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 // The debug location for the "common" instruction is the merged locations of
1578 // all the commoned instructions. We start with the original location of the
1579 // "common" instruction and iteratively merge each location in the loop below.
1580 DILocation *Loc = I0->getDebugLoc();
1582 // Update metadata and IR flags, and merge debug locations.
1583 for (auto *I : Insts)
1585 Loc = DILocation::getMergedLocation(Loc, I->getDebugLoc());
1586 combineMetadataForCSE(I0, I);
1589 if (!isa<CallInst>(I0))
1590 I0->setDebugLoc(Loc);
1592 if (!isa<StoreInst>(I0)) {
1593 // canSinkLastInstruction checked that all instructions were used by
1594 // one and only one PHI node. Find that now, RAUW it to our common
1595 // instruction and nuke it.
1596 assert(I0->hasOneUse());
1597 auto *PN = cast<PHINode>(*I0->user_begin());
1598 PN->replaceAllUsesWith(I0);
1599 PN->eraseFromParent();
1602 // Finally nuke all instructions apart from the common instruction.
1603 for (auto *I : Insts)
1605 I->eraseFromParent();
1612 // LockstepReverseIterator - Iterates through instructions
1613 // in a set of blocks in reverse order from the first non-terminator.
1614 // For example (assume all blocks have size n):
1615 // LockstepReverseIterator I([B1, B2, B3]);
1616 // *I-- = [B1[n], B2[n], B3[n]];
1617 // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1618 // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1620 class LockstepReverseIterator {
1621 ArrayRef<BasicBlock*> Blocks;
1622 SmallVector<Instruction*,4> Insts;
1625 LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) :
1633 for (auto *BB : Blocks) {
1634 Instruction *Inst = BB->getTerminator();
1635 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1636 Inst = Inst->getPrevNode();
1638 // Block wasn't big enough.
1642 Insts.push_back(Inst);
1646 bool isValid() const {
1650 void operator -- () {
1653 for (auto *&Inst : Insts) {
1654 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1655 Inst = Inst->getPrevNode();
1656 // Already at beginning of block.
1664 ArrayRef<Instruction*> operator * () const {
1669 } // end anonymous namespace
1671 /// Given an unconditional branch that goes to BBEnd,
1672 /// check whether BBEnd has only two predecessors and the other predecessor
1673 /// ends with an unconditional branch. If it is true, sink any common code
1674 /// in the two predecessors to BBEnd.
1675 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1676 assert(BI1->isUnconditional());
1677 BasicBlock *BBEnd = BI1->getSuccessor(0);
1679 // We support two situations:
1680 // (1) all incoming arcs are unconditional
1681 // (2) one incoming arc is conditional
1683 // (2) is very common in switch defaults and
1684 // else-if patterns;
1687 // else if (b) f(2);
1700 // [end] has two unconditional predecessor arcs and one conditional. The
1701 // conditional refers to the implicit empty 'else' arc. This conditional
1702 // arc can also be caused by an empty default block in a switch.
1704 // In this case, we attempt to sink code from all *unconditional* arcs.
1705 // If we can sink instructions from these arcs (determined during the scan
1706 // phase below) we insert a common successor for all unconditional arcs and
1707 // connect that to [end], to enable sinking:
1720 SmallVector<BasicBlock*,4> UnconditionalPreds;
1721 Instruction *Cond = nullptr;
1722 for (auto *B : predecessors(BBEnd)) {
1723 auto *T = B->getTerminator();
1724 if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1725 UnconditionalPreds.push_back(B);
1726 else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1731 if (UnconditionalPreds.size() < 2)
1734 bool Changed = false;
1735 // We take a two-step approach to tail sinking. First we scan from the end of
1736 // each block upwards in lockstep. If the n'th instruction from the end of each
1737 // block can be sunk, those instructions are added to ValuesToSink and we
1738 // carry on. If we can sink an instruction but need to PHI-merge some operands
1739 // (because they're not identical in each instruction) we add these to
1741 unsigned ScanIdx = 0;
1742 SmallPtrSet<Value*,4> InstructionsToSink;
1743 DenseMap<Instruction*, SmallVector<Value*,4>> PHIOperands;
1744 LockstepReverseIterator LRI(UnconditionalPreds);
1745 while (LRI.isValid() &&
1746 canSinkInstructions(*LRI, PHIOperands)) {
1747 DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0] << "\n");
1748 InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1753 auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1754 unsigned NumPHIdValues = 0;
1755 for (auto *I : *LRI)
1756 for (auto *V : PHIOperands[I])
1757 if (InstructionsToSink.count(V) == 0)
1759 DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1760 unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1761 if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1764 return NumPHIInsts <= 1;
1767 if (ScanIdx > 0 && Cond) {
1768 // Check if we would actually sink anything first! This mutates the CFG and
1769 // adds an extra block. The goal in doing this is to allow instructions that
1770 // couldn't be sunk before to be sunk - obviously, speculatable instructions
1771 // (such as trunc, add) can be sunk and predicated already. So we check that
1772 // we're going to sink at least one non-speculatable instruction.
1775 bool Profitable = false;
1776 while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1777 if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1787 DEBUG(dbgs() << "SINK: Splitting edge\n");
1788 // We have a conditional edge and we're going to sink some instructions.
1789 // Insert a new block postdominating all blocks we're going to sink from.
1790 if (!SplitBlockPredecessors(BI1->getSuccessor(0), UnconditionalPreds,
1792 // Edges couldn't be split.
1797 // Now that we've analyzed all potential sinking candidates, perform the
1798 // actual sink. We iteratively sink the last non-terminator of the source
1799 // blocks into their common successor unless doing so would require too
1800 // many PHI instructions to be generated (currently only one PHI is allowed
1801 // per sunk instruction).
1803 // We can use InstructionsToSink to discount values needing PHI-merging that will
1804 // actually be sunk in a later iteration. This allows us to be more
1805 // aggressive in what we sink. This does allow a false positive where we
1806 // sink presuming a later value will also be sunk, but stop half way through
1807 // and never actually sink it which means we produce more PHIs than intended.
1808 // This is unlikely in practice though.
1809 for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1810 DEBUG(dbgs() << "SINK: Sink: "
1811 << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1814 // Because we've sunk every instruction in turn, the current instruction to
1815 // sink is always at index 0.
1817 if (!ProfitableToSinkInstruction(LRI)) {
1818 // Too many PHIs would be created.
1819 DEBUG(dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1823 if (!sinkLastInstruction(UnconditionalPreds))
1831 /// \brief Determine if we can hoist sink a sole store instruction out of a
1832 /// conditional block.
1834 /// We are looking for code like the following:
1836 /// store i32 %add, i32* %arrayidx2
1837 /// ... // No other stores or function calls (we could be calling a memory
1838 /// ... // function).
1839 /// %cmp = icmp ult %x, %y
1840 /// br i1 %cmp, label %EndBB, label %ThenBB
1842 /// store i32 %add5, i32* %arrayidx2
1846 /// We are going to transform this into:
1848 /// store i32 %add, i32* %arrayidx2
1850 /// %cmp = icmp ult %x, %y
1851 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1852 /// store i32 %add.add5, i32* %arrayidx2
1855 /// \return The pointer to the value of the previous store if the store can be
1856 /// hoisted into the predecessor block. 0 otherwise.
1857 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1858 BasicBlock *StoreBB, BasicBlock *EndBB) {
1859 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1863 // Volatile or atomic.
1864 if (!StoreToHoist->isSimple())
1867 Value *StorePtr = StoreToHoist->getPointerOperand();
1869 // Look for a store to the same pointer in BrBB.
1870 unsigned MaxNumInstToLookAt = 9;
1871 for (Instruction &CurI : reverse(*BrBB)) {
1872 if (!MaxNumInstToLookAt)
1875 if (isa<DbgInfoIntrinsic>(CurI))
1877 --MaxNumInstToLookAt;
1879 // Could be calling an instruction that affects memory like free().
1880 if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1883 if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1884 // Found the previous store make sure it stores to the same location.
1885 if (SI->getPointerOperand() == StorePtr)
1886 // Found the previous store, return its value operand.
1887 return SI->getValueOperand();
1888 return nullptr; // Unknown store.
1895 /// \brief Speculate a conditional basic block flattening the CFG.
1897 /// Note that this is a very risky transform currently. Speculating
1898 /// instructions like this is most often not desirable. Instead, there is an MI
1899 /// pass which can do it with full awareness of the resource constraints.
1900 /// However, some cases are "obvious" and we should do directly. An example of
1901 /// this is speculating a single, reasonably cheap instruction.
1903 /// There is only one distinct advantage to flattening the CFG at the IR level:
1904 /// it makes very common but simplistic optimizations such as are common in
1905 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1906 /// modeling their effects with easier to reason about SSA value graphs.
1909 /// An illustration of this transform is turning this IR:
1912 /// %cmp = icmp ult %x, %y
1913 /// br i1 %cmp, label %EndBB, label %ThenBB
1915 /// %sub = sub %x, %y
1918 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1925 /// %cmp = icmp ult %x, %y
1926 /// %sub = sub %x, %y
1927 /// %cond = select i1 %cmp, 0, %sub
1931 /// \returns true if the conditional block is removed.
1932 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1933 const TargetTransformInfo &TTI) {
1934 // Be conservative for now. FP select instruction can often be expensive.
1935 Value *BrCond = BI->getCondition();
1936 if (isa<FCmpInst>(BrCond))
1939 BasicBlock *BB = BI->getParent();
1940 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1942 // If ThenBB is actually on the false edge of the conditional branch, remember
1943 // to swap the select operands later.
1944 bool Invert = false;
1945 if (ThenBB != BI->getSuccessor(0)) {
1946 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1949 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1951 // Keep a count of how many times instructions are used within CondBB when
1952 // they are candidates for sinking into CondBB. Specifically:
1953 // - They are defined in BB, and
1954 // - They have no side effects, and
1955 // - All of their uses are in CondBB.
1956 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1958 unsigned SpeculationCost = 0;
1959 Value *SpeculatedStoreValue = nullptr;
1960 StoreInst *SpeculatedStore = nullptr;
1961 for (BasicBlock::iterator BBI = ThenBB->begin(),
1962 BBE = std::prev(ThenBB->end());
1963 BBI != BBE; ++BBI) {
1964 Instruction *I = &*BBI;
1966 if (isa<DbgInfoIntrinsic>(I))
1969 // Only speculatively execute a single instruction (not counting the
1970 // terminator) for now.
1972 if (SpeculationCost > 1)
1975 // Don't hoist the instruction if it's unsafe or expensive.
1976 if (!isSafeToSpeculativelyExecute(I) &&
1977 !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1978 I, BB, ThenBB, EndBB))))
1980 if (!SpeculatedStoreValue &&
1981 ComputeSpeculationCost(I, TTI) >
1982 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
1985 // Store the store speculation candidate.
1986 if (SpeculatedStoreValue)
1987 SpeculatedStore = cast<StoreInst>(I);
1989 // Do not hoist the instruction if any of its operands are defined but not
1990 // used in BB. The transformation will prevent the operand from
1991 // being sunk into the use block.
1992 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
1993 Instruction *OpI = dyn_cast<Instruction>(*i);
1994 if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
1995 continue; // Not a candidate for sinking.
1997 ++SinkCandidateUseCounts[OpI];
2001 // Consider any sink candidates which are only used in CondBB as costs for
2002 // speculation. Note, while we iterate over a DenseMap here, we are summing
2003 // and so iteration order isn't significant.
2004 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator
2005 I = SinkCandidateUseCounts.begin(),
2006 E = SinkCandidateUseCounts.end();
2008 if (I->first->getNumUses() == I->second) {
2010 if (SpeculationCost > 1)
2014 // Check that the PHI nodes can be converted to selects.
2015 bool HaveRewritablePHIs = false;
2016 for (BasicBlock::iterator I = EndBB->begin();
2017 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2018 Value *OrigV = PN->getIncomingValueForBlock(BB);
2019 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
2021 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2022 // Skip PHIs which are trivial.
2026 // Don't convert to selects if we could remove undefined behavior instead.
2027 if (passingValueIsAlwaysUndefined(OrigV, PN) ||
2028 passingValueIsAlwaysUndefined(ThenV, PN))
2031 HaveRewritablePHIs = true;
2032 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2033 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2034 if (!OrigCE && !ThenCE)
2035 continue; // Known safe and cheap.
2037 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2038 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2040 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2041 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2043 2 * PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2044 if (OrigCost + ThenCost > MaxCost)
2047 // Account for the cost of an unfolded ConstantExpr which could end up
2048 // getting expanded into Instructions.
2049 // FIXME: This doesn't account for how many operations are combined in the
2050 // constant expression.
2052 if (SpeculationCost > 1)
2056 // If there are no PHIs to process, bail early. This helps ensure idempotence
2058 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2061 // If we get here, we can hoist the instruction and if-convert.
2062 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2064 // Insert a select of the value of the speculated store.
2065 if (SpeculatedStoreValue) {
2066 IRBuilder<NoFolder> Builder(BI);
2067 Value *TrueV = SpeculatedStore->getValueOperand();
2068 Value *FalseV = SpeculatedStoreValue;
2070 std::swap(TrueV, FalseV);
2071 Value *S = Builder.CreateSelect(
2072 BrCond, TrueV, FalseV, TrueV->getName() + "." + FalseV->getName(), BI);
2073 SpeculatedStore->setOperand(0, S);
2076 // Metadata can be dependent on the condition we are hoisting above.
2077 // Conservatively strip all metadata on the instruction.
2078 for (auto &I : *ThenBB)
2079 I.dropUnknownNonDebugMetadata();
2081 // Hoist the instructions.
2082 BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2083 ThenBB->begin(), std::prev(ThenBB->end()));
2085 // Insert selects and rewrite the PHI operands.
2086 IRBuilder<NoFolder> Builder(BI);
2087 for (BasicBlock::iterator I = EndBB->begin();
2088 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2089 unsigned OrigI = PN->getBasicBlockIndex(BB);
2090 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
2091 Value *OrigV = PN->getIncomingValue(OrigI);
2092 Value *ThenV = PN->getIncomingValue(ThenI);
2094 // Skip PHIs which are trivial.
2098 // Create a select whose true value is the speculatively executed value and
2099 // false value is the preexisting value. Swap them if the branch
2100 // destinations were inverted.
2101 Value *TrueV = ThenV, *FalseV = OrigV;
2103 std::swap(TrueV, FalseV);
2104 Value *V = Builder.CreateSelect(
2105 BrCond, TrueV, FalseV, TrueV->getName() + "." + FalseV->getName(), BI);
2106 PN->setIncomingValue(OrigI, V);
2107 PN->setIncomingValue(ThenI, V);
2114 /// Return true if we can thread a branch across this block.
2115 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
2116 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
2119 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2120 if (isa<DbgInfoIntrinsic>(BBI))
2123 return false; // Don't clone large BB's.
2126 // We can only support instructions that do not define values that are
2127 // live outside of the current basic block.
2128 for (User *U : BBI->users()) {
2129 Instruction *UI = cast<Instruction>(U);
2130 if (UI->getParent() != BB || isa<PHINode>(UI))
2134 // Looks ok, continue checking.
2140 /// If we have a conditional branch on a PHI node value that is defined in the
2141 /// same block as the branch and if any PHI entries are constants, thread edges
2142 /// corresponding to that entry to be branches to their ultimate destination.
2143 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL) {
2144 BasicBlock *BB = BI->getParent();
2145 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2146 // NOTE: we currently cannot transform this case if the PHI node is used
2147 // outside of the block.
2148 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2151 // Degenerate case of a single entry PHI.
2152 if (PN->getNumIncomingValues() == 1) {
2153 FoldSingleEntryPHINodes(PN->getParent());
2157 // Now we know that this block has multiple preds and two succs.
2158 if (!BlockIsSimpleEnoughToThreadThrough(BB))
2161 // Can't fold blocks that contain noduplicate or convergent calls.
