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.getSetSize().ugt(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 // Is it legal to place a variable in operand \c OpIdx of \c I?
1380 // FIXME: This should be promoted to Instruction.
1381 static bool canReplaceOperandWithVariable(const Instruction *I,
1383 // We can't have a PHI with a metadata type.
1384 if (I->getOperand(OpIdx)->getType()->isMetadataTy())
1388 if (!isa<Constant>(I->getOperand(OpIdx)))
1391 switch (I->getOpcode()) {
1394 case Instruction::Call:
1395 case Instruction::Invoke:
1396 // FIXME: many arithmetic intrinsics have no issue taking a
1397 // variable, however it's hard to distingish these from
1398 // specials such as @llvm.frameaddress that require a constant.
1399 if (isa<IntrinsicInst>(I))
1402 // Constant bundle operands may need to retain their constant-ness for
1404 if (ImmutableCallSite(I).isBundleOperand(OpIdx))
1409 case Instruction::ShuffleVector:
1410 // Shufflevector masks are constant.
1412 case Instruction::ExtractValue:
1413 case Instruction::InsertValue:
1414 // All operands apart from the first are constant.
1416 case Instruction::Alloca:
1418 case Instruction::GetElementPtr:
1421 gep_type_iterator It = std::next(gep_type_begin(I), OpIdx - 1);
1422 return It.isSequential();
1426 // All instructions in Insts belong to different blocks that all unconditionally
1427 // branch to a common successor. Analyze each instruction and return true if it
1428 // would be possible to sink them into their successor, creating one common
1429 // instruction instead. For every value that would be required to be provided by
1430 // PHI node (because an operand varies in each input block), add to PHIOperands.
1431 static bool canSinkInstructions(
1432 ArrayRef<Instruction *> Insts,
1433 DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1434 // Prune out obviously bad instructions to move. Any non-store instruction
1435 // must have exactly one use, and we check later that use is by a single,
1436 // common PHI instruction in the successor.
1437 for (auto *I : Insts) {
1438 // These instructions may change or break semantics if moved.
1439 if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1440 I->getType()->isTokenTy())
1443 // Conservatively return false if I is an inline-asm instruction. Sinking
1444 // and merging inline-asm instructions can potentially create arguments
1445 // that cannot satisfy the inline-asm constraints.
1446 if (const auto *C = dyn_cast<CallInst>(I))
1447 if (C->isInlineAsm())
1450 // Everything must have only one use too, apart from stores which
1452 if (!isa<StoreInst>(I) && !I->hasOneUse())
1456 const Instruction *I0 = Insts.front();
1457 for (auto *I : Insts)
1458 if (!I->isSameOperationAs(I0))
1461 // All instructions in Insts are known to be the same opcode. If they aren't
1462 // stores, check the only user of each is a PHI or in the same block as the
1463 // instruction, because if a user is in the same block as an instruction
1464 // we're contemplating sinking, it must already be determined to be sinkable.
1465 if (!isa<StoreInst>(I0)) {
1466 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1467 auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1468 if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1469 auto *U = cast<Instruction>(*I->user_begin());
1471 PNUse->getParent() == Succ &&
1472 PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1473 U->getParent() == I->getParent();
1478 // Because SROA can't handle speculating stores of selects, try not
1479 // to sink loads or stores of allocas when we'd have to create a PHI for
1480 // the address operand. Also, because it is likely that loads or stores
1481 // of allocas will disappear when Mem2Reg/SROA is run, don't sink them.
1482 // This can cause code churn which can have unintended consequences down
1483 // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1484 // FIXME: This is a workaround for a deficiency in SROA - see
1485 // https://llvm.org/bugs/show_bug.cgi?id=30188
1486 if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1487 return isa<AllocaInst>(I->getOperand(1));
1490 if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1491 return isa<AllocaInst>(I->getOperand(0));
1495 for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1496 if (I0->getOperand(OI)->getType()->isTokenTy())
1497 // Don't touch any operand of token type.
1500 auto SameAsI0 = [&I0, OI](const Instruction *I) {
1501 assert(I->getNumOperands() == I0->getNumOperands());
1502 return I->getOperand(OI) == I0->getOperand(OI);
1504 if (!all_of(Insts, SameAsI0)) {
1505 if (!canReplaceOperandWithVariable(I0, OI))
1506 // We can't create a PHI from this GEP.
1508 // Don't create indirect calls! The called value is the final operand.
1509 if ((isa<CallInst>(I0) || isa<InvokeInst>(I0)) && OI == OE - 1) {
1510 // FIXME: if the call was *already* indirect, we should do this.
1513 for (auto *I : Insts)
1514 PHIOperands[I].push_back(I->getOperand(OI));
1520 // Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1521 // instruction of every block in Blocks to their common successor, commoning
1522 // into one instruction.
1523 static bool sinkLastInstruction(ArrayRef<BasicBlock*> Blocks) {
1524 auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1526 // canSinkLastInstruction returning true guarantees that every block has at
1527 // least one non-terminator instruction.
1528 SmallVector<Instruction*,4> Insts;
1529 for (auto *BB : Blocks) {
1530 Instruction *I = BB->getTerminator();
1532 I = I->getPrevNode();
1533 } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1534 if (!isa<DbgInfoIntrinsic>(I))
1538 // The only checking we need to do now is that all users of all instructions
1539 // are the same PHI node. canSinkLastInstruction should have checked this but
1540 // it is slightly over-aggressive - it gets confused by commutative instructions
1541 // so double-check it here.
1542 Instruction *I0 = Insts.front();
1543 if (!isa<StoreInst>(I0)) {
1544 auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1545 if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1546 auto *U = cast<Instruction>(*I->user_begin());
1552 // We don't need to do any more checking here; canSinkLastInstruction should
1553 // have done it all for us.
1554 SmallVector<Value*, 4> NewOperands;
1555 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1556 // This check is different to that in canSinkLastInstruction. There, we
1557 // cared about the global view once simplifycfg (and instcombine) have
1558 // completed - it takes into account PHIs that become trivially
1559 // simplifiable. However here we need a more local view; if an operand
1560 // differs we create a PHI and rely on instcombine to clean up the very
1561 // small mess we may make.
1562 bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1563 return I->getOperand(O) != I0->getOperand(O);
1566 NewOperands.push_back(I0->getOperand(O));
1570 // Create a new PHI in the successor block and populate it.
1571 auto *Op = I0->getOperand(O);
1572 assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1573 auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1574 Op->getName() + ".sink", &BBEnd->front());
1575 for (auto *I : Insts)
1576 PN->addIncoming(I->getOperand(O), I->getParent());
1577 NewOperands.push_back(PN);
1580 // Arbitrarily use I0 as the new "common" instruction; remap its operands
1581 // and move it to the start of the successor block.
1582 for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1583 I0->getOperandUse(O).set(NewOperands[O]);
1584 I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1586 // The debug location for the "common" instruction is the merged locations of
1587 // all the commoned instructions. We start with the original location of the
1588 // "common" instruction and iteratively merge each location in the loop below.
1589 const DILocation *Loc = I0->getDebugLoc();
1591 // Update metadata and IR flags, and merge debug locations.
1592 for (auto *I : Insts)
1594 Loc = DILocation::getMergedLocation(Loc, I->getDebugLoc());
1595 combineMetadataForCSE(I0, I);
1598 if (!isa<CallInst>(I0))
1599 I0->setDebugLoc(Loc);
1601 if (!isa<StoreInst>(I0)) {
1602 // canSinkLastInstruction checked that all instructions were used by
1603 // one and only one PHI node. Find that now, RAUW it to our common
1604 // instruction and nuke it.
1605 assert(I0->hasOneUse());
1606 auto *PN = cast<PHINode>(*I0->user_begin());
1607 PN->replaceAllUsesWith(I0);
1608 PN->eraseFromParent();
1611 // Finally nuke all instructions apart from the common instruction.
1612 for (auto *I : Insts)
1614 I->eraseFromParent();
1621 // LockstepReverseIterator - Iterates through instructions
1622 // in a set of blocks in reverse order from the first non-terminator.
1623 // For example (assume all blocks have size n):
1624 // LockstepReverseIterator I([B1, B2, B3]);
1625 // *I-- = [B1[n], B2[n], B3[n]];
1626 // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1627 // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1629 class LockstepReverseIterator {
1630 ArrayRef<BasicBlock*> Blocks;
1631 SmallVector<Instruction*,4> Insts;
1634 LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) :
1642 for (auto *BB : Blocks) {
1643 Instruction *Inst = BB->getTerminator();
1644 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1645 Inst = Inst->getPrevNode();
1647 // Block wasn't big enough.
1651 Insts.push_back(Inst);
1655 bool isValid() const {
1659 void operator -- () {
1662 for (auto *&Inst : Insts) {
1663 for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1664 Inst = Inst->getPrevNode();
1665 // Already at beginning of block.
1673 ArrayRef<Instruction*> operator * () const {
1678 } // end anonymous namespace
1680 /// Given an unconditional branch that goes to BBEnd,
1681 /// check whether BBEnd has only two predecessors and the other predecessor
1682 /// ends with an unconditional branch. If it is true, sink any common code
1683 /// in the two predecessors to BBEnd.
1684 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1685 assert(BI1->isUnconditional());
1686 BasicBlock *BBEnd = BI1->getSuccessor(0);
1688 // We support two situations:
1689 // (1) all incoming arcs are unconditional
1690 // (2) one incoming arc is conditional
1692 // (2) is very common in switch defaults and
1693 // else-if patterns;
1696 // else if (b) f(2);
1709 // [end] has two unconditional predecessor arcs and one conditional. The
1710 // conditional refers to the implicit empty 'else' arc. This conditional
1711 // arc can also be caused by an empty default block in a switch.
1713 // In this case, we attempt to sink code from all *unconditional* arcs.
1714 // If we can sink instructions from these arcs (determined during the scan
1715 // phase below) we insert a common successor for all unconditional arcs and
1716 // connect that to [end], to enable sinking:
1729 SmallVector<BasicBlock*,4> UnconditionalPreds;
1730 Instruction *Cond = nullptr;
1731 for (auto *B : predecessors(BBEnd)) {
1732 auto *T = B->getTerminator();
1733 if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1734 UnconditionalPreds.push_back(B);
1735 else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1740 if (UnconditionalPreds.size() < 2)
1743 bool Changed = false;
1744 // We take a two-step approach to tail sinking. First we scan from the end of
1745 // each block upwards in lockstep. If the n'th instruction from the end of each
1746 // block can be sunk, those instructions are added to ValuesToSink and we
1747 // carry on. If we can sink an instruction but need to PHI-merge some operands
1748 // (because they're not identical in each instruction) we add these to
1750 unsigned ScanIdx = 0;
1751 SmallPtrSet<Value*,4> InstructionsToSink;
1752 DenseMap<Instruction*, SmallVector<Value*,4>> PHIOperands;
1753 LockstepReverseIterator LRI(UnconditionalPreds);
1754 while (LRI.isValid() &&
1755 canSinkInstructions(*LRI, PHIOperands)) {
1756 DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0] << "\n");
1757 InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1762 auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1763 unsigned NumPHIdValues = 0;
1764 for (auto *I : *LRI)
1765 for (auto *V : PHIOperands[I])
1766 if (InstructionsToSink.count(V) == 0)
1768 DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1769 unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1770 if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1773 return NumPHIInsts <= 1;
1776 if (ScanIdx > 0 && Cond) {
1777 // Check if we would actually sink anything first! This mutates the CFG and
1778 // adds an extra block. The goal in doing this is to allow instructions that
1779 // couldn't be sunk before to be sunk - obviously, speculatable instructions
1780 // (such as trunc, add) can be sunk and predicated already. So we check that
1781 // we're going to sink at least one non-speculatable instruction.
1784 bool Profitable = false;
1785 while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1786 if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1796 DEBUG(dbgs() << "SINK: Splitting edge\n");
1797 // We have a conditional edge and we're going to sink some instructions.
1798 // Insert a new block postdominating all blocks we're going to sink from.
1799 if (!SplitBlockPredecessors(BI1->getSuccessor(0), UnconditionalPreds,
1801 // Edges couldn't be split.
1806 // Now that we've analyzed all potential sinking candidates, perform the
1807 // actual sink. We iteratively sink the last non-terminator of the source
1808 // blocks into their common successor unless doing so would require too
1809 // many PHI instructions to be generated (currently only one PHI is allowed
1810 // per sunk instruction).
1812 // We can use InstructionsToSink to discount values needing PHI-merging that will
1813 // actually be sunk in a later iteration. This allows us to be more
1814 // aggressive in what we sink. This does allow a false positive where we
1815 // sink presuming a later value will also be sunk, but stop half way through
1816 // and never actually sink it which means we produce more PHIs than intended.
1817 // This is unlikely in practice though.
1818 for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1819 DEBUG(dbgs() << "SINK: Sink: "
1820 << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1823 // Because we've sunk every instruction in turn, the current instruction to
1824 // sink is always at index 0.
1826 if (!ProfitableToSinkInstruction(LRI)) {
1827 // Too many PHIs would be created.
1828 DEBUG(dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1832 if (!sinkLastInstruction(UnconditionalPreds))
1840 /// \brief Determine if we can hoist sink a sole store instruction out of a
1841 /// conditional block.
1843 /// We are looking for code like the following:
1845 /// store i32 %add, i32* %arrayidx2
1846 /// ... // No other stores or function calls (we could be calling a memory
1847 /// ... // function).
1848 /// %cmp = icmp ult %x, %y
1849 /// br i1 %cmp, label %EndBB, label %ThenBB
1851 /// store i32 %add5, i32* %arrayidx2
1855 /// We are going to transform this into:
1857 /// store i32 %add, i32* %arrayidx2
1859 /// %cmp = icmp ult %x, %y
1860 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1861 /// store i32 %add.add5, i32* %arrayidx2
1864 /// \return The pointer to the value of the previous store if the store can be
1865 /// hoisted into the predecessor block. 0 otherwise.
1866 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1867 BasicBlock *StoreBB, BasicBlock *EndBB) {
1868 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1872 // Volatile or atomic.
1873 if (!StoreToHoist->isSimple())
1876 Value *StorePtr = StoreToHoist->getPointerOperand();
1878 // Look for a store to the same pointer in BrBB.
1879 unsigned MaxNumInstToLookAt = 9;
1880 for (Instruction &CurI : reverse(*BrBB)) {
1881 if (!MaxNumInstToLookAt)
1884 if (isa<DbgInfoIntrinsic>(CurI))
1886 --MaxNumInstToLookAt;
1888 // Could be calling an instruction that affects memory like free().
1889 if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1892 if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1893 // Found the previous store make sure it stores to the same location.
1894 if (SI->getPointerOperand() == StorePtr)
1895 // Found the previous store, return its value operand.
1896 return SI->getValueOperand();
1897 return nullptr; // Unknown store.
1904 /// \brief Speculate a conditional basic block flattening the CFG.
1906 /// Note that this is a very risky transform currently. Speculating
1907 /// instructions like this is most often not desirable. Instead, there is an MI
1908 /// pass which can do it with full awareness of the resource constraints.
1909 /// However, some cases are "obvious" and we should do directly. An example of
1910 /// this is speculating a single, reasonably cheap instruction.
1912 /// There is only one distinct advantage to flattening the CFG at the IR level:
1913 /// it makes very common but simplistic optimizations such as are common in
1914 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1915 /// modeling their effects with easier to reason about SSA value graphs.
1918 /// An illustration of this transform is turning this IR:
1921 /// %cmp = icmp ult %x, %y
1922 /// br i1 %cmp, label %EndBB, label %ThenBB
1924 /// %sub = sub %x, %y
1927 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1934 /// %cmp = icmp ult %x, %y
1935 /// %sub = sub %x, %y
1936 /// %cond = select i1 %cmp, 0, %sub
1940 /// \returns true if the conditional block is removed.
1941 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1942 const TargetTransformInfo &TTI) {
1943 // Be conservative for now. FP select instruction can often be expensive.
1944 Value *BrCond = BI->getCondition();
1945 if (isa<FCmpInst>(BrCond))
1948 BasicBlock *BB = BI->getParent();
1949 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1951 // If ThenBB is actually on the false edge of the conditional branch, remember
1952 // to swap the select operands later.
1953 bool Invert = false;
1954 if (ThenBB != BI->getSuccessor(0)) {
1955 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1958 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1960 // Keep a count of how many times instructions are used within CondBB when
1961 // they are candidates for sinking into CondBB. Specifically:
1962 // - They are defined in BB, and
1963 // - They have no side effects, and
1964 // - All of their uses are in CondBB.
1965 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1967 unsigned SpeculationCost = 0;
1968 Value *SpeculatedStoreValue = nullptr;
1969 StoreInst *SpeculatedStore = nullptr;
1970 for (BasicBlock::iterator BBI = ThenBB->begin(),
1971 BBE = std::prev(ThenBB->end());
1972 BBI != BBE; ++BBI) {
1973 Instruction *I = &*BBI;
1975 if (isa<DbgInfoIntrinsic>(I))
1978 // Only speculatively execute a single instruction (not counting the
1979 // terminator) for now.
1981 if (SpeculationCost > 1)
1984 // Don't hoist the instruction if it's unsafe or expensive.
