1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Peephole optimize the CFG.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/ADT/APInt.h"
15 #include "llvm/ADT/ArrayRef.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/Optional.h"
18 #include "llvm/ADT/SetOperations.h"
19 #include "llvm/ADT/SetVector.h"
20 #include "llvm/ADT/SmallPtrSet.h"
21 #include "llvm/ADT/SmallSet.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/Analysis/ConstantFolding.h"
26 #include "llvm/Analysis/EHPersonalities.h"
27 #include "llvm/Analysis/InstructionSimplify.h"
28 #include "llvm/Analysis/TargetTransformInfo.h"
29 #include "llvm/Analysis/ValueTracking.h"
30 #include "llvm/IR/BasicBlock.h"
31 #include "llvm/IR/CallSite.h"
32 #include "llvm/IR/CFG.h"
33 #include "llvm/IR/Constant.h"
34 #include "llvm/IR/ConstantRange.h"
35 #include "llvm/IR/Constants.h"
36 #include "llvm/IR/DataLayout.h"
37 #include "llvm/IR/DebugInfo.h"
38 #include "llvm/IR/DerivedTypes.h"
39 #include "llvm/IR/GlobalValue.h"
40 #include "llvm/IR/GlobalVariable.h"
41 #include "llvm/IR/IRBuilder.h"
42 #include "llvm/IR/InstrTypes.h"
43 #include "llvm/IR/Instruction.h"
44 #include "llvm/IR/Instructions.h"
45 #include "llvm/IR/IntrinsicInst.h"
46 #include "llvm/IR/Intrinsics.h"
47 #include "llvm/IR/LLVMContext.h"
48 #include "llvm/IR/MDBuilder.h"
49 #include "llvm/IR/Metadata.h"
50 #include "llvm/IR/Module.h"
51 #include "llvm/IR/NoFolder.h"
52 #include "llvm/IR/Operator.h"
53 #include "llvm/IR/PatternMatch.h"
54 #include "llvm/IR/Type.h"
55 #include "llvm/IR/User.h"
56 #include "llvm/IR/Value.h"
57 #include "llvm/IR/DebugInfo.h"
58 #include "llvm/Support/Casting.h"
59 #include "llvm/Support/CommandLine.h"
60 #include "llvm/Support/Debug.h"
61 #include "llvm/Support/ErrorHandling.h"
62 #include "llvm/Support/MathExtras.h"
63 #include "llvm/Support/raw_ostream.h"
64 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
65 #include "llvm/Transforms/Utils/Local.h"
66 #include "llvm/Transforms/Utils/ValueMapper.h"
79 using namespace PatternMatch;
81 #define DEBUG_TYPE "simplifycfg"
83 // Chosen as 2 so as to be cheap, but still to have enough power to fold
84 // a select, so the "clamp" idiom (of a min followed by a max) will be caught.
85 // To catch this, we need to fold a compare and a select, hence '2' being the
86 // minimum reasonable default.
87 static cl::opt<unsigned> PHINodeFoldingThreshold(
88 "phi-node-folding-threshold", cl::Hidden, cl::init(2),
90 "Control the amount of phi node folding to perform (default = 2)"));
92 static cl::opt<bool> DupRet(
93 "simplifycfg-dup-ret", cl::Hidden, cl::init(false),
94 cl::desc("Duplicate return instructions into unconditional branches"));
97 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
98 cl::desc("Sink common instructions down to the end block"));
100 static cl::opt<bool> HoistCondStores(
101 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
102 cl::desc("Hoist conditional stores if an unconditional store precedes"));
104 static cl::opt<bool> MergeCondStores(
105 "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
106 cl::desc("Hoist conditional stores even if an unconditional store does not "
107 "precede - hoist multiple conditional stores into a single "
108 "predicated store"));
110 static cl::opt<bool> MergeCondStoresAggressively(
111 "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
112 cl::desc("When merging conditional stores, do so even if the resultant "
113 "basic blocks are unlikely to be if-converted as a result"));
115 static cl::opt<bool> SpeculateOneExpensiveInst(
116 "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
117 cl::desc("Allow exactly one expensive instruction to be speculatively "
120 static cl::opt<unsigned> MaxSpeculationDepth(
121 "max-speculation-depth", cl::Hidden, cl::init(10),
122 cl::desc("Limit maximum recursion depth when calculating costs of "
123 "speculatively executed instructions"));
125 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
126 STATISTIC(NumLinearMaps,
127 "Number of switch instructions turned into linear mapping");
128 STATISTIC(NumLookupTables,
129 "Number of switch instructions turned into lookup tables");
131 NumLookupTablesHoles,
132 "Number of switch instructions turned into lookup tables (holes checked)");
133 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
134 STATISTIC(NumSinkCommons,
135 "Number of common instructions sunk down to the end block");
136 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
140 // The first field contains the value that the switch produces when a certain
141 // case group is selected, and the second field is a vector containing the
142 // cases composing the case group.
143 typedef SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>
144 SwitchCaseResultVectorTy;
145 // The first field contains the phi node that generates a result of the switch
146 // and the second field contains the value generated for a certain case in the
147 // switch for that PHI.
148 typedef SmallVector<std::pair<PHINode *, Constant *>, 4> SwitchCaseResultsTy;
150 /// ValueEqualityComparisonCase - Represents a case of a switch.
151 struct ValueEqualityComparisonCase {
155 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
156 : Value(Value), Dest(Dest) {}
158 bool operator<(ValueEqualityComparisonCase RHS) const {
159 // Comparing pointers is ok as we only rely on the order for uniquing.
160 return Value < RHS.Value;
163 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
166 class SimplifyCFGOpt {
167 const TargetTransformInfo &TTI;
168 const DataLayout &DL;
169 unsigned BonusInstThreshold;
171 SmallPtrSetImpl<BasicBlock *> *LoopHeaders;
172 Value *isValueEqualityComparison(TerminatorInst *TI);
173 BasicBlock *GetValueEqualityComparisonCases(
174 TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases);
175 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
177 IRBuilder<> &Builder);
178 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
179 IRBuilder<> &Builder);
181 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
182 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
183 bool SimplifySingleResume(ResumeInst *RI);
184 bool SimplifyCommonResume(ResumeInst *RI);
185 bool SimplifyCleanupReturn(CleanupReturnInst *RI);
186 bool SimplifyUnreachable(UnreachableInst *UI);
187 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
188 bool SimplifyIndirectBr(IndirectBrInst *IBI);
189 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
190 bool SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
193 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
194 unsigned BonusInstThreshold, AssumptionCache *AC,
195 SmallPtrSetImpl<BasicBlock *> *LoopHeaders)
196 : TTI(TTI), DL(DL), BonusInstThreshold(BonusInstThreshold), AC(AC),
197 LoopHeaders(LoopHeaders) {}
199 bool run(BasicBlock *BB);
202 } // end anonymous namespace
204 /// Return true if it is safe to merge these two
205 /// terminator instructions together.
207 SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2,
208 SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
210 return false; // Can't merge with self!
212 // It is not safe to merge these two switch instructions if they have a common
213 // successor, and if that successor has a PHI node, and if *that* PHI node has
214 // conflicting incoming values from the two switch blocks.
215 BasicBlock *SI1BB = SI1->getParent();
216 BasicBlock *SI2BB = SI2->getParent();
218 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
220 for (BasicBlock *Succ : successors(SI2BB))
221 if (SI1Succs.count(Succ))
222 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
223 PHINode *PN = cast<PHINode>(BBI);
224 if (PN->getIncomingValueForBlock(SI1BB) !=
225 PN->getIncomingValueForBlock(SI2BB)) {
227 FailBlocks->insert(Succ);
235 /// Return true if it is safe and profitable to merge these two terminator
236 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
237 /// store all PHI nodes in common successors.
239 isProfitableToFoldUnconditional(BranchInst *SI1, BranchInst *SI2,
241 SmallVectorImpl<PHINode *> &PhiNodes) {
243 return false; // Can't merge with self!
244 assert(SI1->isUnconditional() && SI2->isConditional());
246 // We fold the unconditional branch if we can easily update all PHI nodes in
247 // common successors:
248 // 1> We have a constant incoming value for the conditional branch;
249 // 2> We have "Cond" as the incoming value for the unconditional branch;
250 // 3> SI2->getCondition() and Cond have same operands.
251 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
254 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
255 Cond->getOperand(1) == Ci2->getOperand(1)) &&
256 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
257 Cond->getOperand(1) == Ci2->getOperand(0)))
260 BasicBlock *SI1BB = SI1->getParent();
261 BasicBlock *SI2BB = SI2->getParent();
262 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
263 for (BasicBlock *Succ : successors(SI2BB))
264 if (SI1Succs.count(Succ))
265 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
266 PHINode *PN = cast<PHINode>(BBI);
267 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
268 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
270 PhiNodes.push_back(PN);
275 /// Update PHI nodes in Succ to indicate that there will now be entries in it
276 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
277 /// will be the same as those coming in from ExistPred, an existing predecessor
279 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
280 BasicBlock *ExistPred) {
281 if (!isa<PHINode>(Succ->begin()))
282 return; // Quick exit if nothing to do
285 for (BasicBlock::iterator I = Succ->begin(); (PN = dyn_cast<PHINode>(I)); ++I)
286 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
289 /// Compute an abstract "cost" of speculating the given instruction,
290 /// which is assumed to be safe to speculate. TCC_Free means cheap,
291 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
293 static unsigned ComputeSpeculationCost(const User *I,
294 const TargetTransformInfo &TTI) {
295 assert(isSafeToSpeculativelyExecute(I) &&
296 "Instruction is not safe to speculatively execute!");
297 return TTI.getUserCost(I);
300 /// If we have a merge point of an "if condition" as accepted above,
301 /// return true if the specified value dominates the block. We
302 /// don't handle the true generality of domination here, just a special case
303 /// which works well enough for us.
305 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
306 /// see if V (which must be an instruction) and its recursive operands
307 /// that do not dominate BB have a combined cost lower than CostRemaining and
308 /// are non-trapping. If both are true, the instruction is inserted into the
309 /// set and true is returned.
311 /// The cost for most non-trapping instructions is defined as 1 except for
312 /// Select whose cost is 2.
314 /// After this function returns, CostRemaining is decreased by the cost of
315 /// V plus its non-dominating operands. If that cost is greater than
316 /// CostRemaining, false is returned and CostRemaining is undefined.
317 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
318 SmallPtrSetImpl<Instruction *> *AggressiveInsts,
319 unsigned &CostRemaining,
320 const TargetTransformInfo &TTI,
321 unsigned Depth = 0) {
322 // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
323 // so limit the recursion depth.
324 // TODO: While this recursion limit does prevent pathological behavior, it
325 // would be better to track visited instructions to avoid cycles.
326 if (Depth == MaxSpeculationDepth)
329 Instruction *I = dyn_cast<Instruction>(V);
331 // Non-instructions all dominate instructions, but not all constantexprs
332 // can be executed unconditionally.
333 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
338 BasicBlock *PBB = I->getParent();
340 // We don't want to allow weird loops that might have the "if condition" in
341 // the bottom of this block.
345 // If this instruction is defined in a block that contains an unconditional
346 // branch to BB, then it must be in the 'conditional' part of the "if
347 // statement". If not, it definitely dominates the region.
348 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
349 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
352 // If we aren't allowing aggressive promotion anymore, then don't consider
353 // instructions in the 'if region'.
354 if (!AggressiveInsts)
357 // If we have seen this instruction before, don't count it again.
358 if (AggressiveInsts->count(I))
361 // Okay, it looks like the instruction IS in the "condition". Check to
362 // see if it's a cheap instruction to unconditionally compute, and if it
363 // only uses stuff defined outside of the condition. If so, hoist it out.
364 if (!isSafeToSpeculativelyExecute(I))
367 unsigned Cost = ComputeSpeculationCost(I, TTI);
369 // Allow exactly one instruction to be speculated regardless of its cost
370 // (as long as it is safe to do so).
371 // This is intended to flatten the CFG even if the instruction is a division
372 // or other expensive operation. The speculation of an expensive instruction
373 // is expected to be undone in CodeGenPrepare if the speculation has not
374 // enabled further IR optimizations.
375 if (Cost > CostRemaining &&
376 (!SpeculateOneExpensiveInst || !AggressiveInsts->empty() || Depth > 0))
379 // Avoid unsigned wrap.
380 CostRemaining = (Cost > CostRemaining) ? 0 : CostRemaining - Cost;
382 // Okay, we can only really hoist these out if their operands do
383 // not take us over the cost threshold.
384 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
385 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, TTI,
388 // Okay, it's safe to do this! Remember this instruction.
389 AggressiveInsts->insert(I);
393 /// Extract ConstantInt from value, looking through IntToPtr
394 /// and PointerNullValue. Return NULL if value is not a constant int.
395 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
396 // Normal constant int.
397 ConstantInt *CI = dyn_cast<ConstantInt>(V);
398 if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
401 // This is some kind of pointer constant. Turn it into a pointer-sized
402 // ConstantInt if possible.
403 IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
405 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
406 if (isa<ConstantPointerNull>(V))
407 return ConstantInt::get(PtrTy, 0);
409 // IntToPtr const int.
410 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
411 if (CE->getOpcode() == Instruction::IntToPtr)
412 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
413 // The constant is very likely to have the right type already.
414 if (CI->getType() == PtrTy)
417 return cast<ConstantInt>(
418 ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
425 /// Given a chain of or (||) or and (&&) comparison of a value against a
426 /// constant, this will try to recover the information required for a switch
428 /// It will depth-first traverse the chain of comparison, seeking for patterns
429 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
430 /// representing the different cases for the switch.
431 /// Note that if the chain is composed of '||' it will build the set of elements
432 /// that matches the comparisons (i.e. any of this value validate the chain)
433 /// while for a chain of '&&' it will build the set elements that make the test
435 struct ConstantComparesGatherer {
436 const DataLayout &DL;
437 Value *CompValue; /// Value found for the switch comparison
438 Value *Extra; /// Extra clause to be checked before the switch
439 SmallVector<ConstantInt *, 8> Vals; /// Set of integers to match in switch
440 unsigned UsedICmps; /// Number of comparisons matched in the and/or chain
442 /// Construct and compute the result for the comparison instruction Cond
443 ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL)
444 : DL(DL), CompValue(nullptr), Extra(nullptr), UsedICmps(0) {
449 ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
450 ConstantComparesGatherer &
451 operator=(const ConstantComparesGatherer &) = delete;
454 /// Try to set the current value used for the comparison, it succeeds only if
455 /// it wasn't set before or if the new value is the same as the old one
456 bool setValueOnce(Value *NewVal) {
457 if (CompValue && CompValue != NewVal)
460 return (CompValue != nullptr);
463 /// Try to match Instruction "I" as a comparison against a constant and
464 /// populates the array Vals with the set of values that match (or do not
465 /// match depending on isEQ).
466 /// Return false on failure. On success, the Value the comparison matched
467 /// against is placed in CompValue.
468 /// If CompValue is already set, the function is expected to fail if a match
469 /// is found but the value compared to is different.
470 bool matchInstruction(Instruction *I, bool isEQ) {
471 // If this is an icmp against a constant, handle this as one of the cases.
474 if (!((ICI = dyn_cast<ICmpInst>(I)) &&
475 (C = GetConstantInt(I->getOperand(1), DL)))) {
482 // Pattern match a special case
483 // (x & ~2^z) == y --> x == y || x == y|2^z
484 // This undoes a transformation done by instcombine to fuse 2 compares.
485 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
487 // It's a little bit hard to see why the following transformations are
488 // correct. Here is a CVC3 program to verify them for 64-bit values:
491 ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
495 mask : BITVECTOR(64) = BVSHL(ONE, z);
496 QUERY( (y & ~mask = y) =>
497 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
499 QUERY( (y | mask = y) =>
500 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
504 // Please note that each pattern must be a dual implication (<--> or
505 // iff). One directional implication can create spurious matches. If the
506 // implication is only one-way, an unsatisfiable condition on the left
507 // side can imply a satisfiable condition on the right side. Dual
508 // implication ensures that satisfiable conditions are transformed to
509 // other satisfiable conditions and unsatisfiable conditions are
510 // transformed to other unsatisfiable conditions.
512 // Here is a concrete example of a unsatisfiable condition on the left
513 // implying a satisfiable condition on the right:
516 // (x & ~mask) == y --> (x == y || x == (y | mask))
518 // Substituting y = 3, z = 0 yields:
519 // (x & -2) == 3 --> (x == 3 || x == 2)
521 // Pattern match a special case:
523 QUERY( (y & ~mask = y) =>
524 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
527 if (match(ICI->getOperand(0),
528 m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
530 if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
531 // If we already have a value for the switch, it has to match!
532 if (!setValueOnce(RHSVal))
537 ConstantInt::get(C->getContext(),
538 C->getValue() | Mask));
544 // Pattern match a special case:
546 QUERY( (y | mask = y) =>
547 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
550 if (match(ICI->getOperand(0),
551 m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
553 if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
554 // If we already have a value for the switch, it has to match!
555 if (!setValueOnce(RHSVal))
559 Vals.push_back(ConstantInt::get(C->getContext(),
560 C->getValue() & ~Mask));
566 // If we already have a value for the switch, it has to match!
567 if (!setValueOnce(ICI->getOperand(0)))
572 return ICI->getOperand(0);
575 // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
576 ConstantRange Span = ConstantRange::makeAllowedICmpRegion(
577 ICI->getPredicate(), C->getValue());
579 // Shift the range if the compare is fed by an add. This is the range
580 // compare idiom as emitted by instcombine.
581 Value *CandidateVal = I->getOperand(0);
582 if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
583 Span = Span.subtract(*RHSC);
584 CandidateVal = RHSVal;
587 // If this is an and/!= check, then we are looking to build the set of
588 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
591 Span = Span.inverse();
593 // If there are a ton of values, we don't want to make a ginormous switch.
594 if (Span.getSetSize().ugt(8) || Span.isEmptySet()) {
598 // If we already have a value for the switch, it has to match!
599 if (!setValueOnce(CandidateVal))
602 // Add all values from the range to the set
603 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
604 Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
610 /// Given a potentially 'or'd or 'and'd together collection of icmp
611 /// eq/ne/lt/gt instructions that compare a value against a constant, extract
612 /// the value being compared, and stick the list constants into the Vals
614 /// One "Extra" case is allowed to differ from the other.
615 void gather(Value *V) {
616 Instruction *I = dyn_cast<Instruction>(V);
617 bool isEQ = (I->getOpcode() == Instruction::Or);
619 // Keep a stack (SmallVector for efficiency) for depth-first traversal
620 SmallVector<Value *, 8> DFT;
621 SmallPtrSet<Value *, 8> Visited;
627 while (!DFT.empty()) {
628 V = DFT.pop_back_val();
630 if (Instruction *I = dyn_cast<Instruction>(V)) {
631 // If it is a || (or && depending on isEQ), process the operands.
632 if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
633 if (Visited.insert(I->getOperand(1)).second)
634 DFT.push_back(I->getOperand(1));
635 if (Visited.insert(I->getOperand(0)).second)
636 DFT.push_back(I->getOperand(0));
640 // Try to match the current instruction
641 if (matchInstruction(I, isEQ))
642 // Match succeed, continue the loop
646 // One element of the sequence of || (or &&) could not be match as a
647 // comparison against the same value as the others.
648 // We allow only one "Extra" case to be checked before the switch
653 // Failed to parse a proper sequence, abort now
660 } // end anonymous namespace
662 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
663 Instruction *Cond = nullptr;
664 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
665 Cond = dyn_cast<Instruction>(SI->getCondition());
666 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
667 if (BI->isConditional())
668 Cond = dyn_cast<Instruction>(BI->getCondition());
669 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
670 Cond = dyn_cast<Instruction>(IBI->getAddress());
673 TI->eraseFromParent();
675 RecursivelyDeleteTriviallyDeadInstructions(Cond);
678 /// Return true if the specified terminator checks
679 /// to see if a value is equal to constant integer value.
680 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
682 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
683 // Do not permit merging of large switch instructions into their
684 // predecessors unless there is only one predecessor.
685 if (SI->getNumSuccessors() * std::distance(pred_begin(SI->getParent()),
686 pred_end(SI->getParent())) <=
688 CV = SI->getCondition();
689 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
690 if (BI->isConditional() && BI->getCondition()->hasOneUse())
691 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
692 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
693 CV = ICI->getOperand(0);
696 // Unwrap any lossless ptrtoint cast.
698 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
699 Value *Ptr = PTII->getPointerOperand();
700 if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
707 /// Given a value comparison instruction,
708 /// decode all of the 'cases' that it represents and return the 'default' block.
709 BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
710 TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
711 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
712 Cases.reserve(SI->getNumCases());
713 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e;
716 ValueEqualityComparisonCase(i.getCaseValue(), i.getCaseSuccessor()));
717 return SI->getDefaultDest();
720 BranchInst *BI = cast<BranchInst>(TI);
721 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
722 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
723 Cases.push_back(ValueEqualityComparisonCase(
724 GetConstantInt(ICI->getOperand(1), DL), Succ));
725 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
728 /// Given a vector of bb/value pairs, remove any entries
729 /// in the list that match the specified block.
731 EliminateBlockCases(BasicBlock *BB,
732 std::vector<ValueEqualityComparisonCase> &Cases) {
733 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
736 /// Return true if there are any keys in C1 that exist in C2 as well.
737 static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
738 std::vector<ValueEqualityComparisonCase> &C2) {
739 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
741 // Make V1 be smaller than V2.
742 if (V1->size() > V2->size())
747 if (V1->size() == 1) {
749 ConstantInt *TheVal = (*V1)[0].Value;
750 for (unsigned i = 0, e = V2->size(); i != e; ++i)
751 if (TheVal == (*V2)[i].Value)
755 // Otherwise, just sort both lists and compare element by element.
