1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Peephole optimize the CFG.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/ADT/APInt.h"
15 #include "llvm/ADT/ArrayRef.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/Optional.h"
18 #include "llvm/ADT/SetOperations.h"
19 #include "llvm/ADT/SetVector.h"
20 #include "llvm/ADT/SmallPtrSet.h"
21 #include "llvm/ADT/SmallSet.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/Analysis/AssumptionCache.h"
26 #include "llvm/Analysis/ConstantFolding.h"
27 #include "llvm/Analysis/EHPersonalities.h"
28 #include "llvm/Analysis/InstructionSimplify.h"
29 #include "llvm/Analysis/TargetTransformInfo.h"
30 #include "llvm/Analysis/ValueTracking.h"
31 #include "llvm/IR/BasicBlock.h"
32 #include "llvm/IR/CallSite.h"
33 #include "llvm/IR/CFG.h"
34 #include "llvm/IR/Constant.h"
35 #include "llvm/IR/ConstantRange.h"
36 #include "llvm/IR/Constants.h"
37 #include "llvm/IR/DataLayout.h"
38 #include "llvm/IR/DebugInfo.h"
39 #include "llvm/IR/DerivedTypes.h"
40 #include "llvm/IR/GlobalValue.h"
41 #include "llvm/IR/GlobalVariable.h"
42 #include "llvm/IR/IRBuilder.h"
43 #include "llvm/IR/InstrTypes.h"
44 #include "llvm/IR/Instruction.h"
45 #include "llvm/IR/Instructions.h"
46 #include "llvm/IR/IntrinsicInst.h"
47 #include "llvm/IR/Intrinsics.h"
48 #include "llvm/IR/LLVMContext.h"
49 #include "llvm/IR/MDBuilder.h"
50 #include "llvm/IR/Metadata.h"
51 #include "llvm/IR/Module.h"
52 #include "llvm/IR/NoFolder.h"
53 #include "llvm/IR/Operator.h"
54 #include "llvm/IR/PatternMatch.h"
55 #include "llvm/IR/Type.h"
56 #include "llvm/IR/User.h"
57 #include "llvm/IR/Value.h"
58 #include "llvm/IR/DebugInfo.h"
59 #include "llvm/Support/Casting.h"
60 #include "llvm/Support/CommandLine.h"
61 #include "llvm/Support/Debug.h"
62 #include "llvm/Support/ErrorHandling.h"
63 #include "llvm/Support/MathExtras.h"
64 #include "llvm/Support/raw_ostream.h"
65 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
66 #include "llvm/Transforms/Utils/Local.h"
67 #include "llvm/Transforms/Utils/ValueMapper.h"
80 using namespace PatternMatch;
82 #define DEBUG_TYPE "simplifycfg"
84 // Chosen as 2 so as to be cheap, but still to have enough power to fold
85 // a select, so the "clamp" idiom (of a min followed by a max) will be caught.
86 // To catch this, we need to fold a compare and a select, hence '2' being the
87 // minimum reasonable default.
88 static cl::opt<unsigned> PHINodeFoldingThreshold(
89 "phi-node-folding-threshold", cl::Hidden, cl::init(2),
91 "Control the amount of phi node folding to perform (default = 2)"));
93 static cl::opt<bool> DupRet(
94 "simplifycfg-dup-ret", cl::Hidden, cl::init(false),
95 cl::desc("Duplicate return instructions into unconditional branches"));
98 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
99 cl::desc("Sink common instructions down to the end block"));
101 static cl::opt<bool> HoistCondStores(
102 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
103 cl::desc("Hoist conditional stores if an unconditional store precedes"));
105 static cl::opt<bool> MergeCondStores(
106 "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
107 cl::desc("Hoist conditional stores even if an unconditional store does not "
108 "precede - hoist multiple conditional stores into a single "
109 "predicated store"));
111 static cl::opt<bool> MergeCondStoresAggressively(
112 "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
113 cl::desc("When merging conditional stores, do so even if the resultant "
114 "basic blocks are unlikely to be if-converted as a result"));
116 static cl::opt<bool> SpeculateOneExpensiveInst(
117 "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
118 cl::desc("Allow exactly one expensive instruction to be speculatively "
121 static cl::opt<unsigned> MaxSpeculationDepth(
122 "max-speculation-depth", cl::Hidden, cl::init(10),
123 cl::desc("Limit maximum recursion depth when calculating costs of "
124 "speculatively executed instructions"));
126 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
127 STATISTIC(NumLinearMaps,
128 "Number of switch instructions turned into linear mapping");
129 STATISTIC(NumLookupTables,
130 "Number of switch instructions turned into lookup tables");
132 NumLookupTablesHoles,
133 "Number of switch instructions turned into lookup tables (holes checked)");
134 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
135 STATISTIC(NumSinkCommons,
136 "Number of common instructions sunk down to the end block");
137 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
141 // The first field contains the value that the switch produces when a certain
142 // case group is selected, and the second field is a vector containing the
143 // cases composing the case group.
144 typedef SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>
145 SwitchCaseResultVectorTy;
146 // The first field contains the phi node that generates a result of the switch
147 // and the second field contains the value generated for a certain case in the
148 // switch for that PHI.
149 typedef SmallVector<std::pair<PHINode *, Constant *>, 4> SwitchCaseResultsTy;
151 /// ValueEqualityComparisonCase - Represents a case of a switch.
152 struct ValueEqualityComparisonCase {
156 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
157 : Value(Value), Dest(Dest) {}
159 bool operator<(ValueEqualityComparisonCase RHS) const {
160 // Comparing pointers is ok as we only rely on the order for uniquing.
161 return Value < RHS.Value;
164 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
167 class SimplifyCFGOpt {
168 const TargetTransformInfo &TTI;
169 const DataLayout &DL;
170 unsigned BonusInstThreshold;
172 SmallPtrSetImpl<BasicBlock *> *LoopHeaders;
173 // See comments in SimplifyCFGOpt::SimplifySwitch.
174 bool LateSimplifyCFG;
175 Value *isValueEqualityComparison(TerminatorInst *TI);
176 BasicBlock *GetValueEqualityComparisonCases(
177 TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases);
178 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
180 IRBuilder<> &Builder);
181 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
182 IRBuilder<> &Builder);
184 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
185 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
186 bool SimplifySingleResume(ResumeInst *RI);
187 bool SimplifyCommonResume(ResumeInst *RI);
188 bool SimplifyCleanupReturn(CleanupReturnInst *RI);
189 bool SimplifyUnreachable(UnreachableInst *UI);
190 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
191 bool SimplifyIndirectBr(IndirectBrInst *IBI);
192 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
193 bool SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
196 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
197 unsigned BonusInstThreshold, AssumptionCache *AC,
198 SmallPtrSetImpl<BasicBlock *> *LoopHeaders,
199 bool LateSimplifyCFG)
200 : TTI(TTI), DL(DL), BonusInstThreshold(BonusInstThreshold), AC(AC),
201 LoopHeaders(LoopHeaders), LateSimplifyCFG(LateSimplifyCFG) {}
203 bool run(BasicBlock *BB);
206 } // end anonymous namespace
208 /// Return true if it is safe to merge these two
209 /// terminator instructions together.
211 SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2,
212 SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
214 return false; // Can't merge with self!
216 // It is not safe to merge these two switch instructions if they have a common
217 // successor, and if that successor has a PHI node, and if *that* PHI node has
218 // conflicting incoming values from the two switch blocks.
219 BasicBlock *SI1BB = SI1->getParent();
220 BasicBlock *SI2BB = SI2->getParent();
222 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
224 for (BasicBlock *Succ : successors(SI2BB))
225 if (SI1Succs.count(Succ))
226 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
227 PHINode *PN = cast<PHINode>(BBI);
228 if (PN->getIncomingValueForBlock(SI1BB) !=
229 PN->getIncomingValueForBlock(SI2BB)) {
231 FailBlocks->insert(Succ);
239 /// Return true if it is safe and profitable to merge these two terminator
240 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
241 /// store all PHI nodes in common successors.
243 isProfitableToFoldUnconditional(BranchInst *SI1, BranchInst *SI2,
245 SmallVectorImpl<PHINode *> &PhiNodes) {
247 return false; // Can't merge with self!
248 assert(SI1->isUnconditional() && SI2->isConditional());
250 // We fold the unconditional branch if we can easily update all PHI nodes in
251 // common successors:
252 // 1> We have a constant incoming value for the conditional branch;
253 // 2> We have "Cond" as the incoming value for the unconditional branch;
254 // 3> SI2->getCondition() and Cond have same operands.
255 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
258 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
259 Cond->getOperand(1) == Ci2->getOperand(1)) &&
260 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
261 Cond->getOperand(1) == Ci2->getOperand(0)))
264 BasicBlock *SI1BB = SI1->getParent();
265 BasicBlock *SI2BB = SI2->getParent();
266 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
267 for (BasicBlock *Succ : successors(SI2BB))
268 if (SI1Succs.count(Succ))
269 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
270 PHINode *PN = cast<PHINode>(BBI);
271 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
272 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
274 PhiNodes.push_back(PN);
279 /// Update PHI nodes in Succ to indicate that there will now be entries in it
280 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
281 /// will be the same as those coming in from ExistPred, an existing predecessor
283 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
284 BasicBlock *ExistPred) {
285 if (!isa<PHINode>(Succ->begin()))
286 return; // Quick exit if nothing to do
289 for (BasicBlock::iterator I = Succ->begin(); (PN = dyn_cast<PHINode>(I)); ++I)
290 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
293 /// Compute an abstract "cost" of speculating the given instruction,
294 /// which is assumed to be safe to speculate. TCC_Free means cheap,
295 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
297 static unsigned ComputeSpeculationCost(const User *I,
298 const TargetTransformInfo &TTI) {
299 assert(isSafeToSpeculativelyExecute(I) &&
300 "Instruction is not safe to speculatively execute!");
301 return TTI.getUserCost(I);
304 /// If we have a merge point of an "if condition" as accepted above,
305 /// return true if the specified value dominates the block. We
306 /// don't handle the true generality of domination here, just a special case
307 /// which works well enough for us.
309 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
310 /// see if V (which must be an instruction) and its recursive operands
311 /// that do not dominate BB have a combined cost lower than CostRemaining and
312 /// are non-trapping. If both are true, the instruction is inserted into the
313 /// set and true is returned.
315 /// The cost for most non-trapping instructions is defined as 1 except for
316 /// Select whose cost is 2.
318 /// After this function returns, CostRemaining is decreased by the cost of
319 /// V plus its non-dominating operands. If that cost is greater than
320 /// CostRemaining, false is returned and CostRemaining is undefined.
321 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
322 SmallPtrSetImpl<Instruction *> *AggressiveInsts,
323 unsigned &CostRemaining,
324 const TargetTransformInfo &TTI,
325 unsigned Depth = 0) {
326 // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
327 // so limit the recursion depth.
328 // TODO: While this recursion limit does prevent pathological behavior, it
329 // would be better to track visited instructions to avoid cycles.
330 if (Depth == MaxSpeculationDepth)
333 Instruction *I = dyn_cast<Instruction>(V);
335 // Non-instructions all dominate instructions, but not all constantexprs
336 // can be executed unconditionally.
337 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
342 BasicBlock *PBB = I->getParent();
344 // We don't want to allow weird loops that might have the "if condition" in
345 // the bottom of this block.
349 // If this instruction is defined in a block that contains an unconditional
350 // branch to BB, then it must be in the 'conditional' part of the "if
351 // statement". If not, it definitely dominates the region.
352 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
353 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
356 // If we aren't allowing aggressive promotion anymore, then don't consider
357 // instructions in the 'if region'.
358 if (!AggressiveInsts)
361 // If we have seen this instruction before, don't count it again.
362 if (AggressiveInsts->count(I))
365 // Okay, it looks like the instruction IS in the "condition". Check to
366 // see if it's a cheap instruction to unconditionally compute, and if it
367 // only uses stuff defined outside of the condition. If so, hoist it out.
368 if (!isSafeToSpeculativelyExecute(I))
371 unsigned Cost = ComputeSpeculationCost(I, TTI);
373 // Allow exactly one instruction to be speculated regardless of its cost
374 // (as long as it is safe to do so).
375 // This is intended to flatten the CFG even if the instruction is a division
376 // or other expensive operation. The speculation of an expensive instruction
377 // is expected to be undone in CodeGenPrepare if the speculation has not
378 // enabled further IR optimizations.
379 if (Cost > CostRemaining &&
380 (!SpeculateOneExpensiveInst || !AggressiveInsts->empty() || Depth > 0))
383 // Avoid unsigned wrap.
384 CostRemaining = (Cost > CostRemaining) ? 0 : CostRemaining - Cost;
386 // Okay, we can only really hoist these out if their operands do
387 // not take us over the cost threshold.
388 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
389 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, TTI,
392 // Okay, it's safe to do this! Remember this instruction.
393 AggressiveInsts->insert(I);
397 /// Extract ConstantInt from value, looking through IntToPtr
398 /// and PointerNullValue. Return NULL if value is not a constant int.
399 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
400 // Normal constant int.
401 ConstantInt *CI = dyn_cast<ConstantInt>(V);
402 if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
405 // This is some kind of pointer constant. Turn it into a pointer-sized
406 // ConstantInt if possible.
407 IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
409 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
410 if (isa<ConstantPointerNull>(V))
411 return ConstantInt::get(PtrTy, 0);
413 // IntToPtr const int.
414 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
415 if (CE->getOpcode() == Instruction::IntToPtr)
416 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
417 // The constant is very likely to have the right type already.
418 if (CI->getType() == PtrTy)
421 return cast<ConstantInt>(
422 ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
429 /// Given a chain of or (||) or and (&&) comparison of a value against a
430 /// constant, this will try to recover the information required for a switch
432 /// It will depth-first traverse the chain of comparison, seeking for patterns
433 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
434 /// representing the different cases for the switch.
435 /// Note that if the chain is composed of '||' it will build the set of elements
436 /// that matches the comparisons (i.e. any of this value validate the chain)
437 /// while for a chain of '&&' it will build the set elements that make the test
439 struct ConstantComparesGatherer {
440 const DataLayout &DL;
441 Value *CompValue; /// Value found for the switch comparison
442 Value *Extra; /// Extra clause to be checked before the switch
443 SmallVector<ConstantInt *, 8> Vals; /// Set of integers to match in switch
444 unsigned UsedICmps; /// Number of comparisons matched in the and/or chain
446 /// Construct and compute the result for the comparison instruction Cond
447 ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL)
448 : DL(DL), CompValue(nullptr), Extra(nullptr), UsedICmps(0) {
453 ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
454 ConstantComparesGatherer &
455 operator=(const ConstantComparesGatherer &) = delete;
458 /// Try to set the current value used for the comparison, it succeeds only if
459 /// it wasn't set before or if the new value is the same as the old one
460 bool setValueOnce(Value *NewVal) {
461 if (CompValue && CompValue != NewVal)
464 return (CompValue != nullptr);
467 /// Try to match Instruction "I" as a comparison against a constant and
468 /// populates the array Vals with the set of values that match (or do not
469 /// match depending on isEQ).
470 /// Return false on failure. On success, the Value the comparison matched
471 /// against is placed in CompValue.
472 /// If CompValue is already set, the function is expected to fail if a match
473 /// is found but the value compared to is different.
474 bool matchInstruction(Instruction *I, bool isEQ) {
475 // If this is an icmp against a constant, handle this as one of the cases.
478 if (!((ICI = dyn_cast<ICmpInst>(I)) &&
479 (C = GetConstantInt(I->getOperand(1), DL)))) {
486 // Pattern match a special case
487 // (x & ~2^z) == y --> x == y || x == y|2^z
488 // This undoes a transformation done by instcombine to fuse 2 compares.
489 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
491 // It's a little bit hard to see why the following transformations are
492 // correct. Here is a CVC3 program to verify them for 64-bit values:
495 ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
499 mask : BITVECTOR(64) = BVSHL(ONE, z);
500 QUERY( (y & ~mask = y) =>
501 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
503 QUERY( (y | mask = y) =>
504 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
508 // Please note that each pattern must be a dual implication (<--> or
509 // iff). One directional implication can create spurious matches. If the
510 // implication is only one-way, an unsatisfiable condition on the left
511 // side can imply a satisfiable condition on the right side. Dual
512 // implication ensures that satisfiable conditions are transformed to
513 // other satisfiable conditions and unsatisfiable conditions are
514 // transformed to other unsatisfiable conditions.
516 // Here is a concrete example of a unsatisfiable condition on the left
517 // implying a satisfiable condition on the right:
520 // (x & ~mask) == y --> (x == y || x == (y | mask))
522 // Substituting y = 3, z = 0 yields:
523 // (x & -2) == 3 --> (x == 3 || x == 2)
525 // Pattern match a special case:
527 QUERY( (y & ~mask = y) =>
528 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
531 if (match(ICI->getOperand(0),
532 m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
534 if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
535 // If we already have a value for the switch, it has to match!
536 if (!setValueOnce(RHSVal))
541 ConstantInt::get(C->getContext(),
542 C->getValue() | Mask));
548 // Pattern match a special case:
550 QUERY( (y | mask = y) =>
551 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
554 if (match(ICI->getOperand(0),
555 m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
557 if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
558 // If we already have a value for the switch, it has to match!
559 if (!setValueOnce(RHSVal))
563 Vals.push_back(ConstantInt::get(C->getContext(),
564 C->getValue() & ~Mask));
570 // If we already have a value for the switch, it has to match!
571 if (!setValueOnce(ICI->getOperand(0)))
576 return ICI->getOperand(0);
579 // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
580 ConstantRange Span = ConstantRange::makeAllowedICmpRegion(
581 ICI->getPredicate(), C->getValue());
583 // Shift the range if the compare is fed by an add. This is the range
584 // compare idiom as emitted by instcombine.
585 Value *CandidateVal = I->getOperand(0);
586 if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
587 Span = Span.subtract(*RHSC);
588 CandidateVal = RHSVal;
591 // If this is an and/!= check, then we are looking to build the set of
592 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
595 Span = Span.inverse();
597 // If there are a ton of values, we don't want to make a ginormous switch.
598 if (Span.getSetSize().ugt(8) || Span.isEmptySet()) {
602 // If we already have a value for the switch, it has to match!
603 if (!setValueOnce(CandidateVal))
606 // Add all values from the range to the set
607 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
608 Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
614 /// Given a potentially 'or'd or 'and'd together collection of icmp
615 /// eq/ne/lt/gt instructions that compare a value against a constant, extract
616 /// the value being compared, and stick the list constants into the Vals
618 /// One "Extra" case is allowed to differ from the other.
619 void gather(Value *V) {
620 Instruction *I = dyn_cast<Instruction>(V);
621 bool isEQ = (I->getOpcode() == Instruction::Or);
623 // Keep a stack (SmallVector for efficiency) for depth-first traversal
624 SmallVector<Value *, 8> DFT;
625 SmallPtrSet<Value *, 8> Visited;
631 while (!DFT.empty()) {
632 V = DFT.pop_back_val();
634 if (Instruction *I = dyn_cast<Instruction>(V)) {
635 // If it is a || (or && depending on isEQ), process the operands.
636 if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
637 if (Visited.insert(I->getOperand(1)).second)
638 DFT.push_back(I->getOperand(1));
639 if (Visited.insert(I->getOperand(0)).second)
640 DFT.push_back(I->getOperand(0));
644 // Try to match the current instruction
645 if (matchInstruction(I, isEQ))
646 // Match succeed, continue the loop
650 // One element of the sequence of || (or &&) could not be match as a
651 // comparison against the same value as the others.
652 // We allow only one "Extra" case to be checked before the switch
657 // Failed to parse a proper sequence, abort now
664 } // end anonymous namespace
666 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
667 Instruction *Cond = nullptr;
668 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
669 Cond = dyn_cast<Instruction>(SI->getCondition());
670 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
671 if (BI->isConditional())
672 Cond = dyn_cast<Instruction>(BI->getCondition());
673 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
674 Cond = dyn_cast<Instruction>(IBI->getAddress());
677 TI->eraseFromParent();
679 RecursivelyDeleteTriviallyDeadInstructions(Cond);
682 /// Return true if the specified terminator checks
683 /// to see if a value is equal to constant integer value.
684 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
686 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
687 // Do not permit merging of large switch instructions into their
688 // predecessors unless there is only one predecessor.
689 if (SI->getNumSuccessors() * std::distance(pred_begin(SI->getParent()),
690 pred_end(SI->getParent())) <=
692 CV = SI->getCondition();
693 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
694 if (BI->isConditional() && BI->getCondition()->hasOneUse())
695 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
696 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
697 CV = ICI->getOperand(0);
700 // Unwrap any lossless ptrtoint cast.
702 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
703 Value *Ptr = PTII->getPointerOperand();
704 if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
711 /// Given a value comparison instruction,
712 /// decode all of the 'cases' that it represents and return the 'default' block.
713 BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
714 TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
715 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
716 Cases.reserve(SI->getNumCases());
717 for (auto Case : SI->cases())
718 Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),
719 Case.getCaseSuccessor()));
720 return SI->getDefaultDest();
723 BranchInst *BI = cast<BranchInst>(TI);
724 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
725 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
726 Cases.push_back(ValueEqualityComparisonCase(
727 GetConstantInt(ICI->getOperand(1), DL), Succ));
728 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
731 /// Given a vector of bb/value pairs, remove any entries
732 /// in the list that match the specified block.
734 EliminateBlockCases(BasicBlock *BB,
735 std::vector<ValueEqualityComparisonCase> &Cases) {
736 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
739 /// Return true if there are any keys in C1 that exist in C2 as well.
740 static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
741 std::vector<ValueEqualityComparisonCase> &C2) {
742 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
744 // Make V1 be smaller than V2.
745 if (V1->size() > V2->size())
750 if (V1->size() == 1) {
752 ConstantInt *TheVal = (*V1)[0].Value;
753 for (unsigned i = 0, e = V2->size(); i != e; ++i)
754 if (TheVal == (*V2)[i].Value)
758 // Otherwise, just sort both lists and compare element by element.
