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 // If the debug loc for I1 and I2 are different, as we are combining them
1279 // into one instruction, we do not want to select debug loc randomly from
1281 if (!isa<CallInst>(I1) && I1->getDebugLoc() != I2->getDebugLoc())
1283 DILocation::getMergedLocation(I1->getDebugLoc(), I2->getDebugLoc()));
1285 I2->eraseFromParent();
1290 // Skip debug info if it is not identical.
1291 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1292 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1293 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1294 while (isa<DbgInfoIntrinsic>(I1))
1296 while (isa<DbgInfoIntrinsic>(I2))
1299 } while (I1->isIdenticalToWhenDefined(I2));
1304 // It may not be possible to hoist an invoke.
1305 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1308 for (BasicBlock *Succ : successors(BB1)) {
1310 for (BasicBlock::iterator BBI = Succ->begin();
1311 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1312 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1313 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1317 // Check for passingValueIsAlwaysUndefined here because we would rather
1318 // eliminate undefined control flow then converting it to a select.
1319 if (passingValueIsAlwaysUndefined(BB1V, PN) ||
1320 passingValueIsAlwaysUndefined(BB2V, PN))
1323 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1325 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1330 // Okay, it is safe to hoist the terminator.
1331 Instruction *NT = I1->clone();
1332 BIParent->getInstList().insert(BI->getIterator(), NT);
1333 if (!NT->getType()->isVoidTy()) {
1334 I1->replaceAllUsesWith(NT);
1335 I2->replaceAllUsesWith(NT);
1339 IRBuilder<NoFolder> Builder(NT);
1340 // Hoisting one of the terminators from our successor is a great thing.
1341 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1342 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1343 // nodes, so we insert select instruction to compute the final result.
1344 std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1345 for (BasicBlock *Succ : successors(BB1)) {
1347 for (BasicBlock::iterator BBI = Succ->begin();
1348 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1349 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1350 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1354 // These values do not agree. Insert a select instruction before NT
1355 // that determines the right value.
1356 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1358 SI = cast<SelectInst>(
1359 Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1360 BB1V->getName() + "." + BB2V->getName(), BI));
1362 // Make the PHI node use the select for all incoming values for BB1/BB2
1363 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1364 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1365 PN->setIncomingValue(i, SI);
1369 // Update any PHI nodes in our new successors.
1370 for (BasicBlock *Succ : successors(BB1))
1371 AddPredecessorToBlock(Succ, BIParent, BB1);
1373 EraseTerminatorInstAndDCECond(BI);
1377 // Is it legal to place a variable in operand \c OpIdx of \c I?
1378 // FIXME: This should be promoted to Instruction.
1379 static bool canReplaceOperandWithVariable(const Instruction *I,
1381 // We can't have a PHI with a metadata type.
1382 if (I->getOperand(OpIdx)->getType()->isMetadataTy())
1386 if (!isa<Constant>(I->getOperand(OpIdx)))
1389 switch (I->getOpcode()) {
1392 case Instruction::Call:
1393 case Instruction::Invoke:
1394 // FIXME: many arithmetic intrinsics have no issue taking a
1395 // variable, however it's hard to distingish these from
1396 // specials such as @llvm.frameaddress that require a constant.
1397 if (isa<IntrinsicInst>(I))
1400 // Constant bundle operands may need to retain their constant-ness for
1402 if (ImmutableCallSite(I).isBundleOperand(OpIdx))
1407 case Instruction::ShuffleVector:
1408 // Shufflevector masks are constant.
1410 case Instruction::ExtractValue:
1411 case Instruction::InsertValue:
1412 // All operands apart from the first are constant.
1414 case Instruction::Alloca:
1416 case Instruction::GetElementPtr:
1419 gep_type_iterator It = std::next(gep_type_begin(I), OpIdx - 1);
1420 return It.isSequential();
1424 // All instructions in Insts belong to different blocks that all unconditionally
1425 // branch to a common successor. Analyze each instruction and return true if it
1426 // would be possible to sink them into their successor, creating one common
1427 // instruction instead. For every value that would be required to be provided by
1428 // PHI node (because an operand varies in each input block), add to PHIOperands.
1429 static bool canSinkInstructions(
1430 ArrayRef<Instruction *> Insts,
1431 DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1432 // Prune out obviously bad instructions to move. Any non-store instruction
1433 // must have exactly one use, and we check later that use is by a single,
1434 // common PHI instruction in the successor.
1435 for (auto *I : Insts) {
1436 // These instructions may change or break semantics if moved.
1437 if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1438 I->getType()->isTokenTy())
1440 // Everything must have only one use too, apart from stores which
1442 if (!isa<StoreInst>(I) && !I->hasOneUse())
1446 const Instruction *I0 = Insts.front();
1447 for (auto *I : Insts)
1448 if (!I->isSameOperationAs(I0))
1451 // All instructions in Insts are known to be the same opcode. If they aren't
1452 // stores, check the only user of each is a PHI or in the same block as the
1453 // instruction, because if a user is in the same block as an instruction
1454 // we're contemplating sinking, it must already be determined to be sinkable.
1455 if (!isa<StoreInst>(I0)) {
1456 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1457 auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1458 if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1459 auto *U = cast<Instruction>(*I->user_begin());
1461 PNUse->getParent() == Succ &&
1462 PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1463 U->getParent() == I->getParent();
1468 for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1469 if (I0->getOperand(OI)->getType()->isTokenTy())
1470 // Don't touch any operand of token type.
1473 // Because SROA can't handle speculating stores of selects, try not
1474 // to sink loads or stores of allocas when we'd have to create a PHI for
1475 // the address operand. Also, because it is likely that loads or stores
1476 // of allocas will disappear when Mem2Reg/SROA is run, don't sink them.
1477 // This can cause code churn which can have unintended consequences down
1478 // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1479 // FIXME: This is a workaround for a deficiency in SROA - see
1480 // https://llvm.org/bugs/show_bug.cgi?id=30188
1481 if (OI == 1 && isa<StoreInst>(I0) &&
1482 any_of(Insts, [](const Instruction *I) {
1483 return isa<AllocaInst>(I->getOperand(1));
1486 if (OI == 0 && isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1487 return isa<AllocaInst>(I->getOperand(0));
1491 auto SameAsI0 = [&I0, OI](const Instruction *I) {
1492 assert(I->getNumOperands() == I0->getNumOperands());
1493 return I->getOperand(OI) == I0->getOperand(OI);
1495 if (!all_of(Insts, SameAsI0)) {
1496 if (!canReplaceOperandWithVariable(I0, OI))
1497 // We can't create a PHI from this GEP.
1499 // Don't create indirect calls! The called value is the final operand.
1500 if ((isa<CallInst>(I0) || isa<InvokeInst>(I0)) && OI == OE - 1) {
1501 // FIXME: if the call was *already* indirect, we should do this.
1504 for (auto *I : Insts)
1505 PHIOperands[I].push_back(I->getOperand(OI));
1511 // Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1512 // instruction of every block in Blocks to their common successor, commoning
1513 // into one instruction.
1514 static bool sinkLastInstruction(ArrayRef<BasicBlock*> Blocks) {
1515 auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1517 // canSinkLastInstruction returning true guarantees that every block has at
1518 // least one non-terminator instruction.
1519 SmallVector<Instruction*,4> Insts;
1520 for (auto *BB : Blocks) {
1521 Instruction *I = BB->getTerminator();
1523 I = I->getPrevNode();
1524 } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1525 if (!isa<DbgInfoIntrinsic>(I))
1529 // The only checking we need to do now is that all users of all instructions
1530 // are the same PHI node. canSinkLastInstruction should have checked this but
1531 // it is slightly over-aggressive - it gets confused by commutative instructions
1532 // so double-check it here.
1533 Instruction *I0 = Insts.front();
1534 if (!isa<StoreInst>(I0)) {
1535 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1536 if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1537 auto *U = cast<Instruction>(*I->user_begin());
1543 // We don't need to do any more checking here; canSinkLastInstruction should
1544 // have done it all for us.
1545 SmallVector<Value*, 4> NewOperands;
1546 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1547 // This check is different to that in canSinkLastInstruction. There, we
1548 // cared about the global view once simplifycfg (and instcombine) have
1549 // completed - it takes into account PHIs that become trivially
1550 // simplifiable. However here we need a more local view; if an operand
1551 // differs we create a PHI and rely on instcombine to clean up the very
1552 // small mess we may make.
1553 bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1554 return I->getOperand(O) != I0->getOperand(O);
1557 NewOperands.push_back(I0->getOperand(O));
1561 // Create a new PHI in the successor block and populate it.
1562 auto *Op = I0->getOperand(O);
1563 assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1564 auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1565 Op->getName() + ".sink", &BBEnd->front());
1566 for (auto *I : Insts)
1567 PN->addIncoming(I->getOperand(O), I->getParent());
1568 NewOperands.push_back(PN);
1571 // Arbitrarily use I0 as the new "common" instruction; remap its operands
1572 // and move it to the start of the successor block.
1573 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1574 I0->getOperandUse(O).set(NewOperands[O]);
1575 I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1577 // Update metadata and IR flags.
1578 for (auto *I : Insts)
1580 combineMetadataForCSE(I0, I);
1584 if (!isa<StoreInst>(I0)) {
1585 // canSinkLastInstruction checked that all instructions were used by
1586 // one and only one PHI node. Find that now, RAUW it to our common
1587 // instruction and nuke it.
1588 assert(I0->hasOneUse());
1589 auto *PN = cast<PHINode>(*I0->user_begin());
1590 PN->replaceAllUsesWith(I0);
1591 PN->eraseFromParent();
1594 // Finally nuke all instructions apart from the common instruction.
1595 for (auto *I : Insts)
1597 I->eraseFromParent();
1604 // LockstepReverseIterator - Iterates through instructions
1605 // in a set of blocks in reverse order from the first non-terminator.
1606 // For example (assume all blocks have size n):
1607 // LockstepReverseIterator I([B1, B2, B3]);
1608 // *I-- = [B1[n], B2[n], B3[n]];
1609 // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1610 // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1612 class LockstepReverseIterator {
1613 ArrayRef<BasicBlock*> Blocks;
1614 SmallVector<Instruction*,4> Insts;
1617 LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) :
1625 for (auto *BB : Blocks) {
1626 Instruction *Inst = BB->getTerminator();
1627 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1628 Inst = Inst->getPrevNode();
1630 // Block wasn't big enough.
1634 Insts.push_back(Inst);
1638 bool isValid() const {
1642 void operator -- () {
1645 for (auto *&Inst : Insts) {
1646 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1647 Inst = Inst->getPrevNode();
1648 // Already at beginning of block.
1656 ArrayRef<Instruction*> operator * () const {
1661 } // end anonymous namespace
1663 /// Given an unconditional branch that goes to BBEnd,
1664 /// check whether BBEnd has only two predecessors and the other predecessor
1665 /// ends with an unconditional branch. If it is true, sink any common code
1666 /// in the two predecessors to BBEnd.
1667 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1668 assert(BI1->isUnconditional());
1669 BasicBlock *BBEnd = BI1->getSuccessor(0);
1671 // We support two situations:
1672 // (1) all incoming arcs are unconditional
1673 // (2) one incoming arc is conditional
1675 // (2) is very common in switch defaults and
1676 // else-if patterns;
1679 // else if (b) f(2);
1692 // [end] has two unconditional predecessor arcs and one conditional. The
1693 // conditional refers to the implicit empty 'else' arc. This conditional
1694 // arc can also be caused by an empty default block in a switch.
1696 // In this case, we attempt to sink code from all *unconditional* arcs.
1697 // If we can sink instructions from these arcs (determined during the scan
1698 // phase below) we insert a common successor for all unconditional arcs and
1699 // connect that to [end], to enable sinking:
1712 SmallVector<BasicBlock*,4> UnconditionalPreds;
1713 Instruction *Cond = nullptr;
1714 for (auto *B : predecessors(BBEnd)) {
1715 auto *T = B->getTerminator();
1716 if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1717 UnconditionalPreds.push_back(B);
1718 else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1723 if (UnconditionalPreds.size() < 2)
1726 bool Changed = false;
1727 // We take a two-step approach to tail sinking. First we scan from the end of
1728 // each block upwards in lockstep. If the n'th instruction from the end of each
1729 // block can be sunk, those instructions are added to ValuesToSink and we
1730 // carry on. If we can sink an instruction but need to PHI-merge some operands
1731 // (because they're not identical in each instruction) we add these to
1733 unsigned ScanIdx = 0;
1734 SmallPtrSet<Value*,4> InstructionsToSink;
1735 DenseMap<Instruction*, SmallVector<Value*,4>> PHIOperands;
1736 LockstepReverseIterator LRI(UnconditionalPreds);
1737 while (LRI.isValid() &&
1738 canSinkInstructions(*LRI, PHIOperands)) {
1739 DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0] << "\n");
1740 InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1745 auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1746 unsigned NumPHIdValues = 0;
1747 for (auto *I : *LRI)
1748 for (auto *V : PHIOperands[I])
1749 if (InstructionsToSink.count(V) == 0)
1751 DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1752 unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1753 if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1756 return NumPHIInsts <= 1;
1759 if (ScanIdx > 0 && Cond) {
1760 // Check if we would actually sink anything first! This mutates the CFG and
1761 // adds an extra block. The goal in doing this is to allow instructions that
1762 // couldn't be sunk before to be sunk - obviously, speculatable instructions
1763 // (such as trunc, add) can be sunk and predicated already. So we check that
1764 // we're going to sink at least one non-speculatable instruction.
1767 bool Profitable = false;
1768 while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1769 if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1779 DEBUG(dbgs() << "SINK: Splitting edge\n");
1780 // We have a conditional edge and we're going to sink some instructions.
1781 // Insert a new block postdominating all blocks we're going to sink from.
1782 if (!SplitBlockPredecessors(BI1->getSuccessor(0), UnconditionalPreds,
1784 // Edges couldn't be split.
1789 // Now that we've analyzed all potential sinking candidates, perform the
1790 // actual sink. We iteratively sink the last non-terminator of the source
1791 // blocks into their common successor unless doing so would require too
1792 // many PHI instructions to be generated (currently only one PHI is allowed
1793 // per sunk instruction).
1795 // We can use InstructionsToSink to discount values needing PHI-merging that will
1796 // actually be sunk in a later iteration. This allows us to be more
1797 // aggressive in what we sink. This does allow a false positive where we
1798 // sink presuming a later value will also be sunk, but stop half way through
1799 // and never actually sink it which means we produce more PHIs than intended.
1800 // This is unlikely in practice though.
1801 for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1802 DEBUG(dbgs() << "SINK: Sink: "
1803 << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1806 // Because we've sunk every instruction in turn, the current instruction to
1807 // sink is always at index 0.
1809 if (!ProfitableToSinkInstruction(LRI)) {
1810 // Too many PHIs would be created.
1811 DEBUG(dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1815 if (!sinkLastInstruction(UnconditionalPreds))
1823 /// \brief Determine if we can hoist sink a sole store instruction out of a
1824 /// conditional block.
1826 /// We are looking for code like the following:
1828 /// store i32 %add, i32* %arrayidx2
1829 /// ... // No other stores or function calls (we could be calling a memory
1830 /// ... // function).
1831 /// %cmp = icmp ult %x, %y
1832 /// br i1 %cmp, label %EndBB, label %ThenBB
1834 /// store i32 %add5, i32* %arrayidx2
1838 /// We are going to transform this into:
1840 /// store i32 %add, i32* %arrayidx2
1842 /// %cmp = icmp ult %x, %y
1843 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1844 /// store i32 %add.add5, i32* %arrayidx2
1847 /// \return The pointer to the value of the previous store if the store can be
1848 /// hoisted into the predecessor block. 0 otherwise.
1849 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1850 BasicBlock *StoreBB, BasicBlock *EndBB) {
1851 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1855 // Volatile or atomic.
1856 if (!StoreToHoist->isSimple())
1859 Value *StorePtr = StoreToHoist->getPointerOperand();
1861 // Look for a store to the same pointer in BrBB.
1862 unsigned MaxNumInstToLookAt = 9;
1863 for (Instruction &CurI : reverse(*BrBB)) {
1864 if (!MaxNumInstToLookAt)
1867 if (isa<DbgInfoIntrinsic>(CurI))
1869 --MaxNumInstToLookAt;
1871 // Could be calling an instruction that affects memory like free().
1872 if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1875 if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1876 // Found the previous store make sure it stores to the same location.
1877 if (SI->getPointerOperand() == StorePtr)
1878 // Found the previous store, return its value operand.
1879 return SI->getValueOperand();
1880 return nullptr; // Unknown store.
1887 /// \brief Speculate a conditional basic block flattening the CFG.
1889 /// Note that this is a very risky transform currently. Speculating
1890 /// instructions like this is most often not desirable. Instead, there is an MI
1891 /// pass which can do it with full awareness of the resource constraints.
1892 /// However, some cases are "obvious" and we should do directly. An example of
1893 /// this is speculating a single, reasonably cheap instruction.
1895 /// There is only one distinct advantage to flattening the CFG at the IR level:
1896 /// it makes very common but simplistic optimizations such as are common in
1897 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1898 /// modeling their effects with easier to reason about SSA value graphs.