756 array_pod_sort(V1->begin(), V1->end());
757 array_pod_sort(V2->begin(), V2->end());
758 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
759 while (i1 != e1 && i2 != e2) {
760 if ((*V1)[i1].Value == (*V2)[i2].Value)
762 if ((*V1)[i1].Value < (*V2)[i2].Value)
770 /// If TI is known to be a terminator instruction and its block is known to
771 /// only have a single predecessor block, check to see if that predecessor is
772 /// also a value comparison with the same value, and if that comparison
773 /// determines the outcome of this comparison. If so, simplify TI. This does a
774 /// very limited form of jump threading.
775 bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
776 TerminatorInst *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
777 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
779 return false; // Not a value comparison in predecessor.
781 Value *ThisVal = isValueEqualityComparison(TI);
782 assert(ThisVal && "This isn't a value comparison!!");
783 if (ThisVal != PredVal)
784 return false; // Different predicates.
786 // TODO: Preserve branch weight metadata, similarly to how
787 // FoldValueComparisonIntoPredecessors preserves it.
789 // Find out information about when control will move from Pred to TI's block.
790 std::vector<ValueEqualityComparisonCase> PredCases;
791 BasicBlock *PredDef =
792 GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
793 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
795 // Find information about how control leaves this block.
796 std::vector<ValueEqualityComparisonCase> ThisCases;
797 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
798 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
800 // If TI's block is the default block from Pred's comparison, potentially
801 // simplify TI based on this knowledge.
802 if (PredDef == TI->getParent()) {
803 // If we are here, we know that the value is none of those cases listed in
804 // PredCases. If there are any cases in ThisCases that are in PredCases, we
806 if (!ValuesOverlap(PredCases, ThisCases))
809 if (isa<BranchInst>(TI)) {
810 // Okay, one of the successors of this condbr is dead. Convert it to a
812 assert(ThisCases.size() == 1 && "Branch can only have one case!");
813 // Insert the new branch.
814 Instruction *NI = Builder.CreateBr(ThisDef);
817 // Remove PHI node entries for the dead edge.
818 ThisCases[0].Dest->removePredecessor(TI->getParent());
820 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
821 << "Through successor TI: " << *TI << "Leaving: " << *NI
824 EraseTerminatorInstAndDCECond(TI);
828 SwitchInst *SI = cast<SwitchInst>(TI);
829 // Okay, TI has cases that are statically dead, prune them away.
830 SmallPtrSet<Constant *, 16> DeadCases;
831 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
832 DeadCases.insert(PredCases[i].Value);
834 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
835 << "Through successor TI: " << *TI);
837 // Collect branch weights into a vector.
838 SmallVector<uint32_t, 8> Weights;
839 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
840 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
842 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
844 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
845 Weights.push_back(CI->getValue().getZExtValue());
847 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
849 if (DeadCases.count(i.getCaseValue())) {
851 std::swap(Weights[i.getCaseIndex() + 1], Weights.back());
854 i.getCaseSuccessor()->removePredecessor(TI->getParent());
858 if (HasWeight && Weights.size() >= 2)
859 SI->setMetadata(LLVMContext::MD_prof,
860 MDBuilder(SI->getParent()->getContext())
861 .createBranchWeights(Weights));
863 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
867 // Otherwise, TI's block must correspond to some matched value. Find out
868 // which value (or set of values) this is.
869 ConstantInt *TIV = nullptr;
870 BasicBlock *TIBB = TI->getParent();
871 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
872 if (PredCases[i].Dest == TIBB) {
874 return false; // Cannot handle multiple values coming to this block.
875 TIV = PredCases[i].Value;
877 assert(TIV && "No edge from pred to succ?");
879 // Okay, we found the one constant that our value can be if we get into TI's
880 // BB. Find out which successor will unconditionally be branched to.
881 BasicBlock *TheRealDest = nullptr;
882 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
883 if (ThisCases[i].Value == TIV) {
884 TheRealDest = ThisCases[i].Dest;
888 // If not handled by any explicit cases, it is handled by the default case.
890 TheRealDest = ThisDef;
892 // Remove PHI node entries for dead edges.
893 BasicBlock *CheckEdge = TheRealDest;
894 for (BasicBlock *Succ : successors(TIBB))
895 if (Succ != CheckEdge)
896 Succ->removePredecessor(TIBB);
900 // Insert the new branch.
901 Instruction *NI = Builder.CreateBr(TheRealDest);
904 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
905 << "Through successor TI: " << *TI << "Leaving: " << *NI
908 EraseTerminatorInstAndDCECond(TI);
914 /// This class implements a stable ordering of constant
915 /// integers that does not depend on their address. This is important for
916 /// applications that sort ConstantInt's to ensure uniqueness.
917 struct ConstantIntOrdering {
918 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
919 return LHS->getValue().ult(RHS->getValue());
923 } // end anonymous namespace
925 static int ConstantIntSortPredicate(ConstantInt *const *P1,
926 ConstantInt *const *P2) {
927 const ConstantInt *LHS = *P1;
928 const ConstantInt *RHS = *P2;
931 return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
934 static inline bool HasBranchWeights(const Instruction *I) {
935 MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
936 if (ProfMD && ProfMD->getOperand(0))
937 if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
938 return MDS->getString().equals("branch_weights");
943 /// Get Weights of a given TerminatorInst, the default weight is at the front
944 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
946 static void GetBranchWeights(TerminatorInst *TI,
947 SmallVectorImpl<uint64_t> &Weights) {
948 MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
950 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
951 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
952 Weights.push_back(CI->getValue().getZExtValue());
955 // If TI is a conditional eq, the default case is the false case,
956 // and the corresponding branch-weight data is at index 2. We swap the
957 // default weight to be the first entry.
958 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
959 assert(Weights.size() == 2);
960 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
961 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
962 std::swap(Weights.front(), Weights.back());
966 /// Keep halving the weights until all can fit in uint32_t.
967 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
968 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
969 if (Max > UINT_MAX) {
970 unsigned Offset = 32 - countLeadingZeros(Max);
971 for (uint64_t &I : Weights)
976 /// The specified terminator is a value equality comparison instruction
977 /// (either a switch or a branch on "X == c").
978 /// See if any of the predecessors of the terminator block are value comparisons
979 /// on the same value. If so, and if safe to do so, fold them together.
980 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
981 IRBuilder<> &Builder) {
982 BasicBlock *BB = TI->getParent();
983 Value *CV = isValueEqualityComparison(TI); // CondVal
984 assert(CV && "Not a comparison?");
985 bool Changed = false;
987 SmallVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
988 while (!Preds.empty()) {
989 BasicBlock *Pred = Preds.pop_back_val();
991 // See if the predecessor is a comparison with the same value.
992 TerminatorInst *PTI = Pred->getTerminator();
993 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
995 if (PCV == CV && TI != PTI) {
996 SmallSetVector<BasicBlock*, 4> FailBlocks;
997 if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
998 for (auto *Succ : FailBlocks) {
999 std::vector<BasicBlock*> Blocks = { TI->getParent() };
1000 if (!SplitBlockPredecessors(Succ, Blocks, ".fold.split"))
1005 // Figure out which 'cases' to copy from SI to PSI.
1006 std::vector<ValueEqualityComparisonCase> BBCases;
1007 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1009 std::vector<ValueEqualityComparisonCase> PredCases;
1010 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1012 // Based on whether the default edge from PTI goes to BB or not, fill in
1013 // PredCases and PredDefault with the new switch cases we would like to
1015 SmallVector<BasicBlock *, 8> NewSuccessors;
1017 // Update the branch weight metadata along the way
1018 SmallVector<uint64_t, 8> Weights;
1019 bool PredHasWeights = HasBranchWeights(PTI);
1020 bool SuccHasWeights = HasBranchWeights(TI);
1022 if (PredHasWeights) {
1023 GetBranchWeights(PTI, Weights);
1024 // branch-weight metadata is inconsistent here.
1025 if (Weights.size() != 1 + PredCases.size())
1026 PredHasWeights = SuccHasWeights = false;
1027 } else if (SuccHasWeights)
1028 // If there are no predecessor weights but there are successor weights,
1029 // populate Weights with 1, which will later be scaled to the sum of
1030 // successor's weights
1031 Weights.assign(1 + PredCases.size(), 1);
1033 SmallVector<uint64_t, 8> SuccWeights;
1034 if (SuccHasWeights) {
1035 GetBranchWeights(TI, SuccWeights);
1036 // branch-weight metadata is inconsistent here.
1037 if (SuccWeights.size() != 1 + BBCases.size())
1038 PredHasWeights = SuccHasWeights = false;
1039 } else if (PredHasWeights)
1040 SuccWeights.assign(1 + BBCases.size(), 1);
1042 if (PredDefault == BB) {
1043 // If this is the default destination from PTI, only the edges in TI
1044 // that don't occur in PTI, or that branch to BB will be activated.
1045 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1046 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1047 if (PredCases[i].Dest != BB)
1048 PTIHandled.insert(PredCases[i].Value);
1050 // The default destination is BB, we don't need explicit targets.
1051 std::swap(PredCases[i], PredCases.back());
1053 if (PredHasWeights || SuccHasWeights) {
1054 // Increase weight for the default case.
1055 Weights[0] += Weights[i + 1];
1056 std::swap(Weights[i + 1], Weights.back());
1060 PredCases.pop_back();
1065 // Reconstruct the new switch statement we will be building.
1066 if (PredDefault != BBDefault) {
1067 PredDefault->removePredecessor(Pred);
1068 PredDefault = BBDefault;
1069 NewSuccessors.push_back(BBDefault);
1072 unsigned CasesFromPred = Weights.size();
1073 uint64_t ValidTotalSuccWeight = 0;
1074 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1075 if (!PTIHandled.count(BBCases[i].Value) &&
1076 BBCases[i].Dest != BBDefault) {
1077 PredCases.push_back(BBCases[i]);
1078 NewSuccessors.push_back(BBCases[i].Dest);
1079 if (SuccHasWeights || PredHasWeights) {
1080 // The default weight is at index 0, so weight for the ith case
1081 // should be at index i+1. Scale the cases from successor by
1082 // PredDefaultWeight (Weights[0]).
1083 Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1084 ValidTotalSuccWeight += SuccWeights[i + 1];
1088 if (SuccHasWeights || PredHasWeights) {
1089 ValidTotalSuccWeight += SuccWeights[0];
1090 // Scale the cases from predecessor by ValidTotalSuccWeight.
1091 for (unsigned i = 1; i < CasesFromPred; ++i)
1092 Weights[i] *= ValidTotalSuccWeight;
1093 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1094 Weights[0] *= SuccWeights[0];
1097 // If this is not the default destination from PSI, only the edges
1098 // in SI that occur in PSI with a destination of BB will be
1100 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1101 std::map<ConstantInt *, uint64_t> WeightsForHandled;
1102 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1103 if (PredCases[i].Dest == BB) {
1104 PTIHandled.insert(PredCases[i].Value);
1106 if (PredHasWeights || SuccHasWeights) {
1107 WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1108 std::swap(Weights[i + 1], Weights.back());
1112 std::swap(PredCases[i], PredCases.back());
1113 PredCases.pop_back();
1118 // Okay, now we know which constants were sent to BB from the
1119 // predecessor. Figure out where they will all go now.
1120 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1121 if (PTIHandled.count(BBCases[i].Value)) {
1122 // If this is one we are capable of getting...
1123 if (PredHasWeights || SuccHasWeights)
1124 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1125 PredCases.push_back(BBCases[i]);
1126 NewSuccessors.push_back(BBCases[i].Dest);
1128 BBCases[i].Value); // This constant is taken care of
1131 // If there are any constants vectored to BB that TI doesn't handle,
1132 // they must go to the default destination of TI.
1133 for (ConstantInt *I : PTIHandled) {
1134 if (PredHasWeights || SuccHasWeights)
1135 Weights.push_back(WeightsForHandled[I]);
1136 PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1137 NewSuccessors.push_back(BBDefault);
1141 // Okay, at this point, we know which new successor Pred will get. Make
1142 // sure we update the number of entries in the PHI nodes for these
1144 for (BasicBlock *NewSuccessor : NewSuccessors)
1145 AddPredecessorToBlock(NewSuccessor, Pred, BB);
1147 Builder.SetInsertPoint(PTI);
1148 // Convert pointer to int before we switch.
1149 if (CV->getType()->isPointerTy()) {
1150 CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
1154 // Now that the successors are updated, create the new Switch instruction.
1156 Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1157 NewSI->setDebugLoc(PTI->getDebugLoc());
1158 for (ValueEqualityComparisonCase &V : PredCases)
1159 NewSI->addCase(V.Value, V.Dest);
1161 if (PredHasWeights || SuccHasWeights) {
1162 // Halve the weights if any of them cannot fit in an uint32_t
1163 FitWeights(Weights);
1165 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1168 LLVMContext::MD_prof,
1169 MDBuilder(BB->getContext()).createBranchWeights(MDWeights));
1172 EraseTerminatorInstAndDCECond(PTI);
1174 // Okay, last check. If BB is still a successor of PSI, then we must
1175 // have an infinite loop case. If so, add an infinitely looping block
1176 // to handle the case to preserve the behavior of the code.
1177 BasicBlock *InfLoopBlock = nullptr;
1178 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1179 if (NewSI->getSuccessor(i) == BB) {
1180 if (!InfLoopBlock) {
1181 // Insert it at the end of the function, because it's either code,
1182 // or it won't matter if it's hot. :)
1183 InfLoopBlock = BasicBlock::Create(BB->getContext(), "infloop",
1185 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1187 NewSI->setSuccessor(i, InfLoopBlock);
1196 // If we would need to insert a select that uses the value of this invoke
1197 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1198 // can't hoist the invoke, as there is nowhere to put the select in this case.
1199 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1200 Instruction *I1, Instruction *I2) {
1201 for (BasicBlock *Succ : successors(BB1)) {
1203 for (BasicBlock::iterator BBI = Succ->begin();
1204 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1205 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1206 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1207 if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1215 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I);
1217 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1218 /// in the two blocks up into the branch block. The caller of this function
1219 /// guarantees that BI's block dominates BB1 and BB2.
1220 static bool HoistThenElseCodeToIf(BranchInst *BI,
1221 const TargetTransformInfo &TTI) {
1222 // This does very trivial matching, with limited scanning, to find identical
1223 // instructions in the two blocks. In particular, we don't want to get into
1224 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1225 // such, we currently just scan for obviously identical instructions in an
1227 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1228 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1230 BasicBlock::iterator BB1_Itr = BB1->begin();
1231 BasicBlock::iterator BB2_Itr = BB2->begin();
1233 Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1234 // Skip debug info if it is not identical.
1235 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1236 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1237 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1238 while (isa<DbgInfoIntrinsic>(I1))
1240 while (isa<DbgInfoIntrinsic>(I2))
1243 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1244 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1247 BasicBlock *BIParent = BI->getParent();
1249 bool Changed = false;
1251 // If we are hoisting the terminator instruction, don't move one (making a
1252 // broken BB), instead clone it, and remove BI.
1253 if (isa<TerminatorInst>(I1))
1254 goto HoistTerminator;
1256 if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1259 // For a normal instruction, we just move one to right before the branch,
1260 // then replace all uses of the other with the first. Finally, we remove
1261 // the now redundant second instruction.
1262 BIParent->getInstList().splice(BI->getIterator(), BB1->getInstList(), I1);
1263 if (!I2->use_empty())
1264 I2->replaceAllUsesWith(I1);
1266 unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1267 LLVMContext::MD_range,
1268 LLVMContext::MD_fpmath,
1269 LLVMContext::MD_invariant_load,
1270 LLVMContext::MD_nonnull,
1271 LLVMContext::MD_invariant_group,
1272 LLVMContext::MD_align,
1273 LLVMContext::MD_dereferenceable,
1274 LLVMContext::MD_dereferenceable_or_null,
1275 LLVMContext::MD_mem_parallel_loop_access};
1276 combineMetadata(I1, I2, KnownIDs);
1278 // I1 and I2 are being combined into a single instruction. Its debug
1279 // location is the merged locations of the original instructions.
1280 if (!isa<CallInst>(I1))
1282 DILocation::getMergedLocation(I1->getDebugLoc(), I2->getDebugLoc()));
1284 I2->eraseFromParent();
1289 // Skip debug info if it is not identical.
1290 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1291 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1292 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1293 while (isa<DbgInfoIntrinsic>(I1))
1295 while (isa<DbgInfoIntrinsic>(I2))
1298 } while (I1->isIdenticalToWhenDefined(I2));
1303 // It may not be possible to hoist an invoke.
1304 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1307 for (BasicBlock *Succ : successors(BB1)) {
1309 for (BasicBlock::iterator BBI = Succ->begin();
1310 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1311 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1312 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1316 // Check for passingValueIsAlwaysUndefined here because we would rather
1317 // eliminate undefined control flow then converting it to a select.
1318 if (passingValueIsAlwaysUndefined(BB1V, PN) ||
1319 passingValueIsAlwaysUndefined(BB2V, PN))
1322 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1324 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1329 // Okay, it is safe to hoist the terminator.
1330 Instruction *NT = I1->clone();
1331 BIParent->getInstList().insert(BI->getIterator(), NT);
1332 if (!NT->getType()->isVoidTy()) {
1333 I1->replaceAllUsesWith(NT);
1334 I2->replaceAllUsesWith(NT);
1338 IRBuilder<NoFolder> Builder(NT);
1339 // Hoisting one of the terminators from our successor is a great thing.
1340 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1341 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1342 // nodes, so we insert select instruction to compute the final result.
1343 std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1344 for (BasicBlock *Succ : successors(BB1)) {
1346 for (BasicBlock::iterator BBI = Succ->begin();
1347 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1348 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1349 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1353 // These values do not agree. Insert a select instruction before NT
1354 // that determines the right value.
1355 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1357 SI = cast<SelectInst>(
1358 Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1359 BB1V->getName() + "." + BB2V->getName(), BI));
1361 // Make the PHI node use the select for all incoming values for BB1/BB2
1362 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1363 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1364 PN->setIncomingValue(i, SI);
1368 // Update any PHI nodes in our new successors.
1369 for (BasicBlock *Succ : successors(BB1))
1370 AddPredecessorToBlock(Succ, BIParent, BB1);
1372 EraseTerminatorInstAndDCECond(BI);
1376 // Is it legal to place a variable in operand \c OpIdx of \c I?
1377 // FIXME: This should be promoted to Instruction.
1378 static bool canReplaceOperandWithVariable(const Instruction *I,
1380 // We can't have a PHI with a metadata type.
1381 if (I->getOperand(OpIdx)->getType()->isMetadataTy())
1385 if (!isa<Constant>(I->getOperand(OpIdx)))
1388 switch (I->getOpcode()) {
1391 case Instruction::Call:
1392 case Instruction::Invoke:
1393 // FIXME: many arithmetic intrinsics have no issue taking a
1394 // variable, however it's hard to distingish these from
1395 // specials such as @llvm.frameaddress that require a constant.
1396 if (isa<IntrinsicInst>(I))
1399 // Constant bundle operands may need to retain their constant-ness for
1401 if (ImmutableCallSite(I).isBundleOperand(OpIdx))
1406 case Instruction::ShuffleVector:
1407 // Shufflevector masks are constant.
1409 case Instruction::ExtractValue:
1410 case Instruction::InsertValue:
1411 // All operands apart from the first are constant.
1413 case Instruction::Alloca:
1415 case Instruction::GetElementPtr:
1418 gep_type_iterator It = std::next(gep_type_begin(I), OpIdx - 1);
1419 return It.isSequential();
1423 // All instructions in Insts belong to different blocks that all unconditionally
1424 // branch to a common successor. Analyze each instruction and return true if it
1425 // would be possible to sink them into their successor, creating one common
1426 // instruction instead. For every value that would be required to be provided by
1427 // PHI node (because an operand varies in each input block), add to PHIOperands.
1428 static bool canSinkInstructions(
1429 ArrayRef<Instruction *> Insts,
1430 DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1431 // Prune out obviously bad instructions to move. Any non-store instruction
1432 // must have exactly one use, and we check later that use is by a single,
1433 // common PHI instruction in the successor.
1434 for (auto *I : Insts) {
1435 // These instructions may change or break semantics if moved.
1436 if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1437 I->getType()->isTokenTy())
1440 // Conservatively return false if I is an inline-asm instruction. Sinking
1441 // and merging inline-asm instructions can potentially create arguments
1442 // that cannot satisfy the inline-asm constraints.
1443 if (const auto *C = dyn_cast<CallInst>(I))
1444 if (C->isInlineAsm())
1447 // Everything must have only one use too, apart from stores which
1449 if (!isa<StoreInst>(I) && !I->hasOneUse())
1453 const Instruction *I0 = Insts.front();
1454 for (auto *I : Insts)
1455 if (!I->isSameOperationAs(I0))
1458 // All instructions in Insts are known to be the same opcode. If they aren't
1459 // stores, check the only user of each is a PHI or in the same block as the
1460 // instruction, because if a user is in the same block as an instruction
1461 // we're contemplating sinking, it must already be determined to be sinkable.
1462 if (!isa<StoreInst>(I0)) {
1463 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1464 auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1465 if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1466 auto *U = cast<Instruction>(*I->user_begin());
1468 PNUse->getParent() == Succ &&
1469 PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1470 U->getParent() == I->getParent();
1475 for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1476 if (I0->getOperand(OI)->getType()->isTokenTy())
1477 // Don't touch any operand of token type.
1480 // Because SROA can't handle speculating stores of selects, try not
1481 // to sink loads or stores of allocas when we'd have to create a PHI for
1482 // the address operand. Also, because it is likely that loads or stores
1483 // of allocas will disappear when Mem2Reg/SROA is run, don't sink them.
