1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Peephole optimize the CFG.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Transforms/Utils/Local.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/ADT/SetOperations.h"
18 #include "llvm/ADT/SetVector.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/Analysis/ConstantFolding.h"
23 #include "llvm/Analysis/EHPersonalities.h"
24 #include "llvm/Analysis/InstructionSimplify.h"
25 #include "llvm/Analysis/TargetTransformInfo.h"
26 #include "llvm/Analysis/ValueTracking.h"
27 #include "llvm/IR/CFG.h"
28 #include "llvm/IR/ConstantRange.h"
29 #include "llvm/IR/Constants.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/DerivedTypes.h"
32 #include "llvm/IR/GlobalVariable.h"
33 #include "llvm/IR/IRBuilder.h"
34 #include "llvm/IR/Instructions.h"
35 #include "llvm/IR/IntrinsicInst.h"
36 #include "llvm/IR/LLVMContext.h"
37 #include "llvm/IR/MDBuilder.h"
38 #include "llvm/IR/Metadata.h"
39 #include "llvm/IR/Module.h"
40 #include "llvm/IR/NoFolder.h"
41 #include "llvm/IR/Operator.h"
42 #include "llvm/IR/PatternMatch.h"
43 #include "llvm/IR/Type.h"
44 #include "llvm/Support/CommandLine.h"
45 #include "llvm/Support/Debug.h"
46 #include "llvm/Support/raw_ostream.h"
47 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
48 #include "llvm/Transforms/Utils/ValueMapper.h"
53 using namespace PatternMatch;
55 #define DEBUG_TYPE "simplifycfg"
57 // Chosen as 2 so as to be cheap, but still to have enough power to fold
58 // a select, so the "clamp" idiom (of a min followed by a max) will be caught.
59 // To catch this, we need to fold a compare and a select, hence '2' being the
60 // minimum reasonable default.
61 static cl::opt<unsigned>
62 PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(2),
63 cl::desc("Control the amount of phi node folding to perform (default = 2)"));
66 DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false),
67 cl::desc("Duplicate return instructions into unconditional branches"));
70 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
71 cl::desc("Sink common instructions down to the end block"));
73 static cl::opt<bool> HoistCondStores(
74 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
75 cl::desc("Hoist conditional stores if an unconditional store precedes"));
77 static cl::opt<bool> MergeCondStores(
78 "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
79 cl::desc("Hoist conditional stores even if an unconditional store does not "
80 "precede - hoist multiple conditional stores into a single "
83 static cl::opt<bool> MergeCondStoresAggressively(
84 "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
85 cl::desc("When merging conditional stores, do so even if the resultant "
86 "basic blocks are unlikely to be if-converted as a result"));
88 static cl::opt<bool> SpeculateOneExpensiveInst(
89 "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
90 cl::desc("Allow exactly one expensive instruction to be speculatively "
93 static cl::opt<unsigned> MaxSpeculationDepth(
94 "max-speculation-depth", cl::Hidden, cl::init(10),
95 cl::desc("Limit maximum recursion depth when calculating costs of "
96 "speculatively executed instructions"));
98 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
99 STATISTIC(NumLinearMaps, "Number of switch instructions turned into linear mapping");
100 STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables");
101 STATISTIC(NumLookupTablesHoles, "Number of switch instructions turned into lookup tables (holes checked)");
102 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
103 STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block");
104 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
107 // The first field contains the value that the switch produces when a certain
108 // case group is selected, and the second field is a vector containing the
109 // cases composing the case group.
110 typedef SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>
111 SwitchCaseResultVectorTy;
112 // The first field contains the phi node that generates a result of the switch
113 // and the second field contains the value generated for a certain case in the
114 // switch for that PHI.
115 typedef SmallVector<std::pair<PHINode *, Constant *>, 4> SwitchCaseResultsTy;
117 /// ValueEqualityComparisonCase - Represents a case of a switch.
118 struct ValueEqualityComparisonCase {
122 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
123 : Value(Value), Dest(Dest) {}
125 bool operator<(ValueEqualityComparisonCase RHS) const {
126 // Comparing pointers is ok as we only rely on the order for uniquing.
127 return Value < RHS.Value;
130 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
133 class SimplifyCFGOpt {
134 const TargetTransformInfo &TTI;
135 const DataLayout &DL;
136 unsigned BonusInstThreshold;
138 Value *isValueEqualityComparison(TerminatorInst *TI);
139 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
140 std::vector<ValueEqualityComparisonCase> &Cases);
141 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
143 IRBuilder<> &Builder);
144 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
145 IRBuilder<> &Builder);
147 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
148 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
149 bool SimplifySingleResume(ResumeInst *RI);
150 bool SimplifyCommonResume(ResumeInst *RI);
151 bool SimplifyCleanupReturn(CleanupReturnInst *RI);
152 bool SimplifyUnreachable(UnreachableInst *UI);
153 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
154 bool SimplifyIndirectBr(IndirectBrInst *IBI);
155 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder);
156 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder);
159 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
160 unsigned BonusInstThreshold, AssumptionCache *AC)
161 : TTI(TTI), DL(DL), BonusInstThreshold(BonusInstThreshold), AC(AC) {}
162 bool run(BasicBlock *BB);
166 /// Return true if it is safe to merge these two
167 /// terminator instructions together.
168 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
169 if (SI1 == SI2) return false; // Can't merge with self!
171 // It is not safe to merge these two switch instructions if they have a common
172 // successor, and if that successor has a PHI node, and if *that* PHI node has
173 // conflicting incoming values from the two switch blocks.
174 BasicBlock *SI1BB = SI1->getParent();
175 BasicBlock *SI2BB = SI2->getParent();
176 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
178 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
179 if (SI1Succs.count(*I))
180 for (BasicBlock::iterator BBI = (*I)->begin();
181 isa<PHINode>(BBI); ++BBI) {
182 PHINode *PN = cast<PHINode>(BBI);
183 if (PN->getIncomingValueForBlock(SI1BB) !=
184 PN->getIncomingValueForBlock(SI2BB))
191 /// Return true if it is safe and profitable to merge these two terminator
192 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
193 /// store all PHI nodes in common successors.
194 static bool isProfitableToFoldUnconditional(BranchInst *SI1,
197 SmallVectorImpl<PHINode*> &PhiNodes) {
198 if (SI1 == SI2) return false; // Can't merge with self!
199 assert(SI1->isUnconditional() && SI2->isConditional());
201 // We fold the unconditional branch if we can easily update all PHI nodes in
202 // common successors:
203 // 1> We have a constant incoming value for the conditional branch;
204 // 2> We have "Cond" as the incoming value for the unconditional branch;
205 // 3> SI2->getCondition() and Cond have same operands.
206 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
207 if (!Ci2) return false;
208 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
209 Cond->getOperand(1) == Ci2->getOperand(1)) &&
210 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
211 Cond->getOperand(1) == Ci2->getOperand(0)))
214 BasicBlock *SI1BB = SI1->getParent();
215 BasicBlock *SI2BB = SI2->getParent();
216 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
217 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
218 if (SI1Succs.count(*I))
219 for (BasicBlock::iterator BBI = (*I)->begin();
220 isa<PHINode>(BBI); ++BBI) {
221 PHINode *PN = cast<PHINode>(BBI);
222 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
223 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
225 PhiNodes.push_back(PN);
230 /// Update PHI nodes in Succ to indicate that there will now be entries in it
231 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
232 /// will be the same as those coming in from ExistPred, an existing predecessor
234 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
235 BasicBlock *ExistPred) {
236 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
239 for (BasicBlock::iterator I = Succ->begin();
240 (PN = dyn_cast<PHINode>(I)); ++I)
241 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
244 /// Compute an abstract "cost" of speculating the given instruction,
245 /// which is assumed to be safe to speculate. TCC_Free means cheap,
246 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
248 static unsigned ComputeSpeculationCost(const User *I,
249 const TargetTransformInfo &TTI) {
250 assert(isSafeToSpeculativelyExecute(I) &&
251 "Instruction is not safe to speculatively execute!");
252 return TTI.getUserCost(I);
255 /// If we have a merge point of an "if condition" as accepted above,
256 /// return true if the specified value dominates the block. We
257 /// don't handle the true generality of domination here, just a special case
258 /// which works well enough for us.
260 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
261 /// see if V (which must be an instruction) and its recursive operands
262 /// that do not dominate BB have a combined cost lower than CostRemaining and
263 /// are non-trapping. If both are true, the instruction is inserted into the
264 /// set and true is returned.
266 /// The cost for most non-trapping instructions is defined as 1 except for
267 /// Select whose cost is 2.
269 /// After this function returns, CostRemaining is decreased by the cost of
270 /// V plus its non-dominating operands. If that cost is greater than
271 /// CostRemaining, false is returned and CostRemaining is undefined.
272 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
273 SmallPtrSetImpl<Instruction*> *AggressiveInsts,
274 unsigned &CostRemaining,
275 const TargetTransformInfo &TTI,
276 unsigned Depth = 0) {
277 // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
278 // so limit the recursion depth.
279 // TODO: While this recursion limit does prevent pathological behavior, it
280 // would be better to track visited instructions to avoid cycles.
281 if (Depth == MaxSpeculationDepth)
284 Instruction *I = dyn_cast<Instruction>(V);
286 // Non-instructions all dominate instructions, but not all constantexprs
287 // can be executed unconditionally.
288 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
293 BasicBlock *PBB = I->getParent();
295 // We don't want to allow weird loops that might have the "if condition" in
296 // the bottom of this block.
297 if (PBB == BB) return false;
299 // If this instruction is defined in a block that contains an unconditional
300 // branch to BB, then it must be in the 'conditional' part of the "if
301 // statement". If not, it definitely dominates the region.
302 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
303 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
306 // If we aren't allowing aggressive promotion anymore, then don't consider
307 // instructions in the 'if region'.
308 if (!AggressiveInsts) return false;
310 // If we have seen this instruction before, don't count it again.
311 if (AggressiveInsts->count(I)) return true;
313 // Okay, it looks like the instruction IS in the "condition". Check to
314 // see if it's a cheap instruction to unconditionally compute, and if it
315 // only uses stuff defined outside of the condition. If so, hoist it out.
316 if (!isSafeToSpeculativelyExecute(I))
319 unsigned Cost = ComputeSpeculationCost(I, TTI);
321 // Allow exactly one instruction to be speculated regardless of its cost
322 // (as long as it is safe to do so).
323 // This is intended to flatten the CFG even if the instruction is a division
324 // or other expensive operation. The speculation of an expensive instruction
325 // is expected to be undone in CodeGenPrepare if the speculation has not
326 // enabled further IR optimizations.
327 if (Cost > CostRemaining &&
328 (!SpeculateOneExpensiveInst || !AggressiveInsts->empty() || Depth > 0))
331 // Avoid unsigned wrap.
332 CostRemaining = (Cost > CostRemaining) ? 0 : CostRemaining - Cost;
334 // Okay, we can only really hoist these out if their operands do
335 // not take us over the cost threshold.
336 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
337 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, TTI,
340 // Okay, it's safe to do this! Remember this instruction.
341 AggressiveInsts->insert(I);
345 /// Extract ConstantInt from value, looking through IntToPtr
346 /// and PointerNullValue. Return NULL if value is not a constant int.
347 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
348 // Normal constant int.
349 ConstantInt *CI = dyn_cast<ConstantInt>(V);
350 if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
353 // This is some kind of pointer constant. Turn it into a pointer-sized
354 // ConstantInt if possible.
355 IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
357 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
358 if (isa<ConstantPointerNull>(V))
359 return ConstantInt::get(PtrTy, 0);
361 // IntToPtr const int.
362 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
363 if (CE->getOpcode() == Instruction::IntToPtr)
364 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
365 // The constant is very likely to have the right type already.
366 if (CI->getType() == PtrTy)
369 return cast<ConstantInt>
370 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
377 /// Given a chain of or (||) or and (&&) comparison of a value against a
378 /// constant, this will try to recover the information required for a switch
380 /// It will depth-first traverse the chain of comparison, seeking for patterns
381 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
382 /// representing the different cases for the switch.
383 /// Note that if the chain is composed of '||' it will build the set of elements
384 /// that matches the comparisons (i.e. any of this value validate the chain)
385 /// while for a chain of '&&' it will build the set elements that make the test
387 struct ConstantComparesGatherer {
388 const DataLayout &DL;
389 Value *CompValue; /// Value found for the switch comparison
390 Value *Extra; /// Extra clause to be checked before the switch
391 SmallVector<ConstantInt *, 8> Vals; /// Set of integers to match in switch
392 unsigned UsedICmps; /// Number of comparisons matched in the and/or chain
394 /// Construct and compute the result for the comparison instruction Cond
395 ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL)
396 : DL(DL), CompValue(nullptr), Extra(nullptr), UsedICmps(0) {
401 ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
402 ConstantComparesGatherer &
403 operator=(const ConstantComparesGatherer &) = delete;
407 /// Try to set the current value used for the comparison, it succeeds only if
408 /// it wasn't set before or if the new value is the same as the old one
409 bool setValueOnce(Value *NewVal) {
410 if(CompValue && CompValue != NewVal) return false;
412 return (CompValue != nullptr);
415 /// Try to match Instruction "I" as a comparison against a constant and
416 /// populates the array Vals with the set of values that match (or do not
417 /// match depending on isEQ).
418 /// Return false on failure. On success, the Value the comparison matched
419 /// against is placed in CompValue.
420 /// If CompValue is already set, the function is expected to fail if a match
421 /// is found but the value compared to is different.
422 bool matchInstruction(Instruction *I, bool isEQ) {
423 // If this is an icmp against a constant, handle this as one of the cases.
426 if (!((ICI = dyn_cast<ICmpInst>(I)) &&
427 (C = GetConstantInt(I->getOperand(1), DL)))) {
434 // Pattern match a special case
435 // (x & ~2^x) == y --> x == y || x == y|2^x
436 // This undoes a transformation done by instcombine to fuse 2 compares.
437 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
438 if (match(ICI->getOperand(0),
439 m_And(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
440 APInt Not = ~RHSC->getValue();
441 if (Not.isPowerOf2()) {
442 // If we already have a value for the switch, it has to match!
443 if(!setValueOnce(RHSVal))
447 Vals.push_back(ConstantInt::get(C->getContext(),
448 C->getValue() | Not));
454 // If we already have a value for the switch, it has to match!
455 if(!setValueOnce(ICI->getOperand(0)))
460 return ICI->getOperand(0);
463 // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
464 ConstantRange Span = ConstantRange::makeAllowedICmpRegion(
465 ICI->getPredicate(), C->getValue());
467 // Shift the range if the compare is fed by an add. This is the range
468 // compare idiom as emitted by instcombine.
469 Value *CandidateVal = I->getOperand(0);
470 if(match(I->getOperand(0), m_Add(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
471 Span = Span.subtract(RHSC->getValue());
472 CandidateVal = RHSVal;
475 // If this is an and/!= check, then we are looking to build the set of
476 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
479 Span = Span.inverse();
481 // If there are a ton of values, we don't want to make a ginormous switch.
482 if (Span.getSetSize().ugt(8) || Span.isEmptySet()) {
486 // If we already have a value for the switch, it has to match!
487 if(!setValueOnce(CandidateVal))
490 // Add all values from the range to the set
491 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
492 Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
499 /// Given a potentially 'or'd or 'and'd together collection of icmp
500 /// eq/ne/lt/gt instructions that compare a value against a constant, extract
501 /// the value being compared, and stick the list constants into the Vals
503 /// One "Extra" case is allowed to differ from the other.
504 void gather(Value *V) {
505 Instruction *I = dyn_cast<Instruction>(V);
506 bool isEQ = (I->getOpcode() == Instruction::Or);
508 // Keep a stack (SmallVector for efficiency) for depth-first traversal
509 SmallVector<Value *, 8> DFT;
514 while(!DFT.empty()) {
515 V = DFT.pop_back_val();
517 if (Instruction *I = dyn_cast<Instruction>(V)) {
518 // If it is a || (or && depending on isEQ), process the operands.
519 if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
520 DFT.push_back(I->getOperand(1));
521 DFT.push_back(I->getOperand(0));
525 // Try to match the current instruction
526 if (matchInstruction(I, isEQ))
527 // Match succeed, continue the loop
531 // One element of the sequence of || (or &&) could not be match as a
532 // comparison against the same value as the others.
533 // We allow only one "Extra" case to be checked before the switch
538 // Failed to parse a proper sequence, abort now
547 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
548 Instruction *Cond = nullptr;
549 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
550 Cond = dyn_cast<Instruction>(SI->getCondition());
551 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
552 if (BI->isConditional())
553 Cond = dyn_cast<Instruction>(BI->getCondition());
554 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
555 Cond = dyn_cast<Instruction>(IBI->getAddress());
558 TI->eraseFromParent();
559 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
562 /// Return true if the specified terminator checks
563 /// to see if a value is equal to constant integer value.
564 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
566 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
567 // Do not permit merging of large switch instructions into their
568 // predecessors unless there is only one predecessor.
569 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()),
570 pred_end(SI->getParent())) <= 128)
571 CV = SI->getCondition();
572 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
573 if (BI->isConditional() && BI->getCondition()->hasOneUse())
574 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
575 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
576 CV = ICI->getOperand(0);
579 // Unwrap any lossless ptrtoint cast.
581 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
582 Value *Ptr = PTII->getPointerOperand();
583 if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
590 /// Given a value comparison instruction,
591 /// decode all of the 'cases' that it represents and return the 'default' block.
592 BasicBlock *SimplifyCFGOpt::
593 GetValueEqualityComparisonCases(TerminatorInst *TI,
594 std::vector<ValueEqualityComparisonCase>
596 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
597 Cases.reserve(SI->getNumCases());
598 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
599 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
600 i.getCaseSuccessor()));
601 return SI->getDefaultDest();
604 BranchInst *BI = cast<BranchInst>(TI);
605 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
606 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
607 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
610 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
614 /// Given a vector of bb/value pairs, remove any entries
615 /// in the list that match the specified block.
616 static void EliminateBlockCases(BasicBlock *BB,
617 std::vector<ValueEqualityComparisonCase> &Cases) {
618 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
621 /// Return true if there are any keys in C1 that exist in C2 as well.
623 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
624 std::vector<ValueEqualityComparisonCase > &C2) {
625 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
627 // Make V1 be smaller than V2.
628 if (V1->size() > V2->size())
631 if (V1->size() == 0) return false;
632 if (V1->size() == 1) {
634 ConstantInt *TheVal = (*V1)[0].Value;
635 for (unsigned i = 0, e = V2->size(); i != e; ++i)
636 if (TheVal == (*V2)[i].Value)
640 // Otherwise, just sort both lists and compare element by element.
641 array_pod_sort(V1->begin(), V1->end());
642 array_pod_sort(V2->begin(), V2->end());
643 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
644 while (i1 != e1 && i2 != e2) {
645 if ((*V1)[i1].Value == (*V2)[i2].Value)
647 if ((*V1)[i1].Value < (*V2)[i2].Value)
655 /// If TI is known to be a terminator instruction and its block is known to
656 /// only have a single predecessor block, check to see if that predecessor is
657 /// also a value comparison with the same value, and if that comparison
658 /// determines the outcome of this comparison. If so, simplify TI. This does a
659 /// very limited form of jump threading.
660 bool SimplifyCFGOpt::
661 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
663 IRBuilder<> &Builder) {
664 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
665 if (!PredVal) return false; // Not a value comparison in predecessor.
