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 #define DEBUG_TYPE "simplifycfg"
15 #include "llvm/Transforms/Utils/Local.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SetVector.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/Analysis/InstructionSimplify.h"
23 #include "llvm/Analysis/TargetTransformInfo.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/GlobalVariable.h"
29 #include "llvm/IR/IRBuilder.h"
30 #include "llvm/IR/Instructions.h"
31 #include "llvm/IR/IntrinsicInst.h"
32 #include "llvm/IR/LLVMContext.h"
33 #include "llvm/IR/MDBuilder.h"
34 #include "llvm/IR/Metadata.h"
35 #include "llvm/IR/Module.h"
36 #include "llvm/IR/Operator.h"
37 #include "llvm/IR/Type.h"
38 #include "llvm/Support/CFG.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Support/ConstantRange.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/NoFolder.h"
43 #include "llvm/Support/PatternMatch.h"
44 #include "llvm/Support/raw_ostream.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
50 using namespace PatternMatch;
52 static cl::opt<unsigned>
53 PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1),
54 cl::desc("Control the amount of phi node folding to perform (default = 1)"));
57 DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false),
58 cl::desc("Duplicate return instructions into unconditional branches"));
61 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
62 cl::desc("Sink common instructions down to the end block"));
65 HoistCondStores("simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
66 cl::desc("Hoist conditional stores if an unconditional store preceeds"));
68 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
69 STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables");
70 STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block");
71 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
74 /// ValueEqualityComparisonCase - Represents a case of a switch.
75 struct ValueEqualityComparisonCase {
79 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
80 : Value(Value), Dest(Dest) {}
82 bool operator<(ValueEqualityComparisonCase RHS) const {
83 // Comparing pointers is ok as we only rely on the order for uniquing.
84 return Value < RHS.Value;
87 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
90 class SimplifyCFGOpt {
91 const TargetTransformInfo &TTI;
92 const DataLayout *const TD;
93 Value *isValueEqualityComparison(TerminatorInst *TI);
94 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
95 std::vector<ValueEqualityComparisonCase> &Cases);
96 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
98 IRBuilder<> &Builder);
99 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
100 IRBuilder<> &Builder);
102 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
103 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
104 bool SimplifyUnreachable(UnreachableInst *UI);
105 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
106 bool SimplifyIndirectBr(IndirectBrInst *IBI);
107 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder);
108 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder);
111 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout *TD)
112 : TTI(TTI), TD(TD) {}
113 bool run(BasicBlock *BB);
117 /// SafeToMergeTerminators - Return true if it is safe to merge these two
118 /// terminator instructions together.
120 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
121 if (SI1 == SI2) return false; // Can't merge with self!
123 // It is not safe to merge these two switch instructions if they have a common
124 // successor, and if that successor has a PHI node, and if *that* PHI node has
125 // conflicting incoming values from the two switch blocks.
126 BasicBlock *SI1BB = SI1->getParent();
127 BasicBlock *SI2BB = SI2->getParent();
128 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
130 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
131 if (SI1Succs.count(*I))
132 for (BasicBlock::iterator BBI = (*I)->begin();
133 isa<PHINode>(BBI); ++BBI) {
134 PHINode *PN = cast<PHINode>(BBI);
135 if (PN->getIncomingValueForBlock(SI1BB) !=
136 PN->getIncomingValueForBlock(SI2BB))
143 /// isProfitableToFoldUnconditional - Return true if it is safe and profitable
144 /// to merge these two terminator instructions together, where SI1 is an
145 /// unconditional branch. PhiNodes will store all PHI nodes in common
148 static bool isProfitableToFoldUnconditional(BranchInst *SI1,
151 SmallVectorImpl<PHINode*> &PhiNodes) {
152 if (SI1 == SI2) return false; // Can't merge with self!
153 assert(SI1->isUnconditional() && SI2->isConditional());
155 // We fold the unconditional branch if we can easily update all PHI nodes in
156 // common successors:
157 // 1> We have a constant incoming value for the conditional branch;
158 // 2> We have "Cond" as the incoming value for the unconditional branch;
159 // 3> SI2->getCondition() and Cond have same operands.
160 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
161 if (!Ci2) return false;
162 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
163 Cond->getOperand(1) == Ci2->getOperand(1)) &&
164 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
165 Cond->getOperand(1) == Ci2->getOperand(0)))
168 BasicBlock *SI1BB = SI1->getParent();
169 BasicBlock *SI2BB = SI2->getParent();
170 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
171 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
172 if (SI1Succs.count(*I))
173 for (BasicBlock::iterator BBI = (*I)->begin();
174 isa<PHINode>(BBI); ++BBI) {
175 PHINode *PN = cast<PHINode>(BBI);
176 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
177 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
179 PhiNodes.push_back(PN);
184 /// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will
185 /// now be entries in it from the 'NewPred' block. The values that will be
186 /// flowing into the PHI nodes will be the same as those coming in from
187 /// ExistPred, an existing predecessor of Succ.
188 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
189 BasicBlock *ExistPred) {
190 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
193 for (BasicBlock::iterator I = Succ->begin();
194 (PN = dyn_cast<PHINode>(I)); ++I)
195 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
198 /// ComputeSpeculationCost - Compute an abstract "cost" of speculating the
199 /// given instruction, which is assumed to be safe to speculate. 1 means
200 /// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive.