1901 /// An illustration of this transform is turning this IR:
1904 /// %cmp = icmp ult %x, %y
1905 /// br i1 %cmp, label %EndBB, label %ThenBB
1907 /// %sub = sub %x, %y
1910 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1917 /// %cmp = icmp ult %x, %y
1918 /// %sub = sub %x, %y
1919 /// %cond = select i1 %cmp, 0, %sub
1923 /// \returns true if the conditional block is removed.
1924 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1925 const TargetTransformInfo &TTI) {
1926 // Be conservative for now. FP select instruction can often be expensive.
1927 Value *BrCond = BI->getCondition();
1928 if (isa<FCmpInst>(BrCond))
1931 BasicBlock *BB = BI->getParent();
1932 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1934 // If ThenBB is actually on the false edge of the conditional branch, remember
1935 // to swap the select operands later.
1936 bool Invert = false;
1937 if (ThenBB != BI->getSuccessor(0)) {
1938 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1941 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1943 // Keep a count of how many times instructions are used within CondBB when
1944 // they are candidates for sinking into CondBB. Specifically:
1945 // - They are defined in BB, and
1946 // - They have no side effects, and
1947 // - All of their uses are in CondBB.
1948 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1950 unsigned SpeculationCost = 0;
1951 Value *SpeculatedStoreValue = nullptr;
1952 StoreInst *SpeculatedStore = nullptr;
1953 for (BasicBlock::iterator BBI = ThenBB->begin(),
1954 BBE = std::prev(ThenBB->end());
1955 BBI != BBE; ++BBI) {
1956 Instruction *I = &*BBI;
1958 if (isa<DbgInfoIntrinsic>(I))
1961 // Only speculatively execute a single instruction (not counting the
1962 // terminator) for now.
1964 if (SpeculationCost > 1)
1967 // Don't hoist the instruction if it's unsafe or expensive.
1968 if (!isSafeToSpeculativelyExecute(I) &&
1969 !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1970 I, BB, ThenBB, EndBB))))
1972 if (!SpeculatedStoreValue &&
1973 ComputeSpeculationCost(I, TTI) >
1974 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
1977 // Store the store speculation candidate.
1978 if (SpeculatedStoreValue)
1979 SpeculatedStore = cast<StoreInst>(I);
1981 // Do not hoist the instruction if any of its operands are defined but not
1982 // used in BB. The transformation will prevent the operand from
1983 // being sunk into the use block.
1984 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
1985 Instruction *OpI = dyn_cast<Instruction>(*i);
1986 if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
1987 continue; // Not a candidate for sinking.
1989 ++SinkCandidateUseCounts[OpI];
1993 // Consider any sink candidates which are only used in CondBB as costs for
1994 // speculation. Note, while we iterate over a DenseMap here, we are summing
1995 // and so iteration order isn't significant.
1996 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator
1997 I = SinkCandidateUseCounts.begin(),
1998 E = SinkCandidateUseCounts.end();
2000 if (I->first->getNumUses() == I->second) {
2002 if (SpeculationCost > 1)
2006 // Check that the PHI nodes can be converted to selects.
2007 bool HaveRewritablePHIs = false;
2008 for (BasicBlock::iterator I = EndBB->begin();
2009 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2010 Value *OrigV = PN->getIncomingValueForBlock(BB);
2011 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
2013 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2014 // Skip PHIs which are trivial.
2018 // Don't convert to selects if we could remove undefined behavior instead.
2019 if (passingValueIsAlwaysUndefined(OrigV, PN) ||
2020 passingValueIsAlwaysUndefined(ThenV, PN))
2023 HaveRewritablePHIs = true;
2024 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2025 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2026 if (!OrigCE && !ThenCE)
2027 continue; // Known safe and cheap.
2029 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2030 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2032 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2033 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2035 2 * PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2036 if (OrigCost + ThenCost > MaxCost)
2039 // Account for the cost of an unfolded ConstantExpr which could end up
2040 // getting expanded into Instructions.
2041 // FIXME: This doesn't account for how many operations are combined in the
2042 // constant expression.
2044 if (SpeculationCost > 1)
2048 // If there are no PHIs to process, bail early. This helps ensure idempotence
2050 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2053 // If we get here, we can hoist the instruction and if-convert.
2054 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2056 // Insert a select of the value of the speculated store.
2057 if (SpeculatedStoreValue) {
2058 IRBuilder<NoFolder> Builder(BI);
2059 Value *TrueV = SpeculatedStore->getValueOperand();
2060 Value *FalseV = SpeculatedStoreValue;
2062 std::swap(TrueV, FalseV);
2063 Value *S = Builder.CreateSelect(
2064 BrCond, TrueV, FalseV, TrueV->getName() + "." + FalseV->getName(), BI);
2065 SpeculatedStore->setOperand(0, S);
2068 // Metadata can be dependent on the condition we are hoisting above.
2069 // Conservatively strip all metadata on the instruction.
2070 for (auto &I : *ThenBB)
2071 I.dropUnknownNonDebugMetadata();
2073 // Hoist the instructions.
2074 BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2075 ThenBB->begin(), std::prev(ThenBB->end()));
2077 // Insert selects and rewrite the PHI operands.
2078 IRBuilder<NoFolder> Builder(BI);
2079 for (BasicBlock::iterator I = EndBB->begin();
2080 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2081 unsigned OrigI = PN->getBasicBlockIndex(BB);
2082 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
2083 Value *OrigV = PN->getIncomingValue(OrigI);
2084 Value *ThenV = PN->getIncomingValue(ThenI);
2086 // Skip PHIs which are trivial.
2090 // Create a select whose true value is the speculatively executed value and
2091 // false value is the preexisting value. Swap them if the branch
2092 // destinations were inverted.
2093 Value *TrueV = ThenV, *FalseV = OrigV;
2095 std::swap(TrueV, FalseV);
2096 Value *V = Builder.CreateSelect(
2097 BrCond, TrueV, FalseV, TrueV->getName() + "." + FalseV->getName(), BI);
2098 PN->setIncomingValue(OrigI, V);
2099 PN->setIncomingValue(ThenI, V);
2106 /// Return true if we can thread a branch across this block.
2107 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
2108 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
2111 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2112 if (isa<DbgInfoIntrinsic>(BBI))
2115 return false; // Don't clone large BB's.
2118 // We can only support instructions that do not define values that are
2119 // live outside of the current basic block.
2120 for (User *U : BBI->users()) {
2121 Instruction *UI = cast<Instruction>(U);
2122 if (UI->getParent() != BB || isa<PHINode>(UI))
2126 // Looks ok, continue checking.
2132 /// If we have a conditional branch on a PHI node value that is defined in the
2133 /// same block as the branch and if any PHI entries are constants, thread edges
2134 /// corresponding to that entry to be branches to their ultimate destination.
2135 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL) {
2136 BasicBlock *BB = BI->getParent();
2137 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2138 // NOTE: we currently cannot transform this case if the PHI node is used
2139 // outside of the block.
2140 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2143 // Degenerate case of a single entry PHI.
2144 if (PN->getNumIncomingValues() == 1) {
2145 FoldSingleEntryPHINodes(PN->getParent());
2149 // Now we know that this block has multiple preds and two succs.
2150 if (!BlockIsSimpleEnoughToThreadThrough(BB))
2153 // Can't fold blocks that contain noduplicate or convergent calls.
2154 if (any_of(*BB, [](const Instruction &I) {
2155 const CallInst *CI = dyn_cast<CallInst>(&I);
2156 return CI && (CI->cannotDuplicate() || CI->isConvergent());
2160 // Okay, this is a simple enough basic block. See if any phi values are
2162 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2163 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
2164 if (!CB || !CB->getType()->isIntegerTy(1))
2167 // Okay, we now know that all edges from PredBB should be revectored to
2168 // branch to RealDest.
2169 BasicBlock *PredBB = PN->getIncomingBlock(i);
2170 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2173 continue; // Skip self loops.
2174 // Skip if the predecessor's terminator is an indirect branch.
2175 if (isa<IndirectBrInst>(PredBB->getTerminator()))
2178 // The dest block might have PHI nodes, other predecessors and other
2179 // difficult cases. Instead of being smart about this, just insert a new
2180 // block that jumps to the destination block, effectively splitting
2181 // the edge we are about to create.
2182 BasicBlock *EdgeBB =
2183 BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2184 RealDest->getParent(), RealDest);
2185 BranchInst::Create(RealDest, EdgeBB);
2187 // Update PHI nodes.
2188 AddPredecessorToBlock(RealDest, EdgeBB, BB);
2190 // BB may have instructions that are being threaded over. Clone these
2191 // instructions into EdgeBB. We know that there will be no uses of the
2192 // cloned instructions outside of EdgeBB.
2193 BasicBlock::iterator InsertPt = EdgeBB->begin();
2194 DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2195 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2196 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2197 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2200 // Clone the instruction.
2201 Instruction *N = BBI->clone();
2203 N->setName(BBI->getName() + ".c");
2205 // Update operands due to translation.
2206 for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2207 DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2208 if (PI != TranslateMap.end())
2212 // Check for trivial simplification.
2213 if (Value *V = SimplifyInstruction(N, DL)) {
2214 if (!BBI->use_empty())
2215 TranslateMap[&*BBI] = V;
2216 if (!N->mayHaveSideEffects()) {
2217 delete N; // Instruction folded away, don't need actual inst
2221 if (!BBI->use_empty())
2222 TranslateMap[&*BBI] = N;
2224 // Insert the new instruction into its new home.
2226 EdgeBB->getInstList().insert(InsertPt, N);
2229 // Loop over all of the edges from PredBB to BB, changing them to branch
2230 // to EdgeBB instead.
2231 TerminatorInst *PredBBTI = PredBB->getTerminator();
2232 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2233 if (PredBBTI->getSuccessor(i) == BB) {
2234 BB->removePredecessor(PredBB);
2235 PredBBTI->setSuccessor(i, EdgeBB);
2238 // Recurse, simplifying any other constants.
2239 return FoldCondBranchOnPHI(BI, DL) | true;
2245 /// Given a BB that starts with the specified two-entry PHI node,
2246 /// see if we can eliminate it.
2247 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI,
2248 const DataLayout &DL) {
2249 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2250 // statement", which has a very simple dominance structure. Basically, we
2251 // are trying to find the condition that is being branched on, which
2252 // subsequently causes this merge to happen. We really want control
2253 // dependence information for this check, but simplifycfg can't keep it up
2254 // to date, and this catches most of the cases we care about anyway.
2255 BasicBlock *BB = PN->getParent();
2256 BasicBlock *IfTrue, *IfFalse;
2257 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2259 // Don't bother if the branch will be constant folded trivially.
2260 isa<ConstantInt>(IfCond))
2263 // Okay, we found that we can merge this two-entry phi node into a select.
2264 // Doing so would require us to fold *all* two entry phi nodes in this block.
2265 // At some point this becomes non-profitable (particularly if the target
2266 // doesn't support cmov's). Only do this transformation if there are two or
2267 // fewer PHI nodes in this block.
2268 unsigned NumPhis = 0;
2269 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2273 // Loop over the PHI's seeing if we can promote them all to select
2274 // instructions. While we are at it, keep track of the instructions
2275 // that need to be moved to the dominating block.
2276 SmallPtrSet<Instruction *, 4> AggressiveInsts;
2277 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2278 MaxCostVal1 = PHINodeFoldingThreshold;
2279 MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2280 MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2282 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2283 PHINode *PN = cast<PHINode>(II++);
2284 if (Value *V = SimplifyInstruction(PN, DL)) {
2285 PN->replaceAllUsesWith(V);
2286 PN->eraseFromParent();
2290 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
2291 MaxCostVal0, TTI) ||
2292 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
2297 // If we folded the first phi, PN dangles at this point. Refresh it. If
2298 // we ran out of PHIs then we simplified them all.
2299 PN = dyn_cast<PHINode>(BB->begin());
2303 // Don't fold i1 branches on PHIs which contain binary operators. These can
2304 // often be turned into switches and other things.
2305 if (PN->getType()->isIntegerTy(1) &&
2306 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2307 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2308 isa<BinaryOperator>(IfCond)))
2311 // If all PHI nodes are promotable, check to make sure that all instructions
2312 // in the predecessor blocks can be promoted as well. If not, we won't be able
2313 // to get rid of the control flow, so it's not worth promoting to select
2315 BasicBlock *DomBlock = nullptr;
2316 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2317 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2318 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2321 DomBlock = *pred_begin(IfBlock1);
2322 for (BasicBlock::iterator I = IfBlock1->begin(); !isa<TerminatorInst>(I);
2324 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2325 // This is not an aggressive instruction that we can promote.
2326 // Because of this, we won't be able to get rid of the control flow, so
2327 // the xform is not worth it.
2332 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2335 DomBlock = *pred_begin(IfBlock2);
2336 for (BasicBlock::iterator I = IfBlock2->begin(); !isa<TerminatorInst>(I);
2338 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2339 // This is not an aggressive instruction that we can promote.
2340 // Because of this, we won't be able to get rid of the control flow, so
2341 // the xform is not worth it.
2346 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
2347 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
2349 // If we can still promote the PHI nodes after this gauntlet of tests,
2350 // do all of the PHI's now.
2351 Instruction *InsertPt = DomBlock->getTerminator();
2352 IRBuilder<NoFolder> Builder(InsertPt);
2354 // Move all 'aggressive' instructions, which are defined in the
2355 // conditional parts of the if's up to the dominating block.
2357 for (auto &I : *IfBlock1)
2358 I.dropUnknownNonDebugMetadata();
2359 DomBlock->getInstList().splice(InsertPt->getIterator(),
2360 IfBlock1->getInstList(), IfBlock1->begin(),
2361 IfBlock1->getTerminator()->getIterator());
2364 for (auto &I : *IfBlock2)
2365 I.dropUnknownNonDebugMetadata();
2366 DomBlock->getInstList().splice(InsertPt->getIterator(),
2367 IfBlock2->getInstList(), IfBlock2->begin(),
2368 IfBlock2->getTerminator()->getIterator());
2371 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2372 // Change the PHI node into a select instruction.
2373 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2374 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2376 Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2377 PN->replaceAllUsesWith(Sel);
2379 PN->eraseFromParent();
2382 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2383 // has been flattened. Change DomBlock to jump directly to our new block to
2384 // avoid other simplifycfg's kicking in on the diamond.
2385 TerminatorInst *OldTI = DomBlock->getTerminator();
2386 Builder.SetInsertPoint(OldTI);
2387 Builder.CreateBr(BB);
2388 OldTI->eraseFromParent();
2392 /// If we found a conditional branch that goes to two returning blocks,
2393 /// try to merge them together into one return,
2394 /// introducing a select if the return values disagree.
2395 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
2396 IRBuilder<> &Builder) {
2397 assert(BI->isConditional() && "Must be a conditional branch");
2398 BasicBlock *TrueSucc = BI->getSuccessor(0);
2399 BasicBlock *FalseSucc = BI->getSuccessor(1);
2400 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2401 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2403 // Check to ensure both blocks are empty (just a return) or optionally empty
2404 // with PHI nodes. If there are other instructions, merging would cause extra
2405 // computation on one path or the other.
2406 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2408 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2411 Builder.SetInsertPoint(BI);
2412 // Okay, we found a branch that is going to two return nodes. If
2413 // there is no return value for this function, just change the
2414 // branch into a return.
2415 if (FalseRet->getNumOperands() == 0) {
2416 TrueSucc->removePredecessor(BI->getParent());
2417 FalseSucc->removePredecessor(BI->getParent());
2418 Builder.CreateRetVoid();
2419 EraseTerminatorInstAndDCECond(BI);
2423 // Otherwise, figure out what the true and false return values are
2424 // so we can insert a new select instruction.
2425 Value *TrueValue = TrueRet->getReturnValue();
2426 Value *FalseValue = FalseRet->getReturnValue();
2428 // Unwrap any PHI nodes in the return blocks.
2429 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2430 if (TVPN->getParent() == TrueSucc)
2431 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2432 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2433 if (FVPN->getParent() == FalseSucc)
2434 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2436 // In order for this transformation to be safe, we must be able to
2437 // unconditionally execute both operands to the return. This is
2438 // normally the case, but we could have a potentially-trapping
2439 // constant expression that prevents this transformation from being
2441 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2444 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2448 // Okay, we collected all the mapped values and checked them for sanity, and
2449 // defined to really do this transformation. First, update the CFG.
2450 TrueSucc->removePredecessor(BI->getParent());
2451 FalseSucc->removePredecessor(BI->getParent());
2453 // Insert select instructions where needed.
2454 Value *BrCond = BI->getCondition();
2456 // Insert a select if the results differ.
2457 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2458 } else if (isa<UndefValue>(TrueValue)) {
2459 TrueValue = FalseValue;
2462 Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2467 !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2471 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2472 << "\n " << *BI << "NewRet = " << *RI
2473 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: " << *FalseSucc);
2475 EraseTerminatorInstAndDCECond(BI);
2480 /// Return true if the given instruction is available
2481 /// in its predecessor block. If yes, the instruction will be removed.
2482 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
2483 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2485 for (Instruction &I : *PB) {
2486 Instruction *PBI = &I;
2487 // Check whether Inst and PBI generate the same value.