1985 if (!isSafeToSpeculativelyExecute(I) &&
1986 !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1987 I, BB, ThenBB, EndBB))))
1989 if (!SpeculatedStoreValue &&
1990 ComputeSpeculationCost(I, TTI) >
1991 PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic)
1994 // Store the store speculation candidate.
1995 if (SpeculatedStoreValue)
1996 SpeculatedStore = cast<StoreInst>(I);
1998 // Do not hoist the instruction if any of its operands are defined but not
1999 // used in BB. The transformation will prevent the operand from
2000 // being sunk into the use block.
2001 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
2002 Instruction *OpI = dyn_cast<Instruction>(*i);
2003 if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
2004 continue; // Not a candidate for sinking.
2006 ++SinkCandidateUseCounts[OpI];
2010 // Consider any sink candidates which are only used in CondBB as costs for
2011 // speculation. Note, while we iterate over a DenseMap here, we are summing
2012 // and so iteration order isn't significant.
2013 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator
2014 I = SinkCandidateUseCounts.begin(),
2015 E = SinkCandidateUseCounts.end();
2017 if (I->first->getNumUses() == I->second) {
2019 if (SpeculationCost > 1)
2023 // Check that the PHI nodes can be converted to selects.
2024 bool HaveRewritablePHIs = false;
2025 for (BasicBlock::iterator I = EndBB->begin();
2026 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2027 Value *OrigV = PN->getIncomingValueForBlock(BB);
2028 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
2030 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2031 // Skip PHIs which are trivial.
2035 // Don't convert to selects if we could remove undefined behavior instead.
2036 if (passingValueIsAlwaysUndefined(OrigV, PN) ||
2037 passingValueIsAlwaysUndefined(ThenV, PN))
2040 HaveRewritablePHIs = true;
2041 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2042 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2043 if (!OrigCE && !ThenCE)
2044 continue; // Known safe and cheap.
2046 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2047 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2049 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2050 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2052 2 * PHINodeFoldingThreshold * TargetTransformInfo::TCC_Basic;
2053 if (OrigCost + ThenCost > MaxCost)
2056 // Account for the cost of an unfolded ConstantExpr which could end up
2057 // getting expanded into Instructions.
2058 // FIXME: This doesn't account for how many operations are combined in the
2059 // constant expression.
2061 if (SpeculationCost > 1)
2065 // If there are no PHIs to process, bail early. This helps ensure idempotence
2067 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2070 // If we get here, we can hoist the instruction and if-convert.
2071 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2073 // Insert a select of the value of the speculated store.
2074 if (SpeculatedStoreValue) {
2075 IRBuilder<NoFolder> Builder(BI);
2076 Value *TrueV = SpeculatedStore->getValueOperand();
2077 Value *FalseV = SpeculatedStoreValue;
2079 std::swap(TrueV, FalseV);
2080 Value *S = Builder.CreateSelect(
2081 BrCond, TrueV, FalseV, TrueV->getName() + "." + FalseV->getName(), BI);
2082 SpeculatedStore->setOperand(0, S);
2083 SpeculatedStore->setDebugLoc(
2084 DILocation::getMergedLocation(
2085 BI->getDebugLoc(), SpeculatedStore->getDebugLoc()));
2088 // Metadata can be dependent on the condition we are hoisting above.
2089 // Conservatively strip all metadata on the instruction.
2090 for (auto &I : *ThenBB)
2091 I.dropUnknownNonDebugMetadata();
2093 // Hoist the instructions.
2094 BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2095 ThenBB->begin(), std::prev(ThenBB->end()));
2097 // Insert selects and rewrite the PHI operands.
2098 IRBuilder<NoFolder> Builder(BI);
2099 for (BasicBlock::iterator I = EndBB->begin();
2100 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2101 unsigned OrigI = PN->getBasicBlockIndex(BB);
2102 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
2103 Value *OrigV = PN->getIncomingValue(OrigI);
2104 Value *ThenV = PN->getIncomingValue(ThenI);
2106 // Skip PHIs which are trivial.
2110 // Create a select whose true value is the speculatively executed value and
2111 // false value is the preexisting value. Swap them if the branch
2112 // destinations were inverted.
2113 Value *TrueV = ThenV, *FalseV = OrigV;
2115 std::swap(TrueV, FalseV);
2116 Value *V = Builder.CreateSelect(
2117 BrCond, TrueV, FalseV, TrueV->getName() + "." + FalseV->getName(), BI);
2118 PN->setIncomingValue(OrigI, V);
2119 PN->setIncomingValue(ThenI, V);
2126 /// Return true if we can thread a branch across this block.
2127 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
2128 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
2131 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2132 if (isa<DbgInfoIntrinsic>(BBI))
2135 return false; // Don't clone large BB's.
2138 // We can only support instructions that do not define values that are
2139 // live outside of the current basic block.
2140 for (User *U : BBI->users()) {
2141 Instruction *UI = cast<Instruction>(U);
2142 if (UI->getParent() != BB || isa<PHINode>(UI))
2146 // Looks ok, continue checking.
2152 /// If we have a conditional branch on a PHI node value that is defined in the
2153 /// same block as the branch and if any PHI entries are constants, thread edges
2154 /// corresponding to that entry to be branches to their ultimate destination.
2155 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL,
2156 AssumptionCache *AC) {
2157 BasicBlock *BB = BI->getParent();
2158 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2159 // NOTE: we currently cannot transform this case if the PHI node is used
2160 // outside of the block.
2161 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2164 // Degenerate case of a single entry PHI.
2165 if (PN->getNumIncomingValues() == 1) {
2166 FoldSingleEntryPHINodes(PN->getParent());
2170 // Now we know that this block has multiple preds and two succs.
2171 if (!BlockIsSimpleEnoughToThreadThrough(BB))
2174 // Can't fold blocks that contain noduplicate or convergent calls.
2175 if (any_of(*BB, [](const Instruction &I) {
2176 const CallInst *CI = dyn_cast<CallInst>(&I);
2177 return CI && (CI->cannotDuplicate() || CI->isConvergent());
2181 // Okay, this is a simple enough basic block. See if any phi values are
2183 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2184 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
2185 if (!CB || !CB->getType()->isIntegerTy(1))
2188 // Okay, we now know that all edges from PredBB should be revectored to
2189 // branch to RealDest.
2190 BasicBlock *PredBB = PN->getIncomingBlock(i);
2191 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2194 continue; // Skip self loops.
2195 // Skip if the predecessor's terminator is an indirect branch.
2196 if (isa<IndirectBrInst>(PredBB->getTerminator()))
2199 // The dest block might have PHI nodes, other predecessors and other
2200 // difficult cases. Instead of being smart about this, just insert a new
2201 // block that jumps to the destination block, effectively splitting
2202 // the edge we are about to create.
2203 BasicBlock *EdgeBB =
2204 BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2205 RealDest->getParent(), RealDest);
2206 BranchInst::Create(RealDest, EdgeBB);
2208 // Update PHI nodes.
2209 AddPredecessorToBlock(RealDest, EdgeBB, BB);
2211 // BB may have instructions that are being threaded over. Clone these
2212 // instructions into EdgeBB. We know that there will be no uses of the
2213 // cloned instructions outside of EdgeBB.
2214 BasicBlock::iterator InsertPt = EdgeBB->begin();
2215 DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2216 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2217 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2218 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2221 // Clone the instruction.
2222 Instruction *N = BBI->clone();
2224 N->setName(BBI->getName() + ".c");
2226 // Update operands due to translation.
2227 for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2228 DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2229 if (PI != TranslateMap.end())
2233 // Check for trivial simplification.
2234 if (Value *V = SimplifyInstruction(N, DL)) {
2235 if (!BBI->use_empty())
2236 TranslateMap[&*BBI] = V;
2237 if (!N->mayHaveSideEffects()) {
2238 delete N; // Instruction folded away, don't need actual inst
2242 if (!BBI->use_empty())
2243 TranslateMap[&*BBI] = N;
2245 // Insert the new instruction into its new home.
2247 EdgeBB->getInstList().insert(InsertPt, N);
2249 // Register the new instruction with the assumption cache if necessary.
2250 if (auto *II = dyn_cast_or_null<IntrinsicInst>(N))
2251 if (II->getIntrinsicID() == Intrinsic::assume)
2252 AC->registerAssumption(II);
2255 // Loop over all of the edges from PredBB to BB, changing them to branch
2256 // to EdgeBB instead.
2257 TerminatorInst *PredBBTI = PredBB->getTerminator();
2258 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2259 if (PredBBTI->getSuccessor(i) == BB) {
2260 BB->removePredecessor(PredBB);
2261 PredBBTI->setSuccessor(i, EdgeBB);
2264 // Recurse, simplifying any other constants.
2265 return FoldCondBranchOnPHI(BI, DL, AC) | true;
2271 /// Given a BB that starts with the specified two-entry PHI node,
2272 /// see if we can eliminate it.
2273 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetTransformInfo &TTI,
2274 const DataLayout &DL) {
2275 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2276 // statement", which has a very simple dominance structure. Basically, we
2277 // are trying to find the condition that is being branched on, which
2278 // subsequently causes this merge to happen. We really want control
2279 // dependence information for this check, but simplifycfg can't keep it up
2280 // to date, and this catches most of the cases we care about anyway.
2281 BasicBlock *BB = PN->getParent();
2282 BasicBlock *IfTrue, *IfFalse;
2283 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2285 // Don't bother if the branch will be constant folded trivially.
2286 isa<ConstantInt>(IfCond))
2289 // Okay, we found that we can merge this two-entry phi node into a select.
2290 // Doing so would require us to fold *all* two entry phi nodes in this block.
2291 // At some point this becomes non-profitable (particularly if the target
2292 // doesn't support cmov's). Only do this transformation if there are two or
2293 // fewer PHI nodes in this block.
2294 unsigned NumPhis = 0;
2295 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2299 // Loop over the PHI's seeing if we can promote them all to select
2300 // instructions. While we are at it, keep track of the instructions
2301 // that need to be moved to the dominating block.
2302 SmallPtrSet<Instruction *, 4> AggressiveInsts;
2303 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2304 MaxCostVal1 = PHINodeFoldingThreshold;
2305 MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2306 MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2308 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2309 PHINode *PN = cast<PHINode>(II++);
2310 if (Value *V = SimplifyInstruction(PN, DL)) {
2311 PN->replaceAllUsesWith(V);
2312 PN->eraseFromParent();
2316 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
2317 MaxCostVal0, TTI) ||
2318 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
2323 // If we folded the first phi, PN dangles at this point. Refresh it. If
2324 // we ran out of PHIs then we simplified them all.
2325 PN = dyn_cast<PHINode>(BB->begin());
2329 // Don't fold i1 branches on PHIs which contain binary operators. These can
2330 // often be turned into switches and other things.
2331 if (PN->getType()->isIntegerTy(1) &&
2332 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2333 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2334 isa<BinaryOperator>(IfCond)))
2337 // If all PHI nodes are promotable, check to make sure that all instructions
2338 // in the predecessor blocks can be promoted as well. If not, we won't be able
2339 // to get rid of the control flow, so it's not worth promoting to select
2341 BasicBlock *DomBlock = nullptr;
2342 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2343 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2344 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2347 DomBlock = *pred_begin(IfBlock1);
2348 for (BasicBlock::iterator I = IfBlock1->begin(); !isa<TerminatorInst>(I);
2350 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2351 // This is not an aggressive instruction that we can promote.
2352 // Because of this, we won't be able to get rid of the control flow, so
2353 // the xform is not worth it.
2358 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2361 DomBlock = *pred_begin(IfBlock2);
2362 for (BasicBlock::iterator I = IfBlock2->begin(); !isa<TerminatorInst>(I);
2364 if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2365 // This is not an aggressive instruction that we can promote.
2366 // Because of this, we won't be able to get rid of the control flow, so
2367 // the xform is not worth it.
2372 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
2373 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
2375 // If we can still promote the PHI nodes after this gauntlet of tests,
2376 // do all of the PHI's now.
2377 Instruction *InsertPt = DomBlock->getTerminator();
2378 IRBuilder<NoFolder> Builder(InsertPt);
2380 // Move all 'aggressive' instructions, which are defined in the
2381 // conditional parts of the if's up to the dominating block.
2383 for (auto &I : *IfBlock1)
2384 I.dropUnknownNonDebugMetadata();
2385 DomBlock->getInstList().splice(InsertPt->getIterator(),
2386 IfBlock1->getInstList(), IfBlock1->begin(),
2387 IfBlock1->getTerminator()->getIterator());
2390 for (auto &I : *IfBlock2)
2391 I.dropUnknownNonDebugMetadata();
2392 DomBlock->getInstList().splice(InsertPt->getIterator(),
2393 IfBlock2->getInstList(), IfBlock2->begin(),
2394 IfBlock2->getTerminator()->getIterator());
2397 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2398 // Change the PHI node into a select instruction.
2399 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2400 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2402 Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2403 PN->replaceAllUsesWith(Sel);
2405 PN->eraseFromParent();
2408 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2409 // has been flattened. Change DomBlock to jump directly to our new block to
2410 // avoid other simplifycfg's kicking in on the diamond.
2411 TerminatorInst *OldTI = DomBlock->getTerminator();
2412 Builder.SetInsertPoint(OldTI);
2413 Builder.CreateBr(BB);
2414 OldTI->eraseFromParent();
2418 /// If we found a conditional branch that goes to two returning blocks,
2419 /// try to merge them together into one return,
2420 /// introducing a select if the return values disagree.
2421 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
2422 IRBuilder<> &Builder) {
2423 assert(BI->isConditional() && "Must be a conditional branch");
2424 BasicBlock *TrueSucc = BI->getSuccessor(0);
2425 BasicBlock *FalseSucc = BI->getSuccessor(1);
2426 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2427 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2429 // Check to ensure both blocks are empty (just a return) or optionally empty
2430 // with PHI nodes. If there are other instructions, merging would cause extra
2431 // computation on one path or the other.
2432 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2434 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2437 Builder.SetInsertPoint(BI);
2438 // Okay, we found a branch that is going to two return nodes. If
2439 // there is no return value for this function, just change the
2440 // branch into a return.
2441 if (FalseRet->getNumOperands() == 0) {
2442 TrueSucc->removePredecessor(BI->getParent());
2443 FalseSucc->removePredecessor(BI->getParent());
2444 Builder.CreateRetVoid();
2445 EraseTerminatorInstAndDCECond(BI);
2449 // Otherwise, figure out what the true and false return values are
2450 // so we can insert a new select instruction.
2451 Value *TrueValue = TrueRet->getReturnValue();
2452 Value *FalseValue = FalseRet->getReturnValue();
2454 // Unwrap any PHI nodes in the return blocks.
2455 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2456 if (TVPN->getParent() == TrueSucc)
2457 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2458 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2459 if (FVPN->getParent() == FalseSucc)
2460 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2462 // In order for this transformation to be safe, we must be able to
2463 // unconditionally execute both operands to the return. This is
2464 // normally the case, but we could have a potentially-trapping
2465 // constant expression that prevents this transformation from being
2467 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2470 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2474 // Okay, we collected all the mapped values and checked them for sanity, and
2475 // defined to really do this transformation. First, update the CFG.
2476 TrueSucc->removePredecessor(BI->getParent());
2477 FalseSucc->removePredecessor(BI->getParent());
2479 // Insert select instructions where needed.
2480 Value *BrCond = BI->getCondition();
2482 // Insert a select if the results differ.
2483 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2484 } else if (isa<UndefValue>(TrueValue)) {
2485 TrueValue = FalseValue;
2488 Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2493 !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2497 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2498 << "\n " << *BI << "NewRet = " << *RI
2499 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: " << *FalseSucc);
2501 EraseTerminatorInstAndDCECond(BI);
2506 /// Return true if the given instruction is available
2507 /// in its predecessor block. If yes, the instruction will be removed.
2508 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
2509 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2511 for (Instruction &I : *PB) {
2512 Instruction *PBI = &I;
2513 // Check whether Inst and PBI generate the same value.
2514 if (Inst->isIdenticalTo(PBI)) {
2515 Inst->replaceAllUsesWith(PBI);
2516 Inst->eraseFromParent();
2523 /// Return true if either PBI or BI has branch weight available, and store
2524 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2525 /// not have branch weight, use 1:1 as its weight.
2526 static bool extractPredSuccWeights(BranchInst *PBI, BranchInst *BI,
2527 uint64_t &PredTrueWeight,
2528 uint64_t &PredFalseWeight,
2529 uint64_t &SuccTrueWeight,
2530 uint64_t &SuccFalseWeight) {
2531 bool PredHasWeights =
2532 PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2533 bool SuccHasWeights =
2534 BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2535 if (PredHasWeights || SuccHasWeights) {
2536 if (!PredHasWeights)
2537 PredTrueWeight = PredFalseWeight = 1;
2538 if (!SuccHasWeights)
2539 SuccTrueWeight = SuccFalseWeight = 1;
2546 /// If this basic block is simple enough, and if a predecessor branches to us
2547 /// and one of our successors, fold the block into the predecessor and use
2548 /// logical operations to pick the right destination.
