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/STLExtras.h"
19 #include "llvm/ADT/SetOperations.h"
20 #include "llvm/ADT/SetVector.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/SmallSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/StringRef.h"
26 #include "llvm/Analysis/AssumptionCache.h"
27 #include "llvm/Analysis/ConstantFolding.h"
28 #include "llvm/Analysis/EHPersonalities.h"
29 #include "llvm/Analysis/InstructionSimplify.h"
30 #include "llvm/Analysis/TargetTransformInfo.h"
31 #include "llvm/Analysis/ValueTracking.h"
32 #include "llvm/IR/Attributes.h"
33 #include "llvm/IR/BasicBlock.h"
34 #include "llvm/IR/CFG.h"
35 #include "llvm/IR/CallSite.h"
36 #include "llvm/IR/Constant.h"
37 #include "llvm/IR/ConstantRange.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/DerivedTypes.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/GlobalValue.h"
43 #include "llvm/IR/GlobalVariable.h"
44 #include "llvm/IR/IRBuilder.h"
45 #include "llvm/IR/InstrTypes.h"
46 #include "llvm/IR/Instruction.h"
47 #include "llvm/IR/Instructions.h"
48 #include "llvm/IR/IntrinsicInst.h"
49 #include "llvm/IR/Intrinsics.h"
50 #include "llvm/IR/LLVMContext.h"
51 #include "llvm/IR/MDBuilder.h"
52 #include "llvm/IR/Metadata.h"
53 #include "llvm/IR/Module.h"
54 #include "llvm/IR/NoFolder.h"
55 #include "llvm/IR/Operator.h"
56 #include "llvm/IR/PatternMatch.h"
57 #include "llvm/IR/Type.h"
58 #include "llvm/IR/Use.h"
59 #include "llvm/IR/User.h"
60 #include "llvm/IR/Value.h"
61 #include "llvm/Support/Casting.h"
62 #include "llvm/Support/CommandLine.h"
63 #include "llvm/Support/Debug.h"
64 #include "llvm/Support/ErrorHandling.h"
65 #include "llvm/Support/KnownBits.h"
66 #include "llvm/Support/MathExtras.h"
67 #include "llvm/Support/raw_ostream.h"
68 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
69 #include "llvm/Transforms/Utils/Local.h"
70 #include "llvm/Transforms/Utils/ValueMapper.h"
84 using namespace PatternMatch;
86 #define DEBUG_TYPE "simplifycfg"
88 // Chosen as 2 so as to be cheap, but still to have enough power to fold
89 // a select, so the "clamp" idiom (of a min followed by a max) will be caught.
90 // To catch this, we need to fold a compare and a select, hence '2' being the
91 // minimum reasonable default.
92 static cl::opt<unsigned> PHINodeFoldingThreshold(
93 "phi-node-folding-threshold", cl::Hidden, cl::init(2),
95 "Control the amount of phi node folding to perform (default = 2)"));
97 static cl::opt<bool> DupRet(
98 "simplifycfg-dup-ret", cl::Hidden, cl::init(false),
99 cl::desc("Duplicate return instructions into unconditional branches"));
102 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
103 cl::desc("Sink common instructions down to the end block"));
105 static cl::opt<bool> HoistCondStores(
106 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
107 cl::desc("Hoist conditional stores if an unconditional store precedes"));
109 static cl::opt<bool> MergeCondStores(
110 "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
111 cl::desc("Hoist conditional stores even if an unconditional store does not "
112 "precede - hoist multiple conditional stores into a single "
113 "predicated store"));
115 static cl::opt<bool> MergeCondStoresAggressively(
116 "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
117 cl::desc("When merging conditional stores, do so even if the resultant "
118 "basic blocks are unlikely to be if-converted as a result"));
120 static cl::opt<bool> SpeculateOneExpensiveInst(
121 "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
122 cl::desc("Allow exactly one expensive instruction to be speculatively "
125 static cl::opt<unsigned> MaxSpeculationDepth(
126 "max-speculation-depth", cl::Hidden, cl::init(10),
127 cl::desc("Limit maximum recursion depth when calculating costs of "
128 "speculatively executed instructions"));
130 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
131 STATISTIC(NumLinearMaps,
132 "Number of switch instructions turned into linear mapping");
133 STATISTIC(NumLookupTables,
134 "Number of switch instructions turned into lookup tables");
136 NumLookupTablesHoles,
137 "Number of switch instructions turned into lookup tables (holes checked)");
138 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
139 STATISTIC(NumSinkCommons,
140 "Number of common instructions sunk down to the end block");
141 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
145 // The first field contains the value that the switch produces when a certain
146 // case group is selected, and the second field is a vector containing the
147 // cases composing the case group.
148 using SwitchCaseResultVectorTy =
149 SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>;
151 // The first field contains the phi node that generates a result of the switch
152 // and the second field contains the value generated for a certain case in the
153 // switch for that PHI.
154 using SwitchCaseResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
156 /// ValueEqualityComparisonCase - Represents a case of a switch.
157 struct ValueEqualityComparisonCase {
161 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
162 : Value(Value), Dest(Dest) {}
164 bool operator<(ValueEqualityComparisonCase RHS) const {
165 // Comparing pointers is ok as we only rely on the order for uniquing.
166 return Value < RHS.Value;
169 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
172 class SimplifyCFGOpt {
173 const TargetTransformInfo &TTI;
174 const DataLayout &DL;
175 SmallPtrSetImpl<BasicBlock *> *LoopHeaders;
176 const SimplifyCFGOptions &Options;
178 Value *isValueEqualityComparison(TerminatorInst *TI);
179 BasicBlock *GetValueEqualityComparisonCases(
180 TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases);
181 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
183 IRBuilder<> &Builder);
184 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
185 IRBuilder<> &Builder);
187 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
188 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
189 bool SimplifySingleResume(ResumeInst *RI);
190 bool SimplifyCommonResume(ResumeInst *RI);
191 bool SimplifyCleanupReturn(CleanupReturnInst *RI);
192 bool SimplifyUnreachable(UnreachableInst *UI);
193 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
194 bool SimplifyIndirectBr(IndirectBrInst *IBI);
195 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
196 bool SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
199 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
200 SmallPtrSetImpl<BasicBlock *> *LoopHeaders,
201 const SimplifyCFGOptions &Opts)
202 : TTI(TTI), DL(DL), LoopHeaders(LoopHeaders), Options(Opts) {}
204 bool run(BasicBlock *BB);
207 } // end anonymous namespace
209 /// Return true if it is safe to merge these two
210 /// terminator instructions together.
212 SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2,
213 SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
215 return false; // Can't merge with self!
217 // It is not safe to merge these two switch instructions if they have a common
218 // successor, and if that successor has a PHI node, and if *that* PHI node has
219 // conflicting incoming values from the two switch blocks.
220 BasicBlock *SI1BB = SI1->getParent();
221 BasicBlock *SI2BB = SI2->getParent();
223 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
225 for (BasicBlock *Succ : successors(SI2BB))
226 if (SI1Succs.count(Succ))
227 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
228 PHINode *PN = cast<PHINode>(BBI);
229 if (PN->getIncomingValueForBlock(SI1BB) !=
230 PN->getIncomingValueForBlock(SI2BB)) {
232 FailBlocks->insert(Succ);
240 /// Return true if it is safe and profitable to merge these two terminator
241 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
242 /// store all PHI nodes in common successors.
244 isProfitableToFoldUnconditional(BranchInst *SI1, BranchInst *SI2,
246 SmallVectorImpl<PHINode *> &PhiNodes) {
248 return false; // Can't merge with self!
249 assert(SI1->isUnconditional() && SI2->isConditional());
251 // We fold the unconditional branch if we can easily update all PHI nodes in
252 // common successors:
253 // 1> We have a constant incoming value for the conditional branch;
254 // 2> We have "Cond" as the incoming value for the unconditional branch;
255 // 3> SI2->getCondition() and Cond have same operands.
256 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
259 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
260 Cond->getOperand(1) == Ci2->getOperand(1)) &&
261 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
262 Cond->getOperand(1) == Ci2->getOperand(0)))
265 BasicBlock *SI1BB = SI1->getParent();
266 BasicBlock *SI2BB = SI2->getParent();
267 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
268 for (BasicBlock *Succ : successors(SI2BB))
269 if (SI1Succs.count(Succ))
270 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
271 PHINode *PN = cast<PHINode>(BBI);
272 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
273 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
275 PhiNodes.push_back(PN);
280 /// Update PHI nodes in Succ to indicate that there will now be entries in it
281 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
282 /// will be the same as those coming in from ExistPred, an existing predecessor
284 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
285 BasicBlock *ExistPred) {
286 for (PHINode &PN : Succ->phis())
287 PN.addIncoming(PN.getIncomingValueForBlock(ExistPred), NewPred);
290 /// Compute an abstract "cost" of speculating the given instruction,
291 /// which is assumed to be safe to speculate. TCC_Free means cheap,
292 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
294 static unsigned ComputeSpeculationCost(const User *I,
295 const TargetTransformInfo &TTI) {
296 assert(isSafeToSpeculativelyExecute(I) &&
297 "Instruction is not safe to speculatively execute!");
298 return TTI.getUserCost(I);
301 /// If we have a merge point of an "if condition" as accepted above,
302 /// return true if the specified value dominates the block. We
303 /// don't handle the true generality of domination here, just a special case
304 /// which works well enough for us.
306 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
307 /// see if V (which must be an instruction) and its recursive operands
308 /// that do not dominate BB have a combined cost lower than CostRemaining and
309 /// are non-trapping. If both are true, the instruction is inserted into the
310 /// set and true is returned.
312 /// The cost for most non-trapping instructions is defined as 1 except for
313 /// Select whose cost is 2.
315 /// After this function returns, CostRemaining is decreased by the cost of
316 /// V plus its non-dominating operands. If that cost is greater than
317 /// CostRemaining, false is returned and CostRemaining is undefined.
318 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
319 SmallPtrSetImpl<Instruction *> *AggressiveInsts,
320 unsigned &CostRemaining,
321 const TargetTransformInfo &TTI,
322 unsigned Depth = 0) {
323 // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
324 // so limit the recursion depth.
325 // TODO: While this recursion limit does prevent pathological behavior, it
326 // would be better to track visited instructions to avoid cycles.
327 if (Depth == MaxSpeculationDepth)
330 Instruction *I = dyn_cast<Instruction>(V);
332 // Non-instructions all dominate instructions, but not all constantexprs
333 // can be executed unconditionally.
334 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
339 BasicBlock *PBB = I->getParent();
341 // We don't want to allow weird loops that might have the "if condition" in
342 // the bottom of this block.
346 // If this instruction is defined in a block that contains an unconditional
347 // branch to BB, then it must be in the 'conditional' part of the "if
348 // statement". If not, it definitely dominates the region.
349 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
350 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
353 // If we aren't allowing aggressive promotion anymore, then don't consider
354 // instructions in the 'if region'.
355 if (!AggressiveInsts)
358 // If we have seen this instruction before, don't count it again.
359 if (AggressiveInsts->count(I))
362 // Okay, it looks like the instruction IS in the "condition". Check to
363 // see if it's a cheap instruction to unconditionally compute, and if it
364 // only uses stuff defined outside of the condition. If so, hoist it out.
365 if (!isSafeToSpeculativelyExecute(I))
368 unsigned Cost = ComputeSpeculationCost(I, TTI);
370 // Allow exactly one instruction to be speculated regardless of its cost
371 // (as long as it is safe to do so).
372 // This is intended to flatten the CFG even if the instruction is a division
373 // or other expensive operation. The speculation of an expensive instruction
374 // is expected to be undone in CodeGenPrepare if the speculation has not
375 // enabled further IR optimizations.
376 if (Cost > CostRemaining &&
377 (!SpeculateOneExpensiveInst || !AggressiveInsts->empty() || Depth > 0))
380 // Avoid unsigned wrap.
381 CostRemaining = (Cost > CostRemaining) ? 0 : CostRemaining - Cost;
383 // Okay, we can only really hoist these out if their operands do
384 // not take us over the cost threshold.
385 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
386 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, TTI,
389 // Okay, it's safe to do this! Remember this instruction.
390 AggressiveInsts->insert(I);
394 /// Extract ConstantInt from value, looking through IntToPtr
395 /// and PointerNullValue. Return NULL if value is not a constant int.
396 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
397 // Normal constant int.
398 ConstantInt *CI = dyn_cast<ConstantInt>(V);
399 if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
402 // This is some kind of pointer constant. Turn it into a pointer-sized
403 // ConstantInt if possible.
404 IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
406 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
407 if (isa<ConstantPointerNull>(V))
408 return ConstantInt::get(PtrTy, 0);
410 // IntToPtr const int.
411 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
412 if (CE->getOpcode() == Instruction::IntToPtr)
413 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
414 // The constant is very likely to have the right type already.
415 if (CI->getType() == PtrTy)
418 return cast<ConstantInt>(
419 ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
426 /// Given a chain of or (||) or and (&&) comparison of a value against a
427 /// constant, this will try to recover the information required for a switch
429 /// It will depth-first traverse the chain of comparison, seeking for patterns
430 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
431 /// representing the different cases for the switch.
432 /// Note that if the chain is composed of '||' it will build the set of elements
433 /// that matches the comparisons (i.e. any of this value validate the chain)
434 /// while for a chain of '&&' it will build the set elements that make the test
436 struct ConstantComparesGatherer {
437 const DataLayout &DL;
439 /// Value found for the switch comparison
440 Value *CompValue = nullptr;
442 /// Extra clause to be checked before the switch
443 Value *Extra = nullptr;
445 /// Set of integers to match in switch
446 SmallVector<ConstantInt *, 8> Vals;
448 /// Number of comparisons matched in the and/or chain
449 unsigned UsedICmps = 0;
451 /// Construct and compute the result for the comparison instruction Cond
452 ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL) : DL(DL) {
456 ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
457 ConstantComparesGatherer &
458 operator=(const ConstantComparesGatherer &) = delete;
461 /// Try to set the current value used for the comparison, it succeeds only if
462 /// it wasn't set before or if the new value is the same as the old one
463 bool setValueOnce(Value *NewVal) {
464 if (CompValue && CompValue != NewVal)
467 return (CompValue != nullptr);
470 /// Try to match Instruction "I" as a comparison against a constant and
471 /// populates the array Vals with the set of values that match (or do not
472 /// match depending on isEQ).
473 /// Return false on failure. On success, the Value the comparison matched
474 /// against is placed in CompValue.
475 /// If CompValue is already set, the function is expected to fail if a match
476 /// is found but the value compared to is different.
477 bool matchInstruction(Instruction *I, bool isEQ) {
478 // If this is an icmp against a constant, handle this as one of the cases.
481 if (!((ICI = dyn_cast<ICmpInst>(I)) &&
482 (C = GetConstantInt(I->getOperand(1), DL)))) {
489 // Pattern match a special case
490 // (x & ~2^z) == y --> x == y || x == y|2^z
491 // This undoes a transformation done by instcombine to fuse 2 compares.
492 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
493 // It's a little bit hard to see why the following transformations are
494 // correct. Here is a CVC3 program to verify them for 64-bit values:
497 ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
501 mask : BITVECTOR(64) = BVSHL(ONE, z);
502 QUERY( (y & ~mask = y) =>
503 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
505 QUERY( (y | mask = y) =>
506 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
510 // Please note that each pattern must be a dual implication (<--> or
511 // iff). One directional implication can create spurious matches. If the
512 // implication is only one-way, an unsatisfiable condition on the left
513 // side can imply a satisfiable condition on the right side. Dual
514 // implication ensures that satisfiable conditions are transformed to
515 // other satisfiable conditions and unsatisfiable conditions are
516 // transformed to other unsatisfiable conditions.
518 // Here is a concrete example of a unsatisfiable condition on the left
519 // implying a satisfiable condition on the right:
522 // (x & ~mask) == y --> (x == y || x == (y | mask))
524 // Substituting y = 3, z = 0 yields:
525 // (x & -2) == 3 --> (x == 3 || x == 2)
527 // Pattern match a special case:
529 QUERY( (y & ~mask = y) =>
530 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
533 if (match(ICI->getOperand(0),
534 m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
536 if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
537 // If we already have a value for the switch, it has to match!
538 if (!setValueOnce(RHSVal))
543 ConstantInt::get(C->getContext(),
544 C->getValue() | Mask));
550 // Pattern match a special case:
552 QUERY( (y | mask = y) =>
553 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
556 if (match(ICI->getOperand(0),
557 m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
559 if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
560 // If we already have a value for the switch, it has to match!
561 if (!setValueOnce(RHSVal))
565 Vals.push_back(ConstantInt::get(C->getContext(),
566 C->getValue() & ~Mask));
572 // If we already have a value for the switch, it has to match!
573 if (!setValueOnce(ICI->getOperand(0)))
578 return ICI->getOperand(0);
581 // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
582 ConstantRange Span = ConstantRange::makeAllowedICmpRegion(
583 ICI->getPredicate(), C->getValue());
585 // Shift the range if the compare is fed by an add. This is the range
586 // compare idiom as emitted by instcombine.
587 Value *CandidateVal = I->getOperand(0);
588 if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
589 Span = Span.subtract(*RHSC);
590 CandidateVal = RHSVal;
593 // If this is an and/!= check, then we are looking to build the set of
594 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
597 Span = Span.inverse();
599 // If there are a ton of values, we don't want to make a ginormous switch.
600 if (Span.isSizeLargerThan(8) || Span.isEmptySet()) {
604 // If we already have a value for the switch, it has to match!
605 if (!setValueOnce(CandidateVal))
608 // Add all values from the range to the set
609 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
610 Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
616 /// Given a potentially 'or'd or 'and'd together collection of icmp
617 /// eq/ne/lt/gt instructions that compare a value against a constant, extract
618 /// the value being compared, and stick the list constants into the Vals
620 /// One "Extra" case is allowed to differ from the other.
621 void gather(Value *V) {
622 Instruction *I = dyn_cast<Instruction>(V);
623 bool isEQ = (I->getOpcode() == Instruction::Or);
625 // Keep a stack (SmallVector for efficiency) for depth-first traversal
626 SmallVector<Value *, 8> DFT;
627 SmallPtrSet<Value *, 8> Visited;
633 while (!DFT.empty()) {
634 V = DFT.pop_back_val();
636 if (Instruction *I = dyn_cast<Instruction>(V)) {
637 // If it is a || (or && depending on isEQ), process the operands.
638 if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
639 if (Visited.insert(I->getOperand(1)).second)
640 DFT.push_back(I->getOperand(1));
641 if (Visited.insert(I->getOperand(0)).second)
642 DFT.push_back(I->getOperand(0));
646 // Try to match the current instruction
647 if (matchInstruction(I, isEQ))
648 // Match succeed, continue the loop
652 // One element of the sequence of || (or &&) could not be match as a
653 // comparison against the same value as the others.
654 // We allow only one "Extra" case to be checked before the switch
659 // Failed to parse a proper sequence, abort now
666 } // end anonymous namespace
668 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
669 Instruction *Cond = nullptr;
670 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
671 Cond = dyn_cast<Instruction>(SI->getCondition());
672 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
673 if (BI->isConditional())
674 Cond = dyn_cast<Instruction>(BI->getCondition());
675 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
676 Cond = dyn_cast<Instruction>(IBI->getAddress());
679 TI->eraseFromParent();
681 RecursivelyDeleteTriviallyDeadInstructions(Cond);
684 /// Return true if the specified terminator checks
685 /// to see if a value is equal to constant integer value.
686 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
688 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
689 // Do not permit merging of large switch instructions into their
690 // predecessors unless there is only one predecessor.
691 if (SI->getNumSuccessors() * std::distance(pred_begin(SI->getParent()),
692 pred_end(SI->getParent())) <=
694 CV = SI->getCondition();
695 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
696 if (BI->isConditional() && BI->getCondition()->hasOneUse())
697 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
698 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
699 CV = ICI->getOperand(0);
702 // Unwrap any lossless ptrtoint cast.
704 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
705 Value *Ptr = PTII->getPointerOperand();
706 if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
713 /// Given a value comparison instruction,
714 /// decode all of the 'cases' that it represents and return the 'default' block.
715 BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
716 TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
717 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
718 Cases.reserve(SI->getNumCases());
719 for (auto Case : SI->cases())
720 Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),
721 Case.getCaseSuccessor()));
722 return SI->getDefaultDest();
725 BranchInst *BI = cast<BranchInst>(TI);
726 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
727 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
728 Cases.push_back(ValueEqualityComparisonCase(
729 GetConstantInt(ICI->getOperand(1), DL), Succ));
730 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
733 /// Given a vector of bb/value pairs, remove any entries
734 /// in the list that match the specified block.
736 EliminateBlockCases(BasicBlock *BB,
737 std::vector<ValueEqualityComparisonCase> &Cases) {
738 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
741 /// Return true if there are any keys in C1 that exist in C2 as well.
742 static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
743 std::vector<ValueEqualityComparisonCase> &C2) {
744 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
746 // Make V1 be smaller than V2.
747 if (V1->size() > V2->size())
752 if (V1->size() == 1) {
754 ConstantInt *TheVal = (*V1)[0].Value;
755 for (unsigned i = 0, e = V2->size(); i != e; ++i)
756 if (TheVal == (*V2)[i].Value)
760 // Otherwise, just sort both lists and compare element by element.
761 array_pod_sort(V1->begin(), V1->end());
762 array_pod_sort(V2->begin(), V2->end());
763 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
764 while (i1 != e1 && i2 != e2) {
765 if ((*V1)[i1].Value == (*V2)[i2].Value)
767 if ((*V1)[i1].Value < (*V2)[i2].Value)
775 // Set branch weights on SwitchInst. This sets the metadata if there is at
776 // least one non-zero weight.
777 static void setBranchWeights(SwitchInst *SI, ArrayRef<uint32_t> Weights) {
778 // Check that there is at least one non-zero weight. Otherwise, pass
779 // nullptr to setMetadata which will erase the existing metadata.
781 if (llvm::any_of(Weights, [](uint32_t W) { return W != 0; }))
782 N = MDBuilder(SI->getParent()->getContext()).createBranchWeights(Weights);
783 SI->setMetadata(LLVMContext::MD_prof, N);
786 // Similar to the above, but for branch and select instructions that take
787 // exactly 2 weights.
