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/ADT/APInt.h"
15 #include "llvm/ADT/ArrayRef.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/Optional.h"
18 #include "llvm/ADT/SetOperations.h"
19 #include "llvm/ADT/SetVector.h"
20 #include "llvm/ADT/SmallPtrSet.h"
21 #include "llvm/ADT/SmallSet.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/Analysis/ConstantFolding.h"
26 #include "llvm/Analysis/EHPersonalities.h"
27 #include "llvm/Analysis/InstructionSimplify.h"
28 #include "llvm/Analysis/TargetTransformInfo.h"
29 #include "llvm/Analysis/ValueTracking.h"
30 #include "llvm/IR/BasicBlock.h"
31 #include "llvm/IR/CallSite.h"
32 #include "llvm/IR/CFG.h"
33 #include "llvm/IR/Constant.h"
34 #include "llvm/IR/ConstantRange.h"
35 #include "llvm/IR/Constants.h"
36 #include "llvm/IR/DataLayout.h"
37 #include "llvm/IR/DebugInfo.h"
38 #include "llvm/IR/DerivedTypes.h"
39 #include "llvm/IR/GlobalValue.h"
40 #include "llvm/IR/GlobalVariable.h"
41 #include "llvm/IR/IRBuilder.h"
42 #include "llvm/IR/InstrTypes.h"
43 #include "llvm/IR/Instruction.h"
44 #include "llvm/IR/Instructions.h"
45 #include "llvm/IR/IntrinsicInst.h"
46 #include "llvm/IR/Intrinsics.h"
47 #include "llvm/IR/LLVMContext.h"
48 #include "llvm/IR/MDBuilder.h"
49 #include "llvm/IR/Metadata.h"
50 #include "llvm/IR/Module.h"
51 #include "llvm/IR/NoFolder.h"
52 #include "llvm/IR/Operator.h"
53 #include "llvm/IR/PatternMatch.h"
54 #include "llvm/IR/Type.h"
55 #include "llvm/IR/User.h"
56 #include "llvm/IR/Value.h"
57 #include "llvm/IR/DebugInfo.h"
58 #include "llvm/Support/Casting.h"
59 #include "llvm/Support/CommandLine.h"
60 #include "llvm/Support/Debug.h"
61 #include "llvm/Support/ErrorHandling.h"
62 #include "llvm/Support/MathExtras.h"
63 #include "llvm/Support/raw_ostream.h"
64 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
65 #include "llvm/Transforms/Utils/Local.h"
66 #include "llvm/Transforms/Utils/ValueMapper.h"
79 using namespace PatternMatch;
81 #define DEBUG_TYPE "simplifycfg"
83 // Chosen as 2 so as to be cheap, but still to have enough power to fold
84 // a select, so the "clamp" idiom (of a min followed by a max) will be caught.
85 // To catch this, we need to fold a compare and a select, hence '2' being the
86 // minimum reasonable default.
87 static cl::opt<unsigned> PHINodeFoldingThreshold(
88 "phi-node-folding-threshold", cl::Hidden, cl::init(2),
90 "Control the amount of phi node folding to perform (default = 2)"));
92 static cl::opt<bool> DupRet(
93 "simplifycfg-dup-ret", cl::Hidden, cl::init(false),
94 cl::desc("Duplicate return instructions into unconditional branches"));
97 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
98 cl::desc("Sink common instructions down to the end block"));
100 static cl::opt<bool> HoistCondStores(
101 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
102 cl::desc("Hoist conditional stores if an unconditional store precedes"));
104 static cl::opt<bool> MergeCondStores(
105 "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
106 cl::desc("Hoist conditional stores even if an unconditional store does not "
107 "precede - hoist multiple conditional stores into a single "
108 "predicated store"));
110 static cl::opt<bool> MergeCondStoresAggressively(
111 "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
112 cl::desc("When merging conditional stores, do so even if the resultant "
113 "basic blocks are unlikely to be if-converted as a result"));
115 static cl::opt<bool> SpeculateOneExpensiveInst(
116 "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
117 cl::desc("Allow exactly one expensive instruction to be speculatively "
120 static cl::opt<unsigned> MaxSpeculationDepth(
121 "max-speculation-depth", cl::Hidden, cl::init(10),
122 cl::desc("Limit maximum recursion depth when calculating costs of "
123 "speculatively executed instructions"));
125 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
126 STATISTIC(NumLinearMaps,
127 "Number of switch instructions turned into linear mapping");
128 STATISTIC(NumLookupTables,
129 "Number of switch instructions turned into lookup tables");
131 NumLookupTablesHoles,
132 "Number of switch instructions turned into lookup tables (holes checked)");
133 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
134 STATISTIC(NumSinkCommons,
135 "Number of common instructions sunk down to the end block");
136 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
140 // The first field contains the value that the switch produces when a certain
141 // case group is selected, and the second field is a vector containing the
142 // cases composing the case group.
143 typedef SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>
144 SwitchCaseResultVectorTy;
145 // The first field contains the phi node that generates a result of the switch
146 // and the second field contains the value generated for a certain case in the
147 // switch for that PHI.
148 typedef SmallVector<std::pair<PHINode *, Constant *>, 4> SwitchCaseResultsTy;
150 /// ValueEqualityComparisonCase - Represents a case of a switch.
151 struct ValueEqualityComparisonCase {
155 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
156 : Value(Value), Dest(Dest) {}
158 bool operator<(ValueEqualityComparisonCase RHS) const {
159 // Comparing pointers is ok as we only rely on the order for uniquing.
160 return Value < RHS.Value;
163 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
166 class SimplifyCFGOpt {
167 const TargetTransformInfo &TTI;
168 const DataLayout &DL;
169 unsigned BonusInstThreshold;
171 SmallPtrSetImpl<BasicBlock *> *LoopHeaders;
172 Value *isValueEqualityComparison(TerminatorInst *TI);
173 BasicBlock *GetValueEqualityComparisonCases(
174 TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases);
175 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
177 IRBuilder<> &Builder);
178 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
179 IRBuilder<> &Builder);
181 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
182 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
183 bool SimplifySingleResume(ResumeInst *RI);
184 bool SimplifyCommonResume(ResumeInst *RI);
185 bool SimplifyCleanupReturn(CleanupReturnInst *RI);
186 bool SimplifyUnreachable(UnreachableInst *UI);
187 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
188 bool SimplifyIndirectBr(IndirectBrInst *IBI);
189 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
190 bool SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
193 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
194 unsigned BonusInstThreshold, AssumptionCache *AC,
195 SmallPtrSetImpl<BasicBlock *> *LoopHeaders)
196 : TTI(TTI), DL(DL), BonusInstThreshold(BonusInstThreshold), AC(AC),
197 LoopHeaders(LoopHeaders) {}
199 bool run(BasicBlock *BB);
202 } // end anonymous namespace
204 /// Return true if it is safe to merge these two
205 /// terminator instructions together.
207 SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2,
208 SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
210 return false; // Can't merge with self!
212 // It is not safe to merge these two switch instructions if they have a common
213 // successor, and if that successor has a PHI node, and if *that* PHI node has
214 // conflicting incoming values from the two switch blocks.
215 BasicBlock *SI1BB = SI1->getParent();
216 BasicBlock *SI2BB = SI2->getParent();
218 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
220 for (BasicBlock *Succ : successors(SI2BB))
221 if (SI1Succs.count(Succ))
222 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
223 PHINode *PN = cast<PHINode>(BBI);
224 if (PN->getIncomingValueForBlock(SI1BB) !=
225 PN->getIncomingValueForBlock(SI2BB)) {
227 FailBlocks->insert(Succ);
235 /// Return true if it is safe and profitable to merge these two terminator
236 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
237 /// store all PHI nodes in common successors.
239 isProfitableToFoldUnconditional(BranchInst *SI1, BranchInst *SI2,
241 SmallVectorImpl<PHINode *> &PhiNodes) {
243 return false; // Can't merge with self!
244 assert(SI1->isUnconditional() && SI2->isConditional());
246 // We fold the unconditional branch if we can easily update all PHI nodes in
247 // common successors:
248 // 1> We have a constant incoming value for the conditional branch;
249 // 2> We have "Cond" as the incoming value for the unconditional branch;
250 // 3> SI2->getCondition() and Cond have same operands.
251 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
254 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
255 Cond->getOperand(1) == Ci2->getOperand(1)) &&
256 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
257 Cond->getOperand(1) == Ci2->getOperand(0)))
260 BasicBlock *SI1BB = SI1->getParent();
261 BasicBlock *SI2BB = SI2->getParent();
262 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
263 for (BasicBlock *Succ : successors(SI2BB))
264 if (SI1Succs.count(Succ))
265 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
266 PHINode *PN = cast<PHINode>(BBI);
267 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
268 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
270 PhiNodes.push_back(PN);
275 /// Update PHI nodes in Succ to indicate that there will now be entries in it
276 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
277 /// will be the same as those coming in from ExistPred, an existing predecessor
279 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
280 BasicBlock *ExistPred) {
281 if (!isa<PHINode>(Succ->begin()))
282 return; // Quick exit if nothing to do
285 for (BasicBlock::iterator I = Succ->begin(); (PN = dyn_cast<PHINode>(I)); ++I)
286 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
289 /// Compute an abstract "cost" of speculating the given instruction,
290 /// which is assumed to be safe to speculate. TCC_Free means cheap,
291 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
293 static unsigned ComputeSpeculationCost(const User *I,
294 const TargetTransformInfo &TTI) {
295 assert(isSafeToSpeculativelyExecute(I) &&
296 "Instruction is not safe to speculatively execute!");
297 return TTI.getUserCost(I);
300 /// If we have a merge point of an "if condition" as accepted above,
301 /// return true if the specified value dominates the block. We
302 /// don't handle the true generality of domination here, just a special case
303 /// which works well enough for us.
305 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
306 /// see if V (which must be an instruction) and its recursive operands
307 /// that do not dominate BB have a combined cost lower than CostRemaining and
308 /// are non-trapping. If both are true, the instruction is inserted into the
309 /// set and true is returned.
311 /// The cost for most non-trapping instructions is defined as 1 except for
312 /// Select whose cost is 2.
314 /// After this function returns, CostRemaining is decreased by the cost of
315 /// V plus its non-dominating operands. If that cost is greater than
316 /// CostRemaining, false is returned and CostRemaining is undefined.
317 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
318 SmallPtrSetImpl<Instruction *> *AggressiveInsts,
319 unsigned &CostRemaining,
320 const TargetTransformInfo &TTI,
321 unsigned Depth = 0) {
322 // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
323 // so limit the recursion depth.
324 // TODO: While this recursion limit does prevent pathological behavior, it
325 // would be better to track visited instructions to avoid cycles.
326 if (Depth == MaxSpeculationDepth)
329 Instruction *I = dyn_cast<Instruction>(V);
331 // Non-instructions all dominate instructions, but not all constantexprs
332 // can be executed unconditionally.
333 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
338 BasicBlock *PBB = I->getParent();
340 // We don't want to allow weird loops that might have the "if condition" in
341 // the bottom of this block.
345 // If this instruction is defined in a block that contains an unconditional
346 // branch to BB, then it must be in the 'conditional' part of the "if
347 // statement". If not, it definitely dominates the region.
348 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
349 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
352 // If we aren't allowing aggressive promotion anymore, then don't consider
353 // instructions in the 'if region'.
354 if (!AggressiveInsts)
357 // If we have seen this instruction before, don't count it again.
358 if (AggressiveInsts->count(I))
361 // Okay, it looks like the instruction IS in the "condition". Check to
362 // see if it's a cheap instruction to unconditionally compute, and if it
363 // only uses stuff defined outside of the condition. If so, hoist it out.
364 if (!isSafeToSpeculativelyExecute(I))
367 unsigned Cost = ComputeSpeculationCost(I, TTI);
369 // Allow exactly one instruction to be speculated regardless of its cost
370 // (as long as it is safe to do so).
371 // This is intended to flatten the CFG even if the instruction is a division
372 // or other expensive operation. The speculation of an expensive instruction
373 // is expected to be undone in CodeGenPrepare if the speculation has not
374 // enabled further IR optimizations.
375 if (Cost > CostRemaining &&
376 (!SpeculateOneExpensiveInst || !AggressiveInsts->empty() || Depth > 0))
379 // Avoid unsigned wrap.
380 CostRemaining = (Cost > CostRemaining) ? 0 : CostRemaining - Cost;
382 // Okay, we can only really hoist these out if their operands do
383 // not take us over the cost threshold.
384 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
385 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, TTI,
388 // Okay, it's safe to do this! Remember this instruction.
389 AggressiveInsts->insert(I);
393 /// Extract ConstantInt from value, looking through IntToPtr
394 /// and PointerNullValue. Return NULL if value is not a constant int.
395 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
396 // Normal constant int.
397 ConstantInt *CI = dyn_cast<ConstantInt>(V);
398 if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
401 // This is some kind of pointer constant. Turn it into a pointer-sized
402 // ConstantInt if possible.
403 IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
405 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
406 if (isa<ConstantPointerNull>(V))
407 return ConstantInt::get(PtrTy, 0);
409 // IntToPtr const int.
410 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
411 if (CE->getOpcode() == Instruction::IntToPtr)
412 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
413 // The constant is very likely to have the right type already.
414 if (CI->getType() == PtrTy)
417 return cast<ConstantInt>(
418 ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
425 /// Given a chain of or (||) or and (&&) comparison of a value against a
426 /// constant, this will try to recover the information required for a switch
428 /// It will depth-first traverse the chain of comparison, seeking for patterns
429 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
430 /// representing the different cases for the switch.
431 /// Note that if the chain is composed of '||' it will build the set of elements
432 /// that matches the comparisons (i.e. any of this value validate the chain)
433 /// while for a chain of '&&' it will build the set elements that make the test
435 struct ConstantComparesGatherer {
436 const DataLayout &DL;
437 Value *CompValue; /// Value found for the switch comparison
438 Value *Extra; /// Extra clause to be checked before the switch
439 SmallVector<ConstantInt *, 8> Vals; /// Set of integers to match in switch
440 unsigned UsedICmps; /// Number of comparisons matched in the and/or chain
442 /// Construct and compute the result for the comparison instruction Cond
443 ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL)
444 : DL(DL), CompValue(nullptr), Extra(nullptr), UsedICmps(0) {
449 ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
450 ConstantComparesGatherer &
451 operator=(const ConstantComparesGatherer &) = delete;
454 /// Try to set the current value used for the comparison, it succeeds only if
455 /// it wasn't set before or if the new value is the same as the old one
456 bool setValueOnce(Value *NewVal) {
457 if (CompValue && CompValue != NewVal)
460 return (CompValue != nullptr);
463 /// Try to match Instruction "I" as a comparison against a constant and
464 /// populates the array Vals with the set of values that match (or do not
465 /// match depending on isEQ).
466 /// Return false on failure. On success, the Value the comparison matched
467 /// against is placed in CompValue.
468 /// If CompValue is already set, the function is expected to fail if a match
469 /// is found but the value compared to is different.
470 bool matchInstruction(Instruction *I, bool isEQ) {
471 // If this is an icmp against a constant, handle this as one of the cases.
474 if (!((ICI = dyn_cast<ICmpInst>(I)) &&
475 (C = GetConstantInt(I->getOperand(1), DL)))) {
482 // Pattern match a special case
483 // (x & ~2^z) == y --> x == y || x == y|2^z
484 // This undoes a transformation done by instcombine to fuse 2 compares.
485 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
487 // It's a little bit hard to see why the following transformations are
488 // correct. Here is a CVC3 program to verify them for 64-bit values:
491 ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
495 mask : BITVECTOR(64) = BVSHL(ONE, z);
496 QUERY( (y & ~mask = y) =>
497 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
499 QUERY( (y | mask = y) =>
500 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
504 // Please note that each pattern must be a dual implication (<--> or
505 // iff). One directional implication can create spurious matches. If the
506 // implication is only one-way, an unsatisfiable condition on the left
507 // side can imply a satisfiable condition on the right side. Dual
508 // implication ensures that satisfiable conditions are transformed to
509 // other satisfiable conditions and unsatisfiable conditions are
510 // transformed to other unsatisfiable conditions.
512 // Here is a concrete example of a unsatisfiable condition on the left
513 // implying a satisfiable condition on the right:
516 // (x & ~mask) == y --> (x == y || x == (y | mask))
518 // Substituting y = 3, z = 0 yields:
519 // (x & -2) == 3 --> (x == 3 || x == 2)
521 // Pattern match a special case:
523 QUERY( (y & ~mask = y) =>
524 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
527 if (match(ICI->getOperand(0),
528 m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
530 if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
531 // If we already have a value for the switch, it has to match!
532 if (!setValueOnce(RHSVal))
537 ConstantInt::get(C->getContext(),
538 C->getValue() | Mask));
544 // Pattern match a special case:
546 QUERY( (y | mask = y) =>
547 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
550 if (match(ICI->getOperand(0),
551 m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
553 if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
554 // If we already have a value for the switch, it has to match!
555 if (!setValueOnce(RHSVal))
559 Vals.push_back(ConstantInt::get(C->getContext(),
560 C->getValue() & ~Mask));
566 // If we already have a value for the switch, it has to match!
567 if (!setValueOnce(ICI->getOperand(0)))
572 return ICI->getOperand(0);
575 // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
576 ConstantRange Span = ConstantRange::makeAllowedICmpRegion(
577 ICI->getPredicate(), C->getValue());
579 // Shift the range if the compare is fed by an add. This is the range
580 // compare idiom as emitted by instcombine.
581 Value *CandidateVal = I->getOperand(0);
582 if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
583 Span = Span.subtract(*RHSC);
584 CandidateVal = RHSVal;
587 // If this is an and/!= check, then we are looking to build the set of
588 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
591 Span = Span.inverse();
593 // If there are a ton of values, we don't want to make a ginormous switch.
594 if (Span.getSetSize().ugt(8) || Span.isEmptySet()) {
598 // If we already have a value for the switch, it has to match!
599 if (!setValueOnce(CandidateVal))
602 // Add all values from the range to the set
603 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
604 Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
610 /// Given a potentially 'or'd or 'and'd together collection of icmp
611 /// eq/ne/lt/gt instructions that compare a value against a constant, extract
612 /// the value being compared, and stick the list constants into the Vals
614 /// One "Extra" case is allowed to differ from the other.
615 void gather(Value *V) {
616 Instruction *I = dyn_cast<Instruction>(V);
617 bool isEQ = (I->getOpcode() == Instruction::Or);
619 // Keep a stack (SmallVector for efficiency) for depth-first traversal
620 SmallVector<Value *, 8> DFT;
621 SmallPtrSet<Value *, 8> Visited;
627 while (!DFT.empty()) {
628 V = DFT.pop_back_val();
630 if (Instruction *I = dyn_cast<Instruction>(V)) {
631 // If it is a || (or && depending on isEQ), process the operands.
632 if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
633 if (Visited.insert(I->getOperand(1)).second)
634 DFT.push_back(I->getOperand(1));
635 if (Visited.insert(I->getOperand(0)).second)
636 DFT.push_back(I->getOperand(0));
640 // Try to match the current instruction
641 if (matchInstruction(I, isEQ))
642 // Match succeed, continue the loop
646 // One element of the sequence of || (or &&) could not be match as a
647 // comparison against the same value as the others.
648 // We allow only one "Extra" case to be checked before the switch
653 // Failed to parse a proper sequence, abort now
660 } // end anonymous namespace
662 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
663 Instruction *Cond = nullptr;
664 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
665 Cond = dyn_cast<Instruction>(SI->getCondition());
666 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
667 if (BI->isConditional())
668 Cond = dyn_cast<Instruction>(BI->getCondition());
669 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
670 Cond = dyn_cast<Instruction>(IBI->getAddress());
673 TI->eraseFromParent();
675 RecursivelyDeleteTriviallyDeadInstructions(Cond);
678 /// Return true if the specified terminator checks
679 /// to see if a value is equal to constant integer value.
680 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
682 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
683 // Do not permit merging of large switch instructions into their
684 // predecessors unless there is only one predecessor.
685 if (SI->getNumSuccessors() * std::distance(pred_begin(SI->getParent()),
686 pred_end(SI->getParent())) <=
688 CV = SI->getCondition();
689 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
690 if (BI->isConditional() && BI->getCondition()->hasOneUse())
691 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
692 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
693 CV = ICI->getOperand(0);
696 // Unwrap any lossless ptrtoint cast.
698 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
699 Value *Ptr = PTII->getPointerOperand();
700 if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
707 /// Given a value comparison instruction,
708 /// decode all of the 'cases' that it represents and return the 'default' block.
709 BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
710 TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
711 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
712 Cases.reserve(SI->getNumCases());
713 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e;
716 ValueEqualityComparisonCase(i.getCaseValue(), i.getCaseSuccessor()));
717 return SI->getDefaultDest();
720 BranchInst *BI = cast<BranchInst>(TI);
721 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
722 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
723 Cases.push_back(ValueEqualityComparisonCase(
724 GetConstantInt(ICI->getOperand(1), DL), Succ));
725 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
728 /// Given a vector of bb/value pairs, remove any entries
729 /// in the list that match the specified block.
731 EliminateBlockCases(BasicBlock *BB,
732 std::vector<ValueEqualityComparisonCase> &Cases) {
733 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
736 /// Return true if there are any keys in C1 that exist in C2 as well.
737 static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
738 std::vector<ValueEqualityComparisonCase> &C2) {
739 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
741 // Make V1 be smaller than V2.
742 if (V1->size() > V2->size())
747 if (V1->size() == 1) {
749 ConstantInt *TheVal = (*V1)[0].Value;
750 for (unsigned i = 0, e = V2->size(); i != e; ++i)
751 if (TheVal == (*V2)[i].Value)
755 // Otherwise, just sort both lists and compare element by element.