2162 if (any_of(*BB, [](const Instruction &I) {
2163 const CallInst *CI = dyn_cast<CallInst>(&I);
2164 return CI && (CI->cannotDuplicate() || CI->isConvergent());
2168 // Okay, this is a simple enough basic block. See if any phi values are
2170 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2171 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
2172 if (!CB || !CB->getType()->isIntegerTy(1))
2175 // Okay, we now know that all edges from PredBB should be revectored to
2176 // branch to RealDest.
2177 BasicBlock *PredBB = PN->getIncomingBlock(i);
2178 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2181 continue; // Skip self loops.
2182 // Skip if the predecessor's terminator is an indirect branch.
2183 if (isa<IndirectBrInst>(PredBB->getTerminator()))
2186 // The dest block might have PHI nodes, other predecessors and other
2187 // difficult cases. Instead of being smart about this, just insert a new
2188 // block that jumps to the destination block, effectively splitting
2189 // the edge we are about to create.
2190 BasicBlock *EdgeBB =
2191 BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2192 RealDest->getParent(), RealDest);
2193 BranchInst::Create(RealDest, EdgeBB);
2195 // Update PHI nodes.
2196 AddPredecessorToBlock(RealDest, EdgeBB, BB);
2198 // BB may have instructions that are being threaded over. Clone these
2199 // instructions into EdgeBB. We know that there will be no uses of the
2200 // cloned instructions outside of EdgeBB.
2201 BasicBlock::iterator InsertPt = EdgeBB->begin();
2202 DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2203 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2204 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2205 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2208 // Clone the instruction.
2209 Instruction *N = BBI->clone();
2211 N->setName(BBI->getName() + ".c");
2213 // Update operands due to translation.
2214 for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2215 DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2216 if (PI != TranslateMap.end())
2220 // Check for trivial simplification.
2221 if (Value *V = SimplifyInstruction(N, DL)) {
2222 if (!BBI->use_empty())
2223 TranslateMap[&*BBI] = V;
2224 if (!N->mayHaveSideEffects()) {
2225 delete N; // Instruction folded away, don't need actual inst
2229 if (!BBI->use_empty())
2230 TranslateMap[&*BBI] = N;
2232 // Insert the new instruction into its new home.
2234 EdgeBB->getInstList().insert(InsertPt, N);
2237 // Loop over all of the edges from PredBB to BB, changing them to branch
2238 // to EdgeBB instead.
2239 TerminatorInst *PredBBTI = PredBB->getTerminator();
2240 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2241 if (PredBBTI->getSuccessor(i) == BB) {
2242 BB->removePredecessor(PredBB);
2243 PredBBTI->setSuccessor(i, EdgeBB);
2246 // Recurse, simplifying any other constants.
2247 return FoldCondBranchOnPHI(BI, DL) | true;
2253 /// Given a BB that starts with the specified two-entry PHI node,
2254 /// see if we can eliminate it.
2255 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI,
2256 const DataLayout &DL) {
2257 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2258 // statement", which has a very simple dominance structure. Basically, we
2259 // are trying to find the condition that is being branched on, which
2260 // subsequently causes this merge to happen. We really want control
2261 // dependence information for this check, but simplifycfg can't keep it up
2262 // to date, and this catches most of the cases we care about anyway.
2263 BasicBlock *BB = PN->getParent();
2264 BasicBlock *IfTrue, *IfFalse;
2265 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2267 // Don't bother if the branch will be constant folded trivially.
2268 isa<ConstantInt>(IfCond))
2271 // Okay, we found that we can merge this two-entry phi node into a select.
2272 // Doing so would require us to fold *all* two entry phi nodes in this block.
2273 // At some point this becomes non-profitable (particularly if the target
2274 // doesn't support cmov's). Only do this transformation if there are two or
2275 // fewer PHI nodes in this block.
2276 unsigned NumPhis = 0;
2277 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2281 // Loop over the PHI's seeing if we can promote them all to select
2282 // instructions. While we are at it, keep track of the instructions
2283 // that need to be moved to the dominating block.
2284 SmallPtrSet<Instruction *, 4> AggressiveInsts;
2285 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2286 MaxCostVal1 = PHINodeFoldingThreshold;
2287 MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2288 MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2290 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2291 PHINode *PN = cast<PHINode>(II++);
2292 if (Value *V = SimplifyInstruction(PN, DL)) {
2293 PN->replaceAllUsesWith(V);
2294 PN->eraseFromParent();
2298 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
2299 MaxCostVal0, TTI) ||
2300 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
2305 // If we folded the first phi, PN dangles at this point. Refresh it. If
2306 // we ran out of PHIs then we simplified them all.
2307 PN = dyn_cast<PHINode>(BB->begin());
2311 // Don't fold i1 branches on PHIs which contain binary operators. These can
2312 // often be turned into switches and other things.
2313 if (PN->getType()->isIntegerTy(1) &&
2314 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2315 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2316 isa<BinaryOperator>(IfCond)))
2319 // If all PHI nodes are promotable, check to make sure that all instructions
2320 // in the predecessor blocks can be promoted as well. If not, we won't be able
2321 // to get rid of the control flow, so it's not worth promoting to select
2323 BasicBlock *DomBlock = nullptr;
2324 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2325 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2326 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2329 DomBlock = *pred_begin(IfBlock1);
2330 for (BasicBlock::iterator I = IfBlock1->begin(); !isa<TerminatorInst>(I);
2332 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2333 // This is not an aggressive instruction that we can promote.
2334 // Because of this, we won't be able to get rid of the control flow, so
2335 // the xform is not worth it.
2340 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2343 DomBlock = *pred_begin(IfBlock2);
2344 for (BasicBlock::iterator I = IfBlock2->begin(); !isa<TerminatorInst>(I);
2346 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2347 // This is not an aggressive instruction that we can promote.
2348 // Because of this, we won't be able to get rid of the control flow, so
2349 // the xform is not worth it.
2354 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
2355 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
2357 // If we can still promote the PHI nodes after this gauntlet of tests,
2358 // do all of the PHI's now.
2359 Instruction *InsertPt = DomBlock->getTerminator();
2360 IRBuilder<NoFolder> Builder(InsertPt);
2362 // Move all 'aggressive' instructions, which are defined in the
2363 // conditional parts of the if's up to the dominating block.
2365 for (auto &I : *IfBlock1)
2366 I.dropUnknownNonDebugMetadata();
2367 DomBlock->getInstList().splice(InsertPt->getIterator(),
2368 IfBlock1->getInstList(), IfBlock1->begin(),
2369 IfBlock1->getTerminator()->getIterator());
2372 for (auto &I : *IfBlock2)
2373 I.dropUnknownNonDebugMetadata();
2374 DomBlock->getInstList().splice(InsertPt->getIterator(),
2375 IfBlock2->getInstList(), IfBlock2->begin(),
2376 IfBlock2->getTerminator()->getIterator());
2379 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2380 // Change the PHI node into a select instruction.
2381 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2382 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2384 Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2385 PN->replaceAllUsesWith(Sel);
2387 PN->eraseFromParent();
2390 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2391 // has been flattened. Change DomBlock to jump directly to our new block to
2392 // avoid other simplifycfg's kicking in on the diamond.
2393 TerminatorInst *OldTI = DomBlock->getTerminator();
2394 Builder.SetInsertPoint(OldTI);
2395 Builder.CreateBr(BB);
2396 OldTI->eraseFromParent();
2400 /// If we found a conditional branch that goes to two returning blocks,
2401 /// try to merge them together into one return,
2402 /// introducing a select if the return values disagree.
2403 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
2404 IRBuilder<> &Builder) {
2405 assert(BI->isConditional() && "Must be a conditional branch");
2406 BasicBlock *TrueSucc = BI->getSuccessor(0);
2407 BasicBlock *FalseSucc = BI->getSuccessor(1);
2408 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2409 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2411 // Check to ensure both blocks are empty (just a return) or optionally empty
2412 // with PHI nodes. If there are other instructions, merging would cause extra
2413 // computation on one path or the other.
2414 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2416 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2419 Builder.SetInsertPoint(BI);
2420 // Okay, we found a branch that is going to two return nodes. If
2421 // there is no return value for this function, just change the
2422 // branch into a return.
2423 if (FalseRet->getNumOperands() == 0) {
2424 TrueSucc->removePredecessor(BI->getParent());
2425 FalseSucc->removePredecessor(BI->getParent());
2426 Builder.CreateRetVoid();
2427 EraseTerminatorInstAndDCECond(BI);
2431 // Otherwise, figure out what the true and false return values are
2432 // so we can insert a new select instruction.
2433 Value *TrueValue = TrueRet->getReturnValue();
2434 Value *FalseValue = FalseRet->getReturnValue();
2436 // Unwrap any PHI nodes in the return blocks.
2437 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2438 if (TVPN->getParent() == TrueSucc)
2439 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2440 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2441 if (FVPN->getParent() == FalseSucc)
2442 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2444 // In order for this transformation to be safe, we must be able to
2445 // unconditionally execute both operands to the return. This is
2446 // normally the case, but we could have a potentially-trapping
2447 // constant expression that prevents this transformation from being
2449 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2452 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2456 // Okay, we collected all the mapped values and checked them for sanity, and
2457 // defined to really do this transformation. First, update the CFG.
2458 TrueSucc->removePredecessor(BI->getParent());
2459 FalseSucc->removePredecessor(BI->getParent());
2461 // Insert select instructions where needed.
2462 Value *BrCond = BI->getCondition();
2464 // Insert a select if the results differ.
2465 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2466 } else if (isa<UndefValue>(TrueValue)) {
2467 TrueValue = FalseValue;
2470 Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2475 !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2479 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2480 << "\n " << *BI << "NewRet = " << *RI
2481 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: " << *FalseSucc);
2483 EraseTerminatorInstAndDCECond(BI);
2488 /// Return true if the given instruction is available
2489 /// in its predecessor block. If yes, the instruction will be removed.
2490 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
2491 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2493 for (Instruction &I : *PB) {
2494 Instruction *PBI = &I;
2495 // Check whether Inst and PBI generate the same value.
2496 if (Inst->isIdenticalTo(PBI)) {
2497 Inst->replaceAllUsesWith(PBI);
2498 Inst->eraseFromParent();
2505 /// Return true if either PBI or BI has branch weight available, and store
2506 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2507 /// not have branch weight, use 1:1 as its weight.
2508 static bool extractPredSuccWeights(BranchInst *PBI, BranchInst *BI,
2509 uint64_t &PredTrueWeight,
2510 uint64_t &PredFalseWeight,
2511 uint64_t &SuccTrueWeight,
2512 uint64_t &SuccFalseWeight) {
2513 bool PredHasWeights =
2514 PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2515 bool SuccHasWeights =
2516 BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2517 if (PredHasWeights || SuccHasWeights) {
2518 if (!PredHasWeights)
2519 PredTrueWeight = PredFalseWeight = 1;
2520 if (!SuccHasWeights)
2521 SuccTrueWeight = SuccFalseWeight = 1;
2528 /// If this basic block is simple enough, and if a predecessor branches to us
2529 /// and one of our successors, fold the block into the predecessor and use
2530 /// logical operations to pick the right destination.
2531 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2532 BasicBlock *BB = BI->getParent();
2534 Instruction *Cond = nullptr;
2535 if (BI->isConditional())
2536 Cond = dyn_cast<Instruction>(BI->getCondition());
2538 // For unconditional branch, check for a simple CFG pattern, where
2539 // BB has a single predecessor and BB's successor is also its predecessor's
2540 // successor. If such pattern exisits, check for CSE between BB and its
2542 if (BasicBlock *PB = BB->getSinglePredecessor())
2543 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2544 if (PBI->isConditional() &&
2545 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2546 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2547 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2548 Instruction *Curr = &*I++;
2549 if (isa<CmpInst>(Curr)) {
2553 // Quit if we can't remove this instruction.
2554 if (!checkCSEInPredecessor(Curr, PB))
2563 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2564 Cond->getParent() != BB || !Cond->hasOneUse())
2567 // Make sure the instruction after the condition is the cond branch.
2568 BasicBlock::iterator CondIt = ++Cond->getIterator();
2570 // Ignore dbg intrinsics.
2571 while (isa<DbgInfoIntrinsic>(CondIt))
2577 // Only allow this transformation if computing the condition doesn't involve
2578 // too many instructions and these involved instructions can be executed
2579 // unconditionally. We denote all involved instructions except the condition
2580 // as "bonus instructions", and only allow this transformation when the
2581 // number of the bonus instructions does not exceed a certain threshold.
2582 unsigned NumBonusInsts = 0;
2583 for (auto I = BB->begin(); Cond != &*I; ++I) {
2584 // Ignore dbg intrinsics.
2585 if (isa<DbgInfoIntrinsic>(I))
2587 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2589 // I has only one use and can be executed unconditionally.
2590 Instruction *User = dyn_cast<Instruction>(I->user_back());
2591 if (User == nullptr || User->getParent() != BB)
2593 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2594 // to use any other instruction, User must be an instruction between next(I)
2597 // Early exits once we reach the limit.
2598 if (NumBonusInsts > BonusInstThreshold)
2602 // Cond is known to be a compare or binary operator. Check to make sure that
2603 // neither operand is a potentially-trapping constant expression.
2604 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2607 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2611 // Finally, don't infinitely unroll conditional loops.
2612 BasicBlock *TrueDest = BI->getSuccessor(0);
2613 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2614 if (TrueDest == BB || FalseDest == BB)
2617 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2618 BasicBlock *PredBlock = *PI;
2619 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2621 // Check that we have two conditional branches. If there is a PHI node in
2622 // the common successor, verify that the same value flows in from both
2624 SmallVector<PHINode *, 4> PHIs;
2625 if (!PBI || PBI->isUnconditional() ||
2626 (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2627 (!BI->isConditional() &&
2628 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2631 // Determine if the two branches share a common destination.
2632 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2633 bool InvertPredCond = false;
2635 if (BI->isConditional()) {
2636 if (PBI->getSuccessor(0) == TrueDest) {
2637 Opc = Instruction::Or;
2638 } else if (PBI->getSuccessor(1) == FalseDest) {
2639 Opc = Instruction::And;
2640 } else if (PBI->getSuccessor(0) == FalseDest) {
2641 Opc = Instruction::And;
2642 InvertPredCond = true;
2643 } else if (PBI->getSuccessor(1) == TrueDest) {
2644 Opc = Instruction::Or;
2645 InvertPredCond = true;
2650 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2654 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2655 IRBuilder<> Builder(PBI);
2657 // If we need to invert the condition in the pred block to match, do so now.
2658 if (InvertPredCond) {
2659 Value *NewCond = PBI->getCondition();
2661 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2662 CmpInst *CI = cast<CmpInst>(NewCond);
2663 CI->setPredicate(CI->getInversePredicate());
2666 Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2669 PBI->setCondition(NewCond);
2670 PBI->swapSuccessors();
2673 // If we have bonus instructions, clone them into the predecessor block.
2674 // Note that there may be multiple predecessor blocks, so we cannot move
2675 // bonus instructions to a predecessor block.
2676 ValueToValueMapTy VMap; // maps original values to cloned values
2677 // We already make sure Cond is the last instruction before BI. Therefore,
2678 // all instructions before Cond other than DbgInfoIntrinsic are bonus
2680 for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2681 if (isa<DbgInfoIntrinsic>(BonusInst))
2683 Instruction *NewBonusInst = BonusInst->clone();
2684 RemapInstruction(NewBonusInst, VMap,
2685 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2686 VMap[&*BonusInst] = NewBonusInst;
2688 // If we moved a load, we cannot any longer claim any knowledge about
2689 // its potential value. The previous information might have been valid
2690 // only given the branch precondition.
2691 // For an analogous reason, we must also drop all the metadata whose
2692 // semantics we don't understand.
2693 NewBonusInst->dropUnknownNonDebugMetadata();
2695 PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2696 NewBonusInst->takeName(&*BonusInst);
2697 BonusInst->setName(BonusInst->getName() + ".old");
2700 // Clone Cond into the predecessor basic block, and or/and the
2701 // two conditions together.
