicBlock *BB = UI->getParent();
3216 bool Changed = false;
3218 // If there are any instructions immediately before the unreachable that can
3219 // be removed, do so.
3220 while (UI->getIterator() != BB->begin()) {
3221 BasicBlock::iterator BBI = UI->getIterator();
3223 // Do not delete instructions that can have side effects which might cause
3224 // the unreachable to not be reachable; specifically, calls and volatile
3225 // operations may have this effect.
3226 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
3228 if (BBI->mayHaveSideEffects()) {
3229 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
3230 if (SI->isVolatile())
3232 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
3233 if (LI->isVolatile())
3235 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
3236 if (RMWI->isVolatile())
3238 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
3239 if (CXI->isVolatile())
3241 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
3242 !isa<LandingPadInst>(BBI)) {
3245 // Note that deleting LandingPad's here is in fact okay, although it
3246 // involves a bit of subtle reasoning. If this inst is a LandingPad,
3247 // all the predecessors of this block will be the unwind edges of Invokes,
3248 // and we can therefore guarantee this block will be erased.
3251 // Delete this instruction (any uses are guaranteed to be dead)
3252 if (!BBI->use_empty())
3253 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
3254 BBI->eraseFromParent();
3258 // If the unreachable instruction is the first in the block, take a gander
3259 // at all of the predecessors of this instruction, and simplify them.
3260 if (&BB->front() != UI) return Changed;
3262 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
3263 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
3264 TerminatorInst *TI = Preds[i]->getTerminator();
3265 IRBuilder<> Builder(TI);
3266 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3267 if (BI->isUnconditional()) {
3268 if (BI->getSuccessor(0) == BB) {
3269 new UnreachableInst(TI->getContext(), TI);
3270 TI->eraseFromParent();
3274 if (BI->getSuccessor(0) == BB) {
3275 Builder.CreateBr(BI->getSuccessor(1));
3276 EraseTerminatorInstAndDCECond(BI);
3277 } else if (BI->getSuccessor(1) == BB) {
3278 Builder.CreateBr(BI->getSuccessor(0));
3279 EraseTerminatorInstAndDCECond(BI);
3283 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3284 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3286 if (i.getCaseSuccessor() == BB) {
3287 BB->removePredecessor(SI->getParent());
3292 } else if ((isa<InvokeInst>(TI) &&
3293 cast<InvokeInst>(TI)->getUnwindDest() == BB) ||
3294 isa<CatchEndPadInst>(TI) || isa<TerminatePadInst>(TI)) {
3295 removeUnwindEdge(TI->getParent());
3297 } else if (isa<CleanupReturnInst>(TI) || isa<CleanupEndPadInst>(TI) ||
3298 isa<CatchReturnInst>(TI)) {
3299 new UnreachableInst(TI->getContext(), TI);
3300 TI->eraseFromParent();
3303 // TODO: If TI is a CatchPadInst, then (BB must be its normal dest and)
3304 // we can eliminate it, redirecting its preds to its unwind successor,
3305 // or to the next outer handler if the removed catch is the last for its
3309 // If this block is now dead, remove it.
3310 if (pred_empty(BB) &&
3311 BB != &BB->getParent()->getEntryBlock()) {
3312 // We know there are no successors, so just nuke the block.
3313 BB->eraseFromParent();
3320 static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) {
3321 assert(Cases.size() >= 1);
3323 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3324 for (size_t I = 1, E = Cases.size(); I != E; ++I) {
3325 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
3331 /// Turn a switch with two reachable destinations into an integer range
3332 /// comparison and branch.
3333 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3334 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3337 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
3339 // Partition the cases into two sets with different destinations.
3340 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
3341 BasicBlock *DestB = nullptr;
3342 SmallVector <ConstantInt *, 16> CasesA;
3343 SmallVector <ConstantInt *, 16> CasesB;
3345 for (SwitchInst::CaseIt I : SI->cases()) {
3346 BasicBlock *Dest = I.getCaseSuccessor();
3347 if (!DestA) DestA = Dest;
3348 if (Dest == DestA) {
3349 CasesA.push_back(I.getCaseValue());
3352 if (!DestB) DestB = Dest;
3353 if (Dest == DestB) {
3354 CasesB.push_back(I.getCaseValue());
3357 return false; // More than two destinations.
3360 assert(DestA && DestB && "Single-destination switch should have been folded.");
3361 assert(DestA != DestB);
3362 assert(DestB != SI->getDefaultDest());
3363 assert(!CasesB.empty() && "There must be non-default cases.");
3364 assert(!CasesA.empty() || HasDefault);
3366 // Figure out if one of the sets of cases form a contiguous range.
3367 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
3368 BasicBlock *ContiguousDest = nullptr;
3369 BasicBlock *OtherDest = nullptr;
3370 if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
3371 ContiguousCases = &CasesA;
3372 ContiguousDest = DestA;
3374 } else if (CasesAreContiguous(CasesB)) {
3375 ContiguousCases = &CasesB;
3376 ContiguousDest = DestB;
3381 // Start building the compare and branch.
3383 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
3384 Constant *NumCases = ConstantInt::get(Offset->getType(), ContiguousCases->size());
3386 Value *Sub = SI->getCondition();
3387 if (!Offset->isNullValue())
3388 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
3391 // If NumCases overflowed, then all possible values jump to the successor.
3392 if (NumCases->isNullValue() && !ContiguousCases->empty())
3393 Cmp = ConstantInt::getTrue(SI->getContext());
3395 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3396 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
3398 // Update weight for the newly-created conditional branch.
3399 if (HasBranchWeights(SI)) {
3400 SmallVector<uint64_t, 8> Weights;
3401 GetBranchWeights(SI, Weights);
3402 if (Weights.size() == 1 + SI->getNumCases()) {
3403 uint64_t TrueWeight = 0;
3404 uint64_t FalseWeight = 0;
3405 for (size_t I = 0, E = Weights.size(); I != E; ++I) {
3406 if (SI->getSuccessor(I) == ContiguousDest)
3407 TrueWeight += Weights[I];
3409 FalseWeight += Weights[I];
3411 while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
3415 NewBI->setMetadata(LLVMContext::MD_prof,
3416 MDBuilder(SI->getContext()).createBranchWeights(
3417 (uint32_t)TrueWeight, (uint32_t)FalseWeight));
3421 // Prune obsolete incoming values off the successors' PHI nodes.