788 static void setBranchWeights(Instruction *I, uint32_t TrueWeight,
789 uint32_t FalseWeight) {
790 assert(isa<BranchInst>(I) || isa<SelectInst>(I));
791 // Check that there is at least one non-zero weight. Otherwise, pass
792 // nullptr to setMetadata which will erase the existing metadata.
794 if (TrueWeight || FalseWeight)
795 N = MDBuilder(I->getParent()->getContext())
796 .createBranchWeights(TrueWeight, FalseWeight);
797 I->setMetadata(LLVMContext::MD_prof, N);
800 /// If TI is known to be a terminator instruction and its block is known to
801 /// only have a single predecessor block, check to see if that predecessor is
802 /// also a value comparison with the same value, and if that comparison
803 /// determines the outcome of this comparison. If so, simplify TI. This does a
804 /// very limited form of jump threading.
805 bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
806 TerminatorInst *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
807 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
809 return false; // Not a value comparison in predecessor.
811 Value *ThisVal = isValueEqualityComparison(TI);
812 assert(ThisVal && "This isn't a value comparison!!");
813 if (ThisVal != PredVal)
814 return false; // Different predicates.
816 // TODO: Preserve branch weight metadata, similarly to how
817 // FoldValueComparisonIntoPredecessors preserves it.
819 // Find out information about when control will move from Pred to TI's block.
820 std::vector<ValueEqualityComparisonCase> PredCases;
821 BasicBlock *PredDef =
822 GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
823 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
825 // Find information about how control leaves this block.
826 std::vector<ValueEqualityComparisonCase> ThisCases;
827 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
828 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
830 // If TI's block is the default block from Pred's comparison, potentially
831 // simplify TI based on this knowledge.
832 if (PredDef == TI->getParent()) {
833 // If we are here, we know that the value is none of those cases listed in
834 // PredCases. If there are any cases in ThisCases that are in PredCases, we
836 if (!ValuesOverlap(PredCases, ThisCases))
839 if (isa<BranchInst>(TI)) {
840 // Okay, one of the successors of this condbr is dead. Convert it to a
842 assert(ThisCases.size() == 1 && "Branch can only have one case!");
843 // Insert the new branch.
844 Instruction *NI = Builder.CreateBr(ThisDef);
847 // Remove PHI node entries for the dead edge.
848 ThisCases[0].Dest->removePredecessor(TI->getParent());
850 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
851 << "Through successor TI: " << *TI << "Leaving: " << *NI
854 EraseTerminatorInstAndDCECond(TI);
858 SwitchInst *SI = cast<SwitchInst>(TI);
859 // Okay, TI has cases that are statically dead, prune them away.
860 SmallPtrSet<Constant *, 16> DeadCases;
861 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
862 DeadCases.insert(PredCases[i].Value);
864 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
865 << "Through successor TI: " << *TI);
867 // Collect branch weights into a vector.
868 SmallVector<uint32_t, 8> Weights;
869 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
870 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
872 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
874 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
875 Weights.push_back(CI->getValue().getZExtValue());
877 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
879 if (DeadCases.count(i->getCaseValue())) {
881 std::swap(Weights[i->getCaseIndex() + 1], Weights.back());
884 i->getCaseSuccessor()->removePredecessor(TI->getParent());
888 if (HasWeight && Weights.size() >= 2)
889 setBranchWeights(SI, Weights);
891 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
895 // Otherwise, TI's block must correspond to some matched value. Find out
896 // which value (or set of values) this is.
897 ConstantInt *TIV = nullptr;
898 BasicBlock *TIBB = TI->getParent();
899 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
900 if (PredCases[i].Dest == TIBB) {
902 return false; // Cannot handle multiple values coming to this block.
903 TIV = PredCases[i].Value;
905 assert(TIV && "No edge from pred to succ?");
907 // Okay, we found the one constant that our value can be if we get into TI's
908 // BB. Find out which successor will unconditionally be branched to.
909 BasicBlock *TheRealDest = nullptr;
910 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
911 if (ThisCases[i].Value == TIV) {
912 TheRealDest = ThisCases[i].Dest;
916 // If not handled by any explicit cases, it is handled by the default case.
918 TheRealDest = ThisDef;
920 // Remove PHI node entries for dead edges.
921 BasicBlock *CheckEdge = TheRealDest;
922 for (BasicBlock *Succ : successors(TIBB))
923 if (Succ != CheckEdge)
924 Succ->removePredecessor(TIBB);
928 // Insert the new branch.
929 Instruction *NI = Builder.CreateBr(TheRealDest);
932 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
933 << "Through successor TI: " << *TI << "Leaving: " << *NI
936 EraseTerminatorInstAndDCECond(TI);
942 /// This class implements a stable ordering of constant
943 /// integers that does not depend on their address. This is important for
944 /// applications that sort ConstantInt's to ensure uniqueness.
945 struct ConstantIntOrdering {
946 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
947 return LHS->getValue().ult(RHS->getValue());
951 } // end anonymous namespace
953 static int ConstantIntSortPredicate(ConstantInt *const *P1,
954 ConstantInt *const *P2) {
955 const ConstantInt *LHS = *P1;
956 const ConstantInt *RHS = *P2;
959 return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
962 static inline bool HasBranchWeights(const Instruction *I) {
963 MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
964 if (ProfMD && ProfMD->getOperand(0))
965 if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
966 return MDS->getString().equals("branch_weights");
971 /// Get Weights of a given TerminatorInst, the default weight is at the front
972 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
974 static void GetBranchWeights(TerminatorInst *TI,
975 SmallVectorImpl<uint64_t> &Weights) {
976 MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
978 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
979 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
980 Weights.push_back(CI->getValue().getZExtValue());
983 // If TI is a conditional eq, the default case is the false case,
984 // and the corresponding branch-weight data is at index 2. We swap the
985 // default weight to be the first entry.
986 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
987 assert(Weights.size() == 2);
988 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
989 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
990 std::swap(Weights.front(), Weights.back());
994 /// Keep halving the weights until all can fit in uint32_t.
995 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
996 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
997 if (Max > UINT_MAX) {
998 unsigned Offset = 32 - countLeadingZeros(Max);
999 for (uint64_t &I : Weights)
1004 /// The specified terminator is a value equality comparison instruction
1005 /// (either a switch or a branch on "X == c").
1006 /// See if any of the predecessors of the terminator block are value comparisons
1007 /// on the same value. If so, and if safe to do so, fold them together.
1008 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
1009 IRBuilder<> &Builder) {
1010 BasicBlock *BB = TI->getParent();
1011 Value *CV = isValueEqualityComparison(TI); // CondVal
1012 assert(CV && "Not a comparison?");
1013 bool Changed = false;
1015 SmallVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
1016 while (!Preds.empty()) {
1017 BasicBlock *Pred = Preds.pop_back_val();
1019 // See if the predecessor is a comparison with the same value.
1020 TerminatorInst *PTI = Pred->getTerminator();
1021 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
1023 if (PCV == CV && TI != PTI) {
1024 SmallSetVector<BasicBlock*, 4> FailBlocks;
1025 if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
1026 for (auto *Succ : FailBlocks) {
1027 if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split"))
1032 // Figure out which 'cases' to copy from SI to PSI.
1033 std::vector<ValueEqualityComparisonCase> BBCases;
1034 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1036 std::vector<ValueEqualityComparisonCase> PredCases;
1037 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1039 // Based on whether the default edge from PTI goes to BB or not, fill in
1040 // PredCases and PredDefault with the new switch cases we would like to
1042 SmallVector<BasicBlock *, 8> NewSuccessors;
1044 // Update the branch weight metadata along the way
1045 SmallVector<uint64_t, 8> Weights;
1046 bool PredHasWeights = HasBranchWeights(PTI);
1047 bool SuccHasWeights = HasBranchWeights(TI);
1049 if (PredHasWeights) {
1050 GetBranchWeights(PTI, Weights);
1051 // branch-weight metadata is inconsistent here.
1052 if (Weights.size() != 1 + PredCases.size())
1053 PredHasWeights = SuccHasWeights = false;
1054 } else if (SuccHasWeights)
1055 // If there are no predecessor weights but there are successor weights,
1056 // populate Weights with 1, which will later be scaled to the sum of
1057 // successor's weights
1058 Weights.assign(1 + PredCases.size(), 1);
1060 SmallVector<uint64_t, 8> SuccWeights;
1061 if (SuccHasWeights) {
1062 GetBranchWeights(TI, SuccWeights);
1063 // branch-weight metadata is inconsistent here.
1064 if (SuccWeights.size() != 1 + BBCases.size())
1065 PredHasWeights = SuccHasWeights = false;
1066 } else if (PredHasWeights)
1067 SuccWeights.assign(1 + BBCases.size(), 1);
1069 if (PredDefault == BB) {
1070 // If this is the default destination from PTI, only the edges in TI
1071 // that don't occur in PTI, or that branch to BB will be activated.
1072 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1073 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1074 if (PredCases[i].Dest != BB)
1075 PTIHandled.insert(PredCases[i].Value);
1077 // The default destination is BB, we don't need explicit targets.
1078 std::swap(PredCases[i], PredCases.back());
1080 if (PredHasWeights || SuccHasWeights) {
1081 // Increase weight for the default case.
1082 Weights[0] += Weights[i + 1];
1083 std::swap(Weights[i + 1], Weights.back());
1087 PredCases.pop_back();
1092 // Reconstruct the new switch statement we will be building.
1093 if (PredDefault != BBDefault) {
1094 PredDefault->removePredecessor(Pred);
1095 PredDefault = BBDefault;
1096 NewSuccessors.push_back(BBDefault);
1099 unsigned CasesFromPred = Weights.size();
1100 uint64_t ValidTotalSuccWeight = 0;
1101 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1102 if (!PTIHandled.count(BBCases[i].Value) &&
1103 BBCases[i].Dest != BBDefault) {
1104 PredCases.push_back(BBCases[i]);
1105 NewSuccessors.push_back(BBCases[i].Dest);
1106 if (SuccHasWeights || PredHasWeights) {
1107 // The default weight is at index 0, so weight for the ith case
1108 // should be at index i+1. Scale the cases from successor by
1109 // PredDefaultWeight (Weights[0]).
1110 Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1111 ValidTotalSuccWeight += SuccWeights[i + 1];
1115 if (SuccHasWeights || PredHasWeights) {
1116 ValidTotalSuccWeight += SuccWeights[0];
1117 // Scale the cases from predecessor by ValidTotalSuccWeight.
1118 for (unsigned i = 1; i < CasesFromPred; ++i)
1119 Weights[i] *= ValidTotalSuccWeight;
1120 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1121 Weights[0] *= SuccWeights[0];
1124 // If this is not the default destination from PSI, only the edges
1125 // in SI that occur in PSI with a destination of BB will be
1127 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1128 std::map<ConstantInt *, uint64_t> WeightsForHandled;
1129 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1130 if (PredCases[i].Dest == BB) {
1131 PTIHandled.insert(PredCases[i].Value);
1133 if (PredHasWeights || SuccHasWeights) {
1134 WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1135 std::swap(Weights[i + 1], Weights.back());
1139 std::swap(PredCases[i], PredCases.back());
1140 PredCases.pop_back();
1145 // Okay, now we know which constants were sent to BB from the
1146 // predecessor. Figure out where they will all go now.
1147 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1148 if (PTIHandled.count(BBCases[i].Value)) {
1149 // If this is one we are capable of getting...
1150 if (PredHasWeights || SuccHasWeights)
1151 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1152 PredCases.push_back(BBCases[i]);
1153 NewSuccessors.push_back(BBCases[i].Dest);
1155 BBCases[i].Value); // This constant is taken care of
1158 // If there are any constants vectored to BB that TI doesn't handle,
1159 // they must go to the default destination of TI.
1160 for (ConstantInt *I : PTIHandled) {
1161 if (PredHasWeights || SuccHasWeights)
1162 Weights.push_back(WeightsForHandled[I]);
1163 PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1164 NewSuccessors.push_back(BBDefault);
1168 // Okay, at this point, we know which new successor Pred will get. Make
1169 // sure we update the number of entries in the PHI nodes for these
1171 for (BasicBlock *NewSuccessor : NewSuccessors)
1172 AddPredecessorToBlock(NewSuccessor, Pred, BB);
1174 Builder.SetInsertPoint(PTI);
1175 // Convert pointer to int before we switch.
1176 if (CV->getType()->isPointerTy()) {
1177 CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
1181 // Now that the successors are updated, create the new Switch instruction.
1183 Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1184 NewSI->setDebugLoc(PTI->getDebugLoc());
1185 for (ValueEqualityComparisonCase &V : PredCases)
1186 NewSI->addCase(V.Value, V.Dest);
1188 if (PredHasWeights || SuccHasWeights) {
1189 // Halve the weights if any of them cannot fit in an uint32_t
1190 FitWeights(Weights);
1192 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1194 setBranchWeights(NewSI, MDWeights);
1197 EraseTerminatorInstAndDCECond(PTI);
1199 // Okay, last check. If BB is still a successor of PSI, then we must
1200 // have an infinite loop case. If so, add an infinitely looping block
1201 // to handle the case to preserve the behavior of the code.
1202 BasicBlock *InfLoopBlock = nullptr;
1203 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1204 if (NewSI->getSuccessor(i) == BB) {
1205 if (!InfLoopBlock) {
1206 // Insert it at the end of the function, because it's either code,
1207 // or it won't matter if it's hot. :)
1208 InfLoopBlock = BasicBlock::Create(BB->getContext(), "infloop",
1210 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1212 NewSI->setSuccessor(i, InfLoopBlock);
1221 // If we would need to insert a select that uses the value of this invoke
1222 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1223 // can't hoist the invoke, as there is nowhere to put the select in this case.
1224 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1225 Instruction *I1, Instruction *I2) {
1226 for (BasicBlock *Succ : successors(BB1)) {
1227 for (const PHINode &PN : Succ->phis()) {
1228 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1229 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1230 if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1238 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I);
1240 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1241 /// in the two blocks up into the branch block. The caller of this function
1242 /// guarantees that BI's block dominates BB1 and BB2.
1243 static bool HoistThenElseCodeToIf(BranchInst *BI,
1244 const TargetTransformInfo &TTI) {
1245 // This does very trivial matching, with limited scanning, to find identical
1246 // instructions in the two blocks. In particular, we don't want to get into
1247 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1248 // such, we currently just scan for obviously identical instructions in an
1250 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1251 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1253 BasicBlock::iterator BB1_Itr = BB1->begin();
1254 BasicBlock::iterator BB2_Itr = BB2->begin();
1256 Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1257 // Skip debug info if it is not identical.
1258 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1259 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1260 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1261 while (isa<DbgInfoIntrinsic>(I1))
1263 while (isa<DbgInfoIntrinsic>(I2))
1266 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1267 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1270 BasicBlock *BIParent = BI->getParent();
1272 bool Changed = false;
1274 // If we are hoisting the terminator instruction, don't move one (making a
1275 // broken BB), instead clone it, and remove BI.
1276 if (isa<TerminatorInst>(I1))
1277 goto HoistTerminator;
1279 // If we're going to hoist a call, make sure that the two instructions we're
1280 // commoning/hoisting are both marked with musttail, or neither of them is
1281 // marked as such. Otherwise, we might end up in a situation where we hoist
1282 // from a block where the terminator is a `ret` to a block where the terminator
1283 // is a `br`, and `musttail` calls expect to be followed by a return.
1284 auto *C1 = dyn_cast<CallInst>(I1);
1285 auto *C2 = dyn_cast<CallInst>(I2);
1287 if (C1->isMustTailCall() != C2->isMustTailCall())
1290 if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1293 // For a normal instruction, we just move one to right before the branch,
1294 // then replace all uses of the other with the first. Finally, we remove
1295 // the now redundant second instruction.
1296 BIParent->getInstList().splice(BI->getIterator(), BB1->getInstList(), I1);
1297 if (!I2->use_empty())
1298 I2->replaceAllUsesWith(I1);
1300 unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1301 LLVMContext::MD_range,
1302 LLVMContext::MD_fpmath,
1303 LLVMContext::MD_invariant_load,
1304 LLVMContext::MD_nonnull,
1305 LLVMContext::MD_invariant_group,
1306 LLVMContext::MD_align,
1307 LLVMContext::MD_dereferenceable,
1308 LLVMContext::MD_dereferenceable_or_null,
1309 LLVMContext::MD_mem_parallel_loop_access};
1310 combineMetadata(I1, I2, KnownIDs);
1312 // I1 and I2 are being combined into a single instruction. Its debug
1313 // location is the merged locations of the original instructions.
1314 I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1316 I2->eraseFromParent();
1321 // Skip debug info if it is not identical.
1322 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1323 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1324 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1325 while (isa<DbgInfoIntrinsic>(I1))
1327 while (isa<DbgInfoIntrinsic>(I2))
1330 } while (I1->isIdenticalToWhenDefined(I2));
1335 // It may not be possible to hoist an invoke.
1336 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1339 for (BasicBlock *Succ : successors(BB1)) {
1340 for (PHINode &PN : Succ->phis()) {
1341 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1342 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1346 // Check for passingValueIsAlwaysUndefined here because we would rather
1347 // eliminate undefined control flow then converting it to a select.
1348 if (passingValueIsAlwaysUndefined(BB1V, &PN) ||
1349 passingValueIsAlwaysUndefined(BB2V, &PN))
1352 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1354 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1359 // Okay, it is safe to hoist the terminator.
1360 Instruction *NT = I1->clone();
1361 BIParent->getInstList().insert(BI->getIterator(), NT);
1362 if (!NT->getType()->isVoidTy()) {
1363 I1->replaceAllUsesWith(NT);
1364 I2->replaceAllUsesWith(NT);
1368 IRBuilder<NoFolder> Builder(NT);
1369 // Hoisting one of the terminators from our successor is a great thing.
1370 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1371 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1372 // nodes, so we insert select instruction to compute the final result.
1373 std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1374 for (BasicBlock *Succ : successors(BB1)) {
1375 for (PHINode &PN : Succ->phis()) {
1376 Value *BB1V = PN.getIncomingValueForBlock(BB1);
1377 Value *BB2V = PN.getIncomingValueForBlock(BB2);
1381 // These values do not agree. Insert a select instruction before NT
1382 // that determines the right value.
1383 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1385 SI = cast<SelectInst>(
1386 Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1387 BB1V->getName() + "." + BB2V->getName(), BI));
1389 // Make the PHI node use the select for all incoming values for BB1/BB2
1390 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1391 if (PN.getIncomingBlock(i) == BB1 || PN.getIncomingBlock(i) == BB2)
1392 PN.setIncomingValue(i, SI);
1396 // Update any PHI nodes in our new successors.
1397 for (BasicBlock *Succ : successors(BB1))
1398 AddPredecessorToBlock(Succ, BIParent, BB1);
1400 EraseTerminatorInstAndDCECond(BI);
1404 // All instructions in Insts belong to different blocks that all unconditionally
1405 // branch to a common successor. Analyze each instruction and return true if it
1406 // would be possible to sink them into their successor, creating one common
1407 // instruction instead. For every value that would be required to be provided by
1408 // PHI node (because an operand varies in each input block), add to PHIOperands.
1409 static bool canSinkInstructions(
1410 ArrayRef<Instruction *> Insts,
1411 DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1412 // Prune out obviously bad instructions to move. Any non-store instruction
1413 // must have exactly one use, and we check later that use is by a single,
1414 // common PHI instruction in the successor.
1415 for (auto *I : Insts) {
1416 // These instructions may change or break semantics if moved.
1417 if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1418 I->getType()->isTokenTy())
1421 // Conservatively return false if I is an inline-asm instruction. Sinking
1422 // and merging inline-asm instructions can potentially create arguments
1423 // that cannot satisfy the inline-asm constraints.
1424 if (const auto *C = dyn_cast<CallInst>(I))
1425 if (C->isInlineAsm())
1428 // Everything must have only one use too, apart from stores which
1430 if (!isa<StoreInst>(I) && !I->hasOneUse())
1434 const Instruction *I0 = Insts.front();
1435 for (auto *I : Insts)
1436 if (!I->isSameOperationAs(I0))
1439 // All instructions in Insts are known to be the same opcode. If they aren't
1440 // stores, check the only user of each is a PHI or in the same block as the
1441 // instruction, because if a user is in the same block as an instruction
1442 // we're contemplating sinking, it must already be determined to be sinkable.
1443 if (!isa<StoreInst>(I0)) {
1444 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1445 auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1446 if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1447 auto *U = cast<Instruction>(*I->user_begin());
1449 PNUse->getParent() == Succ &&
1450 PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1451 U->getParent() == I->getParent();
1456 // Because SROA can't handle speculating stores of selects, try not
1457 // to sink loads or stores of allocas when we'd have to create a PHI for
1458 // the address operand. Also, because it is likely that loads or stores
1459 // of allocas will disappear when Mem2Reg/SROA is run, don't sink them.
1460 // This can cause code churn which can have unintended consequences down
1461 // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1462 // FIXME: This is a workaround for a deficiency in SROA - see
1463 // https://llvm.org/bugs/show_bug.cgi?id=30188
1464 if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1465 return isa<AllocaInst>(I->getOperand(1));
1468 if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1469 return isa<AllocaInst>(I->getOperand(0));
1473 for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1474 if (I0->getOperand(OI)->getType()->isTokenTy())
1475 // Don't touch any operand of token type.
1478 auto SameAsI0 = [&I0, OI](const Instruction *I) {
1479 assert(I->getNumOperands() == I0->getNumOperands());
1480 return I->getOperand(OI) == I0->getOperand(OI);
1482 if (!all_of(Insts, SameAsI0)) {
1483 if (!canReplaceOperandWithVariable(I0, OI))
1484 // We can't create a PHI from this GEP.
1486 // Don't create indirect calls! The called value is the final operand.
1487 if ((isa<CallInst>(I0) || isa<InvokeInst>(I0)) && OI == OE - 1) {
1488 // FIXME: if the call was *already* indirect, we should do this.
1491 for (auto *I : Insts)
1492 PHIOperands[I].push_back(I->getOperand(OI));
1498 // Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1499 // instruction of every block in Blocks to their common successor, commoning
1500 // into one instruction.
1501 static bool sinkLastInstruction(ArrayRef<BasicBlock*> Blocks) {
1502 auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1504 // canSinkLastInstruction returning true guarantees that every block has at
1505 // least one non-terminator instruction.
