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/KnownBits.h"
64 #include "llvm/Support/MathExtras.h"
65 #include "llvm/Support/raw_ostream.h"
66 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
67 #include "llvm/Transforms/Utils/Local.h"
68 #include "llvm/Transforms/Utils/ValueMapper.h"
81 using namespace PatternMatch;
83 #define DEBUG_TYPE "simplifycfg"
85 // Chosen as 2 so as to be cheap, but still to have enough power to fold
86 // a select, so the "clamp" idiom (of a min followed by a max) will be caught.
87 // To catch this, we need to fold a compare and a select, hence '2' being the
88 // minimum reasonable default.
89 static cl::opt<unsigned> PHINodeFoldingThreshold(
90 "phi-node-folding-threshold", cl::Hidden, cl::init(2),
92 "Control the amount of phi node folding to perform (default = 2)"));
94 static cl::opt<bool> DupRet(
95 "simplifycfg-dup-ret", cl::Hidden, cl::init(false),
96 cl::desc("Duplicate return instructions into unconditional branches"));
99 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
100 cl::desc("Sink common instructions down to the end block"));
102 static cl::opt<bool> HoistCondStores(
103 "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
104 cl::desc("Hoist conditional stores if an unconditional store precedes"));
106 static cl::opt<bool> MergeCondStores(
107 "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
108 cl::desc("Hoist conditional stores even if an unconditional store does not "
109 "precede - hoist multiple conditional stores into a single "
110 "predicated store"));
112 static cl::opt<bool> MergeCondStoresAggressively(
113 "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
114 cl::desc("When merging conditional stores, do so even if the resultant "
115 "basic blocks are unlikely to be if-converted as a result"));
117 static cl::opt<bool> SpeculateOneExpensiveInst(
118 "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
119 cl::desc("Allow exactly one expensive instruction to be speculatively "
122 static cl::opt<unsigned> MaxSpeculationDepth(
123 "max-speculation-depth", cl::Hidden, cl::init(10),
124 cl::desc("Limit maximum recursion depth when calculating costs of "
125 "speculatively executed instructions"));
127 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
128 STATISTIC(NumLinearMaps,
129 "Number of switch instructions turned into linear mapping");
130 STATISTIC(NumLookupTables,
131 "Number of switch instructions turned into lookup tables");
133 NumLookupTablesHoles,
134 "Number of switch instructions turned into lookup tables (holes checked)");
135 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
136 STATISTIC(NumSinkCommons,
137 "Number of common instructions sunk down to the end block");
138 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
142 // The first field contains the value that the switch produces when a certain
143 // case group is selected, and the second field is a vector containing the
144 // cases composing the case group.
145 typedef SmallVector<std::pair<Constant *, SmallVector<ConstantInt *, 4>>, 2>
146 SwitchCaseResultVectorTy;
147 // The first field contains the phi node that generates a result of the switch
148 // and the second field contains the value generated for a certain case in the
149 // switch for that PHI.
150 typedef SmallVector<std::pair<PHINode *, Constant *>, 4> SwitchCaseResultsTy;
152 /// ValueEqualityComparisonCase - Represents a case of a switch.
153 struct ValueEqualityComparisonCase {
157 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
158 : Value(Value), Dest(Dest) {}
160 bool operator<(ValueEqualityComparisonCase RHS) const {
161 // Comparing pointers is ok as we only rely on the order for uniquing.
162 return Value < RHS.Value;
165 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
168 class SimplifyCFGOpt {
169 const TargetTransformInfo &TTI;
170 const DataLayout &DL;
171 unsigned BonusInstThreshold;
173 SmallPtrSetImpl<BasicBlock *> *LoopHeaders;
174 // See comments in SimplifyCFGOpt::SimplifySwitch.
175 bool LateSimplifyCFG;
176 Value *isValueEqualityComparison(TerminatorInst *TI);
177 BasicBlock *GetValueEqualityComparisonCases(
178 TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases);
179 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
181 IRBuilder<> &Builder);
182 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
183 IRBuilder<> &Builder);
185 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
186 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
187 bool SimplifySingleResume(ResumeInst *RI);
188 bool SimplifyCommonResume(ResumeInst *RI);
189 bool SimplifyCleanupReturn(CleanupReturnInst *RI);
190 bool SimplifyUnreachable(UnreachableInst *UI);
191 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
192 bool SimplifyIndirectBr(IndirectBrInst *IBI);
193 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
194 bool SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
197 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
198 unsigned BonusInstThreshold, AssumptionCache *AC,
199 SmallPtrSetImpl<BasicBlock *> *LoopHeaders,
200 bool LateSimplifyCFG)
201 : TTI(TTI), DL(DL), BonusInstThreshold(BonusInstThreshold), AC(AC),
202 LoopHeaders(LoopHeaders), LateSimplifyCFG(LateSimplifyCFG) {}
204 bool run(BasicBlock *BB);
207 } // end anonymous namespace
209 /// Return true if it is safe to merge these two
210 /// terminator instructions together.
212 SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2,
213 SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
215 return false; // Can't merge with self!
217 // It is not safe to merge these two switch instructions if they have a common
218 // successor, and if that successor has a PHI node, and if *that* PHI node has
219 // conflicting incoming values from the two switch blocks.
220 BasicBlock *SI1BB = SI1->getParent();
221 BasicBlock *SI2BB = SI2->getParent();
223 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
225 for (BasicBlock *Succ : successors(SI2BB))
226 if (SI1Succs.count(Succ))
227 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
228 PHINode *PN = cast<PHINode>(BBI);
229 if (PN->getIncomingValueForBlock(SI1BB) !=
230 PN->getIncomingValueForBlock(SI2BB)) {
232 FailBlocks->insert(Succ);
240 /// Return true if it is safe and profitable to merge these two terminator
241 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
242 /// store all PHI nodes in common successors.
244 isProfitableToFoldUnconditional(BranchInst *SI1, BranchInst *SI2,
246 SmallVectorImpl<PHINode *> &PhiNodes) {
248 return false; // Can't merge with self!
249 assert(SI1->isUnconditional() && SI2->isConditional());
251 // We fold the unconditional branch if we can easily update all PHI nodes in
252 // common successors:
253 // 1> We have a constant incoming value for the conditional branch;
254 // 2> We have "Cond" as the incoming value for the unconditional branch;
255 // 3> SI2->getCondition() and Cond have same operands.
256 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
259 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
260 Cond->getOperand(1) == Ci2->getOperand(1)) &&
261 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
262 Cond->getOperand(1) == Ci2->getOperand(0)))
265 BasicBlock *SI1BB = SI1->getParent();
266 BasicBlock *SI2BB = SI2->getParent();
267 SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
268 for (BasicBlock *Succ : successors(SI2BB))
269 if (SI1Succs.count(Succ))
270 for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
271 PHINode *PN = cast<PHINode>(BBI);
272 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
273 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
275 PhiNodes.push_back(PN);
280 /// Update PHI nodes in Succ to indicate that there will now be entries in it
281 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
282 /// will be the same as those coming in from ExistPred, an existing predecessor
284 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
285 BasicBlock *ExistPred) {
286 if (!isa<PHINode>(Succ->begin()))
287 return; // Quick exit if nothing to do
290 for (BasicBlock::iterator I = Succ->begin(); (PN = dyn_cast<PHINode>(I)); ++I)
291 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
294 /// Compute an abstract "cost" of speculating the given instruction,
295 /// which is assumed to be safe to speculate. TCC_Free means cheap,
296 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
298 static unsigned ComputeSpeculationCost(const User *I,
299 const TargetTransformInfo &TTI) {
300 assert(isSafeToSpeculativelyExecute(I) &&
301 "Instruction is not safe to speculatively execute!");
302 return TTI.getUserCost(I);
305 /// If we have a merge point of an "if condition" as accepted above,
306 /// return true if the specified value dominates the block. We
307 /// don't handle the true generality of domination here, just a special case
308 /// which works well enough for us.
310 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
311 /// see if V (which must be an instruction) and its recursive operands
312 /// that do not dominate BB have a combined cost lower than CostRemaining and
313 /// are non-trapping. If both are true, the instruction is inserted into the
314 /// set and true is returned.
316 /// The cost for most non-trapping instructions is defined as 1 except for
317 /// Select whose cost is 2.
319 /// After this function returns, CostRemaining is decreased by the cost of
320 /// V plus its non-dominating operands. If that cost is greater than
321 /// CostRemaining, false is returned and CostRemaining is undefined.
322 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
323 SmallPtrSetImpl<Instruction *> *AggressiveInsts,
324 unsigned &CostRemaining,
325 const TargetTransformInfo &TTI,
326 unsigned Depth = 0) {
327 // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
328 // so limit the recursion depth.
329 // TODO: While this recursion limit does prevent pathological behavior, it
330 // would be better to track visited instructions to avoid cycles.
331 if (Depth == MaxSpeculationDepth)
334 Instruction *I = dyn_cast<Instruction>(V);
336 // Non-instructions all dominate instructions, but not all constantexprs
337 // can be executed unconditionally.
338 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
343 BasicBlock *PBB = I->getParent();
345 // We don't want to allow weird loops that might have the "if condition" in
346 // the bottom of this block.
350 // If this instruction is defined in a block that contains an unconditional
351 // branch to BB, then it must be in the 'conditional' part of the "if
352 // statement". If not, it definitely dominates the region.
353 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
354 if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
357 // If we aren't allowing aggressive promotion anymore, then don't consider
358 // instructions in the 'if region'.
359 if (!AggressiveInsts)
362 // If we have seen this instruction before, don't count it again.
363 if (AggressiveInsts->count(I))
366 // Okay, it looks like the instruction IS in the "condition". Check to
367 // see if it's a cheap instruction to unconditionally compute, and if it
368 // only uses stuff defined outside of the condition. If so, hoist it out.
369 if (!isSafeToSpeculativelyExecute(I))
372 unsigned Cost = ComputeSpeculationCost(I, TTI);
374 // Allow exactly one instruction to be speculated regardless of its cost
375 // (as long as it is safe to do so).
376 // This is intended to flatten the CFG even if the instruction is a division
377 // or other expensive operation. The speculation of an expensive instruction
378 // is expected to be undone in CodeGenPrepare if the speculation has not
379 // enabled further IR optimizations.
380 if (Cost > CostRemaining &&
381 (!SpeculateOneExpensiveInst || !AggressiveInsts->empty() || Depth > 0))
384 // Avoid unsigned wrap.
385 CostRemaining = (Cost > CostRemaining) ? 0 : CostRemaining - Cost;
387 // Okay, we can only really hoist these out if their operands do
388 // not take us over the cost threshold.
389 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
390 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, TTI,
393 // Okay, it's safe to do this! Remember this instruction.
394 AggressiveInsts->insert(I);
398 /// Extract ConstantInt from value, looking through IntToPtr
399 /// and PointerNullValue. Return NULL if value is not a constant int.
400 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
401 // Normal constant int.
402 ConstantInt *CI = dyn_cast<ConstantInt>(V);
403 if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
406 // This is some kind of pointer constant. Turn it into a pointer-sized
407 // ConstantInt if possible.
408 IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
410 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
411 if (isa<ConstantPointerNull>(V))
412 return ConstantInt::get(PtrTy, 0);
414 // IntToPtr const int.
415 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
416 if (CE->getOpcode() == Instruction::IntToPtr)
417 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
418 // The constant is very likely to have the right type already.
419 if (CI->getType() == PtrTy)
422 return cast<ConstantInt>(
423 ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
430 /// Given a chain of or (||) or and (&&) comparison of a value against a
431 /// constant, this will try to recover the information required for a switch
433 /// It will depth-first traverse the chain of comparison, seeking for patterns
434 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
435 /// representing the different cases for the switch.
436 /// Note that if the chain is composed of '||' it will build the set of elements
437 /// that matches the comparisons (i.e. any of this value validate the chain)
438 /// while for a chain of '&&' it will build the set elements that make the test
440 struct ConstantComparesGatherer {
441 const DataLayout &DL;
442 Value *CompValue; /// Value found for the switch comparison
443 Value *Extra; /// Extra clause to be checked before the switch
444 SmallVector<ConstantInt *, 8> Vals; /// Set of integers to match in switch
445 unsigned UsedICmps; /// Number of comparisons matched in the and/or chain
447 /// Construct and compute the result for the comparison instruction Cond
448 ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL)
449 : DL(DL), CompValue(nullptr), Extra(nullptr), UsedICmps(0) {
454 ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
455 ConstantComparesGatherer &
456 operator=(const ConstantComparesGatherer &) = delete;
459 /// Try to set the current value used for the comparison, it succeeds only if
460 /// it wasn't set before or if the new value is the same as the old one
461 bool setValueOnce(Value *NewVal) {
462 if (CompValue && CompValue != NewVal)
465 return (CompValue != nullptr);
468 /// Try to match Instruction "I" as a comparison against a constant and
469 /// populates the array Vals with the set of values that match (or do not
470 /// match depending on isEQ).
471 /// Return false on failure. On success, the Value the comparison matched
472 /// against is placed in CompValue.
473 /// If CompValue is already set, the function is expected to fail if a match
474 /// is found but the value compared to is different.
475 bool matchInstruction(Instruction *I, bool isEQ) {
476 // If this is an icmp against a constant, handle this as one of the cases.
479 if (!((ICI = dyn_cast<ICmpInst>(I)) &&
480 (C = GetConstantInt(I->getOperand(1), DL)))) {
487 // Pattern match a special case
488 // (x & ~2^z) == y --> x == y || x == y|2^z
489 // This undoes a transformation done by instcombine to fuse 2 compares.
490 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
492 // It's a little bit hard to see why the following transformations are
493 // correct. Here is a CVC3 program to verify them for 64-bit values:
496 ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
500 mask : BITVECTOR(64) = BVSHL(ONE, z);
501 QUERY( (y & ~mask = y) =>
502 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
504 QUERY( (y | mask = y) =>
505 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
509 // Please note that each pattern must be a dual implication (<--> or
510 // iff). One directional implication can create spurious matches. If the
511 // implication is only one-way, an unsatisfiable condition on the left
512 // side can imply a satisfiable condition on the right side. Dual
513 // implication ensures that satisfiable conditions are transformed to
514 // other satisfiable conditions and unsatisfiable conditions are
515 // transformed to other unsatisfiable conditions.
517 // Here is a concrete example of a unsatisfiable condition on the left
518 // implying a satisfiable condition on the right:
521 // (x & ~mask) == y --> (x == y || x == (y | mask))
523 // Substituting y = 3, z = 0 yields:
524 // (x & -2) == 3 --> (x == 3 || x == 2)
526 // Pattern match a special case:
528 QUERY( (y & ~mask = y) =>
529 ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
532 if (match(ICI->getOperand(0),
533 m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
535 if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
536 // If we already have a value for the switch, it has to match!
537 if (!setValueOnce(RHSVal))
542 ConstantInt::get(C->getContext(),
543 C->getValue() | Mask));
549 // Pattern match a special case:
551 QUERY( (y | mask = y) =>
552 ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
555 if (match(ICI->getOperand(0),
556 m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
558 if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
559 // If we already have a value for the switch, it has to match!
560 if (!setValueOnce(RHSVal))
564 Vals.push_back(ConstantInt::get(C->getContext(),
565 C->getValue() & ~Mask));
571 // If we already have a value for the switch, it has to match!
572 if (!setValueOnce(ICI->getOperand(0)))
577 return ICI->getOperand(0);
580 // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
581 ConstantRange Span = ConstantRange::makeAllowedICmpRegion(
582 ICI->getPredicate(), C->getValue());
584 // Shift the range if the compare is fed by an add. This is the range
585 // compare idiom as emitted by instcombine.
586 Value *CandidateVal = I->getOperand(0);
587 if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
588 Span = Span.subtract(*RHSC);
589 CandidateVal = RHSVal;
592 // If this is an and/!= check, then we are looking to build the set of
593 // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
596 Span = Span.inverse();
598 // If there are a ton of values, we don't want to make a ginormous switch.
599 if (Span.isSizeLargerThan(8) || Span.isEmptySet()) {
603 // If we already have a value for the switch, it has to match!
604 if (!setValueOnce(CandidateVal))
607 // Add all values from the range to the set
608 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
609 Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
615 /// Given a potentially 'or'd or 'and'd together collection of icmp
616 /// eq/ne/lt/gt instructions that compare a value against a constant, extract
617 /// the value being compared, and stick the list constants into the Vals
619 /// One "Extra" case is allowed to differ from the other.
620 void gather(Value *V) {
621 Instruction *I = dyn_cast<Instruction>(V);
622 bool isEQ = (I->getOpcode() == Instruction::Or);
624 // Keep a stack (SmallVector for efficiency) for depth-first traversal
625 SmallVector<Value *, 8> DFT;
626 SmallPtrSet<Value *, 8> Visited;
632 while (!DFT.empty()) {
633 V = DFT.pop_back_val();
635 if (Instruction *I = dyn_cast<Instruction>(V)) {
636 // If it is a || (or && depending on isEQ), process the operands.
637 if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
638 if (Visited.insert(I->getOperand(1)).second)
639 DFT.push_back(I->getOperand(1));
640 if (Visited.insert(I->getOperand(0)).second)
641 DFT.push_back(I->getOperand(0));
645 // Try to match the current instruction
646 if (matchInstruction(I, isEQ))
647 // Match succeed, continue the loop
651 // One element of the sequence of || (or &&) could not be match as a
652 // comparison against the same value as the others.
653 // We allow only one "Extra" case to be checked before the switch
658 // Failed to parse a proper sequence, abort now
665 } // end anonymous namespace
667 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
668 Instruction *Cond = nullptr;
669 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
670 Cond = dyn_cast<Instruction>(SI->getCondition());
671 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
672 if (BI->isConditional())
673 Cond = dyn_cast<Instruction>(BI->getCondition());
674 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
675 Cond = dyn_cast<Instruction>(IBI->getAddress());
678 TI->eraseFromParent();
680 RecursivelyDeleteTriviallyDeadInstructions(Cond);
683 /// Return true if the specified terminator checks
684 /// to see if a value is equal to constant integer value.
685 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
687 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
688 // Do not permit merging of large switch instructions into their
689 // predecessors unless there is only one predecessor.
690 if (SI->getNumSuccessors() * std::distance(pred_begin(SI->getParent()),
691 pred_end(SI->getParent())) <=
693 CV = SI->getCondition();
694 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
695 if (BI->isConditional() && BI->getCondition()->hasOneUse())
696 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
697 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
698 CV = ICI->getOperand(0);
701 // Unwrap any lossless ptrtoint cast.
703 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
704 Value *Ptr = PTII->getPointerOperand();
705 if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
712 /// Given a value comparison instruction,
713 /// decode all of the 'cases' that it represents and return the 'default' block.
714 BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
715 TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
716 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
717 Cases.reserve(SI->getNumCases());
718 for (auto Case : SI->cases())
719 Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),
720 Case.getCaseSuccessor()));
721 return SI->getDefaultDest();
724 BranchInst *BI = cast<BranchInst>(TI);
725 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
726 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
727 Cases.push_back(ValueEqualityComparisonCase(
728 GetConstantInt(ICI->getOperand(1), DL), Succ));
729 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
732 /// Given a vector of bb/value pairs, remove any entries
733 /// in the list that match the specified block.
735 EliminateBlockCases(BasicBlock *BB,
736 std::vector<ValueEqualityComparisonCase> &Cases) {
737 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
740 /// Return true if there are any keys in C1 that exist in C2 as well.
741 static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
742 std::vector<ValueEqualityComparisonCase> &C2) {
743 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
745 // Make V1 be smaller than V2.
746 if (V1->size() > V2->size())
751 if (V1->size() == 1) {
753 ConstantInt *TheVal = (*V1)[0].Value;
754 for (unsigned i = 0, e = V2->size(); i != e; ++i)
755 if (TheVal == (*V2)[i].Value)
759 // Otherwise, just sort both lists and compare element by element.
760 array_pod_sort(V1->begin(), V1->end());
761 array_pod_sort(V2->begin(), V2->end());
762 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
763 while (i1 != e1 && i2 != e2) {
764 if ((*V1)[i1].Value == (*V2)[i2].Value)
766 if ((*V1)[i1].Value < (*V2)[i2].Value)
774 /// If TI is known to be a terminator instruction and its block is known to
775 /// only have a single predecessor block, check to see if that predecessor is
776 /// also a value comparison with the same value, and if that comparison
777 /// determines the outcome of this comparison. If so, simplify TI. This does a
778 /// very limited form of jump threading.
779 bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
780 TerminatorInst *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
781 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
783 return false; // Not a value comparison in predecessor.
785 Value *ThisVal = isValueEqualityComparison(TI);
786 assert(ThisVal && "This isn't a value comparison!!");
787 if (ThisVal != PredVal)
788 return false; // Different predicates.
790 // TODO: Preserve branch weight metadata, similarly to how
791 // FoldValueComparisonIntoPredecessors preserves it.
793 // Find out information about when control will move from Pred to TI's block.
794 std::vector<ValueEqualityComparisonCase> PredCases;
795 BasicBlock *PredDef =
796 GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
797 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
799 // Find information about how control leaves this block.
800 std::vector<ValueEqualityComparisonCase> ThisCases;
801 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
802 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
804 // If TI's block is the default block from Pred's comparison, potentially
805 // simplify TI based on this knowledge.
806 if (PredDef == TI->getParent()) {
807 // If we are here, we know that the value is none of those cases listed in
808 // PredCases. If there are any cases in ThisCases that are in PredCases, we
810 if (!ValuesOverlap(PredCases, ThisCases))
813 if (isa<BranchInst>(TI)) {
814 // Okay, one of the successors of this condbr is dead. Convert it to a
816 assert(ThisCases.size() == 1 && "Branch can only have one case!");
817 // Insert the new branch.