756 array_pod_sort(V1->begin(), V1->end());
757 array_pod_sort(V2->begin(), V2->end());
758 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
759 while (i1 != e1 && i2 != e2) {
760 if ((*V1)[i1].Value == (*V2)[i2].Value)
762 if ((*V1)[i1].Value < (*V2)[i2].Value)
770 /// If TI is known to be a terminator instruction and its block is known to
771 /// only have a single predecessor block, check to see if that predecessor is
772 /// also a value comparison with the same value, and if that comparison
773 /// determines the outcome of this comparison. If so, simplify TI. This does a
774 /// very limited form of jump threading.
775 bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
776 TerminatorInst *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
777 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
779 return false; // Not a value comparison in predecessor.
781 Value *ThisVal = isValueEqualityComparison(TI);
782 assert(ThisVal && "This isn't a value comparison!!");
783 if (ThisVal != PredVal)
784 return false; // Different predicates.
786 // TODO: Preserve branch weight metadata, similarly to how
787 // FoldValueComparisonIntoPredecessors preserves it.
789 // Find out information about when control will move from Pred to TI's block.
790 std::vector<ValueEqualityComparisonCase> PredCases;
791 BasicBlock *PredDef =
792 GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
793 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
795 // Find information about how control leaves this block.
796 std::vector<ValueEqualityComparisonCase> ThisCases;
797 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
798 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
800 // If TI's block is the default block from Pred's comparison, potentially
801 // simplify TI based on this knowledge.
802 if (PredDef == TI->getParent()) {
803 // If we are here, we know that the value is none of those cases listed in
804 // PredCases. If there are any cases in ThisCases that are in PredCases, we
806 if (!ValuesOverlap(PredCases, ThisCases))
809 if (isa<BranchInst>(TI)) {
810 // Okay, one of the successors of this condbr is dead. Convert it to a
812 assert(ThisCases.size() == 1 && "Branch can only have one case!");
813 // Insert the new branch.
814 Instruction *NI = Builder.CreateBr(ThisDef);
817 // Remove PHI node entries for the dead edge.
818 ThisCases[0].Dest->removePredecessor(TI->getParent());
820 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
821 << "Through successor TI: " << *TI << "Leaving: " << *NI
824 EraseTerminatorInstAndDCECond(TI);
828 SwitchInst *SI = cast<SwitchInst>(TI);
829 // Okay, TI has cases that are statically dead, prune them away.
830 SmallPtrSet<Constant *, 16> DeadCases;
831 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
832 DeadCases.insert(PredCases[i].Value);
834 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
835 << "Through successor TI: " << *TI);
837 // Collect branch weights into a vector.
838 SmallVector<uint32_t, 8> Weights;
839 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
840 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
842 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
844 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
845 Weights.push_back(CI->getValue().getZExtValue());
847 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
849 if (DeadCases.count(i.getCaseValue())) {
851 std::swap(Weights[i.getCaseIndex() + 1], Weights.back());
854 i.getCaseSuccessor()->removePredecessor(TI->getParent());
858 if (HasWeight && Weights.size() >= 2)
859 SI->setMetadata(LLVMContext::MD_prof,
860 MDBuilder(SI->getParent()->getContext())
861 .createBranchWeights(Weights));
863 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
867 // Otherwise, TI's block must correspond to some matched value. Find out
868 // which value (or set of values) this is.
869 ConstantInt *TIV = nullptr;
870 BasicBlock *TIBB = TI->getParent();
871 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
872 if (PredCases[i].Dest == TIBB) {
874 return false; // Cannot handle multiple values coming to this block.
875 TIV = PredCases[i].Value;
877 assert(TIV && "No edge from pred to succ?");
879 // Okay, we found the one constant that our value can be if we get into TI's
880 // BB. Find out which successor will unconditionally be branched to.
881 BasicBlock *TheRealDest = nullptr;
882 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
883 if (ThisCases[i].Value == TIV) {
884 TheRealDest = ThisCases[i].Dest;
888 // If not handled by any explicit cases, it is handled by the default case.
890 TheRealDest = ThisDef;
892 // Remove PHI node entries for dead edges.
893 BasicBlock *CheckEdge = TheRealDest;
894 for (BasicBlock *Succ : successors(TIBB))
895 if (Succ != CheckEdge)
896 Succ->removePredecessor(TIBB);
900 // Insert the new branch.
901 Instruction *NI = Builder.CreateBr(TheRealDest);
904 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
905 << "Through successor TI: " << *TI << "Leaving: " << *NI
908 EraseTerminatorInstAndDCECond(TI);
914 /// This class implements a stable ordering of constant
915 /// integers that does not depend on their address. This is important for
916 /// applications that sort ConstantInt's to ensure uniqueness.
917 struct ConstantIntOrdering {
918 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
919 return LHS->getValue().ult(RHS->getValue());
923 } // end anonymous namespace
925 static int ConstantIntSortPredicate(ConstantInt *const *P1,
926 ConstantInt *const *P2) {
927 const ConstantInt *LHS = *P1;
928 const ConstantInt *RHS = *P2;
931 return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
934 static inline bool HasBranchWeights(const Instruction *I) {
935 MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
936 if (ProfMD && ProfMD->getOperand(0))
937 if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
938 return MDS->getString().equals("branch_weights");
943 /// Get Weights of a given TerminatorInst, the default weight is at the front
944 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
946 static void GetBranchWeights(TerminatorInst *TI,
947 SmallVectorImpl<uint64_t> &Weights) {
948 MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
950 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
951 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
952 Weights.push_back(CI->getValue().getZExtValue());
955 // If TI is a conditional eq, the default case is the false case,
956 // and the corresponding branch-weight data is at index 2. We swap the
957 // default weight to be the first entry.
958 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
959 assert(Weights.size() == 2);
960 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
961 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
962 std::swap(Weights.front(), Weights.back());
966 /// Keep halving the weights until all can fit in uint32_t.
967 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
968 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
969 if (Max > UINT_MAX) {
970 unsigned Offset = 32 - countLeadingZeros(Max);
971 for (uint64_t &I : Weights)
976 /// The specified terminator is a value equality comparison instruction
977 /// (either a switch or a branch on "X == c").
978 /// See if any of the predecessors of the terminator block are value comparisons
979 /// on the same value. If so, and if safe to do so, fold them together.
980 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
981 IRBuilder<> &Builder) {
982 BasicBlock *BB = TI->getParent();
983 Value *CV = isValueEqualityComparison(TI); // CondVal
984 assert(CV && "Not a comparison?");
985 bool Changed = false;
987 SmallVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
988 while (!Preds.empty()) {
989 BasicBlock *Pred = Preds.pop_back_val();
991 // See if the predecessor is a comparison with the same value.
992 TerminatorInst *PTI = Pred->getTerminator();
993 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
995 if (PCV == CV && TI != PTI) {
996 SmallSetVector<BasicBlock*, 4> FailBlocks;
997 if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
998 for (auto *Succ : FailBlocks) {
999 std::vector<BasicBlock*> Blocks = { TI->getParent() };
1000 if (!SplitBlockPredecessors(Succ, Blocks, ".fold.split"))
1005 // Figure out which 'cases' to copy from SI to PSI.
1006 std::vector<ValueEqualityComparisonCase> BBCases;
1007 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1009 std::vector<ValueEqualityComparisonCase> PredCases;
1010 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1012 // Based on whether the default edge from PTI goes to BB or not, fill in
1013 // PredCases and PredDefault with the new switch cases we would like to
1015 SmallVector<BasicBlock *, 8> NewSuccessors;
1017 // Update the branch weight metadata along the way
1018 SmallVector<uint64_t, 8> Weights;
1019 bool PredHasWeights = HasBranchWeights(PTI);
1020 bool SuccHasWeights = HasBranchWeights(TI);
1022 if (PredHasWeights) {
1023 GetBranchWeights(PTI, Weights);
1024 // branch-weight metadata is inconsistent here.
1025 if (Weights.size() != 1 + PredCases.size())
1026 PredHasWeights = SuccHasWeights = false;
1027 } else if (SuccHasWeights)
1028 // If there are no predecessor weights but there are successor weights,
1029 // populate Weights with 1, which will later be scaled to the sum of
1030 // successor's weights
1031 Weights.assign(1 + PredCases.size(), 1);
1033 SmallVector<uint64_t, 8> SuccWeights;
1034 if (SuccHasWeights) {
1035 GetBranchWeights(TI, SuccWeights);
1036 // branch-weight metadata is inconsistent here.
1037 if (SuccWeights.size() != 1 + BBCases.size())
1038 PredHasWeights = SuccHasWeights = false;
1039 } else if (PredHasWeights)
1040 SuccWeights.assign(1 + BBCases.size(), 1);
1042 if (PredDefault == BB) {
1043 // If this is the default destination from PTI, only the edges in TI
1044 // that don't occur in PTI, or that branch to BB will be activated.
1045 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1046 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1047 if (PredCases[i].Dest != BB)
1048 PTIHandled.insert(PredCases[i].Value);
1050 // The default destination is BB, we don't need explicit targets.
1051 std::swap(PredCases[i], PredCases.back());
1053 if (PredHasWeights || SuccHasWeights) {
1054 // Increase weight for the default case.
1055 Weights[0] += Weights[i + 1];
1056 std::swap(Weights[i + 1], Weights.back());
1060 PredCases.pop_back();
1065 // Reconstruct the new switch statement we will be building.
1066 if (PredDefault != BBDefault) {
1067 PredDefault->removePredecessor(Pred);
1068 PredDefault = BBDefault;
1069 NewSuccessors.push_back(BBDefault);
1072 unsigned CasesFromPred = Weights.size();
1073 uint64_t ValidTotalSuccWeight = 0;
1074 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1075 if (!PTIHandled.count(BBCases[i].Value) &&
1076 BBCases[i].Dest != BBDefault) {
1077 PredCases.push_back(BBCases[i]);
1078 NewSuccessors.push_back(BBCases[i].Dest);
1079 if (SuccHasWeights || PredHasWeights) {
1080 // The default weight is at index 0, so weight for the ith case
1081 // should be at index i+1. Scale the cases from successor by
1082 // PredDefaultWeight (Weights[0]).
1083 Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1084 ValidTotalSuccWeight += SuccWeights[i + 1];
1088 if (SuccHasWeights || PredHasWeights) {
1089 ValidTotalSuccWeight += SuccWeights[0];
1090 // Scale the cases from predecessor by ValidTotalSuccWeight.
1091 for (unsigned i = 1; i < CasesFromPred; ++i)
1092 Weights[i] *= ValidTotalSuccWeight;
1093 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1094 Weights[0] *= SuccWeights[0];
1097 // If this is not the default destination from PSI, only the edges
1098 // in SI that occur in PSI with a destination of BB will be
1100 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1101 std::map<ConstantInt *, uint64_t> WeightsForHandled;
1102 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1103 if (PredCases[i].Dest == BB) {
1104 PTIHandled.insert(PredCases[i].Value);
1106 if (PredHasWeights || SuccHasWeights) {
1107 WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1108 std::swap(Weights[i + 1], Weights.back());
1112 std::swap(PredCases[i], PredCases.back());
1113 PredCases.pop_back();
1118 // Okay, now we know which constants were sent to BB from the
1119 // predecessor. Figure out where they will all go now.
1120 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1121 if (PTIHandled.count(BBCases[i].Value)) {
1122 // If this is one we are capable of getting...
1123 if (PredHasWeights || SuccHasWeights)
1124 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1125 PredCases.push_back(BBCases[i]);
1126 NewSuccessors.push_back(BBCases[i].Dest);
1128 BBCases[i].Value); // This constant is taken care of
1131 // If there are any constants vectored to BB that TI doesn't handle,
1132 // they must go to the default destination of TI.
1133 for (ConstantInt *I : PTIHandled) {
1134 if (PredHasWeights || SuccHasWeights)
1135 Weights.push_back(WeightsForHandled[I]);
1136 PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1137 NewSuccessors.push_back(BBDefault);
1141 // Okay, at this point, we know which new successor Pred will get. Make
1142 // sure we update the number of entries in the PHI nodes for these
1144 for (BasicBlock *NewSuccessor : NewSuccessors)
1145 AddPredecessorToBlock(NewSuccessor, Pred, BB);
1147 Builder.SetInsertPoint(PTI);
1148 // Convert pointer to int before we switch.
1149 if (CV->getType()->isPointerTy()) {
1150 CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
1154 // Now that the successors are updated, create the new Switch instruction.
1156 Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1157 NewSI->setDebugLoc(PTI->getDebugLoc());
1158 for (ValueEqualityComparisonCase &V : PredCases)
1159 NewSI->addCase(V.Value, V.Dest);
1161 if (PredHasWeights || SuccHasWeights) {
1162 // Halve the weights if any of them cannot fit in an uint32_t
1163 FitWeights(Weights);
1165 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1168 LLVMContext::MD_prof,
1169 MDBuilder(BB->getContext()).createBranchWeights(MDWeights));
1172 EraseTerminatorInstAndDCECond(PTI);
1174 // Okay, last check. If BB is still a successor of PSI, then we must
1175 // have an infinite loop case. If so, add an infinitely looping block
1176 // to handle the case to preserve the behavior of the code.
1177 BasicBlock *InfLoopBlock = nullptr;
1178 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1179 if (NewSI->getSuccessor(i) == BB) {
1180 if (!InfLoopBlock) {
1181 // Insert it at the end of the function, because it's either code,
1182 // or it won't matter if it's hot. :)
1183 InfLoopBlock = BasicBlock::Create(BB->getContext(), "infloop",
1185 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1187 NewSI->setSuccessor(i, InfLoopBlock);
1196 // If we would need to insert a select that uses the value of this invoke
1197 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1198 // can't hoist the invoke, as there is nowhere to put the select in this case.
1199 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1200 Instruction *I1, Instruction *I2) {
1201 for (BasicBlock *Succ : successors(BB1)) {
1203 for (BasicBlock::iterator BBI = Succ->begin();
1204 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1205 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1206 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1207 if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1215 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I);
1217 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1218 /// in the two blocks up into the branch block. The caller of this function
1219 /// guarantees that BI's block dominates BB1 and BB2.
1220 static bool HoistThenElseCodeToIf(BranchInst *BI,
1221 const TargetTransformInfo &TTI) {
1222 // This does very trivial matching, with limited scanning, to find identical
1223 // instructions in the two blocks. In particular, we don't want to get into
1224 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1225 // such, we currently just scan for obviously identical instructions in an
1227 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1228 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1230 BasicBlock::iterator BB1_Itr = BB1->begin();
1231 BasicBlock::iterator BB2_Itr = BB2->begin();
1233 Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1234 // Skip debug info if it is not identical.
1235 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1236 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1237 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1238 while (isa<DbgInfoIntrinsic>(I1))
1240 while (isa<DbgInfoIntrinsic>(I2))
1243 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1244 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1247 BasicBlock *BIParent = BI->getParent();
1249 bool Changed = false;
1251 // If we are hoisting the terminator instruction, don't move one (making a
1252 // broken BB), instead clone it, and remove BI.
1253 if (isa<TerminatorInst>(I1))
1254 goto HoistTerminator;
1256 if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1259 // For a normal instruction, we just move one to right before the branch,
1260 // then replace all uses of the other with the first. Finally, we remove
1261 // the now redundant second instruction.
1262 BIParent->getInstList().splice(BI->getIterator(), BB1->getInstList(), I1);
1263 if (!I2->use_empty())
1264 I2->replaceAllUsesWith(I1);
1266 unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1267 LLVMContext::MD_range,
1268 LLVMContext::MD_fpmath,
1269 LLVMContext::MD_invariant_load,
1270 LLVMContext::MD_nonnull,
1271 LLVMContext::MD_invariant_group,
1272 LLVMContext::MD_align,
1273 LLVMContext::MD_dereferenceable,
1274 LLVMContext::MD_dereferenceable_or_null,
1275 LLVMContext::MD_mem_parallel_loop_access};
1276 combineMetadata(I1, I2, KnownIDs);
1278 // I1 and I2 are being combined into a single instruction. Its debug
1279 // location is the merged locations of the original instructions.
1280 if (!isa<CallInst>(I1))
1282 DILocation::getMergedLocation(I1->getDebugLoc(), I2->getDebugLoc()));
1284 I2->eraseFromParent();
1289 // Skip debug info if it is not identical.
1290 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1291 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1292 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1293 while (isa<DbgInfoIntrinsic>(I1))
1295 while (isa<DbgInfoIntrinsic>(I2))
1298 } while (I1->isIdenticalToWhenDefined(I2));
1303 // It may not be possible to hoist an invoke.
1304 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1307 for (BasicBlock *Succ : successors(BB1)) {
1309 for (BasicBlock::iterator BBI = Succ->begin();
1310 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1311 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1312 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1316 // Check for passingValueIsAlwaysUndefined here because we would rather
1317 // eliminate undefined control flow then converting it to a select.
1318 if (passingValueIsAlwaysUndefined(BB1V, PN) ||
1319 passingValueIsAlwaysUndefined(BB2V, PN))
1322 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1324 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1329 // Okay, it is safe to hoist the terminator.
1330 Instruction *NT = I1->clone();
1331 BIParent->getInstList().insert(BI->getIterator(), NT);
1332 if (!NT->getType()->isVoidTy()) {
1333 I1->replaceAllUsesWith(NT);
1334 I2->replaceAllUsesWith(NT);
1338 IRBuilder<NoFolder> Builder(NT);
1339 // Hoisting one of the terminators from our successor is a great thing.
1340 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1341 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1342 // nodes, so we insert select instruction to compute the final result.
1343 std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1344 for (BasicBlock *Succ : successors(BB1)) {
1346 for (BasicBlock::iterator BBI = Succ->begin();
1347 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1348 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1349 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1353 // These values do not agree. Insert a select instruction before NT
1354 // that determines the right value.
1355 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1357 SI = cast<SelectInst>(
1358 Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1359 BB1V->getName() + "." + BB2V->getName(), BI));
1361 // Make the PHI node use the select for all incoming values for BB1/BB2
1362 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1363 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1364 PN->setIncomingValue(i, SI);
1368 // Update any PHI nodes in our new successors.
1369 for (BasicBlock *Succ : successors(BB1))
1370 AddPredecessorToBlock(Succ, BIParent, BB1);
1372 EraseTerminatorInstAndDCECond(BI);
1376 // Is it legal to place a variable in operand \c OpIdx of \c I?
1377 // FIXME: This should be promoted to Instruction.
1378 static bool canReplaceOperandWithVariable(const Instruction *I,
1380 // We can't have a PHI with a metadata type.
1381 if (I->getOperand(OpIdx)->getType()->isMetadataTy())
1385 if (!isa<Constant>(I->getOperand(OpIdx)))
1388 switch (I->getOpcode()) {
1391 case Instruction::Call:
1392 case Instruction::Invoke:
1393 // FIXME: many arithmetic intrinsics have no issue taking a
1394 // variable, however it's hard to distingish these from
1395 // specials such as @llvm.frameaddress that require a constant.
1396 if (isa<IntrinsicInst>(I))
1399 // Constant bundle operands may need to retain their constant-ness for
1401 if (ImmutableCallSite(I).isBundleOperand(OpIdx))
1406 case Instruction::ShuffleVector:
1407 // Shufflevector masks are constant.
1409 case Instruction::ExtractValue:
1410 case Instruction::InsertValue:
1411 // All operands apart from the first are constant.
1413 case Instruction::Alloca:
1415 case Instruction::GetElementPtr:
1418 gep_type_iterator It = std::next(gep_type_begin(I), OpIdx - 1);
1419 return It.isSequential();
1423 // All instructions in Insts belong to different blocks that all unconditionally
1424 // branch to a common successor. Analyze each instruction and return true if it
1425 // would be possible to sink them into their successor, creating one common
1426 // instruction instead. For every value that would be required to be provided by
1427 // PHI node (because an operand varies in each input block), add to PHIOperands.
1428 static bool canSinkInstructions(
1429 ArrayRef<Instruction *> Insts,
1430 DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1431 // Prune out obviously bad instructions to move. Any non-store instruction
1432 // must have exactly one use, and we check later that use is by a single,
1433 // common PHI instruction in the successor.
1434 for (auto *I : Insts) {
1435 // These instructions may change or break semantics if moved.
1436 if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1437 I->getType()->isTokenTy())
1439 // Everything must have only one use too, apart from stores which
1441 if (!isa<StoreInst>(I) && !I->hasOneUse())
1445 const Instruction *I0 = Insts.front();
1446 for (auto *I : Insts)
1447 if (!I->isSameOperationAs(I0))
1450 // All instructions in Insts are known to be the same opcode. If they aren't
1451 // stores, check the only user of each is a PHI or in the same block as the
1452 // instruction, because if a user is in the same block as an instruction
1453 // we're contemplating sinking, it must already be determined to be sinkable.
1454 if (!isa<StoreInst>(I0)) {
1455 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1456 auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1457 if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1458 auto *U = cast<Instruction>(*I->user_begin());
1460 PNUse->getParent() == Succ &&
1461 PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1462 U->getParent() == I->getParent();
1467 for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1468 if (I0->getOperand(OI)->getType()->isTokenTy())
1469 // Don't touch any operand of token type.
1472 // Because SROA can't handle speculating stores of selects, try not
1473 // to sink loads or stores of allocas when we'd have to create a PHI for
1474 // the address operand. Also, because it is likely that loads or stores
1475 // of allocas will disappear when Mem2Reg/SROA is run, don't sink them.
1476 // This can cause code churn which can have unintended consequences down
1477 // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1478 // FIXME: This is a workaround for a deficiency in SROA - see
1479 // https://llvm.org/bugs/show_bug.cgi?id=30188
1480 if (OI == 1 && isa<StoreInst>(I0) &&
1481 any_of(Insts, [](const Instruction *I) {
1482 return isa<AllocaInst>(I->getOperand(1));
1485 if (OI == 0 && isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1486 return isa<AllocaInst>(I->getOperand(0));
1490 auto SameAsI0 = [&I0, OI](const Instruction *I) {
1491 assert(I->getNumOperands() == I0->getNumOperands());
1492 return I->getOperand(OI) == I0->getOperand(OI);
1494 if (!all_of(Insts, SameAsI0)) {
1495 if (!canReplaceOperandWithVariable(I0, OI))
1496 // We can't create a PHI from this GEP.