2702 Instruction *New = Cond->clone();
2703 RemapInstruction(New, VMap,
2704 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2705 PredBlock->getInstList().insert(PBI->getIterator(), New);
2706 New->takeName(Cond);
2707 Cond->setName(New->getName() + ".old");
2709 if (BI->isConditional()) {
2710 Instruction *NewCond = cast<Instruction>(
2711 Builder.CreateBinOp(Opc, PBI->getCondition(), New, "or.cond"));
2712 PBI->setCondition(NewCond);
2714 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2716 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2717 SuccTrueWeight, SuccFalseWeight);
2718 SmallVector<uint64_t, 8> NewWeights;
2720 if (PBI->getSuccessor(0) == BB) {
2722 // PBI: br i1 %x, BB, FalseDest
2723 // BI: br i1 %y, TrueDest, FalseDest
2724 // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2725 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2726 // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2727 // TrueWeight for PBI * FalseWeight for BI.
2728 // We assume that total weights of a BranchInst can fit into 32 bits.
2729 // Therefore, we will not have overflow using 64-bit arithmetic.
2730 NewWeights.push_back(PredFalseWeight *
2731 (SuccFalseWeight + SuccTrueWeight) +
2732 PredTrueWeight * SuccFalseWeight);
2734 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2735 PBI->setSuccessor(0, TrueDest);
2737 if (PBI->getSuccessor(1) == BB) {
2739 // PBI: br i1 %x, TrueDest, BB
2740 // BI: br i1 %y, TrueDest, FalseDest
2741 // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2742 // FalseWeight for PBI * TrueWeight for BI.
2743 NewWeights.push_back(PredTrueWeight *
2744 (SuccFalseWeight + SuccTrueWeight) +
2745 PredFalseWeight * SuccTrueWeight);
2746 // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2747 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2749 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2750 PBI->setSuccessor(1, FalseDest);
2752 if (NewWeights.size() == 2) {
2753 // Halve the weights if any of them cannot fit in an uint32_t
2754 FitWeights(NewWeights);
2756 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2759 LLVMContext::MD_prof,
2760 MDBuilder(BI->getContext()).createBranchWeights(MDWeights));
2762 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2764 // Update PHI nodes in the common successors.
2765 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2766 ConstantInt *PBI_C = cast<ConstantInt>(
2767 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2768 assert(PBI_C->getType()->isIntegerTy(1));
2769 Instruction *MergedCond = nullptr;
2770 if (PBI->getSuccessor(0) == TrueDest) {
2771 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2772 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2773 // is false: !PBI_Cond and BI_Value
2774 Instruction *NotCond = cast<Instruction>(
2775 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2776 MergedCond = cast<Instruction>(
2777 Builder.CreateBinOp(Instruction::And, NotCond, New, "and.cond"));
2779 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2780 Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2782 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2783 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2784 // is false: PBI_Cond and BI_Value
2785 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2786 Instruction::And, PBI->getCondition(), New, "and.cond"));
2787 if (PBI_C->isOne()) {
2788 Instruction *NotCond = cast<Instruction>(
2789 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2790 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2791 Instruction::Or, NotCond, MergedCond, "or.cond"));
2795 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2798 // Change PBI from Conditional to Unconditional.
2799 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2800 EraseTerminatorInstAndDCECond(PBI);
2804 // If BI was a loop latch, it may have had associated loop metadata.
2805 // We need to copy it to the new latch, that is, PBI.
2806 if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2807 PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2809 // TODO: If BB is reachable from all paths through PredBlock, then we
2810 // could replace PBI's branch probabilities with BI's.
2812 // Copy any debug value intrinsics into the end of PredBlock.
2813 for (Instruction &I : *BB)
2814 if (isa<DbgInfoIntrinsic>(I))
2815 I.clone()->insertBefore(PBI);
2822 // If there is only one store in BB1 and BB2, return it, otherwise return
2824 static StoreInst *findUniqueStoreInBlocks(BasicBlock *BB1, BasicBlock *BB2) {
2825 StoreInst *S = nullptr;
2826 for (auto *BB : {BB1, BB2}) {
2830 if (auto *SI = dyn_cast<StoreInst>(&I)) {
2832 // Multiple stores seen.
2841 static Value *ensureValueAvailableInSuccessor(Value *V, BasicBlock *BB,
2842 Value *AlternativeV = nullptr) {
2843 // PHI is going to be a PHI node that allows the value V that is defined in
2844 // BB to be referenced in BB's only successor.
2846 // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2847 // doesn't matter to us what the other operand is (it'll never get used). We
2848 // could just create a new PHI with an undef incoming value, but that could
2849 // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2850 // other PHI. So here we directly look for some PHI in BB's successor with V
2851 // as an incoming operand. If we find one, we use it, else we create a new
2854 // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2855 // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2856 // where OtherBB is the single other predecessor of BB's only successor.
2857 PHINode *PHI = nullptr;
2858 BasicBlock *Succ = BB->getSingleSuccessor();
2860 for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2861 if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2862 PHI = cast<PHINode>(I);
2866 assert(std::distance(pred_begin(Succ), pred_end(Succ)) == 2);
2867 auto PredI = pred_begin(Succ);
2868 BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2869 if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2876 // If V is not an instruction defined in BB, just return it.
2877 if (!AlternativeV &&
2878 (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2881 PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2882 PHI->addIncoming(V, BB);
2883 for (BasicBlock *PredBB : predecessors(Succ))
2886 AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2890 static bool mergeConditionalStoreToAddress(BasicBlock *PTB, BasicBlock *PFB,
2891 BasicBlock *QTB, BasicBlock *QFB,
2892 BasicBlock *PostBB, Value *Address,
2893 bool InvertPCond, bool InvertQCond) {
2894 auto IsaBitcastOfPointerType = [](const Instruction &I) {
2895 return Operator::getOpcode(&I) == Instruction::BitCast &&
2896 I.getType()->isPointerTy();
2899 // If we're not in aggressive mode, we only optimize if we have some
2900 // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2901 auto IsWorthwhile = [&](BasicBlock *BB) {
2904 // Heuristic: if the block can be if-converted/phi-folded and the
2905 // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2906 // thread this store.
2908 for (auto &I : *BB) {
2909 // Cheap instructions viable for folding.
2910 if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2913 // Free instructions.
2914 else if (isa<TerminatorInst>(I) || isa<DbgInfoIntrinsic>(I) ||
2915 IsaBitcastOfPointerType(I))
2920 return N <= PHINodeFoldingThreshold;
2923 if (!MergeCondStoresAggressively &&
2924 (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2925 !IsWorthwhile(QFB)))
2928 // For every pointer, there must be exactly two stores, one coming from
2929 // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2930 // store (to any address) in PTB,PFB or QTB,QFB.
2931 // FIXME: We could relax this restriction with a bit more work and performance
2933 StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2934 StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2935 if (!PStore || !QStore)
2938 // Now check the stores are compatible.
2939 if (!QStore->isUnordered() || !PStore->isUnordered())
2942 // Check that sinking the store won't cause program behavior changes. Sinking
2943 // the store out of the Q blocks won't change any behavior as we're sinking
2944 // from a block to its unconditional successor. But we're moving a store from
2945 // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2946 // So we need to check that there are no aliasing loads or stores in
2947 // QBI, QTB and QFB. We also need to check there are no conflicting memory
2948 // operations between PStore and the end of its parent block.
2950 // The ideal way to do this is to query AliasAnalysis, but we don't
2951 // preserve AA currently so that is dangerous. Be super safe and just
2952 // check there are no other memory operations at all.
2953 for (auto &I : *QFB->getSinglePredecessor())
2954 if (I.mayReadOrWriteMemory())
2956 for (auto &I : *QFB)
2957 if (&I != QStore && I.mayReadOrWriteMemory())
2960 for (auto &I : *QTB)
2961 if (&I != QStore && I.mayReadOrWriteMemory())
2963 for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2965 if (&*I != PStore && I->mayReadOrWriteMemory())
2968 // OK, we're going to sink the stores to PostBB. The store has to be
2969 // conditional though, so first create the predicate.
2970 Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
2972 Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
2975 Value *PPHI = ensureValueAvailableInSuccessor(PStore->getValueOperand(),
2976 PStore->getParent());
2977 Value *QPHI = ensureValueAvailableInSuccessor(QStore->getValueOperand(),
2978 QStore->getParent(), PPHI);
2980 IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
2982 Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
2983 Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
2986 PPred = QB.CreateNot(PPred);
2988 QPred = QB.CreateNot(QPred);
2989 Value *CombinedPred = QB.CreateOr(PPred, QPred);
2992 SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
2993 QB.SetInsertPoint(T);
2994 StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
2996 PStore->getAAMetadata(AAMD, /*Merge=*/false);
2997 PStore->getAAMetadata(AAMD, /*Merge=*/true);
2998 SI->setAAMetadata(AAMD);
3000 QStore->eraseFromParent();
3001 PStore->eraseFromParent();
3006 static bool mergeConditionalStores(BranchInst *PBI, BranchInst *QBI) {
3007 // The intention here is to find diamonds or triangles (see below) where each
3008 // conditional block contains a store to the same address. Both of these
3009 // stores are conditional, so they can't be unconditionally sunk. But it may
3010 // be profitable to speculatively sink the stores into one merged store at the
3011 // end, and predicate the merged store on the union of the two conditions of
3014 // This can reduce the number of stores executed if both of the conditions are
3015 // true, and can allow the blocks to become small enough to be if-converted.
3016 // This optimization will also chain, so that ladders of test-and-set
3017 // sequences can be if-converted away.
3019 // We only deal with simple diamonds or triangles:
3021 // PBI or PBI or a combination of the two
3031 // We model triangles as a type of diamond with a nullptr "true" block.
3032 // Triangles are canonicalized so that the fallthrough edge is represented by
3033 // a true condition, as in the diagram above.
3035 BasicBlock *PTB = PBI->getSuccessor(0);
3036 BasicBlock *PFB = PBI->getSuccessor(1);
3037 BasicBlock *QTB = QBI->getSuccessor(0);
3038 BasicBlock *QFB = QBI->getSuccessor(1);
3039 BasicBlock *PostBB = QFB->getSingleSuccessor();
3041 bool InvertPCond = false, InvertQCond = false;
3042 // Canonicalize fallthroughs to the true branches.
3043 if (PFB == QBI->getParent()) {
3044 std::swap(PFB, PTB);
3047 if (QFB == PostBB) {
3048 std::swap(QFB, QTB);
3052 // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3053 // and QFB may not. Model fallthroughs as a nullptr block.
3054 if (PTB == QBI->getParent())
3059 // Legality bailouts. We must have at least the non-fallthrough blocks and
3060 // the post-dominating block, and the non-fallthroughs must only have one
3062 auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3063 return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3066 !HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3067 !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3069 if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3070 (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3072 if (PostBB->getNumUses() != 2 || QBI->getParent()->getNumUses() != 2)
3075 // OK, this is a sequence of two diamonds or triangles.
3076 // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3077 SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3078 for (auto *BB : {PTB, PFB}) {
3082 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3083 PStoreAddresses.insert(SI->getPointerOperand());
3085 for (auto *BB : {QTB, QFB}) {
3089 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3090 QStoreAddresses.insert(SI->getPointerOperand());
3093 set_intersect(PStoreAddresses, QStoreAddresses);
3094 // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3095 // clear what it contains.
3096 auto &CommonAddresses = PStoreAddresses;
3098 bool Changed = false;
3099 for (auto *Address : CommonAddresses)
3100 Changed |= mergeConditionalStoreToAddress(
3101 PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond);
3105 /// If we have a conditional branch as a predecessor of another block,
3106 /// this function tries to simplify it. We know
3107 /// that PBI and BI are both conditional branches, and BI is in one of the
3108 /// successor blocks of PBI - PBI branches to BI.
3109 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
3110 const DataLayout &DL) {
3111 assert(PBI->isConditional() && BI->isConditional());
3112 BasicBlock *BB = BI->getParent();
3114 // If this block ends with a branch instruction, and if there is a
3115 // predecessor that ends on a branch of the same condition, make
3116 // this conditional branch redundant.
3117 if (PBI->getCondition() == BI->getCondition() &&
3118 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3119 // Okay, the outcome of this conditional branch is statically
3120 // knowable. If this block had a single pred, handle specially.
3121 if (BB->getSinglePredecessor()) {
3122 // Turn this into a branch on constant.
3123 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3125 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3126 return true; // Nuke the branch on constant.
3129 // Otherwise, if there are multiple predecessors, insert a PHI that merges
3130 // in the constant and simplify the block result. Subsequent passes of
3131 // simplifycfg will thread the block.
3132 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
3133 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3134 PHINode *NewPN = PHINode::Create(
3135 Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3136 BI->getCondition()->getName() + ".pr", &BB->front());
3137 // Okay, we're going to insert the PHI node. Since PBI is not the only
3138 // predecessor, compute the PHI'd conditional value for all of the preds.
3139 // Any predecessor where the condition is not computable we keep symbolic.
3140 for (pred_iterator PI = PB; PI != PE; ++PI) {
3141 BasicBlock *P = *PI;
3142 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3143 PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3144 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3145 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3147 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3150 NewPN->addIncoming(BI->getCondition(), P);
3154 BI->setCondition(NewPN);
3159 if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3163 // If both branches are conditional and both contain stores to the same
3164 // address, remove the stores from the conditionals and create a conditional
3165 // merged store at the end.
3166 if (MergeCondStores && mergeConditionalStores(PBI, BI))
3169 // If this is a conditional branch in an empty block, and if any
3170 // predecessors are a conditional branch to one of our destinations,
3171 // fold the conditions into logical ops and one cond br.
3172 BasicBlock::iterator BBI = BB->begin();
3173 // Ignore dbg intrinsics.
3174 while (isa<DbgInfoIntrinsic>(BBI))
3180 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3183 } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3186 } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3189 } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3196 // Check to make sure that the other destination of this branch
3197 // isn't BB itself. If so, this is an infinite loop that will
3198 // keep getting unwound.
3199 if (PBI->getSuccessor(PBIOp) == BB)
3202 // Do not perform this transformation if it would require
3203 // insertion of a large number of select instructions. For targets
3204 // without predication/cmovs, this is a big pessimization.
3206 // Also do not perform this transformation if any phi node in the common
3207 // destination block can trap when reached by BB or PBB (PR17073). In that
3208 // case, it would be unsafe to hoist the operation into a select instruction.
3210 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3211 unsigned NumPhis = 0;
3212 for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3214 if (NumPhis > 2) // Disable this xform.
3217 PHINode *PN = cast<PHINode>(II);
3218 Value *BIV = PN->getIncomingValueForBlock(BB);
3219 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3223 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3224 Value *PBIV = PN->getIncomingValue(PBBIdx);
3225 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3230 // Finally, if everything is ok, fold the branches to logical ops.
3231 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3233 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3234 << "AND: " << *BI->getParent());
3236 // If OtherDest *is* BB, then BB is a basic block with a single conditional
3237 // branch in it, where one edge (OtherDest) goes back to itself but the other
3238 // exits. We don't *know* that the program avoids the infinite loop
3239 // (even though that seems likely). If we do this xform naively, we'll end up
3240 // recursively unpeeling the loop. Since we know that (after the xform is
3241 // done) that the block *is* infinite if reached, we just make it an obviously
3242 // infinite loop with no cond branch.
3243 if (OtherDest == BB) {
3244 // Insert it at the end of the function, because it's either code,
3245 // or it won't matter if it's hot. :)
3246 BasicBlock *InfLoopBlock =
3247 BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3248 BranchInst::Create(InfLoopBlock, InfLoopBlock);
3249 OtherDest = InfLoopBlock;
3252 DEBUG(dbgs() << *PBI->getParent()->getParent());
3254 // BI may have other predecessors. Because of this, we leave
3255 // it alone, but modify PBI.
3257 // Make sure we get to CommonDest on True&True directions.
3258 Value *PBICond = PBI->getCondition();
3259 IRBuilder<NoFolder> Builder(PBI);
3261 PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3263 Value *BICond = BI->getCondition();
3265 BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3267 // Merge the conditions.
3268 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3270 // Modify PBI to branch on the new condition to the new dests.
3271 PBI->setCondition(Cond);
3272 PBI->setSuccessor(0, CommonDest);
3273 PBI->setSuccessor(1, OtherDest);
3275 // Update branch weight for PBI.