818 Instruction *NI = Builder.CreateBr(ThisDef);
821 // Remove PHI node entries for the dead edge.
822 ThisCases[0].Dest->removePredecessor(TI->getParent());
824 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
825 << "Through successor TI: " << *TI << "Leaving: " << *NI
828 EraseTerminatorInstAndDCECond(TI);
832 SwitchInst *SI = cast<SwitchInst>(TI);
833 // Okay, TI has cases that are statically dead, prune them away.
834 SmallPtrSet<Constant *, 16> DeadCases;
835 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
836 DeadCases.insert(PredCases[i].Value);
838 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
839 << "Through successor TI: " << *TI);
841 // Collect branch weights into a vector.
842 SmallVector<uint32_t, 8> Weights;
843 MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
844 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
846 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
848 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
849 Weights.push_back(CI->getValue().getZExtValue());
851 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
853 if (DeadCases.count(i->getCaseValue())) {
855 std::swap(Weights[i->getCaseIndex() + 1], Weights.back());
858 i->getCaseSuccessor()->removePredecessor(TI->getParent());
862 if (HasWeight && Weights.size() >= 2)
863 SI->setMetadata(LLVMContext::MD_prof,
864 MDBuilder(SI->getParent()->getContext())
865 .createBranchWeights(Weights));
867 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
871 // Otherwise, TI's block must correspond to some matched value. Find out
872 // which value (or set of values) this is.
873 ConstantInt *TIV = nullptr;
874 BasicBlock *TIBB = TI->getParent();
875 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
876 if (PredCases[i].Dest == TIBB) {
878 return false; // Cannot handle multiple values coming to this block.
879 TIV = PredCases[i].Value;
881 assert(TIV && "No edge from pred to succ?");
883 // Okay, we found the one constant that our value can be if we get into TI's
884 // BB. Find out which successor will unconditionally be branched to.
885 BasicBlock *TheRealDest = nullptr;
886 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
887 if (ThisCases[i].Value == TIV) {
888 TheRealDest = ThisCases[i].Dest;
892 // If not handled by any explicit cases, it is handled by the default case.
894 TheRealDest = ThisDef;
896 // Remove PHI node entries for dead edges.
897 BasicBlock *CheckEdge = TheRealDest;
898 for (BasicBlock *Succ : successors(TIBB))
899 if (Succ != CheckEdge)
900 Succ->removePredecessor(TIBB);
904 // Insert the new branch.
905 Instruction *NI = Builder.CreateBr(TheRealDest);
908 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
909 << "Through successor TI: " << *TI << "Leaving: " << *NI
912 EraseTerminatorInstAndDCECond(TI);
918 /// This class implements a stable ordering of constant
919 /// integers that does not depend on their address. This is important for
920 /// applications that sort ConstantInt's to ensure uniqueness.
921 struct ConstantIntOrdering {
922 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
923 return LHS->getValue().ult(RHS->getValue());
927 } // end anonymous namespace
929 static int ConstantIntSortPredicate(ConstantInt *const *P1,
930 ConstantInt *const *P2) {
931 const ConstantInt *LHS = *P1;
932 const ConstantInt *RHS = *P2;
935 return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
938 static inline bool HasBranchWeights(const Instruction *I) {
939 MDNode *ProfMD = I->getMetadata(LLVMContext::MD_prof);
940 if (ProfMD && ProfMD->getOperand(0))
941 if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
942 return MDS->getString().equals("branch_weights");
947 /// Get Weights of a given TerminatorInst, the default weight is at the front
948 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
950 static void GetBranchWeights(TerminatorInst *TI,
951 SmallVectorImpl<uint64_t> &Weights) {
952 MDNode *MD = TI->getMetadata(LLVMContext::MD_prof);
954 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
955 ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
956 Weights.push_back(CI->getValue().getZExtValue());
959 // If TI is a conditional eq, the default case is the false case,
960 // and the corresponding branch-weight data is at index 2. We swap the
961 // default weight to be the first entry.
962 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
963 assert(Weights.size() == 2);
964 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
965 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
966 std::swap(Weights.front(), Weights.back());
970 /// Keep halving the weights until all can fit in uint32_t.
971 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
972 uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
973 if (Max > UINT_MAX) {
974 unsigned Offset = 32 - countLeadingZeros(Max);
975 for (uint64_t &I : Weights)
980 /// The specified terminator is a value equality comparison instruction
981 /// (either a switch or a branch on "X == c").
982 /// See if any of the predecessors of the terminator block are value comparisons
983 /// on the same value. If so, and if safe to do so, fold them together.
984 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
985 IRBuilder<> &Builder) {
986 BasicBlock *BB = TI->getParent();
987 Value *CV = isValueEqualityComparison(TI); // CondVal
988 assert(CV && "Not a comparison?");
989 bool Changed = false;
991 SmallVector<BasicBlock *, 16> Preds(pred_begin(BB), pred_end(BB));
992 while (!Preds.empty()) {
993 BasicBlock *Pred = Preds.pop_back_val();
995 // See if the predecessor is a comparison with the same value.
996 TerminatorInst *PTI = Pred->getTerminator();
997 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
999 if (PCV == CV && TI != PTI) {
1000 SmallSetVector<BasicBlock*, 4> FailBlocks;
1001 if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
1002 for (auto *Succ : FailBlocks) {
1003 if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split"))
1008 // Figure out which 'cases' to copy from SI to PSI.
1009 std::vector<ValueEqualityComparisonCase> BBCases;
1010 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1012 std::vector<ValueEqualityComparisonCase> PredCases;
1013 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1015 // Based on whether the default edge from PTI goes to BB or not, fill in
1016 // PredCases and PredDefault with the new switch cases we would like to
1018 SmallVector<BasicBlock *, 8> NewSuccessors;
1020 // Update the branch weight metadata along the way
1021 SmallVector<uint64_t, 8> Weights;
1022 bool PredHasWeights = HasBranchWeights(PTI);
1023 bool SuccHasWeights = HasBranchWeights(TI);
1025 if (PredHasWeights) {
1026 GetBranchWeights(PTI, Weights);
1027 // branch-weight metadata is inconsistent here.
1028 if (Weights.size() != 1 + PredCases.size())
1029 PredHasWeights = SuccHasWeights = false;
1030 } else if (SuccHasWeights)
1031 // If there are no predecessor weights but there are successor weights,
1032 // populate Weights with 1, which will later be scaled to the sum of
1033 // successor's weights
1034 Weights.assign(1 + PredCases.size(), 1);
1036 SmallVector<uint64_t, 8> SuccWeights;
1037 if (SuccHasWeights) {
1038 GetBranchWeights(TI, SuccWeights);
1039 // branch-weight metadata is inconsistent here.
1040 if (SuccWeights.size() != 1 + BBCases.size())
1041 PredHasWeights = SuccHasWeights = false;
1042 } else if (PredHasWeights)
1043 SuccWeights.assign(1 + BBCases.size(), 1);
1045 if (PredDefault == BB) {
1046 // If this is the default destination from PTI, only the edges in TI
1047 // that don't occur in PTI, or that branch to BB will be activated.
1048 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1049 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1050 if (PredCases[i].Dest != BB)
1051 PTIHandled.insert(PredCases[i].Value);
1053 // The default destination is BB, we don't need explicit targets.
1054 std::swap(PredCases[i], PredCases.back());
1056 if (PredHasWeights || SuccHasWeights) {
1057 // Increase weight for the default case.
1058 Weights[0] += Weights[i + 1];
1059 std::swap(Weights[i + 1], Weights.back());
1063 PredCases.pop_back();
1068 // Reconstruct the new switch statement we will be building.
1069 if (PredDefault != BBDefault) {
1070 PredDefault->removePredecessor(Pred);
1071 PredDefault = BBDefault;
1072 NewSuccessors.push_back(BBDefault);
1075 unsigned CasesFromPred = Weights.size();
1076 uint64_t ValidTotalSuccWeight = 0;
1077 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1078 if (!PTIHandled.count(BBCases[i].Value) &&
1079 BBCases[i].Dest != BBDefault) {
1080 PredCases.push_back(BBCases[i]);
1081 NewSuccessors.push_back(BBCases[i].Dest);
1082 if (SuccHasWeights || PredHasWeights) {
1083 // The default weight is at index 0, so weight for the ith case
1084 // should be at index i+1. Scale the cases from successor by
1085 // PredDefaultWeight (Weights[0]).
1086 Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1087 ValidTotalSuccWeight += SuccWeights[i + 1];
1091 if (SuccHasWeights || PredHasWeights) {
1092 ValidTotalSuccWeight += SuccWeights[0];
1093 // Scale the cases from predecessor by ValidTotalSuccWeight.
1094 for (unsigned i = 1; i < CasesFromPred; ++i)
1095 Weights[i] *= ValidTotalSuccWeight;
1096 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1097 Weights[0] *= SuccWeights[0];
1100 // If this is not the default destination from PSI, only the edges
1101 // in SI that occur in PSI with a destination of BB will be
1103 std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1104 std::map<ConstantInt *, uint64_t> WeightsForHandled;
1105 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1106 if (PredCases[i].Dest == BB) {
1107 PTIHandled.insert(PredCases[i].Value);
1109 if (PredHasWeights || SuccHasWeights) {
1110 WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1111 std::swap(Weights[i + 1], Weights.back());
1115 std::swap(PredCases[i], PredCases.back());
1116 PredCases.pop_back();
1121 // Okay, now we know which constants were sent to BB from the
1122 // predecessor. Figure out where they will all go now.
1123 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1124 if (PTIHandled.count(BBCases[i].Value)) {
1125 // If this is one we are capable of getting...
1126 if (PredHasWeights || SuccHasWeights)
1127 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1128 PredCases.push_back(BBCases[i]);
1129 NewSuccessors.push_back(BBCases[i].Dest);
1131 BBCases[i].Value); // This constant is taken care of
1134 // If there are any constants vectored to BB that TI doesn't handle,
1135 // they must go to the default destination of TI.
1136 for (ConstantInt *I : PTIHandled) {
1137 if (PredHasWeights || SuccHasWeights)
1138 Weights.push_back(WeightsForHandled[I]);
1139 PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1140 NewSuccessors.push_back(BBDefault);
1144 // Okay, at this point, we know which new successor Pred will get. Make
1145 // sure we update the number of entries in the PHI nodes for these
1147 for (BasicBlock *NewSuccessor : NewSuccessors)
1148 AddPredecessorToBlock(NewSuccessor, Pred, BB);
1150 Builder.SetInsertPoint(PTI);
1151 // Convert pointer to int before we switch.
1152 if (CV->getType()->isPointerTy()) {
1153 CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
1157 // Now that the successors are updated, create the new Switch instruction.
1159 Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1160 NewSI->setDebugLoc(PTI->getDebugLoc());
1161 for (ValueEqualityComparisonCase &V : PredCases)
1162 NewSI->addCase(V.Value, V.Dest);
1164 if (PredHasWeights || SuccHasWeights) {
1165 // Halve the weights if any of them cannot fit in an uint32_t
1166 FitWeights(Weights);
1168 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1171 LLVMContext::MD_prof,
1172 MDBuilder(BB->getContext()).createBranchWeights(MDWeights));
1175 EraseTerminatorInstAndDCECond(PTI);
1177 // Okay, last check. If BB is still a successor of PSI, then we must
1178 // have an infinite loop case. If so, add an infinitely looping block
1179 // to handle the case to preserve the behavior of the code.
1180 BasicBlock *InfLoopBlock = nullptr;
1181 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1182 if (NewSI->getSuccessor(i) == BB) {
1183 if (!InfLoopBlock) {
1184 // Insert it at the end of the function, because it's either code,
1185 // or it won't matter if it's hot. :)
1186 InfLoopBlock = BasicBlock::Create(BB->getContext(), "infloop",
1188 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1190 NewSI->setSuccessor(i, InfLoopBlock);
1199 // If we would need to insert a select that uses the value of this invoke
1200 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1201 // can't hoist the invoke, as there is nowhere to put the select in this case.
1202 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1203 Instruction *I1, Instruction *I2) {
1204 for (BasicBlock *Succ : successors(BB1)) {
1206 for (BasicBlock::iterator BBI = Succ->begin();
1207 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1208 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1209 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1210 if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1218 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I);
1220 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1221 /// in the two blocks up into the branch block. The caller of this function
1222 /// guarantees that BI's block dominates BB1 and BB2.
1223 static bool HoistThenElseCodeToIf(BranchInst *BI,
1224 const TargetTransformInfo &TTI) {
1225 // This does very trivial matching, with limited scanning, to find identical
1226 // instructions in the two blocks. In particular, we don't want to get into
1227 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1228 // such, we currently just scan for obviously identical instructions in an
1230 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1231 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1233 BasicBlock::iterator BB1_Itr = BB1->begin();
1234 BasicBlock::iterator BB2_Itr = BB2->begin();
1236 Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1237 // Skip debug info if it is not identical.
1238 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1239 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1240 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1241 while (isa<DbgInfoIntrinsic>(I1))
1243 while (isa<DbgInfoIntrinsic>(I2))
1246 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1247 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1250 BasicBlock *BIParent = BI->getParent();
1252 bool Changed = false;
1254 // If we are hoisting the terminator instruction, don't move one (making a
1255 // broken BB), instead clone it, and remove BI.
1256 if (isa<TerminatorInst>(I1))
1257 goto HoistTerminator;
1259 if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1262 // For a normal instruction, we just move one to right before the branch,
1263 // then replace all uses of the other with the first. Finally, we remove
1264 // the now redundant second instruction.
1265 BIParent->getInstList().splice(BI->getIterator(), BB1->getInstList(), I1);
1266 if (!I2->use_empty())
1267 I2->replaceAllUsesWith(I1);
1269 unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1270 LLVMContext::MD_range,
1271 LLVMContext::MD_fpmath,
1272 LLVMContext::MD_invariant_load,
1273 LLVMContext::MD_nonnull,
1274 LLVMContext::MD_invariant_group,
1275 LLVMContext::MD_align,
1276 LLVMContext::MD_dereferenceable,
1277 LLVMContext::MD_dereferenceable_or_null,
1278 LLVMContext::MD_mem_parallel_loop_access};
1279 combineMetadata(I1, I2, KnownIDs);
1281 // I1 and I2 are being combined into a single instruction. Its debug
1282 // location is the merged locations of the original instructions.
1283 if (!isa<CallInst>(I1))
1285 DILocation::getMergedLocation(I1->getDebugLoc(), I2->getDebugLoc()));
1287 I2->eraseFromParent();
1292 // Skip debug info if it is not identical.
1293 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1294 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1295 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1296 while (isa<DbgInfoIntrinsic>(I1))
1298 while (isa<DbgInfoIntrinsic>(I2))
1301 } while (I1->isIdenticalToWhenDefined(I2));
1306 // It may not be possible to hoist an invoke.
1307 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1310 for (BasicBlock *Succ : successors(BB1)) {
1312 for (BasicBlock::iterator BBI = Succ->begin();
1313 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1314 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1315 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1319 // Check for passingValueIsAlwaysUndefined here because we would rather
1320 // eliminate undefined control flow then converting it to a select.
1321 if (passingValueIsAlwaysUndefined(BB1V, PN) ||
1322 passingValueIsAlwaysUndefined(BB2V, PN))
1325 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1327 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1332 // Okay, it is safe to hoist the terminator.
1333 Instruction *NT = I1->clone();
1334 BIParent->getInstList().insert(BI->getIterator(), NT);
1335 if (!NT->getType()->isVoidTy()) {
1336 I1->replaceAllUsesWith(NT);
1337 I2->replaceAllUsesWith(NT);
1341 IRBuilder<NoFolder> Builder(NT);
1342 // Hoisting one of the terminators from our successor is a great thing.
1343 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1344 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1345 // nodes, so we insert select instruction to compute the final result.
1346 std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1347 for (BasicBlock *Succ : successors(BB1)) {
1349 for (BasicBlock::iterator BBI = Succ->begin();
1350 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1351 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1352 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1356 // These values do not agree. Insert a select instruction before NT
1357 // that determines the right value.
1358 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1360 SI = cast<SelectInst>(
1361 Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1362 BB1V->getName() + "." + BB2V->getName(), BI));
1364 // Make the PHI node use the select for all incoming values for BB1/BB2
1365 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1366 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1367 PN->setIncomingValue(i, SI);
1371 // Update any PHI nodes in our new successors.
1372 for (BasicBlock *Succ : successors(BB1))
1373 AddPredecessorToBlock(Succ, BIParent, BB1);
1375 EraseTerminatorInstAndDCECond(BI);
1379 // All instructions in Insts belong to different blocks that all unconditionally
1380 // branch to a common successor. Analyze each instruction and return true if it
1381 // would be possible to sink them into their successor, creating one common
1382 // instruction instead. For every value that would be required to be provided by
1383 // PHI node (because an operand varies in each input block), add to PHIOperands.
1384 static bool canSinkInstructions(
1385 ArrayRef<Instruction *> Insts,
1386 DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1387 // Prune out obviously bad instructions to move. Any non-store instruction
1388 // must have exactly one use, and we check later that use is by a single,
1389 // common PHI instruction in the successor.
1390 for (auto *I : Insts) {
1391 // These instructions may change or break semantics if moved.
1392 if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1393 I->getType()->isTokenTy())
1396 // Conservatively return false if I is an inline-asm instruction. Sinking
1397 // and merging inline-asm instructions can potentially create arguments
1398 // that cannot satisfy the inline-asm constraints.
1399 if (const auto *C = dyn_cast<CallInst>(I))
1400 if (C->isInlineAsm())
1403 // Everything must have only one use too, apart from stores which
1405 if (!isa<StoreInst>(I) && !I->hasOneUse())
1409 const Instruction *I0 = Insts.front();
1410 for (auto *I : Insts)
1411 if (!I->isSameOperationAs(I0))
1414 // All instructions in Insts are known to be the same opcode. If they aren't
1415 // stores, check the only user of each is a PHI or in the same block as the
1416 // instruction, because if a user is in the same block as an instruction
1417 // we're contemplating sinking, it must already be determined to be sinkable.
1418 if (!isa<StoreInst>(I0)) {
1419 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1420 auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1421 if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1422 auto *U = cast<Instruction>(*I->user_begin());
1424 PNUse->getParent() == Succ &&
1425 PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1426 U->getParent() == I->getParent();
1431 // Because SROA can't handle speculating stores of selects, try not
1432 // to sink loads or stores of allocas when we'd have to create a PHI for
1433 // the address operand. Also, because it is likely that loads or stores
1434 // of allocas will disappear when Mem2Reg/SROA is run, don't sink them.
1435 // This can cause code churn which can have unintended consequences down
1436 // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1437 // FIXME: This is a workaround for a deficiency in SROA - see
1438 // https://llvm.org/bugs/show_bug.cgi?id=30188
1439 if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1440 return isa<AllocaInst>(I->getOperand(1));
1443 if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1444 return isa<AllocaInst>(I->getOperand(0));
1448 for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1449 if (I0->getOperand(OI)->getType()->isTokenTy())
1450 // Don't touch any operand of token type.
1453 auto SameAsI0 = [&I0, OI](const Instruction *I) {
1454 assert(I->getNumOperands() == I0->getNumOperands());
1455 return I->getOperand(OI) == I0->getOperand(OI);
1457 if (!all_of(Insts, SameAsI0)) {
1458 if (!canReplaceOperandWithVariable(I0, OI))
1459 // We can't create a PHI from this GEP.
1461 // Don't create indirect calls! The called value is the final operand.
1462 if ((isa<CallInst>(I0) || isa<InvokeInst>(I0)) && OI == OE - 1) {
1463 // FIXME: if the call was *already* indirect, we should do this.
1466 for (auto *I : Insts)
1467 PHIOperands[I].push_back(I->getOperand(OI));
1473 // Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1474 // instruction of every block in Blocks to their common successor, commoning
1475 // into one instruction.
1476 static bool sinkLastInstruction(ArrayRef<BasicBlock*> Blocks) {
1477 auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1479 // canSinkLastInstruction returning true guarantees that every block has at
1480 // least one non-terminator instruction.
1481 SmallVector<Instruction*,4> Insts;
1482 for (auto *BB : Blocks) {
1483 Instruction *I = BB->getTerminator();
1485 I = I->getPrevNode();
1486 } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1487 if (!isa<DbgInfoIntrinsic>(I))
1491 // The only checking we need to do now is that all users of all instructions
1492 // are the same PHI node. canSinkLastInstruction should have checked this but
1493 // it is slightly over-aggressive - it gets confused by commutative instructions
1494 // so double-check it here.
1495 Instruction *I0 = Insts.front();
1496 if (!isa<StoreInst>(I0)) {
1497 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1498 if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1499 auto *U = cast<Instruction>(*I->user_begin());
1505 // We don't need to do any more checking here; canSinkLastInstruction should
1506 // have done it all for us.
1507 SmallVector<Value*, 4> NewOperands;
1508 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1509 // This check is different to that in canSinkLastInstruction. There, we
1510 // cared about the global view once simplifycfg (and instcombine) have
1511 // completed - it takes into account PHIs that become trivially
1512 // simplifiable. However here we need a more local view; if an operand
1513 // differs we create a PHI and rely on instcombine to clean up the very
1514 // small mess we may make.
1515 bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1516 return I->getOperand(O) != I0->getOperand(O);
1519 NewOperands.push_back(I0->getOperand(O));
1523 // Create a new PHI in the successor block and populate it.