1484 // This can cause code churn which can have unintended consequences down
1485 // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1486 // FIXME: This is a workaround for a deficiency in SROA - see
1487 // https://llvm.org/bugs/show_bug.cgi?id=30188
1488 if (OI == 1 && isa<StoreInst>(I0) &&
1489 any_of(Insts, [](const Instruction *I) {
1490 return isa<AllocaInst>(I->getOperand(1));
1493 if (OI == 0 && isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1494 return isa<AllocaInst>(I->getOperand(0));
1498 auto SameAsI0 = [&I0, OI](const Instruction *I) {
1499 assert(I->getNumOperands() == I0->getNumOperands());
1500 return I->getOperand(OI) == I0->getOperand(OI);
1502 if (!all_of(Insts, SameAsI0)) {
1503 if (!canReplaceOperandWithVariable(I0, OI))
1504 // We can't create a PHI from this GEP.
1506 // Don't create indirect calls! The called value is the final operand.
1507 if ((isa<CallInst>(I0) || isa<InvokeInst>(I0)) && OI == OE - 1) {
1508 // FIXME: if the call was *already* indirect, we should do this.
1511 for (auto *I : Insts)
1512 PHIOperands[I].push_back(I->getOperand(OI));
1518 // Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1519 // instruction of every block in Blocks to their common successor, commoning
1520 // into one instruction.
1521 static bool sinkLastInstruction(ArrayRef<BasicBlock*> Blocks) {
1522 auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1524 // canSinkLastInstruction returning true guarantees that every block has at
1525 // least one non-terminator instruction.
1526 SmallVector<Instruction*,4> Insts;
1527 for (auto *BB : Blocks) {
1528 Instruction *I = BB->getTerminator();
1530 I = I->getPrevNode();
1531 } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1532 if (!isa<DbgInfoIntrinsic>(I))
1536 // The only checking we need to do now is that all users of all instructions
1537 // are the same PHI node. canSinkLastInstruction should have checked this but
1538 // it is slightly over-aggressive - it gets confused by commutative instructions
1539 // so double-check it here.
1540 Instruction *I0 = Insts.front();
1541 if (!isa<StoreInst>(I0)) {
1542 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1543 if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1544 auto *U = cast<Instruction>(*I->user_begin());
1550 // We don't need to do any more checking here; canSinkLastInstruction should
1551 // have done it all for us.
1552 SmallVector<Value*, 4> NewOperands;
1553 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1554 // This check is different to that in canSinkLastInstruction. There, we
1555 // cared about the global view once simplifycfg (and instcombine) have
1556 // completed - it takes into account PHIs that become trivially
1557 // simplifiable. However here we need a more local view; if an operand
1558 // differs we create a PHI and rely on instcombine to clean up the very
1559 // small mess we may make.
1560 bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1561 return I->getOperand(O) != I0->getOperand(O);
1564 NewOperands.push_back(I0->getOperand(O));
1568 // Create a new PHI in the successor block and populate it.
1569 auto *Op = I0->getOperand(O);
1570 assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1571 auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1572 Op->getName() + ".sink", &BBEnd->front());
1573 for (auto *I : Insts)
1574 PN->addIncoming(I->getOperand(O), I->getParent());
1575 NewOperands.push_back(PN);
1578 // Arbitrarily use I0 as the new "common" instruction; remap its operands
1579 // and move it to the start of the successor block.
1580 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1581 I0->getOperandUse(O).set(NewOperands[O]);
1582 I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1584 // The debug location for the "common" instruction is the merged locations of
1585 // all the commoned instructions. We start with the original location of the
1586 // "common" instruction and iteratively merge each location in the loop below.
1587 const DILocation *Loc = I0->getDebugLoc();
1589 // Update metadata and IR flags, and merge debug locations.
1590 for (auto *I : Insts)
1592 Loc = DILocation::getMergedLocation(Loc, I->getDebugLoc());
1593 combineMetadataForCSE(I0, I);
1596 if (!isa<CallInst>(I0))
1597 I0->setDebugLoc(Loc);
1599 if (!isa<StoreInst>(I0)) {
1600 // canSinkLastInstruction checked that all instructions were used by
1601 // one and only one PHI node. Find that now, RAUW it to our common
1602 // instruction and nuke it.
1603 assert(I0->hasOneUse());
1604 auto *PN = cast<PHINode>(*I0->user_begin());
1605 PN->replaceAllUsesWith(I0);
1606 PN->eraseFromParent();
1609 // Finally nuke all instructions apart from the common instruction.
1610 for (auto *I : Insts)
1612 I->eraseFromParent();
1619 // LockstepReverseIterator - Iterates through instructions
1620 // in a set of blocks in reverse order from the first non-terminator.
1621 // For example (assume all blocks have size n):
1622 // LockstepReverseIterator I([B1, B2, B3]);
1623 // *I-- = [B1[n], B2[n], B3[n]];
1624 // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1625 // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1627 class LockstepReverseIterator {
1628 ArrayRef<BasicBlock*> Blocks;
1629 SmallVector<Instruction*,4> Insts;
1632 LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) :
1640 for (auto *BB : Blocks) {
1641 Instruction *Inst = BB->getTerminator();
1642 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1643 Inst = Inst->getPrevNode();
1645 // Block wasn't big enough.
1649 Insts.push_back(Inst);
1653 bool isValid() const {
1657 void operator -- () {
1660 for (auto *&Inst : Insts) {
1661 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1662 Inst = Inst->getPrevNode();
1663 // Already at beginning of block.
1671 ArrayRef<Instruction*> operator * () const {
1676 } // end anonymous namespace
1678 /// Given an unconditional branch that goes to BBEnd,
1679 /// check whether BBEnd has only two predecessors and the other predecessor
1680 /// ends with an unconditional branch. If it is true, sink any common code
1681 /// in the two predecessors to BBEnd.
1682 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1683 assert(BI1->isUnconditional());
1684 BasicBlock *BBEnd = BI1->getSuccessor(0);
1686 // We support two situations:
1687 // (1) all incoming arcs are unconditional
1688 // (2) one incoming arc is conditional
1690 // (2) is very common in switch defaults and
1691 // else-if patterns;
1694 // else if (b) f(2);
1707 // [end] has two unconditional predecessor arcs and one conditional. The
1708 // conditional refers to the implicit empty 'else' arc. This conditional
1709 // arc can also be caused by an empty default block in a switch.
1711 // In this case, we attempt to sink code from all *unconditional* arcs.
1712 // If we can sink instructions from these arcs (determined during the scan
1713 // phase below) we insert a common successor for all unconditional arcs and
1714 // connect that to [end], to enable sinking:
1727 SmallVector<BasicBlock*,4> UnconditionalPreds;
1728 Instruction *Cond = nullptr;
1729 for (auto *B : predecessors(BBEnd)) {
1730 auto *T = B->getTerminator();
1731 if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1732 UnconditionalPreds.push_back(B);
1733 else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1738 if (UnconditionalPreds.size() < 2)
1741 bool Changed = false;
1742 // We take a two-step approach to tail sinking. First we scan from the end of
1743 // each block upwards in lockstep. If the n'th instruction from the end of each
1744 // block can be sunk, those instructions are added to ValuesToSink and we
1745 // carry on. If we can sink an instruction but need to PHI-merge some operands
1746 // (because they're not identical in each instruction) we add these to
1748 unsigned ScanIdx = 0;
1749 SmallPtrSet<Value*,4> InstructionsToSink;
1750 DenseMap<Instruction*, SmallVector<Value*,4>> PHIOperands;
1751 LockstepReverseIterator LRI(UnconditionalPreds);
1752 while (LRI.isValid() &&
1753 canSinkInstructions(*LRI, PHIOperands)) {
1754 DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0] << "\n");
1755 InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1760 auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1761 unsigned NumPHIdValues = 0;
1762 for (auto *I : *LRI)
1763 for (auto *V : PHIOperands[I])
1764 if (InstructionsToSink.count(V) == 0)
1766 DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1767 unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1768 if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1771 return NumPHIInsts <= 1;
1774 if (ScanIdx > 0 && Cond) {
1775 // Check if we would actually sink anything first! This mutates the CFG and
1776 // adds an extra block. The goal in doing this is to allow instructions that
1777 // couldn't be sunk before to be sunk - obviously, speculatable instructions
1778 // (such as trunc, add) can be sunk and predicated already. So we check that
1779 // we're going to sink at least one non-speculatable instruction.
1782 bool Profitable = false;
1783 while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1784 if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1794 DEBUG(dbgs() << "SINK: Splitting edge\n");
1795 // We have a conditional edge and we're going to sink some instructions.
1796 // Insert a new block postdominating all blocks we're going to sink from.
1797 if (!SplitBlockPredecessors(BI1->getSuccessor(0), UnconditionalPreds,
1799 // Edges couldn't be split.
1804 // Now that we've analyzed all potential sinking candidates, perform the
1805 // actual sink. We iteratively sink the last non-terminator of the source
1806 // blocks into their common successor unless doing so would require too
1807 // many PHI instructions to be generated (currently only one PHI is allowed
1808 // per sunk instruction).
1810 // We can use InstructionsToSink to discount values needing PHI-merging that will
1811 // actually be sunk in a later iteration. This allows us to be more
1812 // aggressive in what we sink. This does allow a false positive where we
1813 // sink presuming a later value will also be sunk, but stop half way through
1814 // and never actually sink it which means we produce more PHIs than intended.
1815 // This is unlikely in practice though.
1816 for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1817 DEBUG(dbgs() << "SINK: Sink: "
1818 << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1821 // Because we've sunk every instruction in turn, the current instruction to
1822 // sink is always at index 0.
1824 if (!ProfitableToSinkInstruction(LRI)) {
1825 // Too many PHIs would be created.
1826 DEBUG(dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1830 if (!sinkLastInstruction(UnconditionalPreds))
1838 /// \brief Determine if we can hoist sink a sole store instruction out of a
1839 /// conditional block.
1841 /// We are looking for code like the following:
1843 /// store i32 %add, i32* %arrayidx2
1844 /// ... // No other stores or function calls (we could be calling a memory
1845 /// ... // function).
1846 /// %cmp = icmp ult %x, %y
1847 /// br i1 %cmp, label %EndBB, label %ThenBB
1849 /// store i32 %add5, i32* %arrayidx2
1853 /// We are going to transform this into:
1855 /// store i32 %add, i32* %arrayidx2
1857 /// %cmp = icmp ult %x, %y
1858 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1859 /// store i32 %add.add5, i32* %arrayidx2
1862 /// \return The pointer to the value of the previous store if the store can be
1863 /// hoisted into the predecessor block. 0 otherwise.
1864 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1865 BasicBlock *StoreBB, BasicBlock *EndBB) {
1866 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1870 // Volatile or atomic.
1871 if (!StoreToHoist->isSimple())
1874 Value *StorePtr = StoreToHoist->getPointerOperand();
1876 // Look for a store to the same pointer in BrBB.
1877 unsigned MaxNumInstToLookAt = 9;
1878 for (Instruction &CurI : reverse(*BrBB)) {
1879 if (!MaxNumInstToLookAt)
1882 if (isa<DbgInfoIntrinsic>(CurI))
1884 --MaxNumInstToLookAt;
1886 // Could be calling an instruction that affects memory like free().
1887 if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1890 if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1891 // Found the previous store make sure it stores to the same location.
1892 if (SI->getPointerOperand() == StorePtr)
1893 // Found the previous store, return its value operand.
1894 return SI->getValueOperand();
1895 return nullptr; // Unknown store.
1902 /// \brief Speculate a conditional basic block flattening the CFG.
1904 /// Note that this is a very risky transform currently. Speculating
1905 /// instructions like this is most often not desirable. Instead, there is an MI
1906 /// pass which can do it with full awareness of the resource constraints.
1907 /// However, some cases are "obvious" and we should do directly. An example of
1908 /// this is speculating a single, reasonably cheap instruction.
1910 /// There is only one distinct advantage to flattening the CFG at the IR level:
1911 /// it makes very common but simplistic optimizations such as are common in
1912 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1913 /// modeling their effects with easier to reason about SSA value graphs.
1916 /// An illustration of this transform is turning this IR:
1919 /// %cmp = icmp ult %x, %y
1920 /// br i1 %cmp, label %EndBB, label %ThenBB
1922 /// %sub = sub %x, %y
1925 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1932 /// %cmp = icmp ult %x, %y
1933 /// %sub = sub %x, %y
1934 /// %cond = select i1 %cmp, 0, %sub
1938 /// \returns true if the conditional block is removed.
1939 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1940 const TargetTransformInfo &TTI) {
1941 // Be conservative for now. FP select instruction can often be expensive.
1942 Value *BrCond = BI->getCondition();
1943 if (isa<FCmpInst>(BrCond))
1946 BasicBlock *BB = BI->getParent();
1947 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1949 // If ThenBB is actually on the false edge of the conditional branch, remember
1950 // to swap the select operands later.
1951 bool Invert = false;
1952 if (ThenBB != BI->getSuccessor(0)) {
1953 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1956 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1958 // Keep a count of how many times instructions are used within CondBB when
1959 // they are candidates for sinking into CondBB. Specifically:
1960 // - They are defined in BB, and
1961 // - They have no side effects, and
1962 // - All of their uses are in CondBB.
1963 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1965 unsigned SpeculationCost = 0;
1966 Value *SpeculatedStoreValue = nullptr;
1967 StoreInst *SpeculatedStore = nullptr;
1968 for (BasicBlock::iterator BBI = ThenBB->begin(),
1969 BBE = std::prev(ThenBB->end());
1970 BBI != BBE; ++BBI) {
1971 Instruction *I = &*BBI;
1973 if (isa<DbgInfoIntrinsic>(I))
1976 // Only speculatively execute a single instruction (not counting the
1977 // terminator) for now.
1979 if (SpeculationCost > 1)
1982 // Don't hoist the instruction if it's unsafe or expensive.
1983 if (!isSafeToSpeculativelyExecute(I) &&
1984 !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1985 I, BB, ThenBB, EndBB))))
1987 if (!SpeculatedStoreValue &&
1988 ComputeSpeculationCost(I, TTI) >
1989 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
1992 // Store the store speculation candidate.
1993 if (SpeculatedStoreValue)
1994 SpeculatedStore = cast<StoreInst>(I);
1996 // Do not hoist the instruction if any of its operands are defined but not
1997 // used in BB. The transformation will prevent the operand from
1998 // being sunk into the use block.
1999 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
2000 Instruction *OpI = dyn_cast<Instruction>(*i);
2001 if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
2002 continue; // Not a candidate for sinking.
2004 ++SinkCandidateUseCounts[OpI];
2008 // Consider any sink candidates which are only used in CondBB as costs for
2009 // speculation. Note, while we iterate over a DenseMap here, we are summing
2010 // and so iteration order isn't significant.
2011 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator
2012 I = SinkCandidateUseCounts.begin(),
2013 E = SinkCandidateUseCounts.end();
2015 if (I->first->getNumUses() == I->second) {
2017 if (SpeculationCost > 1)
2021 // Check that the PHI nodes can be converted to selects.
2022 bool HaveRewritablePHIs = false;
2023 for (BasicBlock::iterator I = EndBB->begin();
2024 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2025 Value *OrigV = PN->getIncomingValueForBlock(BB);
2026 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
2028 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2029 // Skip PHIs which are trivial.
2033 // Don't convert to selects if we could remove undefined behavior instead.
2034 if (passingValueIsAlwaysUndefined(OrigV, PN) ||
2035 passingValueIsAlwaysUndefined(ThenV, PN))
2038 HaveRewritablePHIs = true;
2039 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2040 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2041 if (!OrigCE && !ThenCE)
2042 continue; // Known safe and cheap.
2044 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2045 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2047 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2048 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2050 2 * PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2051 if (OrigCost + ThenCost > MaxCost)
2054 // Account for the cost of an unfolded ConstantExpr which could end up
2055 // getting expanded into Instructions.
2056 // FIXME: This doesn't account for how many operations are combined in the
2057 // constant expression.
2059 if (SpeculationCost > 1)
2063 // If there are no PHIs to process, bail early. This helps ensure idempotence
2065 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2068 // If we get here, we can hoist the instruction and if-convert.
2069 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2071 // Insert a select of the value of the speculated store.
2072 if (SpeculatedStoreValue) {
2073 IRBuilder<NoFolder> Builder(BI);
2074 Value *TrueV = SpeculatedStore->getValueOperand();
2075 Value *FalseV = SpeculatedStoreValue;
2077 std::swap(TrueV, FalseV);
2078 Value *S = Builder.CreateSelect(
2079 BrCond, TrueV, FalseV, TrueV->getName() + "." + FalseV->getName(), BI);
2080 SpeculatedStore->setOperand(0, S);
2083 // Metadata can be dependent on the condition we are hoisting above.
2084 // Conservatively strip all metadata on the instruction.
2085 for (auto &I : *ThenBB)
2086 I.dropUnknownNonDebugMetadata();
2088 // Hoist the instructions.
2089 BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2090 ThenBB->begin(), std::prev(ThenBB->end()));
2092 // Insert selects and rewrite the PHI operands.
2093 IRBuilder<NoFolder> Builder(BI);
2094 for (BasicBlock::iterator I = EndBB->begin();
2095 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2096 unsigned OrigI = PN->getBasicBlockIndex(BB);
2097 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
2098 Value *OrigV = PN->getIncomingValue(OrigI);
2099 Value *ThenV = PN->getIncomingValue(ThenI);
2101 // Skip PHIs which are trivial.
2105 // Create a select whose true value is the speculatively executed value and
2106 // false value is the preexisting value. Swap them if the branch
2107 // destinations were inverted.
2108 Value *TrueV = ThenV, *FalseV = OrigV;
2110 std::swap(TrueV, FalseV);
2111 Value *V = Builder.CreateSelect(
2112 BrCond, TrueV, FalseV, TrueV->getName() + "." + FalseV->getName(), BI);
2113 PN->setIncomingValue(OrigI, V);
2114 PN->setIncomingValue(ThenI, V);
2121 /// Return true if we can thread a branch across this block.
2122 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
2123 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
2126 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2127 if (isa<DbgInfoIntrinsic>(BBI))
2130 return false; // Don't clone large BB's.
2133 // We can only support instructions that do not define values that are
2134 // live outside of the current basic block.
2135 for (User *U : BBI->users()) {
2136 Instruction *UI = cast<Instruction>(U);
2137 if (UI->getParent() != BB || isa<PHINode>(UI))
2141 // Looks ok, continue checking.
2147 /// If we have a conditional branch on a PHI node value that is defined in the
2148 /// same block as the branch and if any PHI entries are constants, thread edges
2149 /// corresponding to that entry to be branches to their ultimate destination.
2150 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL) {
2151 BasicBlock *BB = BI->getParent();
2152 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2153 // NOTE: we currently cannot transform this case if the PHI node is used
2154 // outside of the block.
2155 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2158 // Degenerate case of a single entry PHI.
2159 if (PN->getNumIncomingValues() == 1) {
2160 FoldSingleEntryPHINodes(PN->getParent());
2164 // Now we know that this block has multiple preds and two succs.
2165 if (!BlockIsSimpleEnoughToThreadThrough(BB))
2168 // Can't fold blocks that contain noduplicate or convergent calls.
2169 if (any_of(*BB, [](const Instruction &I) {
2170 const CallInst *CI = dyn_cast<CallInst>(&I);
2171 return CI && (CI->cannotDuplicate() || CI->isConvergent());
2175 // Okay, this is a simple enough basic block. See if any phi values are
2177 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2178 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
2179 if (!CB || !CB->getType()->isIntegerTy(1))
2182 // Okay, we now know that all edges from PredBB should be revectored to
2183 // branch to RealDest.
2184 BasicBlock *PredBB = PN->getIncomingBlock(i);
2185 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2188 continue; // Skip self loops.
2189 // Skip if the predecessor's terminator is an indirect branch.
2190 if (isa<IndirectBrInst>(PredBB->getTerminator()))
2193 // The dest block might have PHI nodes, other predecessors and other
2194 // difficult cases. Instead of being smart about this, just insert a new
2195 // block that jumps to the destination block, effectively splitting
2196 // the edge we are about to create.
2197 BasicBlock *EdgeBB =
2198 BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2199 RealDest->getParent(), RealDest);
2200 BranchInst::Create(RealDest, EdgeBB);
2202 // Update PHI nodes.
2203 AddPredecessorToBlock(RealDest, EdgeBB, BB);
2205 // BB may have instructions that are being threaded over. Clone these
2206 // instructions into EdgeBB. We know that there will be no uses of the
2207 // cloned instructions outside of EdgeBB.
2208 BasicBlock::iterator InsertPt = EdgeBB->begin();
2209 DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2210 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2211 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2212 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2215 // Clone the instruction.
2216 Instruction *N = BBI->clone();
2218 N->setName(BBI->getName() + ".c");
2220 // Update operands due to translation.
2221 for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2222 DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2223 if (PI != TranslateMap.end())
2227 // Check for trivial simplification.
2228 if (Value *V = SimplifyInstruction(N, DL)) {
2229 if (!BBI->use_empty())
2230 TranslateMap[&*BBI] = V;
2231 if (!N->mayHaveSideEffects()) {
2232 delete N; // Instruction folded away, don't need actual inst
2236 if (!BBI->use_empty())
2237 TranslateMap[&*BBI] = N;
2239 // Insert the new instruction into its new home.
2241 EdgeBB->getInstList().insert(InsertPt, N);
2244 // Loop over all of the edges from PredBB to BB, changing them to branch
2245 // to EdgeBB instead.
2246 TerminatorInst *PredBBTI = PredBB->getTerminator();
2247 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2248 if (PredBBTI->getSuccessor(i) == BB) {
2249 BB->removePredecessor(PredBB);
2250 PredBBTI->setSuccessor(i, EdgeBB);
2253 // Recurse, simplifying any other constants.
2254 return FoldCondBranchOnPHI(BI, DL) | true;
2260 /// Given a BB that starts with the specified two-entry PHI node,
2261 /// see if we can eliminate it.
2262 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI,
2263 const DataLayout &DL) {
2264 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2265 // statement", which has a very simple dominance structure. Basically, we
2266 // are trying to find the condition that is being branched on, which
2267 // subsequently causes this merge to happen. We really want control
2268 // dependence information for this check, but simplifycfg can't keep it up
2269 // to date, and this catches most of the cases we care about anyway.