1506 SmallVector<Instruction*,4> Insts;
1507 for (auto *BB : Blocks) {
1508 Instruction *I = BB->getTerminator();
1510 I = I->getPrevNode();
1511 } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1512 if (!isa<DbgInfoIntrinsic>(I))
1516 // The only checking we need to do now is that all users of all instructions
1517 // are the same PHI node. canSinkLastInstruction should have checked this but
1518 // it is slightly over-aggressive - it gets confused by commutative instructions
1519 // so double-check it here.
1520 Instruction *I0 = Insts.front();
1521 if (!isa<StoreInst>(I0)) {
1522 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1523 if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1524 auto *U = cast<Instruction>(*I->user_begin());
1530 // We don't need to do any more checking here; canSinkLastInstruction should
1531 // have done it all for us.
1532 SmallVector<Value*, 4> NewOperands;
1533 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1534 // This check is different to that in canSinkLastInstruction. There, we
1535 // cared about the global view once simplifycfg (and instcombine) have
1536 // completed - it takes into account PHIs that become trivially
1537 // simplifiable. However here we need a more local view; if an operand
1538 // differs we create a PHI and rely on instcombine to clean up the very
1539 // small mess we may make.
1540 bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1541 return I->getOperand(O) != I0->getOperand(O);
1544 NewOperands.push_back(I0->getOperand(O));
1548 // Create a new PHI in the successor block and populate it.
1549 auto *Op = I0->getOperand(O);
1550 assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1551 auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1552 Op->getName() + ".sink", &BBEnd->front());
1553 for (auto *I : Insts)
1554 PN->addIncoming(I->getOperand(O), I->getParent());
1555 NewOperands.push_back(PN);
1558 // Arbitrarily use I0 as the new "common" instruction; remap its operands
1559 // and move it to the start of the successor block.
1560 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1561 I0->getOperandUse(O).set(NewOperands[O]);
1562 I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1564 // Update metadata and IR flags, and merge debug locations.
1565 for (auto *I : Insts)
1567 // The debug location for the "common" instruction is the merged locations
1568 // of all the commoned instructions. We start with the original location
1569 // of the "common" instruction and iteratively merge each location in the
1571 // This is an N-way merge, which will be inefficient if I0 is a CallInst.
1572 // However, as N-way merge for CallInst is rare, so we use simplified API
1573 // instead of using complex API for N-way merge.
1574 I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());
1575 combineMetadataForCSE(I0, I);
1579 if (!isa<StoreInst>(I0)) {
1580 // canSinkLastInstruction checked that all instructions were used by
1581 // one and only one PHI node. Find that now, RAUW it to our common
1582 // instruction and nuke it.
1583 assert(I0->hasOneUse());
1584 auto *PN = cast<PHINode>(*I0->user_begin());
1585 PN->replaceAllUsesWith(I0);
1586 PN->eraseFromParent();
1589 // Finally nuke all instructions apart from the common instruction.
1590 for (auto *I : Insts)
1592 I->eraseFromParent();
1599 // LockstepReverseIterator - Iterates through instructions
1600 // in a set of blocks in reverse order from the first non-terminator.
1601 // For example (assume all blocks have size n):
1602 // LockstepReverseIterator I([B1, B2, B3]);
1603 // *I-- = [B1[n], B2[n], B3[n]];
1604 // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1605 // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1607 class LockstepReverseIterator {
1608 ArrayRef<BasicBlock*> Blocks;
1609 SmallVector<Instruction*,4> Insts;
1613 LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) {
1620 for (auto *BB : Blocks) {
1621 Instruction *Inst = BB->getTerminator();
1622 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1623 Inst = Inst->getPrevNode();
1625 // Block wasn't big enough.
1629 Insts.push_back(Inst);
1633 bool isValid() const {
1640 for (auto *&Inst : Insts) {
1641 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1642 Inst = Inst->getPrevNode();
1643 // Already at beginning of block.
1651 ArrayRef<Instruction*> operator * () const {
1656 } // end anonymous namespace
1658 /// Check whether BB's predecessors end with unconditional branches. If it is
1659 /// true, sink any common code from the predecessors to BB.
1660 /// We also allow one predecessor to end with conditional branch (but no more
1662 static bool SinkCommonCodeFromPredecessors(BasicBlock *BB) {
1663 // We support two situations:
1664 // (1) all incoming arcs are unconditional
1665 // (2) one incoming arc is conditional
1667 // (2) is very common in switch defaults and
1668 // else-if patterns;
1671 // else if (b) f(2);
1684 // [end] has two unconditional predecessor arcs and one conditional. The
1685 // conditional refers to the implicit empty 'else' arc. This conditional
1686 // arc can also be caused by an empty default block in a switch.
1688 // In this case, we attempt to sink code from all *unconditional* arcs.
1689 // If we can sink instructions from these arcs (determined during the scan
1690 // phase below) we insert a common successor for all unconditional arcs and
1691 // connect that to [end], to enable sinking:
1704 SmallVector<BasicBlock*,4> UnconditionalPreds;
1705 Instruction *Cond = nullptr;
1706 for (auto *B : predecessors(BB)) {
1707 auto *T = B->getTerminator();
1708 if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1709 UnconditionalPreds.push_back(B);
1710 else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1715 if (UnconditionalPreds.size() < 2)
1718 bool Changed = false;
1719 // We take a two-step approach to tail sinking. First we scan from the end of
1720 // each block upwards in lockstep. If the n'th instruction from the end of each
1721 // block can be sunk, those instructions are added to ValuesToSink and we
1722 // carry on. If we can sink an instruction but need to PHI-merge some operands
1723 // (because they're not identical in each instruction) we add these to
1725 unsigned ScanIdx = 0;
1726 SmallPtrSet<Value*,4> InstructionsToSink;
1727 DenseMap<Instruction*, SmallVector<Value*,4>> PHIOperands;
1728 LockstepReverseIterator LRI(UnconditionalPreds);
1729 while (LRI.isValid() &&
1730 canSinkInstructions(*LRI, PHIOperands)) {
1731 DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0] << "\n");
1732 InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1737 auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1738 unsigned NumPHIdValues = 0;
1739 for (auto *I : *LRI)
1740 for (auto *V : PHIOperands[I])
1741 if (InstructionsToSink.count(V) == 0)
1743 DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1744 unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1745 if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1748 return NumPHIInsts <= 1;
1751 if (ScanIdx > 0 && Cond) {
1752 // Check if we would actually sink anything first! This mutates the CFG and
1753 // adds an extra block. The goal in doing this is to allow instructions that
1754 // couldn't be sunk before to be sunk - obviously, speculatable instructions
1755 // (such as trunc, add) can be sunk and predicated already. So we check that
1756 // we're going to sink at least one non-speculatable instruction.
1759 bool Profitable = false;
1760 while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1761 if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1771 DEBUG(dbgs() << "SINK: Splitting edge\n");
1772 // We have a conditional edge and we're going to sink some instructions.
1773 // Insert a new block postdominating all blocks we're going to sink from.
1774 if (!SplitBlockPredecessors(BB, UnconditionalPreds, ".sink.split"))
1775 // Edges couldn't be split.
1780 // Now that we've analyzed all potential sinking candidates, perform the
1781 // actual sink. We iteratively sink the last non-terminator of the source
1782 // blocks into their common successor unless doing so would require too
1783 // many PHI instructions to be generated (currently only one PHI is allowed
1784 // per sunk instruction).
1786 // We can use InstructionsToSink to discount values needing PHI-merging that will
1787 // actually be sunk in a later iteration. This allows us to be more
1788 // aggressive in what we sink. This does allow a false positive where we
1789 // sink presuming a later value will also be sunk, but stop half way through
1790 // and never actually sink it which means we produce more PHIs than intended.
1791 // This is unlikely in practice though.
1792 for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1793 DEBUG(dbgs() << "SINK: Sink: "
1794 << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1797 // Because we've sunk every instruction in turn, the current instruction to
1798 // sink is always at index 0.
1800 if (!ProfitableToSinkInstruction(LRI)) {
1801 // Too many PHIs would be created.
1802 DEBUG(dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1806 if (!sinkLastInstruction(UnconditionalPreds))
1814 /// \brief Determine if we can hoist sink a sole store instruction out of a
1815 /// conditional block.
1817 /// We are looking for code like the following:
1819 /// store i32 %add, i32* %arrayidx2
1820 /// ... // No other stores or function calls (we could be calling a memory
1821 /// ... // function).
1822 /// %cmp = icmp ult %x, %y
1823 /// br i1 %cmp, label %EndBB, label %ThenBB
1825 /// store i32 %add5, i32* %arrayidx2
1829 /// We are going to transform this into:
1831 /// store i32 %add, i32* %arrayidx2
1833 /// %cmp = icmp ult %x, %y
1834 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1835 /// store i32 %add.add5, i32* %arrayidx2
1838 /// \return The pointer to the value of the previous store if the store can be
1839 /// hoisted into the predecessor block. 0 otherwise.
1840 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1841 BasicBlock *StoreBB, BasicBlock *EndBB) {
1842 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1846 // Volatile or atomic.
1847 if (!StoreToHoist->isSimple())
1850 Value *StorePtr = StoreToHoist->getPointerOperand();
1852 // Look for a store to the same pointer in BrBB.
1853 unsigned MaxNumInstToLookAt = 9;
1854 for (Instruction &CurI : reverse(*BrBB)) {
1855 if (!MaxNumInstToLookAt)
1858 if (isa<DbgInfoIntrinsic>(CurI))
1860 --MaxNumInstToLookAt;
1862 // Could be calling an instruction that affects memory like free().
1863 if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1866 if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1867 // Found the previous store make sure it stores to the same location.
1868 if (SI->getPointerOperand() == StorePtr)
1869 // Found the previous store, return its value operand.
1870 return SI->getValueOperand();
1871 return nullptr; // Unknown store.
1878 /// \brief Speculate a conditional basic block flattening the CFG.
1880 /// Note that this is a very risky transform currently. Speculating
1881 /// instructions like this is most often not desirable. Instead, there is an MI
1882 /// pass which can do it with full awareness of the resource constraints.
1883 /// However, some cases are "obvious" and we should do directly. An example of
1884 /// this is speculating a single, reasonably cheap instruction.
1886 /// There is only one distinct advantage to flattening the CFG at the IR level:
1887 /// it makes very common but simplistic optimizations such as are common in
1888 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1889 /// modeling their effects with easier to reason about SSA value graphs.
1892 /// An illustration of this transform is turning this IR:
1895 /// %cmp = icmp ult %x, %y
1896 /// br i1 %cmp, label %EndBB, label %ThenBB
1898 /// %sub = sub %x, %y
1901 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1908 /// %cmp = icmp ult %x, %y
1909 /// %sub = sub %x, %y
1910 /// %cond = select i1 %cmp, 0, %sub
1914 /// \returns true if the conditional block is removed.
1915 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1916 const TargetTransformInfo &TTI) {
1917 // Be conservative for now. FP select instruction can often be expensive.
1918 Value *BrCond = BI->getCondition();
1919 if (isa<FCmpInst>(BrCond))
1922 BasicBlock *BB = BI->getParent();
1923 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1925 // If ThenBB is actually on the false edge of the conditional branch, remember
1926 // to swap the select operands later.
1927 bool Invert = false;
1928 if (ThenBB != BI->getSuccessor(0)) {
1929 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1932 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1934 // Keep a count of how many times instructions are used within CondBB when
1935 // they are candidates for sinking into CondBB. Specifically:
1936 // - They are defined in BB, and
1937 // - They have no side effects, and
1938 // - All of their uses are in CondBB.
1939 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1941 SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics;
1943 unsigned SpeculationCost = 0;
1944 Value *SpeculatedStoreValue = nullptr;
1945 StoreInst *SpeculatedStore = nullptr;
1946 for (BasicBlock::iterator BBI = ThenBB->begin(),
1947 BBE = std::prev(ThenBB->end());
1948 BBI != BBE; ++BBI) {
1949 Instruction *I = &*BBI;
1951 if (isa<DbgInfoIntrinsic>(I)) {
1952 SpeculatedDbgIntrinsics.push_back(I);
1956 // Only speculatively execute a single instruction (not counting the
1957 // terminator) for now.
1959 if (SpeculationCost > 1)
1962 // Don't hoist the instruction if it's unsafe or expensive.
1963 if (!isSafeToSpeculativelyExecute(I) &&
1964 !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1965 I, BB, ThenBB, EndBB))))
1967 if (!SpeculatedStoreValue &&
1968 ComputeSpeculationCost(I, TTI) >
1969 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
1972 // Store the store speculation candidate.
1973 if (SpeculatedStoreValue)
1974 SpeculatedStore = cast<StoreInst>(I);
1976 // Do not hoist the instruction if any of its operands are defined but not
1977 // used in BB. The transformation will prevent the operand from
1978 // being sunk into the use block.
1979 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
1980 Instruction *OpI = dyn_cast<Instruction>(*i);
1981 if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
1982 continue; // Not a candidate for sinking.
1984 ++SinkCandidateUseCounts[OpI];
1988 // Consider any sink candidates which are only used in CondBB as costs for
1989 // speculation. Note, while we iterate over a DenseMap here, we are summing
1990 // and so iteration order isn't significant.
1991 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator
1992 I = SinkCandidateUseCounts.begin(),
1993 E = SinkCandidateUseCounts.end();
1995 if (I->first->getNumUses() == I->second) {
1997 if (SpeculationCost > 1)
2001 // Check that the PHI nodes can be converted to selects.
2002 bool HaveRewritablePHIs = false;
2003 for (PHINode &PN : EndBB->phis()) {
2004 Value *OrigV = PN.getIncomingValueForBlock(BB);
2005 Value *ThenV = PN.getIncomingValueForBlock(ThenBB);
2007 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2008 // Skip PHIs which are trivial.
2012 // Don't convert to selects if we could remove undefined behavior instead.
2013 if (passingValueIsAlwaysUndefined(OrigV, &PN) ||
2014 passingValueIsAlwaysUndefined(ThenV, &PN))
2017 HaveRewritablePHIs = true;
2018 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2019 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2020 if (!OrigCE && !ThenCE)
2021 continue; // Known safe and cheap.
2023 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2024 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2026 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2027 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2029 2 * PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2030 if (OrigCost + ThenCost > MaxCost)
2033 // Account for the cost of an unfolded ConstantExpr which could end up
2034 // getting expanded into Instructions.
2035 // FIXME: This doesn't account for how many operations are combined in the
2036 // constant expression.
2038 if (SpeculationCost > 1)
2042 // If there are no PHIs to process, bail early. This helps ensure idempotence
2044 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2047 // If we get here, we can hoist the instruction and if-convert.
2048 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2050 // Insert a select of the value of the speculated store.
2051 if (SpeculatedStoreValue) {
2052 IRBuilder<NoFolder> Builder(BI);
2053 Value *TrueV = SpeculatedStore->getValueOperand();
2054 Value *FalseV = SpeculatedStoreValue;
2056 std::swap(TrueV, FalseV);
2057 Value *S = Builder.CreateSelect(
2058 BrCond, TrueV, FalseV, "spec.store.select", BI);
2059 SpeculatedStore->setOperand(0, S);
2060 SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),
2061 SpeculatedStore->getDebugLoc());
2064 // Metadata can be dependent on the condition we are hoisting above.
2065 // Conservatively strip all metadata on the instruction.
2066 for (auto &I : *ThenBB)
2067 I.dropUnknownNonDebugMetadata();
2069 // Hoist the instructions.
2070 BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2071 ThenBB->begin(), std::prev(ThenBB->end()));
2073 // Insert selects and rewrite the PHI operands.
2074 IRBuilder<NoFolder> Builder(BI);
2075 for (PHINode &PN : EndBB->phis()) {
2076 unsigned OrigI = PN.getBasicBlockIndex(BB);
2077 unsigned ThenI = PN.getBasicBlockIndex(ThenBB);
2078 Value *OrigV = PN.getIncomingValue(OrigI);
2079 Value *ThenV = PN.getIncomingValue(ThenI);
2081 // Skip PHIs which are trivial.
2085 // Create a select whose true value is the speculatively executed value and
2086 // false value is the preexisting value. Swap them if the branch
2087 // destinations were inverted.
2088 Value *TrueV = ThenV, *FalseV = OrigV;
2090 std::swap(TrueV, FalseV);
2091 Value *V = Builder.CreateSelect(
2092 BrCond, TrueV, FalseV, "spec.select", BI);
2093 PN.setIncomingValue(OrigI, V);
2094 PN.setIncomingValue(ThenI, V);
2097 // Remove speculated dbg intrinsics.
2098 // FIXME: Is it possible to do this in a more elegant way? Moving/merging the
2099 // dbg value for the different flows and inserting it after the select.
2100 for (Instruction *I : SpeculatedDbgIntrinsics)
2101 I->eraseFromParent();
2107 /// Return true if we can thread a branch across this block.
2108 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
2109 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
2112 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2113 if (isa<DbgInfoIntrinsic>(BBI))
2116 return false; // Don't clone large BB's.
2119 // We can only support instructions that do not define values that are
2120 // live outside of the current basic block.
2121 for (User *U : BBI->users()) {
2122 Instruction *UI = cast<Instruction>(U);
2123 if (UI->getParent() != BB || isa<PHINode>(UI))
2127 // Looks ok, continue checking.
2133 /// If we have a conditional branch on a PHI node value that is defined in the
2134 /// same block as the branch and if any PHI entries are constants, thread edges
2135 /// corresponding to that entry to be branches to their ultimate destination.
2136 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL,
2137 AssumptionCache *AC) {
2138 BasicBlock *BB = BI->getParent();
2139 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2140 // NOTE: we currently cannot transform this case if the PHI node is used
2141 // outside of the block.
2142 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2145 // Degenerate case of a single entry PHI.
2146 if (PN->getNumIncomingValues() == 1) {
2147 FoldSingleEntryPHINodes(PN->getParent());
2151 // Now we know that this block has multiple preds and two succs.
2152 if (!BlockIsSimpleEnoughToThreadThrough(BB))
2155 // Can't fold blocks that contain noduplicate or convergent calls.
2156 if (any_of(*BB, [](const Instruction &I) {
2157 const CallInst *CI = dyn_cast<CallInst>(&I);
2158 return CI && (CI->cannotDuplicate() || CI->isConvergent());
2162 // Okay, this is a simple enough basic block. See if any phi values are
2164 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2165 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
2166 if (!CB || !CB->getType()->isIntegerTy(1))
2169 // Okay, we now know that all edges from PredBB should be revectored to
2170 // branch to RealDest.
2171 BasicBlock *PredBB = PN->getIncomingBlock(i);
2172 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2175 continue; // Skip self loops.
2176 // Skip if the predecessor's terminator is an indirect branch.
2177 if (isa<IndirectBrInst>(PredBB->getTerminator()))
2180 // The dest block might have PHI nodes, other predecessors and other
2181 // difficult cases. Instead of being smart about this, just insert a new
2182 // block that jumps to the destination block, effectively splitting
2183 // the edge we are about to create.
2184 BasicBlock *EdgeBB =
2185 BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2186 RealDest->getParent(), RealDest);
2187 BranchInst::Create(RealDest, EdgeBB);
2189 // Update PHI nodes.
2190 AddPredecessorToBlock(RealDest, EdgeBB, BB);
2192 // BB may have instructions that are being threaded over. Clone these
2193 // instructions into EdgeBB. We know that there will be no uses of the
2194 // cloned instructions outside of EdgeBB.
2195 BasicBlock::iterator InsertPt = EdgeBB->begin();
2196 DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2197 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2198 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2199 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2202 // Clone the instruction.
2203 Instruction *N = BBI->clone();
2205 N->setName(BBI->getName() + ".c");
2207 // Update operands due to translation.
2208 for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2209 DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2210 if (PI != TranslateMap.end())
2214 // Check for trivial simplification.
2215 if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2216 if (!BBI->use_empty())
2217 TranslateMap[&*BBI] = V;
2218 if (!N->mayHaveSideEffects()) {
2219 N->deleteValue(); // Instruction folded away, don't need actual inst
2223 if (!BBI->use_empty())
2224 TranslateMap[&*BBI] = N;
2226 // Insert the new instruction into its new home.
2228 EdgeBB->getInstList().insert(InsertPt, N);
2230 // Register the new instruction with the assumption cache if necessary.
2231 if (auto *II = dyn_cast_or_null<IntrinsicInst>(N))
2232 if (II->getIntrinsicID() == Intrinsic::assume)
2233 AC->registerAssumption(II);
2236 // Loop over all of the edges from PredBB to BB, changing them to branch
2237 // to EdgeBB instead.
2238 TerminatorInst *PredBBTI = PredBB->getTerminator();
2239 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2240 if (PredBBTI->getSuccessor(i) == BB) {
2241 BB->removePredecessor(PredBB);
2242 PredBBTI->setSuccessor(i, EdgeBB);
2245 // Recurse, simplifying any other constants.
2246 return FoldCondBranchOnPHI(BI, DL, AC) | true;
2252 /// Given a BB that starts with the specified two-entry PHI node,
2253 /// see if we can eliminate it.
2254 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI,
2255 const DataLayout &DL) {
2256 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2257 // statement", which has a very simple dominance structure. Basically, we
2258 // are trying to find the condition that is being branched on, which
2259 // subsequently causes this merge to happen. We really want control
2260 // dependence information for this check, but simplifycfg can't keep it up
2261 // to date, and this catches most of the cases we care about anyway.
2262 BasicBlock *BB = PN->getParent();
2263 BasicBlock *IfTrue, *IfFalse;
2264 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2266 // Don't bother if the branch will be constant folded trivially.
2267 isa<ConstantInt>(IfCond))
2270 // Okay, we found that we can merge this two-entry phi node into a select.
2271 // Doing so would require us to fold *all* two entry phi nodes in this block.
2272 // At some point this becomes non-profitable (particularly if the target
2273 // doesn't support cmov's). Only do this transformation if there are two or
2274 // fewer PHI nodes in this block.
2275 unsigned NumPhis = 0;
2276 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2280 // Loop over the PHI's seeing if we can promote them all to select
2281 // instructions. While we are at it, keep track of the instructions
2282 // that need to be moved to the dominating block.
2283 SmallPtrSet<Instruction *, 4> AggressiveInsts;
2284 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2285 MaxCostVal1 = PHINodeFoldingThreshold;
2286 MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2287 MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2289 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2290 PHINode *PN = cast<PHINode>(II++);
2291 if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2292 PN->replaceAllUsesWith(V);
2293 PN->eraseFromParent();
2297 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
2298 MaxCostVal0, TTI) ||
2299 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
2304 // If we folded the first phi, PN dangles at this point. Refresh it. If
2305 // we ran out of PHIs then we simplified them all.