201 static unsigned ComputeSpeculationCost(const User *I) {
202 assert(isSafeToSpeculativelyExecute(I) &&
203 "Instruction is not safe to speculatively execute!");
204 switch (Operator::getOpcode(I)) {
206 // In doubt, be conservative.
208 case Instruction::GetElementPtr:
209 // GEPs are cheap if all indices are constant.
210 if (!cast<GEPOperator>(I)->hasAllConstantIndices())
213 case Instruction::Load:
214 case Instruction::Add:
215 case Instruction::Sub:
216 case Instruction::And:
217 case Instruction::Or:
218 case Instruction::Xor:
219 case Instruction::Shl:
220 case Instruction::LShr:
221 case Instruction::AShr:
222 case Instruction::ICmp:
223 case Instruction::Trunc:
224 case Instruction::ZExt:
225 case Instruction::SExt:
226 return 1; // These are all cheap.
228 case Instruction::Call:
229 case Instruction::Select:
234 /// DominatesMergePoint - If we have a merge point of an "if condition" as
235 /// accepted above, return true if the specified value dominates the block. We
236 /// don't handle the true generality of domination here, just a special case
237 /// which works well enough for us.
239 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
240 /// see if V (which must be an instruction) and its recursive operands
241 /// that do not dominate BB have a combined cost lower than CostRemaining and
242 /// are non-trapping. If both are true, the instruction is inserted into the
243 /// set and true is returned.
245 /// The cost for most non-trapping instructions is defined as 1 except for
246 /// Select whose cost is 2.
248 /// After this function returns, CostRemaining is decreased by the cost of
249 /// V plus its non-dominating operands. If that cost is greater than
250 /// CostRemaining, false is returned and CostRemaining is undefined.
251 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
252 SmallPtrSet<Instruction*, 4> *AggressiveInsts,
253 unsigned &CostRemaining) {
254 Instruction *I = dyn_cast<Instruction>(V);
256 // Non-instructions all dominate instructions, but not all constantexprs
257 // can be executed unconditionally.
258 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
263 BasicBlock *PBB = I->getParent();
265 // We don't want to allow weird loops that might have the "if condition" in
266 // the bottom of this block.
267 if (PBB == BB) return false;
269 // If this instruction is defined in a block that contains an unconditional
270 // branch to BB, then it must be in the 'conditional' part of the "if
271 // statement". If not, it definitely dominates the region.
272 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
273 if (BI == 0 || BI->isConditional() || BI->getSuccessor(0) != BB)
276 // If we aren't allowing aggressive promotion anymore, then don't consider
277 // instructions in the 'if region'.
278 if (AggressiveInsts == 0) return false;
280 // If we have seen this instruction before, don't count it again.
281 if (AggressiveInsts->count(I)) return true;
283 // Okay, it looks like the instruction IS in the "condition". Check to
284 // see if it's a cheap instruction to unconditionally compute, and if it
285 // only uses stuff defined outside of the condition. If so, hoist it out.
286 if (!isSafeToSpeculativelyExecute(I))
289 unsigned Cost = ComputeSpeculationCost(I);
291 if (Cost > CostRemaining)
294 CostRemaining -= Cost;
296 // Okay, we can only really hoist these out if their operands do
297 // not take us over the cost threshold.
298 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
299 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining))
301 // Okay, it's safe to do this! Remember this instruction.
302 AggressiveInsts->insert(I);
306 /// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr
307 /// and PointerNullValue. Return NULL if value is not a constant int.
308 static ConstantInt *GetConstantInt(Value *V, const DataLayout *TD) {
309 // Normal constant int.
310 ConstantInt *CI = dyn_cast<ConstantInt>(V);
311 if (CI || !TD || !isa<Constant>(V) || !V->getType()->isPointerTy())
314 // This is some kind of pointer constant. Turn it into a pointer-sized
315 // ConstantInt if possible.
316 IntegerType *PtrTy = cast<IntegerType>(TD->getIntPtrType(V->getType()));
318 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
319 if (isa<ConstantPointerNull>(V))
320 return ConstantInt::get(PtrTy, 0);
322 // IntToPtr const int.
323 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
324 if (CE->getOpcode() == Instruction::IntToPtr)
325 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
326 // The constant is very likely to have the right type already.
327 if (CI->getType() == PtrTy)
330 return cast<ConstantInt>
331 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
336 /// GatherConstantCompares - Given a potentially 'or'd or 'and'd together
337 /// collection of icmp eq/ne instructions that compare a value against a
338 /// constant, return the value being compared, and stick the constant into the
341 GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra,
342 const DataLayout *TD, bool isEQ, unsigned &UsedICmps) {
343 Instruction *I = dyn_cast<Instruction>(V);
344 if (I == 0) return 0;
346 // If this is an icmp against a constant, handle this as one of the cases.
347 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) {
348 if (ConstantInt *C = GetConstantInt(I->getOperand(1), TD)) {
352 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
353 // (x & ~2^x) == y --> x == y || x == y|2^x
354 // This undoes a transformation done by instcombine to fuse 2 compares.
355 if (match(ICI->getOperand(0),
356 m_And(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
357 APInt Not = ~RHSC->getValue();
358 if (Not.isPowerOf2()) {
361 ConstantInt::get(C->getContext(), C->getValue() | Not));
369 return I->getOperand(0);
372 // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to
375 ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue());
377 // Shift the range if the compare is fed by an add. This is the range
378 // compare idiom as emitted by instcombine.