1524 auto *Op = I0->getOperand(O);
1525 assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1526 auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1527 Op->getName() + ".sink", &BBEnd->front());
1528 for (auto *I : Insts)
1529 PN->addIncoming(I->getOperand(O), I->getParent());
1530 NewOperands.push_back(PN);
1533 // Arbitrarily use I0 as the new "common" instruction; remap its operands
1534 // and move it to the start of the successor block.
1535 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1536 I0->getOperandUse(O).set(NewOperands[O]);
1537 I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1539 // The debug location for the "common" instruction is the merged locations of
1540 // all the commoned instructions. We start with the original location of the
1541 // "common" instruction and iteratively merge each location in the loop below.
1542 const DILocation *Loc = I0->getDebugLoc();
1544 // Update metadata and IR flags, and merge debug locations.
1545 for (auto *I : Insts)
1547 Loc = DILocation::getMergedLocation(Loc, I->getDebugLoc());
1548 combineMetadataForCSE(I0, I);
1551 if (!isa<CallInst>(I0))
1552 I0->setDebugLoc(Loc);
1554 if (!isa<StoreInst>(I0)) {
1555 // canSinkLastInstruction checked that all instructions were used by
1556 // one and only one PHI node. Find that now, RAUW it to our common
1557 // instruction and nuke it.
1558 assert(I0->hasOneUse());
1559 auto *PN = cast<PHINode>(*I0->user_begin());
1560 PN->replaceAllUsesWith(I0);
1561 PN->eraseFromParent();
1564 // Finally nuke all instructions apart from the common instruction.
1565 for (auto *I : Insts)
1567 I->eraseFromParent();
1574 // LockstepReverseIterator - Iterates through instructions
1575 // in a set of blocks in reverse order from the first non-terminator.
1576 // For example (assume all blocks have size n):
1577 // LockstepReverseIterator I([B1, B2, B3]);
1578 // *I-- = [B1[n], B2[n], B3[n]];
1579 // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1580 // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1582 class LockstepReverseIterator {
1583 ArrayRef<BasicBlock*> Blocks;
1584 SmallVector<Instruction*,4> Insts;
1587 LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) :
1595 for (auto *BB : Blocks) {
1596 Instruction *Inst = BB->getTerminator();
1597 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1598 Inst = Inst->getPrevNode();
1600 // Block wasn't big enough.
1604 Insts.push_back(Inst);
1608 bool isValid() const {
1612 void operator -- () {
1615 for (auto *&Inst : Insts) {
1616 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1617 Inst = Inst->getPrevNode();
1618 // Already at beginning of block.
1626 ArrayRef<Instruction*> operator * () const {
1631 } // end anonymous namespace
1633 /// Given an unconditional branch that goes to BBEnd,
1634 /// check whether BBEnd has only two predecessors and the other predecessor
1635 /// ends with an unconditional branch. If it is true, sink any common code
1636 /// in the two predecessors to BBEnd.
1637 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1638 assert(BI1->isUnconditional());
1639 BasicBlock *BBEnd = BI1->getSuccessor(0);
1641 // We support two situations:
1642 // (1) all incoming arcs are unconditional
1643 // (2) one incoming arc is conditional
1645 // (2) is very common in switch defaults and
1646 // else-if patterns;
1649 // else if (b) f(2);
1662 // [end] has two unconditional predecessor arcs and one conditional. The
1663 // conditional refers to the implicit empty 'else' arc. This conditional
1664 // arc can also be caused by an empty default block in a switch.
1666 // In this case, we attempt to sink code from all *unconditional* arcs.
1667 // If we can sink instructions from these arcs (determined during the scan
1668 // phase below) we insert a common successor for all unconditional arcs and
1669 // connect that to [end], to enable sinking:
1682 SmallVector<BasicBlock*,4> UnconditionalPreds;
1683 Instruction *Cond = nullptr;
1684 for (auto *B : predecessors(BBEnd)) {
1685 auto *T = B->getTerminator();
1686 if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1687 UnconditionalPreds.push_back(B);
1688 else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1693 if (UnconditionalPreds.size() < 2)
1696 bool Changed = false;
1697 // We take a two-step approach to tail sinking. First we scan from the end of
1698 // each block upwards in lockstep. If the n'th instruction from the end of each
1699 // block can be sunk, those instructions are added to ValuesToSink and we
1700 // carry on. If we can sink an instruction but need to PHI-merge some operands
1701 // (because they're not identical in each instruction) we add these to
1703 unsigned ScanIdx = 0;
1704 SmallPtrSet<Value*,4> InstructionsToSink;
1705 DenseMap<Instruction*, SmallVector<Value*,4>> PHIOperands;
1706 LockstepReverseIterator LRI(UnconditionalPreds);
1707 while (LRI.isValid() &&
1708 canSinkInstructions(*LRI, PHIOperands)) {
1709 DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0] << "\n");
1710 InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1715 auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1716 unsigned NumPHIdValues = 0;
1717 for (auto *I : *LRI)
1718 for (auto *V : PHIOperands[I])
1719 if (InstructionsToSink.count(V) == 0)
1721 DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1722 unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1723 if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1726 return NumPHIInsts <= 1;
1729 if (ScanIdx > 0 && Cond) {
1730 // Check if we would actually sink anything first! This mutates the CFG and
1731 // adds an extra block. The goal in doing this is to allow instructions that
1732 // couldn't be sunk before to be sunk - obviously, speculatable instructions
1733 // (such as trunc, add) can be sunk and predicated already. So we check that
1734 // we're going to sink at least one non-speculatable instruction.
1737 bool Profitable = false;
1738 while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1739 if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1749 DEBUG(dbgs() << "SINK: Splitting edge\n");
1750 // We have a conditional edge and we're going to sink some instructions.
1751 // Insert a new block postdominating all blocks we're going to sink from.
1752 if (!SplitBlockPredecessors(BI1->getSuccessor(0), UnconditionalPreds,
1754 // Edges couldn't be split.
1759 // Now that we've analyzed all potential sinking candidates, perform the
1760 // actual sink. We iteratively sink the last non-terminator of the source
1761 // blocks into their common successor unless doing so would require too
1762 // many PHI instructions to be generated (currently only one PHI is allowed
1763 // per sunk instruction).
1765 // We can use InstructionsToSink to discount values needing PHI-merging that will
1766 // actually be sunk in a later iteration. This allows us to be more
1767 // aggressive in what we sink. This does allow a false positive where we
1768 // sink presuming a later value will also be sunk, but stop half way through
1769 // and never actually sink it which means we produce more PHIs than intended.
1770 // This is unlikely in practice though.
1771 for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1772 DEBUG(dbgs() << "SINK: Sink: "
1773 << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1776 // Because we've sunk every instruction in turn, the current instruction to
1777 // sink is always at index 0.
1779 if (!ProfitableToSinkInstruction(LRI)) {
1780 // Too many PHIs would be created.
1781 DEBUG(dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1785 if (!sinkLastInstruction(UnconditionalPreds))
1793 /// \brief Determine if we can hoist sink a sole store instruction out of a
1794 /// conditional block.
1796 /// We are looking for code like the following:
1798 /// store i32 %add, i32* %arrayidx2
1799 /// ... // No other stores or function calls (we could be calling a memory
1800 /// ... // function).
1801 /// %cmp = icmp ult %x, %y
1802 /// br i1 %cmp, label %EndBB, label %ThenBB
1804 /// store i32 %add5, i32* %arrayidx2
1808 /// We are going to transform this into:
1810 /// store i32 %add, i32* %arrayidx2
1812 /// %cmp = icmp ult %x, %y
1813 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1814 /// store i32 %add.add5, i32* %arrayidx2
1817 /// \return The pointer to the value of the previous store if the store can be
1818 /// hoisted into the predecessor block. 0 otherwise.
1819 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1820 BasicBlock *StoreBB, BasicBlock *EndBB) {
1821 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1825 // Volatile or atomic.
1826 if (!StoreToHoist->isSimple())
1829 Value *StorePtr = StoreToHoist->getPointerOperand();
1831 // Look for a store to the same pointer in BrBB.
1832 unsigned MaxNumInstToLookAt = 9;
1833 for (Instruction &CurI : reverse(*BrBB)) {
1834 if (!MaxNumInstToLookAt)
1837 if (isa<DbgInfoIntrinsic>(CurI))
1839 --MaxNumInstToLookAt;
1841 // Could be calling an instruction that affects memory like free().
1842 if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1845 if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1846 // Found the previous store make sure it stores to the same location.
1847 if (SI->getPointerOperand() == StorePtr)
1848 // Found the previous store, return its value operand.
1849 return SI->getValueOperand();
1850 return nullptr; // Unknown store.
1857 /// \brief Speculate a conditional basic block flattening the CFG.
1859 /// Note that this is a very risky transform currently. Speculating
1860 /// instructions like this is most often not desirable. Instead, there is an MI
1861 /// pass which can do it with full awareness of the resource constraints.
1862 /// However, some cases are "obvious" and we should do directly. An example of
1863 /// this is speculating a single, reasonably cheap instruction.
1865 /// There is only one distinct advantage to flattening the CFG at the IR level:
1866 /// it makes very common but simplistic optimizations such as are common in
1867 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1868 /// modeling their effects with easier to reason about SSA value graphs.
1871 /// An illustration of this transform is turning this IR:
1874 /// %cmp = icmp ult %x, %y
1875 /// br i1 %cmp, label %EndBB, label %ThenBB
1877 /// %sub = sub %x, %y
1880 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1887 /// %cmp = icmp ult %x, %y
1888 /// %sub = sub %x, %y
1889 /// %cond = select i1 %cmp, 0, %sub
1893 /// \returns true if the conditional block is removed.
1894 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1895 const TargetTransformInfo &TTI) {
1896 // Be conservative for now. FP select instruction can often be expensive.
1897 Value *BrCond = BI->getCondition();
1898 if (isa<FCmpInst>(BrCond))
1901 BasicBlock *BB = BI->getParent();
1902 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1904 // If ThenBB is actually on the false edge of the conditional branch, remember
1905 // to swap the select operands later.
1906 bool Invert = false;
1907 if (ThenBB != BI->getSuccessor(0)) {
1908 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1911 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1913 // Keep a count of how many times instructions are used within CondBB when
1914 // they are candidates for sinking into CondBB. Specifically:
1915 // - They are defined in BB, and
1916 // - They have no side effects, and
1917 // - All of their uses are in CondBB.
1918 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1920 unsigned SpeculationCost = 0;
1921 Value *SpeculatedStoreValue = nullptr;
1922 StoreInst *SpeculatedStore = nullptr;
1923 for (BasicBlock::iterator BBI = ThenBB->begin(),
1924 BBE = std::prev(ThenBB->end());
1925 BBI != BBE; ++BBI) {
1926 Instruction *I = &*BBI;
1928 if (isa<DbgInfoIntrinsic>(I))
1931 // Only speculatively execute a single instruction (not counting the
1932 // terminator) for now.
1934 if (SpeculationCost > 1)
1937 // Don't hoist the instruction if it's unsafe or expensive.
1938 if (!isSafeToSpeculativelyExecute(I) &&
1939 !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1940 I, BB, ThenBB, EndBB))))
1942 if (!SpeculatedStoreValue &&
1943 ComputeSpeculationCost(I, TTI) >
1944 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
1947 // Store the store speculation candidate.
1948 if (SpeculatedStoreValue)
1949 SpeculatedStore = cast<StoreInst>(I);
1951 // Do not hoist the instruction if any of its operands are defined but not
1952 // used in BB. The transformation will prevent the operand from
1953 // being sunk into the use block.
1954 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
1955 Instruction *OpI = dyn_cast<Instruction>(*i);
1956 if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
1957 continue; // Not a candidate for sinking.
1959 ++SinkCandidateUseCounts[OpI];
1963 // Consider any sink candidates which are only used in CondBB as costs for
1964 // speculation. Note, while we iterate over a DenseMap here, we are summing
1965 // and so iteration order isn't significant.
1966 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator
1967 I = SinkCandidateUseCounts.begin(),
1968 E = SinkCandidateUseCounts.end();
1970 if (I->first->getNumUses() == I->second) {
1972 if (SpeculationCost > 1)
1976 // Check that the PHI nodes can be converted to selects.
1977 bool HaveRewritablePHIs = false;
1978 for (BasicBlock::iterator I = EndBB->begin();
1979 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1980 Value *OrigV = PN->getIncomingValueForBlock(BB);
1981 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1983 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1984 // Skip PHIs which are trivial.
1988 // Don't convert to selects if we could remove undefined behavior instead.
1989 if (passingValueIsAlwaysUndefined(OrigV, PN) ||
1990 passingValueIsAlwaysUndefined(ThenV, PN))
1993 HaveRewritablePHIs = true;
1994 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1995 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1996 if (!OrigCE && !ThenCE)
1997 continue; // Known safe and cheap.
1999 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2000 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2002 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2003 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2005 2 * PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2006 if (OrigCost + ThenCost > MaxCost)
2009 // Account for the cost of an unfolded ConstantExpr which could end up
2010 // getting expanded into Instructions.
2011 // FIXME: This doesn't account for how many operations are combined in the
2012 // constant expression.
2014 if (SpeculationCost > 1)
2018 // If there are no PHIs to process, bail early. This helps ensure idempotence
2020 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2023 // If we get here, we can hoist the instruction and if-convert.
2024 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2026 // Insert a select of the value of the speculated store.
2027 if (SpeculatedStoreValue) {
2028 IRBuilder<NoFolder> Builder(BI);
2029 Value *TrueV = SpeculatedStore->getValueOperand();
2030 Value *FalseV = SpeculatedStoreValue;
2032 std::swap(TrueV, FalseV);
2033 Value *S = Builder.CreateSelect(
2034 BrCond, TrueV, FalseV, TrueV->getName() + "." + FalseV->getName(), BI);
2035 SpeculatedStore->setOperand(0, S);
2036 SpeculatedStore->setDebugLoc(
2037 DILocation::getMergedLocation(
2038 BI->getDebugLoc(), SpeculatedStore->getDebugLoc()));
2041 // Metadata can be dependent on the condition we are hoisting above.
2042 // Conservatively strip all metadata on the instruction.
2043 for (auto &I : *ThenBB)
2044 I.dropUnknownNonDebugMetadata();
2046 // Hoist the instructions.
2047 BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2048 ThenBB->begin(), std::prev(ThenBB->end()));
2050 // Insert selects and rewrite the PHI operands.
2051 IRBuilder<NoFolder> Builder(BI);
2052 for (BasicBlock::iterator I = EndBB->begin();
2053 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2054 unsigned OrigI = PN->getBasicBlockIndex(BB);
2055 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
2056 Value *OrigV = PN->getIncomingValue(OrigI);
2057 Value *ThenV = PN->getIncomingValue(ThenI);
2059 // Skip PHIs which are trivial.
2063 // Create a select whose true value is the speculatively executed value and
2064 // false value is the preexisting value. Swap them if the branch
2065 // destinations were inverted.
2066 Value *TrueV = ThenV, *FalseV = OrigV;
2068 std::swap(TrueV, FalseV);
2069 Value *V = Builder.CreateSelect(
2070 BrCond, TrueV, FalseV, TrueV->getName() + "." + FalseV->getName(), BI);
2071 PN->setIncomingValue(OrigI, V);
2072 PN->setIncomingValue(ThenI, V);
2079 /// Return true if we can thread a branch across this block.
2080 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
2081 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
2084 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2085 if (isa<DbgInfoIntrinsic>(BBI))
2088 return false; // Don't clone large BB's.
2091 // We can only support instructions that do not define values that are
2092 // live outside of the current basic block.
2093 for (User *U : BBI->users()) {
2094 Instruction *UI = cast<Instruction>(U);
2095 if (UI->getParent() != BB || isa<PHINode>(UI))
2099 // Looks ok, continue checking.
2105 /// If we have a conditional branch on a PHI node value that is defined in the
2106 /// same block as the branch and if any PHI entries are constants, thread edges
2107 /// corresponding to that entry to be branches to their ultimate destination.
2108 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL,
2109 AssumptionCache *AC) {
2110 BasicBlock *BB = BI->getParent();
2111 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2112 // NOTE: we currently cannot transform this case if the PHI node is used
2113 // outside of the block.
2114 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2117 // Degenerate case of a single entry PHI.
2118 if (PN->getNumIncomingValues() == 1) {
2119 FoldSingleEntryPHINodes(PN->getParent());
2123 // Now we know that this block has multiple preds and two succs.
2124 if (!BlockIsSimpleEnoughToThreadThrough(BB))
2127 // Can't fold blocks that contain noduplicate or convergent calls.
2128 if (any_of(*BB, [](const Instruction &I) {
2129 const CallInst *CI = dyn_cast<CallInst>(&I);
2130 return CI && (CI->cannotDuplicate() || CI->isConvergent());
2134 // Okay, this is a simple enough basic block. See if any phi values are
2136 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2137 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
2138 if (!CB || !CB->getType()->isIntegerTy(1))
2141 // Okay, we now know that all edges from PredBB should be revectored to
2142 // branch to RealDest.
2143 BasicBlock *PredBB = PN->getIncomingBlock(i);
2144 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2147 continue; // Skip self loops.
2148 // Skip if the predecessor's terminator is an indirect branch.
2149 if (isa<IndirectBrInst>(PredBB->getTerminator()))
2152 // The dest block might have PHI nodes, other predecessors and other
2153 // difficult cases. Instead of being smart about this, just insert a new
2154 // block that jumps to the destination block, effectively splitting
2155 // the edge we are about to create.
2156 BasicBlock *EdgeBB =
2157 BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2158 RealDest->getParent(), RealDest);
2159 BranchInst::Create(RealDest, EdgeBB);
2161 // Update PHI nodes.
2162 AddPredecessorToBlock(RealDest, EdgeBB, BB);
2164 // BB may have instructions that are being threaded over. Clone these
2165 // instructions into EdgeBB. We know that there will be no uses of the
2166 // cloned instructions outside of EdgeBB.
2167 BasicBlock::iterator InsertPt = EdgeBB->begin();
2168 DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2169 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2170 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2171 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2174 // Clone the instruction.
2175 Instruction *N = BBI->clone();
2177 N->setName(BBI->getName() + ".c");
2179 // Update operands due to translation.
2180 for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2181 DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2182 if (PI != TranslateMap.end())
2186 // Check for trivial simplification.
2187 if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2188 if (!BBI->use_empty())
2189 TranslateMap[&*BBI] = V;
2190 if (!N->mayHaveSideEffects()) {
2191 N->deleteValue(); // Instruction folded away, don't need actual inst
2195 if (!BBI->use_empty())
2196 TranslateMap[&*BBI] = N;
2198 // Insert the new instruction into its new home.
2200 EdgeBB->getInstList().insert(InsertPt, N);
2202 // Register the new instruction with the assumption cache if necessary.
2203 if (auto *II = dyn_cast_or_null<IntrinsicInst>(N))
2204 if (II->getIntrinsicID() == Intrinsic::assume)
2205 AC->registerAssumption(II);
2208 // Loop over all of the edges from PredBB to BB, changing them to branch
2209 // to EdgeBB instead.
2210 TerminatorInst *PredBBTI = PredBB->getTerminator();
2211 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2212 if (PredBBTI->getSuccessor(i) == BB) {
2213 BB->removePredecessor(PredBB);
2214 PredBBTI->setSuccessor(i, EdgeBB);
2217 // Recurse, simplifying any other constants.
2218 return FoldCondBranchOnPHI(BI, DL, AC) | true;
2224 /// Given a BB that starts with the specified two-entry PHI node,
2225 /// see if we can eliminate it.
2226 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI,
2227 const DataLayout &DL) {
2228 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2229 // statement", which has a very simple dominance structure. Basically, we
2230 // are trying to find the condition that is being branched on, which
2231 // subsequently causes this merge to happen. We really want control
2232 // dependence information for this check, but simplifycfg can't keep it up
2233 // to date, and this catches most of the cases we care about anyway.
2234 BasicBlock *BB = PN->getParent();
2235 BasicBlock *IfTrue, *IfFalse;
2236 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2238 // Don't bother if the branch will be constant folded trivially.
2239 isa<ConstantInt>(IfCond))
2242 // Okay, we found that we can merge this two-entry phi node into a select.
2243 // Doing so would require us to fold *all* two entry phi nodes in this block.
2244 // At some point this becomes non-profitable (particularly if the target
2245 // doesn't support cmov's). Only do this transformation if there are two or
2246 // fewer PHI nodes in this block.
2247 unsigned NumPhis = 0;
2248 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2252 // Loop over the PHI's seeing if we can promote them all to select
2253 // instructions. While we are at it, keep track of the instructions
2254 // that need to be moved to the dominating block.
2255 SmallPtrSet<Instruction *, 4> AggressiveInsts;
2256 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2257 MaxCostVal1 = PHINodeFoldingThreshold;
2258 MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2259 MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2261 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2262 PHINode *PN = cast<PHINode>(II++);
2263 if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2264 PN->replaceAllUsesWith(V);
2265 PN->eraseFromParent();
2269 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
2270 MaxCostVal0, TTI) ||
2271 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
2276 // If we folded the first phi, PN dangles at this point. Refresh it. If
2277 // we ran out of PHIs then we simplified them all.
2278 PN = dyn_cast<PHINode>(BB->begin());
2282 // Don't fold i1 branches on PHIs which contain binary operators. These can
2283 // often be turned into switches and other things.