2488 if (Inst->isIdenticalTo(PBI)) {
2489 Inst->replaceAllUsesWith(PBI);
2490 Inst->eraseFromParent();
2497 /// Return true if either PBI or BI has branch weight available, and store
2498 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2499 /// not have branch weight, use 1:1 as its weight.
2500 static bool extractPredSuccWeights(BranchInst *PBI, BranchInst *BI,
2501 uint64_t &PredTrueWeight,
2502 uint64_t &PredFalseWeight,
2503 uint64_t &SuccTrueWeight,
2504 uint64_t &SuccFalseWeight) {
2505 bool PredHasWeights =
2506 PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2507 bool SuccHasWeights =
2508 BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2509 if (PredHasWeights || SuccHasWeights) {
2510 if (!PredHasWeights)
2511 PredTrueWeight = PredFalseWeight = 1;
2512 if (!SuccHasWeights)
2513 SuccTrueWeight = SuccFalseWeight = 1;
2520 /// If this basic block is simple enough, and if a predecessor branches to us
2521 /// and one of our successors, fold the block into the predecessor and use
2522 /// logical operations to pick the right destination.
2523 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2524 BasicBlock *BB = BI->getParent();
2526 Instruction *Cond = nullptr;
2527 if (BI->isConditional())
2528 Cond = dyn_cast<Instruction>(BI->getCondition());
2530 // For unconditional branch, check for a simple CFG pattern, where
2531 // BB has a single predecessor and BB's successor is also its predecessor's
2532 // successor. If such pattern exisits, check for CSE between BB and its
2534 if (BasicBlock *PB = BB->getSinglePredecessor())
2535 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2536 if (PBI->isConditional() &&
2537 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2538 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2539 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2540 Instruction *Curr = &*I++;
2541 if (isa<CmpInst>(Curr)) {
2545 // Quit if we can't remove this instruction.
2546 if (!checkCSEInPredecessor(Curr, PB))
2555 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2556 Cond->getParent() != BB || !Cond->hasOneUse())
2559 // Make sure the instruction after the condition is the cond branch.
2560 BasicBlock::iterator CondIt = ++Cond->getIterator();
2562 // Ignore dbg intrinsics.
2563 while (isa<DbgInfoIntrinsic>(CondIt))
2569 // Only allow this transformation if computing the condition doesn't involve
2570 // too many instructions and these involved instructions can be executed
2571 // unconditionally. We denote all involved instructions except the condition
2572 // as "bonus instructions", and only allow this transformation when the
2573 // number of the bonus instructions does not exceed a certain threshold.
2574 unsigned NumBonusInsts = 0;
2575 for (auto I = BB->begin(); Cond != &*I; ++I) {
2576 // Ignore dbg intrinsics.
2577 if (isa<DbgInfoIntrinsic>(I))
2579 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2581 // I has only one use and can be executed unconditionally.
2582 Instruction *User = dyn_cast<Instruction>(I->user_back());
2583 if (User == nullptr || User->getParent() != BB)
2585 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2586 // to use any other instruction, User must be an instruction between next(I)
2589 // Early exits once we reach the limit.
2590 if (NumBonusInsts > BonusInstThreshold)
2594 // Cond is known to be a compare or binary operator. Check to make sure that
2595 // neither operand is a potentially-trapping constant expression.
2596 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2599 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2603 // Finally, don't infinitely unroll conditional loops.
2604 BasicBlock *TrueDest = BI->getSuccessor(0);
2605 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2606 if (TrueDest == BB || FalseDest == BB)
2609 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2610 BasicBlock *PredBlock = *PI;
2611 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2613 // Check that we have two conditional branches. If there is a PHI node in
2614 // the common successor, verify that the same value flows in from both
2616 SmallVector<PHINode *, 4> PHIs;
2617 if (!PBI || PBI->isUnconditional() ||
2618 (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2619 (!BI->isConditional() &&
2620 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2623 // Determine if the two branches share a common destination.
2624 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2625 bool InvertPredCond = false;
2627 if (BI->isConditional()) {
2628 if (PBI->getSuccessor(0) == TrueDest) {
2629 Opc = Instruction::Or;
2630 } else if (PBI->getSuccessor(1) == FalseDest) {
2631 Opc = Instruction::And;
2632 } else if (PBI->getSuccessor(0) == FalseDest) {
2633 Opc = Instruction::And;
2634 InvertPredCond = true;
2635 } else if (PBI->getSuccessor(1) == TrueDest) {
2636 Opc = Instruction::Or;
2637 InvertPredCond = true;
2642 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2646 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2647 IRBuilder<> Builder(PBI);
2649 // If we need to invert the condition in the pred block to match, do so now.
2650 if (InvertPredCond) {
2651 Value *NewCond = PBI->getCondition();
2653 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2654 CmpInst *CI = cast<CmpInst>(NewCond);
2655 CI->setPredicate(CI->getInversePredicate());
2658 Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2661 PBI->setCondition(NewCond);
2662 PBI->swapSuccessors();
2665 // If we have bonus instructions, clone them into the predecessor block.
2666 // Note that there may be multiple predecessor blocks, so we cannot move
2667 // bonus instructions to a predecessor block.
2668 ValueToValueMapTy VMap; // maps original values to cloned values
2669 // We already make sure Cond is the last instruction before BI. Therefore,
2670 // all instructions before Cond other than DbgInfoIntrinsic are bonus
2672 for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2673 if (isa<DbgInfoIntrinsic>(BonusInst))
2675 Instruction *NewBonusInst = BonusInst->clone();
2676 RemapInstruction(NewBonusInst, VMap,
2677 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2678 VMap[&*BonusInst] = NewBonusInst;
2680 // If we moved a load, we cannot any longer claim any knowledge about
2681 // its potential value. The previous information might have been valid
2682 // only given the branch precondition.
2683 // For an analogous reason, we must also drop all the metadata whose
2684 // semantics we don't understand.
2685 NewBonusInst->dropUnknownNonDebugMetadata();
2687 PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2688 NewBonusInst->takeName(&*BonusInst);
2689 BonusInst->setName(BonusInst->getName() + ".old");
2692 // Clone Cond into the predecessor basic block, and or/and the
2693 // two conditions together.
2694 Instruction *New = Cond->clone();
2695 RemapInstruction(New, VMap,
2696 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2697 PredBlock->getInstList().insert(PBI->getIterator(), New);
2698 New->takeName(Cond);
2699 Cond->setName(New->getName() + ".old");
2701 if (BI->isConditional()) {
2702 Instruction *NewCond = cast<Instruction>(
2703 Builder.CreateBinOp(Opc, PBI->getCondition(), New, "or.cond"));
2704 PBI->setCondition(NewCond);
2706 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2708 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2709 SuccTrueWeight, SuccFalseWeight);
2710 SmallVector<uint64_t, 8> NewWeights;
2712 if (PBI->getSuccessor(0) == BB) {
2714 // PBI: br i1 %x, BB, FalseDest
2715 // BI: br i1 %y, TrueDest, FalseDest
2716 // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2717 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2718 // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2719 // TrueWeight for PBI * FalseWeight for BI.
2720 // We assume that total weights of a BranchInst can fit into 32 bits.
2721 // Therefore, we will not have overflow using 64-bit arithmetic.
2722 NewWeights.push_back(PredFalseWeight *
2723 (SuccFalseWeight + SuccTrueWeight) +
2724 PredTrueWeight * SuccFalseWeight);
2726 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2727 PBI->setSuccessor(0, TrueDest);
2729 if (PBI->getSuccessor(1) == BB) {
2731 // PBI: br i1 %x, TrueDest, BB
2732 // BI: br i1 %y, TrueDest, FalseDest
2733 // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2734 // FalseWeight for PBI * TrueWeight for BI.
2735 NewWeights.push_back(PredTrueWeight *
2736 (SuccFalseWeight + SuccTrueWeight) +
2737 PredFalseWeight * SuccTrueWeight);
2738 // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2739 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2741 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2742 PBI->setSuccessor(1, FalseDest);
2744 if (NewWeights.size() == 2) {
2745 // Halve the weights if any of them cannot fit in an uint32_t
2746 FitWeights(NewWeights);
2748 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2751 LLVMContext::MD_prof,
2752 MDBuilder(BI->getContext()).createBranchWeights(MDWeights));
2754 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2756 // Update PHI nodes in the common successors.
2757 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2758 ConstantInt *PBI_C = cast<ConstantInt>(
2759 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2760 assert(PBI_C->getType()->isIntegerTy(1));
2761 Instruction *MergedCond = nullptr;
2762 if (PBI->getSuccessor(0) == TrueDest) {
2763 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2764 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2765 // is false: !PBI_Cond and BI_Value
2766 Instruction *NotCond = cast<Instruction>(
2767 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2768 MergedCond = cast<Instruction>(
2769 Builder.CreateBinOp(Instruction::And, NotCond, New, "and.cond"));
2771 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2772 Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2774 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2775 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2776 // is false: PBI_Cond and BI_Value
2777 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2778 Instruction::And, PBI->getCondition(), New, "and.cond"));
2779 if (PBI_C->isOne()) {
2780 Instruction *NotCond = cast<Instruction>(
2781 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2782 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2783 Instruction::Or, NotCond, MergedCond, "or.cond"));
2787 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2790 // Change PBI from Conditional to Unconditional.
2791 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2792 EraseTerminatorInstAndDCECond(PBI);
2796 // If BI was a loop latch, it may have had associated loop metadata.
2797 // We need to copy it to the new latch, that is, PBI.
2798 if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2799 PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2801 // TODO: If BB is reachable from all paths through PredBlock, then we
2802 // could replace PBI's branch probabilities with BI's.
2804 // Copy any debug value intrinsics into the end of PredBlock.
2805 for (Instruction &I : *BB)
2806 if (isa<DbgInfoIntrinsic>(I))
2807 I.clone()->insertBefore(PBI);
2814 // If there is only one store in BB1 and BB2, return it, otherwise return
2816 static StoreInst *findUniqueStoreInBlocks(BasicBlock *BB1, BasicBlock *BB2) {
2817 StoreInst *S = nullptr;
2818 for (auto *BB : {BB1, BB2}) {
2822 if (auto *SI = dyn_cast<StoreInst>(&I)) {
2824 // Multiple stores seen.
2833 static Value *ensureValueAvailableInSuccessor(Value *V, BasicBlock *BB,
2834 Value *AlternativeV = nullptr) {
2835 // PHI is going to be a PHI node that allows the value V that is defined in
2836 // BB to be referenced in BB's only successor.
2838 // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2839 // doesn't matter to us what the other operand is (it'll never get used). We
2840 // could just create a new PHI with an undef incoming value, but that could
2841 // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2842 // other PHI. So here we directly look for some PHI in BB's successor with V
2843 // as an incoming operand. If we find one, we use it, else we create a new
2846 // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2847 // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2848 // where OtherBB is the single other predecessor of BB's only successor.
2849 PHINode *PHI = nullptr;
2850 BasicBlock *Succ = BB->getSingleSuccessor();
2852 for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2853 if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2854 PHI = cast<PHINode>(I);
2858 assert(std::distance(pred_begin(Succ), pred_end(Succ)) == 2);
2859 auto PredI = pred_begin(Succ);
2860 BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2861 if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2868 // If V is not an instruction defined in BB, just return it.
2869 if (!AlternativeV &&
2870 (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2873 PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2874 PHI->addIncoming(V, BB);
2875 for (BasicBlock *PredBB : predecessors(Succ))
2878 AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2882 static bool mergeConditionalStoreToAddress(BasicBlock *PTB, BasicBlock *PFB,
2883 BasicBlock *QTB, BasicBlock *QFB,
2884 BasicBlock *PostBB, Value *Address,
2885 bool InvertPCond, bool InvertQCond) {
2886 auto IsaBitcastOfPointerType = [](const Instruction &I) {
2887 return Operator::getOpcode(&I) == Instruction::BitCast &&
2888 I.getType()->isPointerTy();
2891 // If we're not in aggressive mode, we only optimize if we have some
2892 // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2893 auto IsWorthwhile = [&](BasicBlock *BB) {
2896 // Heuristic: if the block can be if-converted/phi-folded and the
2897 // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2898 // thread this store.
2900 for (auto &I : *BB) {
2901 // Cheap instructions viable for folding.
2902 if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2905 // Free instructions.
2906 else if (isa<TerminatorInst>(I) || isa<DbgInfoIntrinsic>(I) ||
2907 IsaBitcastOfPointerType(I))
2912 return N <= PHINodeFoldingThreshold;
2915 if (!MergeCondStoresAggressively &&
2916 (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2917 !IsWorthwhile(QFB)))
2920 // For every pointer, there must be exactly two stores, one coming from
2921 // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2922 // store (to any address) in PTB,PFB or QTB,QFB.
2923 // FIXME: We could relax this restriction with a bit more work and performance
2925 StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2926 StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2927 if (!PStore || !QStore)
2930 // Now check the stores are compatible.
2931 if (!QStore->isUnordered() || !PStore->isUnordered())
2934 // Check that sinking the store won't cause program behavior changes. Sinking
2935 // the store out of the Q blocks won't change any behavior as we're sinking
2936 // from a block to its unconditional successor. But we're moving a store from
2937 // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2938 // So we need to check that there are no aliasing loads or stores in
2939 // QBI, QTB and QFB. We also need to check there are no conflicting memory
2940 // operations between PStore and the end of its parent block.
2942 // The ideal way to do this is to query AliasAnalysis, but we don't
2943 // preserve AA currently so that is dangerous. Be super safe and just
2944 // check there are no other memory operations at all.
2945 for (auto &I : *QFB->getSinglePredecessor())
2946 if (I.mayReadOrWriteMemory())
2948 for (auto &I : *QFB)
2949 if (&I != QStore && I.mayReadOrWriteMemory())
2952 for (auto &I : *QTB)
2953 if (&I != QStore && I.mayReadOrWriteMemory())
2955 for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2957 if (&*I != PStore && I->mayReadOrWriteMemory())
2960 // OK, we're going to sink the stores to PostBB. The store has to be
2961 // conditional though, so first create the predicate.
2962 Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
2964 Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
2967 Value *PPHI = ensureValueAvailableInSuccessor(PStore->getValueOperand(),
2968 PStore->getParent());
2969 Value *QPHI = ensureValueAvailableInSuccessor(QStore->getValueOperand(),
2970 QStore->getParent(), PPHI);
2972 IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
2974 Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
2975 Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
2978 PPred = QB.CreateNot(PPred);
2980 QPred = QB.CreateNot(QPred);
2981 Value *CombinedPred = QB.CreateOr(PPred, QPred);
2984 SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
2985 QB.SetInsertPoint(T);
2986 StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
2988 PStore->getAAMetadata(AAMD, /*Merge=*/false);
2989 PStore->getAAMetadata(AAMD, /*Merge=*/true);
2990 SI->setAAMetadata(AAMD);
2992 QStore->eraseFromParent();
2993 PStore->eraseFromParent();
2998 static bool mergeConditionalStores(BranchInst *PBI, BranchInst *QBI) {
2999 // The intention here is to find diamonds or triangles (see below) where each
3000 // conditional block contains a store to the same address. Both of these
3001 // stores are conditional, so they can't be unconditionally sunk. But it may
3002 // be profitable to speculatively sink the stores into one merged store at the
3003 // end, and predicate the merged store on the union of the two conditions of
3006 // This can reduce the number of stores executed if both of the conditions are
3007 // true, and can allow the blocks to become small enough to be if-converted.
3008 // This optimization will also chain, so that ladders of test-and-set
3009 // sequences can be if-converted away.
3011 // We only deal with simple diamonds or triangles:
3013 // PBI or PBI or a combination of the two
3023 // We model triangles as a type of diamond with a nullptr "true" block.
3024 // Triangles are canonicalized so that the fallthrough edge is represented by
3025 // a true condition, as in the diagram above.
3027 BasicBlock *PTB = PBI->getSuccessor(0);
3028 BasicBlock *PFB = PBI->getSuccessor(1);
3029 BasicBlock *QTB = QBI->getSuccessor(0);
3030 BasicBlock *QFB = QBI->getSuccessor(1);
3031 BasicBlock *PostBB = QFB->getSingleSuccessor();
3033 bool InvertPCond = false, InvertQCond = false;
3034 // Canonicalize fallthroughs to the true branches.
3035 if (PFB == QBI->getParent()) {
3036 std::swap(PFB, PTB);
3039 if (QFB == PostBB) {
3040 std::swap(QFB, QTB);
3044 // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3045 // and QFB may not. Model fallthroughs as a nullptr block.
3046 if (PTB == QBI->getParent())
3051 // Legality bailouts. We must have at least the non-fallthrough blocks and
3052 // the post-dominating block, and the non-fallthroughs must only have one
3054 auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3055 return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3058 !HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3059 !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3061 if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3062 (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3064 if (PostBB->getNumUses() != 2 || QBI->getParent()->getNumUses() != 2)
3067 // OK, this is a sequence of two diamonds or triangles.
3068 // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3069 SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3070 for (auto *BB : {PTB, PFB}) {
3074 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3075 PStoreAddresses.insert(SI->getPointerOperand());
3077 for (auto *BB : {QTB, QFB}) {
3081 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3082 QStoreAddresses.insert(SI->getPointerOperand());
3085 set_intersect(PStoreAddresses, QStoreAddresses);
3086 // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3087 // clear what it contains.