380 match(I->getOperand(0), m_Add(m_Value(RHSVal), m_ConstantInt(RHSC)));
382 Span = Span.subtract(RHSC->getValue());
384 // If this is an and/!= check then we want to optimize "x ugt 2" into
387 Span = Span.inverse();
389 // If there are a ton of values, we don't want to make a ginormous switch.
390 if (Span.getSetSize().ugt(8) || Span.isEmptySet())
393 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
394 Vals.push_back(ConstantInt::get(V->getContext(), Tmp));
396 return hasAdd ? RHSVal : I->getOperand(0);
401 // Otherwise, we can only handle an | or &, depending on isEQ.
402 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And))
405 unsigned NumValsBeforeLHS = Vals.size();
406 unsigned UsedICmpsBeforeLHS = UsedICmps;
407 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, TD,
409 unsigned NumVals = Vals.size();
410 unsigned UsedICmpsBeforeRHS = UsedICmps;
411 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD,
415 Vals.resize(NumVals);
416 UsedICmps = UsedICmpsBeforeRHS;
419 // The RHS of the or/and can't be folded in and we haven't used "Extra" yet,
420 // set it and return success.
421 if (Extra == 0 || Extra == I->getOperand(1)) {
422 Extra = I->getOperand(1);
426 Vals.resize(NumValsBeforeLHS);
427 UsedICmps = UsedICmpsBeforeLHS;
431 // If the LHS can't be folded in, but Extra is available and RHS can, try to
433 if (Extra == 0 || Extra == I->getOperand(0)) {
434 Value *OldExtra = Extra;
435 Extra = I->getOperand(0);
436 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD,
439 assert(Vals.size() == NumValsBeforeLHS);
446 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
447 Instruction *Cond = 0;
448 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
449 Cond = dyn_cast<Instruction>(SI->getCondition());
450 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
451 if (BI->isConditional())
452 Cond = dyn_cast<Instruction>(BI->getCondition());
453 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
454 Cond = dyn_cast<Instruction>(IBI->getAddress());
457 TI->eraseFromParent();
458 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
461 /// isValueEqualityComparison - Return true if the specified terminator checks
462 /// to see if a value is equal to constant integer value.
463 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
465 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
466 // Do not permit merging of large switch instructions into their
467 // predecessors unless there is only one predecessor.
468 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()),
469 pred_end(SI->getParent())) <= 128)
470 CV = SI->getCondition();
471 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
472 if (BI->isConditional() && BI->getCondition()->hasOneUse())
473 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition()))
474 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), TD))
475 CV = ICI->getOperand(0);
477 // Unwrap any lossless ptrtoint cast.
478 if (TD && CV && CV->getType() == TD->getIntPtrType(CV->getContext()))
479 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV))
480 CV = PTII->getOperand(0);
484 /// GetValueEqualityComparisonCases - Given a value comparison instruction,
485 /// decode all of the 'cases' that it represents and return the 'default' block.
486 BasicBlock *SimplifyCFGOpt::
487 GetValueEqualityComparisonCases(TerminatorInst *TI,
488 std::vector<ValueEqualityComparisonCase>
490 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
491 Cases.reserve(SI->getNumCases());
492 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
493 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
494 i.getCaseSuccessor()));
495 return SI->getDefaultDest();
498 BranchInst *BI = cast<BranchInst>(TI);
499 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
500 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
501 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
504 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
508 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries
509 /// in the list that match the specified block.
510 static void EliminateBlockCases(BasicBlock *BB,
511 std::vector<ValueEqualityComparisonCase> &Cases) {
512 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
515 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as
518 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
519 std::vector<ValueEqualityComparisonCase > &C2) {
520 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
522 // Make V1 be smaller than V2.
523 if (V1->size() > V2->size())
526 if (V1->size() == 0) return false;
527 if (V1->size() == 1) {
529 ConstantInt *TheVal = (*V1)[0].Value;
530 for (unsigned i = 0, e = V2->size(); i != e; ++i)
531 if (TheVal == (*V2)[i].Value)
535 // Otherwise, just sort both lists and compare element by element.
536 array_pod_sort(V1->begin(), V1->end());
537 array_pod_sort(V2->begin(), V2->end());
538 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
539 while (i1 != e1 && i2 != e2) {
540 if ((*V1)[i1].Value == (*V2)[i2].Value)
542 if ((*V1)[i1].Value < (*V2)[i2].Value)
550 /// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a
551 /// terminator instruction and its block is known to only have a single
552 /// predecessor block, check to see if that predecessor is also a value
553 /// comparison with the same value, and if that comparison determines the
554 /// outcome of this comparison. If so, simplify TI. This does a very limited
555 /// form of jump threading.
556 bool SimplifyCFGOpt::
557 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
559 IRBuilder<> &Builder) {
560 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
561 if (!PredVal) return false; // Not a value comparison in predecessor.
563 Value *ThisVal = isValueEqualityComparison(TI);
564 assert(ThisVal && "This isn't a value comparison!!");
565 if (ThisVal != PredVal) return false; // Different predicates.
567 // TODO: Preserve branch weight metadata, similarly to how
568 // FoldValueComparisonIntoPredecessors preserves it.
570 // Find out information about when control will move from Pred to TI's block.
571 std::vector<ValueEqualityComparisonCase> PredCases;
572 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
574 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
576 // Find information about how control leaves this block.