2270 BasicBlock *BB = PN->getParent();
2271 BasicBlock *IfTrue, *IfFalse;
2272 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2274 // Don't bother if the branch will be constant folded trivially.
2275 isa<ConstantInt>(IfCond))
2278 // Okay, we found that we can merge this two-entry phi node into a select.
2279 // Doing so would require us to fold *all* two entry phi nodes in this block.
2280 // At some point this becomes non-profitable (particularly if the target
2281 // doesn't support cmov's). Only do this transformation if there are two or
2282 // fewer PHI nodes in this block.
2283 unsigned NumPhis = 0;
2284 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2288 // Loop over the PHI's seeing if we can promote them all to select
2289 // instructions. While we are at it, keep track of the instructions
2290 // that need to be moved to the dominating block.
2291 SmallPtrSet<Instruction *, 4> AggressiveInsts;
2292 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2293 MaxCostVal1 = PHINodeFoldingThreshold;
2294 MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2295 MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2297 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2298 PHINode *PN = cast<PHINode>(II++);
2299 if (Value *V = SimplifyInstruction(PN, DL)) {
2300 PN->replaceAllUsesWith(V);
2301 PN->eraseFromParent();
2305 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
2306 MaxCostVal0, TTI) ||
2307 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
2312 // If we folded the first phi, PN dangles at this point. Refresh it. If
2313 // we ran out of PHIs then we simplified them all.
2314 PN = dyn_cast<PHINode>(BB->begin());
2318 // Don't fold i1 branches on PHIs which contain binary operators. These can
2319 // often be turned into switches and other things.
2320 if (PN->getType()->isIntegerTy(1) &&
2321 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2322 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2323 isa<BinaryOperator>(IfCond)))
2326 // If all PHI nodes are promotable, check to make sure that all instructions
2327 // in the predecessor blocks can be promoted as well. If not, we won't be able
2328 // to get rid of the control flow, so it's not worth promoting to select
2330 BasicBlock *DomBlock = nullptr;
2331 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2332 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2333 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2336 DomBlock = *pred_begin(IfBlock1);
2337 for (BasicBlock::iterator I = IfBlock1->begin(); !isa<TerminatorInst>(I);
2339 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2340 // This is not an aggressive instruction that we can promote.
2341 // Because of this, we won't be able to get rid of the control flow, so
2342 // the xform is not worth it.
2347 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2350 DomBlock = *pred_begin(IfBlock2);
2351 for (BasicBlock::iterator I = IfBlock2->begin(); !isa<TerminatorInst>(I);
2353 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2354 // This is not an aggressive instruction that we can promote.
2355 // Because of this, we won't be able to get rid of the control flow, so
2356 // the xform is not worth it.
2361 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
2362 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
2364 // If we can still promote the PHI nodes after this gauntlet of tests,
2365 // do all of the PHI's now.
2366 Instruction *InsertPt = DomBlock->getTerminator();
2367 IRBuilder<NoFolder> Builder(InsertPt);
2369 // Move all 'aggressive' instructions, which are defined in the
2370 // conditional parts of the if's up to the dominating block.
2372 for (auto &I : *IfBlock1)
2373 I.dropUnknownNonDebugMetadata();
2374 DomBlock->getInstList().splice(InsertPt->getIterator(),
2375 IfBlock1->getInstList(), IfBlock1->begin(),
2376 IfBlock1->getTerminator()->getIterator());
2379 for (auto &I : *IfBlock2)
2380 I.dropUnknownNonDebugMetadata();
2381 DomBlock->getInstList().splice(InsertPt->getIterator(),
2382 IfBlock2->getInstList(), IfBlock2->begin(),
2383 IfBlock2->getTerminator()->getIterator());
2386 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2387 // Change the PHI node into a select instruction.
2388 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2389 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2391 Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2392 PN->replaceAllUsesWith(Sel);
2394 PN->eraseFromParent();
2397 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2398 // has been flattened. Change DomBlock to jump directly to our new block to
2399 // avoid other simplifycfg's kicking in on the diamond.
2400 TerminatorInst *OldTI = DomBlock->getTerminator();
2401 Builder.SetInsertPoint(OldTI);
2402 Builder.CreateBr(BB);
2403 OldTI->eraseFromParent();
2407 /// If we found a conditional branch that goes to two returning blocks,
2408 /// try to merge them together into one return,
2409 /// introducing a select if the return values disagree.
2410 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
2411 IRBuilder<> &Builder) {
2412 assert(BI->isConditional() && "Must be a conditional branch");
2413 BasicBlock *TrueSucc = BI->getSuccessor(0);
2414 BasicBlock *FalseSucc = BI->getSuccessor(1);
2415 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2416 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2418 // Check to ensure both blocks are empty (just a return) or optionally empty
2419 // with PHI nodes. If there are other instructions, merging would cause extra
2420 // computation on one path or the other.
2421 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2423 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2426 Builder.SetInsertPoint(BI);
2427 // Okay, we found a branch that is going to two return nodes. If
2428 // there is no return value for this function, just change the
2429 // branch into a return.
2430 if (FalseRet->getNumOperands() == 0) {
2431 TrueSucc->removePredecessor(BI->getParent());
2432 FalseSucc->removePredecessor(BI->getParent());
2433 Builder.CreateRetVoid();
2434 EraseTerminatorInstAndDCECond(BI);
2438 // Otherwise, figure out what the true and false return values are
2439 // so we can insert a new select instruction.
2440 Value *TrueValue = TrueRet->getReturnValue();
2441 Value *FalseValue = FalseRet->getReturnValue();
2443 // Unwrap any PHI nodes in the return blocks.
2444 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2445 if (TVPN->getParent() == TrueSucc)
2446 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2447 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2448 if (FVPN->getParent() == FalseSucc)
2449 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2451 // In order for this transformation to be safe, we must be able to
2452 // unconditionally execute both operands to the return. This is
2453 // normally the case, but we could have a potentially-trapping
2454 // constant expression that prevents this transformation from being
2456 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2459 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2463 // Okay, we collected all the mapped values and checked them for sanity, and
2464 // defined to really do this transformation. First, update the CFG.
2465 TrueSucc->removePredecessor(BI->getParent());
2466 FalseSucc->removePredecessor(BI->getParent());
2468 // Insert select instructions where needed.
2469 Value *BrCond = BI->getCondition();
2471 // Insert a select if the results differ.
2472 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2473 } else if (isa<UndefValue>(TrueValue)) {
2474 TrueValue = FalseValue;
2477 Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2482 !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2486 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2487 << "\n " << *BI << "NewRet = " << *RI
2488 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: " << *FalseSucc);
2490 EraseTerminatorInstAndDCECond(BI);
2495 /// Return true if the given instruction is available
2496 /// in its predecessor block. If yes, the instruction will be removed.
2497 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
2498 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2500 for (Instruction &I : *PB) {
2501 Instruction *PBI = &I;
2502 // Check whether Inst and PBI generate the same value.
2503 if (Inst->isIdenticalTo(PBI)) {
2504 Inst->replaceAllUsesWith(PBI);
2505 Inst->eraseFromParent();
2512 /// Return true if either PBI or BI has branch weight available, and store
2513 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2514 /// not have branch weight, use 1:1 as its weight.
2515 static bool extractPredSuccWeights(BranchInst *PBI, BranchInst *BI,
2516 uint64_t &PredTrueWeight,
2517 uint64_t &PredFalseWeight,
2518 uint64_t &SuccTrueWeight,
2519 uint64_t &SuccFalseWeight) {
2520 bool PredHasWeights =
2521 PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2522 bool SuccHasWeights =
2523 BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2524 if (PredHasWeights || SuccHasWeights) {
2525 if (!PredHasWeights)
2526 PredTrueWeight = PredFalseWeight = 1;
2527 if (!SuccHasWeights)
2528 SuccTrueWeight = SuccFalseWeight = 1;
2535 /// If this basic block is simple enough, and if a predecessor branches to us
2536 /// and one of our successors, fold the block into the predecessor and use
2537 /// logical operations to pick the right destination.
2538 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2539 BasicBlock *BB = BI->getParent();
2541 Instruction *Cond = nullptr;
2542 if (BI->isConditional())
2543 Cond = dyn_cast<Instruction>(BI->getCondition());
2545 // For unconditional branch, check for a simple CFG pattern, where
2546 // BB has a single predecessor and BB's successor is also its predecessor's
2547 // successor. If such pattern exisits, check for CSE between BB and its
2549 if (BasicBlock *PB = BB->getSinglePredecessor())
2550 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2551 if (PBI->isConditional() &&
2552 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2553 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2554 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2555 Instruction *Curr = &*I++;
2556 if (isa<CmpInst>(Curr)) {
2560 // Quit if we can't remove this instruction.
2561 if (!checkCSEInPredecessor(Curr, PB))
2570 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2571 Cond->getParent() != BB || !Cond->hasOneUse())
2574 // Make sure the instruction after the condition is the cond branch.
2575 BasicBlock::iterator CondIt = ++Cond->getIterator();
2577 // Ignore dbg intrinsics.
2578 while (isa<DbgInfoIntrinsic>(CondIt))
2584 // Only allow this transformation if computing the condition doesn't involve
2585 // too many instructions and these involved instructions can be executed
2586 // unconditionally. We denote all involved instructions except the condition
2587 // as "bonus instructions", and only allow this transformation when the
2588 // number of the bonus instructions does not exceed a certain threshold.
2589 unsigned NumBonusInsts = 0;
2590 for (auto I = BB->begin(); Cond != &*I; ++I) {
2591 // Ignore dbg intrinsics.
2592 if (isa<DbgInfoIntrinsic>(I))
2594 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2596 // I has only one use and can be executed unconditionally.
2597 Instruction *User = dyn_cast<Instruction>(I->user_back());
2598 if (User == nullptr || User->getParent() != BB)
2600 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2601 // to use any other instruction, User must be an instruction between next(I)
2604 // Early exits once we reach the limit.
2605 if (NumBonusInsts > BonusInstThreshold)
2609 // Cond is known to be a compare or binary operator. Check to make sure that
2610 // neither operand is a potentially-trapping constant expression.
2611 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2614 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2618 // Finally, don't infinitely unroll conditional loops.
2619 BasicBlock *TrueDest = BI->getSuccessor(0);
2620 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2621 if (TrueDest == BB || FalseDest == BB)
2624 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2625 BasicBlock *PredBlock = *PI;
2626 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2628 // Check that we have two conditional branches. If there is a PHI node in
2629 // the common successor, verify that the same value flows in from both
2631 SmallVector<PHINode *, 4> PHIs;
2632 if (!PBI || PBI->isUnconditional() ||
2633 (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2634 (!BI->isConditional() &&
2635 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2638 // Determine if the two branches share a common destination.
2639 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2640 bool InvertPredCond = false;
2642 if (BI->isConditional()) {
2643 if (PBI->getSuccessor(0) == TrueDest) {
2644 Opc = Instruction::Or;
2645 } else if (PBI->getSuccessor(1) == FalseDest) {
2646 Opc = Instruction::And;
2647 } else if (PBI->getSuccessor(0) == FalseDest) {
2648 Opc = Instruction::And;
2649 InvertPredCond = true;
2650 } else if (PBI->getSuccessor(1) == TrueDest) {
2651 Opc = Instruction::Or;
2652 InvertPredCond = true;
2657 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2661 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2662 IRBuilder<> Builder(PBI);
2664 // If we need to invert the condition in the pred block to match, do so now.
2665 if (InvertPredCond) {
2666 Value *NewCond = PBI->getCondition();
2668 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2669 CmpInst *CI = cast<CmpInst>(NewCond);
2670 CI->setPredicate(CI->getInversePredicate());
2673 Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2676 PBI->setCondition(NewCond);
2677 PBI->swapSuccessors();
2680 // If we have bonus instructions, clone them into the predecessor block.
2681 // Note that there may be multiple predecessor blocks, so we cannot move
2682 // bonus instructions to a predecessor block.
2683 ValueToValueMapTy VMap; // maps original values to cloned values
2684 // We already make sure Cond is the last instruction before BI. Therefore,
2685 // all instructions before Cond other than DbgInfoIntrinsic are bonus
2687 for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2688 if (isa<DbgInfoIntrinsic>(BonusInst))
2690 Instruction *NewBonusInst = BonusInst->clone();
2691 RemapInstruction(NewBonusInst, VMap,
2692 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2693 VMap[&*BonusInst] = NewBonusInst;
2695 // If we moved a load, we cannot any longer claim any knowledge about
2696 // its potential value. The previous information might have been valid
2697 // only given the branch precondition.
2698 // For an analogous reason, we must also drop all the metadata whose
2699 // semantics we don't understand.
2700 NewBonusInst->dropUnknownNonDebugMetadata();
2702 PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2703 NewBonusInst->takeName(&*BonusInst);
2704 BonusInst->setName(BonusInst->getName() + ".old");
2707 // Clone Cond into the predecessor basic block, and or/and the
2708 // two conditions together.
2709 Instruction *New = Cond->clone();
2710 RemapInstruction(New, VMap,
2711 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2712 PredBlock->getInstList().insert(PBI->getIterator(), New);
2713 New->takeName(Cond);
2714 Cond->setName(New->getName() + ".old");
2716 if (BI->isConditional()) {
2717 Instruction *NewCond = cast<Instruction>(
2718 Builder.CreateBinOp(Opc, PBI->getCondition(), New, "or.cond"));
2719 PBI->setCondition(NewCond);
2721 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2723 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2724 SuccTrueWeight, SuccFalseWeight);
2725 SmallVector<uint64_t, 8> NewWeights;
2727 if (PBI->getSuccessor(0) == BB) {
2729 // PBI: br i1 %x, BB, FalseDest
2730 // BI: br i1 %y, TrueDest, FalseDest
2731 // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2732 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2733 // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2734 // TrueWeight for PBI * FalseWeight for BI.
2735 // We assume that total weights of a BranchInst can fit into 32 bits.
2736 // Therefore, we will not have overflow using 64-bit arithmetic.
2737 NewWeights.push_back(PredFalseWeight *
2738 (SuccFalseWeight + SuccTrueWeight) +
2739 PredTrueWeight * SuccFalseWeight);
2741 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2742 PBI->setSuccessor(0, TrueDest);
2744 if (PBI->getSuccessor(1) == BB) {
2746 // PBI: br i1 %x, TrueDest, BB
2747 // BI: br i1 %y, TrueDest, FalseDest
2748 // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2749 // FalseWeight for PBI * TrueWeight for BI.
2750 NewWeights.push_back(PredTrueWeight *
2751 (SuccFalseWeight + SuccTrueWeight) +
2752 PredFalseWeight * SuccTrueWeight);
2753 // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2754 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2756 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2757 PBI->setSuccessor(1, FalseDest);
2759 if (NewWeights.size() == 2) {
2760 // Halve the weights if any of them cannot fit in an uint32_t
2761 FitWeights(NewWeights);
2763 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2766 LLVMContext::MD_prof,
2767 MDBuilder(BI->getContext()).createBranchWeights(MDWeights));
2769 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2771 // Update PHI nodes in the common successors.
2772 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2773 ConstantInt *PBI_C = cast<ConstantInt>(
2774 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2775 assert(PBI_C->getType()->isIntegerTy(1));
2776 Instruction *MergedCond = nullptr;
2777 if (PBI->getSuccessor(0) == TrueDest) {
2778 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2779 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2780 // is false: !PBI_Cond and BI_Value
2781 Instruction *NotCond = cast<Instruction>(
2782 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2783 MergedCond = cast<Instruction>(
2784 Builder.CreateBinOp(Instruction::And, NotCond, New, "and.cond"));
2786 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2787 Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2789 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2790 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2791 // is false: PBI_Cond and BI_Value
2792 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2793 Instruction::And, PBI->getCondition(), New, "and.cond"));
2794 if (PBI_C->isOne()) {
2795 Instruction *NotCond = cast<Instruction>(
2796 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2797 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2798 Instruction::Or, NotCond, MergedCond, "or.cond"));
2802 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2805 // Change PBI from Conditional to Unconditional.
2806 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2807 EraseTerminatorInstAndDCECond(PBI);
2811 // If BI was a loop latch, it may have had associated loop metadata.
2812 // We need to copy it to the new latch, that is, PBI.
2813 if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2814 PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2816 // TODO: If BB is reachable from all paths through PredBlock, then we
2817 // could replace PBI's branch probabilities with BI's.
2819 // Copy any debug value intrinsics into the end of PredBlock.
2820 for (Instruction &I : *BB)
2821 if (isa<DbgInfoIntrinsic>(I))
2822 I.clone()->insertBefore(PBI);
2829 // If there is only one store in BB1 and BB2, return it, otherwise return
2831 static StoreInst *findUniqueStoreInBlocks(BasicBlock *BB1, BasicBlock *BB2) {
2832 StoreInst *S = nullptr;
2833 for (auto *BB : {BB1, BB2}) {
2837 if (auto *SI = dyn_cast<StoreInst>(&I)) {
2839 // Multiple stores seen.
2848 static Value *ensureValueAvailableInSuccessor(Value *V, BasicBlock *BB,
2849 Value *AlternativeV = nullptr) {
2850 // PHI is going to be a PHI node that allows the value V that is defined in
2851 // BB to be referenced in BB's only successor.
2853 // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2854 // doesn't matter to us what the other operand is (it'll never get used). We
2855 // could just create a new PHI with an undef incoming value, but that could
2856 // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2857 // other PHI. So here we directly look for some PHI in BB's successor with V
2858 // as an incoming operand. If we find one, we use it, else we create a new
2861 // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2862 // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2863 // where OtherBB is the single other predecessor of BB's only successor.
2864 PHINode *PHI = nullptr;
2865 BasicBlock *Succ = BB->getSingleSuccessor();
2867 for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2868 if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2869 PHI = cast<PHINode>(I);
2873 assert(std::distance(pred_begin(Succ), pred_end(Succ)) == 2);
2874 auto PredI = pred_begin(Succ);
2875 BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2876 if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2883 // If V is not an instruction defined in BB, just return it.
2884 if (!AlternativeV &&
2885 (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2888 PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2889 PHI->addIncoming(V, BB);
2890 for (BasicBlock *PredBB : predecessors(Succ))
2893 AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2897 static bool mergeConditionalStoreToAddress(BasicBlock *PTB, BasicBlock *PFB,
2898 BasicBlock *QTB, BasicBlock *QFB,
2899 BasicBlock *PostBB, Value *Address,
2900 bool InvertPCond, bool InvertQCond) {
2901 auto IsaBitcastOfPointerType = [](const Instruction &I) {
2902 return Operator::getOpcode(&I) == Instruction::BitCast &&
2903 I.getType()->isPointerTy();
2906 // If we're not in aggressive mode, we only optimize if we have some
2907 // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2908 auto IsWorthwhile = [&](BasicBlock *BB) {
2911 // Heuristic: if the block can be if-converted/phi-folded and the
2912 // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2913 // thread this store.
2915 for (auto &I : *BB) {
2916 // Cheap instructions viable for folding.
2917 if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2920 // Free instructions.
2921 else if (isa<TerminatorInst>(I) || isa<DbgInfoIntrinsic>(I) ||
2922 IsaBitcastOfPointerType(I))
2927 return N <= PHINodeFoldingThreshold;
2930 if (!MergeCondStoresAggressively &&
2931 (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2932 !IsWorthwhile(QFB)))
2935 // For every pointer, there must be exactly two stores, one coming from
2936 // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2937 // store (to any address) in PTB,PFB or QTB,QFB.
2938 // FIXME: We could relax this restriction with a bit more work and performance
2940 StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2941 StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2942 if (!PStore || !QStore)
2945 // Now check the stores are compatible.
2946 if (!QStore->isUnordered() || !PStore->isUnordered())
2949 // Check that sinking the store won't cause program behavior changes. Sinking
2950 // the store out of the Q blocks won't change any behavior as we're sinking
2951 // from a block to its unconditional successor. But we're moving a store from
2952 // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2953 // So we need to check that there are no aliasing loads or stores in
2954 // QBI, QTB and QFB. We also need to check there are no conflicting memory
2955 // operations between PStore and the end of its parent block.
2957 // The ideal way to do this is to query AliasAnalysis, but we don't
2958 // preserve AA currently so that is dangerous. Be super safe and just
2959 // check there are no other memory operations at all.
2960 for (auto &I : *QFB->getSinglePredecessor())
2961 if (I.mayReadOrWriteMemory())
2963 for (auto &I : *QFB)
2964 if (&I != QStore && I.mayReadOrWriteMemory())
2967 for (auto &I : *QTB)
2968 if (&I != QStore && I.mayReadOrWriteMemory())
2970 for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2972 if (&*I != PStore && I->mayReadOrWriteMemory())
2975 // OK, we're going to sink the stores to PostBB. The store has to be
2976 // conditional though, so first create the predicate.
2977 Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
2979 Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
2982 Value *PPHI = ensureValueAvailableInSuccessor(PStore->getValueOperand(),
2983 PStore->getParent());
2984 Value *QPHI = ensureValueAvailableInSuccessor(QStore->getValueOperand(),
2985 QStore->getParent(), PPHI);
2987 IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
2989 Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
2990 Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
2993 PPred = QB.CreateNot(PPred);
2995 QPred = QB.CreateNot(QPred);
2996 Value *CombinedPred = QB.CreateOr(PPred, QPred);
2999 SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
3000 QB.SetInsertPoint(T);
3001 StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
3003 PStore->getAAMetadata(AAMD, /*Merge=*/false);
3004 PStore->getAAMetadata(AAMD, /*Merge=*/true);
3005 SI->setAAMetadata(AAMD);
3007 QStore->eraseFromParent();
3008 PStore->eraseFromParent();
3013 static bool mergeConditionalStores(BranchInst *PBI, BranchInst *QBI) {
3014 // The intention here is to find diamonds or triangles (see below) where each
3015 // conditional block contains a store to the same address. Both of these
3016 // stores are conditional, so they can't be unconditionally sunk. But it may
3017 // be profitable to speculatively sink the stores into one merged store at the
3018 // end, and predicate the merged store on the union of the two conditions of
3021 // This can reduce the number of stores executed if both of the conditions are
3022 // true, and can allow the blocks to become small enough to be if-converted.