2306 PN = dyn_cast<PHINode>(BB->begin());
2310 // Don't fold i1 branches on PHIs which contain binary operators. These can
2311 // often be turned into switches and other things.
2312 if (PN->getType()->isIntegerTy(1) &&
2313 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2314 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2315 isa<BinaryOperator>(IfCond)))
2318 // If all PHI nodes are promotable, check to make sure that all instructions
2319 // in the predecessor blocks can be promoted as well. If not, we won't be able
2320 // to get rid of the control flow, so it's not worth promoting to select
2322 BasicBlock *DomBlock = nullptr;
2323 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2324 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2325 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2328 DomBlock = *pred_begin(IfBlock1);
2329 for (BasicBlock::iterator I = IfBlock1->begin(); !isa<TerminatorInst>(I);
2331 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2332 // This is not an aggressive instruction that we can promote.
2333 // Because of this, we won't be able to get rid of the control flow, so
2334 // the xform is not worth it.
2339 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2342 DomBlock = *pred_begin(IfBlock2);
2343 for (BasicBlock::iterator I = IfBlock2->begin(); !isa<TerminatorInst>(I);
2345 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2346 // This is not an aggressive instruction that we can promote.
2347 // Because of this, we won't be able to get rid of the control flow, so
2348 // the xform is not worth it.
2353 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
2354 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
2356 // If we can still promote the PHI nodes after this gauntlet of tests,
2357 // do all of the PHI's now.
2358 Instruction *InsertPt = DomBlock->getTerminator();
2359 IRBuilder<NoFolder> Builder(InsertPt);
2361 // Move all 'aggressive' instructions, which are defined in the
2362 // conditional parts of the if's up to the dominating block.
2364 for (auto &I : *IfBlock1)
2365 I.dropUnknownNonDebugMetadata();
2366 DomBlock->getInstList().splice(InsertPt->getIterator(),
2367 IfBlock1->getInstList(), IfBlock1->begin(),
2368 IfBlock1->getTerminator()->getIterator());
2371 for (auto &I : *IfBlock2)
2372 I.dropUnknownNonDebugMetadata();
2373 DomBlock->getInstList().splice(InsertPt->getIterator(),
2374 IfBlock2->getInstList(), IfBlock2->begin(),
2375 IfBlock2->getTerminator()->getIterator());
2378 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2379 // Change the PHI node into a select instruction.
2380 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2381 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2383 Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2384 PN->replaceAllUsesWith(Sel);
2386 PN->eraseFromParent();
2389 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2390 // has been flattened. Change DomBlock to jump directly to our new block to
2391 // avoid other simplifycfg's kicking in on the diamond.
2392 TerminatorInst *OldTI = DomBlock->getTerminator();
2393 Builder.SetInsertPoint(OldTI);
2394 Builder.CreateBr(BB);
2395 OldTI->eraseFromParent();
2399 /// If we found a conditional branch that goes to two returning blocks,
2400 /// try to merge them together into one return,
2401 /// introducing a select if the return values disagree.
2402 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
2403 IRBuilder<> &Builder) {
2404 assert(BI->isConditional() && "Must be a conditional branch");
2405 BasicBlock *TrueSucc = BI->getSuccessor(0);
2406 BasicBlock *FalseSucc = BI->getSuccessor(1);
2407 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2408 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2410 // Check to ensure both blocks are empty (just a return) or optionally empty
2411 // with PHI nodes. If there are other instructions, merging would cause extra
2412 // computation on one path or the other.
2413 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2415 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2418 Builder.SetInsertPoint(BI);
2419 // Okay, we found a branch that is going to two return nodes. If
2420 // there is no return value for this function, just change the
2421 // branch into a return.
2422 if (FalseRet->getNumOperands() == 0) {
2423 TrueSucc->removePredecessor(BI->getParent());
2424 FalseSucc->removePredecessor(BI->getParent());
2425 Builder.CreateRetVoid();
2426 EraseTerminatorInstAndDCECond(BI);
2430 // Otherwise, figure out what the true and false return values are
2431 // so we can insert a new select instruction.
2432 Value *TrueValue = TrueRet->getReturnValue();
2433 Value *FalseValue = FalseRet->getReturnValue();
2435 // Unwrap any PHI nodes in the return blocks.
2436 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2437 if (TVPN->getParent() == TrueSucc)
2438 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2439 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2440 if (FVPN->getParent() == FalseSucc)
2441 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2443 // In order for this transformation to be safe, we must be able to
2444 // unconditionally execute both operands to the return. This is
2445 // normally the case, but we could have a potentially-trapping
2446 // constant expression that prevents this transformation from being
2448 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2451 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2455 // Okay, we collected all the mapped values and checked them for sanity, and
2456 // defined to really do this transformation. First, update the CFG.
2457 TrueSucc->removePredecessor(BI->getParent());
2458 FalseSucc->removePredecessor(BI->getParent());
2460 // Insert select instructions where needed.
2461 Value *BrCond = BI->getCondition();
2463 // Insert a select if the results differ.
2464 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2465 } else if (isa<UndefValue>(TrueValue)) {
2466 TrueValue = FalseValue;
2469 Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2474 !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2478 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2479 << "\n " << *BI << "NewRet = " << *RI
2480 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: " << *FalseSucc);
2482 EraseTerminatorInstAndDCECond(BI);
2487 /// Return true if the given instruction is available
2488 /// in its predecessor block. If yes, the instruction will be removed.
2489 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
2490 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2492 for (Instruction &I : *PB) {
2493 Instruction *PBI = &I;
2494 // Check whether Inst and PBI generate the same value.
2495 if (Inst->isIdenticalTo(PBI)) {
2496 Inst->replaceAllUsesWith(PBI);
2497 Inst->eraseFromParent();
2504 /// Return true if either PBI or BI has branch weight available, and store
2505 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2506 /// not have branch weight, use 1:1 as its weight.
2507 static bool extractPredSuccWeights(BranchInst *PBI, BranchInst *BI,
2508 uint64_t &PredTrueWeight,
2509 uint64_t &PredFalseWeight,
2510 uint64_t &SuccTrueWeight,
2511 uint64_t &SuccFalseWeight) {
2512 bool PredHasWeights =
2513 PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2514 bool SuccHasWeights =
2515 BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2516 if (PredHasWeights || SuccHasWeights) {
2517 if (!PredHasWeights)
2518 PredTrueWeight = PredFalseWeight = 1;
2519 if (!SuccHasWeights)
2520 SuccTrueWeight = SuccFalseWeight = 1;
2527 /// If this basic block is simple enough, and if a predecessor branches to us
2528 /// and one of our successors, fold the block into the predecessor and use
2529 /// logical operations to pick the right destination.
2530 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2531 BasicBlock *BB = BI->getParent();
2533 Instruction *Cond = nullptr;
2534 if (BI->isConditional())
2535 Cond = dyn_cast<Instruction>(BI->getCondition());
2537 // For unconditional branch, check for a simple CFG pattern, where
2538 // BB has a single predecessor and BB's successor is also its predecessor's
2539 // successor. If such pattern exists, check for CSE between BB and its
2541 if (BasicBlock *PB = BB->getSinglePredecessor())
2542 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2543 if (PBI->isConditional() &&
2544 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2545 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2546 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2547 Instruction *Curr = &*I++;
2548 if (isa<CmpInst>(Curr)) {
2552 // Quit if we can't remove this instruction.
2553 if (!checkCSEInPredecessor(Curr, PB))
2562 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2563 Cond->getParent() != BB || !Cond->hasOneUse())
2566 // Make sure the instruction after the condition is the cond branch.
2567 BasicBlock::iterator CondIt = ++Cond->getIterator();
2569 // Ignore dbg intrinsics.
2570 while (isa<DbgInfoIntrinsic>(CondIt))
2576 // Only allow this transformation if computing the condition doesn't involve
2577 // too many instructions and these involved instructions can be executed
2578 // unconditionally. We denote all involved instructions except the condition
2579 // as "bonus instructions", and only allow this transformation when the
2580 // number of the bonus instructions does not exceed a certain threshold.
2581 unsigned NumBonusInsts = 0;
2582 for (auto I = BB->begin(); Cond != &*I; ++I) {
2583 // Ignore dbg intrinsics.
2584 if (isa<DbgInfoIntrinsic>(I))
2586 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2588 // I has only one use and can be executed unconditionally.
2589 Instruction *User = dyn_cast<Instruction>(I->user_back());
2590 if (User == nullptr || User->getParent() != BB)
2592 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2593 // to use any other instruction, User must be an instruction between next(I)
2596 // Early exits once we reach the limit.
2597 if (NumBonusInsts > BonusInstThreshold)
2601 // Cond is known to be a compare or binary operator. Check to make sure that
2602 // neither operand is a potentially-trapping constant expression.
2603 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2606 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2610 // Finally, don't infinitely unroll conditional loops.
2611 BasicBlock *TrueDest = BI->getSuccessor(0);
2612 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2613 if (TrueDest == BB || FalseDest == BB)
2616 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2617 BasicBlock *PredBlock = *PI;
2618 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2620 // Check that we have two conditional branches. If there is a PHI node in
2621 // the common successor, verify that the same value flows in from both
2623 SmallVector<PHINode *, 4> PHIs;
2624 if (!PBI || PBI->isUnconditional() ||
2625 (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2626 (!BI->isConditional() &&
2627 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2630 // Determine if the two branches share a common destination.
2631 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2632 bool InvertPredCond = false;
2634 if (BI->isConditional()) {
2635 if (PBI->getSuccessor(0) == TrueDest) {
2636 Opc = Instruction::Or;
2637 } else if (PBI->getSuccessor(1) == FalseDest) {
2638 Opc = Instruction::And;
2639 } else if (PBI->getSuccessor(0) == FalseDest) {
2640 Opc = Instruction::And;
2641 InvertPredCond = true;
2642 } else if (PBI->getSuccessor(1) == TrueDest) {
2643 Opc = Instruction::Or;
2644 InvertPredCond = true;
2649 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2653 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2654 IRBuilder<> Builder(PBI);
2656 // If we need to invert the condition in the pred block to match, do so now.
2657 if (InvertPredCond) {
2658 Value *NewCond = PBI->getCondition();
2660 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2661 CmpInst *CI = cast<CmpInst>(NewCond);
2662 CI->setPredicate(CI->getInversePredicate());
2665 Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2668 PBI->setCondition(NewCond);
2669 PBI->swapSuccessors();
2672 // If we have bonus instructions, clone them into the predecessor block.
2673 // Note that there may be multiple predecessor blocks, so we cannot move
2674 // bonus instructions to a predecessor block.
2675 ValueToValueMapTy VMap; // maps original values to cloned values
2676 // We already make sure Cond is the last instruction before BI. Therefore,
2677 // all instructions before Cond other than DbgInfoIntrinsic are bonus
2679 for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2680 if (isa<DbgInfoIntrinsic>(BonusInst))
2682 Instruction *NewBonusInst = BonusInst->clone();
2683 RemapInstruction(NewBonusInst, VMap,
2684 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2685 VMap[&*BonusInst] = NewBonusInst;
2687 // If we moved a load, we cannot any longer claim any knowledge about
2688 // its potential value. The previous information might have been valid
2689 // only given the branch precondition.
2690 // For an analogous reason, we must also drop all the metadata whose
2691 // semantics we don't understand.
2692 NewBonusInst->dropUnknownNonDebugMetadata();
2694 PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2695 NewBonusInst->takeName(&*BonusInst);
2696 BonusInst->setName(BonusInst->getName() + ".old");
2699 // Clone Cond into the predecessor basic block, and or/and the
2700 // two conditions together.
2701 Instruction *New = Cond->clone();
2702 RemapInstruction(New, VMap,
2703 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2704 PredBlock->getInstList().insert(PBI->getIterator(), New);
2705 New->takeName(Cond);
2706 Cond->setName(New->getName() + ".old");
2708 if (BI->isConditional()) {
2709 Instruction *NewCond = cast<Instruction>(
2710 Builder.CreateBinOp(Opc, PBI->getCondition(), New, "or.cond"));
2711 PBI->setCondition(NewCond);
2713 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2715 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2716 SuccTrueWeight, SuccFalseWeight);
2717 SmallVector<uint64_t, 8> NewWeights;
2719 if (PBI->getSuccessor(0) == BB) {
2721 // PBI: br i1 %x, BB, FalseDest
2722 // BI: br i1 %y, TrueDest, FalseDest
2723 // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2724 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2725 // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2726 // TrueWeight for PBI * FalseWeight for BI.
2727 // We assume that total weights of a BranchInst can fit into 32 bits.
2728 // Therefore, we will not have overflow using 64-bit arithmetic.
2729 NewWeights.push_back(PredFalseWeight *
2730 (SuccFalseWeight + SuccTrueWeight) +
2731 PredTrueWeight * SuccFalseWeight);
2733 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2734 PBI->setSuccessor(0, TrueDest);
2736 if (PBI->getSuccessor(1) == BB) {
2738 // PBI: br i1 %x, TrueDest, BB
2739 // BI: br i1 %y, TrueDest, FalseDest
2740 // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2741 // FalseWeight for PBI * TrueWeight for BI.
2742 NewWeights.push_back(PredTrueWeight *
2743 (SuccFalseWeight + SuccTrueWeight) +
2744 PredFalseWeight * SuccTrueWeight);
2745 // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2746 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2748 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2749 PBI->setSuccessor(1, FalseDest);
2751 if (NewWeights.size() == 2) {
2752 // Halve the weights if any of them cannot fit in an uint32_t
2753 FitWeights(NewWeights);
2755 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2757 setBranchWeights(PBI, MDWeights[0], MDWeights[1]);
2759 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2761 // Update PHI nodes in the common successors.
2762 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2763 ConstantInt *PBI_C = cast<ConstantInt>(
2764 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2765 assert(PBI_C->getType()->isIntegerTy(1));
2766 Instruction *MergedCond = nullptr;
2767 if (PBI->getSuccessor(0) == TrueDest) {
2768 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2769 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2770 // is false: !PBI_Cond and BI_Value
2771 Instruction *NotCond = cast<Instruction>(
2772 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2773 MergedCond = cast<Instruction>(
2774 Builder.CreateBinOp(Instruction::And, NotCond, New, "and.cond"));
2776 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2777 Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2779 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2780 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2781 // is false: PBI_Cond and BI_Value
2782 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2783 Instruction::And, PBI->getCondition(), New, "and.cond"));
2784 if (PBI_C->isOne()) {
2785 Instruction *NotCond = cast<Instruction>(
2786 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2787 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2788 Instruction::Or, NotCond, MergedCond, "or.cond"));
2792 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2795 // Change PBI from Conditional to Unconditional.
2796 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2797 EraseTerminatorInstAndDCECond(PBI);
2801 // If BI was a loop latch, it may have had associated loop metadata.
2802 // We need to copy it to the new latch, that is, PBI.
2803 if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2804 PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2806 // TODO: If BB is reachable from all paths through PredBlock, then we
2807 // could replace PBI's branch probabilities with BI's.
2809 // Copy any debug value intrinsics into the end of PredBlock.
2810 for (Instruction &I : *BB)
2811 if (isa<DbgInfoIntrinsic>(I))
2812 I.clone()->insertBefore(PBI);
2819 // If there is only one store in BB1 and BB2, return it, otherwise return
2821 static StoreInst *findUniqueStoreInBlocks(BasicBlock *BB1, BasicBlock *BB2) {
2822 StoreInst *S = nullptr;
2823 for (auto *BB : {BB1, BB2}) {
2827 if (auto *SI = dyn_cast<StoreInst>(&I)) {
2829 // Multiple stores seen.
2838 static Value *ensureValueAvailableInSuccessor(Value *V, BasicBlock *BB,
2839 Value *AlternativeV = nullptr) {
2840 // PHI is going to be a PHI node that allows the value V that is defined in
2841 // BB to be referenced in BB's only successor.
2843 // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2844 // doesn't matter to us what the other operand is (it'll never get used). We
2845 // could just create a new PHI with an undef incoming value, but that could
2846 // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2847 // other PHI. So here we directly look for some PHI in BB's successor with V
2848 // as an incoming operand. If we find one, we use it, else we create a new
2851 // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2852 // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2853 // where OtherBB is the single other predecessor of BB's only successor.
2854 PHINode *PHI = nullptr;
2855 BasicBlock *Succ = BB->getSingleSuccessor();
2857 for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2858 if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2859 PHI = cast<PHINode>(I);
2863 assert(std::distance(pred_begin(Succ), pred_end(Succ)) == 2);
2864 auto PredI = pred_begin(Succ);
2865 BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2866 if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2873 // If V is not an instruction defined in BB, just return it.
2874 if (!AlternativeV &&
2875 (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2878 PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2879 PHI->addIncoming(V, BB);
2880 for (BasicBlock *PredBB : predecessors(Succ))
2883 AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2887 static bool mergeConditionalStoreToAddress(BasicBlock *PTB, BasicBlock *PFB,
2888 BasicBlock *QTB, BasicBlock *QFB,
2889 BasicBlock *PostBB, Value *Address,
2890 bool InvertPCond, bool InvertQCond,
2891 const DataLayout &DL) {
2892 auto IsaBitcastOfPointerType = [](const Instruction &I) {
2893 return Operator::getOpcode(&I) == Instruction::BitCast &&
2894 I.getType()->isPointerTy();
2897 // If we're not in aggressive mode, we only optimize if we have some
2898 // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2899 auto IsWorthwhile = [&](BasicBlock *BB) {
2902 // Heuristic: if the block can be if-converted/phi-folded and the
2903 // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2904 // thread this store.
2906 for (auto &I : *BB) {
2907 // Cheap instructions viable for folding.
2908 if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2911 // Free instructions.
2912 else if (isa<TerminatorInst>(I) || isa<DbgInfoIntrinsic>(I) ||
2913 IsaBitcastOfPointerType(I))
2918 // The store we want to merge is counted in N, so add 1 to make sure
2919 // we're counting the instructions that would be left.
2920 return N <= (PHINodeFoldingThreshold + 1);
2923 if (!MergeCondStoresAggressively &&
2924 (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2925 !IsWorthwhile(QFB)))
2928 // For every pointer, there must be exactly two stores, one coming from
2929 // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2930 // store (to any address) in PTB,PFB or QTB,QFB.
2931 // FIXME: We could relax this restriction with a bit more work and performance
2933 StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2934 StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2935 if (!PStore || !QStore)
2938 // Now check the stores are compatible.
2939 if (!QStore->isUnordered() || !PStore->isUnordered())
2942 // Check that sinking the store won't cause program behavior changes. Sinking
2943 // the store out of the Q blocks won't change any behavior as we're sinking
2944 // from a block to its unconditional successor. But we're moving a store from
2945 // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2946 // So we need to check that there are no aliasing loads or stores in
2947 // QBI, QTB and QFB. We also need to check there are no conflicting memory
2948 // operations between PStore and the end of its parent block.
2950 // The ideal way to do this is to query AliasAnalysis, but we don't
2951 // preserve AA currently so that is dangerous. Be super safe and just
2952 // check there are no other memory operations at all.
2953 for (auto &I : *QFB->getSinglePredecessor())
2954 if (I.mayReadOrWriteMemory())
2956 for (auto &I : *QFB)
2957 if (&I != QStore && I.mayReadOrWriteMemory())
2960 for (auto &I : *QTB)
2961 if (&I != QStore && I.mayReadOrWriteMemory())
2963 for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2965 if (&*I != PStore && I->mayReadOrWriteMemory())
2968 // OK, we're going to sink the stores to PostBB. The store has to be
2969 // conditional though, so first create the predicate.
2970 Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
2972 Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
2975 Value *PPHI = ensureValueAvailableInSuccessor(PStore->getValueOperand(),
2976 PStore->getParent());
2977 Value *QPHI = ensureValueAvailableInSuccessor(QStore->getValueOperand(),
2978 QStore->getParent(), PPHI);
2980 IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
2982 Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
2983 Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
2986 PPred = QB.CreateNot(PPred);
2988 QPred = QB.CreateNot(QPred);
2989 Value *CombinedPred = QB.CreateOr(PPred, QPred);
2992 SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
2993 QB.SetInsertPoint(T);
2994 StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
2996 PStore->getAAMetadata(AAMD, /*Merge=*/false);
2997 PStore->getAAMetadata(AAMD, /*Merge=*/true);
2998 SI->setAAMetadata(AAMD);
2999 unsigned PAlignment = PStore->getAlignment();
3000 unsigned QAlignment = QStore->getAlignment();
3001 unsigned TypeAlignment =
3002 DL.getABITypeAlignment(SI->getValueOperand()->getType());
3003 unsigned MinAlignment;
3004 unsigned MaxAlignment;
3005 std::tie(MinAlignment, MaxAlignment) = std::minmax(PAlignment, QAlignment);
3006 // Choose the minimum alignment. If we could prove both stores execute, we
3007 // could use biggest one. In this case, though, we only know that one of the
3008 // stores executes. And we don't know it's safe to take the alignment from a
3009 // store that doesn't execute.
3010 if (MinAlignment != 0) {
3011 // Choose the minimum of all non-zero alignments.
3012 SI->setAlignment(MinAlignment);
3013 } else if (MaxAlignment != 0) {
3014 // Choose the minimal alignment between the non-zero alignment and the ABI
3015 // default alignment for the type of the stored value.
3016 SI->setAlignment(std::min(MaxAlignment, TypeAlignment));
3018 // If both alignments are zero, use ABI default alignment for the type of
3019 // the stored value.
3020 SI->setAlignment(TypeAlignment);
3023 QStore->eraseFromParent();
3024 PStore->eraseFromParent();
3029 static bool mergeConditionalStores(BranchInst *PBI, BranchInst *QBI,
3030 const DataLayout &DL) {
3031 // The intention here is to find diamonds or triangles (see below) where each
3032 // conditional block contains a store to the same address. Both of these
3033 // stores are conditional, so they can't be unconditionally sunk. But it may
3034 // be profitable to speculatively sink the stores into one merged store at the
3035 // end, and predicate the merged store on the union of the two conditions of
3038 // This can reduce the number of stores executed if both of the conditions are
3039 // true, and can allow the blocks to become small enough to be if-converted.
3040 // This optimization will also chain, so that ladders of test-and-set
3041 // sequences can be if-converted away.
3043 // We only deal with simple diamonds or triangles:
3045 // PBI or PBI or a combination of the two
3055 // We model triangles as a type of diamond with a nullptr "true" block.