759 array_pod_sort(V1->begin(), V1->end());
760 array_pod_sort(V2->begin(), V2->end());
761 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
762 while (i1 != e1 && i2 != e2) {
763 if ((*V1)[i1].Value == (*V2)[i2].Value)
765 if ((*V1)[i1].Value < (*V2)[i2].Value)
773 /// If TI is known to be a terminator instruction and its block is known to
774 /// only have a single predecessor block, check to see if that predecessor is
775 /// also a value comparison with the same value, and if that comparison
776 /// determines the outcome of this comparison. If so, simplify TI. This does a
777 /// very limited form of jump threading.
778 bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
779 TerminatorInst *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
780 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
782 return false; // Not a value comparison in predecessor.
784 Value *ThisVal = isValueEqualityComparison(TI);
785 assert(ThisVal && "This isn't a value comparison!!");
786 if (ThisVal != PredVal)
787 return false; // Different predicates.
789 // TODO: Preserve branch weight metadata, similarly to how
790 // FoldValueComparisonIntoPredecessors preserves it.
792 // Find out information about when control will move from Pred to TI's block.
793 std::vector<ValueEqualityComparisonCase> PredCases;
794 BasicBlock *PredDef =
795 GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
796 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
798 // Find information about how control leaves this block.
799 std::vector<ValueEqualityComparisonCase> ThisCases;
800 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
801 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
803 // If TI's block is the default block from Pred's comparison, potentially
804 // simplify TI based on this knowledge.
805 if (PredDef == TI->getParent()) {
806 // If we are here, we know that the value is none of those cases listed in
807 // PredCases. If there are any cases in ThisCases that are in PredCases, we
809 if (!ValuesOverlap(PredCases, ThisCases))
812 if (isa<BranchInst>(TI)) {
813 // Okay, one of the successors of this condbr is dead. Convert it to a
815 assert(ThisCases.size() == 1 && "Branch can only have one case!");
816 // Insert the new branch.
817 Instruction *NI = Builder.CreateBr(ThisDef);
820 // Remove PHI node entries for the dead edge.
821 ThisCases[0].Dest->removePredecessor(TI->getParent());
823 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
824 << "Through successor TI: " << *TI << "Leaving: " << *NI
827 EraseTerminatorInstAndDCECond(TI);
831 SwitchInst *SI = cast<SwitchInst>(TI);
832 // Okay, TI has cases that are statically dead, prune them away.
833 SmallPtrSet<Constant *, 16> DeadCases;
834 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
835 DeadCases.insert(PredCases[i].Value);
837 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
838 << "Through successor TI: " << *TI);
840 // Collect branch weights into a vector.
841 SmallVector<uint32_t, 8> Weights;
842 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
843 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
845 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
847 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
848 Weights.push_back(CI->getValue().getZExtValue());
850 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
852 if (DeadCases.count(i->getCaseValue())) {
854 std::swap(Weights[i->getCaseIndex() + 1], Weights.back());
857 i->getCaseSuccessor()->removePredecessor(TI->getParent());
861 if (HasWeight && Weights.size() >= 2)
862 SI->setMetadata(LLVMContext::MD_prof,
863 MDBuilder(SI->getParent()->getContext())
864 .createBranchWeights(Weights));
866 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
870 // Otherwise, TI's block must correspond to some matched value. Find out
871 // which value (or set of values) this is.
872 ConstantInt *TIV = nullptr;
873 BasicBlock *TIBB = TI->getParent();
874 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
875 if (PredCases[i].Dest == TIBB) {
877 return false; // Cannot handle multiple values coming to this block.
878 TIV = PredCases[i].Value;
880 assert(TIV && "No edge from pred to succ?");
882 // Okay, we found the one constant that our value can be if we get into TI's
883 // BB. Find out which successor will unconditionally be branched to.
884 BasicBlock *TheRealDest = nullptr;
885 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
886 if (ThisCases[i].Value == TIV) {
887 TheRealDest = ThisCases[i].Dest;
891 // If not handled by any explicit cases, it is handled by the default case.
893 TheRealDest = ThisDef;
895 // Remove PHI node entries for dead edges.
896 BasicBlock *CheckEdge = TheRealDest;
897 for (BasicBlock *Succ : successors(TIBB))
898 if (Succ != CheckEdge)
899 Succ->removePredecessor(TIBB);
903 // Insert the new branch.
904 Instruction *NI = Builder.CreateBr(TheRealDest);
907 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
908 << "Through successor TI: " << *TI << "Leaving: " << *NI
911 EraseTerminatorInstAndDCECond(TI);
917 /// This class implements a stable ordering of constant
918 /// integers that does not depend on their address. This is important for
919 /// applications that sort ConstantInt's to ensure uniqueness.
920 struct ConstantIntOrdering {
921 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
922 return LHS->getValue().ult(RHS->getValue());
926 } // end anonymous namespace
928 static int ConstantIntSortPredicate(ConstantInt *const *P1,
929 ConstantInt *const *P2) {
930 const ConstantInt *LHS = *P1;
931 const ConstantInt *RHS = *P2;
934 return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
937 static inline bool HasBranchWeights(const Instruction *I) {
938 MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
939 if (ProfMD && ProfMD->getOperand(0))
940 if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
941 return MDS->getString().equals("branch_weights");
946 /// Get Weights of a given TerminatorInst, the default weight is at the front
947 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
949 static void GetBranchWeights(TerminatorInst *TI,
950 SmallVectorImpl<uint64_t> &Weights) {
951 MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
953 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
954 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
955 Weights.push_back(CI->getValue().getZExtValue());
958 // If TI is a conditional eq, the default case is the false case,
959 // and the corresponding branch-weight data is at index 2. We swap the
960 // default weight to be the first entry.
961 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
962 assert(Weights.size() == 2);
963 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
964 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
965 std::swap(Weights.front(), Weights.back());
969 /// Keep halving the weights until all can fit in uint32_t.
970 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
971 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
972 if (Max > UINT_MAX) {
973 unsigned Offset = 32 - countLeadingZeros(Max);
974 for (uint64_t &I : Weights)
979 /// The specified terminator is a value equality comparison instruction
980 /// (either a switch or a branch on "X == c").
981 /// See if any of the predecessors of the terminator block are value comparisons
982 /// on the same value. If so, and if safe to do so, fold them together.
983 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
984 IRBuilder<> &Builder) {
985 BasicBlock *BB = TI->getParent();
986 Value *CV = isValueEqualityComparison(TI); // CondVal
987 assert(CV && "Not a comparison?");
988 bool Changed = false;
990 SmallVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
991 while (!Preds.empty()) {
992 BasicBlock *Pred = Preds.pop_back_val();
994 // See if the predecessor is a comparison with the same value.
995 TerminatorInst *PTI = Pred->getTerminator();
996 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
998 if (PCV == CV && TI != PTI) {
999 SmallSetVector<BasicBlock*, 4> FailBlocks;
1000 if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
1001 for (auto *Succ : FailBlocks) {
1002 if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split"))
1007 // Figure out which 'cases' to copy from SI to PSI.
1008 std::vector<ValueEqualityComparisonCase> BBCases;
1009 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1011 std::vector<ValueEqualityComparisonCase> PredCases;
1012 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1014 // Based on whether the default edge from PTI goes to BB or not, fill in
1015 // PredCases and PredDefault with the new switch cases we would like to
1017 SmallVector<BasicBlock *, 8> NewSuccessors;
1019 // Update the branch weight metadata along the way
1020 SmallVector<uint64_t, 8> Weights;
1021 bool PredHasWeights = HasBranchWeights(PTI);
1022 bool SuccHasWeights = HasBranchWeights(TI);
1024 if (PredHasWeights) {
1025 GetBranchWeights(PTI, Weights);
1026 // branch-weight metadata is inconsistent here.
1027 if (Weights.size() != 1 + PredCases.size())
1028 PredHasWeights = SuccHasWeights = false;
1029 } else if (SuccHasWeights)
1030 // If there are no predecessor weights but there are successor weights,
1031 // populate Weights with 1, which will later be scaled to the sum of
1032 // successor's weights
1033 Weights.assign(1 + PredCases.size(), 1);
1035 SmallVector<uint64_t, 8> SuccWeights;
1036 if (SuccHasWeights) {
1037 GetBranchWeights(TI, SuccWeights);
1038 // branch-weight metadata is inconsistent here.
1039 if (SuccWeights.size() != 1 + BBCases.size())
1040 PredHasWeights = SuccHasWeights = false;
1041 } else if (PredHasWeights)
1042 SuccWeights.assign(1 + BBCases.size(), 1);
1044 if (PredDefault == BB) {
1045 // If this is the default destination from PTI, only the edges in TI
1046 // that don't occur in PTI, or that branch to BB will be activated.
1047 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1048 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1049 if (PredCases[i].Dest != BB)
1050 PTIHandled.insert(PredCases[i].Value);
1052 // The default destination is BB, we don't need explicit targets.
1053 std::swap(PredCases[i], PredCases.back());
1055 if (PredHasWeights || SuccHasWeights) {
1056 // Increase weight for the default case.
1057 Weights[0] += Weights[i + 1];
1058 std::swap(Weights[i + 1], Weights.back());
1062 PredCases.pop_back();
1067 // Reconstruct the new switch statement we will be building.
1068 if (PredDefault != BBDefault) {
1069 PredDefault->removePredecessor(Pred);
1070 PredDefault = BBDefault;
1071 NewSuccessors.push_back(BBDefault);
1074 unsigned CasesFromPred = Weights.size();
1075 uint64_t ValidTotalSuccWeight = 0;
1076 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1077 if (!PTIHandled.count(BBCases[i].Value) &&
1078 BBCases[i].Dest != BBDefault) {
1079 PredCases.push_back(BBCases[i]);
1080 NewSuccessors.push_back(BBCases[i].Dest);
1081 if (SuccHasWeights || PredHasWeights) {
1082 // The default weight is at index 0, so weight for the ith case
1083 // should be at index i+1. Scale the cases from successor by
1084 // PredDefaultWeight (Weights[0]).
1085 Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1086 ValidTotalSuccWeight += SuccWeights[i + 1];
1090 if (SuccHasWeights || PredHasWeights) {
1091 ValidTotalSuccWeight += SuccWeights[0];
1092 // Scale the cases from predecessor by ValidTotalSuccWeight.
1093 for (unsigned i = 1; i < CasesFromPred; ++i)
1094 Weights[i] *= ValidTotalSuccWeight;
1095 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1096 Weights[0] *= SuccWeights[0];
1099 // If this is not the default destination from PSI, only the edges
1100 // in SI that occur in PSI with a destination of BB will be
1102 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1103 std::map<ConstantInt *, uint64_t> WeightsForHandled;
1104 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1105 if (PredCases[i].Dest == BB) {
1106 PTIHandled.insert(PredCases[i].Value);
1108 if (PredHasWeights || SuccHasWeights) {
1109 WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1110 std::swap(Weights[i + 1], Weights.back());
1114 std::swap(PredCases[i], PredCases.back());
1115 PredCases.pop_back();
1120 // Okay, now we know which constants were sent to BB from the
1121 // predecessor. Figure out where they will all go now.
1122 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1123 if (PTIHandled.count(BBCases[i].Value)) {
1124 // If this is one we are capable of getting...
1125 if (PredHasWeights || SuccHasWeights)
1126 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1127 PredCases.push_back(BBCases[i]);
1128 NewSuccessors.push_back(BBCases[i].Dest);
1130 BBCases[i].Value); // This constant is taken care of
1133 // If there are any constants vectored to BB that TI doesn't handle,
1134 // they must go to the default destination of TI.
1135 for (ConstantInt *I : PTIHandled) {
1136 if (PredHasWeights || SuccHasWeights)
1137 Weights.push_back(WeightsForHandled[I]);
1138 PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1139 NewSuccessors.push_back(BBDefault);
1143 // Okay, at this point, we know which new successor Pred will get. Make
1144 // sure we update the number of entries in the PHI nodes for these
1146 for (BasicBlock *NewSuccessor : NewSuccessors)
1147 AddPredecessorToBlock(NewSuccessor, Pred, BB);
1149 Builder.SetInsertPoint(PTI);
1150 // Convert pointer to int before we switch.
1151 if (CV->getType()->isPointerTy()) {
1152 CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
1156 // Now that the successors are updated, create the new Switch instruction.
1158 Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1159 NewSI->setDebugLoc(PTI->getDebugLoc());
1160 for (ValueEqualityComparisonCase &V : PredCases)
1161 NewSI->addCase(V.Value, V.Dest);
1163 if (PredHasWeights || SuccHasWeights) {
1164 // Halve the weights if any of them cannot fit in an uint32_t
1165 FitWeights(Weights);
1167 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1170 LLVMContext::MD_prof,
1171 MDBuilder(BB->getContext()).createBranchWeights(MDWeights));
1174 EraseTerminatorInstAndDCECond(PTI);
1176 // Okay, last check. If BB is still a successor of PSI, then we must
1177 // have an infinite loop case. If so, add an infinitely looping block
1178 // to handle the case to preserve the behavior of the code.
1179 BasicBlock *InfLoopBlock = nullptr;
1180 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1181 if (NewSI->getSuccessor(i) == BB) {
1182 if (!InfLoopBlock) {
1183 // Insert it at the end of the function, because it's either code,
1184 // or it won't matter if it's hot. :)
1185 InfLoopBlock = BasicBlock::Create(BB->getContext(), "infloop",
1187 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1189 NewSI->setSuccessor(i, InfLoopBlock);
1198 // If we would need to insert a select that uses the value of this invoke
1199 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1200 // can't hoist the invoke, as there is nowhere to put the select in this case.
1201 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1202 Instruction *I1, Instruction *I2) {
1203 for (BasicBlock *Succ : successors(BB1)) {
1205 for (BasicBlock::iterator BBI = Succ->begin();
1206 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1207 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1208 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1209 if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1217 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I);
1219 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1220 /// in the two blocks up into the branch block. The caller of this function
1221 /// guarantees that BI's block dominates BB1 and BB2.
1222 static bool HoistThenElseCodeToIf(BranchInst *BI,
1223 const TargetTransformInfo &TTI) {
1224 // This does very trivial matching, with limited scanning, to find identical
1225 // instructions in the two blocks. In particular, we don't want to get into
1226 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1227 // such, we currently just scan for obviously identical instructions in an
1229 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1230 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1232 BasicBlock::iterator BB1_Itr = BB1->begin();
1233 BasicBlock::iterator BB2_Itr = BB2->begin();
1235 Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1236 // Skip debug info if it is not identical.
1237 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1238 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1239 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1240 while (isa<DbgInfoIntrinsic>(I1))
1242 while (isa<DbgInfoIntrinsic>(I2))
1245 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1246 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1249 BasicBlock *BIParent = BI->getParent();
1251 bool Changed = false;
1253 // If we are hoisting the terminator instruction, don't move one (making a
1254 // broken BB), instead clone it, and remove BI.
1255 if (isa<TerminatorInst>(I1))
1256 goto HoistTerminator;
1258 if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1261 // For a normal instruction, we just move one to right before the branch,
1262 // then replace all uses of the other with the first. Finally, we remove
1263 // the now redundant second instruction.
1264 BIParent->getInstList().splice(BI->getIterator(), BB1->getInstList(), I1);
1265 if (!I2->use_empty())
1266 I2->replaceAllUsesWith(I1);
1268 unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1269 LLVMContext::MD_range,
1270 LLVMContext::MD_fpmath,
1271 LLVMContext::MD_invariant_load,
1272 LLVMContext::MD_nonnull,
1273 LLVMContext::MD_invariant_group,
1274 LLVMContext::MD_align,
1275 LLVMContext::MD_dereferenceable,
1276 LLVMContext::MD_dereferenceable_or_null,
1277 LLVMContext::MD_mem_parallel_loop_access};
1278 combineMetadata(I1, I2, KnownIDs);
1280 // I1 and I2 are being combined into a single instruction. Its debug
1281 // location is the merged locations of the original instructions.
1282 if (!isa<CallInst>(I1))
1284 DILocation::getMergedLocation(I1->getDebugLoc(), I2->getDebugLoc()));
1286 I2->eraseFromParent();
1291 // Skip debug info if it is not identical.
1292 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1293 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1294 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1295 while (isa<DbgInfoIntrinsic>(I1))
1297 while (isa<DbgInfoIntrinsic>(I2))
1300 } while (I1->isIdenticalToWhenDefined(I2));
1305 // It may not be possible to hoist an invoke.
1306 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1309 for (BasicBlock *Succ : successors(BB1)) {
1311 for (BasicBlock::iterator BBI = Succ->begin();
1312 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1313 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1314 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1318 // Check for passingValueIsAlwaysUndefined here because we would rather
1319 // eliminate undefined control flow then converting it to a select.
1320 if (passingValueIsAlwaysUndefined(BB1V, PN) ||
1321 passingValueIsAlwaysUndefined(BB2V, PN))
1324 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1326 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1331 // Okay, it is safe to hoist the terminator.
1332 Instruction *NT = I1->clone();
1333 BIParent->getInstList().insert(BI->getIterator(), NT);
1334 if (!NT->getType()->isVoidTy()) {
1335 I1->replaceAllUsesWith(NT);
1336 I2->replaceAllUsesWith(NT);
1340 IRBuilder<NoFolder> Builder(NT);
1341 // Hoisting one of the terminators from our successor is a great thing.
1342 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1343 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1344 // nodes, so we insert select instruction to compute the final result.
1345 std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1346 for (BasicBlock *Succ : successors(BB1)) {
1348 for (BasicBlock::iterator BBI = Succ->begin();
1349 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1350 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1351 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1355 // These values do not agree. Insert a select instruction before NT
1356 // that determines the right value.
1357 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1359 SI = cast<SelectInst>(
1360 Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1361 BB1V->getName() + "." + BB2V->getName(), BI));
1363 // Make the PHI node use the select for all incoming values for BB1/BB2
1364 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1365 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1366 PN->setIncomingValue(i, SI);
1370 // Update any PHI nodes in our new successors.
1371 for (BasicBlock *Succ : successors(BB1))
1372 AddPredecessorToBlock(Succ, BIParent, BB1);
1374 EraseTerminatorInstAndDCECond(BI);
1378 // Is it legal to place a variable in operand \c OpIdx of \c I?
1379 // FIXME: This should be promoted to Instruction.
1380 static bool canReplaceOperandWithVariable(const Instruction *I,
1382 // We can't have a PHI with a metadata type.
1383 if (I->getOperand(OpIdx)->getType()->isMetadataTy())
1387 if (!isa<Constant>(I->getOperand(OpIdx)))
1390 switch (I->getOpcode()) {
1393 case Instruction::Call:
1394 case Instruction::Invoke:
1395 // FIXME: many arithmetic intrinsics have no issue taking a
1396 // variable, however it's hard to distingish these from
1397 // specials such as @llvm.frameaddress that require a constant.
1398 if (isa<IntrinsicInst>(I))
1401 // Constant bundle operands may need to retain their constant-ness for
1403 if (ImmutableCallSite(I).isBundleOperand(OpIdx))
1408 case Instruction::ShuffleVector:
1409 // Shufflevector masks are constant.
1411 case Instruction::ExtractValue:
1412 case Instruction::InsertValue:
1413 // All operands apart from the first are constant.
1415 case Instruction::Alloca:
1417 case Instruction::GetElementPtr:
1420 gep_type_iterator It = std::next(gep_type_begin(I), OpIdx - 1);
1421 return It.isSequential();
1425 // All instructions in Insts belong to different blocks that all unconditionally
1426 // branch to a common successor. Analyze each instruction and return true if it
1427 // would be possible to sink them into their successor, creating one common
1428 // instruction instead. For every value that would be required to be provided by
1429 // PHI node (because an operand varies in each input block), add to PHIOperands.
1430 static bool canSinkInstructions(
1431 ArrayRef<Instruction *> Insts,
1432 DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1433 // Prune out obviously bad instructions to move. Any non-store instruction
1434 // must have exactly one use, and we check later that use is by a single,
1435 // common PHI instruction in the successor.
1436 for (auto *I : Insts) {
1437 // These instructions may change or break semantics if moved.
1438 if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1439 I->getType()->isTokenTy())
1442 // Conservatively return false if I is an inline-asm instruction. Sinking
1443 // and merging inline-asm instructions can potentially create arguments
1444 // that cannot satisfy the inline-asm constraints.
1445 if (const auto *C = dyn_cast<CallInst>(I))
1446 if (C->isInlineAsm())
1449 // Everything must have only one use too, apart from stores which
1451 if (!isa<StoreInst>(I) && !I->hasOneUse())
1455 const Instruction *I0 = Insts.front();
1456 for (auto *I : Insts)
1457 if (!I->isSameOperationAs(I0))
1460 // All instructions in Insts are known to be the same opcode. If they aren't
1461 // stores, check the only user of each is a PHI or in the same block as the
1462 // instruction, because if a user is in the same block as an instruction
1463 // we're contemplating sinking, it must already be determined to be sinkable.
1464 if (!isa<StoreInst>(I0)) {
1465 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1466 auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1467 if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1468 auto *U = cast<Instruction>(*I->user_begin());
1470 PNUse->getParent() == Succ &&
1471 PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1472 U->getParent() == I->getParent();
1477 // Because SROA can't handle speculating stores of selects, try not
1478 // to sink loads or stores of allocas when we'd have to create a PHI for
1479 // the address operand. Also, because it is likely that loads or stores
1480 // of allocas will disappear when Mem2Reg/SROA is run, don't sink them.
1481 // This can cause code churn which can have unintended consequences down
1482 // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1483 // FIXME: This is a workaround for a deficiency in SROA - see
1484 // https://llvm.org/bugs/show_bug.cgi?id=30188
1485 if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1486 return isa<AllocaInst>(I->getOperand(1));
1489 if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1490 return isa<AllocaInst>(I->getOperand(0));
1494 for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1495 if (I0->getOperand(OI)->getType()->isTokenTy())
1496 // Don't touch any operand of token type.