1498 // Don't create indirect calls! The called value is the final operand.
1499 if ((isa<CallInst>(I0) || isa<InvokeInst>(I0)) && OI == OE - 1) {
1500 // FIXME: if the call was *already* indirect, we should do this.
1503 for (auto *I : Insts)
1504 PHIOperands[I].push_back(I->getOperand(OI));
1510 // Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1511 // instruction of every block in Blocks to their common successor, commoning
1512 // into one instruction.
1513 static bool sinkLastInstruction(ArrayRef<BasicBlock*> Blocks) {
1514 auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1516 // canSinkLastInstruction returning true guarantees that every block has at
1517 // least one non-terminator instruction.
1518 SmallVector<Instruction*,4> Insts;
1519 for (auto *BB : Blocks) {
1520 Instruction *I = BB->getTerminator();
1522 I = I->getPrevNode();
1523 } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1524 if (!isa<DbgInfoIntrinsic>(I))
1528 // The only checking we need to do now is that all users of all instructions
1529 // are the same PHI node. canSinkLastInstruction should have checked this but
1530 // it is slightly over-aggressive - it gets confused by commutative instructions
1531 // so double-check it here.
1532 Instruction *I0 = Insts.front();
1533 if (!isa<StoreInst>(I0)) {
1534 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1535 if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1536 auto *U = cast<Instruction>(*I->user_begin());
1542 // We don't need to do any more checking here; canSinkLastInstruction should
1543 // have done it all for us.
1544 SmallVector<Value*, 4> NewOperands;
1545 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1546 // This check is different to that in canSinkLastInstruction. There, we
1547 // cared about the global view once simplifycfg (and instcombine) have
1548 // completed - it takes into account PHIs that become trivially
1549 // simplifiable. However here we need a more local view; if an operand
1550 // differs we create a PHI and rely on instcombine to clean up the very
1551 // small mess we may make.
1552 bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1553 return I->getOperand(O) != I0->getOperand(O);
1556 NewOperands.push_back(I0->getOperand(O));
1560 // Create a new PHI in the successor block and populate it.
1561 auto *Op = I0->getOperand(O);
1562 assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1563 auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1564 Op->getName() + ".sink", &BBEnd->front());
1565 for (auto *I : Insts)
1566 PN->addIncoming(I->getOperand(O), I->getParent());
1567 NewOperands.push_back(PN);
1570 // Arbitrarily use I0 as the new "common" instruction; remap its operands
1571 // and move it to the start of the successor block.
1572 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1573 I0->getOperandUse(O).set(NewOperands[O]);
1574 I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1576 // The debug location for the "common" instruction is the merged locations of
1577 // all the commoned instructions. We start with the original location of the
1578 // "common" instruction and iteratively merge each location in the loop below.
1579 const DILocation *Loc = I0->getDebugLoc();
1581 // Update metadata and IR flags, and merge debug locations.
1582 for (auto *I : Insts)
1584 Loc = DILocation::getMergedLocation(Loc, I->getDebugLoc());
1585 combineMetadataForCSE(I0, I);
1588 if (!isa<CallInst>(I0))
1589 I0->setDebugLoc(Loc);
1591 if (!isa<StoreInst>(I0)) {
1592 // canSinkLastInstruction checked that all instructions were used by
1593 // one and only one PHI node. Find that now, RAUW it to our common
1594 // instruction and nuke it.
1595 assert(I0->hasOneUse());
1596 auto *PN = cast<PHINode>(*I0->user_begin());
1597 PN->replaceAllUsesWith(I0);
1598 PN->eraseFromParent();
1601 // Finally nuke all instructions apart from the common instruction.
1602 for (auto *I : Insts)
1604 I->eraseFromParent();
1611 // LockstepReverseIterator - Iterates through instructions
1612 // in a set of blocks in reverse order from the first non-terminator.
1613 // For example (assume all blocks have size n):
1614 // LockstepReverseIterator I([B1, B2, B3]);
1615 // *I-- = [B1[n], B2[n], B3[n]];
1616 // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1617 // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1619 class LockstepReverseIterator {
1620 ArrayRef<BasicBlock*> Blocks;
1621 SmallVector<Instruction*,4> Insts;
1624 LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) :
1632 for (auto *BB : Blocks) {
1633 Instruction *Inst = BB->getTerminator();
1634 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1635 Inst = Inst->getPrevNode();
1637 // Block wasn't big enough.
1641 Insts.push_back(Inst);
1645 bool isValid() const {
1649 void operator -- () {
1652 for (auto *&Inst : Insts) {
1653 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1654 Inst = Inst->getPrevNode();
1655 // Already at beginning of block.
1663 ArrayRef<Instruction*> operator * () const {
1668 } // end anonymous namespace
1670 /// Given an unconditional branch that goes to BBEnd,
1671 /// check whether BBEnd has only two predecessors and the other predecessor
1672 /// ends with an unconditional branch. If it is true, sink any common code
1673 /// in the two predecessors to BBEnd.
1674 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1675 assert(BI1->isUnconditional());
1676 BasicBlock *BBEnd = BI1->getSuccessor(0);
1678 // We support two situations:
1679 // (1) all incoming arcs are unconditional
1680 // (2) one incoming arc is conditional
1682 // (2) is very common in switch defaults and
1683 // else-if patterns;
1686 // else if (b) f(2);
1699 // [end] has two unconditional predecessor arcs and one conditional. The
1700 // conditional refers to the implicit empty 'else' arc. This conditional
1701 // arc can also be caused by an empty default block in a switch.
1703 // In this case, we attempt to sink code from all *unconditional* arcs.
1704 // If we can sink instructions from these arcs (determined during the scan
1705 // phase below) we insert a common successor for all unconditional arcs and
1706 // connect that to [end], to enable sinking:
1719 SmallVector<BasicBlock*,4> UnconditionalPreds;
1720 Instruction *Cond = nullptr;
1721 for (auto *B : predecessors(BBEnd)) {
1722 auto *T = B->getTerminator();
1723 if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1724 UnconditionalPreds.push_back(B);
1725 else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1730 if (UnconditionalPreds.size() < 2)
1733 bool Changed = false;
1734 // We take a two-step approach to tail sinking. First we scan from the end of
1735 // each block upwards in lockstep. If the n'th instruction from the end of each
1736 // block can be sunk, those instructions are added to ValuesToSink and we
1737 // carry on. If we can sink an instruction but need to PHI-merge some operands
1738 // (because they're not identical in each instruction) we add these to
1740 unsigned ScanIdx = 0;
1741 SmallPtrSet<Value*,4> InstructionsToSink;
1742 DenseMap<Instruction*, SmallVector<Value*,4>> PHIOperands;
1743 LockstepReverseIterator LRI(UnconditionalPreds);
1744 while (LRI.isValid() &&
1745 canSinkInstructions(*LRI, PHIOperands)) {
1746 DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0] << "\n");
1747 InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1752 auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1753 unsigned NumPHIdValues = 0;
1754 for (auto *I : *LRI)
1755 for (auto *V : PHIOperands[I])
1756 if (InstructionsToSink.count(V) == 0)
1758 DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1759 unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1760 if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1763 return NumPHIInsts <= 1;
1766 if (ScanIdx > 0 && Cond) {
1767 // Check if we would actually sink anything first! This mutates the CFG and
1768 // adds an extra block. The goal in doing this is to allow instructions that
1769 // couldn't be sunk before to be sunk - obviously, speculatable instructions
1770 // (such as trunc, add) can be sunk and predicated already. So we check that
1771 // we're going to sink at least one non-speculatable instruction.
1774 bool Profitable = false;
1775 while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1776 if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1786 DEBUG(dbgs() << "SINK: Splitting edge\n");
1787 // We have a conditional edge and we're going to sink some instructions.
1788 // Insert a new block postdominating all blocks we're going to sink from.
1789 if (!SplitBlockPredecessors(BI1->getSuccessor(0), UnconditionalPreds,
1791 // Edges couldn't be split.
1796 // Now that we've analyzed all potential sinking candidates, perform the
1797 // actual sink. We iteratively sink the last non-terminator of the source
1798 // blocks into their common successor unless doing so would require too
1799 // many PHI instructions to be generated (currently only one PHI is allowed
1800 // per sunk instruction).
1802 // We can use InstructionsToSink to discount values needing PHI-merging that will
1803 // actually be sunk in a later iteration. This allows us to be more
1804 // aggressive in what we sink. This does allow a false positive where we
1805 // sink presuming a later value will also be sunk, but stop half way through
1806 // and never actually sink it which means we produce more PHIs than intended.
1807 // This is unlikely in practice though.
1808 for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1809 DEBUG(dbgs() << "SINK: Sink: "
1810 << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1813 // Because we've sunk every instruction in turn, the current instruction to
1814 // sink is always at index 0.
1816 if (!ProfitableToSinkInstruction(LRI)) {
1817 // Too many PHIs would be created.
1818 DEBUG(dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1822 if (!sinkLastInstruction(UnconditionalPreds))
1830 /// \brief Determine if we can hoist sink a sole store instruction out of a
1831 /// conditional block.
1833 /// We are looking for code like the following:
1835 /// store i32 %add, i32* %arrayidx2
1836 /// ... // No other stores or function calls (we could be calling a memory
1837 /// ... // function).
1838 /// %cmp = icmp ult %x, %y
1839 /// br i1 %cmp, label %EndBB, label %ThenBB
1841 /// store i32 %add5, i32* %arrayidx2
1845 /// We are going to transform this into:
1847 /// store i32 %add, i32* %arrayidx2
1849 /// %cmp = icmp ult %x, %y
1850 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1851 /// store i32 %add.add5, i32* %arrayidx2
1854 /// \return The pointer to the value of the previous store if the store can be
1855 /// hoisted into the predecessor block. 0 otherwise.
1856 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1857 BasicBlock *StoreBB, BasicBlock *EndBB) {
1858 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1862 // Volatile or atomic.
1863 if (!StoreToHoist->isSimple())
1866 Value *StorePtr = StoreToHoist->getPointerOperand();
1868 // Look for a store to the same pointer in BrBB.
1869 unsigned MaxNumInstToLookAt = 9;
1870 for (Instruction &CurI : reverse(*BrBB)) {
1871 if (!MaxNumInstToLookAt)
1874 if (isa<DbgInfoIntrinsic>(CurI))
1876 --MaxNumInstToLookAt;
1878 // Could be calling an instruction that affects memory like free().
1879 if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1882 if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1883 // Found the previous store make sure it stores to the same location.
1884 if (SI->getPointerOperand() == StorePtr)
1885 // Found the previous store, return its value operand.
1886 return SI->getValueOperand();
1887 return nullptr; // Unknown store.
1894 /// \brief Speculate a conditional basic block flattening the CFG.
1896 /// Note that this is a very risky transform currently. Speculating
1897 /// instructions like this is most often not desirable. Instead, there is an MI
1898 /// pass which can do it with full awareness of the resource constraints.
1899 /// However, some cases are "obvious" and we should do directly. An example of
1900 /// this is speculating a single, reasonably cheap instruction.
1902 /// There is only one distinct advantage to flattening the CFG at the IR level:
1903 /// it makes very common but simplistic optimizations such as are common in
1904 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1905 /// modeling their effects with easier to reason about SSA value graphs.
1908 /// An illustration of this transform is turning this IR:
1911 /// %cmp = icmp ult %x, %y
1912 /// br i1 %cmp, label %EndBB, label %ThenBB
1914 /// %sub = sub %x, %y
1917 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1924 /// %cmp = icmp ult %x, %y
1925 /// %sub = sub %x, %y
1926 /// %cond = select i1 %cmp, 0, %sub
1930 /// \returns true if the conditional block is removed.
1931 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1932 const TargetTransformInfo &TTI) {
1933 // Be conservative for now. FP select instruction can often be expensive.
1934 Value *BrCond = BI->getCondition();
1935 if (isa<FCmpInst>(BrCond))
1938 BasicBlock *BB = BI->getParent();
1939 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1941 // If ThenBB is actually on the false edge of the conditional branch, remember
1942 // to swap the select operands later.
1943 bool Invert = false;
1944 if (ThenBB != BI->getSuccessor(0)) {
1945 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1948 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1950 // Keep a count of how many times instructions are used within CondBB when
1951 // they are candidates for sinking into CondBB. Specifically:
1952 // - They are defined in BB, and
1953 // - They have no side effects, and
1954 // - All of their uses are in CondBB.
1955 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1957 unsigned SpeculationCost = 0;
1958 Value *SpeculatedStoreValue = nullptr;
1959 StoreInst *SpeculatedStore = nullptr;
1960 for (BasicBlock::iterator BBI = ThenBB->begin(),
1961 BBE = std::prev(ThenBB->end());
1962 BBI != BBE; ++BBI) {
1963 Instruction *I = &*BBI;
1965 if (isa<DbgInfoIntrinsic>(I))
1968 // Only speculatively execute a single instruction (not counting the
1969 // terminator) for now.
1971 if (SpeculationCost > 1)
1974 // Don't hoist the instruction if it's unsafe or expensive.
1975 if (!isSafeToSpeculativelyExecute(I) &&
1976 !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1977 I, BB, ThenBB, EndBB))))
1979 if (!SpeculatedStoreValue &&
1980 ComputeSpeculationCost(I, TTI) >
1981 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
1984 // Store the store speculation candidate.
1985 if (SpeculatedStoreValue)
1986 SpeculatedStore = cast<StoreInst>(I);
1988 // Do not hoist the instruction if any of its operands are defined but not
1989 // used in BB. The transformation will prevent the operand from
1990 // being sunk into the use block.
1991 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
1992 Instruction *OpI = dyn_cast<Instruction>(*i);
1993 if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
1994 continue; // Not a candidate for sinking.
1996 ++SinkCandidateUseCounts[OpI];
2000 // Consider any sink candidates which are only used in CondBB as costs for
2001 // speculation. Note, while we iterate over a DenseMap here, we are summing
2002 // and so iteration order isn't significant.
2003 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator
2004 I = SinkCandidateUseCounts.begin(),
2005 E = SinkCandidateUseCounts.end();
2007 if (I->first->getNumUses() == I->second) {
2009 if (SpeculationCost > 1)
2013 // Check that the PHI nodes can be converted to selects.
2014 bool HaveRewritablePHIs = false;
2015 for (BasicBlock::iterator I = EndBB->begin();
2016 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2017 Value *OrigV = PN->getIncomingValueForBlock(BB);
2018 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
2020 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2021 // Skip PHIs which are trivial.
2025 // Don't convert to selects if we could remove undefined behavior instead.
2026 if (passingValueIsAlwaysUndefined(OrigV, PN) ||
2027 passingValueIsAlwaysUndefined(ThenV, PN))
2030 HaveRewritablePHIs = true;
2031 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2032 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2033 if (!OrigCE && !ThenCE)
2034 continue; // Known safe and cheap.
2036 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2037 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2039 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2040 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2042 2 * PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2043 if (OrigCost + ThenCost > MaxCost)
2046 // Account for the cost of an unfolded ConstantExpr which could end up
2047 // getting expanded into Instructions.
2048 // FIXME: This doesn't account for how many operations are combined in the
2049 // constant expression.
2051 if (SpeculationCost > 1)
2055 // If there are no PHIs to process, bail early. This helps ensure idempotence
2057 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2060 // If we get here, we can hoist the instruction and if-convert.
2061 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2063 // Insert a select of the value of the speculated store.
2064 if (SpeculatedStoreValue) {
2065 IRBuilder<NoFolder> Builder(BI);
2066 Value *TrueV = SpeculatedStore->getValueOperand();
2067 Value *FalseV = SpeculatedStoreValue;
2069 std::swap(TrueV, FalseV);
2070 Value *S = Builder.CreateSelect(
2071 BrCond, TrueV, FalseV, TrueV->getName() + "." + FalseV->getName(), BI);
2072 SpeculatedStore->setOperand(0, S);
2075 // Metadata can be dependent on the condition we are hoisting above.
2076 // Conservatively strip all metadata on the instruction.
2077 for (auto &I : *ThenBB)
2078 I.dropUnknownNonDebugMetadata();
2080 // Hoist the instructions.
2081 BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2082 ThenBB->begin(), std::prev(ThenBB->end()));
2084 // Insert selects and rewrite the PHI operands.
2085 IRBuilder<NoFolder> Builder(BI);
2086 for (BasicBlock::iterator I = EndBB->begin();
2087 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2088 unsigned OrigI = PN->getBasicBlockIndex(BB);
2089 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
2090 Value *OrigV = PN->getIncomingValue(OrigI);
2091 Value *ThenV = PN->getIncomingValue(ThenI);
2093 // Skip PHIs which are trivial.
2097 // Create a select whose true value is the speculatively executed value and
2098 // false value is the preexisting value. Swap them if the branch
2099 // destinations were inverted.
2100 Value *TrueV = ThenV, *FalseV = OrigV;
2102 std::swap(TrueV, FalseV);
2103 Value *V = Builder.CreateSelect(
2104 BrCond, TrueV, FalseV, TrueV->getName() + "." + FalseV->getName(), BI);
2105 PN->setIncomingValue(OrigI, V);
2106 PN->setIncomingValue(ThenI, V);
2113 /// Return true if we can thread a branch across this block.
2114 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
2115 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
2118 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2119 if (isa<DbgInfoIntrinsic>(BBI))
2122 return false; // Don't clone large BB's.
2125 // We can only support instructions that do not define values that are
2126 // live outside of the current basic block.
2127 for (User *U : BBI->users()) {
2128 Instruction *UI = cast<Instruction>(U);
2129 if (UI->getParent() != BB || isa<PHINode>(UI))
2133 // Looks ok, continue checking.
2139 /// If we have a conditional branch on a PHI node value that is defined in the
2140 /// same block as the branch and if any PHI entries are constants, thread edges
2141 /// corresponding to that entry to be branches to their ultimate destination.
2142 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL) {
2143 BasicBlock *BB = BI->getParent();
2144 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2145 // NOTE: we currently cannot transform this case if the PHI node is used
2146 // outside of the block.
2147 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2150 // Degenerate case of a single entry PHI.
2151 if (PN->getNumIncomingValues() == 1) {
2152 FoldSingleEntryPHINodes(PN->getParent());
2156 // Now we know that this block has multiple preds and two succs.
2157 if (!BlockIsSimpleEnoughToThreadThrough(BB))
2160 // Can't fold blocks that contain noduplicate or convergent calls.
2161 if (any_of(*BB, [](const Instruction &I) {
2162 const CallInst *CI = dyn_cast<CallInst>(&I);
2163 return CI && (CI->cannotDuplicate() || CI->isConvergent());
2167 // Okay, this is a simple enough basic block. See if any phi values are
2169 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2170 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
2171 if (!CB || !CB->getType()->isIntegerTy(1))
2174 // Okay, we now know that all edges from PredBB should be revectored to
2175 // branch to RealDest.
2176 BasicBlock *PredBB = PN->getIncomingBlock(i);
2177 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2180 continue; // Skip self loops.
2181 // Skip if the predecessor's terminator is an indirect branch.
2182 if (isa<IndirectBrInst>(PredBB->getTerminator()))
2185 // The dest block might have PHI nodes, other predecessors and other
2186 // difficult cases. Instead of being smart about this, just insert a new
2187 // block that jumps to the destination block, effectively splitting
2188 // the edge we are about to create.
2189 BasicBlock *EdgeBB =
2190 BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2191 RealDest->getParent(), RealDest);
2192 BranchInst::Create(RealDest, EdgeBB);
2194 // Update PHI nodes.
2195 AddPredecessorToBlock(RealDest, EdgeBB, BB);
2197 // BB may have instructions that are being threaded over. Clone these
2198 // instructions into EdgeBB. We know that there will be no uses of the
2199 // cloned instructions outside of EdgeBB.
2200 BasicBlock::iterator InsertPt = EdgeBB->begin();
2201 DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2202 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2203 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2204 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2207 // Clone the instruction.
2208 Instruction *N = BBI->clone();
2210 N->setName(BBI->getName() + ".c");
2212 // Update operands due to translation.
2213 for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2214 DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2215 if (PI != TranslateMap.end())
2219 // Check for trivial simplification.
2220 if (Value *V = SimplifyInstruction(N, DL)) {
2221 if (!BBI->use_empty())
2222 TranslateMap[&*BBI] = V;
2223 if (!N->mayHaveSideEffects()) {
2224 delete N; // Instruction folded away, don't need actual inst
2228 if (!BBI->use_empty())
2229 TranslateMap[&*BBI] = N;
2231 // Insert the new instruction into its new home.
2233 EdgeBB->getInstList().insert(InsertPt, N);
2236 // Loop over all of the edges from PredBB to BB, changing them to branch
2237 // to EdgeBB instead.
2238 TerminatorInst *PredBBTI = PredBB->getTerminator();
2239 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2240 if (PredBBTI->getSuccessor(i) == BB) {
2241 BB->removePredecessor(PredBB);
2242 PredBBTI->setSuccessor(i, EdgeBB);
2245 // Recurse, simplifying any other constants.
2246 return FoldCondBranchOnPHI(BI, DL) | true;
2252 /// Given a BB that starts with the specified two-entry PHI node,
2253 /// see if we can eliminate it.
2254 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI,
2255 const DataLayout &DL) {
2256 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2257 // statement", which has a very simple dominance structure. Basically, we
2258 // are trying to find the condition that is being branched on, which
2259 // subsequently causes this merge to happen. We really want control
2260 // dependence information for this check, but simplifycfg can't keep it up
2261 // to date, and this catches most of the cases we care about anyway.
2262 BasicBlock *BB = PN->getParent();
2263 BasicBlock *IfTrue, *IfFalse;
2264 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2266 // Don't bother if the branch will be constant folded trivially.