3422 for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
3423 unsigned PreviousEdges = ContiguousCases->size();
3424 if (ContiguousDest == SI->getDefaultDest()) ++PreviousEdges;
3425 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
3426 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3428 for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
3429 unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
3430 if (OtherDest == SI->getDefaultDest()) ++PreviousEdges;
3431 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
3432 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3436 SI->eraseFromParent();
3441 /// Compute masked bits for the condition of a switch
3442 /// and use it to remove dead cases.
3443 static bool EliminateDeadSwitchCases(SwitchInst *SI, AssumptionCache *AC,
3444 const DataLayout &DL) {
3445 Value *Cond = SI->getCondition();
3446 unsigned Bits = Cond->getType()->getIntegerBitWidth();
3447 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3448 computeKnownBits(Cond, KnownZero, KnownOne, DL, 0, AC, SI);
3450 // Gather dead cases.
3451 SmallVector<ConstantInt*, 8> DeadCases;
3452 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3453 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3454 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3455 DeadCases.push_back(I.getCaseValue());
3456 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3457 << I.getCaseValue() << "' is dead.\n");
3461 // If we can prove that the cases must cover all possible values, the
3462 // default destination becomes dead and we can remove it. If we know some
3463 // of the bits in the value, we can use that to more precisely compute the
3464 // number of possible unique case values.
3466 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
3467 const unsigned NumUnknownBits = Bits -
3468 (KnownZero.Or(KnownOne)).countPopulation();
3469 assert(NumUnknownBits <= Bits);
3470 if (HasDefault && DeadCases.empty() &&
3471 NumUnknownBits < 64 /* avoid overflow */ &&
3472 SI->getNumCases() == (1ULL << NumUnknownBits)) {
3473 DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
3474 BasicBlock *NewDefault = SplitBlockPredecessors(SI->getDefaultDest(),
3475 SI->getParent(), "");
3476 SI->setDefaultDest(&*NewDefault);
3477 SplitBlock(&*NewDefault, &NewDefault->front());
3478 auto *OldTI = NewDefault->getTerminator();
3479 new UnreachableInst(SI->getContext(), OldTI);
3480 EraseTerminatorInstAndDCECond(OldTI);
3484 SmallVector<uint64_t, 8> Weights;
3485 bool HasWeight = HasBranchWeights(SI);
3487 GetBranchWeights(SI, Weights);
3488 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3491 // Remove dead cases from the switch.
3492 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3493 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3494 assert(Case != SI->case_default() &&
3495 "Case was not found. Probably mistake in DeadCases forming.");
3497 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3501 // Prune unused values from PHI nodes.
3502 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3503 SI->removeCase(Case);
3505 if (HasWeight && Weights.size() >= 2) {
3506 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3507 SI->setMetadata(LLVMContext::MD_prof,
3508 MDBuilder(SI->getParent()->getContext()).
3509 createBranchWeights(MDWeights));
3512 return !DeadCases.empty();
3515 /// If BB would be eligible for simplification by
3516 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3517 /// by an unconditional branch), look at the phi node for BB in the successor
3518 /// block and see if the incoming value is equal to CaseValue. If so, return
3519 /// the phi node, and set PhiIndex to BB's index in the phi node.
3520 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3523 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3524 return nullptr; // BB must be empty to be a candidate for simplification.
3525 if (!BB->getSinglePredecessor())
3526 return nullptr; // BB must be dominated by the switch.
3528 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3529 if (!Branch || !Branch->isUnconditional())
3530 return nullptr; // Terminator must be unconditional branch.
3532 BasicBlock *Succ = Branch->getSuccessor(0);
3534 BasicBlock::iterator I = Succ->begin();
3535 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3536 int Idx = PHI->getBasicBlockIndex(BB);
3537 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3539 Value *InValue = PHI->getIncomingValue(Idx);
3540 if (InValue != CaseValue) continue;
3549 /// Try to forward the condition of a switch instruction to a phi node
3550 /// dominated by the switch, if that would mean that some of the destination
3551 /// blocks of the switch can be folded away.
3552 /// Returns true if a change is made.
3553 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3554 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3555 ForwardingNodesMap ForwardingNodes;
3557 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3558 ConstantInt *CaseValue = I.getCaseValue();
3559 BasicBlock *CaseDest = I.getCaseSuccessor();
3562 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3566 ForwardingNodes[PHI].push_back(PhiIndex);
3569 bool Changed = false;
3571 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3572 E = ForwardingNodes.end(); I != E; ++I) {
3573 PHINode *Phi = I->first;
3574 SmallVectorImpl<int> &Indexes = I->second;
3576 if (Indexes.size() < 2) continue;
3578 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3579 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3586 /// Return true if the backend will be able to handle
3587 /// initializing an array of constants like C.
3588 static bool ValidLookupTableConstant(Constant *C) {
3589 if (C->isThreadDependent())
3591 if (C->isDLLImportDependent())
3594 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3595 return CE->isGEPWithNoNotionalOverIndexing();
3597 return isa<ConstantFP>(C) ||
3598 isa<ConstantInt>(C) ||
3599 isa<ConstantPointerNull>(C) ||
3600 isa<GlobalValue>(C) ||
3604 /// If V is a Constant, return it. Otherwise, try to look up
3605 /// its constant value in ConstantPool, returning 0 if it's not there.
3606 static Constant *LookupConstant(Value *V,
3607 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3608 if (Constant *C = dyn_cast<Constant>(V))
3610 return ConstantPool.lookup(V);
3613 /// Try to fold instruction I into a constant. This works for
3614 /// simple instructions such as binary operations where both operands are
3615 /// constant or can be replaced by constants from the ConstantPool. Returns the
3616 /// resulting constant on success, 0 otherwise.
3618 ConstantFold(Instruction *I, const DataLayout &DL,
3619 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
3620 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3621 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3624 if (A->isAllOnesValue())
3625 return LookupConstant(Select->getTrueValue(), ConstantPool);
3626 if (A->isNullValue())
3627 return LookupConstant(Select->getFalseValue(), ConstantPool);
3631 SmallVector<Constant *, 4> COps;
3632 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
3633 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
3639 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
3640 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
3644 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), COps, DL);
3647 /// Try to determine the resulting constant values in phi nodes
3648 /// at the common destination basic block, *CommonDest, for one of the case
3649 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3650 /// case), of a switch instruction SI.