667 Value *ThisVal = isValueEqualityComparison(TI);
668 assert(ThisVal && "This isn't a value comparison!!");
669 if (ThisVal != PredVal) return false; // Different predicates.
671 // TODO: Preserve branch weight metadata, similarly to how
672 // FoldValueComparisonIntoPredecessors preserves it.
674 // Find out information about when control will move from Pred to TI's block.
675 std::vector<ValueEqualityComparisonCase> PredCases;
676 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
678 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
680 // Find information about how control leaves this block.
681 std::vector<ValueEqualityComparisonCase> ThisCases;
682 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
683 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
685 // If TI's block is the default block from Pred's comparison, potentially
686 // simplify TI based on this knowledge.
687 if (PredDef == TI->getParent()) {
688 // If we are here, we know that the value is none of those cases listed in
689 // PredCases. If there are any cases in ThisCases that are in PredCases, we
691 if (!ValuesOverlap(PredCases, ThisCases))
694 if (isa<BranchInst>(TI)) {
695 // Okay, one of the successors of this condbr is dead. Convert it to a
697 assert(ThisCases.size() == 1 && "Branch can only have one case!");
698 // Insert the new branch.
699 Instruction *NI = Builder.CreateBr(ThisDef);
702 // Remove PHI node entries for the dead edge.
703 ThisCases[0].Dest->removePredecessor(TI->getParent());
705 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
706 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
708 EraseTerminatorInstAndDCECond(TI);
712 SwitchInst *SI = cast<SwitchInst>(TI);
713 // Okay, TI has cases that are statically dead, prune them away.
714 SmallPtrSet<Constant*, 16> DeadCases;
715 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
716 DeadCases.insert(PredCases[i].Value);
718 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
719 << "Through successor TI: " << *TI);
721 // Collect branch weights into a vector.
722 SmallVector<uint32_t, 8> Weights;
723 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
724 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
726 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
728 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
729 Weights.push_back(CI->getValue().getZExtValue());
731 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
733 if (DeadCases.count(i.getCaseValue())) {
735 std::swap(Weights[i.getCaseIndex()+1], Weights.back());
738 i.getCaseSuccessor()->removePredecessor(TI->getParent());
742 if (HasWeight && Weights.size() >= 2)
743 SI->setMetadata(LLVMContext::MD_prof,
744 MDBuilder(SI->getParent()->getContext()).
745 createBranchWeights(Weights));
747 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
751 // Otherwise, TI's block must correspond to some matched value. Find out
752 // which value (or set of values) this is.
753 ConstantInt *TIV = nullptr;
754 BasicBlock *TIBB = TI->getParent();
755 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
756 if (PredCases[i].Dest == TIBB) {
758 return false; // Cannot handle multiple values coming to this block.
759 TIV = PredCases[i].Value;
761 assert(TIV && "No edge from pred to succ?");
763 // Okay, we found the one constant that our value can be if we get into TI's
764 // BB. Find out which successor will unconditionally be branched to.
765 BasicBlock *TheRealDest = nullptr;
766 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
767 if (ThisCases[i].Value == TIV) {
768 TheRealDest = ThisCases[i].Dest;
772 // If not handled by any explicit cases, it is handled by the default case.
773 if (!TheRealDest) TheRealDest = ThisDef;
775 // Remove PHI node entries for dead edges.
776 BasicBlock *CheckEdge = TheRealDest;
777 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
778 if (*SI != CheckEdge)
779 (*SI)->removePredecessor(TIBB);
783 // Insert the new branch.
784 Instruction *NI = Builder.CreateBr(TheRealDest);
787 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
788 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
790 EraseTerminatorInstAndDCECond(TI);
795 /// This class implements a stable ordering of constant
796 /// integers that does not depend on their address. This is important for
797 /// applications that sort ConstantInt's to ensure uniqueness.
798 struct ConstantIntOrdering {
799 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
800 return LHS->getValue().ult(RHS->getValue());
805 static int ConstantIntSortPredicate(ConstantInt *const *P1,
806 ConstantInt *const *P2) {
807 const ConstantInt *LHS = *P1;
808 const ConstantInt *RHS = *P2;
809 if (LHS->getValue().ult(RHS->getValue()))
811 if (LHS->getValue() == RHS->getValue())
816 static inline bool HasBranchWeights(const Instruction* I) {
817 MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
818 if (ProfMD && ProfMD->getOperand(0))
819 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
820 return MDS->getString().equals("branch_weights");
825 /// Get Weights of a given TerminatorInst, the default weight is at the front
826 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
828 static void GetBranchWeights(TerminatorInst *TI,
829 SmallVectorImpl<uint64_t> &Weights) {
830 MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
832 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
833 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
834 Weights.push_back(CI->getValue().getZExtValue());
837 // If TI is a conditional eq, the default case is the false case,
838 // and the corresponding branch-weight data is at index 2. We swap the
839 // default weight to be the first entry.
840 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
841 assert(Weights.size() == 2);
842 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
843 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
844 std::swap(Weights.front(), Weights.back());
848 /// Keep halving the weights until all can fit in uint32_t.
849 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
850 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
851 if (Max > UINT_MAX) {
852 unsigned Offset = 32 - countLeadingZeros(Max);
853 for (uint64_t &I : Weights)
858 /// The specified terminator is a value equality comparison instruction
859 /// (either a switch or a branch on "X == c").
860 /// See if any of the predecessors of the terminator block are value comparisons
861 /// on the same value. If so, and if safe to do so, fold them together.
862 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
863 IRBuilder<> &Builder) {
864 BasicBlock *BB = TI->getParent();
865 Value *CV = isValueEqualityComparison(TI); // CondVal
866 assert(CV && "Not a comparison?");
867 bool Changed = false;
869 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
870 while (!Preds.empty()) {
871 BasicBlock *Pred = Preds.pop_back_val();
873 // See if the predecessor is a comparison with the same value.
874 TerminatorInst *PTI = Pred->getTerminator();
875 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
877 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
878 // Figure out which 'cases' to copy from SI to PSI.
879 std::vector<ValueEqualityComparisonCase> BBCases;
880 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
882 std::vector<ValueEqualityComparisonCase> PredCases;
883 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
885 // Based on whether the default edge from PTI goes to BB or not, fill in
886 // PredCases and PredDefault with the new switch cases we would like to
888 SmallVector<BasicBlock*, 8> NewSuccessors;
890 // Update the branch weight metadata along the way
891 SmallVector<uint64_t, 8> Weights;
892 bool PredHasWeights = HasBranchWeights(PTI);
893 bool SuccHasWeights = HasBranchWeights(TI);
895 if (PredHasWeights) {
896 GetBranchWeights(PTI, Weights);
897 // branch-weight metadata is inconsistent here.
898 if (Weights.size() != 1 + PredCases.size())
899 PredHasWeights = SuccHasWeights = false;
900 } else if (SuccHasWeights)
901 // If there are no predecessor weights but there are successor weights,
902 // populate Weights with 1, which will later be scaled to the sum of
903 // successor's weights
904 Weights.assign(1 + PredCases.size(), 1);
906 SmallVector<uint64_t, 8> SuccWeights;
907 if (SuccHasWeights) {
908 GetBranchWeights(TI, SuccWeights);
909 // branch-weight metadata is inconsistent here.
910 if (SuccWeights.size() != 1 + BBCases.size())
911 PredHasWeights = SuccHasWeights = false;
912 } else if (PredHasWeights)
913 SuccWeights.assign(1 + BBCases.size(), 1);
915 if (PredDefault == BB) {
916 // If this is the default destination from PTI, only the edges in TI
917 // that don't occur in PTI, or that branch to BB will be activated.
918 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
919 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
920 if (PredCases[i].Dest != BB)
921 PTIHandled.insert(PredCases[i].Value);
923 // The default destination is BB, we don't need explicit targets.
924 std::swap(PredCases[i], PredCases.back());
926 if (PredHasWeights || SuccHasWeights) {
927 // Increase weight for the default case.
928 Weights[0] += Weights[i+1];
929 std::swap(Weights[i+1], Weights.back());
933 PredCases.pop_back();
937 // Reconstruct the new switch statement we will be building.
938 if (PredDefault != BBDefault) {
939 PredDefault->removePredecessor(Pred);
940 PredDefault = BBDefault;
941 NewSuccessors.push_back(BBDefault);
944 unsigned CasesFromPred = Weights.size();
945 uint64_t ValidTotalSuccWeight = 0;
946 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
947 if (!PTIHandled.count(BBCases[i].Value) &&
948 BBCases[i].Dest != BBDefault) {
949 PredCases.push_back(BBCases[i]);
950 NewSuccessors.push_back(BBCases[i].Dest);
951 if (SuccHasWeights || PredHasWeights) {
952 // The default weight is at index 0, so weight for the ith case
953 // should be at index i+1. Scale the cases from successor by
954 // PredDefaultWeight (Weights[0]).
955 Weights.push_back(Weights[0] * SuccWeights[i+1]);
956 ValidTotalSuccWeight += SuccWeights[i+1];
960 if (SuccHasWeights || PredHasWeights) {
961 ValidTotalSuccWeight += SuccWeights[0];
962 // Scale the cases from predecessor by ValidTotalSuccWeight.
963 for (unsigned i = 1; i < CasesFromPred; ++i)
964 Weights[i] *= ValidTotalSuccWeight;
965 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
966 Weights[0] *= SuccWeights[0];
969 // If this is not the default destination from PSI, only the edges
970 // in SI that occur in PSI with a destination of BB will be
972 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
973 std::map<ConstantInt*, uint64_t> WeightsForHandled;
974 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
975 if (PredCases[i].Dest == BB) {
976 PTIHandled.insert(PredCases[i].Value);
978 if (PredHasWeights || SuccHasWeights) {
979 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
980 std::swap(Weights[i+1], Weights.back());
984 std::swap(PredCases[i], PredCases.back());
985 PredCases.pop_back();
989 // Okay, now we know which constants were sent to BB from the
990 // predecessor. Figure out where they will all go now.
991 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
992 if (PTIHandled.count(BBCases[i].Value)) {
993 // If this is one we are capable of getting...
994 if (PredHasWeights || SuccHasWeights)
995 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
996 PredCases.push_back(BBCases[i]);
997 NewSuccessors.push_back(BBCases[i].Dest);
998 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
1001 // If there are any constants vectored to BB that TI doesn't handle,
1002 // they must go to the default destination of TI.
1003 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
1005 E = PTIHandled.end(); I != E; ++I) {
1006 if (PredHasWeights || SuccHasWeights)
1007 Weights.push_back(WeightsForHandled[*I]);
1008 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
1009 NewSuccessors.push_back(BBDefault);
1013 // Okay, at this point, we know which new successor Pred will get. Make
1014 // sure we update the number of entries in the PHI nodes for these
1016 for (BasicBlock *NewSuccessor : NewSuccessors)
1017 AddPredecessorToBlock(NewSuccessor, Pred, BB);
1019 Builder.SetInsertPoint(PTI);
1020 // Convert pointer to int before we switch.
1021 if (CV->getType()->isPointerTy()) {
1022 CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
1026 // Now that the successors are updated, create the new Switch instruction.
1027 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
1029 NewSI->setDebugLoc(PTI->getDebugLoc());
1030 for (ValueEqualityComparisonCase &V : PredCases)
1031 NewSI->addCase(V.Value, V.Dest);
1033 if (PredHasWeights || SuccHasWeights) {
1034 // Halve the weights if any of them cannot fit in an uint32_t
1035 FitWeights(Weights);
1037 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1039 NewSI->setMetadata(LLVMContext::MD_prof,
1040 MDBuilder(BB->getContext()).
1041 createBranchWeights(MDWeights));
1044 EraseTerminatorInstAndDCECond(PTI);
1046 // Okay, last check. If BB is still a successor of PSI, then we must
1047 // have an infinite loop case. If so, add an infinitely looping block
1048 // to handle the case to preserve the behavior of the code.
1049 BasicBlock *InfLoopBlock = nullptr;
1050 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1051 if (NewSI->getSuccessor(i) == BB) {
1052 if (!InfLoopBlock) {
1053 // Insert it at the end of the function, because it's either code,
1054 // or it won't matter if it's hot. :)
1055 InfLoopBlock = BasicBlock::Create(BB->getContext(),
1056 "infloop", BB->getParent());
1057 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1059 NewSI->setSuccessor(i, InfLoopBlock);
1068 // If we would need to insert a select that uses the value of this invoke
1069 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1070 // can't hoist the invoke, as there is nowhere to put the select in this case.
1071 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1072 Instruction *I1, Instruction *I2) {
1073 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1075 for (BasicBlock::iterator BBI = SI->begin();
1076 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1077 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1078 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1079 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
1087 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I);
1089 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1090 /// in the two blocks up into the branch block. The caller of this function
1091 /// guarantees that BI's block dominates BB1 and BB2.
1092 static bool HoistThenElseCodeToIf(BranchInst *BI,
1093 const TargetTransformInfo &TTI) {
1094 // This does very trivial matching, with limited scanning, to find identical
1095 // instructions in the two blocks. In particular, we don't want to get into
1096 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1097 // such, we currently just scan for obviously identical instructions in an
1099 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1100 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1102 BasicBlock::iterator BB1_Itr = BB1->begin();
1103 BasicBlock::iterator BB2_Itr = BB2->begin();
1105 Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1106 // Skip debug info if it is not identical.
1107 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1108 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1109 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1110 while (isa<DbgInfoIntrinsic>(I1))
1112 while (isa<DbgInfoIntrinsic>(I2))
1115 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1116 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1119 BasicBlock *BIParent = BI->getParent();
1121 bool Changed = false;
1123 // If we are hoisting the terminator instruction, don't move one (making a
1124 // broken BB), instead clone it, and remove BI.
1125 if (isa<TerminatorInst>(I1))
1126 goto HoistTerminator;
1128 if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1131 // For a normal instruction, we just move one to right before the branch,
1132 // then replace all uses of the other with the first. Finally, we remove
1133 // the now redundant second instruction.
1134 BIParent->getInstList().splice(BI->getIterator(), BB1->getInstList(), I1);
1135 if (!I2->use_empty())
1136 I2->replaceAllUsesWith(I1);
1137 I1->intersectOptionalDataWith(I2);
1138 unsigned KnownIDs[] = {
1139 LLVMContext::MD_tbaa, LLVMContext::MD_range,
1140 LLVMContext::MD_fpmath, LLVMContext::MD_invariant_load,
1141 LLVMContext::MD_nonnull, LLVMContext::MD_invariant_group,
1142 LLVMContext::MD_align, LLVMContext::MD_dereferenceable,
1143 LLVMContext::MD_dereferenceable_or_null};
1144 combineMetadata(I1, I2, KnownIDs);
1145 I2->eraseFromParent();
1150 // Skip debug info if it is not identical.
1151 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1152 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1153 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1154 while (isa<DbgInfoIntrinsic>(I1))
1156 while (isa<DbgInfoIntrinsic>(I2))
1159 } while (I1->isIdenticalToWhenDefined(I2));
1164 // It may not be possible to hoist an invoke.
1165 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1168 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1170 for (BasicBlock::iterator BBI = SI->begin();
1171 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1172 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1173 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1177 // Check for passingValueIsAlwaysUndefined here because we would rather
1178 // eliminate undefined control flow then converting it to a select.
1179 if (passingValueIsAlwaysUndefined(BB1V, PN) ||
1180 passingValueIsAlwaysUndefined(BB2V, PN))
1183 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1185 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1190 // Okay, it is safe to hoist the terminator.
1191 Instruction *NT = I1->clone();
1192 BIParent->getInstList().insert(BI->getIterator(), NT);
1193 if (!NT->getType()->isVoidTy()) {
1194 I1->replaceAllUsesWith(NT);
1195 I2->replaceAllUsesWith(NT);
1199 IRBuilder<true, NoFolder> Builder(NT);
1200 // Hoisting one of the terminators from our successor is a great thing.
1201 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1202 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1203 // nodes, so we insert select instruction to compute the final result.
1204 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1205 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1207 for (BasicBlock::iterator BBI = SI->begin();
1208 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1209 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1210 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1211 if (BB1V == BB2V) continue;
1213 // These values do not agree. Insert a select instruction before NT
1214 // that determines the right value.
1215 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1217 SI = cast<SelectInst>
1218 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1219 BB1V->getName()+"."+BB2V->getName()));
1221 // Make the PHI node use the select for all incoming values for BB1/BB2
1222 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1223 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1224 PN->setIncomingValue(i, SI);
1228 // Update any PHI nodes in our new successors.
1229 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1230 AddPredecessorToBlock(*SI, BIParent, BB1);
1232 EraseTerminatorInstAndDCECond(BI);
1236 /// Given an unconditional branch that goes to BBEnd,
1237 /// check whether BBEnd has only two predecessors and the other predecessor
1238 /// ends with an unconditional branch. If it is true, sink any common code
1239 /// in the two predecessors to BBEnd.
1240 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1241 assert(BI1->isUnconditional());
1242 BasicBlock *BB1 = BI1->getParent();
1243 BasicBlock *BBEnd = BI1->getSuccessor(0);
1245 // Check that BBEnd has two predecessors and the other predecessor ends with
1246 // an unconditional branch.
1247 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1248 BasicBlock *Pred0 = *PI++;
1249 if (PI == PE) // Only one predecessor.
1251 BasicBlock *Pred1 = *PI++;
1252 if (PI != PE) // More than two predecessors.
1254 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1255 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1256 if (!BI2 || !BI2->isUnconditional())
1259 // Gather the PHI nodes in BBEnd.
1260 SmallDenseMap<std::pair<Value *, Value *>, PHINode *> JointValueMap;
1261 Instruction *FirstNonPhiInBBEnd = nullptr;
1262 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end(); I != E; ++I) {
1263 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1264 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1265 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1266 JointValueMap[std::make_pair(BB1V, BB2V)] = PN;
1268 FirstNonPhiInBBEnd = &*I;
1272 if (!FirstNonPhiInBBEnd)
1275 // This does very trivial matching, with limited scanning, to find identical
1276 // instructions in the two blocks. We scan backward for obviously identical
1277 // instructions in an identical order.
1278 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1279 RE1 = BB1->getInstList().rend(),
1280 RI2 = BB2->getInstList().rbegin(),
1281 RE2 = BB2->getInstList().rend();
1283 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1286 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1289 // Skip the unconditional branches.
1293 bool Changed = false;
1294 while (RI1 != RE1 && RI2 != RE2) {
1296 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1299 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1303 Instruction *I1 = &*RI1, *I2 = &*RI2;
1304 auto InstPair = std::make_pair(I1, I2);
1305 // I1 and I2 should have a single use in the same PHI node, and they
1306 // perform the same operation.
1307 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1308 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1309 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1310 I1->isEHPad() || I2->isEHPad() ||
1311 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1312 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1313 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1314 !I1->hasOneUse() || !I2->hasOneUse() ||
1315 !JointValueMap.count(InstPair))
1318 // Check whether we should swap the operands of ICmpInst.
1319 // TODO: Add support of communativity.
1320 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1321 bool SwapOpnds = false;
1322 if (ICmp1 && ICmp2 &&
1323 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1324 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1325 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1326 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1327 ICmp2->swapOperands();
1330 if (!I1->isSameOperationAs(I2)) {
1332 ICmp2->swapOperands();
1336 // The operands should be either the same or they need to be generated
1337 // with a PHI node after sinking. We only handle the case where there is
1338 // a single pair of different operands.