2267 isa<ConstantInt>(IfCond))
2270 // Okay, we found that we can merge this two-entry phi node into a select.
2271 // Doing so would require us to fold *all* two entry phi nodes in this block.
2272 // At some point this becomes non-profitable (particularly if the target
2273 // doesn't support cmov's). Only do this transformation if there are two or
2274 // fewer PHI nodes in this block.
2275 unsigned NumPhis = 0;
2276 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2280 // Loop over the PHI's seeing if we can promote them all to select
2281 // instructions. While we are at it, keep track of the instructions
2282 // that need to be moved to the dominating block.
2283 SmallPtrSet<Instruction *, 4> AggressiveInsts;
2284 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2285 MaxCostVal1 = PHINodeFoldingThreshold;
2286 MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2287 MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2289 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2290 PHINode *PN = cast<PHINode>(II++);
2291 if (Value *V = SimplifyInstruction(PN, DL)) {
2292 PN->replaceAllUsesWith(V);
2293 PN->eraseFromParent();
2297 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
2298 MaxCostVal0, TTI) ||
2299 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
2304 // If we folded the first phi, PN dangles at this point. Refresh it. If
2305 // we ran out of PHIs then we simplified them all.
2306 PN = dyn_cast<PHINode>(BB->begin());
2310 // Don't fold i1 branches on PHIs which contain binary operators. These can
2311 // often be turned into switches and other things.
2312 if (PN->getType()->isIntegerTy(1) &&
2313 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2314 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2315 isa<BinaryOperator>(IfCond)))
2318 // If all PHI nodes are promotable, check to make sure that all instructions
2319 // in the predecessor blocks can be promoted as well. If not, we won't be able
2320 // to get rid of the control flow, so it's not worth promoting to select
2322 BasicBlock *DomBlock = nullptr;
2323 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2324 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2325 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2328 DomBlock = *pred_begin(IfBlock1);
2329 for (BasicBlock::iterator I = IfBlock1->begin(); !isa<TerminatorInst>(I);
2331 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2332 // This is not an aggressive instruction that we can promote.
2333 // Because of this, we won't be able to get rid of the control flow, so
2334 // the xform is not worth it.
2339 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2342 DomBlock = *pred_begin(IfBlock2);
2343 for (BasicBlock::iterator I = IfBlock2->begin(); !isa<TerminatorInst>(I);
2345 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2346 // This is not an aggressive instruction that we can promote.
2347 // Because of this, we won't be able to get rid of the control flow, so
2348 // the xform is not worth it.
2353 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
2354 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
2356 // If we can still promote the PHI nodes after this gauntlet of tests,
2357 // do all of the PHI's now.
2358 Instruction *InsertPt = DomBlock->getTerminator();
2359 IRBuilder<NoFolder> Builder(InsertPt);
2361 // Move all 'aggressive' instructions, which are defined in the
2362 // conditional parts of the if's up to the dominating block.
2364 for (auto &I : *IfBlock1)
2365 I.dropUnknownNonDebugMetadata();
2366 DomBlock->getInstList().splice(InsertPt->getIterator(),
2367 IfBlock1->getInstList(), IfBlock1->begin(),
2368 IfBlock1->getTerminator()->getIterator());
2371 for (auto &I : *IfBlock2)
2372 I.dropUnknownNonDebugMetadata();
2373 DomBlock->getInstList().splice(InsertPt->getIterator(),
2374 IfBlock2->getInstList(), IfBlock2->begin(),
2375 IfBlock2->getTerminator()->getIterator());
2378 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2379 // Change the PHI node into a select instruction.
2380 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2381 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2383 Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2384 PN->replaceAllUsesWith(Sel);
2386 PN->eraseFromParent();
2389 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2390 // has been flattened. Change DomBlock to jump directly to our new block to
2391 // avoid other simplifycfg's kicking in on the diamond.
2392 TerminatorInst *OldTI = DomBlock->getTerminator();
2393 Builder.SetInsertPoint(OldTI);
2394 Builder.CreateBr(BB);
2395 OldTI->eraseFromParent();
2399 /// If we found a conditional branch that goes to two returning blocks,
2400 /// try to merge them together into one return,
2401 /// introducing a select if the return values disagree.
2402 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
2403 IRBuilder<> &Builder) {
2404 assert(BI->isConditional() && "Must be a conditional branch");
2405 BasicBlock *TrueSucc = BI->getSuccessor(0);
2406 BasicBlock *FalseSucc = BI->getSuccessor(1);
2407 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2408 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2410 // Check to ensure both blocks are empty (just a return) or optionally empty
2411 // with PHI nodes. If there are other instructions, merging would cause extra
2412 // computation on one path or the other.
2413 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2415 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2418 Builder.SetInsertPoint(BI);
2419 // Okay, we found a branch that is going to two return nodes. If
2420 // there is no return value for this function, just change the
2421 // branch into a return.
2422 if (FalseRet->getNumOperands() == 0) {
2423 TrueSucc->removePredecessor(BI->getParent());
2424 FalseSucc->removePredecessor(BI->getParent());
2425 Builder.CreateRetVoid();
2426 EraseTerminatorInstAndDCECond(BI);
2430 // Otherwise, figure out what the true and false return values are
2431 // so we can insert a new select instruction.
2432 Value *TrueValue = TrueRet->getReturnValue();
2433 Value *FalseValue = FalseRet->getReturnValue();
2435 // Unwrap any PHI nodes in the return blocks.
2436 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2437 if (TVPN->getParent() == TrueSucc)
2438 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2439 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2440 if (FVPN->getParent() == FalseSucc)
2441 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2443 // In order for this transformation to be safe, we must be able to
2444 // unconditionally execute both operands to the return. This is
2445 // normally the case, but we could have a potentially-trapping
2446 // constant expression that prevents this transformation from being
2448 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2451 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2455 // Okay, we collected all the mapped values and checked them for sanity, and
2456 // defined to really do this transformation. First, update the CFG.
2457 TrueSucc->removePredecessor(BI->getParent());
2458 FalseSucc->removePredecessor(BI->getParent());
2460 // Insert select instructions where needed.
2461 Value *BrCond = BI->getCondition();
2463 // Insert a select if the results differ.
2464 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2465 } else if (isa<UndefValue>(TrueValue)) {
2466 TrueValue = FalseValue;
2469 Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2474 !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2478 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2479 << "\n " << *BI << "NewRet = " << *RI
2480 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: " << *FalseSucc);
2482 EraseTerminatorInstAndDCECond(BI);
2487 /// Return true if the given instruction is available
2488 /// in its predecessor block. If yes, the instruction will be removed.
2489 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
2490 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2492 for (Instruction &I : *PB) {
2493 Instruction *PBI = &I;
2494 // Check whether Inst and PBI generate the same value.
2495 if (Inst->isIdenticalTo(PBI)) {
2496 Inst->replaceAllUsesWith(PBI);
2497 Inst->eraseFromParent();
2504 /// Return true if either PBI or BI has branch weight available, and store
2505 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2506 /// not have branch weight, use 1:1 as its weight.
2507 static bool extractPredSuccWeights(BranchInst *PBI, BranchInst *BI,
2508 uint64_t &PredTrueWeight,
2509 uint64_t &PredFalseWeight,
2510 uint64_t &SuccTrueWeight,
2511 uint64_t &SuccFalseWeight) {
2512 bool PredHasWeights =
2513 PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2514 bool SuccHasWeights =
2515 BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2516 if (PredHasWeights || SuccHasWeights) {
2517 if (!PredHasWeights)
2518 PredTrueWeight = PredFalseWeight = 1;
2519 if (!SuccHasWeights)
2520 SuccTrueWeight = SuccFalseWeight = 1;
2527 /// If this basic block is simple enough, and if a predecessor branches to us
2528 /// and one of our successors, fold the block into the predecessor and use
2529 /// logical operations to pick the right destination.
2530 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2531 BasicBlock *BB = BI->getParent();
2533 Instruction *Cond = nullptr;
2534 if (BI->isConditional())
2535 Cond = dyn_cast<Instruction>(BI->getCondition());
2537 // For unconditional branch, check for a simple CFG pattern, where
2538 // BB has a single predecessor and BB's successor is also its predecessor's
2539 // successor. If such pattern exisits, check for CSE between BB and its
2541 if (BasicBlock *PB = BB->getSinglePredecessor())
2542 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2543 if (PBI->isConditional() &&
2544 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2545 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2546 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2547 Instruction *Curr = &*I++;
2548 if (isa<CmpInst>(Curr)) {
2552 // Quit if we can't remove this instruction.
2553 if (!checkCSEInPredecessor(Curr, PB))
2562 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2563 Cond->getParent() != BB || !Cond->hasOneUse())
2566 // Make sure the instruction after the condition is the cond branch.
2567 BasicBlock::iterator CondIt = ++Cond->getIterator();
2569 // Ignore dbg intrinsics.
2570 while (isa<DbgInfoIntrinsic>(CondIt))
2576 // Only allow this transformation if computing the condition doesn't involve
2577 // too many instructions and these involved instructions can be executed
2578 // unconditionally. We denote all involved instructions except the condition
2579 // as "bonus instructions", and only allow this transformation when the
2580 // number of the bonus instructions does not exceed a certain threshold.
2581 unsigned NumBonusInsts = 0;
2582 for (auto I = BB->begin(); Cond != &*I; ++I) {
2583 // Ignore dbg intrinsics.
2584 if (isa<DbgInfoIntrinsic>(I))
2586 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2588 // I has only one use and can be executed unconditionally.
2589 Instruction *User = dyn_cast<Instruction>(I->user_back());
2590 if (User == nullptr || User->getParent() != BB)
2592 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2593 // to use any other instruction, User must be an instruction between next(I)
2596 // Early exits once we reach the limit.
2597 if (NumBonusInsts > BonusInstThreshold)
2601 // Cond is known to be a compare or binary operator. Check to make sure that
2602 // neither operand is a potentially-trapping constant expression.
2603 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2606 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2610 // Finally, don't infinitely unroll conditional loops.
2611 BasicBlock *TrueDest = BI->getSuccessor(0);
2612 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2613 if (TrueDest == BB || FalseDest == BB)
2616 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2617 BasicBlock *PredBlock = *PI;
2618 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2620 // Check that we have two conditional branches. If there is a PHI node in
2621 // the common successor, verify that the same value flows in from both
2623 SmallVector<PHINode *, 4> PHIs;
2624 if (!PBI || PBI->isUnconditional() ||
2625 (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2626 (!BI->isConditional() &&
2627 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2630 // Determine if the two branches share a common destination.
2631 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2632 bool InvertPredCond = false;
2634 if (BI->isConditional()) {
2635 if (PBI->getSuccessor(0) == TrueDest) {
2636 Opc = Instruction::Or;
2637 } else if (PBI->getSuccessor(1) == FalseDest) {
2638 Opc = Instruction::And;
2639 } else if (PBI->getSuccessor(0) == FalseDest) {
2640 Opc = Instruction::And;
2641 InvertPredCond = true;
2642 } else if (PBI->getSuccessor(1) == TrueDest) {
2643 Opc = Instruction::Or;
2644 InvertPredCond = true;
2649 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2653 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2654 IRBuilder<> Builder(PBI);
2656 // If we need to invert the condition in the pred block to match, do so now.
2657 if (InvertPredCond) {
2658 Value *NewCond = PBI->getCondition();
2660 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2661 CmpInst *CI = cast<CmpInst>(NewCond);
2662 CI->setPredicate(CI->getInversePredicate());
2665 Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2668 PBI->setCondition(NewCond);
2669 PBI->swapSuccessors();
2672 // If we have bonus instructions, clone them into the predecessor block.
2673 // Note that there may be multiple predecessor blocks, so we cannot move
2674 // bonus instructions to a predecessor block.
2675 ValueToValueMapTy VMap; // maps original values to cloned values
2676 // We already make sure Cond is the last instruction before BI. Therefore,
2677 // all instructions before Cond other than DbgInfoIntrinsic are bonus
2679 for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2680 if (isa<DbgInfoIntrinsic>(BonusInst))
2682 Instruction *NewBonusInst = BonusInst->clone();
2683 RemapInstruction(NewBonusInst, VMap,
2684 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2685 VMap[&*BonusInst] = NewBonusInst;
2687 // If we moved a load, we cannot any longer claim any knowledge about
2688 // its potential value. The previous information might have been valid
2689 // only given the branch precondition.
2690 // For an analogous reason, we must also drop all the metadata whose
2691 // semantics we don't understand.
2692 NewBonusInst->dropUnknownNonDebugMetadata();
2694 PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2695 NewBonusInst->takeName(&*BonusInst);
2696 BonusInst->setName(BonusInst->getName() + ".old");
2699 // Clone Cond into the predecessor basic block, and or/and the
2700 // two conditions together.
2701 Instruction *New = Cond->clone();
2702 RemapInstruction(New, VMap,
2703 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2704 PredBlock->getInstList().insert(PBI->getIterator(), New);
2705 New->takeName(Cond);
2706 Cond->setName(New->getName() + ".old");
2708 if (BI->isConditional()) {
2709 Instruction *NewCond = cast<Instruction>(
2710 Builder.CreateBinOp(Opc, PBI->getCondition(), New, "or.cond"));
2711 PBI->setCondition(NewCond);
2713 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2715 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2716 SuccTrueWeight, SuccFalseWeight);
2717 SmallVector<uint64_t, 8> NewWeights;
2719 if (PBI->getSuccessor(0) == BB) {
2721 // PBI: br i1 %x, BB, FalseDest
2722 // BI: br i1 %y, TrueDest, FalseDest
2723 // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2724 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2725 // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2726 // TrueWeight for PBI * FalseWeight for BI.
2727 // We assume that total weights of a BranchInst can fit into 32 bits.
2728 // Therefore, we will not have overflow using 64-bit arithmetic.
2729 NewWeights.push_back(PredFalseWeight *
2730 (SuccFalseWeight + SuccTrueWeight) +
2731 PredTrueWeight * SuccFalseWeight);
2733 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2734 PBI->setSuccessor(0, TrueDest);
2736 if (PBI->getSuccessor(1) == BB) {
2738 // PBI: br i1 %x, TrueDest, BB
2739 // BI: br i1 %y, TrueDest, FalseDest
2740 // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2741 // FalseWeight for PBI * TrueWeight for BI.
2742 NewWeights.push_back(PredTrueWeight *
2743 (SuccFalseWeight + SuccTrueWeight) +
2744 PredFalseWeight * SuccTrueWeight);
2745 // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2746 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2748 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2749 PBI->setSuccessor(1, FalseDest);
2751 if (NewWeights.size() == 2) {
2752 // Halve the weights if any of them cannot fit in an uint32_t
2753 FitWeights(NewWeights);
2755 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2758 LLVMContext::MD_prof,
2759 MDBuilder(BI->getContext()).createBranchWeights(MDWeights));
2761 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2763 // Update PHI nodes in the common successors.
2764 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2765 ConstantInt *PBI_C = cast<ConstantInt>(
2766 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2767 assert(PBI_C->getType()->isIntegerTy(1));
2768 Instruction *MergedCond = nullptr;
2769 if (PBI->getSuccessor(0) == TrueDest) {
2770 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2771 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2772 // is false: !PBI_Cond and BI_Value
2773 Instruction *NotCond = cast<Instruction>(
2774 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2775 MergedCond = cast<Instruction>(
2776 Builder.CreateBinOp(Instruction::And, NotCond, New, "and.cond"));
2778 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2779 Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2781 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2782 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2783 // is false: PBI_Cond and BI_Value
2784 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2785 Instruction::And, PBI->getCondition(), New, "and.cond"));
2786 if (PBI_C->isOne()) {
2787 Instruction *NotCond = cast<Instruction>(
2788 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2789 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2790 Instruction::Or, NotCond, MergedCond, "or.cond"));
2794 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2797 // Change PBI from Conditional to Unconditional.
2798 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2799 EraseTerminatorInstAndDCECond(PBI);
2803 // If BI was a loop latch, it may have had associated loop metadata.
2804 // We need to copy it to the new latch, that is, PBI.
2805 if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2806 PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2808 // TODO: If BB is reachable from all paths through PredBlock, then we
2809 // could replace PBI's branch probabilities with BI's.
2811 // Copy any debug value intrinsics into the end of PredBlock.
2812 for (Instruction &I : *BB)
2813 if (isa<DbgInfoIntrinsic>(I))
2814 I.clone()->insertBefore(PBI);
2821 // If there is only one store in BB1 and BB2, return it, otherwise return
2823 static StoreInst *findUniqueStoreInBlocks(BasicBlock *BB1, BasicBlock *BB2) {
2824 StoreInst *S = nullptr;
2825 for (auto *BB : {BB1, BB2}) {
2829 if (auto *SI = dyn_cast<StoreInst>(&I)) {
2831 // Multiple stores seen.
2840 static Value *ensureValueAvailableInSuccessor(Value *V, BasicBlock *BB,
2841 Value *AlternativeV = nullptr) {
2842 // PHI is going to be a PHI node that allows the value V that is defined in
2843 // BB to be referenced in BB's only successor.
2845 // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2846 // doesn't matter to us what the other operand is (it'll never get used). We
2847 // could just create a new PHI with an undef incoming value, but that could
2848 // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2849 // other PHI. So here we directly look for some PHI in BB's successor with V
2850 // as an incoming operand. If we find one, we use it, else we create a new
2853 // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2854 // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2855 // where OtherBB is the single other predecessor of BB's only successor.
2856 PHINode *PHI = nullptr;
2857 BasicBlock *Succ = BB->getSingleSuccessor();
2859 for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2860 if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2861 PHI = cast<PHINode>(I);
2865 assert(std::distance(pred_begin(Succ), pred_end(Succ)) == 2);
2866 auto PredI = pred_begin(Succ);
2867 BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2868 if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2875 // If V is not an instruction defined in BB, just return it.
2876 if (!AlternativeV &&
2877 (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2880 PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2881 PHI->addIncoming(V, BB);
2882 for (BasicBlock *PredBB : predecessors(Succ))
2885 AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2889 static bool mergeConditionalStoreToAddress(BasicBlock *PTB, BasicBlock *PFB,
2890 BasicBlock *QTB, BasicBlock *QFB,
2891 BasicBlock *PostBB, Value *Address,
2892 bool InvertPCond, bool InvertQCond) {
2893 auto IsaBitcastOfPointerType = [](const Instruction &I) {
2894 return Operator::getOpcode(&I) == Instruction::BitCast &&
2895 I.getType()->isPointerTy();
2898 // If we're not in aggressive mode, we only optimize if we have some
2899 // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2900 auto IsWorthwhile = [&](BasicBlock *BB) {
2903 // Heuristic: if the block can be if-converted/phi-folded and the
2904 // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2905 // thread this store.
2907 for (auto &I : *BB) {
2908 // Cheap instructions viable for folding.
2909 if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2912 // Free instructions.
2913 else if (isa<TerminatorInst>(I) || isa<DbgInfoIntrinsic>(I) ||
2914 IsaBitcastOfPointerType(I))
2919 return N <= PHINodeFoldingThreshold;
2922 if (!MergeCondStoresAggressively &&
2923 (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2924 !IsWorthwhile(QFB)))
2927 // For every pointer, there must be exactly two stores, one coming from
2928 // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2929 // store (to any address) in PTB,PFB or QTB,QFB.
2930 // FIXME: We could relax this restriction with a bit more work and performance
2932 StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2933 StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2934 if (!PStore || !QStore)
2937 // Now check the stores are compatible.
2938 if (!QStore->isUnordered() || !PStore->isUnordered())
2941 // Check that sinking the store won't cause program behavior changes. Sinking
2942 // the store out of the Q blocks won't change any behavior as we're sinking
2943 // from a block to its unconditional successor. But we're moving a store from
2944 // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2945 // So we need to check that there are no aliasing loads or stores in
2946 // QBI, QTB and QFB. We also need to check there are no conflicting memory
2947 // operations between PStore and the end of its parent block.
2949 // The ideal way to do this is to query AliasAnalysis, but we don't
2950 // preserve AA currently so that is dangerous. Be super safe and just
2951 // check there are no other memory operations at all.
2952 for (auto &I : *QFB->getSinglePredecessor())
2953 if (I.mayReadOrWriteMemory())
2955 for (auto &I : *QFB)
2956 if (&I != QStore && I.mayReadOrWriteMemory())
2959 for (auto &I : *QTB)
2960 if (&I != QStore && I.mayReadOrWriteMemory())
2962 for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2964 if (&*I != PStore && I->mayReadOrWriteMemory())
2967 // OK, we're going to sink the stores to PostBB. The store has to be
2968 // conditional though, so first create the predicate.
2969 Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
2971 Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
2974 Value *PPHI = ensureValueAvailableInSuccessor(PStore->getValueOperand(),
2975 PStore->getParent());
2976 Value *QPHI = ensureValueAvailableInSuccessor(QStore->getValueOperand(),
2977 QStore->getParent(), PPHI);
2979 IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
2981 Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
2982 Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
2985 PPred = QB.CreateNot(PPred);
2987 QPred = QB.CreateNot(QPred);
2988 Value *CombinedPred = QB.CreateOr(PPred, QPred);
2991 SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
2992 QB.SetInsertPoint(T);
2993 StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
2995 PStore->getAAMetadata(AAMD, /*Merge=*/false);
2996 PStore->getAAMetadata(AAMD, /*Merge=*/true);
2997 SI->setAAMetadata(AAMD);
2999 QStore->eraseFromParent();
3000 PStore->eraseFromParent();
3005 static bool mergeConditionalStores(BranchInst *PBI, BranchInst *QBI) {
3006 // The intention here is to find diamonds or triangles (see below) where each
3007 // conditional block contains a store to the same address. Both of these
3008 // stores are conditional, so they can't be unconditionally sunk. But it may
3009 // be profitable to speculatively sink the stores into one merged store at the
3010 // end, and predicate the merged store on the union of the two conditions of
3013 // This can reduce the number of stores executed if both of the conditions are
3014 // true, and can allow the blocks to become small enough to be if-converted.