3652 GetCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest,
3653 BasicBlock **CommonDest,
3654 SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
3655 const DataLayout &DL) {
3656 // The block from which we enter the common destination.
3657 BasicBlock *Pred = SI->getParent();
3659 // If CaseDest is empty except for some side-effect free instructions through
3660 // which we can constant-propagate the CaseVal, continue to its successor.
3661 SmallDenseMap<Value*, Constant*> ConstantPool;
3662 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3663 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3665 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3666 // If the terminator is a simple branch, continue to the next block.
3667 if (T->getNumSuccessors() != 1)
3670 CaseDest = T->getSuccessor(0);
3671 } else if (isa<DbgInfoIntrinsic>(I)) {
3672 // Skip debug intrinsic.
3674 } else if (Constant *C = ConstantFold(&*I, DL, ConstantPool)) {
3675 // Instruction is side-effect free and constant.
3677 // If the instruction has uses outside this block or a phi node slot for
3678 // the block, it is not safe to bypass the instruction since it would then
3679 // no longer dominate all its uses.
3680 for (auto &Use : I->uses()) {
3681 User *User = Use.getUser();
3682 if (Instruction *I = dyn_cast<Instruction>(User))
3683 if (I->getParent() == CaseDest)
3685 if (PHINode *Phi = dyn_cast<PHINode>(User))
3686 if (Phi->getIncomingBlock(Use) == CaseDest)
3691 ConstantPool.insert(std::make_pair(&*I, C));
3697 // If we did not have a CommonDest before, use the current one.
3699 *CommonDest = CaseDest;
3700 // If the destination isn't the common one, abort.
3701 if (CaseDest != *CommonDest)
3704 // Get the values for this case from phi nodes in the destination block.
3705 BasicBlock::iterator I = (*CommonDest)->begin();
3706 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3707 int Idx = PHI->getBasicBlockIndex(Pred);
3711 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3716 // Be conservative about which kinds of constants we support.
3717 if (!ValidLookupTableConstant(ConstVal))
3720 Res.push_back(std::make_pair(PHI, ConstVal));
3723 return Res.size() > 0;
3726 // Helper function used to add CaseVal to the list of cases that generate
3728 static void MapCaseToResult(ConstantInt *CaseVal,
3729 SwitchCaseResultVectorTy &UniqueResults,
3731 for (auto &I : UniqueResults) {
3732 if (I.first == Result) {
3733 I.second.push_back(CaseVal);
3737 UniqueResults.push_back(std::make_pair(Result,
3738 SmallVector<ConstantInt*, 4>(1, CaseVal)));
3741 // Helper function that initializes a map containing
3742 // results for the PHI node of the common destination block for a switch
3743 // instruction. Returns false if multiple PHI nodes have been found or if
3744 // there is not a common destination block for the switch.
3745 static bool InitializeUniqueCases(SwitchInst *SI, PHINode *&PHI,
3746 BasicBlock *&CommonDest,
3747 SwitchCaseResultVectorTy &UniqueResults,
3748 Constant *&DefaultResult,
3749 const DataLayout &DL) {
3750 for (auto &I : SI->cases()) {
3751 ConstantInt *CaseVal = I.getCaseValue();
3753 // Resulting value at phi nodes for this case value.
3754 SwitchCaseResultsTy Results;
3755 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
3759 // Only one value per case is permitted
3760 if (Results.size() > 1)
3762 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
3764 // Check the PHI consistency.
3766 PHI = Results[0].first;
3767 else if (PHI != Results[0].first)
3770 // Find the default result value.
3771 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
3772 BasicBlock *DefaultDest = SI->getDefaultDest();
3773 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
3775 // If the default value is not found abort unless the default destination
3778 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
3779 if ((!DefaultResult &&
3780 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
3786 // Helper function that checks if it is possible to transform a switch with only
3787 // two cases (or two cases + default) that produces a result into a select.
3790 // case 10: %0 = icmp eq i32 %a, 10
3791 // return 10; %1 = select i1 %0, i32 10, i32 4
3792 // case 20: ----> %2 = icmp eq i32 %a, 20
3793 // return 2; %3 = select i1 %2, i32 2, i32 %1
3798 ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
3799 Constant *DefaultResult, Value *Condition,
3800 IRBuilder<> &Builder) {
3801 assert(ResultVector.size() == 2 &&
3802 "We should have exactly two unique results at this point");
3803 // If we are selecting between only two cases transform into a simple
3804 // select or a two-way select if default is possible.
3805 if (ResultVector[0].second.size() == 1 &&
3806 ResultVector[1].second.size() == 1) {
3807 ConstantInt *const FirstCase = ResultVector[0].second[0];
3808 ConstantInt *const SecondCase = ResultVector[1].second[0];
3810 bool DefaultCanTrigger = DefaultResult;
3811 Value *SelectValue = ResultVector[1].first;
3812 if (DefaultCanTrigger) {
3813 Value *const ValueCompare =
3814 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
3815 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
3816 DefaultResult, "switch.select");
3818 Value *const ValueCompare =
3819 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
3820 return Builder.CreateSelect(ValueCompare, ResultVector[0].first, SelectValue,
3827 // Helper function to cleanup a switch instruction that has been converted into
3828 // a select, fixing up PHI nodes and basic blocks.
3829 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
3831 IRBuilder<> &Builder) {
3832 BasicBlock *SelectBB = SI->getParent();
3833 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
3834 PHI->removeIncomingValue(SelectBB);
3835 PHI->addIncoming(SelectValue, SelectBB);
3837 Builder.CreateBr(PHI->getParent());
3839 // Remove the switch.
3840 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
3841 BasicBlock *Succ = SI->getSuccessor(i);
3843 if (Succ == PHI->getParent())
3845 Succ->removePredecessor(SelectBB);
3847 SI->eraseFromParent();
3850 /// If the switch is only used to initialize one or more
3851 /// phi nodes in a common successor block with only two different
3852 /// constant values, replace the switch with select.
3853 static bool SwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
3854 AssumptionCache *AC, const DataLayout &DL) {
3855 Value *const Cond = SI->getCondition();
3856 PHINode *PHI = nullptr;
3857 BasicBlock *CommonDest = nullptr;
3858 Constant *DefaultResult;
3859 SwitchCaseResultVectorTy UniqueResults;
3860 // Collect all the cases that will deliver the same value from the switch.