1339 Value *DifferentOp1 = nullptr, *DifferentOp2 = nullptr;
1340 unsigned Op1Idx = ~0U;
1341 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1342 if (I1->getOperand(I) == I2->getOperand(I))
1344 // Early exit if we have more-than one pair of different operands or if
1345 // we need a PHI node to replace a constant.
1346 if (Op1Idx != ~0U ||
1347 isa<Constant>(I1->getOperand(I)) ||
1348 isa<Constant>(I2->getOperand(I))) {
1349 // If we can't sink the instructions, undo the swapping.
1351 ICmp2->swapOperands();
1354 DifferentOp1 = I1->getOperand(I);
1356 DifferentOp2 = I2->getOperand(I);
1359 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n");
1360 DEBUG(dbgs() << " " << *I2 << "\n");
1362 // We insert the pair of different operands to JointValueMap and
1363 // remove (I1, I2) from JointValueMap.
1364 if (Op1Idx != ~0U) {
1365 auto &NewPN = JointValueMap[std::make_pair(DifferentOp1, DifferentOp2)];
1368 PHINode::Create(DifferentOp1->getType(), 2,
1369 DifferentOp1->getName() + ".sink", &BBEnd->front());
1370 NewPN->addIncoming(DifferentOp1, BB1);
1371 NewPN->addIncoming(DifferentOp2, BB2);
1372 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1374 // I1 should use NewPN instead of DifferentOp1.
1375 I1->setOperand(Op1Idx, NewPN);
1377 PHINode *OldPN = JointValueMap[InstPair];
1378 JointValueMap.erase(InstPair);
1380 // We need to update RE1 and RE2 if we are going to sink the first
1381 // instruction in the basic block down.
1382 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1383 // Sink the instruction.
1384 BBEnd->getInstList().splice(FirstNonPhiInBBEnd->getIterator(),
1385 BB1->getInstList(), I1);
1386 if (!OldPN->use_empty())
1387 OldPN->replaceAllUsesWith(I1);
1388 OldPN->eraseFromParent();
1390 if (!I2->use_empty())
1391 I2->replaceAllUsesWith(I1);
1392 I1->intersectOptionalDataWith(I2);
1393 // TODO: Use combineMetadata here to preserve what metadata we can
1394 // (analogous to the hoisting case above).
1395 I2->eraseFromParent();
1398 RE1 = BB1->getInstList().rend();
1400 RE2 = BB2->getInstList().rend();
1401 FirstNonPhiInBBEnd = &*I1;
1408 /// \brief Determine if we can hoist sink a sole store instruction out of a
1409 /// conditional block.
1411 /// We are looking for code like the following:
1413 /// store i32 %add, i32* %arrayidx2
1414 /// ... // No other stores or function calls (we could be calling a memory
1415 /// ... // function).
1416 /// %cmp = icmp ult %x, %y
1417 /// br i1 %cmp, label %EndBB, label %ThenBB
1419 /// store i32 %add5, i32* %arrayidx2
1423 /// We are going to transform this into:
1425 /// store i32 %add, i32* %arrayidx2
1427 /// %cmp = icmp ult %x, %y
1428 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1429 /// store i32 %add.add5, i32* %arrayidx2
1432 /// \return The pointer to the value of the previous store if the store can be
1433 /// hoisted into the predecessor block. 0 otherwise.
1434 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1435 BasicBlock *StoreBB, BasicBlock *EndBB) {
1436 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1440 // Volatile or atomic.
1441 if (!StoreToHoist->isSimple())
1444 Value *StorePtr = StoreToHoist->getPointerOperand();
1446 // Look for a store to the same pointer in BrBB.
1447 unsigned MaxNumInstToLookAt = 10;
1448 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(),
1449 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) {
1450 Instruction *CurI = &*RI;
1452 // Could be calling an instruction that effects memory like free().
1453 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI))
1456 StoreInst *SI = dyn_cast<StoreInst>(CurI);
1457 // Found the previous store make sure it stores to the same location.
1458 if (SI && SI->getPointerOperand() == StorePtr)
1459 // Found the previous store, return its value operand.
1460 return SI->getValueOperand();
1462 return nullptr; // Unknown store.
1468 /// \brief Speculate a conditional basic block flattening the CFG.
1470 /// Note that this is a very risky transform currently. Speculating
1471 /// instructions like this is most often not desirable. Instead, there is an MI
1472 /// pass which can do it with full awareness of the resource constraints.
1473 /// However, some cases are "obvious" and we should do directly. An example of
1474 /// this is speculating a single, reasonably cheap instruction.
1476 /// There is only one distinct advantage to flattening the CFG at the IR level:
1477 /// it makes very common but simplistic optimizations such as are common in
1478 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1479 /// modeling their effects with easier to reason about SSA value graphs.
1482 /// An illustration of this transform is turning this IR:
1485 /// %cmp = icmp ult %x, %y
1486 /// br i1 %cmp, label %EndBB, label %ThenBB
1488 /// %sub = sub %x, %y
1491 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1498 /// %cmp = icmp ult %x, %y
1499 /// %sub = sub %x, %y
1500 /// %cond = select i1 %cmp, 0, %sub
1504 /// \returns true if the conditional block is removed.
1505 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1506 const TargetTransformInfo &TTI) {
1507 // Be conservative for now. FP select instruction can often be expensive.
1508 Value *BrCond = BI->getCondition();
1509 if (isa<FCmpInst>(BrCond))
1512 BasicBlock *BB = BI->getParent();
1513 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1515 // If ThenBB is actually on the false edge of the conditional branch, remember
1516 // to swap the select operands later.
1517 bool Invert = false;
1518 if (ThenBB != BI->getSuccessor(0)) {
1519 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1522 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1524 // Keep a count of how many times instructions are used within CondBB when
1525 // they are candidates for sinking into CondBB. Specifically:
1526 // - They are defined in BB, and
1527 // - They have no side effects, and
1528 // - All of their uses are in CondBB.
1529 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1531 unsigned SpeculationCost = 0;
1532 Value *SpeculatedStoreValue = nullptr;
1533 StoreInst *SpeculatedStore = nullptr;
1534 for (BasicBlock::iterator BBI = ThenBB->begin(),
1535 BBE = std::prev(ThenBB->end());
1536 BBI != BBE; ++BBI) {
1537 Instruction *I = &*BBI;
1539 if (isa<DbgInfoIntrinsic>(I))
1542 // Only speculatively execute a single instruction (not counting the
1543 // terminator) for now.
1545 if (SpeculationCost > 1)
1548 // Don't hoist the instruction if it's unsafe or expensive.
1549 if (!isSafeToSpeculativelyExecute(I) &&
1550 !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1551 I, BB, ThenBB, EndBB))))
1553 if (!SpeculatedStoreValue &&
1554 ComputeSpeculationCost(I, TTI) >
1555 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
1558 // Store the store speculation candidate.
1559 if (SpeculatedStoreValue)
1560 SpeculatedStore = cast<StoreInst>(I);
1562 // Do not hoist the instruction if any of its operands are defined but not
1563 // used in BB. The transformation will prevent the operand from
1564 // being sunk into the use block.
1565 for (User::op_iterator i = I->op_begin(), e = I->op_end();
1567 Instruction *OpI = dyn_cast<Instruction>(*i);
1568 if (!OpI || OpI->getParent() != BB ||
1569 OpI->mayHaveSideEffects())
1570 continue; // Not a candidate for sinking.
1572 ++SinkCandidateUseCounts[OpI];
1576 // Consider any sink candidates which are only used in CondBB as costs for
1577 // speculation. Note, while we iterate over a DenseMap here, we are summing
1578 // and so iteration order isn't significant.
1579 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I =
1580 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end();
1582 if (I->first->getNumUses() == I->second) {
1584 if (SpeculationCost > 1)
1588 // Check that the PHI nodes can be converted to selects.
1589 bool HaveRewritablePHIs = false;
1590 for (BasicBlock::iterator I = EndBB->begin();
1591 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1592 Value *OrigV = PN->getIncomingValueForBlock(BB);
1593 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1595 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1596 // Skip PHIs which are trivial.
1600 // Don't convert to selects if we could remove undefined behavior instead.
1601 if (passingValueIsAlwaysUndefined(OrigV, PN) ||
1602 passingValueIsAlwaysUndefined(ThenV, PN))
1605 HaveRewritablePHIs = true;
1606 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1607 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1608 if (!OrigCE && !ThenCE)
1609 continue; // Known safe and cheap.
1611 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
1612 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
1614 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
1615 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
1616 unsigned MaxCost = 2 * PHINodeFoldingThreshold *
1617 TargetTransformInfo::TCC_Basic;
1618 if (OrigCost + ThenCost > MaxCost)
1621 // Account for the cost of an unfolded ConstantExpr which could end up
1622 // getting expanded into Instructions.
1623 // FIXME: This doesn't account for how many operations are combined in the
1624 // constant expression.
1626 if (SpeculationCost > 1)
1630 // If there are no PHIs to process, bail early. This helps ensure idempotence
1632 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
1635 // If we get here, we can hoist the instruction and if-convert.
1636 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1638 // Insert a select of the value of the speculated store.
1639 if (SpeculatedStoreValue) {
1640 IRBuilder<true, NoFolder> Builder(BI);
1641 Value *TrueV = SpeculatedStore->getValueOperand();
1642 Value *FalseV = SpeculatedStoreValue;
1644 std::swap(TrueV, FalseV);
1645 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() +
1646 "." + FalseV->getName());
1647 SpeculatedStore->setOperand(0, S);
1650 // Metadata can be dependent on the condition we are hoisting above.
1651 // Conservatively strip all metadata on the instruction.
1652 for (auto &I: *ThenBB)
1653 I.dropUnknownNonDebugMetadata();
1655 // Hoist the instructions.
1656 BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
1657 ThenBB->begin(), std::prev(ThenBB->end()));
1659 // Insert selects and rewrite the PHI operands.
1660 IRBuilder<true, NoFolder> Builder(BI);
1661 for (BasicBlock::iterator I = EndBB->begin();
1662 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1663 unsigned OrigI = PN->getBasicBlockIndex(BB);
1664 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1665 Value *OrigV = PN->getIncomingValue(OrigI);
1666 Value *ThenV = PN->getIncomingValue(ThenI);
1668 // Skip PHIs which are trivial.
1672 // Create a select whose true value is the speculatively executed value and
1673 // false value is the preexisting value. Swap them if the branch
1674 // destinations were inverted.
1675 Value *TrueV = ThenV, *FalseV = OrigV;
1677 std::swap(TrueV, FalseV);
1678 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV,
1679 TrueV->getName() + "." + FalseV->getName());
1680 PN->setIncomingValue(OrigI, V);
1681 PN->setIncomingValue(ThenI, V);
1688 /// \returns True if this block contains a CallInst with the NoDuplicate
1690 static bool HasNoDuplicateCall(const BasicBlock *BB) {
1691 for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
1692 const CallInst *CI = dyn_cast<CallInst>(I);
1695 if (CI->cannotDuplicate())
1701 /// Return true if we can thread a branch across this block.
1702 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1703 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1706 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1707 if (isa<DbgInfoIntrinsic>(BBI))
1709 if (Size > 10) return false; // Don't clone large BB's.
1712 // We can only support instructions that do not define values that are
1713 // live outside of the current basic block.
1714 for (User *U : BBI->users()) {
1715 Instruction *UI = cast<Instruction>(U);
1716 if (UI->getParent() != BB || isa<PHINode>(UI)) return false;
1719 // Looks ok, continue checking.
1725 /// If we have a conditional branch on a PHI node value that is defined in the
1726 /// same block as the branch and if any PHI entries are constants, thread edges
1727 /// corresponding to that entry to be branches to their ultimate destination.
1728 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL) {
1729 BasicBlock *BB = BI->getParent();
1730 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1731 // NOTE: we currently cannot transform this case if the PHI node is used
1732 // outside of the block.
1733 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1736 // Degenerate case of a single entry PHI.
1737 if (PN->getNumIncomingValues() == 1) {
1738 FoldSingleEntryPHINodes(PN->getParent());
1742 // Now we know that this block has multiple preds and two succs.
1743 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1745 if (HasNoDuplicateCall(BB)) return false;
1747 // Okay, this is a simple enough basic block. See if any phi values are
1749 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1750 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1751 if (!CB || !CB->getType()->isIntegerTy(1)) continue;
1753 // Okay, we now know that all edges from PredBB should be revectored to
1754 // branch to RealDest.
1755 BasicBlock *PredBB = PN->getIncomingBlock(i);
1756 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1758 if (RealDest == BB) continue; // Skip self loops.
1759 // Skip if the predecessor's terminator is an indirect branch.
1760 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1762 // The dest block might have PHI nodes, other predecessors and other
1763 // difficult cases. Instead of being smart about this, just insert a new
1764 // block that jumps to the destination block, effectively splitting
1765 // the edge we are about to create.
1766 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1767 RealDest->getName()+".critedge",
1768 RealDest->getParent(), RealDest);
1769 BranchInst::Create(RealDest, EdgeBB);
1771 // Update PHI nodes.
1772 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1774 // BB may have instructions that are being threaded over. Clone these
1775 // instructions into EdgeBB. We know that there will be no uses of the
1776 // cloned instructions outside of EdgeBB.
1777 BasicBlock::iterator InsertPt = EdgeBB->begin();
1778 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1779 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1780 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1781 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1784 // Clone the instruction.
1785 Instruction *N = BBI->clone();
1786 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1788 // Update operands due to translation.
1789 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1791 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1792 if (PI != TranslateMap.end())
1796 // Check for trivial simplification.
1797 if (Value *V = SimplifyInstruction(N, DL)) {
1798 TranslateMap[&*BBI] = V;
1799 delete N; // Instruction folded away, don't need actual inst
1801 // Insert the new instruction into its new home.
1802 EdgeBB->getInstList().insert(InsertPt, N);
1803 if (!BBI->use_empty())
1804 TranslateMap[&*BBI] = N;
1808 // Loop over all of the edges from PredBB to BB, changing them to branch
1809 // to EdgeBB instead.
1810 TerminatorInst *PredBBTI = PredBB->getTerminator();
1811 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1812 if (PredBBTI->getSuccessor(i) == BB) {
1813 BB->removePredecessor(PredBB);
1814 PredBBTI->setSuccessor(i, EdgeBB);
1817 // Recurse, simplifying any other constants.
1818 return FoldCondBranchOnPHI(BI, DL) | true;
1824 /// Given a BB that starts with the specified two-entry PHI node,
1825 /// see if we can eliminate it.
1826 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI,
1827 const DataLayout &DL) {
1828 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1829 // statement", which has a very simple dominance structure. Basically, we
1830 // are trying to find the condition that is being branched on, which
1831 // subsequently causes this merge to happen. We really want control
1832 // dependence information for this check, but simplifycfg can't keep it up
1833 // to date, and this catches most of the cases we care about anyway.
1834 BasicBlock *BB = PN->getParent();
1835 BasicBlock *IfTrue, *IfFalse;
1836 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1838 // Don't bother if the branch will be constant folded trivially.
1839 isa<ConstantInt>(IfCond))
1842 // Okay, we found that we can merge this two-entry phi node into a select.
1843 // Doing so would require us to fold *all* two entry phi nodes in this block.
1844 // At some point this becomes non-profitable (particularly if the target
1845 // doesn't support cmov's). Only do this transformation if there are two or
1846 // fewer PHI nodes in this block.
1847 unsigned NumPhis = 0;
1848 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1852 // Loop over the PHI's seeing if we can promote them all to select
1853 // instructions. While we are at it, keep track of the instructions
1854 // that need to be moved to the dominating block.
1855 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1856 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1857 MaxCostVal1 = PHINodeFoldingThreshold;
1858 MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
1859 MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
1861 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1862 PHINode *PN = cast<PHINode>(II++);
1863 if (Value *V = SimplifyInstruction(PN, DL)) {
1864 PN->replaceAllUsesWith(V);
1865 PN->eraseFromParent();
1869 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1870 MaxCostVal0, TTI) ||
1871 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1876 // If we folded the first phi, PN dangles at this point. Refresh it. If
1877 // we ran out of PHIs then we simplified them all.
1878 PN = dyn_cast<PHINode>(BB->begin());
1879 if (!PN) return true;
1881 // Don't fold i1 branches on PHIs which contain binary operators. These can
1882 // often be turned into switches and other things.
1883 if (PN->getType()->isIntegerTy(1) &&
1884 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1885 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1886 isa<BinaryOperator>(IfCond)))
1889 // If we all PHI nodes are promotable, check to make sure that all
1890 // instructions in the predecessor blocks can be promoted as well. If
1891 // not, we won't be able to get rid of the control flow, so it's not
1892 // worth promoting to select instructions.
1893 BasicBlock *DomBlock = nullptr;
1894 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1895 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1896 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1899 DomBlock = *pred_begin(IfBlock1);
1900 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1901 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
1902 // This is not an aggressive instruction that we can promote.
1903 // Because of this, we won't be able to get rid of the control
1904 // flow, so the xform is not worth it.
1909 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1912 DomBlock = *pred_begin(IfBlock2);
1913 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1914 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
1915 // This is not an aggressive instruction that we can promote.
1916 // Because of this, we won't be able to get rid of the control
1917 // flow, so the xform is not worth it.
1922 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1923 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1925 // If we can still promote the PHI nodes after this gauntlet of tests,
1926 // do all of the PHI's now.
1927 Instruction *InsertPt = DomBlock->getTerminator();
1928 IRBuilder<true, NoFolder> Builder(InsertPt);
1930 // Move all 'aggressive' instructions, which are defined in the
1931 // conditional parts of the if's up to the dominating block.
1933 DomBlock->getInstList().splice(InsertPt->getIterator(),
1934 IfBlock1->getInstList(), IfBlock1->begin(),
1935 IfBlock1->getTerminator()->getIterator());
1937 DomBlock->getInstList().splice(InsertPt->getIterator(),
1938 IfBlock2->getInstList(), IfBlock2->begin(),
1939 IfBlock2->getTerminator()->getIterator());
1941 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1942 // Change the PHI node into a select instruction.
1943 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1944 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1947 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1948 PN->replaceAllUsesWith(NV);
1950 PN->eraseFromParent();
1953 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1954 // has been flattened. Change DomBlock to jump directly to our new block to
1955 // avoid other simplifycfg's kicking in on the diamond.
1956 TerminatorInst *OldTI = DomBlock->getTerminator();
1957 Builder.SetInsertPoint(OldTI);
1958 Builder.CreateBr(BB);
1959 OldTI->eraseFromParent();
1963 /// If we found a conditional branch that goes to two returning blocks,
1964 /// try to merge them together into one return,
1965 /// introducing a select if the return values disagree.
1966 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1967 IRBuilder<> &Builder) {
1968 assert(BI->isConditional() && "Must be a conditional branch");
1969 BasicBlock *TrueSucc = BI->getSuccessor(0);
1970 BasicBlock *FalseSucc = BI->getSuccessor(1);
1971 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1972 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1974 // Check to ensure both blocks are empty (just a return) or optionally empty
1975 // with PHI nodes. If there are other instructions, merging would cause extra
1976 // computation on one path or the other.
1977 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1979 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1982 Builder.SetInsertPoint(BI);
1983 // Okay, we found a branch that is going to two return nodes. If
1984 // there is no return value for this function, just change the
1985 // branch into a return.
1986 if (FalseRet->getNumOperands() == 0) {
1987 TrueSucc->removePredecessor(BI->getParent());
1988 FalseSucc->removePredecessor(BI->getParent());
1989 Builder.CreateRetVoid();
1990 EraseTerminatorInstAndDCECond(BI);
1994 // Otherwise, figure out what the true and false return values are
1995 // so we can insert a new select instruction.