3015 // This optimization will also chain, so that ladders of test-and-set
3016 // sequences can be if-converted away.
3018 // We only deal with simple diamonds or triangles:
3020 // PBI or PBI or a combination of the two
3030 // We model triangles as a type of diamond with a nullptr "true" block.
3031 // Triangles are canonicalized so that the fallthrough edge is represented by
3032 // a true condition, as in the diagram above.
3034 BasicBlock *PTB = PBI->getSuccessor(0);
3035 BasicBlock *PFB = PBI->getSuccessor(1);
3036 BasicBlock *QTB = QBI->getSuccessor(0);
3037 BasicBlock *QFB = QBI->getSuccessor(1);
3038 BasicBlock *PostBB = QFB->getSingleSuccessor();
3040 bool InvertPCond = false, InvertQCond = false;
3041 // Canonicalize fallthroughs to the true branches.
3042 if (PFB == QBI->getParent()) {
3043 std::swap(PFB, PTB);
3046 if (QFB == PostBB) {
3047 std::swap(QFB, QTB);
3051 // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3052 // and QFB may not. Model fallthroughs as a nullptr block.
3053 if (PTB == QBI->getParent())
3058 // Legality bailouts. We must have at least the non-fallthrough blocks and
3059 // the post-dominating block, and the non-fallthroughs must only have one
3061 auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3062 return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3065 !HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3066 !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3068 if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3069 (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3071 if (PostBB->getNumUses() != 2 || QBI->getParent()->getNumUses() != 2)
3074 // OK, this is a sequence of two diamonds or triangles.
3075 // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3076 SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3077 for (auto *BB : {PTB, PFB}) {
3081 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3082 PStoreAddresses.insert(SI->getPointerOperand());
3084 for (auto *BB : {QTB, QFB}) {
3088 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3089 QStoreAddresses.insert(SI->getPointerOperand());
3092 set_intersect(PStoreAddresses, QStoreAddresses);
3093 // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3094 // clear what it contains.
3095 auto &CommonAddresses = PStoreAddresses;
3097 bool Changed = false;
3098 for (auto *Address : CommonAddresses)
3099 Changed |= mergeConditionalStoreToAddress(
3100 PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond);
3104 /// If we have a conditional branch as a predecessor of another block,
3105 /// this function tries to simplify it. We know
3106 /// that PBI and BI are both conditional branches, and BI is in one of the
3107 /// successor blocks of PBI - PBI branches to BI.
3108 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
3109 const DataLayout &DL) {
3110 assert(PBI->isConditional() && BI->isConditional());
3111 BasicBlock *BB = BI->getParent();
3113 // If this block ends with a branch instruction, and if there is a
3114 // predecessor that ends on a branch of the same condition, make
3115 // this conditional branch redundant.
3116 if (PBI->getCondition() == BI->getCondition() &&
3117 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3118 // Okay, the outcome of this conditional branch is statically
3119 // knowable. If this block had a single pred, handle specially.
3120 if (BB->getSinglePredecessor()) {
3121 // Turn this into a branch on constant.
3122 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3124 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3125 return true; // Nuke the branch on constant.
3128 // Otherwise, if there are multiple predecessors, insert a PHI that merges
3129 // in the constant and simplify the block result. Subsequent passes of
3130 // simplifycfg will thread the block.
3131 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
3132 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3133 PHINode *NewPN = PHINode::Create(
3134 Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3135 BI->getCondition()->getName() + ".pr", &BB->front());
3136 // Okay, we're going to insert the PHI node. Since PBI is not the only
3137 // predecessor, compute the PHI'd conditional value for all of the preds.
3138 // Any predecessor where the condition is not computable we keep symbolic.
3139 for (pred_iterator PI = PB; PI != PE; ++PI) {
3140 BasicBlock *P = *PI;
3141 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3142 PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3143 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3144 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3146 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3149 NewPN->addIncoming(BI->getCondition(), P);
3153 BI->setCondition(NewPN);
3158 if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3162 // If both branches are conditional and both contain stores to the same
3163 // address, remove the stores from the conditionals and create a conditional
3164 // merged store at the end.
3165 if (MergeCondStores && mergeConditionalStores(PBI, BI))
3168 // If this is a conditional branch in an empty block, and if any
3169 // predecessors are a conditional branch to one of our destinations,
3170 // fold the conditions into logical ops and one cond br.
3171 BasicBlock::iterator BBI = BB->begin();
3172 // Ignore dbg intrinsics.
3173 while (isa<DbgInfoIntrinsic>(BBI))
3179 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3182 } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3185 } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3188 } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3195 // Check to make sure that the other destination of this branch
3196 // isn't BB itself. If so, this is an infinite loop that will
3197 // keep getting unwound.
3198 if (PBI->getSuccessor(PBIOp) == BB)
3201 // Do not perform this transformation if it would require
3202 // insertion of a large number of select instructions. For targets
3203 // without predication/cmovs, this is a big pessimization.
3205 // Also do not perform this transformation if any phi node in the common
3206 // destination block can trap when reached by BB or PBB (PR17073). In that
3207 // case, it would be unsafe to hoist the operation into a select instruction.
3209 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3210 unsigned NumPhis = 0;
3211 for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3213 if (NumPhis > 2) // Disable this xform.
3216 PHINode *PN = cast<PHINode>(II);
3217 Value *BIV = PN->getIncomingValueForBlock(BB);
3218 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3222 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3223 Value *PBIV = PN->getIncomingValue(PBBIdx);
3224 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3229 // Finally, if everything is ok, fold the branches to logical ops.
3230 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3232 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3233 << "AND: " << *BI->getParent());
3235 // If OtherDest *is* BB, then BB is a basic block with a single conditional
3236 // branch in it, where one edge (OtherDest) goes back to itself but the other
3237 // exits. We don't *know* that the program avoids the infinite loop
3238 // (even though that seems likely). If we do this xform naively, we'll end up
3239 // recursively unpeeling the loop. Since we know that (after the xform is
3240 // done) that the block *is* infinite if reached, we just make it an obviously
3241 // infinite loop with no cond branch.
3242 if (OtherDest == BB) {
3243 // Insert it at the end of the function, because it's either code,
3244 // or it won't matter if it's hot. :)
3245 BasicBlock *InfLoopBlock =
3246 BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3247 BranchInst::Create(InfLoopBlock, InfLoopBlock);
3248 OtherDest = InfLoopBlock;
3251 DEBUG(dbgs() << *PBI->getParent()->getParent());
3253 // BI may have other predecessors. Because of this, we leave
3254 // it alone, but modify PBI.
3256 // Make sure we get to CommonDest on True&True directions.
3257 Value *PBICond = PBI->getCondition();
3258 IRBuilder<NoFolder> Builder(PBI);
3260 PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3262 Value *BICond = BI->getCondition();
3264 BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3266 // Merge the conditions.
3267 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3269 // Modify PBI to branch on the new condition to the new dests.
3270 PBI->setCondition(Cond);
3271 PBI->setSuccessor(0, CommonDest);
3272 PBI->setSuccessor(1, OtherDest);
3274 // Update branch weight for PBI.
3275 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3276 uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3278 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3279 SuccTrueWeight, SuccFalseWeight);
3281 PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3282 PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3283 SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3284 SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3285 // The weight to CommonDest should be PredCommon * SuccTotal +
3286 // PredOther * SuccCommon.
3287 // The weight to OtherDest should be PredOther * SuccOther.
3288 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3289 PredOther * SuccCommon,
3290 PredOther * SuccOther};
3291 // Halve the weights if any of them cannot fit in an uint32_t
3292 FitWeights(NewWeights);
3294 PBI->setMetadata(LLVMContext::MD_prof,
3295 MDBuilder(BI->getContext())
3296 .createBranchWeights(NewWeights[0], NewWeights[1]));
3299 // OtherDest may have phi nodes. If so, add an entry from PBI's
3300 // block that are identical to the entries for BI's block.
3301 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3303 // We know that the CommonDest already had an edge from PBI to
3304 // it. If it has PHIs though, the PHIs may have different
3305 // entries for BB and PBI's BB. If so, insert a select to make
3308 for (BasicBlock::iterator II = CommonDest->begin();
3309 (PN = dyn_cast<PHINode>(II)); ++II) {
3310 Value *BIV = PN->getIncomingValueForBlock(BB);
3311 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3312 Value *PBIV = PN->getIncomingValue(PBBIdx);
3314 // Insert a select in PBI to pick the right value.
3315 SelectInst *NV = cast<SelectInst>(
3316 Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3317 PN->setIncomingValue(PBBIdx, NV);
3318 // Although the select has the same condition as PBI, the original branch
3319 // weights for PBI do not apply to the new select because the select's
3320 // 'logical' edges are incoming edges of the phi that is eliminated, not
3321 // the outgoing edges of PBI.
3323 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3324 uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3325 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3326 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3327 // The weight to PredCommonDest should be PredCommon * SuccTotal.
3328 // The weight to PredOtherDest should be PredOther * SuccCommon.
3329 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3330 PredOther * SuccCommon};
3332 FitWeights(NewWeights);
3334 NV->setMetadata(LLVMContext::MD_prof,
3335 MDBuilder(BI->getContext())
3336 .createBranchWeights(NewWeights[0], NewWeights[1]));
3341 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3342 DEBUG(dbgs() << *PBI->getParent()->getParent());
3344 // This basic block is probably dead. We know it has at least
3345 // one fewer predecessor.
3349 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3350 // true or to FalseBB if Cond is false.
3351 // Takes care of updating the successors and removing the old terminator.
3352 // Also makes sure not to introduce new successors by assuming that edges to
3353 // non-successor TrueBBs and FalseBBs aren't reachable.
3354 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
3355 BasicBlock *TrueBB, BasicBlock *FalseBB,
3356 uint32_t TrueWeight,
3357 uint32_t FalseWeight) {
3358 // Remove any superfluous successor edges from the CFG.
3359 // First, figure out which successors to preserve.
3360 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3362 BasicBlock *KeepEdge1 = TrueBB;
3363 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3365 // Then remove the rest.
3366 for (BasicBlock *Succ : OldTerm->successors()) {
3367 // Make sure only to keep exactly one copy of each edge.
3368 if (Succ == KeepEdge1)
3369 KeepEdge1 = nullptr;
3370 else if (Succ == KeepEdge2)
3371 KeepEdge2 = nullptr;
3373 Succ->removePredecessor(OldTerm->getParent(),
3374 /*DontDeleteUselessPHIs=*/true);
3377 IRBuilder<> Builder(OldTerm);
3378 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3380 // Insert an appropriate new terminator.
3381 if (!KeepEdge1 && !KeepEdge2) {
3382 if (TrueBB == FalseBB)
3383 // We were only looking for one successor, and it was present.
3384 // Create an unconditional branch to it.
3385 Builder.CreateBr(TrueBB);
3387 // We found both of the successors we were looking for.
3388 // Create a conditional branch sharing the condition of the select.
3389 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3390 if (TrueWeight != FalseWeight)
3391 NewBI->setMetadata(LLVMContext::MD_prof,
3392 MDBuilder(OldTerm->getContext())
3393 .createBranchWeights(TrueWeight, FalseWeight));
3395 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3396 // Neither of the selected blocks were successors, so this
3397 // terminator must be unreachable.
3398 new UnreachableInst(OldTerm->getContext(), OldTerm);
3400 // One of the selected values was a successor, but the other wasn't.
3401 // Insert an unconditional branch to the one that was found;
3402 // the edge to the one that wasn't must be unreachable.
3404 // Only TrueBB was found.
3405 Builder.CreateBr(TrueBB);
3407 // Only FalseBB was found.
3408 Builder.CreateBr(FalseBB);
3411 EraseTerminatorInstAndDCECond(OldTerm);
3416 // (switch (select cond, X, Y)) on constant X, Y
3417 // with a branch - conditional if X and Y lead to distinct BBs,
3418 // unconditional otherwise.
3419 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
3420 // Check for constant integer values in the select.
3421 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3422 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3423 if (!TrueVal || !FalseVal)
3426 // Find the relevant condition and destinations.
3427 Value *Condition = Select->getCondition();
3428 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
3429 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
3431 // Get weight for TrueBB and FalseBB.
3432 uint32_t TrueWeight = 0, FalseWeight = 0;
3433 SmallVector<uint64_t, 8> Weights;
3434 bool HasWeights = HasBranchWeights(SI);
3436 GetBranchWeights(SI, Weights);
3437 if (Weights.size() == 1 + SI->getNumCases()) {
3439 (uint32_t)Weights[SI->findCaseValue(TrueVal).getSuccessorIndex()];
3441 (uint32_t)Weights[SI->findCaseValue(FalseVal).getSuccessorIndex()];
3445 // Perform the actual simplification.
3446 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3451 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3452 // blockaddress(@fn, BlockB)))
3454 // (br cond, BlockA, BlockB).
3455 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
3456 // Check that both operands of the select are block addresses.
3457 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
3458 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
3462 // Extract the actual blocks.
3463 BasicBlock *TrueBB = TBA->getBasicBlock();
3464 BasicBlock *FalseBB = FBA->getBasicBlock();
3466 // Perform the actual simplification.
3467 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3471 /// This is called when we find an icmp instruction
3472 /// (a seteq/setne with a constant) as the only instruction in a
3473 /// block that ends with an uncond branch. We are looking for a very specific
3474 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3475 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3476 /// default value goes to an uncond block with a seteq in it, we get something
3479 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3481 /// %tmp = icmp eq i8 %A, 92
3484 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3486 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3487 /// the PHI, merging the third icmp into the switch.
3488 static bool TryToSimplifyUncondBranchWithICmpInIt(
3489 ICmpInst *ICI, IRBuilder<> &Builder, const DataLayout &DL,
3490 const TargetTransformInfo &TTI, unsigned BonusInstThreshold,
3491 AssumptionCache *AC) {
3492 BasicBlock *BB = ICI->getParent();
3494 // If the block has any PHIs in it or the icmp has multiple uses, it is too
3496 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3499 Value *V = ICI->getOperand(0);
3500 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3502 // The pattern we're looking for is where our only predecessor is a switch on
3503 // 'V' and this block is the default case for the switch. In this case we can
3504 // fold the compared value into the switch to simplify things.
3505 BasicBlock *Pred = BB->getSinglePredecessor();
3506 if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3509 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3510 if (SI->getCondition() != V)
3513 // If BB is reachable on a non-default case, then we simply know the value of
3514 // V in this block. Substitute it and constant fold the icmp instruction
3516 if (SI->getDefaultDest() != BB) {
3517 ConstantInt *VVal = SI->findCaseDest(BB);
3518 assert(VVal && "Should have a unique destination value");
3519 ICI->setOperand(0, VVal);
3521 if (Value *V = SimplifyInstruction(ICI, DL)) {
3522 ICI->replaceAllUsesWith(V);
3523 ICI->eraseFromParent();
3525 // BB is now empty, so it is likely to simplify away.
3526 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3529 // Ok, the block is reachable from the default dest. If the constant we're
3530 // comparing exists in one of the other edges, then we can constant fold ICI
3532 if (SI->findCaseValue(Cst) != SI->case_default()) {
3534 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3535 V = ConstantInt::getFalse(BB->getContext());
3537 V = ConstantInt::getTrue(BB->getContext());
3539 ICI->replaceAllUsesWith(V);
3540 ICI->eraseFromParent();
3541 // BB is now empty, so it is likely to simplify away.
3542 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3545 // The use of the icmp has to be in the 'end' block, by the only PHI node in
3547 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3548 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3549 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3550 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3553 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3555 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3556 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3558 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3559 std::swap(DefaultCst, NewCst);
3561 // Replace ICI (which is used by the PHI for the default value) with true or
3562 // false depending on if it is EQ or NE.
3563 ICI->replaceAllUsesWith(DefaultCst);
3564 ICI->eraseFromParent();
3566 // Okay, the switch goes to this block on a default value. Add an edge from
3567 // the switch to the merge point on the compared value.
3569 BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3570 SmallVector<uint64_t, 8> Weights;
3571 bool HasWeights = HasBranchWeights(SI);
3573 GetBranchWeights(SI, Weights);
3574 if (Weights.size() == 1 + SI->getNumCases()) {
3575 // Split weight for default case to case for "Cst".
3576 Weights[0] = (Weights[0] + 1) >> 1;
3577 Weights.push_back(Weights[0]);
3579 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3581 LLVMContext::MD_prof,
3582 MDBuilder(SI->getContext()).createBranchWeights(MDWeights));
3585 SI->addCase(Cst, NewBB);
3587 // NewBB branches to the phi block, add the uncond branch and the phi entry.
3588 Builder.SetInsertPoint(NewBB);
3589 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3590 Builder.CreateBr(SuccBlock);
3591 PHIUse->addIncoming(NewCst, NewBB);
3595 /// The specified branch is a conditional branch.
3596 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3597 /// fold it into a switch instruction if so.
3598 static bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
3599 const DataLayout &DL) {
3600 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3604 // Change br (X == 0 | X == 1), T, F into a switch instruction.
3605 // If this is a bunch of seteq's or'd together, or if it's a bunch of
3606 // 'setne's and'ed together, collect them.
3608 // Try to gather values from a chain of and/or to be turned into a switch
3609 ConstantComparesGatherer ConstantCompare(Cond, DL);
3610 // Unpack the result
3611 SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3612 Value *CompVal = ConstantCompare.CompValue;
3613 unsigned UsedICmps = ConstantCompare.UsedICmps;
3614 Value *ExtraCase = ConstantCompare.Extra;
3616 // If we didn't have a multiply compared value, fail.
3620 // Avoid turning single icmps into a switch.
3624 bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3626 // There might be duplicate constants in the list, which the switch
3627 // instruction can't handle, remove them now.
3628 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3629 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3631 // If Extra was used, we require at least two switch values to do the
3632 // transformation. A switch with one value is just a conditional branch.
3633 if (ExtraCase && Values.size() < 2)
3636 // TODO: Preserve branch weight metadata, similarly to how
3637 // FoldValueComparisonIntoPredecessors preserves it.
3639 // Figure out which block is which destination.
3640 BasicBlock *DefaultBB = BI->getSuccessor(1);
3641 BasicBlock *EdgeBB = BI->getSuccessor(0);
3643 std::swap(DefaultBB, EdgeBB);
3645 BasicBlock *BB = BI->getParent();
3647 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3648 << " cases into SWITCH. BB is:\n"
3651 // If there are any extra values that couldn't be folded into the switch
3652 // then we evaluate them with an explicit branch first. Split the block
3653 // right before the condbr to handle it.
3656 BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3657 // Remove the uncond branch added to the old block.
3658 TerminatorInst *OldTI = BB->getTerminator();
3659 Builder.SetInsertPoint(OldTI);
3662 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3664 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3666 OldTI->eraseFromParent();
3668 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3669 // for the edge we just added.
3670 AddPredecessorToBlock(EdgeBB, BB, NewBB);
3672 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3673 << "\nEXTRABB = " << *BB);
3677 Builder.SetInsertPoint(BI);
3678 // Convert pointer to int before we switch.
3679 if (CompVal->getType()->isPointerTy()) {
3680 CompVal = Builder.CreatePtrToInt(
3681 CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3684 // Create the new switch instruction now.
3685 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3687 // Add all of the 'cases' to the switch instruction.
3688 for (unsigned i = 0, e = Values.size(); i != e; ++i)
3689 New->addCase(Values[i], EdgeBB);
3691 // We added edges from PI to the EdgeBB. As such, if there were any
3692 // PHI nodes in EdgeBB, they need entries to be added corresponding to
3693 // the number of edges added.
3694 for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3695 PHINode *PN = cast<PHINode>(BBI);
3696 Value *InVal = PN->getIncomingValueForBlock(BB);
3697 for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3698 PN->addIncoming(InVal, BB);
3701 // Erase the old branch instruction.
3702 EraseTerminatorInstAndDCECond(BI);
3704 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3708 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3709 if (isa<PHINode>(RI->getValue()))
3710 return SimplifyCommonResume(RI);
3711 else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3712 RI->getValue() == RI->getParent()->getFirstNonPHI())
3713 // The resume must unwind the exception that caused control to branch here.
3714 return SimplifySingleResume(RI);
3719 // Simplify resume that is shared by several landing pads (phi of landing pad).
3720 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3721 BasicBlock *BB = RI->getParent();
3723 // Check that there are no other instructions except for debug intrinsics
3724 // between the phi of landing pads (RI->getValue()) and resume instruction.
3725 BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3726 E = RI->getIterator();
3728 if (!isa<DbgInfoIntrinsic>(I))
3731 SmallSet<BasicBlock *, 4> TrivialUnwindBlocks;
3732 auto *PhiLPInst = cast<PHINode>(RI->getValue());
3734 // Check incoming blocks to see if any of them are trivial.
3735 for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3737 auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3738 auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3740 // If the block has other successors, we can not delete it because
3741 // it has other dependents.
3742 if (IncomingBB->getUniqueSuccessor() != BB)
3745 auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3746 // Not the landing pad that caused the control to branch here.
3747 if (IncomingValue != LandingPad)
3750 bool isTrivial = true;
3752 I = IncomingBB->getFirstNonPHI()->getIterator();
3753 E = IncomingBB->getTerminator()->getIterator();
3755 if (!isa<DbgInfoIntrinsic>(I)) {
3761 TrivialUnwindBlocks.insert(IncomingBB);
3764 // If no trivial unwind blocks, don't do any simplifications.
3765 if (TrivialUnwindBlocks.empty())
3768 // Turn all invokes that unwind here into calls.
3769 for (auto *TrivialBB : TrivialUnwindBlocks) {
3770 // Blocks that will be simplified should be removed from the phi node.
3771 // Note there could be multiple edges to the resume block, and we need
3772 // to remove them all.