3861 if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
3864 // Selects choose between maximum two values.
3865 if (UniqueResults.size() != 2)
3867 assert(PHI != nullptr && "PHI for value select not found");
3869 Builder.SetInsertPoint(SI);
3870 Value *SelectValue = ConvertTwoCaseSwitch(
3872 DefaultResult, Cond, Builder);
3874 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
3877 // The switch couldn't be converted into a select.
3882 /// This class represents a lookup table that can be used to replace a switch.
3883 class SwitchLookupTable {
3885 /// Create a lookup table to use as a switch replacement with the contents
3886 /// of Values, using DefaultValue to fill any holes in the table.
3888 Module &M, uint64_t TableSize, ConstantInt *Offset,
3889 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
3890 Constant *DefaultValue, const DataLayout &DL);
3892 /// Build instructions with Builder to retrieve the value at
3893 /// the position given by Index in the lookup table.
3894 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3896 /// Return true if a table with TableSize elements of
3897 /// type ElementType would fit in a target-legal register.
3898 static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
3902 // Depending on the contents of the table, it can be represented in
3905 // For tables where each element contains the same value, we just have to
3906 // store that single value and return it for each lookup.
3909 // For tables where there is a linear relationship between table index
3910 // and values. We calculate the result with a simple multiplication
3911 // and addition instead of a table lookup.
3914 // For small tables with integer elements, we can pack them into a bitmap
3915 // that fits into a target-legal register. Values are retrieved by
3916 // shift and mask operations.
3919 // The table is stored as an array of values. Values are retrieved by load
3920 // instructions from the table.
3924 // For SingleValueKind, this is the single value.
3925 Constant *SingleValue;
3927 // For BitMapKind, this is the bitmap.
3928 ConstantInt *BitMap;
3929 IntegerType *BitMapElementTy;
3931 // For LinearMapKind, these are the constants used to derive the value.
3932 ConstantInt *LinearOffset;
3933 ConstantInt *LinearMultiplier;
3935 // For ArrayKind, this is the array.
3936 GlobalVariable *Array;
3940 SwitchLookupTable::SwitchLookupTable(
3941 Module &M, uint64_t TableSize, ConstantInt *Offset,
3942 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
3943 Constant *DefaultValue, const DataLayout &DL)
3944 : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
3945 LinearOffset(nullptr), LinearMultiplier(nullptr), Array(nullptr) {
3946 assert(Values.size() && "Can't build lookup table without values!");
3947 assert(TableSize >= Values.size() && "Can't fit values in table!");
3949 // If all values in the table are equal, this is that value.
3950 SingleValue = Values.begin()->second;
3952 Type *ValueType = Values.begin()->second->getType();
3954 // Build up the table contents.
3955 SmallVector<Constant*, 64> TableContents(TableSize);
3956 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3957 ConstantInt *CaseVal = Values[I].first;
3958 Constant *CaseRes = Values[I].second;
3959 assert(CaseRes->getType() == ValueType);
3961 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3963 TableContents[Idx] = CaseRes;
3965 if (CaseRes != SingleValue)
3966 SingleValue = nullptr;
3969 // Fill in any holes in the table with the default result.
3970 if (Values.size() < TableSize) {
3971 assert(DefaultValue &&
3972 "Need a default value to fill the lookup table holes.");
3973 assert(DefaultValue->getType() == ValueType);
3974 for (uint64_t I = 0; I < TableSize; ++I) {
3975 if (!TableContents[I])
3976 TableContents[I] = DefaultValue;
3979 if (DefaultValue != SingleValue)
3980 SingleValue = nullptr;
3983 // If each element in the table contains the same value, we only need to store
3984 // that single value.
3986 Kind = SingleValueKind;
3990 // Check if we can derive the value with a linear transformation from the
3992 if (isa<IntegerType>(ValueType)) {
3993 bool LinearMappingPossible = true;
3996 assert(TableSize >= 2 && "Should be a SingleValue table.");
3997 // Check if there is the same distance between two consecutive values.
3998 for (uint64_t I = 0; I < TableSize; ++I) {
3999 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4001 // This is an undef. We could deal with it, but undefs in lookup tables
4002 // are very seldom. It's probably not worth the additional complexity.
4003 LinearMappingPossible = false;
4006 APInt Val = ConstVal->getValue();
4008 APInt Dist = Val - PrevVal;
4011 } else if (Dist != DistToPrev) {
4012 LinearMappingPossible = false;
4018 if (LinearMappingPossible) {
4019 LinearOffset = cast<ConstantInt>(TableContents[0]);
4020 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
4021 Kind = LinearMapKind;
4027 // If the type is integer and the table fits in a register, build a bitmap.
4028 if (WouldFitInRegister(DL, TableSize, ValueType)) {
4029 IntegerType *IT = cast<IntegerType>(ValueType);
4030 APInt TableInt(TableSize * IT->getBitWidth(), 0);
4031 for (uint64_t I = TableSize; I > 0; --I) {
4032 TableInt <<= IT->getBitWidth();
4033 // Insert values into the bitmap. Undef values are set to zero.
4034 if (!isa<UndefValue>(TableContents[I - 1])) {
4035 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
4036 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
4039 BitMap = ConstantInt::get(M.getContext(), TableInt);
4040 BitMapElementTy = IT;
4046 // Store the table in an array.
4047 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
4048 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
4050 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
4051 GlobalVariable::PrivateLinkage,
4054 Array->setUnnamedAddr(true);
4058 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
4060 case SingleValueKind:
4062 case LinearMapKind: {
4063 // Derive the result value from the input value.
4064 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
4065 false, "switch.idx.cast");
4066 if (!LinearMultiplier->isOne())
4067 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
4068 if (!LinearOffset->isZero())
4069 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
4073 // Type of the bitmap (e.g. i59).
4074 IntegerType *MapTy = BitMap->getType();
4076 // Cast Index to the same type as the bitmap.
4077 // Note: The Index is <= the number of elements in the table, so
4078 // truncating it to the width of the bitmask is safe.
4079 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
4081 // Multiply the shift amount by the element width.
4082 ShiftAmt = Builder.CreateMul(ShiftAmt,
4083 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
4087 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
4088 "switch.downshift");
4090 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
4094 // Make sure the table index will not overflow when treated as signed.