2284 if (PN->getType()->isIntegerTy(1) &&
2285 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2286 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2287 isa<BinaryOperator>(IfCond)))
2290 // If all PHI nodes are promotable, check to make sure that all instructions
2291 // in the predecessor blocks can be promoted as well. If not, we won't be able
2292 // to get rid of the control flow, so it's not worth promoting to select
2294 BasicBlock *DomBlock = nullptr;
2295 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2296 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2297 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2300 DomBlock = *pred_begin(IfBlock1);
2301 for (BasicBlock::iterator I = IfBlock1->begin(); !isa<TerminatorInst>(I);
2303 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2304 // This is not an aggressive instruction that we can promote.
2305 // Because of this, we won't be able to get rid of the control flow, so
2306 // the xform is not worth it.
2311 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2314 DomBlock = *pred_begin(IfBlock2);
2315 for (BasicBlock::iterator I = IfBlock2->begin(); !isa<TerminatorInst>(I);
2317 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2318 // This is not an aggressive instruction that we can promote.
2319 // Because of this, we won't be able to get rid of the control flow, so
2320 // the xform is not worth it.
2325 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
2326 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
2328 // If we can still promote the PHI nodes after this gauntlet of tests,
2329 // do all of the PHI's now.
2330 Instruction *InsertPt = DomBlock->getTerminator();
2331 IRBuilder<NoFolder> Builder(InsertPt);
2333 // Move all 'aggressive' instructions, which are defined in the
2334 // conditional parts of the if's up to the dominating block.
2336 for (auto &I : *IfBlock1)
2337 I.dropUnknownNonDebugMetadata();
2338 DomBlock->getInstList().splice(InsertPt->getIterator(),
2339 IfBlock1->getInstList(), IfBlock1->begin(),
2340 IfBlock1->getTerminator()->getIterator());
2343 for (auto &I : *IfBlock2)
2344 I.dropUnknownNonDebugMetadata();
2345 DomBlock->getInstList().splice(InsertPt->getIterator(),
2346 IfBlock2->getInstList(), IfBlock2->begin(),
2347 IfBlock2->getTerminator()->getIterator());
2350 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2351 // Change the PHI node into a select instruction.
2352 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2353 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2355 Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2356 PN->replaceAllUsesWith(Sel);
2358 PN->eraseFromParent();
2361 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2362 // has been flattened. Change DomBlock to jump directly to our new block to
2363 // avoid other simplifycfg's kicking in on the diamond.
2364 TerminatorInst *OldTI = DomBlock->getTerminator();
2365 Builder.SetInsertPoint(OldTI);
2366 Builder.CreateBr(BB);
2367 OldTI->eraseFromParent();
2371 /// If we found a conditional branch that goes to two returning blocks,
2372 /// try to merge them together into one return,
2373 /// introducing a select if the return values disagree.
2374 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
2375 IRBuilder<> &Builder) {
2376 assert(BI->isConditional() && "Must be a conditional branch");
2377 BasicBlock *TrueSucc = BI->getSuccessor(0);
2378 BasicBlock *FalseSucc = BI->getSuccessor(1);
2379 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2380 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2382 // Check to ensure both blocks are empty (just a return) or optionally empty
2383 // with PHI nodes. If there are other instructions, merging would cause extra
2384 // computation on one path or the other.
2385 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2387 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2390 Builder.SetInsertPoint(BI);
2391 // Okay, we found a branch that is going to two return nodes. If
2392 // there is no return value for this function, just change the
2393 // branch into a return.
2394 if (FalseRet->getNumOperands() == 0) {
2395 TrueSucc->removePredecessor(BI->getParent());
2396 FalseSucc->removePredecessor(BI->getParent());
2397 Builder.CreateRetVoid();
2398 EraseTerminatorInstAndDCECond(BI);
2402 // Otherwise, figure out what the true and false return values are
2403 // so we can insert a new select instruction.
2404 Value *TrueValue = TrueRet->getReturnValue();
2405 Value *FalseValue = FalseRet->getReturnValue();
2407 // Unwrap any PHI nodes in the return blocks.
2408 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2409 if (TVPN->getParent() == TrueSucc)
2410 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2411 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2412 if (FVPN->getParent() == FalseSucc)
2413 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2415 // In order for this transformation to be safe, we must be able to
2416 // unconditionally execute both operands to the return. This is
2417 // normally the case, but we could have a potentially-trapping
2418 // constant expression that prevents this transformation from being
2420 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2423 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2427 // Okay, we collected all the mapped values and checked them for sanity, and
2428 // defined to really do this transformation. First, update the CFG.
2429 TrueSucc->removePredecessor(BI->getParent());
2430 FalseSucc->removePredecessor(BI->getParent());
2432 // Insert select instructions where needed.
2433 Value *BrCond = BI->getCondition();
2435 // Insert a select if the results differ.
2436 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2437 } else if (isa<UndefValue>(TrueValue)) {
2438 TrueValue = FalseValue;
2441 Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2446 !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2450 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2451 << "\n " << *BI << "NewRet = " << *RI
2452 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: " << *FalseSucc);
2454 EraseTerminatorInstAndDCECond(BI);
2459 /// Return true if the given instruction is available
2460 /// in its predecessor block. If yes, the instruction will be removed.
2461 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
2462 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2464 for (Instruction &I : *PB) {
2465 Instruction *PBI = &I;
2466 // Check whether Inst and PBI generate the same value.
2467 if (Inst->isIdenticalTo(PBI)) {
2468 Inst->replaceAllUsesWith(PBI);
2469 Inst->eraseFromParent();
2476 /// Return true if either PBI or BI has branch weight available, and store
2477 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2478 /// not have branch weight, use 1:1 as its weight.
2479 static bool extractPredSuccWeights(BranchInst *PBI, BranchInst *BI,
2480 uint64_t &PredTrueWeight,
2481 uint64_t &PredFalseWeight,
2482 uint64_t &SuccTrueWeight,
2483 uint64_t &SuccFalseWeight) {
2484 bool PredHasWeights =
2485 PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2486 bool SuccHasWeights =
2487 BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2488 if (PredHasWeights || SuccHasWeights) {
2489 if (!PredHasWeights)
2490 PredTrueWeight = PredFalseWeight = 1;
2491 if (!SuccHasWeights)
2492 SuccTrueWeight = SuccFalseWeight = 1;
2499 /// If this basic block is simple enough, and if a predecessor branches to us
2500 /// and one of our successors, fold the block into the predecessor and use
2501 /// logical operations to pick the right destination.
2502 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2503 BasicBlock *BB = BI->getParent();
2505 Instruction *Cond = nullptr;
2506 if (BI->isConditional())
2507 Cond = dyn_cast<Instruction>(BI->getCondition());
2509 // For unconditional branch, check for a simple CFG pattern, where
2510 // BB has a single predecessor and BB's successor is also its predecessor's
2511 // successor. If such pattern exisits, check for CSE between BB and its
2513 if (BasicBlock *PB = BB->getSinglePredecessor())
2514 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2515 if (PBI->isConditional() &&
2516 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2517 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2518 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2519 Instruction *Curr = &*I++;
2520 if (isa<CmpInst>(Curr)) {
2524 // Quit if we can't remove this instruction.
2525 if (!checkCSEInPredecessor(Curr, PB))
2534 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2535 Cond->getParent() != BB || !Cond->hasOneUse())
2538 // Make sure the instruction after the condition is the cond branch.
2539 BasicBlock::iterator CondIt = ++Cond->getIterator();
2541 // Ignore dbg intrinsics.
2542 while (isa<DbgInfoIntrinsic>(CondIt))
2548 // Only allow this transformation if computing the condition doesn't involve
2549 // too many instructions and these involved instructions can be executed
2550 // unconditionally. We denote all involved instructions except the condition
2551 // as "bonus instructions", and only allow this transformation when the
2552 // number of the bonus instructions does not exceed a certain threshold.
2553 unsigned NumBonusInsts = 0;
2554 for (auto I = BB->begin(); Cond != &*I; ++I) {
2555 // Ignore dbg intrinsics.
2556 if (isa<DbgInfoIntrinsic>(I))
2558 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2560 // I has only one use and can be executed unconditionally.
2561 Instruction *User = dyn_cast<Instruction>(I->user_back());
2562 if (User == nullptr || User->getParent() != BB)
2564 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2565 // to use any other instruction, User must be an instruction between next(I)
2568 // Early exits once we reach the limit.
2569 if (NumBonusInsts > BonusInstThreshold)
2573 // Cond is known to be a compare or binary operator. Check to make sure that
2574 // neither operand is a potentially-trapping constant expression.
2575 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2578 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2582 // Finally, don't infinitely unroll conditional loops.
2583 BasicBlock *TrueDest = BI->getSuccessor(0);
2584 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2585 if (TrueDest == BB || FalseDest == BB)
2588 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2589 BasicBlock *PredBlock = *PI;
2590 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2592 // Check that we have two conditional branches. If there is a PHI node in
2593 // the common successor, verify that the same value flows in from both
2595 SmallVector<PHINode *, 4> PHIs;
2596 if (!PBI || PBI->isUnconditional() ||
2597 (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2598 (!BI->isConditional() &&
2599 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2602 // Determine if the two branches share a common destination.
2603 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2604 bool InvertPredCond = false;
2606 if (BI->isConditional()) {
2607 if (PBI->getSuccessor(0) == TrueDest) {
2608 Opc = Instruction::Or;
2609 } else if (PBI->getSuccessor(1) == FalseDest) {
2610 Opc = Instruction::And;
2611 } else if (PBI->getSuccessor(0) == FalseDest) {
2612 Opc = Instruction::And;
2613 InvertPredCond = true;
2614 } else if (PBI->getSuccessor(1) == TrueDest) {
2615 Opc = Instruction::Or;
2616 InvertPredCond = true;
2621 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2625 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2626 IRBuilder<> Builder(PBI);
2628 // If we need to invert the condition in the pred block to match, do so now.
2629 if (InvertPredCond) {
2630 Value *NewCond = PBI->getCondition();
2632 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2633 CmpInst *CI = cast<CmpInst>(NewCond);
2634 CI->setPredicate(CI->getInversePredicate());
2637 Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2640 PBI->setCondition(NewCond);
2641 PBI->swapSuccessors();
2644 // If we have bonus instructions, clone them into the predecessor block.
2645 // Note that there may be multiple predecessor blocks, so we cannot move
2646 // bonus instructions to a predecessor block.
2647 ValueToValueMapTy VMap; // maps original values to cloned values
2648 // We already make sure Cond is the last instruction before BI. Therefore,
2649 // all instructions before Cond other than DbgInfoIntrinsic are bonus
2651 for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2652 if (isa<DbgInfoIntrinsic>(BonusInst))
2654 Instruction *NewBonusInst = BonusInst->clone();
2655 RemapInstruction(NewBonusInst, VMap,
2656 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2657 VMap[&*BonusInst] = NewBonusInst;
2659 // If we moved a load, we cannot any longer claim any knowledge about
2660 // its potential value. The previous information might have been valid
2661 // only given the branch precondition.
2662 // For an analogous reason, we must also drop all the metadata whose
2663 // semantics we don't understand.
2664 NewBonusInst->dropUnknownNonDebugMetadata();
2666 PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2667 NewBonusInst->takeName(&*BonusInst);
2668 BonusInst->setName(BonusInst->getName() + ".old");
2671 // Clone Cond into the predecessor basic block, and or/and the
2672 // two conditions together.
2673 Instruction *New = Cond->clone();
2674 RemapInstruction(New, VMap,
2675 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2676 PredBlock->getInstList().insert(PBI->getIterator(), New);
2677 New->takeName(Cond);
2678 Cond->setName(New->getName() + ".old");
2680 if (BI->isConditional()) {
2681 Instruction *NewCond = cast<Instruction>(
2682 Builder.CreateBinOp(Opc, PBI->getCondition(), New, "or.cond"));
2683 PBI->setCondition(NewCond);
2685 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2687 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2688 SuccTrueWeight, SuccFalseWeight);
2689 SmallVector<uint64_t, 8> NewWeights;
2691 if (PBI->getSuccessor(0) == BB) {
2693 // PBI: br i1 %x, BB, FalseDest
2694 // BI: br i1 %y, TrueDest, FalseDest
2695 // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2696 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2697 // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2698 // TrueWeight for PBI * FalseWeight for BI.
2699 // We assume that total weights of a BranchInst can fit into 32 bits.
2700 // Therefore, we will not have overflow using 64-bit arithmetic.
2701 NewWeights.push_back(PredFalseWeight *
2702 (SuccFalseWeight + SuccTrueWeight) +
2703 PredTrueWeight * SuccFalseWeight);
2705 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2706 PBI->setSuccessor(0, TrueDest);
2708 if (PBI->getSuccessor(1) == BB) {
2710 // PBI: br i1 %x, TrueDest, BB
2711 // BI: br i1 %y, TrueDest, FalseDest
2712 // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2713 // FalseWeight for PBI * TrueWeight for BI.
2714 NewWeights.push_back(PredTrueWeight *
2715 (SuccFalseWeight + SuccTrueWeight) +
2716 PredFalseWeight * SuccTrueWeight);
2717 // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2718 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2720 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2721 PBI->setSuccessor(1, FalseDest);
2723 if (NewWeights.size() == 2) {
2724 // Halve the weights if any of them cannot fit in an uint32_t
2725 FitWeights(NewWeights);
2727 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2730 LLVMContext::MD_prof,
2731 MDBuilder(BI->getContext()).createBranchWeights(MDWeights));
2733 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2735 // Update PHI nodes in the common successors.
2736 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2737 ConstantInt *PBI_C = cast<ConstantInt>(
2738 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2739 assert(PBI_C->getType()->isIntegerTy(1));
2740 Instruction *MergedCond = nullptr;
2741 if (PBI->getSuccessor(0) == TrueDest) {
2742 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2743 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2744 // is false: !PBI_Cond and BI_Value
2745 Instruction *NotCond = cast<Instruction>(
2746 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2747 MergedCond = cast<Instruction>(
2748 Builder.CreateBinOp(Instruction::And, NotCond, New, "and.cond"));
2750 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2751 Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2753 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2754 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2755 // is false: PBI_Cond and BI_Value
2756 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2757 Instruction::And, PBI->getCondition(), New, "and.cond"));
2758 if (PBI_C->isOne()) {
2759 Instruction *NotCond = cast<Instruction>(
2760 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2761 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2762 Instruction::Or, NotCond, MergedCond, "or.cond"));
2766 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2769 // Change PBI from Conditional to Unconditional.
2770 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2771 EraseTerminatorInstAndDCECond(PBI);
2775 // If BI was a loop latch, it may have had associated loop metadata.
2776 // We need to copy it to the new latch, that is, PBI.
2777 if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2778 PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2780 // TODO: If BB is reachable from all paths through PredBlock, then we
2781 // could replace PBI's branch probabilities with BI's.
2783 // Copy any debug value intrinsics into the end of PredBlock.
2784 for (Instruction &I : *BB)
2785 if (isa<DbgInfoIntrinsic>(I))
2786 I.clone()->insertBefore(PBI);
2793 // If there is only one store in BB1 and BB2, return it, otherwise return
2795 static StoreInst *findUniqueStoreInBlocks(BasicBlock *BB1, BasicBlock *BB2) {
2796 StoreInst *S = nullptr;
2797 for (auto *BB : {BB1, BB2}) {
2801 if (auto *SI = dyn_cast<StoreInst>(&I)) {
2803 // Multiple stores seen.
2812 static Value *ensureValueAvailableInSuccessor(Value *V, BasicBlock *BB,
2813 Value *AlternativeV = nullptr) {
2814 // PHI is going to be a PHI node that allows the value V that is defined in
2815 // BB to be referenced in BB's only successor.
2817 // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2818 // doesn't matter to us what the other operand is (it'll never get used). We
2819 // could just create a new PHI with an undef incoming value, but that could
2820 // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2821 // other PHI. So here we directly look for some PHI in BB's successor with V
2822 // as an incoming operand. If we find one, we use it, else we create a new
2825 // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2826 // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2827 // where OtherBB is the single other predecessor of BB's only successor.
2828 PHINode *PHI = nullptr;
2829 BasicBlock *Succ = BB->getSingleSuccessor();
2831 for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2832 if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2833 PHI = cast<PHINode>(I);
2837 assert(std::distance(pred_begin(Succ), pred_end(Succ)) == 2);
2838 auto PredI = pred_begin(Succ);
2839 BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2840 if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2847 // If V is not an instruction defined in BB, just return it.
2848 if (!AlternativeV &&
2849 (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2852 PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2853 PHI->addIncoming(V, BB);
2854 for (BasicBlock *PredBB : predecessors(Succ))
2857 AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2861 static bool mergeConditionalStoreToAddress(BasicBlock *PTB, BasicBlock *PFB,
2862 BasicBlock *QTB, BasicBlock *QFB,
2863 BasicBlock *PostBB, Value *Address,
2864 bool InvertPCond, bool InvertQCond) {
2865 auto IsaBitcastOfPointerType = [](const Instruction &I) {
2866 return Operator::getOpcode(&I) == Instruction::BitCast &&
2867 I.getType()->isPointerTy();
2870 // If we're not in aggressive mode, we only optimize if we have some
2871 // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2872 auto IsWorthwhile = [&](BasicBlock *BB) {
2875 // Heuristic: if the block can be if-converted/phi-folded and the
2876 // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2877 // thread this store.
2879 for (auto &I : *BB) {
2880 // Cheap instructions viable for folding.
2881 if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2884 // Free instructions.
2885 else if (isa<TerminatorInst>(I) || isa<DbgInfoIntrinsic>(I) ||
2886 IsaBitcastOfPointerType(I))
2891 return N <= PHINodeFoldingThreshold;
2894 if (!MergeCondStoresAggressively &&
2895 (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2896 !IsWorthwhile(QFB)))
2899 // For every pointer, there must be exactly two stores, one coming from
2900 // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2901 // store (to any address) in PTB,PFB or QTB,QFB.
2902 // FIXME: We could relax this restriction with a bit more work and performance
2904 StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2905 StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2906 if (!PStore || !QStore)
2909 // Now check the stores are compatible.
2910 if (!QStore->isUnordered() || !PStore->isUnordered())
2913 // Check that sinking the store won't cause program behavior changes. Sinking
2914 // the store out of the Q blocks won't change any behavior as we're sinking
2915 // from a block to its unconditional successor. But we're moving a store from
2916 // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2917 // So we need to check that there are no aliasing loads or stores in
2918 // QBI, QTB and QFB. We also need to check there are no conflicting memory
2919 // operations between PStore and the end of its parent block.
2921 // The ideal way to do this is to query AliasAnalysis, but we don't
2922 // preserve AA currently so that is dangerous. Be super safe and just
2923 // check there are no other memory operations at all.
2924 for (auto &I : *QFB->getSinglePredecessor())
2925 if (I.mayReadOrWriteMemory())
2927 for (auto &I : *QFB)
2928 if (&I != QStore && I.mayReadOrWriteMemory())
2931 for (auto &I : *QTB)
2932 if (&I != QStore && I.mayReadOrWriteMemory())
2934 for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2936 if (&*I != PStore && I->mayReadOrWriteMemory())
2939 // OK, we're going to sink the stores to PostBB. The store has to be
2940 // conditional though, so first create the predicate.
2941 Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
2943 Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
2946 Value *PPHI = ensureValueAvailableInSuccessor(PStore->getValueOperand(),
2947 PStore->getParent());
2948 Value *QPHI = ensureValueAvailableInSuccessor(QStore->getValueOperand(),
2949 QStore->getParent(), PPHI);
2951 IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
2953 Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
2954 Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
2957 PPred = QB.CreateNot(PPred);
2959 QPred = QB.CreateNot(QPred);
2960 Value *CombinedPred = QB.CreateOr(PPred, QPred);
2963 SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
2964 QB.SetInsertPoint(T);
2965 StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
2967 PStore->getAAMetadata(AAMD, /*Merge=*/false);
2968 PStore->getAAMetadata(AAMD, /*Merge=*/true);
2969 SI->setAAMetadata(AAMD);
2971 QStore->eraseFromParent();
2972 PStore->eraseFromParent();
2977 static bool mergeConditionalStores(BranchInst *PBI, BranchInst *QBI) {
2978 // The intention here is to find diamonds or triangles (see below) where each
2979 // conditional block contains a store to the same address. Both of these
2980 // stores are conditional, so they can't be unconditionally sunk. But it may
2981 // be profitable to speculatively sink the stores into one merged store at the
2982 // end, and predicate the merged store on the union of the two conditions of
2985 // This can reduce the number of stores executed if both of the conditions are
2986 // true, and can allow the blocks to become small enough to be if-converted.
2987 // This optimization will also chain, so that ladders of test-and-set
2988 // sequences can be if-converted away.
2990 // We only deal with simple diamonds or triangles:
2992 // PBI or PBI or a combination of the two
3002 // We model triangles as a type of diamond with a nullptr "true" block.
3003 // Triangles are canonicalized so that the fallthrough edge is represented by
3004 // a true condition, as in the diagram above.
3006 BasicBlock *PTB = PBI->getSuccessor(0);
3007 BasicBlock *PFB = PBI->getSuccessor(1);
3008 BasicBlock *QTB = QBI->getSuccessor(0);
3009 BasicBlock *QFB = QBI->getSuccessor(1);
3010 BasicBlock *PostBB = QFB->getSingleSuccessor();
3012 // Make sure we have a good guess for PostBB. If QTB's only successor is
3013 // QFB, then QFB is a better PostBB.
3014 if (QTB->getSingleSuccessor() == QFB)
3017 // If we couldn't find a good PostBB, stop.
3021 bool InvertPCond = false, InvertQCond = false;
3022 // Canonicalize fallthroughs to the true branches.
3023 if (PFB == QBI->getParent()) {
3024 std::swap(PFB, PTB);
3027 if (QFB == PostBB) {
3028 std::swap(QFB, QTB);
3032 // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3033 // and QFB may not. Model fallthroughs as a nullptr block.