2549 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2550 BasicBlock *BB = BI->getParent();
2552 Instruction *Cond = nullptr;
2553 if (BI->isConditional())
2554 Cond = dyn_cast<Instruction>(BI->getCondition());
2556 // For unconditional branch, check for a simple CFG pattern, where
2557 // BB has a single predecessor and BB's successor is also its predecessor's
2558 // successor. If such pattern exisits, check for CSE between BB and its
2560 if (BasicBlock *PB = BB->getSinglePredecessor())
2561 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2562 if (PBI->isConditional() &&
2563 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2564 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2565 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2566 Instruction *Curr = &*I++;
2567 if (isa<CmpInst>(Curr)) {
2571 // Quit if we can't remove this instruction.
2572 if (!checkCSEInPredecessor(Curr, PB))
2581 if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2582 Cond->getParent() != BB || !Cond->hasOneUse())
2585 // Make sure the instruction after the condition is the cond branch.
2586 BasicBlock::iterator CondIt = ++Cond->getIterator();
2588 // Ignore dbg intrinsics.
2589 while (isa<DbgInfoIntrinsic>(CondIt))
2595 // Only allow this transformation if computing the condition doesn't involve
2596 // too many instructions and these involved instructions can be executed
2597 // unconditionally. We denote all involved instructions except the condition
2598 // as "bonus instructions", and only allow this transformation when the
2599 // number of the bonus instructions does not exceed a certain threshold.
2600 unsigned NumBonusInsts = 0;
2601 for (auto I = BB->begin(); Cond != &*I; ++I) {
2602 // Ignore dbg intrinsics.
2603 if (isa<DbgInfoIntrinsic>(I))
2605 if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2607 // I has only one use and can be executed unconditionally.
2608 Instruction *User = dyn_cast<Instruction>(I->user_back());
2609 if (User == nullptr || User->getParent() != BB)
2611 // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2612 // to use any other instruction, User must be an instruction between next(I)
2615 // Early exits once we reach the limit.
2616 if (NumBonusInsts > BonusInstThreshold)
2620 // Cond is known to be a compare or binary operator. Check to make sure that
2621 // neither operand is a potentially-trapping constant expression.
2622 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2625 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2629 // Finally, don't infinitely unroll conditional loops.
2630 BasicBlock *TrueDest = BI->getSuccessor(0);
2631 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2632 if (TrueDest == BB || FalseDest == BB)
2635 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2636 BasicBlock *PredBlock = *PI;
2637 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2639 // Check that we have two conditional branches. If there is a PHI node in
2640 // the common successor, verify that the same value flows in from both
2642 SmallVector<PHINode *, 4> PHIs;
2643 if (!PBI || PBI->isUnconditional() ||
2644 (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2645 (!BI->isConditional() &&
2646 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2649 // Determine if the two branches share a common destination.
2650 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2651 bool InvertPredCond = false;
2653 if (BI->isConditional()) {
2654 if (PBI->getSuccessor(0) == TrueDest) {
2655 Opc = Instruction::Or;
2656 } else if (PBI->getSuccessor(1) == FalseDest) {
2657 Opc = Instruction::And;
2658 } else if (PBI->getSuccessor(0) == FalseDest) {
2659 Opc = Instruction::And;
2660 InvertPredCond = true;
2661 } else if (PBI->getSuccessor(1) == TrueDest) {
2662 Opc = Instruction::Or;
2663 InvertPredCond = true;
2668 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2672 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2673 IRBuilder<> Builder(PBI);
2675 // If we need to invert the condition in the pred block to match, do so now.
2676 if (InvertPredCond) {
2677 Value *NewCond = PBI->getCondition();
2679 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2680 CmpInst *CI = cast<CmpInst>(NewCond);
2681 CI->setPredicate(CI->getInversePredicate());
2684 Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2687 PBI->setCondition(NewCond);
2688 PBI->swapSuccessors();
2691 // If we have bonus instructions, clone them into the predecessor block.
2692 // Note that there may be multiple predecessor blocks, so we cannot move
2693 // bonus instructions to a predecessor block.
2694 ValueToValueMapTy VMap; // maps original values to cloned values
2695 // We already make sure Cond is the last instruction before BI. Therefore,
2696 // all instructions before Cond other than DbgInfoIntrinsic are bonus
2698 for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2699 if (isa<DbgInfoIntrinsic>(BonusInst))
2701 Instruction *NewBonusInst = BonusInst->clone();
2702 RemapInstruction(NewBonusInst, VMap,
2703 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2704 VMap[&*BonusInst] = NewBonusInst;
2706 // If we moved a load, we cannot any longer claim any knowledge about
2707 // its potential value. The previous information might have been valid
2708 // only given the branch precondition.
2709 // For an analogous reason, we must also drop all the metadata whose
2710 // semantics we don't understand.
2711 NewBonusInst->dropUnknownNonDebugMetadata();
2713 PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2714 NewBonusInst->takeName(&*BonusInst);
2715 BonusInst->setName(BonusInst->getName() + ".old");
2718 // Clone Cond into the predecessor basic block, and or/and the
2719 // two conditions together.
2720 Instruction *New = Cond->clone();
2721 RemapInstruction(New, VMap,
2722 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
2723 PredBlock->getInstList().insert(PBI->getIterator(), New);
2724 New->takeName(Cond);
2725 Cond->setName(New->getName() + ".old");
2727 if (BI->isConditional()) {
2728 Instruction *NewCond = cast<Instruction>(
2729 Builder.CreateBinOp(Opc, PBI->getCondition(), New, "or.cond"));
2730 PBI->setCondition(NewCond);
2732 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2734 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2735 SuccTrueWeight, SuccFalseWeight);
2736 SmallVector<uint64_t, 8> NewWeights;
2738 if (PBI->getSuccessor(0) == BB) {
2740 // PBI: br i1 %x, BB, FalseDest
2741 // BI: br i1 %y, TrueDest, FalseDest
2742 // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2743 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2744 // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2745 // TrueWeight for PBI * FalseWeight for BI.
2746 // We assume that total weights of a BranchInst can fit into 32 bits.
2747 // Therefore, we will not have overflow using 64-bit arithmetic.
2748 NewWeights.push_back(PredFalseWeight *
2749 (SuccFalseWeight + SuccTrueWeight) +
2750 PredTrueWeight * SuccFalseWeight);
2752 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2753 PBI->setSuccessor(0, TrueDest);
2755 if (PBI->getSuccessor(1) == BB) {
2757 // PBI: br i1 %x, TrueDest, BB
2758 // BI: br i1 %y, TrueDest, FalseDest
2759 // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2760 // FalseWeight for PBI * TrueWeight for BI.
2761 NewWeights.push_back(PredTrueWeight *
2762 (SuccFalseWeight + SuccTrueWeight) +
2763 PredFalseWeight * SuccTrueWeight);
2764 // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2765 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2767 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2768 PBI->setSuccessor(1, FalseDest);
2770 if (NewWeights.size() == 2) {
2771 // Halve the weights if any of them cannot fit in an uint32_t
2772 FitWeights(NewWeights);
2774 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2777 LLVMContext::MD_prof,
2778 MDBuilder(BI->getContext()).createBranchWeights(MDWeights));
2780 PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2782 // Update PHI nodes in the common successors.
2783 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2784 ConstantInt *PBI_C = cast<ConstantInt>(
2785 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2786 assert(PBI_C->getType()->isIntegerTy(1));
2787 Instruction *MergedCond = nullptr;
2788 if (PBI->getSuccessor(0) == TrueDest) {
2789 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2790 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2791 // is false: !PBI_Cond and BI_Value
2792 Instruction *NotCond = cast<Instruction>(
2793 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2794 MergedCond = cast<Instruction>(
2795 Builder.CreateBinOp(Instruction::And, NotCond, New, "and.cond"));
2797 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2798 Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2800 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2801 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2802 // is false: PBI_Cond and BI_Value
2803 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2804 Instruction::And, PBI->getCondition(), New, "and.cond"));
2805 if (PBI_C->isOne()) {
2806 Instruction *NotCond = cast<Instruction>(
2807 Builder.CreateNot(PBI->getCondition(), "not.cond"));
2808 MergedCond = cast<Instruction>(Builder.CreateBinOp(
2809 Instruction::Or, NotCond, MergedCond, "or.cond"));
2813 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2816 // Change PBI from Conditional to Unconditional.
2817 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2818 EraseTerminatorInstAndDCECond(PBI);
2822 // If BI was a loop latch, it may have had associated loop metadata.
2823 // We need to copy it to the new latch, that is, PBI.
2824 if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2825 PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2827 // TODO: If BB is reachable from all paths through PredBlock, then we
2828 // could replace PBI's branch probabilities with BI's.
2830 // Copy any debug value intrinsics into the end of PredBlock.
2831 for (Instruction &I : *BB)
2832 if (isa<DbgInfoIntrinsic>(I))
2833 I.clone()->insertBefore(PBI);
2840 // If there is only one store in BB1 and BB2, return it, otherwise return
2842 static StoreInst *findUniqueStoreInBlocks(BasicBlock *BB1, BasicBlock *BB2) {
2843 StoreInst *S = nullptr;
2844 for (auto *BB : {BB1, BB2}) {
2848 if (auto *SI = dyn_cast<StoreInst>(&I)) {
2850 // Multiple stores seen.
2859 static Value *ensureValueAvailableInSuccessor(Value *V, BasicBlock *BB,
2860 Value *AlternativeV = nullptr) {
2861 // PHI is going to be a PHI node that allows the value V that is defined in
2862 // BB to be referenced in BB's only successor.
2864 // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2865 // doesn't matter to us what the other operand is (it'll never get used). We
2866 // could just create a new PHI with an undef incoming value, but that could
2867 // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2868 // other PHI. So here we directly look for some PHI in BB's successor with V
2869 // as an incoming operand. If we find one, we use it, else we create a new
2872 // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2873 // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2874 // where OtherBB is the single other predecessor of BB's only successor.
2875 PHINode *PHI = nullptr;
2876 BasicBlock *Succ = BB->getSingleSuccessor();
2878 for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2879 if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2880 PHI = cast<PHINode>(I);
2884 assert(std::distance(pred_begin(Succ), pred_end(Succ)) == 2);
2885 auto PredI = pred_begin(Succ);
2886 BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2887 if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2894 // If V is not an instruction defined in BB, just return it.
2895 if (!AlternativeV &&
2896 (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2899 PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2900 PHI->addIncoming(V, BB);
2901 for (BasicBlock *PredBB : predecessors(Succ))
2904 AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2908 static bool mergeConditionalStoreToAddress(BasicBlock *PTB, BasicBlock *PFB,
2909 BasicBlock *QTB, BasicBlock *QFB,
2910 BasicBlock *PostBB, Value *Address,
2911 bool InvertPCond, bool InvertQCond) {
2912 auto IsaBitcastOfPointerType = [](const Instruction &I) {
2913 return Operator::getOpcode(&I) == Instruction::BitCast &&
2914 I.getType()->isPointerTy();
2917 // If we're not in aggressive mode, we only optimize if we have some
2918 // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2919 auto IsWorthwhile = [&](BasicBlock *BB) {
2922 // Heuristic: if the block can be if-converted/phi-folded and the
2923 // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2924 // thread this store.
2926 for (auto &I : *BB) {
2927 // Cheap instructions viable for folding.
2928 if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2931 // Free instructions.
2932 else if (isa<TerminatorInst>(I) || isa<DbgInfoIntrinsic>(I) ||
2933 IsaBitcastOfPointerType(I))
2938 return N <= PHINodeFoldingThreshold;
2941 if (!MergeCondStoresAggressively &&
2942 (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2943 !IsWorthwhile(QFB)))
2946 // For every pointer, there must be exactly two stores, one coming from
2947 // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2948 // store (to any address) in PTB,PFB or QTB,QFB.
2949 // FIXME: We could relax this restriction with a bit more work and performance
2951 StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2952 StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2953 if (!PStore || !QStore)
2956 // Now check the stores are compatible.
2957 if (!QStore->isUnordered() || !PStore->isUnordered())
2960 // Check that sinking the store won't cause program behavior changes. Sinking
2961 // the store out of the Q blocks won't change any behavior as we're sinking
2962 // from a block to its unconditional successor. But we're moving a store from
2963 // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2964 // So we need to check that there are no aliasing loads or stores in
2965 // QBI, QTB and QFB. We also need to check there are no conflicting memory
2966 // operations between PStore and the end of its parent block.
2968 // The ideal way to do this is to query AliasAnalysis, but we don't
2969 // preserve AA currently so that is dangerous. Be super safe and just
2970 // check there are no other memory operations at all.
2971 for (auto &I : *QFB->getSinglePredecessor())
2972 if (I.mayReadOrWriteMemory())
2974 for (auto &I : *QFB)
2975 if (&I != QStore && I.mayReadOrWriteMemory())
2978 for (auto &I : *QTB)
2979 if (&I != QStore && I.mayReadOrWriteMemory())
2981 for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2983 if (&*I != PStore && I->mayReadOrWriteMemory())
2986 // OK, we're going to sink the stores to PostBB. The store has to be
2987 // conditional though, so first create the predicate.
2988 Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
2990 Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
2993 Value *PPHI = ensureValueAvailableInSuccessor(PStore->getValueOperand(),
2994 PStore->getParent());
2995 Value *QPHI = ensureValueAvailableInSuccessor(QStore->getValueOperand(),
2996 QStore->getParent(), PPHI);
2998 IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
3000 Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
3001 Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
3004 PPred = QB.CreateNot(PPred);
3006 QPred = QB.CreateNot(QPred);
3007 Value *CombinedPred = QB.CreateOr(PPred, QPred);
3010 SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
3011 QB.SetInsertPoint(T);
3012 StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
3014 PStore->getAAMetadata(AAMD, /*Merge=*/false);
3015 PStore->getAAMetadata(AAMD, /*Merge=*/true);
3016 SI->setAAMetadata(AAMD);
3018 QStore->eraseFromParent();
3019 PStore->eraseFromParent();
3024 static bool mergeConditionalStores(BranchInst *PBI, BranchInst *QBI) {
3025 // The intention here is to find diamonds or triangles (see below) where each
3026 // conditional block contains a store to the same address. Both of these
3027 // stores are conditional, so they can't be unconditionally sunk. But it may
3028 // be profitable to speculatively sink the stores into one merged store at the
3029 // end, and predicate the merged store on the union of the two conditions of
3032 // This can reduce the number of stores executed if both of the conditions are
3033 // true, and can allow the blocks to become small enough to be if-converted.
3034 // This optimization will also chain, so that ladders of test-and-set
3035 // sequences can be if-converted away.
3037 // We only deal with simple diamonds or triangles:
3039 // PBI or PBI or a combination of the two
3049 // We model triangles as a type of diamond with a nullptr "true" block.
3050 // Triangles are canonicalized so that the fallthrough edge is represented by
3051 // a true condition, as in the diagram above.
3053 BasicBlock *PTB = PBI->getSuccessor(0);
3054 BasicBlock *PFB = PBI->getSuccessor(1);
3055 BasicBlock *QTB = QBI->getSuccessor(0);
3056 BasicBlock *QFB = QBI->getSuccessor(1);
3057 BasicBlock *PostBB = QFB->getSingleSuccessor();
3059 // Make sure we have a good guess for PostBB. If QTB's only successor is
3060 // QFB, then QFB is a better PostBB.
3061 if (QTB->getSingleSuccessor() == QFB)
3064 // If we couldn't find a good PostBB, stop.
3068 bool InvertPCond = false, InvertQCond = false;
3069 // Canonicalize fallthroughs to the true branches.
3070 if (PFB == QBI->getParent()) {
3071 std::swap(PFB, PTB);
3074 if (QFB == PostBB) {
3075 std::swap(QFB, QTB);
3079 // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3080 // and QFB may not. Model fallthroughs as a nullptr block.
3081 if (PTB == QBI->getParent())
3086 // Legality bailouts. We must have at least the non-fallthrough blocks and
3087 // the post-dominating block, and the non-fallthroughs must only have one
3089 auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3090 return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3092 if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3093 !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3095 if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3096 (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3098 if (!PostBB->hasNUses(2) || !QBI->getParent()->hasNUses(2))
3101 // OK, this is a sequence of two diamonds or triangles.
3102 // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3103 SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3104 for (auto *BB : {PTB, PFB}) {
3108 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3109 PStoreAddresses.insert(SI->getPointerOperand());
3111 for (auto *BB : {QTB, QFB}) {
3115 if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3116 QStoreAddresses.insert(SI->getPointerOperand());
3119 set_intersect(PStoreAddresses, QStoreAddresses);
3120 // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3121 // clear what it contains.
3122 auto &CommonAddresses = PStoreAddresses;
3124 bool Changed = false;
3125 for (auto *Address : CommonAddresses)
3126 Changed |= mergeConditionalStoreToAddress(
3127 PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond);
3131 /// If we have a conditional branch as a predecessor of another block,
3132 /// this function tries to simplify it. We know
3133 /// that PBI and BI are both conditional branches, and BI is in one of the
3134 /// successor blocks of PBI - PBI branches to BI.
3135 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI,
3136 const DataLayout &DL) {
3137 assert(PBI->isConditional() && BI->isConditional());
3138 BasicBlock *BB = BI->getParent();
3140 // If this block ends with a branch instruction, and if there is a
3141 // predecessor that ends on a branch of the same condition, make
3142 // this conditional branch redundant.