1996 Value *TrueValue = TrueRet->getReturnValue();
1997 Value *FalseValue = FalseRet->getReturnValue();
1999 // Unwrap any PHI nodes in the return blocks.
2000 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2001 if (TVPN->getParent() == TrueSucc)
2002 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2003 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2004 if (FVPN->getParent() == FalseSucc)
2005 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2007 // In order for this transformation to be safe, we must be able to
2008 // unconditionally execute both operands to the return. This is
2009 // normally the case, but we could have a potentially-trapping
2010 // constant expression that prevents this transformation from being
2012 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2015 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2019 // Okay, we collected all the mapped values and checked them for sanity, and
2020 // defined to really do this transformation. First, update the CFG.
2021 TrueSucc->removePredecessor(BI->getParent());
2022 FalseSucc->removePredecessor(BI->getParent());
2024 // Insert select instructions where needed.
2025 Value *BrCond = BI->getCondition();
2027 // Insert a select if the results differ.
2028 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2029 } else if (isa<UndefValue>(TrueValue)) {
2030 TrueValue = FalseValue;
2032 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
2033 FalseValue, "retval");
2037 Value *RI = !TrueValue ?
2038 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2042 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2043 << "\n " << *BI << "NewRet = " << *RI
2044 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
2046 EraseTerminatorInstAndDCECond(BI);
2051 /// Given a conditional BranchInstruction, retrieve the probabilities of the
2052 /// branch taking each edge. Fills in the two APInt parameters and returns true,
2053 /// or returns false if no or invalid metadata was found.
2054 static bool ExtractBranchMetadata(BranchInst *BI,
2055 uint64_t &ProbTrue, uint64_t &ProbFalse) {
2056 assert(BI->isConditional() &&
2057 "Looking for probabilities on unconditional branch?");
2058 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
2059 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
2060 ConstantInt *CITrue =
2061 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
2062 ConstantInt *CIFalse =
2063 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
2064 if (!CITrue || !CIFalse) return false;
2065 ProbTrue = CITrue->getValue().getZExtValue();
2066 ProbFalse = CIFalse->getValue().getZExtValue();
2070 /// Return true if the given instruction is available
2071 /// in its predecessor block. If yes, the instruction will be removed.
2072 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
2073 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2075 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
2076 Instruction *PBI = &*I;
2077 // Check whether Inst and PBI generate the same value.
2078 if (Inst->isIdenticalTo(PBI)) {
2079 Inst->replaceAllUsesWith(PBI);
2080 Inst->eraseFromParent();
2087 /// If this basic block is simple enough, and if a predecessor branches to us
2088 /// and one of our successors, fold the block into the predecessor and use
2089 /// logical operations to pick the right destination.
2090 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2091 BasicBlock *BB = BI->getParent();
2093 Instruction *Cond = nullptr;
2094 if (BI->isConditional())
2095 Cond = dyn_cast<Instruction>(BI->getCondition());
2097 // For unconditional branch, check for a simple CFG pattern, where
2098 // BB has a single predecessor and BB's successor is also its predecessor's
2099 // successor. If such pattern exisits, check for CSE between BB and its
2101 if (BasicBlock *PB = BB->getSinglePredecessor())
2102 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2103 if (PBI->isConditional() &&
2104 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2105 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2106 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
2108 Instruction *Curr = &*I++;
2109 if (isa<CmpInst>(Curr)) {
2113 // Quit if we can't remove this instruction.
2114 if (!checkCSEInPredecessor(Curr, PB))
2123 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2124 Cond->getParent() != BB || !Cond->hasOneUse())
2127 // Make sure the instruction after the condition is the cond branch.
2128 BasicBlock::iterator CondIt = ++Cond->getIterator();
2130 // Ignore dbg intrinsics.
2131 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
2136 // Only allow this transformation if computing the condition doesn't involve
2137 // too many instructions and these involved instructions can be executed
2138 // unconditionally. We denote all involved instructions except the condition
2139 // as "bonus instructions", and only allow this transformation when the
2140 // number of the bonus instructions does not exceed a certain threshold.
2141 unsigned NumBonusInsts = 0;
2142 for (auto I = BB->begin(); Cond != I; ++I) {
2143 // Ignore dbg intrinsics.
2144 if (isa<DbgInfoIntrinsic>(I))
2146 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2148 // I has only one use and can be executed unconditionally.
2149 Instruction *User = dyn_cast<Instruction>(I->user_back());
2150 if (User == nullptr || User->getParent() != BB)
2152 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2153 // to use any other instruction, User must be an instruction between next(I)
2156 // Early exits once we reach the limit.
2157 if (NumBonusInsts > BonusInstThreshold)
2161 // Cond is known to be a compare or binary operator. Check to make sure that
2162 // neither operand is a potentially-trapping constant expression.
2163 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2166 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2170 // Finally, don't infinitely unroll conditional loops.
2171 BasicBlock *TrueDest = BI->getSuccessor(0);
2172 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2173 if (TrueDest == BB || FalseDest == BB)
2176 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2177 BasicBlock *PredBlock = *PI;
2178 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2180 // Check that we have two conditional branches. If there is a PHI node in
2181 // the common successor, verify that the same value flows in from both
2183 SmallVector<PHINode*, 4> PHIs;
2184 if (!PBI || PBI->isUnconditional() ||
2185 (BI->isConditional() &&
2186 !SafeToMergeTerminators(BI, PBI)) ||
2187 (!BI->isConditional() &&
2188 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2191 // Determine if the two branches share a common destination.
2192 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2193 bool InvertPredCond = false;
2195 if (BI->isConditional()) {
2196 if (PBI->getSuccessor(0) == TrueDest)
2197 Opc = Instruction::Or;
2198 else if (PBI->getSuccessor(1) == FalseDest)
2199 Opc = Instruction::And;
2200 else if (PBI->getSuccessor(0) == FalseDest)
2201 Opc = Instruction::And, InvertPredCond = true;
2202 else if (PBI->getSuccessor(1) == TrueDest)
2203 Opc = Instruction::Or, InvertPredCond = true;
2207 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2211 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2212 IRBuilder<> Builder(PBI);
2214 // If we need to invert the condition in the pred block to match, do so now.
2215 if (InvertPredCond) {
2216 Value *NewCond = PBI->getCondition();
2218 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2219 CmpInst *CI = cast<CmpInst>(NewCond);
2220 CI->setPredicate(CI->getInversePredicate());
2222 NewCond = Builder.CreateNot(NewCond,
2223 PBI->getCondition()->getName()+".not");
2226 PBI->setCondition(NewCond);
2227 PBI->swapSuccessors();
2230 // If we have bonus instructions, clone them into the predecessor block.
2231 // Note that there may be multiple predecessor blocks, so we cannot move
2232 // bonus instructions to a predecessor block.
2233 ValueToValueMapTy VMap; // maps original values to cloned values
2234 // We already make sure Cond is the last instruction before BI. Therefore,
2235 // all instructions before Cond other than DbgInfoIntrinsic are bonus
2237 for (auto BonusInst = BB->begin(); Cond != BonusInst; ++BonusInst) {
2238 if (isa<DbgInfoIntrinsic>(BonusInst))
2240 Instruction *NewBonusInst = BonusInst->clone();
2241 RemapInstruction(NewBonusInst, VMap,
2242 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2243 VMap[&*BonusInst] = NewBonusInst;
2245 // If we moved a load, we cannot any longer claim any knowledge about
2246 // its potential value. The previous information might have been valid
2247 // only given the branch precondition.
2248 // For an analogous reason, we must also drop all the metadata whose
2249 // semantics we don't understand.
2250 NewBonusInst->dropUnknownNonDebugMetadata();
2252 PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2253 NewBonusInst->takeName(&*BonusInst);
2254 BonusInst->setName(BonusInst->getName() + ".old");
2257 // Clone Cond into the predecessor basic block, and or/and the
2258 // two conditions together.
2259 Instruction *New = Cond->clone();
2260 RemapInstruction(New, VMap,
2261 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
2262 PredBlock->getInstList().insert(PBI->getIterator(), New);
2263 New->takeName(Cond);
2264 Cond->setName(New->getName() + ".old");
2266 if (BI->isConditional()) {
2267 Instruction *NewCond =
2268 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2270 PBI->setCondition(NewCond);
2272 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2273 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2275 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2277 SmallVector<uint64_t, 8> NewWeights;
2279 if (PBI->getSuccessor(0) == BB) {
2280 if (PredHasWeights && SuccHasWeights) {
2281 // PBI: br i1 %x, BB, FalseDest
2282 // BI: br i1 %y, TrueDest, FalseDest
2283 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2284 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2285 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2286 // TrueWeight for PBI * FalseWeight for BI.
2287 // We assume that total weights of a BranchInst can fit into 32 bits.
2288 // Therefore, we will not have overflow using 64-bit arithmetic.
2289 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2290 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2292 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2293 PBI->setSuccessor(0, TrueDest);
2295 if (PBI->getSuccessor(1) == BB) {
2296 if (PredHasWeights && SuccHasWeights) {
2297 // PBI: br i1 %x, TrueDest, BB
2298 // BI: br i1 %y, TrueDest, FalseDest
2299 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2300 // FalseWeight for PBI * TrueWeight for BI.
2301 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2302 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2303 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2304 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2306 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2307 PBI->setSuccessor(1, FalseDest);
2309 if (NewWeights.size() == 2) {
2310 // Halve the weights if any of them cannot fit in an uint32_t
2311 FitWeights(NewWeights);
2313 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2314 PBI->setMetadata(LLVMContext::MD_prof,
2315 MDBuilder(BI->getContext()).
2316 createBranchWeights(MDWeights));
2318 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2320 // Update PHI nodes in the common successors.
2321 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2322 ConstantInt *PBI_C = cast<ConstantInt>(
2323 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2324 assert(PBI_C->getType()->isIntegerTy(1));
2325 Instruction *MergedCond = nullptr;
2326 if (PBI->getSuccessor(0) == TrueDest) {
2327 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2328 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2329 // is false: !PBI_Cond and BI_Value
2330 Instruction *NotCond =
2331 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2334 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2339 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2340 PBI->getCondition(), MergedCond,
2343 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2344 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2345 // is false: PBI_Cond and BI_Value
2347 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2348 PBI->getCondition(), New,
2350 if (PBI_C->isOne()) {
2351 Instruction *NotCond =
2352 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2355 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2356 NotCond, MergedCond,
2361 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2364 // Change PBI from Conditional to Unconditional.
2365 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2366 EraseTerminatorInstAndDCECond(PBI);
2370 // TODO: If BB is reachable from all paths through PredBlock, then we
2371 // could replace PBI's branch probabilities with BI's.
2373 // Copy any debug value intrinsics into the end of PredBlock.
2374 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2375 if (isa<DbgInfoIntrinsic>(*I))
2376 I->clone()->insertBefore(PBI);
2383 // If there is only one store in BB1 and BB2, return it, otherwise return
2385 static StoreInst *findUniqueStoreInBlocks(BasicBlock *BB1, BasicBlock *BB2) {
2386 StoreInst *S = nullptr;
2387 for (auto *BB : {BB1, BB2}) {
2391 if (auto *SI = dyn_cast<StoreInst>(&I)) {
2393 // Multiple stores seen.
2402 static Value *ensureValueAvailableInSuccessor(Value *V, BasicBlock *BB,
2403 Value *AlternativeV = nullptr) {
2404 // PHI is going to be a PHI node that allows the value V that is defined in
2405 // BB to be referenced in BB's only successor.
2407 // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2408 // doesn't matter to us what the other operand is (it'll never get used). We
2409 // could just create a new PHI with an undef incoming value, but that could
2410 // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2411 // other PHI. So here we directly look for some PHI in BB's successor with V
2412 // as an incoming operand. If we find one, we use it, else we create a new
2415 // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2416 // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2417 // where OtherBB is the single other predecessor of BB's only successor.
2418 PHINode *PHI = nullptr;
2419 BasicBlock *Succ = BB->getSingleSuccessor();
2421 for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2422 if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2423 PHI = cast<PHINode>(I);
2427 assert(std::distance(pred_begin(Succ), pred_end(Succ)) == 2);
2428 auto PredI = pred_begin(Succ);
2429 BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2430 if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2437 // If V is not an instruction defined in BB, just return it.
2438 if (!AlternativeV &&
2439 (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2442 PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2443 PHI->addIncoming(V, BB);
2444 for (BasicBlock *PredBB : predecessors(Succ))
2446 PHI->addIncoming(AlternativeV ? AlternativeV : UndefValue::get(V->getType()),
2451 static bool mergeConditionalStoreToAddress(BasicBlock *PTB, BasicBlock *PFB,
2452 BasicBlock *QTB, BasicBlock *QFB,
2453 BasicBlock *PostBB, Value *Address,
2454 bool InvertPCond, bool InvertQCond) {
2455 auto IsaBitcastOfPointerType = [](const Instruction &I) {
2456 return Operator::getOpcode(&I) == Instruction::BitCast &&
2457 I.getType()->isPointerTy();
2460 // If we're not in aggressive mode, we only optimize if we have some
2461 // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2462 auto IsWorthwhile = [&](BasicBlock *BB) {
2465 // Heuristic: if the block can be if-converted/phi-folded and the
2466 // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2467 // thread this store.
2469 for (auto &I : *BB) {
2470 // Cheap instructions viable for folding.
2471 if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2474 // Free instructions.
2475 else if (isa<TerminatorInst>(I) || isa<DbgInfoIntrinsic>(I) ||
2476 IsaBitcastOfPointerType(I))
2481 return N <= PHINodeFoldingThreshold;
2484 if (!MergeCondStoresAggressively && (!IsWorthwhile(PTB) ||
2485 !IsWorthwhile(PFB) ||
2486 !IsWorthwhile(QTB) ||
2487 !IsWorthwhile(QFB)))
2490 // For every pointer, there must be exactly two stores, one coming from
2491 // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2492 // store (to any address) in PTB,PFB or QTB,QFB.
2493 // FIXME: We could relax this restriction with a bit more work and performance
2495 StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2496 StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2497 if (!PStore || !QStore)
2500 // Now check the stores are compatible.
2501 if (!QStore->isUnordered() || !PStore->isUnordered())
2504 // Check that sinking the store won't cause program behavior changes. Sinking
2505 // the store out of the Q blocks won't change any behavior as we're sinking
2506 // from a block to its unconditional successor. But we're moving a store from
2507 // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2508 // So we need to check that there are no aliasing loads or stores in
2509 // QBI, QTB and QFB. We also need to check there are no conflicting memory
2510 // operations between PStore and the end of its parent block.
2512 // The ideal way to do this is to query AliasAnalysis, but we don't
2513 // preserve AA currently so that is dangerous. Be super safe and just
2514 // check there are no other memory operations at all.
2515 for (auto &I : *QFB->getSinglePredecessor())
2516 if (I.mayReadOrWriteMemory())
2518 for (auto &I : *QFB)
2519 if (&I != QStore && I.mayReadOrWriteMemory())
2522 for (auto &I : *QTB)
2523 if (&I != QStore && I.mayReadOrWriteMemory())
2525 for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2527 if (&*I != PStore && I->mayReadOrWriteMemory())
2530 // OK, we're going to sink the stores to PostBB. The store has to be
2531 // conditional though, so first create the predicate.
2532 Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
2534 Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
2537 Value *PPHI = ensureValueAvailableInSuccessor(PStore->getValueOperand(),
2538 PStore->getParent());
2539 Value *QPHI = ensureValueAvailableInSuccessor(QStore->getValueOperand(),
2540 QStore->getParent(), PPHI);
2542 IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
2544 Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
2545 Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
2548 PPred = QB.CreateNot(PPred);
2550 QPred = QB.CreateNot(QPred);
2551 Value *CombinedPred = QB.CreateOr(PPred, QPred);
2554 SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
2555 QB.SetInsertPoint(T);
2556 StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
2558 PStore->getAAMetadata(AAMD, /*Merge=*/false);
2559 PStore->getAAMetadata(AAMD, /*Merge=*/true);
2560 SI->setAAMetadata(AAMD);
2562 QStore->eraseFromParent();
2563 PStore->eraseFromParent();
2568 static bool mergeConditionalStores(BranchInst *PBI, BranchInst *QBI) {
2569 // The intention here is to find diamonds or triangles (see below) where each
2570 // conditional block contains a store to the same address. Both of these
2571 // stores are conditional, so they can't be unconditionally sunk. But it may
2572 // be profitable to speculatively sink the stores into one merged store at the
2573 // end, and predicate the merged store on the union of the two conditions of
2576 // This can reduce the number of stores executed if both of the conditions are
2577 // true, and can allow the blocks to become small enough to be if-converted.
2578 // This optimization will also chain, so that ladders of test-and-set
2579 // sequences can be if-converted away.
2581 // We only deal with simple diamonds or triangles:
2583 // PBI or PBI or a combination of the two
2593 // We model triangles as a type of diamond with a nullptr "true" block.
2594 // Triangles are canonicalized so that the fallthrough edge is represented by
2595 // a true condition, as in the diagram above.
2597 BasicBlock *PTB = PBI->getSuccessor(0);
2598 BasicBlock *PFB = PBI->getSuccessor(1);
2599 BasicBlock *QTB = QBI->getSuccessor(0);
2600 BasicBlock *QFB = QBI->getSuccessor(1);
2601 BasicBlock *PostBB = QFB->getSingleSuccessor();
2603 bool InvertPCond = false, InvertQCond = false;
2604 // Canonicalize fallthroughs to the true branches.
2605 if (PFB == QBI->getParent()) {
2606 std::swap(PFB, PTB);
2609 if (QFB == PostBB) {
2610 std::swap(QFB, QTB);
2614 // From this point on we can assume PTB or QTB may be fallthroughs but PFB
2615 // and QFB may not. Model fallthroughs as a nullptr block.
2616 if (PTB == QBI->getParent())
2621 // Legality bailouts. We must have at least the non-fallthrough blocks and
2622 // the post-dominating block, and the non-fallthroughs must only have one
2624 auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
2625 return BB->getSinglePredecessor() == P &&
2626 BB->getSingleSuccessor() == S;
2629 !HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
2630 !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
2632 if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
2633 (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
2635 if (PostBB->getNumUses() != 2 || QBI->getParent()->getNumUses() != 2)
2638 // OK, this is a sequence of two diamonds or triangles.
2639 // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
2640 SmallPtrSet<Value *,4> PStoreAddresses, QStoreAddresses;
2641 for (auto *BB : {PTB, PFB}) {
2645 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
2646 PStoreAddresses.insert(SI->getPointerOperand());
2648 for (auto *BB : {QTB, QFB}) {
2652 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
2653 QStoreAddresses.insert(SI->getPointerOperand());
2656 set_intersect(PStoreAddresses, QStoreAddresses);
2657 // set_intersect mutates PStoreAddresses in place. Rename it here to make it
2658 // clear what it contains.
2659 auto &CommonAddresses = PStoreAddresses;
2661 bool Changed = false;
2662 for (auto *Address : CommonAddresses)
2663 Changed |= mergeConditionalStoreToAddress(
2664 PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond);
2668 /// If we have a conditional branch as a predecessor of another block,
2669 /// this function tries to simplify it. We know
2670 /// that PBI and BI are both conditional branches, and BI is in one of the
2671 /// successor blocks of PBI - PBI branches to BI.