3034 if (PTB == QBI->getParent())
3039 // Legality bailouts. We must have at least the non-fallthrough blocks and
3040 // the post-dominating block, and the non-fallthroughs must only have one
3042 auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3043 return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3045 if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3046 !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3048 if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3049 (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3051 if (!PostBB->hasNUses(2) || !QBI->getParent()->hasNUses(2))
3054 // OK, this is a sequence of two diamonds or triangles.
3055 // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3056 SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3057 for (auto *BB : {PTB, PFB}) {
3061 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3062 PStoreAddresses.insert(SI->getPointerOperand());
3064 for (auto *BB : {QTB, QFB}) {
3068 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3069 QStoreAddresses.insert(SI->getPointerOperand());
3072 set_intersect(PStoreAddresses, QStoreAddresses);
3073 // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3074 // clear what it contains.
3075 auto &CommonAddresses = PStoreAddresses;
3077 bool Changed = false;
3078 for (auto *Address : CommonAddresses)
3079 Changed |= mergeConditionalStoreToAddress(
3080 PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond);
3084 /// If we have a conditional branch as a predecessor of another block,
3085 /// this function tries to simplify it. We know
3086 /// that PBI and BI are both conditional branches, and BI is in one of the
3087 /// successor blocks of PBI - PBI branches to BI.
3088 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
3089 const DataLayout &DL) {
3090 assert(PBI->isConditional() && BI->isConditional());
3091 BasicBlock *BB = BI->getParent();
3093 // If this block ends with a branch instruction, and if there is a
3094 // predecessor that ends on a branch of the same condition, make
3095 // this conditional branch redundant.
3096 if (PBI->getCondition() == BI->getCondition() &&
3097 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3098 // Okay, the outcome of this conditional branch is statically
3099 // knowable. If this block had a single pred, handle specially.
3100 if (BB->getSinglePredecessor()) {
3101 // Turn this into a branch on constant.
3102 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3104 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3105 return true; // Nuke the branch on constant.
3108 // Otherwise, if there are multiple predecessors, insert a PHI that merges
3109 // in the constant and simplify the block result. Subsequent passes of
3110 // simplifycfg will thread the block.
3111 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
3112 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3113 PHINode *NewPN = PHINode::Create(
3114 Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3115 BI->getCondition()->getName() + ".pr", &BB->front());
3116 // Okay, we're going to insert the PHI node. Since PBI is not the only
3117 // predecessor, compute the PHI'd conditional value for all of the preds.
3118 // Any predecessor where the condition is not computable we keep symbolic.
3119 for (pred_iterator PI = PB; PI != PE; ++PI) {
3120 BasicBlock *P = *PI;
3121 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3122 PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3123 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3124 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3126 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3129 NewPN->addIncoming(BI->getCondition(), P);
3133 BI->setCondition(NewPN);
3138 if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3142 // If both branches are conditional and both contain stores to the same
3143 // address, remove the stores from the conditionals and create a conditional
3144 // merged store at the end.
3145 if (MergeCondStores && mergeConditionalStores(PBI, BI))
3148 // If this is a conditional branch in an empty block, and if any
3149 // predecessors are a conditional branch to one of our destinations,
3150 // fold the conditions into logical ops and one cond br.
3151 BasicBlock::iterator BBI = BB->begin();
3152 // Ignore dbg intrinsics.
3153 while (isa<DbgInfoIntrinsic>(BBI))
3159 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3162 } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3165 } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3168 } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3175 // Check to make sure that the other destination of this branch
3176 // isn't BB itself. If so, this is an infinite loop that will
3177 // keep getting unwound.
3178 if (PBI->getSuccessor(PBIOp) == BB)
3181 // Do not perform this transformation if it would require
3182 // insertion of a large number of select instructions. For targets
3183 // without predication/cmovs, this is a big pessimization.
3185 // Also do not perform this transformation if any phi node in the common
3186 // destination block can trap when reached by BB or PBB (PR17073). In that
3187 // case, it would be unsafe to hoist the operation into a select instruction.
3189 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3190 unsigned NumPhis = 0;
3191 for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3193 if (NumPhis > 2) // Disable this xform.
3196 PHINode *PN = cast<PHINode>(II);
3197 Value *BIV = PN->getIncomingValueForBlock(BB);
3198 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3202 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3203 Value *PBIV = PN->getIncomingValue(PBBIdx);
3204 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3209 // Finally, if everything is ok, fold the branches to logical ops.
3210 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3212 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3213 << "AND: " << *BI->getParent());
3215 // If OtherDest *is* BB, then BB is a basic block with a single conditional
3216 // branch in it, where one edge (OtherDest) goes back to itself but the other
3217 // exits. We don't *know* that the program avoids the infinite loop
3218 // (even though that seems likely). If we do this xform naively, we'll end up
3219 // recursively unpeeling the loop. Since we know that (after the xform is
3220 // done) that the block *is* infinite if reached, we just make it an obviously
3221 // infinite loop with no cond branch.
3222 if (OtherDest == BB) {
3223 // Insert it at the end of the function, because it's either code,
3224 // or it won't matter if it's hot. :)
3225 BasicBlock *InfLoopBlock =
3226 BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3227 BranchInst::Create(InfLoopBlock, InfLoopBlock);
3228 OtherDest = InfLoopBlock;
3231 DEBUG(dbgs() << *PBI->getParent()->getParent());
3233 // BI may have other predecessors. Because of this, we leave
3234 // it alone, but modify PBI.
3236 // Make sure we get to CommonDest on True&True directions.
3237 Value *PBICond = PBI->getCondition();
3238 IRBuilder<NoFolder> Builder(PBI);
3240 PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3242 Value *BICond = BI->getCondition();
3244 BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3246 // Merge the conditions.
3247 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3249 // Modify PBI to branch on the new condition to the new dests.
3250 PBI->setCondition(Cond);
3251 PBI->setSuccessor(0, CommonDest);
3252 PBI->setSuccessor(1, OtherDest);
3254 // Update branch weight for PBI.
3255 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3256 uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3258 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3259 SuccTrueWeight, SuccFalseWeight);
3261 PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3262 PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3263 SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3264 SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3265 // The weight to CommonDest should be PredCommon * SuccTotal +
3266 // PredOther * SuccCommon.
3267 // The weight to OtherDest should be PredOther * SuccOther.
3268 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3269 PredOther * SuccCommon,
3270 PredOther * SuccOther};
3271 // Halve the weights if any of them cannot fit in an uint32_t
3272 FitWeights(NewWeights);
3274 PBI->setMetadata(LLVMContext::MD_prof,
3275 MDBuilder(BI->getContext())
3276 .createBranchWeights(NewWeights[0], NewWeights[1]));
3279 // OtherDest may have phi nodes. If so, add an entry from PBI's
3280 // block that are identical to the entries for BI's block.
3281 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3283 // We know that the CommonDest already had an edge from PBI to
3284 // it. If it has PHIs though, the PHIs may have different
3285 // entries for BB and PBI's BB. If so, insert a select to make
3288 for (BasicBlock::iterator II = CommonDest->begin();
3289 (PN = dyn_cast<PHINode>(II)); ++II) {
3290 Value *BIV = PN->getIncomingValueForBlock(BB);
3291 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3292 Value *PBIV = PN->getIncomingValue(PBBIdx);
3294 // Insert a select in PBI to pick the right value.
3295 SelectInst *NV = cast<SelectInst>(
3296 Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3297 PN->setIncomingValue(PBBIdx, NV);
3298 // Although the select has the same condition as PBI, the original branch
3299 // weights for PBI do not apply to the new select because the select's
3300 // 'logical' edges are incoming edges of the phi that is eliminated, not
3301 // the outgoing edges of PBI.
3303 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3304 uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3305 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3306 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3307 // The weight to PredCommonDest should be PredCommon * SuccTotal.
3308 // The weight to PredOtherDest should be PredOther * SuccCommon.
3309 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3310 PredOther * SuccCommon};
3312 FitWeights(NewWeights);
3314 NV->setMetadata(LLVMContext::MD_prof,
3315 MDBuilder(BI->getContext())
3316 .createBranchWeights(NewWeights[0], NewWeights[1]));
3321 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3322 DEBUG(dbgs() << *PBI->getParent()->getParent());
3324 // This basic block is probably dead. We know it has at least
3325 // one fewer predecessor.
3329 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3330 // true or to FalseBB if Cond is false.
3331 // Takes care of updating the successors and removing the old terminator.
3332 // Also makes sure not to introduce new successors by assuming that edges to
3333 // non-successor TrueBBs and FalseBBs aren't reachable.
3334 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
3335 BasicBlock *TrueBB, BasicBlock *FalseBB,
3336 uint32_t TrueWeight,
3337 uint32_t FalseWeight) {
3338 // Remove any superfluous successor edges from the CFG.
3339 // First, figure out which successors to preserve.
3340 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3342 BasicBlock *KeepEdge1 = TrueBB;
3343 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3345 // Then remove the rest.
3346 for (BasicBlock *Succ : OldTerm->successors()) {
3347 // Make sure only to keep exactly one copy of each edge.
3348 if (Succ == KeepEdge1)
3349 KeepEdge1 = nullptr;
3350 else if (Succ == KeepEdge2)
3351 KeepEdge2 = nullptr;
3353 Succ->removePredecessor(OldTerm->getParent(),
3354 /*DontDeleteUselessPHIs=*/true);
3357 IRBuilder<> Builder(OldTerm);
3358 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3360 // Insert an appropriate new terminator.
3361 if (!KeepEdge1 && !KeepEdge2) {
3362 if (TrueBB == FalseBB)
3363 // We were only looking for one successor, and it was present.
3364 // Create an unconditional branch to it.
3365 Builder.CreateBr(TrueBB);
3367 // We found both of the successors we were looking for.
3368 // Create a conditional branch sharing the condition of the select.
3369 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3370 if (TrueWeight != FalseWeight)
3371 NewBI->setMetadata(LLVMContext::MD_prof,
3372 MDBuilder(OldTerm->getContext())
3373 .createBranchWeights(TrueWeight, FalseWeight));
3375 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3376 // Neither of the selected blocks were successors, so this
3377 // terminator must be unreachable.
3378 new UnreachableInst(OldTerm->getContext(), OldTerm);
3380 // One of the selected values was a successor, but the other wasn't.
3381 // Insert an unconditional branch to the one that was found;
3382 // the edge to the one that wasn't must be unreachable.
3384 // Only TrueBB was found.
3385 Builder.CreateBr(TrueBB);
3387 // Only FalseBB was found.
3388 Builder.CreateBr(FalseBB);
3391 EraseTerminatorInstAndDCECond(OldTerm);
3396 // (switch (select cond, X, Y)) on constant X, Y
3397 // with a branch - conditional if X and Y lead to distinct BBs,
3398 // unconditional otherwise.
3399 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
3400 // Check for constant integer values in the select.
3401 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3402 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3403 if (!TrueVal || !FalseVal)
3406 // Find the relevant condition and destinations.
3407 Value *Condition = Select->getCondition();
3408 BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3409 BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3411 // Get weight for TrueBB and FalseBB.
3412 uint32_t TrueWeight = 0, FalseWeight = 0;
3413 SmallVector<uint64_t, 8> Weights;
3414 bool HasWeights = HasBranchWeights(SI);
3416 GetBranchWeights(SI, Weights);
3417 if (Weights.size() == 1 + SI->getNumCases()) {
3419 (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3421 (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3425 // Perform the actual simplification.
3426 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3431 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3432 // blockaddress(@fn, BlockB)))
3434 // (br cond, BlockA, BlockB).
3435 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
3436 // Check that both operands of the select are block addresses.
3437 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
3438 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
3442 // Extract the actual blocks.
3443 BasicBlock *TrueBB = TBA->getBasicBlock();
3444 BasicBlock *FalseBB = FBA->getBasicBlock();
3446 // Perform the actual simplification.
3447 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3451 /// This is called when we find an icmp instruction
3452 /// (a seteq/setne with a constant) as the only instruction in a
3453 /// block that ends with an uncond branch. We are looking for a very specific
3454 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3455 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3456 /// default value goes to an uncond block with a seteq in it, we get something
3459 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3461 /// %tmp = icmp eq i8 %A, 92
3464 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3466 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3467 /// the PHI, merging the third icmp into the switch.
3468 static bool TryToSimplifyUncondBranchWithICmpInIt(
3469 ICmpInst *ICI, IRBuilder<> &Builder, const DataLayout &DL,
3470 const TargetTransformInfo &TTI, unsigned BonusInstThreshold,
3471 AssumptionCache *AC) {
3472 BasicBlock *BB = ICI->getParent();
3474 // If the block has any PHIs in it or the icmp has multiple uses, it is too
3476 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3479 Value *V = ICI->getOperand(0);
3480 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3482 // The pattern we're looking for is where our only predecessor is a switch on
3483 // 'V' and this block is the default case for the switch. In this case we can
3484 // fold the compared value into the switch to simplify things.
3485 BasicBlock *Pred = BB->getSinglePredecessor();
3486 if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3489 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3490 if (SI->getCondition() != V)
3493 // If BB is reachable on a non-default case, then we simply know the value of
3494 // V in this block. Substitute it and constant fold the icmp instruction
3496 if (SI->getDefaultDest() != BB) {
3497 ConstantInt *VVal = SI->findCaseDest(BB);
3498 assert(VVal && "Should have a unique destination value");
3499 ICI->setOperand(0, VVal);
3501 if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
3502 ICI->replaceAllUsesWith(V);
3503 ICI->eraseFromParent();
3505 // BB is now empty, so it is likely to simplify away.
3506 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3509 // Ok, the block is reachable from the default dest. If the constant we're
3510 // comparing exists in one of the other edges, then we can constant fold ICI
3512 if (SI->findCaseValue(Cst) != SI->case_default()) {
3514 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3515 V = ConstantInt::getFalse(BB->getContext());
3517 V = ConstantInt::getTrue(BB->getContext());
3519 ICI->replaceAllUsesWith(V);
3520 ICI->eraseFromParent();
3521 // BB is now empty, so it is likely to simplify away.
3522 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3525 // The use of the icmp has to be in the 'end' block, by the only PHI node in
3527 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3528 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3529 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3530 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3533 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3535 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3536 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3538 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3539 std::swap(DefaultCst, NewCst);
3541 // Replace ICI (which is used by the PHI for the default value) with true or
3542 // false depending on if it is EQ or NE.
3543 ICI->replaceAllUsesWith(DefaultCst);
3544 ICI->eraseFromParent();
3546 // Okay, the switch goes to this block on a default value. Add an edge from
3547 // the switch to the merge point on the compared value.
3549 BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3550 SmallVector<uint64_t, 8> Weights;
3551 bool HasWeights = HasBranchWeights(SI);
3553 GetBranchWeights(SI, Weights);
3554 if (Weights.size() == 1 + SI->getNumCases()) {
3555 // Split weight for default case to case for "Cst".
3556 Weights[0] = (Weights[0] + 1) >> 1;
3557 Weights.push_back(Weights[0]);
3559 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3561 LLVMContext::MD_prof,
3562 MDBuilder(SI->getContext()).createBranchWeights(MDWeights));
3565 SI->addCase(Cst, NewBB);
3567 // NewBB branches to the phi block, add the uncond branch and the phi entry.
3568 Builder.SetInsertPoint(NewBB);
3569 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3570 Builder.CreateBr(SuccBlock);
3571 PHIUse->addIncoming(NewCst, NewBB);
3575 /// The specified branch is a conditional branch.
3576 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3577 /// fold it into a switch instruction if so.
3578 static bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
3579 const DataLayout &DL) {
3580 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3584 // Change br (X == 0 | X == 1), T, F into a switch instruction.
3585 // If this is a bunch of seteq's or'd together, or if it's a bunch of
3586 // 'setne's and'ed together, collect them.
3588 // Try to gather values from a chain of and/or to be turned into a switch
3589 ConstantComparesGatherer ConstantCompare(Cond, DL);
3590 // Unpack the result
3591 SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3592 Value *CompVal = ConstantCompare.CompValue;
3593 unsigned UsedICmps = ConstantCompare.UsedICmps;
3594 Value *ExtraCase = ConstantCompare.Extra;
3596 // If we didn't have a multiply compared value, fail.
3600 // Avoid turning single icmps into a switch.
3604 bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3606 // There might be duplicate constants in the list, which the switch
3607 // instruction can't handle, remove them now.
3608 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3609 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3611 // If Extra was used, we require at least two switch values to do the
3612 // transformation. A switch with one value is just a conditional branch.
3613 if (ExtraCase && Values.size() < 2)
3616 // TODO: Preserve branch weight metadata, similarly to how
3617 // FoldValueComparisonIntoPredecessors preserves it.
3619 // Figure out which block is which destination.
3620 BasicBlock *DefaultBB = BI->getSuccessor(1);
3621 BasicBlock *EdgeBB = BI->getSuccessor(0);
3623 std::swap(DefaultBB, EdgeBB);
3625 BasicBlock *BB = BI->getParent();
3627 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3628 << " cases into SWITCH. BB is:\n"
3631 // If there are any extra values that couldn't be folded into the switch
3632 // then we evaluate them with an explicit branch first. Split the block
3633 // right before the condbr to handle it.
3636 BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3637 // Remove the uncond branch added to the old block.
3638 TerminatorInst *OldTI = BB->getTerminator();
3639 Builder.SetInsertPoint(OldTI);
3642 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3644 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3646 OldTI->eraseFromParent();
3648 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3649 // for the edge we just added.
3650 AddPredecessorToBlock(EdgeBB, BB, NewBB);
3652 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3653 << "\nEXTRABB = " << *BB);
3657 Builder.SetInsertPoint(BI);
3658 // Convert pointer to int before we switch.
3659 if (CompVal->getType()->isPointerTy()) {
3660 CompVal = Builder.CreatePtrToInt(
3661 CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3664 // Create the new switch instruction now.
3665 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3667 // Add all of the 'cases' to the switch instruction.
3668 for (unsigned i = 0, e = Values.size(); i != e; ++i)
3669 New->addCase(Values[i], EdgeBB);
3671 // We added edges from PI to the EdgeBB. As such, if there were any
3672 // PHI nodes in EdgeBB, they need entries to be added corresponding to
3673 // the number of edges added.
3674 for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3675 PHINode *PN = cast<PHINode>(BBI);
3676 Value *InVal = PN->getIncomingValueForBlock(BB);
3677 for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3678 PN->addIncoming(InVal, BB);
3681 // Erase the old branch instruction.
3682 EraseTerminatorInstAndDCECond(BI);
3684 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3688 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3689 if (isa<PHINode>(RI->getValue()))
3690 return SimplifyCommonResume(RI);
3691 else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3692 RI->getValue() == RI->getParent()->getFirstNonPHI())
3693 // The resume must unwind the exception that caused control to branch here.
3694 return SimplifySingleResume(RI);
3699 // Simplify resume that is shared by several landing pads (phi of landing pad).
3700 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3701 BasicBlock *BB = RI->getParent();
3703 // Check that there are no other instructions except for debug intrinsics
3704 // between the phi of landing pads (RI->getValue()) and resume instruction.
3705 BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3706 E = RI->getIterator();
3708 if (!isa<DbgInfoIntrinsic>(I))
3711 SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
3712 auto *PhiLPInst = cast<PHINode>(RI->getValue());
3714 // Check incoming blocks to see if any of them are trivial.
3715 for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3717 auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3718 auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3720 // If the block has other successors, we can not delete it because
3721 // it has other dependents.
3722 if (IncomingBB->getUniqueSuccessor() != BB)
3725 auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3726 // Not the landing pad that caused the control to branch here.
3727 if (IncomingValue != LandingPad)
3730 bool isTrivial = true;
3732 I = IncomingBB->getFirstNonPHI()->getIterator();
3733 E = IncomingBB->getTerminator()->getIterator();
3735 if (!isa<DbgInfoIntrinsic>(I)) {
3741 TrivialUnwindBlocks.insert(IncomingBB);
3744 // If no trivial unwind blocks, don't do any simplifications.
3745 if (TrivialUnwindBlocks.empty())
3748 // Turn all invokes that unwind here into calls.
3749 for (auto *TrivialBB : TrivialUnwindBlocks) {
3750 // Blocks that will be simplified should be removed from the phi node.
3751 // Note there could be multiple edges to the resume block, and we need
3752 // to remove them all.
3753 while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3754 BB->removePredecessor(TrivialBB, true);
3756 for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3758 BasicBlock *Pred = *PI++;
3759 removeUnwindEdge(Pred);
3762 // In each SimplifyCFG run, only the current processed block can be erased.
3763 // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3764 // of erasing TrivialBB, we only remove the branch to the common resume
3765 // block so that we can later erase the resume block since it has no
3767 TrivialBB->getTerminator()->eraseFromParent();
3768 new UnreachableInst(RI->getContext(), TrivialBB);
3771 // Delete the resume block if all its predecessors have been removed.
3773 BB->eraseFromParent();
3775 return !TrivialUnwindBlocks.empty();
3778 // Simplify resume that is only used by a single (non-phi) landing pad.
3779 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3780 BasicBlock *BB = RI->getParent();
3781 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
3782 assert(RI->getValue() == LPInst &&
3783 "Resume must unwind the exception that caused control to here");
3785 // Check that there are no other instructions except for debug intrinsics.
3786 BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3788 if (!isa<DbgInfoIntrinsic>(I))
3791 // Turn all invokes that unwind here into calls and delete the basic block.
3792 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3793 BasicBlock *Pred = *PI++;
3794 removeUnwindEdge(Pred);
3797 // The landingpad is now unreachable. Zap it.