1499 auto SameAsI0 = [&I0, OI](const Instruction *I) {
1500 assert(I->getNumOperands() == I0->getNumOperands());
1501 return I->getOperand(OI) == I0->getOperand(OI);
1503 if (!all_of(Insts, SameAsI0)) {
1504 if (!canReplaceOperandWithVariable(I0, OI))
1505 // We can't create a PHI from this GEP.
1507 // Don't create indirect calls! The called value is the final operand.
1508 if ((isa<CallInst>(I0) || isa<InvokeInst>(I0)) && OI == OE - 1) {
1509 // FIXME: if the call was *already* indirect, we should do this.
1512 for (auto *I : Insts)
1513 PHIOperands[I].push_back(I->getOperand(OI));
1519 // Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1520 // instruction of every block in Blocks to their common successor, commoning
1521 // into one instruction.
1522 static bool sinkLastInstruction(ArrayRef<BasicBlock*> Blocks) {
1523 auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1525 // canSinkLastInstruction returning true guarantees that every block has at
1526 // least one non-terminator instruction.
1527 SmallVector<Instruction*,4> Insts;
1528 for (auto *BB : Blocks) {
1529 Instruction *I = BB->getTerminator();
1531 I = I->getPrevNode();
1532 } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1533 if (!isa<DbgInfoIntrinsic>(I))
1537 // The only checking we need to do now is that all users of all instructions
1538 // are the same PHI node. canSinkLastInstruction should have checked this but
1539 // it is slightly over-aggressive - it gets confused by commutative instructions
1540 // so double-check it here.
1541 Instruction *I0 = Insts.front();
1542 if (!isa<StoreInst>(I0)) {
1543 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1544 if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1545 auto *U = cast<Instruction>(*I->user_begin());
1551 // We don't need to do any more checking here; canSinkLastInstruction should
1552 // have done it all for us.
1553 SmallVector<Value*, 4> NewOperands;
1554 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1555 // This check is different to that in canSinkLastInstruction. There, we
1556 // cared about the global view once simplifycfg (and instcombine) have
1557 // completed - it takes into account PHIs that become trivially
1558 // simplifiable. However here we need a more local view; if an operand
1559 // differs we create a PHI and rely on instcombine to clean up the very
1560 // small mess we may make.
1561 bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1562 return I->getOperand(O) != I0->getOperand(O);
1565 NewOperands.push_back(I0->getOperand(O));
1569 // Create a new PHI in the successor block and populate it.
1570 auto *Op = I0->getOperand(O);
1571 assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1572 auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1573 Op->getName() + ".sink", &BBEnd->front());
1574 for (auto *I : Insts)
1575 PN->addIncoming(I->getOperand(O), I->getParent());
1576 NewOperands.push_back(PN);
1579 // Arbitrarily use I0 as the new "common" instruction; remap its operands
1580 // and move it to the start of the successor block.
1581 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1582 I0->getOperandUse(O).set(NewOperands[O]);
1583 I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1585 // The debug location for the "common" instruction is the merged locations of
1586 // all the commoned instructions. We start with the original location of the
1587 // "common" instruction and iteratively merge each location in the loop below.
1588 const DILocation *Loc = I0->getDebugLoc();
1590 // Update metadata and IR flags, and merge debug locations.
1591 for (auto *I : Insts)
1593 Loc = DILocation::getMergedLocation(Loc, I->getDebugLoc());
1594 combineMetadataForCSE(I0, I);
1597 if (!isa<CallInst>(I0))
1598 I0->setDebugLoc(Loc);
1600 if (!isa<StoreInst>(I0)) {
1601 // canSinkLastInstruction checked that all instructions were used by
1602 // one and only one PHI node. Find that now, RAUW it to our common
1603 // instruction and nuke it.
1604 assert(I0->hasOneUse());
1605 auto *PN = cast<PHINode>(*I0->user_begin());
1606 PN->replaceAllUsesWith(I0);
1607 PN->eraseFromParent();
1610 // Finally nuke all instructions apart from the common instruction.
1611 for (auto *I : Insts)
1613 I->eraseFromParent();
1620 // LockstepReverseIterator - Iterates through instructions
1621 // in a set of blocks in reverse order from the first non-terminator.
1622 // For example (assume all blocks have size n):
1623 // LockstepReverseIterator I([B1, B2, B3]);
1624 // *I-- = [B1[n], B2[n], B3[n]];
1625 // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1626 // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1628 class LockstepReverseIterator {
1629 ArrayRef<BasicBlock*> Blocks;
1630 SmallVector<Instruction*,4> Insts;
1633 LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) :
1641 for (auto *BB : Blocks) {
1642 Instruction *Inst = BB->getTerminator();
1643 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1644 Inst = Inst->getPrevNode();
1646 // Block wasn't big enough.
1650 Insts.push_back(Inst);
1654 bool isValid() const {
1658 void operator -- () {
1661 for (auto *&Inst : Insts) {
1662 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1663 Inst = Inst->getPrevNode();
1664 // Already at beginning of block.
1672 ArrayRef<Instruction*> operator * () const {
1677 } // end anonymous namespace
1679 /// Given an unconditional branch that goes to BBEnd,
1680 /// check whether BBEnd has only two predecessors and the other predecessor
1681 /// ends with an unconditional branch. If it is true, sink any common code
1682 /// in the two predecessors to BBEnd.
1683 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1684 assert(BI1->isUnconditional());
1685 BasicBlock *BBEnd = BI1->getSuccessor(0);
1687 // We support two situations:
1688 // (1) all incoming arcs are unconditional
1689 // (2) one incoming arc is conditional
1691 // (2) is very common in switch defaults and
1692 // else-if patterns;
1695 // else if (b) f(2);
1708 // [end] has two unconditional predecessor arcs and one conditional. The
1709 // conditional refers to the implicit empty 'else' arc. This conditional
1710 // arc can also be caused by an empty default block in a switch.
1712 // In this case, we attempt to sink code from all *unconditional* arcs.
1713 // If we can sink instructions from these arcs (determined during the scan
1714 // phase below) we insert a common successor for all unconditional arcs and
1715 // connect that to [end], to enable sinking:
1728 SmallVector<BasicBlock*,4> UnconditionalPreds;
1729 Instruction *Cond = nullptr;
1730 for (auto *B : predecessors(BBEnd)) {
1731 auto *T = B->getTerminator();
1732 if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1733 UnconditionalPreds.push_back(B);
1734 else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1739 if (UnconditionalPreds.size() < 2)
1742 bool Changed = false;
1743 // We take a two-step approach to tail sinking. First we scan from the end of
1744 // each block upwards in lockstep. If the n'th instruction from the end of each
1745 // block can be sunk, those instructions are added to ValuesToSink and we
1746 // carry on. If we can sink an instruction but need to PHI-merge some operands
1747 // (because they're not identical in each instruction) we add these to
1749 unsigned ScanIdx = 0;
1750 SmallPtrSet<Value*,4> InstructionsToSink;
1751 DenseMap<Instruction*, SmallVector<Value*,4>> PHIOperands;
1752 LockstepReverseIterator LRI(UnconditionalPreds);
1753 while (LRI.isValid() &&
1754 canSinkInstructions(*LRI, PHIOperands)) {
1755 DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0] << "\n");
1756 InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1761 auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1762 unsigned NumPHIdValues = 0;
1763 for (auto *I : *LRI)
1764 for (auto *V : PHIOperands[I])
1765 if (InstructionsToSink.count(V) == 0)
1767 DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1768 unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1769 if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1772 return NumPHIInsts <= 1;
1775 if (ScanIdx > 0 && Cond) {
1776 // Check if we would actually sink anything first! This mutates the CFG and
1777 // adds an extra block. The goal in doing this is to allow instructions that
1778 // couldn't be sunk before to be sunk - obviously, speculatable instructions
1779 // (such as trunc, add) can be sunk and predicated already. So we check that
1780 // we're going to sink at least one non-speculatable instruction.
1783 bool Profitable = false;
1784 while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1785 if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1795 DEBUG(dbgs() << "SINK: Splitting edge\n");
1796 // We have a conditional edge and we're going to sink some instructions.
1797 // Insert a new block postdominating all blocks we're going to sink from.
1798 if (!SplitBlockPredecessors(BI1->getSuccessor(0), UnconditionalPreds,
1800 // Edges couldn't be split.
1805 // Now that we've analyzed all potential sinking candidates, perform the
1806 // actual sink. We iteratively sink the last non-terminator of the source
1807 // blocks into their common successor unless doing so would require too
1808 // many PHI instructions to be generated (currently only one PHI is allowed
1809 // per sunk instruction).
1811 // We can use InstructionsToSink to discount values needing PHI-merging that will
1812 // actually be sunk in a later iteration. This allows us to be more
1813 // aggressive in what we sink. This does allow a false positive where we
1814 // sink presuming a later value will also be sunk, but stop half way through
1815 // and never actually sink it which means we produce more PHIs than intended.
1816 // This is unlikely in practice though.
1817 for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1818 DEBUG(dbgs() << "SINK: Sink: "
1819 << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1822 // Because we've sunk every instruction in turn, the current instruction to
1823 // sink is always at index 0.
1825 if (!ProfitableToSinkInstruction(LRI)) {
1826 // Too many PHIs would be created.
1827 DEBUG(dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1831 if (!sinkLastInstruction(UnconditionalPreds))
1839 /// \brief Determine if we can hoist sink a sole store instruction out of a
1840 /// conditional block.
1842 /// We are looking for code like the following:
1844 /// store i32 %add, i32* %arrayidx2
1845 /// ... // No other stores or function calls (we could be calling a memory
1846 /// ... // function).
1847 /// %cmp = icmp ult %x, %y
1848 /// br i1 %cmp, label %EndBB, label %ThenBB
1850 /// store i32 %add5, i32* %arrayidx2
1854 /// We are going to transform this into:
1856 /// store i32 %add, i32* %arrayidx2
1858 /// %cmp = icmp ult %x, %y
1859 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1860 /// store i32 %add.add5, i32* %arrayidx2
1863 /// \return The pointer to the value of the previous store if the store can be
1864 /// hoisted into the predecessor block. 0 otherwise.
1865 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1866 BasicBlock *StoreBB, BasicBlock *EndBB) {
1867 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1871 // Volatile or atomic.
1872 if (!StoreToHoist->isSimple())
1875 Value *StorePtr = StoreToHoist->getPointerOperand();
1877 // Look for a store to the same pointer in BrBB.
1878 unsigned MaxNumInstToLookAt = 9;
1879 for (Instruction &CurI : reverse(*BrBB)) {
1880 if (!MaxNumInstToLookAt)
1883 if (isa<DbgInfoIntrinsic>(CurI))
1885 --MaxNumInstToLookAt;
1887 // Could be calling an instruction that affects memory like free().
1888 if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1891 if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1892 // Found the previous store make sure it stores to the same location.
1893 if (SI->getPointerOperand() == StorePtr)
1894 // Found the previous store, return its value operand.
1895 return SI->getValueOperand();
1896 return nullptr; // Unknown store.
1903 /// \brief Speculate a conditional basic block flattening the CFG.
1905 /// Note that this is a very risky transform currently. Speculating
1906 /// instructions like this is most often not desirable. Instead, there is an MI
1907 /// pass which can do it with full awareness of the resource constraints.
1908 /// However, some cases are "obvious" and we should do directly. An example of
1909 /// this is speculating a single, reasonably cheap instruction.
1911 /// There is only one distinct advantage to flattening the CFG at the IR level:
1912 /// it makes very common but simplistic optimizations such as are common in
1913 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1914 /// modeling their effects with easier to reason about SSA value graphs.
1917 /// An illustration of this transform is turning this IR:
1920 /// %cmp = icmp ult %x, %y
1921 /// br i1 %cmp, label %EndBB, label %ThenBB
1923 /// %sub = sub %x, %y
1926 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1933 /// %cmp = icmp ult %x, %y
1934 /// %sub = sub %x, %y
1935 /// %cond = select i1 %cmp, 0, %sub
1939 /// \returns true if the conditional block is removed.
1940 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1941 const TargetTransformInfo &TTI) {
1942 // Be conservative for now. FP select instruction can often be expensive.
1943 Value *BrCond = BI->getCondition();
1944 if (isa<FCmpInst>(BrCond))
1947 BasicBlock *BB = BI->getParent();
1948 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1950 // If ThenBB is actually on the false edge of the conditional branch, remember
1951 // to swap the select operands later.
1952 bool Invert = false;
1953 if (ThenBB != BI->getSuccessor(0)) {
1954 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1957 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1959 // Keep a count of how many times instructions are used within CondBB when
1960 // they are candidates for sinking into CondBB. Specifically:
1961 // - They are defined in BB, and
1962 // - They have no side effects, and
1963 // - All of their uses are in CondBB.
1964 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1966 unsigned SpeculationCost = 0;
1967 Value *SpeculatedStoreValue = nullptr;
1968 StoreInst *SpeculatedStore = nullptr;
1969 for (BasicBlock::iterator BBI = ThenBB->begin(),
1970 BBE = std::prev(ThenBB->end());
1971 BBI != BBE; ++BBI) {
1972 Instruction *I = &*BBI;
1974 if (isa<DbgInfoIntrinsic>(I))
1977 // Only speculatively execute a single instruction (not counting the
1978 // terminator) for now.
1980 if (SpeculationCost > 1)
1983 // Don't hoist the instruction if it's unsafe or expensive.
1984 if (!isSafeToSpeculativelyExecute(I) &&
1985 !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1986 I, BB, ThenBB, EndBB))))
1988 if (!SpeculatedStoreValue &&
1989 ComputeSpeculationCost(I, TTI) >
1990 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
1993 // Store the store speculation candidate.
1994 if (SpeculatedStoreValue)
1995 SpeculatedStore = cast<StoreInst>(I);
1997 // Do not hoist the instruction if any of its operands are defined but not
1998 // used in BB. The transformation will prevent the operand from
1999 // being sunk into the use block.
2000 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
2001 Instruction *OpI = dyn_cast<Instruction>(*i);
2002 if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
2003 continue; // Not a candidate for sinking.
2005 ++SinkCandidateUseCounts[OpI];
2009 // Consider any sink candidates which are only used in CondBB as costs for
2010 // speculation. Note, while we iterate over a DenseMap here, we are summing
2011 // and so iteration order isn't significant.
2012 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator
2013 I = SinkCandidateUseCounts.begin(),
2014 E = SinkCandidateUseCounts.end();
2016 if (I->first->getNumUses() == I->second) {
2018 if (SpeculationCost > 1)
2022 // Check that the PHI nodes can be converted to selects.
2023 bool HaveRewritablePHIs = false;
2024 for (BasicBlock::iterator I = EndBB->begin();
2025 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2026 Value *OrigV = PN->getIncomingValueForBlock(BB);
2027 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
2029 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2030 // Skip PHIs which are trivial.
2034 // Don't convert to selects if we could remove undefined behavior instead.
2035 if (passingValueIsAlwaysUndefined(OrigV, PN) ||
2036 passingValueIsAlwaysUndefined(ThenV, PN))
2039 HaveRewritablePHIs = true;
2040 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2041 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2042 if (!OrigCE && !ThenCE)
2043 continue; // Known safe and cheap.
2045 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2046 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2048 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2049 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2051 2 * PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2052 if (OrigCost + ThenCost > MaxCost)
2055 // Account for the cost of an unfolded ConstantExpr which could end up
2056 // getting expanded into Instructions.
2057 // FIXME: This doesn't account for how many operations are combined in the
2058 // constant expression.
2060 if (SpeculationCost > 1)
2064 // If there are no PHIs to process, bail early. This helps ensure idempotence
2066 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2069 // If we get here, we can hoist the instruction and if-convert.
2070 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2072 // Insert a select of the value of the speculated store.
2073 if (SpeculatedStoreValue) {
2074 IRBuilder<NoFolder> Builder(BI);
2075 Value *TrueV = SpeculatedStore->getValueOperand();
2076 Value *FalseV = SpeculatedStoreValue;
2078 std::swap(TrueV, FalseV);
2079 Value *S = Builder.CreateSelect(
2080 BrCond, TrueV, FalseV, TrueV->getName() + "." + FalseV->getName(), BI);
2081 SpeculatedStore->setOperand(0, S);
2082 SpeculatedStore->setDebugLoc(
2083 DILocation::getMergedLocation(
2084 BI->getDebugLoc(), SpeculatedStore->getDebugLoc()));
2087 // Metadata can be dependent on the condition we are hoisting above.
2088 // Conservatively strip all metadata on the instruction.
2089 for (auto &I : *ThenBB)
2090 I.dropUnknownNonDebugMetadata();
2092 // Hoist the instructions.
2093 BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2094 ThenBB->begin(), std::prev(ThenBB->end()));
2096 // Insert selects and rewrite the PHI operands.
2097 IRBuilder<NoFolder> Builder(BI);
2098 for (BasicBlock::iterator I = EndBB->begin();
2099 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2100 unsigned OrigI = PN->getBasicBlockIndex(BB);
2101 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
2102 Value *OrigV = PN->getIncomingValue(OrigI);
2103 Value *ThenV = PN->getIncomingValue(ThenI);
2105 // Skip PHIs which are trivial.
2109 // Create a select whose true value is the speculatively executed value and
2110 // false value is the preexisting value. Swap them if the branch
2111 // destinations were inverted.
2112 Value *TrueV = ThenV, *FalseV = OrigV;
2114 std::swap(TrueV, FalseV);
2115 Value *V = Builder.CreateSelect(
2116 BrCond, TrueV, FalseV, TrueV->getName() + "." + FalseV->getName(), BI);
2117 PN->setIncomingValue(OrigI, V);
2118 PN->setIncomingValue(ThenI, V);
2125 /// Return true if we can thread a branch across this block.
2126 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
2127 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
2130 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2131 if (isa<DbgInfoIntrinsic>(BBI))
2134 return false; // Don't clone large BB's.
2137 // We can only support instructions that do not define values that are
2138 // live outside of the current basic block.
2139 for (User *U : BBI->users()) {
2140 Instruction *UI = cast<Instruction>(U);
2141 if (UI->getParent() != BB || isa<PHINode>(UI))
2145 // Looks ok, continue checking.
2151 /// If we have a conditional branch on a PHI node value that is defined in the
2152 /// same block as the branch and if any PHI entries are constants, thread edges
2153 /// corresponding to that entry to be branches to their ultimate destination.
2154 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL,
2155 AssumptionCache *AC) {
2156 BasicBlock *BB = BI->getParent();
2157 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2158 // NOTE: we currently cannot transform this case if the PHI node is used
2159 // outside of the block.
2160 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2163 // Degenerate case of a single entry PHI.
2164 if (PN->getNumIncomingValues() == 1) {
2165 FoldSingleEntryPHINodes(PN->getParent());
2169 // Now we know that this block has multiple preds and two succs.
2170 if (!BlockIsSimpleEnoughToThreadThrough(BB))
2173 // Can't fold blocks that contain noduplicate or convergent calls.
2174 if (any_of(*BB, [](const Instruction &I) {
2175 const CallInst *CI = dyn_cast<CallInst>(&I);
2176 return CI && (CI->cannotDuplicate() || CI->isConvergent());
2180 // Okay, this is a simple enough basic block. See if any phi values are
2182 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2183 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
2184 if (!CB || !CB->getType()->isIntegerTy(1))
2187 // Okay, we now know that all edges from PredBB should be revectored to
2188 // branch to RealDest.
2189 BasicBlock *PredBB = PN->getIncomingBlock(i);
2190 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2193 continue; // Skip self loops.
2194 // Skip if the predecessor's terminator is an indirect branch.
2195 if (isa<IndirectBrInst>(PredBB->getTerminator()))
2198 // The dest block might have PHI nodes, other predecessors and other
2199 // difficult cases. Instead of being smart about this, just insert a new
2200 // block that jumps to the destination block, effectively splitting
2201 // the edge we are about to create.
2202 BasicBlock *EdgeBB =
2203 BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2204 RealDest->getParent(), RealDest);
2205 BranchInst::Create(RealDest, EdgeBB);
2207 // Update PHI nodes.
2208 AddPredecessorToBlock(RealDest, EdgeBB, BB);
2210 // BB may have instructions that are being threaded over. Clone these
2211 // instructions into EdgeBB. We know that there will be no uses of the
2212 // cloned instructions outside of EdgeBB.
2213 BasicBlock::iterator InsertPt = EdgeBB->begin();
2214 DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2215 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2216 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2217 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2220 // Clone the instruction.
2221 Instruction *N = BBI->clone();
2223 N->setName(BBI->getName() + ".c");
2225 // Update operands due to translation.
2226 for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2227 DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2228 if (PI != TranslateMap.end())
2232 // Check for trivial simplification.
2233 if (Value *V = SimplifyInstruction(N, DL)) {
2234 if (!BBI->use_empty())
2235 TranslateMap[&*BBI] = V;
2236 if (!N->mayHaveSideEffects()) {
2237 delete N; // Instruction folded away, don't need actual inst
2241 if (!BBI->use_empty())
2242 TranslateMap[&*BBI] = N;
2244 // Insert the new instruction into its new home.
2246 EdgeBB->getInstList().insert(InsertPt, N);
2248 // Register the new instruction with the assumption cache if necessary.
2249 if (auto *II = dyn_cast_or_null<IntrinsicInst>(N))
2250 if (II->getIntrinsicID() == Intrinsic::assume)
2251 AC->registerAssumption(II);
2254 // Loop over all of the edges from PredBB to BB, changing them to branch
2255 // to EdgeBB instead.
2256 TerminatorInst *PredBBTI = PredBB->getTerminator();
2257 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2258 if (PredBBTI->getSuccessor(i) == BB) {
2259 BB->removePredecessor(PredBB);
2260 PredBBTI->setSuccessor(i, EdgeBB);
2263 // Recurse, simplifying any other constants.