577 std::vector<ValueEqualityComparisonCase> ThisCases;
578 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
579 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
581 // If TI's block is the default block from Pred's comparison, potentially
582 // simplify TI based on this knowledge.
583 if (PredDef == TI->getParent()) {
584 // If we are here, we know that the value is none of those cases listed in
585 // PredCases. If there are any cases in ThisCases that are in PredCases, we
587 if (!ValuesOverlap(PredCases, ThisCases))
590 if (isa<BranchInst>(TI)) {
591 // Okay, one of the successors of this condbr is dead. Convert it to a
593 assert(ThisCases.size() == 1 && "Branch can only have one case!");
594 // Insert the new branch.
595 Instruction *NI = Builder.CreateBr(ThisDef);
598 // Remove PHI node entries for the dead edge.
599 ThisCases[0].Dest->removePredecessor(TI->getParent());
601 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
602 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
604 EraseTerminatorInstAndDCECond(TI);
608 SwitchInst *SI = cast<SwitchInst>(TI);
609 // Okay, TI has cases that are statically dead, prune them away.
610 SmallPtrSet<Constant*, 16> DeadCases;
611 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
612 DeadCases.insert(PredCases[i].Value);
614 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
615 << "Through successor TI: " << *TI);
617 // Collect branch weights into a vector.
618 SmallVector<uint32_t, 8> Weights;
619 MDNode* MD = SI->getMetadata(LLVMContext::MD_prof);
620 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
622 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
624 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i));
626 Weights.push_back(CI->getValue().getZExtValue());
628 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
630 if (DeadCases.count(i.getCaseValue())) {
632 std::swap(Weights[i.getCaseIndex()+1], Weights.back());
635 i.getCaseSuccessor()->removePredecessor(TI->getParent());
639 if (HasWeight && Weights.size() >= 2)
640 SI->setMetadata(LLVMContext::MD_prof,
641 MDBuilder(SI->getParent()->getContext()).
642 createBranchWeights(Weights));
644 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
648 // Otherwise, TI's block must correspond to some matched value. Find out
649 // which value (or set of values) this is.
650 ConstantInt *TIV = 0;
651 BasicBlock *TIBB = TI->getParent();
652 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
653 if (PredCases[i].Dest == TIBB) {
655 return false; // Cannot handle multiple values coming to this block.
656 TIV = PredCases[i].Value;
658 assert(TIV && "No edge from pred to succ?");
660 // Okay, we found the one constant that our value can be if we get into TI's
661 // BB. Find out which successor will unconditionally be branched to.
662 BasicBlock *TheRealDest = 0;
663 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
664 if (ThisCases[i].Value == TIV) {
665 TheRealDest = ThisCases[i].Dest;
669 // If not handled by any explicit cases, it is handled by the default case.
670 if (TheRealDest == 0) TheRealDest = ThisDef;
672 // Remove PHI node entries for dead edges.
673 BasicBlock *CheckEdge = TheRealDest;
674 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
675 if (*SI != CheckEdge)
676 (*SI)->removePredecessor(TIBB);
680 // Insert the new branch.
681 Instruction *NI = Builder.CreateBr(TheRealDest);
684 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
685 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
687 EraseTerminatorInstAndDCECond(TI);
692 /// ConstantIntOrdering - This class implements a stable ordering of constant
693 /// integers that does not depend on their address. This is important for
694 /// applications that sort ConstantInt's to ensure uniqueness.
695 struct ConstantIntOrdering {
696 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
697 return LHS->getValue().ult(RHS->getValue());
702 static int ConstantIntSortPredicate(ConstantInt *const *P1,
703 ConstantInt *const *P2) {
704 const ConstantInt *LHS = *P1;
705 const ConstantInt *RHS = *P2;
706 if (LHS->getValue().ult(RHS->getValue()))
708 if (LHS->getValue() == RHS->getValue())
713 static inline bool HasBranchWeights(const Instruction* I) {
714 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof);
715 if (ProfMD && ProfMD->getOperand(0))
716 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
717 return MDS->getString().equals("branch_weights");
722 /// Get Weights of a given TerminatorInst, the default weight is at the front
723 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
725 static void GetBranchWeights(TerminatorInst *TI,
726 SmallVectorImpl<uint64_t> &Weights) {
727 MDNode* MD = TI->getMetadata(LLVMContext::MD_prof);
729 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
730 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(i));
732 Weights.push_back(CI->getValue().getZExtValue());
735 // If TI is a conditional eq, the default case is the false case,
736 // and the corresponding branch-weight data is at index 2. We swap the
737 // default weight to be the first entry.
738 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
739 assert(Weights.size() == 2);
740 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
741 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
742 std::swap(Weights.front(), Weights.back());
746 /// Sees if any of the weights are too big for a uint32_t, and halves all the
747 /// weights if any are.
748 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
750 for (unsigned i = 0; i < Weights.size(); ++i)
751 if (Weights[i] > UINT_MAX) {
759 for (unsigned i = 0; i < Weights.size(); ++i)
763 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value
764 /// equality comparison instruction (either a switch or a branch on "X == c").
765 /// See if any of the predecessors of the terminator block are value comparisons
766 /// on the same value. If so, and if safe to do so, fold them together.