3088 auto &CommonAddresses = PStoreAddresses;
3090 bool Changed = false;
3091 for (auto *Address : CommonAddresses)
3092 Changed |= mergeConditionalStoreToAddress(
3093 PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond);
3097 /// If we have a conditional branch as a predecessor of another block,
3098 /// this function tries to simplify it. We know
3099 /// that PBI and BI are both conditional branches, and BI is in one of the
3100 /// successor blocks of PBI - PBI branches to BI.
3101 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
3102 const DataLayout &DL) {
3103 assert(PBI->isConditional() && BI->isConditional());
3104 BasicBlock *BB = BI->getParent();
3106 // If this block ends with a branch instruction, and if there is a
3107 // predecessor that ends on a branch of the same condition, make
3108 // this conditional branch redundant.
3109 if (PBI->getCondition() == BI->getCondition() &&
3110 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3111 // Okay, the outcome of this conditional branch is statically
3112 // knowable. If this block had a single pred, handle specially.
3113 if (BB->getSinglePredecessor()) {
3114 // Turn this into a branch on constant.
3115 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3117 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3118 return true; // Nuke the branch on constant.
3121 // Otherwise, if there are multiple predecessors, insert a PHI that merges
3122 // in the constant and simplify the block result. Subsequent passes of
3123 // simplifycfg will thread the block.
3124 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
3125 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3126 PHINode *NewPN = PHINode::Create(
3127 Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3128 BI->getCondition()->getName() + ".pr", &BB->front());
3129 // Okay, we're going to insert the PHI node. Since PBI is not the only
3130 // predecessor, compute the PHI'd conditional value for all of the preds.
3131 // Any predecessor where the condition is not computable we keep symbolic.
3132 for (pred_iterator PI = PB; PI != PE; ++PI) {
3133 BasicBlock *P = *PI;
3134 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3135 PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3136 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3137 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3139 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3142 NewPN->addIncoming(BI->getCondition(), P);
3146 BI->setCondition(NewPN);
3151 if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3155 // If both branches are conditional and both contain stores to the same
3156 // address, remove the stores from the conditionals and create a conditional
3157 // merged store at the end.
3158 if (MergeCondStores && mergeConditionalStores(PBI, BI))
3161 // If this is a conditional branch in an empty block, and if any
3162 // predecessors are a conditional branch to one of our destinations,
3163 // fold the conditions into logical ops and one cond br.
3164 BasicBlock::iterator BBI = BB->begin();
3165 // Ignore dbg intrinsics.
3166 while (isa<DbgInfoIntrinsic>(BBI))
3172 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3175 } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3178 } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3181 } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3188 // Check to make sure that the other destination of this branch
3189 // isn't BB itself. If so, this is an infinite loop that will
3190 // keep getting unwound.
3191 if (PBI->getSuccessor(PBIOp) == BB)
3194 // Do not perform this transformation if it would require
3195 // insertion of a large number of select instructions. For targets
3196 // without predication/cmovs, this is a big pessimization.
3198 // Also do not perform this transformation if any phi node in the common
3199 // destination block can trap when reached by BB or PBB (PR17073). In that
3200 // case, it would be unsafe to hoist the operation into a select instruction.
3202 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3203 unsigned NumPhis = 0;
3204 for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3206 if (NumPhis > 2) // Disable this xform.
3209 PHINode *PN = cast<PHINode>(II);
3210 Value *BIV = PN->getIncomingValueForBlock(BB);
3211 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3215 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3216 Value *PBIV = PN->getIncomingValue(PBBIdx);
3217 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3222 // Finally, if everything is ok, fold the branches to logical ops.
3223 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3225 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3226 << "AND: " << *BI->getParent());
3228 // If OtherDest *is* BB, then BB is a basic block with a single conditional
3229 // branch in it, where one edge (OtherDest) goes back to itself but the other
3230 // exits. We don't *know* that the program avoids the infinite loop
3231 // (even though that seems likely). If we do this xform naively, we'll end up
3232 // recursively unpeeling the loop. Since we know that (after the xform is
3233 // done) that the block *is* infinite if reached, we just make it an obviously
3234 // infinite loop with no cond branch.
3235 if (OtherDest == BB) {
3236 // Insert it at the end of the function, because it's either code,
3237 // or it won't matter if it's hot. :)
3238 BasicBlock *InfLoopBlock =
3239 BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3240 BranchInst::Create(InfLoopBlock, InfLoopBlock);
3241 OtherDest = InfLoopBlock;
3244 DEBUG(dbgs() << *PBI->getParent()->getParent());
3246 // BI may have other predecessors. Because of this, we leave
3247 // it alone, but modify PBI.
3249 // Make sure we get to CommonDest on True&True directions.
3250 Value *PBICond = PBI->getCondition();
3251 IRBuilder<NoFolder> Builder(PBI);
3253 PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3255 Value *BICond = BI->getCondition();
3257 BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3259 // Merge the conditions.
3260 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3262 // Modify PBI to branch on the new condition to the new dests.
3263 PBI->setCondition(Cond);
3264 PBI->setSuccessor(0, CommonDest);
3265 PBI->setSuccessor(1, OtherDest);
3267 // Update branch weight for PBI.
3268 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3269 uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3271 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3272 SuccTrueWeight, SuccFalseWeight);
3274 PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3275 PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3276 SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3277 SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3278 // The weight to CommonDest should be PredCommon * SuccTotal +
3279 // PredOther * SuccCommon.
3280 // The weight to OtherDest should be PredOther * SuccOther.
3281 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3282 PredOther * SuccCommon,
3283 PredOther * SuccOther};
3284 // Halve the weights if any of them cannot fit in an uint32_t
3285 FitWeights(NewWeights);
3287 PBI->setMetadata(LLVMContext::MD_prof,
3288 MDBuilder(BI->getContext())
3289 .createBranchWeights(NewWeights[0], NewWeights[1]));
3292 // OtherDest may have phi nodes. If so, add an entry from PBI's
3293 // block that are identical to the entries for BI's block.
3294 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3296 // We know that the CommonDest already had an edge from PBI to
3297 // it. If it has PHIs though, the PHIs may have different
3298 // entries for BB and PBI's BB. If so, insert a select to make
3301 for (BasicBlock::iterator II = CommonDest->begin();
3302 (PN = dyn_cast<PHINode>(II)); ++II) {
3303 Value *BIV = PN->getIncomingValueForBlock(BB);
3304 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3305 Value *PBIV = PN->getIncomingValue(PBBIdx);
3307 // Insert a select in PBI to pick the right value.
3308 SelectInst *NV = cast<SelectInst>(
3309 Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3310 PN->setIncomingValue(PBBIdx, NV);
3311 // Although the select has the same condition as PBI, the original branch
3312 // weights for PBI do not apply to the new select because the select's
3313 // 'logical' edges are incoming edges of the phi that is eliminated, not
3314 // the outgoing edges of PBI.
3316 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3317 uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3318 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3319 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3320 // The weight to PredCommonDest should be PredCommon * SuccTotal.
3321 // The weight to PredOtherDest should be PredOther * SuccCommon.
3322 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3323 PredOther * SuccCommon};
3325 FitWeights(NewWeights);
3327 NV->setMetadata(LLVMContext::MD_prof,
3328 MDBuilder(BI->getContext())
3329 .createBranchWeights(NewWeights[0], NewWeights[1]));
3334 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3335 DEBUG(dbgs() << *PBI->getParent()->getParent());
3337 // This basic block is probably dead. We know it has at least
3338 // one fewer predecessor.
3342 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3343 // true or to FalseBB if Cond is false.
3344 // Takes care of updating the successors and removing the old terminator.
3345 // Also makes sure not to introduce new successors by assuming that edges to
3346 // non-successor TrueBBs and FalseBBs aren't reachable.
3347 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
3348 BasicBlock *TrueBB, BasicBlock *FalseBB,
3349 uint32_t TrueWeight,
3350 uint32_t FalseWeight) {
3351 // Remove any superfluous successor edges from the CFG.
3352 // First, figure out which successors to preserve.
3353 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3355 BasicBlock *KeepEdge1 = TrueBB;
3356 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3358 // Then remove the rest.
3359 for (BasicBlock *Succ : OldTerm->successors()) {
3360 // Make sure only to keep exactly one copy of each edge.
3361 if (Succ == KeepEdge1)
3362 KeepEdge1 = nullptr;
3363 else if (Succ == KeepEdge2)
3364 KeepEdge2 = nullptr;
3366 Succ->removePredecessor(OldTerm->getParent(),
3367 /*DontDeleteUselessPHIs=*/true);
3370 IRBuilder<> Builder(OldTerm);
3371 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3373 // Insert an appropriate new terminator.
3374 if (!KeepEdge1 && !KeepEdge2) {
3375 if (TrueBB == FalseBB)
3376 // We were only looking for one successor, and it was present.
3377 // Create an unconditional branch to it.
3378 Builder.CreateBr(TrueBB);
3380 // We found both of the successors we were looking for.
3381 // Create a conditional branch sharing the condition of the select.
3382 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3383 if (TrueWeight != FalseWeight)
3384 NewBI->setMetadata(LLVMContext::MD_prof,
3385 MDBuilder(OldTerm->getContext())
3386 .createBranchWeights(TrueWeight, FalseWeight));
3388 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3389 // Neither of the selected blocks were successors, so this
3390 // terminator must be unreachable.
3391 new UnreachableInst(OldTerm->getContext(), OldTerm);
3393 // One of the selected values was a successor, but the other wasn't.
3394 // Insert an unconditional branch to the one that was found;
3395 // the edge to the one that wasn't must be unreachable.
3397 // Only TrueBB was found.
3398 Builder.CreateBr(TrueBB);
3400 // Only FalseBB was found.
3401 Builder.CreateBr(FalseBB);
3404 EraseTerminatorInstAndDCECond(OldTerm);
3409 // (switch (select cond, X, Y)) on constant X, Y
3410 // with a branch - conditional if X and Y lead to distinct BBs,
3411 // unconditional otherwise.
3412 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
3413 // Check for constant integer values in the select.
3414 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3415 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3416 if (!TrueVal || !FalseVal)
3419 // Find the relevant condition and destinations.
3420 Value *Condition = Select->getCondition();
3421 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
3422 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
3424 // Get weight for TrueBB and FalseBB.
3425 uint32_t TrueWeight = 0, FalseWeight = 0;
3426 SmallVector<uint64_t, 8> Weights;
3427 bool HasWeights = HasBranchWeights(SI);
3429 GetBranchWeights(SI, Weights);
3430 if (Weights.size() == 1 + SI->getNumCases()) {
3432 (uint32_t)Weights[SI->findCaseValue(TrueVal).getSuccessorIndex()];
3434 (uint32_t)Weights[SI->findCaseValue(FalseVal).getSuccessorIndex()];
3438 // Perform the actual simplification.
3439 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3444 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3445 // blockaddress(@fn, BlockB)))
3447 // (br cond, BlockA, BlockB).
3448 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
3449 // Check that both operands of the select are block addresses.
3450 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
3451 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
3455 // Extract the actual blocks.
3456 BasicBlock *TrueBB = TBA->getBasicBlock();
3457 BasicBlock *FalseBB = FBA->getBasicBlock();
3459 // Perform the actual simplification.
3460 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3464 /// This is called when we find an icmp instruction
3465 /// (a seteq/setne with a constant) as the only instruction in a
3466 /// block that ends with an uncond branch. We are looking for a very specific
3467 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3468 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3469 /// default value goes to an uncond block with a seteq in it, we get something
3472 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3474 /// %tmp = icmp eq i8 %A, 92
3477 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3479 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3480 /// the PHI, merging the third icmp into the switch.
3481 static bool TryToSimplifyUncondBranchWithICmpInIt(
3482 ICmpInst *ICI, IRBuilder<> &Builder, const DataLayout &DL,
3483 const TargetTransformInfo &TTI, unsigned BonusInstThreshold,
3484 AssumptionCache *AC) {
3485 BasicBlock *BB = ICI->getParent();
3487 // If the block has any PHIs in it or the icmp has multiple uses, it is too
3489 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3492 Value *V = ICI->getOperand(0);
3493 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3495 // The pattern we're looking for is where our only predecessor is a switch on
3496 // 'V' and this block is the default case for the switch. In this case we can
3497 // fold the compared value into the switch to simplify things.
3498 BasicBlock *Pred = BB->getSinglePredecessor();
3499 if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3502 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3503 if (SI->getCondition() != V)
3506 // If BB is reachable on a non-default case, then we simply know the value of
3507 // V in this block. Substitute it and constant fold the icmp instruction
3509 if (SI->getDefaultDest() != BB) {
3510 ConstantInt *VVal = SI->findCaseDest(BB);
3511 assert(VVal && "Should have a unique destination value");
3512 ICI->setOperand(0, VVal);
3514 if (Value *V = SimplifyInstruction(ICI, DL)) {
3515 ICI->replaceAllUsesWith(V);
3516 ICI->eraseFromParent();
3518 // BB is now empty, so it is likely to simplify away.
3519 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3522 // Ok, the block is reachable from the default dest. If the constant we're
3523 // comparing exists in one of the other edges, then we can constant fold ICI
3525 if (SI->findCaseValue(Cst) != SI->case_default()) {
3527 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3528 V = ConstantInt::getFalse(BB->getContext());
3530 V = ConstantInt::getTrue(BB->getContext());
3532 ICI->replaceAllUsesWith(V);
3533 ICI->eraseFromParent();
3534 // BB is now empty, so it is likely to simplify away.
3535 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3538 // The use of the icmp has to be in the 'end' block, by the only PHI node in
3540 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3541 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3542 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3543 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3546 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3548 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3549 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3551 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3552 std::swap(DefaultCst, NewCst);
3554 // Replace ICI (which is used by the PHI for the default value) with true or
3555 // false depending on if it is EQ or NE.
3556 ICI->replaceAllUsesWith(DefaultCst);
3557 ICI->eraseFromParent();
3559 // Okay, the switch goes to this block on a default value. Add an edge from
3560 // the switch to the merge point on the compared value.
3562 BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3563 SmallVector<uint64_t, 8> Weights;
3564 bool HasWeights = HasBranchWeights(SI);
3566 GetBranchWeights(SI, Weights);
3567 if (Weights.size() == 1 + SI->getNumCases()) {
3568 // Split weight for default case to case for "Cst".
3569 Weights[0] = (Weights[0] + 1) >> 1;
3570 Weights.push_back(Weights[0]);
3572 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3574 LLVMContext::MD_prof,
3575 MDBuilder(SI->getContext()).createBranchWeights(MDWeights));
3578 SI->addCase(Cst, NewBB);
3580 // NewBB branches to the phi block, add the uncond branch and the phi entry.
3581 Builder.SetInsertPoint(NewBB);
3582 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3583 Builder.CreateBr(SuccBlock);
3584 PHIUse->addIncoming(NewCst, NewBB);
3588 /// The specified branch is a conditional branch.
3589 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3590 /// fold it into a switch instruction if so.
3591 static bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
3592 const DataLayout &DL) {
3593 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3597 // Change br (X == 0 | X == 1), T, F into a switch instruction.
3598 // If this is a bunch of seteq's or'd together, or if it's a bunch of
3599 // 'setne's and'ed together, collect them.
3601 // Try to gather values from a chain of and/or to be turned into a switch
3602 ConstantComparesGatherer ConstantCompare(Cond, DL);
3603 // Unpack the result
3604 SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3605 Value *CompVal = ConstantCompare.CompValue;
3606 unsigned UsedICmps = ConstantCompare.UsedICmps;
3607 Value *ExtraCase = ConstantCompare.Extra;
3609 // If we didn't have a multiply compared value, fail.
3613 // Avoid turning single icmps into a switch.
3617 bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3619 // There might be duplicate constants in the list, which the switch
3620 // instruction can't handle, remove them now.
3621 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3622 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3624 // If Extra was used, we require at least two switch values to do the
3625 // transformation. A switch with one value is just a conditional branch.
3626 if (ExtraCase && Values.size() < 2)
3629 // TODO: Preserve branch weight metadata, similarly to how
3630 // FoldValueComparisonIntoPredecessors preserves it.
3632 // Figure out which block is which destination.
3633 BasicBlock *DefaultBB = BI->getSuccessor(1);
3634 BasicBlock *EdgeBB = BI->getSuccessor(0);
3636 std::swap(DefaultBB, EdgeBB);
3638 BasicBlock *BB = BI->getParent();
3640 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3641 << " cases into SWITCH. BB is:\n"
3644 // If there are any extra values that couldn't be folded into the switch
3645 // then we evaluate them with an explicit branch first. Split the block
3646 // right before the condbr to handle it.
3649 BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3650 // Remove the uncond branch added to the old block.
3651 TerminatorInst *OldTI = BB->getTerminator();
3652 Builder.SetInsertPoint(OldTI);
3655 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3657 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3659 OldTI->eraseFromParent();
3661 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3662 // for the edge we just added.