3143 if (PBI->getCondition() == BI->getCondition() &&
3144 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3145 // Okay, the outcome of this conditional branch is statically
3146 // knowable. If this block had a single pred, handle specially.
3147 if (BB->getSinglePredecessor()) {
3148 // Turn this into a branch on constant.
3149 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3151 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3152 return true; // Nuke the branch on constant.
3155 // Otherwise, if there are multiple predecessors, insert a PHI that merges
3156 // in the constant and simplify the block result. Subsequent passes of
3157 // simplifycfg will thread the block.
3158 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
3159 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3160 PHINode *NewPN = PHINode::Create(
3161 Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3162 BI->getCondition()->getName() + ".pr", &BB->front());
3163 // Okay, we're going to insert the PHI node. Since PBI is not the only
3164 // predecessor, compute the PHI'd conditional value for all of the preds.
3165 // Any predecessor where the condition is not computable we keep symbolic.
3166 for (pred_iterator PI = PB; PI != PE; ++PI) {
3167 BasicBlock *P = *PI;
3168 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3169 PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3170 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3171 bool CondIsTrue = PBI->getSuccessor(0) == BB;
3173 ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3176 NewPN->addIncoming(BI->getCondition(), P);
3180 BI->setCondition(NewPN);
3185 if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3189 // If both branches are conditional and both contain stores to the same
3190 // address, remove the stores from the conditionals and create a conditional
3191 // merged store at the end.
3192 if (MergeCondStores && mergeConditionalStores(PBI, BI))
3195 // If this is a conditional branch in an empty block, and if any
3196 // predecessors are a conditional branch to one of our destinations,
3197 // fold the conditions into logical ops and one cond br.
3198 BasicBlock::iterator BBI = BB->begin();
3199 // Ignore dbg intrinsics.
3200 while (isa<DbgInfoIntrinsic>(BBI))
3206 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3209 } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3212 } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3215 } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3222 // Check to make sure that the other destination of this branch
3223 // isn't BB itself. If so, this is an infinite loop that will
3224 // keep getting unwound.
3225 if (PBI->getSuccessor(PBIOp) == BB)
3228 // Do not perform this transformation if it would require
3229 // insertion of a large number of select instructions. For targets
3230 // without predication/cmovs, this is a big pessimization.
3232 // Also do not perform this transformation if any phi node in the common
3233 // destination block can trap when reached by BB or PBB (PR17073). In that
3234 // case, it would be unsafe to hoist the operation into a select instruction.
3236 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3237 unsigned NumPhis = 0;
3238 for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3240 if (NumPhis > 2) // Disable this xform.
3243 PHINode *PN = cast<PHINode>(II);
3244 Value *BIV = PN->getIncomingValueForBlock(BB);
3245 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3249 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3250 Value *PBIV = PN->getIncomingValue(PBBIdx);
3251 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3256 // Finally, if everything is ok, fold the branches to logical ops.
3257 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3259 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3260 << "AND: " << *BI->getParent());
3262 // If OtherDest *is* BB, then BB is a basic block with a single conditional
3263 // branch in it, where one edge (OtherDest) goes back to itself but the other
3264 // exits. We don't *know* that the program avoids the infinite loop
3265 // (even though that seems likely). If we do this xform naively, we'll end up
3266 // recursively unpeeling the loop. Since we know that (after the xform is
3267 // done) that the block *is* infinite if reached, we just make it an obviously
3268 // infinite loop with no cond branch.
3269 if (OtherDest == BB) {
3270 // Insert it at the end of the function, because it's either code,
3271 // or it won't matter if it's hot. :)
3272 BasicBlock *InfLoopBlock =
3273 BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3274 BranchInst::Create(InfLoopBlock, InfLoopBlock);
3275 OtherDest = InfLoopBlock;
3278 DEBUG(dbgs() << *PBI->getParent()->getParent());
3280 // BI may have other predecessors. Because of this, we leave
3281 // it alone, but modify PBI.
3283 // Make sure we get to CommonDest on True&True directions.
3284 Value *PBICond = PBI->getCondition();
3285 IRBuilder<NoFolder> Builder(PBI);
3287 PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3289 Value *BICond = BI->getCondition();
3291 BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3293 // Merge the conditions.
3294 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3296 // Modify PBI to branch on the new condition to the new dests.
3297 PBI->setCondition(Cond);
3298 PBI->setSuccessor(0, CommonDest);
3299 PBI->setSuccessor(1, OtherDest);
3301 // Update branch weight for PBI.
3302 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3303 uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3305 extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3306 SuccTrueWeight, SuccFalseWeight);
3308 PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3309 PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3310 SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3311 SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3312 // The weight to CommonDest should be PredCommon * SuccTotal +
3313 // PredOther * SuccCommon.
3314 // The weight to OtherDest should be PredOther * SuccOther.
3315 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3316 PredOther * SuccCommon,
3317 PredOther * SuccOther};
3318 // Halve the weights if any of them cannot fit in an uint32_t
3319 FitWeights(NewWeights);
3321 PBI->setMetadata(LLVMContext::MD_prof,
3322 MDBuilder(BI->getContext())
3323 .createBranchWeights(NewWeights[0], NewWeights[1]));
3326 // OtherDest may have phi nodes. If so, add an entry from PBI's
3327 // block that are identical to the entries for BI's block.
3328 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3330 // We know that the CommonDest already had an edge from PBI to
3331 // it. If it has PHIs though, the PHIs may have different
3332 // entries for BB and PBI's BB. If so, insert a select to make
3335 for (BasicBlock::iterator II = CommonDest->begin();
3336 (PN = dyn_cast<PHINode>(II)); ++II) {
3337 Value *BIV = PN->getIncomingValueForBlock(BB);
3338 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3339 Value *PBIV = PN->getIncomingValue(PBBIdx);
3341 // Insert a select in PBI to pick the right value.
3342 SelectInst *NV = cast<SelectInst>(
3343 Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3344 PN->setIncomingValue(PBBIdx, NV);
3345 // Although the select has the same condition as PBI, the original branch
3346 // weights for PBI do not apply to the new select because the select's
3347 // 'logical' edges are incoming edges of the phi that is eliminated, not
3348 // the outgoing edges of PBI.
3350 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3351 uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3352 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3353 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3354 // The weight to PredCommonDest should be PredCommon * SuccTotal.
3355 // The weight to PredOtherDest should be PredOther * SuccCommon.
3356 uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3357 PredOther * SuccCommon};
3359 FitWeights(NewWeights);
3361 NV->setMetadata(LLVMContext::MD_prof,
3362 MDBuilder(BI->getContext())
3363 .createBranchWeights(NewWeights[0], NewWeights[1]));
3368 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3369 DEBUG(dbgs() << *PBI->getParent()->getParent());
3371 // This basic block is probably dead. We know it has at least
3372 // one fewer predecessor.
3376 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3377 // true or to FalseBB if Cond is false.
3378 // Takes care of updating the successors and removing the old terminator.
3379 // Also makes sure not to introduce new successors by assuming that edges to
3380 // non-successor TrueBBs and FalseBBs aren't reachable.
3381 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
3382 BasicBlock *TrueBB, BasicBlock *FalseBB,
3383 uint32_t TrueWeight,
3384 uint32_t FalseWeight) {
3385 // Remove any superfluous successor edges from the CFG.
3386 // First, figure out which successors to preserve.
3387 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3389 BasicBlock *KeepEdge1 = TrueBB;
3390 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3392 // Then remove the rest.
3393 for (BasicBlock *Succ : OldTerm->successors()) {
3394 // Make sure only to keep exactly one copy of each edge.
3395 if (Succ == KeepEdge1)
3396 KeepEdge1 = nullptr;
3397 else if (Succ == KeepEdge2)
3398 KeepEdge2 = nullptr;
3400 Succ->removePredecessor(OldTerm->getParent(),
3401 /*DontDeleteUselessPHIs=*/true);
3404 IRBuilder<> Builder(OldTerm);
3405 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3407 // Insert an appropriate new terminator.
3408 if (!KeepEdge1 && !KeepEdge2) {
3409 if (TrueBB == FalseBB)
3410 // We were only looking for one successor, and it was present.
3411 // Create an unconditional branch to it.
3412 Builder.CreateBr(TrueBB);
3414 // We found both of the successors we were looking for.
3415 // Create a conditional branch sharing the condition of the select.
3416 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3417 if (TrueWeight != FalseWeight)
3418 NewBI->setMetadata(LLVMContext::MD_prof,
3419 MDBuilder(OldTerm->getContext())
3420 .createBranchWeights(TrueWeight, FalseWeight));
3422 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3423 // Neither of the selected blocks were successors, so this
3424 // terminator must be unreachable.
3425 new UnreachableInst(OldTerm->getContext(), OldTerm);
3427 // One of the selected values was a successor, but the other wasn't.
3428 // Insert an unconditional branch to the one that was found;
3429 // the edge to the one that wasn't must be unreachable.
3431 // Only TrueBB was found.
3432 Builder.CreateBr(TrueBB);
3434 // Only FalseBB was found.
3435 Builder.CreateBr(FalseBB);
3438 EraseTerminatorInstAndDCECond(OldTerm);
3443 // (switch (select cond, X, Y)) on constant X, Y
3444 // with a branch - conditional if X and Y lead to distinct BBs,
3445 // unconditional otherwise.
3446 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
3447 // Check for constant integer values in the select.
3448 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3449 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3450 if (!TrueVal || !FalseVal)
3453 // Find the relevant condition and destinations.
3454 Value *Condition = Select->getCondition();
3455 BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3456 BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3458 // Get weight for TrueBB and FalseBB.
3459 uint32_t TrueWeight = 0, FalseWeight = 0;
3460 SmallVector<uint64_t, 8> Weights;
3461 bool HasWeights = HasBranchWeights(SI);
3463 GetBranchWeights(SI, Weights);
3464 if (Weights.size() == 1 + SI->getNumCases()) {
3466 (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3468 (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3472 // Perform the actual simplification.
3473 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3478 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3479 // blockaddress(@fn, BlockB)))
3481 // (br cond, BlockA, BlockB).
3482 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
3483 // Check that both operands of the select are block addresses.
3484 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
3485 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
3489 // Extract the actual blocks.
3490 BasicBlock *TrueBB = TBA->getBasicBlock();
3491 BasicBlock *FalseBB = FBA->getBasicBlock();
3493 // Perform the actual simplification.
3494 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3498 /// This is called when we find an icmp instruction
3499 /// (a seteq/setne with a constant) as the only instruction in a
3500 /// block that ends with an uncond branch. We are looking for a very specific
3501 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3502 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3503 /// default value goes to an uncond block with a seteq in it, we get something
3506 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3508 /// %tmp = icmp eq i8 %A, 92
3511 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3513 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3514 /// the PHI, merging the third icmp into the switch.
3515 static bool TryToSimplifyUncondBranchWithICmpInIt(
3516 ICmpInst *ICI, IRBuilder<> &Builder, const DataLayout &DL,
3517 const TargetTransformInfo &TTI, unsigned BonusInstThreshold,
3518 AssumptionCache *AC) {
3519 BasicBlock *BB = ICI->getParent();
3521 // If the block has any PHIs in it or the icmp has multiple uses, it is too
3523 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3526 Value *V = ICI->getOperand(0);
3527 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3529 // The pattern we're looking for is where our only predecessor is a switch on
3530 // 'V' and this block is the default case for the switch. In this case we can
3531 // fold the compared value into the switch to simplify things.
3532 BasicBlock *Pred = BB->getSinglePredecessor();
3533 if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3536 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3537 if (SI->getCondition() != V)
3540 // If BB is reachable on a non-default case, then we simply know the value of
3541 // V in this block. Substitute it and constant fold the icmp instruction
3543 if (SI->getDefaultDest() != BB) {
3544 ConstantInt *VVal = SI->findCaseDest(BB);
3545 assert(VVal && "Should have a unique destination value");
3546 ICI->setOperand(0, VVal);
3548 if (Value *V = SimplifyInstruction(ICI, DL)) {
3549 ICI->replaceAllUsesWith(V);
3550 ICI->eraseFromParent();
3552 // BB is now empty, so it is likely to simplify away.
3553 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3556 // Ok, the block is reachable from the default dest. If the constant we're
3557 // comparing exists in one of the other edges, then we can constant fold ICI
3559 if (SI->findCaseValue(Cst) != SI->case_default()) {
3561 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3562 V = ConstantInt::getFalse(BB->getContext());
3564 V = ConstantInt::getTrue(BB->getContext());
3566 ICI->replaceAllUsesWith(V);
3567 ICI->eraseFromParent();
3568 // BB is now empty, so it is likely to simplify away.
3569 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3572 // The use of the icmp has to be in the 'end' block, by the only PHI node in
3574 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3575 PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3576 if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3577 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3580 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3582 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3583 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3585 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3586 std::swap(DefaultCst, NewCst);
3588 // Replace ICI (which is used by the PHI for the default value) with true or
3589 // false depending on if it is EQ or NE.
3590 ICI->replaceAllUsesWith(DefaultCst);
3591 ICI->eraseFromParent();
3593 // Okay, the switch goes to this block on a default value. Add an edge from
3594 // the switch to the merge point on the compared value.
3596 BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3597 SmallVector<uint64_t, 8> Weights;
3598 bool HasWeights = HasBranchWeights(SI);
3600 GetBranchWeights(SI, Weights);
3601 if (Weights.size() == 1 + SI->getNumCases()) {
3602 // Split weight for default case to case for "Cst".
3603 Weights[0] = (Weights[0] + 1) >> 1;
3604 Weights.push_back(Weights[0]);
3606 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3608 LLVMContext::MD_prof,
3609 MDBuilder(SI->getContext()).createBranchWeights(MDWeights));
3612 SI->addCase(Cst, NewBB);
3614 // NewBB branches to the phi block, add the uncond branch and the phi entry.
3615 Builder.SetInsertPoint(NewBB);
3616 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3617 Builder.CreateBr(SuccBlock);
3618 PHIUse->addIncoming(NewCst, NewBB);
3622 /// The specified branch is a conditional branch.
3623 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3624 /// fold it into a switch instruction if so.
3625 static bool SimplifyBranchOnICmpChain(BranchInst *BI, IRBuilder<> &Builder,
3626 const DataLayout &DL) {
3627 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3631 // Change br (X == 0 | X == 1), T, F into a switch instruction.
3632 // If this is a bunch of seteq's or'd together, or if it's a bunch of
3633 // 'setne's and'ed together, collect them.
3635 // Try to gather values from a chain of and/or to be turned into a switch
3636 ConstantComparesGatherer ConstantCompare(Cond, DL);
3637 // Unpack the result
3638 SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3639 Value *CompVal = ConstantCompare.CompValue;
3640 unsigned UsedICmps = ConstantCompare.UsedICmps;
3641 Value *ExtraCase = ConstantCompare.Extra;
3643 // If we didn't have a multiply compared value, fail.
3647 // Avoid turning single icmps into a switch.
3651 bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3653 // There might be duplicate constants in the list, which the switch
3654 // instruction can't handle, remove them now.
3655 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3656 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3658 // If Extra was used, we require at least two switch values to do the
3659 // transformation. A switch with one value is just a conditional branch.
3660 if (ExtraCase && Values.size() < 2)
3663 // TODO: Preserve branch weight metadata, similarly to how
3664 // FoldValueComparisonIntoPredecessors preserves it.
3666 // Figure out which block is which destination.
3667 BasicBlock *DefaultBB = BI->getSuccessor(1);
3668 BasicBlock *EdgeBB = BI->getSuccessor(0);
3670 std::swap(DefaultBB, EdgeBB);
3672 BasicBlock *BB = BI->getParent();
3674 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3675 << " cases into SWITCH. BB is:\n"
3678 // If there are any extra values that couldn't be folded into the switch
3679 // then we evaluate them with an explicit branch first. Split the block
3680 // right before the condbr to handle it.
3683 BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3684 // Remove the uncond branch added to the old block.
3685 TerminatorInst *OldTI = BB->getTerminator();
3686 Builder.SetInsertPoint(OldTI);
3689 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3691 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3693 OldTI->eraseFromParent();
3695 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3696 // for the edge we just added.
3697 AddPredecessorToBlock(EdgeBB, BB, NewBB);
3699 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3700 << "\nEXTRABB = " << *BB);
3704 Builder.SetInsertPoint(BI);
3705 // Convert pointer to int before we switch.
3706 if (CompVal->getType()->isPointerTy()) {
3707 CompVal = Builder.CreatePtrToInt(
3708 CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3711 // Create the new switch instruction now.
3712 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3714 // Add all of the 'cases' to the switch instruction.
3715 for (unsigned i = 0, e = Values.size(); i != e; ++i)
3716 New->addCase(Values[i], EdgeBB);
3718 // We added edges from PI to the EdgeBB. As such, if there were any
3719 // PHI nodes in EdgeBB, they need entries to be added corresponding to
3720 // the number of edges added.
3721 for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3722 PHINode *PN = cast<PHINode>(BBI);
3723 Value *InVal = PN->getIncomingValueForBlock(BB);
3724 for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3725 PN->addIncoming(InVal, BB);
3728 // Erase the old branch instruction.