2264 return FoldCondBranchOnPHI(BI, DL, AC) | true;
2270 /// Given a BB that starts with the specified two-entry PHI node,
2271 /// see if we can eliminate it.
2272 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI,
2273 const DataLayout &DL) {
2274 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2275 // statement", which has a very simple dominance structure. Basically, we
2276 // are trying to find the condition that is being branched on, which
2277 // subsequently causes this merge to happen. We really want control
2278 // dependence information for this check, but simplifycfg can't keep it up
2279 // to date, and this catches most of the cases we care about anyway.
2280 BasicBlock *BB = PN->getParent();
2281 BasicBlock *IfTrue, *IfFalse;
2282 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2284 // Don't bother if the branch will be constant folded trivially.
2285 isa<ConstantInt>(IfCond))
2288 // Okay, we found that we can merge this two-entry phi node into a select.
2289 // Doing so would require us to fold *all* two entry phi nodes in this block.
2290 // At some point this becomes non-profitable (particularly if the target
2291 // doesn't support cmov's). Only do this transformation if there are two or
2292 // fewer PHI nodes in this block.
2293 unsigned NumPhis = 0;
2294 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2298 // Loop over the PHI's seeing if we can promote them all to select
2299 // instructions. While we are at it, keep track of the instructions
2300 // that need to be moved to the dominating block.
2301 SmallPtrSet<Instruction *, 4> AggressiveInsts;
2302 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2303 MaxCostVal1 = PHINodeFoldingThreshold;
2304 MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2305 MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2307 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2308 PHINode *PN = cast<PHINode>(II++);
2309 if (Value *V = SimplifyInstruction(PN, DL)) {
2310 PN->replaceAllUsesWith(V);
2311 PN->eraseFromParent();
2315 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
2316 MaxCostVal0, TTI) ||
2317 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
2322 // If we folded the first phi, PN dangles at this point. Refresh it. If
2323 // we ran out of PHIs then we simplified them all.
2324 PN = dyn_cast<PHINode>(BB->begin());
2328 // Don't fold i1 branches on PHIs which contain binary operators. These can
2329 // often be turned into switches and other things.
2330 if (PN->getType()->isIntegerTy(1) &&
2331 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2332 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2333 isa<BinaryOperator>(IfCond)))
2336 // If all PHI nodes are promotable, check to make sure that all instructions
2337 // in the predecessor blocks can be promoted as well. If not, we won't be able
2338 // to get rid of the control flow, so it's not worth promoting to select
2340 BasicBlock *DomBlock = nullptr;
2341 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2342 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2343 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2346 DomBlock = *pred_begin(IfBlock1);
2347 for (BasicBlock::iterator I = IfBlock1->begin(); !isa<TerminatorInst>(I);
2349 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2350 // This is not an aggressive instruction that we can promote.
2351 // Because of this, we won't be able to get rid of the control flow, so
2352 // the xform is not worth it.
2357 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2360 DomBlock = *pred_begin(IfBlock2);
2361 for (BasicBlock::iterator I = IfBlock2->begin(); !isa<TerminatorInst>(I);
2363 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2364 // This is not an aggressive instruction that we can promote.
2365 // Because of this, we won't be able to get rid of the control flow, so
2366 // the xform is not worth it.
2371 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
2372 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
2374 // If we can still promote the PHI nodes after this gauntlet of tests,
2375 // do all of the PHI's now.
2376 Instruction *InsertPt = DomBlock->getTerminator();
2377 IRBuilder<NoFolder> Builder(InsertPt);
2379 // Move all 'aggressive' instructions, which are defined in the
2380 // conditional parts of the if's up to the dominating block.
2382 for (auto &I : *IfBlock1)
2383 I.dropUnknownNonDebugMetadata();
2384 DomBlock->getInstList().splice(InsertPt->getIterator(),
2385 IfBlock1->getInstList(), IfBlock1->begin(),
2386 IfBlock1->getTerminator()->getIterator());
2389 for (auto &I : *IfBlock2)
2390 I.dropUnknownNonDebugMetadata();
2391 DomBlock->getInstList().splice(InsertPt->getIterator(),
2392 IfBlock2->getInstList(), IfBlock2->begin(),
2393 IfBlock2->getTerminator()->getIterator());
2396 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2397 // Change the PHI node into a select instruction.
2398 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2399 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2401 Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2402 PN->replaceAllUsesWith(Sel);
2404 PN->eraseFromParent();
2407 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2408 // has been flattened. Change DomBlock to jump directly to our new block to
2409 // avoid other simplifycfg's kicking in on the diamond.
2410 TerminatorInst *OldTI = DomBlock->getTerminator();
2411 Builder.SetInsertPoint(OldTI);
2412 Builder.CreateBr(BB);
2413 OldTI->eraseFromParent();
2417 /// If we found a conditional branch that goes to two returning blocks,
2418 /// try to merge them together into one return,
2419 /// introducing a select if the return values disagree.
2420 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
2421 IRBuilder<> &Builder) {
2422 assert(BI->isConditional() && "Must be a conditional branch");
2423 BasicBlock *TrueSucc = BI->getSuccessor(0);
2424 BasicBlock *FalseSucc = BI->getSuccessor(1);
2425 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2426 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2428 // Check to ensure both blocks are empty (just a return) or optionally empty
2429 // with PHI nodes. If there are other instructions, merging would cause extra
2430 // computation on one path or the other.
2431 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2433 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2436 Builder.SetInsertPoint(BI);
2437 // Okay, we found a branch that is going to two return nodes. If
2438 // there is no return value for this function, just change the
2439 // branch into a return.
2440 if (FalseRet->getNumOperands() == 0) {
2441 TrueSucc->removePredecessor(BI->getParent());
2442 FalseSucc->removePredecessor(BI->getParent());
2443 Builder.CreateRetVoid();
2444 EraseTerminatorInstAndDCECond(BI);
2448 // Otherwise, figure out what the true and false return values are
2449 // so we can insert a new select instruction.
2450 Value *TrueValue = TrueRet->getReturnValue();
2451 Value *FalseValue = FalseRet->getReturnValue();
2453 // Unwrap any PHI nodes in the return blocks.
2454 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2455 if (TVPN->getParent() == TrueSucc)
2456 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2457 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2458 if (FVPN->getParent() == FalseSucc)
2459 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2461 // In order for this transformation to be safe, we must be able to
2462 // unconditionally execute both operands to the return. This is
2463 // normally the case, but we could have a potentially-trapping
2464 // constant expression that prevents this transformation from being
2466 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2469 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2473 // Okay, we collected all the mapped values and checked them for sanity, and
2474 // defined to really do this transformation. First, update the CFG.
2475 TrueSucc->removePredecessor(BI->getParent());
2476 FalseSucc->removePredecessor(BI->getParent());
2478 // Insert select instructions where needed.
2479 Value *BrCond = BI->getCondition();
2481 // Insert a select if the results differ.
2482 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2483 } else if (isa<UndefValue>(TrueValue)) {
2484 TrueValue = FalseValue;
2487 Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2492 !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2496 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2497 << "\n " << *BI << "NewRet = " << *RI
2498 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: " << *FalseSucc);
2500 EraseTerminatorInstAndDCECond(BI);
2505 /// Return true if the given instruction is available
2506 /// in its predecessor block. If yes, the instruction will be removed.
2507 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
2508 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2510 for (Instruction &I : *PB) {
2511 Instruction *PBI = &I;
2512 // Check whether Inst and PBI generate the same value.
2513 if (Inst->isIdenticalTo(PBI)) {
2514 Inst->replaceAllUsesWith(PBI);
2515 Inst->eraseFromParent();
2522 /// Return true if either PBI or BI has branch weight available, and store
2523 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2524 /// not have branch weight, use 1:1 as its weight.
2525 static bool extractPredSuccWeights(BranchInst *PBI, BranchInst *BI,
2526 uint64_t &PredTrueWeight,
2527 uint64_t &PredFalseWeight,
2528 uint64_t &SuccTrueWeight,
2529 uint64_t &SuccFalseWeight) {
2530 bool PredHasWeights =
2531 PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2532 bool SuccHasWeights =
2533 BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2534 if (PredHasWeights || SuccHasWeights) {
2535 if (!PredHasWeights)
2536 PredTrueWeight = PredFalseWeight = 1;
2537 if (!SuccHasWeights)
2538 SuccTrueWeight = SuccFalseWeight = 1;
2545 /// If this basic block is simple enough, and if a predecessor branches to us
2546 /// and one of our successors, fold the block into the predecessor and use
2547 /// logical operations to pick the right destination.
2548 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2549 BasicBlock *BB = BI->getParent();
2551 Instruction *Cond = nullptr;
2552 if (BI->isConditional())
2553 Cond = dyn_cast<Instruction>(BI->getCondition());
2555 // For unconditional branch, check for a simple CFG pattern, where
2556 // BB has a single predecessor and BB's successor is also its predecessor's
2557 // successor. If such pattern exisits, check for CSE between BB and its
2559 if (BasicBlock *PB = BB->getSinglePredecessor())
2560 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2561 if (PBI->isConditional() &&
2562 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2563 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2564 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2565 Instruction *Curr = &*I++;
2566 if (isa<CmpInst>(Curr)) {
2570 // Quit if we can't remove this instruction.
2571 if (!checkCSEInPredecessor(Curr, PB))
2580 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2581 Cond->getParent() != BB || !Cond->hasOneUse())
2584 // Make sure the instruction after the condition is the cond branch.
2585 BasicBlock::iterator CondIt = ++Cond->getIterator();
2587 // Ignore dbg intrinsics.
2588 while (isa<DbgInfoIntrinsic>(CondIt))
2594 // Only allow this transformation if computing the condition doesn't involve
2595 // too many instructions and these involved instructions can be executed
2596 // unconditionally. We denote all involved instructions except the condition
2597 // as "bonus instructions", and only allow this transformation when the
2598 // number of the bonus instructions does not exceed a certain threshold.
2599 unsigned NumBonusInsts = 0;
2600 for (auto I = BB->begin(); Cond != &*I; ++I) {
2601 // Ignore dbg intrinsics.
2602 if (isa<DbgInfoIntrinsic>(I))
2604 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2606 // I has only one use and can be executed unconditionally.
2607 Instruction *User = dyn_cast<Instruction>(I->user_back());
2608 if (User == nullptr || User->getParent() != BB)
2610 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2611 // to use any other instruction, User must be an instruction between next(I)
2614 // Early exits once we reach the limit.
2615 if (NumBonusInsts > BonusInstThreshold)
2619 // Cond is known to be a compare or binary operator. Check to make sure that
2620 // neither operand is a potentially-trapping constant expression.
2621 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2624 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2628 // Finally, don't infinitely unroll conditional loops.
2629 BasicBlock *TrueDest = BI->getSuccessor(0);
2630 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2631 if (TrueDest == BB || FalseDest == BB)
2634 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2635 BasicBlock *PredBlock = *PI;
2636 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2638 // Check that we have two conditional branches. If there is a PHI node in
2639 // the common successor, verify that the same value flows in from both
2641 SmallVector<PHINode *, 4> PHIs;
2642 if (!PBI || PBI->isUnconditional() ||
2643 (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2644 (!BI->isConditional() &&
2645 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2648 // Determine if the two branches share a common destination.
2649 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2650 bool InvertPredCond = false;
2652 if (BI->isConditional()) {
2653 if (PBI->getSuccessor(0) == TrueDest) {
2654 Opc = Instruction::Or;
2655 } else if (PBI->getSuccessor(1) == FalseDest) {
2656 Opc = Instruction::And;
2657 } else if (PBI->getSuccessor(0) == FalseDest) {
2658 Opc = Instruction::And;
2659 InvertPredCond = true;
2660 } else if (PBI->getSuccessor(1) == TrueDest) {
2661 Opc = Instruction::Or;
2662 InvertPredCond = true;
2667 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2671 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2672 IRBuilder<> Builder(PBI);
2674 // If we need to invert the condition in the pred block to match, do so now.
2675 if (InvertPredCond) {
2676 Value *NewCond = PBI->getCondition();
2678 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2679 CmpInst *CI = cast<CmpInst>(NewCond);
2680 CI->setPredicate(CI->getInversePredicate());
2683 Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2686 PBI->setCondition(NewCond);
2687 PBI->swapSuccessors();
2690 // If we have bonus instructions, clone them into the predecessor block.
2691 // Note that there may be multiple predecessor blocks, so we cannot move
2692 // bonus instructions to a predecessor block.
2693 ValueToValueMapTy VMap; // maps original values to cloned values
2694 // We already make sure Cond is the last instruction before BI. Therefore,
2695 // all instructions before Cond other than DbgInfoIntrinsic are bonus
2697 for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2698 if (isa<DbgInfoIntrinsic>(BonusInst))
2700 Instruction *NewBonusInst = BonusInst->clone();
2701 RemapInstruction(NewBonusInst, VMap,
2702 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2703 VMap[&*BonusInst] = NewBonusInst;
2705 // If we moved a load, we cannot any longer claim any knowledge about
2706 // its potential value. The previous information might have been valid
2707 // only given the branch precondition.
2708 // For an analogous reason, we must also drop all the metadata whose
2709 // semantics we don't understand.
2710 NewBonusInst->dropUnknownNonDebugMetadata();
2712 PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2713 NewBonusInst->takeName(&*BonusInst);
2714 BonusInst->setName(BonusInst->getName() + ".old");
2717 // Clone Cond into the predecessor basic block, and or/and the
2718 // two conditions together.
2719 Instruction *New = Cond->clone();
2720 RemapInstruction(New, VMap,
2721 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2722 PredBlock->getInstList().insert(PBI->getIterator(), New);
2723 New->takeName(Cond);
2724 Cond->setName(New->getName() + ".old");
2726 if (BI->isConditional()) {
2727 Instruction *NewCond = cast<Instruction>(
2728 Builder.CreateBinOp(Opc, PBI->getCondition(), New, "or.cond"));
2729 PBI->setCondition(NewCond);
2731 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2733 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2734 SuccTrueWeight, SuccFalseWeight);
2735 SmallVector<uint64_t, 8> NewWeights;
2737 if (PBI->getSuccessor(0) == BB) {
2739 // PBI: br i1 %x, BB, FalseDest
2740 // BI: br i1 %y, TrueDest, FalseDest
2741 // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2742 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2743 // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2744 // TrueWeight for PBI * FalseWeight for BI.
2745 // We assume that total weights of a BranchInst can fit into 32 bits.
2746 // Therefore, we will not have overflow using 64-bit arithmetic.
2747 NewWeights.push_back(PredFalseWeight *
2748 (SuccFalseWeight + SuccTrueWeight) +
2749 PredTrueWeight * SuccFalseWeight);
2751 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2752 PBI->setSuccessor(0, TrueDest);
2754 if (PBI->getSuccessor(1) == BB) {
2756 // PBI: br i1 %x, TrueDest, BB
2757 // BI: br i1 %y, TrueDest, FalseDest
2758 // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2759 // FalseWeight for PBI * TrueWeight for BI.
2760 NewWeights.push_back(PredTrueWeight *
2761 (SuccFalseWeight + SuccTrueWeight) +
2762 PredFalseWeight * SuccTrueWeight);
2763 // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2764 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2766 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2767 PBI->setSuccessor(1, FalseDest);
2769 if (NewWeights.size() == 2) {
2770 // Halve the weights if any of them cannot fit in an uint32_t
2771 FitWeights(NewWeights);
2773 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2776 LLVMContext::MD_prof,
2777 MDBuilder(BI->getContext()).createBranchWeights(MDWeights));
2779 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2781 // Update PHI nodes in the common successors.
2782 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2783 ConstantInt *PBI_C = cast<ConstantInt>(
2784 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2785 assert(PBI_C->getType()->isIntegerTy(1));
2786 Instruction *MergedCond = nullptr;
2787 if (PBI->getSuccessor(0) == TrueDest) {
2788 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2789 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2790 // is false: !PBI_Cond and BI_Value
2791 Instruction *NotCond = cast<Instruction>(
2792 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2793 MergedCond = cast<Instruction>(
2794 Builder.CreateBinOp(Instruction::And, NotCond, New, "and.cond"));
2796 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2797 Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2799 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2800 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2801 // is false: PBI_Cond and BI_Value
2802 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2803 Instruction::And, PBI->getCondition(), New, "and.cond"));
2804 if (PBI_C->isOne()) {
2805 Instruction *NotCond = cast<Instruction>(
2806 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2807 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2808 Instruction::Or, NotCond, MergedCond, "or.cond"));
2812 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2815 // Change PBI from Conditional to Unconditional.
2816 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2817 EraseTerminatorInstAndDCECond(PBI);
2821 // If BI was a loop latch, it may have had associated loop metadata.
2822 // We need to copy it to the new latch, that is, PBI.
2823 if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2824 PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2826 // TODO: If BB is reachable from all paths through PredBlock, then we
2827 // could replace PBI's branch probabilities with BI's.
2829 // Copy any debug value intrinsics into the end of PredBlock.
2830 for (Instruction &I : *BB)
2831 if (isa<DbgInfoIntrinsic>(I))
2832 I.clone()->insertBefore(PBI);
2839 // If there is only one store in BB1 and BB2, return it, otherwise return
2841 static StoreInst *findUniqueStoreInBlocks(BasicBlock *BB1, BasicBlock *BB2) {
2842 StoreInst *S = nullptr;
2843 for (auto *BB : {BB1, BB2}) {
2847 if (auto *SI = dyn_cast<StoreInst>(&I)) {
2849 // Multiple stores seen.
2858 static Value *ensureValueAvailableInSuccessor(Value *V, BasicBlock *BB,
2859 Value *AlternativeV = nullptr) {
2860 // PHI is going to be a PHI node that allows the value V that is defined in
2861 // BB to be referenced in BB's only successor.
2863 // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2864 // doesn't matter to us what the other operand is (it'll never get used). We
2865 // could just create a new PHI with an undef incoming value, but that could
2866 // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2867 // other PHI. So here we directly look for some PHI in BB's successor with V
2868 // as an incoming operand. If we find one, we use it, else we create a new
2871 // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2872 // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2873 // where OtherBB is the single other predecessor of BB's only successor.
2874 PHINode *PHI = nullptr;
2875 BasicBlock *Succ = BB->getSingleSuccessor();
2877 for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2878 if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2879 PHI = cast<PHINode>(I);
2883 assert(std::distance(pred_begin(Succ), pred_end(Succ)) == 2);
2884 auto PredI = pred_begin(Succ);
2885 BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2886 if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2893 // If V is not an instruction defined in BB, just return it.
2894 if (!AlternativeV &&
2895 (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2898 PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2899 PHI->addIncoming(V, BB);
2900 for (BasicBlock *PredBB : predecessors(Succ))
2903 AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2907 static bool mergeConditionalStoreToAddress(BasicBlock *PTB, BasicBlock *PFB,
2908 BasicBlock *QTB, BasicBlock *QFB,
2909 BasicBlock *PostBB, Value *Address,
2910 bool InvertPCond, bool InvertQCond) {
2911 auto IsaBitcastOfPointerType = [](const Instruction &I) {
2912 return Operator::getOpcode(&I) == Instruction::BitCast &&
2913 I.getType()->isPointerTy();
2916 // If we're not in aggressive mode, we only optimize if we have some
2917 // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2918 auto IsWorthwhile = [&](BasicBlock *BB) {
2921 // Heuristic: if the block can be if-converted/phi-folded and the
2922 // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2923 // thread this store.
2925 for (auto &I : *BB) {
2926 // Cheap instructions viable for folding.
2927 if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2930 // Free instructions.
2931 else if (isa<TerminatorInst>(I) || isa<DbgInfoIntrinsic>(I) ||
2932 IsaBitcastOfPointerType(I))
2937 return N <= PHINodeFoldingThreshold;
2940 if (!MergeCondStoresAggressively &&
2941 (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2942 !IsWorthwhile(QFB)))
2945 // For every pointer, there must be exactly two stores, one coming from
2946 // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2947 // store (to any address) in PTB,PFB or QTB,QFB.
2948 // FIXME: We could relax this restriction with a bit more work and performance
2950 StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2951 StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2952 if (!PStore || !QStore)
2955 // Now check the stores are compatible.
2956 if (!QStore->isUnordered() || !PStore->isUnordered())
2959 // Check that sinking the store won't cause program behavior changes. Sinking
2960 // the store out of the Q blocks won't change any behavior as we're sinking
2961 // from a block to its unconditional successor. But we're moving a store from
2962 // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2963 // So we need to check that there are no aliasing loads or stores in
2964 // QBI, QTB and QFB. We also need to check there are no conflicting memory
2965 // operations between PStore and the end of its parent block.
2967 // The ideal way to do this is to query AliasAnalysis, but we don't
2968 // preserve AA currently so that is dangerous. Be super safe and just
2969 // check there are no other memory operations at all.
2970 for (auto &I : *QFB->getSinglePredecessor())
2971 if (I.mayReadOrWriteMemory())
2973 for (auto &I : *QFB)
2974 if (&I != QStore && I.mayReadOrWriteMemory())
2977 for (auto &I : *QTB)
2978 if (&I != QStore && I.mayReadOrWriteMemory())
2980 for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2982 if (&*I != PStore && I->mayReadOrWriteMemory())
2985 // OK, we're going to sink the stores to PostBB. The store has to be
2986 // conditional though, so first create the predicate.