3056 // Triangles are canonicalized so that the fallthrough edge is represented by
3057 // a true condition, as in the diagram above.
3058 BasicBlock *PTB = PBI->getSuccessor(0);
3059 BasicBlock *PFB = PBI->getSuccessor(1);
3060 BasicBlock *QTB = QBI->getSuccessor(0);
3061 BasicBlock *QFB = QBI->getSuccessor(1);
3062 BasicBlock *PostBB = QFB->getSingleSuccessor();
3064 // Make sure we have a good guess for PostBB. If QTB's only successor is
3065 // QFB, then QFB is a better PostBB.
3066 if (QTB->getSingleSuccessor() == QFB)
3069 // If we couldn't find a good PostBB, stop.
3073 bool InvertPCond = false, InvertQCond = false;
3074 // Canonicalize fallthroughs to the true branches.
3075 if (PFB == QBI->getParent()) {
3076 std::swap(PFB, PTB);
3079 if (QFB == PostBB) {
3080 std::swap(QFB, QTB);
3084 // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3085 // and QFB may not. Model fallthroughs as a nullptr block.
3086 if (PTB == QBI->getParent())
3091 // Legality bailouts. We must have at least the non-fallthrough blocks and
3092 // the post-dominating block, and the non-fallthroughs must only have one
3094 auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3095 return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3097 if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3098 !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3100 if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3101 (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3103 if (!PostBB->hasNUses(2) || !QBI->getParent()->hasNUses(2))
3106 // OK, this is a sequence of two diamonds or triangles.
3107 // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3108 SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3109 for (auto *BB : {PTB, PFB}) {
3113 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3114 PStoreAddresses.insert(SI->getPointerOperand());
3116 for (auto *BB : {QTB, QFB}) {
3120 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3121 QStoreAddresses.insert(SI->getPointerOperand());
3124 set_intersect(PStoreAddresses, QStoreAddresses);
3125 // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3126 // clear what it contains.
3127 auto &CommonAddresses = PStoreAddresses;
3129 bool Changed = false;
3130 for (auto *Address : CommonAddresses)
3131 Changed |= mergeConditionalStoreToAddress(
3132 PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond, DL);
3136 /// If we have a conditional branch as a predecessor of another block,
3137 /// this function tries to simplify it. We know
3138 /// that PBI and BI are both conditional branches, and BI is in one of the
3139 /// successor blocks of PBI - PBI branches to BI.
3140 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
3141 const DataLayout &DL) {
3142 assert(PBI->isConditional() && BI->isConditional());
3143 BasicBlock *BB = BI->getParent();
3145 // If this block ends with a branch instruction, and if there is a
3146 // predecessor that ends on a branch of the same condition, make
3147 // this conditional branch redundant.
3148 if (PBI->getCondition() == BI->getCondition() &&
3149 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3150 // Okay, the outcome of this conditional branch is statically
3151 // knowable. If this block had a single pred, handle specially.
3152 if (BB->getSinglePredecessor()) {
3153 // Turn this into a branch on constant.
3154 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3156 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3157 return true; // Nuke the branch on constant.
3160 // Otherwise, if there are multiple predecessors, insert a PHI that merges
3161 // in the constant and simplify the block result. Subsequent passes of
3162 // simplifycfg will thread the block.
3163 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
3164 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3165 PHINode *NewPN = PHINode::Create(
3166 Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3167 BI->getCondition()->getName() + ".pr", &BB->front());
3168 // Okay, we're going to insert the PHI node. Since PBI is not the only
3169 // predecessor, compute the PHI'd conditional value for all of the preds.
3170 // Any predecessor where the condition is not computable we keep symbolic.
3171 for (pred_iterator PI = PB; PI != PE; ++PI) {
3172 BasicBlock *P = *PI;
3173 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3174 PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3175 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3176 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3178 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3181 NewPN->addIncoming(BI->getCondition(), P);
3185 BI->setCondition(NewPN);
3190 if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3194 // If both branches are conditional and both contain stores to the same
3195 // address, remove the stores from the conditionals and create a conditional
3196 // merged store at the end.
3197 if (MergeCondStores && mergeConditionalStores(PBI, BI, DL))
3200 // If this is a conditional branch in an empty block, and if any
3201 // predecessors are a conditional branch to one of our destinations,
3202 // fold the conditions into logical ops and one cond br.
3203 BasicBlock::iterator BBI = BB->begin();
3204 // Ignore dbg intrinsics.
3205 while (isa<DbgInfoIntrinsic>(BBI))
3211 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3214 } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3217 } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3220 } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3227 // Check to make sure that the other destination of this branch
3228 // isn't BB itself. If so, this is an infinite loop that will
3229 // keep getting unwound.
3230 if (PBI->getSuccessor(PBIOp) == BB)
3233 // Do not perform this transformation if it would require
3234 // insertion of a large number of select instructions. For targets
3235 // without predication/cmovs, this is a big pessimization.
3237 // Also do not perform this transformation if any phi node in the common
3238 // destination block can trap when reached by BB or PBB (PR17073). In that
3239 // case, it would be unsafe to hoist the operation into a select instruction.
3241 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3242 unsigned NumPhis = 0;
3243 for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3245 if (NumPhis > 2) // Disable this xform.
3248 PHINode *PN = cast<PHINode>(II);
3249 Value *BIV = PN->getIncomingValueForBlock(BB);
3250 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3254 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3255 Value *PBIV = PN->getIncomingValue(PBBIdx);
3256 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3261 // Finally, if everything is ok, fold the branches to logical ops.
3262 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3264 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3265 << "AND: " << *BI->getParent());
3267 // If OtherDest *is* BB, then BB is a basic block with a single conditional
3268 // branch in it, where one edge (OtherDest) goes back to itself but the other
3269 // exits. We don't *know* that the program avoids the infinite loop
3270 // (even though that seems likely). If we do this xform naively, we'll end up
3271 // recursively unpeeling the loop. Since we know that (after the xform is
3272 // done) that the block *is* infinite if reached, we just make it an obviously
3273 // infinite loop with no cond branch.
3274 if (OtherDest == BB) {
3275 // Insert it at the end of the function, because it's either code,
3276 // or it won't matter if it's hot. :)
3277 BasicBlock *InfLoopBlock =
3278 BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3279 BranchInst::Create(InfLoopBlock, InfLoopBlock);
3280 OtherDest = InfLoopBlock;
3283 DEBUG(dbgs() << *PBI->getParent()->getParent());
3285 // BI may have other predecessors. Because of this, we leave
3286 // it alone, but modify PBI.
3288 // Make sure we get to CommonDest on True&True directions.
3289 Value *PBICond = PBI->getCondition();
3290 IRBuilder<NoFolder> Builder(PBI);
3292 PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3294 Value *BICond = BI->getCondition();
3296 BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3298 // Merge the conditions.
3299 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3301 // Modify PBI to branch on the new condition to the new dests.
3302 PBI->setCondition(Cond);
3303 PBI->setSuccessor(0, CommonDest);
3304 PBI->setSuccessor(1, OtherDest);
3306 // Update branch weight for PBI.
3307 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3308 uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3310 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3311 SuccTrueWeight, SuccFalseWeight);
3313 PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3314 PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3315 SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3316 SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3317 // The weight to CommonDest should be PredCommon * SuccTotal +
3318 // PredOther * SuccCommon.
3319 // The weight to OtherDest should be PredOther * SuccOther.
3320 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3321 PredOther * SuccCommon,
3322 PredOther * SuccOther};
3323 // Halve the weights if any of them cannot fit in an uint32_t
3324 FitWeights(NewWeights);
3326 setBranchWeights(PBI, NewWeights[0], NewWeights[1]);
3329 // OtherDest may have phi nodes. If so, add an entry from PBI's
3330 // block that are identical to the entries for BI's block.
3331 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3333 // We know that the CommonDest already had an edge from PBI to
3334 // it. If it has PHIs though, the PHIs may have different
3335 // entries for BB and PBI's BB. If so, insert a select to make
3337 for (PHINode &PN : CommonDest->phis()) {
3338 Value *BIV = PN.getIncomingValueForBlock(BB);
3339 unsigned PBBIdx = PN.getBasicBlockIndex(PBI->getParent());
3340 Value *PBIV = PN.getIncomingValue(PBBIdx);
3342 // Insert a select in PBI to pick the right value.
3343 SelectInst *NV = cast<SelectInst>(
3344 Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3345 PN.setIncomingValue(PBBIdx, NV);
3346 // Although the select has the same condition as PBI, the original branch
3347 // weights for PBI do not apply to the new select because the select's
3348 // 'logical' edges are incoming edges of the phi that is eliminated, not
3349 // the outgoing edges of PBI.
3351 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3352 uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3353 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3354 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3355 // The weight to PredCommonDest should be PredCommon * SuccTotal.
3356 // The weight to PredOtherDest should be PredOther * SuccCommon.
3357 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3358 PredOther * SuccCommon};
3360 FitWeights(NewWeights);
3362 setBranchWeights(NV, NewWeights[0], NewWeights[1]);
3367 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3368 DEBUG(dbgs() << *PBI->getParent()->getParent());
3370 // This basic block is probably dead. We know it has at least
3371 // one fewer predecessor.
3375 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3376 // true or to FalseBB if Cond is false.
3377 // Takes care of updating the successors and removing the old terminator.
3378 // Also makes sure not to introduce new successors by assuming that edges to
3379 // non-successor TrueBBs and FalseBBs aren't reachable.
3380 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
3381 BasicBlock *TrueBB, BasicBlock *FalseBB,
3382 uint32_t TrueWeight,
3383 uint32_t FalseWeight) {
3384 // Remove any superfluous successor edges from the CFG.
3385 // First, figure out which successors to preserve.
3386 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3388 BasicBlock *KeepEdge1 = TrueBB;
3389 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3391 // Then remove the rest.
3392 for (BasicBlock *Succ : OldTerm->successors()) {
3393 // Make sure only to keep exactly one copy of each edge.
3394 if (Succ == KeepEdge1)
3395 KeepEdge1 = nullptr;
3396 else if (Succ == KeepEdge2)
3397 KeepEdge2 = nullptr;
3399 Succ->removePredecessor(OldTerm->getParent(),
3400 /*DontDeleteUselessPHIs=*/true);
3403 IRBuilder<> Builder(OldTerm);
3404 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3406 // Insert an appropriate new terminator.
3407 if (!KeepEdge1 && !KeepEdge2) {
3408 if (TrueBB == FalseBB)
3409 // We were only looking for one successor, and it was present.
3410 // Create an unconditional branch to it.
3411 Builder.CreateBr(TrueBB);
3413 // We found both of the successors we were looking for.
3414 // Create a conditional branch sharing the condition of the select.
3415 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3416 if (TrueWeight != FalseWeight)
3417 setBranchWeights(NewBI, TrueWeight, FalseWeight);
3419 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3420 // Neither of the selected blocks were successors, so this
3421 // terminator must be unreachable.
3422 new UnreachableInst(OldTerm->getContext(), OldTerm);
3424 // One of the selected values was a successor, but the other wasn't.
3425 // Insert an unconditional branch to the one that was found;
3426 // the edge to the one that wasn't must be unreachable.
3428 // Only TrueBB was found.
3429 Builder.CreateBr(TrueBB);
3431 // Only FalseBB was found.
3432 Builder.CreateBr(FalseBB);
3435 EraseTerminatorInstAndDCECond(OldTerm);
3440 // (switch (select cond, X, Y)) on constant X, Y
3441 // with a branch - conditional if X and Y lead to distinct BBs,
3442 // unconditional otherwise.
3443 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
3444 // Check for constant integer values in the select.
3445 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3446 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3447 if (!TrueVal || !FalseVal)
3450 // Find the relevant condition and destinations.
3451 Value *Condition = Select->getCondition();
3452 BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3453 BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3455 // Get weight for TrueBB and FalseBB.
3456 uint32_t TrueWeight = 0, FalseWeight = 0;
3457 SmallVector<uint64_t, 8> Weights;
3458 bool HasWeights = HasBranchWeights(SI);
3460 GetBranchWeights(SI, Weights);
3461 if (Weights.size() == 1 + SI->getNumCases()) {
3463 (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3465 (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3469 // Perform the actual simplification.
3470 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3475 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3476 // blockaddress(@fn, BlockB)))
3478 // (br cond, BlockA, BlockB).
3479 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
3480 // Check that both operands of the select are block addresses.
3481 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
3482 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
3486 // Extract the actual blocks.
3487 BasicBlock *TrueBB = TBA->getBasicBlock();
3488 BasicBlock *FalseBB = FBA->getBasicBlock();
3490 // Perform the actual simplification.
3491 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3495 /// This is called when we find an icmp instruction
3496 /// (a seteq/setne with a constant) as the only instruction in a
3497 /// block that ends with an uncond branch. We are looking for a very specific
3498 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3499 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3500 /// default value goes to an uncond block with a seteq in it, we get something
3503 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3505 /// %tmp = icmp eq i8 %A, 92
3508 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3510 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3511 /// the PHI, merging the third icmp into the switch.
3512 static bool tryToSimplifyUncondBranchWithICmpInIt(
3513 ICmpInst *ICI, IRBuilder<> &Builder, const DataLayout &DL,
3514 const TargetTransformInfo &TTI, const SimplifyCFGOptions &Options) {
3515 BasicBlock *BB = ICI->getParent();
3517 // If the block has any PHIs in it or the icmp has multiple uses, it is too
3519 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3522 Value *V = ICI->getOperand(0);
3523 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3525 // The pattern we're looking for is where our only predecessor is a switch on
3526 // 'V' and this block is the default case for the switch. In this case we can
3527 // fold the compared value into the switch to simplify things.
3528 BasicBlock *Pred = BB->getSinglePredecessor();
3529 if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3532 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3533 if (SI->getCondition() != V)
3536 // If BB is reachable on a non-default case, then we simply know the value of
3537 // V in this block. Substitute it and constant fold the icmp instruction
3539 if (SI->getDefaultDest() != BB) {
3540 ConstantInt *VVal = SI->findCaseDest(BB);
3541 assert(VVal && "Should have a unique destination value");
3542 ICI->setOperand(0, VVal);
3544 if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
3545 ICI->replaceAllUsesWith(V);
3546 ICI->eraseFromParent();
3548 // BB is now empty, so it is likely to simplify away.
3549 return simplifyCFG(BB, TTI, Options) | true;
3552 // Ok, the block is reachable from the default dest. If the constant we're
3553 // comparing exists in one of the other edges, then we can constant fold ICI
3555 if (SI->findCaseValue(Cst) != SI->case_default()) {
3557 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3558 V = ConstantInt::getFalse(BB->getContext());
3560 V = ConstantInt::getTrue(BB->getContext());
3562 ICI->replaceAllUsesWith(V);
3563 ICI->eraseFromParent();
3564 // BB is now empty, so it is likely to simplify away.
3565 return simplifyCFG(BB, TTI, Options) | true;
3568 // The use of the icmp has to be in the 'end' block, by the only PHI node in
3570 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3571 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3572 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3573 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3576 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3578 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3579 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3581 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3582 std::swap(DefaultCst, NewCst);
3584 // Replace ICI (which is used by the PHI for the default value) with true or
3585 // false depending on if it is EQ or NE.
3586 ICI->replaceAllUsesWith(DefaultCst);
3587 ICI->eraseFromParent();
3589 // Okay, the switch goes to this block on a default value. Add an edge from
3590 // the switch to the merge point on the compared value.
3592 BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3593 SmallVector<uint64_t, 8> Weights;
3594 bool HasWeights = HasBranchWeights(SI);
3596 GetBranchWeights(SI, Weights);
3597 if (Weights.size() == 1 + SI->getNumCases()) {
3598 // Split weight for default case to case for "Cst".
3599 Weights[0] = (Weights[0] + 1) >> 1;
3600 Weights.push_back(Weights[0]);
3602 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3603 setBranchWeights(SI, MDWeights);
3606 SI->addCase(Cst, NewBB);
3608 // NewBB branches to the phi block, add the uncond branch and the phi entry.
3609 Builder.SetInsertPoint(NewBB);
3610 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3611 Builder.CreateBr(SuccBlock);
3612 PHIUse->addIncoming(NewCst, NewBB);
3616 /// The specified branch is a conditional branch.
3617 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3618 /// fold it into a switch instruction if so.
3619 static bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
3620 const DataLayout &DL) {
3621 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3625 // Change br (X == 0 | X == 1), T, F into a switch instruction.
3626 // If this is a bunch of seteq's or'd together, or if it's a bunch of
3627 // 'setne's and'ed together, collect them.
3629 // Try to gather values from a chain of and/or to be turned into a switch
3630 ConstantComparesGatherer ConstantCompare(Cond, DL);
3631 // Unpack the result
3632 SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3633 Value *CompVal = ConstantCompare.CompValue;
3634 unsigned UsedICmps = ConstantCompare.UsedICmps;
3635 Value *ExtraCase = ConstantCompare.Extra;
3637 // If we didn't have a multiply compared value, fail.
3641 // Avoid turning single icmps into a switch.
3645 bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3647 // There might be duplicate constants in the list, which the switch
3648 // instruction can't handle, remove them now.
3649 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3650 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3652 // If Extra was used, we require at least two switch values to do the
3653 // transformation. A switch with one value is just a conditional branch.
3654 if (ExtraCase && Values.size() < 2)
3657 // TODO: Preserve branch weight metadata, similarly to how
3658 // FoldValueComparisonIntoPredecessors preserves it.
3660 // Figure out which block is which destination.
3661 BasicBlock *DefaultBB = BI->getSuccessor(1);
3662 BasicBlock *EdgeBB = BI->getSuccessor(0);
3664 std::swap(DefaultBB, EdgeBB);
3666 BasicBlock *BB = BI->getParent();
3668 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3669 << " cases into SWITCH. BB is:\n"
3672 // If there are any extra values that couldn't be folded into the switch
3673 // then we evaluate them with an explicit branch first. Split the block
3674 // right before the condbr to handle it.
3677 BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3678 // Remove the uncond branch added to the old block.
3679 TerminatorInst *OldTI = BB->getTerminator();
3680 Builder.SetInsertPoint(OldTI);
3683 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3685 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3687 OldTI->eraseFromParent();
3689 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3690 // for the edge we just added.
3691 AddPredecessorToBlock(EdgeBB, BB, NewBB);
3693 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3694 << "\nEXTRABB = " << *BB);
3698 Builder.SetInsertPoint(BI);
3699 // Convert pointer to int before we switch.
3700 if (CompVal->getType()->isPointerTy()) {
3701 CompVal = Builder.CreatePtrToInt(
3702 CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3705 // Create the new switch instruction now.
3706 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3708 // Add all of the 'cases' to the switch instruction.
3709 for (unsigned i = 0, e = Values.size(); i != e; ++i)
3710 New->addCase(Values[i], EdgeBB);
3712 // We added edges from PI to the EdgeBB. As such, if there were any
3713 // PHI nodes in EdgeBB, they need entries to be added corresponding to
3714 // the number of edges added.
3715 for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3716 PHINode *PN = cast<PHINode>(BBI);
3717 Value *InVal = PN->getIncomingValueForBlock(BB);
3718 for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3719 PN->addIncoming(InVal, BB);
3722 // Erase the old branch instruction.
3723 EraseTerminatorInstAndDCECond(BI);
3725 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3729 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3730 if (isa<PHINode>(RI->getValue()))
3731 return SimplifyCommonResume(RI);
3732 else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3733 RI->getValue() == RI->getParent()->getFirstNonPHI())
3734 // The resume must unwind the exception that caused control to branch here.
3735 return SimplifySingleResume(RI);
3740 // Simplify resume that is shared by several landing pads (phi of landing pad).
3741 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3742 BasicBlock *BB = RI->getParent();
3744 // Check that there are no other instructions except for debug intrinsics
3745 // between the phi of landing pads (RI->getValue()) and resume instruction.
3746 BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3747 E = RI->getIterator();
3749 if (!isa<DbgInfoIntrinsic>(I))
3752 SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
3753 auto *PhiLPInst = cast<PHINode>(RI->getValue());
3755 // Check incoming blocks to see if any of them are trivial.
3756 for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3758 auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3759 auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3761 // If the block has other successors, we can not delete it because
3762 // it has other dependents.
3763 if (IncomingBB->getUniqueSuccessor() != BB)
3766 auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3767 // Not the landing pad that caused the control to branch here.
3768 if (IncomingValue != LandingPad)
3771 bool isTrivial = true;
3773 I = IncomingBB->getFirstNonPHI()->getIterator();
3774 E = IncomingBB->getTerminator()->getIterator();
3776 if (!isa<DbgInfoIntrinsic>(I)) {
3782 TrivialUnwindBlocks.insert(IncomingBB);
3785 // If no trivial unwind blocks, don't do any simplifications.
3786 if (TrivialUnwindBlocks.empty())
3789 // Turn all invokes that unwind here into calls.
3790 for (auto *TrivialBB : TrivialUnwindBlocks) {
3791 // Blocks that will be simplified should be removed from the phi node.
3792 // Note there could be multiple edges to the resume block, and we need
3793 // to remove them all.
3794 while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3795 BB->removePredecessor(TrivialBB, true);
3797 for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3799 BasicBlock *Pred = *PI++;
3800 removeUnwindEdge(Pred);
3803 // In each SimplifyCFG run, only the current processed block can be erased.
3804 // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3805 // of erasing TrivialBB, we only remove the branch to the common resume
3806 // block so that we can later erase the resume block since it has no
3808 TrivialBB->getTerminator()->eraseFromParent();
3809 new UnreachableInst(RI->getContext(), TrivialBB);
3812 // Delete the resume block if all its predecessors have been removed.
3814 BB->eraseFromParent();
3816 return !TrivialUnwindBlocks.empty();
3819 // Simplify resume that is only used by a single (non-phi) landing pad.
3820 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3821 BasicBlock *BB = RI->getParent();
3822 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
3823 assert(RI->getValue() == LPInst &&
3824 "Resume must unwind the exception that caused control to here");
3826 // Check that there are no other instructions except for debug intrinsics.
3827 BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3829 if (!isa<DbgInfoIntrinsic>(I))
3832 // Turn all invokes that unwind here into calls and delete the basic block.
3833 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3834 BasicBlock *Pred = *PI++;
3835 removeUnwindEdge(Pred);
3838 // The landingpad is now unreachable. Zap it.