3729 EraseTerminatorInstAndDCECond(BI);
3731 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3735 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3736 if (isa<PHINode>(RI->getValue()))
3737 return SimplifyCommonResume(RI);
3738 else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3739 RI->getValue() == RI->getParent()->getFirstNonPHI())
3740 // The resume must unwind the exception that caused control to branch here.
3741 return SimplifySingleResume(RI);
3746 // Simplify resume that is shared by several landing pads (phi of landing pad).
3747 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3748 BasicBlock *BB = RI->getParent();
3750 // Check that there are no other instructions except for debug intrinsics
3751 // between the phi of landing pads (RI->getValue()) and resume instruction.
3752 BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3753 E = RI->getIterator();
3755 if (!isa<DbgInfoIntrinsic>(I))
3758 SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
3759 auto *PhiLPInst = cast<PHINode>(RI->getValue());
3761 // Check incoming blocks to see if any of them are trivial.
3762 for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3764 auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3765 auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3767 // If the block has other successors, we can not delete it because
3768 // it has other dependents.
3769 if (IncomingBB->getUniqueSuccessor() != BB)
3772 auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3773 // Not the landing pad that caused the control to branch here.
3774 if (IncomingValue != LandingPad)
3777 bool isTrivial = true;
3779 I = IncomingBB->getFirstNonPHI()->getIterator();
3780 E = IncomingBB->getTerminator()->getIterator();
3782 if (!isa<DbgInfoIntrinsic>(I)) {
3788 TrivialUnwindBlocks.insert(IncomingBB);
3791 // If no trivial unwind blocks, don't do any simplifications.
3792 if (TrivialUnwindBlocks.empty())
3795 // Turn all invokes that unwind here into calls.
3796 for (auto *TrivialBB : TrivialUnwindBlocks) {
3797 // Blocks that will be simplified should be removed from the phi node.
3798 // Note there could be multiple edges to the resume block, and we need
3799 // to remove them all.
3800 while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3801 BB->removePredecessor(TrivialBB, true);
3803 for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3805 BasicBlock *Pred = *PI++;
3806 removeUnwindEdge(Pred);
3809 // In each SimplifyCFG run, only the current processed block can be erased.
3810 // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3811 // of erasing TrivialBB, we only remove the branch to the common resume
3812 // block so that we can later erase the resume block since it has no
3814 TrivialBB->getTerminator()->eraseFromParent();
3815 new UnreachableInst(RI->getContext(), TrivialBB);
3818 // Delete the resume block if all its predecessors have been removed.
3820 BB->eraseFromParent();
3822 return !TrivialUnwindBlocks.empty();
3825 // Simplify resume that is only used by a single (non-phi) landing pad.
3826 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3827 BasicBlock *BB = RI->getParent();
3828 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
3829 assert(RI->getValue() == LPInst &&
3830 "Resume must unwind the exception that caused control to here");
3832 // Check that there are no other instructions except for debug intrinsics.
3833 BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3835 if (!isa<DbgInfoIntrinsic>(I))
3838 // Turn all invokes that unwind here into calls and delete the basic block.
3839 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3840 BasicBlock *Pred = *PI++;
3841 removeUnwindEdge(Pred);
3844 // The landingpad is now unreachable. Zap it.
3845 BB->eraseFromParent();
3847 LoopHeaders->erase(BB);
3851 static bool removeEmptyCleanup(CleanupReturnInst *RI) {
3852 // If this is a trivial cleanup pad that executes no instructions, it can be
3853 // eliminated. If the cleanup pad continues to the caller, any predecessor
3854 // that is an EH pad will be updated to continue to the caller and any
3855 // predecessor that terminates with an invoke instruction will have its invoke
3856 // instruction converted to a call instruction. If the cleanup pad being
3857 // simplified does not continue to the caller, each predecessor will be
3858 // updated to continue to the unwind destination of the cleanup pad being
3860 BasicBlock *BB = RI->getParent();
3861 CleanupPadInst *CPInst = RI->getCleanupPad();
3862 if (CPInst->getParent() != BB)
3863 // This isn't an empty cleanup.
3866 // We cannot kill the pad if it has multiple uses. This typically arises
3867 // from unreachable basic blocks.
3868 if (!CPInst->hasOneUse())
3871 // Check that there are no other instructions except for benign intrinsics.
3872 BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3874 auto *II = dyn_cast<IntrinsicInst>(I);
3878 Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3879 switch (IntrinsicID) {
3880 case Intrinsic::dbg_declare:
3881 case Intrinsic::dbg_value:
3882 case Intrinsic::lifetime_end:
3889 // If the cleanup return we are simplifying unwinds to the caller, this will
3890 // set UnwindDest to nullptr.
3891 BasicBlock *UnwindDest = RI->getUnwindDest();
3892 Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3894 // We're about to remove BB from the control flow. Before we do, sink any
3895 // PHINodes into the unwind destination. Doing this before changing the
3896 // control flow avoids some potentially slow checks, since we can currently
3897 // be certain that UnwindDest and BB have no common predecessors (since they
3898 // are both EH pads).
3900 // First, go through the PHI nodes in UnwindDest and update any nodes that
3901 // reference the block we are removing
3902 for (BasicBlock::iterator I = UnwindDest->begin(),
3903 IE = DestEHPad->getIterator();
3905 PHINode *DestPN = cast<PHINode>(I);
3907 int Idx = DestPN->getBasicBlockIndex(BB);
3908 // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3910 // This PHI node has an incoming value that corresponds to a control
3911 // path through the cleanup pad we are removing. If the incoming
3912 // value is in the cleanup pad, it must be a PHINode (because we
3913 // verified above that the block is otherwise empty). Otherwise, the
3914 // value is either a constant or a value that dominates the cleanup
3915 // pad being removed.
3917 // Because BB and UnwindDest are both EH pads, all of their
3918 // predecessors must unwind to these blocks, and since no instruction
3919 // can have multiple unwind destinations, there will be no overlap in
3920 // incoming blocks between SrcPN and DestPN.
3921 Value *SrcVal = DestPN->getIncomingValue(Idx);
3922 PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3924 // Remove the entry for the block we are deleting.
3925 DestPN->removeIncomingValue(Idx, false);
3927 if (SrcPN && SrcPN->getParent() == BB) {
3928 // If the incoming value was a PHI node in the cleanup pad we are
3929 // removing, we need to merge that PHI node's incoming values into
3931 for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3932 SrcIdx != SrcE; ++SrcIdx) {
3933 DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3934 SrcPN->getIncomingBlock(SrcIdx));
3937 // Otherwise, the incoming value came from above BB and
3938 // so we can just reuse it. We must associate all of BB's
3939 // predecessors with this value.
3940 for (auto *pred : predecessors(BB)) {
3941 DestPN->addIncoming(SrcVal, pred);
3946 // Sink any remaining PHI nodes directly into UnwindDest.
3947 Instruction *InsertPt = DestEHPad;
3948 for (BasicBlock::iterator I = BB->begin(),
3949 IE = BB->getFirstNonPHI()->getIterator();
3951 // The iterator must be incremented here because the instructions are
3952 // being moved to another block.
3953 PHINode *PN = cast<PHINode>(I++);
3954 if (PN->use_empty())
3955 // If the PHI node has no uses, just leave it. It will be erased
3956 // when we erase BB below.
3959 // Otherwise, sink this PHI node into UnwindDest.
3960 // Any predecessors to UnwindDest which are not already represented
3961 // must be back edges which inherit the value from the path through
3962 // BB. In this case, the PHI value must reference itself.
3963 for (auto *pred : predecessors(UnwindDest))
3965 PN->addIncoming(PN, pred);
3966 PN->moveBefore(InsertPt);
3970 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3971 // The iterator must be updated here because we are removing this pred.
3972 BasicBlock *PredBB = *PI++;
3973 if (UnwindDest == nullptr) {
3974 removeUnwindEdge(PredBB);
3976 TerminatorInst *TI = PredBB->getTerminator();
3977 TI->replaceUsesOfWith(BB, UnwindDest);
3981 // The cleanup pad is now unreachable. Zap it.
3982 BB->eraseFromParent();
3986 // Try to merge two cleanuppads together.
3987 static bool mergeCleanupPad(CleanupReturnInst *RI) {
3988 // Skip any cleanuprets which unwind to caller, there is nothing to merge
3990 BasicBlock *UnwindDest = RI->getUnwindDest();
3994 // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
3995 // be safe to merge without code duplication.
3996 if (UnwindDest->getSinglePredecessor() != RI->getParent())
3999 // Verify that our cleanuppad's unwind destination is another cleanuppad.
4000 auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
4001 if (!SuccessorCleanupPad)
4004 CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
4005 // Replace any uses of the successor cleanupad with the predecessor pad
4006 // The only cleanuppad uses should be this cleanupret, it's cleanupret and
4007 // funclet bundle operands.
4008 SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4009 // Remove the old cleanuppad.
4010 SuccessorCleanupPad->eraseFromParent();
4011 // Now, we simply replace the cleanupret with a branch to the unwind
4013 BranchInst::Create(UnwindDest, RI->getParent());
4014 RI->eraseFromParent();
4019 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
4020 // It is possible to transiantly have an undef cleanuppad operand because we
4021 // have deleted some, but not all, dead blocks.
4022 // Eventually, this block will be deleted.
4023 if (isa<UndefValue>(RI->getOperand(0)))
4026 if (mergeCleanupPad(RI))
4029 if (removeEmptyCleanup(RI))
4035 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4036 BasicBlock *BB = RI->getParent();
4037 if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4040 // Find predecessors that end with branches.
4041 SmallVector<BasicBlock *, 8> UncondBranchPreds;
4042 SmallVector<BranchInst *, 8> CondBranchPreds;
4043 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4044 BasicBlock *P = *PI;
4045 TerminatorInst *PTI = P->getTerminator();
4046 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4047 if (BI->isUnconditional())
4048 UncondBranchPreds.push_back(P);
4050 CondBranchPreds.push_back(BI);
4054 // If we found some, do the transformation!
4055 if (!UncondBranchPreds.empty() && DupRet) {
4056 while (!UncondBranchPreds.empty()) {
4057 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4058 DEBUG(dbgs() << "FOLDING: " << *BB
4059 << "INTO UNCOND BRANCH PRED: " << *Pred);
4060 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4063 // If we eliminated all predecessors of the block, delete the block now.
4064 if (pred_empty(BB)) {
4065 // We know there are no successors, so just nuke the block.
4066 BB->eraseFromParent();
4068 LoopHeaders->erase(BB);
4074 // Check out all of the conditional branches going to this return
4075 // instruction. If any of them just select between returns, change the
4076 // branch itself into a select/return pair.
4077 while (!CondBranchPreds.empty()) {
4078 BranchInst *BI = CondBranchPreds.pop_back_val();
4080 // Check to see if the non-BB successor is also a return block.
4081 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4082 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4083 SimplifyCondBranchToTwoReturns(BI, Builder))
4089 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4090 BasicBlock *BB = UI->getParent();
4092 bool Changed = false;
4094 // If there are any instructions immediately before the unreachable that can
4095 // be removed, do so.
4096 while (UI->getIterator() != BB->begin()) {
4097 BasicBlock::iterator BBI = UI->getIterator();
4099 // Do not delete instructions that can have side effects which might cause
4100 // the unreachable to not be reachable; specifically, calls and volatile
4101 // operations may have this effect.
4102 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4105 if (BBI->mayHaveSideEffects()) {
4106 if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4107 if (SI->isVolatile())
4109 } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4110 if (LI->isVolatile())
4112 } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4113 if (RMWI->isVolatile())
4115 } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4116 if (CXI->isVolatile())
4118 } else if (isa<CatchPadInst>(BBI)) {
4119 // A catchpad may invoke exception object constructors and such, which
4120 // in some languages can be arbitrary code, so be conservative by
4122 // For CoreCLR, it just involves a type test, so can be removed.
4123 if (classifyEHPersonality(BB->getParent()->getPersonalityFn()) !=
4124 EHPersonality::CoreCLR)
4126 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4127 !isa<LandingPadInst>(BBI)) {
4130 // Note that deleting LandingPad's here is in fact okay, although it
4131 // involves a bit of subtle reasoning. If this inst is a LandingPad,
4132 // all the predecessors of this block will be the unwind edges of Invokes,
4133 // and we can therefore guarantee this block will be erased.
4136 // Delete this instruction (any uses are guaranteed to be dead)
4137 if (!BBI->use_empty())
4138 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4139 BBI->eraseFromParent();
4143 // If the unreachable instruction is the first in the block, take a gander
4144 // at all of the predecessors of this instruction, and simplify them.
4145 if (&BB->front() != UI)
4148 SmallVector<BasicBlock *, 8> Preds(pred_begin(BB), pred_end(BB));
4149 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4150 TerminatorInst *TI = Preds[i]->getTerminator();
4151 IRBuilder<> Builder(TI);
4152 if (auto *BI = dyn_cast<BranchInst>(TI)) {
4153 if (BI->isUnconditional()) {
4154 if (BI->getSuccessor(0) == BB) {
4155 new UnreachableInst(TI->getContext(), TI);
4156 TI->eraseFromParent();
4160 if (BI->getSuccessor(0) == BB) {
4161 Builder.CreateBr(BI->getSuccessor(1));
4162 EraseTerminatorInstAndDCECond(BI);
4163 } else if (BI->getSuccessor(1) == BB) {
4164 Builder.CreateBr(BI->getSuccessor(0));
4165 EraseTerminatorInstAndDCECond(BI);
4169 } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4170 for (auto i = SI->case_begin(), e = SI->case_end(); i != e;) {
4171 if (i->getCaseSuccessor() != BB) {
4175 BB->removePredecessor(SI->getParent());
4176 i = SI->removeCase(i);
4180 } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4181 if (II->getUnwindDest() == BB) {
4182 removeUnwindEdge(TI->getParent());
4185 } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4186 if (CSI->getUnwindDest() == BB) {
4187 removeUnwindEdge(TI->getParent());
4192 for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4193 E = CSI->handler_end();
4196 CSI->removeHandler(I);
4202 if (CSI->getNumHandlers() == 0) {
4203 BasicBlock *CatchSwitchBB = CSI->getParent();
4204 if (CSI->hasUnwindDest()) {
4205 // Redirect preds to the unwind dest
4206 CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4208 // Rewrite all preds to unwind to caller (or from invoke to call).
4209 SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4210 for (BasicBlock *EHPred : EHPreds)
4211 removeUnwindEdge(EHPred);
4213 // The catchswitch is no longer reachable.
4214 new UnreachableInst(CSI->getContext(), CSI);
4215 CSI->eraseFromParent();
4218 } else if (isa<CleanupReturnInst>(TI)) {
4219 new UnreachableInst(TI->getContext(), TI);
4220 TI->eraseFromParent();
4225 // If this block is now dead, remove it.
4226 if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4227 // We know there are no successors, so just nuke the block.
4228 BB->eraseFromParent();
4230 LoopHeaders->erase(BB);
4237 static bool CasesAreContiguous(SmallVectorImpl<ConstantInt *> &Cases) {
4238 assert(Cases.size() >= 1);
4240 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4241 for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4242 if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4248 /// Turn a switch with two reachable destinations into an integer range
4249 /// comparison and branch.
4250 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
4251 assert(SI->getNumCases() > 1 && "Degenerate switch?");
4254 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4256 // Partition the cases into two sets with different destinations.
4257 BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4258 BasicBlock *DestB = nullptr;
4259 SmallVector<ConstantInt *, 16> CasesA;
4260 SmallVector<ConstantInt *, 16> CasesB;
4262 for (auto Case : SI->cases()) {
4263 BasicBlock *Dest = Case.getCaseSuccessor();
4266 if (Dest == DestA) {
4267 CasesA.push_back(Case.getCaseValue());
4272 if (Dest == DestB) {
4273 CasesB.push_back(Case.getCaseValue());
4276 return false; // More than two destinations.
4279 assert(DestA && DestB &&
4280 "Single-destination switch should have been folded.");
4281 assert(DestA != DestB);
4282 assert(DestB != SI->getDefaultDest());
4283 assert(!CasesB.empty() && "There must be non-default cases.");
4284 assert(!CasesA.empty() || HasDefault);
4286 // Figure out if one of the sets of cases form a contiguous range.
4287 SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4288 BasicBlock *ContiguousDest = nullptr;
4289 BasicBlock *OtherDest = nullptr;
4290 if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4291 ContiguousCases = &CasesA;
4292 ContiguousDest = DestA;
4294 } else if (CasesAreContiguous(CasesB)) {
4295 ContiguousCases = &CasesB;
4296 ContiguousDest = DestB;
4301 // Start building the compare and branch.
4303 Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4304 Constant *NumCases =
4305 ConstantInt::get(Offset->getType(), ContiguousCases->size());
4307 Value *Sub = SI->getCondition();
4308 if (!Offset->isNullValue())
4309 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4312 // If NumCases overflowed, then all possible values jump to the successor.
4313 if (NumCases->isNullValue() && !ContiguousCases->empty())
4314 Cmp = ConstantInt::getTrue(SI->getContext());
4316 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4317 BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4319 // Update weight for the newly-created conditional branch.