3663 AddPredecessorToBlock(EdgeBB, BB, NewBB);
3665 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3666 << "\nEXTRABB = " << *BB);
3670 Builder.SetInsertPoint(BI);
3671 // Convert pointer to int before we switch.
3672 if (CompVal->getType()->isPointerTy()) {
3673 CompVal = Builder.CreatePtrToInt(
3674 CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3677 // Create the new switch instruction now.
3678 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3680 // Add all of the 'cases' to the switch instruction.
3681 for (unsigned i = 0, e = Values.size(); i != e; ++i)
3682 New->addCase(Values[i], EdgeBB);
3684 // We added edges from PI to the EdgeBB. As such, if there were any
3685 // PHI nodes in EdgeBB, they need entries to be added corresponding to
3686 // the number of edges added.
3687 for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3688 PHINode *PN = cast<PHINode>(BBI);
3689 Value *InVal = PN->getIncomingValueForBlock(BB);
3690 for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3691 PN->addIncoming(InVal, BB);
3694 // Erase the old branch instruction.
3695 EraseTerminatorInstAndDCECond(BI);
3697 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3701 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3702 if (isa<PHINode>(RI->getValue()))
3703 return SimplifyCommonResume(RI);
3704 else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3705 RI->getValue() == RI->getParent()->getFirstNonPHI())
3706 // The resume must unwind the exception that caused control to branch here.
3707 return SimplifySingleResume(RI);
3712 // Simplify resume that is shared by several landing pads (phi of landing pad).
3713 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3714 BasicBlock *BB = RI->getParent();
3716 // Check that there are no other instructions except for debug intrinsics
3717 // between the phi of landing pads (RI->getValue()) and resume instruction.
3718 BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3719 E = RI->getIterator();
3721 if (!isa<DbgInfoIntrinsic>(I))
3724 SmallSet<BasicBlock *, 4> TrivialUnwindBlocks;
3725 auto *PhiLPInst = cast<PHINode>(RI->getValue());
3727 // Check incoming blocks to see if any of them are trivial.
3728 for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3730 auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3731 auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3733 // If the block has other successors, we can not delete it because
3734 // it has other dependents.
3735 if (IncomingBB->getUniqueSuccessor() != BB)
3738 auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3739 // Not the landing pad that caused the control to branch here.
3740 if (IncomingValue != LandingPad)
3743 bool isTrivial = true;
3745 I = IncomingBB->getFirstNonPHI()->getIterator();
3746 E = IncomingBB->getTerminator()->getIterator();
3748 if (!isa<DbgInfoIntrinsic>(I)) {
3754 TrivialUnwindBlocks.insert(IncomingBB);
3757 // If no trivial unwind blocks, don't do any simplifications.
3758 if (TrivialUnwindBlocks.empty())
3761 // Turn all invokes that unwind here into calls.
3762 for (auto *TrivialBB : TrivialUnwindBlocks) {
3763 // Blocks that will be simplified should be removed from the phi node.
3764 // Note there could be multiple edges to the resume block, and we need
3765 // to remove them all.
3766 while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3767 BB->removePredecessor(TrivialBB, true);
3769 for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3771 BasicBlock *Pred = *PI++;
3772 removeUnwindEdge(Pred);
3775 // In each SimplifyCFG run, only the current processed block can be erased.
3776 // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3777 // of erasing TrivialBB, we only remove the branch to the common resume
3778 // block so that we can later erase the resume block since it has no
3780 TrivialBB->getTerminator()->eraseFromParent();
3781 new UnreachableInst(RI->getContext(), TrivialBB);
3784 // Delete the resume block if all its predecessors have been removed.
3786 BB->eraseFromParent();
3788 return !TrivialUnwindBlocks.empty();
3791 // Simplify resume that is only used by a single (non-phi) landing pad.
3792 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3793 BasicBlock *BB = RI->getParent();
3794 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
3795 assert(RI->getValue() == LPInst &&
3796 "Resume must unwind the exception that caused control to here");
3798 // Check that there are no other instructions except for debug intrinsics.
3799 BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3801 if (!isa<DbgInfoIntrinsic>(I))
3804 // Turn all invokes that unwind here into calls and delete the basic block.
3805 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3806 BasicBlock *Pred = *PI++;
3807 removeUnwindEdge(Pred);
3810 // The landingpad is now unreachable. Zap it.
3811 BB->eraseFromParent();
3813 LoopHeaders->erase(BB);
3817 static bool removeEmptyCleanup(CleanupReturnInst *RI) {
3818 // If this is a trivial cleanup pad that executes no instructions, it can be
3819 // eliminated. If the cleanup pad continues to the caller, any predecessor
3820 // that is an EH pad will be updated to continue to the caller and any
3821 // predecessor that terminates with an invoke instruction will have its invoke
3822 // instruction converted to a call instruction. If the cleanup pad being
3823 // simplified does not continue to the caller, each predecessor will be
3824 // updated to continue to the unwind destination of the cleanup pad being
3826 BasicBlock *BB = RI->getParent();
3827 CleanupPadInst *CPInst = RI->getCleanupPad();
3828 if (CPInst->getParent() != BB)
3829 // This isn't an empty cleanup.
3832 // We cannot kill the pad if it has multiple uses. This typically arises
3833 // from unreachable basic blocks.
3834 if (!CPInst->hasOneUse())
3837 // Check that there are no other instructions except for benign intrinsics.
3838 BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3840 auto *II = dyn_cast<IntrinsicInst>(I);
3844 Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3845 switch (IntrinsicID) {
3846 case Intrinsic::dbg_declare:
3847 case Intrinsic::dbg_value:
3848 case Intrinsic::lifetime_end:
3855 // If the cleanup return we are simplifying unwinds to the caller, this will
3856 // set UnwindDest to nullptr.
3857 BasicBlock *UnwindDest = RI->getUnwindDest();
3858 Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3860 // We're about to remove BB from the control flow. Before we do, sink any
3861 // PHINodes into the unwind destination. Doing this before changing the
3862 // control flow avoids some potentially slow checks, since we can currently
3863 // be certain that UnwindDest and BB have no common predecessors (since they
3864 // are both EH pads).
3866 // First, go through the PHI nodes in UnwindDest and update any nodes that
3867 // reference the block we are removing
3868 for (BasicBlock::iterator I = UnwindDest->begin(),
3869 IE = DestEHPad->getIterator();
3871 PHINode *DestPN = cast<PHINode>(I);
3873 int Idx = DestPN->getBasicBlockIndex(BB);
3874 // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3876 // This PHI node has an incoming value that corresponds to a control
3877 // path through the cleanup pad we are removing. If the incoming
3878 // value is in the cleanup pad, it must be a PHINode (because we
3879 // verified above that the block is otherwise empty). Otherwise, the
3880 // value is either a constant or a value that dominates the cleanup
3881 // pad being removed.
3883 // Because BB and UnwindDest are both EH pads, all of their
3884 // predecessors must unwind to these blocks, and since no instruction
3885 // can have multiple unwind destinations, there will be no overlap in
3886 // incoming blocks between SrcPN and DestPN.
3887 Value *SrcVal = DestPN->getIncomingValue(Idx);
3888 PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3890 // Remove the entry for the block we are deleting.
3891 DestPN->removeIncomingValue(Idx, false);
3893 if (SrcPN && SrcPN->getParent() == BB) {
3894 // If the incoming value was a PHI node in the cleanup pad we are
3895 // removing, we need to merge that PHI node's incoming values into
3897 for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3898 SrcIdx != SrcE; ++SrcIdx) {
3899 DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3900 SrcPN->getIncomingBlock(SrcIdx));
3903 // Otherwise, the incoming value came from above BB and
3904 // so we can just reuse it. We must associate all of BB's
3905 // predecessors with this value.
3906 for (auto *pred : predecessors(BB)) {
3907 DestPN->addIncoming(SrcVal, pred);
3912 // Sink any remaining PHI nodes directly into UnwindDest.
3913 Instruction *InsertPt = DestEHPad;
3914 for (BasicBlock::iterator I = BB->begin(),
3915 IE = BB->getFirstNonPHI()->getIterator();
3917 // The iterator must be incremented here because the instructions are
3918 // being moved to another block.
3919 PHINode *PN = cast<PHINode>(I++);
3920 if (PN->use_empty())
3921 // If the PHI node has no uses, just leave it. It will be erased
3922 // when we erase BB below.
3925 // Otherwise, sink this PHI node into UnwindDest.
3926 // Any predecessors to UnwindDest which are not already represented
3927 // must be back edges which inherit the value from the path through
3928 // BB. In this case, the PHI value must reference itself.
3929 for (auto *pred : predecessors(UnwindDest))
3931 PN->addIncoming(PN, pred);
3932 PN->moveBefore(InsertPt);
3936 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3937 // The iterator must be updated here because we are removing this pred.
3938 BasicBlock *PredBB = *PI++;
3939 if (UnwindDest == nullptr) {
3940 removeUnwindEdge(PredBB);
3942 TerminatorInst *TI = PredBB->getTerminator();
3943 TI->replaceUsesOfWith(BB, UnwindDest);
3947 // The cleanup pad is now unreachable. Zap it.
3948 BB->eraseFromParent();
3952 // Try to merge two cleanuppads together.
3953 static bool mergeCleanupPad(CleanupReturnInst *RI) {
3954 // Skip any cleanuprets which unwind to caller, there is nothing to merge
3956 BasicBlock *UnwindDest = RI->getUnwindDest();
3960 // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
3961 // be safe to merge without code duplication.
3962 if (UnwindDest->getSinglePredecessor() != RI->getParent())
3965 // Verify that our cleanuppad's unwind destination is another cleanuppad.
3966 auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
3967 if (!SuccessorCleanupPad)
3970 CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
3971 // Replace any uses of the successor cleanupad with the predecessor pad
3972 // The only cleanuppad uses should be this cleanupret, it's cleanupret and
3973 // funclet bundle operands.
3974 SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
3975 // Remove the old cleanuppad.
3976 SuccessorCleanupPad->eraseFromParent();
3977 // Now, we simply replace the cleanupret with a branch to the unwind
3979 BranchInst::Create(UnwindDest, RI->getParent());
3980 RI->eraseFromParent();
3985 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
3986 // It is possible to transiantly have an undef cleanuppad operand because we
3987 // have deleted some, but not all, dead blocks.
3988 // Eventually, this block will be deleted.
3989 if (isa<UndefValue>(RI->getOperand(0)))
3992 if (mergeCleanupPad(RI))
3995 if (removeEmptyCleanup(RI))
4001 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4002 BasicBlock *BB = RI->getParent();
4003 if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4006 // Find predecessors that end with branches.
4007 SmallVector<BasicBlock *, 8> UncondBranchPreds;
4008 SmallVector<BranchInst *, 8> CondBranchPreds;
4009 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4010 BasicBlock *P = *PI;
4011 TerminatorInst *PTI = P->getTerminator();
4012 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4013 if (BI->isUnconditional())
4014 UncondBranchPreds.push_back(P);
4016 CondBranchPreds.push_back(BI);
4020 // If we found some, do the transformation!
4021 if (!UncondBranchPreds.empty() && DupRet) {
4022 while (!UncondBranchPreds.empty()) {
4023 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4024 DEBUG(dbgs() << "FOLDING: " << *BB
4025 << "INTO UNCOND BRANCH PRED: " << *Pred);
4026 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4029 // If we eliminated all predecessors of the block, delete the block now.
4030 if (pred_empty(BB)) {
4031 // We know there are no successors, so just nuke the block.
4032 BB->eraseFromParent();
4034 LoopHeaders->erase(BB);
4040 // Check out all of the conditional branches going to this return
4041 // instruction. If any of them just select between returns, change the
4042 // branch itself into a select/return pair.
4043 while (!CondBranchPreds.empty()) {
4044 BranchInst *BI = CondBranchPreds.pop_back_val();
4046 // Check to see if the non-BB successor is also a return block.
4047 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4048 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4049 SimplifyCondBranchToTwoReturns(BI, Builder))
4055 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4056 BasicBlock *BB = UI->getParent();
4058 bool Changed = false;
4060 // If there are any instructions immediately before the unreachable that can
4061 // be removed, do so.
4062 while (UI->getIterator() != BB->begin()) {
4063 BasicBlock::iterator BBI = UI->getIterator();
4065 // Do not delete instructions that can have side effects which might cause
4066 // the unreachable to not be reachable; specifically, calls and volatile
4067 // operations may have this effect.
4068 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4071 if (BBI->mayHaveSideEffects()) {
4072 if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4073 if (SI->isVolatile())
4075 } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4076 if (LI->isVolatile())
4078 } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4079 if (RMWI->isVolatile())
4081 } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4082 if (CXI->isVolatile())
4084 } else if (isa<CatchPadInst>(BBI)) {
4085 // A catchpad may invoke exception object constructors and such, which
4086 // in some languages can be arbitrary code, so be conservative by
4088 // For CoreCLR, it just involves a type test, so can be removed.
4089 if (classifyEHPersonality(BB->getParent()->getPersonalityFn()) !=
4090 EHPersonality::CoreCLR)
4092 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4093 !isa<LandingPadInst>(BBI)) {
4096 // Note that deleting LandingPad's here is in fact okay, although it
4097 // involves a bit of subtle reasoning. If this inst is a LandingPad,
4098 // all the predecessors of this block will be the unwind edges of Invokes,
4099 // and we can therefore guarantee this block will be erased.
4102 // Delete this instruction (any uses are guaranteed to be dead)
4103 if (!BBI->use_empty())
4104 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4105 BBI->eraseFromParent();
4109 // If the unreachable instruction is the first in the block, take a gander
4110 // at all of the predecessors of this instruction, and simplify them.
4111 if (&BB->front() != UI)
4114 SmallVector<BasicBlock *, 8> Preds(pred_begin(BB), pred_end(BB));
4115 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4116 TerminatorInst *TI = Preds[i]->getTerminator();
4117 IRBuilder<> Builder(TI);
4118 if (auto *BI = dyn_cast<BranchInst>(TI)) {
4119 if (BI->isUnconditional()) {
4120 if (BI->getSuccessor(0) == BB) {
4121 new UnreachableInst(TI->getContext(), TI);
4122 TI->eraseFromParent();
4126 if (BI->getSuccessor(0) == BB) {
4127 Builder.CreateBr(BI->getSuccessor(1));
4128 EraseTerminatorInstAndDCECond(BI);
4129 } else if (BI->getSuccessor(1) == BB) {
4130 Builder.CreateBr(BI->getSuccessor(0));
4131 EraseTerminatorInstAndDCECond(BI);
4135 } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4136 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e;
4138 if (i.getCaseSuccessor() == BB) {
4139 BB->removePredecessor(SI->getParent());
4145 } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4146 if (II->getUnwindDest() == BB) {
4147 removeUnwindEdge(TI->getParent());
4150 } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4151 if (CSI->getUnwindDest() == BB) {
4152 removeUnwindEdge(TI->getParent());
4157 for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4158 E = CSI->handler_end();
4161 CSI->removeHandler(I);
4167 if (CSI->getNumHandlers() == 0) {
4168 BasicBlock *CatchSwitchBB = CSI->getParent();
4169 if (CSI->hasUnwindDest()) {
4170 // Redirect preds to the unwind dest
4171 CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4173 // Rewrite all preds to unwind to caller (or from invoke to call).
4174 SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4175 for (BasicBlock *EHPred : EHPreds)
4176 removeUnwindEdge(EHPred);
4178 // The catchswitch is no longer reachable.
4179 new UnreachableInst(CSI->getContext(), CSI);
4180 CSI->eraseFromParent();
4183 } else if (isa<CleanupReturnInst>(TI)) {
4184 new UnreachableInst(TI->getContext(), TI);
4185 TI->eraseFromParent();
4190 // If this block is now dead, remove it.
4191 if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4192 // We know there are no successors, so just nuke the block.
4193 BB->eraseFromParent();
4195 LoopHeaders->erase(BB);
4202 static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) {
4203 assert(Cases.size() >= 1);
4205 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4206 for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4207 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4213 /// Turn a switch with two reachable destinations into an integer range
4214 /// comparison and branch.
4215 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
4216 assert(SI->getNumCases() > 1 && "Degenerate switch?");
4219 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4221 // Partition the cases into two sets with different destinations.
4222 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4223 BasicBlock *DestB = nullptr;
4224 SmallVector<ConstantInt *, 16> CasesA;
4225 SmallVector<ConstantInt *, 16> CasesB;
4227 for (SwitchInst::CaseIt I : SI->cases()) {
4228 BasicBlock *Dest = I.getCaseSuccessor();
4231 if (Dest == DestA) {
4232 CasesA.push_back(I.getCaseValue());
4237 if (Dest == DestB) {
4238 CasesB.push_back(I.getCaseValue());
4241 return false; // More than two destinations.
4244 assert(DestA && DestB &&
4245 "Single-destination switch should have been folded.");
4246 assert(DestA != DestB);
4247 assert(DestB != SI->getDefaultDest());
4248 assert(!CasesB.empty() && "There must be non-default cases.");
4249 assert(!CasesA.empty() || HasDefault);
4251 // Figure out if one of the sets of cases form a contiguous range.