767 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
768 IRBuilder<> &Builder) {
769 BasicBlock *BB = TI->getParent();
770 Value *CV = isValueEqualityComparison(TI); // CondVal
771 assert(CV && "Not a comparison?");
772 bool Changed = false;
774 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
775 while (!Preds.empty()) {
776 BasicBlock *Pred = Preds.pop_back_val();
778 // See if the predecessor is a comparison with the same value.
779 TerminatorInst *PTI = Pred->getTerminator();
780 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
782 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
783 // Figure out which 'cases' to copy from SI to PSI.
784 std::vector<ValueEqualityComparisonCase> BBCases;
785 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
787 std::vector<ValueEqualityComparisonCase> PredCases;
788 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
790 // Based on whether the default edge from PTI goes to BB or not, fill in
791 // PredCases and PredDefault with the new switch cases we would like to
793 SmallVector<BasicBlock*, 8> NewSuccessors;
795 // Update the branch weight metadata along the way
796 SmallVector<uint64_t, 8> Weights;
797 bool PredHasWeights = HasBranchWeights(PTI);
798 bool SuccHasWeights = HasBranchWeights(TI);
800 if (PredHasWeights) {
801 GetBranchWeights(PTI, Weights);
802 // branch-weight metadata is inconsistent here.
803 if (Weights.size() != 1 + PredCases.size())
804 PredHasWeights = SuccHasWeights = false;
805 } else if (SuccHasWeights)
806 // If there are no predecessor weights but there are successor weights,
807 // populate Weights with 1, which will later be scaled to the sum of
808 // successor's weights
809 Weights.assign(1 + PredCases.size(), 1);
811 SmallVector<uint64_t, 8> SuccWeights;
812 if (SuccHasWeights) {
813 GetBranchWeights(TI, SuccWeights);
814 // branch-weight metadata is inconsistent here.
815 if (SuccWeights.size() != 1 + BBCases.size())
816 PredHasWeights = SuccHasWeights = false;
817 } else if (PredHasWeights)
818 SuccWeights.assign(1 + BBCases.size(), 1);
820 if (PredDefault == BB) {
821 // If this is the default destination from PTI, only the edges in TI
822 // that don't occur in PTI, or that branch to BB will be activated.
823 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
824 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
825 if (PredCases[i].Dest != BB)
826 PTIHandled.insert(PredCases[i].Value);
828 // The default destination is BB, we don't need explicit targets.
829 std::swap(PredCases[i], PredCases.back());
831 if (PredHasWeights || SuccHasWeights) {
832 // Increase weight for the default case.
833 Weights[0] += Weights[i+1];
834 std::swap(Weights[i+1], Weights.back());
838 PredCases.pop_back();
842 // Reconstruct the new switch statement we will be building.
843 if (PredDefault != BBDefault) {
844 PredDefault->removePredecessor(Pred);
845 PredDefault = BBDefault;
846 NewSuccessors.push_back(BBDefault);
849 unsigned CasesFromPred = Weights.size();
850 uint64_t ValidTotalSuccWeight = 0;
851 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
852 if (!PTIHandled.count(BBCases[i].Value) &&
853 BBCases[i].Dest != BBDefault) {
854 PredCases.push_back(BBCases[i]);
855 NewSuccessors.push_back(BBCases[i].Dest);
856 if (SuccHasWeights || PredHasWeights) {
857 // The default weight is at index 0, so weight for the ith case
858 // should be at index i+1. Scale the cases from successor by
859 // PredDefaultWeight (Weights[0]).
860 Weights.push_back(Weights[0] * SuccWeights[i+1]);
861 ValidTotalSuccWeight += SuccWeights[i+1];
865 if (SuccHasWeights || PredHasWeights) {
866 ValidTotalSuccWeight += SuccWeights[0];
867 // Scale the cases from predecessor by ValidTotalSuccWeight.
868 for (unsigned i = 1; i < CasesFromPred; ++i)
869 Weights[i] *= ValidTotalSuccWeight;
870 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
871 Weights[0] *= SuccWeights[0];
874 // If this is not the default destination from PSI, only the edges
875 // in SI that occur in PSI with a destination of BB will be
877 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
878 std::map<ConstantInt*, uint64_t> WeightsForHandled;
879 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
880 if (PredCases[i].Dest == BB) {
881 PTIHandled.insert(PredCases[i].Value);
883 if (PredHasWeights || SuccHasWeights) {
884 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
885 std::swap(Weights[i+1], Weights.back());
889 std::swap(PredCases[i], PredCases.back());
890 PredCases.pop_back();
894 // Okay, now we know which constants were sent to BB from the
895 // predecessor. Figure out where they will all go now.
896 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
897 if (PTIHandled.count(BBCases[i].Value)) {
898 // If this is one we are capable of getting...
899 if (PredHasWeights || SuccHasWeights)
900 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
901 PredCases.push_back(BBCases[i]);
902 NewSuccessors.push_back(BBCases[i].Dest);
903 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
906 // If there are any constants vectored to BB that TI doesn't handle,
907 // they must go to the default destination of TI.
908 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
910 E = PTIHandled.end(); I != E; ++I) {
911 if (PredHasWeights || SuccHasWeights)
912 Weights.push_back(WeightsForHandled[*I]);
913 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
914 NewSuccessors.push_back(BBDefault);
918 // Okay, at this point, we know which new successor Pred will get. Make
919 // sure we update the number of entries in the PHI nodes for these
921 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i)
922 AddPredecessorToBlock(NewSuccessors[i], Pred, BB);
924 Builder.SetInsertPoint(PTI);
925 // Convert pointer to int before we switch.