4320 if (HasBranchWeights(SI)) {
4321 SmallVector<uint64_t, 8> Weights;
4322 GetBranchWeights(SI, Weights);
4323 if (Weights.size() == 1 + SI->getNumCases()) {
4324 uint64_t TrueWeight = 0;
4325 uint64_t FalseWeight = 0;
4326 for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4327 if (SI->getSuccessor(I) == ContiguousDest)
4328 TrueWeight += Weights[I];
4330 FalseWeight += Weights[I];
4332 while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4336 NewBI->setMetadata(LLVMContext::MD_prof,
4337 MDBuilder(SI->getContext())
4338 .createBranchWeights((uint32_t)TrueWeight,
4339 (uint32_t)FalseWeight));
4343 // Prune obsolete incoming values off the successors' PHI nodes.
4344 for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4345 unsigned PreviousEdges = ContiguousCases->size();
4346 if (ContiguousDest == SI->getDefaultDest())
4348 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4349 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4351 for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4352 unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4353 if (OtherDest == SI->getDefaultDest())
4355 for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4356 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4360 SI->eraseFromParent();
4365 /// Compute masked bits for the condition of a switch
4366 /// and use it to remove dead cases.
4367 static bool EliminateDeadSwitchCases(SwitchInst *SI, AssumptionCache *AC,
4368 const DataLayout &DL) {
4369 Value *Cond = SI->getCondition();
4370 unsigned Bits = Cond->getType()->getIntegerBitWidth();
4371 KnownBits Known(Bits);
4372 computeKnownBits(Cond, Known, DL, 0, AC, SI);
4374 // We can also eliminate cases by determining that their values are outside of
4375 // the limited range of the condition based on how many significant (non-sign)
4376 // bits are in the condition value.
4377 unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4378 unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4380 // Gather dead cases.
4381 SmallVector<ConstantInt *, 8> DeadCases;
4382 for (auto &Case : SI->cases()) {
4383 APInt CaseVal = Case.getCaseValue()->getValue();
4384 if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4385 (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4386 DeadCases.push_back(Case.getCaseValue());
4387 DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal << " is dead.\n");
4391 // If we can prove that the cases must cover all possible values, the
4392 // default destination becomes dead and we can remove it. If we know some
4393 // of the bits in the value, we can use that to more precisely compute the
4394 // number of possible unique case values.
4396 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4397 const unsigned NumUnknownBits =
4398 Bits - (Known.Zero | Known.One).countPopulation();
4399 assert(NumUnknownBits <= Bits);
4400 if (HasDefault && DeadCases.empty() &&
4401 NumUnknownBits < 64 /* avoid overflow */ &&
4402 SI->getNumCases() == (1ULL << NumUnknownBits)) {
4403 DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4404 BasicBlock *NewDefault =
4405 SplitBlockPredecessors(SI->getDefaultDest(), SI->getParent(), "");
4406 SI->setDefaultDest(&*NewDefault);
4407 SplitBlock(&*NewDefault, &NewDefault->front());
4408 auto *OldTI = NewDefault->getTerminator();
4409 new UnreachableInst(SI->getContext(), OldTI);
4410 EraseTerminatorInstAndDCECond(OldTI);
4414 SmallVector<uint64_t, 8> Weights;
4415 bool HasWeight = HasBranchWeights(SI);
4417 GetBranchWeights(SI, Weights);
4418 HasWeight = (Weights.size() == 1 + SI->getNumCases());
4421 // Remove dead cases from the switch.
4422 for (ConstantInt *DeadCase : DeadCases) {
4423 SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4424 assert(CaseI != SI->case_default() &&
4425 "Case was not found. Probably mistake in DeadCases forming.");
4427 std::swap(Weights[CaseI->getCaseIndex() + 1], Weights.back());
4431 // Prune unused values from PHI nodes.
4432 CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4433 SI->removeCase(CaseI);
4435 if (HasWeight && Weights.size() >= 2) {
4436 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
4437 SI->setMetadata(LLVMContext::MD_prof,
4438 MDBuilder(SI->getParent()->getContext())
4439 .createBranchWeights(MDWeights));
4442 return !DeadCases.empty();
4445 /// If BB would be eligible for simplification by
4446 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4447 /// by an unconditional branch), look at the phi node for BB in the successor
4448 /// block and see if the incoming value is equal to CaseValue. If so, return
4449 /// the phi node, and set PhiIndex to BB's index in the phi node.
4450 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
4451 BasicBlock *BB, int *PhiIndex) {
4452 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4453 return nullptr; // BB must be empty to be a candidate for simplification.
4454 if (!BB->getSinglePredecessor())
4455 return nullptr; // BB must be dominated by the switch.
4457 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
4458 if (!Branch || !Branch->isUnconditional())
4459 return nullptr; // Terminator must be unconditional branch.
4461 BasicBlock *Succ = Branch->getSuccessor(0);
4463 BasicBlock::iterator I = Succ->begin();
4464 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
4465 int Idx = PHI->getBasicBlockIndex(BB);
4466 assert(Idx >= 0 && "PHI has no entry for predecessor?");
4468 Value *InValue = PHI->getIncomingValue(Idx);
4469 if (InValue != CaseValue)
4479 /// Try to forward the condition of a switch instruction to a phi node
4480 /// dominated by the switch, if that would mean that some of the destination
4481 /// blocks of the switch can be folded away.
4482 /// Returns true if a change is made.
4483 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
4484 typedef DenseMap<PHINode *, SmallVector<int, 4>> ForwardingNodesMap;
4485 ForwardingNodesMap ForwardingNodes;
4487 for (auto Case : SI->cases()) {
4488 ConstantInt *CaseValue = Case.getCaseValue();
4489 BasicBlock *CaseDest = Case.getCaseSuccessor();
4493 FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIndex);
4497 ForwardingNodes[PHI].push_back(PhiIndex);
4500 bool Changed = false;
4502 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
4503 E = ForwardingNodes.end();
4505 PHINode *Phi = I->first;
4506 SmallVectorImpl<int> &Indexes = I->second;
4508 if (Indexes.size() < 2)
4511 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
4512 Phi->setIncomingValue(Indexes[I], SI->getCondition());
4519 /// Return true if the backend will be able to handle
4520 /// initializing an array of constants like C.
4521 static bool ValidLookupTableConstant(Constant *C, const TargetTransformInfo &TTI) {
4522 if (C->isThreadDependent())
4524 if (C->isDLLImportDependent())
4527 if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4528 !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4529 !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4532 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4533 if (!CE->isGEPWithNoNotionalOverIndexing())
4535 if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4539 if (!TTI.shouldBuildLookupTablesForConstant(C))
4545 /// If V is a Constant, return it. Otherwise, try to look up
4546 /// its constant value in ConstantPool, returning 0 if it's not there.
4548 LookupConstant(Value *V,
4549 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4550 if (Constant *C = dyn_cast<Constant>(V))
4552 return ConstantPool.lookup(V);
4555 /// Try to fold instruction I into a constant. This works for
4556 /// simple instructions such as binary operations where both operands are
4557 /// constant or can be replaced by constants from the ConstantPool. Returns the
4558 /// resulting constant on success, 0 otherwise.
4560 ConstantFold(Instruction *I, const DataLayout &DL,
4561 const SmallDenseMap<Value *, Constant *> &ConstantPool) {
4562 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4563 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4566 if (A->isAllOnesValue())
4567 return LookupConstant(Select->getTrueValue(), ConstantPool);
4568 if (A->isNullValue())
4569 return LookupConstant(Select->getFalseValue(), ConstantPool);
4573 SmallVector<Constant *, 4> COps;
4574 for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4575 if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
4581 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4582 return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4586 return ConstantFoldInstOperands(I, COps, DL);
4589 /// Try to determine the resulting constant values in phi nodes
4590 /// at the common destination basic block, *CommonDest, for one of the case
4591 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4592 /// case), of a switch instruction SI.
4594 GetCaseResults(SwitchInst *SI, ConstantInt *CaseVal, BasicBlock *CaseDest,
4595 BasicBlock **CommonDest,
4596 SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4597 const DataLayout &DL, const TargetTransformInfo &TTI) {
4598 // The block from which we enter the common destination.
4599 BasicBlock *Pred = SI->getParent();
4601 // If CaseDest is empty except for some side-effect free instructions through
4602 // which we can constant-propagate the CaseVal, continue to its successor.
4603 SmallDenseMap<Value *, Constant *> ConstantPool;
4604 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4605 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
4607 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
4608 // If the terminator is a simple branch, continue to the next block.
4609 if (T->getNumSuccessors() != 1 || T->isExceptional())
4612 CaseDest = T->getSuccessor(0);
4613 } else if (isa<DbgInfoIntrinsic>(I)) {
4614 // Skip debug intrinsic.
4616 } else if (Constant *C = ConstantFold(&*I, DL, ConstantPool)) {
4617 // Instruction is side-effect free and constant.
4619 // If the instruction has uses outside this block or a phi node slot for
4620 // the block, it is not safe to bypass the instruction since it would then
4621 // no longer dominate all its uses.
4622 for (auto &Use : I->uses()) {
4623 User *User = Use.getUser();
4624 if (Instruction *I = dyn_cast<Instruction>(User))
4625 if (I->getParent() == CaseDest)
4627 if (PHINode *Phi = dyn_cast<PHINode>(User))
4628 if (Phi->getIncomingBlock(Use) == CaseDest)
4633 ConstantPool.insert(std::make_pair(&*I, C));
4639 // If we did not have a CommonDest before, use the current one.
4641 *CommonDest = CaseDest;
4642 // If the destination isn't the common one, abort.
4643 if (CaseDest != *CommonDest)
4646 // Get the values for this case from phi nodes in the destination block.
4647 BasicBlock::iterator I = (*CommonDest)->begin();
4648 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
4649 int Idx = PHI->getBasicBlockIndex(Pred);
4653 Constant *ConstVal =
4654 LookupConstant(PHI->getIncomingValue(Idx), ConstantPool);
4658 // Be conservative about which kinds of constants we support.
4659 if (!ValidLookupTableConstant(ConstVal, TTI))
4662 Res.push_back(std::make_pair(PHI, ConstVal));
4665 return Res.size() > 0;
4668 // Helper function used to add CaseVal to the list of cases that generate
4670 static void MapCaseToResult(ConstantInt *CaseVal,
4671 SwitchCaseResultVectorTy &UniqueResults,
4673 for (auto &I : UniqueResults) {
4674 if (I.first == Result) {
4675 I.second.push_back(CaseVal);
4679 UniqueResults.push_back(
4680 std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4683 // Helper function that initializes a map containing
4684 // results for the PHI node of the common destination block for a switch
4685 // instruction. Returns false if multiple PHI nodes have been found or if
4686 // there is not a common destination block for the switch.
4687 static bool InitializeUniqueCases(SwitchInst *SI, PHINode *&PHI,
4688 BasicBlock *&CommonDest,
4689 SwitchCaseResultVectorTy &UniqueResults,
4690 Constant *&DefaultResult,
4691 const DataLayout &DL,
4692 const TargetTransformInfo &TTI) {
4693 for (auto &I : SI->cases()) {
4694 ConstantInt *CaseVal = I.getCaseValue();
4696 // Resulting value at phi nodes for this case value.
4697 SwitchCaseResultsTy Results;
4698 if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4702 // Only one value per case is permitted
4703 if (Results.size() > 1)
4705 MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4707 // Check the PHI consistency.
4709 PHI = Results[0].first;
4710 else if (PHI != Results[0].first)
4713 // Find the default result value.
4714 SmallVector<std::pair<PHINode *, Constant *>, 1> DefaultResults;
4715 BasicBlock *DefaultDest = SI->getDefaultDest();
4716 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4718 // If the default value is not found abort unless the default destination
4721 DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4722 if ((!DefaultResult &&
4723 !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4729 // Helper function that checks if it is possible to transform a switch with only
4730 // two cases (or two cases + default) that produces a result into a select.
4733 // case 10: %0 = icmp eq i32 %a, 10
4734 // return 10; %1 = select i1 %0, i32 10, i32 4
4735 // case 20: ----> %2 = icmp eq i32 %a, 20
4736 // return 2; %3 = select i1 %2, i32 2, i32 %1
4740 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4741 Constant *DefaultResult, Value *Condition,
4742 IRBuilder<> &Builder) {
4743 assert(ResultVector.size() == 2 &&
4744 "We should have exactly two unique results at this point");
4745 // If we are selecting between only two cases transform into a simple
4746 // select or a two-way select if default is possible.
4747 if (ResultVector[0].second.size() == 1 &&
4748 ResultVector[1].second.size() == 1) {
4749 ConstantInt *const FirstCase = ResultVector[0].second[0];
4750 ConstantInt *const SecondCase = ResultVector[1].second[0];
4752 bool DefaultCanTrigger = DefaultResult;
4753 Value *SelectValue = ResultVector[1].first;
4754 if (DefaultCanTrigger) {
4755 Value *const ValueCompare =
4756 Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4757 SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4758 DefaultResult, "switch.select");
4760 Value *const ValueCompare =
4761 Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4762 return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4763 SelectValue, "switch.select");
4769 // Helper function to cleanup a switch instruction that has been converted into
4770 // a select, fixing up PHI nodes and basic blocks.
4771 static void RemoveSwitchAfterSelectConversion(SwitchInst *SI, PHINode *PHI,
4773 IRBuilder<> &Builder) {
4774 BasicBlock *SelectBB = SI->getParent();
4775 while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4776 PHI->removeIncomingValue(SelectBB);
4777 PHI->addIncoming(SelectValue, SelectBB);
4779 Builder.CreateBr(PHI->getParent());
4781 // Remove the switch.
4782 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4783 BasicBlock *Succ = SI->getSuccessor(i);
4785 if (Succ == PHI->getParent())
4787 Succ->removePredecessor(SelectBB);
4789 SI->eraseFromParent();
4792 /// If the switch is only used to initialize one or more
4793 /// phi nodes in a common successor block with only two different
4794 /// constant values, replace the switch with select.
4795 static bool SwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4796 AssumptionCache *AC, const DataLayout &DL,
4797 const TargetTransformInfo &TTI) {
4798 Value *const Cond = SI->getCondition();
4799 PHINode *PHI = nullptr;
4800 BasicBlock *CommonDest = nullptr;
4801 Constant *DefaultResult;
4802 SwitchCaseResultVectorTy UniqueResults;
4803 // Collect all the cases that will deliver the same value from the switch.
4804 if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4807 // Selects choose between maximum two values.
4808 if (UniqueResults.size() != 2)
4810 assert(PHI != nullptr && "PHI for value select not found");
4812 Builder.SetInsertPoint(SI);
4813 Value *SelectValue =
4814 ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4816 RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4819 // The switch couldn't be converted into a select.
4825 /// This class represents a lookup table that can be used to replace a switch.
4826 class SwitchLookupTable {
4828 /// Create a lookup table to use as a switch replacement with the contents
4829 /// of Values, using DefaultValue to fill any holes in the table.
4831 Module &M, uint64_t TableSize, ConstantInt *Offset,
4832 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4833 Constant *DefaultValue, const DataLayout &DL);
4835 /// Build instructions with Builder to retrieve the value at
4836 /// the position given by Index in the lookup table.
4837 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4839 /// Return true if a table with TableSize elements of
4840 /// type ElementType would fit in a target-legal register.
4841 static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4845 // Depending on the contents of the table, it can be represented in
4848 // For tables where each element contains the same value, we just have to
4849 // store that single value and return it for each lookup.
4852 // For tables where there is a linear relationship between table index
4853 // and values. We calculate the result with a simple multiplication
4854 // and addition instead of a table lookup.
4857 // For small tables with integer elements, we can pack them into a bitmap
4858 // that fits into a target-legal register. Values are retrieved by
4859 // shift and mask operations.
4862 // The table is stored as an array of values. Values are retrieved by load
4863 // instructions from the table.
4867 // For SingleValueKind, this is the single value.
4868 Constant *SingleValue;
4870 // For BitMapKind, this is the bitmap.
4871 ConstantInt *BitMap;
4872 IntegerType *BitMapElementTy;
4874 // For LinearMapKind, these are the constants used to derive the value.
4875 ConstantInt *LinearOffset;
4876 ConstantInt *LinearMultiplier;
4878 // For ArrayKind, this is the array.
4879 GlobalVariable *Array;
4882 } // end anonymous namespace
4884 SwitchLookupTable::SwitchLookupTable(
4885 Module &M, uint64_t TableSize, ConstantInt *Offset,
4886 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4887 Constant *DefaultValue, const DataLayout &DL)
4888 : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
4889 LinearOffset(nullptr), LinearMultiplier(nullptr), Array(nullptr) {
4890 assert(Values.size() && "Can't build lookup table without values!");
4891 assert(TableSize >= Values.size() && "Can't fit values in table!");
4893 // If all values in the table are equal, this is that value.
4894 SingleValue = Values.begin()->second;
4896 Type *ValueType = Values.begin()->second->getType();
4898 // Build up the table contents.
4899 SmallVector<Constant *, 64> TableContents(TableSize);
4900 for (size_t I = 0, E = Values.size(); I != E; ++I) {
4901 ConstantInt *CaseVal = Values[I].first;
4902 Constant *CaseRes = Values[I].second;
4903 assert(CaseRes->getType() == ValueType);
4905 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4906 TableContents[Idx] = CaseRes;
4908 if (CaseRes != SingleValue)
4909 SingleValue = nullptr;
4912 // Fill in any holes in the table with the default result.