4252 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4253 BasicBlock *ContiguousDest = nullptr;
4254 BasicBlock *OtherDest = nullptr;
4255 if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4256 ContiguousCases = &CasesA;
4257 ContiguousDest = DestA;
4259 } else if (CasesAreContiguous(CasesB)) {
4260 ContiguousCases = &CasesB;
4261 ContiguousDest = DestB;
4266 // Start building the compare and branch.
4268 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4269 Constant *NumCases =
4270 ConstantInt::get(Offset->getType(), ContiguousCases->size());
4272 Value *Sub = SI->getCondition();
4273 if (!Offset->isNullValue())
4274 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4277 // If NumCases overflowed, then all possible values jump to the successor.
4278 if (NumCases->isNullValue() && !ContiguousCases->empty())
4279 Cmp = ConstantInt::getTrue(SI->getContext());
4281 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4282 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4284 // Update weight for the newly-created conditional branch.
4285 if (HasBranchWeights(SI)) {
4286 SmallVector<uint64_t, 8> Weights;
4287 GetBranchWeights(SI, Weights);
4288 if (Weights.size() == 1 + SI->getNumCases()) {
4289 uint64_t TrueWeight = 0;
4290 uint64_t FalseWeight = 0;
4291 for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4292 if (SI->getSuccessor(I) == ContiguousDest)
4293 TrueWeight += Weights[I];
4295 FalseWeight += Weights[I];
4297 while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4301 NewBI->setMetadata(LLVMContext::MD_prof,
4302 MDBuilder(SI->getContext())
4303 .createBranchWeights((uint32_t)TrueWeight,
4304 (uint32_t)FalseWeight));
4308 // Prune obsolete incoming values off the successors' PHI nodes.
4309 for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4310 unsigned PreviousEdges = ContiguousCases->size();
4311 if (ContiguousDest == SI->getDefaultDest())
4313 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4314 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4316 for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4317 unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4318 if (OtherDest == SI->getDefaultDest())
4320 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4321 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4325 SI->eraseFromParent();
4330 /// Compute masked bits for the condition of a switch
4331 /// and use it to remove dead cases.
4332 static bool EliminateDeadSwitchCases(SwitchInst *SI, AssumptionCache *AC,
4333 const DataLayout &DL) {
4334 Value *Cond = SI->getCondition();
4335 unsigned Bits = Cond->getType()->getIntegerBitWidth();
4336 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
4337 computeKnownBits(Cond, KnownZero, KnownOne, DL, 0, AC, SI);
4339 // We can also eliminate cases by determining that their values are outside of
4340 // the limited range of the condition based on how many significant (non-sign)
4341 // bits are in the condition value.
4342 unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4343 unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4345 // Gather dead cases.
4346 SmallVector<ConstantInt *, 8> DeadCases;
4347 for (auto &Case : SI->cases()) {
4348 APInt CaseVal = Case.getCaseValue()->getValue();
4349 if ((CaseVal & KnownZero) != 0 || (CaseVal & KnownOne) != KnownOne ||
4350 (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4351 DeadCases.push_back(Case.getCaseValue());
4352 DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal << " is dead.\n");
4356 // If we can prove that the cases must cover all possible values, the
4357 // default destination becomes dead and we can remove it. If we know some
4358 // of the bits in the value, we can use that to more precisely compute the
4359 // number of possible unique case values.
4361 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4362 const unsigned NumUnknownBits =
4363 Bits - (KnownZero.Or(KnownOne)).countPopulation();
4364 assert(NumUnknownBits <= Bits);
4365 if (HasDefault && DeadCases.empty() &&
4366 NumUnknownBits < 64 /* avoid overflow */ &&
4367 SI->getNumCases() == (1ULL << NumUnknownBits)) {
4368 DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4369 BasicBlock *NewDefault =
4370 SplitBlockPredecessors(SI->getDefaultDest(), SI->getParent(), "");
4371 SI->setDefaultDest(&*NewDefault);
4372 SplitBlock(&*NewDefault, &NewDefault->front());
4373 auto *OldTI = NewDefault->getTerminator();
4374 new UnreachableInst(SI->getContext(), OldTI);
4375 EraseTerminatorInstAndDCECond(OldTI);
4379 SmallVector<uint64_t, 8> Weights;
4380 bool HasWeight = HasBranchWeights(SI);
4382 GetBranchWeights(SI, Weights);
4383 HasWeight = (Weights.size() == 1 + SI->getNumCases());
4386 // Remove dead cases from the switch.
4387 for (ConstantInt *DeadCase : DeadCases) {
4388 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCase);
4389 assert(Case != SI->case_default() &&
4390 "Case was not found. Probably mistake in DeadCases forming.");
4392 std::swap(Weights[Case.getCaseIndex() + 1], Weights.back());
4396 // Prune unused values from PHI nodes.
4397 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
4398 SI->removeCase(Case);
4400 if (HasWeight && Weights.size() >= 2) {
4401 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
4402 SI->setMetadata(LLVMContext::MD_prof,
4403 MDBuilder(SI->getParent()->getContext())
4404 .createBranchWeights(MDWeights));
4407 return !DeadCases.empty();
4410 /// If BB would be eligible for simplification by
4411 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4412 /// by an unconditional branch), look at the phi node for BB in the successor
4413 /// block and see if the incoming value is equal to CaseValue. If so, return
4414 /// the phi node, and set PhiIndex to BB's index in the phi node.
4415 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
4416 BasicBlock *BB, int *PhiIndex) {
4417 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4418 return nullptr; // BB must be empty to be a candidate for simplification.
4419 if (!BB->getSinglePredecessor())
4420 return nullptr; // BB must be dominated by the switch.
4422 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
4423 if (!Branch || !Branch->isUnconditional())
4424 return nullptr; // Terminator must be unconditional branch.
4426 BasicBlock *Succ = Branch->getSuccessor(0);
4428 BasicBlock::iterator I = Succ->begin();
4429 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
4430 int Idx = PHI->getBasicBlockIndex(BB);
4431 assert(Idx >= 0 && "PHI has no entry for predecessor?");
4433 Value *InValue = PHI->getIncomingValue(Idx);
4434 if (InValue != CaseValue)
4444 /// Try to forward the condition of a switch instruction to a phi node
4445 /// dominated by the switch, if that would mean that some of the destination
4446 /// blocks of the switch can be folded away.
4447 /// Returns true if a change is made.
4448 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
4449 typedef DenseMap<PHINode *, SmallVector<int, 4>> ForwardingNodesMap;
4450 ForwardingNodesMap ForwardingNodes;
4452 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E;
4454 ConstantInt *CaseValue = I.getCaseValue();
4455 BasicBlock *CaseDest = I.getCaseSuccessor();
4459 FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIndex);
4463 ForwardingNodes[PHI].push_back(PhiIndex);
4466 bool Changed = false;
4468 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
4469 E = ForwardingNodes.end();
4471 PHINode *Phi = I->first;
4472 SmallVectorImpl<int> &Indexes = I->second;
4474 if (Indexes.size() < 2)
4477 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
4478 Phi->setIncomingValue(Indexes[I], SI->getCondition());
4485 /// Return true if the backend will be able to handle
4486 /// initializing an array of constants like C.
4487 static bool ValidLookupTableConstant(Constant *C, const TargetTransformInfo &TTI) {
4488 if (C->isThreadDependent())
4490 if (C->isDLLImportDependent())
4493 if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4494 !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4495 !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4498 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4499 if (!CE->isGEPWithNoNotionalOverIndexing())
4501 if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4505 if (!TTI.shouldBuildLookupTablesForConstant(C))
4511 /// If V is a Constant, return it. Otherwise, try to look up
4512 /// its constant value in ConstantPool, returning 0 if it's not there.
4514 LookupConstant(Value *V,
4515 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4516 if (Constant *C = dyn_cast<Constant>(V))
4518 return ConstantPool.lookup(V);
4521 /// Try to fold instruction I into a constant. This works for
4522 /// simple instructions such as binary operations where both operands are
4523 /// constant or can be replaced by constants from the ConstantPool. Returns the
4524 /// resulting constant on success, 0 otherwise.
4526 ConstantFold(Instruction *I, const DataLayout &DL,
4527 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4528 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4529 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4532 if (A->isAllOnesValue())
4533 return LookupConstant(Select->getTrueValue(), ConstantPool);
4534 if (A->isNullValue())
4535 return LookupConstant(Select->getFalseValue(), ConstantPool);
4539 SmallVector<Constant *, 4> COps;
4540 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4541 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
4547 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4548 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4552 return ConstantFoldInstOperands(I, COps, DL);
4555 /// Try to determine the resulting constant values in phi nodes
4556 /// at the common destination basic block, *CommonDest, for one of the case
4557 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4558 /// case), of a switch instruction SI.
4560 GetCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest,
4561 BasicBlock **CommonDest,
4562 SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4563 const DataLayout &DL, const TargetTransformInfo &TTI) {
4564 // The block from which we enter the common destination.
4565 BasicBlock *Pred = SI->getParent();
4567 // If CaseDest is empty except for some side-effect free instructions through
4568 // which we can constant-propagate the CaseVal, continue to its successor.
4569 SmallDenseMap<Value *, Constant *> ConstantPool;
4570 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4571 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
4573 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
4574 // If the terminator is a simple branch, continue to the next block.
4575 if (T->getNumSuccessors() != 1 || T->isExceptional())
4578 CaseDest = T->getSuccessor(0);
4579 } else if (isa<DbgInfoIntrinsic>(I)) {
4580 // Skip debug intrinsic.
4582 } else if (Constant *C = ConstantFold(&*I, DL, ConstantPool)) {
4583 // Instruction is side-effect free and constant.
4585 // If the instruction has uses outside this block or a phi node slot for
4586 // the block, it is not safe to bypass the instruction since it would then
4587 // no longer dominate all its uses.
4588 for (auto &Use : I->uses()) {
4589 User *User = Use.getUser();
4590 if (Instruction *I = dyn_cast<Instruction>(User))
4591 if (I->getParent() == CaseDest)
4593 if (PHINode *Phi = dyn_cast<PHINode>(User))
4594 if (Phi->getIncomingBlock(Use) == CaseDest)
4599 ConstantPool.insert(std::make_pair(&*I, C));
4605 // If we did not have a CommonDest before, use the current one.
4607 *CommonDest = CaseDest;
4608 // If the destination isn't the common one, abort.
4609 if (CaseDest != *CommonDest)
4612 // Get the values for this case from phi nodes in the destination block.
4613 BasicBlock::iterator I = (*CommonDest)->begin();
4614 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
4615 int Idx = PHI->getBasicBlockIndex(Pred);
4619 Constant *ConstVal =
4620 LookupConstant(PHI->getIncomingValue(Idx), ConstantPool);
4624 // Be conservative about which kinds of constants we support.
4625 if (!ValidLookupTableConstant(ConstVal, TTI))
4628 Res.push_back(std::make_pair(PHI, ConstVal));
4631 return Res.size() > 0;
4634 // Helper function used to add CaseVal to the list of cases that generate
4636 static void MapCaseToResult(ConstantInt *CaseVal,
4637 SwitchCaseResultVectorTy &UniqueResults,
4639 for (auto &I : UniqueResults) {
4640 if (I.first == Result) {
4641 I.second.push_back(CaseVal);
4645 UniqueResults.push_back(
4646 std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4649 // Helper function that initializes a map containing
4650 // results for the PHI node of the common destination block for a switch
4651 // instruction. Returns false if multiple PHI nodes have been found or if
4652 // there is not a common destination block for the switch.
4653 static bool InitializeUniqueCases(SwitchInst *SI, PHINode *&PHI,
4654 BasicBlock *&CommonDest,
4655 SwitchCaseResultVectorTy &UniqueResults,
4656 Constant *&DefaultResult,
4657 const DataLayout &DL,
4658 const TargetTransformInfo &TTI) {
4659 for (auto &I : SI->cases()) {
4660 ConstantInt *CaseVal = I.getCaseValue();
4662 // Resulting value at phi nodes for this case value.
4663 SwitchCaseResultsTy Results;
4664 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4668 // Only one value per case is permitted
4669 if (Results.size() > 1)
4671 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4673 // Check the PHI consistency.
4675 PHI = Results[0].first;
4676 else if (PHI != Results[0].first)
4679 // Find the default result value.
4680 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
4681 BasicBlock *DefaultDest = SI->getDefaultDest();
4682 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4684 // If the default value is not found abort unless the default destination
4687 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4688 if ((!DefaultResult &&
4689 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4695 // Helper function that checks if it is possible to transform a switch with only
4696 // two cases (or two cases + default) that produces a result into a select.
4699 // case 10: %0 = icmp eq i32 %a, 10
4700 // return 10; %1 = select i1 %0, i32 10, i32 4
4701 // case 20: ----> %2 = icmp eq i32 %a, 20
4702 // return 2; %3 = select i1 %2, i32 2, i32 %1
4706 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4707 Constant *DefaultResult, Value *Condition,
4708 IRBuilder<> &Builder) {
4709 assert(ResultVector.size() == 2 &&
4710 "We should have exactly two unique results at this point");
4711 // If we are selecting between only two cases transform into a simple
4712 // select or a two-way select if default is possible.
4713 if (ResultVector[0].second.size() == 1 &&
4714 ResultVector[1].second.size() == 1) {
4715 ConstantInt *const FirstCase = ResultVector[0].second[0];
4716 ConstantInt *const SecondCase = ResultVector[1].second[0];
4718 bool DefaultCanTrigger = DefaultResult;
4719 Value *SelectValue = ResultVector[1].first;
4720 if (DefaultCanTrigger) {
4721 Value *const ValueCompare =
4722 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4723 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4724 DefaultResult, "switch.select");
4726 Value *const ValueCompare =
4727 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4728 return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4729 SelectValue, "switch.select");
4735 // Helper function to cleanup a switch instruction that has been converted into
4736 // a select, fixing up PHI nodes and basic blocks.
4737 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
4739 IRBuilder<> &Builder) {
4740 BasicBlock *SelectBB = SI->getParent();
4741 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4742 PHI->removeIncomingValue(SelectBB);
4743 PHI->addIncoming(SelectValue, SelectBB);
4745 Builder.CreateBr(PHI->getParent());
4747 // Remove the switch.
4748 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4749 BasicBlock *Succ = SI->getSuccessor(i);
4751 if (Succ == PHI->getParent())
4753 Succ->removePredecessor(SelectBB);
4755 SI->eraseFromParent();
4758 /// If the switch is only used to initialize one or more
4759 /// phi nodes in a common successor block with only two different
4760 /// constant values, replace the switch with select.
4761 static bool SwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4762 AssumptionCache *AC, const DataLayout &DL,
4763 const TargetTransformInfo &TTI) {
4764 Value *const Cond = SI->getCondition();
4765 PHINode *PHI = nullptr;
4766 BasicBlock *CommonDest = nullptr;
4767 Constant *DefaultResult;
4768 SwitchCaseResultVectorTy UniqueResults;
4769 // Collect all the cases that will deliver the same value from the switch.
4770 if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4773 // Selects choose between maximum two values.
4774 if (UniqueResults.size() != 2)
4776 assert(PHI != nullptr && "PHI for value select not found");
4778 Builder.SetInsertPoint(SI);
4779 Value *SelectValue =
4780 ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4782 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4785 // The switch couldn't be converted into a select.
4791 /// This class represents a lookup table that can be used to replace a switch.
4792 class SwitchLookupTable {
4794 /// Create a lookup table to use as a switch replacement with the contents
4795 /// of Values, using DefaultValue to fill any holes in the table.
4797 Module &M, uint64_t TableSize, ConstantInt *Offset,
4798 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4799 Constant *DefaultValue, const DataLayout &DL);
4801 /// Build instructions with Builder to retrieve the value at
4802 /// the position given by Index in the lookup table.
4803 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4805 /// Return true if a table with TableSize elements of
4806 /// type ElementType would fit in a target-legal register.
4807 static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4811 // Depending on the contents of the table, it can be represented in
4814 // For tables where each element contains the same value, we just have to
4815 // store that single value and return it for each lookup.
4818 // For tables where there is a linear relationship between table index
4819 // and values. We calculate the result with a simple multiplication
4820 // and addition instead of a table lookup.
4823 // For small tables with integer elements, we can pack them into a bitmap
4824 // that fits into a target-legal register. Values are retrieved by
4825 // shift and mask operations.
4828 // The table is stored as an array of values. Values are retrieved by load
4829 // instructions from the table.
4833 // For SingleValueKind, this is the single value.
4834 Constant *SingleValue;
4836 // For BitMapKind, this is the bitmap.
4837 ConstantInt *BitMap;
4838 IntegerType *BitMapElementTy;
4840 // For LinearMapKind, these are the constants used to derive the value.
4841 ConstantInt *LinearOffset;
4842 ConstantInt *LinearMultiplier;
4844 // For ArrayKind, this is the array.
4845 GlobalVariable *Array;
4848 } // end anonymous namespace
4850 SwitchLookupTable::SwitchLookupTable(
4851 Module &M, uint64_t TableSize, ConstantInt *Offset,
4852 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4853 Constant *DefaultValue, const DataLayout &DL)
4854 : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
4855 LinearOffset(nullptr), LinearMultiplier(nullptr), Array(nullptr) {
4856 assert(Values.size() && "Can't build lookup table without values!");
4857 assert(TableSize >= Values.size() && "Can't fit values in table!");
4859 // If all values in the table are equal, this is that value.