3773 while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3774 BB->removePredecessor(TrivialBB, true);
3776 for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3778 BasicBlock *Pred = *PI++;
3779 removeUnwindEdge(Pred);
3782 // In each SimplifyCFG run, only the current processed block can be erased.
3783 // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3784 // of erasing TrivialBB, we only remove the branch to the common resume
3785 // block so that we can later erase the resume block since it has no
3787 TrivialBB->getTerminator()->eraseFromParent();
3788 new UnreachableInst(RI->getContext(), TrivialBB);
3791 // Delete the resume block if all its predecessors have been removed.
3793 BB->eraseFromParent();
3795 return !TrivialUnwindBlocks.empty();
3798 // Simplify resume that is only used by a single (non-phi) landing pad.
3799 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3800 BasicBlock *BB = RI->getParent();
3801 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
3802 assert(RI->getValue() == LPInst &&
3803 "Resume must unwind the exception that caused control to here");
3805 // Check that there are no other instructions except for debug intrinsics.
3806 BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3808 if (!isa<DbgInfoIntrinsic>(I))
3811 // Turn all invokes that unwind here into calls and delete the basic block.
3812 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3813 BasicBlock *Pred = *PI++;
3814 removeUnwindEdge(Pred);
3817 // The landingpad is now unreachable. Zap it.
3818 BB->eraseFromParent();
3820 LoopHeaders->erase(BB);
3824 static bool removeEmptyCleanup(CleanupReturnInst *RI) {
3825 // If this is a trivial cleanup pad that executes no instructions, it can be
3826 // eliminated. If the cleanup pad continues to the caller, any predecessor
3827 // that is an EH pad will be updated to continue to the caller and any
3828 // predecessor that terminates with an invoke instruction will have its invoke
3829 // instruction converted to a call instruction. If the cleanup pad being
3830 // simplified does not continue to the caller, each predecessor will be
3831 // updated to continue to the unwind destination of the cleanup pad being
3833 BasicBlock *BB = RI->getParent();
3834 CleanupPadInst *CPInst = RI->getCleanupPad();
3835 if (CPInst->getParent() != BB)
3836 // This isn't an empty cleanup.
3839 // We cannot kill the pad if it has multiple uses. This typically arises
3840 // from unreachable basic blocks.
3841 if (!CPInst->hasOneUse())
3844 // Check that there are no other instructions except for benign intrinsics.
3845 BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3847 auto *II = dyn_cast<IntrinsicInst>(I);
3851 Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3852 switch (IntrinsicID) {
3853 case Intrinsic::dbg_declare:
3854 case Intrinsic::dbg_value:
3855 case Intrinsic::lifetime_end:
3862 // If the cleanup return we are simplifying unwinds to the caller, this will
3863 // set UnwindDest to nullptr.
3864 BasicBlock *UnwindDest = RI->getUnwindDest();
3865 Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3867 // We're about to remove BB from the control flow. Before we do, sink any
3868 // PHINodes into the unwind destination. Doing this before changing the
3869 // control flow avoids some potentially slow checks, since we can currently
3870 // be certain that UnwindDest and BB have no common predecessors (since they
3871 // are both EH pads).
3873 // First, go through the PHI nodes in UnwindDest and update any nodes that
3874 // reference the block we are removing
3875 for (BasicBlock::iterator I = UnwindDest->begin(),
3876 IE = DestEHPad->getIterator();
3878 PHINode *DestPN = cast<PHINode>(I);
3880 int Idx = DestPN->getBasicBlockIndex(BB);
3881 // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3883 // This PHI node has an incoming value that corresponds to a control
3884 // path through the cleanup pad we are removing. If the incoming
3885 // value is in the cleanup pad, it must be a PHINode (because we
3886 // verified above that the block is otherwise empty). Otherwise, the
3887 // value is either a constant or a value that dominates the cleanup
3888 // pad being removed.
3890 // Because BB and UnwindDest are both EH pads, all of their
3891 // predecessors must unwind to these blocks, and since no instruction
3892 // can have multiple unwind destinations, there will be no overlap in
3893 // incoming blocks between SrcPN and DestPN.
3894 Value *SrcVal = DestPN->getIncomingValue(Idx);
3895 PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3897 // Remove the entry for the block we are deleting.
3898 DestPN->removeIncomingValue(Idx, false);
3900 if (SrcPN && SrcPN->getParent() == BB) {
3901 // If the incoming value was a PHI node in the cleanup pad we are
3902 // removing, we need to merge that PHI node's incoming values into
3904 for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3905 SrcIdx != SrcE; ++SrcIdx) {
3906 DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3907 SrcPN->getIncomingBlock(SrcIdx));
3910 // Otherwise, the incoming value came from above BB and
3911 // so we can just reuse it. We must associate all of BB's
3912 // predecessors with this value.
3913 for (auto *pred : predecessors(BB)) {
3914 DestPN->addIncoming(SrcVal, pred);
3919 // Sink any remaining PHI nodes directly into UnwindDest.
3920 Instruction *InsertPt = DestEHPad;
3921 for (BasicBlock::iterator I = BB->begin(),
3922 IE = BB->getFirstNonPHI()->getIterator();
3924 // The iterator must be incremented here because the instructions are
3925 // being moved to another block.
3926 PHINode *PN = cast<PHINode>(I++);
3927 if (PN->use_empty())
3928 // If the PHI node has no uses, just leave it. It will be erased
3929 // when we erase BB below.
3932 // Otherwise, sink this PHI node into UnwindDest.
3933 // Any predecessors to UnwindDest which are not already represented
3934 // must be back edges which inherit the value from the path through
3935 // BB. In this case, the PHI value must reference itself.
3936 for (auto *pred : predecessors(UnwindDest))
3938 PN->addIncoming(PN, pred);
3939 PN->moveBefore(InsertPt);
3943 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3944 // The iterator must be updated here because we are removing this pred.
3945 BasicBlock *PredBB = *PI++;
3946 if (UnwindDest == nullptr) {
3947 removeUnwindEdge(PredBB);
3949 TerminatorInst *TI = PredBB->getTerminator();
3950 TI->replaceUsesOfWith(BB, UnwindDest);
3954 // The cleanup pad is now unreachable. Zap it.
3955 BB->eraseFromParent();
3959 // Try to merge two cleanuppads together.
3960 static bool mergeCleanupPad(CleanupReturnInst *RI) {
3961 // Skip any cleanuprets which unwind to caller, there is nothing to merge
3963 BasicBlock *UnwindDest = RI->getUnwindDest();
3967 // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
3968 // be safe to merge without code duplication.
3969 if (UnwindDest->getSinglePredecessor() != RI->getParent())
3972 // Verify that our cleanuppad's unwind destination is another cleanuppad.
3973 auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
3974 if (!SuccessorCleanupPad)
3977 CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
3978 // Replace any uses of the successor cleanupad with the predecessor pad
3979 // The only cleanuppad uses should be this cleanupret, it's cleanupret and
3980 // funclet bundle operands.
3981 SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
3982 // Remove the old cleanuppad.
3983 SuccessorCleanupPad->eraseFromParent();
3984 // Now, we simply replace the cleanupret with a branch to the unwind
3986 BranchInst::Create(UnwindDest, RI->getParent());
3987 RI->eraseFromParent();
3992 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
3993 // It is possible to transiantly have an undef cleanuppad operand because we
3994 // have deleted some, but not all, dead blocks.
3995 // Eventually, this block will be deleted.
3996 if (isa<UndefValue>(RI->getOperand(0)))
3999 if (mergeCleanupPad(RI))
4002 if (removeEmptyCleanup(RI))
4008 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4009 BasicBlock *BB = RI->getParent();
4010 if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4013 // Find predecessors that end with branches.
4014 SmallVector<BasicBlock *, 8> UncondBranchPreds;
4015 SmallVector<BranchInst *, 8> CondBranchPreds;
4016 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4017 BasicBlock *P = *PI;
4018 TerminatorInst *PTI = P->getTerminator();
4019 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4020 if (BI->isUnconditional())
4021 UncondBranchPreds.push_back(P);
4023 CondBranchPreds.push_back(BI);
4027 // If we found some, do the transformation!
4028 if (!UncondBranchPreds.empty() && DupRet) {
4029 while (!UncondBranchPreds.empty()) {
4030 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4031 DEBUG(dbgs() << "FOLDING: " << *BB
4032 << "INTO UNCOND BRANCH PRED: " << *Pred);
4033 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4036 // If we eliminated all predecessors of the block, delete the block now.
4037 if (pred_empty(BB)) {
4038 // We know there are no successors, so just nuke the block.
4039 BB->eraseFromParent();
4041 LoopHeaders->erase(BB);
4047 // Check out all of the conditional branches going to this return
4048 // instruction. If any of them just select between returns, change the
4049 // branch itself into a select/return pair.
4050 while (!CondBranchPreds.empty()) {
4051 BranchInst *BI = CondBranchPreds.pop_back_val();
4053 // Check to see if the non-BB successor is also a return block.
4054 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4055 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4056 SimplifyCondBranchToTwoReturns(BI, Builder))
4062 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4063 BasicBlock *BB = UI->getParent();
4065 bool Changed = false;
4067 // If there are any instructions immediately before the unreachable that can
4068 // be removed, do so.
4069 while (UI->getIterator() != BB->begin()) {
4070 BasicBlock::iterator BBI = UI->getIterator();
4072 // Do not delete instructions that can have side effects which might cause
4073 // the unreachable to not be reachable; specifically, calls and volatile
4074 // operations may have this effect.
4075 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4078 if (BBI->mayHaveSideEffects()) {
4079 if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4080 if (SI->isVolatile())
4082 } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4083 if (LI->isVolatile())
4085 } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4086 if (RMWI->isVolatile())
4088 } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4089 if (CXI->isVolatile())
4091 } else if (isa<CatchPadInst>(BBI)) {
4092 // A catchpad may invoke exception object constructors and such, which
4093 // in some languages can be arbitrary code, so be conservative by
4095 // For CoreCLR, it just involves a type test, so can be removed.
4096 if (classifyEHPersonality(BB->getParent()->getPersonalityFn()) !=
4097 EHPersonality::CoreCLR)
4099 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4100 !isa<LandingPadInst>(BBI)) {
4103 // Note that deleting LandingPad's here is in fact okay, although it
4104 // involves a bit of subtle reasoning. If this inst is a LandingPad,
4105 // all the predecessors of this block will be the unwind edges of Invokes,
4106 // and we can therefore guarantee this block will be erased.
4109 // Delete this instruction (any uses are guaranteed to be dead)
4110 if (!BBI->use_empty())
4111 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4112 BBI->eraseFromParent();
4116 // If the unreachable instruction is the first in the block, take a gander
4117 // at all of the predecessors of this instruction, and simplify them.
4118 if (&BB->front() != UI)
4121 SmallVector<BasicBlock *, 8> Preds(pred_begin(BB), pred_end(BB));
4122 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4123 TerminatorInst *TI = Preds[i]->getTerminator();
4124 IRBuilder<> Builder(TI);
4125 if (auto *BI = dyn_cast<BranchInst>(TI)) {
4126 if (BI->isUnconditional()) {
4127 if (BI->getSuccessor(0) == BB) {
4128 new UnreachableInst(TI->getContext(), TI);
4129 TI->eraseFromParent();
4133 if (BI->getSuccessor(0) == BB) {
4134 Builder.CreateBr(BI->getSuccessor(1));
4135 EraseTerminatorInstAndDCECond(BI);
4136 } else if (BI->getSuccessor(1) == BB) {
4137 Builder.CreateBr(BI->getSuccessor(0));
4138 EraseTerminatorInstAndDCECond(BI);
4142 } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4143 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e;
4145 if (i.getCaseSuccessor() == BB) {
4146 BB->removePredecessor(SI->getParent());
4152 } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4153 if (II->getUnwindDest() == BB) {
4154 removeUnwindEdge(TI->getParent());
4157 } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4158 if (CSI->getUnwindDest() == BB) {
4159 removeUnwindEdge(TI->getParent());
4164 for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4165 E = CSI->handler_end();
4168 CSI->removeHandler(I);
4174 if (CSI->getNumHandlers() == 0) {
4175 BasicBlock *CatchSwitchBB = CSI->getParent();
4176 if (CSI->hasUnwindDest()) {
4177 // Redirect preds to the unwind dest
4178 CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4180 // Rewrite all preds to unwind to caller (or from invoke to call).
4181 SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4182 for (BasicBlock *EHPred : EHPreds)
4183 removeUnwindEdge(EHPred);
4185 // The catchswitch is no longer reachable.
4186 new UnreachableInst(CSI->getContext(), CSI);
4187 CSI->eraseFromParent();
4190 } else if (isa<CleanupReturnInst>(TI)) {
4191 new UnreachableInst(TI->getContext(), TI);
4192 TI->eraseFromParent();
4197 // If this block is now dead, remove it.
4198 if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4199 // We know there are no successors, so just nuke the block.
4200 BB->eraseFromParent();
4202 LoopHeaders->erase(BB);
4209 static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) {
4210 assert(Cases.size() >= 1);
4212 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4213 for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4214 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4220 /// Turn a switch with two reachable destinations into an integer range
4221 /// comparison and branch.
4222 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
4223 assert(SI->getNumCases() > 1 && "Degenerate switch?");
4226 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4228 // Partition the cases into two sets with different destinations.
4229 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4230 BasicBlock *DestB = nullptr;
4231 SmallVector<ConstantInt *, 16> CasesA;
4232 SmallVector<ConstantInt *, 16> CasesB;
4234 for (SwitchInst::CaseIt I : SI->cases()) {
4235 BasicBlock *Dest = I.getCaseSuccessor();
4238 if (Dest == DestA) {
4239 CasesA.push_back(I.getCaseValue());
4244 if (Dest == DestB) {
4245 CasesB.push_back(I.getCaseValue());
4248 return false; // More than two destinations.
4251 assert(DestA && DestB &&
4252 "Single-destination switch should have been folded.");
4253 assert(DestA != DestB);
4254 assert(DestB != SI->getDefaultDest());
4255 assert(!CasesB.empty() && "There must be non-default cases.");
4256 assert(!CasesA.empty() || HasDefault);
4258 // Figure out if one of the sets of cases form a contiguous range.
4259 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4260 BasicBlock *ContiguousDest = nullptr;
4261 BasicBlock *OtherDest = nullptr;
4262 if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4263 ContiguousCases = &CasesA;
4264 ContiguousDest = DestA;
4266 } else if (CasesAreContiguous(CasesB)) {
4267 ContiguousCases = &CasesB;
4268 ContiguousDest = DestB;
4273 // Start building the compare and branch.
4275 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4276 Constant *NumCases =
4277 ConstantInt::get(Offset->getType(), ContiguousCases->size());
4279 Value *Sub = SI->getCondition();
4280 if (!Offset->isNullValue())
4281 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4284 // If NumCases overflowed, then all possible values jump to the successor.
4285 if (NumCases->isNullValue() && !ContiguousCases->empty())
4286 Cmp = ConstantInt::getTrue(SI->getContext());
4288 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4289 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4291 // Update weight for the newly-created conditional branch.
4292 if (HasBranchWeights(SI)) {
4293 SmallVector<uint64_t, 8> Weights;
4294 GetBranchWeights(SI, Weights);
4295 if (Weights.size() == 1 + SI->getNumCases()) {
4296 uint64_t TrueWeight = 0;
4297 uint64_t FalseWeight = 0;
4298 for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4299 if (SI->getSuccessor(I) == ContiguousDest)
4300 TrueWeight += Weights[I];
4302 FalseWeight += Weights[I];
4304 while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4308 NewBI->setMetadata(LLVMContext::MD_prof,
4309 MDBuilder(SI->getContext())
4310 .createBranchWeights((uint32_t)TrueWeight,
4311 (uint32_t)FalseWeight));
4315 // Prune obsolete incoming values off the successors' PHI nodes.
4316 for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4317 unsigned PreviousEdges = ContiguousCases->size();
4318 if (ContiguousDest == SI->getDefaultDest())
4320 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4321 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4323 for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4324 unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4325 if (OtherDest == SI->getDefaultDest())
4327 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4328 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4332 SI->eraseFromParent();
4337 /// Compute masked bits for the condition of a switch
4338 /// and use it to remove dead cases.
4339 static bool EliminateDeadSwitchCases(SwitchInst *SI, AssumptionCache *AC,
4340 const DataLayout &DL) {
4341 Value *Cond = SI->getCondition();
4342 unsigned Bits = Cond->getType()->getIntegerBitWidth();
4343 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
4344 computeKnownBits(Cond, KnownZero, KnownOne, DL, 0, AC, SI);
4346 // We can also eliminate cases by determining that their values are outside of
4347 // the limited range of the condition based on how many significant (non-sign)
4348 // bits are in the condition value.
4349 unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4350 unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4352 // Gather dead cases.
4353 SmallVector<ConstantInt *, 8> DeadCases;
4354 for (auto &Case : SI->cases()) {
4355 APInt CaseVal = Case.getCaseValue()->getValue();
4356 if ((CaseVal & KnownZero) != 0 || (CaseVal & KnownOne) != KnownOne ||
4357 (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4358 DeadCases.push_back(Case.getCaseValue());
4359 DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal << " is dead.\n");
4363 // If we can prove that the cases must cover all possible values, the
4364 // default destination becomes dead and we can remove it. If we know some
4365 // of the bits in the value, we can use that to more precisely compute the
4366 // number of possible unique case values.
4368 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4369 const unsigned NumUnknownBits =
4370 Bits - (KnownZero.Or(KnownOne)).countPopulation();
4371 assert(NumUnknownBits <= Bits);
4372 if (HasDefault && DeadCases.empty() &&
4373 NumUnknownBits < 64 /* avoid overflow */ &&
4374 SI->getNumCases() == (1ULL << NumUnknownBits)) {
4375 DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4376 BasicBlock *NewDefault =
4377 SplitBlockPredecessors(SI->getDefaultDest(), SI->getParent(), "");
4378 SI->setDefaultDest(&*NewDefault);
4379 SplitBlock(&*NewDefault, &NewDefault->front());
4380 auto *OldTI = NewDefault->getTerminator();
4381 new UnreachableInst(SI->getContext(), OldTI);
4382 EraseTerminatorInstAndDCECond(OldTI);
4386 SmallVector<uint64_t, 8> Weights;
4387 bool HasWeight = HasBranchWeights(SI);
4389 GetBranchWeights(SI, Weights);
4390 HasWeight = (Weights.size() == 1 + SI->getNumCases());
4393 // Remove dead cases from the switch.
4394 for (ConstantInt *DeadCase : DeadCases) {
4395 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCase);
4396 assert(Case != SI->case_default() &&
4397 "Case was not found. Probably mistake in DeadCases forming.");
4399 std::swap(Weights[Case.getCaseIndex() + 1], Weights.back());
4403 // Prune unused values from PHI nodes.
4404 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
4405 SI->removeCase(Case);
4407 if (HasWeight && Weights.size() >= 2) {
4408 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
4409 SI->setMetadata(LLVMContext::MD_prof,
4410 MDBuilder(SI->getParent()->getContext())
4411 .createBranchWeights(MDWeights));
4414 return !DeadCases.empty();
4417 /// If BB would be eligible for simplification by
4418 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4419 /// by an unconditional branch), look at the phi node for BB in the successor
4420 /// block and see if the incoming value is equal to CaseValue. If so, return
4421 /// the phi node, and set PhiIndex to BB's index in the phi node.
4422 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
4423 BasicBlock *BB, int *PhiIndex) {
4424 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4425 return nullptr; // BB must be empty to be a candidate for simplification.
4426 if (!BB->getSinglePredecessor())
4427 return nullptr; // BB must be dominated by the switch.
4429 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
4430 if (!Branch || !Branch->isUnconditional())
4431 return nullptr; // Terminator must be unconditional branch.
4433 BasicBlock *Succ = Branch->getSuccessor(0);
4435 BasicBlock::iterator I = Succ->begin();
4436 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
4437 int Idx = PHI->getBasicBlockIndex(BB);
4438 assert(Idx >= 0 && "PHI has no entry for predecessor?");
4440 Value *InValue = PHI->getIncomingValue(Idx);
4441 if (InValue != CaseValue)
4451 /// Try to forward the condition of a switch instruction to a phi node
4452 /// dominated by the switch, if that would mean that some of the destination
4453 /// blocks of the switch can be folded away.
4454 /// Returns true if a change is made.
4455 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
4456 typedef DenseMap<PHINode *, SmallVector<int, 4>> ForwardingNodesMap;
4457 ForwardingNodesMap ForwardingNodes;
4459 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E;
4461 ConstantInt *CaseValue = I.getCaseValue();
4462 BasicBlock *CaseDest = I.getCaseSuccessor();
4466 FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIndex);
4470 ForwardingNodes[PHI].push_back(PhiIndex);
4473 bool Changed = false;
4475 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
4476 E = ForwardingNodes.end();
4478 PHINode *Phi = I->first;
4479 SmallVectorImpl<int> &Indexes = I->second;
4481 if (Indexes.size() < 2)
4484 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
4485 Phi->setIncomingValue(Indexes[I], SI->getCondition());
4492 /// Return true if the backend will be able to handle
4493 /// initializing an array of constants like C.
4494 static bool ValidLookupTableConstant(Constant *C, const TargetTransformInfo &TTI) {
4495 if (C->isThreadDependent())
4497 if (C->isDLLImportDependent())
4500 if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4501 !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4502 !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4505 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4506 if (!CE->isGEPWithNoNotionalOverIndexing())
4508 if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4512 if (!TTI.shouldBuildLookupTablesForConstant(C))
4518 /// If V is a Constant, return it. Otherwise, try to look up
4519 /// its constant value in ConstantPool, returning 0 if it's not there.
4521 LookupConstant(Value *V,
4522 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4523 if (Constant *C = dyn_cast<Constant>(V))
4525 return ConstantPool.lookup(V);
4528 /// Try to fold instruction I into a constant. This works for
4529 /// simple instructions such as binary operations where both operands are
4530 /// constant or can be replaced by constants from the ConstantPool. Returns the
4531 /// resulting constant on success, 0 otherwise.