4913 if (Values.size() < TableSize) {
4914 assert(DefaultValue &&
4915 "Need a default value to fill the lookup table holes.");
4916 assert(DefaultValue->getType() == ValueType);
4917 for (uint64_t I = 0; I < TableSize; ++I) {
4918 if (!TableContents[I])
4919 TableContents[I] = DefaultValue;
4922 if (DefaultValue != SingleValue)
4923 SingleValue = nullptr;
4926 // If each element in the table contains the same value, we only need to store
4927 // that single value.
4929 Kind = SingleValueKind;
4933 // Check if we can derive the value with a linear transformation from the
4935 if (isa<IntegerType>(ValueType)) {
4936 bool LinearMappingPossible = true;
4939 assert(TableSize >= 2 && "Should be a SingleValue table.");
4940 // Check if there is the same distance between two consecutive values.
4941 for (uint64_t I = 0; I < TableSize; ++I) {
4942 ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4944 // This is an undef. We could deal with it, but undefs in lookup tables
4945 // are very seldom. It's probably not worth the additional complexity.
4946 LinearMappingPossible = false;
4949 APInt Val = ConstVal->getValue();
4951 APInt Dist = Val - PrevVal;
4954 } else if (Dist != DistToPrev) {
4955 LinearMappingPossible = false;
4961 if (LinearMappingPossible) {
4962 LinearOffset = cast<ConstantInt>(TableContents[0]);
4963 LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
4964 Kind = LinearMapKind;
4970 // If the type is integer and the table fits in a register, build a bitmap.
4971 if (WouldFitInRegister(DL, TableSize, ValueType)) {
4972 IntegerType *IT = cast<IntegerType>(ValueType);
4973 APInt TableInt(TableSize * IT->getBitWidth(), 0);
4974 for (uint64_t I = TableSize; I > 0; --I) {
4975 TableInt <<= IT->getBitWidth();
4976 // Insert values into the bitmap. Undef values are set to zero.
4977 if (!isa<UndefValue>(TableContents[I - 1])) {
4978 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
4979 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
4982 BitMap = ConstantInt::get(M.getContext(), TableInt);
4983 BitMapElementTy = IT;
4989 // Store the table in an array.
4990 ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
4991 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
4993 Array = new GlobalVariable(M, ArrayTy, /*constant=*/true,
4994 GlobalVariable::PrivateLinkage, Initializer,
4996 Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
5000 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
5002 case SingleValueKind:
5004 case LinearMapKind: {
5005 // Derive the result value from the input value.
5006 Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
5007 false, "switch.idx.cast");
5008 if (!LinearMultiplier->isOne())
5009 Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
5010 if (!LinearOffset->isZero())
5011 Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
5015 // Type of the bitmap (e.g. i59).
5016 IntegerType *MapTy = BitMap->getType();
5018 // Cast Index to the same type as the bitmap.
5019 // Note: The Index is <= the number of elements in the table, so
5020 // truncating it to the width of the bitmask is safe.
5021 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
5023 // Multiply the shift amount by the element width.
5024 ShiftAmt = Builder.CreateMul(
5025 ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5029 Value *DownShifted =
5030 Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5032 return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5035 // Make sure the table index will not overflow when treated as signed.
5036 IntegerType *IT = cast<IntegerType>(Index->getType());
5037 uint64_t TableSize =
5038 Array->getInitializer()->getType()->getArrayNumElements();
5039 if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5040 Index = Builder.CreateZExt(
5041 Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5042 "switch.tableidx.zext");
5044 Value *GEPIndices[] = {Builder.getInt32(0), Index};
5045 Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5046 GEPIndices, "switch.gep");
5047 return Builder.CreateLoad(GEP, "switch.load");
5050 llvm_unreachable("Unknown lookup table kind!");
5053 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5055 Type *ElementType) {
5056 auto *IT = dyn_cast<IntegerType>(ElementType);
5059 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5060 // are <= 15, we could try to narrow the type.
5062 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5063 if (TableSize >= UINT_MAX / IT->getBitWidth())
5065 return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5068 /// Determine whether a lookup table should be built for this switch, based on
5069 /// the number of cases, size of the table, and the types of the results.
5071 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5072 const TargetTransformInfo &TTI, const DataLayout &DL,
5073 const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5074 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5075 return false; // TableSize overflowed, or mul below might overflow.
5077 bool AllTablesFitInRegister = true;
5078 bool HasIllegalType = false;
5079 for (const auto &I : ResultTypes) {
5080 Type *Ty = I.second;
5082 // Saturate this flag to true.
5083 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5085 // Saturate this flag to false.
5086 AllTablesFitInRegister =
5087 AllTablesFitInRegister &&
5088 SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5090 // If both flags saturate, we're done. NOTE: This *only* works with
5091 // saturating flags, and all flags have to saturate first due to the
5092 // non-deterministic behavior of iterating over a dense map.
5093 if (HasIllegalType && !AllTablesFitInRegister)
5097 // If each table would fit in a register, we should build it anyway.
5098 if (AllTablesFitInRegister)
5101 // Don't build a table that doesn't fit in-register if it has illegal types.
5105 // The table density should be at least 40%. This is the same criterion as for
5106 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5107 // FIXME: Find the best cut-off.
5108 return SI->getNumCases() * 10 >= TableSize * 4;
5111 /// Try to reuse the switch table index compare. Following pattern:
5113 /// if (idx < tablesize)
5114 /// r = table[idx]; // table does not contain default_value
5116 /// r = default_value;
5117 /// if (r != default_value)
5120 /// Is optimized to:
5122 /// cond = idx < tablesize;
5126 /// r = default_value;
5130 /// Jump threading will then eliminate the second if(cond).
5131 static void reuseTableCompare(
5132 User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5133 Constant *DefaultValue,
5134 const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5136 ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5140 // We require that the compare is in the same block as the phi so that jump
5141 // threading can do its work afterwards.
5142 if (CmpInst->getParent() != PhiBlock)
5145 Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5149 Value *RangeCmp = RangeCheckBranch->getCondition();
5150 Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5151 Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5153 // Check if the compare with the default value is constant true or false.
5154 Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5155 DefaultValue, CmpOp1, true);
5156 if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5159 // Check if the compare with the case values is distinct from the default
5161 for (auto ValuePair : Values) {
5162 Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5163 ValuePair.second, CmpOp1, true);
5164 if (!CaseConst || CaseConst == DefaultConst)
5166 assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5167 "Expect true or false as compare result.");
5170 // Check if the branch instruction dominates the phi node. It's a simple
5171 // dominance check, but sufficient for our needs.
5172 // Although this check is invariant in the calling loops, it's better to do it
5173 // at this late stage. Practically we do it at most once for a switch.
5174 BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5175 for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5176 BasicBlock *Pred = *PI;
5177 if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5181 if (DefaultConst == FalseConst) {
5182 // The compare yields the same result. We can replace it.
5183 CmpInst->replaceAllUsesWith(RangeCmp);
5184 ++NumTableCmpReuses;
5186 // The compare yields the same result, just inverted. We can replace it.
5187 Value *InvertedTableCmp = BinaryOperator::CreateXor(
5188 RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5190 CmpInst->replaceAllUsesWith(InvertedTableCmp);
5191 ++NumTableCmpReuses;
5195 /// If the switch is only used to initialize one or more phi nodes in a common
5196 /// successor block with different constant values, replace the switch with
5198 static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
5199 const DataLayout &DL,
5200 const TargetTransformInfo &TTI) {
5201 assert(SI->getNumCases() > 1 && "Degenerate switch?");
5203 // Only build lookup table when we have a target that supports it.
5204 if (!TTI.shouldBuildLookupTables())
5207 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5208 // split off a dense part and build a lookup table for that.
5210 // FIXME: This creates arrays of GEPs to constant strings, which means each
5211 // GEP needs a runtime relocation in PIC code. We should just build one big
5212 // string and lookup indices into that.
5214 // Ignore switches with less than three cases. Lookup tables will not make
5216 // faster, so we don't analyze them.
5217 if (SI->getNumCases() < 3)
5220 // Figure out the corresponding result for each case value and phi node in the
5221 // common destination, as well as the min and max case values.
5222 assert(SI->case_begin() != SI->case_end());
5223 SwitchInst::CaseIt CI = SI->case_begin();
5224 ConstantInt *MinCaseVal = CI->getCaseValue();
5225 ConstantInt *MaxCaseVal = CI->getCaseValue();
5227 BasicBlock *CommonDest = nullptr;
5228 typedef SmallVector<std::pair<ConstantInt *, Constant *>, 4> ResultListTy;
5229 SmallDenseMap<PHINode *, ResultListTy> ResultLists;
5230 SmallDenseMap<PHINode *, Constant *> DefaultResults;
5231 SmallDenseMap<PHINode *, Type *> ResultTypes;
5232 SmallVector<PHINode *, 4> PHIs;
5234 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5235 ConstantInt *CaseVal = CI->getCaseValue();
5236 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5237 MinCaseVal = CaseVal;
5238 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5239 MaxCaseVal = CaseVal;
5241 // Resulting value at phi nodes for this case value.
5242 typedef SmallVector<std::pair<PHINode *, Constant *>, 4> ResultsTy;
5244 if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5248 // Append the result from this case to the list for each phi.
5249 for (const auto &I : Results) {
5250 PHINode *PHI = I.first;
5251 Constant *Value = I.second;
5252 if (!ResultLists.count(PHI))
5253 PHIs.push_back(PHI);
5254 ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5258 // Keep track of the result types.
5259 for (PHINode *PHI : PHIs) {
5260 ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5263 uint64_t NumResults = ResultLists[PHIs[0]].size();
5264 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5265 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5266 bool TableHasHoles = (NumResults < TableSize);
5268 // If the table has holes, we need a constant result for the default case
5269 // or a bitmask that fits in a register.
5270 SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5271 bool HasDefaultResults =
5272 GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5273 DefaultResultsList, DL, TTI);
5275 bool NeedMask = (TableHasHoles && !HasDefaultResults);
5277 // As an extra penalty for the validity test we require more cases.
5278 if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5280 if (!DL.fitsInLegalInteger(TableSize))
5284 for (const auto &I : DefaultResultsList) {
5285 PHINode *PHI = I.first;
5286 Constant *Result = I.second;
5287 DefaultResults[PHI] = Result;
5290 if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5293 // Create the BB that does the lookups.
5294 Module &Mod = *CommonDest->getParent()->getParent();
5295 BasicBlock *LookupBB = BasicBlock::Create(
5296 Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5298 // Compute the table index value.
5299 Builder.SetInsertPoint(SI);
5301 Builder.CreateSub(SI->getCondition(), MinCaseVal, "switch.tableidx");
5303 // Compute the maximum table size representable by the integer type we are
5305 unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5306 uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5307 assert(MaxTableSize >= TableSize &&
5308 "It is impossible for a switch to have more entries than the max "
5309 "representable value of its input integer type's size.");
5311 // If the default destination is unreachable, or if the lookup table covers
5312 // all values of the conditional variable, branch directly to the lookup table
5313 // BB. Otherwise, check that the condition is within the case range.
5314 const bool DefaultIsReachable =
5315 !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5316 const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5317 BranchInst *RangeCheckBranch = nullptr;
5319 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5320 Builder.CreateBr(LookupBB);
5321 // Note: We call removeProdecessor later since we need to be able to get the
5322 // PHI value for the default case in case we're using a bit mask.
5324 Value *Cmp = Builder.CreateICmpULT(
5325 TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5327 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5330 // Populate the BB that does the lookups.
5331 Builder.SetInsertPoint(LookupBB);
5334 // Before doing the lookup we do the hole check.
5335 // The LookupBB is therefore re-purposed to do the hole check
5336 // and we create a new LookupBB.
5337 BasicBlock *MaskBB = LookupBB;
5338 MaskBB->setName("switch.hole_check");
5339 LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5340 CommonDest->getParent(), CommonDest);
5342 // Make the mask's bitwidth at least 8bit and a power-of-2 to avoid
5343 // unnecessary illegal types.
5344 uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5345 APInt MaskInt(TableSizePowOf2, 0);
5346 APInt One(TableSizePowOf2, 1);
5347 // Build bitmask; fill in a 1 bit for every case.
5348 const ResultListTy &ResultList = ResultLists[PHIs[0]];
5349 for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5350 uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5352 MaskInt |= One << Idx;
5354 ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5356 // Get the TableIndex'th bit of the bitmask.
5357 // If this bit is 0 (meaning hole) jump to the default destination,
5358 // else continue with table lookup.
5359 IntegerType *MapTy = TableMask->getType();
5361 Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5362 Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5363 Value *LoBit = Builder.CreateTrunc(
5364 Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5365 Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5367 Builder.SetInsertPoint(LookupBB);
5368 AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5371 if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5372 // We cached PHINodes in PHIs, to avoid accessing deleted PHINodes later,
5373 // do not delete PHINodes here.
5374 SI->getDefaultDest()->removePredecessor(SI->getParent(),
5375 /*DontDeleteUselessPHIs=*/true);
5378 bool ReturnedEarly = false;
5379 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
5380 PHINode *PHI = PHIs[I];
5381 const ResultListTy &ResultList = ResultLists[PHI];
5383 // If using a bitmask, use any value to fill the lookup table holes.
5384 Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5385 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL);
5387 Value *Result = Table.BuildLookup(TableIndex, Builder);
5389 // If the result is used to return immediately from the function, we want to
5390 // do that right here.
5391 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5392 PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5393 Builder.CreateRet(Result);
5394 ReturnedEarly = true;
5398 // Do a small peephole optimization: re-use the switch table compare if
5400 if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5401 BasicBlock *PhiBlock = PHI->getParent();
5402 // Search for compare instructions which use the phi.
5403 for (auto *User : PHI->users()) {
5404 reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5408 PHI->addIncoming(Result, LookupBB);
5412 Builder.CreateBr(CommonDest);
5414 // Remove the switch.
5415 for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5416 BasicBlock *Succ = SI->getSuccessor(i);
5418 if (Succ == SI->getDefaultDest())
5420 Succ->removePredecessor(SI->getParent());
5422 SI->eraseFromParent();
5426 ++NumLookupTablesHoles;
5430 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5431 // See also SelectionDAGBuilder::isDense(), which this function was based on.
5432 uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5433 uint64_t Range = Diff + 1;
5434 uint64_t NumCases = Values.size();
5435 // 40% is the default density for building a jump table in optsize/minsize mode.
5436 uint64_t MinDensity = 40;
5438 return NumCases * 100 >= Range * MinDensity;
5441 // Try and transform a switch that has "holes" in it to a contiguous sequence
5444 // A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5445 // range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5447 // This converts a sparse switch into a dense switch which allows better
5448 // lowering and could also allow transforming into a lookup table.
5449 static bool ReduceSwitchRange(SwitchInst *SI, IRBuilder<> &Builder,
5450 const DataLayout &DL,
5451 const TargetTransformInfo &TTI) {
5452 auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5453 if (CondTy->getIntegerBitWidth() > 64 ||
5454 !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5456 // Only bother with this optimization if there are more than 3 switch cases;
5457 // SDAG will only bother creating jump tables for 4 or more cases.
5458 if (SI->getNumCases() < 4)
5461 // This transform is agnostic to the signedness of the input or case values. We
5462 // can treat the case values as signed or unsigned. We can optimize more common
5463 // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5465 SmallVector<int64_t,4> Values;
5466 for (auto &C : SI->cases())
5467 Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5468 std::sort(Values.begin(), Values.end());
5470 // If the switch is already dense, there's nothing useful to do here.
5471 if (isSwitchDense(Values))
5474 // First, transform the values such that they start at zero and ascend.
5475 int64_t Base = Values[0];
5476 for (auto &V : Values)
5479 // Now we have signed numbers that have been shifted so that, given enough
5480 // precision, there are no negative values. Since the rest of the transform
5481 // is bitwise only, we switch now to an unsigned representation.
5483 for (auto &V : Values)
5484 GCD = GreatestCommonDivisor64(GCD, (uint64_t)V);
5486 // This transform can be done speculatively because it is so cheap - it results
5487 // in a single rotate operation being inserted. This can only happen if the
5488 // factor extracted is a power of 2.
5489 // FIXME: If the GCD is an odd number we can multiply by the multiplicative
5490 // inverse of GCD and then perform this transform.
5491 // FIXME: It's possible that optimizing a switch on powers of two might also
5492 // be beneficial - flag values are often powers of two and we could use a CLZ
5493 // as the key function.
5494 if (GCD <= 1 || !isPowerOf2_64(GCD))
5495 // No common divisor found or too expensive to compute key function.
5498 unsigned Shift = Log2_64(GCD);
5499 for (auto &V : Values)
5500 V = (int64_t)((uint64_t)V >> Shift);
5502 if (!isSwitchDense(Values))
5503 // Transform didn't create a dense switch.
5506 // The obvious transform is to shift the switch condition right and emit a
5507 // check that the condition actually cleanly divided by GCD, i.e.