3276 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3277 uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3279 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3280 SuccTrueWeight, SuccFalseWeight);
3282 PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3283 PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3284 SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3285 SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3286 // The weight to CommonDest should be PredCommon * SuccTotal +
3287 // PredOther * SuccCommon.
3288 // The weight to OtherDest should be PredOther * SuccOther.
3289 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3290 PredOther * SuccCommon,
3291 PredOther * SuccOther};
3292 // Halve the weights if any of them cannot fit in an uint32_t
3293 FitWeights(NewWeights);
3295 PBI->setMetadata(LLVMContext::MD_prof,
3296 MDBuilder(BI->getContext())
3297 .createBranchWeights(NewWeights[0], NewWeights[1]));
3300 // OtherDest may have phi nodes. If so, add an entry from PBI's
3301 // block that are identical to the entries for BI's block.
3302 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3304 // We know that the CommonDest already had an edge from PBI to
3305 // it. If it has PHIs though, the PHIs may have different
3306 // entries for BB and PBI's BB. If so, insert a select to make
3309 for (BasicBlock::iterator II = CommonDest->begin();
3310 (PN = dyn_cast<PHINode>(II)); ++II) {
3311 Value *BIV = PN->getIncomingValueForBlock(BB);
3312 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3313 Value *PBIV = PN->getIncomingValue(PBBIdx);
3315 // Insert a select in PBI to pick the right value.
3316 SelectInst *NV = cast<SelectInst>(
3317 Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3318 PN->setIncomingValue(PBBIdx, NV);
3319 // Although the select has the same condition as PBI, the original branch
3320 // weights for PBI do not apply to the new select because the select's
3321 // 'logical' edges are incoming edges of the phi that is eliminated, not
3322 // the outgoing edges of PBI.
3324 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3325 uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3326 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3327 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3328 // The weight to PredCommonDest should be PredCommon * SuccTotal.
3329 // The weight to PredOtherDest should be PredOther * SuccCommon.
3330 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3331 PredOther * SuccCommon};
3333 FitWeights(NewWeights);
3335 NV->setMetadata(LLVMContext::MD_prof,
3336 MDBuilder(BI->getContext())
3337 .createBranchWeights(NewWeights[0], NewWeights[1]));
3342 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3343 DEBUG(dbgs() << *PBI->getParent()->getParent());
3345 // This basic block is probably dead. We know it has at least
3346 // one fewer predecessor.
3350 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3351 // true or to FalseBB if Cond is false.
3352 // Takes care of updating the successors and removing the old terminator.
3353 // Also makes sure not to introduce new successors by assuming that edges to
3354 // non-successor TrueBBs and FalseBBs aren't reachable.
3355 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
3356 BasicBlock *TrueBB, BasicBlock *FalseBB,
3357 uint32_t TrueWeight,
3358 uint32_t FalseWeight) {
3359 // Remove any superfluous successor edges from the CFG.
3360 // First, figure out which successors to preserve.
3361 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3363 BasicBlock *KeepEdge1 = TrueBB;
3364 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3366 // Then remove the rest.
3367 for (BasicBlock *Succ : OldTerm->successors()) {
3368 // Make sure only to keep exactly one copy of each edge.
3369 if (Succ == KeepEdge1)
3370 KeepEdge1 = nullptr;
3371 else if (Succ == KeepEdge2)
3372 KeepEdge2 = nullptr;
3374 Succ->removePredecessor(OldTerm->getParent(),
3375 /*DontDeleteUselessPHIs=*/true);
3378 IRBuilder<> Builder(OldTerm);
3379 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3381 // Insert an appropriate new terminator.
3382 if (!KeepEdge1 && !KeepEdge2) {
3383 if (TrueBB == FalseBB)
3384 // We were only looking for one successor, and it was present.
3385 // Create an unconditional branch to it.
3386 Builder.CreateBr(TrueBB);
3388 // We found both of the successors we were looking for.
3389 // Create a conditional branch sharing the condition of the select.
3390 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3391 if (TrueWeight != FalseWeight)
3392 NewBI->setMetadata(LLVMContext::MD_prof,
3393 MDBuilder(OldTerm->getContext())
3394 .createBranchWeights(TrueWeight, FalseWeight));
3396 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3397 // Neither of the selected blocks were successors, so this
3398 // terminator must be unreachable.
3399 new UnreachableInst(OldTerm->getContext(), OldTerm);
3401 // One of the selected values was a successor, but the other wasn't.
3402 // Insert an unconditional branch to the one that was found;
3403 // the edge to the one that wasn't must be unreachable.
3405 // Only TrueBB was found.
3406 Builder.CreateBr(TrueBB);
3408 // Only FalseBB was found.
3409 Builder.CreateBr(FalseBB);
3412 EraseTerminatorInstAndDCECond(OldTerm);
3417 // (switch (select cond, X, Y)) on constant X, Y
3418 // with a branch - conditional if X and Y lead to distinct BBs,
3419 // unconditional otherwise.
3420 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
3421 // Check for constant integer values in the select.
3422 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3423 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3424 if (!TrueVal || !FalseVal)
3427 // Find the relevant condition and destinations.
3428 Value *Condition = Select->getCondition();
3429 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
3430 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
3432 // Get weight for TrueBB and FalseBB.
3433 uint32_t TrueWeight = 0, FalseWeight = 0;
3434 SmallVector<uint64_t, 8> Weights;
3435 bool HasWeights = HasBranchWeights(SI);
3437 GetBranchWeights(SI, Weights);
3438 if (Weights.size() == 1 + SI->getNumCases()) {
3440 (uint32_t)Weights[SI->findCaseValue(TrueVal).getSuccessorIndex()];
3442 (uint32_t)Weights[SI->findCaseValue(FalseVal).getSuccessorIndex()];
3446 // Perform the actual simplification.
3447 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3452 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3453 // blockaddress(@fn, BlockB)))
3455 // (br cond, BlockA, BlockB).
3456 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
3457 // Check that both operands of the select are block addresses.
3458 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
3459 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
3463 // Extract the actual blocks.
3464 BasicBlock *TrueBB = TBA->getBasicBlock();
3465 BasicBlock *FalseBB = FBA->getBasicBlock();
3467 // Perform the actual simplification.
3468 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3472 /// This is called when we find an icmp instruction
3473 /// (a seteq/setne with a constant) as the only instruction in a
3474 /// block that ends with an uncond branch. We are looking for a very specific
3475 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3476 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3477 /// default value goes to an uncond block with a seteq in it, we get something
3480 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3482 /// %tmp = icmp eq i8 %A, 92
3485 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3487 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3488 /// the PHI, merging the third icmp into the switch.
3489 static bool TryToSimplifyUncondBranchWithICmpInIt(
3490 ICmpInst *ICI, IRBuilder<> &Builder, const DataLayout &DL,
3491 const TargetTransformInfo &TTI, unsigned BonusInstThreshold,
3492 AssumptionCache *AC) {
3493 BasicBlock *BB = ICI->getParent();
3495 // If the block has any PHIs in it or the icmp has multiple uses, it is too
3497 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3500 Value *V = ICI->getOperand(0);
3501 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3503 // The pattern we're looking for is where our only predecessor is a switch on
3504 // 'V' and this block is the default case for the switch. In this case we can
3505 // fold the compared value into the switch to simplify things.
3506 BasicBlock *Pred = BB->getSinglePredecessor();
3507 if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3510 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3511 if (SI->getCondition() != V)
3514 // If BB is reachable on a non-default case, then we simply know the value of
3515 // V in this block. Substitute it and constant fold the icmp instruction
3517 if (SI->getDefaultDest() != BB) {
3518 ConstantInt *VVal = SI->findCaseDest(BB);
3519 assert(VVal && "Should have a unique destination value");
3520 ICI->setOperand(0, VVal);
3522 if (Value *V = SimplifyInstruction(ICI, DL)) {
3523 ICI->replaceAllUsesWith(V);
3524 ICI->eraseFromParent();
3526 // BB is now empty, so it is likely to simplify away.
3527 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3530 // Ok, the block is reachable from the default dest. If the constant we're
3531 // comparing exists in one of the other edges, then we can constant fold ICI
3533 if (SI->findCaseValue(Cst) != SI->case_default()) {
3535 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3536 V = ConstantInt::getFalse(BB->getContext());
3538 V = ConstantInt::getTrue(BB->getContext());
3540 ICI->replaceAllUsesWith(V);
3541 ICI->eraseFromParent();
3542 // BB is now empty, so it is likely to simplify away.
3543 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3546 // The use of the icmp has to be in the 'end' block, by the only PHI node in
3548 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3549 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3550 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3551 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3554 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3556 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3557 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3559 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3560 std::swap(DefaultCst, NewCst);
3562 // Replace ICI (which is used by the PHI for the default value) with true or
3563 // false depending on if it is EQ or NE.
3564 ICI->replaceAllUsesWith(DefaultCst);
3565 ICI->eraseFromParent();
3567 // Okay, the switch goes to this block on a default value. Add an edge from
3568 // the switch to the merge point on the compared value.
3570 BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3571 SmallVector<uint64_t, 8> Weights;
3572 bool HasWeights = HasBranchWeights(SI);
3574 GetBranchWeights(SI, Weights);
3575 if (Weights.size() == 1 + SI->getNumCases()) {
3576 // Split weight for default case to case for "Cst".
3577 Weights[0] = (Weights[0] + 1) >> 1;
3578 Weights.push_back(Weights[0]);
3580 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3582 LLVMContext::MD_prof,
3583 MDBuilder(SI->getContext()).createBranchWeights(MDWeights));
3586 SI->addCase(Cst, NewBB);
3588 // NewBB branches to the phi block, add the uncond branch and the phi entry.
3589 Builder.SetInsertPoint(NewBB);
3590 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3591 Builder.CreateBr(SuccBlock);
3592 PHIUse->addIncoming(NewCst, NewBB);
3596 /// The specified branch is a conditional branch.
3597 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3598 /// fold it into a switch instruction if so.
3599 static bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
3600 const DataLayout &DL) {
3601 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3605 // Change br (X == 0 | X == 1), T, F into a switch instruction.
3606 // If this is a bunch of seteq's or'd together, or if it's a bunch of
3607 // 'setne's and'ed together, collect them.
3609 // Try to gather values from a chain of and/or to be turned into a switch
3610 ConstantComparesGatherer ConstantCompare(Cond, DL);
3611 // Unpack the result
3612 SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3613 Value *CompVal = ConstantCompare.CompValue;
3614 unsigned UsedICmps = ConstantCompare.UsedICmps;
3615 Value *ExtraCase = ConstantCompare.Extra;
3617 // If we didn't have a multiply compared value, fail.
3621 // Avoid turning single icmps into a switch.
3625 bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3627 // There might be duplicate constants in the list, which the switch
3628 // instruction can't handle, remove them now.
3629 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3630 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3632 // If Extra was used, we require at least two switch values to do the
3633 // transformation. A switch with one value is just a conditional branch.
3634 if (ExtraCase && Values.size() < 2)
3637 // TODO: Preserve branch weight metadata, similarly to how
3638 // FoldValueComparisonIntoPredecessors preserves it.
3640 // Figure out which block is which destination.
3641 BasicBlock *DefaultBB = BI->getSuccessor(1);
3642 BasicBlock *EdgeBB = BI->getSuccessor(0);
3644 std::swap(DefaultBB, EdgeBB);
3646 BasicBlock *BB = BI->getParent();
3648 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3649 << " cases into SWITCH. BB is:\n"
3652 // If there are any extra values that couldn't be folded into the switch
3653 // then we evaluate them with an explicit branch first. Split the block
3654 // right before the condbr to handle it.
3657 BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3658 // Remove the uncond branch added to the old block.
3659 TerminatorInst *OldTI = BB->getTerminator();
3660 Builder.SetInsertPoint(OldTI);
3663 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3665 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3667 OldTI->eraseFromParent();
3669 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3670 // for the edge we just added.
3671 AddPredecessorToBlock(EdgeBB, BB, NewBB);
3673 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3674 << "\nEXTRABB = " << *BB);
3678 Builder.SetInsertPoint(BI);
3679 // Convert pointer to int before we switch.
3680 if (CompVal->getType()->isPointerTy()) {
3681 CompVal = Builder.CreatePtrToInt(
3682 CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3685 // Create the new switch instruction now.
3686 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3688 // Add all of the 'cases' to the switch instruction.
3689 for (unsigned i = 0, e = Values.size(); i != e; ++i)
3690 New->addCase(Values[i], EdgeBB);
3692 // We added edges from PI to the EdgeBB. As such, if there were any
3693 // PHI nodes in EdgeBB, they need entries to be added corresponding to
3694 // the number of edges added.
3695 for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3696 PHINode *PN = cast<PHINode>(BBI);
3697 Value *InVal = PN->getIncomingValueForBlock(BB);
3698 for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3699 PN->addIncoming(InVal, BB);
3702 // Erase the old branch instruction.
3703 EraseTerminatorInstAndDCECond(BI);
3705 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3709 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3710 if (isa<PHINode>(RI->getValue()))
3711 return SimplifyCommonResume(RI);
3712 else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3713 RI->getValue() == RI->getParent()->getFirstNonPHI())
3714 // The resume must unwind the exception that caused control to branch here.
3715 return SimplifySingleResume(RI);
3720 // Simplify resume that is shared by several landing pads (phi of landing pad).
3721 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3722 BasicBlock *BB = RI->getParent();
3724 // Check that there are no other instructions except for debug intrinsics
3725 // between the phi of landing pads (RI->getValue()) and resume instruction.
3726 BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3727 E = RI->getIterator();
3729 if (!isa<DbgInfoIntrinsic>(I))
3732 SmallSet<BasicBlock *, 4> TrivialUnwindBlocks;
3733 auto *PhiLPInst = cast<PHINode>(RI->getValue());
3735 // Check incoming blocks to see if any of them are trivial.
3736 for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3738 auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3739 auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3741 // If the block has other successors, we can not delete it because
3742 // it has other dependents.
3743 if (IncomingBB->getUniqueSuccessor() != BB)
3746 auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3747 // Not the landing pad that caused the control to branch here.
3748 if (IncomingValue != LandingPad)
3751 bool isTrivial = true;
3753 I = IncomingBB->getFirstNonPHI()->getIterator();
3754 E = IncomingBB->getTerminator()->getIterator();
3756 if (!isa<DbgInfoIntrinsic>(I)) {
3762 TrivialUnwindBlocks.insert(IncomingBB);
3765 // If no trivial unwind blocks, don't do any simplifications.
3766 if (TrivialUnwindBlocks.empty())
3769 // Turn all invokes that unwind here into calls.
3770 for (auto *TrivialBB : TrivialUnwindBlocks) {
3771 // Blocks that will be simplified should be removed from the phi node.
3772 // Note there could be multiple edges to the resume block, and we need
3773 // to remove them all.
3774 while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3775 BB->removePredecessor(TrivialBB, true);
3777 for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3779 BasicBlock *Pred = *PI++;
3780 removeUnwindEdge(Pred);
3783 // In each SimplifyCFG run, only the current processed block can be erased.
3784 // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3785 // of erasing TrivialBB, we only remove the branch to the common resume
3786 // block so that we can later erase the resume block since it has no
3788 TrivialBB->getTerminator()->eraseFromParent();
3789 new UnreachableInst(RI->getContext(), TrivialBB);
3792 // Delete the resume block if all its predecessors have been removed.
3794 BB->eraseFromParent();
3796 return !TrivialUnwindBlocks.empty();
3799 // Simplify resume that is only used by a single (non-phi) landing pad.
3800 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3801 BasicBlock *BB = RI->getParent();
3802 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
3803 assert(RI->getValue() == LPInst &&
3804 "Resume must unwind the exception that caused control to here");
3806 // Check that there are no other instructions except for debug intrinsics.
3807 BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3809 if (!isa<DbgInfoIntrinsic>(I))
3812 // Turn all invokes that unwind here into calls and delete the basic block.
3813 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3814 BasicBlock *Pred = *PI++;
3815 removeUnwindEdge(Pred);
3818 // The landingpad is now unreachable. Zap it.