926 if (CV->getType()->isPointerTy()) {
927 assert(TD && "Cannot switch on pointer without DataLayout");
928 CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getContext()),
932 // Now that the successors are updated, create the new Switch instruction.
933 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
935 NewSI->setDebugLoc(PTI->getDebugLoc());
936 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
937 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest);
939 if (PredHasWeights || SuccHasWeights) {
940 // Halve the weights if any of them cannot fit in an uint32_t
943 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
945 NewSI->setMetadata(LLVMContext::MD_prof,
946 MDBuilder(BB->getContext()).
947 createBranchWeights(MDWeights));
950 EraseTerminatorInstAndDCECond(PTI);
952 // Okay, last check. If BB is still a successor of PSI, then we must
953 // have an infinite loop case. If so, add an infinitely looping block
954 // to handle the case to preserve the behavior of the code.
955 BasicBlock *InfLoopBlock = 0;
956 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
957 if (NewSI->getSuccessor(i) == BB) {
958 if (InfLoopBlock == 0) {
959 // Insert it at the end of the function, because it's either code,
960 // or it won't matter if it's hot. :)
961 InfLoopBlock = BasicBlock::Create(BB->getContext(),
962 "infloop", BB->getParent());
963 BranchInst::Create(InfLoopBlock, InfLoopBlock);
965 NewSI->setSuccessor(i, InfLoopBlock);
974 // isSafeToHoistInvoke - If we would need to insert a select that uses the
975 // value of this invoke (comments in HoistThenElseCodeToIf explain why we
976 // would need to do this), we can't hoist the invoke, as there is nowhere
977 // to put the select in this case.
978 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
979 Instruction *I1, Instruction *I2) {
980 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
982 for (BasicBlock::iterator BBI = SI->begin();
983 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
984 Value *BB1V = PN->getIncomingValueForBlock(BB1);
985 Value *BB2V = PN->getIncomingValueForBlock(BB2);
986 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
994 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
995 /// BB2, hoist any common code in the two blocks up into the branch block. The
996 /// caller of this function guarantees that BI's block dominates BB1 and BB2.
997 static bool HoistThenElseCodeToIf(BranchInst *BI) {
998 // This does very trivial matching, with limited scanning, to find identical
999 // instructions in the two blocks. In particular, we don't want to get into
1000 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1001 // such, we currently just scan for obviously identical instructions in an
1003 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1004 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1006 BasicBlock::iterator BB1_Itr = BB1->begin();
1007 BasicBlock::iterator BB2_Itr = BB2->begin();
1009 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++;
1010 // Skip debug info if it is not identical.
1011 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1012 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1013 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1014 while (isa<DbgInfoIntrinsic>(I1))
1016 while (isa<DbgInfoIntrinsic>(I2))
1019 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1020 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1023 BasicBlock *BIParent = BI->getParent();
1025 bool Changed = false;
1027 // If we are hoisting the terminator instruction, don't move one (making a
1028 // broken BB), instead clone it, and remove BI.
1029 if (isa<TerminatorInst>(I1))
1030 goto HoistTerminator;
1032 // For a normal instruction, we just move one to right before the branch,
1033 // then replace all uses of the other with the first. Finally, we remove
1034 // the now redundant second instruction.
1035 BIParent->getInstList().splice(BI, BB1->getInstList(), I1);
1036 if (!I2->use_empty())
1037 I2->replaceAllUsesWith(I1);
1038 I1->intersectOptionalDataWith(I2);
1039 I2->eraseFromParent();
1044 // Skip debug info if it is not identical.
1045 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1046 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1047 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1048 while (isa<DbgInfoIntrinsic>(I1))
1050 while (isa<DbgInfoIntrinsic>(I2))
1053 } while (I1->isIdenticalToWhenDefined(I2));
1058 // It may not be possible to hoist an invoke.
1059 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1062 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1064 for (BasicBlock::iterator BBI = SI->begin();
1065 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1066 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1067 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1071 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1073 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1078 // Okay, it is safe to hoist the terminator.
1079 Instruction *NT = I1->clone();
1080 BIParent->getInstList().insert(BI, NT);
1081 if (!NT->getType()->isVoidTy()) {
1082 I1->replaceAllUsesWith(NT);
1083 I2->replaceAllUsesWith(NT);
1087 IRBuilder<true, NoFolder> Builder(NT);
1088 // Hoisting one of the terminators from our successor is a great thing.
1089 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1090 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1091 // nodes, so we insert select instruction to compute the final result.
1092 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1093 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1095 for (BasicBlock::iterator BBI = SI->begin();
1096 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1097 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1098 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1099 if (BB1V == BB2V) continue;
1101 // These values do not agree. Insert a select instruction before NT
1102 // that determines the right value.
1103 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1105 SI = cast<SelectInst>
1106 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1107 BB1V->getName()+"."+BB2V->getName()));
1109 // Make the PHI node use the select for all incoming values for BB1/BB2
1110 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1111 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1112 PN->setIncomingValue(i, SI);
1116 // Update any PHI nodes in our new successors.
1117 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1118 AddPredecessorToBlock(*SI, BIParent, BB1);
1120 EraseTerminatorInstAndDCECond(BI);
1124 /// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd,
1125 /// check whether BBEnd has only two predecessors and the other predecessor
1126 /// ends with an unconditional branch. If it is true, sink any common code
1127 /// in the two predecessors to BBEnd.