2987 Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
2989 Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
2992 Value *PPHI = ensureValueAvailableInSuccessor(PStore->getValueOperand(),
2993 PStore->getParent());
2994 Value *QPHI = ensureValueAvailableInSuccessor(QStore->getValueOperand(),
2995 QStore->getParent(), PPHI);
2997 IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
2999 Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
3000 Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
3003 PPred = QB.CreateNot(PPred);
3005 QPred = QB.CreateNot(QPred);
3006 Value *CombinedPred = QB.CreateOr(PPred, QPred);
3009 SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
3010 QB.SetInsertPoint(T);
3011 StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
3013 PStore->getAAMetadata(AAMD, /*Merge=*/false);
3014 PStore->getAAMetadata(AAMD, /*Merge=*/true);
3015 SI->setAAMetadata(AAMD);
3017 QStore->eraseFromParent();
3018 PStore->eraseFromParent();
3023 static bool mergeConditionalStores(BranchInst *PBI, BranchInst *QBI) {
3024 // The intention here is to find diamonds or triangles (see below) where each
3025 // conditional block contains a store to the same address. Both of these
3026 // stores are conditional, so they can't be unconditionally sunk. But it may
3027 // be profitable to speculatively sink the stores into one merged store at the
3028 // end, and predicate the merged store on the union of the two conditions of
3031 // This can reduce the number of stores executed if both of the conditions are
3032 // true, and can allow the blocks to become small enough to be if-converted.
3033 // This optimization will also chain, so that ladders of test-and-set
3034 // sequences can be if-converted away.
3036 // We only deal with simple diamonds or triangles:
3038 // PBI or PBI or a combination of the two
3048 // We model triangles as a type of diamond with a nullptr "true" block.
3049 // Triangles are canonicalized so that the fallthrough edge is represented by
3050 // a true condition, as in the diagram above.
3052 BasicBlock *PTB = PBI->getSuccessor(0);
3053 BasicBlock *PFB = PBI->getSuccessor(1);
3054 BasicBlock *QTB = QBI->getSuccessor(0);
3055 BasicBlock *QFB = QBI->getSuccessor(1);
3056 BasicBlock *PostBB = QFB->getSingleSuccessor();
3058 bool InvertPCond = false, InvertQCond = false;
3059 // Canonicalize fallthroughs to the true branches.
3060 if (PFB == QBI->getParent()) {
3061 std::swap(PFB, PTB);
3064 if (QFB == PostBB) {
3065 std::swap(QFB, QTB);
3069 // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3070 // and QFB may not. Model fallthroughs as a nullptr block.
3071 if (PTB == QBI->getParent())
3076 // Legality bailouts. We must have at least the non-fallthrough blocks and
3077 // the post-dominating block, and the non-fallthroughs must only have one
3079 auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3080 return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3083 !HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3084 !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3086 if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3087 (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3089 if (!PostBB->hasNUses(2) || !QBI->getParent()->hasNUses(2))
3092 // OK, this is a sequence of two diamonds or triangles.
3093 // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3094 SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3095 for (auto *BB : {PTB, PFB}) {
3099 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3100 PStoreAddresses.insert(SI->getPointerOperand());
3102 for (auto *BB : {QTB, QFB}) {
3106 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3107 QStoreAddresses.insert(SI->getPointerOperand());
3110 set_intersect(PStoreAddresses, QStoreAddresses);
3111 // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3112 // clear what it contains.
3113 auto &CommonAddresses = PStoreAddresses;
3115 bool Changed = false;
3116 for (auto *Address : CommonAddresses)
3117 Changed |= mergeConditionalStoreToAddress(
3118 PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond);
3122 /// If we have a conditional branch as a predecessor of another block,
3123 /// this function tries to simplify it. We know
3124 /// that PBI and BI are both conditional branches, and BI is in one of the
3125 /// successor blocks of PBI - PBI branches to BI.
3126 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
3127 const DataLayout &DL) {
3128 assert(PBI->isConditional() && BI->isConditional());
3129 BasicBlock *BB = BI->getParent();
3131 // If this block ends with a branch instruction, and if there is a
3132 // predecessor that ends on a branch of the same condition, make
3133 // this conditional branch redundant.
3134 if (PBI->getCondition() == BI->getCondition() &&
3135 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3136 // Okay, the outcome of this conditional branch is statically
3137 // knowable. If this block had a single pred, handle specially.
3138 if (BB->getSinglePredecessor()) {
3139 // Turn this into a branch on constant.
3140 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3142 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3143 return true; // Nuke the branch on constant.
3146 // Otherwise, if there are multiple predecessors, insert a PHI that merges
3147 // in the constant and simplify the block result. Subsequent passes of
3148 // simplifycfg will thread the block.
3149 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
3150 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3151 PHINode *NewPN = PHINode::Create(
3152 Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3153 BI->getCondition()->getName() + ".pr", &BB->front());
3154 // Okay, we're going to insert the PHI node. Since PBI is not the only
3155 // predecessor, compute the PHI'd conditional value for all of the preds.
3156 // Any predecessor where the condition is not computable we keep symbolic.
3157 for (pred_iterator PI = PB; PI != PE; ++PI) {
3158 BasicBlock *P = *PI;
3159 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3160 PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3161 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3162 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3164 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3167 NewPN->addIncoming(BI->getCondition(), P);
3171 BI->setCondition(NewPN);
3176 if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3180 // If both branches are conditional and both contain stores to the same
3181 // address, remove the stores from the conditionals and create a conditional
3182 // merged store at the end.
3183 if (MergeCondStores && mergeConditionalStores(PBI, BI))
3186 // If this is a conditional branch in an empty block, and if any
3187 // predecessors are a conditional branch to one of our destinations,
3188 // fold the conditions into logical ops and one cond br.
3189 BasicBlock::iterator BBI = BB->begin();
3190 // Ignore dbg intrinsics.
3191 while (isa<DbgInfoIntrinsic>(BBI))
3197 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3200 } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3203 } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3206 } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3213 // Check to make sure that the other destination of this branch
3214 // isn't BB itself. If so, this is an infinite loop that will
3215 // keep getting unwound.
3216 if (PBI->getSuccessor(PBIOp) == BB)
3219 // Do not perform this transformation if it would require
3220 // insertion of a large number of select instructions. For targets
3221 // without predication/cmovs, this is a big pessimization.
3223 // Also do not perform this transformation if any phi node in the common
3224 // destination block can trap when reached by BB or PBB (PR17073). In that
3225 // case, it would be unsafe to hoist the operation into a select instruction.
3227 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3228 unsigned NumPhis = 0;
3229 for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3231 if (NumPhis > 2) // Disable this xform.
3234 PHINode *PN = cast<PHINode>(II);
3235 Value *BIV = PN->getIncomingValueForBlock(BB);
3236 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3240 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3241 Value *PBIV = PN->getIncomingValue(PBBIdx);
3242 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3247 // Finally, if everything is ok, fold the branches to logical ops.
3248 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3250 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3251 << "AND: " << *BI->getParent());
3253 // If OtherDest *is* BB, then BB is a basic block with a single conditional
3254 // branch in it, where one edge (OtherDest) goes back to itself but the other
3255 // exits. We don't *know* that the program avoids the infinite loop
3256 // (even though that seems likely). If we do this xform naively, we'll end up
3257 // recursively unpeeling the loop. Since we know that (after the xform is
3258 // done) that the block *is* infinite if reached, we just make it an obviously
3259 // infinite loop with no cond branch.
3260 if (OtherDest == BB) {
3261 // Insert it at the end of the function, because it's either code,
3262 // or it won't matter if it's hot. :)
3263 BasicBlock *InfLoopBlock =
3264 BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3265 BranchInst::Create(InfLoopBlock, InfLoopBlock);
3266 OtherDest = InfLoopBlock;
3269 DEBUG(dbgs() << *PBI->getParent()->getParent());
3271 // BI may have other predecessors. Because of this, we leave
3272 // it alone, but modify PBI.
3274 // Make sure we get to CommonDest on True&True directions.
3275 Value *PBICond = PBI->getCondition();
3276 IRBuilder<NoFolder> Builder(PBI);
3278 PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3280 Value *BICond = BI->getCondition();
3282 BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3284 // Merge the conditions.
3285 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3287 // Modify PBI to branch on the new condition to the new dests.
3288 PBI->setCondition(Cond);
3289 PBI->setSuccessor(0, CommonDest);
3290 PBI->setSuccessor(1, OtherDest);
3292 // Update branch weight for PBI.
3293 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3294 uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3296 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3297 SuccTrueWeight, SuccFalseWeight);
3299 PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3300 PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3301 SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3302 SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3303 // The weight to CommonDest should be PredCommon * SuccTotal +
3304 // PredOther * SuccCommon.
3305 // The weight to OtherDest should be PredOther * SuccOther.
3306 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3307 PredOther * SuccCommon,
3308 PredOther * SuccOther};
3309 // Halve the weights if any of them cannot fit in an uint32_t
3310 FitWeights(NewWeights);
3312 PBI->setMetadata(LLVMContext::MD_prof,
3313 MDBuilder(BI->getContext())
3314 .createBranchWeights(NewWeights[0], NewWeights[1]));
3317 // OtherDest may have phi nodes. If so, add an entry from PBI's
3318 // block that are identical to the entries for BI's block.
3319 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3321 // We know that the CommonDest already had an edge from PBI to
3322 // it. If it has PHIs though, the PHIs may have different
3323 // entries for BB and PBI's BB. If so, insert a select to make
3326 for (BasicBlock::iterator II = CommonDest->begin();
3327 (PN = dyn_cast<PHINode>(II)); ++II) {
3328 Value *BIV = PN->getIncomingValueForBlock(BB);
3329 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3330 Value *PBIV = PN->getIncomingValue(PBBIdx);
3332 // Insert a select in PBI to pick the right value.
3333 SelectInst *NV = cast<SelectInst>(
3334 Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3335 PN->setIncomingValue(PBBIdx, NV);
3336 // Although the select has the same condition as PBI, the original branch
3337 // weights for PBI do not apply to the new select because the select's
3338 // 'logical' edges are incoming edges of the phi that is eliminated, not
3339 // the outgoing edges of PBI.
3341 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3342 uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3343 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3344 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3345 // The weight to PredCommonDest should be PredCommon * SuccTotal.
3346 // The weight to PredOtherDest should be PredOther * SuccCommon.
3347 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3348 PredOther * SuccCommon};
3350 FitWeights(NewWeights);
3352 NV->setMetadata(LLVMContext::MD_prof,
3353 MDBuilder(BI->getContext())
3354 .createBranchWeights(NewWeights[0], NewWeights[1]));
3359 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3360 DEBUG(dbgs() << *PBI->getParent()->getParent());
3362 // This basic block is probably dead. We know it has at least
3363 // one fewer predecessor.
3367 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3368 // true or to FalseBB if Cond is false.
3369 // Takes care of updating the successors and removing the old terminator.
3370 // Also makes sure not to introduce new successors by assuming that edges to
3371 // non-successor TrueBBs and FalseBBs aren't reachable.
3372 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
3373 BasicBlock *TrueBB, BasicBlock *FalseBB,
3374 uint32_t TrueWeight,
3375 uint32_t FalseWeight) {
3376 // Remove any superfluous successor edges from the CFG.
3377 // First, figure out which successors to preserve.
3378 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3380 BasicBlock *KeepEdge1 = TrueBB;
3381 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3383 // Then remove the rest.
3384 for (BasicBlock *Succ : OldTerm->successors()) {
3385 // Make sure only to keep exactly one copy of each edge.
3386 if (Succ == KeepEdge1)
3387 KeepEdge1 = nullptr;
3388 else if (Succ == KeepEdge2)
3389 KeepEdge2 = nullptr;
3391 Succ->removePredecessor(OldTerm->getParent(),
3392 /*DontDeleteUselessPHIs=*/true);
3395 IRBuilder<> Builder(OldTerm);
3396 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3398 // Insert an appropriate new terminator.
3399 if (!KeepEdge1 && !KeepEdge2) {
3400 if (TrueBB == FalseBB)
3401 // We were only looking for one successor, and it was present.
3402 // Create an unconditional branch to it.
3403 Builder.CreateBr(TrueBB);
3405 // We found both of the successors we were looking for.
3406 // Create a conditional branch sharing the condition of the select.
3407 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3408 if (TrueWeight != FalseWeight)
3409 NewBI->setMetadata(LLVMContext::MD_prof,
3410 MDBuilder(OldTerm->getContext())
3411 .createBranchWeights(TrueWeight, FalseWeight));
3413 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3414 // Neither of the selected blocks were successors, so this
3415 // terminator must be unreachable.
3416 new UnreachableInst(OldTerm->getContext(), OldTerm);
3418 // One of the selected values was a successor, but the other wasn't.
3419 // Insert an unconditional branch to the one that was found;
3420 // the edge to the one that wasn't must be unreachable.
3422 // Only TrueBB was found.
3423 Builder.CreateBr(TrueBB);
3425 // Only FalseBB was found.
3426 Builder.CreateBr(FalseBB);
3429 EraseTerminatorInstAndDCECond(OldTerm);
3434 // (switch (select cond, X, Y)) on constant X, Y
3435 // with a branch - conditional if X and Y lead to distinct BBs,
3436 // unconditional otherwise.
3437 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
3438 // Check for constant integer values in the select.
3439 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3440 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3441 if (!TrueVal || !FalseVal)
3444 // Find the relevant condition and destinations.
3445 Value *Condition = Select->getCondition();
3446 BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3447 BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3449 // Get weight for TrueBB and FalseBB.
3450 uint32_t TrueWeight = 0, FalseWeight = 0;
3451 SmallVector<uint64_t, 8> Weights;
3452 bool HasWeights = HasBranchWeights(SI);
3454 GetBranchWeights(SI, Weights);
3455 if (Weights.size() == 1 + SI->getNumCases()) {
3457 (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3459 (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3463 // Perform the actual simplification.
3464 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3469 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3470 // blockaddress(@fn, BlockB)))
3472 // (br cond, BlockA, BlockB).
3473 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
3474 // Check that both operands of the select are block addresses.
3475 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
3476 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
3480 // Extract the actual blocks.
3481 BasicBlock *TrueBB = TBA->getBasicBlock();
3482 BasicBlock *FalseBB = FBA->getBasicBlock();
3484 // Perform the actual simplification.
3485 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3489 /// This is called when we find an icmp instruction
3490 /// (a seteq/setne with a constant) as the only instruction in a
3491 /// block that ends with an uncond branch. We are looking for a very specific
3492 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3493 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3494 /// default value goes to an uncond block with a seteq in it, we get something
3497 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3499 /// %tmp = icmp eq i8 %A, 92
3502 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3504 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3505 /// the PHI, merging the third icmp into the switch.
3506 static bool TryToSimplifyUncondBranchWithICmpInIt(
3507 ICmpInst *ICI, IRBuilder<> &Builder, const DataLayout &DL,
3508 const TargetTransformInfo &TTI, unsigned BonusInstThreshold,
3509 AssumptionCache *AC) {
3510 BasicBlock *BB = ICI->getParent();
3512 // If the block has any PHIs in it or the icmp has multiple uses, it is too
3514 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3517 Value *V = ICI->getOperand(0);
3518 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3520 // The pattern we're looking for is where our only predecessor is a switch on
3521 // 'V' and this block is the default case for the switch. In this case we can
3522 // fold the compared value into the switch to simplify things.
3523 BasicBlock *Pred = BB->getSinglePredecessor();
3524 if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3527 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3528 if (SI->getCondition() != V)
3531 // If BB is reachable on a non-default case, then we simply know the value of
3532 // V in this block. Substitute it and constant fold the icmp instruction
3534 if (SI->getDefaultDest() != BB) {
3535 ConstantInt *VVal = SI->findCaseDest(BB);
3536 assert(VVal && "Should have a unique destination value");
3537 ICI->setOperand(0, VVal);
3539 if (Value *V = SimplifyInstruction(ICI, DL)) {
3540 ICI->replaceAllUsesWith(V);
3541 ICI->eraseFromParent();
3543 // BB is now empty, so it is likely to simplify away.
3544 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3547 // Ok, the block is reachable from the default dest. If the constant we're
3548 // comparing exists in one of the other edges, then we can constant fold ICI
3550 if (SI->findCaseValue(Cst) != SI->case_default()) {
3552 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3553 V = ConstantInt::getFalse(BB->getContext());
3555 V = ConstantInt::getTrue(BB->getContext());
3557 ICI->replaceAllUsesWith(V);
3558 ICI->eraseFromParent();
3559 // BB is now empty, so it is likely to simplify away.
3560 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3563 // The use of the icmp has to be in the 'end' block, by the only PHI node in
3565 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3566 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3567 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3568 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3571 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3573 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3574 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3576 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3577 std::swap(DefaultCst, NewCst);
3579 // Replace ICI (which is used by the PHI for the default value) with true or
3580 // false depending on if it is EQ or NE.
3581 ICI->replaceAllUsesWith(DefaultCst);
3582 ICI->eraseFromParent();
3584 // Okay, the switch goes to this block on a default value. Add an edge from
3585 // the switch to the merge point on the compared value.
3587 BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3588 SmallVector<uint64_t, 8> Weights;
3589 bool HasWeights = HasBranchWeights(SI);
3591 GetBranchWeights(SI, Weights);
3592 if (Weights.size() == 1 + SI->getNumCases()) {
3593 // Split weight for default case to case for "Cst".
3594 Weights[0] = (Weights[0] + 1) >> 1;
3595 Weights.push_back(Weights[0]);
3597 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3599 LLVMContext::MD_prof,
3600 MDBuilder(SI->getContext()).createBranchWeights(MDWeights));
3603 SI->addCase(Cst, NewBB);
3605 // NewBB branches to the phi block, add the uncond branch and the phi entry.
3606 Builder.SetInsertPoint(NewBB);
3607 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3608 Builder.CreateBr(SuccBlock);
3609 PHIUse->addIncoming(NewCst, NewBB);
3613 /// The specified branch is a conditional branch.
3614 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3615 /// fold it into a switch instruction if so.
3616 static bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
3617 const DataLayout &DL) {
3618 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3622 // Change br (X == 0 | X == 1), T, F into a switch instruction.
3623 // If this is a bunch of seteq's or'd together, or if it's a bunch of
3624 // 'setne's and'ed together, collect them.
3626 // Try to gather values from a chain of and/or to be turned into a switch
3627 ConstantComparesGatherer ConstantCompare(Cond, DL);
3628 // Unpack the result
3629 SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3630 Value *CompVal = ConstantCompare.CompValue;
3631 unsigned UsedICmps = ConstantCompare.UsedICmps;
3632 Value *ExtraCase = ConstantCompare.Extra;
3634 // If we didn't have a multiply compared value, fail.
3638 // Avoid turning single icmps into a switch.
3642 bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3644 // There might be duplicate constants in the list, which the switch
3645 // instruction can't handle, remove them now.
3646 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3647 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3649 // If Extra was used, we require at least two switch values to do the
3650 // transformation. A switch with one value is just a conditional branch.
3651 if (ExtraCase && Values.size() < 2)
3654 // TODO: Preserve branch weight metadata, similarly to how
3655 // FoldValueComparisonIntoPredecessors preserves it.
3657 // Figure out which block is which destination.
3658 BasicBlock *DefaultBB = BI->getSuccessor(1);
3659 BasicBlock *EdgeBB = BI->getSuccessor(0);
3661 std::swap(DefaultBB, EdgeBB);
3663 BasicBlock *BB = BI->getParent();
3665 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3666 << " cases into SWITCH. BB is:\n"
3669 // If there are any extra values that couldn't be folded into the switch
3670 // then we evaluate them with an explicit branch first. Split the block
3671 // right before the condbr to handle it.
3674 BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3675 // Remove the uncond branch added to the old block.
3676 TerminatorInst *OldTI = BB->getTerminator();
3677 Builder.SetInsertPoint(OldTI);
3680 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3682 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3684 OldTI->eraseFromParent();
3686 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3687 // for the edge we just added.
3688 AddPredecessorToBlock(EdgeBB, BB, NewBB);
3690 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3691 << "\nEXTRABB = " << *BB);
3695 Builder.SetInsertPoint(BI);
3696 // Convert pointer to int before we switch.
3697 if (CompVal->getType()->isPointerTy()) {
3698 CompVal = Builder.CreatePtrToInt(
3699 CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3702 // Create the new switch instruction now.
3703 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3705 // Add all of the 'cases' to the switch instruction.
3706 for (unsigned i = 0, e = Values.size(); i != e; ++i)
3707 New->addCase(Values[i], EdgeBB);
3709 // We added edges from PI to the EdgeBB. As such, if there were any
3710 // PHI nodes in EdgeBB, they need entries to be added corresponding to
3711 // the number of edges added.
3712 for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3713 PHINode *PN = cast<PHINode>(BBI);
3714 Value *InVal = PN->getIncomingValueForBlock(BB);
3715 for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3716 PN->addIncoming(InVal, BB);
3719 // Erase the old branch instruction.
3720 EraseTerminatorInstAndDCECond(BI);
3722 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3726 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3727 if (isa<PHINode>(RI->getValue()))
3728 return SimplifyCommonResume(RI);
3729 else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3730 RI->getValue() == RI->getParent()->getFirstNonPHI())
3731 // The resume must unwind the exception that caused control to branch here.
3732 return SimplifySingleResume(RI);
3737 // Simplify resume that is shared by several landing pads (phi of landing pad).
3738 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3739 BasicBlock *BB = RI->getParent();
3741 // Check that there are no other instructions except for debug intrinsics
3742 // between the phi of landing pads (RI->getValue()) and resume instruction.
3743 BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3744 E = RI->getIterator();
3746 if (!isa<DbgInfoIntrinsic>(I))
3749 SmallSet<BasicBlock *, 4> TrivialUnwindBlocks;
3750 auto *PhiLPInst = cast<PHINode>(RI->getValue());
3752 // Check incoming blocks to see if any of them are trivial.
3753 for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3755 auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3756 auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3758 // If the block has other successors, we can not delete it because
3759 // it has other dependents.
3760 if (IncomingBB->getUniqueSuccessor() != BB)
3763 auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3764 // Not the landing pad that caused the control to branch here.
3765 if (IncomingValue != LandingPad)
3768 bool isTrivial = true;
3770 I = IncomingBB->getFirstNonPHI()->getIterator();
3771 E = IncomingBB->getTerminator()->getIterator();
3773 if (!isa<DbgInfoIntrinsic>(I)) {
3779 TrivialUnwindBlocks.insert(IncomingBB);
3782 // If no trivial unwind blocks, don't do any simplifications.