2672 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
2673 const DataLayout &DL) {
2674 assert(PBI->isConditional() && BI->isConditional());
2675 BasicBlock *BB = BI->getParent();
2677 // If this block ends with a branch instruction, and if there is a
2678 // predecessor that ends on a branch of the same condition, make
2679 // this conditional branch redundant.
2680 if (PBI->getCondition() == BI->getCondition() &&
2681 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2682 // Okay, the outcome of this conditional branch is statically
2683 // knowable. If this block had a single pred, handle specially.
2684 if (BB->getSinglePredecessor()) {
2685 // Turn this into a branch on constant.
2686 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2687 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2689 return true; // Nuke the branch on constant.
2692 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2693 // in the constant and simplify the block result. Subsequent passes of
2694 // simplifycfg will thread the block.
2695 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2696 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2697 PHINode *NewPN = PHINode::Create(
2698 Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
2699 BI->getCondition()->getName() + ".pr", &BB->front());
2700 // Okay, we're going to insert the PHI node. Since PBI is not the only
2701 // predecessor, compute the PHI'd conditional value for all of the preds.
2702 // Any predecessor where the condition is not computable we keep symbolic.
2703 for (pred_iterator PI = PB; PI != PE; ++PI) {
2704 BasicBlock *P = *PI;
2705 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2706 PBI != BI && PBI->isConditional() &&
2707 PBI->getCondition() == BI->getCondition() &&
2708 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2709 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2710 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2713 NewPN->addIncoming(BI->getCondition(), P);
2717 BI->setCondition(NewPN);
2722 if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2726 // If BI is reached from the true path of PBI and PBI's condition implies
2727 // BI's condition, we know the direction of the BI branch.
2728 if (PBI->getSuccessor(0) == BI->getParent() &&
2729 isImpliedCondition(PBI->getCondition(), BI->getCondition(), DL) &&
2730 PBI->getSuccessor(0) != PBI->getSuccessor(1) &&
2731 BB->getSinglePredecessor()) {
2732 // Turn this into a branch on constant.
2733 auto *OldCond = BI->getCondition();
2734 BI->setCondition(ConstantInt::getTrue(BB->getContext()));
2735 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
2736 return true; // Nuke the branch on constant.
2739 // If both branches are conditional and both contain stores to the same
2740 // address, remove the stores from the conditionals and create a conditional
2741 // merged store at the end.
2742 if (MergeCondStores && mergeConditionalStores(PBI, BI))
2745 // If this is a conditional branch in an empty block, and if any
2746 // predecessors are a conditional branch to one of our destinations,
2747 // fold the conditions into logical ops and one cond br.
2748 BasicBlock::iterator BBI = BB->begin();
2749 // Ignore dbg intrinsics.
2750 while (isa<DbgInfoIntrinsic>(BBI))
2756 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2758 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2759 PBIOp = 0, BIOp = 1;
2760 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2761 PBIOp = 1, BIOp = 0;
2762 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2767 // Check to make sure that the other destination of this branch
2768 // isn't BB itself. If so, this is an infinite loop that will
2769 // keep getting unwound.
2770 if (PBI->getSuccessor(PBIOp) == BB)
2773 // Do not perform this transformation if it would require
2774 // insertion of a large number of select instructions. For targets
2775 // without predication/cmovs, this is a big pessimization.
2777 // Also do not perform this transformation if any phi node in the common
2778 // destination block can trap when reached by BB or PBB (PR17073). In that
2779 // case, it would be unsafe to hoist the operation into a select instruction.
2781 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2782 unsigned NumPhis = 0;
2783 for (BasicBlock::iterator II = CommonDest->begin();
2784 isa<PHINode>(II); ++II, ++NumPhis) {
2785 if (NumPhis > 2) // Disable this xform.
2788 PHINode *PN = cast<PHINode>(II);
2789 Value *BIV = PN->getIncomingValueForBlock(BB);
2790 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
2794 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2795 Value *PBIV = PN->getIncomingValue(PBBIdx);
2796 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
2801 // Finally, if everything is ok, fold the branches to logical ops.
2802 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2804 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2805 << "AND: " << *BI->getParent());
2808 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2809 // branch in it, where one edge (OtherDest) goes back to itself but the other
2810 // exits. We don't *know* that the program avoids the infinite loop
2811 // (even though that seems likely). If we do this xform naively, we'll end up
2812 // recursively unpeeling the loop. Since we know that (after the xform is
2813 // done) that the block *is* infinite if reached, we just make it an obviously
2814 // infinite loop with no cond branch.
2815 if (OtherDest == BB) {
2816 // Insert it at the end of the function, because it's either code,
2817 // or it won't matter if it's hot. :)
2818 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2819 "infloop", BB->getParent());
2820 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2821 OtherDest = InfLoopBlock;
2824 DEBUG(dbgs() << *PBI->getParent()->getParent());
2826 // BI may have other predecessors. Because of this, we leave
2827 // it alone, but modify PBI.
2829 // Make sure we get to CommonDest on True&True directions.
2830 Value *PBICond = PBI->getCondition();
2831 IRBuilder<true, NoFolder> Builder(PBI);
2833 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2835 Value *BICond = BI->getCondition();
2837 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2839 // Merge the conditions.
2840 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2842 // Modify PBI to branch on the new condition to the new dests.
2843 PBI->setCondition(Cond);
2844 PBI->setSuccessor(0, CommonDest);
2845 PBI->setSuccessor(1, OtherDest);
2847 // Update branch weight for PBI.
2848 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2849 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2851 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2853 if (PredHasWeights && SuccHasWeights) {
2854 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2855 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2856 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2857 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2858 // The weight to CommonDest should be PredCommon * SuccTotal +
2859 // PredOther * SuccCommon.
2860 // The weight to OtherDest should be PredOther * SuccOther.
2861 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
2862 PredOther * SuccCommon,
2863 PredOther * SuccOther};
2864 // Halve the weights if any of them cannot fit in an uint32_t
2865 FitWeights(NewWeights);
2867 PBI->setMetadata(LLVMContext::MD_prof,
2868 MDBuilder(BI->getContext())
2869 .createBranchWeights(NewWeights[0], NewWeights[1]));
2872 // OtherDest may have phi nodes. If so, add an entry from PBI's
2873 // block that are identical to the entries for BI's block.
2874 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2876 // We know that the CommonDest already had an edge from PBI to
2877 // it. If it has PHIs though, the PHIs may have different
2878 // entries for BB and PBI's BB. If so, insert a select to make
2881 for (BasicBlock::iterator II = CommonDest->begin();
2882 (PN = dyn_cast<PHINode>(II)); ++II) {
2883 Value *BIV = PN->getIncomingValueForBlock(BB);
2884 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2885 Value *PBIV = PN->getIncomingValue(PBBIdx);
2887 // Insert a select in PBI to pick the right value.
2888 Value *NV = cast<SelectInst>
2889 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2890 PN->setIncomingValue(PBBIdx, NV);
2894 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2895 DEBUG(dbgs() << *PBI->getParent()->getParent());
2897 // This basic block is probably dead. We know it has at least
2898 // one fewer predecessor.
2902 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
2903 // true or to FalseBB if Cond is false.
2904 // Takes care of updating the successors and removing the old terminator.
2905 // Also makes sure not to introduce new successors by assuming that edges to
2906 // non-successor TrueBBs and FalseBBs aren't reachable.
2907 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2908 BasicBlock *TrueBB, BasicBlock *FalseBB,
2909 uint32_t TrueWeight,
2910 uint32_t FalseWeight){
2911 // Remove any superfluous successor edges from the CFG.
2912 // First, figure out which successors to preserve.
2913 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2915 BasicBlock *KeepEdge1 = TrueBB;
2916 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
2918 // Then remove the rest.
2919 for (BasicBlock *Succ : OldTerm->successors()) {
2920 // Make sure only to keep exactly one copy of each edge.
2921 if (Succ == KeepEdge1)
2922 KeepEdge1 = nullptr;
2923 else if (Succ == KeepEdge2)
2924 KeepEdge2 = nullptr;
2926 Succ->removePredecessor(OldTerm->getParent(),
2927 /*DontDeleteUselessPHIs=*/true);
2930 IRBuilder<> Builder(OldTerm);
2931 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2933 // Insert an appropriate new terminator.
2934 if (!KeepEdge1 && !KeepEdge2) {
2935 if (TrueBB == FalseBB)
2936 // We were only looking for one successor, and it was present.
2937 // Create an unconditional branch to it.
2938 Builder.CreateBr(TrueBB);
2940 // We found both of the successors we were looking for.
2941 // Create a conditional branch sharing the condition of the select.
2942 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2943 if (TrueWeight != FalseWeight)
2944 NewBI->setMetadata(LLVMContext::MD_prof,
2945 MDBuilder(OldTerm->getContext()).
2946 createBranchWeights(TrueWeight, FalseWeight));
2948 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2949 // Neither of the selected blocks were successors, so this
2950 // terminator must be unreachable.
2951 new UnreachableInst(OldTerm->getContext(), OldTerm);
2953 // One of the selected values was a successor, but the other wasn't.
2954 // Insert an unconditional branch to the one that was found;
2955 // the edge to the one that wasn't must be unreachable.
2957 // Only TrueBB was found.
2958 Builder.CreateBr(TrueBB);
2960 // Only FalseBB was found.
2961 Builder.CreateBr(FalseBB);
2964 EraseTerminatorInstAndDCECond(OldTerm);
2969 // (switch (select cond, X, Y)) on constant X, Y
2970 // with a branch - conditional if X and Y lead to distinct BBs,
2971 // unconditional otherwise.
2972 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2973 // Check for constant integer values in the select.
2974 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2975 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2976 if (!TrueVal || !FalseVal)
2979 // Find the relevant condition and destinations.
2980 Value *Condition = Select->getCondition();
2981 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2982 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2984 // Get weight for TrueBB and FalseBB.
2985 uint32_t TrueWeight = 0, FalseWeight = 0;
2986 SmallVector<uint64_t, 8> Weights;
2987 bool HasWeights = HasBranchWeights(SI);
2989 GetBranchWeights(SI, Weights);
2990 if (Weights.size() == 1 + SI->getNumCases()) {
2991 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2992 getSuccessorIndex()];
2993 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2994 getSuccessorIndex()];
2998 // Perform the actual simplification.
2999 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
3000 TrueWeight, FalseWeight);
3004 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3005 // blockaddress(@fn, BlockB)))
3007 // (br cond, BlockA, BlockB).
3008 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
3009 // Check that both operands of the select are block addresses.
3010 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
3011 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
3015 // Extract the actual blocks.
3016 BasicBlock *TrueBB = TBA->getBasicBlock();
3017 BasicBlock *FalseBB = FBA->getBasicBlock();
3019 // Perform the actual simplification.
3020 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
3024 /// This is called when we find an icmp instruction
3025 /// (a seteq/setne with a constant) as the only instruction in a
3026 /// block that ends with an uncond branch. We are looking for a very specific
3027 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3028 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3029 /// default value goes to an uncond block with a seteq in it, we get something
3032 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3034 /// %tmp = icmp eq i8 %A, 92
3037 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3039 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3040 /// the PHI, merging the third icmp into the switch.
3041 static bool TryToSimplifyUncondBranchWithICmpInIt(
3042 ICmpInst *ICI, IRBuilder<> &Builder, const DataLayout &DL,
3043 const TargetTransformInfo &TTI, unsigned BonusInstThreshold,
3044 AssumptionCache *AC) {
3045 BasicBlock *BB = ICI->getParent();
3047 // If the block has any PHIs in it or the icmp has multiple uses, it is too
3049 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
3051 Value *V = ICI->getOperand(0);
3052 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3054 // The pattern we're looking for is where our only predecessor is a switch on
3055 // 'V' and this block is the default case for the switch. In this case we can
3056 // fold the compared value into the switch to simplify things.
3057 BasicBlock *Pred = BB->getSinglePredecessor();
3058 if (!Pred || !isa<SwitchInst>(Pred->getTerminator())) return false;
3060 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3061 if (SI->getCondition() != V)
3064 // If BB is reachable on a non-default case, then we simply know the value of
3065 // V in this block. Substitute it and constant fold the icmp instruction
3067 if (SI->getDefaultDest() != BB) {
3068 ConstantInt *VVal = SI->findCaseDest(BB);
3069 assert(VVal && "Should have a unique destination value");
3070 ICI->setOperand(0, VVal);
3072 if (Value *V = SimplifyInstruction(ICI, DL)) {
3073 ICI->replaceAllUsesWith(V);
3074 ICI->eraseFromParent();
3076 // BB is now empty, so it is likely to simplify away.
3077 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3080 // Ok, the block is reachable from the default dest. If the constant we're
3081 // comparing exists in one of the other edges, then we can constant fold ICI
3083 if (SI->findCaseValue(Cst) != SI->case_default()) {
3085 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3086 V = ConstantInt::getFalse(BB->getContext());
3088 V = ConstantInt::getTrue(BB->getContext());
3090 ICI->replaceAllUsesWith(V);
3091 ICI->eraseFromParent();
3092 // BB is now empty, so it is likely to simplify away.
3093 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3096 // The use of the icmp has to be in the 'end' block, by the only PHI node in
3098 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3099 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3100 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3101 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3104 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3106 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3107 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3109 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3110 std::swap(DefaultCst, NewCst);
3112 // Replace ICI (which is used by the PHI for the default value) with true or
3113 // false depending on if it is EQ or NE.
3114 ICI->replaceAllUsesWith(DefaultCst);
3115 ICI->eraseFromParent();
3117 // Okay, the switch goes to this block on a default value. Add an edge from
3118 // the switch to the merge point on the compared value.
3119 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
3120 BB->getParent(), BB);
3121 SmallVector<uint64_t, 8> Weights;
3122 bool HasWeights = HasBranchWeights(SI);
3124 GetBranchWeights(SI, Weights);
3125 if (Weights.size() == 1 + SI->getNumCases()) {
3126 // Split weight for default case to case for "Cst".
3127 Weights[0] = (Weights[0]+1) >> 1;
3128 Weights.push_back(Weights[0]);
3130 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3131 SI->setMetadata(LLVMContext::MD_prof,
3132 MDBuilder(SI->getContext()).
3133 createBranchWeights(MDWeights));
3136 SI->addCase(Cst, NewBB);
3138 // NewBB branches to the phi block, add the uncond branch and the phi entry.
3139 Builder.SetInsertPoint(NewBB);
3140 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3141 Builder.CreateBr(SuccBlock);
3142 PHIUse->addIncoming(NewCst, NewBB);
3146 /// The specified branch is a conditional branch.
3147 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3148 /// fold it into a switch instruction if so.
3149 static bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
3150 const DataLayout &DL) {
3151 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3152 if (!Cond) return false;
3154 // Change br (X == 0 | X == 1), T, F into a switch instruction.
3155 // If this is a bunch of seteq's or'd together, or if it's a bunch of
3156 // 'setne's and'ed together, collect them.
3158 // Try to gather values from a chain of and/or to be turned into a switch
3159 ConstantComparesGatherer ConstantCompare(Cond, DL);
3160 // Unpack the result
3161 SmallVectorImpl<ConstantInt*> &Values = ConstantCompare.Vals;
3162 Value *CompVal = ConstantCompare.CompValue;
3163 unsigned UsedICmps = ConstantCompare.UsedICmps;
3164 Value *ExtraCase = ConstantCompare.Extra;
3166 // If we didn't have a multiply compared value, fail.
3167 if (!CompVal) return false;
3169 // Avoid turning single icmps into a switch.
3173 bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3175 // There might be duplicate constants in the list, which the switch
3176 // instruction can't handle, remove them now.
3177 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3178 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3180 // If Extra was used, we require at least two switch values to do the
3181 // transformation. A switch with one value is just a conditional branch.
3182 if (ExtraCase && Values.size() < 2) return false;
3184 // TODO: Preserve branch weight metadata, similarly to how
3185 // FoldValueComparisonIntoPredecessors preserves it.
3187 // Figure out which block is which destination.
3188 BasicBlock *DefaultBB = BI->getSuccessor(1);
3189 BasicBlock *EdgeBB = BI->getSuccessor(0);
3190 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
3192 BasicBlock *BB = BI->getParent();
3194 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3195 << " cases into SWITCH. BB is:\n" << *BB);
3197 // If there are any extra values that couldn't be folded into the switch
3198 // then we evaluate them with an explicit branch first. Split the block
3199 // right before the condbr to handle it.
3202 BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3203 // Remove the uncond branch added to the old block.
3204 TerminatorInst *OldTI = BB->getTerminator();
3205 Builder.SetInsertPoint(OldTI);
3208 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3210 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3212 OldTI->eraseFromParent();
3214 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3215 // for the edge we just added.
3216 AddPredecessorToBlock(EdgeBB, BB, NewBB);
3218 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3219 << "\nEXTRABB = " << *BB);
3223 Builder.SetInsertPoint(BI);
3224 // Convert pointer to int before we switch.
3225 if (CompVal->getType()->isPointerTy()) {
3226 CompVal = Builder.CreatePtrToInt(
3227 CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3230 // Create the new switch instruction now.
3231 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3233 // Add all of the 'cases' to the switch instruction.
3234 for (unsigned i = 0, e = Values.size(); i != e; ++i)
3235 New->addCase(Values[i], EdgeBB);
3237 // We added edges from PI to the EdgeBB. As such, if there were any
3238 // PHI nodes in EdgeBB, they need entries to be added corresponding to
3239 // the number of edges added.
3240 for (BasicBlock::iterator BBI = EdgeBB->begin();
3241 isa<PHINode>(BBI); ++BBI) {
3242 PHINode *PN = cast<PHINode>(BBI);
3243 Value *InVal = PN->getIncomingValueForBlock(BB);
3244 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
3245 PN->addIncoming(InVal, BB);
3248 // Erase the old branch instruction.
3249 EraseTerminatorInstAndDCECond(BI);
3251 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3255 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3256 if (isa<PHINode>(RI->getValue()))
3257 return SimplifyCommonResume(RI);
3258 else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3259 RI->getValue() == RI->getParent()->getFirstNonPHI())
3260 // The resume must unwind the exception that caused control to branch here.
3261 return SimplifySingleResume(RI);
3266 // Simplify resume that is shared by several landing pads (phi of landing pad).
3267 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3268 BasicBlock *BB = RI->getParent();
3270 // Check that there are no other instructions except for debug intrinsics
3271 // between the phi of landing pads (RI->getValue()) and resume instruction.
3272 BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3273 E = RI->getIterator();
3275 if (!isa<DbgInfoIntrinsic>(I))
3278 SmallSet<BasicBlock *, 4> TrivialUnwindBlocks;
3279 auto *PhiLPInst = cast<PHINode>(RI->getValue());
3281 // Check incoming blocks to see if any of them are trivial.
3282 for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues();
3283 Idx != End; Idx++) {
3284 auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3285 auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3287 // If the block has other successors, we can not delete it because
3288 // it has other dependents.
3289 if (IncomingBB->getUniqueSuccessor() != BB)
3293 dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3294 // Not the landing pad that caused the control to branch here.
3295 if (IncomingValue != LandingPad)
3298 bool isTrivial = true;
3300 I = IncomingBB->getFirstNonPHI()->getIterator();
3301 E = IncomingBB->getTerminator()->getIterator();
3303 if (!isa<DbgInfoIntrinsic>(I)) {
3309 TrivialUnwindBlocks.insert(IncomingBB);
3312 // If no trivial unwind blocks, don't do any simplifications.
3313 if (TrivialUnwindBlocks.empty()) return false;
3315 // Turn all invokes that unwind here into calls.
3316 for (auto *TrivialBB : TrivialUnwindBlocks) {
3317 // Blocks that will be simplified should be removed from the phi node.