4860 SingleValue = Values.begin()->second;
4862 Type *ValueType = Values.begin()->second->getType();
4864 // Build up the table contents.
4865 SmallVector<Constant *, 64> TableContents(TableSize);
4866 for (size_t I = 0, E = Values.size(); I != E; ++I) {
4867 ConstantInt *CaseVal = Values[I].first;
4868 Constant *CaseRes = Values[I].second;
4869 assert(CaseRes->getType() == ValueType);
4871 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4872 TableContents[Idx] = CaseRes;
4874 if (CaseRes != SingleValue)
4875 SingleValue = nullptr;
4878 // Fill in any holes in the table with the default result.
4879 if (Values.size() < TableSize) {
4880 assert(DefaultValue &&
4881 "Need a default value to fill the lookup table holes.");
4882 assert(DefaultValue->getType() == ValueType);
4883 for (uint64_t I = 0; I < TableSize; ++I) {
4884 if (!TableContents[I])
4885 TableContents[I] = DefaultValue;
4888 if (DefaultValue != SingleValue)
4889 SingleValue = nullptr;
4892 // If each element in the table contains the same value, we only need to store
4893 // that single value.
4895 Kind = SingleValueKind;
4899 // Check if we can derive the value with a linear transformation from the
4901 if (isa<IntegerType>(ValueType)) {
4902 bool LinearMappingPossible = true;
4905 assert(TableSize >= 2 && "Should be a SingleValue table.");
4906 // Check if there is the same distance between two consecutive values.
4907 for (uint64_t I = 0; I < TableSize; ++I) {
4908 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4910 // This is an undef. We could deal with it, but undefs in lookup tables
4911 // are very seldom. It's probably not worth the additional complexity.
4912 LinearMappingPossible = false;
4915 APInt Val = ConstVal->getValue();
4917 APInt Dist = Val - PrevVal;
4920 } else if (Dist != DistToPrev) {
4921 LinearMappingPossible = false;
4927 if (LinearMappingPossible) {
4928 LinearOffset = cast<ConstantInt>(TableContents[0]);
4929 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
4930 Kind = LinearMapKind;
4936 // If the type is integer and the table fits in a register, build a bitmap.
4937 if (WouldFitInRegister(DL, TableSize, ValueType)) {
4938 IntegerType *IT = cast<IntegerType>(ValueType);
4939 APInt TableInt(TableSize * IT->getBitWidth(), 0);
4940 for (uint64_t I = TableSize; I > 0; --I) {
4941 TableInt <<= IT->getBitWidth();
4942 // Insert values into the bitmap. Undef values are set to zero.
4943 if (!isa<UndefValue>(TableContents[I - 1])) {
4944 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
4945 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
4948 BitMap = ConstantInt::get(M.getContext(), TableInt);
4949 BitMapElementTy = IT;
4955 // Store the table in an array.
4956 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
4957 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
4959 Array = new GlobalVariable(M, ArrayTy, /*constant=*/true,
4960 GlobalVariable::PrivateLinkage, Initializer,
4962 Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
4966 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
4968 case SingleValueKind:
4970 case LinearMapKind: {
4971 // Derive the result value from the input value.
4972 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
4973 false, "switch.idx.cast");
4974 if (!LinearMultiplier->isOne())
4975 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
4976 if (!LinearOffset->isZero())
4977 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
4981 // Type of the bitmap (e.g. i59).
4982 IntegerType *MapTy = BitMap->getType();
4984 // Cast Index to the same type as the bitmap.
4985 // Note: The Index is <= the number of elements in the table, so
4986 // truncating it to the width of the bitmask is safe.
4987 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
4989 // Multiply the shift amount by the element width.
4990 ShiftAmt = Builder.CreateMul(
4991 ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
4995 Value *DownShifted =
4996 Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
4998 return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5001 // Make sure the table index will not overflow when treated as signed.
5002 IntegerType *IT = cast<IntegerType>(Index->getType());
5003 uint64_t TableSize =
5004 Array->getInitializer()->getType()->getArrayNumElements();
5005 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5006 Index = Builder.CreateZExt(
5007 Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5008 "switch.tableidx.zext");
5010 Value *GEPIndices[] = {Builder.getInt32(0), Index};
5011 Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5012 GEPIndices, "switch.gep");
5013 return Builder.CreateLoad(GEP, "switch.load");
5016 llvm_unreachable("Unknown lookup table kind!");
5019 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5021 Type *ElementType) {
5022 auto *IT = dyn_cast<IntegerType>(ElementType);
5025 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5026 // are <= 15, we could try to narrow the type.
5028 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5029 if (TableSize >= UINT_MAX / IT->getBitWidth())
5031 return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5034 /// Determine whether a lookup table should be built for this switch, based on
5035 /// the number of cases, size of the table, and the types of the results.
5037 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5038 const TargetTransformInfo &TTI, const DataLayout &DL,
5039 const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5040 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5041 return false; // TableSize overflowed, or mul below might overflow.
5043 bool AllTablesFitInRegister = true;
5044 bool HasIllegalType = false;
5045 for (const auto &I : ResultTypes) {
5046 Type *Ty = I.second;
5048 // Saturate this flag to true.
5049 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5051 // Saturate this flag to false.
5052 AllTablesFitInRegister =
5053 AllTablesFitInRegister &&
5054 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5056 // If both flags saturate, we're done. NOTE: This *only* works with
5057 // saturating flags, and all flags have to saturate first due to the
5058 // non-deterministic behavior of iterating over a dense map.
5059 if (HasIllegalType && !AllTablesFitInRegister)
5063 // If each table would fit in a register, we should build it anyway.
5064 if (AllTablesFitInRegister)
5067 // Don't build a table that doesn't fit in-register if it has illegal types.
5071 // The table density should be at least 40%. This is the same criterion as for
5072 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5073 // FIXME: Find the best cut-off.
5074 return SI->getNumCases() * 10 >= TableSize * 4;
5077 /// Try to reuse the switch table index compare. Following pattern:
5079 /// if (idx < tablesize)
5080 /// r = table[idx]; // table does not contain default_value
5082 /// r = default_value;
5083 /// if (r != default_value)
5086 /// Is optimized to:
5088 /// cond = idx < tablesize;
5092 /// r = default_value;
5096 /// Jump threading will then eliminate the second if(cond).
5097 static void reuseTableCompare(
5098 User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5099 Constant *DefaultValue,
5100 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5102 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5106 // We require that the compare is in the same block as the phi so that jump
5107 // threading can do its work afterwards.
5108 if (CmpInst->getParent() != PhiBlock)
5111 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5115 Value *RangeCmp = RangeCheckBranch->getCondition();
5116 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5117 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5119 // Check if the compare with the default value is constant true or false.
5120 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5121 DefaultValue, CmpOp1, true);
5122 if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5125 // Check if the compare with the case values is distinct from the default
5127 for (auto ValuePair : Values) {
5128 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5129 ValuePair.second, CmpOp1, true);
5130 if (!CaseConst || CaseConst == DefaultConst)
5132 assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5133 "Expect true or false as compare result.");
5136 // Check if the branch instruction dominates the phi node. It's a simple
5137 // dominance check, but sufficient for our needs.
5138 // Although this check is invariant in the calling loops, it's better to do it
5139 // at this late stage. Practically we do it at most once for a switch.
5140 BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5141 for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5142 BasicBlock *Pred = *PI;
5143 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5147 if (DefaultConst == FalseConst) {
5148 // The compare yields the same result. We can replace it.
5149 CmpInst->replaceAllUsesWith(RangeCmp);
5150 ++NumTableCmpReuses;
5152 // The compare yields the same result, just inverted. We can replace it.
5153 Value *InvertedTableCmp = BinaryOperator::CreateXor(
5154 RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5156 CmpInst->replaceAllUsesWith(InvertedTableCmp);
5157 ++NumTableCmpReuses;
5161 /// If the switch is only used to initialize one or more phi nodes in a common
5162 /// successor block with different constant values, replace the switch with
5164 static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
5165 const DataLayout &DL,
5166 const TargetTransformInfo &TTI) {
5167 assert(SI->getNumCases() > 1 && "Degenerate switch?");
5169 // Only build lookup table when we have a target that supports it.
5170 if (!TTI.shouldBuildLookupTables())
5173 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5174 // split off a dense part and build a lookup table for that.
5176 // FIXME: This creates arrays of GEPs to constant strings, which means each
5177 // GEP needs a runtime relocation in PIC code. We should just build one big
5178 // string and lookup indices into that.
5180 // Ignore switches with less than three cases. Lookup tables will not make
5182 // faster, so we don't analyze them.
5183 if (SI->getNumCases() < 3)
5186 // Figure out the corresponding result for each case value and phi node in the
5187 // common destination, as well as the min and max case values.
5188 assert(SI->case_begin() != SI->case_end());
5189 SwitchInst::CaseIt CI = SI->case_begin();
5190 ConstantInt *MinCaseVal = CI.getCaseValue();
5191 ConstantInt *MaxCaseVal = CI.getCaseValue();
5193 BasicBlock *CommonDest = nullptr;
5194 typedef SmallVector<std::pair<ConstantInt *, Constant *>, 4> ResultListTy;
5195 SmallDenseMap<PHINode *, ResultListTy> ResultLists;
5196 SmallDenseMap<PHINode *, Constant *> DefaultResults;
5197 SmallDenseMap<PHINode *, Type *> ResultTypes;
5198 SmallVector<PHINode *, 4> PHIs;
5200 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5201 ConstantInt *CaseVal = CI.getCaseValue();
5202 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5203 MinCaseVal = CaseVal;
5204 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5205 MaxCaseVal = CaseVal;
5207 // Resulting value at phi nodes for this case value.
5208 typedef SmallVector<std::pair<PHINode *, Constant *>, 4> ResultsTy;
5210 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
5214 // Append the result from this case to the list for each phi.
5215 for (const auto &I : Results) {
5216 PHINode *PHI = I.first;
5217 Constant *Value = I.second;
5218 if (!ResultLists.count(PHI))
5219 PHIs.push_back(PHI);
5220 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5224 // Keep track of the result types.
5225 for (PHINode *PHI : PHIs) {
5226 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5229 uint64_t NumResults = ResultLists[PHIs[0]].size();
5230 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5231 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5232 bool TableHasHoles = (NumResults < TableSize);
5234 // If the table has holes, we need a constant result for the default case
5235 // or a bitmask that fits in a register.
5236 SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5237 bool HasDefaultResults =
5238 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5239 DefaultResultsList, DL, TTI);
5241 bool NeedMask = (TableHasHoles && !HasDefaultResults);
5243 // As an extra penalty for the validity test we require more cases.
5244 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5246 if (!DL.fitsInLegalInteger(TableSize))
5250 for (const auto &I : DefaultResultsList) {
5251 PHINode *PHI = I.first;
5252 Constant *Result = I.second;
5253 DefaultResults[PHI] = Result;
5256 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5259 // Create the BB that does the lookups.
5260 Module &Mod = *CommonDest->getParent()->getParent();
5261 BasicBlock *LookupBB = BasicBlock::Create(
5262 Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5264 // Compute the table index value.
5265 Builder.SetInsertPoint(SI);
5267 Builder.CreateSub(SI->getCondition(), MinCaseVal, "switch.tableidx");
5269 // Compute the maximum table size representable by the integer type we are
5271 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5272 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5273 assert(MaxTableSize >= TableSize &&
5274 "It is impossible for a switch to have more entries than the max "
5275 "representable value of its input integer type's size.");
5277 // If the default destination is unreachable, or if the lookup table covers
5278 // all values of the conditional variable, branch directly to the lookup table
5279 // BB. Otherwise, check that the condition is within the case range.
5280 const bool DefaultIsReachable =
5281 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5282 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5283 BranchInst *RangeCheckBranch = nullptr;
5285 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5286 Builder.CreateBr(LookupBB);
5287 // Note: We call removeProdecessor later since we need to be able to get the
5288 // PHI value for the default case in case we're using a bit mask.
5290 Value *Cmp = Builder.CreateICmpULT(
5291 TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5293 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5296 // Populate the BB that does the lookups.
5297 Builder.SetInsertPoint(LookupBB);
5300 // Before doing the lookup we do the hole check.
5301 // The LookupBB is therefore re-purposed to do the hole check
5302 // and we create a new LookupBB.
5303 BasicBlock *MaskBB = LookupBB;
5304 MaskBB->setName("switch.hole_check");
5305 LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5306 CommonDest->getParent(), CommonDest);
5308 // Make the mask's bitwidth at least 8bit and a power-of-2 to avoid
5309 // unnecessary illegal types.
5310 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5311 APInt MaskInt(TableSizePowOf2, 0);
5312 APInt One(TableSizePowOf2, 1);
5313 // Build bitmask; fill in a 1 bit for every case.
5314 const ResultListTy &ResultList = ResultLists[PHIs[0]];
5315 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5316 uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5318 MaskInt |= One << Idx;
5320 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5322 // Get the TableIndex'th bit of the bitmask.
5323 // If this bit is 0 (meaning hole) jump to the default destination,
5324 // else continue with table lookup.
5325 IntegerType *MapTy = TableMask->getType();
5327 Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5328 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5329 Value *LoBit = Builder.CreateTrunc(
5330 Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5331 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5333 Builder.SetInsertPoint(LookupBB);
5334 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5337 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5338 // We cached PHINodes in PHIs, to avoid accessing deleted PHINodes later,
5339 // do not delete PHINodes here.
5340 SI->getDefaultDest()->removePredecessor(SI->getParent(),
5341 /*DontDeleteUselessPHIs=*/true);
5344 bool ReturnedEarly = false;
5345 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
5346 PHINode *PHI = PHIs[I];
5347 const ResultListTy &ResultList = ResultLists[PHI];
5349 // If using a bitmask, use any value to fill the lookup table holes.
5350 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5351 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL);
5353 Value *Result = Table.BuildLookup(TableIndex, Builder);
5355 // If the result is used to return immediately from the function, we want to
5356 // do that right here.
5357 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5358 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5359 Builder.CreateRet(Result);
5360 ReturnedEarly = true;
5364 // Do a small peephole optimization: re-use the switch table compare if
5366 if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5367 BasicBlock *PhiBlock = PHI->getParent();
5368 // Search for compare instructions which use the phi.
5369 for (auto *User : PHI->users()) {
5370 reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5374 PHI->addIncoming(Result, LookupBB);
5378 Builder.CreateBr(CommonDest);
5380 // Remove the switch.
5381 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5382 BasicBlock *Succ = SI->getSuccessor(i);
5384 if (Succ == SI->getDefaultDest())
5386 Succ->removePredecessor(SI->getParent());
5388 SI->eraseFromParent();
5392 ++NumLookupTablesHoles;
5396 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5397 // See also SelectionDAGBuilder::isDense(), which this function was based on.
5398 uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5399 uint64_t Range = Diff + 1;
5400 uint64_t NumCases = Values.size();
5401 // 40% is the default density for building a jump table in optsize/minsize mode.
5402 uint64_t MinDensity = 40;
5404 return NumCases * 100 >= Range * MinDensity;
5407 // Try and transform a switch that has "holes" in it to a contiguous sequence
5410 // A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5411 // range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5413 // This converts a sparse switch into a dense switch which allows better
5414 // lowering and could also allow transforming into a lookup table.
5415 static bool ReduceSwitchRange(SwitchInst *SI, IRBuilder<> &Builder,
5416 const DataLayout &DL,
5417 const TargetTransformInfo &TTI) {
5418 auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5419 if (CondTy->getIntegerBitWidth() > 64 ||
5420 !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5422 // Only bother with this optimization if there are more than 3 switch cases;
5423 // SDAG will only bother creating jump tables for 4 or more cases.
5424 if (SI->getNumCases() < 4)
5427 // This transform is agnostic to the signedness of the input or case values. We
5428 // can treat the case values as signed or unsigned. We can optimize more common
5429 // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5431 SmallVector<int64_t,4> Values;
5432 for (auto &C : SI->cases())
5433 Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5434 std::sort(Values.begin(), Values.end());
5436 // If the switch is already dense, there's nothing useful to do here.
5437 if (isSwitchDense(Values))
5440 // First, transform the values such that they start at zero and ascend.
5441 int64_t Base = Values[0];
5442 for (auto &V : Values)
5445 // Now we have signed numbers that have been shifted so that, given enough
5446 // precision, there are no negative values. Since the rest of the transform
5447 // is bitwise only, we switch now to an unsigned representation.
5449 for (auto &V : Values)
5450 GCD = GreatestCommonDivisor64(GCD, (uint64_t)V);
5452 // This transform can be done speculatively because it is so cheap - it results
5453 // in a single rotate operation being inserted. This can only happen if the
5454 // factor extracted is a power of 2.
5455 // FIXME: If the GCD is an odd number we can multiply by the multiplicative
5456 // inverse of GCD and then perform this transform.
5457 // FIXME: It's possible that optimizing a switch on powers of two might also
5458 // be beneficial - flag values are often powers of two and we could use a CLZ
5459 // as the key function.
5460 if (GCD <= 1 || !isPowerOf2_64(GCD))
5461 // No common divisor found or too expensive to compute key function.