3783 if (TrivialUnwindBlocks.empty())
3786 // Turn all invokes that unwind here into calls.
3787 for (auto *TrivialBB : TrivialUnwindBlocks) {
3788 // Blocks that will be simplified should be removed from the phi node.
3789 // Note there could be multiple edges to the resume block, and we need
3790 // to remove them all.
3791 while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3792 BB->removePredecessor(TrivialBB, true);
3794 for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3796 BasicBlock *Pred = *PI++;
3797 removeUnwindEdge(Pred);
3800 // In each SimplifyCFG run, only the current processed block can be erased.
3801 // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3802 // of erasing TrivialBB, we only remove the branch to the common resume
3803 // block so that we can later erase the resume block since it has no
3805 TrivialBB->getTerminator()->eraseFromParent();
3806 new UnreachableInst(RI->getContext(), TrivialBB);
3809 // Delete the resume block if all its predecessors have been removed.
3811 BB->eraseFromParent();
3813 return !TrivialUnwindBlocks.empty();
3816 // Simplify resume that is only used by a single (non-phi) landing pad.
3817 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3818 BasicBlock *BB = RI->getParent();
3819 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
3820 assert(RI->getValue() == LPInst &&
3821 "Resume must unwind the exception that caused control to here");
3823 // Check that there are no other instructions except for debug intrinsics.
3824 BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3826 if (!isa<DbgInfoIntrinsic>(I))
3829 // Turn all invokes that unwind here into calls and delete the basic block.
3830 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3831 BasicBlock *Pred = *PI++;
3832 removeUnwindEdge(Pred);
3835 // The landingpad is now unreachable. Zap it.
3836 BB->eraseFromParent();
3838 LoopHeaders->erase(BB);
3842 static bool removeEmptyCleanup(CleanupReturnInst *RI) {
3843 // If this is a trivial cleanup pad that executes no instructions, it can be
3844 // eliminated. If the cleanup pad continues to the caller, any predecessor
3845 // that is an EH pad will be updated to continue to the caller and any
3846 // predecessor that terminates with an invoke instruction will have its invoke
3847 // instruction converted to a call instruction. If the cleanup pad being
3848 // simplified does not continue to the caller, each predecessor will be
3849 // updated to continue to the unwind destination of the cleanup pad being
3851 BasicBlock *BB = RI->getParent();
3852 CleanupPadInst *CPInst = RI->getCleanupPad();
3853 if (CPInst->getParent() != BB)
3854 // This isn't an empty cleanup.
3857 // We cannot kill the pad if it has multiple uses. This typically arises
3858 // from unreachable basic blocks.
3859 if (!CPInst->hasOneUse())
3862 // Check that there are no other instructions except for benign intrinsics.
3863 BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3865 auto *II = dyn_cast<IntrinsicInst>(I);
3869 Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3870 switch (IntrinsicID) {
3871 case Intrinsic::dbg_declare:
3872 case Intrinsic::dbg_value:
3873 case Intrinsic::lifetime_end:
3880 // If the cleanup return we are simplifying unwinds to the caller, this will
3881 // set UnwindDest to nullptr.
3882 BasicBlock *UnwindDest = RI->getUnwindDest();
3883 Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3885 // We're about to remove BB from the control flow. Before we do, sink any
3886 // PHINodes into the unwind destination. Doing this before changing the
3887 // control flow avoids some potentially slow checks, since we can currently
3888 // be certain that UnwindDest and BB have no common predecessors (since they
3889 // are both EH pads).
3891 // First, go through the PHI nodes in UnwindDest and update any nodes that
3892 // reference the block we are removing
3893 for (BasicBlock::iterator I = UnwindDest->begin(),
3894 IE = DestEHPad->getIterator();
3896 PHINode *DestPN = cast<PHINode>(I);
3898 int Idx = DestPN->getBasicBlockIndex(BB);
3899 // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3901 // This PHI node has an incoming value that corresponds to a control
3902 // path through the cleanup pad we are removing. If the incoming
3903 // value is in the cleanup pad, it must be a PHINode (because we
3904 // verified above that the block is otherwise empty). Otherwise, the
3905 // value is either a constant or a value that dominates the cleanup
3906 // pad being removed.
3908 // Because BB and UnwindDest are both EH pads, all of their
3909 // predecessors must unwind to these blocks, and since no instruction
3910 // can have multiple unwind destinations, there will be no overlap in
3911 // incoming blocks between SrcPN and DestPN.
3912 Value *SrcVal = DestPN->getIncomingValue(Idx);
3913 PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3915 // Remove the entry for the block we are deleting.
3916 DestPN->removeIncomingValue(Idx, false);
3918 if (SrcPN && SrcPN->getParent() == BB) {
3919 // If the incoming value was a PHI node in the cleanup pad we are
3920 // removing, we need to merge that PHI node's incoming values into
3922 for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3923 SrcIdx != SrcE; ++SrcIdx) {
3924 DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3925 SrcPN->getIncomingBlock(SrcIdx));
3928 // Otherwise, the incoming value came from above BB and
3929 // so we can just reuse it. We must associate all of BB's
3930 // predecessors with this value.
3931 for (auto *pred : predecessors(BB)) {
3932 DestPN->addIncoming(SrcVal, pred);
3937 // Sink any remaining PHI nodes directly into UnwindDest.
3938 Instruction *InsertPt = DestEHPad;
3939 for (BasicBlock::iterator I = BB->begin(),
3940 IE = BB->getFirstNonPHI()->getIterator();
3942 // The iterator must be incremented here because the instructions are
3943 // being moved to another block.
3944 PHINode *PN = cast<PHINode>(I++);
3945 if (PN->use_empty())
3946 // If the PHI node has no uses, just leave it. It will be erased
3947 // when we erase BB below.
3950 // Otherwise, sink this PHI node into UnwindDest.
3951 // Any predecessors to UnwindDest which are not already represented
3952 // must be back edges which inherit the value from the path through
3953 // BB. In this case, the PHI value must reference itself.
3954 for (auto *pred : predecessors(UnwindDest))
3956 PN->addIncoming(PN, pred);
3957 PN->moveBefore(InsertPt);
3961 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3962 // The iterator must be updated here because we are removing this pred.
3963 BasicBlock *PredBB = *PI++;
3964 if (UnwindDest == nullptr) {
3965 removeUnwindEdge(PredBB);
3967 TerminatorInst *TI = PredBB->getTerminator();
3968 TI->replaceUsesOfWith(BB, UnwindDest);
3972 // The cleanup pad is now unreachable. Zap it.
3973 BB->eraseFromParent();
3977 // Try to merge two cleanuppads together.
3978 static bool mergeCleanupPad(CleanupReturnInst *RI) {
3979 // Skip any cleanuprets which unwind to caller, there is nothing to merge
3981 BasicBlock *UnwindDest = RI->getUnwindDest();
3985 // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
3986 // be safe to merge without code duplication.
3987 if (UnwindDest->getSinglePredecessor() != RI->getParent())
3990 // Verify that our cleanuppad's unwind destination is another cleanuppad.
3991 auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
3992 if (!SuccessorCleanupPad)
3995 CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
3996 // Replace any uses of the successor cleanupad with the predecessor pad
3997 // The only cleanuppad uses should be this cleanupret, it's cleanupret and
3998 // funclet bundle operands.
3999 SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4000 // Remove the old cleanuppad.
4001 SuccessorCleanupPad->eraseFromParent();
4002 // Now, we simply replace the cleanupret with a branch to the unwind
4004 BranchInst::Create(UnwindDest, RI->getParent());
4005 RI->eraseFromParent();
4010 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
4011 // It is possible to transiantly have an undef cleanuppad operand because we
4012 // have deleted some, but not all, dead blocks.
4013 // Eventually, this block will be deleted.
4014 if (isa<UndefValue>(RI->getOperand(0)))
4017 if (mergeCleanupPad(RI))
4020 if (removeEmptyCleanup(RI))
4026 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4027 BasicBlock *BB = RI->getParent();
4028 if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4031 // Find predecessors that end with branches.
4032 SmallVector<BasicBlock *, 8> UncondBranchPreds;
4033 SmallVector<BranchInst *, 8> CondBranchPreds;
4034 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4035 BasicBlock *P = *PI;
4036 TerminatorInst *PTI = P->getTerminator();
4037 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4038 if (BI->isUnconditional())
4039 UncondBranchPreds.push_back(P);
4041 CondBranchPreds.push_back(BI);
4045 // If we found some, do the transformation!
4046 if (!UncondBranchPreds.empty() && DupRet) {
4047 while (!UncondBranchPreds.empty()) {
4048 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4049 DEBUG(dbgs() << "FOLDING: " << *BB
4050 << "INTO UNCOND BRANCH PRED: " << *Pred);
4051 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4054 // If we eliminated all predecessors of the block, delete the block now.
4055 if (pred_empty(BB)) {
4056 // We know there are no successors, so just nuke the block.
4057 BB->eraseFromParent();
4059 LoopHeaders->erase(BB);
4065 // Check out all of the conditional branches going to this return
4066 // instruction. If any of them just select between returns, change the
4067 // branch itself into a select/return pair.
4068 while (!CondBranchPreds.empty()) {
4069 BranchInst *BI = CondBranchPreds.pop_back_val();
4071 // Check to see if the non-BB successor is also a return block.
4072 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4073 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4074 SimplifyCondBranchToTwoReturns(BI, Builder))
4080 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4081 BasicBlock *BB = UI->getParent();
4083 bool Changed = false;
4085 // If there are any instructions immediately before the unreachable that can
4086 // be removed, do so.
4087 while (UI->getIterator() != BB->begin()) {
4088 BasicBlock::iterator BBI = UI->getIterator();
4090 // Do not delete instructions that can have side effects which might cause
4091 // the unreachable to not be reachable; specifically, calls and volatile
4092 // operations may have this effect.
4093 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4096 if (BBI->mayHaveSideEffects()) {
4097 if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4098 if (SI->isVolatile())
4100 } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4101 if (LI->isVolatile())
4103 } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4104 if (RMWI->isVolatile())
4106 } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4107 if (CXI->isVolatile())
4109 } else if (isa<CatchPadInst>(BBI)) {
4110 // A catchpad may invoke exception object constructors and such, which
4111 // in some languages can be arbitrary code, so be conservative by
4113 // For CoreCLR, it just involves a type test, so can be removed.
4114 if (classifyEHPersonality(BB->getParent()->getPersonalityFn()) !=
4115 EHPersonality::CoreCLR)
4117 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4118 !isa<LandingPadInst>(BBI)) {
4121 // Note that deleting LandingPad's here is in fact okay, although it
4122 // involves a bit of subtle reasoning. If this inst is a LandingPad,
4123 // all the predecessors of this block will be the unwind edges of Invokes,
4124 // and we can therefore guarantee this block will be erased.
4127 // Delete this instruction (any uses are guaranteed to be dead)
4128 if (!BBI->use_empty())
4129 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4130 BBI->eraseFromParent();
4134 // If the unreachable instruction is the first in the block, take a gander
4135 // at all of the predecessors of this instruction, and simplify them.
4136 if (&BB->front() != UI)
4139 SmallVector<BasicBlock *, 8> Preds(pred_begin(BB), pred_end(BB));
4140 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4141 TerminatorInst *TI = Preds[i]->getTerminator();
4142 IRBuilder<> Builder(TI);
4143 if (auto *BI = dyn_cast<BranchInst>(TI)) {
4144 if (BI->isUnconditional()) {
4145 if (BI->getSuccessor(0) == BB) {
4146 new UnreachableInst(TI->getContext(), TI);
4147 TI->eraseFromParent();
4151 if (BI->getSuccessor(0) == BB) {
4152 Builder.CreateBr(BI->getSuccessor(1));
4153 EraseTerminatorInstAndDCECond(BI);
4154 } else if (BI->getSuccessor(1) == BB) {
4155 Builder.CreateBr(BI->getSuccessor(0));
4156 EraseTerminatorInstAndDCECond(BI);
4160 } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4161 for (auto i = SI->case_begin(), e = SI->case_end(); i != e;) {
4162 if (i->getCaseSuccessor() != BB) {
4166 BB->removePredecessor(SI->getParent());
4167 i = SI->removeCase(i);
4171 } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4172 if (II->getUnwindDest() == BB) {
4173 removeUnwindEdge(TI->getParent());
4176 } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4177 if (CSI->getUnwindDest() == BB) {
4178 removeUnwindEdge(TI->getParent());
4183 for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4184 E = CSI->handler_end();
4187 CSI->removeHandler(I);
4193 if (CSI->getNumHandlers() == 0) {
4194 BasicBlock *CatchSwitchBB = CSI->getParent();
4195 if (CSI->hasUnwindDest()) {
4196 // Redirect preds to the unwind dest
4197 CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4199 // Rewrite all preds to unwind to caller (or from invoke to call).
4200 SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4201 for (BasicBlock *EHPred : EHPreds)
4202 removeUnwindEdge(EHPred);
4204 // The catchswitch is no longer reachable.
4205 new UnreachableInst(CSI->getContext(), CSI);
4206 CSI->eraseFromParent();
4209 } else if (isa<CleanupReturnInst>(TI)) {
4210 new UnreachableInst(TI->getContext(), TI);
4211 TI->eraseFromParent();
4216 // If this block is now dead, remove it.
4217 if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4218 // We know there are no successors, so just nuke the block.
4219 BB->eraseFromParent();
4221 LoopHeaders->erase(BB);
4228 static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) {
4229 assert(Cases.size() >= 1);
4231 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4232 for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4233 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4239 /// Turn a switch with two reachable destinations into an integer range
4240 /// comparison and branch.
4241 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
4242 assert(SI->getNumCases() > 1 && "Degenerate switch?");
4245 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4247 // Partition the cases into two sets with different destinations.
4248 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4249 BasicBlock *DestB = nullptr;
4250 SmallVector<ConstantInt *, 16> CasesA;
4251 SmallVector<ConstantInt *, 16> CasesB;
4253 for (auto Case : SI->cases()) {
4254 BasicBlock *Dest = Case.getCaseSuccessor();
4257 if (Dest == DestA) {
4258 CasesA.push_back(Case.getCaseValue());
4263 if (Dest == DestB) {
4264 CasesB.push_back(Case.getCaseValue());
4267 return false; // More than two destinations.
4270 assert(DestA && DestB &&
4271 "Single-destination switch should have been folded.");
4272 assert(DestA != DestB);
4273 assert(DestB != SI->getDefaultDest());
4274 assert(!CasesB.empty() && "There must be non-default cases.");
4275 assert(!CasesA.empty() || HasDefault);
4277 // Figure out if one of the sets of cases form a contiguous range.
4278 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4279 BasicBlock *ContiguousDest = nullptr;
4280 BasicBlock *OtherDest = nullptr;
4281 if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4282 ContiguousCases = &CasesA;
4283 ContiguousDest = DestA;
4285 } else if (CasesAreContiguous(CasesB)) {
4286 ContiguousCases = &CasesB;
4287 ContiguousDest = DestB;
4292 // Start building the compare and branch.
4294 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4295 Constant *NumCases =
4296 ConstantInt::get(Offset->getType(), ContiguousCases->size());
4298 Value *Sub = SI->getCondition();
4299 if (!Offset->isNullValue())
4300 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4303 // If NumCases overflowed, then all possible values jump to the successor.
4304 if (NumCases->isNullValue() && !ContiguousCases->empty())
4305 Cmp = ConstantInt::getTrue(SI->getContext());
4307 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4308 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4310 // Update weight for the newly-created conditional branch.
4311 if (HasBranchWeights(SI)) {
4312 SmallVector<uint64_t, 8> Weights;
4313 GetBranchWeights(SI, Weights);
4314 if (Weights.size() == 1 + SI->getNumCases()) {
4315 uint64_t TrueWeight = 0;
4316 uint64_t FalseWeight = 0;
4317 for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4318 if (SI->getSuccessor(I) == ContiguousDest)
4319 TrueWeight += Weights[I];
4321 FalseWeight += Weights[I];
4323 while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4327 NewBI->setMetadata(LLVMContext::MD_prof,
4328 MDBuilder(SI->getContext())
4329 .createBranchWeights((uint32_t)TrueWeight,
4330 (uint32_t)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 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
4363 computeKnownBits(Cond, KnownZero, KnownOne, DL, 0, AC, SI);
4365 // We can also eliminate cases by determining that their values are outside of
4366 // the limited range of the condition based on how many significant (non-sign)
4367 // bits are in the condition value.
4368 unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4369 unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4371 // Gather dead cases.
4372 SmallVector<ConstantInt *, 8> DeadCases;
4373 for (auto &Case : SI->cases()) {
4374 APInt CaseVal = Case.getCaseValue()->getValue();
4375 if ((CaseVal & KnownZero) != 0 || (CaseVal & KnownOne) != KnownOne ||
4376 (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4377 DeadCases.push_back(Case.getCaseValue());
4378 DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal << " is dead.\n");
4382 // If we can prove that the cases must cover all possible values, the
4383 // default destination becomes dead and we can remove it. If we know some
4384 // of the bits in the value, we can use that to more precisely compute the
4385 // number of possible unique case values.
4387 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4388 const unsigned NumUnknownBits =
4389 Bits - (KnownZero | KnownOne).countPopulation();
4390 assert(NumUnknownBits <= Bits);
4391 if (HasDefault && DeadCases.empty() &&
4392 NumUnknownBits < 64 /* avoid overflow */ &&
4393 SI->getNumCases() == (1ULL << NumUnknownBits)) {
4394 DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4395 BasicBlock *NewDefault =
4396 SplitBlockPredecessors(SI->getDefaultDest(), SI->getParent(), "");
4397 SI->setDefaultDest(&*NewDefault);
4398 SplitBlock(&*NewDefault, &NewDefault->front());
4399 auto *OldTI = NewDefault->getTerminator();
4400 new UnreachableInst(SI->getContext(), OldTI);
4401 EraseTerminatorInstAndDCECond(OldTI);
4405 SmallVector<uint64_t, 8> Weights;
4406 bool HasWeight = HasBranchWeights(SI);
4408 GetBranchWeights(SI, Weights);
4409 HasWeight = (Weights.size() == 1 + SI->getNumCases());
4412 // Remove dead cases from the switch.
4413 for (ConstantInt *DeadCase : DeadCases) {
4414 SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4415 assert(CaseI != SI->case_default() &&
4416 "Case was not found. Probably mistake in DeadCases forming.");
4418 std::swap(Weights[CaseI->getCaseIndex() + 1], Weights.back());
4422 // Prune unused values from PHI nodes.
4423 CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4424 SI->removeCase(CaseI);
4426 if (HasWeight && Weights.size() >= 2) {
4427 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
4428 SI->setMetadata(LLVMContext::MD_prof,
4429 MDBuilder(SI->getParent()->getContext())
4430 .createBranchWeights(MDWeights));
4433 return !DeadCases.empty();
4436 /// If BB would be eligible for simplification by
4437 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4438 /// by an unconditional branch), look at the phi node for BB in the successor
4439 /// block and see if the incoming value is equal to CaseValue. If so, return
4440 /// the phi node, and set PhiIndex to BB's index in the phi node.
4441 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
4442 BasicBlock *BB, int *PhiIndex) {
4443 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4444 return nullptr; // BB must be empty to be a candidate for simplification.
4445 if (!BB->getSinglePredecessor())
4446 return nullptr; // BB must be dominated by the switch.
4448 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
4449 if (!Branch || !Branch->isUnconditional())
4450 return nullptr; // Terminator must be unconditional branch.
4452 BasicBlock *Succ = Branch->getSuccessor(0);
4454 BasicBlock::iterator I = Succ->begin();
4455 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
4456 int Idx = PHI->getBasicBlockIndex(BB);
4457 assert(Idx >= 0 && "PHI has no entry for predecessor?");
4459 Value *InValue = PHI->getIncomingValue(Idx);
4460 if (InValue != CaseValue)
4470 /// Try to forward the condition of a switch instruction to a phi node
4471 /// dominated by the switch, if that would mean that some of the destination
4472 /// blocks of the switch can be folded away.
4473 /// Returns true if a change is made.
4474 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
4475 typedef DenseMap<PHINode *, SmallVector<int, 4>> ForwardingNodesMap;
4476 ForwardingNodesMap ForwardingNodes;
4478 for (auto Case : SI->cases()) {
4479 ConstantInt *CaseValue = Case.getCaseValue();
4480 BasicBlock *CaseDest = Case.getCaseSuccessor();
4484 FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIndex);
4488 ForwardingNodes[PHI].push_back(PhiIndex);
4491 bool Changed = false;
4493 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
4494 E = ForwardingNodes.end();
4496 PHINode *Phi = I->first;
4497 SmallVectorImpl<int> &Indexes = I->second;
4499 if (Indexes.size() < 2)
4502 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
4503 Phi->setIncomingValue(Indexes[I], SI->getCondition());
4510 /// Return true if the backend will be able to handle
4511 /// initializing an array of constants like C.
4512 static bool ValidLookupTableConstant(Constant *C, const TargetTransformInfo &TTI) {
4513 if (C->isThreadDependent())
4515 if (C->isDLLImportDependent())
4518 if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4519 !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4520 !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4523 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4524 if (!CE->isGEPWithNoNotionalOverIndexing())
4526 if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4530 if (!TTI.shouldBuildLookupTablesForConstant(C))
4536 /// If V is a Constant, return it. Otherwise, try to look up
4537 /// its constant value in ConstantPool, returning 0 if it's not there.
4539 LookupConstant(Value *V,
4540 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4541 if (Constant *C = dyn_cast<Constant>(V))
4543 return ConstantPool.lookup(V);
4546 /// Try to fold instruction I into a constant. This works for
4547 /// simple instructions such as binary operations where both operands are
4548 /// constant or can be replaced by constants from the ConstantPool. Returns the
4549 /// resulting constant on success, 0 otherwise.