3318 // Note there could be multiple edges to the resume block, and we need
3319 // to remove them all.
3320 while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3321 BB->removePredecessor(TrivialBB, true);
3323 for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3325 BasicBlock *Pred = *PI++;
3326 removeUnwindEdge(Pred);
3329 // In each SimplifyCFG run, only the current processed block can be erased.
3330 // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3331 // of erasing TrivialBB, we only remove the branch to the common resume
3332 // block so that we can later erase the resume block since it has no
3334 TrivialBB->getTerminator()->eraseFromParent();
3335 new UnreachableInst(RI->getContext(), TrivialBB);
3338 // Delete the resume block if all its predecessors have been removed.
3340 BB->eraseFromParent();
3342 return !TrivialUnwindBlocks.empty();
3345 // Simplify resume that is only used by a single (non-phi) landing pad.
3346 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3347 BasicBlock *BB = RI->getParent();
3348 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
3349 assert (RI->getValue() == LPInst &&
3350 "Resume must unwind the exception that caused control to here");
3352 // Check that there are no other instructions except for debug intrinsics.
3353 BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3355 if (!isa<DbgInfoIntrinsic>(I))
3358 // Turn all invokes that unwind here into calls and delete the basic block.
3359 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3360 BasicBlock *Pred = *PI++;
3361 removeUnwindEdge(Pred);
3364 // The landingpad is now unreachable. Zap it.
3365 BB->eraseFromParent();
3369 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
3370 // If this is a trivial cleanup pad that executes no instructions, it can be
3371 // eliminated. If the cleanup pad continues to the caller, any predecessor
3372 // that is an EH pad will be updated to continue to the caller and any
3373 // predecessor that terminates with an invoke instruction will have its invoke
3374 // instruction converted to a call instruction. If the cleanup pad being
3375 // simplified does not continue to the caller, each predecessor will be
3376 // updated to continue to the unwind destination of the cleanup pad being
3378 BasicBlock *BB = RI->getParent();
3379 CleanupPadInst *CPInst = RI->getCleanupPad();
3380 if (CPInst->getParent() != BB)
3381 // This isn't an empty cleanup.
3384 // Check that there are no other instructions except for debug intrinsics.
3385 BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3387 if (!isa<DbgInfoIntrinsic>(I))
3390 // If the cleanup return we are simplifying unwinds to the caller, this will
3391 // set UnwindDest to nullptr.
3392 BasicBlock *UnwindDest = RI->getUnwindDest();
3393 Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3395 // We're about to remove BB from the control flow. Before we do, sink any
3396 // PHINodes into the unwind destination. Doing this before changing the
3397 // control flow avoids some potentially slow checks, since we can currently
3398 // be certain that UnwindDest and BB have no common predecessors (since they
3399 // are both EH pads).
3401 // First, go through the PHI nodes in UnwindDest and update any nodes that
3402 // reference the block we are removing
3403 for (BasicBlock::iterator I = UnwindDest->begin(),
3404 IE = DestEHPad->getIterator();
3406 PHINode *DestPN = cast<PHINode>(I);
3408 int Idx = DestPN->getBasicBlockIndex(BB);
3409 // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3411 // This PHI node has an incoming value that corresponds to a control
3412 // path through the cleanup pad we are removing. If the incoming
3413 // value is in the cleanup pad, it must be a PHINode (because we
3414 // verified above that the block is otherwise empty). Otherwise, the
3415 // value is either a constant or a value that dominates the cleanup
3416 // pad being removed.
3418 // Because BB and UnwindDest are both EH pads, all of their
3419 // predecessors must unwind to these blocks, and since no instruction
3420 // can have multiple unwind destinations, there will be no overlap in
3421 // incoming blocks between SrcPN and DestPN.
3422 Value *SrcVal = DestPN->getIncomingValue(Idx);
3423 PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3425 // Remove the entry for the block we are deleting.
3426 DestPN->removeIncomingValue(Idx, false);
3428 if (SrcPN && SrcPN->getParent() == BB) {
3429 // If the incoming value was a PHI node in the cleanup pad we are
3430 // removing, we need to merge that PHI node's incoming values into
3432 for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3433 SrcIdx != SrcE; ++SrcIdx) {
3434 DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3435 SrcPN->getIncomingBlock(SrcIdx));
3438 // Otherwise, the incoming value came from above BB and
3439 // so we can just reuse it. We must associate all of BB's
3440 // predecessors with this value.
3441 for (auto *pred : predecessors(BB)) {
3442 DestPN->addIncoming(SrcVal, pred);
3447 // Sink any remaining PHI nodes directly into UnwindDest.
3448 Instruction *InsertPt = DestEHPad;
3449 for (BasicBlock::iterator I = BB->begin(),
3450 IE = BB->getFirstNonPHI()->getIterator();
3452 // The iterator must be incremented here because the instructions are
3453 // being moved to another block.
3454 PHINode *PN = cast<PHINode>(I++);
3455 if (PN->use_empty())
3456 // If the PHI node has no uses, just leave it. It will be erased
3457 // when we erase BB below.
3460 // Otherwise, sink this PHI node into UnwindDest.
3461 // Any predecessors to UnwindDest which are not already represented
3462 // must be back edges which inherit the value from the path through
3463 // BB. In this case, the PHI value must reference itself.
3464 for (auto *pred : predecessors(UnwindDest))
3466 PN->addIncoming(PN, pred);
3467 PN->moveBefore(InsertPt);
3471 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3472 // The iterator must be updated here because we are removing this pred.
3473 BasicBlock *PredBB = *PI++;
3474 if (UnwindDest == nullptr) {
3475 removeUnwindEdge(PredBB);
3477 TerminatorInst *TI = PredBB->getTerminator();
3478 TI->replaceUsesOfWith(BB, UnwindDest);
3482 // The cleanup pad is now unreachable. Zap it.
3483 BB->eraseFromParent();
3487 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
3488 BasicBlock *BB = RI->getParent();
3489 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
3491 // Find predecessors that end with branches.
3492 SmallVector<BasicBlock*, 8> UncondBranchPreds;
3493 SmallVector<BranchInst*, 8> CondBranchPreds;
3494 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
3495 BasicBlock *P = *PI;
3496 TerminatorInst *PTI = P->getTerminator();
3497 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
3498 if (BI->isUnconditional())
3499 UncondBranchPreds.push_back(P);
3501 CondBranchPreds.push_back(BI);
3505 // If we found some, do the transformation!
3506 if (!UncondBranchPreds.empty() && DupRet) {
3507 while (!UncondBranchPreds.empty()) {
3508 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
3509 DEBUG(dbgs() << "FOLDING: " << *BB
3510 << "INTO UNCOND BRANCH PRED: " << *Pred);
3511 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
3514 // If we eliminated all predecessors of the block, delete the block now.
3516 // We know there are no successors, so just nuke the block.
3517 BB->eraseFromParent();
3522 // Check out all of the conditional branches going to this return
3523 // instruction. If any of them just select between returns, change the
3524 // branch itself into a select/return pair.
3525 while (!CondBranchPreds.empty()) {
3526 BranchInst *BI = CondBranchPreds.pop_back_val();
3528 // Check to see if the non-BB successor is also a return block.
3529 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
3530 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
3531 SimplifyCondBranchToTwoReturns(BI, Builder))
3537 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
3538 BasicBlock *BB = UI->getParent();
3540 bool Changed = false;
3542 // If there are any instructions immediately before the unreachable that can
3543 // be removed, do so.
3544 while (UI->getIterator() != BB->begin()) {
3545 BasicBlock::iterator BBI = UI->getIterator();
3547 // Do not delete instructions that can have side effects which might cause
3548 // the unreachable to not be reachable; specifically, calls and volatile
3549 // operations may have this effect.
3550 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
3552 if (BBI->mayHaveSideEffects()) {
3553 if (auto *SI = dyn_cast<StoreInst>(BBI)) {
3554 if (SI->isVolatile())
3556 } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
3557 if (LI->isVolatile())
3559 } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
3560 if (RMWI->isVolatile())
3562 } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
3563 if (CXI->isVolatile())
3565 } else if (isa<CatchPadInst>(BBI)) {
3566 // A catchpad may invoke exception object constructors and such, which
3567 // in some languages can be arbitrary code, so be conservative by
3569 // For CoreCLR, it just involves a type test, so can be removed.
3570 if (classifyEHPersonality(BB->getParent()->getPersonalityFn()) !=
3571 EHPersonality::CoreCLR)
3573 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
3574 !isa<LandingPadInst>(BBI)) {
3577 // Note that deleting LandingPad's here is in fact okay, although it
3578 // involves a bit of subtle reasoning. If this inst is a LandingPad,
3579 // all the predecessors of this block will be the unwind edges of Invokes,
3580 // and we can therefore guarantee this block will be erased.
3583 // Delete this instruction (any uses are guaranteed to be dead)
3584 if (!BBI->use_empty())
3585 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
3586 BBI->eraseFromParent();
3590 // If the unreachable instruction is the first in the block, take a gander
3591 // at all of the predecessors of this instruction, and simplify them.
3592 if (&BB->front() != UI) return Changed;
3594 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
3595 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
3596 TerminatorInst *TI = Preds[i]->getTerminator();
3597 IRBuilder<> Builder(TI);
3598 if (auto *BI = dyn_cast<BranchInst>(TI)) {
3599 if (BI->isUnconditional()) {
3600 if (BI->getSuccessor(0) == BB) {
3601 new UnreachableInst(TI->getContext(), TI);
3602 TI->eraseFromParent();
3606 if (BI->getSuccessor(0) == BB) {
3607 Builder.CreateBr(BI->getSuccessor(1));
3608 EraseTerminatorInstAndDCECond(BI);
3609 } else if (BI->getSuccessor(1) == BB) {
3610 Builder.CreateBr(BI->getSuccessor(0));
3611 EraseTerminatorInstAndDCECond(BI);
3615 } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
3616 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3618 if (i.getCaseSuccessor() == BB) {
3619 BB->removePredecessor(SI->getParent());
3624 } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
3625 if (II->getUnwindDest() == BB) {
3626 removeUnwindEdge(TI->getParent());
3629 } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
3630 if (CSI->getUnwindDest() == BB) {
3631 removeUnwindEdge(TI->getParent());
3636 for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
3637 E = CSI->handler_end();
3640 CSI->removeHandler(I);
3646 if (CSI->getNumHandlers() == 0) {
3647 BasicBlock *CatchSwitchBB = CSI->getParent();
3648 if (CSI->hasUnwindDest()) {
3649 // Redirect preds to the unwind dest
3650 CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
3652 // Rewrite all preds to unwind to caller (or from invoke to call).
3653 SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
3654 for (BasicBlock *EHPred : EHPreds)
3655 removeUnwindEdge(EHPred);
3657 // The catchswitch is no longer reachable.
3658 new UnreachableInst(CSI->getContext(), CSI);
3659 CSI->eraseFromParent();
3662 } else if (isa<CleanupReturnInst>(TI)) {
3663 new UnreachableInst(TI->getContext(), TI);
3664 TI->eraseFromParent();
3669 // If this block is now dead, remove it.
3670 if (pred_empty(BB) &&
3671 BB != &BB->getParent()->getEntryBlock()) {
3672 // We know there are no successors, so just nuke the block.
3673 BB->eraseFromParent();
3680 static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) {
3681 assert(Cases.size() >= 1);
3683 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3684 for (size_t I = 1, E = Cases.size(); I != E; ++I) {
3685 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
3691 /// Turn a switch with two reachable destinations into an integer range
3692 /// comparison and branch.
3693 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3694 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3697 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
3699 // Partition the cases into two sets with different destinations.
3700 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
3701 BasicBlock *DestB = nullptr;
3702 SmallVector <ConstantInt *, 16> CasesA;
3703 SmallVector <ConstantInt *, 16> CasesB;
3705 for (SwitchInst::CaseIt I : SI->cases()) {
3706 BasicBlock *Dest = I.getCaseSuccessor();
3707 if (!DestA) DestA = Dest;
3708 if (Dest == DestA) {
3709 CasesA.push_back(I.getCaseValue());
3712 if (!DestB) DestB = Dest;
3713 if (Dest == DestB) {
3714 CasesB.push_back(I.getCaseValue());
3717 return false; // More than two destinations.
3720 assert(DestA && DestB && "Single-destination switch should have been folded.");
3721 assert(DestA != DestB);
3722 assert(DestB != SI->getDefaultDest());
3723 assert(!CasesB.empty() && "There must be non-default cases.");
3724 assert(!CasesA.empty() || HasDefault);
3726 // Figure out if one of the sets of cases form a contiguous range.
3727 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
3728 BasicBlock *ContiguousDest = nullptr;
3729 BasicBlock *OtherDest = nullptr;
3730 if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
3731 ContiguousCases = &CasesA;
3732 ContiguousDest = DestA;
3734 } else if (CasesAreContiguous(CasesB)) {
3735 ContiguousCases = &CasesB;
3736 ContiguousDest = DestB;
3741 // Start building the compare and branch.
3743 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
3744 Constant *NumCases = ConstantInt::get(Offset->getType(), ContiguousCases->size());
3746 Value *Sub = SI->getCondition();
3747 if (!Offset->isNullValue())
3748 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
3751 // If NumCases overflowed, then all possible values jump to the successor.
3752 if (NumCases->isNullValue() && !ContiguousCases->empty())
3753 Cmp = ConstantInt::getTrue(SI->getContext());
3755 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3756 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
3758 // Update weight for the newly-created conditional branch.
3759 if (HasBranchWeights(SI)) {
3760 SmallVector<uint64_t, 8> Weights;
3761 GetBranchWeights(SI, Weights);
3762 if (Weights.size() == 1 + SI->getNumCases()) {
3763 uint64_t TrueWeight = 0;
3764 uint64_t FalseWeight = 0;
3765 for (size_t I = 0, E = Weights.size(); I != E; ++I) {
3766 if (SI->getSuccessor(I) == ContiguousDest)
3767 TrueWeight += Weights[I];
3769 FalseWeight += Weights[I];
3771 while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
3775 NewBI->setMetadata(LLVMContext::MD_prof,
3776 MDBuilder(SI->getContext()).createBranchWeights(
3777 (uint32_t)TrueWeight, (uint32_t)FalseWeight));
3781 // Prune obsolete incoming values off the successors' PHI nodes.
3782 for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
3783 unsigned PreviousEdges = ContiguousCases->size();
3784 if (ContiguousDest == SI->getDefaultDest()) ++PreviousEdges;
3785 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
3786 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3788 for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
3789 unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
3790 if (OtherDest == SI->getDefaultDest()) ++PreviousEdges;
3791 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
3792 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3796 SI->eraseFromParent();
3801 /// Compute masked bits for the condition of a switch
3802 /// and use it to remove dead cases.
3803 static bool EliminateDeadSwitchCases(SwitchInst *SI, AssumptionCache *AC,
3804 const DataLayout &DL) {
3805 Value *Cond = SI->getCondition();
3806 unsigned Bits = Cond->getType()->getIntegerBitWidth();
3807 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3808 computeKnownBits(Cond, KnownZero, KnownOne, DL, 0, AC, SI);
3810 // Gather dead cases.
3811 SmallVector<ConstantInt*, 8> DeadCases;
3812 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3813 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3814 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3815 DeadCases.push_back(I.getCaseValue());
3816 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3817 << I.getCaseValue() << "' is dead.\n");
3821 // If we can prove that the cases must cover all possible values, the
3822 // default destination becomes dead and we can remove it. If we know some
3823 // of the bits in the value, we can use that to more precisely compute the
3824 // number of possible unique case values.
3826 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
3827 const unsigned NumUnknownBits = Bits -
3828 (KnownZero.Or(KnownOne)).countPopulation();
3829 assert(NumUnknownBits <= Bits);
3830 if (HasDefault && DeadCases.empty() &&
3831 NumUnknownBits < 64 /* avoid overflow */ &&
3832 SI->getNumCases() == (1ULL << NumUnknownBits)) {
3833 DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
3834 BasicBlock *NewDefault = SplitBlockPredecessors(SI->getDefaultDest(),
3835 SI->getParent(), "");
3836 SI->setDefaultDest(&*NewDefault);
3837 SplitBlock(&*NewDefault, &NewDefault->front());
3838 auto *OldTI = NewDefault->getTerminator();
3839 new UnreachableInst(SI->getContext(), OldTI);
3840 EraseTerminatorInstAndDCECond(OldTI);
3844 SmallVector<uint64_t, 8> Weights;
3845 bool HasWeight = HasBranchWeights(SI);
3847 GetBranchWeights(SI, Weights);
3848 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3851 // Remove dead cases from the switch.
3852 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3853 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3854 assert(Case != SI->case_default() &&
3855 "Case was not found. Probably mistake in DeadCases forming.");
3857 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3861 // Prune unused values from PHI nodes.
3862 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3863 SI->removeCase(Case);
3865 if (HasWeight && Weights.size() >= 2) {
3866 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3867 SI->setMetadata(LLVMContext::MD_prof,
3868 MDBuilder(SI->getParent()->getContext()).
3869 createBranchWeights(MDWeights));
3872 return !DeadCases.empty();
3875 /// If BB would be eligible for simplification by
3876 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3877 /// by an unconditional branch), look at the phi node for BB in the successor
3878 /// block and see if the incoming value is equal to CaseValue. If so, return
3879 /// the phi node, and set PhiIndex to BB's index in the phi node.
3880 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3883 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3884 return nullptr; // BB must be empty to be a candidate for simplification.
3885 if (!BB->getSinglePredecessor())
3886 return nullptr; // BB must be dominated by the switch.
3888 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3889 if (!Branch || !Branch->isUnconditional())
3890 return nullptr; // Terminator must be unconditional branch.
3892 BasicBlock *Succ = Branch->getSuccessor(0);
3894 BasicBlock::iterator I = Succ->begin();
3895 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3896 int Idx = PHI->getBasicBlockIndex(BB);
3897 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3899 Value *InValue = PHI->getIncomingValue(Idx);
3900 if (InValue != CaseValue) continue;
3909 /// Try to forward the condition of a switch instruction to a phi node
3910 /// dominated by the switch, if that would mean that some of the destination
3911 /// blocks of the switch can be folded away.
3912 /// Returns true if a change is made.
3913 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3914 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3915 ForwardingNodesMap ForwardingNodes;
3917 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3918 ConstantInt *CaseValue = I.getCaseValue();
3919 BasicBlock *CaseDest = I.getCaseSuccessor();
3922 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3926 ForwardingNodes[PHI].push_back(PhiIndex);
3929 bool Changed = false;
3931 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3932 E = ForwardingNodes.end(); I != E; ++I) {
3933 PHINode *Phi = I->first;
3934 SmallVectorImpl<int> &Indexes = I->second;
3936 if (Indexes.size() < 2) continue;
3938 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3939 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3946 /// Return true if the backend will be able to handle
3947 /// initializing an array of constants like C.
3948 static bool ValidLookupTableConstant(Constant *C) {
3949 if (C->isThreadDependent())
3951 if (C->isDLLImportDependent())
3954 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3955 return CE->isGEPWithNoNotionalOverIndexing();
3957 return isa<ConstantFP>(C) ||
3958 isa<ConstantInt>(C) ||
3959 isa<ConstantPointerNull>(C) ||
3960 isa<GlobalValue>(C) ||
3964 /// If V is a Constant, return it. Otherwise, try to look up
3965 /// its constant value in ConstantPool, returning 0 if it's not there.