3798 BB->eraseFromParent();
3800 LoopHeaders->erase(BB);
3804 static bool removeEmptyCleanup(CleanupReturnInst *RI) {
3805 // If this is a trivial cleanup pad that executes no instructions, it can be
3806 // eliminated. If the cleanup pad continues to the caller, any predecessor
3807 // that is an EH pad will be updated to continue to the caller and any
3808 // predecessor that terminates with an invoke instruction will have its invoke
3809 // instruction converted to a call instruction. If the cleanup pad being
3810 // simplified does not continue to the caller, each predecessor will be
3811 // updated to continue to the unwind destination of the cleanup pad being
3813 BasicBlock *BB = RI->getParent();
3814 CleanupPadInst *CPInst = RI->getCleanupPad();
3815 if (CPInst->getParent() != BB)
3816 // This isn't an empty cleanup.
3819 // We cannot kill the pad if it has multiple uses. This typically arises
3820 // from unreachable basic blocks.
3821 if (!CPInst->hasOneUse())
3824 // Check that there are no other instructions except for benign intrinsics.
3825 BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3827 auto *II = dyn_cast<IntrinsicInst>(I);
3831 Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3832 switch (IntrinsicID) {
3833 case Intrinsic::dbg_declare:
3834 case Intrinsic::dbg_value:
3835 case Intrinsic::lifetime_end:
3842 // If the cleanup return we are simplifying unwinds to the caller, this will
3843 // set UnwindDest to nullptr.
3844 BasicBlock *UnwindDest = RI->getUnwindDest();
3845 Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3847 // We're about to remove BB from the control flow. Before we do, sink any
3848 // PHINodes into the unwind destination. Doing this before changing the
3849 // control flow avoids some potentially slow checks, since we can currently
3850 // be certain that UnwindDest and BB have no common predecessors (since they
3851 // are both EH pads).
3853 // First, go through the PHI nodes in UnwindDest and update any nodes that
3854 // reference the block we are removing
3855 for (BasicBlock::iterator I = UnwindDest->begin(),
3856 IE = DestEHPad->getIterator();
3858 PHINode *DestPN = cast<PHINode>(I);
3860 int Idx = DestPN->getBasicBlockIndex(BB);
3861 // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3863 // This PHI node has an incoming value that corresponds to a control
3864 // path through the cleanup pad we are removing. If the incoming
3865 // value is in the cleanup pad, it must be a PHINode (because we
3866 // verified above that the block is otherwise empty). Otherwise, the
3867 // value is either a constant or a value that dominates the cleanup
3868 // pad being removed.
3870 // Because BB and UnwindDest are both EH pads, all of their
3871 // predecessors must unwind to these blocks, and since no instruction
3872 // can have multiple unwind destinations, there will be no overlap in
3873 // incoming blocks between SrcPN and DestPN.
3874 Value *SrcVal = DestPN->getIncomingValue(Idx);
3875 PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3877 // Remove the entry for the block we are deleting.
3878 DestPN->removeIncomingValue(Idx, false);
3880 if (SrcPN && SrcPN->getParent() == BB) {
3881 // If the incoming value was a PHI node in the cleanup pad we are
3882 // removing, we need to merge that PHI node's incoming values into
3884 for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3885 SrcIdx != SrcE; ++SrcIdx) {
3886 DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3887 SrcPN->getIncomingBlock(SrcIdx));
3890 // Otherwise, the incoming value came from above BB and
3891 // so we can just reuse it. We must associate all of BB's
3892 // predecessors with this value.
3893 for (auto *pred : predecessors(BB)) {
3894 DestPN->addIncoming(SrcVal, pred);
3899 // Sink any remaining PHI nodes directly into UnwindDest.
3900 Instruction *InsertPt = DestEHPad;
3901 for (BasicBlock::iterator I = BB->begin(),
3902 IE = BB->getFirstNonPHI()->getIterator();
3904 // The iterator must be incremented here because the instructions are
3905 // being moved to another block.
3906 PHINode *PN = cast<PHINode>(I++);
3907 if (PN->use_empty())
3908 // If the PHI node has no uses, just leave it. It will be erased
3909 // when we erase BB below.
3912 // Otherwise, sink this PHI node into UnwindDest.
3913 // Any predecessors to UnwindDest which are not already represented
3914 // must be back edges which inherit the value from the path through
3915 // BB. In this case, the PHI value must reference itself.
3916 for (auto *pred : predecessors(UnwindDest))
3918 PN->addIncoming(PN, pred);
3919 PN->moveBefore(InsertPt);
3923 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3924 // The iterator must be updated here because we are removing this pred.
3925 BasicBlock *PredBB = *PI++;
3926 if (UnwindDest == nullptr) {
3927 removeUnwindEdge(PredBB);
3929 TerminatorInst *TI = PredBB->getTerminator();
3930 TI->replaceUsesOfWith(BB, UnwindDest);
3934 // The cleanup pad is now unreachable. Zap it.
3935 BB->eraseFromParent();
3939 // Try to merge two cleanuppads together.
3940 static bool mergeCleanupPad(CleanupReturnInst *RI) {
3941 // Skip any cleanuprets which unwind to caller, there is nothing to merge
3943 BasicBlock *UnwindDest = RI->getUnwindDest();
3947 // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
3948 // be safe to merge without code duplication.
3949 if (UnwindDest->getSinglePredecessor() != RI->getParent())
3952 // Verify that our cleanuppad's unwind destination is another cleanuppad.
3953 auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
3954 if (!SuccessorCleanupPad)
3957 CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
3958 // Replace any uses of the successor cleanupad with the predecessor pad
3959 // The only cleanuppad uses should be this cleanupret, it's cleanupret and
3960 // funclet bundle operands.
3961 SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
3962 // Remove the old cleanuppad.
3963 SuccessorCleanupPad->eraseFromParent();
3964 // Now, we simply replace the cleanupret with a branch to the unwind
3966 BranchInst::Create(UnwindDest, RI->getParent());
3967 RI->eraseFromParent();
3972 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
3973 // It is possible to transiantly have an undef cleanuppad operand because we
3974 // have deleted some, but not all, dead blocks.
3975 // Eventually, this block will be deleted.
3976 if (isa<UndefValue>(RI->getOperand(0)))
3979 if (mergeCleanupPad(RI))
3982 if (removeEmptyCleanup(RI))
3988 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
3989 BasicBlock *BB = RI->getParent();
3990 if (!BB->getFirstNonPHIOrDbg()->isTerminator())
3993 // Find predecessors that end with branches.
3994 SmallVector<BasicBlock *, 8> UncondBranchPreds;
3995 SmallVector<BranchInst *, 8> CondBranchPreds;
3996 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
3997 BasicBlock *P = *PI;
3998 TerminatorInst *PTI = P->getTerminator();
3999 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4000 if (BI->isUnconditional())
4001 UncondBranchPreds.push_back(P);
4003 CondBranchPreds.push_back(BI);
4007 // If we found some, do the transformation!
4008 if (!UncondBranchPreds.empty() && DupRet) {
4009 while (!UncondBranchPreds.empty()) {
4010 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4011 DEBUG(dbgs() << "FOLDING: " << *BB
4012 << "INTO UNCOND BRANCH PRED: " << *Pred);
4013 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4016 // If we eliminated all predecessors of the block, delete the block now.
4017 if (pred_empty(BB)) {
4018 // We know there are no successors, so just nuke the block.
4019 BB->eraseFromParent();
4021 LoopHeaders->erase(BB);
4027 // Check out all of the conditional branches going to this return
4028 // instruction. If any of them just select between returns, change the
4029 // branch itself into a select/return pair.
4030 while (!CondBranchPreds.empty()) {
4031 BranchInst *BI = CondBranchPreds.pop_back_val();
4033 // Check to see if the non-BB successor is also a return block.
4034 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4035 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4036 SimplifyCondBranchToTwoReturns(BI, Builder))
4042 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4043 BasicBlock *BB = UI->getParent();
4045 bool Changed = false;
4047 // If there are any instructions immediately before the unreachable that can
4048 // be removed, do so.
4049 while (UI->getIterator() != BB->begin()) {
4050 BasicBlock::iterator BBI = UI->getIterator();
4052 // Do not delete instructions that can have side effects which might cause
4053 // the unreachable to not be reachable; specifically, calls and volatile
4054 // operations may have this effect.
4055 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4058 if (BBI->mayHaveSideEffects()) {
4059 if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4060 if (SI->isVolatile())
4062 } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4063 if (LI->isVolatile())
4065 } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4066 if (RMWI->isVolatile())
4068 } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4069 if (CXI->isVolatile())
4071 } else if (isa<CatchPadInst>(BBI)) {
4072 // A catchpad may invoke exception object constructors and such, which
4073 // in some languages can be arbitrary code, so be conservative by
4075 // For CoreCLR, it just involves a type test, so can be removed.
4076 if (classifyEHPersonality(BB->getParent()->getPersonalityFn()) !=
4077 EHPersonality::CoreCLR)
4079 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4080 !isa<LandingPadInst>(BBI)) {
4083 // Note that deleting LandingPad's here is in fact okay, although it
4084 // involves a bit of subtle reasoning. If this inst is a LandingPad,
4085 // all the predecessors of this block will be the unwind edges of Invokes,
4086 // and we can therefore guarantee this block will be erased.
4089 // Delete this instruction (any uses are guaranteed to be dead)
4090 if (!BBI->use_empty())
4091 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4092 BBI->eraseFromParent();
4096 // If the unreachable instruction is the first in the block, take a gander
4097 // at all of the predecessors of this instruction, and simplify them.
4098 if (&BB->front() != UI)
4101 SmallVector<BasicBlock *, 8> Preds(pred_begin(BB), pred_end(BB));
4102 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4103 TerminatorInst *TI = Preds[i]->getTerminator();
4104 IRBuilder<> Builder(TI);
4105 if (auto *BI = dyn_cast<BranchInst>(TI)) {
4106 if (BI->isUnconditional()) {
4107 if (BI->getSuccessor(0) == BB) {
4108 new UnreachableInst(TI->getContext(), TI);
4109 TI->eraseFromParent();
4113 if (BI->getSuccessor(0) == BB) {
4114 Builder.CreateBr(BI->getSuccessor(1));
4115 EraseTerminatorInstAndDCECond(BI);
4116 } else if (BI->getSuccessor(1) == BB) {
4117 Builder.CreateBr(BI->getSuccessor(0));
4118 EraseTerminatorInstAndDCECond(BI);
4122 } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4123 for (auto i = SI->case_begin(), e = SI->case_end(); i != e;) {
4124 if (i->getCaseSuccessor() != BB) {
4128 BB->removePredecessor(SI->getParent());
4129 i = SI->removeCase(i);
4133 } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4134 if (II->getUnwindDest() == BB) {
4135 removeUnwindEdge(TI->getParent());
4138 } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4139 if (CSI->getUnwindDest() == BB) {
4140 removeUnwindEdge(TI->getParent());
4145 for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4146 E = CSI->handler_end();
4149 CSI->removeHandler(I);
4155 if (CSI->getNumHandlers() == 0) {
4156 BasicBlock *CatchSwitchBB = CSI->getParent();
4157 if (CSI->hasUnwindDest()) {
4158 // Redirect preds to the unwind dest
4159 CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4161 // Rewrite all preds to unwind to caller (or from invoke to call).
4162 SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4163 for (BasicBlock *EHPred : EHPreds)
4164 removeUnwindEdge(EHPred);
4166 // The catchswitch is no longer reachable.
4167 new UnreachableInst(CSI->getContext(), CSI);
4168 CSI->eraseFromParent();
4171 } else if (isa<CleanupReturnInst>(TI)) {
4172 new UnreachableInst(TI->getContext(), TI);
4173 TI->eraseFromParent();
4178 // If this block is now dead, remove it.
4179 if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4180 // We know there are no successors, so just nuke the block.
4181 BB->eraseFromParent();
4183 LoopHeaders->erase(BB);
4190 static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) {
4191 assert(Cases.size() >= 1);
4193 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4194 for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4195 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4201 /// Turn a switch with two reachable destinations into an integer range
4202 /// comparison and branch.
4203 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
4204 assert(SI->getNumCases() > 1 && "Degenerate switch?");
4207 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4209 // Partition the cases into two sets with different destinations.
4210 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4211 BasicBlock *DestB = nullptr;
4212 SmallVector<ConstantInt *, 16> CasesA;
4213 SmallVector<ConstantInt *, 16> CasesB;
4215 for (auto Case : SI->cases()) {
4216 BasicBlock *Dest = Case.getCaseSuccessor();
4219 if (Dest == DestA) {
4220 CasesA.push_back(Case.getCaseValue());
4225 if (Dest == DestB) {
4226 CasesB.push_back(Case.getCaseValue());
4229 return false; // More than two destinations.
4232 assert(DestA && DestB &&
4233 "Single-destination switch should have been folded.");
4234 assert(DestA != DestB);
4235 assert(DestB != SI->getDefaultDest());
4236 assert(!CasesB.empty() && "There must be non-default cases.");
4237 assert(!CasesA.empty() || HasDefault);
4239 // Figure out if one of the sets of cases form a contiguous range.
4240 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4241 BasicBlock *ContiguousDest = nullptr;
4242 BasicBlock *OtherDest = nullptr;
4243 if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4244 ContiguousCases = &CasesA;
4245 ContiguousDest = DestA;
4247 } else if (CasesAreContiguous(CasesB)) {
4248 ContiguousCases = &CasesB;
4249 ContiguousDest = DestB;
4254 // Start building the compare and branch.
4256 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4257 Constant *NumCases =
4258 ConstantInt::get(Offset->getType(), ContiguousCases->size());
4260 Value *Sub = SI->getCondition();
4261 if (!Offset->isNullValue())
4262 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4265 // If NumCases overflowed, then all possible values jump to the successor.
4266 if (NumCases->isNullValue() && !ContiguousCases->empty())
4267 Cmp = ConstantInt::getTrue(SI->getContext());
4269 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4270 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4272 // Update weight for the newly-created conditional branch.
4273 if (HasBranchWeights(SI)) {
4274 SmallVector<uint64_t, 8> Weights;
4275 GetBranchWeights(SI, Weights);
4276 if (Weights.size() == 1 + SI->getNumCases()) {
4277 uint64_t TrueWeight = 0;
4278 uint64_t FalseWeight = 0;
4279 for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4280 if (SI->getSuccessor(I) == ContiguousDest)
4281 TrueWeight += Weights[I];
4283 FalseWeight += Weights[I];
4285 while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4289 NewBI->setMetadata(LLVMContext::MD_prof,
4290 MDBuilder(SI->getContext())
4291 .createBranchWeights((uint32_t)TrueWeight,
4292 (uint32_t)FalseWeight));
4296 // Prune obsolete incoming values off the successors' PHI nodes.
4297 for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4298 unsigned PreviousEdges = ContiguousCases->size();
4299 if (ContiguousDest == SI->getDefaultDest())
4301 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4302 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4304 for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4305 unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4306 if (OtherDest == SI->getDefaultDest())
4308 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4309 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4313 SI->eraseFromParent();
4318 /// Compute masked bits for the condition of a switch
4319 /// and use it to remove dead cases.
4320 static bool EliminateDeadSwitchCases(SwitchInst *SI, AssumptionCache *AC,
4321 const DataLayout &DL) {
4322 Value *Cond = SI->getCondition();
4323 unsigned Bits = Cond->getType()->getIntegerBitWidth();
4324 KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4326 // We can also eliminate cases by determining that their values are outside of
4327 // the limited range of the condition based on how many significant (non-sign)
4328 // bits are in the condition value.
4329 unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4330 unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4332 // Gather dead cases.
4333 SmallVector<ConstantInt *, 8> DeadCases;
4334 for (auto &Case : SI->cases()) {
4335 const APInt &CaseVal = Case.getCaseValue()->getValue();
4336 if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4337 (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4338 DeadCases.push_back(Case.getCaseValue());
4339 DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal << " is dead.\n");
4343 // If we can prove that the cases must cover all possible values, the
4344 // default destination becomes dead and we can remove it. If we know some
4345 // of the bits in the value, we can use that to more precisely compute the
4346 // number of possible unique case values.
4348 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4349 const unsigned NumUnknownBits =
4350 Bits - (Known.Zero | Known.One).countPopulation();
4351 assert(NumUnknownBits <= Bits);
4352 if (HasDefault && DeadCases.empty() &&
4353 NumUnknownBits < 64 /* avoid overflow */ &&
4354 SI->getNumCases() == (1ULL << NumUnknownBits)) {
4355 DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4356 BasicBlock *NewDefault =
4357 SplitBlockPredecessors(SI->getDefaultDest(), SI->getParent(), "");
4358 SI->setDefaultDest(&*NewDefault);
4359 SplitBlock(&*NewDefault, &NewDefault->front());
4360 auto *OldTI = NewDefault->getTerminator();
4361 new UnreachableInst(SI->getContext(), OldTI);
4362 EraseTerminatorInstAndDCECond(OldTI);
4366 SmallVector<uint64_t, 8> Weights;
4367 bool HasWeight = HasBranchWeights(SI);
4369 GetBranchWeights(SI, Weights);
4370 HasWeight = (Weights.size() == 1 + SI->getNumCases());
4373 // Remove dead cases from the switch.
4374 for (ConstantInt *DeadCase : DeadCases) {
4375 SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4376 assert(CaseI != SI->case_default() &&
4377 "Case was not found. Probably mistake in DeadCases forming.");
4379 std::swap(Weights[CaseI->getCaseIndex() + 1], Weights.back());
4383 // Prune unused values from PHI nodes.
4384 CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4385 SI->removeCase(CaseI);
4387 if (HasWeight && Weights.size() >= 2) {
4388 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
4389 SI->setMetadata(LLVMContext::MD_prof,
4390 MDBuilder(SI->getParent()->getContext())
4391 .createBranchWeights(MDWeights));
4394 return !DeadCases.empty();
4397 /// If BB would be eligible for simplification by
4398 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4399 /// by an unconditional branch), look at the phi node for BB in the successor
4400 /// block and see if the incoming value is equal to CaseValue. If so, return
4401 /// the phi node, and set PhiIndex to BB's index in the phi node.
4402 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
4403 BasicBlock *BB, int *PhiIndex) {
4404 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4405 return nullptr; // BB must be empty to be a candidate for simplification.
4406 if (!BB->getSinglePredecessor())
4407 return nullptr; // BB must be dominated by the switch.
4409 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
4410 if (!Branch || !Branch->isUnconditional())
4411 return nullptr; // Terminator must be unconditional branch.
4413 BasicBlock *Succ = Branch->getSuccessor(0);
4415 BasicBlock::iterator I = Succ->begin();
4416 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
4417 int Idx = PHI->getBasicBlockIndex(BB);
4418 assert(Idx >= 0 && "PHI has no entry for predecessor?");
4420 Value *InValue = PHI->getIncomingValue(Idx);
4421 if (InValue != CaseValue)
4431 /// Try to forward the condition of a switch instruction to a phi node
4432 /// dominated by the switch, if that would mean that some of the destination
4433 /// blocks of the switch can be folded away.
4434 /// Returns true if a change is made.
4435 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
4436 typedef DenseMap<PHINode *, SmallVector<int, 4>> ForwardingNodesMap;
4437 ForwardingNodesMap ForwardingNodes;
4439 for (auto Case : SI->cases()) {
4440 ConstantInt *CaseValue = Case.getCaseValue();
4441 BasicBlock *CaseDest = Case.getCaseSuccessor();
4445 FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIndex);
4449 ForwardingNodes[PHI].push_back(PhiIndex);
4452 bool Changed = false;
4454 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
4455 E = ForwardingNodes.end();
4457 PHINode *Phi = I->first;
4458 SmallVectorImpl<int> &Indexes = I->second;
4460 if (Indexes.size() < 2)
4463 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
4464 Phi->setIncomingValue(Indexes[I], SI->getCondition());
4471 /// Return true if the backend will be able to handle
4472 /// initializing an array of constants like C.
4473 static bool ValidLookupTableConstant(Constant *C, const TargetTransformInfo &TTI) {
4474 if (C->isThreadDependent())
4476 if (C->isDLLImportDependent())
4479 if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4480 !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4481 !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4484 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4485 if (!CE->isGEPWithNoNotionalOverIndexing())
4487 if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4491 if (!TTI.shouldBuildLookupTablesForConstant(C))
4497 /// If V is a Constant, return it. Otherwise, try to look up
4498 /// its constant value in ConstantPool, returning 0 if it's not there.
4500 LookupConstant(Value *V,
4501 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4502 if (Constant *C = dyn_cast<Constant>(V))
4504 return ConstantPool.lookup(V);
4507 /// Try to fold instruction I into a constant. This works for
4508 /// simple instructions such as binary operations where both operands are
4509 /// constant or can be replaced by constants from the ConstantPool. Returns the
4510 /// resulting constant on success, 0 otherwise.
4512 ConstantFold(Instruction *I, const DataLayout &DL,
4513 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4514 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4515 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4518 if (A->isAllOnesValue())
4519 return LookupConstant(Select->getTrueValue(), ConstantPool);
4520 if (A->isNullValue())
4521 return LookupConstant(Select->getFalseValue(), ConstantPool);
4525 SmallVector<Constant *, 4> COps;
4526 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4527 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
4533 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4534 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4538 return ConstantFoldInstOperands(I, COps, DL);
4541 /// Try to determine the resulting constant values in phi nodes
4542 /// at the common destination basic block, *CommonDest, for one of the case
4543 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4544 /// case), of a switch instruction SI.