3839 BB->eraseFromParent();
3841 LoopHeaders->erase(BB);
3845 static bool removeEmptyCleanup(CleanupReturnInst *RI) {
3846 // If this is a trivial cleanup pad that executes no instructions, it can be
3847 // eliminated. If the cleanup pad continues to the caller, any predecessor
3848 // that is an EH pad will be updated to continue to the caller and any
3849 // predecessor that terminates with an invoke instruction will have its invoke
3850 // instruction converted to a call instruction. If the cleanup pad being
3851 // simplified does not continue to the caller, each predecessor will be
3852 // updated to continue to the unwind destination of the cleanup pad being
3854 BasicBlock *BB = RI->getParent();
3855 CleanupPadInst *CPInst = RI->getCleanupPad();
3856 if (CPInst->getParent() != BB)
3857 // This isn't an empty cleanup.
3860 // We cannot kill the pad if it has multiple uses. This typically arises
3861 // from unreachable basic blocks.
3862 if (!CPInst->hasOneUse())
3865 // Check that there are no other instructions except for benign intrinsics.
3866 BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3868 auto *II = dyn_cast<IntrinsicInst>(I);
3872 Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3873 switch (IntrinsicID) {
3874 case Intrinsic::dbg_declare:
3875 case Intrinsic::dbg_value:
3876 case Intrinsic::lifetime_end:
3883 // If the cleanup return we are simplifying unwinds to the caller, this will
3884 // set UnwindDest to nullptr.
3885 BasicBlock *UnwindDest = RI->getUnwindDest();
3886 Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3888 // We're about to remove BB from the control flow. Before we do, sink any
3889 // PHINodes into the unwind destination. Doing this before changing the
3890 // control flow avoids some potentially slow checks, since we can currently
3891 // be certain that UnwindDest and BB have no common predecessors (since they
3892 // are both EH pads).
3894 // First, go through the PHI nodes in UnwindDest and update any nodes that
3895 // reference the block we are removing
3896 for (BasicBlock::iterator I = UnwindDest->begin(),
3897 IE = DestEHPad->getIterator();
3899 PHINode *DestPN = cast<PHINode>(I);
3901 int Idx = DestPN->getBasicBlockIndex(BB);
3902 // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3904 // This PHI node has an incoming value that corresponds to a control
3905 // path through the cleanup pad we are removing. If the incoming
3906 // value is in the cleanup pad, it must be a PHINode (because we
3907 // verified above that the block is otherwise empty). Otherwise, the
3908 // value is either a constant or a value that dominates the cleanup
3909 // pad being removed.
3911 // Because BB and UnwindDest are both EH pads, all of their
3912 // predecessors must unwind to these blocks, and since no instruction
3913 // can have multiple unwind destinations, there will be no overlap in
3914 // incoming blocks between SrcPN and DestPN.
3915 Value *SrcVal = DestPN->getIncomingValue(Idx);
3916 PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3918 // Remove the entry for the block we are deleting.
3919 DestPN->removeIncomingValue(Idx, false);
3921 if (SrcPN && SrcPN->getParent() == BB) {
3922 // If the incoming value was a PHI node in the cleanup pad we are
3923 // removing, we need to merge that PHI node's incoming values into
3925 for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3926 SrcIdx != SrcE; ++SrcIdx) {
3927 DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3928 SrcPN->getIncomingBlock(SrcIdx));
3931 // Otherwise, the incoming value came from above BB and
3932 // so we can just reuse it. We must associate all of BB's
3933 // predecessors with this value.
3934 for (auto *pred : predecessors(BB)) {
3935 DestPN->addIncoming(SrcVal, pred);
3940 // Sink any remaining PHI nodes directly into UnwindDest.
3941 Instruction *InsertPt = DestEHPad;
3942 for (BasicBlock::iterator I = BB->begin(),
3943 IE = BB->getFirstNonPHI()->getIterator();
3945 // The iterator must be incremented here because the instructions are
3946 // being moved to another block.
3947 PHINode *PN = cast<PHINode>(I++);
3948 if (PN->use_empty())
3949 // If the PHI node has no uses, just leave it. It will be erased
3950 // when we erase BB below.
3953 // Otherwise, sink this PHI node into UnwindDest.
3954 // Any predecessors to UnwindDest which are not already represented
3955 // must be back edges which inherit the value from the path through
3956 // BB. In this case, the PHI value must reference itself.
3957 for (auto *pred : predecessors(UnwindDest))
3959 PN->addIncoming(PN, pred);
3960 PN->moveBefore(InsertPt);
3964 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3965 // The iterator must be updated here because we are removing this pred.
3966 BasicBlock *PredBB = *PI++;
3967 if (UnwindDest == nullptr) {
3968 removeUnwindEdge(PredBB);
3970 TerminatorInst *TI = PredBB->getTerminator();
3971 TI->replaceUsesOfWith(BB, UnwindDest);
3975 // The cleanup pad is now unreachable. Zap it.
3976 BB->eraseFromParent();
3980 // Try to merge two cleanuppads together.
3981 static bool mergeCleanupPad(CleanupReturnInst *RI) {
3982 // Skip any cleanuprets which unwind to caller, there is nothing to merge
3984 BasicBlock *UnwindDest = RI->getUnwindDest();
3988 // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
3989 // be safe to merge without code duplication.
3990 if (UnwindDest->getSinglePredecessor() != RI->getParent())
3993 // Verify that our cleanuppad's unwind destination is another cleanuppad.
3994 auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
3995 if (!SuccessorCleanupPad)
3998 CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
3999 // Replace any uses of the successor cleanupad with the predecessor pad
4000 // The only cleanuppad uses should be this cleanupret, it's cleanupret and
4001 // funclet bundle operands.
4002 SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4003 // Remove the old cleanuppad.
4004 SuccessorCleanupPad->eraseFromParent();
4005 // Now, we simply replace the cleanupret with a branch to the unwind
4007 BranchInst::Create(UnwindDest, RI->getParent());
4008 RI->eraseFromParent();
4013 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
4014 // It is possible to transiantly have an undef cleanuppad operand because we
4015 // have deleted some, but not all, dead blocks.
4016 // Eventually, this block will be deleted.
4017 if (isa<UndefValue>(RI->getOperand(0)))
4020 if (mergeCleanupPad(RI))
4023 if (removeEmptyCleanup(RI))
4029 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4030 BasicBlock *BB = RI->getParent();
4031 if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4034 // Find predecessors that end with branches.
4035 SmallVector<BasicBlock *, 8> UncondBranchPreds;
4036 SmallVector<BranchInst *, 8> CondBranchPreds;
4037 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4038 BasicBlock *P = *PI;
4039 TerminatorInst *PTI = P->getTerminator();
4040 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4041 if (BI->isUnconditional())
4042 UncondBranchPreds.push_back(P);
4044 CondBranchPreds.push_back(BI);
4048 // If we found some, do the transformation!
4049 if (!UncondBranchPreds.empty() && DupRet) {
4050 while (!UncondBranchPreds.empty()) {
4051 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4052 DEBUG(dbgs() << "FOLDING: " << *BB
4053 << "INTO UNCOND BRANCH PRED: " << *Pred);
4054 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4057 // If we eliminated all predecessors of the block, delete the block now.
4058 if (pred_empty(BB)) {
4059 // We know there are no successors, so just nuke the block.
4060 BB->eraseFromParent();
4062 LoopHeaders->erase(BB);
4068 // Check out all of the conditional branches going to this return
4069 // instruction. If any of them just select between returns, change the
4070 // branch itself into a select/return pair.
4071 while (!CondBranchPreds.empty()) {
4072 BranchInst *BI = CondBranchPreds.pop_back_val();
4074 // Check to see if the non-BB successor is also a return block.
4075 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4076 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4077 SimplifyCondBranchToTwoReturns(BI, Builder))
4083 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4084 BasicBlock *BB = UI->getParent();
4086 bool Changed = false;
4088 // If there are any instructions immediately before the unreachable that can
4089 // be removed, do so.
4090 while (UI->getIterator() != BB->begin()) {
4091 BasicBlock::iterator BBI = UI->getIterator();
4093 // Do not delete instructions that can have side effects which might cause
4094 // the unreachable to not be reachable; specifically, calls and volatile
4095 // operations may have this effect.
4096 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4099 if (BBI->mayHaveSideEffects()) {
4100 if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4101 if (SI->isVolatile())
4103 } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4104 if (LI->isVolatile())
4106 } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4107 if (RMWI->isVolatile())
4109 } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4110 if (CXI->isVolatile())
4112 } else if (isa<CatchPadInst>(BBI)) {
4113 // A catchpad may invoke exception object constructors and such, which
4114 // in some languages can be arbitrary code, so be conservative by
4116 // For CoreCLR, it just involves a type test, so can be removed.
4117 if (classifyEHPersonality(BB->getParent()->getPersonalityFn()) !=
4118 EHPersonality::CoreCLR)
4120 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4121 !isa<LandingPadInst>(BBI)) {
4124 // Note that deleting LandingPad's here is in fact okay, although it
4125 // involves a bit of subtle reasoning. If this inst is a LandingPad,
4126 // all the predecessors of this block will be the unwind edges of Invokes,
4127 // and we can therefore guarantee this block will be erased.
4130 // Delete this instruction (any uses are guaranteed to be dead)
4131 if (!BBI->use_empty())
4132 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4133 BBI->eraseFromParent();
4137 // If the unreachable instruction is the first in the block, take a gander
4138 // at all of the predecessors of this instruction, and simplify them.
4139 if (&BB->front() != UI)
4142 SmallVector<BasicBlock *, 8> Preds(pred_begin(BB), pred_end(BB));
4143 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4144 TerminatorInst *TI = Preds[i]->getTerminator();
4145 IRBuilder<> Builder(TI);
4146 if (auto *BI = dyn_cast<BranchInst>(TI)) {
4147 if (BI->isUnconditional()) {
4148 if (BI->getSuccessor(0) == BB) {
4149 new UnreachableInst(TI->getContext(), TI);
4150 TI->eraseFromParent();
4154 if (BI->getSuccessor(0) == BB) {
4155 Builder.CreateBr(BI->getSuccessor(1));
4156 EraseTerminatorInstAndDCECond(BI);
4157 } else if (BI->getSuccessor(1) == BB) {
4158 Builder.CreateBr(BI->getSuccessor(0));
4159 EraseTerminatorInstAndDCECond(BI);
4163 } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4164 for (auto i = SI->case_begin(), e = SI->case_end(); i != e;) {
4165 if (i->getCaseSuccessor() != BB) {
4169 BB->removePredecessor(SI->getParent());
4170 i = SI->removeCase(i);
4174 } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4175 if (II->getUnwindDest() == BB) {
4176 removeUnwindEdge(TI->getParent());
4179 } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4180 if (CSI->getUnwindDest() == BB) {
4181 removeUnwindEdge(TI->getParent());
4186 for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4187 E = CSI->handler_end();
4190 CSI->removeHandler(I);
4196 if (CSI->getNumHandlers() == 0) {
4197 BasicBlock *CatchSwitchBB = CSI->getParent();
4198 if (CSI->hasUnwindDest()) {
4199 // Redirect preds to the unwind dest
4200 CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4202 // Rewrite all preds to unwind to caller (or from invoke to call).
4203 SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4204 for (BasicBlock *EHPred : EHPreds)
4205 removeUnwindEdge(EHPred);
4207 // The catchswitch is no longer reachable.
4208 new UnreachableInst(CSI->getContext(), CSI);
4209 CSI->eraseFromParent();
4212 } else if (isa<CleanupReturnInst>(TI)) {
4213 new UnreachableInst(TI->getContext(), TI);
4214 TI->eraseFromParent();
4219 // If this block is now dead, remove it.
4220 if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4221 // We know there are no successors, so just nuke the block.
4222 BB->eraseFromParent();
4224 LoopHeaders->erase(BB);
4231 static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) {
4232 assert(Cases.size() >= 1);
4234 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4235 for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4236 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4242 /// Turn a switch with two reachable destinations into an integer range
4243 /// comparison and branch.
4244 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
4245 assert(SI->getNumCases() > 1 && "Degenerate switch?");
4248 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4250 // Partition the cases into two sets with different destinations.
4251 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4252 BasicBlock *DestB = nullptr;
4253 SmallVector<ConstantInt *, 16> CasesA;
4254 SmallVector<ConstantInt *, 16> CasesB;
4256 for (auto Case : SI->cases()) {
4257 BasicBlock *Dest = Case.getCaseSuccessor();
4260 if (Dest == DestA) {
4261 CasesA.push_back(Case.getCaseValue());
4266 if (Dest == DestB) {
4267 CasesB.push_back(Case.getCaseValue());
4270 return false; // More than two destinations.
4273 assert(DestA && DestB &&
4274 "Single-destination switch should have been folded.");
4275 assert(DestA != DestB);
4276 assert(DestB != SI->getDefaultDest());
4277 assert(!CasesB.empty() && "There must be non-default cases.");
4278 assert(!CasesA.empty() || HasDefault);
4280 // Figure out if one of the sets of cases form a contiguous range.
4281 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4282 BasicBlock *ContiguousDest = nullptr;
4283 BasicBlock *OtherDest = nullptr;
4284 if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4285 ContiguousCases = &CasesA;
4286 ContiguousDest = DestA;
4288 } else if (CasesAreContiguous(CasesB)) {
4289 ContiguousCases = &CasesB;
4290 ContiguousDest = DestB;
4295 // Start building the compare and branch.
4297 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4298 Constant *NumCases =
4299 ConstantInt::get(Offset->getType(), ContiguousCases->size());
4301 Value *Sub = SI->getCondition();
4302 if (!Offset->isNullValue())
4303 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4306 // If NumCases overflowed, then all possible values jump to the successor.
4307 if (NumCases->isNullValue() && !ContiguousCases->empty())
4308 Cmp = ConstantInt::getTrue(SI->getContext());
4310 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4311 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4313 // Update weight for the newly-created conditional branch.
4314 if (HasBranchWeights(SI)) {
4315 SmallVector<uint64_t, 8> Weights;
4316 GetBranchWeights(SI, Weights);
4317 if (Weights.size() == 1 + SI->getNumCases()) {
4318 uint64_t TrueWeight = 0;
4319 uint64_t FalseWeight = 0;
4320 for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4321 if (SI->getSuccessor(I) == ContiguousDest)
4322 TrueWeight += Weights[I];
4324 FalseWeight += Weights[I];
4326 while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4330 setBranchWeights(NewBI, TrueWeight, FalseWeight);
4334 // Prune obsolete incoming values off the successors' PHI nodes.
4335 for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4336 unsigned PreviousEdges = ContiguousCases->size();
4337 if (ContiguousDest == SI->getDefaultDest())
4339 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4340 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4342 for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4343 unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4344 if (OtherDest == SI->getDefaultDest())
4346 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4347 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4351 SI->eraseFromParent();
4356 /// Compute masked bits for the condition of a switch
4357 /// and use it to remove dead cases.
4358 static bool eliminateDeadSwitchCases(SwitchInst *SI, AssumptionCache *AC,
4359 const DataLayout &DL) {
4360 Value *Cond = SI->getCondition();
4361 unsigned Bits = Cond->getType()->getIntegerBitWidth();
4362 KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4364 // We can also eliminate cases by determining that their values are outside of
4365 // the limited range of the condition based on how many significant (non-sign)
4366 // bits are in the condition value.
4367 unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4368 unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4370 // Gather dead cases.
4371 SmallVector<ConstantInt *, 8> DeadCases;
4372 for (auto &Case : SI->cases()) {
4373 const APInt &CaseVal = Case.getCaseValue()->getValue();
4374 if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4375 (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4376 DeadCases.push_back(Case.getCaseValue());
4377 DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal << " is dead.\n");
4381 // If we can prove that the cases must cover all possible values, the
4382 // default destination becomes dead and we can remove it. If we know some
4383 // of the bits in the value, we can use that to more precisely compute the
4384 // number of possible unique case values.
4386 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4387 const unsigned NumUnknownBits =
4388 Bits - (Known.Zero | Known.One).countPopulation();
4389 assert(NumUnknownBits <= Bits);
4390 if (HasDefault && DeadCases.empty() &&
4391 NumUnknownBits < 64 /* avoid overflow */ &&
4392 SI->getNumCases() == (1ULL << NumUnknownBits)) {
4393 DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4394 BasicBlock *NewDefault =
4395 SplitBlockPredecessors(SI->getDefaultDest(), SI->getParent(), "");
4396 SI->setDefaultDest(&*NewDefault);
4397 SplitBlock(&*NewDefault, &NewDefault->front());
4398 auto *OldTI = NewDefault->getTerminator();
4399 new UnreachableInst(SI->getContext(), OldTI);
4400 EraseTerminatorInstAndDCECond(OldTI);
4404 SmallVector<uint64_t, 8> Weights;
4405 bool HasWeight = HasBranchWeights(SI);
4407 GetBranchWeights(SI, Weights);
4408 HasWeight = (Weights.size() == 1 + SI->getNumCases());
4411 // Remove dead cases from the switch.
4412 for (ConstantInt *DeadCase : DeadCases) {
4413 SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4414 assert(CaseI != SI->case_default() &&
4415 "Case was not found. Probably mistake in DeadCases forming.");
4417 std::swap(Weights[CaseI->getCaseIndex() + 1], Weights.back());
4421 // Prune unused values from PHI nodes.
4422 CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4423 SI->removeCase(CaseI);
4425 if (HasWeight && Weights.size() >= 2) {
4426 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
4427 setBranchWeights(SI, MDWeights);
4430 return !DeadCases.empty();
4433 /// If BB would be eligible for simplification by
4434 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4435 /// by an unconditional branch), look at the phi node for BB in the successor
4436 /// block and see if the incoming value is equal to CaseValue. If so, return
4437 /// the phi node, and set PhiIndex to BB's index in the phi node.
4438 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
4439 BasicBlock *BB, int *PhiIndex) {
4440 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4441 return nullptr; // BB must be empty to be a candidate for simplification.
4442 if (!BB->getSinglePredecessor())
4443 return nullptr; // BB must be dominated by the switch.
4445 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
4446 if (!Branch || !Branch->isUnconditional())
4447 return nullptr; // Terminator must be unconditional branch.
4449 BasicBlock *Succ = Branch->getSuccessor(0);
4451 for (PHINode &PHI : Succ->phis()) {
4452 int Idx = PHI.getBasicBlockIndex(BB);
4453 assert(Idx >= 0 && "PHI has no entry for predecessor?");
4455 Value *InValue = PHI.getIncomingValue(Idx);
4456 if (InValue != CaseValue)
4466 /// Try to forward the condition of a switch instruction to a phi node
4467 /// dominated by the switch, if that would mean that some of the destination
4468 /// blocks of the switch can be folded away. Return true if a change is made.
4469 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
4470 using ForwardingNodesMap = DenseMap<PHINode *, SmallVector<int, 4>>;
4472 ForwardingNodesMap ForwardingNodes;
4473 BasicBlock *SwitchBlock = SI->getParent();
4474 bool Changed = false;
4475 for (auto &Case : SI->cases()) {
4476 ConstantInt *CaseValue = Case.getCaseValue();
4477 BasicBlock *CaseDest = Case.getCaseSuccessor();
4479 // Replace phi operands in successor blocks that are using the constant case
4480 // value rather than the switch condition variable:
4482 // switch i32 %x, label %default [
4483 // i32 17, label %succ
4486 // %r = phi i32 ... [ 17, %switchbb ] ...
4488 // %r = phi i32 ... [ %x, %switchbb ] ...
4490 for (PHINode &Phi : CaseDest->phis()) {
4491 // This only works if there is exactly 1 incoming edge from the switch to
4492 // a phi. If there is >1, that means multiple cases of the switch map to 1
4493 // value in the phi, and that phi value is not the switch condition. Thus,
4494 // this transform would not make sense (the phi would be invalid because
4495 // a phi can't have different incoming values from the same block).
4496 int SwitchBBIdx = Phi.getBasicBlockIndex(SwitchBlock);
4497 if (Phi.getIncomingValue(SwitchBBIdx) == CaseValue &&
4498 count(Phi.blocks(), SwitchBlock) == 1) {
4499 Phi.setIncomingValue(SwitchBBIdx, SI->getCondition());
4504 // Collect phi nodes that are indirectly using this switch's case constants.
4506 if (auto *Phi = FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIdx))
4507 ForwardingNodes[Phi].push_back(PhiIdx);
4510 for (auto &ForwardingNode : ForwardingNodes) {
4511 PHINode *Phi = ForwardingNode.first;
4512 SmallVectorImpl<int> &Indexes = ForwardingNode.second;
4513 if (Indexes.size() < 2)
4516 for (int Index : Indexes)
4517 Phi->setIncomingValue(Index, SI->getCondition());
4524 /// Return true if the backend will be able to handle
4525 /// initializing an array of constants like C.
4526 static bool ValidLookupTableConstant(Constant *C, const TargetTransformInfo &TTI) {
4527 if (C->isThreadDependent())
4529 if (C->isDLLImportDependent())
4532 if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4533 !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4534 !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4537 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4538 if (!CE->isGEPWithNoNotionalOverIndexing())
4540 if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4544 if (!TTI.shouldBuildLookupTablesForConstant(C))
4550 /// If V is a Constant, return it. Otherwise, try to look up
4551 /// its constant value in ConstantPool, returning 0 if it's not there.
4553 LookupConstant(Value *V,
4554 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4555 if (Constant *C = dyn_cast<Constant>(V))
4557 return ConstantPool.lookup(V);
4560 /// Try to fold instruction I into a constant. This works for
4561 /// simple instructions such as binary operations where both operands are
4562 /// constant or can be replaced by constants from the ConstantPool. Returns the
4563 /// resulting constant on success, 0 otherwise.
4565 ConstantFold(Instruction *I, const DataLayout &DL,
4566 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4567 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4568 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4571 if (A->isAllOnesValue())
4572 return LookupConstant(Select->getTrueValue(), ConstantPool);
4573 if (A->isNullValue())
4574 return LookupConstant(Select->getFalseValue(), ConstantPool);
4578 SmallVector<Constant *, 4> COps;
4579 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4580 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
4586 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4587 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4591 return ConstantFoldInstOperands(I, COps, DL);
4594 /// Try to determine the resulting constant values in phi nodes
4595 /// at the common destination basic block, *CommonDest, for one of the case
4596 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4597 /// case), of a switch instruction SI.
4599 GetCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest,
4600 BasicBlock **CommonDest,
4601 SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4602 const DataLayout &DL, const TargetTransformInfo &TTI) {
4603 // The block from which we enter the common destination.