3966 static Constant *LookupConstant(Value *V,
3967 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3968 if (Constant *C = dyn_cast<Constant>(V))
3970 return ConstantPool.lookup(V);
3973 /// Try to fold instruction I into a constant. This works for
3974 /// simple instructions such as binary operations where both operands are
3975 /// constant or can be replaced by constants from the ConstantPool. Returns the
3976 /// resulting constant on success, 0 otherwise.
3978 ConstantFold(Instruction *I, const DataLayout &DL,
3979 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
3980 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3981 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3984 if (A->isAllOnesValue())
3985 return LookupConstant(Select->getTrueValue(), ConstantPool);
3986 if (A->isNullValue())
3987 return LookupConstant(Select->getFalseValue(), ConstantPool);
3991 SmallVector<Constant *, 4> COps;
3992 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
3993 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
3999 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4000 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4004 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), COps, DL);
4007 /// Try to determine the resulting constant values in phi nodes
4008 /// at the common destination basic block, *CommonDest, for one of the case
4009 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4010 /// case), of a switch instruction SI.
4012 GetCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest,
4013 BasicBlock **CommonDest,
4014 SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4015 const DataLayout &DL) {
4016 // The block from which we enter the common destination.
4017 BasicBlock *Pred = SI->getParent();
4019 // If CaseDest is empty except for some side-effect free instructions through
4020 // which we can constant-propagate the CaseVal, continue to its successor.
4021 SmallDenseMap<Value*, Constant*> ConstantPool;
4022 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4023 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
4025 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
4026 // If the terminator is a simple branch, continue to the next block.
4027 if (T->getNumSuccessors() != 1)
4030 CaseDest = T->getSuccessor(0);
4031 } else if (isa<DbgInfoIntrinsic>(I)) {
4032 // Skip debug intrinsic.
4034 } else if (Constant *C = ConstantFold(&*I, DL, ConstantPool)) {
4035 // Instruction is side-effect free and constant.
4037 // If the instruction has uses outside this block or a phi node slot for
4038 // the block, it is not safe to bypass the instruction since it would then
4039 // no longer dominate all its uses.
4040 for (auto &Use : I->uses()) {
4041 User *User = Use.getUser();
4042 if (Instruction *I = dyn_cast<Instruction>(User))
4043 if (I->getParent() == CaseDest)
4045 if (PHINode *Phi = dyn_cast<PHINode>(User))
4046 if (Phi->getIncomingBlock(Use) == CaseDest)
4051 ConstantPool.insert(std::make_pair(&*I, C));
4057 // If we did not have a CommonDest before, use the current one.
4059 *CommonDest = CaseDest;
4060 // If the destination isn't the common one, abort.
4061 if (CaseDest != *CommonDest)
4064 // Get the values for this case from phi nodes in the destination block.
4065 BasicBlock::iterator I = (*CommonDest)->begin();
4066 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
4067 int Idx = PHI->getBasicBlockIndex(Pred);
4071 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
4076 // Be conservative about which kinds of constants we support.
4077 if (!ValidLookupTableConstant(ConstVal))
4080 Res.push_back(std::make_pair(PHI, ConstVal));
4083 return Res.size() > 0;
4086 // Helper function used to add CaseVal to the list of cases that generate
4088 static void MapCaseToResult(ConstantInt *CaseVal,
4089 SwitchCaseResultVectorTy &UniqueResults,
4091 for (auto &I : UniqueResults) {
4092 if (I.first == Result) {
4093 I.second.push_back(CaseVal);
4097 UniqueResults.push_back(std::make_pair(Result,
4098 SmallVector<ConstantInt*, 4>(1, CaseVal)));
4101 // Helper function that initializes a map containing
4102 // results for the PHI node of the common destination block for a switch
4103 // instruction. Returns false if multiple PHI nodes have been found or if
4104 // there is not a common destination block for the switch.
4105 static bool InitializeUniqueCases(SwitchInst *SI, PHINode *&PHI,
4106 BasicBlock *&CommonDest,
4107 SwitchCaseResultVectorTy &UniqueResults,
4108 Constant *&DefaultResult,
4109 const DataLayout &DL) {
4110 for (auto &I : SI->cases()) {
4111 ConstantInt *CaseVal = I.getCaseValue();
4113 // Resulting value at phi nodes for this case value.
4114 SwitchCaseResultsTy Results;
4115 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4119 // Only one value per case is permitted
4120 if (Results.size() > 1)
4122 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4124 // Check the PHI consistency.
4126 PHI = Results[0].first;
4127 else if (PHI != Results[0].first)
4130 // Find the default result value.
4131 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
4132 BasicBlock *DefaultDest = SI->getDefaultDest();
4133 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4135 // If the default value is not found abort unless the default destination
4138 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4139 if ((!DefaultResult &&
4140 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4146 // Helper function that checks if it is possible to transform a switch with only
4147 // two cases (or two cases + default) that produces a result into a select.
4150 // case 10: %0 = icmp eq i32 %a, 10
4151 // return 10; %1 = select i1 %0, i32 10, i32 4
4152 // case 20: ----> %2 = icmp eq i32 %a, 20
4153 // return 2; %3 = select i1 %2, i32 2, i32 %1
4158 ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4159 Constant *DefaultResult, Value *Condition,
4160 IRBuilder<> &Builder) {
4161 assert(ResultVector.size() == 2 &&
4162 "We should have exactly two unique results at this point");
4163 // If we are selecting between only two cases transform into a simple
4164 // select or a two-way select if default is possible.
4165 if (ResultVector[0].second.size() == 1 &&
4166 ResultVector[1].second.size() == 1) {
4167 ConstantInt *const FirstCase = ResultVector[0].second[0];
4168 ConstantInt *const SecondCase = ResultVector[1].second[0];
4170 bool DefaultCanTrigger = DefaultResult;
4171 Value *SelectValue = ResultVector[1].first;
4172 if (DefaultCanTrigger) {
4173 Value *const ValueCompare =
4174 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4175 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4176 DefaultResult, "switch.select");
4178 Value *const ValueCompare =
4179 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4180 return Builder.CreateSelect(ValueCompare, ResultVector[0].first, SelectValue,
4187 // Helper function to cleanup a switch instruction that has been converted into
4188 // a select, fixing up PHI nodes and basic blocks.
4189 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
4191 IRBuilder<> &Builder) {
4192 BasicBlock *SelectBB = SI->getParent();
4193 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4194 PHI->removeIncomingValue(SelectBB);
4195 PHI->addIncoming(SelectValue, SelectBB);
4197 Builder.CreateBr(PHI->getParent());
4199 // Remove the switch.
4200 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4201 BasicBlock *Succ = SI->getSuccessor(i);
4203 if (Succ == PHI->getParent())
4205 Succ->removePredecessor(SelectBB);
4207 SI->eraseFromParent();
4210 /// If the switch is only used to initialize one or more
4211 /// phi nodes in a common successor block with only two different
4212 /// constant values, replace the switch with select.
4213 static bool SwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4214 AssumptionCache *AC, const DataLayout &DL) {
4215 Value *const Cond = SI->getCondition();
4216 PHINode *PHI = nullptr;
4217 BasicBlock *CommonDest = nullptr;
4218 Constant *DefaultResult;
4219 SwitchCaseResultVectorTy UniqueResults;
4220 // Collect all the cases that will deliver the same value from the switch.
4221 if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4224 // Selects choose between maximum two values.
4225 if (UniqueResults.size() != 2)
4227 assert(PHI != nullptr && "PHI for value select not found");
4229 Builder.SetInsertPoint(SI);
4230 Value *SelectValue = ConvertTwoCaseSwitch(
4232 DefaultResult, Cond, Builder);
4234 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4237 // The switch couldn't be converted into a select.
4242 /// This class represents a lookup table that can be used to replace a switch.
4243 class SwitchLookupTable {
4245 /// Create a lookup table to use as a switch replacement with the contents
4246 /// of Values, using DefaultValue to fill any holes in the table.
4248 Module &M, uint64_t TableSize, ConstantInt *Offset,
4249 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4250 Constant *DefaultValue, const DataLayout &DL);
4252 /// Build instructions with Builder to retrieve the value at
4253 /// the position given by Index in the lookup table.
4254 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4256 /// Return true if a table with TableSize elements of
4257 /// type ElementType would fit in a target-legal register.
4258 static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4262 // Depending on the contents of the table, it can be represented in
4265 // For tables where each element contains the same value, we just have to
4266 // store that single value and return it for each lookup.
4269 // For tables where there is a linear relationship between table index
4270 // and values. We calculate the result with a simple multiplication
4271 // and addition instead of a table lookup.
4274 // For small tables with integer elements, we can pack them into a bitmap
4275 // that fits into a target-legal register. Values are retrieved by
4276 // shift and mask operations.
4279 // The table is stored as an array of values. Values are retrieved by load
4280 // instructions from the table.
4284 // For SingleValueKind, this is the single value.
4285 Constant *SingleValue;
4287 // For BitMapKind, this is the bitmap.
4288 ConstantInt *BitMap;
4289 IntegerType *BitMapElementTy;
4291 // For LinearMapKind, these are the constants used to derive the value.
4292 ConstantInt *LinearOffset;
4293 ConstantInt *LinearMultiplier;
4295 // For ArrayKind, this is the array.
4296 GlobalVariable *Array;
4300 SwitchLookupTable::SwitchLookupTable(
4301 Module &M, uint64_t TableSize, ConstantInt *Offset,
4302 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4303 Constant *DefaultValue, const DataLayout &DL)
4304 : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
4305 LinearOffset(nullptr), LinearMultiplier(nullptr), Array(nullptr) {
4306 assert(Values.size() && "Can't build lookup table without values!");
4307 assert(TableSize >= Values.size() && "Can't fit values in table!");
4309 // If all values in the table are equal, this is that value.
4310 SingleValue = Values.begin()->second;
4312 Type *ValueType = Values.begin()->second->getType();
4314 // Build up the table contents.
4315 SmallVector<Constant*, 64> TableContents(TableSize);
4316 for (size_t I = 0, E = Values.size(); I != E; ++I) {
4317 ConstantInt *CaseVal = Values[I].first;
4318 Constant *CaseRes = Values[I].second;
4319 assert(CaseRes->getType() == ValueType);
4321 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
4323 TableContents[Idx] = CaseRes;
4325 if (CaseRes != SingleValue)
4326 SingleValue = nullptr;
4329 // Fill in any holes in the table with the default result.
4330 if (Values.size() < TableSize) {
4331 assert(DefaultValue &&
4332 "Need a default value to fill the lookup table holes.");
4333 assert(DefaultValue->getType() == ValueType);
4334 for (uint64_t I = 0; I < TableSize; ++I) {
4335 if (!TableContents[I])
4336 TableContents[I] = DefaultValue;
4339 if (DefaultValue != SingleValue)
4340 SingleValue = nullptr;
4343 // If each element in the table contains the same value, we only need to store
4344 // that single value.
4346 Kind = SingleValueKind;
4350 // Check if we can derive the value with a linear transformation from the
4352 if (isa<IntegerType>(ValueType)) {
4353 bool LinearMappingPossible = true;
4356 assert(TableSize >= 2 && "Should be a SingleValue table.");
4357 // Check if there is the same distance between two consecutive values.
4358 for (uint64_t I = 0; I < TableSize; ++I) {
4359 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4361 // This is an undef. We could deal with it, but undefs in lookup tables
4362 // are very seldom. It's probably not worth the additional complexity.
4363 LinearMappingPossible = false;
4366 APInt Val = ConstVal->getValue();
4368 APInt Dist = Val - PrevVal;
4371 } else if (Dist != DistToPrev) {
4372 LinearMappingPossible = false;
4378 if (LinearMappingPossible) {
4379 LinearOffset = cast<ConstantInt>(TableContents[0]);
4380 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
4381 Kind = LinearMapKind;
4387 // If the type is integer and the table fits in a register, build a bitmap.
4388 if (WouldFitInRegister(DL, TableSize, ValueType)) {
4389 IntegerType *IT = cast<IntegerType>(ValueType);
4390 APInt TableInt(TableSize * IT->getBitWidth(), 0);
4391 for (uint64_t I = TableSize; I > 0; --I) {
4392 TableInt <<= IT->getBitWidth();
4393 // Insert values into the bitmap. Undef values are set to zero.
4394 if (!isa<UndefValue>(TableContents[I - 1])) {
4395 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
4396 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
4399 BitMap = ConstantInt::get(M.getContext(), TableInt);
4400 BitMapElementTy = IT;
4406 // Store the table in an array.
4407 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
4408 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
4410 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
4411 GlobalVariable::PrivateLinkage,
4414 Array->setUnnamedAddr(true);
4418 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
4420 case SingleValueKind:
4422 case LinearMapKind: {
4423 // Derive the result value from the input value.
4424 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
4425 false, "switch.idx.cast");
4426 if (!LinearMultiplier->isOne())
4427 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
4428 if (!LinearOffset->isZero())
4429 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
4433 // Type of the bitmap (e.g. i59).
4434 IntegerType *MapTy = BitMap->getType();
4436 // Cast Index to the same type as the bitmap.
4437 // Note: The Index is <= the number of elements in the table, so
4438 // truncating it to the width of the bitmask is safe.
4439 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
4441 // Multiply the shift amount by the element width.
4442 ShiftAmt = Builder.CreateMul(ShiftAmt,
4443 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
4447 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
4448 "switch.downshift");
4450 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
4454 // Make sure the table index will not overflow when treated as signed.
4455 IntegerType *IT = cast<IntegerType>(Index->getType());
4456 uint64_t TableSize = Array->getInitializer()->getType()
4457 ->getArrayNumElements();
4458 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
4459 Index = Builder.CreateZExt(Index,
4460 IntegerType::get(IT->getContext(),
4461 IT->getBitWidth() + 1),
4462 "switch.tableidx.zext");
4464 Value *GEPIndices[] = { Builder.getInt32(0), Index };
4465 Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
4466 GEPIndices, "switch.gep");
4467 return Builder.CreateLoad(GEP, "switch.load");
4470 llvm_unreachable("Unknown lookup table kind!");
4473 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
4475 Type *ElementType) {
4476 auto *IT = dyn_cast<IntegerType>(ElementType);
4479 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
4480 // are <= 15, we could try to narrow the type.
4482 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
4483 if (TableSize >= UINT_MAX/IT->getBitWidth())
4485 return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
4488 /// Determine whether a lookup table should be built for this switch, based on
4489 /// the number of cases, size of the table, and the types of the results.
4491 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
4492 const TargetTransformInfo &TTI, const DataLayout &DL,
4493 const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
4494 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
4495 return false; // TableSize overflowed, or mul below might overflow.
4497 bool AllTablesFitInRegister = true;
4498 bool HasIllegalType = false;
4499 for (const auto &I : ResultTypes) {
4500 Type *Ty = I.second;
4502 // Saturate this flag to true.
4503 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
4505 // Saturate this flag to false.
4506 AllTablesFitInRegister = AllTablesFitInRegister &&
4507 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
4509 // If both flags saturate, we're done. NOTE: This *only* works with
4510 // saturating flags, and all flags have to saturate first due to the
4511 // non-deterministic behavior of iterating over a dense map.
4512 if (HasIllegalType && !AllTablesFitInRegister)
4516 // If each table would fit in a register, we should build it anyway.
4517 if (AllTablesFitInRegister)
4520 // Don't build a table that doesn't fit in-register if it has illegal types.
4524 // The table density should be at least 40%. This is the same criterion as for
4525 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
4526 // FIXME: Find the best cut-off.
4527 return SI->getNumCases() * 10 >= TableSize * 4;
4530 /// Try to reuse the switch table index compare. Following pattern:
4532 /// if (idx < tablesize)
4533 /// r = table[idx]; // table does not contain default_value
4535 /// r = default_value;
4536 /// if (r != default_value)
4539 /// Is optimized to:
4541 /// cond = idx < tablesize;
4545 /// r = default_value;
4549 /// Jump threading will then eliminate the second if(cond).
4550 static void reuseTableCompare(User *PhiUser, BasicBlock *PhiBlock,
4551 BranchInst *RangeCheckBranch, Constant *DefaultValue,
4552 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values) {
4554 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
4558 // We require that the compare is in the same block as the phi so that jump
4559 // threading can do its work afterwards.
4560 if (CmpInst->getParent() != PhiBlock)
4563 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
4567 Value *RangeCmp = RangeCheckBranch->getCondition();
4568 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
4569 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
4571 // Check if the compare with the default value is constant true or false.
4572 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
4573 DefaultValue, CmpOp1, true);
4574 if (DefaultConst != TrueConst && DefaultConst != FalseConst)
4577 // Check if the compare with the case values is distinct from the default
4579 for (auto ValuePair : Values) {
4580 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
4581 ValuePair.second, CmpOp1, true);
4582 if (!CaseConst || CaseConst == DefaultConst)
4584 assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
4585 "Expect true or false as compare result.");
4588 // Check if the branch instruction dominates the phi node. It's a simple
4589 // dominance check, but sufficient for our needs.
4590 // Although this check is invariant in the calling loops, it's better to do it
4591 // at this late stage. Practically we do it at most once for a switch.
4592 BasicBlock *BranchBlock = RangeCheckBranch->getParent();
4593 for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
4594 BasicBlock *Pred = *PI;
4595 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
4599 if (DefaultConst == FalseConst) {
4600 // The compare yields the same result. We can replace it.
4601 CmpInst->replaceAllUsesWith(RangeCmp);
4602 ++NumTableCmpReuses;
4604 // The compare yields the same result, just inverted. We can replace it.
4605 Value *InvertedTableCmp = BinaryOperator::CreateXor(RangeCmp,
4606 ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
4608 CmpInst->replaceAllUsesWith(InvertedTableCmp);
4609 ++NumTableCmpReuses;
4613 /// If the switch is only used to initialize one or more phi nodes in a common
4614 /// successor block with different constant values, replace the switch with
4616 static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
4617 const DataLayout &DL,
4618 const TargetTransformInfo &TTI) {
4619 assert(SI->getNumCases() > 1 && "Degenerate switch?");
4621 // Only build lookup table when we have a target that supports it.
4622 if (!TTI.shouldBuildLookupTables())
4625 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
4626 // split off a dense part and build a lookup table for that.
4628 // FIXME: This creates arrays of GEPs to constant strings, which means each
4629 // GEP needs a runtime relocation in PIC code. We should just build one big
4630 // string and lookup indices into that.
4632 // Ignore switches with less than three cases. Lookup tables will not make them
4633 // faster, so we don't analyze them.
4634 if (SI->getNumCases() < 3)
4637 // Figure out the corresponding result for each case value and phi node in the
4638 // common destination, as well as the min and max case values.
4639 assert(SI->case_begin() != SI->case_end());
4640 SwitchInst::CaseIt CI = SI->case_begin();
4641 ConstantInt *MinCaseVal = CI.getCaseValue();
4642 ConstantInt *MaxCaseVal = CI.getCaseValue();
4644 BasicBlock *CommonDest = nullptr;
4645 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
4646 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
4647 SmallDenseMap<PHINode*, Constant*> DefaultResults;
4648 SmallDenseMap<PHINode*, Type*> ResultTypes;
4649 SmallVector<PHINode*, 4> PHIs;
4651 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
4652 ConstantInt *CaseVal = CI.getCaseValue();
4653 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
4654 MinCaseVal = CaseVal;
4655 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
4656 MaxCaseVal = CaseVal;
4658 // Resulting value at phi nodes for this case value.
4659 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
4661 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
4665 // Append the result from this case to the list for each phi.