3023 // This optimization will also chain, so that ladders of test-and-set
3024 // sequences can be if-converted away.
3026 // We only deal with simple diamonds or triangles:
3028 // PBI or PBI or a combination of the two
3038 // We model triangles as a type of diamond with a nullptr "true" block.
3039 // Triangles are canonicalized so that the fallthrough edge is represented by
3040 // a true condition, as in the diagram above.
3042 BasicBlock *PTB = PBI->getSuccessor(0);
3043 BasicBlock *PFB = PBI->getSuccessor(1);
3044 BasicBlock *QTB = QBI->getSuccessor(0);
3045 BasicBlock *QFB = QBI->getSuccessor(1);
3046 BasicBlock *PostBB = QFB->getSingleSuccessor();
3048 bool InvertPCond = false, InvertQCond = false;
3049 // Canonicalize fallthroughs to the true branches.
3050 if (PFB == QBI->getParent()) {
3051 std::swap(PFB, PTB);
3054 if (QFB == PostBB) {
3055 std::swap(QFB, QTB);
3059 // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3060 // and QFB may not. Model fallthroughs as a nullptr block.
3061 if (PTB == QBI->getParent())
3066 // Legality bailouts. We must have at least the non-fallthrough blocks and
3067 // the post-dominating block, and the non-fallthroughs must only have one
3069 auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3070 return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3073 !HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3074 !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3076 if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3077 (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3079 if (PostBB->getNumUses() != 2 || QBI->getParent()->getNumUses() != 2)
3082 // OK, this is a sequence of two diamonds or triangles.
3083 // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3084 SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3085 for (auto *BB : {PTB, PFB}) {
3089 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3090 PStoreAddresses.insert(SI->getPointerOperand());
3092 for (auto *BB : {QTB, QFB}) {
3096 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3097 QStoreAddresses.insert(SI->getPointerOperand());
3100 set_intersect(PStoreAddresses, QStoreAddresses);
3101 // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3102 // clear what it contains.
3103 auto &CommonAddresses = PStoreAddresses;
3105 bool Changed = false;
3106 for (auto *Address : CommonAddresses)
3107 Changed |= mergeConditionalStoreToAddress(
3108 PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond);
3112 /// If we have a conditional branch as a predecessor of another block,
3113 /// this function tries to simplify it. We know
3114 /// that PBI and BI are both conditional branches, and BI is in one of the
3115 /// successor blocks of PBI - PBI branches to BI.
3116 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
3117 const DataLayout &DL) {
3118 assert(PBI->isConditional() && BI->isConditional());
3119 BasicBlock *BB = BI->getParent();
3121 // If this block ends with a branch instruction, and if there is a
3122 // predecessor that ends on a branch of the same condition, make
3123 // this conditional branch redundant.
3124 if (PBI->getCondition() == BI->getCondition() &&
3125 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3126 // Okay, the outcome of this conditional branch is statically
3127 // knowable. If this block had a single pred, handle specially.
3128 if (BB->getSinglePredecessor()) {
3129 // Turn this into a branch on constant.
3130 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3132 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3133 return true; // Nuke the branch on constant.
3136 // Otherwise, if there are multiple predecessors, insert a PHI that merges
3137 // in the constant and simplify the block result. Subsequent passes of
3138 // simplifycfg will thread the block.
3139 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
3140 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3141 PHINode *NewPN = PHINode::Create(
3142 Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3143 BI->getCondition()->getName() + ".pr", &BB->front());
3144 // Okay, we're going to insert the PHI node. Since PBI is not the only
3145 // predecessor, compute the PHI'd conditional value for all of the preds.
3146 // Any predecessor where the condition is not computable we keep symbolic.
3147 for (pred_iterator PI = PB; PI != PE; ++PI) {
3148 BasicBlock *P = *PI;
3149 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3150 PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3151 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3152 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3154 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3157 NewPN->addIncoming(BI->getCondition(), P);
3161 BI->setCondition(NewPN);
3166 if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3170 // If both branches are conditional and both contain stores to the same
3171 // address, remove the stores from the conditionals and create a conditional
3172 // merged store at the end.
3173 if (MergeCondStores && mergeConditionalStores(PBI, BI))
3176 // If this is a conditional branch in an empty block, and if any
3177 // predecessors are a conditional branch to one of our destinations,
3178 // fold the conditions into logical ops and one cond br.
3179 BasicBlock::iterator BBI = BB->begin();
3180 // Ignore dbg intrinsics.
3181 while (isa<DbgInfoIntrinsic>(BBI))
3187 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3190 } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3193 } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3196 } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3203 // Check to make sure that the other destination of this branch
3204 // isn't BB itself. If so, this is an infinite loop that will
3205 // keep getting unwound.
3206 if (PBI->getSuccessor(PBIOp) == BB)
3209 // Do not perform this transformation if it would require
3210 // insertion of a large number of select instructions. For targets
3211 // without predication/cmovs, this is a big pessimization.
3213 // Also do not perform this transformation if any phi node in the common
3214 // destination block can trap when reached by BB or PBB (PR17073). In that
3215 // case, it would be unsafe to hoist the operation into a select instruction.
3217 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3218 unsigned NumPhis = 0;
3219 for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3221 if (NumPhis > 2) // Disable this xform.
3224 PHINode *PN = cast<PHINode>(II);
3225 Value *BIV = PN->getIncomingValueForBlock(BB);
3226 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3230 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3231 Value *PBIV = PN->getIncomingValue(PBBIdx);
3232 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3237 // Finally, if everything is ok, fold the branches to logical ops.
3238 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3240 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3241 << "AND: " << *BI->getParent());
3243 // If OtherDest *is* BB, then BB is a basic block with a single conditional
3244 // branch in it, where one edge (OtherDest) goes back to itself but the other
3245 // exits. We don't *know* that the program avoids the infinite loop
3246 // (even though that seems likely). If we do this xform naively, we'll end up
3247 // recursively unpeeling the loop. Since we know that (after the xform is
3248 // done) that the block *is* infinite if reached, we just make it an obviously
3249 // infinite loop with no cond branch.
3250 if (OtherDest == BB) {
3251 // Insert it at the end of the function, because it's either code,
3252 // or it won't matter if it's hot. :)
3253 BasicBlock *InfLoopBlock =
3254 BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3255 BranchInst::Create(InfLoopBlock, InfLoopBlock);
3256 OtherDest = InfLoopBlock;
3259 DEBUG(dbgs() << *PBI->getParent()->getParent());
3261 // BI may have other predecessors. Because of this, we leave
3262 // it alone, but modify PBI.
3264 // Make sure we get to CommonDest on True&True directions.
3265 Value *PBICond = PBI->getCondition();
3266 IRBuilder<NoFolder> Builder(PBI);
3268 PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3270 Value *BICond = BI->getCondition();
3272 BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3274 // Merge the conditions.
3275 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3277 // Modify PBI to branch on the new condition to the new dests.
3278 PBI->setCondition(Cond);
3279 PBI->setSuccessor(0, CommonDest);
3280 PBI->setSuccessor(1, OtherDest);
3282 // Update branch weight for PBI.
3283 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3284 uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3286 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3287 SuccTrueWeight, SuccFalseWeight);
3289 PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3290 PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3291 SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3292 SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3293 // The weight to CommonDest should be PredCommon * SuccTotal +
3294 // PredOther * SuccCommon.
3295 // The weight to OtherDest should be PredOther * SuccOther.
3296 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3297 PredOther * SuccCommon,
3298 PredOther * SuccOther};
3299 // Halve the weights if any of them cannot fit in an uint32_t
3300 FitWeights(NewWeights);
3302 PBI->setMetadata(LLVMContext::MD_prof,
3303 MDBuilder(BI->getContext())
3304 .createBranchWeights(NewWeights[0], NewWeights[1]));
3307 // OtherDest may have phi nodes. If so, add an entry from PBI's
3308 // block that are identical to the entries for BI's block.
3309 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3311 // We know that the CommonDest already had an edge from PBI to
3312 // it. If it has PHIs though, the PHIs may have different
3313 // entries for BB and PBI's BB. If so, insert a select to make
3316 for (BasicBlock::iterator II = CommonDest->begin();
3317 (PN = dyn_cast<PHINode>(II)); ++II) {
3318 Value *BIV = PN->getIncomingValueForBlock(BB);
3319 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3320 Value *PBIV = PN->getIncomingValue(PBBIdx);
3322 // Insert a select in PBI to pick the right value.
3323 SelectInst *NV = cast<SelectInst>(
3324 Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3325 PN->setIncomingValue(PBBIdx, NV);
3326 // Although the select has the same condition as PBI, the original branch
3327 // weights for PBI do not apply to the new select because the select's
3328 // 'logical' edges are incoming edges of the phi that is eliminated, not
3329 // the outgoing edges of PBI.
3331 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3332 uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3333 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3334 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3335 // The weight to PredCommonDest should be PredCommon * SuccTotal.
3336 // The weight to PredOtherDest should be PredOther * SuccCommon.
3337 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3338 PredOther * SuccCommon};
3340 FitWeights(NewWeights);
3342 NV->setMetadata(LLVMContext::MD_prof,
3343 MDBuilder(BI->getContext())
3344 .createBranchWeights(NewWeights[0], NewWeights[1]));
3349 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3350 DEBUG(dbgs() << *PBI->getParent()->getParent());
3352 // This basic block is probably dead. We know it has at least
3353 // one fewer predecessor.
3357 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3358 // true or to FalseBB if Cond is false.
3359 // Takes care of updating the successors and removing the old terminator.
3360 // Also makes sure not to introduce new successors by assuming that edges to
3361 // non-successor TrueBBs and FalseBBs aren't reachable.
3362 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
3363 BasicBlock *TrueBB, BasicBlock *FalseBB,
3364 uint32_t TrueWeight,
3365 uint32_t FalseWeight) {
3366 // Remove any superfluous successor edges from the CFG.
3367 // First, figure out which successors to preserve.
3368 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3370 BasicBlock *KeepEdge1 = TrueBB;
3371 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3373 // Then remove the rest.
3374 for (BasicBlock *Succ : OldTerm->successors()) {
3375 // Make sure only to keep exactly one copy of each edge.
3376 if (Succ == KeepEdge1)
3377 KeepEdge1 = nullptr;
3378 else if (Succ == KeepEdge2)
3379 KeepEdge2 = nullptr;
3381 Succ->removePredecessor(OldTerm->getParent(),
3382 /*DontDeleteUselessPHIs=*/true);
3385 IRBuilder<> Builder(OldTerm);
3386 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3388 // Insert an appropriate new terminator.
3389 if (!KeepEdge1 && !KeepEdge2) {
3390 if (TrueBB == FalseBB)
3391 // We were only looking for one successor, and it was present.
3392 // Create an unconditional branch to it.
3393 Builder.CreateBr(TrueBB);
3395 // We found both of the successors we were looking for.
3396 // Create a conditional branch sharing the condition of the select.
3397 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3398 if (TrueWeight != FalseWeight)
3399 NewBI->setMetadata(LLVMContext::MD_prof,
3400 MDBuilder(OldTerm->getContext())
3401 .createBranchWeights(TrueWeight, FalseWeight));
3403 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3404 // Neither of the selected blocks were successors, so this
3405 // terminator must be unreachable.
3406 new UnreachableInst(OldTerm->getContext(), OldTerm);
3408 // One of the selected values was a successor, but the other wasn't.
3409 // Insert an unconditional branch to the one that was found;
3410 // the edge to the one that wasn't must be unreachable.
3412 // Only TrueBB was found.
3413 Builder.CreateBr(TrueBB);
3415 // Only FalseBB was found.
3416 Builder.CreateBr(FalseBB);
3419 EraseTerminatorInstAndDCECond(OldTerm);
3424 // (switch (select cond, X, Y)) on constant X, Y
3425 // with a branch - conditional if X and Y lead to distinct BBs,
3426 // unconditional otherwise.
3427 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
3428 // Check for constant integer values in the select.
3429 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3430 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3431 if (!TrueVal || !FalseVal)
3434 // Find the relevant condition and destinations.
3435 Value *Condition = Select->getCondition();
3436 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
3437 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
3439 // Get weight for TrueBB and FalseBB.
3440 uint32_t TrueWeight = 0, FalseWeight = 0;
3441 SmallVector<uint64_t, 8> Weights;
3442 bool HasWeights = HasBranchWeights(SI);
3444 GetBranchWeights(SI, Weights);
3445 if (Weights.size() == 1 + SI->getNumCases()) {
3447 (uint32_t)Weights[SI->findCaseValue(TrueVal).getSuccessorIndex()];
3449 (uint32_t)Weights[SI->findCaseValue(FalseVal).getSuccessorIndex()];
3453 // Perform the actual simplification.
3454 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3459 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3460 // blockaddress(@fn, BlockB)))
3462 // (br cond, BlockA, BlockB).
3463 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
3464 // Check that both operands of the select are block addresses.
3465 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
3466 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
3470 // Extract the actual blocks.
3471 BasicBlock *TrueBB = TBA->getBasicBlock();
3472 BasicBlock *FalseBB = FBA->getBasicBlock();
3474 // Perform the actual simplification.
3475 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3479 /// This is called when we find an icmp instruction
3480 /// (a seteq/setne with a constant) as the only instruction in a
3481 /// block that ends with an uncond branch. We are looking for a very specific
3482 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3483 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3484 /// default value goes to an uncond block with a seteq in it, we get something
3487 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3489 /// %tmp = icmp eq i8 %A, 92
3492 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3494 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3495 /// the PHI, merging the third icmp into the switch.
3496 static bool TryToSimplifyUncondBranchWithICmpInIt(
3497 ICmpInst *ICI, IRBuilder<> &Builder, const DataLayout &DL,
3498 const TargetTransformInfo &TTI, unsigned BonusInstThreshold,
3499 AssumptionCache *AC) {
3500 BasicBlock *BB = ICI->getParent();
3502 // If the block has any PHIs in it or the icmp has multiple uses, it is too
3504 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3507 Value *V = ICI->getOperand(0);
3508 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3510 // The pattern we're looking for is where our only predecessor is a switch on
3511 // 'V' and this block is the default case for the switch. In this case we can
3512 // fold the compared value into the switch to simplify things.
3513 BasicBlock *Pred = BB->getSinglePredecessor();
3514 if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3517 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3518 if (SI->getCondition() != V)
3521 // If BB is reachable on a non-default case, then we simply know the value of
3522 // V in this block. Substitute it and constant fold the icmp instruction
3524 if (SI->getDefaultDest() != BB) {
3525 ConstantInt *VVal = SI->findCaseDest(BB);
3526 assert(VVal && "Should have a unique destination value");
3527 ICI->setOperand(0, VVal);
3529 if (Value *V = SimplifyInstruction(ICI, DL)) {
3530 ICI->replaceAllUsesWith(V);
3531 ICI->eraseFromParent();
3533 // BB is now empty, so it is likely to simplify away.
3534 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3537 // Ok, the block is reachable from the default dest. If the constant we're
3538 // comparing exists in one of the other edges, then we can constant fold ICI
3540 if (SI->findCaseValue(Cst) != SI->case_default()) {
3542 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3543 V = ConstantInt::getFalse(BB->getContext());
3545 V = ConstantInt::getTrue(BB->getContext());
3547 ICI->replaceAllUsesWith(V);
3548 ICI->eraseFromParent();
3549 // BB is now empty, so it is likely to simplify away.
3550 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3553 // The use of the icmp has to be in the 'end' block, by the only PHI node in
3555 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3556 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3557 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3558 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3561 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3563 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3564 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3566 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3567 std::swap(DefaultCst, NewCst);
3569 // Replace ICI (which is used by the PHI for the default value) with true or
3570 // false depending on if it is EQ or NE.
3571 ICI->replaceAllUsesWith(DefaultCst);
3572 ICI->eraseFromParent();
3574 // Okay, the switch goes to this block on a default value. Add an edge from
3575 // the switch to the merge point on the compared value.
3577 BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3578 SmallVector<uint64_t, 8> Weights;
3579 bool HasWeights = HasBranchWeights(SI);
3581 GetBranchWeights(SI, Weights);
3582 if (Weights.size() == 1 + SI->getNumCases()) {
3583 // Split weight for default case to case for "Cst".
3584 Weights[0] = (Weights[0] + 1) >> 1;
3585 Weights.push_back(Weights[0]);
3587 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3589 LLVMContext::MD_prof,
3590 MDBuilder(SI->getContext()).createBranchWeights(MDWeights));
3593 SI->addCase(Cst, NewBB);
3595 // NewBB branches to the phi block, add the uncond branch and the phi entry.
3596 Builder.SetInsertPoint(NewBB);
3597 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3598 Builder.CreateBr(SuccBlock);
3599 PHIUse->addIncoming(NewCst, NewBB);
3603 /// The specified branch is a conditional branch.
3604 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3605 /// fold it into a switch instruction if so.
3606 static bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
3607 const DataLayout &DL) {
3608 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3612 // Change br (X == 0 | X == 1), T, F into a switch instruction.
3613 // If this is a bunch of seteq's or'd together, or if it's a bunch of
3614 // 'setne's and'ed together, collect them.
3616 // Try to gather values from a chain of and/or to be turned into a switch
3617 ConstantComparesGatherer ConstantCompare(Cond, DL);
3618 // Unpack the result
3619 SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3620 Value *CompVal = ConstantCompare.CompValue;
3621 unsigned UsedICmps = ConstantCompare.UsedICmps;
3622 Value *ExtraCase = ConstantCompare.Extra;
3624 // If we didn't have a multiply compared value, fail.
3628 // Avoid turning single icmps into a switch.
3632 bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3634 // There might be duplicate constants in the list, which the switch
3635 // instruction can't handle, remove them now.
3636 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3637 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3639 // If Extra was used, we require at least two switch values to do the
3640 // transformation. A switch with one value is just a conditional branch.
3641 if (ExtraCase && Values.size() < 2)
3644 // TODO: Preserve branch weight metadata, similarly to how
3645 // FoldValueComparisonIntoPredecessors preserves it.
3647 // Figure out which block is which destination.
3648 BasicBlock *DefaultBB = BI->getSuccessor(1);
3649 BasicBlock *EdgeBB = BI->getSuccessor(0);
3651 std::swap(DefaultBB, EdgeBB);
3653 BasicBlock *BB = BI->getParent();
3655 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3656 << " cases into SWITCH. BB is:\n"
3659 // If there are any extra values that couldn't be folded into the switch
3660 // then we evaluate them with an explicit branch first. Split the block
3661 // right before the condbr to handle it.
3664 BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3665 // Remove the uncond branch added to the old block.
3666 TerminatorInst *OldTI = BB->getTerminator();
3667 Builder.SetInsertPoint(OldTI);
3670 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3672 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3674 OldTI->eraseFromParent();
3676 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3677 // for the edge we just added.
3678 AddPredecessorToBlock(EdgeBB, BB, NewBB);
3680 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3681 << "\nEXTRABB = " << *BB);
3685 Builder.SetInsertPoint(BI);
3686 // Convert pointer to int before we switch.
3687 if (CompVal->getType()->isPointerTy()) {
3688 CompVal = Builder.CreatePtrToInt(
3689 CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3692 // Create the new switch instruction now.
3693 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3695 // Add all of the 'cases' to the switch instruction.
3696 for (unsigned i = 0, e = Values.size(); i != e; ++i)
3697 New->addCase(Values[i], EdgeBB);
3699 // We added edges from PI to the EdgeBB. As such, if there were any
3700 // PHI nodes in EdgeBB, they need entries to be added corresponding to
3701 // the number of edges added.
3702 for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3703 PHINode *PN = cast<PHINode>(BBI);
3704 Value *InVal = PN->getIncomingValueForBlock(BB);
3705 for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3706 PN->addIncoming(InVal, BB);
3709 // Erase the old branch instruction.
3710 EraseTerminatorInstAndDCECond(BI);
3712 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3716 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3717 if (isa<PHINode>(RI->getValue()))
3718 return SimplifyCommonResume(RI);
3719 else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3720 RI->getValue() == RI->getParent()->getFirstNonPHI())
3721 // The resume must unwind the exception that caused control to branch here.
3722 return SimplifySingleResume(RI);
3727 // Simplify resume that is shared by several landing pads (phi of landing pad).
3728 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3729 BasicBlock *BB = RI->getParent();
3731 // Check that there are no other instructions except for debug intrinsics
3732 // between the phi of landing pads (RI->getValue()) and resume instruction.
3733 BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3734 E = RI->getIterator();
3736 if (!isa<DbgInfoIntrinsic>(I))
3739 SmallSet<BasicBlock *, 4> TrivialUnwindBlocks;
3740 auto *PhiLPInst = cast<PHINode>(RI->getValue());
3742 // Check incoming blocks to see if any of them are trivial.
3743 for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3745 auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3746 auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3748 // If the block has other successors, we can not delete it because
3749 // it has other dependents.
3750 if (IncomingBB->getUniqueSuccessor() != BB)
3753 auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3754 // Not the landing pad that caused the control to branch here.
3755 if (IncomingValue != LandingPad)
3758 bool isTrivial = true;
3760 I = IncomingBB->getFirstNonPHI()->getIterator();
3761 E = IncomingBB->getTerminator()->getIterator();
3763 if (!isa<DbgInfoIntrinsic>(I)) {
3769 TrivialUnwindBlocks.insert(IncomingBB);
3772 // If no trivial unwind blocks, don't do any simplifications.
3773 if (TrivialUnwindBlocks.empty())
3776 // Turn all invokes that unwind here into calls.
3777 for (auto *TrivialBB : TrivialUnwindBlocks) {
3778 // Blocks that will be simplified should be removed from the phi node.