4604 BasicBlock *Pred = SI->getParent();
4606 // If CaseDest is empty except for some side-effect free instructions through
4607 // which we can constant-propagate the CaseVal, continue to its successor.
4608 SmallDenseMap<Value *, Constant *> ConstantPool;
4609 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4610 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
4612 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
4613 // If the terminator is a simple branch, continue to the next block.
4614 if (T->getNumSuccessors() != 1 || T->isExceptional())
4617 CaseDest = T->getSuccessor(0);
4618 } else if (isa<DbgInfoIntrinsic>(I)) {
4619 // Skip debug intrinsic.
4621 } else if (Constant *C = ConstantFold(&*I, DL, ConstantPool)) {
4622 // Instruction is side-effect free and constant.
4624 // If the instruction has uses outside this block or a phi node slot for
4625 // the block, it is not safe to bypass the instruction since it would then
4626 // no longer dominate all its uses.
4627 for (auto &Use : I->uses()) {
4628 User *User = Use.getUser();
4629 if (Instruction *I = dyn_cast<Instruction>(User))
4630 if (I->getParent() == CaseDest)
4632 if (PHINode *Phi = dyn_cast<PHINode>(User))
4633 if (Phi->getIncomingBlock(Use) == CaseDest)
4638 ConstantPool.insert(std::make_pair(&*I, C));
4644 // If we did not have a CommonDest before, use the current one.
4646 *CommonDest = CaseDest;
4647 // If the destination isn't the common one, abort.
4648 if (CaseDest != *CommonDest)
4651 // Get the values for this case from phi nodes in the destination block.
4652 for (PHINode &PHI : (*CommonDest)->phis()) {
4653 int Idx = PHI.getBasicBlockIndex(Pred);
4657 Constant *ConstVal =
4658 LookupConstant(PHI.getIncomingValue(Idx), ConstantPool);
4662 // Be conservative about which kinds of constants we support.
4663 if (!ValidLookupTableConstant(ConstVal, TTI))
4666 Res.push_back(std::make_pair(&PHI, ConstVal));
4669 return Res.size() > 0;
4672 // Helper function used to add CaseVal to the list of cases that generate
4674 static void MapCaseToResult(ConstantInt *CaseVal,
4675 SwitchCaseResultVectorTy &UniqueResults,
4677 for (auto &I : UniqueResults) {
4678 if (I.first == Result) {
4679 I.second.push_back(CaseVal);
4683 UniqueResults.push_back(
4684 std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4687 // Helper function that initializes a map containing
4688 // results for the PHI node of the common destination block for a switch
4689 // instruction. Returns false if multiple PHI nodes have been found or if
4690 // there is not a common destination block for the switch.
4691 static bool InitializeUniqueCases(SwitchInst *SI, PHINode *&PHI,
4692 BasicBlock *&CommonDest,
4693 SwitchCaseResultVectorTy &UniqueResults,
4694 Constant *&DefaultResult,
4695 const DataLayout &DL,
4696 const TargetTransformInfo &TTI) {
4697 for (auto &I : SI->cases()) {
4698 ConstantInt *CaseVal = I.getCaseValue();
4700 // Resulting value at phi nodes for this case value.
4701 SwitchCaseResultsTy Results;
4702 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4706 // Only one value per case is permitted
4707 if (Results.size() > 1)
4709 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4711 // Check the PHI consistency.
4713 PHI = Results[0].first;
4714 else if (PHI != Results[0].first)
4717 // Find the default result value.
4718 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
4719 BasicBlock *DefaultDest = SI->getDefaultDest();
4720 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4722 // If the default value is not found abort unless the default destination
4725 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4726 if ((!DefaultResult &&
4727 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4733 // Helper function that checks if it is possible to transform a switch with only
4734 // two cases (or two cases + default) that produces a result into a select.
4737 // case 10: %0 = icmp eq i32 %a, 10
4738 // return 10; %1 = select i1 %0, i32 10, i32 4
4739 // case 20: ----> %2 = icmp eq i32 %a, 20
4740 // return 2; %3 = select i1 %2, i32 2, i32 %1
4744 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4745 Constant *DefaultResult, Value *Condition,
4746 IRBuilder<> &Builder) {
4747 assert(ResultVector.size() == 2 &&
4748 "We should have exactly two unique results at this point");
4749 // If we are selecting between only two cases transform into a simple
4750 // select or a two-way select if default is possible.
4751 if (ResultVector[0].second.size() == 1 &&
4752 ResultVector[1].second.size() == 1) {
4753 ConstantInt *const FirstCase = ResultVector[0].second[0];
4754 ConstantInt *const SecondCase = ResultVector[1].second[0];
4756 bool DefaultCanTrigger = DefaultResult;
4757 Value *SelectValue = ResultVector[1].first;
4758 if (DefaultCanTrigger) {
4759 Value *const ValueCompare =
4760 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4761 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4762 DefaultResult, "switch.select");
4764 Value *const ValueCompare =
4765 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4766 return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4767 SelectValue, "switch.select");
4773 // Helper function to cleanup a switch instruction that has been converted into
4774 // a select, fixing up PHI nodes and basic blocks.
4775 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
4777 IRBuilder<> &Builder) {
4778 BasicBlock *SelectBB = SI->getParent();
4779 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4780 PHI->removeIncomingValue(SelectBB);
4781 PHI->addIncoming(SelectValue, SelectBB);
4783 Builder.CreateBr(PHI->getParent());
4785 // Remove the switch.
4786 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4787 BasicBlock *Succ = SI->getSuccessor(i);
4789 if (Succ == PHI->getParent())
4791 Succ->removePredecessor(SelectBB);
4793 SI->eraseFromParent();
4796 /// If the switch is only used to initialize one or more
4797 /// phi nodes in a common successor block with only two different
4798 /// constant values, replace the switch with select.
4799 static bool switchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4800 const DataLayout &DL,
4801 const TargetTransformInfo &TTI) {
4802 Value *const Cond = SI->getCondition();
4803 PHINode *PHI = nullptr;
4804 BasicBlock *CommonDest = nullptr;
4805 Constant *DefaultResult;
4806 SwitchCaseResultVectorTy UniqueResults;
4807 // Collect all the cases that will deliver the same value from the switch.
4808 if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4811 // Selects choose between maximum two values.
4812 if (UniqueResults.size() != 2)
4814 assert(PHI != nullptr && "PHI for value select not found");
4816 Builder.SetInsertPoint(SI);
4817 Value *SelectValue =
4818 ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4820 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4823 // The switch couldn't be converted into a select.
4829 /// This class represents a lookup table that can be used to replace a switch.
4830 class SwitchLookupTable {
4832 /// Create a lookup table to use as a switch replacement with the contents
4833 /// of Values, using DefaultValue to fill any holes in the table.
4835 Module &M, uint64_t TableSize, ConstantInt *Offset,
4836 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4837 Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName);
4839 /// Build instructions with Builder to retrieve the value at
4840 /// the position given by Index in the lookup table.
4841 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4843 /// Return true if a table with TableSize elements of
4844 /// type ElementType would fit in a target-legal register.
4845 static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4849 // Depending on the contents of the table, it can be represented in
4852 // For tables where each element contains the same value, we just have to
4853 // store that single value and return it for each lookup.
4856 // For tables where there is a linear relationship between table index
4857 // and values. We calculate the result with a simple multiplication
4858 // and addition instead of a table lookup.
4861 // For small tables with integer elements, we can pack them into a bitmap
4862 // that fits into a target-legal register. Values are retrieved by
4863 // shift and mask operations.
4866 // The table is stored as an array of values. Values are retrieved by load
4867 // instructions from the table.
4871 // For SingleValueKind, this is the single value.
4872 Constant *SingleValue = nullptr;
4874 // For BitMapKind, this is the bitmap.
4875 ConstantInt *BitMap = nullptr;
4876 IntegerType *BitMapElementTy = nullptr;
4878 // For LinearMapKind, these are the constants used to derive the value.
4879 ConstantInt *LinearOffset = nullptr;
4880 ConstantInt *LinearMultiplier = nullptr;
4882 // For ArrayKind, this is the array.
4883 GlobalVariable *Array = nullptr;
4886 } // end anonymous namespace
4888 SwitchLookupTable::SwitchLookupTable(
4889 Module &M, uint64_t TableSize, ConstantInt *Offset,
4890 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4891 Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName) {
4892 assert(Values.size() && "Can't build lookup table without values!");
4893 assert(TableSize >= Values.size() && "Can't fit values in table!");
4895 // If all values in the table are equal, this is that value.
4896 SingleValue = Values.begin()->second;
4898 Type *ValueType = Values.begin()->second->getType();
4900 // Build up the table contents.
4901 SmallVector<Constant *, 64> TableContents(TableSize);
4902 for (size_t I = 0, E = Values.size(); I != E; ++I) {
4903 ConstantInt *CaseVal = Values[I].first;
4904 Constant *CaseRes = Values[I].second;
4905 assert(CaseRes->getType() == ValueType);
4907 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4908 TableContents[Idx] = CaseRes;
4910 if (CaseRes != SingleValue)
4911 SingleValue = nullptr;
4914 // Fill in any holes in the table with the default result.
4915 if (Values.size() < TableSize) {
4916 assert(DefaultValue &&
4917 "Need a default value to fill the lookup table holes.");
4918 assert(DefaultValue->getType() == ValueType);
4919 for (uint64_t I = 0; I < TableSize; ++I) {
4920 if (!TableContents[I])
4921 TableContents[I] = DefaultValue;
4924 if (DefaultValue != SingleValue)
4925 SingleValue = nullptr;
4928 // If each element in the table contains the same value, we only need to store
4929 // that single value.
4931 Kind = SingleValueKind;
4935 // Check if we can derive the value with a linear transformation from the
4937 if (isa<IntegerType>(ValueType)) {
4938 bool LinearMappingPossible = true;
4941 assert(TableSize >= 2 && "Should be a SingleValue table.");
4942 // Check if there is the same distance between two consecutive values.
4943 for (uint64_t I = 0; I < TableSize; ++I) {
4944 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4946 // This is an undef. We could deal with it, but undefs in lookup tables
4947 // are very seldom. It's probably not worth the additional complexity.
4948 LinearMappingPossible = false;
4951 const APInt &Val = ConstVal->getValue();
4953 APInt Dist = Val - PrevVal;
4956 } else if (Dist != DistToPrev) {
4957 LinearMappingPossible = false;
4963 if (LinearMappingPossible) {
4964 LinearOffset = cast<ConstantInt>(TableContents[0]);
4965 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
4966 Kind = LinearMapKind;
4972 // If the type is integer and the table fits in a register, build a bitmap.
4973 if (WouldFitInRegister(DL, TableSize, ValueType)) {
4974 IntegerType *IT = cast<IntegerType>(ValueType);
4975 APInt TableInt(TableSize * IT->getBitWidth(), 0);
4976 for (uint64_t I = TableSize; I > 0; --I) {
4977 TableInt <<= IT->getBitWidth();
4978 // Insert values into the bitmap. Undef values are set to zero.
4979 if (!isa<UndefValue>(TableContents[I - 1])) {
4980 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
4981 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
4984 BitMap = ConstantInt::get(M.getContext(), TableInt);
4985 BitMapElementTy = IT;
4991 // Store the table in an array.
4992 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
4993 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
4995 Array = new GlobalVariable(M, ArrayTy, /*constant=*/true,
4996 GlobalVariable::PrivateLinkage, Initializer,
4997 "switch.table." + FuncName);
4998 Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
5002 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
5004 case SingleValueKind:
5006 case LinearMapKind: {
5007 // Derive the result value from the input value.
5008 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
5009 false, "switch.idx.cast");
5010 if (!LinearMultiplier->isOne())
5011 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
5012 if (!LinearOffset->isZero())
5013 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
5017 // Type of the bitmap (e.g. i59).
5018 IntegerType *MapTy = BitMap->getType();
5020 // Cast Index to the same type as the bitmap.
5021 // Note: The Index is <= the number of elements in the table, so
5022 // truncating it to the width of the bitmask is safe.
5023 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
5025 // Multiply the shift amount by the element width.
5026 ShiftAmt = Builder.CreateMul(
5027 ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5031 Value *DownShifted =
5032 Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5034 return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5037 // Make sure the table index will not overflow when treated as signed.
5038 IntegerType *IT = cast<IntegerType>(Index->getType());
5039 uint64_t TableSize =
5040 Array->getInitializer()->getType()->getArrayNumElements();
5041 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5042 Index = Builder.CreateZExt(
5043 Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5044 "switch.tableidx.zext");
5046 Value *GEPIndices[] = {Builder.getInt32(0), Index};
5047 Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5048 GEPIndices, "switch.gep");
5049 return Builder.CreateLoad(GEP, "switch.load");
5052 llvm_unreachable("Unknown lookup table kind!");
5055 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5057 Type *ElementType) {
5058 auto *IT = dyn_cast<IntegerType>(ElementType);
5061 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5062 // are <= 15, we could try to narrow the type.
5064 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5065 if (TableSize >= UINT_MAX / IT->getBitWidth())
5067 return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5070 /// Determine whether a lookup table should be built for this switch, based on
5071 /// the number of cases, size of the table, and the types of the results.
5073 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5074 const TargetTransformInfo &TTI, const DataLayout &DL,
5075 const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5076 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5077 return false; // TableSize overflowed, or mul below might overflow.
5079 bool AllTablesFitInRegister = true;
5080 bool HasIllegalType = false;
5081 for (const auto &I : ResultTypes) {
5082 Type *Ty = I.second;
5084 // Saturate this flag to true.
5085 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5087 // Saturate this flag to false.
5088 AllTablesFitInRegister =
5089 AllTablesFitInRegister &&
5090 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5092 // If both flags saturate, we're done. NOTE: This *only* works with
5093 // saturating flags, and all flags have to saturate first due to the
5094 // non-deterministic behavior of iterating over a dense map.
5095 if (HasIllegalType && !AllTablesFitInRegister)
5099 // If each table would fit in a register, we should build it anyway.
5100 if (AllTablesFitInRegister)
5103 // Don't build a table that doesn't fit in-register if it has illegal types.
5107 // The table density should be at least 40%. This is the same criterion as for
5108 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5109 // FIXME: Find the best cut-off.
5110 return SI->getNumCases() * 10 >= TableSize * 4;
5113 /// Try to reuse the switch table index compare. Following pattern:
5115 /// if (idx < tablesize)
5116 /// r = table[idx]; // table does not contain default_value
5118 /// r = default_value;
5119 /// if (r != default_value)
5122 /// Is optimized to:
5124 /// cond = idx < tablesize;
5128 /// r = default_value;
5132 /// Jump threading will then eliminate the second if(cond).
5133 static void reuseTableCompare(
5134 User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5135 Constant *DefaultValue,
5136 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5137 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5141 // We require that the compare is in the same block as the phi so that jump
5142 // threading can do its work afterwards.
5143 if (CmpInst->getParent() != PhiBlock)
5146 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5150 Value *RangeCmp = RangeCheckBranch->getCondition();
5151 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5152 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5154 // Check if the compare with the default value is constant true or false.
5155 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5156 DefaultValue, CmpOp1, true);
5157 if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5160 // Check if the compare with the case values is distinct from the default
5162 for (auto ValuePair : Values) {
5163 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5164 ValuePair.second, CmpOp1, true);
5165 if (!CaseConst || CaseConst == DefaultConst || isa<UndefValue>(CaseConst))
5167 assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5168 "Expect true or false as compare result.");
5171 // Check if the branch instruction dominates the phi node. It's a simple
5172 // dominance check, but sufficient for our needs.
5173 // Although this check is invariant in the calling loops, it's better to do it
5174 // at this late stage. Practically we do it at most once for a switch.
5175 BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5176 for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5177 BasicBlock *Pred = *PI;
5178 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5182 if (DefaultConst == FalseConst) {
5183 // The compare yields the same result. We can replace it.
5184 CmpInst->replaceAllUsesWith(RangeCmp);
5185 ++NumTableCmpReuses;
5187 // The compare yields the same result, just inverted. We can replace it.
5188 Value *InvertedTableCmp = BinaryOperator::CreateXor(
5189 RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5191 CmpInst->replaceAllUsesWith(InvertedTableCmp);
5192 ++NumTableCmpReuses;
5196 /// If the switch is only used to initialize one or more phi nodes in a common
5197 /// successor block with different constant values, replace the switch with
5199 static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
5200 const DataLayout &DL,
5201 const TargetTransformInfo &TTI) {
5202 assert(SI->getNumCases() > 1 && "Degenerate switch?");
5204 Function *Fn = SI->getParent()->getParent();
5205 // Only build lookup table when we have a target that supports it or the
5206 // attribute is not set.
5207 if (!TTI.shouldBuildLookupTables() ||
5208 (Fn->getFnAttribute("no-jump-tables").getValueAsString() == "true"))
5211 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5212 // split off a dense part and build a lookup table for that.
5214 // FIXME: This creates arrays of GEPs to constant strings, which means each
5215 // GEP needs a runtime relocation in PIC code. We should just build one big
5216 // string and lookup indices into that.
5218 // Ignore switches with less than three cases. Lookup tables will not make
5219 // them faster, so we don't analyze them.
5220 if (SI->getNumCases() < 3)
5223 // Figure out the corresponding result for each case value and phi node in the
5224 // common destination, as well as the min and max case values.
5225 assert(SI->case_begin() != SI->case_end());
5226 SwitchInst::CaseIt CI = SI->case_begin();
5227 ConstantInt *MinCaseVal = CI->getCaseValue();
5228 ConstantInt *MaxCaseVal = CI->getCaseValue();
5230 BasicBlock *CommonDest = nullptr;
5232 using ResultListTy = SmallVector<std::pair<ConstantInt *, Constant *>, 4>;
5233 SmallDenseMap<PHINode *, ResultListTy> ResultLists;
5235 SmallDenseMap<PHINode *, Constant *> DefaultResults;
5236 SmallDenseMap<PHINode *, Type *> ResultTypes;
5237 SmallVector<PHINode *, 4> PHIs;
5239 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5240 ConstantInt *CaseVal = CI->getCaseValue();
5241 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5242 MinCaseVal = CaseVal;
5243 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5244 MaxCaseVal = CaseVal;
5246 // Resulting value at phi nodes for this case value.
5247 using ResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
5249 if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5253 // Append the result from this case to the list for each phi.
5254 for (const auto &I : Results) {
5255 PHINode *PHI = I.first;
5256 Constant *Value = I.second;
5257 if (!ResultLists.count(PHI))
5258 PHIs.push_back(PHI);
5259 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5263 // Keep track of the result types.
5264 for (PHINode *PHI : PHIs) {
5265 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5268 uint64_t NumResults = ResultLists[PHIs[0]].size();
5269 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5270 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5271 bool TableHasHoles = (NumResults < TableSize);
5273 // If the table has holes, we need a constant result for the default case
5274 // or a bitmask that fits in a register.
5275 SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5276 bool HasDefaultResults =
5277 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5278 DefaultResultsList, DL, TTI);
5280 bool NeedMask = (TableHasHoles && !HasDefaultResults);
5282 // As an extra penalty for the validity test we require more cases.
5283 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5285 if (!DL.fitsInLegalInteger(TableSize))
5289 for (const auto &I : DefaultResultsList) {
5290 PHINode *PHI = I.first;
5291 Constant *Result = I.second;
5292 DefaultResults[PHI] = Result;
5295 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5298 // Create the BB that does the lookups.
5299 Module &Mod = *CommonDest->getParent()->getParent();
5300 BasicBlock *LookupBB = BasicBlock::Create(
5301 Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5303 // Compute the table index value.
5304 Builder.SetInsertPoint(SI);
5306 if (MinCaseVal->isNullValue())
5307 TableIndex = SI->getCondition();
5309 TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
5312 // Compute the maximum table size representable by the integer type we are
5314 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5315 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5316 assert(MaxTableSize >= TableSize &&
5317 "It is impossible for a switch to have more entries than the max "
5318 "representable value of its input integer type's size.");
5320 // If the default destination is unreachable, or if the lookup table covers
5321 // all values of the conditional variable, branch directly to the lookup table
5322 // BB. Otherwise, check that the condition is within the case range.
5323 const bool DefaultIsReachable =
5324 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5325 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5326 BranchInst *RangeCheckBranch = nullptr;
5328 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5329 Builder.CreateBr(LookupBB);
5330 // Note: We call removeProdecessor later since we need to be able to get the
5331 // PHI value for the default case in case we're using a bit mask.
5333 Value *Cmp = Builder.CreateICmpULT(
5334 TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5336 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5339 // Populate the BB that does the lookups.
5340 Builder.SetInsertPoint(LookupBB);
5343 // Before doing the lookup, we do the hole check. The LookupBB is therefore
5344 // re-purposed to do the hole check, and we create a new LookupBB.
5345 BasicBlock *MaskBB = LookupBB;
5346 MaskBB->setName("switch.hole_check");
5347 LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5348 CommonDest->getParent(), CommonDest);
5350 // Make the mask's bitwidth at least 8-bit and a power-of-2 to avoid
5351 // unnecessary illegal types.
5352 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5353 APInt MaskInt(TableSizePowOf2, 0);
5354 APInt One(TableSizePowOf2, 1);
5355 // Build bitmask; fill in a 1 bit for every case.
5356 const ResultListTy &ResultList = ResultLists[PHIs[0]];
5357 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5358 uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5360 MaskInt |= One << Idx;
5362 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5364 // Get the TableIndex'th bit of the bitmask.
5365 // If this bit is 0 (meaning hole) jump to the default destination,
5366 // else continue with table lookup.
5367 IntegerType *MapTy = TableMask->getType();
5369 Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5370 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5371 Value *LoBit = Builder.CreateTrunc(
5372 Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5373 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5375 Builder.SetInsertPoint(LookupBB);
5376 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5379 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5380 // We cached PHINodes in PHIs. To avoid accessing deleted PHINodes later,
5381 // do not delete PHINodes here.
5382 SI->getDefaultDest()->removePredecessor(SI->getParent(),
5383 /*DontDeleteUselessPHIs=*/true);
5386 bool ReturnedEarly = false;
5387 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
5388 PHINode *PHI = PHIs[I];
5389 const ResultListTy &ResultList = ResultLists[PHI];
5391 // If using a bitmask, use any value to fill the lookup table holes.