1128 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1129 assert(BI1->isUnconditional());
1130 BasicBlock *BB1 = BI1->getParent();
1131 BasicBlock *BBEnd = BI1->getSuccessor(0);
1133 // Check that BBEnd has two predecessors and the other predecessor ends with
1134 // an unconditional branch.
1135 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1136 BasicBlock *Pred0 = *PI++;
1137 if (PI == PE) // Only one predecessor.
1139 BasicBlock *Pred1 = *PI++;
1140 if (PI != PE) // More than two predecessors.
1142 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1143 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1144 if (!BI2 || !BI2->isUnconditional())
1147 // Gather the PHI nodes in BBEnd.
1148 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2;
1149 Instruction *FirstNonPhiInBBEnd = 0;
1150 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end();
1152 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1153 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1154 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1155 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN);
1157 FirstNonPhiInBBEnd = &*I;
1161 if (!FirstNonPhiInBBEnd)
1165 // This does very trivial matching, with limited scanning, to find identical
1166 // instructions in the two blocks. We scan backward for obviously identical
1167 // instructions in an identical order.
1168 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1169 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(),
1170 RE2 = BB2->getInstList().rend();
1172 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1175 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1178 // Skip the unconditional branches.
1182 bool Changed = false;
1183 while (RI1 != RE1 && RI2 != RE2) {
1185 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1188 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1192 Instruction *I1 = &*RI1, *I2 = &*RI2;
1193 // I1 and I2 should have a single use in the same PHI node, and they
1194 // perform the same operation.
1195 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1196 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1197 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1198 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) ||
1199 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1200 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1201 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1202 !I1->hasOneUse() || !I2->hasOneUse() ||
1203 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() ||
1204 MapValueFromBB1ToBB2[I1].first != I2)
1207 // Check whether we should swap the operands of ICmpInst.
1208 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1209 bool SwapOpnds = false;
1210 if (ICmp1 && ICmp2 &&
1211 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1212 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1213 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1214 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1215 ICmp2->swapOperands();
1218 if (!I1->isSameOperationAs(I2)) {
1220 ICmp2->swapOperands();
1224 // The operands should be either the same or they need to be generated
1225 // with a PHI node after sinking. We only handle the case where there is
1226 // a single pair of different operands.
1227 Value *DifferentOp1 = 0, *DifferentOp2 = 0;
1228 unsigned Op1Idx = 0;
1229 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1230 if (I1->getOperand(I) == I2->getOperand(I))
1232 // Early exit if we have more-than one pair of different operands or
1233 // the different operand is already in MapValueFromBB1ToBB2.
1234 // Early exit if we need a PHI node to replace a constant.
1236 MapValueFromBB1ToBB2.find(I1->getOperand(I)) !=
1237 MapValueFromBB1ToBB2.end() ||
1238 isa<Constant>(I1->getOperand(I)) ||
1239 isa<Constant>(I2->getOperand(I))) {
1240 // If we can't sink the instructions, undo the swapping.
1242 ICmp2->swapOperands();
1245 DifferentOp1 = I1->getOperand(I);
1247 DifferentOp2 = I2->getOperand(I);
1250 // We insert the pair of different operands to MapValueFromBB1ToBB2 and
1251 // remove (I1, I2) from MapValueFromBB1ToBB2.
1253 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2,
1254 DifferentOp1->getName() + ".sink",
1256 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN);
1257 // I1 should use NewPN instead of DifferentOp1.
1258 I1->setOperand(Op1Idx, NewPN);
1259 NewPN->addIncoming(DifferentOp1, BB1);
1260 NewPN->addIncoming(DifferentOp2, BB2);
1261 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1263 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second;
1264 MapValueFromBB1ToBB2.erase(I1);
1266 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";);
1267 DEBUG(dbgs() << " " << *I2 << "\n";);
1268 // We need to update RE1 and RE2 if we are going to sink the first
1269 // instruction in the basic block down.
1270 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1271 // Sink the instruction.
1272 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1);
1273 if (!OldPN->use_empty())
1274 OldPN->replaceAllUsesWith(I1);
1275 OldPN->eraseFromParent();
1277 if (!I2->use_empty())
1278 I2->replaceAllUsesWith(I1);
1279 I1->intersectOptionalDataWith(I2);
1280 I2->eraseFromParent();
1283 RE1 = BB1->getInstList().rend();
1285 RE2 = BB2->getInstList().rend();
1286 FirstNonPhiInBBEnd = I1;
1293 /// \brief Determine if we can hoist sink a sole store instruction out of a
1294 /// conditional block.
1296 /// We are looking for code like the following:
1298 /// store i32 %add, i32* %arrayidx2
1299 /// ... // No other stores or function calls (we could be calling a memory
1300 /// ... // function).
1301 /// %cmp = icmp ult %x, %y
1302 /// br i1 %cmp, label %EndBB, label %ThenBB
1304 /// store i32 %add5, i32* %arrayidx2
1308 /// We are going to transform this into:
1310 /// store i32 %add, i32* %arrayidx2
1312 /// %cmp = icmp ult %x, %y
1313 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1314 /// store i32 %add.add5, i32* %arrayidx2
1317 /// \return The pointer to the value of the previous store if the store can be
1318 /// hoisted into the predecessor block. 0 otherwise.
1319 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1320 BasicBlock *StoreBB, BasicBlock *EndBB) {
1321 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1325 // Volatile or atomic.