4095 IntegerType *IT = cast<IntegerType>(Index->getType());
4096 uint64_t TableSize = Array->getInitializer()->getType()
4097 ->getArrayNumElements();
4098 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
4099 Index = Builder.CreateZExt(Index,
4100 IntegerType::get(IT->getContext(),
4101 IT->getBitWidth() + 1),
4102 "switch.tableidx.zext");
4104 Value *GEPIndices[] = { Builder.getInt32(0), Index };
4105 Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
4106 GEPIndices, "switch.gep");
4107 return Builder.CreateLoad(GEP, "switch.load");
4110 llvm_unreachable("Unknown lookup table kind!");
4113 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
4115 Type *ElementType) {
4116 auto *IT = dyn_cast<IntegerType>(ElementType);
4119 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
4120 // are <= 15, we could try to narrow the type.
4122 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
4123 if (TableSize >= UINT_MAX/IT->getBitWidth())
4125 return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
4128 /// Determine whether a lookup table should be built for this switch, based on
4129 /// the number of cases, size of the table, and the types of the results.
4131 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
4132 const TargetTransformInfo &TTI, const DataLayout &DL,
4133 const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
4134 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
4135 return false; // TableSize overflowed, or mul below might overflow.
4137 bool AllTablesFitInRegister = true;
4138 bool HasIllegalType = false;
4139 for (const auto &I : ResultTypes) {
4140 Type *Ty = I.second;
4142 // Saturate this flag to true.
4143 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
4145 // Saturate this flag to false.
4146 AllTablesFitInRegister = AllTablesFitInRegister &&
4147 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
4149 // If both flags saturate, we're done. NOTE: This *only* works with
4150 // saturating flags, and all flags have to saturate first due to the
4151 // non-deterministic behavior of iterating over a dense map.
4152 if (HasIllegalType && !AllTablesFitInRegister)
4156 // If each table would fit in a register, we should build it anyway.
4157 if (AllTablesFitInRegister)
4160 // Don't build a table that doesn't fit in-register if it has illegal types.
4164 // The table density should be at least 40%. This is the same criterion as for
4165 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
4166 // FIXME: Find the best cut-off.
4167 return SI->getNumCases() * 10 >= TableSize * 4;
4170 /// Try to reuse the switch table index compare. Following pattern:
4172 /// if (idx < tablesize)
4173 /// r = table[idx]; // table does not contain default_value
4175 /// r = default_value;
4176 /// if (r != default_value)
4179 /// Is optimized to:
4181 /// cond = idx < tablesize;
4185 /// r = default_value;
4189 /// Jump threading will then eliminate the second if(cond).
4190 static void reuseTableCompare(User *PhiUser, BasicBlock *PhiBlock,
4191 BranchInst *RangeCheckBranch, Constant *DefaultValue,
4192 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values) {
4194 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
4198 // We require that the compare is in the same block as the phi so that jump
4199 // threading can do its work afterwards.
4200 if (CmpInst->getParent() != PhiBlock)
4203 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
4207 Value *RangeCmp = RangeCheckBranch->getCondition();
4208 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
4209 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
4211 // Check if the compare with the default value is constant true or false.
4212 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
4213 DefaultValue, CmpOp1, true);
4214 if (DefaultConst != TrueConst && DefaultConst != FalseConst)
4217 // Check if the compare with the case values is distinct from the default
4219 for (auto ValuePair : Values) {
4220 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
4221 ValuePair.second, CmpOp1, true);
4222 if (!CaseConst || CaseConst == DefaultConst)
4224 assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
4225 "Expect true or false as compare result.");
4228 // Check if the branch instruction dominates the phi node. It's a simple
4229 // dominance check, but sufficient for our needs.
4230 // Although this check is invariant in the calling loops, it's better to do it
4231 // at this late stage. Practically we do it at most once for a switch.
4232 BasicBlock *BranchBlock = RangeCheckBranch->getParent();
4233 for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
4234 BasicBlock *Pred = *PI;
4235 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
4239 if (DefaultConst == FalseConst) {
4240 // The compare yields the same result. We can replace it.
4241 CmpInst->replaceAllUsesWith(RangeCmp);
4242 ++NumTableCmpReuses;
4244 // The compare yields the same result, just inverted. We can replace it.
4245 Value *InvertedTableCmp = BinaryOperator::CreateXor(RangeCmp,
4246 ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
4248 CmpInst->replaceAllUsesWith(InvertedTableCmp);
4249 ++NumTableCmpReuses;
4253 /// If the switch is only used to initialize one or more phi nodes in a common
4254 /// successor block with different constant values, replace the switch with
4256 static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
4257 const DataLayout &DL,
4258 const TargetTransformInfo &TTI) {
4259 assert(SI->getNumCases() > 1 && "Degenerate switch?");
4261 // Only build lookup table when we have a target that supports it.
4262 if (!TTI.shouldBuildLookupTables())
4265 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
4266 // split off a dense part and build a lookup table for that.
4268 // FIXME: This creates arrays of GEPs to constant strings, which means each
4269 // GEP needs a runtime relocation in PIC code. We should just build one big
4270 // string and lookup indices into that.
4272 // Ignore switches with less than three cases. Lookup tables will not make them
4273 // faster, so we don't analyze them.
4274 if (SI->getNumCases() < 3)
4277 // Figure out the corresponding result for each case value and phi node in the
4278 // common destination, as well as the min and max case values.
4279 assert(SI->case_begin() != SI->case_end());
4280 SwitchInst::CaseIt CI = SI->case_begin();
4281 ConstantInt *MinCaseVal = CI.getCaseValue();
4282 ConstantInt *MaxCaseVal = CI.getCaseValue();
4284 BasicBlock *CommonDest = nullptr;
4285 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
4286 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
4287 SmallDenseMap<PHINode*, Constant*> DefaultResults;
4288 SmallDenseMap<PHINode*, Type*> ResultTypes;
4289 SmallVector<PHINode*, 4> PHIs;
4291 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
4292 ConstantInt *CaseVal = CI.getCaseValue();
4293 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
4294 MinCaseVal = CaseVal;
4295 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
4296 MaxCaseVal = CaseVal;
4298 // Resulting value at phi nodes for this case value.
4299 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
4301 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
4305 // Append the result from this case to the list for each phi.