5392 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5393 StringRef FuncName = Fn->getName();
5394 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL,
5397 Value *Result = Table.BuildLookup(TableIndex, Builder);
5399 // If the result is used to return immediately from the function, we want to
5400 // do that right here.
5401 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5402 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5403 Builder.CreateRet(Result);
5404 ReturnedEarly = true;
5408 // Do a small peephole optimization: re-use the switch table compare if
5410 if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5411 BasicBlock *PhiBlock = PHI->getParent();
5412 // Search for compare instructions which use the phi.
5413 for (auto *User : PHI->users()) {
5414 reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5418 PHI->addIncoming(Result, LookupBB);
5422 Builder.CreateBr(CommonDest);
5424 // Remove the switch.
5425 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5426 BasicBlock *Succ = SI->getSuccessor(i);
5428 if (Succ == SI->getDefaultDest())
5430 Succ->removePredecessor(SI->getParent());
5432 SI->eraseFromParent();
5436 ++NumLookupTablesHoles;
5440 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5441 // See also SelectionDAGBuilder::isDense(), which this function was based on.
5442 uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5443 uint64_t Range = Diff + 1;
5444 uint64_t NumCases = Values.size();
5445 // 40% is the default density for building a jump table in optsize/minsize mode.
5446 uint64_t MinDensity = 40;
5448 return NumCases * 100 >= Range * MinDensity;
5451 /// Try to transform a switch that has "holes" in it to a contiguous sequence
5454 /// A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5455 /// range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5457 /// This converts a sparse switch into a dense switch which allows better
5458 /// lowering and could also allow transforming into a lookup table.
5459 static bool ReduceSwitchRange(SwitchInst *SI, IRBuilder<> &Builder,
5460 const DataLayout &DL,
5461 const TargetTransformInfo &TTI) {
5462 auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5463 if (CondTy->getIntegerBitWidth() > 64 ||
5464 !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5466 // Only bother with this optimization if there are more than 3 switch cases;
5467 // SDAG will only bother creating jump tables for 4 or more cases.
5468 if (SI->getNumCases() < 4)
5471 // This transform is agnostic to the signedness of the input or case values. We
5472 // can treat the case values as signed or unsigned. We can optimize more common
5473 // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5475 SmallVector<int64_t,4> Values;
5476 for (auto &C : SI->cases())
5477 Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5478 std::sort(Values.begin(), Values.end());
5480 // If the switch is already dense, there's nothing useful to do here.
5481 if (isSwitchDense(Values))
5484 // First, transform the values such that they start at zero and ascend.
5485 int64_t Base = Values[0];
5486 for (auto &V : Values)
5487 V -= (uint64_t)(Base);
5489 // Now we have signed numbers that have been shifted so that, given enough
5490 // precision, there are no negative values. Since the rest of the transform
5491 // is bitwise only, we switch now to an unsigned representation.
5493 for (auto &V : Values)
5494 GCD = GreatestCommonDivisor64(GCD, (uint64_t)V);
5496 // This transform can be done speculatively because it is so cheap - it results
5497 // in a single rotate operation being inserted. This can only happen if the
5498 // factor extracted is a power of 2.
5499 // FIXME: If the GCD is an odd number we can multiply by the multiplicative
5500 // inverse of GCD and then perform this transform.
5501 // FIXME: It's possible that optimizing a switch on powers of two might also
5502 // be beneficial - flag values are often powers of two and we could use a CLZ
5503 // as the key function.
5504 if (GCD <= 1 || !isPowerOf2_64(GCD))
5505 // No common divisor found or too expensive to compute key function.
5508 unsigned Shift = Log2_64(GCD);
5509 for (auto &V : Values)
5510 V = (int64_t)((uint64_t)V >> Shift);
5512 if (!isSwitchDense(Values))
5513 // Transform didn't create a dense switch.
5516 // The obvious transform is to shift the switch condition right and emit a
5517 // check that the condition actually cleanly divided by GCD, i.e.
5518 // C & (1 << Shift - 1) == 0
5519 // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5521 // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5522 // shift and puts the shifted-off bits in the uppermost bits. If any of these
5523 // are nonzero then the switch condition will be very large and will hit the
5526 auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5527 Builder.SetInsertPoint(SI);
5528 auto *ShiftC = ConstantInt::get(Ty, Shift);
5529 auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5530 auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5531 auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5532 auto *Rot = Builder.CreateOr(LShr, Shl);
5533 SI->replaceUsesOfWith(SI->getCondition(), Rot);
5535 for (auto Case : SI->cases()) {
5536 auto *Orig = Case.getCaseValue();
5537 auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5539 cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5544 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
5545 BasicBlock *BB = SI->getParent();
5547 if (isValueEqualityComparison(SI)) {
5548 // If we only have one predecessor, and if it is a branch on this value,
5549 // see if that predecessor totally determines the outcome of this switch.
5550 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5551 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
5552 return simplifyCFG(BB, TTI, Options) | true;
5554 Value *Cond = SI->getCondition();
5555 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
5556 if (SimplifySwitchOnSelect(SI, Select))
5557 return simplifyCFG(BB, TTI, Options) | true;
5559 // If the block only contains the switch, see if we can fold the block
5560 // away into any preds.
5561 BasicBlock::iterator BBI = BB->begin();
5562 // Ignore dbg intrinsics.
5563 while (isa<DbgInfoIntrinsic>(BBI))
5566 if (FoldValueComparisonIntoPredecessors(SI, Builder))
5567 return simplifyCFG(BB, TTI, Options) | true;
5570 // Try to transform the switch into an icmp and a branch.
5571 if (TurnSwitchRangeIntoICmp(SI, Builder))
5572 return simplifyCFG(BB, TTI, Options) | true;
5574 // Remove unreachable cases.
5575 if (eliminateDeadSwitchCases(SI, Options.AC, DL))
5576 return simplifyCFG(BB, TTI, Options) | true;
5578 if (switchToSelect(SI, Builder, DL, TTI))
5579 return simplifyCFG(BB, TTI, Options) | true;
5581 if (Options.ForwardSwitchCondToPhi && ForwardSwitchConditionToPHI(SI))
5582 return simplifyCFG(BB, TTI, Options) | true;
5584 // The conversion from switch to lookup tables results in difficult-to-analyze
5585 // code and makes pruning branches much harder. This is a problem if the
5586 // switch expression itself can still be restricted as a result of inlining or
5587 // CVP. Therefore, only apply this transformation during late stages of the
5588 // optimisation pipeline.
5589 if (Options.ConvertSwitchToLookupTable &&
5590 SwitchToLookupTable(SI, Builder, DL, TTI))
5591 return simplifyCFG(BB, TTI, Options) | true;
5593 if (ReduceSwitchRange(SI, Builder, DL, TTI))
5594 return simplifyCFG(BB, TTI, Options) | true;
5599 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
5600 BasicBlock *BB = IBI->getParent();
5601 bool Changed = false;
5603 // Eliminate redundant destinations.
5604 SmallPtrSet<Value *, 8> Succs;
5605 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
5606 BasicBlock *Dest = IBI->getDestination(i);
5607 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
5608 Dest->removePredecessor(BB);
5609 IBI->removeDestination(i);
5616 if (IBI->getNumDestinations() == 0) {
5617 // If the indirectbr has no successors, change it to unreachable.
5618 new UnreachableInst(IBI->getContext(), IBI);
5619 EraseTerminatorInstAndDCECond(IBI);
5623 if (IBI->getNumDestinations() == 1) {
5624 // If the indirectbr has one successor, change it to a direct branch.
5625 BranchInst::Create(IBI->getDestination(0), IBI);
5626 EraseTerminatorInstAndDCECond(IBI);
5630 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
5631 if (SimplifyIndirectBrOnSelect(IBI, SI))
5632 return simplifyCFG(BB, TTI, Options) | true;
5637 /// Given an block with only a single landing pad and a unconditional branch
5638 /// try to find another basic block which this one can be merged with. This
5639 /// handles cases where we have multiple invokes with unique landing pads, but
5640 /// a shared handler.
5642 /// We specifically choose to not worry about merging non-empty blocks
5643 /// here. That is a PRE/scheduling problem and is best solved elsewhere. In
5644 /// practice, the optimizer produces empty landing pad blocks quite frequently
5645 /// when dealing with exception dense code. (see: instcombine, gvn, if-else
5646 /// sinking in this file)
5648 /// This is primarily a code size optimization. We need to avoid performing
5649 /// any transform which might inhibit optimization (such as our ability to
5650 /// specialize a particular handler via tail commoning). We do this by not
5651 /// merging any blocks which require us to introduce a phi. Since the same
5652 /// values are flowing through both blocks, we don't loose any ability to
5653 /// specialize. If anything, we make such specialization more likely.
5655 /// TODO - This transformation could remove entries from a phi in the target
5656 /// block when the inputs in the phi are the same for the two blocks being
5657 /// merged. In some cases, this could result in removal of the PHI entirely.
5658 static bool TryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI,
5660 auto Succ = BB->getUniqueSuccessor();
5662 // If there's a phi in the successor block, we'd likely have to introduce
5663 // a phi into the merged landing pad block.
5664 if (isa<PHINode>(*Succ->begin()))
5667 for (BasicBlock *OtherPred : predecessors(Succ)) {
5668 if (BB == OtherPred)
5670 BasicBlock::iterator I = OtherPred->begin();
5671 LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(I);
5672 if (!LPad2 || !LPad2->isIdenticalTo(LPad))
5674 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5676 BranchInst *BI2 = dyn_cast<BranchInst>(I);
5677 if (!BI2 || !BI2->isIdenticalTo(BI))
5680 // We've found an identical block. Update our predecessors to take that
5681 // path instead and make ourselves dead.
5682 SmallSet<BasicBlock *, 16> Preds;
5683 Preds.insert(pred_begin(BB), pred_end(BB));
5684 for (BasicBlock *Pred : Preds) {
5685 InvokeInst *II = cast<InvokeInst>(Pred->getTerminator());
5686 assert(II->getNormalDest() != BB && II->getUnwindDest() == BB &&
5687 "unexpected successor");
5688 II->setUnwindDest(OtherPred);
5691 // The debug info in OtherPred doesn't cover the merged control flow that
5692 // used to go through BB. We need to delete it or update it.
5693 for (auto I = OtherPred->begin(), E = OtherPred->end(); I != E;) {
5694 Instruction &Inst = *I;
5696 if (isa<DbgInfoIntrinsic>(Inst))
5697 Inst.eraseFromParent();
5700 SmallSet<BasicBlock *, 16> Succs;
5701 Succs.insert(succ_begin(BB), succ_end(BB));
5702 for (BasicBlock *Succ : Succs) {
5703 Succ->removePredecessor(BB);
5706 IRBuilder<> Builder(BI);
5707 Builder.CreateUnreachable();
5708 BI->eraseFromParent();
5714 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI,
5715 IRBuilder<> &Builder) {
5716 BasicBlock *BB = BI->getParent();
5717 BasicBlock *Succ = BI->getSuccessor(0);
5719 // If the Terminator is the only non-phi instruction, simplify the block.
5720 // If LoopHeader is provided, check if the block or its successor is a loop
5721 // header. (This is for early invocations before loop simplify and
5722 // vectorization to keep canonical loop forms for nested loops. These blocks
5723 // can be eliminated when the pass is invoked later in the back-end.)
5724 bool NeedCanonicalLoop =
5725 Options.NeedCanonicalLoop &&
5726 (LoopHeaders && (LoopHeaders->count(BB) || LoopHeaders->count(Succ)));
5727 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg()->getIterator();
5728 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
5729 !NeedCanonicalLoop && TryToSimplifyUncondBranchFromEmptyBlock(BB))
5732 // If the only instruction in the block is a seteq/setne comparison against a
5733 // constant, try to simplify the block.
5734 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
5735 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
5736 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5738 if (I->isTerminator() &&
5739 tryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, DL, TTI, Options))
5743 // See if we can merge an empty landing pad block with another which is
5745 if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(I)) {
5746 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5748 if (I->isTerminator() && TryToMergeLandingPad(LPad, BI, BB))
5752 // If this basic block is ONLY a compare and a branch, and if a predecessor
5753 // branches to us and our successor, fold the comparison into the
5754 // predecessor and use logical operations to update the incoming value
5755 // for PHI nodes in common successor.
5756 if (FoldBranchToCommonDest(BI, Options.BonusInstThreshold))
5757 return simplifyCFG(BB, TTI, Options) | true;
5761 static BasicBlock *allPredecessorsComeFromSameSource(BasicBlock *BB) {
5762 BasicBlock *PredPred = nullptr;
5763 for (auto *P : predecessors(BB)) {
5764 BasicBlock *PPred = P->getSinglePredecessor();
5765 if (!PPred || (PredPred && PredPred != PPred))
5772 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
5773 BasicBlock *BB = BI->getParent();
5775 // Conditional branch
5776 if (isValueEqualityComparison(BI)) {
5777 // If we only have one predecessor, and if it is a branch on this value,
5778 // see if that predecessor totally determines the outcome of this
5780 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5781 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
5782 return simplifyCFG(BB, TTI, Options) | true;
5784 // This block must be empty, except for the setcond inst, if it exists.
5785 // Ignore dbg intrinsics.
5786 BasicBlock::iterator I = BB->begin();
5787 // Ignore dbg intrinsics.
5788 while (isa<DbgInfoIntrinsic>(I))
5791 if (FoldValueComparisonIntoPredecessors(BI, Builder))
5792 return simplifyCFG(BB, TTI, Options) | true;
5793 } else if (&*I == cast<Instruction>(BI->getCondition())) {
5795 // Ignore dbg intrinsics.
5796 while (isa<DbgInfoIntrinsic>(I))
5798 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
5799 return simplifyCFG(BB, TTI, Options) | true;
5803 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
5804 if (SimplifyBranchOnICmpChain(BI, Builder, DL))
5807 // If this basic block has a single dominating predecessor block and the
5808 // dominating block's condition implies BI's condition, we know the direction
5809 // of the BI branch.
5810 if (BasicBlock *Dom = BB->getSinglePredecessor()) {
5811 auto *PBI = dyn_cast_or_null<BranchInst>(Dom->getTerminator());
5812 if (PBI && PBI->isConditional() &&
5813 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
5814 assert(PBI->getSuccessor(0) == BB || PBI->getSuccessor(1) == BB);
5815 bool CondIsTrue = PBI->getSuccessor(0) == BB;
5816 Optional<bool> Implication = isImpliedCondition(
5817 PBI->getCondition(), BI->getCondition(), DL, CondIsTrue);
5819 // Turn this into a branch on constant.
5820 auto *OldCond = BI->getCondition();
5821 ConstantInt *CI = *Implication
5822 ? ConstantInt::getTrue(BB->getContext())
5823 : ConstantInt::getFalse(BB->getContext());
5824 BI->setCondition(CI);
5825 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
5826 return simplifyCFG(BB, TTI, Options) | true;
5831 // If this basic block is ONLY a compare and a branch, and if a predecessor
5832 // branches to us and one of our successors, fold the comparison into the
5833 // predecessor and use logical operations to pick the right destination.
5834 if (FoldBranchToCommonDest(BI, Options.BonusInstThreshold))
5835 return simplifyCFG(BB, TTI, Options) | true;
5837 // We have a conditional branch to two blocks that are only reachable
5838 // from BI. We know that the condbr dominates the two blocks, so see if
5839 // there is any identical code in the "then" and "else" blocks. If so, we
5840 // can hoist it up to the branching block.
5841 if (BI->getSuccessor(0)->getSinglePredecessor()) {
5842 if (BI->getSuccessor(1)->getSinglePredecessor()) {
5843 if (HoistThenElseCodeToIf(BI, TTI))
5844 return simplifyCFG(BB, TTI, Options) | true;
5846 // If Successor #1 has multiple preds, we may be able to conditionally
5847 // execute Successor #0 if it branches to Successor #1.
5848 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
5849 if (Succ0TI->getNumSuccessors() == 1 &&
5850 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
5851 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), TTI))
5852 return simplifyCFG(BB, TTI, Options) | true;
5854 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
5855 // If Successor #0 has multiple preds, we may be able to conditionally
5856 // execute Successor #1 if it branches to Successor #0.
5857 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
5858 if (Succ1TI->getNumSuccessors() == 1 &&
5859 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
5860 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), TTI))
5861 return simplifyCFG(BB, TTI, Options) | true;
5864 // If this is a branch on a phi node in the current block, thread control
5865 // through this block if any PHI node entries are constants.
5866 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
5867 if (PN->getParent() == BI->getParent())
5868 if (FoldCondBranchOnPHI(BI, DL, Options.AC))
5869 return simplifyCFG(BB, TTI, Options) | true;
5871 // Scan predecessor blocks for conditional branches.
5872 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
5873 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
5874 if (PBI != BI && PBI->isConditional())
5875 if (SimplifyCondBranchToCondBranch(PBI, BI, DL))
5876 return simplifyCFG(BB, TTI, Options) | true;
5878 // Look for diamond patterns.
5879 if (MergeCondStores)
5880 if (BasicBlock *PrevBB = allPredecessorsComeFromSameSource(BB))
5881 if (BranchInst *PBI = dyn_cast<BranchInst>(PrevBB->getTerminator()))
5882 if (PBI != BI && PBI->isConditional())
5883 if (mergeConditionalStores(PBI, BI, DL))
5884 return simplifyCFG(BB, TTI, Options) | true;
5889 /// Check if passing a value to an instruction will cause undefined behavior.
5890 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
5891 Constant *C = dyn_cast<Constant>(V);
5898 if (C->isNullValue() || isa<UndefValue>(C)) {
5899 // Only look at the first use, avoid hurting compile time with long uselists
5900 User *Use = *I->user_begin();
5902 // Now make sure that there are no instructions in between that can alter
5903 // control flow (eg. calls)
5904 for (BasicBlock::iterator
5905 i = ++BasicBlock::iterator(I),
5906 UI = BasicBlock::iterator(dyn_cast<Instruction>(Use));
5908 if (i == I->getParent()->end() || i->mayHaveSideEffects())
5911 // Look through GEPs. A load from a GEP derived from NULL is still undefined
5912 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
5913 if (GEP->getPointerOperand() == I)
5914 return passingValueIsAlwaysUndefined(V, GEP);
5916 // Look through bitcasts.
5917 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
5918 return passingValueIsAlwaysUndefined(V, BC);
5920 // Load from null is undefined.
5921 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
5922 if (!LI->isVolatile())
5923 return LI->getPointerAddressSpace() == 0;
5925 // Store to null is undefined.
5926 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
5927 if (!SI->isVolatile())
5928 return SI->getPointerAddressSpace() == 0 &&
5929 SI->getPointerOperand() == I;
5931 // A call to null is undefined.
5932 if (auto CS = CallSite(Use))
5933 return CS.getCalledValue() == I;
5938 /// If BB has an incoming value that will always trigger undefined behavior
5939 /// (eg. null pointer dereference), remove the branch leading here.
5940 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
5941 for (PHINode &PHI : BB->phis())
5942 for (unsigned i = 0, e = PHI.getNumIncomingValues(); i != e; ++i)
5943 if (passingValueIsAlwaysUndefined(PHI.getIncomingValue(i), &PHI)) {
5944 TerminatorInst *T = PHI.getIncomingBlock(i)->getTerminator();
5945 IRBuilder<> Builder(T);
5946 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
5947 BB->removePredecessor(PHI.getIncomingBlock(i));
5948 // Turn uncoditional branches into unreachables and remove the dead
5949 // destination from conditional branches.
5950 if (BI->isUnconditional())
5951 Builder.CreateUnreachable();
5953 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1)
5954 : BI->getSuccessor(0));
5955 BI->eraseFromParent();
5958 // TODO: SwitchInst.
5964 bool SimplifyCFGOpt::run(BasicBlock *BB) {
5965 bool Changed = false;
5967 assert(BB && BB->getParent() && "Block not embedded in function!");
5968 assert(BB->getTerminator() && "Degenerate basic block encountered!");
5970 // Remove basic blocks that have no predecessors (except the entry block)...
5971 // or that just have themself as a predecessor. These are unreachable.
5972 if ((pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) ||
5973 BB->getSinglePredecessor() == BB) {
5974 DEBUG(dbgs() << "Removing BB: \n" << *BB);
5975 DeleteDeadBlock(BB);
5979 // Check to see if we can constant propagate this terminator instruction
5981 Changed |= ConstantFoldTerminator(BB, true);
5983 // Check for and eliminate duplicate PHI nodes in this block.
5984 Changed |= EliminateDuplicatePHINodes(BB);
5986 // Check for and remove branches that will always cause undefined behavior.
5987 Changed |= removeUndefIntroducingPredecessor(BB);
5989 // Merge basic blocks into their predecessor if there is only one distinct
5990 // pred, and if there is only one distinct successor of the predecessor, and
5991 // if there are no PHI nodes.
5992 if (MergeBlockIntoPredecessor(BB))
5995 if (SinkCommon && Options.SinkCommonInsts)
5996 Changed |= SinkCommonCodeFromPredecessors(BB);
5998 IRBuilder<> Builder(BB);
6000 // If there is a trivial two-entry PHI node in this basic block, and we can
6001 // eliminate it, do so now.
6002 if (auto *PN = dyn_cast<PHINode>(BB->begin()))
6003 if (PN->getNumIncomingValues() == 2)
6004 Changed |= FoldTwoEntryPHINode(PN, TTI, DL);
6006 Builder.SetInsertPoint(BB->getTerminator());
6007 if (auto *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
6008 if (BI->isUnconditional()) {
6009 if (SimplifyUncondBranch(BI, Builder))
6012 if (SimplifyCondBranch(BI, Builder))
6015 } else if (auto *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
6016 if (SimplifyReturn(RI, Builder))
6018 } else if (auto *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
6019 if (SimplifyResume(RI, Builder))
6021 } else if (auto *RI = dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
6022 if (SimplifyCleanupReturn(RI))
6024 } else if (auto *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
6025 if (SimplifySwitch(SI, Builder))
6027 } else if (auto *UI = dyn_cast<UnreachableInst>(BB->getTerminator())) {
6028 if (SimplifyUnreachable(UI))
6030 } else if (auto *IBI = dyn_cast<IndirectBrInst>(BB->getTerminator())) {
6031 if (SimplifyIndirectBr(IBI))
6038 bool llvm::simplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
6039 const SimplifyCFGOptions &Options,
6040 SmallPtrSetImpl<BasicBlock *> *LoopHeaders) {
6041 return SimplifyCFGOpt(TTI, BB->getModule()->getDataLayout(), LoopHeaders,