4666 for (const auto &I : Results) {
4667 PHINode *PHI = I.first;
4668 Constant *Value = I.second;
4669 if (!ResultLists.count(PHI))
4670 PHIs.push_back(PHI);
4671 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
4675 // Keep track of the result types.
4676 for (PHINode *PHI : PHIs) {
4677 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
4680 uint64_t NumResults = ResultLists[PHIs[0]].size();
4681 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
4682 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
4683 bool TableHasHoles = (NumResults < TableSize);
4685 // If the table has holes, we need a constant result for the default case
4686 // or a bitmask that fits in a register.
4687 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
4688 bool HasDefaultResults = GetCaseResults(SI, nullptr, SI->getDefaultDest(),
4689 &CommonDest, DefaultResultsList, DL);
4691 bool NeedMask = (TableHasHoles && !HasDefaultResults);
4693 // As an extra penalty for the validity test we require more cases.
4694 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
4696 if (!DL.fitsInLegalInteger(TableSize))
4700 for (const auto &I : DefaultResultsList) {
4701 PHINode *PHI = I.first;
4702 Constant *Result = I.second;
4703 DefaultResults[PHI] = Result;
4706 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
4709 // Create the BB that does the lookups.
4710 Module &Mod = *CommonDest->getParent()->getParent();
4711 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
4713 CommonDest->getParent(),
4716 // Compute the table index value.
4717 Builder.SetInsertPoint(SI);
4718 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
4721 // Compute the maximum table size representable by the integer type we are
4723 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
4724 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
4725 assert(MaxTableSize >= TableSize &&
4726 "It is impossible for a switch to have more entries than the max "
4727 "representable value of its input integer type's size.");
4729 // If the default destination is unreachable, or if the lookup table covers
4730 // all values of the conditional variable, branch directly to the lookup table
4731 // BB. Otherwise, check that the condition is within the case range.
4732 const bool DefaultIsReachable =
4733 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4734 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
4735 BranchInst *RangeCheckBranch = nullptr;
4737 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
4738 Builder.CreateBr(LookupBB);
4739 // Note: We call removeProdecessor later since we need to be able to get the
4740 // PHI value for the default case in case we're using a bit mask.
4742 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
4743 MinCaseVal->getType(), TableSize));
4744 RangeCheckBranch = Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
4747 // Populate the BB that does the lookups.
4748 Builder.SetInsertPoint(LookupBB);
4751 // Before doing the lookup we do the hole check.
4752 // The LookupBB is therefore re-purposed to do the hole check
4753 // and we create a new LookupBB.
4754 BasicBlock *MaskBB = LookupBB;
4755 MaskBB->setName("switch.hole_check");
4756 LookupBB = BasicBlock::Create(Mod.getContext(),
4758 CommonDest->getParent(),
4761 // Make the mask's bitwidth at least 8bit and a power-of-2 to avoid
4762 // unnecessary illegal types.
4763 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
4764 APInt MaskInt(TableSizePowOf2, 0);
4765 APInt One(TableSizePowOf2, 1);
4766 // Build bitmask; fill in a 1 bit for every case.
4767 const ResultListTy &ResultList = ResultLists[PHIs[0]];
4768 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
4769 uint64_t Idx = (ResultList[I].first->getValue() -
4770 MinCaseVal->getValue()).getLimitedValue();
4771 MaskInt |= One << Idx;
4773 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
4775 // Get the TableIndex'th bit of the bitmask.
4776 // If this bit is 0 (meaning hole) jump to the default destination,
4777 // else continue with table lookup.
4778 IntegerType *MapTy = TableMask->getType();
4779 Value *MaskIndex = Builder.CreateZExtOrTrunc(TableIndex, MapTy,
4780 "switch.maskindex");
4781 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex,
4783 Value *LoBit = Builder.CreateTrunc(Shifted,
4784 Type::getInt1Ty(Mod.getContext()),
4786 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
4788 Builder.SetInsertPoint(LookupBB);
4789 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
4792 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
4793 // We cached PHINodes in PHIs, to avoid accessing deleted PHINodes later,
4794 // do not delete PHINodes here.
4795 SI->getDefaultDest()->removePredecessor(SI->getParent(),
4796 /*DontDeleteUselessPHIs=*/true);
4799 bool ReturnedEarly = false;
4800 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
4801 PHINode *PHI = PHIs[I];
4802 const ResultListTy &ResultList = ResultLists[PHI];
4804 // If using a bitmask, use any value to fill the lookup table holes.
4805 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
4806 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL);
4808 Value *Result = Table.BuildLookup(TableIndex, Builder);
4810 // If the result is used to return immediately from the function, we want to
4811 // do that right here.
4812 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
4813 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
4814 Builder.CreateRet(Result);
4815 ReturnedEarly = true;
4819 // Do a small peephole optimization: re-use the switch table compare if
4821 if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
4822 BasicBlock *PhiBlock = PHI->getParent();
4823 // Search for compare instructions which use the phi.
4824 for (auto *User : PHI->users()) {
4825 reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
4829 PHI->addIncoming(Result, LookupBB);
4833 Builder.CreateBr(CommonDest);
4835 // Remove the switch.
4836 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4837 BasicBlock *Succ = SI->getSuccessor(i);
4839 if (Succ == SI->getDefaultDest())
4841 Succ->removePredecessor(SI->getParent());
4843 SI->eraseFromParent();
4847 ++NumLookupTablesHoles;
4851 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
4852 BasicBlock *BB = SI->getParent();
4854 if (isValueEqualityComparison(SI)) {
4855 // If we only have one predecessor, and if it is a branch on this value,
4856 // see if that predecessor totally determines the outcome of this switch.
4857 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4858 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
4859 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4861 Value *Cond = SI->getCondition();
4862 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
4863 if (SimplifySwitchOnSelect(SI, Select))
4864 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4866 // If the block only contains the switch, see if we can fold the block
4867 // away into any preds.
4868 BasicBlock::iterator BBI = BB->begin();
4869 // Ignore dbg intrinsics.
4870 while (isa<DbgInfoIntrinsic>(BBI))
4873 if (FoldValueComparisonIntoPredecessors(SI, Builder))
4874 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4877 // Try to transform the switch into an icmp and a branch.
4878 if (TurnSwitchRangeIntoICmp(SI, Builder))
4879 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4881 // Remove unreachable cases.
4882 if (EliminateDeadSwitchCases(SI, AC, DL))
4883 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4885 if (SwitchToSelect(SI, Builder, AC, DL))
4886 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4888 if (ForwardSwitchConditionToPHI(SI))
4889 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4891 if (SwitchToLookupTable(SI, Builder, DL, TTI))
4892 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4897 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
4898 BasicBlock *BB = IBI->getParent();
4899 bool Changed = false;
4901 // Eliminate redundant destinations.
4902 SmallPtrSet<Value *, 8> Succs;
4903 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
4904 BasicBlock *Dest = IBI->getDestination(i);
4905 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
4906 Dest->removePredecessor(BB);
4907 IBI->removeDestination(i);
4913 if (IBI->getNumDestinations() == 0) {
4914 // If the indirectbr has no successors, change it to unreachable.
4915 new UnreachableInst(IBI->getContext(), IBI);
4916 EraseTerminatorInstAndDCECond(IBI);
4920 if (IBI->getNumDestinations() == 1) {
4921 // If the indirectbr has one successor, change it to a direct branch.
4922 BranchInst::Create(IBI->getDestination(0), IBI);
4923 EraseTerminatorInstAndDCECond(IBI);
4927 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
4928 if (SimplifyIndirectBrOnSelect(IBI, SI))
4929 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
4934 /// Given an block with only a single landing pad and a unconditional branch
4935 /// try to find another basic block which this one can be merged with. This
4936 /// handles cases where we have multiple invokes with unique landing pads, but
4937 /// a shared handler.
4939 /// We specifically choose to not worry about merging non-empty blocks
4940 /// here. That is a PRE/scheduling problem and is best solved elsewhere. In
4941 /// practice, the optimizer produces empty landing pad blocks quite frequently
4942 /// when dealing with exception dense code. (see: instcombine, gvn, if-else
4943 /// sinking in this file)
4945 /// This is primarily a code size optimization. We need to avoid performing
4946 /// any transform which might inhibit optimization (such as our ability to
4947 /// specialize a particular handler via tail commoning). We do this by not
4948 /// merging any blocks which require us to introduce a phi. Since the same
4949 /// values are flowing through both blocks, we don't loose any ability to
4950 /// specialize. If anything, we make such specialization more likely.
4952 /// TODO - This transformation could remove entries from a phi in the target
4953 /// block when the inputs in the phi are the same for the two blocks being
4954 /// merged. In some cases, this could result in removal of the PHI entirely.
4955 static bool TryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI,
4957 auto Succ = BB->getUniqueSuccessor();
4959 // If there's a phi in the successor block, we'd likely have to introduce
4960 // a phi into the merged landing pad block.
4961 if (isa<PHINode>(*Succ->begin()))
4964 for (BasicBlock *OtherPred : predecessors(Succ)) {
4965 if (BB == OtherPred)
4967 BasicBlock::iterator I = OtherPred->begin();
4968 LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(I);
4969 if (!LPad2 || !LPad2->isIdenticalTo(LPad))
4971 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {}
4972 BranchInst *BI2 = dyn_cast<BranchInst>(I);
4973 if (!BI2 || !BI2->isIdenticalTo(BI))
4976 // We've found an identical block. Update our predeccessors to take that
4977 // path instead and make ourselves dead.
4978 SmallSet<BasicBlock *, 16> Preds;
4979 Preds.insert(pred_begin(BB), pred_end(BB));
4980 for (BasicBlock *Pred : Preds) {
4981 InvokeInst *II = cast<InvokeInst>(Pred->getTerminator());
4982 assert(II->getNormalDest() != BB &&
4983 II->getUnwindDest() == BB && "unexpected successor");
4984 II->setUnwindDest(OtherPred);
4987 // The debug info in OtherPred doesn't cover the merged control flow that
4988 // used to go through BB. We need to delete it or update it.
4989 for (auto I = OtherPred->begin(), E = OtherPred->end();
4991 Instruction &Inst = *I; I++;
4992 if (isa<DbgInfoIntrinsic>(Inst))
4993 Inst.eraseFromParent();
4996 SmallSet<BasicBlock *, 16> Succs;
4997 Succs.insert(succ_begin(BB), succ_end(BB));
4998 for (BasicBlock *Succ : Succs) {
4999 Succ->removePredecessor(BB);
5002 IRBuilder<> Builder(BI);
5003 Builder.CreateUnreachable();
5004 BI->eraseFromParent();
5010 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
5011 BasicBlock *BB = BI->getParent();
5013 if (SinkCommon && SinkThenElseCodeToEnd(BI))
5016 // If the Terminator is the only non-phi instruction, simplify the block.
5017 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg()->getIterator();
5018 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
5019 TryToSimplifyUncondBranchFromEmptyBlock(BB))
5022 // If the only instruction in the block is a seteq/setne comparison
5023 // against a constant, try to simplify the block.
5024 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
5025 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
5026 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5028 if (I->isTerminator() &&
5029 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, DL, TTI,
5030 BonusInstThreshold, AC))
5034 // See if we can merge an empty landing pad block with another which is
5036 if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(I)) {
5037 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {}
5038 if (I->isTerminator() &&
5039 TryToMergeLandingPad(LPad, BI, BB))
5043 // If this basic block is ONLY a compare and a branch, and if a predecessor
5044 // branches to us and our successor, fold the comparison into the
5045 // predecessor and use logical operations to update the incoming value
5046 // for PHI nodes in common successor.
5047 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
5048 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5052 static BasicBlock *allPredecessorsComeFromSameSource(BasicBlock *BB) {
5053 BasicBlock *PredPred = nullptr;
5054 for (auto *P : predecessors(BB)) {
5055 BasicBlock *PPred = P->getSinglePredecessor();
5056 if (!PPred || (PredPred && PredPred != PPred))
5063 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
5064 BasicBlock *BB = BI->getParent();
5066 // Conditional branch
5067 if (isValueEqualityComparison(BI)) {
5068 // If we only have one predecessor, and if it is a branch on this value,
5069 // see if that predecessor totally determines the outcome of this
5071 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5072 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
5073 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5075 // This block must be empty, except for the setcond inst, if it exists.
5076 // Ignore dbg intrinsics.
5077 BasicBlock::iterator I = BB->begin();
5078 // Ignore dbg intrinsics.
5079 while (isa<DbgInfoIntrinsic>(I))
5082 if (FoldValueComparisonIntoPredecessors(BI, Builder))
5083 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5084 } else if (&*I == cast<Instruction>(BI->getCondition())){
5086 // Ignore dbg intrinsics.
5087 while (isa<DbgInfoIntrinsic>(I))
5089 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
5090 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5094 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
5095 if (SimplifyBranchOnICmpChain(BI, Builder, DL))
5098 // If this basic block is ONLY a compare and a branch, and if a predecessor
5099 // branches to us and one of our successors, fold the comparison into the
5100 // predecessor and use logical operations to pick the right destination.
5101 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
5102 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5104 // We have a conditional branch to two blocks that are only reachable
5105 // from BI. We know that the condbr dominates the two blocks, so see if
5106 // there is any identical code in the "then" and "else" blocks. If so, we
5107 // can hoist it up to the branching block.
5108 if (BI->getSuccessor(0)->getSinglePredecessor()) {
5109 if (BI->getSuccessor(1)->getSinglePredecessor()) {
5110 if (HoistThenElseCodeToIf(BI, TTI))
5111 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5113 // If Successor #1 has multiple preds, we may be able to conditionally
5114 // execute Successor #0 if it branches to Successor #1.
5115 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
5116 if (Succ0TI->getNumSuccessors() == 1 &&
5117 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
5118 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), TTI))
5119 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5121 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
5122 // If Successor #0 has multiple preds, we may be able to conditionally
5123 // execute Successor #1 if it branches to Successor #0.
5124 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
5125 if (Succ1TI->getNumSuccessors() == 1 &&
5126 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
5127 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), TTI))
5128 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5131 // If this is a branch on a phi node in the current block, thread control
5132 // through this block if any PHI node entries are constants.
5133 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
5134 if (PN->getParent() == BI->getParent())
5135 if (FoldCondBranchOnPHI(BI, DL))
5136 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5138 // Scan predecessor blocks for conditional branches.
5139 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
5140 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
5141 if (PBI != BI && PBI->isConditional())
5142 if (SimplifyCondBranchToCondBranch(PBI, BI, DL))
5143 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5145 // Look for diamond patterns.
5146 if (MergeCondStores)
5147 if (BasicBlock *PrevBB = allPredecessorsComeFromSameSource(BB))
5148 if (BranchInst *PBI = dyn_cast<BranchInst>(PrevBB->getTerminator()))
5149 if (PBI != BI && PBI->isConditional())
5150 if (mergeConditionalStores(PBI, BI))
5151 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5156 /// Check if passing a value to an instruction will cause undefined behavior.
5157 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
5158 Constant *C = dyn_cast<Constant>(V);
5165 if (C->isNullValue()) {
5166 // Only look at the first use, avoid hurting compile time with long uselists
5167 User *Use = *I->user_begin();
5169 // Now make sure that there are no instructions in between that can alter
5170 // control flow (eg. calls)
5171 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
5172 if (i == I->getParent()->end() || i->mayHaveSideEffects())
5175 // Look through GEPs. A load from a GEP derived from NULL is still undefined
5176 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
5177 if (GEP->getPointerOperand() == I)
5178 return passingValueIsAlwaysUndefined(V, GEP);
5180 // Look through bitcasts.
5181 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
5182 return passingValueIsAlwaysUndefined(V, BC);
5184 // Load from null is undefined.
5185 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
5186 if (!LI->isVolatile())
5187 return LI->getPointerAddressSpace() == 0;
5189 // Store to null is undefined.
5190 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
5191 if (!SI->isVolatile())
5192 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
5197 /// If BB has an incoming value that will always trigger undefined behavior
5198 /// (eg. null pointer dereference), remove the branch leading here.
5199 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
5200 for (BasicBlock::iterator i = BB->begin();
5201 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
5202 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
5203 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
5204 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
5205 IRBuilder<> Builder(T);
5206 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
5207 BB->removePredecessor(PHI->getIncomingBlock(i));
5208 // Turn uncoditional branches into unreachables and remove the dead
5209 // destination from conditional branches.
5210 if (BI->isUnconditional())
5211 Builder.CreateUnreachable();
5213 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
5214 BI->getSuccessor(0));
5215 BI->eraseFromParent();
5218 // TODO: SwitchInst.
5224 bool SimplifyCFGOpt::run(BasicBlock *BB) {
5225 bool Changed = false;
5227 assert(BB && BB->getParent() && "Block not embedded in function!");
5228 assert(BB->getTerminator() && "Degenerate basic block encountered!");
5230 // Remove basic blocks that have no predecessors (except the entry block)...
5231 // or that just have themself as a predecessor. These are unreachable.
5232 if ((pred_empty(BB) &&
5233 BB != &BB->getParent()->getEntryBlock()) ||
5234 BB->getSinglePredecessor() == BB) {
5235 DEBUG(dbgs() << "Removing BB: \n" << *BB);
5236 DeleteDeadBlock(BB);
5240 // Check to see if we can constant propagate this terminator instruction
5242 Changed |= ConstantFoldTerminator(BB, true);
5244 // Check for and eliminate duplicate PHI nodes in this block.
5245 Changed |= EliminateDuplicatePHINodes(BB);
5247 // Check for and remove branches that will always cause undefined behavior.
5248 Changed |= removeUndefIntroducingPredecessor(BB);
5250 // Merge basic blocks into their predecessor if there is only one distinct
5251 // pred, and if there is only one distinct successor of the predecessor, and
5252 // if there are no PHI nodes.
5254 if (MergeBlockIntoPredecessor(BB))
5257 IRBuilder<> Builder(BB);
5259 // If there is a trivial two-entry PHI node in this basic block, and we can
5260 // eliminate it, do so now.
5261 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
5262 if (PN->getNumIncomingValues() == 2)
5263 Changed |= FoldTwoEntryPHINode(PN, TTI, DL);
5265 Builder.SetInsertPoint(BB->getTerminator());
5266 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
5267 if (BI->isUnconditional()) {
5268 if (SimplifyUncondBranch(BI, Builder)) return true;
5270 if (SimplifyCondBranch(BI, Builder)) return true;
5272 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
5273 if (SimplifyReturn(RI, Builder)) return true;
5274 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
5275 if (SimplifyResume(RI, Builder)) return true;
5276 } else if (CleanupReturnInst *RI =
5277 dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
5278 if (SimplifyCleanupReturn(RI)) return true;
5279 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
5280 if (SimplifySwitch(SI, Builder)) return true;
5281 } else if (UnreachableInst *UI =
5282 dyn_cast<UnreachableInst>(BB->getTerminator())) {
5283 if (SimplifyUnreachable(UI)) return true;
5284 } else if (IndirectBrInst *IBI =
5285 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
5286 if (SimplifyIndirectBr(IBI)) return true;
5292 /// This function is used to do simplification of a CFG.
5293 /// For example, it adjusts branches to branches to eliminate the extra hop,
5294 /// eliminates unreachable basic blocks, and does other "peephole" optimization
5295 /// of the CFG. It returns true if a modification was made.
5297 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
5298 unsigned BonusInstThreshold, AssumptionCache *AC) {
5299 return SimplifyCFGOpt(TTI, BB->getModule()->getDataLayout(),
5300 BonusInstThreshold, AC).run(BB);