4306 for (const auto &I : Results) {
4307 PHINode *PHI = I.first;
4308 Constant *Value = I.second;
4309 if (!ResultLists.count(PHI))
4310 PHIs.push_back(PHI);
4311 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
4315 // Keep track of the result types.
4316 for (PHINode *PHI : PHIs) {
4317 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
4320 uint64_t NumResults = ResultLists[PHIs[0]].size();
4321 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
4322 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
4323 bool TableHasHoles = (NumResults < TableSize);
4325 // If the table has holes, we need a constant result for the default case
4326 // or a bitmask that fits in a register.
4327 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
4328 bool HasDefaultResults = GetCaseResults(SI, nullptr, SI->getDefaultDest(),
4329 &CommonDest, DefaultResultsList, DL);
4331 bool NeedMask = (TableHasHoles && !HasDefaultResults);
4333 // As an extra penalty for the validity test we require more cases.
4334 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
4336 if (!DL.fitsInLegalInteger(TableSize))
4340 for (const auto &I : DefaultResultsList) {
4341 PHINode *PHI = I.first;
4342 Constant *Result = I.second;
4343 DefaultResults[PHI] = Result;
4346 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
4349 // Create the BB that does the lookups.
4350 Module &Mod = *CommonDest->getParent()->getParent();
4351 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
4353 CommonDest->getParent(),
4356 // Compute the table index value.
4357 Builder.SetInsertPoint(SI);
4358 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
4361 // Compute the maximum table size representable by the integer type we are
4363 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
4364 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
4365 assert(MaxTableSize >= TableSize &&
4366 "It is impossible for a switch to have more entries than the max "
4367 "representable value of its input integer type's size.");
4369 // If the default destination is unreachable, or if the lookup table covers
4370 // all values of the conditional variable, branch directly to the lookup table
4371 // BB. Otherwise, check that the condition is within the case range.
4372 const bool DefaultIsReachable =
4373 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4374 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
4375 BranchInst *RangeCheckBranch = nullptr;
4377 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
4378 Builder.CreateBr(LookupBB);
4379 // Note: We call removeProdecessor later since we need to be able to get the
4380 // PHI value for the default case in case we're using a bit mask.
4382 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
4383 MinCaseVal->getType(), TableSize));
4384 RangeCheckBranch = Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
4387 // Populate the BB that does the lookups.
4388 Builder.SetInsertPoint(LookupBB);
4391 // Before doing the lookup we do the hole check.
4392 // The LookupBB is therefore re-purposed to do the hole check
4393 // and we create a new LookupBB.
4394 BasicBlock *MaskBB = LookupBB;
4395 MaskBB->setName("switch.hole_check");
4396 LookupBB = BasicBlock::Create(Mod.getContext(),
4398 CommonDest->getParent(),
4401 // Make the mask's bitwidth at least 8bit and a power-of-2 to avoid
4402 // unnecessary illegal types.
4403 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
4404 APInt MaskInt(TableSizePowOf2, 0);
4405 APInt One(TableSizePowOf2, 1);
4406 // Build bitmask; fill in a 1 bit for every case.
4407 const ResultListTy &ResultList = ResultLists[PHIs[0]];
4408 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
4409 uint64_t Idx = (ResultList[I].first->getValue() -
4410 MinCaseVal->getValue()).getLimitedValue();
4411 MaskInt |= One << Idx;
4413 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
4415 // Get the TableIndex'th bit of the bitmask.
4416 // If this bit is 0 (meaning hole) jump to the default destination,
4417 // else continue with table lookup.
4418 IntegerType *MapTy = TableMask->getType();
4419 Value *MaskIndex = Builder.CreateZExtOrTrunc(TableIndex, MapTy,
4420 "switch.maskindex");
4421 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex,
4423 Value *LoBit = Builder.CreateTrunc(Shifted,
4424 Type::getInt1Ty(Mod.getContext()),
4426 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
4428 Builder.SetInsertPoint(LookupBB);
4429 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
4432 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
4433 // We cached PHINodes in PHIs, to avoid accessing deleted PHINodes later,
4434 // do not delete PHINodes here.
4435 SI->getDefaultDest()->removePredecessor(SI->getParent(),
4436 /*DontDeleteUselessPHIs=*/true);
4439 bool ReturnedEarly = false;
4440 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
4441 PHINode *PHI = PHIs[I];
4442 const ResultListTy &ResultList = ResultLists[PHI];
4444 // If using a bitmask, use any value to fill the lookup table holes.
4445 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
4446 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL);
4448 Value *Result = Table.BuildLookup(TableIndex, Builder);
4450 // If the result is used to return immediately from the function, we want to
4451 // do that right here.
4452 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
4453 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
4454 Builder.CreateRet(Result);
4455 ReturnedEarly = true;
4459 // Do a small peephole optimization: re-use the switch table compare if
4461 if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
4462 BasicBlock *PhiBlock = PHI->getParent();
4463 // Search for compare instructions which use the phi.
4464 for (auto *User : PHI->users()) {
4465 reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
4469 PHI->addIncoming(Result, LookupBB);
4473 Builder.CreateBr(CommonDest);
4475 // Remove the switch.
4476 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4477 BasicBlock *Succ = SI->getSuccessor(i);
4479 if (Succ == SI->getDefaultDest())
4481 Succ->removePredecessor(SI->getParent());
4483 SI->eraseFromParent();
4487 ++NumLookupTablesHoles;
4491 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
4492 BasicBlock *BB = SI->getParent();
4494 if (isValueEqualityComparison(SI)) {
4495 // If we only have one predecessor, and if it is a branch on this value,
4496 // see if that predecessor totally determines the outcome of this switch.
4497 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4498 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
4499 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4501 Value *Cond = SI->getCondition();
4502 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
4503 if (SimplifySwitchOnSelect(SI, Select))
4504 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4506 // If the block only contains the switch, see if we can fold the block
4507 // away into any preds.
4508 BasicBlock::iterator BBI = BB->begin();
4509 // Ignore dbg intrinsics.
4510 while (isa<DbgInfoIntrinsic>(BBI))
4513 if (FoldValueComparisonIntoPredecessors(SI, Builder))
4514 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4517 // Try to transform the switch into an icmp and a branch.
4518 if (TurnSwitchRangeIntoICmp(SI, Builder))
4519 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4521 // Remove unreachable cases.