5464 unsigned Shift = Log2_64(GCD);
5465 for (auto &V : Values)
5466 V = (int64_t)((uint64_t)V >> Shift);
5468 if (!isSwitchDense(Values))
5469 // Transform didn't create a dense switch.
5472 // The obvious transform is to shift the switch condition right and emit a
5473 // check that the condition actually cleanly divided by GCD, i.e.
5474 // C & (1 << Shift - 1) == 0
5475 // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5477 // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5478 // shift and puts the shifted-off bits in the uppermost bits. If any of these
5479 // are nonzero then the switch condition will be very large and will hit the
5482 auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5483 Builder.SetInsertPoint(SI);
5484 auto *ShiftC = ConstantInt::get(Ty, Shift);
5485 auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5486 auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5487 auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5488 auto *Rot = Builder.CreateOr(LShr, Shl);
5489 SI->replaceUsesOfWith(SI->getCondition(), Rot);
5491 for (SwitchInst::CaseIt C = SI->case_begin(), E = SI->case_end(); C != E;
5493 auto *Orig = C.getCaseValue();
5494 auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5496 cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5501 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
5502 BasicBlock *BB = SI->getParent();
5504 if (isValueEqualityComparison(SI)) {
5505 // If we only have one predecessor, and if it is a branch on this value,
5506 // see if that predecessor totally determines the outcome of this switch.
5507 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5508 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
5509 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5511 Value *Cond = SI->getCondition();
5512 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
5513 if (SimplifySwitchOnSelect(SI, Select))
5514 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5516 // If the block only contains the switch, see if we can fold the block
5517 // away into any preds.
5518 BasicBlock::iterator BBI = BB->begin();
5519 // Ignore dbg intrinsics.
5520 while (isa<DbgInfoIntrinsic>(BBI))
5523 if (FoldValueComparisonIntoPredecessors(SI, Builder))
5524 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5527 // Try to transform the switch into an icmp and a branch.
5528 if (TurnSwitchRangeIntoICmp(SI, Builder))
5529 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5531 // Remove unreachable cases.
5532 if (EliminateDeadSwitchCases(SI, AC, DL))
5533 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5535 if (SwitchToSelect(SI, Builder, AC, DL, TTI))
5536 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5538 if (ForwardSwitchConditionToPHI(SI))
5539 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5541 if (SwitchToLookupTable(SI, Builder, DL, TTI))
5542 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5544 if (ReduceSwitchRange(SI, Builder, DL, TTI))
5545 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5550 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
5551 BasicBlock *BB = IBI->getParent();
5552 bool Changed = false;
5554 // Eliminate redundant destinations.
5555 SmallPtrSet<Value *, 8> Succs;
5556 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
5557 BasicBlock *Dest = IBI->getDestination(i);
5558 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
5559 Dest->removePredecessor(BB);
5560 IBI->removeDestination(i);
5567 if (IBI->getNumDestinations() == 0) {
5568 // If the indirectbr has no successors, change it to unreachable.
5569 new UnreachableInst(IBI->getContext(), IBI);
5570 EraseTerminatorInstAndDCECond(IBI);
5574 if (IBI->getNumDestinations() == 1) {
5575 // If the indirectbr has one successor, change it to a direct branch.
5576 BranchInst::Create(IBI->getDestination(0), IBI);
5577 EraseTerminatorInstAndDCECond(IBI);
5581 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
5582 if (SimplifyIndirectBrOnSelect(IBI, SI))
5583 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5588 /// Given an block with only a single landing pad and a unconditional branch
5589 /// try to find another basic block which this one can be merged with. This
5590 /// handles cases where we have multiple invokes with unique landing pads, but
5591 /// a shared handler.
5593 /// We specifically choose to not worry about merging non-empty blocks
5594 /// here. That is a PRE/scheduling problem and is best solved elsewhere. In
5595 /// practice, the optimizer produces empty landing pad blocks quite frequently
5596 /// when dealing with exception dense code. (see: instcombine, gvn, if-else
5597 /// sinking in this file)
5599 /// This is primarily a code size optimization. We need to avoid performing
5600 /// any transform which might inhibit optimization (such as our ability to
5601 /// specialize a particular handler via tail commoning). We do this by not
5602 /// merging any blocks which require us to introduce a phi. Since the same
5603 /// values are flowing through both blocks, we don't loose any ability to
5604 /// specialize. If anything, we make such specialization more likely.
5606 /// TODO - This transformation could remove entries from a phi in the target
5607 /// block when the inputs in the phi are the same for the two blocks being
5608 /// merged. In some cases, this could result in removal of the PHI entirely.
5609 static bool TryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI,
5611 auto Succ = BB->getUniqueSuccessor();
5613 // If there's a phi in the successor block, we'd likely have to introduce
5614 // a phi into the merged landing pad block.
5615 if (isa<PHINode>(*Succ->begin()))
5618 for (BasicBlock *OtherPred : predecessors(Succ)) {
5619 if (BB == OtherPred)
5621 BasicBlock::iterator I = OtherPred->begin();
5622 LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(I);
5623 if (!LPad2 || !LPad2->isIdenticalTo(LPad))
5625 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {
5627 BranchInst *BI2 = dyn_cast<BranchInst>(I);
5628 if (!BI2 || !BI2->isIdenticalTo(BI))
5631 // We've found an identical block. Update our predecessors to take that
5632 // path instead and make ourselves dead.
5633 SmallSet<BasicBlock *, 16> Preds;
5634 Preds.insert(pred_begin(BB), pred_end(BB));
5635 for (BasicBlock *Pred : Preds) {
5636 InvokeInst *II = cast<InvokeInst>(Pred->getTerminator());
5637 assert(II->getNormalDest() != BB && II->getUnwindDest() == BB &&
5638 "unexpected successor");
5639 II->setUnwindDest(OtherPred);
5642 // The debug info in OtherPred doesn't cover the merged control flow that
5643 // used to go through BB. We need to delete it or update it.
5644 for (auto I = OtherPred->begin(), E = OtherPred->end(); I != E;) {
5645 Instruction &Inst = *I;
5647 if (isa<DbgInfoIntrinsic>(Inst))
5648 Inst.eraseFromParent();
5651 SmallSet<BasicBlock *, 16> Succs;
5652 Succs.insert(succ_begin(BB), succ_end(BB));
5653 for (BasicBlock *Succ : Succs) {
5654 Succ->removePredecessor(BB);
5657 IRBuilder<> Builder(BI);
5658 Builder.CreateUnreachable();
5659 BI->eraseFromParent();
5665 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI,
5666 IRBuilder<> &Builder) {
5667 BasicBlock *BB = BI->getParent();
5669 if (SinkCommon && SinkThenElseCodeToEnd(BI))
5672 // If the Terminator is the only non-phi instruction, simplify the block.
5673 // if LoopHeader is provided, check if the block is a loop header
5674 // (This is for early invocations before loop simplify and vectorization
5675 // to keep canonical loop forms for nested loops.
5676 // These blocks can be eliminated when the pass is invoked later
5677 // in the back-end.)
5678 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg()->getIterator();
5679 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
5680 (!LoopHeaders || !LoopHeaders->count(BB)) &&
5681 TryToSimplifyUncondBranchFromEmptyBlock(BB))
5684 // If the only instruction in the block is a seteq/setne comparison
5685 // against a constant, try to simplify the block.
5686 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
5687 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
5688 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5690 if (I->isTerminator() &&
5691 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, DL, TTI,
5692 BonusInstThreshold, AC))
5696 // See if we can merge an empty landing pad block with another which is
5698 if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(I)) {
5699 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {
5701 if (I->isTerminator() && TryToMergeLandingPad(LPad, BI, BB))
5705 // If this basic block is ONLY a compare and a branch, and if a predecessor
5706 // branches to us and our successor, fold the comparison into the
5707 // predecessor and use logical operations to update the incoming value
5708 // for PHI nodes in common successor.
5709 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
5710 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5714 static BasicBlock *allPredecessorsComeFromSameSource(BasicBlock *BB) {
5715 BasicBlock *PredPred = nullptr;
5716 for (auto *P : predecessors(BB)) {
5717 BasicBlock *PPred = P->getSinglePredecessor();
5718 if (!PPred || (PredPred && PredPred != PPred))
5725 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
5726 BasicBlock *BB = BI->getParent();
5728 // Conditional branch
5729 if (isValueEqualityComparison(BI)) {
5730 // If we only have one predecessor, and if it is a branch on this value,
5731 // see if that predecessor totally determines the outcome of this
5733 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5734 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
5735 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5737 // This block must be empty, except for the setcond inst, if it exists.
5738 // Ignore dbg intrinsics.
5739 BasicBlock::iterator I = BB->begin();
5740 // Ignore dbg intrinsics.
5741 while (isa<DbgInfoIntrinsic>(I))
5744 if (FoldValueComparisonIntoPredecessors(BI, Builder))
5745 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5746 } else if (&*I == cast<Instruction>(BI->getCondition())) {
5748 // Ignore dbg intrinsics.
5749 while (isa<DbgInfoIntrinsic>(I))
5751 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
5752 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5756 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
5757 if (SimplifyBranchOnICmpChain(BI, Builder, DL))
5760 // If this basic block has a single dominating predecessor block and the
5761 // dominating block's condition implies BI's condition, we know the direction
5762 // of the BI branch.
5763 if (BasicBlock *Dom = BB->getSinglePredecessor()) {
5764 auto *PBI = dyn_cast_or_null<BranchInst>(Dom->getTerminator());
5765 if (PBI && PBI->isConditional() &&
5766 PBI->getSuccessor(0) != PBI->getSuccessor(1) &&
5767 (PBI->getSuccessor(0) == BB || PBI->getSuccessor(1) == BB)) {
5768 bool CondIsFalse = PBI->getSuccessor(1) == BB;
5769 Optional<bool> Implication = isImpliedCondition(
5770 PBI->getCondition(), BI->getCondition(), DL, CondIsFalse);
5772 // Turn this into a branch on constant.
5773 auto *OldCond = BI->getCondition();
5774 ConstantInt *CI = *Implication
5775 ? ConstantInt::getTrue(BB->getContext())
5776 : ConstantInt::getFalse(BB->getContext());
5777 BI->setCondition(CI);
5778 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
5779 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5784 // If this basic block is ONLY a compare and a branch, and if a predecessor
5785 // branches to us and one of our successors, fold the comparison into the
5786 // predecessor and use logical operations to pick the right destination.
5787 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
5788 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5790 // We have a conditional branch to two blocks that are only reachable
5791 // from BI. We know that the condbr dominates the two blocks, so see if
5792 // there is any identical code in the "then" and "else" blocks. If so, we
5793 // can hoist it up to the branching block.
5794 if (BI->getSuccessor(0)->getSinglePredecessor()) {
5795 if (BI->getSuccessor(1)->getSinglePredecessor()) {
5796 if (HoistThenElseCodeToIf(BI, TTI))
5797 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5799 // If Successor #1 has multiple preds, we may be able to conditionally
5800 // execute Successor #0 if it branches to Successor #1.
5801 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
5802 if (Succ0TI->getNumSuccessors() == 1 &&
5803 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
5804 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), TTI))
5805 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5807 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
5808 // If Successor #0 has multiple preds, we may be able to conditionally
5809 // execute Successor #1 if it branches to Successor #0.
5810 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
5811 if (Succ1TI->getNumSuccessors() == 1 &&
5812 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
5813 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), TTI))
5814 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5817 // If this is a branch on a phi node in the current block, thread control
5818 // through this block if any PHI node entries are constants.
5819 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
5820 if (PN->getParent() == BI->getParent())
5821 if (FoldCondBranchOnPHI(BI, DL))
5822 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5824 // Scan predecessor blocks for conditional branches.
5825 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
5826 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
5827 if (PBI != BI && PBI->isConditional())
5828 if (SimplifyCondBranchToCondBranch(PBI, BI, DL))
5829 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5831 // Look for diamond patterns.
5832 if (MergeCondStores)
5833 if (BasicBlock *PrevBB = allPredecessorsComeFromSameSource(BB))
5834 if (BranchInst *PBI = dyn_cast<BranchInst>(PrevBB->getTerminator()))
5835 if (PBI != BI && PBI->isConditional())
5836 if (mergeConditionalStores(PBI, BI))
5837 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5842 /// Check if passing a value to an instruction will cause undefined behavior.
5843 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
5844 Constant *C = dyn_cast<Constant>(V);
5851 if (C->isNullValue() || isa<UndefValue>(C)) {
5852 // Only look at the first use, avoid hurting compile time with long uselists
5853 User *Use = *I->user_begin();
5855 // Now make sure that there are no instructions in between that can alter
5856 // control flow (eg. calls)
5857 for (BasicBlock::iterator
5858 i = ++BasicBlock::iterator(I),
5859 UI = BasicBlock::iterator(dyn_cast<Instruction>(Use));
5861 if (i == I->getParent()->end() || i->mayHaveSideEffects())
5864 // Look through GEPs. A load from a GEP derived from NULL is still undefined
5865 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
5866 if (GEP->getPointerOperand() == I)
5867 return passingValueIsAlwaysUndefined(V, GEP);
5869 // Look through bitcasts.
5870 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
5871 return passingValueIsAlwaysUndefined(V, BC);
5873 // Load from null is undefined.
5874 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
5875 if (!LI->isVolatile())
5876 return LI->getPointerAddressSpace() == 0;
5878 // Store to null is undefined.
5879 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
5880 if (!SI->isVolatile())
5881 return SI->getPointerAddressSpace() == 0 &&
5882 SI->getPointerOperand() == I;
5884 // A call to null is undefined.
5885 if (auto CS = CallSite(Use))
5886 return CS.getCalledValue() == I;
5891 /// If BB has an incoming value that will always trigger undefined behavior
5892 /// (eg. null pointer dereference), remove the branch leading here.
5893 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
5894 for (BasicBlock::iterator i = BB->begin();
5895 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
5896 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
5897 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
5898 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
5899 IRBuilder<> Builder(T);
5900 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
5901 BB->removePredecessor(PHI->getIncomingBlock(i));
5902 // Turn uncoditional branches into unreachables and remove the dead
5903 // destination from conditional branches.
5904 if (BI->isUnconditional())
5905 Builder.CreateUnreachable();
5907 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1)
5908 : BI->getSuccessor(0));
5909 BI->eraseFromParent();
5912 // TODO: SwitchInst.
5918 bool SimplifyCFGOpt::run(BasicBlock *BB) {
5919 bool Changed = false;
5921 assert(BB && BB->getParent() && "Block not embedded in function!");
5922 assert(BB->getTerminator() && "Degenerate basic block encountered!");
5924 // Remove basic blocks that have no predecessors (except the entry block)...
5925 // or that just have themself as a predecessor. These are unreachable.
5926 if ((pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) ||
5927 BB->getSinglePredecessor() == BB) {
5928 DEBUG(dbgs() << "Removing BB: \n" << *BB);
5929 DeleteDeadBlock(BB);
5933 // Check to see if we can constant propagate this terminator instruction
5935 Changed |= ConstantFoldTerminator(BB, true);
5937 // Check for and eliminate duplicate PHI nodes in this block.
5938 Changed |= EliminateDuplicatePHINodes(BB);
5940 // Check for and remove branches that will always cause undefined behavior.
5941 Changed |= removeUndefIntroducingPredecessor(BB);
5943 // Merge basic blocks into their predecessor if there is only one distinct
5944 // pred, and if there is only one distinct successor of the predecessor, and
5945 // if there are no PHI nodes.
5947 if (MergeBlockIntoPredecessor(BB))
5950 IRBuilder<> Builder(BB);
5952 // If there is a trivial two-entry PHI node in this basic block, and we can
5953 // eliminate it, do so now.
5954 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
5955 if (PN->getNumIncomingValues() == 2)
5956 Changed |= FoldTwoEntryPHINode(PN, TTI, DL);
5958 Builder.SetInsertPoint(BB->getTerminator());
5959 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
5960 if (BI->isUnconditional()) {
5961 if (SimplifyUncondBranch(BI, Builder))
5964 if (SimplifyCondBranch(BI, Builder))
5967 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
5968 if (SimplifyReturn(RI, Builder))
5970 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
5971 if (SimplifyResume(RI, Builder))
5973 } else if (CleanupReturnInst *RI =
5974 dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
5975 if (SimplifyCleanupReturn(RI))
5977 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
5978 if (SimplifySwitch(SI, Builder))
5980 } else if (UnreachableInst *UI =
5981 dyn_cast<UnreachableInst>(BB->getTerminator())) {
5982 if (SimplifyUnreachable(UI))
5984 } else if (IndirectBrInst *IBI =
5985 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
5986 if (SimplifyIndirectBr(IBI))
5993 /// This function is used to do simplification of a CFG.
5994 /// For example, it adjusts branches to branches to eliminate the extra hop,
5995 /// eliminates unreachable basic blocks, and does other "peephole" optimization
5996 /// of the CFG. It returns true if a modification was made.
5998 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
5999 unsigned BonusInstThreshold, AssumptionCache *AC,
6000 SmallPtrSetImpl<BasicBlock *> *LoopHeaders) {
6001 return SimplifyCFGOpt(TTI, BB->getModule()->getDataLayout(),
6002 BonusInstThreshold, AC, LoopHeaders)