3779 // Note there could be multiple edges to the resume block, and we need
3780 // to remove them all.
3781 while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3782 BB->removePredecessor(TrivialBB, true);
3784 for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3786 BasicBlock *Pred = *PI++;
3787 removeUnwindEdge(Pred);
3790 // In each SimplifyCFG run, only the current processed block can be erased.
3791 // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3792 // of erasing TrivialBB, we only remove the branch to the common resume
3793 // block so that we can later erase the resume block since it has no
3795 TrivialBB->getTerminator()->eraseFromParent();
3796 new UnreachableInst(RI->getContext(), TrivialBB);
3799 // Delete the resume block if all its predecessors have been removed.
3801 BB->eraseFromParent();
3803 return !TrivialUnwindBlocks.empty();
3806 // Simplify resume that is only used by a single (non-phi) landing pad.
3807 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3808 BasicBlock *BB = RI->getParent();
3809 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
3810 assert(RI->getValue() == LPInst &&
3811 "Resume must unwind the exception that caused control to here");
3813 // Check that there are no other instructions except for debug intrinsics.
3814 BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3816 if (!isa<DbgInfoIntrinsic>(I))
3819 // Turn all invokes that unwind here into calls and delete the basic block.
3820 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3821 BasicBlock *Pred = *PI++;
3822 removeUnwindEdge(Pred);
3825 // The landingpad is now unreachable. Zap it.
3826 BB->eraseFromParent();
3828 LoopHeaders->erase(BB);
3832 static bool removeEmptyCleanup(CleanupReturnInst *RI) {
3833 // If this is a trivial cleanup pad that executes no instructions, it can be
3834 // eliminated. If the cleanup pad continues to the caller, any predecessor
3835 // that is an EH pad will be updated to continue to the caller and any
3836 // predecessor that terminates with an invoke instruction will have its invoke
3837 // instruction converted to a call instruction. If the cleanup pad being
3838 // simplified does not continue to the caller, each predecessor will be
3839 // updated to continue to the unwind destination of the cleanup pad being
3841 BasicBlock *BB = RI->getParent();
3842 CleanupPadInst *CPInst = RI->getCleanupPad();
3843 if (CPInst->getParent() != BB)
3844 // This isn't an empty cleanup.
3847 // We cannot kill the pad if it has multiple uses. This typically arises
3848 // from unreachable basic blocks.
3849 if (!CPInst->hasOneUse())
3852 // Check that there are no other instructions except for benign intrinsics.
3853 BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3855 auto *II = dyn_cast<IntrinsicInst>(I);
3859 Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3860 switch (IntrinsicID) {
3861 case Intrinsic::dbg_declare:
3862 case Intrinsic::dbg_value:
3863 case Intrinsic::lifetime_end:
3870 // If the cleanup return we are simplifying unwinds to the caller, this will
3871 // set UnwindDest to nullptr.
3872 BasicBlock *UnwindDest = RI->getUnwindDest();
3873 Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3875 // We're about to remove BB from the control flow. Before we do, sink any
3876 // PHINodes into the unwind destination. Doing this before changing the
3877 // control flow avoids some potentially slow checks, since we can currently
3878 // be certain that UnwindDest and BB have no common predecessors (since they
3879 // are both EH pads).
3881 // First, go through the PHI nodes in UnwindDest and update any nodes that
3882 // reference the block we are removing
3883 for (BasicBlock::iterator I = UnwindDest->begin(),
3884 IE = DestEHPad->getIterator();
3886 PHINode *DestPN = cast<PHINode>(I);
3888 int Idx = DestPN->getBasicBlockIndex(BB);
3889 // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3891 // This PHI node has an incoming value that corresponds to a control
3892 // path through the cleanup pad we are removing. If the incoming
3893 // value is in the cleanup pad, it must be a PHINode (because we
3894 // verified above that the block is otherwise empty). Otherwise, the
3895 // value is either a constant or a value that dominates the cleanup
3896 // pad being removed.
3898 // Because BB and UnwindDest are both EH pads, all of their
3899 // predecessors must unwind to these blocks, and since no instruction
3900 // can have multiple unwind destinations, there will be no overlap in
3901 // incoming blocks between SrcPN and DestPN.
3902 Value *SrcVal = DestPN->getIncomingValue(Idx);
3903 PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3905 // Remove the entry for the block we are deleting.
3906 DestPN->removeIncomingValue(Idx, false);
3908 if (SrcPN && SrcPN->getParent() == BB) {
3909 // If the incoming value was a PHI node in the cleanup pad we are
3910 // removing, we need to merge that PHI node's incoming values into
3912 for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3913 SrcIdx != SrcE; ++SrcIdx) {
3914 DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3915 SrcPN->getIncomingBlock(SrcIdx));
3918 // Otherwise, the incoming value came from above BB and
3919 // so we can just reuse it. We must associate all of BB's
3920 // predecessors with this value.
3921 for (auto *pred : predecessors(BB)) {
3922 DestPN->addIncoming(SrcVal, pred);
3927 // Sink any remaining PHI nodes directly into UnwindDest.
3928 Instruction *InsertPt = DestEHPad;
3929 for (BasicBlock::iterator I = BB->begin(),
3930 IE = BB->getFirstNonPHI()->getIterator();
3932 // The iterator must be incremented here because the instructions are
3933 // being moved to another block.
3934 PHINode *PN = cast<PHINode>(I++);
3935 if (PN->use_empty())
3936 // If the PHI node has no uses, just leave it. It will be erased
3937 // when we erase BB below.
3940 // Otherwise, sink this PHI node into UnwindDest.
3941 // Any predecessors to UnwindDest which are not already represented
3942 // must be back edges which inherit the value from the path through
3943 // BB. In this case, the PHI value must reference itself.
3944 for (auto *pred : predecessors(UnwindDest))
3946 PN->addIncoming(PN, pred);
3947 PN->moveBefore(InsertPt);
3951 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3952 // The iterator must be updated here because we are removing this pred.
3953 BasicBlock *PredBB = *PI++;
3954 if (UnwindDest == nullptr) {
3955 removeUnwindEdge(PredBB);
3957 TerminatorInst *TI = PredBB->getTerminator();
3958 TI->replaceUsesOfWith(BB, UnwindDest);
3962 // The cleanup pad is now unreachable. Zap it.
3963 BB->eraseFromParent();
3967 // Try to merge two cleanuppads together.
3968 static bool mergeCleanupPad(CleanupReturnInst *RI) {
3969 // Skip any cleanuprets which unwind to caller, there is nothing to merge
3971 BasicBlock *UnwindDest = RI->getUnwindDest();
3975 // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
3976 // be safe to merge without code duplication.
3977 if (UnwindDest->getSinglePredecessor() != RI->getParent())
3980 // Verify that our cleanuppad's unwind destination is another cleanuppad.
3981 auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
3982 if (!SuccessorCleanupPad)
3985 CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
3986 // Replace any uses of the successor cleanupad with the predecessor pad
3987 // The only cleanuppad uses should be this cleanupret, it's cleanupret and
3988 // funclet bundle operands.
3989 SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
3990 // Remove the old cleanuppad.
3991 SuccessorCleanupPad->eraseFromParent();
3992 // Now, we simply replace the cleanupret with a branch to the unwind
3994 BranchInst::Create(UnwindDest, RI->getParent());
3995 RI->eraseFromParent();
4000 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
4001 // It is possible to transiantly have an undef cleanuppad operand because we
4002 // have deleted some, but not all, dead blocks.
4003 // Eventually, this block will be deleted.
4004 if (isa<UndefValue>(RI->getOperand(0)))
4007 if (mergeCleanupPad(RI))
4010 if (removeEmptyCleanup(RI))
4016 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4017 BasicBlock *BB = RI->getParent();
4018 if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4021 // Find predecessors that end with branches.
4022 SmallVector<BasicBlock *, 8> UncondBranchPreds;
4023 SmallVector<BranchInst *, 8> CondBranchPreds;
4024 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4025 BasicBlock *P = *PI;
4026 TerminatorInst *PTI = P->getTerminator();
4027 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4028 if (BI->isUnconditional())
4029 UncondBranchPreds.push_back(P);
4031 CondBranchPreds.push_back(BI);
4035 // If we found some, do the transformation!
4036 if (!UncondBranchPreds.empty() && DupRet) {
4037 while (!UncondBranchPreds.empty()) {
4038 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4039 DEBUG(dbgs() << "FOLDING: " << *BB
4040 << "INTO UNCOND BRANCH PRED: " << *Pred);
4041 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4044 // If we eliminated all predecessors of the block, delete the block now.
4045 if (pred_empty(BB)) {
4046 // We know there are no successors, so just nuke the block.
4047 BB->eraseFromParent();
4049 LoopHeaders->erase(BB);
4055 // Check out all of the conditional branches going to this return
4056 // instruction. If any of them just select between returns, change the
4057 // branch itself into a select/return pair.
4058 while (!CondBranchPreds.empty()) {
4059 BranchInst *BI = CondBranchPreds.pop_back_val();
4061 // Check to see if the non-BB successor is also a return block.
4062 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4063 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4064 SimplifyCondBranchToTwoReturns(BI, Builder))
4070 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4071 BasicBlock *BB = UI->getParent();
4073 bool Changed = false;
4075 // If there are any instructions immediately before the unreachable that can
4076 // be removed, do so.
4077 while (UI->getIterator() != BB->begin()) {
4078 BasicBlock::iterator BBI = UI->getIterator();
4080 // Do not delete instructions that can have side effects which might cause
4081 // the unreachable to not be reachable; specifically, calls and volatile
4082 // operations may have this effect.
4083 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4086 if (BBI->mayHaveSideEffects()) {
4087 if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4088 if (SI->isVolatile())
4090 } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4091 if (LI->isVolatile())
4093 } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4094 if (RMWI->isVolatile())
4096 } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4097 if (CXI->isVolatile())
4099 } else if (isa<CatchPadInst>(BBI)) {
4100 // A catchpad may invoke exception object constructors and such, which
4101 // in some languages can be arbitrary code, so be conservative by
4103 // For CoreCLR, it just involves a type test, so can be removed.
4104 if (classifyEHPersonality(BB->getParent()->getPersonalityFn()) !=
4105 EHPersonality::CoreCLR)
4107 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4108 !isa<LandingPadInst>(BBI)) {
4111 // Note that deleting LandingPad's here is in fact okay, although it
4112 // involves a bit of subtle reasoning. If this inst is a LandingPad,
4113 // all the predecessors of this block will be the unwind edges of Invokes,
4114 // and we can therefore guarantee this block will be erased.
4117 // Delete this instruction (any uses are guaranteed to be dead)
4118 if (!BBI->use_empty())
4119 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4120 BBI->eraseFromParent();
4124 // If the unreachable instruction is the first in the block, take a gander
4125 // at all of the predecessors of this instruction, and simplify them.
4126 if (&BB->front() != UI)
4129 SmallVector<BasicBlock *, 8> Preds(pred_begin(BB), pred_end(BB));
4130 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4131 TerminatorInst *TI = Preds[i]->getTerminator();
4132 IRBuilder<> Builder(TI);
4133 if (auto *BI = dyn_cast<BranchInst>(TI)) {
4134 if (BI->isUnconditional()) {
4135 if (BI->getSuccessor(0) == BB) {
4136 new UnreachableInst(TI->getContext(), TI);
4137 TI->eraseFromParent();
4141 if (BI->getSuccessor(0) == BB) {
4142 Builder.CreateBr(BI->getSuccessor(1));
4143 EraseTerminatorInstAndDCECond(BI);
4144 } else if (BI->getSuccessor(1) == BB) {
4145 Builder.CreateBr(BI->getSuccessor(0));
4146 EraseTerminatorInstAndDCECond(BI);
4150 } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4151 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e;
4153 if (i.getCaseSuccessor() == BB) {
4154 BB->removePredecessor(SI->getParent());
4160 } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4161 if (II->getUnwindDest() == BB) {
4162 removeUnwindEdge(TI->getParent());
4165 } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4166 if (CSI->getUnwindDest() == BB) {
4167 removeUnwindEdge(TI->getParent());
4172 for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4173 E = CSI->handler_end();
4176 CSI->removeHandler(I);
4182 if (CSI->getNumHandlers() == 0) {
4183 BasicBlock *CatchSwitchBB = CSI->getParent();
4184 if (CSI->hasUnwindDest()) {
4185 // Redirect preds to the unwind dest
4186 CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4188 // Rewrite all preds to unwind to caller (or from invoke to call).
4189 SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4190 for (BasicBlock *EHPred : EHPreds)
4191 removeUnwindEdge(EHPred);
4193 // The catchswitch is no longer reachable.
4194 new UnreachableInst(CSI->getContext(), CSI);
4195 CSI->eraseFromParent();
4198 } else if (isa<CleanupReturnInst>(TI)) {
4199 new UnreachableInst(TI->getContext(), TI);
4200 TI->eraseFromParent();
4205 // If this block is now dead, remove it.
4206 if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4207 // We know there are no successors, so just nuke the block.
4208 BB->eraseFromParent();
4210 LoopHeaders->erase(BB);
4217 static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) {
4218 assert(Cases.size() >= 1);
4220 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4221 for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4222 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4228 /// Turn a switch with two reachable destinations into an integer range
4229 /// comparison and branch.
4230 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
4231 assert(SI->getNumCases() > 1 && "Degenerate switch?");
4234 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4236 // Partition the cases into two sets with different destinations.
4237 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4238 BasicBlock *DestB = nullptr;
4239 SmallVector<ConstantInt *, 16> CasesA;
4240 SmallVector<ConstantInt *, 16> CasesB;
4242 for (SwitchInst::CaseIt I : SI->cases()) {
4243 BasicBlock *Dest = I.getCaseSuccessor();
4246 if (Dest == DestA) {
4247 CasesA.push_back(I.getCaseValue());
4252 if (Dest == DestB) {
4253 CasesB.push_back(I.getCaseValue());
4256 return false; // More than two destinations.
4259 assert(DestA && DestB &&
4260 "Single-destination switch should have been folded.");
4261 assert(DestA != DestB);
4262 assert(DestB != SI->getDefaultDest());
4263 assert(!CasesB.empty() && "There must be non-default cases.");
4264 assert(!CasesA.empty() || HasDefault);
4266 // Figure out if one of the sets of cases form a contiguous range.
4267 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4268 BasicBlock *ContiguousDest = nullptr;
4269 BasicBlock *OtherDest = nullptr;
4270 if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4271 ContiguousCases = &CasesA;
4272 ContiguousDest = DestA;
4274 } else if (CasesAreContiguous(CasesB)) {
4275 ContiguousCases = &CasesB;
4276 ContiguousDest = DestB;
4281 // Start building the compare and branch.
4283 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4284 Constant *NumCases =
4285 ConstantInt::get(Offset->getType(), ContiguousCases->size());
4287 Value *Sub = SI->getCondition();
4288 if (!Offset->isNullValue())
4289 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4292 // If NumCases overflowed, then all possible values jump to the successor.
4293 if (NumCases->isNullValue() && !ContiguousCases->empty())
4294 Cmp = ConstantInt::getTrue(SI->getContext());
4296 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4297 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4299 // Update weight for the newly-created conditional branch.
4300 if (HasBranchWeights(SI)) {
4301 SmallVector<uint64_t, 8> Weights;
4302 GetBranchWeights(SI, Weights);
4303 if (Weights.size() == 1 + SI->getNumCases()) {
4304 uint64_t TrueWeight = 0;
4305 uint64_t FalseWeight = 0;
4306 for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4307 if (SI->getSuccessor(I) == ContiguousDest)
4308 TrueWeight += Weights[I];
4310 FalseWeight += Weights[I];
4312 while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4316 NewBI->setMetadata(LLVMContext::MD_prof,
4317 MDBuilder(SI->getContext())
4318 .createBranchWeights((uint32_t)TrueWeight,
4319 (uint32_t)FalseWeight));
4323 // Prune obsolete incoming values off the successors' PHI nodes.
4324 for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4325 unsigned PreviousEdges = ContiguousCases->size();
4326 if (ContiguousDest == SI->getDefaultDest())
4328 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4329 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4331 for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4332 unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4333 if (OtherDest == SI->getDefaultDest())
4335 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4336 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4340 SI->eraseFromParent();
4345 /// Compute masked bits for the condition of a switch
4346 /// and use it to remove dead cases.
4347 static bool EliminateDeadSwitchCases(SwitchInst *SI, AssumptionCache *AC,
4348 const DataLayout &DL) {
4349 Value *Cond = SI->getCondition();
4350 unsigned Bits = Cond->getType()->getIntegerBitWidth();
4351 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
4352 computeKnownBits(Cond, KnownZero, KnownOne, DL, 0, AC, SI);
4354 // We can also eliminate cases by determining that their values are outside of
4355 // the limited range of the condition based on how many significant (non-sign)
4356 // bits are in the condition value.
4357 unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4358 unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4360 // Gather dead cases.
4361 SmallVector<ConstantInt *, 8> DeadCases;
4362 for (auto &Case : SI->cases()) {
4363 APInt CaseVal = Case.getCaseValue()->getValue();
4364 if ((CaseVal & KnownZero) != 0 || (CaseVal & KnownOne) != KnownOne ||
4365 (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4366 DeadCases.push_back(Case.getCaseValue());
4367 DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal << " is dead.\n");
4371 // If we can prove that the cases must cover all possible values, the
4372 // default destination becomes dead and we can remove it. If we know some
4373 // of the bits in the value, we can use that to more precisely compute the
4374 // number of possible unique case values.
4376 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4377 const unsigned NumUnknownBits =
4378 Bits - (KnownZero.Or(KnownOne)).countPopulation();
4379 assert(NumUnknownBits <= Bits);
4380 if (HasDefault && DeadCases.empty() &&
4381 NumUnknownBits < 64 /* avoid overflow */ &&
4382 SI->getNumCases() == (1ULL << NumUnknownBits)) {
4383 DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4384 BasicBlock *NewDefault =
4385 SplitBlockPredecessors(SI->getDefaultDest(), SI->getParent(), "");
4386 SI->setDefaultDest(&*NewDefault);
4387 SplitBlock(&*NewDefault, &NewDefault->front());
4388 auto *OldTI = NewDefault->getTerminator();
4389 new UnreachableInst(SI->getContext(), OldTI);
4390 EraseTerminatorInstAndDCECond(OldTI);
4394 SmallVector<uint64_t, 8> Weights;
4395 bool HasWeight = HasBranchWeights(SI);
4397 GetBranchWeights(SI, Weights);
4398 HasWeight = (Weights.size() == 1 + SI->getNumCases());
4401 // Remove dead cases from the switch.
4402 for (ConstantInt *DeadCase : DeadCases) {
4403 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCase);
4404 assert(Case != SI->case_default() &&
4405 "Case was not found. Probably mistake in DeadCases forming.");
4407 std::swap(Weights[Case.getCaseIndex() + 1], Weights.back());
4411 // Prune unused values from PHI nodes.
4412 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
4413 SI->removeCase(Case);
4415 if (HasWeight && Weights.size() >= 2) {
4416 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
4417 SI->setMetadata(LLVMContext::MD_prof,
4418 MDBuilder(SI->getParent()->getContext())
4419 .createBranchWeights(MDWeights));
4422 return !DeadCases.empty();
4425 /// If BB would be eligible for simplification by
4426 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4427 /// by an unconditional branch), look at the phi node for BB in the successor
4428 /// block and see if the incoming value is equal to CaseValue. If so, return
4429 /// the phi node, and set PhiIndex to BB's index in the phi node.
4430 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
4431 BasicBlock *BB, int *PhiIndex) {
4432 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4433 return nullptr; // BB must be empty to be a candidate for simplification.
4434 if (!BB->getSinglePredecessor())
4435 return nullptr; // BB must be dominated by the switch.
4437 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
4438 if (!Branch || !Branch->isUnconditional())
4439 return nullptr; // Terminator must be unconditional branch.
4441 BasicBlock *Succ = Branch->getSuccessor(0);
4443 BasicBlock::iterator I = Succ->begin();
4444 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
4445 int Idx = PHI->getBasicBlockIndex(BB);
4446 assert(Idx >= 0 && "PHI has no entry for predecessor?");
4448 Value *InValue = PHI->getIncomingValue(Idx);
4449 if (InValue != CaseValue)
4459 /// Try to forward the condition of a switch instruction to a phi node
4460 /// dominated by the switch, if that would mean that some of the destination
4461 /// blocks of the switch can be folded away.
4462 /// Returns true if a change is made.
4463 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
4464 typedef DenseMap<PHINode *, SmallVector<int, 4>> ForwardingNodesMap;
4465 ForwardingNodesMap ForwardingNodes;
4467 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E;
4469 ConstantInt *CaseValue = I.getCaseValue();
4470 BasicBlock *CaseDest = I.getCaseSuccessor();
4474 FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIndex);
4478 ForwardingNodes[PHI].push_back(PhiIndex);
4481 bool Changed = false;
4483 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
4484 E = ForwardingNodes.end();
4486 PHINode *Phi = I->first;
4487 SmallVectorImpl<int> &Indexes = I->second;
4489 if (Indexes.size() < 2)
4492 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
4493 Phi->setIncomingValue(Indexes[I], SI->getCondition());
4500 /// Return true if the backend will be able to handle
4501 /// initializing an array of constants like C.
4502 static bool ValidLookupTableConstant(Constant *C, const TargetTransformInfo &TTI) {
4503 if (C->isThreadDependent())
4505 if (C->isDLLImportDependent())
4508 if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4509 !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4510 !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4513 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4514 if (!CE->isGEPWithNoNotionalOverIndexing())
4516 if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4520 if (!TTI.shouldBuildLookupTablesForConstant(C))
4526 /// If V is a Constant, return it. Otherwise, try to look up
4527 /// its constant value in ConstantPool, returning 0 if it's not there.