4522 if (EliminateDeadSwitchCases(SI, AC, DL))
4523 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4525 if (SwitchToSelect(SI, Builder, AC, DL))
4526 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4528 if (ForwardSwitchConditionToPHI(SI))
4529 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4531 if (SwitchToLookupTable(SI, Builder, DL, TTI))
4532 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4537 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
4538 BasicBlock *BB = IBI->getParent();
4539 bool Changed = false;
4541 // Eliminate redundant destinations.
4542 SmallPtrSet<Value *, 8> Succs;
4543 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
4544 BasicBlock *Dest = IBI->getDestination(i);
4545 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
4546 Dest->removePredecessor(BB);
4547 IBI->removeDestination(i);
4553 if (IBI->getNumDestinations() == 0) {
4554 // If the indirectbr has no successors, change it to unreachable.
4555 new UnreachableInst(IBI->getContext(), IBI);
4556 EraseTerminatorInstAndDCECond(IBI);
4560 if (IBI->getNumDestinations() == 1) {
4561 // If the indirectbr has one successor, change it to a direct branch.
4562 BranchInst::Create(IBI->getDestination(0), IBI);
4563 EraseTerminatorInstAndDCECond(IBI);
4567 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
4568 if (SimplifyIndirectBrOnSelect(IBI, SI))
4569 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4574 /// Given an block with only a single landing pad and a unconditional branch
4575 /// try to find another basic block which this one can be merged with. This
4576 /// handles cases where we have multiple invokes with unique landing pads, but
4577 /// a shared handler.
4579 /// We specifically choose to not worry about merging non-empty blocks
4580 /// here. That is a PRE/scheduling problem and is best solved elsewhere. In
4581 /// practice, the optimizer produces empty landing pad blocks quite frequently
4582 /// when dealing with exception dense code. (see: instcombine, gvn, if-else
4583 /// sinking in this file)
4585 /// This is primarily a code size optimization. We need to avoid performing
4586 /// any transform which might inhibit optimization (such as our ability to
4587 /// specialize a particular handler via tail commoning). We do this by not
4588 /// merging any blocks which require us to introduce a phi. Since the same
4589 /// values are flowing through both blocks, we don't loose any ability to
4590 /// specialize. If anything, we make such specialization more likely.
4592 /// TODO - This transformation could remove entries from a phi in the target
4593 /// block when the inputs in the phi are the same for the two blocks being
4594 /// merged. In some cases, this could result in removal of the PHI entirely.
4595 static bool TryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI,
4597 auto Succ = BB->getUniqueSuccessor();
4599 // If there's a phi in the successor block, we'd likely have to introduce
4600 // a phi into the merged landing pad block.
4601 if (isa<PHINode>(*Succ->begin()))
4604 for (BasicBlock *OtherPred : predecessors(Succ)) {
4605 if (BB == OtherPred)
4607 BasicBlock::iterator I = OtherPred->begin();
4608 LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(I);
4609 if (!LPad2 || !LPad2->isIdenticalTo(LPad))
4611 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {}
4612 BranchInst *BI2 = dyn_cast<BranchInst>(I);
4613 if (!BI2 || !BI2->isIdenticalTo(BI))
4616 // We've found an identical block. Update our predeccessors to take that
4617 // path instead and make ourselves dead.
4618 SmallSet<BasicBlock *, 16> Preds;
4619 Preds.insert(pred_begin(BB), pred_end(BB));
4620 for (BasicBlock *Pred : Preds) {
4621 InvokeInst *II = cast<InvokeInst>(Pred->getTerminator());
4622 assert(II->getNormalDest() != BB &&
4623 II->getUnwindDest() == BB && "unexpected successor");
4624 II->setUnwindDest(OtherPred);
4627 // The debug info in OtherPred doesn't cover the merged control flow that
4628 // used to go through BB. We need to delete it or update it.
4629 for (auto I = OtherPred->begin(), E = OtherPred->end();
4631 Instruction &Inst = *I; I++;
4632 if (isa<DbgInfoIntrinsic>(Inst))
4633 Inst.eraseFromParent();
4636 SmallSet<BasicBlock *, 16> Succs;
4637 Succs.insert(succ_begin(BB), succ_end(BB));
4638 for (BasicBlock *Succ : Succs) {
4639 Succ->removePredecessor(BB);
4642 IRBuilder<> Builder(BI);
4643 Builder.CreateUnreachable();
4644 BI->eraseFromParent();
4650 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
4651 BasicBlock *BB = BI->getParent();
4653 if (SinkCommon && SinkThenElseCodeToEnd(BI))
4656 // If the Terminator is the only non-phi instruction, simplify the block.
4657 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg()->getIterator();
4658 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
4659 TryToSimplifyUncondBranchFromEmptyBlock(BB))
4662 // If the only instruction in the block is a seteq/setne comparison
4663 // against a constant, try to simplify the block.
4664 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
4665 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
4666 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
4668 if (I->isTerminator() &&
4669 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, DL, TTI,
4670 BonusInstThreshold, AC))
4674 // See if we can merge an empty landing pad block with another which is
4676 if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(I)) {
4677 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {}
4678 if (I->isTerminator() &&
4679 TryToMergeLandingPad(LPad, BI, BB))
4683 // If this basic block is ONLY a compare and a branch, and if a predecessor
4684 // branches to us and our successor, fold the comparison into the
4685 // predecessor and use logical operations to update the incoming value
4686 // for PHI nodes in common successor.
4687 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
4688 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4693 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
4694 BasicBlock *BB = BI->getParent();
4696 // Conditional branch
4697 if (isValueEqualityComparison(BI)) {
4698 // If we only have one predecessor, and if it is a branch on this value,
4699 // see if that predecessor totally determines the outcome of this
4701 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4702 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
4703 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4705 // This block must be empty, except for the setcond inst, if it exists.
4706 // Ignore dbg intrinsics.
4707 BasicBlock::iterator I = BB->begin();
4708 // Ignore dbg intrinsics.
4709 while (isa<DbgInfoIntrinsic>(I))
4712 if (FoldValueComparisonIntoPredecessors(BI, Builder))
4713 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4714 } else if (&*I == cast<Instruction>(BI->getCondition())){
4716 // Ignore dbg intrinsics.
4717 while (isa<DbgInfoIntrinsic>(I))
4719 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
4720 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4724 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
4725 if (SimplifyBranchOnICmpChain(BI, Builder, DL))
4728 // If this basic block is ONLY a compare and a branch, and if a predecessor
4729 // branches to us and one of our successors, fold the comparison into the
4730 // predecessor and use logical operations to pick the right destination.