3819 BB->eraseFromParent();
3821 LoopHeaders->erase(BB);
3825 static bool removeEmptyCleanup(CleanupReturnInst *RI) {
3826 // If this is a trivial cleanup pad that executes no instructions, it can be
3827 // eliminated. If the cleanup pad continues to the caller, any predecessor
3828 // that is an EH pad will be updated to continue to the caller and any
3829 // predecessor that terminates with an invoke instruction will have its invoke
3830 // instruction converted to a call instruction. If the cleanup pad being
3831 // simplified does not continue to the caller, each predecessor will be
3832 // updated to continue to the unwind destination of the cleanup pad being
3834 BasicBlock *BB = RI->getParent();
3835 CleanupPadInst *CPInst = RI->getCleanupPad();
3836 if (CPInst->getParent() != BB)
3837 // This isn't an empty cleanup.
3840 // We cannot kill the pad if it has multiple uses. This typically arises
3841 // from unreachable basic blocks.
3842 if (!CPInst->hasOneUse())
3845 // Check that there are no other instructions except for benign intrinsics.
3846 BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3848 auto *II = dyn_cast<IntrinsicInst>(I);
3852 Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3853 switch (IntrinsicID) {
3854 case Intrinsic::dbg_declare:
3855 case Intrinsic::dbg_value:
3856 case Intrinsic::lifetime_end:
3863 // If the cleanup return we are simplifying unwinds to the caller, this will
3864 // set UnwindDest to nullptr.
3865 BasicBlock *UnwindDest = RI->getUnwindDest();
3866 Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3868 // We're about to remove BB from the control flow. Before we do, sink any
3869 // PHINodes into the unwind destination. Doing this before changing the
3870 // control flow avoids some potentially slow checks, since we can currently
3871 // be certain that UnwindDest and BB have no common predecessors (since they
3872 // are both EH pads).
3874 // First, go through the PHI nodes in UnwindDest and update any nodes that
3875 // reference the block we are removing
3876 for (BasicBlock::iterator I = UnwindDest->begin(),
3877 IE = DestEHPad->getIterator();
3879 PHINode *DestPN = cast<PHINode>(I);
3881 int Idx = DestPN->getBasicBlockIndex(BB);
3882 // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3884 // This PHI node has an incoming value that corresponds to a control
3885 // path through the cleanup pad we are removing. If the incoming
3886 // value is in the cleanup pad, it must be a PHINode (because we
3887 // verified above that the block is otherwise empty). Otherwise, the
3888 // value is either a constant or a value that dominates the cleanup
3889 // pad being removed.
3891 // Because BB and UnwindDest are both EH pads, all of their
3892 // predecessors must unwind to these blocks, and since no instruction
3893 // can have multiple unwind destinations, there will be no overlap in
3894 // incoming blocks between SrcPN and DestPN.
3895 Value *SrcVal = DestPN->getIncomingValue(Idx);
3896 PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3898 // Remove the entry for the block we are deleting.
3899 DestPN->removeIncomingValue(Idx, false);
3901 if (SrcPN && SrcPN->getParent() == BB) {
3902 // If the incoming value was a PHI node in the cleanup pad we are
3903 // removing, we need to merge that PHI node's incoming values into
3905 for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3906 SrcIdx != SrcE; ++SrcIdx) {
3907 DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3908 SrcPN->getIncomingBlock(SrcIdx));
3911 // Otherwise, the incoming value came from above BB and
3912 // so we can just reuse it. We must associate all of BB's
3913 // predecessors with this value.
3914 for (auto *pred : predecessors(BB)) {
3915 DestPN->addIncoming(SrcVal, pred);
3920 // Sink any remaining PHI nodes directly into UnwindDest.
3921 Instruction *InsertPt = DestEHPad;
3922 for (BasicBlock::iterator I = BB->begin(),
3923 IE = BB->getFirstNonPHI()->getIterator();
3925 // The iterator must be incremented here because the instructions are
3926 // being moved to another block.
3927 PHINode *PN = cast<PHINode>(I++);
3928 if (PN->use_empty())
3929 // If the PHI node has no uses, just leave it. It will be erased
3930 // when we erase BB below.
3933 // Otherwise, sink this PHI node into UnwindDest.
3934 // Any predecessors to UnwindDest which are not already represented
3935 // must be back edges which inherit the value from the path through
3936 // BB. In this case, the PHI value must reference itself.
3937 for (auto *pred : predecessors(UnwindDest))
3939 PN->addIncoming(PN, pred);
3940 PN->moveBefore(InsertPt);
3944 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3945 // The iterator must be updated here because we are removing this pred.
3946 BasicBlock *PredBB = *PI++;
3947 if (UnwindDest == nullptr) {
3948 removeUnwindEdge(PredBB);
3950 TerminatorInst *TI = PredBB->getTerminator();
3951 TI->replaceUsesOfWith(BB, UnwindDest);
3955 // The cleanup pad is now unreachable. Zap it.
3956 BB->eraseFromParent();
3960 // Try to merge two cleanuppads together.
3961 static bool mergeCleanupPad(CleanupReturnInst *RI) {
3962 // Skip any cleanuprets which unwind to caller, there is nothing to merge
3964 BasicBlock *UnwindDest = RI->getUnwindDest();
3968 // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
3969 // be safe to merge without code duplication.
3970 if (UnwindDest->getSinglePredecessor() != RI->getParent())
3973 // Verify that our cleanuppad's unwind destination is another cleanuppad.
3974 auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
3975 if (!SuccessorCleanupPad)
3978 CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
3979 // Replace any uses of the successor cleanupad with the predecessor pad
3980 // The only cleanuppad uses should be this cleanupret, it's cleanupret and
3981 // funclet bundle operands.
3982 SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
3983 // Remove the old cleanuppad.
3984 SuccessorCleanupPad->eraseFromParent();
3985 // Now, we simply replace the cleanupret with a branch to the unwind
3987 BranchInst::Create(UnwindDest, RI->getParent());
3988 RI->eraseFromParent();
3993 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
3994 // It is possible to transiantly have an undef cleanuppad operand because we
3995 // have deleted some, but not all, dead blocks.
3996 // Eventually, this block will be deleted.
3997 if (isa<UndefValue>(RI->getOperand(0)))
4000 if (mergeCleanupPad(RI))
4003 if (removeEmptyCleanup(RI))
4009 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4010 BasicBlock *BB = RI->getParent();
4011 if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4014 // Find predecessors that end with branches.
4015 SmallVector<BasicBlock *, 8> UncondBranchPreds;
4016 SmallVector<BranchInst *, 8> CondBranchPreds;
4017 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4018 BasicBlock *P = *PI;
4019 TerminatorInst *PTI = P->getTerminator();
4020 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4021 if (BI->isUnconditional())
4022 UncondBranchPreds.push_back(P);
4024 CondBranchPreds.push_back(BI);
4028 // If we found some, do the transformation!
4029 if (!UncondBranchPreds.empty() && DupRet) {
4030 while (!UncondBranchPreds.empty()) {
4031 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4032 DEBUG(dbgs() << "FOLDING: " << *BB
4033 << "INTO UNCOND BRANCH PRED: " << *Pred);
4034 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4037 // If we eliminated all predecessors of the block, delete the block now.
4038 if (pred_empty(BB)) {
4039 // We know there are no successors, so just nuke the block.
4040 BB->eraseFromParent();
4042 LoopHeaders->erase(BB);
4048 // Check out all of the conditional branches going to this return
4049 // instruction. If any of them just select between returns, change the
4050 // branch itself into a select/return pair.
4051 while (!CondBranchPreds.empty()) {
4052 BranchInst *BI = CondBranchPreds.pop_back_val();
4054 // Check to see if the non-BB successor is also a return block.
4055 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4056 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4057 SimplifyCondBranchToTwoReturns(BI, Builder))
4063 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4064 BasicBlock *BB = UI->getParent();
4066 bool Changed = false;
4068 // If there are any instructions immediately before the unreachable that can
4069 // be removed, do so.
4070 while (UI->getIterator() != BB->begin()) {
4071 BasicBlock::iterator BBI = UI->getIterator();
4073 // Do not delete instructions that can have side effects which might cause
4074 // the unreachable to not be reachable; specifically, calls and volatile
4075 // operations may have this effect.
4076 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4079 if (BBI->mayHaveSideEffects()) {
4080 if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4081 if (SI->isVolatile())
4083 } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4084 if (LI->isVolatile())
4086 } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4087 if (RMWI->isVolatile())
4089 } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4090 if (CXI->isVolatile())
4092 } else if (isa<CatchPadInst>(BBI)) {
4093 // A catchpad may invoke exception object constructors and such, which
4094 // in some languages can be arbitrary code, so be conservative by
4096 // For CoreCLR, it just involves a type test, so can be removed.
4097 if (classifyEHPersonality(BB->getParent()->getPersonalityFn()) !=
4098 EHPersonality::CoreCLR)
4100 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4101 !isa<LandingPadInst>(BBI)) {
4104 // Note that deleting LandingPad's here is in fact okay, although it
4105 // involves a bit of subtle reasoning. If this inst is a LandingPad,
4106 // all the predecessors of this block will be the unwind edges of Invokes,
4107 // and we can therefore guarantee this block will be erased.
4110 // Delete this instruction (any uses are guaranteed to be dead)
4111 if (!BBI->use_empty())
4112 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4113 BBI->eraseFromParent();
4117 // If the unreachable instruction is the first in the block, take a gander
4118 // at all of the predecessors of this instruction, and simplify them.
4119 if (&BB->front() != UI)
4122 SmallVector<BasicBlock *, 8> Preds(pred_begin(BB), pred_end(BB));
4123 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4124 TerminatorInst *TI = Preds[i]->getTerminator();
4125 IRBuilder<> Builder(TI);
4126 if (auto *BI = dyn_cast<BranchInst>(TI)) {
4127 if (BI->isUnconditional()) {
4128 if (BI->getSuccessor(0) == BB) {
4129 new UnreachableInst(TI->getContext(), TI);
4130 TI->eraseFromParent();
4134 if (BI->getSuccessor(0) == BB) {
4135 Builder.CreateBr(BI->getSuccessor(1));
4136 EraseTerminatorInstAndDCECond(BI);
4137 } else if (BI->getSuccessor(1) == BB) {
4138 Builder.CreateBr(BI->getSuccessor(0));
4139 EraseTerminatorInstAndDCECond(BI);
4143 } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4144 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e;
4146 if (i.getCaseSuccessor() == BB) {
4147 BB->removePredecessor(SI->getParent());
4153 } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4154 if (II->getUnwindDest() == BB) {
4155 removeUnwindEdge(TI->getParent());
4158 } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4159 if (CSI->getUnwindDest() == BB) {
4160 removeUnwindEdge(TI->getParent());
4165 for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4166 E = CSI->handler_end();
4169 CSI->removeHandler(I);
4175 if (CSI->getNumHandlers() == 0) {
4176 BasicBlock *CatchSwitchBB = CSI->getParent();
4177 if (CSI->hasUnwindDest()) {
4178 // Redirect preds to the unwind dest
4179 CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4181 // Rewrite all preds to unwind to caller (or from invoke to call).
4182 SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4183 for (BasicBlock *EHPred : EHPreds)
4184 removeUnwindEdge(EHPred);
4186 // The catchswitch is no longer reachable.
4187 new UnreachableInst(CSI->getContext(), CSI);
4188 CSI->eraseFromParent();
4191 } else if (isa<CleanupReturnInst>(TI)) {
4192 new UnreachableInst(TI->getContext(), TI);
4193 TI->eraseFromParent();
4198 // If this block is now dead, remove it.
4199 if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4200 // We know there are no successors, so just nuke the block.
4201 BB->eraseFromParent();
4203 LoopHeaders->erase(BB);
4210 static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) {
4211 assert(Cases.size() >= 1);
4213 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4214 for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4215 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4221 /// Turn a switch with two reachable destinations into an integer range
4222 /// comparison and branch.
4223 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
4224 assert(SI->getNumCases() > 1 && "Degenerate switch?");
4227 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4229 // Partition the cases into two sets with different destinations.
4230 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4231 BasicBlock *DestB = nullptr;
4232 SmallVector<ConstantInt *, 16> CasesA;
4233 SmallVector<ConstantInt *, 16> CasesB;
4235 for (SwitchInst::CaseIt I : SI->cases()) {
4236 BasicBlock *Dest = I.getCaseSuccessor();
4239 if (Dest == DestA) {
4240 CasesA.push_back(I.getCaseValue());
4245 if (Dest == DestB) {
4246 CasesB.push_back(I.getCaseValue());
4249 return false; // More than two destinations.
4252 assert(DestA && DestB &&
4253 "Single-destination switch should have been folded.");
4254 assert(DestA != DestB);
4255 assert(DestB != SI->getDefaultDest());
4256 assert(!CasesB.empty() && "There must be non-default cases.");
4257 assert(!CasesA.empty() || HasDefault);
4259 // Figure out if one of the sets of cases form a contiguous range.
4260 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4261 BasicBlock *ContiguousDest = nullptr;
4262 BasicBlock *OtherDest = nullptr;
4263 if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4264 ContiguousCases = &CasesA;
4265 ContiguousDest = DestA;
4267 } else if (CasesAreContiguous(CasesB)) {
4268 ContiguousCases = &CasesB;
4269 ContiguousDest = DestB;
4274 // Start building the compare and branch.
4276 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4277 Constant *NumCases =
4278 ConstantInt::get(Offset->getType(), ContiguousCases->size());
4280 Value *Sub = SI->getCondition();
4281 if (!Offset->isNullValue())
4282 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4285 // If NumCases overflowed, then all possible values jump to the successor.
4286 if (NumCases->isNullValue() && !ContiguousCases->empty())
4287 Cmp = ConstantInt::getTrue(SI->getContext());
4289 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4290 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4292 // Update weight for the newly-created conditional branch.
4293 if (HasBranchWeights(SI)) {
4294 SmallVector<uint64_t, 8> Weights;
4295 GetBranchWeights(SI, Weights);
4296 if (Weights.size() == 1 + SI->getNumCases()) {
4297 uint64_t TrueWeight = 0;
4298 uint64_t FalseWeight = 0;
4299 for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4300 if (SI->getSuccessor(I) == ContiguousDest)
4301 TrueWeight += Weights[I];
4303 FalseWeight += Weights[I];
4305 while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4309 NewBI->setMetadata(LLVMContext::MD_prof,
4310 MDBuilder(SI->getContext())
4311 .createBranchWeights((uint32_t)TrueWeight,
4312 (uint32_t)FalseWeight));
4316 // Prune obsolete incoming values off the successors' PHI nodes.
4317 for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4318 unsigned PreviousEdges = ContiguousCases->size();
4319 if (ContiguousDest == SI->getDefaultDest())
4321 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4322 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4324 for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4325 unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4326 if (OtherDest == SI->getDefaultDest())
4328 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4329 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4333 SI->eraseFromParent();
4338 /// Compute masked bits for the condition of a switch
4339 /// and use it to remove dead cases.
4340 static bool EliminateDeadSwitchCases(SwitchInst *SI, AssumptionCache *AC,
4341 const DataLayout &DL) {
4342 Value *Cond = SI->getCondition();
4343 unsigned Bits = Cond->getType()->getIntegerBitWidth();
4344 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
4345 computeKnownBits(Cond, KnownZero, KnownOne, DL, 0, AC, SI);
4347 // We can also eliminate cases by determining that their values are outside of
4348 // the limited range of the condition based on how many significant (non-sign)
4349 // bits are in the condition value.
4350 unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4351 unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4353 // Gather dead cases.
4354 SmallVector<ConstantInt *, 8> DeadCases;
4355 for (auto &Case : SI->cases()) {
4356 APInt CaseVal = Case.getCaseValue()->getValue();
4357 if ((CaseVal & KnownZero) != 0 || (CaseVal & KnownOne) != KnownOne ||
4358 (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4359 DeadCases.push_back(Case.getCaseValue());
4360 DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal << " is dead.\n");
4364 // If we can prove that the cases must cover all possible values, the
4365 // default destination becomes dead and we can remove it. If we know some
4366 // of the bits in the value, we can use that to more precisely compute the
4367 // number of possible unique case values.