5508 // C & (1 << Shift - 1) == 0
5509 // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5511 // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5512 // shift and puts the shifted-off bits in the uppermost bits. If any of these
5513 // are nonzero then the switch condition will be very large and will hit the
5516 auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5517 Builder.SetInsertPoint(SI);
5518 auto *ShiftC = ConstantInt::get(Ty, Shift);
5519 auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5520 auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5521 auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5522 auto *Rot = Builder.CreateOr(LShr, Shl);
5523 SI->replaceUsesOfWith(SI->getCondition(), Rot);
5525 for (auto Case : SI->cases()) {
5526 auto *Orig = Case.getCaseValue();
5527 auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5529 cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5534 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
5535 BasicBlock *BB = SI->getParent();
5537 if (isValueEqualityComparison(SI)) {
5538 // If we only have one predecessor, and if it is a branch on this value,
5539 // see if that predecessor totally determines the outcome of this switch.
5540 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5541 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
5542 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5544 Value *Cond = SI->getCondition();
5545 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
5546 if (SimplifySwitchOnSelect(SI, Select))
5547 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5549 // If the block only contains the switch, see if we can fold the block
5550 // away into any preds.
5551 BasicBlock::iterator BBI = BB->begin();
5552 // Ignore dbg intrinsics.
5553 while (isa<DbgInfoIntrinsic>(BBI))
5556 if (FoldValueComparisonIntoPredecessors(SI, Builder))
5557 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5560 // Try to transform the switch into an icmp and a branch.
5561 if (TurnSwitchRangeIntoICmp(SI, Builder))
5562 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5564 // Remove unreachable cases.
5565 if (EliminateDeadSwitchCases(SI, AC, DL))
5566 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5568 if (SwitchToSelect(SI, Builder, AC, DL, TTI))
5569 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5571 if (ForwardSwitchConditionToPHI(SI))
5572 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5574 // The conversion from switch to lookup tables results in difficult
5575 // to analyze code and makes pruning branches much harder.
5576 // This is a problem of the switch expression itself can still be
5577 // restricted as a result of inlining or CVP. There only apply this
5578 // transformation during late steps of the optimisation chain.
5579 if (LateSimplifyCFG && SwitchToLookupTable(SI, Builder, DL, TTI))
5580 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5582 if (ReduceSwitchRange(SI, Builder, DL, TTI))
5583 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5588 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
5589 BasicBlock *BB = IBI->getParent();
5590 bool Changed = false;
5592 // Eliminate redundant destinations.
5593 SmallPtrSet<Value *, 8> Succs;
5594 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
5595 BasicBlock *Dest = IBI->getDestination(i);
5596 if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
5597 Dest->removePredecessor(BB);
5598 IBI->removeDestination(i);
5605 if (IBI->getNumDestinations() == 0) {
5606 // If the indirectbr has no successors, change it to unreachable.
5607 new UnreachableInst(IBI->getContext(), IBI);
5608 EraseTerminatorInstAndDCECond(IBI);
5612 if (IBI->getNumDestinations() == 1) {
5613 // If the indirectbr has one successor, change it to a direct branch.
5614 BranchInst::Create(IBI->getDestination(0), IBI);
5615 EraseTerminatorInstAndDCECond(IBI);
5619 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
5620 if (SimplifyIndirectBrOnSelect(IBI, SI))
5621 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5626 /// Given an block with only a single landing pad and a unconditional branch
5627 /// try to find another basic block which this one can be merged with. This
5628 /// handles cases where we have multiple invokes with unique landing pads, but
5629 /// a shared handler.
5631 /// We specifically choose to not worry about merging non-empty blocks
5632 /// here. That is a PRE/scheduling problem and is best solved elsewhere. In
5633 /// practice, the optimizer produces empty landing pad blocks quite frequently
5634 /// when dealing with exception dense code. (see: instcombine, gvn, if-else
5635 /// sinking in this file)
5637 /// This is primarily a code size optimization. We need to avoid performing
5638 /// any transform which might inhibit optimization (such as our ability to
5639 /// specialize a particular handler via tail commoning). We do this by not
5640 /// merging any blocks which require us to introduce a phi. Since the same
5641 /// values are flowing through both blocks, we don't loose any ability to
5642 /// specialize. If anything, we make such specialization more likely.
5644 /// TODO - This transformation could remove entries from a phi in the target
5645 /// block when the inputs in the phi are the same for the two blocks being
5646 /// merged. In some cases, this could result in removal of the PHI entirely.
5647 static bool TryToMergeLandingPad(LandingPadInst *LPad, BranchInst *BI,
5649 auto Succ = BB->getUniqueSuccessor();
5651 // If there's a phi in the successor block, we'd likely have to introduce
5652 // a phi into the merged landing pad block.
5653 if (isa<PHINode>(*Succ->begin()))
5656 for (BasicBlock *OtherPred : predecessors(Succ)) {
5657 if (BB == OtherPred)
5659 BasicBlock::iterator I = OtherPred->begin();
5660 LandingPadInst *LPad2 = dyn_cast<LandingPadInst>(I);
5661 if (!LPad2 || !LPad2->isIdenticalTo(LPad))
5663 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {
5665 BranchInst *BI2 = dyn_cast<BranchInst>(I);
5666 if (!BI2 || !BI2->isIdenticalTo(BI))
5669 // We've found an identical block. Update our predecessors to take that
5670 // path instead and make ourselves dead.
5671 SmallSet<BasicBlock *, 16> Preds;
5672 Preds.insert(pred_begin(BB), pred_end(BB));
5673 for (BasicBlock *Pred : Preds) {
5674 InvokeInst *II = cast<InvokeInst>(Pred->getTerminator());
5675 assert(II->getNormalDest() != BB && II->getUnwindDest() == BB &&
5676 "unexpected successor");
5677 II->setUnwindDest(OtherPred);
5680 // The debug info in OtherPred doesn't cover the merged control flow that
5681 // used to go through BB. We need to delete it or update it.
5682 for (auto I = OtherPred->begin(), E = OtherPred->end(); I != E;) {
5683 Instruction &Inst = *I;
5685 if (isa<DbgInfoIntrinsic>(Inst))
5686 Inst.eraseFromParent();
5689 SmallSet<BasicBlock *, 16> Succs;
5690 Succs.insert(succ_begin(BB), succ_end(BB));
5691 for (BasicBlock *Succ : Succs) {
5692 Succ->removePredecessor(BB);
5695 IRBuilder<> Builder(BI);
5696 Builder.CreateUnreachable();
5697 BI->eraseFromParent();
5703 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI,
5704 IRBuilder<> &Builder) {
5705 BasicBlock *BB = BI->getParent();
5707 if (SinkCommon && SinkThenElseCodeToEnd(BI))
5710 // If the Terminator is the only non-phi instruction, simplify the block.
5711 // if LoopHeader is provided, check if the block is a loop header
5712 // (This is for early invocations before loop simplify and vectorization
5713 // to keep canonical loop forms for nested loops.
5714 // These blocks can be eliminated when the pass is invoked later
5715 // in the back-end.)
5716 BasicBlock::iterator I = BB->getFirstNonPHIOrDbg()->getIterator();
5717 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
5718 (!LoopHeaders || !LoopHeaders->count(BB)) &&
5719 TryToSimplifyUncondBranchFromEmptyBlock(BB))
5722 // If the only instruction in the block is a seteq/setne comparison
5723 // against a constant, try to simplify the block.
5724 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
5725 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
5726 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
5728 if (I->isTerminator() &&
5729 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, DL, TTI,
5730 BonusInstThreshold, AC))
5734 // See if we can merge an empty landing pad block with another which is
5736 if (LandingPadInst *LPad = dyn_cast<LandingPadInst>(I)) {
5737 for (++I; isa<DbgInfoIntrinsic>(I); ++I) {
5739 if (I->isTerminator() && TryToMergeLandingPad(LPad, BI, BB))
5743 // If this basic block is ONLY a compare and a branch, and if a predecessor
5744 // branches to us and our successor, fold the comparison into the
5745 // predecessor and use logical operations to update the incoming value
5746 // for PHI nodes in common successor.
5747 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
5748 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5752 static BasicBlock *allPredecessorsComeFromSameSource(BasicBlock *BB) {
5753 BasicBlock *PredPred = nullptr;
5754 for (auto *P : predecessors(BB)) {
5755 BasicBlock *PPred = P->getSinglePredecessor();
5756 if (!PPred || (PredPred && PredPred != PPred))
5763 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
5764 BasicBlock *BB = BI->getParent();
5766 // Conditional branch
5767 if (isValueEqualityComparison(BI)) {
5768 // If we only have one predecessor, and if it is a branch on this value,
5769 // see if that predecessor totally determines the outcome of this
5771 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5772 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
5773 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5775 // This block must be empty, except for the setcond inst, if it exists.
5776 // Ignore dbg intrinsics.
5777 BasicBlock::iterator I = BB->begin();
5778 // Ignore dbg intrinsics.
5779 while (isa<DbgInfoIntrinsic>(I))
5782 if (FoldValueComparisonIntoPredecessors(BI, Builder))
5783 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5784 } else if (&*I == cast<Instruction>(BI->getCondition())) {
5786 // Ignore dbg intrinsics.
5787 while (isa<DbgInfoIntrinsic>(I))
5789 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
5790 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5794 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
5795 if (SimplifyBranchOnICmpChain(BI, Builder, DL))
5798 // If this basic block has a single dominating predecessor block and the
5799 // dominating block's condition implies BI's condition, we know the direction
5800 // of the BI branch.
5801 if (BasicBlock *Dom = BB->getSinglePredecessor()) {
5802 auto *PBI = dyn_cast_or_null<BranchInst>(Dom->getTerminator());
5803 if (PBI && PBI->isConditional() &&
5804 PBI->getSuccessor(0) != PBI->getSuccessor(1) &&
5805 (PBI->getSuccessor(0) == BB || PBI->getSuccessor(1) == BB)) {
5806 bool CondIsFalse = PBI->getSuccessor(1) == BB;
5807 Optional<bool> Implication = isImpliedCondition(
5808 PBI->getCondition(), BI->getCondition(), DL, CondIsFalse);
5810 // Turn this into a branch on constant.
5811 auto *OldCond = BI->getCondition();
5812 ConstantInt *CI = *Implication
5813 ? ConstantInt::getTrue(BB->getContext())
5814 : ConstantInt::getFalse(BB->getContext());
5815 BI->setCondition(CI);
5816 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
5817 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5822 // If this basic block is ONLY a compare and a branch, and if a predecessor
5823 // branches to us and one of our successors, fold the comparison into the
5824 // predecessor and use logical operations to pick the right destination.
5825 if (FoldBranchToCommonDest(BI, BonusInstThreshold))
5826 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5828 // We have a conditional branch to two blocks that are only reachable
5829 // from BI. We know that the condbr dominates the two blocks, so see if
5830 // there is any identical code in the "then" and "else" blocks. If so, we
5831 // can hoist it up to the branching block.
5832 if (BI->getSuccessor(0)->getSinglePredecessor()) {
5833 if (BI->getSuccessor(1)->getSinglePredecessor()) {
5834 if (HoistThenElseCodeToIf(BI, TTI))
5835 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5837 // If Successor #1 has multiple preds, we may be able to conditionally
5838 // execute Successor #0 if it branches to Successor #1.
5839 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
5840 if (Succ0TI->getNumSuccessors() == 1 &&
5841 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
5842 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), TTI))
5843 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5845 } else if (BI->getSuccessor(1)->getSinglePredecessor()) {
5846 // If Successor #0 has multiple preds, we may be able to conditionally
5847 // execute Successor #1 if it branches to Successor #0.
5848 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
5849 if (Succ1TI->getNumSuccessors() == 1 &&
5850 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
5851 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), TTI))
5852 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5855 // If this is a branch on a phi node in the current block, thread control
5856 // through this block if any PHI node entries are constants.
5857 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
5858 if (PN->getParent() == BI->getParent())
5859 if (FoldCondBranchOnPHI(BI, DL, AC))
5860 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5862 // Scan predecessor blocks for conditional branches.
5863 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
5864 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
5865 if (PBI != BI && PBI->isConditional())
5866 if (SimplifyCondBranchToCondBranch(PBI, BI, DL))
5867 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5869 // Look for diamond patterns.
5870 if (MergeCondStores)
5871 if (BasicBlock *PrevBB = allPredecessorsComeFromSameSource(BB))
5872 if (BranchInst *PBI = dyn_cast<BranchInst>(PrevBB->getTerminator()))
5873 if (PBI != BI && PBI->isConditional())
5874 if (mergeConditionalStores(PBI, BI))
5875 return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5880 /// Check if passing a value to an instruction will cause undefined behavior.
5881 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
5882 Constant *C = dyn_cast<Constant>(V);
5889 if (C->isNullValue() || isa<UndefValue>(C)) {
5890 // Only look at the first use, avoid hurting compile time with long uselists
5891 User *Use = *I->user_begin();
5893 // Now make sure that there are no instructions in between that can alter
5894 // control flow (eg. calls)
5895 for (BasicBlock::iterator
5896 i = ++BasicBlock::iterator(I),
5897 UI = BasicBlock::iterator(dyn_cast<Instruction>(Use));
5899 if (i == I->getParent()->end() || i->mayHaveSideEffects())
5902 // Look through GEPs. A load from a GEP derived from NULL is still undefined
5903 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
5904 if (GEP->getPointerOperand() == I)
5905 return passingValueIsAlwaysUndefined(V, GEP);
5907 // Look through bitcasts.
5908 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
5909 return passingValueIsAlwaysUndefined(V, BC);
5911 // Load from null is undefined.
5912 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
5913 if (!LI->isVolatile())
5914 return LI->getPointerAddressSpace() == 0;
5916 // Store to null is undefined.
5917 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
5918 if (!SI->isVolatile())
5919 return SI->getPointerAddressSpace() == 0 &&
5920 SI->getPointerOperand() == I;
5922 // A call to null is undefined.
5923 if (auto CS = CallSite(Use))
5924 return CS.getCalledValue() == I;
5929 /// If BB has an incoming value that will always trigger undefined behavior
5930 /// (eg. null pointer dereference), remove the branch leading here.
5931 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
5932 for (BasicBlock::iterator i = BB->begin();
5933 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
5934 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
5935 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
5936 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
5937 IRBuilder<> Builder(T);
5938 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
5939 BB->removePredecessor(PHI->getIncomingBlock(i));
5940 // Turn uncoditional branches into unreachables and remove the dead
5941 // destination from conditional branches.
5942 if (BI->isUnconditional())
5943 Builder.CreateUnreachable();
5945 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1)
5946 : BI->getSuccessor(0));
5947 BI->eraseFromParent();
5950 // TODO: SwitchInst.
5956 bool SimplifyCFGOpt::run(BasicBlock *BB) {
5957 bool Changed = false;
5959 assert(BB && BB->getParent() && "Block not embedded in function!");
5960 assert(BB->getTerminator() && "Degenerate basic block encountered!");
5962 // Remove basic blocks that have no predecessors (except the entry block)...
5963 // or that just have themself as a predecessor. These are unreachable.
5964 if ((pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) ||
5965 BB->getSinglePredecessor() == BB) {
5966 DEBUG(dbgs() << "Removing BB: \n" << *BB);
5967 DeleteDeadBlock(BB);
5971 // Check to see if we can constant propagate this terminator instruction
5973 Changed |= ConstantFoldTerminator(BB, true);
5975 // Check for and eliminate duplicate PHI nodes in this block.
5976 Changed |= EliminateDuplicatePHINodes(BB);
5978 // Check for and remove branches that will always cause undefined behavior.
5979 Changed |= removeUndefIntroducingPredecessor(BB);
5981 // Merge basic blocks into their predecessor if there is only one distinct
5982 // pred, and if there is only one distinct successor of the predecessor, and
5983 // if there are no PHI nodes.
5985 if (MergeBlockIntoPredecessor(BB))
5988 IRBuilder<> Builder(BB);
5990 // If there is a trivial two-entry PHI node in this basic block, and we can
5991 // eliminate it, do so now.
5992 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
5993 if (PN->getNumIncomingValues() == 2)
5994 Changed |= FoldTwoEntryPHINode(PN, TTI, DL);
5996 Builder.SetInsertPoint(BB->getTerminator());
5997 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
5998 if (BI->isUnconditional()) {
5999 if (SimplifyUncondBranch(BI, Builder))
6002 if (SimplifyCondBranch(BI, Builder))
6005 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
6006 if (SimplifyReturn(RI, Builder))
6008 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
6009 if (SimplifyResume(RI, Builder))
6011 } else if (CleanupReturnInst *RI =
6012 dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
6013 if (SimplifyCleanupReturn(RI))
6015 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
6016 if (SimplifySwitch(SI, Builder))
6018 } else if (UnreachableInst *UI =
6019 dyn_cast<UnreachableInst>(BB->getTerminator())) {
6020 if (SimplifyUnreachable(UI))
6022 } else if (IndirectBrInst *IBI =
6023 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
6024 if (SimplifyIndirectBr(IBI))
6031 /// This function is used to do simplification of a CFG.
6032 /// For example, it adjusts branches to branches to eliminate the extra hop,
6033 /// eliminates unreachable basic blocks, and does other "peephole" optimization
6034 /// of the CFG. It returns true if a modification was made.
6036 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
6037 unsigned BonusInstThreshold, AssumptionCache *AC,
6038 SmallPtrSetImpl<BasicBlock *> *LoopHeaders,
6039 bool LateSimplifyCFG) {
6040 return SimplifyCFGOpt(TTI, BB->getModule()->getDataLayout(),
6041 BonusInstThreshold, AC, LoopHeaders, LateSimplifyCFG)