4731 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
4732 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4734 // We have a conditional branch to two blocks that are only reachable
4735 // from BI. We know that the condbr dominates the two blocks, so see if
4736 // there is any identical code in the "then" and "else" blocks. If so, we
4737 // can hoist it up to the branching block.
4738 if (BI->getSuccessor(0)->getSinglePredecessor()) {
4739 if (BI->getSuccessor(1)->getSinglePredecessor()) {
4740 if (HoistThenElseCodeToIf(BI, TTI))
4741 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4743 // If Successor #1 has multiple preds, we may be able to conditionally
4744 // execute Successor #0 if it branches to Successor #1.
4745 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
4746 if (Succ0TI->getNumSuccessors() == 1 &&
4747 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
4748 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), TTI))
4749 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4751 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
4752 // If Successor #0 has multiple preds, we may be able to conditionally
4753 // execute Successor #1 if it branches to Successor #0.
4754 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
4755 if (Succ1TI->getNumSuccessors() == 1 &&
4756 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
4757 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), TTI))
4758 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4761 // If this is a branch on a phi node in the current block, thread control
4762 // through this block if any PHI node entries are constants.
4763 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
4764 if (PN->getParent() == BI->getParent())
4765 if (FoldCondBranchOnPHI(BI, DL))
4766 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4768 // Scan predecessor blocks for conditional branches.
4769 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
4770 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
4771 if (PBI != BI && PBI->isConditional())
4772 if (SimplifyCondBranchToCondBranch(PBI, BI, DL))
4773 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4778 /// Check if passing a value to an instruction will cause undefined behavior.
4779 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
4780 Constant *C = dyn_cast<Constant>(V);
4787 if (C->isNullValue()) {
4788 // Only look at the first use, avoid hurting compile time with long uselists
4789 User *Use = *I->user_begin();
4791 // Now make sure that there are no instructions in between that can alter
4792 // control flow (eg. calls)
4793 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
4794 if (i == I->getParent()->end() || i->mayHaveSideEffects())
4797 // Look through GEPs. A load from a GEP derived from NULL is still undefined
4798 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
4799 if (GEP->getPointerOperand() == I)
4800 return passingValueIsAlwaysUndefined(V, GEP);
4802 // Look through bitcasts.
4803 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
4804 return passingValueIsAlwaysUndefined(V, BC);
4806 // Load from null is undefined.
4807 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
4808 if (!LI->isVolatile())
4809 return LI->getPointerAddressSpace() == 0;
4811 // Store to null is undefined.
4812 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
4813 if (!SI->isVolatile())
4814 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
4819 /// If BB has an incoming value that will always trigger undefined behavior
4820 /// (eg. null pointer dereference), remove the branch leading here.
4821 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
4822 for (BasicBlock::iterator i = BB->begin();
4823 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
4824 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
4825 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
4826 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
4827 IRBuilder<> Builder(T);
4828 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
4829 BB->removePredecessor(PHI->getIncomingBlock(i));
4830 // Turn uncoditional branches into unreachables and remove the dead
4831 // destination from conditional branches.
4832 if (BI->isUnconditional())
4833 Builder.CreateUnreachable();
4835 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
4836 BI->getSuccessor(0));
4837 BI->eraseFromParent();
4840 // TODO: SwitchInst.
4846 bool SimplifyCFGOpt::run(BasicBlock *BB) {
4847 bool Changed = false;
4849 assert(BB && BB->getParent() && "Block not embedded in function!");
4850 assert(BB->getTerminator() && "Degenerate basic block encountered!");
4852 // Remove basic blocks that have no predecessors (except the entry block)...
4853 // or that just have themself as a predecessor. These are unreachable.
4854 if ((pred_empty(BB) &&
4855 BB != &BB->getParent()->getEntryBlock()) ||
4856 BB->getSinglePredecessor() == BB) {
4857 DEBUG(dbgs() << "Removing BB: \n" << *BB);
4858 DeleteDeadBlock(BB);
4862 // Check to see if we can constant propagate this terminator instruction
4864 Changed |= ConstantFoldTerminator(BB, true);
4866 // Check for and eliminate duplicate PHI nodes in this block.
4867 Changed |= EliminateDuplicatePHINodes(BB);
4869 // Check for and remove branches that will always cause undefined behavior.
4870 Changed |= removeUndefIntroducingPredecessor(BB);
4872 // Merge basic blocks into their predecessor if there is only one distinct
4873 // pred, and if there is only one distinct successor of the predecessor, and
4874 // if there are no PHI nodes.
4876 if (MergeBlockIntoPredecessor(BB))
4879 IRBuilder<> Builder(BB);
4881 // If there is a trivial two-entry PHI node in this basic block, and we can
4882 // eliminate it, do so now.
4883 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
4884 if (PN->getNumIncomingValues() == 2)
4885 Changed |= FoldTwoEntryPHINode(PN, TTI, DL);
4887 Builder.SetInsertPoint(BB->getTerminator());
4888 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
4889 if (BI->isUnconditional()) {
4890 if (SimplifyUncondBranch(BI, Builder)) return true;
4892 if (SimplifyCondBranch(BI, Builder)) return true;
4894 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
4895 if (SimplifyReturn(RI, Builder)) return true;
4896 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
4897 if (SimplifyResume(RI, Builder)) return true;
4898 } else if (CleanupReturnInst *RI =
4899 dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
4900 if (SimplifyCleanupReturn(RI)) return true;
4901 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
4902 if (SimplifySwitch(SI, Builder)) return true;
4903 } else if (UnreachableInst *UI =
4904 dyn_cast<UnreachableInst>(BB->getTerminator())) {
4905 if (SimplifyUnreachable(UI)) return true;
4906 } else if (IndirectBrInst *IBI =
4907 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
4908 if (SimplifyIndirectBr(IBI)) return true;
4914 /// This function is used to do simplification of a CFG.
4915 /// For example, it adjusts branches to branches to eliminate the extra hop,
4916 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4917 /// of the CFG. It returns true if a modification was made.
4919 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
4920 unsigned BonusInstThreshold, AssumptionCache *AC) {
4921 return SimplifyCFGOpt(TTI, BB->getModule()->getDataLayout(),
4922 BonusInstThreshold, AC).run(BB);