4533 ConstantFold(Instruction *I, const DataLayout &DL,
4534 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4535 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4536 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4539 if (A->isAllOnesValue())
4540 return LookupConstant(Select->getTrueValue(), ConstantPool);
4541 if (A->isNullValue())
4542 return LookupConstant(Select->getFalseValue(), ConstantPool);
4546 SmallVector<Constant *, 4> COps;
4547 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4548 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
4554 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4555 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4559 return ConstantFoldInstOperands(I, COps, DL);
4562 /// Try to determine the resulting constant values in phi nodes
4563 /// at the common destination basic block, *CommonDest, for one of the case
4564 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4565 /// case), of a switch instruction SI.
4567 GetCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest,
4568 BasicBlock **CommonDest,
4569 SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4570 const DataLayout &DL, const TargetTransformInfo &TTI) {
4571 // The block from which we enter the common destination.
4572 BasicBlock *Pred = SI->getParent();
4574 // If CaseDest is empty except for some side-effect free instructions through
4575 // which we can constant-propagate the CaseVal, continue to its successor.
4576 SmallDenseMap<Value *, Constant *> ConstantPool;
4577 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4578 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
4580 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
4581 // If the terminator is a simple branch, continue to the next block.
4582 if (T->getNumSuccessors() != 1 || T->isExceptional())
4585 CaseDest = T->getSuccessor(0);
4586 } else if (isa<DbgInfoIntrinsic>(I)) {
4587 // Skip debug intrinsic.
4589 } else if (Constant *C = ConstantFold(&*I, DL, ConstantPool)) {
4590 // Instruction is side-effect free and constant.
4592 // If the instruction has uses outside this block or a phi node slot for
4593 // the block, it is not safe to bypass the instruction since it would then
4594 // no longer dominate all its uses.
4595 for (auto &Use : I->uses()) {
4596 User *User = Use.getUser();
4597 if (Instruction *I = dyn_cast<Instruction>(User))
4598 if (I->getParent() == CaseDest)
4600 if (PHINode *Phi = dyn_cast<PHINode>(User))
4601 if (Phi->getIncomingBlock(Use) == CaseDest)
4606 ConstantPool.insert(std::make_pair(&*I, C));
4612 // If we did not have a CommonDest before, use the current one.
4614 *CommonDest = CaseDest;
4615 // If the destination isn't the common one, abort.
4616 if (CaseDest != *CommonDest)
4619 // Get the values for this case from phi nodes in the destination block.
4620 BasicBlock::iterator I = (*CommonDest)->begin();
4621 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
4622 int Idx = PHI->getBasicBlockIndex(Pred);
4626 Constant *ConstVal =
4627 LookupConstant(PHI->getIncomingValue(Idx), ConstantPool);
4631 // Be conservative about which kinds of constants we support.
4632 if (!ValidLookupTableConstant(ConstVal, TTI))
4635 Res.push_back(std::make_pair(PHI, ConstVal));
4638 return Res.size() > 0;
4641 // Helper function used to add CaseVal to the list of cases that generate
4643 static void MapCaseToResult(ConstantInt *CaseVal,
4644 SwitchCaseResultVectorTy &UniqueResults,
4646 for (auto &I : UniqueResults) {
4647 if (I.first == Result) {
4648 I.second.push_back(CaseVal);
4652 UniqueResults.push_back(
4653 std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4656 // Helper function that initializes a map containing
4657 // results for the PHI node of the common destination block for a switch
4658 // instruction. Returns false if multiple PHI nodes have been found or if
4659 // there is not a common destination block for the switch.
4660 static bool InitializeUniqueCases(SwitchInst *SI, PHINode *&PHI,
4661 BasicBlock *&CommonDest,
4662 SwitchCaseResultVectorTy &UniqueResults,
4663 Constant *&DefaultResult,
4664 const DataLayout &DL,
4665 const TargetTransformInfo &TTI) {
4666 for (auto &I : SI->cases()) {
4667 ConstantInt *CaseVal = I.getCaseValue();
4669 // Resulting value at phi nodes for this case value.
4670 SwitchCaseResultsTy Results;
4671 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4675 // Only one value per case is permitted
4676 if (Results.size() > 1)
4678 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4680 // Check the PHI consistency.
4682 PHI = Results[0].first;
4683 else if (PHI != Results[0].first)
4686 // Find the default result value.
4687 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
4688 BasicBlock *DefaultDest = SI->getDefaultDest();
4689 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4691 // If the default value is not found abort unless the default destination
4694 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4695 if ((!DefaultResult &&
4696 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4702 // Helper function that checks if it is possible to transform a switch with only
4703 // two cases (or two cases + default) that produces a result into a select.
4706 // case 10: %0 = icmp eq i32 %a, 10
4707 // return 10; %1 = select i1 %0, i32 10, i32 4
4708 // case 20: ----> %2 = icmp eq i32 %a, 20
4709 // return 2; %3 = select i1 %2, i32 2, i32 %1
4713 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4714 Constant *DefaultResult, Value *Condition,
4715 IRBuilder<> &Builder) {
4716 assert(ResultVector.size() == 2 &&
4717 "We should have exactly two unique results at this point");
4718 // If we are selecting between only two cases transform into a simple
4719 // select or a two-way select if default is possible.
4720 if (ResultVector[0].second.size() == 1 &&
4721 ResultVector[1].second.size() == 1) {
4722 ConstantInt *const FirstCase = ResultVector[0].second[0];
4723 ConstantInt *const SecondCase = ResultVector[1].second[0];
4725 bool DefaultCanTrigger = DefaultResult;
4726 Value *SelectValue = ResultVector[1].first;
4727 if (DefaultCanTrigger) {
4728 Value *const ValueCompare =
4729 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4730 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4731 DefaultResult, "switch.select");
4733 Value *const ValueCompare =
4734 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4735 return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4736 SelectValue, "switch.select");
4742 // Helper function to cleanup a switch instruction that has been converted into
4743 // a select, fixing up PHI nodes and basic blocks.
4744 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
4746 IRBuilder<> &Builder) {
4747 BasicBlock *SelectBB = SI->getParent();
4748 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4749 PHI->removeIncomingValue(SelectBB);
4750 PHI->addIncoming(SelectValue, SelectBB);
4752 Builder.CreateBr(PHI->getParent());
4754 // Remove the switch.
4755 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4756 BasicBlock *Succ = SI->getSuccessor(i);
4758 if (Succ == PHI->getParent())
4760 Succ->removePredecessor(SelectBB);
4762 SI->eraseFromParent();
4765 /// If the switch is only used to initialize one or more
4766 /// phi nodes in a common successor block with only two different
4767 /// constant values, replace the switch with select.
4768 static bool SwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4769 AssumptionCache *AC, const DataLayout &DL,
4770 const TargetTransformInfo &TTI) {
4771 Value *const Cond = SI->getCondition();
4772 PHINode *PHI = nullptr;
4773 BasicBlock *CommonDest = nullptr;
4774 Constant *DefaultResult;
4775 SwitchCaseResultVectorTy UniqueResults;
4776 // Collect all the cases that will deliver the same value from the switch.
4777 if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4780 // Selects choose between maximum two values.
4781 if (UniqueResults.size() != 2)
4783 assert(PHI != nullptr && "PHI for value select not found");
4785 Builder.SetInsertPoint(SI);
4786 Value *SelectValue =
4787 ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4789 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4792 // The switch couldn't be converted into a select.
4798 /// This class represents a lookup table that can be used to replace a switch.
4799 class SwitchLookupTable {
4801 /// Create a lookup table to use as a switch replacement with the contents
4802 /// of Values, using DefaultValue to fill any holes in the table.
4804 Module &M, uint64_t TableSize, ConstantInt *Offset,
4805 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4806 Constant *DefaultValue, const DataLayout &DL);
4808 /// Build instructions with Builder to retrieve the value at
4809 /// the position given by Index in the lookup table.
4810 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4812 /// Return true if a table with TableSize elements of
4813 /// type ElementType would fit in a target-legal register.
4814 static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4818 // Depending on the contents of the table, it can be represented in
4821 // For tables where each element contains the same value, we just have to
4822 // store that single value and return it for each lookup.
4825 // For tables where there is a linear relationship between table index
4826 // and values. We calculate the result with a simple multiplication
4827 // and addition instead of a table lookup.
4830 // For small tables with integer elements, we can pack them into a bitmap
4831 // that fits into a target-legal register. Values are retrieved by
4832 // shift and mask operations.
4835 // The table is stored as an array of values. Values are retrieved by load
4836 // instructions from the table.
4840 // For SingleValueKind, this is the single value.
4841 Constant *SingleValue;
4843 // For BitMapKind, this is the bitmap.
4844 ConstantInt *BitMap;
4845 IntegerType *BitMapElementTy;
4847 // For LinearMapKind, these are the constants used to derive the value.
4848 ConstantInt *LinearOffset;
4849 ConstantInt *LinearMultiplier;
4851 // For ArrayKind, this is the array.
4852 GlobalVariable *Array;
4855 } // end anonymous namespace
4857 SwitchLookupTable::SwitchLookupTable(
4858 Module &M, uint64_t TableSize, ConstantInt *Offset,
4859 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4860 Constant *DefaultValue, const DataLayout &DL)
4861 : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
4862 LinearOffset(nullptr), LinearMultiplier(nullptr), Array(nullptr) {
4863 assert(Values.size() && "Can't build lookup table without values!");
4864 assert(TableSize >= Values.size() && "Can't fit values in table!");
4866 // If all values in the table are equal, this is that value.
4867 SingleValue = Values.begin()->second;
4869 Type *ValueType = Values.begin()->second->getType();
4871 // Build up the table contents.
4872 SmallVector<Constant *, 64> TableContents(TableSize);
4873 for (size_t I = 0, E = Values.size(); I != E; ++I) {
4874 ConstantInt *CaseVal = Values[I].first;
4875 Constant *CaseRes = Values[I].second;
4876 assert(CaseRes->getType() == ValueType);
4878 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4879 TableContents[Idx] = CaseRes;
4881 if (CaseRes != SingleValue)
4882 SingleValue = nullptr;
4885 // Fill in any holes in the table with the default result.
4886 if (Values.size() < TableSize) {
4887 assert(DefaultValue &&
4888 "Need a default value to fill the lookup table holes.");
4889 assert(DefaultValue->getType() == ValueType);
4890 for (uint64_t I = 0; I < TableSize; ++I) {
4891 if (!TableContents[I])
4892 TableContents[I] = DefaultValue;
4895 if (DefaultValue != SingleValue)
4896 SingleValue = nullptr;
4899 // If each element in the table contains the same value, we only need to store
4900 // that single value.
4902 Kind = SingleValueKind;
4906 // Check if we can derive the value with a linear transformation from the
4908 if (isa<IntegerType>(ValueType)) {
4909 bool LinearMappingPossible = true;
4912 assert(TableSize >= 2 && "Should be a SingleValue table.");
4913 // Check if there is the same distance between two consecutive values.
4914 for (uint64_t I = 0; I < TableSize; ++I) {
4915 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4917 // This is an undef. We could deal with it, but undefs in lookup tables
4918 // are very seldom. It's probably not worth the additional complexity.
4919 LinearMappingPossible = false;
4922 APInt Val = ConstVal->getValue();
4924 APInt Dist = Val - PrevVal;
4927 } else if (Dist != DistToPrev) {
4928 LinearMappingPossible = false;
4934 if (LinearMappingPossible) {
4935 LinearOffset = cast<ConstantInt>(TableContents[0]);
4936 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
4937 Kind = LinearMapKind;
4943 // If the type is integer and the table fits in a register, build a bitmap.
4944 if (WouldFitInRegister(DL, TableSize, ValueType)) {
4945 IntegerType *IT = cast<IntegerType>(ValueType);
4946 APInt TableInt(TableSize * IT->getBitWidth(), 0);
4947 for (uint64_t I = TableSize; I > 0; --I) {
4948 TableInt <<= IT->getBitWidth();
4949 // Insert values into the bitmap. Undef values are set to zero.
4950 if (!isa<UndefValue>(TableContents[I - 1])) {
4951 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
4952 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
4955 BitMap = ConstantInt::get(M.getContext(), TableInt);
4956 BitMapElementTy = IT;
4962 // Store the table in an array.
4963 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
4964 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
4966 Array = new GlobalVariable(M, ArrayTy, /*constant=*/true,
4967 GlobalVariable::PrivateLinkage, Initializer,
4969 Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
4973 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
4975 case SingleValueKind:
4977 case LinearMapKind: {
4978 // Derive the result value from the input value.
4979 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
4980 false, "switch.idx.cast");
4981 if (!LinearMultiplier->isOne())
4982 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
4983 if (!LinearOffset->isZero())
4984 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
4988 // Type of the bitmap (e.g. i59).
4989 IntegerType *MapTy = BitMap->getType();
4991 // Cast Index to the same type as the bitmap.
4992 // Note: The Index is <= the number of elements in the table, so
4993 // truncating it to the width of the bitmask is safe.
4994 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
4996 // Multiply the shift amount by the element width.
4997 ShiftAmt = Builder.CreateMul(
4998 ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5002 Value *DownShifted =
5003 Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5005 return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5008 // Make sure the table index will not overflow when treated as signed.
5009 IntegerType *IT = cast<IntegerType>(Index->getType());
5010 uint64_t TableSize =
5011 Array->getInitializer()->getType()->getArrayNumElements();
5012 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5013 Index = Builder.CreateZExt(
5014 Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5015 "switch.tableidx.zext");
5017 Value *GEPIndices[] = {Builder.getInt32(0), Index};
5018 Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5019 GEPIndices, "switch.gep");
5020 return Builder.CreateLoad(GEP, "switch.load");
5023 llvm_unreachable("Unknown lookup table kind!");
5026 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5028 Type *ElementType) {
5029 auto *IT = dyn_cast<IntegerType>(ElementType);
5032 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5033 // are <= 15, we could try to narrow the type.
5035 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5036 if (TableSize >= UINT_MAX / IT->getBitWidth())
5038 return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5041 /// Determine whether a lookup table should be built for this switch, based on
5042 /// the number of cases, size of the table, and the types of the results.
5044 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5045 const TargetTransformInfo &TTI, const DataLayout &DL,
5046 const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5047 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5048 return false; // TableSize overflowed, or mul below might overflow.
5050 bool AllTablesFitInRegister = true;
5051 bool HasIllegalType = false;
5052 for (const auto &I : ResultTypes) {
5053 Type *Ty = I.second;
5055 // Saturate this flag to true.
5056 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5058 // Saturate this flag to false.
5059 AllTablesFitInRegister =
5060 AllTablesFitInRegister &&
5061 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5063 // If both flags saturate, we're done. NOTE: This *only* works with
5064 // saturating flags, and all flags have to saturate first due to the
5065 // non-deterministic behavior of iterating over a dense map.
5066 if (HasIllegalType && !AllTablesFitInRegister)
5070 // If each table would fit in a register, we should build it anyway.
5071 if (AllTablesFitInRegister)
5074 // Don't build a table that doesn't fit in-register if it has illegal types.
5078 // The table density should be at least 40%. This is the same criterion as for
5079 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5080 // FIXME: Find the best cut-off.
5081 return SI->getNumCases() * 10 >= TableSize * 4;
5084 /// Try to reuse the switch table index compare. Following pattern:
5086 /// if (idx < tablesize)
5087 /// r = table[idx]; // table does not contain default_value
5089 /// r = default_value;
5090 /// if (r != default_value)
5093 /// Is optimized to:
5095 /// cond = idx < tablesize;
5099 /// r = default_value;
5103 /// Jump threading will then eliminate the second if(cond).
5104 static void reuseTableCompare(
5105 User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5106 Constant *DefaultValue,
5107 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5109 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5113 // We require that the compare is in the same block as the phi so that jump
5114 // threading can do its work afterwards.
5115 if (CmpInst->getParent() != PhiBlock)
5118 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5122 Value *RangeCmp = RangeCheckBranch->getCondition();
5123 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5124 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5126 // Check if the compare with the default value is constant true or false.
5127 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5128 DefaultValue, CmpOp1, true);
5129 if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5132 // Check if the compare with the case values is distinct from the default
5134 for (auto ValuePair : Values) {
5135 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5136 ValuePair.second, CmpOp1, true);
5137 if (!CaseConst || CaseConst == DefaultConst)
5139 assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5140 "Expect true or false as compare result.");
5143 // Check if the branch instruction dominates the phi node. It's a simple
5144 // dominance check, but sufficient for our needs.
5145 // Although this check is invariant in the calling loops, it's better to do it
5146 // at this late stage. Practically we do it at most once for a switch.
5147 BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5148 for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5149 BasicBlock *Pred = *PI;
5150 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5154 if (DefaultConst == FalseConst) {
5155 // The compare yields the same result. We can replace it.
5156 CmpInst->replaceAllUsesWith(RangeCmp);
5157 ++NumTableCmpReuses;
5159 // The compare yields the same result, just inverted. We can replace it.
5160 Value *InvertedTableCmp = BinaryOperator::CreateXor(
5161 RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5163 CmpInst->replaceAllUsesWith(InvertedTableCmp);
5164 ++NumTableCmpReuses;
5168 /// If the switch is only used to initialize one or more phi nodes in a common
5169 /// successor block with different constant values, replace the switch with
5171 static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
5172 const DataLayout &DL,
5173 const TargetTransformInfo &TTI) {
5174 assert(SI->getNumCases() > 1 && "Degenerate switch?");
5176 // Only build lookup table when we have a target that supports it.
5177 if (!TTI.shouldBuildLookupTables())
5180 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5181 // split off a dense part and build a lookup table for that.
5183 // FIXME: This creates arrays of GEPs to constant strings, which means each
5184 // GEP needs a runtime relocation in PIC code. We should just build one big
5185 // string and lookup indices into that.
5187 // Ignore switches with less than three cases. Lookup tables will not make
5189 // faster, so we don't analyze them.
5190 if (SI->getNumCases() < 3)
5193 // Figure out the corresponding result for each case value and phi node in the
5194 // common destination, as well as the min and max case values.
5195 assert(SI->case_begin() != SI->case_end());
5196 SwitchInst::CaseIt CI = SI->case_begin();
5197 ConstantInt *MinCaseVal = CI.getCaseValue();
5198 ConstantInt *MaxCaseVal = CI.getCaseValue();
5200 BasicBlock *CommonDest = nullptr;
5201 typedef SmallVector<std::pair<ConstantInt *, Constant *>, 4> ResultListTy;
5202 SmallDenseMap<PHINode *, ResultListTy> ResultLists;
5203 SmallDenseMap<PHINode *, Constant *> DefaultResults;
5204 SmallDenseMap<PHINode *, Type *> ResultTypes;
5205 SmallVector<PHINode *, 4> PHIs;
5207 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5208 ConstantInt *CaseVal = CI.getCaseValue();
5209 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5210 MinCaseVal = CaseVal;
5211 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5212 MaxCaseVal = CaseVal;
5214 // Resulting value at phi nodes for this case value.
5215 typedef SmallVector<std::pair<PHINode *, Constant *>, 4> ResultsTy;
5217 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
5221 // Append the result from this case to the list for each phi.
5222 for (const auto &I : Results) {
5223 PHINode *PHI = I.first;
5224 Constant *Value = I.second;
5225 if (!ResultLists.count(PHI))
5226 PHIs.push_back(PHI);
5227 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5231 // Keep track of the result types.
5232 for (PHINode *PHI : PHIs) {
5233 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5236 uint64_t NumResults = ResultLists[PHIs[0]].size();
5237 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5238 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5239 bool TableHasHoles = (NumResults < TableSize);
5241 // If the table has holes, we need a constant result for the default case
5242 // or a bitmask that fits in a register.
5243 SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5244 bool HasDefaultResults =
5245 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5246 DefaultResultsList, DL, TTI);
5248 bool NeedMask = (TableHasHoles && !HasDefaultResults);
5250 // As an extra penalty for the validity test we require more cases.
5251 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5253 if (!DL.fitsInLegalInteger(TableSize))
5257 for (const auto &I : DefaultResultsList) {
5258 PHINode *PHI = I.first;
5259 Constant *Result = I.second;
5260 DefaultResults[PHI] = Result;
5263 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5266 // Create the BB that does the lookups.
5267 Module &Mod = *CommonDest->getParent()->getParent();
5268 BasicBlock *LookupBB = BasicBlock::Create(
5269 Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5271 // Compute the table index value.
5272 Builder.SetInsertPoint(SI);
5274 Builder.CreateSub(SI->getCondition(), MinCaseVal, "switch.tableidx");
5276 // Compute the maximum table size representable by the integer type we are
5278 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5279 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5280 assert(MaxTableSize >= TableSize &&
5281 "It is impossible for a switch to have more entries than the max "
5282 "representable value of its input integer type's size.");
5284 // If the default destination is unreachable, or if the lookup table covers
5285 // all values of the conditional variable, branch directly to the lookup table
5286 // BB. Otherwise, check that the condition is within the case range.
5287 const bool DefaultIsReachable =
5288 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5289 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5290 BranchInst *RangeCheckBranch = nullptr;
5292 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5293 Builder.CreateBr(LookupBB);
5294 // Note: We call removeProdecessor later since we need to be able to get the
5295 // PHI value for the default case in case we're using a bit mask.
5297 Value *Cmp = Builder.CreateICmpULT(
5298 TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5300 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5303 // Populate the BB that does the lookups.
5304 Builder.SetInsertPoint(LookupBB);
5307 // Before doing the lookup we do the hole check.
5308 // The LookupBB is therefore re-purposed to do the hole check
5309 // and we create a new LookupBB.
5310 BasicBlock *MaskBB = LookupBB;
5311 MaskBB->setName("switch.hole_check");
5312 LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5313 CommonDest->getParent(), CommonDest);
5315 // Make the mask's bitwidth at least 8bit and a power-of-2 to avoid
5316 // unnecessary illegal types.