4546 GetCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest,
4547 BasicBlock **CommonDest,
4548 SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4549 const DataLayout &DL, const TargetTransformInfo &TTI) {
4550 // The block from which we enter the common destination.
4551 BasicBlock *Pred = SI->getParent();
4553 // If CaseDest is empty except for some side-effect free instructions through
4554 // which we can constant-propagate the CaseVal, continue to its successor.
4555 SmallDenseMap<Value *, Constant *> ConstantPool;
4556 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4557 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
4559 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
4560 // If the terminator is a simple branch, continue to the next block.
4561 if (T->getNumSuccessors() != 1 || T->isExceptional())
4564 CaseDest = T->getSuccessor(0);
4565 } else if (isa<DbgInfoIntrinsic>(I)) {
4566 // Skip debug intrinsic.
4568 } else if (Constant *C = ConstantFold(&*I, DL, ConstantPool)) {
4569 // Instruction is side-effect free and constant.
4571 // If the instruction has uses outside this block or a phi node slot for
4572 // the block, it is not safe to bypass the instruction since it would then
4573 // no longer dominate all its uses.
4574 for (auto &Use : I->uses()) {
4575 User *User = Use.getUser();
4576 if (Instruction *I = dyn_cast<Instruction>(User))
4577 if (I->getParent() == CaseDest)
4579 if (PHINode *Phi = dyn_cast<PHINode>(User))
4580 if (Phi->getIncomingBlock(Use) == CaseDest)
4585 ConstantPool.insert(std::make_pair(&*I, C));
4591 // If we did not have a CommonDest before, use the current one.
4593 *CommonDest = CaseDest;
4594 // If the destination isn't the common one, abort.
4595 if (CaseDest != *CommonDest)
4598 // Get the values for this case from phi nodes in the destination block.
4599 BasicBlock::iterator I = (*CommonDest)->begin();
4600 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
4601 int Idx = PHI->getBasicBlockIndex(Pred);
4605 Constant *ConstVal =
4606 LookupConstant(PHI->getIncomingValue(Idx), ConstantPool);
4610 // Be conservative about which kinds of constants we support.
4611 if (!ValidLookupTableConstant(ConstVal, TTI))
4614 Res.push_back(std::make_pair(PHI, ConstVal));
4617 return Res.size() > 0;
4620 // Helper function used to add CaseVal to the list of cases that generate
4622 static void MapCaseToResult(ConstantInt *CaseVal,
4623 SwitchCaseResultVectorTy &UniqueResults,
4625 for (auto &I : UniqueResults) {
4626 if (I.first == Result) {
4627 I.second.push_back(CaseVal);
4631 UniqueResults.push_back(
4632 std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4635 // Helper function that initializes a map containing
4636 // results for the PHI node of the common destination block for a switch
4637 // instruction. Returns false if multiple PHI nodes have been found or if
4638 // there is not a common destination block for the switch.
4639 static bool InitializeUniqueCases(SwitchInst *SI, PHINode *&PHI,
4640 BasicBlock *&CommonDest,
4641 SwitchCaseResultVectorTy &UniqueResults,
4642 Constant *&DefaultResult,
4643 const DataLayout &DL,
4644 const TargetTransformInfo &TTI) {
4645 for (auto &I : SI->cases()) {
4646 ConstantInt *CaseVal = I.getCaseValue();
4648 // Resulting value at phi nodes for this case value.
4649 SwitchCaseResultsTy Results;
4650 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4654 // Only one value per case is permitted
4655 if (Results.size() > 1)
4657 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4659 // Check the PHI consistency.
4661 PHI = Results[0].first;
4662 else if (PHI != Results[0].first)
4665 // Find the default result value.
4666 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
4667 BasicBlock *DefaultDest = SI->getDefaultDest();
4668 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4670 // If the default value is not found abort unless the default destination
4673 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4674 if ((!DefaultResult &&
4675 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4681 // Helper function that checks if it is possible to transform a switch with only
4682 // two cases (or two cases + default) that produces a result into a select.
4685 // case 10: %0 = icmp eq i32 %a, 10
4686 // return 10; %1 = select i1 %0, i32 10, i32 4
4687 // case 20: ----> %2 = icmp eq i32 %a, 20
4688 // return 2; %3 = select i1 %2, i32 2, i32 %1
4692 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4693 Constant *DefaultResult, Value *Condition,
4694 IRBuilder<> &Builder) {
4695 assert(ResultVector.size() == 2 &&
4696 "We should have exactly two unique results at this point");
4697 // If we are selecting between only two cases transform into a simple
4698 // select or a two-way select if default is possible.
4699 if (ResultVector[0].second.size() == 1 &&
4700 ResultVector[1].second.size() == 1) {
4701 ConstantInt *const FirstCase = ResultVector[0].second[0];
4702 ConstantInt *const SecondCase = ResultVector[1].second[0];
4704 bool DefaultCanTrigger = DefaultResult;
4705 Value *SelectValue = ResultVector[1].first;
4706 if (DefaultCanTrigger) {
4707 Value *const ValueCompare =
4708 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4709 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4710 DefaultResult, "switch.select");
4712 Value *const ValueCompare =
4713 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4714 return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4715 SelectValue, "switch.select");
4721 // Helper function to cleanup a switch instruction that has been converted into
4722 // a select, fixing up PHI nodes and basic blocks.
4723 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
4725 IRBuilder<> &Builder) {
4726 BasicBlock *SelectBB = SI->getParent();
4727 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4728 PHI->removeIncomingValue(SelectBB);
4729 PHI->addIncoming(SelectValue, SelectBB);
4731 Builder.CreateBr(PHI->getParent());
4733 // Remove the switch.
4734 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4735 BasicBlock *Succ = SI->getSuccessor(i);
4737 if (Succ == PHI->getParent())
4739 Succ->removePredecessor(SelectBB);
4741 SI->eraseFromParent();
4744 /// If the switch is only used to initialize one or more
4745 /// phi nodes in a common successor block with only two different
4746 /// constant values, replace the switch with select.
4747 static bool SwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4748 AssumptionCache *AC, const DataLayout &DL,
4749 const TargetTransformInfo &TTI) {
4750 Value *const Cond = SI->getCondition();
4751 PHINode *PHI = nullptr;
4752 BasicBlock *CommonDest = nullptr;
4753 Constant *DefaultResult;
4754 SwitchCaseResultVectorTy UniqueResults;
4755 // Collect all the cases that will deliver the same value from the switch.
4756 if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4759 // Selects choose between maximum two values.
4760 if (UniqueResults.size() != 2)
4762 assert(PHI != nullptr && "PHI for value select not found");
4764 Builder.SetInsertPoint(SI);
4765 Value *SelectValue =
4766 ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4768 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4771 // The switch couldn't be converted into a select.
4777 /// This class represents a lookup table that can be used to replace a switch.
4778 class SwitchLookupTable {
4780 /// Create a lookup table to use as a switch replacement with the contents
4781 /// of Values, using DefaultValue to fill any holes in the table.
4783 Module &M, uint64_t TableSize, ConstantInt *Offset,
4784 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4785 Constant *DefaultValue, const DataLayout &DL);
4787 /// Build instructions with Builder to retrieve the value at
4788 /// the position given by Index in the lookup table.
4789 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4791 /// Return true if a table with TableSize elements of
4792 /// type ElementType would fit in a target-legal register.
4793 static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4797 // Depending on the contents of the table, it can be represented in
4800 // For tables where each element contains the same value, we just have to
4801 // store that single value and return it for each lookup.
4804 // For tables where there is a linear relationship between table index
4805 // and values. We calculate the result with a simple multiplication
4806 // and addition instead of a table lookup.
4809 // For small tables with integer elements, we can pack them into a bitmap
4810 // that fits into a target-legal register. Values are retrieved by
4811 // shift and mask operations.
4814 // The table is stored as an array of values. Values are retrieved by load
4815 // instructions from the table.
4819 // For SingleValueKind, this is the single value.
4820 Constant *SingleValue;
4822 // For BitMapKind, this is the bitmap.
4823 ConstantInt *BitMap;
4824 IntegerType *BitMapElementTy;
4826 // For LinearMapKind, these are the constants used to derive the value.
4827 ConstantInt *LinearOffset;
4828 ConstantInt *LinearMultiplier;
4830 // For ArrayKind, this is the array.
4831 GlobalVariable *Array;
4834 } // end anonymous namespace
4836 SwitchLookupTable::SwitchLookupTable(
4837 Module &M, uint64_t TableSize, ConstantInt *Offset,
4838 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4839 Constant *DefaultValue, const DataLayout &DL)
4840 : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
4841 LinearOffset(nullptr), LinearMultiplier(nullptr), Array(nullptr) {
4842 assert(Values.size() && "Can't build lookup table without values!");
4843 assert(TableSize >= Values.size() && "Can't fit values in table!");
4845 // If all values in the table are equal, this is that value.
4846 SingleValue = Values.begin()->second;
4848 Type *ValueType = Values.begin()->second->getType();
4850 // Build up the table contents.
4851 SmallVector<Constant *, 64> TableContents(TableSize);
4852 for (size_t I = 0, E = Values.size(); I != E; ++I) {
4853 ConstantInt *CaseVal = Values[I].first;
4854 Constant *CaseRes = Values[I].second;
4855 assert(CaseRes->getType() == ValueType);
4857 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4858 TableContents[Idx] = CaseRes;
4860 if (CaseRes != SingleValue)
4861 SingleValue = nullptr;
4864 // Fill in any holes in the table with the default result.
4865 if (Values.size() < TableSize) {
4866 assert(DefaultValue &&
4867 "Need a default value to fill the lookup table holes.");
4868 assert(DefaultValue->getType() == ValueType);
4869 for (uint64_t I = 0; I < TableSize; ++I) {
4870 if (!TableContents[I])
4871 TableContents[I] = DefaultValue;
4874 if (DefaultValue != SingleValue)
4875 SingleValue = nullptr;
4878 // If each element in the table contains the same value, we only need to store
4879 // that single value.
4881 Kind = SingleValueKind;
4885 // Check if we can derive the value with a linear transformation from the
4887 if (isa<IntegerType>(ValueType)) {
4888 bool LinearMappingPossible = true;
4891 assert(TableSize >= 2 && "Should be a SingleValue table.");
4892 // Check if there is the same distance between two consecutive values.
4893 for (uint64_t I = 0; I < TableSize; ++I) {
4894 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4896 // This is an undef. We could deal with it, but undefs in lookup tables
4897 // are very seldom. It's probably not worth the additional complexity.
4898 LinearMappingPossible = false;
4901 const APInt &Val = ConstVal->getValue();
4903 APInt Dist = Val - PrevVal;
4906 } else if (Dist != DistToPrev) {
4907 LinearMappingPossible = false;
4913 if (LinearMappingPossible) {
4914 LinearOffset = cast<ConstantInt>(TableContents[0]);
4915 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
4916 Kind = LinearMapKind;
4922 // If the type is integer and the table fits in a register, build a bitmap.
4923 if (WouldFitInRegister(DL, TableSize, ValueType)) {
4924 IntegerType *IT = cast<IntegerType>(ValueType);
4925 APInt TableInt(TableSize * IT->getBitWidth(), 0);
4926 for (uint64_t I = TableSize; I > 0; --I) {
4927 TableInt <<= IT->getBitWidth();
4928 // Insert values into the bitmap. Undef values are set to zero.
4929 if (!isa<UndefValue>(TableContents[I - 1])) {
4930 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
4931 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
4934 BitMap = ConstantInt::get(M.getContext(), TableInt);
4935 BitMapElementTy = IT;
4941 // Store the table in an array.
4942 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
4943 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
4945 Array = new GlobalVariable(M, ArrayTy, /*constant=*/true,
4946 GlobalVariable::PrivateLinkage, Initializer,
4948 Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
4952 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
4954 case SingleValueKind:
4956 case LinearMapKind: {
4957 // Derive the result value from the input value.
4958 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
4959 false, "switch.idx.cast");
4960 if (!LinearMultiplier->isOne())
4961 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
4962 if (!LinearOffset->isZero())
4963 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
4967 // Type of the bitmap (e.g. i59).
4968 IntegerType *MapTy = BitMap->getType();
4970 // Cast Index to the same type as the bitmap.
4971 // Note: The Index is <= the number of elements in the table, so
4972 // truncating it to the width of the bitmask is safe.
4973 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
4975 // Multiply the shift amount by the element width.
4976 ShiftAmt = Builder.CreateMul(
4977 ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
4981 Value *DownShifted =
4982 Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
4984 return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
4987 // Make sure the table index will not overflow when treated as signed.
4988 IntegerType *IT = cast<IntegerType>(Index->getType());
4989 uint64_t TableSize =
4990 Array->getInitializer()->getType()->getArrayNumElements();
4991 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
4992 Index = Builder.CreateZExt(
4993 Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
4994 "switch.tableidx.zext");
4996 Value *GEPIndices[] = {Builder.getInt32(0), Index};
4997 Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
4998 GEPIndices, "switch.gep");
4999 return Builder.CreateLoad(GEP, "switch.load");
5002 llvm_unreachable("Unknown lookup table kind!");
5005 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5007 Type *ElementType) {
5008 auto *IT = dyn_cast<IntegerType>(ElementType);
5011 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5012 // are <= 15, we could try to narrow the type.
5014 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5015 if (TableSize >= UINT_MAX / IT->getBitWidth())
5017 return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5020 /// Determine whether a lookup table should be built for this switch, based on
5021 /// the number of cases, size of the table, and the types of the results.
5023 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5024 const TargetTransformInfo &TTI, const DataLayout &DL,
5025 const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5026 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5027 return false; // TableSize overflowed, or mul below might overflow.
5029 bool AllTablesFitInRegister = true;
5030 bool HasIllegalType = false;
5031 for (const auto &I : ResultTypes) {
5032 Type *Ty = I.second;
5034 // Saturate this flag to true.
5035 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5037 // Saturate this flag to false.
5038 AllTablesFitInRegister =
5039 AllTablesFitInRegister &&
5040 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5042 // If both flags saturate, we're done. NOTE: This *only* works with
5043 // saturating flags, and all flags have to saturate first due to the
5044 // non-deterministic behavior of iterating over a dense map.
5045 if (HasIllegalType && !AllTablesFitInRegister)
5049 // If each table would fit in a register, we should build it anyway.
5050 if (AllTablesFitInRegister)
5053 // Don't build a table that doesn't fit in-register if it has illegal types.
5057 // The table density should be at least 40%. This is the same criterion as for
5058 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5059 // FIXME: Find the best cut-off.
5060 return SI->getNumCases() * 10 >= TableSize * 4;
5063 /// Try to reuse the switch table index compare. Following pattern:
5065 /// if (idx < tablesize)
5066 /// r = table[idx]; // table does not contain default_value
5068 /// r = default_value;
5069 /// if (r != default_value)
5072 /// Is optimized to:
5074 /// cond = idx < tablesize;
5078 /// r = default_value;
5082 /// Jump threading will then eliminate the second if(cond).
5083 static void reuseTableCompare(
5084 User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5085 Constant *DefaultValue,
5086 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5088 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5092 // We require that the compare is in the same block as the phi so that jump
5093 // threading can do its work afterwards.
5094 if (CmpInst->getParent() != PhiBlock)
5097 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5101 Value *RangeCmp = RangeCheckBranch->getCondition();
5102 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5103 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5105 // Check if the compare with the default value is constant true or false.
5106 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5107 DefaultValue, CmpOp1, true);
5108 if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5111 // Check if the compare with the case values is distinct from the default
5113 for (auto ValuePair : Values) {
5114 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5115 ValuePair.second, CmpOp1, true);
5116 if (!CaseConst || CaseConst == DefaultConst)
5118 assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5119 "Expect true or false as compare result.");
5122 // Check if the branch instruction dominates the phi node. It's a simple
5123 // dominance check, but sufficient for our needs.
5124 // Although this check is invariant in the calling loops, it's better to do it
5125 // at this late stage. Practically we do it at most once for a switch.
5126 BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5127 for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5128 BasicBlock *Pred = *PI;
5129 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5133 if (DefaultConst == FalseConst) {
5134 // The compare yields the same result. We can replace it.
5135 CmpInst->replaceAllUsesWith(RangeCmp);
5136 ++NumTableCmpReuses;
5138 // The compare yields the same result, just inverted. We can replace it.
5139 Value *InvertedTableCmp = BinaryOperator::CreateXor(
5140 RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5142 CmpInst->replaceAllUsesWith(InvertedTableCmp);
5143 ++NumTableCmpReuses;
5147 /// If the switch is only used to initialize one or more phi nodes in a common
5148 /// successor block with different constant values, replace the switch with
5150 static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
5151 const DataLayout &DL,
5152 const TargetTransformInfo &TTI) {
5153 assert(SI->getNumCases() > 1 && "Degenerate switch?");
5155 // Only build lookup table when we have a target that supports it.
5156 if (!TTI.shouldBuildLookupTables())
5159 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5160 // split off a dense part and build a lookup table for that.
5162 // FIXME: This creates arrays of GEPs to constant strings, which means each
5163 // GEP needs a runtime relocation in PIC code. We should just build one big
5164 // string and lookup indices into that.
5166 // Ignore switches with less than three cases. Lookup tables will not make
5168 // faster, so we don't analyze them.
5169 if (SI->getNumCases() < 3)
5172 // Figure out the corresponding result for each case value and phi node in the
5173 // common destination, as well as the min and max case values.
5174 assert(SI->case_begin() != SI->case_end());
5175 SwitchInst::CaseIt CI = SI->case_begin();
5176 ConstantInt *MinCaseVal = CI->getCaseValue();
5177 ConstantInt *MaxCaseVal = CI->getCaseValue();
5179 BasicBlock *CommonDest = nullptr;
5180 typedef SmallVector<std::pair<ConstantInt *, Constant *>, 4> ResultListTy;
5181 SmallDenseMap<PHINode *, ResultListTy> ResultLists;
5182 SmallDenseMap<PHINode *, Constant *> DefaultResults;
5183 SmallDenseMap<PHINode *, Type *> ResultTypes;
5184 SmallVector<PHINode *, 4> PHIs;
5186 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5187 ConstantInt *CaseVal = CI->getCaseValue();
5188 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5189 MinCaseVal = CaseVal;
5190 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5191 MaxCaseVal = CaseVal;
5193 // Resulting value at phi nodes for this case value.
5194 typedef SmallVector<std::pair<PHINode *, Constant *>, 4> ResultsTy;
5196 if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5200 // Append the result from this case to the list for each phi.
5201 for (const auto &I : Results) {
5202 PHINode *PHI = I.first;
5203 Constant *Value = I.second;
5204 if (!ResultLists.count(PHI))
5205 PHIs.push_back(PHI);
5206 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5210 // Keep track of the result types.
5211 for (PHINode *PHI : PHIs) {
5212 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5215 uint64_t NumResults = ResultLists[PHIs[0]].size();
5216 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5217 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5218 bool TableHasHoles = (NumResults < TableSize);
5220 // If the table has holes, we need a constant result for the default case
5221 // or a bitmask that fits in a register.
5222 SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5223 bool HasDefaultResults =
5224 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5225 DefaultResultsList, DL, TTI);
5227 bool NeedMask = (TableHasHoles && !HasDefaultResults);
5229 // As an extra penalty for the validity test we require more cases.
5230 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5232 if (!DL.fitsInLegalInteger(TableSize))
5236 for (const auto &I : DefaultResultsList) {
5237 PHINode *PHI = I.first;
5238 Constant *Result = I.second;
5239 DefaultResults[PHI] = Result;
5242 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5245 // Create the BB that does the lookups.
5246 Module &Mod = *CommonDest->getParent()->getParent();
5247 BasicBlock *LookupBB = BasicBlock::Create(
5248 Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5250 // Compute the table index value.
5251 Builder.SetInsertPoint(SI);
5253 Builder.CreateSub(SI->getCondition(), MinCaseVal, "switch.tableidx");
5255 // Compute the maximum table size representable by the integer type we are
5257 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5258 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5259 assert(MaxTableSize >= TableSize &&
5260 "It is impossible for a switch to have more entries than the max "
5261 "representable value of its input integer type's size.");
5263 // If the default destination is unreachable, or if the lookup table covers
5264 // all values of the conditional variable, branch directly to the lookup table
5265 // BB. Otherwise, check that the condition is within the case range.
5266 const bool DefaultIsReachable =
5267 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5268 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5269 BranchInst *RangeCheckBranch = nullptr;
5271 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5272 Builder.CreateBr(LookupBB);
5273 // Note: We call removeProdecessor later since we need to be able to get the
5274 // PHI value for the default case in case we're using a bit mask.
5276 Value *Cmp = Builder.CreateICmpULT(
5277 TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5279 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5282 // Populate the BB that does the lookups.
5283 Builder.SetInsertPoint(LookupBB);
5286 // Before doing the lookup we do the hole check.
5287 // The LookupBB is therefore re-purposed to do the hole check
5288 // and we create a new LookupBB.
5289 BasicBlock *MaskBB = LookupBB;
5290 MaskBB->setName("switch.hole_check");
5291 LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5292 CommonDest->getParent(), CommonDest);
5294 // Make the mask's bitwidth at least 8bit and a power-of-2 to avoid
5295 // unnecessary illegal types.
5296 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5297 APInt MaskInt(TableSizePowOf2, 0);
5298 APInt One(TableSizePowOf2, 1);
5299 // Build bitmask; fill in a 1 bit for every case.
5300 const ResultListTy &ResultList = ResultLists[PHIs[0]];
5301 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5302 uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5304 MaskInt |= One << Idx;
5306 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5308 // Get the TableIndex'th bit of the bitmask.
5309 // If this bit is 0 (meaning hole) jump to the default destination,
5310 // else continue with table lookup.