4551 ConstantFold(Instruction *I, const DataLayout &DL,
4552 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4553 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4554 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4557 if (A->isAllOnesValue())
4558 return LookupConstant(Select->getTrueValue(), ConstantPool);
4559 if (A->isNullValue())
4560 return LookupConstant(Select->getFalseValue(), ConstantPool);
4564 SmallVector<Constant *, 4> COps;
4565 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4566 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
4572 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4573 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4577 return ConstantFoldInstOperands(I, COps, DL);
4580 /// Try to determine the resulting constant values in phi nodes
4581 /// at the common destination basic block, *CommonDest, for one of the case
4582 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4583 /// case), of a switch instruction SI.
4585 GetCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest,
4586 BasicBlock **CommonDest,
4587 SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4588 const DataLayout &DL, const TargetTransformInfo &TTI) {
4589 // The block from which we enter the common destination.
4590 BasicBlock *Pred = SI->getParent();
4592 // If CaseDest is empty except for some side-effect free instructions through
4593 // which we can constant-propagate the CaseVal, continue to its successor.
4594 SmallDenseMap<Value *, Constant *> ConstantPool;
4595 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4596 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
4598 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
4599 // If the terminator is a simple branch, continue to the next block.
4600 if (T->getNumSuccessors() != 1 || T->isExceptional())
4603 CaseDest = T->getSuccessor(0);
4604 } else if (isa<DbgInfoIntrinsic>(I)) {
4605 // Skip debug intrinsic.
4607 } else if (Constant *C = ConstantFold(&*I, DL, ConstantPool)) {
4608 // Instruction is side-effect free and constant.
4610 // If the instruction has uses outside this block or a phi node slot for
4611 // the block, it is not safe to bypass the instruction since it would then
4612 // no longer dominate all its uses.
4613 for (auto &Use : I->uses()) {
4614 User *User = Use.getUser();
4615 if (Instruction *I = dyn_cast<Instruction>(User))
4616 if (I->getParent() == CaseDest)
4618 if (PHINode *Phi = dyn_cast<PHINode>(User))
4619 if (Phi->getIncomingBlock(Use) == CaseDest)
4624 ConstantPool.insert(std::make_pair(&*I, C));
4630 // If we did not have a CommonDest before, use the current one.
4632 *CommonDest = CaseDest;
4633 // If the destination isn't the common one, abort.
4634 if (CaseDest != *CommonDest)
4637 // Get the values for this case from phi nodes in the destination block.
4638 BasicBlock::iterator I = (*CommonDest)->begin();
4639 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
4640 int Idx = PHI->getBasicBlockIndex(Pred);
4644 Constant *ConstVal =
4645 LookupConstant(PHI->getIncomingValue(Idx), ConstantPool);
4649 // Be conservative about which kinds of constants we support.
4650 if (!ValidLookupTableConstant(ConstVal, TTI))
4653 Res.push_back(std::make_pair(PHI, ConstVal));
4656 return Res.size() > 0;
4659 // Helper function used to add CaseVal to the list of cases that generate
4661 static void MapCaseToResult(ConstantInt *CaseVal,
4662 SwitchCaseResultVectorTy &UniqueResults,
4664 for (auto &I : UniqueResults) {
4665 if (I.first == Result) {
4666 I.second.push_back(CaseVal);
4670 UniqueResults.push_back(
4671 std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4674 // Helper function that initializes a map containing
4675 // results for the PHI node of the common destination block for a switch
4676 // instruction. Returns false if multiple PHI nodes have been found or if
4677 // there is not a common destination block for the switch.
4678 static bool InitializeUniqueCases(SwitchInst *SI, PHINode *&PHI,
4679 BasicBlock *&CommonDest,
4680 SwitchCaseResultVectorTy &UniqueResults,
4681 Constant *&DefaultResult,
4682 const DataLayout &DL,
4683 const TargetTransformInfo &TTI) {
4684 for (auto &I : SI->cases()) {
4685 ConstantInt *CaseVal = I.getCaseValue();
4687 // Resulting value at phi nodes for this case value.
4688 SwitchCaseResultsTy Results;
4689 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4693 // Only one value per case is permitted
4694 if (Results.size() > 1)
4696 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4698 // Check the PHI consistency.
4700 PHI = Results[0].first;
4701 else if (PHI != Results[0].first)
4704 // Find the default result value.
4705 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
4706 BasicBlock *DefaultDest = SI->getDefaultDest();
4707 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4709 // If the default value is not found abort unless the default destination
4712 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4713 if ((!DefaultResult &&
4714 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4720 // Helper function that checks if it is possible to transform a switch with only
4721 // two cases (or two cases + default) that produces a result into a select.
4724 // case 10: %0 = icmp eq i32 %a, 10
4725 // return 10; %1 = select i1 %0, i32 10, i32 4
4726 // case 20: ----> %2 = icmp eq i32 %a, 20
4727 // return 2; %3 = select i1 %2, i32 2, i32 %1
4731 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4732 Constant *DefaultResult, Value *Condition,
4733 IRBuilder<> &Builder) {
4734 assert(ResultVector.size() == 2 &&
4735 "We should have exactly two unique results at this point");
4736 // If we are selecting between only two cases transform into a simple
4737 // select or a two-way select if default is possible.
4738 if (ResultVector[0].second.size() == 1 &&
4739 ResultVector[1].second.size() == 1) {
4740 ConstantInt *const FirstCase = ResultVector[0].second[0];
4741 ConstantInt *const SecondCase = ResultVector[1].second[0];
4743 bool DefaultCanTrigger = DefaultResult;
4744 Value *SelectValue = ResultVector[1].first;
4745 if (DefaultCanTrigger) {
4746 Value *const ValueCompare =
4747 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4748 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4749 DefaultResult, "switch.select");
4751 Value *const ValueCompare =
4752 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4753 return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4754 SelectValue, "switch.select");
4760 // Helper function to cleanup a switch instruction that has been converted into
4761 // a select, fixing up PHI nodes and basic blocks.
4762 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
4764 IRBuilder<> &Builder) {
4765 BasicBlock *SelectBB = SI->getParent();
4766 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4767 PHI->removeIncomingValue(SelectBB);
4768 PHI->addIncoming(SelectValue, SelectBB);
4770 Builder.CreateBr(PHI->getParent());
4772 // Remove the switch.
4773 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4774 BasicBlock *Succ = SI->getSuccessor(i);
4776 if (Succ == PHI->getParent())
4778 Succ->removePredecessor(SelectBB);
4780 SI->eraseFromParent();
4783 /// If the switch is only used to initialize one or more
4784 /// phi nodes in a common successor block with only two different
4785 /// constant values, replace the switch with select.
4786 static bool SwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4787 AssumptionCache *AC, const DataLayout &DL,
4788 const TargetTransformInfo &TTI) {
4789 Value *const Cond = SI->getCondition();
4790 PHINode *PHI = nullptr;
4791 BasicBlock *CommonDest = nullptr;
4792 Constant *DefaultResult;
4793 SwitchCaseResultVectorTy UniqueResults;
4794 // Collect all the cases that will deliver the same value from the switch.
4795 if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4798 // Selects choose between maximum two values.
4799 if (UniqueResults.size() != 2)
4801 assert(PHI != nullptr && "PHI for value select not found");
4803 Builder.SetInsertPoint(SI);
4804 Value *SelectValue =
4805 ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4807 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4810 // The switch couldn't be converted into a select.
4816 /// This class represents a lookup table that can be used to replace a switch.
4817 class SwitchLookupTable {
4819 /// Create a lookup table to use as a switch replacement with the contents
4820 /// of Values, using DefaultValue to fill any holes in the table.
4822 Module &M, uint64_t TableSize, ConstantInt *Offset,
4823 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4824 Constant *DefaultValue, const DataLayout &DL);
4826 /// Build instructions with Builder to retrieve the value at
4827 /// the position given by Index in the lookup table.
4828 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4830 /// Return true if a table with TableSize elements of
4831 /// type ElementType would fit in a target-legal register.
4832 static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4836 // Depending on the contents of the table, it can be represented in
4839 // For tables where each element contains the same value, we just have to
4840 // store that single value and return it for each lookup.
4843 // For tables where there is a linear relationship between table index
4844 // and values. We calculate the result with a simple multiplication
4845 // and addition instead of a table lookup.
4848 // For small tables with integer elements, we can pack them into a bitmap
4849 // that fits into a target-legal register. Values are retrieved by
4850 // shift and mask operations.
4853 // The table is stored as an array of values. Values are retrieved by load
4854 // instructions from the table.
4858 // For SingleValueKind, this is the single value.
4859 Constant *SingleValue;
4861 // For BitMapKind, this is the bitmap.
4862 ConstantInt *BitMap;
4863 IntegerType *BitMapElementTy;
4865 // For LinearMapKind, these are the constants used to derive the value.
4866 ConstantInt *LinearOffset;
4867 ConstantInt *LinearMultiplier;
4869 // For ArrayKind, this is the array.
4870 GlobalVariable *Array;
4873 } // end anonymous namespace
4875 SwitchLookupTable::SwitchLookupTable(
4876 Module &M, uint64_t TableSize, ConstantInt *Offset,
4877 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4878 Constant *DefaultValue, const DataLayout &DL)
4879 : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
4880 LinearOffset(nullptr), LinearMultiplier(nullptr), Array(nullptr) {
4881 assert(Values.size() && "Can't build lookup table without values!");
4882 assert(TableSize >= Values.size() && "Can't fit values in table!");
4884 // If all values in the table are equal, this is that value.
4885 SingleValue = Values.begin()->second;
4887 Type *ValueType = Values.begin()->second->getType();
4889 // Build up the table contents.
4890 SmallVector<Constant *, 64> TableContents(TableSize);
4891 for (size_t I = 0, E = Values.size(); I != E; ++I) {
4892 ConstantInt *CaseVal = Values[I].first;
4893 Constant *CaseRes = Values[I].second;
4894 assert(CaseRes->getType() == ValueType);
4896 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4897 TableContents[Idx] = CaseRes;
4899 if (CaseRes != SingleValue)
4900 SingleValue = nullptr;
4903 // Fill in any holes in the table with the default result.
4904 if (Values.size() < TableSize) {
4905 assert(DefaultValue &&
4906 "Need a default value to fill the lookup table holes.");
4907 assert(DefaultValue->getType() == ValueType);
4908 for (uint64_t I = 0; I < TableSize; ++I) {
4909 if (!TableContents[I])
4910 TableContents[I] = DefaultValue;
4913 if (DefaultValue != SingleValue)
4914 SingleValue = nullptr;
4917 // If each element in the table contains the same value, we only need to store
4918 // that single value.
4920 Kind = SingleValueKind;
4924 // Check if we can derive the value with a linear transformation from the
4926 if (isa<IntegerType>(ValueType)) {
4927 bool LinearMappingPossible = true;
4930 assert(TableSize >= 2 && "Should be a SingleValue table.");
4931 // Check if there is the same distance between two consecutive values.
4932 for (uint64_t I = 0; I < TableSize; ++I) {
4933 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4935 // This is an undef. We could deal with it, but undefs in lookup tables
4936 // are very seldom. It's probably not worth the additional complexity.
4937 LinearMappingPossible = false;
4940 APInt Val = ConstVal->getValue();
4942 APInt Dist = Val - PrevVal;
4945 } else if (Dist != DistToPrev) {
4946 LinearMappingPossible = false;
4952 if (LinearMappingPossible) {
4953 LinearOffset = cast<ConstantInt>(TableContents[0]);
4954 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
4955 Kind = LinearMapKind;
4961 // If the type is integer and the table fits in a register, build a bitmap.
4962 if (WouldFitInRegister(DL, TableSize, ValueType)) {
4963 IntegerType *IT = cast<IntegerType>(ValueType);
4964 APInt TableInt(TableSize * IT->getBitWidth(), 0);
4965 for (uint64_t I = TableSize; I > 0; --I) {
4966 TableInt <<= IT->getBitWidth();
4967 // Insert values into the bitmap. Undef values are set to zero.
4968 if (!isa<UndefValue>(TableContents[I - 1])) {
4969 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
4970 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
4973 BitMap = ConstantInt::get(M.getContext(), TableInt);
4974 BitMapElementTy = IT;
4980 // Store the table in an array.
4981 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
4982 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
4984 Array = new GlobalVariable(M, ArrayTy, /*constant=*/true,
4985 GlobalVariable::PrivateLinkage, Initializer,
4987 Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
4991 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
4993 case SingleValueKind:
4995 case LinearMapKind: {
4996 // Derive the result value from the input value.
4997 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
4998 false, "switch.idx.cast");
4999 if (!LinearMultiplier->isOne())
5000 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
5001 if (!LinearOffset->isZero())
5002 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
5006 // Type of the bitmap (e.g. i59).
5007 IntegerType *MapTy = BitMap->getType();
5009 // Cast Index to the same type as the bitmap.
5010 // Note: The Index is <= the number of elements in the table, so
5011 // truncating it to the width of the bitmask is safe.
5012 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
5014 // Multiply the shift amount by the element width.
5015 ShiftAmt = Builder.CreateMul(
5016 ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5020 Value *DownShifted =
5021 Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5023 return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5026 // Make sure the table index will not overflow when treated as signed.
5027 IntegerType *IT = cast<IntegerType>(Index->getType());
5028 uint64_t TableSize =
5029 Array->getInitializer()->getType()->getArrayNumElements();
5030 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5031 Index = Builder.CreateZExt(
5032 Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5033 "switch.tableidx.zext");
5035 Value *GEPIndices[] = {Builder.getInt32(0), Index};
5036 Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5037 GEPIndices, "switch.gep");
5038 return Builder.CreateLoad(GEP, "switch.load");
5041 llvm_unreachable("Unknown lookup table kind!");
5044 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5046 Type *ElementType) {
5047 auto *IT = dyn_cast<IntegerType>(ElementType);
5050 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5051 // are <= 15, we could try to narrow the type.
5053 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5054 if (TableSize >= UINT_MAX / IT->getBitWidth())
5056 return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5059 /// Determine whether a lookup table should be built for this switch, based on
5060 /// the number of cases, size of the table, and the types of the results.
5062 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5063 const TargetTransformInfo &TTI, const DataLayout &DL,
5064 const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5065 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5066 return false; // TableSize overflowed, or mul below might overflow.
5068 bool AllTablesFitInRegister = true;
5069 bool HasIllegalType = false;
5070 for (const auto &I : ResultTypes) {
5071 Type *Ty = I.second;
5073 // Saturate this flag to true.
5074 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5076 // Saturate this flag to false.
5077 AllTablesFitInRegister =
5078 AllTablesFitInRegister &&
5079 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5081 // If both flags saturate, we're done. NOTE: This *only* works with
5082 // saturating flags, and all flags have to saturate first due to the
5083 // non-deterministic behavior of iterating over a dense map.
5084 if (HasIllegalType && !AllTablesFitInRegister)
5088 // If each table would fit in a register, we should build it anyway.
5089 if (AllTablesFitInRegister)
5092 // Don't build a table that doesn't fit in-register if it has illegal types.
5096 // The table density should be at least 40%. This is the same criterion as for
5097 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5098 // FIXME: Find the best cut-off.
5099 return SI->getNumCases() * 10 >= TableSize * 4;
5102 /// Try to reuse the switch table index compare. Following pattern:
5104 /// if (idx < tablesize)
5105 /// r = table[idx]; // table does not contain default_value
5107 /// r = default_value;
5108 /// if (r != default_value)
5111 /// Is optimized to:
5113 /// cond = idx < tablesize;
5117 /// r = default_value;
5121 /// Jump threading will then eliminate the second if(cond).
5122 static void reuseTableCompare(
5123 User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5124 Constant *DefaultValue,
5125 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5127 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5131 // We require that the compare is in the same block as the phi so that jump
5132 // threading can do its work afterwards.
5133 if (CmpInst->getParent() != PhiBlock)
5136 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5140 Value *RangeCmp = RangeCheckBranch->getCondition();
5141 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5142 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5144 // Check if the compare with the default value is constant true or false.
5145 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5146 DefaultValue, CmpOp1, true);
5147 if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5150 // Check if the compare with the case values is distinct from the default
5152 for (auto ValuePair : Values) {
5153 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5154 ValuePair.second, CmpOp1, true);
5155 if (!CaseConst || CaseConst == DefaultConst)
5157 assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5158 "Expect true or false as compare result.");
5161 // Check if the branch instruction dominates the phi node. It's a simple
5162 // dominance check, but sufficient for our needs.
5163 // Although this check is invariant in the calling loops, it's better to do it
5164 // at this late stage. Practically we do it at most once for a switch.
5165 BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5166 for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5167 BasicBlock *Pred = *PI;
5168 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5172 if (DefaultConst == FalseConst) {
5173 // The compare yields the same result. We can replace it.
5174 CmpInst->replaceAllUsesWith(RangeCmp);
5175 ++NumTableCmpReuses;
5177 // The compare yields the same result, just inverted. We can replace it.
5178 Value *InvertedTableCmp = BinaryOperator::CreateXor(
5179 RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5181 CmpInst->replaceAllUsesWith(InvertedTableCmp);
5182 ++NumTableCmpReuses;
5186 /// If the switch is only used to initialize one or more phi nodes in a common
5187 /// successor block with different constant values, replace the switch with
5189 static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
5190 const DataLayout &DL,
5191 const TargetTransformInfo &TTI) {
5192 assert(SI->getNumCases() > 1 && "Degenerate switch?");
5194 // Only build lookup table when we have a target that supports it.
5195 if (!TTI.shouldBuildLookupTables())
5198 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5199 // split off a dense part and build a lookup table for that.
5201 // FIXME: This creates arrays of GEPs to constant strings, which means each
5202 // GEP needs a runtime relocation in PIC code. We should just build one big
5203 // string and lookup indices into that.
5205 // Ignore switches with less than three cases. Lookup tables will not make
5207 // faster, so we don't analyze them.
5208 if (SI->getNumCases() < 3)
5211 // Figure out the corresponding result for each case value and phi node in the
5212 // common destination, as well as the min and max case values.
5213 assert(SI->case_begin() != SI->case_end());
5214 SwitchInst::CaseIt CI = SI->case_begin();
5215 ConstantInt *MinCaseVal = CI->getCaseValue();
5216 ConstantInt *MaxCaseVal = CI->getCaseValue();
5218 BasicBlock *CommonDest = nullptr;
5219 typedef SmallVector<std::pair<ConstantInt *, Constant *>, 4> ResultListTy;
5220 SmallDenseMap<PHINode *, ResultListTy> ResultLists;
5221 SmallDenseMap<PHINode *, Constant *> DefaultResults;
5222 SmallDenseMap<PHINode *, Type *> ResultTypes;
5223 SmallVector<PHINode *, 4> PHIs;
5225 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5226 ConstantInt *CaseVal = CI->getCaseValue();
5227 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5228 MinCaseVal = CaseVal;
5229 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5230 MaxCaseVal = CaseVal;
5232 // Resulting value at phi nodes for this case value.
5233 typedef SmallVector<std::pair<PHINode *, Constant *>, 4> ResultsTy;
5235 if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5239 // Append the result from this case to the list for each phi.
5240 for (const auto &I : Results) {
5241 PHINode *PHI = I.first;
5242 Constant *Value = I.second;
5243 if (!ResultLists.count(PHI))
5244 PHIs.push_back(PHI);
5245 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5249 // Keep track of the result types.
5250 for (PHINode *PHI : PHIs) {
5251 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5254 uint64_t NumResults = ResultLists[PHIs[0]].size();
5255 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5256 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5257 bool TableHasHoles = (NumResults < TableSize);
5259 // If the table has holes, we need a constant result for the default case
5260 // or a bitmask that fits in a register.
5261 SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5262 bool HasDefaultResults =
5263 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5264 DefaultResultsList, DL, TTI);
5266 bool NeedMask = (TableHasHoles && !HasDefaultResults);
5268 // As an extra penalty for the validity test we require more cases.
5269 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5271 if (!DL.fitsInLegalInteger(TableSize))
5275 for (const auto &I : DefaultResultsList) {
5276 PHINode *PHI = I.first;
5277 Constant *Result = I.second;
5278 DefaultResults[PHI] = Result;
5281 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5284 // Create the BB that does the lookups.
5285 Module &Mod = *CommonDest->getParent()->getParent();
5286 BasicBlock *LookupBB = BasicBlock::Create(
5287 Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5289 // Compute the table index value.
5290 Builder.SetInsertPoint(SI);
5292 Builder.CreateSub(SI->getCondition(), MinCaseVal, "switch.tableidx");
5294 // Compute the maximum table size representable by the integer type we are
5296 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5297 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5298 assert(MaxTableSize >= TableSize &&
5299 "It is impossible for a switch to have more entries than the max "
5300 "representable value of its input integer type's size.");
5302 // If the default destination is unreachable, or if the lookup table covers
5303 // all values of the conditional variable, branch directly to the lookup table
5304 // BB. Otherwise, check that the condition is within the case range.
5305 const bool DefaultIsReachable =
5306 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5307 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5308 BranchInst *RangeCheckBranch = nullptr;
5310 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5311 Builder.CreateBr(LookupBB);
5312 // Note: We call removeProdecessor later since we need to be able to get the
5313 // PHI value for the default case in case we're using a bit mask.
5315 Value *Cmp = Builder.CreateICmpULT(
5316 TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5318 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5321 // Populate the BB that does the lookups.
5322 Builder.SetInsertPoint(LookupBB);
5325 // Before doing the lookup we do the hole check.
5326 // The LookupBB is therefore re-purposed to do the hole check
5327 // and we create a new LookupBB.
5328 BasicBlock *MaskBB = LookupBB;
5329 MaskBB->setName("switch.hole_check");
5330 LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5331 CommonDest->getParent(), CommonDest);
5333 // Make the mask's bitwidth at least 8bit and a power-of-2 to avoid
5334 // unnecessary illegal types.
5335 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5336 APInt MaskInt(TableSizePowOf2, 0);
5337 APInt One(TableSizePowOf2, 1);
5338 // Build bitmask; fill in a 1 bit for every case.