4529 LookupConstant(Value *V,
4530 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4531 if (Constant *C = dyn_cast<Constant>(V))
4533 return ConstantPool.lookup(V);
4536 /// Try to fold instruction I into a constant. This works for
4537 /// simple instructions such as binary operations where both operands are
4538 /// constant or can be replaced by constants from the ConstantPool. Returns the
4539 /// resulting constant on success, 0 otherwise.
4541 ConstantFold(Instruction *I, const DataLayout &DL,
4542 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4543 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4544 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4547 if (A->isAllOnesValue())
4548 return LookupConstant(Select->getTrueValue(), ConstantPool);
4549 if (A->isNullValue())
4550 return LookupConstant(Select->getFalseValue(), ConstantPool);
4554 SmallVector<Constant *, 4> COps;
4555 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4556 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
4562 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4563 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4567 return ConstantFoldInstOperands(I, COps, DL);
4570 /// Try to determine the resulting constant values in phi nodes
4571 /// at the common destination basic block, *CommonDest, for one of the case
4572 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4573 /// case), of a switch instruction SI.
4575 GetCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest,
4576 BasicBlock **CommonDest,
4577 SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4578 const DataLayout &DL, const TargetTransformInfo &TTI) {
4579 // The block from which we enter the common destination.
4580 BasicBlock *Pred = SI->getParent();
4582 // If CaseDest is empty except for some side-effect free instructions through
4583 // which we can constant-propagate the CaseVal, continue to its successor.
4584 SmallDenseMap<Value *, Constant *> ConstantPool;
4585 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4586 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
4588 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
4589 // If the terminator is a simple branch, continue to the next block.
4590 if (T->getNumSuccessors() != 1 || T->isExceptional())
4593 CaseDest = T->getSuccessor(0);
4594 } else if (isa<DbgInfoIntrinsic>(I)) {
4595 // Skip debug intrinsic.
4597 } else if (Constant *C = ConstantFold(&*I, DL, ConstantPool)) {
4598 // Instruction is side-effect free and constant.
4600 // If the instruction has uses outside this block or a phi node slot for
4601 // the block, it is not safe to bypass the instruction since it would then
4602 // no longer dominate all its uses.
4603 for (auto &Use : I->uses()) {
4604 User *User = Use.getUser();
4605 if (Instruction *I = dyn_cast<Instruction>(User))
4606 if (I->getParent() == CaseDest)
4608 if (PHINode *Phi = dyn_cast<PHINode>(User))
4609 if (Phi->getIncomingBlock(Use) == CaseDest)
4614 ConstantPool.insert(std::make_pair(&*I, C));
4620 // If we did not have a CommonDest before, use the current one.
4622 *CommonDest = CaseDest;
4623 // If the destination isn't the common one, abort.
4624 if (CaseDest != *CommonDest)
4627 // Get the values for this case from phi nodes in the destination block.
4628 BasicBlock::iterator I = (*CommonDest)->begin();
4629 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
4630 int Idx = PHI->getBasicBlockIndex(Pred);
4634 Constant *ConstVal =
4635 LookupConstant(PHI->getIncomingValue(Idx), ConstantPool);
4639 // Be conservative about which kinds of constants we support.
4640 if (!ValidLookupTableConstant(ConstVal, TTI))
4643 Res.push_back(std::make_pair(PHI, ConstVal));
4646 return Res.size() > 0;
4649 // Helper function used to add CaseVal to the list of cases that generate
4651 static void MapCaseToResult(ConstantInt *CaseVal,
4652 SwitchCaseResultVectorTy &UniqueResults,
4654 for (auto &I : UniqueResults) {
4655 if (I.first == Result) {
4656 I.second.push_back(CaseVal);
4660 UniqueResults.push_back(
4661 std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4664 // Helper function that initializes a map containing
4665 // results for the PHI node of the common destination block for a switch
4666 // instruction. Returns false if multiple PHI nodes have been found or if
4667 // there is not a common destination block for the switch.
4668 static bool InitializeUniqueCases(SwitchInst *SI, PHINode *&PHI,
4669 BasicBlock *&CommonDest,
4670 SwitchCaseResultVectorTy &UniqueResults,
4671 Constant *&DefaultResult,
4672 const DataLayout &DL,
4673 const TargetTransformInfo &TTI) {
4674 for (auto &I : SI->cases()) {
4675 ConstantInt *CaseVal = I.getCaseValue();
4677 // Resulting value at phi nodes for this case value.
4678 SwitchCaseResultsTy Results;
4679 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4683 // Only one value per case is permitted
4684 if (Results.size() > 1)
4686 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4688 // Check the PHI consistency.
4690 PHI = Results[0].first;
4691 else if (PHI != Results[0].first)
4694 // Find the default result value.
4695 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
4696 BasicBlock *DefaultDest = SI->getDefaultDest();
4697 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4699 // If the default value is not found abort unless the default destination
4702 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4703 if ((!DefaultResult &&
4704 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4710 // Helper function that checks if it is possible to transform a switch with only
4711 // two cases (or two cases + default) that produces a result into a select.
4714 // case 10: %0 = icmp eq i32 %a, 10
4715 // return 10; %1 = select i1 %0, i32 10, i32 4
4716 // case 20: ----> %2 = icmp eq i32 %a, 20
4717 // return 2; %3 = select i1 %2, i32 2, i32 %1
4721 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4722 Constant *DefaultResult, Value *Condition,
4723 IRBuilder<> &Builder) {
4724 assert(ResultVector.size() == 2 &&
4725 "We should have exactly two unique results at this point");
4726 // If we are selecting between only two cases transform into a simple
4727 // select or a two-way select if default is possible.
4728 if (ResultVector[0].second.size() == 1 &&
4729 ResultVector[1].second.size() == 1) {
4730 ConstantInt *const FirstCase = ResultVector[0].second[0];
4731 ConstantInt *const SecondCase = ResultVector[1].second[0];
4733 bool DefaultCanTrigger = DefaultResult;
4734 Value *SelectValue = ResultVector[1].first;
4735 if (DefaultCanTrigger) {
4736 Value *const ValueCompare =
4737 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4738 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4739 DefaultResult, "switch.select");
4741 Value *const ValueCompare =
4742 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4743 return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4744 SelectValue, "switch.select");
4750 // Helper function to cleanup a switch instruction that has been converted into
4751 // a select, fixing up PHI nodes and basic blocks.
4752 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
4754 IRBuilder<> &Builder) {
4755 BasicBlock *SelectBB = SI->getParent();
4756 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4757 PHI->removeIncomingValue(SelectBB);
4758 PHI->addIncoming(SelectValue, SelectBB);
4760 Builder.CreateBr(PHI->getParent());
4762 // Remove the switch.
4763 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4764 BasicBlock *Succ = SI->getSuccessor(i);
4766 if (Succ == PHI->getParent())
4768 Succ->removePredecessor(SelectBB);
4770 SI->eraseFromParent();
4773 /// If the switch is only used to initialize one or more
4774 /// phi nodes in a common successor block with only two different
4775 /// constant values, replace the switch with select.
4776 static bool SwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4777 AssumptionCache *AC, const DataLayout &DL,
4778 const TargetTransformInfo &TTI) {
4779 Value *const Cond = SI->getCondition();
4780 PHINode *PHI = nullptr;
4781 BasicBlock *CommonDest = nullptr;
4782 Constant *DefaultResult;
4783 SwitchCaseResultVectorTy UniqueResults;
4784 // Collect all the cases that will deliver the same value from the switch.
4785 if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4788 // Selects choose between maximum two values.
4789 if (UniqueResults.size() != 2)
4791 assert(PHI != nullptr && "PHI for value select not found");
4793 Builder.SetInsertPoint(SI);
4794 Value *SelectValue =
4795 ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4797 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4800 // The switch couldn't be converted into a select.
4806 /// This class represents a lookup table that can be used to replace a switch.
4807 class SwitchLookupTable {
4809 /// Create a lookup table to use as a switch replacement with the contents
4810 /// of Values, using DefaultValue to fill any holes in the table.
4812 Module &M, uint64_t TableSize, ConstantInt *Offset,
4813 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4814 Constant *DefaultValue, const DataLayout &DL);
4816 /// Build instructions with Builder to retrieve the value at
4817 /// the position given by Index in the lookup table.
4818 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4820 /// Return true if a table with TableSize elements of
4821 /// type ElementType would fit in a target-legal register.
4822 static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4826 // Depending on the contents of the table, it can be represented in
4829 // For tables where each element contains the same value, we just have to
4830 // store that single value and return it for each lookup.
4833 // For tables where there is a linear relationship between table index
4834 // and values. We calculate the result with a simple multiplication
4835 // and addition instead of a table lookup.
4838 // For small tables with integer elements, we can pack them into a bitmap
4839 // that fits into a target-legal register. Values are retrieved by
4840 // shift and mask operations.
4843 // The table is stored as an array of values. Values are retrieved by load
4844 // instructions from the table.
4848 // For SingleValueKind, this is the single value.
4849 Constant *SingleValue;
4851 // For BitMapKind, this is the bitmap.
4852 ConstantInt *BitMap;
4853 IntegerType *BitMapElementTy;
4855 // For LinearMapKind, these are the constants used to derive the value.
4856 ConstantInt *LinearOffset;
4857 ConstantInt *LinearMultiplier;
4859 // For ArrayKind, this is the array.
4860 GlobalVariable *Array;
4863 } // end anonymous namespace
4865 SwitchLookupTable::SwitchLookupTable(
4866 Module &M, uint64_t TableSize, ConstantInt *Offset,
4867 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4868 Constant *DefaultValue, const DataLayout &DL)
4869 : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
4870 LinearOffset(nullptr), LinearMultiplier(nullptr), Array(nullptr) {
4871 assert(Values.size() && "Can't build lookup table without values!");
4872 assert(TableSize >= Values.size() && "Can't fit values in table!");
4874 // If all values in the table are equal, this is that value.
4875 SingleValue = Values.begin()->second;
4877 Type *ValueType = Values.begin()->second->getType();
4879 // Build up the table contents.
4880 SmallVector<Constant *, 64> TableContents(TableSize);
4881 for (size_t I = 0, E = Values.size(); I != E; ++I) {
4882 ConstantInt *CaseVal = Values[I].first;
4883 Constant *CaseRes = Values[I].second;
4884 assert(CaseRes->getType() == ValueType);
4886 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4887 TableContents[Idx] = CaseRes;
4889 if (CaseRes != SingleValue)
4890 SingleValue = nullptr;
4893 // Fill in any holes in the table with the default result.
4894 if (Values.size() < TableSize) {
4895 assert(DefaultValue &&
4896 "Need a default value to fill the lookup table holes.");
4897 assert(DefaultValue->getType() == ValueType);
4898 for (uint64_t I = 0; I < TableSize; ++I) {
4899 if (!TableContents[I])
4900 TableContents[I] = DefaultValue;
4903 if (DefaultValue != SingleValue)
4904 SingleValue = nullptr;
4907 // If each element in the table contains the same value, we only need to store
4908 // that single value.
4910 Kind = SingleValueKind;
4914 // Check if we can derive the value with a linear transformation from the
4916 if (isa<IntegerType>(ValueType)) {
4917 bool LinearMappingPossible = true;
4920 assert(TableSize >= 2 && "Should be a SingleValue table.");
4921 // Check if there is the same distance between two consecutive values.
4922 for (uint64_t I = 0; I < TableSize; ++I) {
4923 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4925 // This is an undef. We could deal with it, but undefs in lookup tables
4926 // are very seldom. It's probably not worth the additional complexity.
4927 LinearMappingPossible = false;
4930 APInt Val = ConstVal->getValue();
4932 APInt Dist = Val - PrevVal;
4935 } else if (Dist != DistToPrev) {
4936 LinearMappingPossible = false;
4942 if (LinearMappingPossible) {
4943 LinearOffset = cast<ConstantInt>(TableContents[0]);
4944 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
4945 Kind = LinearMapKind;
4951 // If the type is integer and the table fits in a register, build a bitmap.
4952 if (WouldFitInRegister(DL, TableSize, ValueType)) {
4953 IntegerType *IT = cast<IntegerType>(ValueType);
4954 APInt TableInt(TableSize * IT->getBitWidth(), 0);
4955 for (uint64_t I = TableSize; I > 0; --I) {
4956 TableInt <<= IT->getBitWidth();
4957 // Insert values into the bitmap. Undef values are set to zero.
4958 if (!isa<UndefValue>(TableContents[I - 1])) {
4959 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
4960 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
4963 BitMap = ConstantInt::get(M.getContext(), TableInt);
4964 BitMapElementTy = IT;
4970 // Store the table in an array.
4971 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
4972 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
4974 Array = new GlobalVariable(M, ArrayTy, /*constant=*/true,
4975 GlobalVariable::PrivateLinkage, Initializer,
4977 Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
4981 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
4983 case SingleValueKind:
4985 case LinearMapKind: {
4986 // Derive the result value from the input value.
4987 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
4988 false, "switch.idx.cast");
4989 if (!LinearMultiplier->isOne())
4990 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
4991 if (!LinearOffset->isZero())
4992 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
4996 // Type of the bitmap (e.g. i59).
4997 IntegerType *MapTy = BitMap->getType();
4999 // Cast Index to the same type as the bitmap.
5000 // Note: The Index is <= the number of elements in the table, so
5001 // truncating it to the width of the bitmask is safe.
5002 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
5004 // Multiply the shift amount by the element width.
5005 ShiftAmt = Builder.CreateMul(
5006 ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5010 Value *DownShifted =
5011 Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5013 return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5016 // Make sure the table index will not overflow when treated as signed.
5017 IntegerType *IT = cast<IntegerType>(Index->getType());
5018 uint64_t TableSize =
5019 Array->getInitializer()->getType()->getArrayNumElements();
5020 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5021 Index = Builder.CreateZExt(
5022 Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5023 "switch.tableidx.zext");
5025 Value *GEPIndices[] = {Builder.getInt32(0), Index};
5026 Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5027 GEPIndices, "switch.gep");
5028 return Builder.CreateLoad(GEP, "switch.load");
5031 llvm_unreachable("Unknown lookup table kind!");
5034 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5036 Type *ElementType) {
5037 auto *IT = dyn_cast<IntegerType>(ElementType);
5040 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5041 // are <= 15, we could try to narrow the type.
5043 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5044 if (TableSize >= UINT_MAX / IT->getBitWidth())
5046 return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5049 /// Determine whether a lookup table should be built for this switch, based on
5050 /// the number of cases, size of the table, and the types of the results.
5052 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5053 const TargetTransformInfo &TTI, const DataLayout &DL,
5054 const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5055 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5056 return false; // TableSize overflowed, or mul below might overflow.
5058 bool AllTablesFitInRegister = true;
5059 bool HasIllegalType = false;
5060 for (const auto &I : ResultTypes) {
5061 Type *Ty = I.second;
5063 // Saturate this flag to true.
5064 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5066 // Saturate this flag to false.
5067 AllTablesFitInRegister =
5068 AllTablesFitInRegister &&
5069 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5071 // If both flags saturate, we're done. NOTE: This *only* works with
5072 // saturating flags, and all flags have to saturate first due to the
5073 // non-deterministic behavior of iterating over a dense map.
5074 if (HasIllegalType && !AllTablesFitInRegister)
5078 // If each table would fit in a register, we should build it anyway.
5079 if (AllTablesFitInRegister)
5082 // Don't build a table that doesn't fit in-register if it has illegal types.
5086 // The table density should be at least 40%. This is the same criterion as for
5087 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5088 // FIXME: Find the best cut-off.
5089 return SI->getNumCases() * 10 >= TableSize * 4;
5092 /// Try to reuse the switch table index compare. Following pattern:
5094 /// if (idx < tablesize)
5095 /// r = table[idx]; // table does not contain default_value
5097 /// r = default_value;
5098 /// if (r != default_value)
5101 /// Is optimized to:
5103 /// cond = idx < tablesize;
5107 /// r = default_value;
5111 /// Jump threading will then eliminate the second if(cond).
5112 static void reuseTableCompare(
5113 User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5114 Constant *DefaultValue,
5115 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5117 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5121 // We require that the compare is in the same block as the phi so that jump
5122 // threading can do its work afterwards.
5123 if (CmpInst->getParent() != PhiBlock)
5126 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5130 Value *RangeCmp = RangeCheckBranch->getCondition();
5131 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5132 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5134 // Check if the compare with the default value is constant true or false.
5135 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5136 DefaultValue, CmpOp1, true);
5137 if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5140 // Check if the compare with the case values is distinct from the default
5142 for (auto ValuePair : Values) {
5143 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5144 ValuePair.second, CmpOp1, true);
5145 if (!CaseConst || CaseConst == DefaultConst)
5147 assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5148 "Expect true or false as compare result.");
5151 // Check if the branch instruction dominates the phi node. It's a simple
5152 // dominance check, but sufficient for our needs.
5153 // Although this check is invariant in the calling loops, it's better to do it
5154 // at this late stage. Practically we do it at most once for a switch.
5155 BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5156 for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5157 BasicBlock *Pred = *PI;
5158 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5162 if (DefaultConst == FalseConst) {
5163 // The compare yields the same result. We can replace it.
5164 CmpInst->replaceAllUsesWith(RangeCmp);
5165 ++NumTableCmpReuses;
5167 // The compare yields the same result, just inverted. We can replace it.
5168 Value *InvertedTableCmp = BinaryOperator::CreateXor(
5169 RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5171 CmpInst->replaceAllUsesWith(InvertedTableCmp);
5172 ++NumTableCmpReuses;
5176 /// If the switch is only used to initialize one or more phi nodes in a common
5177 /// successor block with different constant values, replace the switch with
5179 static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
5180 const DataLayout &DL,
5181 const TargetTransformInfo &TTI) {
5182 assert(SI->getNumCases() > 1 && "Degenerate switch?");
5184 // Only build lookup table when we have a target that supports it.
5185 if (!TTI.shouldBuildLookupTables())
5188 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5189 // split off a dense part and build a lookup table for that.
5191 // FIXME: This creates arrays of GEPs to constant strings, which means each
5192 // GEP needs a runtime relocation in PIC code. We should just build one big
5193 // string and lookup indices into that.
5195 // Ignore switches with less than three cases. Lookup tables will not make
5197 // faster, so we don't analyze them.
5198 if (SI->getNumCases() < 3)
5201 // Figure out the corresponding result for each case value and phi node in the
5202 // common destination, as well as the min and max case values.
5203 assert(SI->case_begin() != SI->case_end());
5204 SwitchInst::CaseIt CI = SI->case_begin();
5205 ConstantInt *MinCaseVal = CI.getCaseValue();
5206 ConstantInt *MaxCaseVal = CI.getCaseValue();
5208 BasicBlock *CommonDest = nullptr;
5209 typedef SmallVector<std::pair<ConstantInt *, Constant *>, 4> ResultListTy;
5210 SmallDenseMap<PHINode *, ResultListTy> ResultLists;
5211 SmallDenseMap<PHINode *, Constant *> DefaultResults;
5212 SmallDenseMap<PHINode *, Type *> ResultTypes;
5213 SmallVector<PHINode *, 4> PHIs;
5215 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5216 ConstantInt *CaseVal = CI.getCaseValue();
5217 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5218 MinCaseVal = CaseVal;
5219 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5220 MaxCaseVal = CaseVal;
5222 // Resulting value at phi nodes for this case value.
5223 typedef SmallVector<std::pair<PHINode *, Constant *>, 4> ResultsTy;
5225 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
5229 // Append the result from this case to the list for each phi.
5230 for (const auto &I : Results) {
5231 PHINode *PHI = I.first;
5232 Constant *Value = I.second;
5233 if (!ResultLists.count(PHI))
5234 PHIs.push_back(PHI);
5235 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5239 // Keep track of the result types.
5240 for (PHINode *PHI : PHIs) {
5241 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5244 uint64_t NumResults = ResultLists[PHIs[0]].size();
5245 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5246 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5247 bool TableHasHoles = (NumResults < TableSize);
5249 // If the table has holes, we need a constant result for the default case
5250 // or a bitmask that fits in a register.
5251 SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5252 bool HasDefaultResults =
5253 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5254 DefaultResultsList, DL, TTI);
5256 bool NeedMask = (TableHasHoles && !HasDefaultResults);
5258 // As an extra penalty for the validity test we require more cases.
5259 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5261 if (!DL.fitsInLegalInteger(TableSize))
5265 for (const auto &I : DefaultResultsList) {
5266 PHINode *PHI = I.first;
5267 Constant *Result = I.second;
5268 DefaultResults[PHI] = Result;
5271 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5274 // Create the BB that does the lookups.
5275 Module &Mod = *CommonDest->getParent()->getParent();
5276 BasicBlock *LookupBB = BasicBlock::Create(
5277 Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5279 // Compute the table index value.
5280 Builder.SetInsertPoint(SI);
5282 Builder.CreateSub(SI->getCondition(), MinCaseVal, "switch.tableidx");
5284 // Compute the maximum table size representable by the integer type we are
5286 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5287 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5288 assert(MaxTableSize >= TableSize &&
5289 "It is impossible for a switch to have more entries than the max "
5290 "representable value of its input integer type's size.");
5292 // If the default destination is unreachable, or if the lookup table covers
5293 // all values of the conditional variable, branch directly to the lookup table
5294 // BB. Otherwise, check that the condition is within the case range.
5295 const bool DefaultIsReachable =
5296 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5297 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5298 BranchInst *RangeCheckBranch = nullptr;
5300 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5301 Builder.CreateBr(LookupBB);
5302 // Note: We call removeProdecessor later since we need to be able to get the
5303 // PHI value for the default case in case we're using a bit mask.
5305 Value *Cmp = Builder.CreateICmpULT(
5306 TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5308 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5311 // Populate the BB that does the lookups.
5312 Builder.SetInsertPoint(LookupBB);
5315 // Before doing the lookup we do the hole check.
5316 // The LookupBB is therefore re-purposed to do the hole check
5317 // and we create a new LookupBB.