4369 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4370 const unsigned NumUnknownBits =
4371 Bits - (KnownZero.Or(KnownOne)).countPopulation();
4372 assert(NumUnknownBits <= Bits);
4373 if (HasDefault && DeadCases.empty() &&
4374 NumUnknownBits < 64 /* avoid overflow */ &&
4375 SI->getNumCases() == (1ULL << NumUnknownBits)) {
4376 DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4377 BasicBlock *NewDefault =
4378 SplitBlockPredecessors(SI->getDefaultDest(), SI->getParent(), "");
4379 SI->setDefaultDest(&*NewDefault);
4380 SplitBlock(&*NewDefault, &NewDefault->front());
4381 auto *OldTI = NewDefault->getTerminator();
4382 new UnreachableInst(SI->getContext(), OldTI);
4383 EraseTerminatorInstAndDCECond(OldTI);
4387 SmallVector<uint64_t, 8> Weights;
4388 bool HasWeight = HasBranchWeights(SI);
4390 GetBranchWeights(SI, Weights);
4391 HasWeight = (Weights.size() == 1 + SI->getNumCases());
4394 // Remove dead cases from the switch.
4395 for (ConstantInt *DeadCase : DeadCases) {
4396 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCase);
4397 assert(Case != SI->case_default() &&
4398 "Case was not found. Probably mistake in DeadCases forming.");
4400 std::swap(Weights[Case.getCaseIndex() + 1], Weights.back());
4404 // Prune unused values from PHI nodes.
4405 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
4406 SI->removeCase(Case);
4408 if (HasWeight && Weights.size() >= 2) {
4409 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
4410 SI->setMetadata(LLVMContext::MD_prof,
4411 MDBuilder(SI->getParent()->getContext())
4412 .createBranchWeights(MDWeights));
4415 return !DeadCases.empty();
4418 /// If BB would be eligible for simplification by
4419 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4420 /// by an unconditional branch), look at the phi node for BB in the successor
4421 /// block and see if the incoming value is equal to CaseValue. If so, return
4422 /// the phi node, and set PhiIndex to BB's index in the phi node.
4423 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
4424 BasicBlock *BB, int *PhiIndex) {
4425 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4426 return nullptr; // BB must be empty to be a candidate for simplification.
4427 if (!BB->getSinglePredecessor())
4428 return nullptr; // BB must be dominated by the switch.
4430 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
4431 if (!Branch || !Branch->isUnconditional())
4432 return nullptr; // Terminator must be unconditional branch.
4434 BasicBlock *Succ = Branch->getSuccessor(0);
4436 BasicBlock::iterator I = Succ->begin();
4437 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
4438 int Idx = PHI->getBasicBlockIndex(BB);
4439 assert(Idx >= 0 && "PHI has no entry for predecessor?");
4441 Value *InValue = PHI->getIncomingValue(Idx);
4442 if (InValue != CaseValue)
4452 /// Try to forward the condition of a switch instruction to a phi node
4453 /// dominated by the switch, if that would mean that some of the destination
4454 /// blocks of the switch can be folded away.
4455 /// Returns true if a change is made.
4456 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
4457 typedef DenseMap<PHINode *, SmallVector<int, 4>> ForwardingNodesMap;
4458 ForwardingNodesMap ForwardingNodes;
4460 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E;
4462 ConstantInt *CaseValue = I.getCaseValue();
4463 BasicBlock *CaseDest = I.getCaseSuccessor();
4467 FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIndex);
4471 ForwardingNodes[PHI].push_back(PhiIndex);
4474 bool Changed = false;
4476 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
4477 E = ForwardingNodes.end();
4479 PHINode *Phi = I->first;
4480 SmallVectorImpl<int> &Indexes = I->second;
4482 if (Indexes.size() < 2)
4485 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
4486 Phi->setIncomingValue(Indexes[I], SI->getCondition());
4493 /// Return true if the backend will be able to handle
4494 /// initializing an array of constants like C.
4495 static bool ValidLookupTableConstant(Constant *C, const TargetTransformInfo &TTI) {
4496 if (C->isThreadDependent())
4498 if (C->isDLLImportDependent())
4501 if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4502 !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4503 !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4506 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4507 if (!CE->isGEPWithNoNotionalOverIndexing())
4509 if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4513 if (!TTI.shouldBuildLookupTablesForConstant(C))
4519 /// If V is a Constant, return it. Otherwise, try to look up
4520 /// its constant value in ConstantPool, returning 0 if it's not there.
4522 LookupConstant(Value *V,
4523 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4524 if (Constant *C = dyn_cast<Constant>(V))
4526 return ConstantPool.lookup(V);
4529 /// Try to fold instruction I into a constant. This works for
4530 /// simple instructions such as binary operations where both operands are
4531 /// constant or can be replaced by constants from the ConstantPool. Returns the
4532 /// resulting constant on success, 0 otherwise.
4534 ConstantFold(Instruction *I, const DataLayout &DL,
4535 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4536 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4537 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4540 if (A->isAllOnesValue())
4541 return LookupConstant(Select->getTrueValue(), ConstantPool);
4542 if (A->isNullValue())
4543 return LookupConstant(Select->getFalseValue(), ConstantPool);
4547 SmallVector<Constant *, 4> COps;
4548 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4549 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
4555 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4556 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4560 return ConstantFoldInstOperands(I, COps, DL);
4563 /// Try to determine the resulting constant values in phi nodes
4564 /// at the common destination basic block, *CommonDest, for one of the case
4565 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4566 /// case), of a switch instruction SI.
4568 GetCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest,
4569 BasicBlock **CommonDest,
4570 SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4571 const DataLayout &DL, const TargetTransformInfo &TTI) {
4572 // The block from which we enter the common destination.
4573 BasicBlock *Pred = SI->getParent();
4575 // If CaseDest is empty except for some side-effect free instructions through
4576 // which we can constant-propagate the CaseVal, continue to its successor.
4577 SmallDenseMap<Value *, Constant *> ConstantPool;
4578 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4579 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
4581 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
4582 // If the terminator is a simple branch, continue to the next block.
4583 if (T->getNumSuccessors() != 1 || T->isExceptional())
4586 CaseDest = T->getSuccessor(0);
4587 } else if (isa<DbgInfoIntrinsic>(I)) {
4588 // Skip debug intrinsic.
4590 } else if (Constant *C = ConstantFold(&*I, DL, ConstantPool)) {
4591 // Instruction is side-effect free and constant.
4593 // If the instruction has uses outside this block or a phi node slot for
4594 // the block, it is not safe to bypass the instruction since it would then
4595 // no longer dominate all its uses.
4596 for (auto &Use : I->uses()) {
4597 User *User = Use.getUser();
4598 if (Instruction *I = dyn_cast<Instruction>(User))
4599 if (I->getParent() == CaseDest)
4601 if (PHINode *Phi = dyn_cast<PHINode>(User))
4602 if (Phi->getIncomingBlock(Use) == CaseDest)
4607 ConstantPool.insert(std::make_pair(&*I, C));
4613 // If we did not have a CommonDest before, use the current one.
4615 *CommonDest = CaseDest;
4616 // If the destination isn't the common one, abort.
4617 if (CaseDest != *CommonDest)
4620 // Get the values for this case from phi nodes in the destination block.
4621 BasicBlock::iterator I = (*CommonDest)->begin();
4622 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
4623 int Idx = PHI->getBasicBlockIndex(Pred);
4627 Constant *ConstVal =
4628 LookupConstant(PHI->getIncomingValue(Idx), ConstantPool);
4632 // Be conservative about which kinds of constants we support.
4633 if (!ValidLookupTableConstant(ConstVal, TTI))
4636 Res.push_back(std::make_pair(PHI, ConstVal));
4639 return Res.size() > 0;
4642 // Helper function used to add CaseVal to the list of cases that generate
4644 static void MapCaseToResult(ConstantInt *CaseVal,
4645 SwitchCaseResultVectorTy &UniqueResults,
4647 for (auto &I : UniqueResults) {
4648 if (I.first == Result) {
4649 I.second.push_back(CaseVal);
4653 UniqueResults.push_back(
4654 std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4657 // Helper function that initializes a map containing
4658 // results for the PHI node of the common destination block for a switch
4659 // instruction. Returns false if multiple PHI nodes have been found or if
4660 // there is not a common destination block for the switch.
4661 static bool InitializeUniqueCases(SwitchInst *SI, PHINode *&PHI,
4662 BasicBlock *&CommonDest,
4663 SwitchCaseResultVectorTy &UniqueResults,
4664 Constant *&DefaultResult,
4665 const DataLayout &DL,
4666 const TargetTransformInfo &TTI) {
4667 for (auto &I : SI->cases()) {
4668 ConstantInt *CaseVal = I.getCaseValue();
4670 // Resulting value at phi nodes for this case value.
4671 SwitchCaseResultsTy Results;
4672 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4676 // Only one value per case is permitted
4677 if (Results.size() > 1)
4679 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4681 // Check the PHI consistency.
4683 PHI = Results[0].first;
4684 else if (PHI != Results[0].first)
4687 // Find the default result value.
4688 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
4689 BasicBlock *DefaultDest = SI->getDefaultDest();
4690 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4692 // If the default value is not found abort unless the default destination
4695 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4696 if ((!DefaultResult &&
4697 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4703 // Helper function that checks if it is possible to transform a switch with only
4704 // two cases (or two cases + default) that produces a result into a select.
4707 // case 10: %0 = icmp eq i32 %a, 10
4708 // return 10; %1 = select i1 %0, i32 10, i32 4
4709 // case 20: ----> %2 = icmp eq i32 %a, 20
4710 // return 2; %3 = select i1 %2, i32 2, i32 %1
4714 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4715 Constant *DefaultResult, Value *Condition,
4716 IRBuilder<> &Builder) {
4717 assert(ResultVector.size() == 2 &&
4718 "We should have exactly two unique results at this point");
4719 // If we are selecting between only two cases transform into a simple
4720 // select or a two-way select if default is possible.
4721 if (ResultVector[0].second.size() == 1 &&
4722 ResultVector[1].second.size() == 1) {
4723 ConstantInt *const FirstCase = ResultVector[0].second[0];
4724 ConstantInt *const SecondCase = ResultVector[1].second[0];
4726 bool DefaultCanTrigger = DefaultResult;
4727 Value *SelectValue = ResultVector[1].first;
4728 if (DefaultCanTrigger) {
4729 Value *const ValueCompare =
4730 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4731 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4732 DefaultResult, "switch.select");
4734 Value *const ValueCompare =
4735 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4736 return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4737 SelectValue, "switch.select");
4743 // Helper function to cleanup a switch instruction that has been converted into
4744 // a select, fixing up PHI nodes and basic blocks.
4745 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
4747 IRBuilder<> &Builder) {
4748 BasicBlock *SelectBB = SI->getParent();
4749 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4750 PHI->removeIncomingValue(SelectBB);
4751 PHI->addIncoming(SelectValue, SelectBB);
4753 Builder.CreateBr(PHI->getParent());
4755 // Remove the switch.
4756 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4757 BasicBlock *Succ = SI->getSuccessor(i);
4759 if (Succ == PHI->getParent())
4761 Succ->removePredecessor(SelectBB);
4763 SI->eraseFromParent();
4766 /// If the switch is only used to initialize one or more
4767 /// phi nodes in a common successor block with only two different
4768 /// constant values, replace the switch with select.
4769 static bool SwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4770 AssumptionCache *AC, const DataLayout &DL,
4771 const TargetTransformInfo &TTI) {
4772 Value *const Cond = SI->getCondition();
4773 PHINode *PHI = nullptr;
4774 BasicBlock *CommonDest = nullptr;
4775 Constant *DefaultResult;
4776 SwitchCaseResultVectorTy UniqueResults;
4777 // Collect all the cases that will deliver the same value from the switch.
4778 if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4781 // Selects choose between maximum two values.
4782 if (UniqueResults.size() != 2)
4784 assert(PHI != nullptr && "PHI for value select not found");
4786 Builder.SetInsertPoint(SI);
4787 Value *SelectValue =
4788 ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4790 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4793 // The switch couldn't be converted into a select.
4799 /// This class represents a lookup table that can be used to replace a switch.
4800 class SwitchLookupTable {
4802 /// Create a lookup table to use as a switch replacement with the contents
4803 /// of Values, using DefaultValue to fill any holes in the table.
4805 Module &M, uint64_t TableSize, ConstantInt *Offset,
4806 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4807 Constant *DefaultValue, const DataLayout &DL);
4809 /// Build instructions with Builder to retrieve the value at
4810 /// the position given by Index in the lookup table.
4811 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4813 /// Return true if a table with TableSize elements of
4814 /// type ElementType would fit in a target-legal register.
4815 static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4819 // Depending on the contents of the table, it can be represented in
4822 // For tables where each element contains the same value, we just have to
4823 // store that single value and return it for each lookup.
4826 // For tables where there is a linear relationship between table index
4827 // and values. We calculate the result with a simple multiplication
4828 // and addition instead of a table lookup.
4831 // For small tables with integer elements, we can pack them into a bitmap
4832 // that fits into a target-legal register. Values are retrieved by
4833 // shift and mask operations.
4836 // The table is stored as an array of values. Values are retrieved by load
4837 // instructions from the table.
4841 // For SingleValueKind, this is the single value.
4842 Constant *SingleValue;
4844 // For BitMapKind, this is the bitmap.
4845 ConstantInt *BitMap;
4846 IntegerType *BitMapElementTy;
4848 // For LinearMapKind, these are the constants used to derive the value.
4849 ConstantInt *LinearOffset;
4850 ConstantInt *LinearMultiplier;
4852 // For ArrayKind, this is the array.
4853 GlobalVariable *Array;
4856 } // end anonymous namespace
4858 SwitchLookupTable::SwitchLookupTable(
4859 Module &M, uint64_t TableSize, ConstantInt *Offset,
4860 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4861 Constant *DefaultValue, const DataLayout &DL)
4862 : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
4863 LinearOffset(nullptr), LinearMultiplier(nullptr), Array(nullptr) {
4864 assert(Values.size() && "Can't build lookup table without values!");
4865 assert(TableSize >= Values.size() && "Can't fit values in table!");
4867 // If all values in the table are equal, this is that value.
4868 SingleValue = Values.begin()->second;
4870 Type *ValueType = Values.begin()->second->getType();
4872 // Build up the table contents.
4873 SmallVector<Constant *, 64> TableContents(TableSize);
4874 for (size_t I = 0, E = Values.size(); I != E; ++I) {
4875 ConstantInt *CaseVal = Values[I].first;
4876 Constant *CaseRes = Values[I].second;
4877 assert(CaseRes->getType() == ValueType);
4879 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4880 TableContents[Idx] = CaseRes;
4882 if (CaseRes != SingleValue)
4883 SingleValue = nullptr;
4886 // Fill in any holes in the table with the default result.
4887 if (Values.size() < TableSize) {
4888 assert(DefaultValue &&
4889 "Need a default value to fill the lookup table holes.");
4890 assert(DefaultValue->getType() == ValueType);
4891 for (uint64_t I = 0; I < TableSize; ++I) {
4892 if (!TableContents[I])
4893 TableContents[I] = DefaultValue;
4896 if (DefaultValue != SingleValue)
4897 SingleValue = nullptr;
4900 // If each element in the table contains the same value, we only need to store
4901 // that single value.
4903 Kind = SingleValueKind;
4907 // Check if we can derive the value with a linear transformation from the
4909 if (isa<IntegerType>(ValueType)) {
4910 bool LinearMappingPossible = true;
4913 assert(TableSize >= 2 && "Should be a SingleValue table.");
4914 // Check if there is the same distance between two consecutive values.
4915 for (uint64_t I = 0; I < TableSize; ++I) {
4916 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4918 // This is an undef. We could deal with it, but undefs in lookup tables
4919 // are very seldom. It's probably not worth the additional complexity.
4920 LinearMappingPossible = false;
4923 APInt Val = ConstVal->getValue();
4925 APInt Dist = Val - PrevVal;
4928 } else if (Dist != DistToPrev) {
4929 LinearMappingPossible = false;
4935 if (LinearMappingPossible) {
4936 LinearOffset = cast<ConstantInt>(TableContents[0]);
4937 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
4938 Kind = LinearMapKind;
4944 // If the type is integer and the table fits in a register, build a bitmap.