5339 const ResultListTy &ResultList = ResultLists[PHIs[0]];
5340 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5341 uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5343 MaskInt |= One << Idx;
5345 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5347 // Get the TableIndex'th bit of the bitmask.
5348 // If this bit is 0 (meaning hole) jump to the default destination,
5349 // else continue with table lookup.
5350 IntegerType *MapTy = TableMask->getType();
5352 Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5353 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5354 Value *LoBit = Builder.CreateTrunc(
5355 Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5356 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5358 Builder.SetInsertPoint(LookupBB);
5359 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5362 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5363 // We cached PHINodes in PHIs, to avoid accessing deleted PHINodes later,
5364 // do not delete PHINodes here.
5365 SI->getDefaultDest()->removePredecessor(SI->getParent(),
5366 /*DontDeleteUselessPHIs=*/true);
5369 bool ReturnedEarly = false;
5370 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
5371 PHINode *PHI = PHIs[I];
5372 const ResultListTy &ResultList = ResultLists[PHI];
5374 // If using a bitmask, use any value to fill the lookup table holes.
5375 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5376 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL);
5378 Value *Result = Table.BuildLookup(TableIndex, Builder);
5380 // If the result is used to return immediately from the function, we want to
5381 // do that right here.
5382 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5383 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5384 Builder.CreateRet(Result);
5385 ReturnedEarly = true;
5389 // Do a small peephole optimization: re-use the switch table compare if
5391 if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5392 BasicBlock *PhiBlock = PHI->getParent();
5393 // Search for compare instructions which use the phi.
5394 for (auto *User : PHI->users()) {
5395 reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5399 PHI->addIncoming(Result, LookupBB);
5403 Builder.CreateBr(CommonDest);
5405 // Remove the switch.
5406 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5407 BasicBlock *Succ = SI->getSuccessor(i);
5409 if (Succ == SI->getDefaultDest())
5411 Succ->removePredecessor(SI->getParent());
5413 SI->eraseFromParent();
5417 ++NumLookupTablesHoles;
5421 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5422 // See also SelectionDAGBuilder::isDense(), which this function was based on.
5423 uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5424 uint64_t Range = Diff + 1;
5425 uint64_t NumCases = Values.size();
5426 // 40% is the default density for building a jump table in optsize/minsize mode.
5427 uint64_t MinDensity = 40;
5429 return NumCases * 100 >= Range * MinDensity;
5432 // Try and transform a switch that has "holes" in it to a contiguous sequence
5435 // A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5436 // range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5438 // This converts a sparse switch into a dense switch which allows better
5439 // lowering and could also allow transforming into a lookup table.
5440 static bool ReduceSwitchRange(SwitchInst *SI, IRBuilder<> &Builder,
5441 const DataLayout &DL,
5442 const TargetTransformInfo &TTI) {
5443 auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5444 if (CondTy->getIntegerBitWidth() > 64 ||
5445 !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5447 // Only bother with this optimization if there are more than 3 switch cases;
5448 // SDAG will only bother creating jump tables for 4 or more cases.
5449 if (SI->getNumCases() < 4)
5452 // This transform is agnostic to the signedness of the input or case values. We
5453 // can treat the case values as signed or unsigned. We can optimize more common
5454 // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5456 SmallVector<int64_t,4> Values;
5457 for (auto &C : SI->cases())
5458 Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5459 std::sort(Values.begin(), Values.end());
5461 // If the switch is already dense, there's nothing useful to do here.
5462 if (isSwitchDense(Values))
5465 // First, transform the values such that they start at zero and ascend.
5466 int64_t Base = Values[0];
5467 for (auto &V : Values)
5470 // Now we have signed numbers that have been shifted so that, given enough
5471 // precision, there are no negative values. Since the rest of the transform
5472 // is bitwise only, we switch now to an unsigned representation.
5474 for (auto &V : Values)
5475 GCD = GreatestCommonDivisor64(GCD, (uint64_t)V);
5477 // This transform can be done speculatively because it is so cheap - it results
5478 // in a single rotate operation being inserted. This can only happen if the
5479 // factor extracted is a power of 2.
5480 // FIXME: If the GCD is an odd number we can multiply by the multiplicative
5481 // inverse of GCD and then perform this transform.
5482 // FIXME: It's possible that optimizing a switch on powers of two might also
5483 // be beneficial - flag values are often powers of two and we could use a CLZ
5484 // as the key function.
5485 if (GCD <= 1 || !isPowerOf2_64(GCD))
5486 // No common divisor found or too expensive to compute key function.
5489 unsigned Shift = Log2_64(GCD);
5490 for (auto &V : Values)
5491 V = (int64_t)((uint64_t)V >> Shift);
5493 if (!isSwitchDense(Values))
5494 // Transform didn't create a dense switch.
5497 // The obvious transform is to shift the switch condition right and emit a
5498 // check that the condition actually cleanly divided by GCD, i.e.
5499 // C & (1 << Shift - 1) == 0
5500 // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5502 // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5503 // shift and puts the shifted-off bits in the uppermost bits. If any of these
5504 // are nonzero then the switch condition will be very large and will hit the
5507 auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5508 Builder.SetInsertPoint(SI);
5509 auto *ShiftC = ConstantInt::get(Ty, Shift);
5510 auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5511 auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5512 auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5513 auto *Rot = Builder.CreateOr(LShr, Shl);
5514 SI->replaceUsesOfWith(SI->getCondition(), Rot);
5516 for (auto Case : SI->cases()) {
5517 auto *Orig = Case.getCaseValue();
5518 auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5520 cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5525 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
5526 BasicBlock *BB = SI->getParent();
5528 if (isValueEqualityComparison(SI)) {
5529 // If we only have one predecessor, and if it is a branch on this value,
5530 // see if that predecessor totally determines the outcome of this switch.
5531 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5532 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
5533 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5535 Value *Cond = SI->getCondition();
5536 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
5537 if (SimplifySwitchOnSelect(SI, Select))
5538 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5540 // If the block only contains the switch, see if we can fold the block
5541 // away into any preds.
5542 BasicBlock::iterator BBI = BB->begin();
5543 // Ignore dbg intrinsics.
5544 while (isa<DbgInfoIntrinsic>(BBI))
5547 if (FoldValueComparisonIntoPredecessors(SI, Builder))
5548 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5551 // Try to transform the switch into an icmp and a branch.
5552 if (TurnSwitchRangeIntoICmp(SI, Builder))
5553 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5555 // Remove unreachable cases.
5556 if (EliminateDeadSwitchCases(SI, AC, DL))
5557 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5559 if (SwitchToSelect(SI, Builder, AC, DL, TTI))
5560 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5562 if (ForwardSwitchConditionToPHI(SI))
5563 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5565 // The conversion from switch to lookup tables results in difficult
5566 // to analyze code and makes pruning branches much harder.
5567 // This is a problem of the switch expression itself can still be
5568 // restricted as a result of inlining or CVP. There only apply this
5569 // transformation during late steps of the optimisation chain.
5570 if (LateSimplifyCFG && SwitchToLookupTable(SI, Builder, DL, TTI))
5571 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5573 if (ReduceSwitchRange(SI, Builder, DL, TTI))
5574 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5579 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
5580 BasicBlock *BB = IBI->getParent();
5581 bool Changed = false;
5583 // Eliminate redundant destinations.
5584 SmallPtrSet<Value *, 8> Succs;
5585 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
5586 BasicBlock *Dest = IBI->getDestination(i);
5587 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
5588 Dest->removePredecessor(BB);
5589 IBI->removeDestination(i);
5596 if (IBI->getNumDestinations() == 0) {
5597 // If the indirectbr has no successors, change it to unreachable.
5598 new UnreachableInst(IBI->getContext(), IBI);
5599 EraseTerminatorInstAndDCECond(IBI);
5603 if (IBI->getNumDestinations() == 1) {
5604 // If the indirectbr has one successor, change it to a direct branch.
5605 BranchInst::Create(IBI->getDestination(0), IBI);
5606 EraseTerminatorInstAndDCECond(IBI);
5610 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
5611 if (SimplifyIndirectBrOnSelect(IBI, SI))
5612 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5617 /// Given an block with only a single landing pad and a unconditional branch
5618 /// try to find another basic block which this one can be merged with. This
5619 /// handles cases where we have multiple invokes with unique landing pads, but
5620 /// a shared handler.
5622 /// We specifically choose to not worry about merging non-empty blocks
5623 /// here. That is a PRE/scheduling problem and is best solved elsewhere. In
5624 /// practice, the optimizer produces empty landing pad blocks quite frequently
5625 /// when dealing with exception dense code. (see: instcombine, gvn, if-else
5626 /// sinking in this file)
5628 /// This is primarily a code size optimization. We need to avoid performing
5629 /// any transform which might inhibit optimization (such as our ability to
5630 /// specialize a particular handler via tail commoning). We do this by not
5631 /// merging any blocks which require us to introduce a phi. Since the same
5632 /// values are flowing through both blocks, we don't loose any ability to
5633 /// specialize. If anything, we make such specialization more likely.
5635 /// TODO - This transformation could remove entries from a phi in the target
5636 /// block when the inputs in the phi are the same for the two blocks being
5637 /// merged. In some cases, this could result in removal of the PHI entirely.
5638 static bool TryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI,
5640 auto Succ = BB->getUniqueSuccessor();
5642 // If there's a phi in the successor block, we'd likely have to introduce
5643 // a phi into the merged landing pad block.
5644 if (isa<PHINode>(*Succ->begin()))
5647 for (BasicBlock *OtherPred : predecessors(Succ)) {
5648 if (BB == OtherPred)
5650 BasicBlock::iterator I = OtherPred->begin();
5651 LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(I);
5652 if (!LPad2 || !LPad2->isIdenticalTo(LPad))
5654 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {
5656 BranchInst *BI2 = dyn_cast<BranchInst>(I);
5657 if (!BI2 || !BI2->isIdenticalTo(BI))
5660 // We've found an identical block. Update our predecessors to take that
5661 // path instead and make ourselves dead.
5662 SmallSet<BasicBlock *, 16> Preds;
5663 Preds.insert(pred_begin(BB), pred_end(BB));
5664 for (BasicBlock *Pred : Preds) {
5665 InvokeInst *II = cast<InvokeInst>(Pred->getTerminator());
5666 assert(II->getNormalDest() != BB && II->getUnwindDest() == BB &&
5667 "unexpected successor");
5668 II->setUnwindDest(OtherPred);
5671 // The debug info in OtherPred doesn't cover the merged control flow that
5672 // used to go through BB. We need to delete it or update it.
5673 for (auto I = OtherPred->begin(), E = OtherPred->end(); I != E;) {
5674 Instruction &Inst = *I;
5676 if (isa<DbgInfoIntrinsic>(Inst))
5677 Inst.eraseFromParent();
5680 SmallSet<BasicBlock *, 16> Succs;
5681 Succs.insert(succ_begin(BB), succ_end(BB));
5682 for (BasicBlock *Succ : Succs) {
5683 Succ->removePredecessor(BB);
5686 IRBuilder<> Builder(BI);
5687 Builder.CreateUnreachable();
5688 BI->eraseFromParent();
5694 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI,
5695 IRBuilder<> &Builder) {
5696 BasicBlock *BB = BI->getParent();
5698 if (SinkCommon && SinkThenElseCodeToEnd(BI))
5701 // If the Terminator is the only non-phi instruction, simplify the block.
5702 // if LoopHeader is provided, check if the block is a loop header
5703 // (This is for early invocations before loop simplify and vectorization
5704 // to keep canonical loop forms for nested loops.
5705 // These blocks can be eliminated when the pass is invoked later
5706 // in the back-end.)
5707 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg()->getIterator();
5708 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
5709 (!LoopHeaders || !LoopHeaders->count(BB)) &&
5710 TryToSimplifyUncondBranchFromEmptyBlock(BB))
5713 // If the only instruction in the block is a seteq/setne comparison
5714 // against a constant, try to simplify the block.
5715 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
5716 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
5717 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5719 if (I->isTerminator() &&
5720 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, DL, TTI,
5721 BonusInstThreshold, AC))
5725 // See if we can merge an empty landing pad block with another which is
5727 if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(I)) {
5728 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {
5730 if (I->isTerminator() && TryToMergeLandingPad(LPad, BI, BB))
5734 // If this basic block is ONLY a compare and a branch, and if a predecessor
5735 // branches to us and our successor, fold the comparison into the
5736 // predecessor and use logical operations to update the incoming value
5737 // for PHI nodes in common successor.
5738 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
5739 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5743 static BasicBlock *allPredecessorsComeFromSameSource(BasicBlock *BB) {
5744 BasicBlock *PredPred = nullptr;
5745 for (auto *P : predecessors(BB)) {
5746 BasicBlock *PPred = P->getSinglePredecessor();
5747 if (!PPred || (PredPred && PredPred != PPred))
5754 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
5755 BasicBlock *BB = BI->getParent();
5757 // Conditional branch
5758 if (isValueEqualityComparison(BI)) {
5759 // If we only have one predecessor, and if it is a branch on this value,
5760 // see if that predecessor totally determines the outcome of this
5762 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5763 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
5764 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5766 // This block must be empty, except for the setcond inst, if it exists.
5767 // Ignore dbg intrinsics.
5768 BasicBlock::iterator I = BB->begin();
5769 // Ignore dbg intrinsics.
5770 while (isa<DbgInfoIntrinsic>(I))
5773 if (FoldValueComparisonIntoPredecessors(BI, Builder))
5774 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5775 } else if (&*I == cast<Instruction>(BI->getCondition())) {
5777 // Ignore dbg intrinsics.
5778 while (isa<DbgInfoIntrinsic>(I))
5780 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
5781 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5785 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
5786 if (SimplifyBranchOnICmpChain(BI, Builder, DL))
5789 // If this basic block has a single dominating predecessor block and the
5790 // dominating block's condition implies BI's condition, we know the direction
5791 // of the BI branch.
5792 if (BasicBlock *Dom = BB->getSinglePredecessor()) {
5793 auto *PBI = dyn_cast_or_null<BranchInst>(Dom->getTerminator());
5794 if (PBI && PBI->isConditional() &&
5795 PBI->getSuccessor(0) != PBI->getSuccessor(1) &&
5796 (PBI->getSuccessor(0) == BB || PBI->getSuccessor(1) == BB)) {
5797 bool CondIsFalse = PBI->getSuccessor(1) == BB;
5798 Optional<bool> Implication = isImpliedCondition(
5799 PBI->getCondition(), BI->getCondition(), DL, CondIsFalse);
5801 // Turn this into a branch on constant.
5802 auto *OldCond = BI->getCondition();
5803 ConstantInt *CI = *Implication
5804 ? ConstantInt::getTrue(BB->getContext())
5805 : ConstantInt::getFalse(BB->getContext());
5806 BI->setCondition(CI);
5807 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
5808 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5813 // If this basic block is ONLY a compare and a branch, and if a predecessor
5814 // branches to us and one of our successors, fold the comparison into the
5815 // predecessor and use logical operations to pick the right destination.
5816 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
5817 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5819 // We have a conditional branch to two blocks that are only reachable
5820 // from BI. We know that the condbr dominates the two blocks, so see if
5821 // there is any identical code in the "then" and "else" blocks. If so, we
5822 // can hoist it up to the branching block.
5823 if (BI->getSuccessor(0)->getSinglePredecessor()) {
5824 if (BI->getSuccessor(1)->getSinglePredecessor()) {
5825 if (HoistThenElseCodeToIf(BI, TTI))
5826 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5828 // If Successor #1 has multiple preds, we may be able to conditionally
5829 // execute Successor #0 if it branches to Successor #1.
5830 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
5831 if (Succ0TI->getNumSuccessors() == 1 &&
5832 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
5833 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), TTI))
5834 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5836 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
5837 // If Successor #0 has multiple preds, we may be able to conditionally
5838 // execute Successor #1 if it branches to Successor #0.
5839 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
5840 if (Succ1TI->getNumSuccessors() == 1 &&
5841 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
5842 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), TTI))
5843 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5846 // If this is a branch on a phi node in the current block, thread control
5847 // through this block if any PHI node entries are constants.
5848 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
5849 if (PN->getParent() == BI->getParent())
5850 if (FoldCondBranchOnPHI(BI, DL, AC))
5851 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5853 // Scan predecessor blocks for conditional branches.
5854 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
5855 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
5856 if (PBI != BI && PBI->isConditional())
5857 if (SimplifyCondBranchToCondBranch(PBI, BI, DL))
5858 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5860 // Look for diamond patterns.
5861 if (MergeCondStores)
5862 if (BasicBlock *PrevBB = allPredecessorsComeFromSameSource(BB))
5863 if (BranchInst *PBI = dyn_cast<BranchInst>(PrevBB->getTerminator()))
5864 if (PBI != BI && PBI->isConditional())
5865 if (mergeConditionalStores(PBI, BI))
5866 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5871 /// Check if passing a value to an instruction will cause undefined behavior.
5872 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
5873 Constant *C = dyn_cast<Constant>(V);
5880 if (C->isNullValue() || isa<UndefValue>(C)) {
5881 // Only look at the first use, avoid hurting compile time with long uselists
5882 User *Use = *I->user_begin();
5884 // Now make sure that there are no instructions in between that can alter
5885 // control flow (eg. calls)
5886 for (BasicBlock::iterator
5887 i = ++BasicBlock::iterator(I),
5888 UI = BasicBlock::iterator(dyn_cast<Instruction>(Use));
5890 if (i == I->getParent()->end() || i->mayHaveSideEffects())
5893 // Look through GEPs. A load from a GEP derived from NULL is still undefined
5894 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
5895 if (GEP->getPointerOperand() == I)
5896 return passingValueIsAlwaysUndefined(V, GEP);
5898 // Look through bitcasts.
5899 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
5900 return passingValueIsAlwaysUndefined(V, BC);
5902 // Load from null is undefined.
5903 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
5904 if (!LI->isVolatile())
5905 return LI->getPointerAddressSpace() == 0;
5907 // Store to null is undefined.
5908 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
5909 if (!SI->isVolatile())
5910 return SI->getPointerAddressSpace() == 0 &&
5911 SI->getPointerOperand() == I;
5913 // A call to null is undefined.
5914 if (auto CS = CallSite(Use))
5915 return CS.getCalledValue() == I;
5920 /// If BB has an incoming value that will always trigger undefined behavior
5921 /// (eg. null pointer dereference), remove the branch leading here.
5922 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
5923 for (BasicBlock::iterator i = BB->begin();
5924 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
5925 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
5926 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
5927 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
5928 IRBuilder<> Builder(T);
5929 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
5930 BB->removePredecessor(PHI->getIncomingBlock(i));
5931 // Turn uncoditional branches into unreachables and remove the dead
5932 // destination from conditional branches.
5933 if (BI->isUnconditional())
5934 Builder.CreateUnreachable();
5936 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1)
5937 : BI->getSuccessor(0));
5938 BI->eraseFromParent();
5941 // TODO: SwitchInst.
5947 bool SimplifyCFGOpt::run(BasicBlock *BB) {
5948 bool Changed = false;
5950 assert(BB && BB->getParent() && "Block not embedded in function!");
5951 assert(BB->getTerminator() && "Degenerate basic block encountered!");
5953 // Remove basic blocks that have no predecessors (except the entry block)...
5954 // or that just have themself as a predecessor. These are unreachable.
5955 if ((pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) ||
5956 BB->getSinglePredecessor() == BB) {
5957 DEBUG(dbgs() << "Removing BB: \n" << *BB);
5958 DeleteDeadBlock(BB);
5962 // Check to see if we can constant propagate this terminator instruction
5964 Changed |= ConstantFoldTerminator(BB, true);
5966 // Check for and eliminate duplicate PHI nodes in this block.
5967 Changed |= EliminateDuplicatePHINodes(BB);
5969 // Check for and remove branches that will always cause undefined behavior.
5970 Changed |= removeUndefIntroducingPredecessor(BB);
5972 // Merge basic blocks into their predecessor if there is only one distinct
5973 // pred, and if there is only one distinct successor of the predecessor, and
5974 // if there are no PHI nodes.
5976 if (MergeBlockIntoPredecessor(BB))
5979 IRBuilder<> Builder(BB);
5981 // If there is a trivial two-entry PHI node in this basic block, and we can
5982 // eliminate it, do so now.
5983 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
5984 if (PN->getNumIncomingValues() == 2)
5985 Changed |= FoldTwoEntryPHINode(PN, TTI, DL);
5987 Builder.SetInsertPoint(BB->getTerminator());
5988 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
5989 if (BI->isUnconditional()) {
5990 if (SimplifyUncondBranch(BI, Builder))
5993 if (SimplifyCondBranch(BI, Builder))
5996 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
5997 if (SimplifyReturn(RI, Builder))
5999 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
6000 if (SimplifyResume(RI, Builder))
6002 } else if (CleanupReturnInst *RI =
6003 dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
6004 if (SimplifyCleanupReturn(RI))
6006 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
6007 if (SimplifySwitch(SI, Builder))
6009 } else if (UnreachableInst *UI =
6010 dyn_cast<UnreachableInst>(BB->getTerminator())) {
6011 if (SimplifyUnreachable(UI))
6013 } else if (IndirectBrInst *IBI =
6014 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
6015 if (SimplifyIndirectBr(IBI))
6022 /// This function is used to do simplification of a CFG.
6023 /// For example, it adjusts branches to branches to eliminate the extra hop,
6024 /// eliminates unreachable basic blocks, and does other "peephole" optimization
6025 /// of the CFG. It returns true if a modification was made.
6027 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
6028 unsigned BonusInstThreshold, AssumptionCache *AC,
6029 SmallPtrSetImpl<BasicBlock *> *LoopHeaders,
6030 bool LateSimplifyCFG) {
6031 return SimplifyCFGOpt(TTI, BB->getModule()->getDataLayout(),
6032 BonusInstThreshold, AC, LoopHeaders, LateSimplifyCFG)