5311 IntegerType *MapTy = TableMask->getType();
5313 Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5314 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5315 Value *LoBit = Builder.CreateTrunc(
5316 Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5317 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5319 Builder.SetInsertPoint(LookupBB);
5320 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5323 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5324 // We cached PHINodes in PHIs, to avoid accessing deleted PHINodes later,
5325 // do not delete PHINodes here.
5326 SI->getDefaultDest()->removePredecessor(SI->getParent(),
5327 /*DontDeleteUselessPHIs=*/true);
5330 bool ReturnedEarly = false;
5331 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
5332 PHINode *PHI = PHIs[I];
5333 const ResultListTy &ResultList = ResultLists[PHI];
5335 // If using a bitmask, use any value to fill the lookup table holes.
5336 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5337 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL);
5339 Value *Result = Table.BuildLookup(TableIndex, Builder);
5341 // If the result is used to return immediately from the function, we want to
5342 // do that right here.
5343 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5344 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5345 Builder.CreateRet(Result);
5346 ReturnedEarly = true;
5350 // Do a small peephole optimization: re-use the switch table compare if
5352 if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5353 BasicBlock *PhiBlock = PHI->getParent();
5354 // Search for compare instructions which use the phi.
5355 for (auto *User : PHI->users()) {
5356 reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5360 PHI->addIncoming(Result, LookupBB);
5364 Builder.CreateBr(CommonDest);
5366 // Remove the switch.
5367 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5368 BasicBlock *Succ = SI->getSuccessor(i);
5370 if (Succ == SI->getDefaultDest())
5372 Succ->removePredecessor(SI->getParent());
5374 SI->eraseFromParent();
5378 ++NumLookupTablesHoles;
5382 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5383 // See also SelectionDAGBuilder::isDense(), which this function was based on.
5384 uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5385 uint64_t Range = Diff + 1;
5386 uint64_t NumCases = Values.size();
5387 // 40% is the default density for building a jump table in optsize/minsize mode.
5388 uint64_t MinDensity = 40;
5390 return NumCases * 100 >= Range * MinDensity;
5393 // Try and transform a switch that has "holes" in it to a contiguous sequence
5396 // A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5397 // range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5399 // This converts a sparse switch into a dense switch which allows better
5400 // lowering and could also allow transforming into a lookup table.
5401 static bool ReduceSwitchRange(SwitchInst *SI, IRBuilder<> &Builder,
5402 const DataLayout &DL,
5403 const TargetTransformInfo &TTI) {
5404 auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5405 if (CondTy->getIntegerBitWidth() > 64 ||
5406 !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5408 // Only bother with this optimization if there are more than 3 switch cases;
5409 // SDAG will only bother creating jump tables for 4 or more cases.
5410 if (SI->getNumCases() < 4)
5413 // This transform is agnostic to the signedness of the input or case values. We
5414 // can treat the case values as signed or unsigned. We can optimize more common
5415 // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5417 SmallVector<int64_t,4> Values;
5418 for (auto &C : SI->cases())
5419 Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5420 std::sort(Values.begin(), Values.end());
5422 // If the switch is already dense, there's nothing useful to do here.
5423 if (isSwitchDense(Values))
5426 // First, transform the values such that they start at zero and ascend.
5427 int64_t Base = Values[0];
5428 for (auto &V : Values)
5431 // Now we have signed numbers that have been shifted so that, given enough
5432 // precision, there are no negative values. Since the rest of the transform
5433 // is bitwise only, we switch now to an unsigned representation.
5435 for (auto &V : Values)
5436 GCD = GreatestCommonDivisor64(GCD, (uint64_t)V);
5438 // This transform can be done speculatively because it is so cheap - it results
5439 // in a single rotate operation being inserted. This can only happen if the
5440 // factor extracted is a power of 2.
5441 // FIXME: If the GCD is an odd number we can multiply by the multiplicative
5442 // inverse of GCD and then perform this transform.
5443 // FIXME: It's possible that optimizing a switch on powers of two might also
5444 // be beneficial - flag values are often powers of two and we could use a CLZ
5445 // as the key function.
5446 if (GCD <= 1 || !isPowerOf2_64(GCD))
5447 // No common divisor found or too expensive to compute key function.
5450 unsigned Shift = Log2_64(GCD);
5451 for (auto &V : Values)
5452 V = (int64_t)((uint64_t)V >> Shift);
5454 if (!isSwitchDense(Values))
5455 // Transform didn't create a dense switch.
5458 // The obvious transform is to shift the switch condition right and emit a
5459 // check that the condition actually cleanly divided by GCD, i.e.
5460 // C & (1 << Shift - 1) == 0
5461 // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5463 // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5464 // shift and puts the shifted-off bits in the uppermost bits. If any of these
5465 // are nonzero then the switch condition will be very large and will hit the
5468 auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5469 Builder.SetInsertPoint(SI);
5470 auto *ShiftC = ConstantInt::get(Ty, Shift);
5471 auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5472 auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5473 auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5474 auto *Rot = Builder.CreateOr(LShr, Shl);
5475 SI->replaceUsesOfWith(SI->getCondition(), Rot);
5477 for (auto Case : SI->cases()) {
5478 auto *Orig = Case.getCaseValue();
5479 auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5481 cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5486 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
5487 BasicBlock *BB = SI->getParent();
5489 if (isValueEqualityComparison(SI)) {
5490 // If we only have one predecessor, and if it is a branch on this value,
5491 // see if that predecessor totally determines the outcome of this switch.
5492 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5493 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
5494 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5496 Value *Cond = SI->getCondition();
5497 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
5498 if (SimplifySwitchOnSelect(SI, Select))
5499 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5501 // If the block only contains the switch, see if we can fold the block
5502 // away into any preds.
5503 BasicBlock::iterator BBI = BB->begin();
5504 // Ignore dbg intrinsics.
5505 while (isa<DbgInfoIntrinsic>(BBI))
5508 if (FoldValueComparisonIntoPredecessors(SI, Builder))
5509 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5512 // Try to transform the switch into an icmp and a branch.
5513 if (TurnSwitchRangeIntoICmp(SI, Builder))
5514 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5516 // Remove unreachable cases.
5517 if (EliminateDeadSwitchCases(SI, AC, DL))
5518 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5520 if (SwitchToSelect(SI, Builder, AC, DL, TTI))
5521 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5523 if (ForwardSwitchConditionToPHI(SI))
5524 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5526 // The conversion from switch to lookup tables results in difficult
5527 // to analyze code and makes pruning branches much harder.
5528 // This is a problem of the switch expression itself can still be
5529 // restricted as a result of inlining or CVP. There only apply this
5530 // transformation during late steps of the optimisation chain.
5531 if (LateSimplifyCFG && SwitchToLookupTable(SI, Builder, DL, TTI))
5532 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5534 if (ReduceSwitchRange(SI, Builder, DL, TTI))
5535 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5540 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
5541 BasicBlock *BB = IBI->getParent();
5542 bool Changed = false;
5544 // Eliminate redundant destinations.
5545 SmallPtrSet<Value *, 8> Succs;
5546 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
5547 BasicBlock *Dest = IBI->getDestination(i);
5548 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
5549 Dest->removePredecessor(BB);
5550 IBI->removeDestination(i);
5557 if (IBI->getNumDestinations() == 0) {
5558 // If the indirectbr has no successors, change it to unreachable.
5559 new UnreachableInst(IBI->getContext(), IBI);
5560 EraseTerminatorInstAndDCECond(IBI);
5564 if (IBI->getNumDestinations() == 1) {
5565 // If the indirectbr has one successor, change it to a direct branch.
5566 BranchInst::Create(IBI->getDestination(0), IBI);
5567 EraseTerminatorInstAndDCECond(IBI);
5571 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
5572 if (SimplifyIndirectBrOnSelect(IBI, SI))
5573 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5578 /// Given an block with only a single landing pad and a unconditional branch
5579 /// try to find another basic block which this one can be merged with. This
5580 /// handles cases where we have multiple invokes with unique landing pads, but
5581 /// a shared handler.
5583 /// We specifically choose to not worry about merging non-empty blocks
5584 /// here. That is a PRE/scheduling problem and is best solved elsewhere. In
5585 /// practice, the optimizer produces empty landing pad blocks quite frequently
5586 /// when dealing with exception dense code. (see: instcombine, gvn, if-else
5587 /// sinking in this file)
5589 /// This is primarily a code size optimization. We need to avoid performing
5590 /// any transform which might inhibit optimization (such as our ability to
5591 /// specialize a particular handler via tail commoning). We do this by not
5592 /// merging any blocks which require us to introduce a phi. Since the same
5593 /// values are flowing through both blocks, we don't loose any ability to
5594 /// specialize. If anything, we make such specialization more likely.
5596 /// TODO - This transformation could remove entries from a phi in the target
5597 /// block when the inputs in the phi are the same for the two blocks being
5598 /// merged. In some cases, this could result in removal of the PHI entirely.
5599 static bool TryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI,
5601 auto Succ = BB->getUniqueSuccessor();
5603 // If there's a phi in the successor block, we'd likely have to introduce
5604 // a phi into the merged landing pad block.
5605 if (isa<PHINode>(*Succ->begin()))
5608 for (BasicBlock *OtherPred : predecessors(Succ)) {
5609 if (BB == OtherPred)
5611 BasicBlock::iterator I = OtherPred->begin();
5612 LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(I);
5613 if (!LPad2 || !LPad2->isIdenticalTo(LPad))
5615 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {
5617 BranchInst *BI2 = dyn_cast<BranchInst>(I);
5618 if (!BI2 || !BI2->isIdenticalTo(BI))
5621 // We've found an identical block. Update our predecessors to take that
5622 // path instead and make ourselves dead.
5623 SmallSet<BasicBlock *, 16> Preds;
5624 Preds.insert(pred_begin(BB), pred_end(BB));
5625 for (BasicBlock *Pred : Preds) {
5626 InvokeInst *II = cast<InvokeInst>(Pred->getTerminator());
5627 assert(II->getNormalDest() != BB && II->getUnwindDest() == BB &&
5628 "unexpected successor");
5629 II->setUnwindDest(OtherPred);
5632 // The debug info in OtherPred doesn't cover the merged control flow that
5633 // used to go through BB. We need to delete it or update it.
5634 for (auto I = OtherPred->begin(), E = OtherPred->end(); I != E;) {
5635 Instruction &Inst = *I;
5637 if (isa<DbgInfoIntrinsic>(Inst))
5638 Inst.eraseFromParent();
5641 SmallSet<BasicBlock *, 16> Succs;
5642 Succs.insert(succ_begin(BB), succ_end(BB));
5643 for (BasicBlock *Succ : Succs) {
5644 Succ->removePredecessor(BB);
5647 IRBuilder<> Builder(BI);
5648 Builder.CreateUnreachable();
5649 BI->eraseFromParent();
5655 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI,
5656 IRBuilder<> &Builder) {
5657 BasicBlock *BB = BI->getParent();
5659 if (SinkCommon && SinkThenElseCodeToEnd(BI))
5662 // If the Terminator is the only non-phi instruction, simplify the block.
5663 // if LoopHeader is provided, check if the block is a loop header
5664 // (This is for early invocations before loop simplify and vectorization
5665 // to keep canonical loop forms for nested loops.
5666 // These blocks can be eliminated when the pass is invoked later
5667 // in the back-end.)
5668 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg()->getIterator();
5669 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
5670 (!LoopHeaders || !LoopHeaders->count(BB)) &&
5671 TryToSimplifyUncondBranchFromEmptyBlock(BB))
5674 // If the only instruction in the block is a seteq/setne comparison
5675 // against a constant, try to simplify the block.
5676 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
5677 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
5678 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5680 if (I->isTerminator() &&
5681 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, DL, TTI,
5682 BonusInstThreshold, AC))
5686 // See if we can merge an empty landing pad block with another which is
5688 if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(I)) {
5689 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {
5691 if (I->isTerminator() && TryToMergeLandingPad(LPad, BI, BB))
5695 // If this basic block is ONLY a compare and a branch, and if a predecessor
5696 // branches to us and our successor, fold the comparison into the
5697 // predecessor and use logical operations to update the incoming value
5698 // for PHI nodes in common successor.
5699 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
5700 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5704 static BasicBlock *allPredecessorsComeFromSameSource(BasicBlock *BB) {
5705 BasicBlock *PredPred = nullptr;
5706 for (auto *P : predecessors(BB)) {
5707 BasicBlock *PPred = P->getSinglePredecessor();
5708 if (!PPred || (PredPred && PredPred != PPred))
5715 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
5716 BasicBlock *BB = BI->getParent();
5718 // Conditional branch
5719 if (isValueEqualityComparison(BI)) {
5720 // If we only have one predecessor, and if it is a branch on this value,
5721 // see if that predecessor totally determines the outcome of this
5723 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5724 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
5725 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5727 // This block must be empty, except for the setcond inst, if it exists.
5728 // Ignore dbg intrinsics.
5729 BasicBlock::iterator I = BB->begin();
5730 // Ignore dbg intrinsics.
5731 while (isa<DbgInfoIntrinsic>(I))
5734 if (FoldValueComparisonIntoPredecessors(BI, Builder))
5735 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5736 } else if (&*I == cast<Instruction>(BI->getCondition())) {
5738 // Ignore dbg intrinsics.
5739 while (isa<DbgInfoIntrinsic>(I))
5741 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
5742 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5746 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
5747 if (SimplifyBranchOnICmpChain(BI, Builder, DL))
5750 // If this basic block has a single dominating predecessor block and the
5751 // dominating block's condition implies BI's condition, we know the direction
5752 // of the BI branch.
5753 if (BasicBlock *Dom = BB->getSinglePredecessor()) {
5754 auto *PBI = dyn_cast_or_null<BranchInst>(Dom->getTerminator());
5755 if (PBI && PBI->isConditional() &&
5756 PBI->getSuccessor(0) != PBI->getSuccessor(1) &&
5757 (PBI->getSuccessor(0) == BB || PBI->getSuccessor(1) == BB)) {
5758 bool CondIsFalse = PBI->getSuccessor(1) == BB;
5759 Optional<bool> Implication = isImpliedCondition(
5760 PBI->getCondition(), BI->getCondition(), DL, CondIsFalse);
5762 // Turn this into a branch on constant.
5763 auto *OldCond = BI->getCondition();
5764 ConstantInt *CI = *Implication
5765 ? ConstantInt::getTrue(BB->getContext())
5766 : ConstantInt::getFalse(BB->getContext());
5767 BI->setCondition(CI);
5768 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
5769 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5774 // If this basic block is ONLY a compare and a branch, and if a predecessor
5775 // branches to us and one of our successors, fold the comparison into the
5776 // predecessor and use logical operations to pick the right destination.
5777 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
5778 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5780 // We have a conditional branch to two blocks that are only reachable
5781 // from BI. We know that the condbr dominates the two blocks, so see if
5782 // there is any identical code in the "then" and "else" blocks. If so, we
5783 // can hoist it up to the branching block.
5784 if (BI->getSuccessor(0)->getSinglePredecessor()) {
5785 if (BI->getSuccessor(1)->getSinglePredecessor()) {
5786 if (HoistThenElseCodeToIf(BI, TTI))
5787 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5789 // If Successor #1 has multiple preds, we may be able to conditionally
5790 // execute Successor #0 if it branches to Successor #1.
5791 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
5792 if (Succ0TI->getNumSuccessors() == 1 &&
5793 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
5794 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), TTI))
5795 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5797 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
5798 // If Successor #0 has multiple preds, we may be able to conditionally
5799 // execute Successor #1 if it branches to Successor #0.
5800 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
5801 if (Succ1TI->getNumSuccessors() == 1 &&
5802 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
5803 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), TTI))
5804 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5807 // If this is a branch on a phi node in the current block, thread control
5808 // through this block if any PHI node entries are constants.
5809 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
5810 if (PN->getParent() == BI->getParent())
5811 if (FoldCondBranchOnPHI(BI, DL, AC))
5812 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5814 // Scan predecessor blocks for conditional branches.
5815 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
5816 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
5817 if (PBI != BI && PBI->isConditional())
5818 if (SimplifyCondBranchToCondBranch(PBI, BI, DL))
5819 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5821 // Look for diamond patterns.
5822 if (MergeCondStores)
5823 if (BasicBlock *PrevBB = allPredecessorsComeFromSameSource(BB))
5824 if (BranchInst *PBI = dyn_cast<BranchInst>(PrevBB->getTerminator()))
5825 if (PBI != BI && PBI->isConditional())
5826 if (mergeConditionalStores(PBI, BI))
5827 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5832 /// Check if passing a value to an instruction will cause undefined behavior.
5833 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
5834 Constant *C = dyn_cast<Constant>(V);
5841 if (C->isNullValue() || isa<UndefValue>(C)) {
5842 // Only look at the first use, avoid hurting compile time with long uselists
5843 User *Use = *I->user_begin();
5845 // Now make sure that there are no instructions in between that can alter
5846 // control flow (eg. calls)
5847 for (BasicBlock::iterator
5848 i = ++BasicBlock::iterator(I),
5849 UI = BasicBlock::iterator(dyn_cast<Instruction>(Use));
5851 if (i == I->getParent()->end() || i->mayHaveSideEffects())
5854 // Look through GEPs. A load from a GEP derived from NULL is still undefined
5855 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
5856 if (GEP->getPointerOperand() == I)
5857 return passingValueIsAlwaysUndefined(V, GEP);
5859 // Look through bitcasts.
5860 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
5861 return passingValueIsAlwaysUndefined(V, BC);
5863 // Load from null is undefined.
5864 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
5865 if (!LI->isVolatile())
5866 return LI->getPointerAddressSpace() == 0;
5868 // Store to null is undefined.
5869 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
5870 if (!SI->isVolatile())
5871 return SI->getPointerAddressSpace() == 0 &&
5872 SI->getPointerOperand() == I;
5874 // A call to null is undefined.
5875 if (auto CS = CallSite(Use))
5876 return CS.getCalledValue() == I;
5881 /// If BB has an incoming value that will always trigger undefined behavior
5882 /// (eg. null pointer dereference), remove the branch leading here.
5883 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
5884 for (BasicBlock::iterator i = BB->begin();
5885 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
5886 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
5887 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
5888 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
5889 IRBuilder<> Builder(T);
5890 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
5891 BB->removePredecessor(PHI->getIncomingBlock(i));
5892 // Turn uncoditional branches into unreachables and remove the dead
5893 // destination from conditional branches.
5894 if (BI->isUnconditional())
5895 Builder.CreateUnreachable();
5897 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1)
5898 : BI->getSuccessor(0));
5899 BI->eraseFromParent();
5902 // TODO: SwitchInst.
5908 bool SimplifyCFGOpt::run(BasicBlock *BB) {
5909 bool Changed = false;
5911 assert(BB && BB->getParent() && "Block not embedded in function!");
5912 assert(BB->getTerminator() && "Degenerate basic block encountered!");
5914 // Remove basic blocks that have no predecessors (except the entry block)...
5915 // or that just have themself as a predecessor. These are unreachable.
5916 if ((pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) ||
5917 BB->getSinglePredecessor() == BB) {
5918 DEBUG(dbgs() << "Removing BB: \n" << *BB);
5919 DeleteDeadBlock(BB);
5923 // Check to see if we can constant propagate this terminator instruction
5925 Changed |= ConstantFoldTerminator(BB, true);
5927 // Check for and eliminate duplicate PHI nodes in this block.
5928 Changed |= EliminateDuplicatePHINodes(BB);
5930 // Check for and remove branches that will always cause undefined behavior.
5931 Changed |= removeUndefIntroducingPredecessor(BB);
5933 // Merge basic blocks into their predecessor if there is only one distinct
5934 // pred, and if there is only one distinct successor of the predecessor, and
5935 // if there are no PHI nodes.
5937 if (MergeBlockIntoPredecessor(BB))
5940 IRBuilder<> Builder(BB);
5942 // If there is a trivial two-entry PHI node in this basic block, and we can
5943 // eliminate it, do so now.
5944 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
5945 if (PN->getNumIncomingValues() == 2)
5946 Changed |= FoldTwoEntryPHINode(PN, TTI, DL);
5948 Builder.SetInsertPoint(BB->getTerminator());
5949 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
5950 if (BI->isUnconditional()) {
5951 if (SimplifyUncondBranch(BI, Builder))
5954 if (SimplifyCondBranch(BI, Builder))
5957 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
5958 if (SimplifyReturn(RI, Builder))
5960 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
5961 if (SimplifyResume(RI, Builder))
5963 } else if (CleanupReturnInst *RI =
5964 dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
5965 if (SimplifyCleanupReturn(RI))
5967 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
5968 if (SimplifySwitch(SI, Builder))
5970 } else if (UnreachableInst *UI =
5971 dyn_cast<UnreachableInst>(BB->getTerminator())) {
5972 if (SimplifyUnreachable(UI))
5974 } else if (IndirectBrInst *IBI =
5975 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
5976 if (SimplifyIndirectBr(IBI))
5983 /// This function is used to do simplification of a CFG.
5984 /// For example, it adjusts branches to branches to eliminate the extra hop,
5985 /// eliminates unreachable basic blocks, and does other "peephole" optimization
5986 /// of the CFG. It returns true if a modification was made.
5988 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
5989 unsigned BonusInstThreshold, AssumptionCache *AC,
5990 SmallPtrSetImpl<BasicBlock *> *LoopHeaders,
5991 bool LateSimplifyCFG) {
5992 return SimplifyCFGOpt(TTI, BB->getModule()->getDataLayout(),
5993 BonusInstThreshold, AC, LoopHeaders, LateSimplifyCFG)