4945 if (WouldFitInRegister(DL, TableSize, ValueType)) {
4946 IntegerType *IT = cast<IntegerType>(ValueType);
4947 APInt TableInt(TableSize * IT->getBitWidth(), 0);
4948 for (uint64_t I = TableSize; I > 0; --I) {
4949 TableInt <<= IT->getBitWidth();
4950 // Insert values into the bitmap. Undef values are set to zero.
4951 if (!isa<UndefValue>(TableContents[I - 1])) {
4952 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
4953 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
4956 BitMap = ConstantInt::get(M.getContext(), TableInt);
4957 BitMapElementTy = IT;
4963 // Store the table in an array.
4964 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
4965 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
4967 Array = new GlobalVariable(M, ArrayTy, /*constant=*/true,
4968 GlobalVariable::PrivateLinkage, Initializer,
4970 Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
4974 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
4976 case SingleValueKind:
4978 case LinearMapKind: {
4979 // Derive the result value from the input value.
4980 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
4981 false, "switch.idx.cast");
4982 if (!LinearMultiplier->isOne())
4983 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
4984 if (!LinearOffset->isZero())
4985 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
4989 // Type of the bitmap (e.g. i59).
4990 IntegerType *MapTy = BitMap->getType();
4992 // Cast Index to the same type as the bitmap.
4993 // Note: The Index is <= the number of elements in the table, so
4994 // truncating it to the width of the bitmask is safe.
4995 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
4997 // Multiply the shift amount by the element width.
4998 ShiftAmt = Builder.CreateMul(
4999 ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5003 Value *DownShifted =
5004 Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5006 return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5009 // Make sure the table index will not overflow when treated as signed.
5010 IntegerType *IT = cast<IntegerType>(Index->getType());
5011 uint64_t TableSize =
5012 Array->getInitializer()->getType()->getArrayNumElements();
5013 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5014 Index = Builder.CreateZExt(
5015 Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5016 "switch.tableidx.zext");
5018 Value *GEPIndices[] = {Builder.getInt32(0), Index};
5019 Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5020 GEPIndices, "switch.gep");
5021 return Builder.CreateLoad(GEP, "switch.load");
5024 llvm_unreachable("Unknown lookup table kind!");
5027 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5029 Type *ElementType) {
5030 auto *IT = dyn_cast<IntegerType>(ElementType);
5033 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5034 // are <= 15, we could try to narrow the type.
5036 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5037 if (TableSize >= UINT_MAX / IT->getBitWidth())
5039 return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5042 /// Determine whether a lookup table should be built for this switch, based on
5043 /// the number of cases, size of the table, and the types of the results.
5045 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5046 const TargetTransformInfo &TTI, const DataLayout &DL,
5047 const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5048 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5049 return false; // TableSize overflowed, or mul below might overflow.
5051 bool AllTablesFitInRegister = true;
5052 bool HasIllegalType = false;
5053 for (const auto &I : ResultTypes) {
5054 Type *Ty = I.second;
5056 // Saturate this flag to true.
5057 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5059 // Saturate this flag to false.
5060 AllTablesFitInRegister =
5061 AllTablesFitInRegister &&
5062 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5064 // If both flags saturate, we're done. NOTE: This *only* works with
5065 // saturating flags, and all flags have to saturate first due to the
5066 // non-deterministic behavior of iterating over a dense map.
5067 if (HasIllegalType && !AllTablesFitInRegister)
5071 // If each table would fit in a register, we should build it anyway.
5072 if (AllTablesFitInRegister)
5075 // Don't build a table that doesn't fit in-register if it has illegal types.
5079 // The table density should be at least 40%. This is the same criterion as for
5080 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5081 // FIXME: Find the best cut-off.
5082 return SI->getNumCases() * 10 >= TableSize * 4;
5085 /// Try to reuse the switch table index compare. Following pattern:
5087 /// if (idx < tablesize)
5088 /// r = table[idx]; // table does not contain default_value
5090 /// r = default_value;
5091 /// if (r != default_value)
5094 /// Is optimized to:
5096 /// cond = idx < tablesize;
5100 /// r = default_value;
5104 /// Jump threading will then eliminate the second if(cond).
5105 static void reuseTableCompare(
5106 User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5107 Constant *DefaultValue,
5108 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5110 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5114 // We require that the compare is in the same block as the phi so that jump
5115 // threading can do its work afterwards.
5116 if (CmpInst->getParent() != PhiBlock)
5119 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5123 Value *RangeCmp = RangeCheckBranch->getCondition();
5124 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5125 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5127 // Check if the compare with the default value is constant true or false.
5128 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5129 DefaultValue, CmpOp1, true);
5130 if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5133 // Check if the compare with the case values is distinct from the default
5135 for (auto ValuePair : Values) {
5136 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5137 ValuePair.second, CmpOp1, true);
5138 if (!CaseConst || CaseConst == DefaultConst)
5140 assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5141 "Expect true or false as compare result.");
5144 // Check if the branch instruction dominates the phi node. It's a simple
5145 // dominance check, but sufficient for our needs.
5146 // Although this check is invariant in the calling loops, it's better to do it
5147 // at this late stage. Practically we do it at most once for a switch.
5148 BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5149 for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5150 BasicBlock *Pred = *PI;
5151 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5155 if (DefaultConst == FalseConst) {
5156 // The compare yields the same result. We can replace it.
5157 CmpInst->replaceAllUsesWith(RangeCmp);
5158 ++NumTableCmpReuses;
5160 // The compare yields the same result, just inverted. We can replace it.
5161 Value *InvertedTableCmp = BinaryOperator::CreateXor(
5162 RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5164 CmpInst->replaceAllUsesWith(InvertedTableCmp);
5165 ++NumTableCmpReuses;
5169 /// If the switch is only used to initialize one or more phi nodes in a common
5170 /// successor block with different constant values, replace the switch with
5172 static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
5173 const DataLayout &DL,
5174 const TargetTransformInfo &TTI) {
5175 assert(SI->getNumCases() > 1 && "Degenerate switch?");
5177 // Only build lookup table when we have a target that supports it.
5178 if (!TTI.shouldBuildLookupTables())
5181 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5182 // split off a dense part and build a lookup table for that.
5184 // FIXME: This creates arrays of GEPs to constant strings, which means each
5185 // GEP needs a runtime relocation in PIC code. We should just build one big
5186 // string and lookup indices into that.
5188 // Ignore switches with less than three cases. Lookup tables will not make
5190 // faster, so we don't analyze them.
5191 if (SI->getNumCases() < 3)
5194 // Figure out the corresponding result for each case value and phi node in the
5195 // common destination, as well as the min and max case values.
5196 assert(SI->case_begin() != SI->case_end());
5197 SwitchInst::CaseIt CI = SI->case_begin();
5198 ConstantInt *MinCaseVal = CI.getCaseValue();
5199 ConstantInt *MaxCaseVal = CI.getCaseValue();
5201 BasicBlock *CommonDest = nullptr;
5202 typedef SmallVector<std::pair<ConstantInt *, Constant *>, 4> ResultListTy;
5203 SmallDenseMap<PHINode *, ResultListTy> ResultLists;
5204 SmallDenseMap<PHINode *, Constant *> DefaultResults;
5205 SmallDenseMap<PHINode *, Type *> ResultTypes;
5206 SmallVector<PHINode *, 4> PHIs;
5208 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5209 ConstantInt *CaseVal = CI.getCaseValue();
5210 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5211 MinCaseVal = CaseVal;
5212 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5213 MaxCaseVal = CaseVal;
5215 // Resulting value at phi nodes for this case value.
5216 typedef SmallVector<std::pair<PHINode *, Constant *>, 4> ResultsTy;
5218 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
5222 // Append the result from this case to the list for each phi.
5223 for (const auto &I : Results) {
5224 PHINode *PHI = I.first;
5225 Constant *Value = I.second;
5226 if (!ResultLists.count(PHI))
5227 PHIs.push_back(PHI);
5228 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5232 // Keep track of the result types.
5233 for (PHINode *PHI : PHIs) {
5234 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5237 uint64_t NumResults = ResultLists[PHIs[0]].size();
5238 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5239 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5240 bool TableHasHoles = (NumResults < TableSize);
5242 // If the table has holes, we need a constant result for the default case
5243 // or a bitmask that fits in a register.
5244 SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5245 bool HasDefaultResults =
5246 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5247 DefaultResultsList, DL, TTI);
5249 bool NeedMask = (TableHasHoles && !HasDefaultResults);
5251 // As an extra penalty for the validity test we require more cases.
5252 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5254 if (!DL.fitsInLegalInteger(TableSize))
5258 for (const auto &I : DefaultResultsList) {
5259 PHINode *PHI = I.first;
5260 Constant *Result = I.second;
5261 DefaultResults[PHI] = Result;
5264 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5267 // Create the BB that does the lookups.
5268 Module &Mod = *CommonDest->getParent()->getParent();
5269 BasicBlock *LookupBB = BasicBlock::Create(
5270 Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5272 // Compute the table index value.
5273 Builder.SetInsertPoint(SI);
5275 Builder.CreateSub(SI->getCondition(), MinCaseVal, "switch.tableidx");
5277 // Compute the maximum table size representable by the integer type we are
5279 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5280 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5281 assert(MaxTableSize >= TableSize &&
5282 "It is impossible for a switch to have more entries than the max "
5283 "representable value of its input integer type's size.");
5285 // If the default destination is unreachable, or if the lookup table covers
5286 // all values of the conditional variable, branch directly to the lookup table
5287 // BB. Otherwise, check that the condition is within the case range.
5288 const bool DefaultIsReachable =
5289 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5290 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5291 BranchInst *RangeCheckBranch = nullptr;
5293 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5294 Builder.CreateBr(LookupBB);
5295 // Note: We call removeProdecessor later since we need to be able to get the
5296 // PHI value for the default case in case we're using a bit mask.
5298 Value *Cmp = Builder.CreateICmpULT(
5299 TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5301 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5304 // Populate the BB that does the lookups.
5305 Builder.SetInsertPoint(LookupBB);
5308 // Before doing the lookup we do the hole check.
5309 // The LookupBB is therefore re-purposed to do the hole check
5310 // and we create a new LookupBB.
5311 BasicBlock *MaskBB = LookupBB;
5312 MaskBB->setName("switch.hole_check");
5313 LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5314 CommonDest->getParent(), CommonDest);
5316 // Make the mask's bitwidth at least 8bit and a power-of-2 to avoid
5317 // unnecessary illegal types.
5318 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5319 APInt MaskInt(TableSizePowOf2, 0);
5320 APInt One(TableSizePowOf2, 1);
5321 // Build bitmask; fill in a 1 bit for every case.
5322 const ResultListTy &ResultList = ResultLists[PHIs[0]];
5323 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5324 uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5326 MaskInt |= One << Idx;
5328 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5330 // Get the TableIndex'th bit of the bitmask.
5331 // If this bit is 0 (meaning hole) jump to the default destination,
5332 // else continue with table lookup.
5333 IntegerType *MapTy = TableMask->getType();
5335 Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5336 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5337 Value *LoBit = Builder.CreateTrunc(
5338 Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5339 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5341 Builder.SetInsertPoint(LookupBB);
5342 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5345 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5346 // We cached PHINodes in PHIs, to avoid accessing deleted PHINodes later,
5347 // do not delete PHINodes here.
5348 SI->getDefaultDest()->removePredecessor(SI->getParent(),
5349 /*DontDeleteUselessPHIs=*/true);
5352 bool ReturnedEarly = false;
5353 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
5354 PHINode *PHI = PHIs[I];
5355 const ResultListTy &ResultList = ResultLists[PHI];
5357 // If using a bitmask, use any value to fill the lookup table holes.
5358 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5359 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL);
5361 Value *Result = Table.BuildLookup(TableIndex, Builder);
5363 // If the result is used to return immediately from the function, we want to
5364 // do that right here.
5365 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5366 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5367 Builder.CreateRet(Result);
5368 ReturnedEarly = true;
5372 // Do a small peephole optimization: re-use the switch table compare if
5374 if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5375 BasicBlock *PhiBlock = PHI->getParent();
5376 // Search for compare instructions which use the phi.
5377 for (auto *User : PHI->users()) {
5378 reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5382 PHI->addIncoming(Result, LookupBB);
5386 Builder.CreateBr(CommonDest);
5388 // Remove the switch.
5389 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5390 BasicBlock *Succ = SI->getSuccessor(i);
5392 if (Succ == SI->getDefaultDest())
5394 Succ->removePredecessor(SI->getParent());
5396 SI->eraseFromParent();
5400 ++NumLookupTablesHoles;
5404 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5405 // See also SelectionDAGBuilder::isDense(), which this function was based on.
5406 uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5407 uint64_t Range = Diff + 1;
5408 uint64_t NumCases = Values.size();
5409 // 40% is the default density for building a jump table in optsize/minsize mode.
5410 uint64_t MinDensity = 40;
5412 return NumCases * 100 >= Range * MinDensity;
5415 // Try and transform a switch that has "holes" in it to a contiguous sequence
5418 // A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5419 // range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5421 // This converts a sparse switch into a dense switch which allows better
5422 // lowering and could also allow transforming into a lookup table.
5423 static bool ReduceSwitchRange(SwitchInst *SI, IRBuilder<> &Builder,
5424 const DataLayout &DL,
5425 const TargetTransformInfo &TTI) {
5426 auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5427 if (CondTy->getIntegerBitWidth() > 64 ||
5428 !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5430 // Only bother with this optimization if there are more than 3 switch cases;
5431 // SDAG will only bother creating jump tables for 4 or more cases.
5432 if (SI->getNumCases() < 4)
5435 // This transform is agnostic to the signedness of the input or case values. We
5436 // can treat the case values as signed or unsigned. We can optimize more common
5437 // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5439 SmallVector<int64_t,4> Values;
5440 for (auto &C : SI->cases())
5441 Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5442 std::sort(Values.begin(), Values.end());
5444 // If the switch is already dense, there's nothing useful to do here.
5445 if (isSwitchDense(Values))
5448 // First, transform the values such that they start at zero and ascend.
5449 int64_t Base = Values[0];
5450 for (auto &V : Values)
5453 // Now we have signed numbers that have been shifted so that, given enough
5454 // precision, there are no negative values. Since the rest of the transform
5455 // is bitwise only, we switch now to an unsigned representation.
5457 for (auto &V : Values)
5458 GCD = GreatestCommonDivisor64(GCD, (uint64_t)V);
5460 // This transform can be done speculatively because it is so cheap - it results
5461 // in a single rotate operation being inserted. This can only happen if the
5462 // factor extracted is a power of 2.
5463 // FIXME: If the GCD is an odd number we can multiply by the multiplicative
5464 // inverse of GCD and then perform this transform.
5465 // FIXME: It's possible that optimizing a switch on powers of two might also
5466 // be beneficial - flag values are often powers of two and we could use a CLZ
5467 // as the key function.
5468 if (GCD <= 1 || !isPowerOf2_64(GCD))
5469 // No common divisor found or too expensive to compute key function.
5472 unsigned Shift = Log2_64(GCD);
5473 for (auto &V : Values)
5474 V = (int64_t)((uint64_t)V >> Shift);
5476 if (!isSwitchDense(Values))
5477 // Transform didn't create a dense switch.
5480 // The obvious transform is to shift the switch condition right and emit a
5481 // check that the condition actually cleanly divided by GCD, i.e.
5482 // C & (1 << Shift - 1) == 0
5483 // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5485 // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5486 // shift and puts the shifted-off bits in the uppermost bits. If any of these
5487 // are nonzero then the switch condition will be very large and will hit the
5490 auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5491 Builder.SetInsertPoint(SI);
5492 auto *ShiftC = ConstantInt::get(Ty, Shift);
5493 auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5494 auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5495 auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5496 auto *Rot = Builder.CreateOr(LShr, Shl);
5497 SI->replaceUsesOfWith(SI->getCondition(), Rot);
5499 for (SwitchInst::CaseIt C = SI->case_begin(), E = SI->case_end(); C != E;
5501 auto *Orig = C.getCaseValue();
5502 auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5504 cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5509 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
5510 BasicBlock *BB = SI->getParent();
5512 if (isValueEqualityComparison(SI)) {
5513 // If we only have one predecessor, and if it is a branch on this value,
5514 // see if that predecessor totally determines the outcome of this switch.