5318 BasicBlock *MaskBB = LookupBB;
5319 MaskBB->setName("switch.hole_check");
5320 LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5321 CommonDest->getParent(), CommonDest);
5323 // Make the mask's bitwidth at least 8bit and a power-of-2 to avoid
5324 // unnecessary illegal types.
5325 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5326 APInt MaskInt(TableSizePowOf2, 0);
5327 APInt One(TableSizePowOf2, 1);
5328 // Build bitmask; fill in a 1 bit for every case.
5329 const ResultListTy &ResultList = ResultLists[PHIs[0]];
5330 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5331 uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5333 MaskInt |= One << Idx;
5335 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5337 // Get the TableIndex'th bit of the bitmask.
5338 // If this bit is 0 (meaning hole) jump to the default destination,
5339 // else continue with table lookup.
5340 IntegerType *MapTy = TableMask->getType();
5342 Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5343 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5344 Value *LoBit = Builder.CreateTrunc(
5345 Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5346 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5348 Builder.SetInsertPoint(LookupBB);
5349 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5352 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5353 // We cached PHINodes in PHIs, to avoid accessing deleted PHINodes later,
5354 // do not delete PHINodes here.
5355 SI->getDefaultDest()->removePredecessor(SI->getParent(),
5356 /*DontDeleteUselessPHIs=*/true);
5359 bool ReturnedEarly = false;
5360 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
5361 PHINode *PHI = PHIs[I];
5362 const ResultListTy &ResultList = ResultLists[PHI];
5364 // If using a bitmask, use any value to fill the lookup table holes.
5365 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5366 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL);
5368 Value *Result = Table.BuildLookup(TableIndex, Builder);
5370 // If the result is used to return immediately from the function, we want to
5371 // do that right here.
5372 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5373 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5374 Builder.CreateRet(Result);
5375 ReturnedEarly = true;
5379 // Do a small peephole optimization: re-use the switch table compare if
5381 if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5382 BasicBlock *PhiBlock = PHI->getParent();
5383 // Search for compare instructions which use the phi.
5384 for (auto *User : PHI->users()) {
5385 reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5389 PHI->addIncoming(Result, LookupBB);
5393 Builder.CreateBr(CommonDest);
5395 // Remove the switch.
5396 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5397 BasicBlock *Succ = SI->getSuccessor(i);
5399 if (Succ == SI->getDefaultDest())
5401 Succ->removePredecessor(SI->getParent());
5403 SI->eraseFromParent();
5407 ++NumLookupTablesHoles;
5411 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5412 // See also SelectionDAGBuilder::isDense(), which this function was based on.
5413 uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5414 uint64_t Range = Diff + 1;
5415 uint64_t NumCases = Values.size();
5416 // 40% is the default density for building a jump table in optsize/minsize mode.
5417 uint64_t MinDensity = 40;
5419 return NumCases * 100 >= Range * MinDensity;
5422 // Try and transform a switch that has "holes" in it to a contiguous sequence
5425 // A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5426 // range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5428 // This converts a sparse switch into a dense switch which allows better
5429 // lowering and could also allow transforming into a lookup table.
5430 static bool ReduceSwitchRange(SwitchInst *SI, IRBuilder<> &Builder,
5431 const DataLayout &DL,
5432 const TargetTransformInfo &TTI) {
5433 auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5434 if (CondTy->getIntegerBitWidth() > 64 ||
5435 !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5437 // Only bother with this optimization if there are more than 3 switch cases;
5438 // SDAG will only bother creating jump tables for 4 or more cases.
5439 if (SI->getNumCases() < 4)
5442 // This transform is agnostic to the signedness of the input or case values. We
5443 // can treat the case values as signed or unsigned. We can optimize more common
5444 // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5446 SmallVector<int64_t,4> Values;
5447 for (auto &C : SI->cases())
5448 Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5449 std::sort(Values.begin(), Values.end());
5451 // If the switch is already dense, there's nothing useful to do here.
5452 if (isSwitchDense(Values))
5455 // First, transform the values such that they start at zero and ascend.
5456 int64_t Base = Values[0];
5457 for (auto &V : Values)
5460 // Now we have signed numbers that have been shifted so that, given enough
5461 // precision, there are no negative values. Since the rest of the transform
5462 // is bitwise only, we switch now to an unsigned representation.
5464 for (auto &V : Values)
5465 GCD = GreatestCommonDivisor64(GCD, (uint64_t)V);
5467 // This transform can be done speculatively because it is so cheap - it results
5468 // in a single rotate operation being inserted. This can only happen if the
5469 // factor extracted is a power of 2.
5470 // FIXME: If the GCD is an odd number we can multiply by the multiplicative
5471 // inverse of GCD and then perform this transform.
5472 // FIXME: It's possible that optimizing a switch on powers of two might also
5473 // be beneficial - flag values are often powers of two and we could use a CLZ
5474 // as the key function.
5475 if (GCD <= 1 || !isPowerOf2_64(GCD))
5476 // No common divisor found or too expensive to compute key function.
5479 unsigned Shift = Log2_64(GCD);
5480 for (auto &V : Values)
5481 V = (int64_t)((uint64_t)V >> Shift);
5483 if (!isSwitchDense(Values))
5484 // Transform didn't create a dense switch.
5487 // The obvious transform is to shift the switch condition right and emit a
5488 // check that the condition actually cleanly divided by GCD, i.e.
5489 // C & (1 << Shift - 1) == 0
5490 // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5492 // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5493 // shift and puts the shifted-off bits in the uppermost bits. If any of these
5494 // are nonzero then the switch condition will be very large and will hit the
5497 auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5498 Builder.SetInsertPoint(SI);
5499 auto *ShiftC = ConstantInt::get(Ty, Shift);
5500 auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5501 auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5502 auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5503 auto *Rot = Builder.CreateOr(LShr, Shl);
5504 SI->replaceUsesOfWith(SI->getCondition(), Rot);
5506 for (SwitchInst::CaseIt C = SI->case_begin(), E = SI->case_end(); C != E;
5508 auto *Orig = C.getCaseValue();
5509 auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5511 cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5516 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
5517 BasicBlock *BB = SI->getParent();
5519 if (isValueEqualityComparison(SI)) {
5520 // If we only have one predecessor, and if it is a branch on this value,
5521 // see if that predecessor totally determines the outcome of this switch.
5522 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5523 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
5524 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5526 Value *Cond = SI->getCondition();
5527 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
5528 if (SimplifySwitchOnSelect(SI, Select))
5529 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5531 // If the block only contains the switch, see if we can fold the block
5532 // away into any preds.
5533 BasicBlock::iterator BBI = BB->begin();
5534 // Ignore dbg intrinsics.
5535 while (isa<DbgInfoIntrinsic>(BBI))
5538 if (FoldValueComparisonIntoPredecessors(SI, Builder))
5539 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5542 // Try to transform the switch into an icmp and a branch.
5543 if (TurnSwitchRangeIntoICmp(SI, Builder))
5544 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5546 // Remove unreachable cases.
5547 if (EliminateDeadSwitchCases(SI, AC, DL))
5548 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5550 if (SwitchToSelect(SI, Builder, AC, DL, TTI))
5551 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5553 if (ForwardSwitchConditionToPHI(SI))
5554 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5556 if (SwitchToLookupTable(SI, Builder, DL, TTI))
5557 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5559 if (ReduceSwitchRange(SI, Builder, DL, TTI))
5560 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5565 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
5566 BasicBlock *BB = IBI->getParent();
5567 bool Changed = false;
5569 // Eliminate redundant destinations.
5570 SmallPtrSet<Value *, 8> Succs;
5571 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
5572 BasicBlock *Dest = IBI->getDestination(i);
5573 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
5574 Dest->removePredecessor(BB);
5575 IBI->removeDestination(i);
5582 if (IBI->getNumDestinations() == 0) {
5583 // If the indirectbr has no successors, change it to unreachable.
5584 new UnreachableInst(IBI->getContext(), IBI);
5585 EraseTerminatorInstAndDCECond(IBI);
5589 if (IBI->getNumDestinations() == 1) {
5590 // If the indirectbr has one successor, change it to a direct branch.
5591 BranchInst::Create(IBI->getDestination(0), IBI);
5592 EraseTerminatorInstAndDCECond(IBI);
5596 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
5597 if (SimplifyIndirectBrOnSelect(IBI, SI))
5598 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5603 /// Given an block with only a single landing pad and a unconditional branch
5604 /// try to find another basic block which this one can be merged with. This
5605 /// handles cases where we have multiple invokes with unique landing pads, but
5606 /// a shared handler.
5608 /// We specifically choose to not worry about merging non-empty blocks
5609 /// here. That is a PRE/scheduling problem and is best solved elsewhere. In
5610 /// practice, the optimizer produces empty landing pad blocks quite frequently
5611 /// when dealing with exception dense code. (see: instcombine, gvn, if-else
5612 /// sinking in this file)
5614 /// This is primarily a code size optimization. We need to avoid performing
5615 /// any transform which might inhibit optimization (such as our ability to
5616 /// specialize a particular handler via tail commoning). We do this by not
5617 /// merging any blocks which require us to introduce a phi. Since the same
5618 /// values are flowing through both blocks, we don't loose any ability to
5619 /// specialize. If anything, we make such specialization more likely.
5621 /// TODO - This transformation could remove entries from a phi in the target
5622 /// block when the inputs in the phi are the same for the two blocks being
5623 /// merged. In some cases, this could result in removal of the PHI entirely.
5624 static bool TryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI,
5626 auto Succ = BB->getUniqueSuccessor();
5628 // If there's a phi in the successor block, we'd likely have to introduce
5629 // a phi into the merged landing pad block.
5630 if (isa<PHINode>(*Succ->begin()))
5633 for (BasicBlock *OtherPred : predecessors(Succ)) {
5634 if (BB == OtherPred)
5636 BasicBlock::iterator I = OtherPred->begin();
5637 LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(I);
5638 if (!LPad2 || !LPad2->isIdenticalTo(LPad))
5640 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {
5642 BranchInst *BI2 = dyn_cast<BranchInst>(I);
5643 if (!BI2 || !BI2->isIdenticalTo(BI))
5646 // We've found an identical block. Update our predecessors to take that
5647 // path instead and make ourselves dead.
5648 SmallSet<BasicBlock *, 16> Preds;
5649 Preds.insert(pred_begin(BB), pred_end(BB));
5650 for (BasicBlock *Pred : Preds) {
5651 InvokeInst *II = cast<InvokeInst>(Pred->getTerminator());
5652 assert(II->getNormalDest() != BB && II->getUnwindDest() == BB &&
5653 "unexpected successor");
5654 II->setUnwindDest(OtherPred);
5657 // The debug info in OtherPred doesn't cover the merged control flow that
5658 // used to go through BB. We need to delete it or update it.
5659 for (auto I = OtherPred->begin(), E = OtherPred->end(); I != E;) {
5660 Instruction &Inst = *I;
5662 if (isa<DbgInfoIntrinsic>(Inst))
5663 Inst.eraseFromParent();
5666 SmallSet<BasicBlock *, 16> Succs;
5667 Succs.insert(succ_begin(BB), succ_end(BB));
5668 for (BasicBlock *Succ : Succs) {
5669 Succ->removePredecessor(BB);
5672 IRBuilder<> Builder(BI);
5673 Builder.CreateUnreachable();
5674 BI->eraseFromParent();
5680 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI,
5681 IRBuilder<> &Builder) {
5682 BasicBlock *BB = BI->getParent();
5684 if (SinkCommon && SinkThenElseCodeToEnd(BI))
5687 // If the Terminator is the only non-phi instruction, simplify the block.
5688 // if LoopHeader is provided, check if the block is a loop header
5689 // (This is for early invocations before loop simplify and vectorization
5690 // to keep canonical loop forms for nested loops.
5691 // These blocks can be eliminated when the pass is invoked later
5692 // in the back-end.)
5693 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg()->getIterator();
5694 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
5695 (!LoopHeaders || !LoopHeaders->count(BB)) &&
5696 TryToSimplifyUncondBranchFromEmptyBlock(BB))
5699 // If the only instruction in the block is a seteq/setne comparison
5700 // against a constant, try to simplify the block.
5701 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
5702 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
5703 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5705 if (I->isTerminator() &&
5706 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, DL, TTI,
5707 BonusInstThreshold, AC))
5711 // See if we can merge an empty landing pad block with another which is
5713 if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(I)) {
5714 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {
5716 if (I->isTerminator() && TryToMergeLandingPad(LPad, BI, BB))
5720 // If this basic block is ONLY a compare and a branch, and if a predecessor
5721 // branches to us and our successor, fold the comparison into the
5722 // predecessor and use logical operations to update the incoming value
5723 // for PHI nodes in common successor.
5724 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
5725 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5729 static BasicBlock *allPredecessorsComeFromSameSource(BasicBlock *BB) {
5730 BasicBlock *PredPred = nullptr;
5731 for (auto *P : predecessors(BB)) {
5732 BasicBlock *PPred = P->getSinglePredecessor();
5733 if (!PPred || (PredPred && PredPred != PPred))
5740 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
5741 BasicBlock *BB = BI->getParent();
5743 // Conditional branch
5744 if (isValueEqualityComparison(BI)) {
5745 // If we only have one predecessor, and if it is a branch on this value,
5746 // see if that predecessor totally determines the outcome of this
5748 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5749 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
5750 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5752 // This block must be empty, except for the setcond inst, if it exists.
5753 // Ignore dbg intrinsics.
5754 BasicBlock::iterator I = BB->begin();
5755 // Ignore dbg intrinsics.
5756 while (isa<DbgInfoIntrinsic>(I))
5759 if (FoldValueComparisonIntoPredecessors(BI, Builder))
5760 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5761 } else if (&*I == cast<Instruction>(BI->getCondition())) {
5763 // Ignore dbg intrinsics.
5764 while (isa<DbgInfoIntrinsic>(I))
5766 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
5767 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5771 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
5772 if (SimplifyBranchOnICmpChain(BI, Builder, DL))
5775 // If this basic block has a single dominating predecessor block and the
5776 // dominating block's condition implies BI's condition, we know the direction
5777 // of the BI branch.
5778 if (BasicBlock *Dom = BB->getSinglePredecessor()) {
5779 auto *PBI = dyn_cast_or_null<BranchInst>(Dom->getTerminator());
5780 if (PBI && PBI->isConditional() &&
5781 PBI->getSuccessor(0) != PBI->getSuccessor(1) &&
5782 (PBI->getSuccessor(0) == BB || PBI->getSuccessor(1) == BB)) {
5783 bool CondIsFalse = PBI->getSuccessor(1) == BB;
5784 Optional<bool> Implication = isImpliedCondition(
5785 PBI->getCondition(), BI->getCondition(), DL, CondIsFalse);
5787 // Turn this into a branch on constant.
5788 auto *OldCond = BI->getCondition();
5789 ConstantInt *CI = *Implication
5790 ? ConstantInt::getTrue(BB->getContext())
5791 : ConstantInt::getFalse(BB->getContext());
5792 BI->setCondition(CI);
5793 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
5794 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5799 // If this basic block is ONLY a compare and a branch, and if a predecessor
5800 // branches to us and one of our successors, fold the comparison into the
5801 // predecessor and use logical operations to pick the right destination.
5802 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
5803 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5805 // We have a conditional branch to two blocks that are only reachable
5806 // from BI. We know that the condbr dominates the two blocks, so see if
5807 // there is any identical code in the "then" and "else" blocks. If so, we
5808 // can hoist it up to the branching block.
5809 if (BI->getSuccessor(0)->getSinglePredecessor()) {
5810 if (BI->getSuccessor(1)->getSinglePredecessor()) {
5811 if (HoistThenElseCodeToIf(BI, TTI))
5812 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5814 // If Successor #1 has multiple preds, we may be able to conditionally
5815 // execute Successor #0 if it branches to Successor #1.
5816 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
5817 if (Succ0TI->getNumSuccessors() == 1 &&
5818 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
5819 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), TTI))
5820 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5822 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
5823 // If Successor #0 has multiple preds, we may be able to conditionally
5824 // execute Successor #1 if it branches to Successor #0.
5825 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
5826 if (Succ1TI->getNumSuccessors() == 1 &&
5827 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
5828 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), TTI))
5829 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5832 // If this is a branch on a phi node in the current block, thread control
5833 // through this block if any PHI node entries are constants.
5834 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
5835 if (PN->getParent() == BI->getParent())
5836 if (FoldCondBranchOnPHI(BI, DL))
5837 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5839 // Scan predecessor blocks for conditional branches.
5840 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
5841 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
5842 if (PBI != BI && PBI->isConditional())
5843 if (SimplifyCondBranchToCondBranch(PBI, BI, DL))
5844 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5846 // Look for diamond patterns.
5847 if (MergeCondStores)
5848 if (BasicBlock *PrevBB = allPredecessorsComeFromSameSource(BB))
5849 if (BranchInst *PBI = dyn_cast<BranchInst>(PrevBB->getTerminator()))
5850 if (PBI != BI && PBI->isConditional())
5851 if (mergeConditionalStores(PBI, BI))
5852 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5857 /// Check if passing a value to an instruction will cause undefined behavior.
5858 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
5859 Constant *C = dyn_cast<Constant>(V);
5866 if (C->isNullValue() || isa<UndefValue>(C)) {
5867 // Only look at the first use, avoid hurting compile time with long uselists
5868 User *Use = *I->user_begin();
5870 // Now make sure that there are no instructions in between that can alter
5871 // control flow (eg. calls)
5872 for (BasicBlock::iterator
5873 i = ++BasicBlock::iterator(I),
5874 UI = BasicBlock::iterator(dyn_cast<Instruction>(Use));
5876 if (i == I->getParent()->end() || i->mayHaveSideEffects())
5879 // Look through GEPs. A load from a GEP derived from NULL is still undefined
5880 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
5881 if (GEP->getPointerOperand() == I)
5882 return passingValueIsAlwaysUndefined(V, GEP);
5884 // Look through bitcasts.
5885 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
5886 return passingValueIsAlwaysUndefined(V, BC);
5888 // Load from null is undefined.
5889 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
5890 if (!LI->isVolatile())
5891 return LI->getPointerAddressSpace() == 0;
5893 // Store to null is undefined.
5894 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
5895 if (!SI->isVolatile())
5896 return SI->getPointerAddressSpace() == 0 &&
5897 SI->getPointerOperand() == I;
5899 // A call to null is undefined.
5900 if (auto CS = CallSite(Use))
5901 return CS.getCalledValue() == I;
5906 /// If BB has an incoming value that will always trigger undefined behavior
5907 /// (eg. null pointer dereference), remove the branch leading here.
5908 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
5909 for (BasicBlock::iterator i = BB->begin();
5910 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
5911 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
5912 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
5913 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
5914 IRBuilder<> Builder(T);
5915 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
5916 BB->removePredecessor(PHI->getIncomingBlock(i));
5917 // Turn uncoditional branches into unreachables and remove the dead
5918 // destination from conditional branches.
5919 if (BI->isUnconditional())
5920 Builder.CreateUnreachable();
5922 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1)
5923 : BI->getSuccessor(0));
5924 BI->eraseFromParent();
5927 // TODO: SwitchInst.
5933 bool SimplifyCFGOpt::run(BasicBlock *BB) {
5934 bool Changed = false;
5936 assert(BB && BB->getParent() && "Block not embedded in function!");
5937 assert(BB->getTerminator() && "Degenerate basic block encountered!");
5939 // Remove basic blocks that have no predecessors (except the entry block)...
5940 // or that just have themself as a predecessor. These are unreachable.
5941 if ((pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) ||
5942 BB->getSinglePredecessor() == BB) {
5943 DEBUG(dbgs() << "Removing BB: \n" << *BB);
5944 DeleteDeadBlock(BB);
5948 // Check to see if we can constant propagate this terminator instruction
5950 Changed |= ConstantFoldTerminator(BB, true);
5952 // Check for and eliminate duplicate PHI nodes in this block.
5953 Changed |= EliminateDuplicatePHINodes(BB);
5955 // Check for and remove branches that will always cause undefined behavior.
5956 Changed |= removeUndefIntroducingPredecessor(BB);
5958 // Merge basic blocks into their predecessor if there is only one distinct
5959 // pred, and if there is only one distinct successor of the predecessor, and
5960 // if there are no PHI nodes.
5962 if (MergeBlockIntoPredecessor(BB))
5965 IRBuilder<> Builder(BB);
5967 // If there is a trivial two-entry PHI node in this basic block, and we can
5968 // eliminate it, do so now.
5969 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
5970 if (PN->getNumIncomingValues() == 2)
5971 Changed |= FoldTwoEntryPHINode(PN, TTI, DL);
5973 Builder.SetInsertPoint(BB->getTerminator());
5974 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
5975 if (BI->isUnconditional()) {
5976 if (SimplifyUncondBranch(BI, Builder))
5979 if (SimplifyCondBranch(BI, Builder))
5982 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
5983 if (SimplifyReturn(RI, Builder))
5985 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
5986 if (SimplifyResume(RI, Builder))
5988 } else if (CleanupReturnInst *RI =
5989 dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
5990 if (SimplifyCleanupReturn(RI))
5992 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
5993 if (SimplifySwitch(SI, Builder))
5995 } else if (UnreachableInst *UI =
5996 dyn_cast<UnreachableInst>(BB->getTerminator())) {
5997 if (SimplifyUnreachable(UI))
5999 } else if (IndirectBrInst *IBI =
6000 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
6001 if (SimplifyIndirectBr(IBI))
6008 /// This function is used to do simplification of a CFG.
6009 /// For example, it adjusts branches to branches to eliminate the extra hop,
6010 /// eliminates unreachable basic blocks, and does other "peephole" optimization
6011 /// of the CFG. It returns true if a modification was made.
6013 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
6014 unsigned BonusInstThreshold, AssumptionCache *AC,
6015 SmallPtrSetImpl<BasicBlock *> *LoopHeaders) {
6016 return SimplifyCFGOpt(TTI, BB->getModule()->getDataLayout(),
6017 BonusInstThreshold, AC, LoopHeaders)