5317 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5318 APInt MaskInt(TableSizePowOf2, 0);
5319 APInt One(TableSizePowOf2, 1);
5320 // Build bitmask; fill in a 1 bit for every case.
5321 const ResultListTy &ResultList = ResultLists[PHIs[0]];
5322 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5323 uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5325 MaskInt |= One << Idx;
5327 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5329 // Get the TableIndex'th bit of the bitmask.
5330 // If this bit is 0 (meaning hole) jump to the default destination,
5331 // else continue with table lookup.
5332 IntegerType *MapTy = TableMask->getType();
5334 Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5335 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5336 Value *LoBit = Builder.CreateTrunc(
5337 Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5338 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5340 Builder.SetInsertPoint(LookupBB);
5341 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5344 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5345 // We cached PHINodes in PHIs, to avoid accessing deleted PHINodes later,
5346 // do not delete PHINodes here.
5347 SI->getDefaultDest()->removePredecessor(SI->getParent(),
5348 /*DontDeleteUselessPHIs=*/true);
5351 bool ReturnedEarly = false;
5352 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
5353 PHINode *PHI = PHIs[I];
5354 const ResultListTy &ResultList = ResultLists[PHI];
5356 // If using a bitmask, use any value to fill the lookup table holes.
5357 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5358 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL);
5360 Value *Result = Table.BuildLookup(TableIndex, Builder);
5362 // If the result is used to return immediately from the function, we want to
5363 // do that right here.
5364 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5365 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5366 Builder.CreateRet(Result);
5367 ReturnedEarly = true;
5371 // Do a small peephole optimization: re-use the switch table compare if
5373 if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5374 BasicBlock *PhiBlock = PHI->getParent();
5375 // Search for compare instructions which use the phi.
5376 for (auto *User : PHI->users()) {
5377 reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5381 PHI->addIncoming(Result, LookupBB);
5385 Builder.CreateBr(CommonDest);
5387 // Remove the switch.
5388 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5389 BasicBlock *Succ = SI->getSuccessor(i);
5391 if (Succ == SI->getDefaultDest())
5393 Succ->removePredecessor(SI->getParent());
5395 SI->eraseFromParent();
5399 ++NumLookupTablesHoles;
5403 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5404 // See also SelectionDAGBuilder::isDense(), which this function was based on.
5405 uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5406 uint64_t Range = Diff + 1;
5407 uint64_t NumCases = Values.size();
5408 // 40% is the default density for building a jump table in optsize/minsize mode.
5409 uint64_t MinDensity = 40;
5411 return NumCases * 100 >= Range * MinDensity;
5414 // Try and transform a switch that has "holes" in it to a contiguous sequence
5417 // A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5418 // range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5420 // This converts a sparse switch into a dense switch which allows better
5421 // lowering and could also allow transforming into a lookup table.
5422 static bool ReduceSwitchRange(SwitchInst *SI, IRBuilder<> &Builder,
5423 const DataLayout &DL,
5424 const TargetTransformInfo &TTI) {
5425 auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5426 if (CondTy->getIntegerBitWidth() > 64 ||
5427 !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5429 // Only bother with this optimization if there are more than 3 switch cases;
5430 // SDAG will only bother creating jump tables for 4 or more cases.
5431 if (SI->getNumCases() < 4)
5434 // This transform is agnostic to the signedness of the input or case values. We
5435 // can treat the case values as signed or unsigned. We can optimize more common
5436 // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5438 SmallVector<int64_t,4> Values;
5439 for (auto &C : SI->cases())
5440 Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5441 std::sort(Values.begin(), Values.end());
5443 // If the switch is already dense, there's nothing useful to do here.
5444 if (isSwitchDense(Values))
5447 // First, transform the values such that they start at zero and ascend.
5448 int64_t Base = Values[0];
5449 for (auto &V : Values)
5452 // Now we have signed numbers that have been shifted so that, given enough
5453 // precision, there are no negative values. Since the rest of the transform
5454 // is bitwise only, we switch now to an unsigned representation.
5456 for (auto &V : Values)
5457 GCD = GreatestCommonDivisor64(GCD, (uint64_t)V);
5459 // This transform can be done speculatively because it is so cheap - it results
5460 // in a single rotate operation being inserted. This can only happen if the
5461 // factor extracted is a power of 2.
5462 // FIXME: If the GCD is an odd number we can multiply by the multiplicative
5463 // inverse of GCD and then perform this transform.
5464 // FIXME: It's possible that optimizing a switch on powers of two might also
5465 // be beneficial - flag values are often powers of two and we could use a CLZ
5466 // as the key function.
5467 if (GCD <= 1 || !isPowerOf2_64(GCD))
5468 // No common divisor found or too expensive to compute key function.
5471 unsigned Shift = Log2_64(GCD);
5472 for (auto &V : Values)
5473 V = (int64_t)((uint64_t)V >> Shift);
5475 if (!isSwitchDense(Values))
5476 // Transform didn't create a dense switch.
5479 // The obvious transform is to shift the switch condition right and emit a
5480 // check that the condition actually cleanly divided by GCD, i.e.
5481 // C & (1 << Shift - 1) == 0
5482 // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5484 // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5485 // shift and puts the shifted-off bits in the uppermost bits. If any of these
5486 // are nonzero then the switch condition will be very large and will hit the
5489 auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5490 Builder.SetInsertPoint(SI);
5491 auto *ShiftC = ConstantInt::get(Ty, Shift);
5492 auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5493 auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5494 auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5495 auto *Rot = Builder.CreateOr(LShr, Shl);
5496 SI->replaceUsesOfWith(SI->getCondition(), Rot);
5498 for (SwitchInst::CaseIt C = SI->case_begin(), E = SI->case_end(); C != E;
5500 auto *Orig = C.getCaseValue();
5501 auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5503 cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5508 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
5509 BasicBlock *BB = SI->getParent();
5511 if (isValueEqualityComparison(SI)) {
5512 // If we only have one predecessor, and if it is a branch on this value,
5513 // see if that predecessor totally determines the outcome of this switch.
5514 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5515 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
5516 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5518 Value *Cond = SI->getCondition();
5519 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
5520 if (SimplifySwitchOnSelect(SI, Select))
5521 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5523 // If the block only contains the switch, see if we can fold the block
5524 // away into any preds.
5525 BasicBlock::iterator BBI = BB->begin();
5526 // Ignore dbg intrinsics.
5527 while (isa<DbgInfoIntrinsic>(BBI))
5530 if (FoldValueComparisonIntoPredecessors(SI, Builder))
5531 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5534 // Try to transform the switch into an icmp and a branch.
5535 if (TurnSwitchRangeIntoICmp(SI, Builder))
5536 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5538 // Remove unreachable cases.
5539 if (EliminateDeadSwitchCases(SI, AC, DL))
5540 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5542 if (SwitchToSelect(SI, Builder, AC, DL, TTI))
5543 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5545 if (ForwardSwitchConditionToPHI(SI))
5546 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5548 if (SwitchToLookupTable(SI, Builder, DL, TTI))
5549 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5551 if (ReduceSwitchRange(SI, Builder, DL, TTI))
5552 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5557 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
5558 BasicBlock *BB = IBI->getParent();
5559 bool Changed = false;
5561 // Eliminate redundant destinations.
5562 SmallPtrSet<Value *, 8> Succs;
5563 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
5564 BasicBlock *Dest = IBI->getDestination(i);
5565 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
5566 Dest->removePredecessor(BB);
5567 IBI->removeDestination(i);
5574 if (IBI->getNumDestinations() == 0) {
5575 // If the indirectbr has no successors, change it to unreachable.
5576 new UnreachableInst(IBI->getContext(), IBI);
5577 EraseTerminatorInstAndDCECond(IBI);
5581 if (IBI->getNumDestinations() == 1) {
5582 // If the indirectbr has one successor, change it to a direct branch.
5583 BranchInst::Create(IBI->getDestination(0), IBI);
5584 EraseTerminatorInstAndDCECond(IBI);
5588 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
5589 if (SimplifyIndirectBrOnSelect(IBI, SI))
5590 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5595 /// Given an block with only a single landing pad and a unconditional branch
5596 /// try to find another basic block which this one can be merged with. This
5597 /// handles cases where we have multiple invokes with unique landing pads, but
5598 /// a shared handler.
5600 /// We specifically choose to not worry about merging non-empty blocks
5601 /// here. That is a PRE/scheduling problem and is best solved elsewhere. In
5602 /// practice, the optimizer produces empty landing pad blocks quite frequently
5603 /// when dealing with exception dense code. (see: instcombine, gvn, if-else
5604 /// sinking in this file)
5606 /// This is primarily a code size optimization. We need to avoid performing
5607 /// any transform which might inhibit optimization (such as our ability to
5608 /// specialize a particular handler via tail commoning). We do this by not
5609 /// merging any blocks which require us to introduce a phi. Since the same
5610 /// values are flowing through both blocks, we don't loose any ability to
5611 /// specialize. If anything, we make such specialization more likely.
5613 /// TODO - This transformation could remove entries from a phi in the target
5614 /// block when the inputs in the phi are the same for the two blocks being
5615 /// merged. In some cases, this could result in removal of the PHI entirely.
5616 static bool TryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI,
5618 auto Succ = BB->getUniqueSuccessor();
5620 // If there's a phi in the successor block, we'd likely have to introduce
5621 // a phi into the merged landing pad block.
5622 if (isa<PHINode>(*Succ->begin()))
5625 for (BasicBlock *OtherPred : predecessors(Succ)) {
5626 if (BB == OtherPred)
5628 BasicBlock::iterator I = OtherPred->begin();
5629 LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(I);
5630 if (!LPad2 || !LPad2->isIdenticalTo(LPad))
5632 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {
5634 BranchInst *BI2 = dyn_cast<BranchInst>(I);
5635 if (!BI2 || !BI2->isIdenticalTo(BI))
5638 // We've found an identical block. Update our predecessors to take that
5639 // path instead and make ourselves dead.
5640 SmallSet<BasicBlock *, 16> Preds;
5641 Preds.insert(pred_begin(BB), pred_end(BB));
5642 for (BasicBlock *Pred : Preds) {
5643 InvokeInst *II = cast<InvokeInst>(Pred->getTerminator());
5644 assert(II->getNormalDest() != BB && II->getUnwindDest() == BB &&
5645 "unexpected successor");
5646 II->setUnwindDest(OtherPred);
5649 // The debug info in OtherPred doesn't cover the merged control flow that
5650 // used to go through BB. We need to delete it or update it.
5651 for (auto I = OtherPred->begin(), E = OtherPred->end(); I != E;) {
5652 Instruction &Inst = *I;
5654 if (isa<DbgInfoIntrinsic>(Inst))
5655 Inst.eraseFromParent();
5658 SmallSet<BasicBlock *, 16> Succs;
5659 Succs.insert(succ_begin(BB), succ_end(BB));
5660 for (BasicBlock *Succ : Succs) {
5661 Succ->removePredecessor(BB);
5664 IRBuilder<> Builder(BI);
5665 Builder.CreateUnreachable();
5666 BI->eraseFromParent();
5672 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI,
5673 IRBuilder<> &Builder) {
5674 BasicBlock *BB = BI->getParent();
5676 if (SinkCommon && SinkThenElseCodeToEnd(BI))
5679 // If the Terminator is the only non-phi instruction, simplify the block.
5680 // if LoopHeader is provided, check if the block is a loop header
5681 // (This is for early invocations before loop simplify and vectorization
5682 // to keep canonical loop forms for nested loops.
5683 // These blocks can be eliminated when the pass is invoked later
5684 // in the back-end.)
5685 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg()->getIterator();
5686 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
5687 (!LoopHeaders || !LoopHeaders->count(BB)) &&
5688 TryToSimplifyUncondBranchFromEmptyBlock(BB))
5691 // If the only instruction in the block is a seteq/setne comparison
5692 // against a constant, try to simplify the block.
5693 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
5694 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
5695 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5697 if (I->isTerminator() &&
5698 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, DL, TTI,
5699 BonusInstThreshold, AC))
5703 // See if we can merge an empty landing pad block with another which is
5705 if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(I)) {
5706 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {
5708 if (I->isTerminator() && TryToMergeLandingPad(LPad, BI, BB))
5712 // If this basic block is ONLY a compare and a branch, and if a predecessor
5713 // branches to us and our successor, fold the comparison into the
5714 // predecessor and use logical operations to update the incoming value
5715 // for PHI nodes in common successor.
5716 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
5717 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5721 static BasicBlock *allPredecessorsComeFromSameSource(BasicBlock *BB) {
5722 BasicBlock *PredPred = nullptr;
5723 for (auto *P : predecessors(BB)) {
5724 BasicBlock *PPred = P->getSinglePredecessor();
5725 if (!PPred || (PredPred && PredPred != PPred))
5732 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
5733 BasicBlock *BB = BI->getParent();
5735 // Conditional branch
5736 if (isValueEqualityComparison(BI)) {
5737 // If we only have one predecessor, and if it is a branch on this value,
5738 // see if that predecessor totally determines the outcome of this
5740 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5741 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
5742 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5744 // This block must be empty, except for the setcond inst, if it exists.
5745 // Ignore dbg intrinsics.
5746 BasicBlock::iterator I = BB->begin();
5747 // Ignore dbg intrinsics.
5748 while (isa<DbgInfoIntrinsic>(I))
5751 if (FoldValueComparisonIntoPredecessors(BI, Builder))
5752 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5753 } else if (&*I == cast<Instruction>(BI->getCondition())) {
5755 // Ignore dbg intrinsics.
5756 while (isa<DbgInfoIntrinsic>(I))
5758 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
5759 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5763 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
5764 if (SimplifyBranchOnICmpChain(BI, Builder, DL))
5767 // If this basic block has a single dominating predecessor block and the
5768 // dominating block's condition implies BI's condition, we know the direction
5769 // of the BI branch.
5770 if (BasicBlock *Dom = BB->getSinglePredecessor()) {
5771 auto *PBI = dyn_cast_or_null<BranchInst>(Dom->getTerminator());
5772 if (PBI && PBI->isConditional() &&
5773 PBI->getSuccessor(0) != PBI->getSuccessor(1) &&
5774 (PBI->getSuccessor(0) == BB || PBI->getSuccessor(1) == BB)) {
5775 bool CondIsFalse = PBI->getSuccessor(1) == BB;
5776 Optional<bool> Implication = isImpliedCondition(
5777 PBI->getCondition(), BI->getCondition(), DL, CondIsFalse);
5779 // Turn this into a branch on constant.
5780 auto *OldCond = BI->getCondition();
5781 ConstantInt *CI = *Implication
5782 ? ConstantInt::getTrue(BB->getContext())
5783 : ConstantInt::getFalse(BB->getContext());
5784 BI->setCondition(CI);
5785 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
5786 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5791 // If this basic block is ONLY a compare and a branch, and if a predecessor
5792 // branches to us and one of our successors, fold the comparison into the
5793 // predecessor and use logical operations to pick the right destination.
5794 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
5795 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5797 // We have a conditional branch to two blocks that are only reachable
5798 // from BI. We know that the condbr dominates the two blocks, so see if
5799 // there is any identical code in the "then" and "else" blocks. If so, we
5800 // can hoist it up to the branching block.
5801 if (BI->getSuccessor(0)->getSinglePredecessor()) {
5802 if (BI->getSuccessor(1)->getSinglePredecessor()) {
5803 if (HoistThenElseCodeToIf(BI, TTI))
5804 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5806 // If Successor #1 has multiple preds, we may be able to conditionally
5807 // execute Successor #0 if it branches to Successor #1.
5808 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
5809 if (Succ0TI->getNumSuccessors() == 1 &&
5810 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
5811 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), TTI))
5812 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5814 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
5815 // If Successor #0 has multiple preds, we may be able to conditionally
5816 // execute Successor #1 if it branches to Successor #0.
5817 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
5818 if (Succ1TI->getNumSuccessors() == 1 &&
5819 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
5820 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), TTI))
5821 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5824 // If this is a branch on a phi node in the current block, thread control
5825 // through this block if any PHI node entries are constants.
5826 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
5827 if (PN->getParent() == BI->getParent())
5828 if (FoldCondBranchOnPHI(BI, DL))
5829 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5831 // Scan predecessor blocks for conditional branches.
5832 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
5833 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
5834 if (PBI != BI && PBI->isConditional())
5835 if (SimplifyCondBranchToCondBranch(PBI, BI, DL))
5836 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5838 // Look for diamond patterns.
5839 if (MergeCondStores)
5840 if (BasicBlock *PrevBB = allPredecessorsComeFromSameSource(BB))
5841 if (BranchInst *PBI = dyn_cast<BranchInst>(PrevBB->getTerminator()))
5842 if (PBI != BI && PBI->isConditional())
5843 if (mergeConditionalStores(PBI, BI))
5844 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5849 /// Check if passing a value to an instruction will cause undefined behavior.
5850 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
5851 Constant *C = dyn_cast<Constant>(V);
5858 if (C->isNullValue() || isa<UndefValue>(C)) {
5859 // Only look at the first use, avoid hurting compile time with long uselists
5860 User *Use = *I->user_begin();
5862 // Now make sure that there are no instructions in between that can alter
5863 // control flow (eg. calls)
5864 for (BasicBlock::iterator
5865 i = ++BasicBlock::iterator(I),
5866 UI = BasicBlock::iterator(dyn_cast<Instruction>(Use));
5868 if (i == I->getParent()->end() || i->mayHaveSideEffects())
5871 // Look through GEPs. A load from a GEP derived from NULL is still undefined
5872 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
5873 if (GEP->getPointerOperand() == I)
5874 return passingValueIsAlwaysUndefined(V, GEP);
5876 // Look through bitcasts.
5877 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
5878 return passingValueIsAlwaysUndefined(V, BC);
5880 // Load from null is undefined.
5881 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
5882 if (!LI->isVolatile())
5883 return LI->getPointerAddressSpace() == 0;
5885 // Store to null is undefined.
5886 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
5887 if (!SI->isVolatile())
5888 return SI->getPointerAddressSpace() == 0 &&
5889 SI->getPointerOperand() == I;
5891 // A call to null is undefined.
5892 if (auto CS = CallSite(Use))
5893 return CS.getCalledValue() == I;
5898 /// If BB has an incoming value that will always trigger undefined behavior
5899 /// (eg. null pointer dereference), remove the branch leading here.
5900 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
5901 for (BasicBlock::iterator i = BB->begin();
5902 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
5903 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
5904 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
5905 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
5906 IRBuilder<> Builder(T);
5907 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
5908 BB->removePredecessor(PHI->getIncomingBlock(i));
5909 // Turn uncoditional branches into unreachables and remove the dead
5910 // destination from conditional branches.
5911 if (BI->isUnconditional())
5912 Builder.CreateUnreachable();
5914 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1)
5915 : BI->getSuccessor(0));
5916 BI->eraseFromParent();
5919 // TODO: SwitchInst.
5925 bool SimplifyCFGOpt::run(BasicBlock *BB) {
5926 bool Changed = false;
5928 assert(BB && BB->getParent() && "Block not embedded in function!");
5929 assert(BB->getTerminator() && "Degenerate basic block encountered!");
5931 // Remove basic blocks that have no predecessors (except the entry block)...
5932 // or that just have themself as a predecessor. These are unreachable.
5933 if ((pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) ||
5934 BB->getSinglePredecessor() == BB) {
5935 DEBUG(dbgs() << "Removing BB: \n" << *BB);
5936 DeleteDeadBlock(BB);
5940 // Check to see if we can constant propagate this terminator instruction
5942 Changed |= ConstantFoldTerminator(BB, true);
5944 // Check for and eliminate duplicate PHI nodes in this block.
5945 Changed |= EliminateDuplicatePHINodes(BB);
5947 // Check for and remove branches that will always cause undefined behavior.
5948 Changed |= removeUndefIntroducingPredecessor(BB);
5950 // Merge basic blocks into their predecessor if there is only one distinct
5951 // pred, and if there is only one distinct successor of the predecessor, and
5952 // if there are no PHI nodes.
5954 if (MergeBlockIntoPredecessor(BB))
5957 IRBuilder<> Builder(BB);
5959 // If there is a trivial two-entry PHI node in this basic block, and we can
5960 // eliminate it, do so now.
5961 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
5962 if (PN->getNumIncomingValues() == 2)
5963 Changed |= FoldTwoEntryPHINode(PN, TTI, DL);
5965 Builder.SetInsertPoint(BB->getTerminator());
5966 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
5967 if (BI->isUnconditional()) {
5968 if (SimplifyUncondBranch(BI, Builder))
5971 if (SimplifyCondBranch(BI, Builder))
5974 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
5975 if (SimplifyReturn(RI, Builder))
5977 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
5978 if (SimplifyResume(RI, Builder))
5980 } else if (CleanupReturnInst *RI =
5981 dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
5982 if (SimplifyCleanupReturn(RI))
5984 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
5985 if (SimplifySwitch(SI, Builder))
5987 } else if (UnreachableInst *UI =
5988 dyn_cast<UnreachableInst>(BB->getTerminator())) {
5989 if (SimplifyUnreachable(UI))
5991 } else if (IndirectBrInst *IBI =
5992 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
5993 if (SimplifyIndirectBr(IBI))
6000 /// This function is used to do simplification of a CFG.
6001 /// For example, it adjusts branches to branches to eliminate the extra hop,
6002 /// eliminates unreachable basic blocks, and does other "peephole" optimization
6003 /// of the CFG. It returns true if a modification was made.
6005 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
6006 unsigned BonusInstThreshold, AssumptionCache *AC,
6007 SmallPtrSetImpl<BasicBlock *> *LoopHeaders) {
6008 return SimplifyCFGOpt(TTI, BB->getModule()->getDataLayout(),
6009 BonusInstThreshold, AC, LoopHeaders)