5515 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5516 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
5517 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5519 Value *Cond = SI->getCondition();
5520 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
5521 if (SimplifySwitchOnSelect(SI, Select))
5522 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5524 // If the block only contains the switch, see if we can fold the block
5525 // away into any preds.
5526 BasicBlock::iterator BBI = BB->begin();
5527 // Ignore dbg intrinsics.
5528 while (isa<DbgInfoIntrinsic>(BBI))
5531 if (FoldValueComparisonIntoPredecessors(SI, Builder))
5532 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5535 // Try to transform the switch into an icmp and a branch.
5536 if (TurnSwitchRangeIntoICmp(SI, Builder))
5537 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5539 // Remove unreachable cases.
5540 if (EliminateDeadSwitchCases(SI, AC, DL))
5541 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5543 if (SwitchToSelect(SI, Builder, AC, DL, TTI))
5544 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5546 if (ForwardSwitchConditionToPHI(SI))
5547 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5549 if (SwitchToLookupTable(SI, Builder, DL, TTI))
5550 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5552 if (ReduceSwitchRange(SI, Builder, DL, TTI))
5553 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5558 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
5559 BasicBlock *BB = IBI->getParent();
5560 bool Changed = false;
5562 // Eliminate redundant destinations.
5563 SmallPtrSet<Value *, 8> Succs;
5564 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
5565 BasicBlock *Dest = IBI->getDestination(i);
5566 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
5567 Dest->removePredecessor(BB);
5568 IBI->removeDestination(i);
5575 if (IBI->getNumDestinations() == 0) {
5576 // If the indirectbr has no successors, change it to unreachable.
5577 new UnreachableInst(IBI->getContext(), IBI);
5578 EraseTerminatorInstAndDCECond(IBI);
5582 if (IBI->getNumDestinations() == 1) {
5583 // If the indirectbr has one successor, change it to a direct branch.
5584 BranchInst::Create(IBI->getDestination(0), IBI);
5585 EraseTerminatorInstAndDCECond(IBI);
5589 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
5590 if (SimplifyIndirectBrOnSelect(IBI, SI))
5591 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5596 /// Given an block with only a single landing pad and a unconditional branch
5597 /// try to find another basic block which this one can be merged with. This
5598 /// handles cases where we have multiple invokes with unique landing pads, but
5599 /// a shared handler.
5601 /// We specifically choose to not worry about merging non-empty blocks
5602 /// here. That is a PRE/scheduling problem and is best solved elsewhere. In
5603 /// practice, the optimizer produces empty landing pad blocks quite frequently
5604 /// when dealing with exception dense code. (see: instcombine, gvn, if-else
5605 /// sinking in this file)
5607 /// This is primarily a code size optimization. We need to avoid performing
5608 /// any transform which might inhibit optimization (such as our ability to
5609 /// specialize a particular handler via tail commoning). We do this by not
5610 /// merging any blocks which require us to introduce a phi. Since the same
5611 /// values are flowing through both blocks, we don't loose any ability to
5612 /// specialize. If anything, we make such specialization more likely.
5614 /// TODO - This transformation could remove entries from a phi in the target
5615 /// block when the inputs in the phi are the same for the two blocks being
5616 /// merged. In some cases, this could result in removal of the PHI entirely.
5617 static bool TryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI,
5619 auto Succ = BB->getUniqueSuccessor();
5621 // If there's a phi in the successor block, we'd likely have to introduce
5622 // a phi into the merged landing pad block.
5623 if (isa<PHINode>(*Succ->begin()))
5626 for (BasicBlock *OtherPred : predecessors(Succ)) {
5627 if (BB == OtherPred)
5629 BasicBlock::iterator I = OtherPred->begin();
5630 LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(I);
5631 if (!LPad2 || !LPad2->isIdenticalTo(LPad))
5633 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {
5635 BranchInst *BI2 = dyn_cast<BranchInst>(I);
5636 if (!BI2 || !BI2->isIdenticalTo(BI))
5639 // We've found an identical block. Update our predecessors to take that
5640 // path instead and make ourselves dead.
5641 SmallSet<BasicBlock *, 16> Preds;
5642 Preds.insert(pred_begin(BB), pred_end(BB));
5643 for (BasicBlock *Pred : Preds) {
5644 InvokeInst *II = cast<InvokeInst>(Pred->getTerminator());
5645 assert(II->getNormalDest() != BB && II->getUnwindDest() == BB &&
5646 "unexpected successor");
5647 II->setUnwindDest(OtherPred);
5650 // The debug info in OtherPred doesn't cover the merged control flow that
5651 // used to go through BB. We need to delete it or update it.
5652 for (auto I = OtherPred->begin(), E = OtherPred->end(); I != E;) {
5653 Instruction &Inst = *I;
5655 if (isa<DbgInfoIntrinsic>(Inst))
5656 Inst.eraseFromParent();
5659 SmallSet<BasicBlock *, 16> Succs;
5660 Succs.insert(succ_begin(BB), succ_end(BB));
5661 for (BasicBlock *Succ : Succs) {
5662 Succ->removePredecessor(BB);
5665 IRBuilder<> Builder(BI);
5666 Builder.CreateUnreachable();
5667 BI->eraseFromParent();
5673 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI,
5674 IRBuilder<> &Builder) {
5675 BasicBlock *BB = BI->getParent();
5677 if (SinkCommon && SinkThenElseCodeToEnd(BI))
5680 // If the Terminator is the only non-phi instruction, simplify the block.
5681 // if LoopHeader is provided, check if the block is a loop header
5682 // (This is for early invocations before loop simplify and vectorization
5683 // to keep canonical loop forms for nested loops.
5684 // These blocks can be eliminated when the pass is invoked later
5685 // in the back-end.)
5686 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg()->getIterator();
5687 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
5688 (!LoopHeaders || !LoopHeaders->count(BB)) &&
5689 TryToSimplifyUncondBranchFromEmptyBlock(BB))
5692 // If the only instruction in the block is a seteq/setne comparison
5693 // against a constant, try to simplify the block.
5694 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
5695 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
5696 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5698 if (I->isTerminator() &&
5699 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, DL, TTI,
5700 BonusInstThreshold, AC))
5704 // See if we can merge an empty landing pad block with another which is
5706 if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(I)) {
5707 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {
5709 if (I->isTerminator() && TryToMergeLandingPad(LPad, BI, BB))
5713 // If this basic block is ONLY a compare and a branch, and if a predecessor
5714 // branches to us and our successor, fold the comparison into the
5715 // predecessor and use logical operations to update the incoming value
5716 // for PHI nodes in common successor.
5717 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
5718 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5722 static BasicBlock *allPredecessorsComeFromSameSource(BasicBlock *BB) {
5723 BasicBlock *PredPred = nullptr;
5724 for (auto *P : predecessors(BB)) {
5725 BasicBlock *PPred = P->getSinglePredecessor();
5726 if (!PPred || (PredPred && PredPred != PPred))
5733 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
5734 BasicBlock *BB = BI->getParent();
5736 // Conditional branch
5737 if (isValueEqualityComparison(BI)) {
5738 // If we only have one predecessor, and if it is a branch on this value,
5739 // see if that predecessor totally determines the outcome of this
5741 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5742 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
5743 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5745 // This block must be empty, except for the setcond inst, if it exists.
5746 // Ignore dbg intrinsics.
5747 BasicBlock::iterator I = BB->begin();
5748 // Ignore dbg intrinsics.
5749 while (isa<DbgInfoIntrinsic>(I))
5752 if (FoldValueComparisonIntoPredecessors(BI, Builder))
5753 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5754 } else if (&*I == cast<Instruction>(BI->getCondition())) {
5756 // Ignore dbg intrinsics.
5757 while (isa<DbgInfoIntrinsic>(I))
5759 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
5760 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5764 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
5765 if (SimplifyBranchOnICmpChain(BI, Builder, DL))
5768 // If this basic block has a single dominating predecessor block and the
5769 // dominating block's condition implies BI's condition, we know the direction
5770 // of the BI branch.
5771 if (BasicBlock *Dom = BB->getSinglePredecessor()) {
5772 auto *PBI = dyn_cast_or_null<BranchInst>(Dom->getTerminator());
5773 if (PBI && PBI->isConditional() &&
5774 PBI->getSuccessor(0) != PBI->getSuccessor(1) &&
5775 (PBI->getSuccessor(0) == BB || PBI->getSuccessor(1) == BB)) {
5776 bool CondIsFalse = PBI->getSuccessor(1) == BB;
5777 Optional<bool> Implication = isImpliedCondition(
5778 PBI->getCondition(), BI->getCondition(), DL, CondIsFalse);
5780 // Turn this into a branch on constant.
5781 auto *OldCond = BI->getCondition();
5782 ConstantInt *CI = *Implication
5783 ? ConstantInt::getTrue(BB->getContext())
5784 : ConstantInt::getFalse(BB->getContext());
5785 BI->setCondition(CI);
5786 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
5787 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5792 // If this basic block is ONLY a compare and a branch, and if a predecessor
5793 // branches to us and one of our successors, fold the comparison into the
5794 // predecessor and use logical operations to pick the right destination.
5795 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
5796 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5798 // We have a conditional branch to two blocks that are only reachable
5799 // from BI. We know that the condbr dominates the two blocks, so see if
5800 // there is any identical code in the "then" and "else" blocks. If so, we
5801 // can hoist it up to the branching block.
5802 if (BI->getSuccessor(0)->getSinglePredecessor()) {
5803 if (BI->getSuccessor(1)->getSinglePredecessor()) {
5804 if (HoistThenElseCodeToIf(BI, TTI))
5805 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5807 // If Successor #1 has multiple preds, we may be able to conditionally
5808 // execute Successor #0 if it branches to Successor #1.
5809 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
5810 if (Succ0TI->getNumSuccessors() == 1 &&
5811 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
5812 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), TTI))
5813 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5815 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
5816 // If Successor #0 has multiple preds, we may be able to conditionally
5817 // execute Successor #1 if it branches to Successor #0.
5818 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
5819 if (Succ1TI->getNumSuccessors() == 1 &&
5820 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
5821 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), TTI))
5822 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5825 // If this is a branch on a phi node in the current block, thread control
5826 // through this block if any PHI node entries are constants.
5827 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
5828 if (PN->getParent() == BI->getParent())
5829 if (FoldCondBranchOnPHI(BI, DL))
5830 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5832 // Scan predecessor blocks for conditional branches.
5833 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
5834 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
5835 if (PBI != BI && PBI->isConditional())
5836 if (SimplifyCondBranchToCondBranch(PBI, BI, DL))
5837 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5839 // Look for diamond patterns.
5840 if (MergeCondStores)
5841 if (BasicBlock *PrevBB = allPredecessorsComeFromSameSource(BB))
5842 if (BranchInst *PBI = dyn_cast<BranchInst>(PrevBB->getTerminator()))
5843 if (PBI != BI && PBI->isConditional())
5844 if (mergeConditionalStores(PBI, BI))
5845 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5850 /// Check if passing a value to an instruction will cause undefined behavior.
5851 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
5852 Constant *C = dyn_cast<Constant>(V);
5859 if (C->isNullValue() || isa<UndefValue>(C)) {
5860 // Only look at the first use, avoid hurting compile time with long uselists
5861 User *Use = *I->user_begin();
5863 // Now make sure that there are no instructions in between that can alter
5864 // control flow (eg. calls)
5865 for (BasicBlock::iterator
5866 i = ++BasicBlock::iterator(I),
5867 UI = BasicBlock::iterator(dyn_cast<Instruction>(Use));
5869 if (i == I->getParent()->end() || i->mayHaveSideEffects())
5872 // Look through GEPs. A load from a GEP derived from NULL is still undefined
5873 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
5874 if (GEP->getPointerOperand() == I)
5875 return passingValueIsAlwaysUndefined(V, GEP);
5877 // Look through bitcasts.
5878 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
5879 return passingValueIsAlwaysUndefined(V, BC);
5881 // Load from null is undefined.
5882 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
5883 if (!LI->isVolatile())
5884 return LI->getPointerAddressSpace() == 0;
5886 // Store to null is undefined.
5887 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
5888 if (!SI->isVolatile())
5889 return SI->getPointerAddressSpace() == 0 &&
5890 SI->getPointerOperand() == I;
5892 // A call to null is undefined.
5893 if (auto CS = CallSite(Use))
5894 return CS.getCalledValue() == I;
5899 /// If BB has an incoming value that will always trigger undefined behavior
5900 /// (eg. null pointer dereference), remove the branch leading here.
5901 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
5902 for (BasicBlock::iterator i = BB->begin();
5903 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
5904 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
5905 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
5906 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
5907 IRBuilder<> Builder(T);
5908 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
5909 BB->removePredecessor(PHI->getIncomingBlock(i));
5910 // Turn uncoditional branches into unreachables and remove the dead
5911 // destination from conditional branches.
5912 if (BI->isUnconditional())
5913 Builder.CreateUnreachable();
5915 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1)
5916 : BI->getSuccessor(0));
5917 BI->eraseFromParent();
5920 // TODO: SwitchInst.
5926 bool SimplifyCFGOpt::run(BasicBlock *BB) {
5927 bool Changed = false;
5929 assert(BB && BB->getParent() && "Block not embedded in function!");
5930 assert(BB->getTerminator() && "Degenerate basic block encountered!");
5932 // Remove basic blocks that have no predecessors (except the entry block)...
5933 // or that just have themself as a predecessor. These are unreachable.
5934 if ((pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) ||
5935 BB->getSinglePredecessor() == BB) {
5936 DEBUG(dbgs() << "Removing BB: \n" << *BB);
5937 DeleteDeadBlock(BB);
5941 // Check to see if we can constant propagate this terminator instruction
5943 Changed |= ConstantFoldTerminator(BB, true);
5945 // Check for and eliminate duplicate PHI nodes in this block.
5946 Changed |= EliminateDuplicatePHINodes(BB);
5948 // Check for and remove branches that will always cause undefined behavior.
5949 Changed |= removeUndefIntroducingPredecessor(BB);
5951 // Merge basic blocks into their predecessor if there is only one distinct
5952 // pred, and if there is only one distinct successor of the predecessor, and
5953 // if there are no PHI nodes.
5955 if (MergeBlockIntoPredecessor(BB))
5958 IRBuilder<> Builder(BB);
5960 // If there is a trivial two-entry PHI node in this basic block, and we can
5961 // eliminate it, do so now.
5962 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
5963 if (PN->getNumIncomingValues() == 2)
5964 Changed |= FoldTwoEntryPHINode(PN, TTI, DL);
5966 Builder.SetInsertPoint(BB->getTerminator());
5967 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
5968 if (BI->isUnconditional()) {
5969 if (SimplifyUncondBranch(BI, Builder))
5972 if (SimplifyCondBranch(BI, Builder))
5975 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
5976 if (SimplifyReturn(RI, Builder))
5978 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
5979 if (SimplifyResume(RI, Builder))
5981 } else if (CleanupReturnInst *RI =
5982 dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
5983 if (SimplifyCleanupReturn(RI))
5985 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
5986 if (SimplifySwitch(SI, Builder))
5988 } else if (UnreachableInst *UI =
5989 dyn_cast<UnreachableInst>(BB->getTerminator())) {
5990 if (SimplifyUnreachable(UI))
5992 } else if (IndirectBrInst *IBI =
5993 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
5994 if (SimplifyIndirectBr(IBI))
6001 /// This function is used to do simplification of a CFG.
6002 /// For example, it adjusts branches to branches to eliminate the extra hop,
6003 /// eliminates unreachable basic blocks, and does other "peephole" optimization
6004 /// of the CFG. It returns true if a modification was made.
6006 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
6007 unsigned BonusInstThreshold, AssumptionCache *AC,
6008 SmallPtrSetImpl<BasicBlock *> *LoopHeaders) {
6009 return SimplifyCFGOpt(TTI, BB->getModule()->getDataLayout(),
6010 BonusInstThreshold, AC, LoopHeaders)