1326 if (!StoreToHoist->isSimple())
1329 Value *StorePtr = StoreToHoist->getPointerOperand();
1331 // Look for a store to the same pointer in BrBB.
1332 unsigned MaxNumInstToLookAt = 10;
1333 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(),
1334 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) {
1335 Instruction *CurI = &*RI;
1337 // Could be calling an instruction that effects memory like free().
1338 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI))
1341 StoreInst *SI = dyn_cast<StoreInst>(CurI);
1342 // Found the previous store make sure it stores to the same location.
1343 if (SI && SI->getPointerOperand() == StorePtr)
1344 // Found the previous store, return its value operand.
1345 return SI->getValueOperand();
1347 return 0; // Unknown store.
1353 /// \brief Speculate a conditional basic block flattening the CFG.
1355 /// Note that this is a very risky transform currently. Speculating
1356 /// instructions like this is most often not desirable. Instead, there is an MI
1357 /// pass which can do it with full awareness of the resource constraints.
1358 /// However, some cases are "obvious" and we should do directly. An example of
1359 /// this is speculating a single, reasonably cheap instruction.
1361 /// There is only one distinct advantage to flattening the CFG at the IR level:
1362 /// it makes very common but simplistic optimizations such as are common in
1363 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1364 /// modeling their effects with easier to reason about SSA value graphs.
1367 /// An illustration of this transform is turning this IR:
1370 /// %cmp = icmp ult %x, %y
1371 /// br i1 %cmp, label %EndBB, label %ThenBB
1373 /// %sub = sub %x, %y
1376 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1383 /// %cmp = icmp ult %x, %y
1384 /// %sub = sub %x, %y
1385 /// %cond = select i1 %cmp, 0, %sub
1389 /// \returns true if the conditional block is removed.
1390 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB) {
1391 // Be conservative for now. FP select instruction can often be expensive.
1392 Value *BrCond = BI->getCondition();
1393 if (isa<FCmpInst>(BrCond))
1396 BasicBlock *BB = BI->getParent();
1397 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1399 // If ThenBB is actually on the false edge of the conditional branch, remember
1400 // to swap the select operands later.
1401 bool Invert = false;
1402 if (ThenBB != BI->getSuccessor(0)) {
1403 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1406 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1408 // Keep a count of how many times instructions are used within CondBB when
1409 // they are candidates for sinking into CondBB. Specifically:
1410 // - They are defined in BB, and
1411 // - They have no side effects, and
1412 // - All of their uses are in CondBB.
1413 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1415 unsigned SpeculationCost = 0;
1416 Value *SpeculatedStoreValue = 0;
1417 StoreInst *SpeculatedStore = 0;
1418 for (BasicBlock::iterator BBI = ThenBB->begin(),
1419 BBE = llvm::prior(ThenBB->end());
1420 BBI != BBE; ++BBI) {
1421 Instruction *I = BBI;
1423 if (isa<DbgInfoIntrinsic>(I))
1426 // Only speculatively execution a single instruction (not counting the
1427 // terminator) for now.
1429 if (SpeculationCost > 1)
1432 // Don't hoist the instruction if it's unsafe or expensive.
1433 if (!isSafeToSpeculativelyExecute(I) &&
1434 !(HoistCondStores &&
1435 (SpeculatedStoreValue = isSafeToSpeculateStore(I, BB, ThenBB,
1438 if (!SpeculatedStoreValue &&
1439 ComputeSpeculationCost(I) > PHINodeFoldingThreshold)
1442 // Store the store speculation candidate.
1443 if (SpeculatedStoreValue)
1444 SpeculatedStore = cast<StoreInst>(I);
1446 // Do not hoist the instruction if any of its operands are defined but not
1447 // used in BB. The transformation will prevent the operand from
1448 // being sunk into the use block.
1449 for (User::op_iterator i = I->op_begin(), e = I->op_end();
1451 Instruction *OpI = dyn_cast<Instruction>(*i);
1452 if (!OpI || OpI->getParent() != BB ||
1453 OpI->mayHaveSideEffects())
1454 continue; // Not a candidate for sinking.
1456 ++SinkCandidateUseCounts[OpI];
1460 // Consider any sink candidates which are only used in CondBB as costs for
1461 // speculation. Note, while we iterate over a DenseMap here, we are summing
1462 // and so iteration order isn't significant.
1463 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I =
1464 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end();
1466 if (I->first->getNumUses() == I->second) {
1468 if (SpeculationCost > 1)
1472 // Check that the PHI nodes can be converted to selects.
1473 bool HaveRewritablePHIs = false;
1474 for (BasicBlock::iterator I = EndBB->begin();
1475 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1476 Value *OrigV = PN->getIncomingValueForBlock(BB);
1477 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1479 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1480 // Skip PHIs which are trivial.
1484 HaveRewritablePHIs = true;
1485 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1486 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1487 if (!OrigCE && !ThenCE)
1488 continue; // Known safe and cheap.
1490 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
1491 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
1493 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE) : 0;
1494 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE) : 0;
1495 if (OrigCost + ThenCost > 2 * PHINodeFoldingThreshold)
1498 // Account for the cost of an unfolded ConstantExpr which could end up
1499 // getting expanded into Instructions.
1500 // FIXME: This doesn't account for how many operations are combined in the
1501 // constant expression.
1503 if (SpeculationCost > 1)
1507 // If there are no PHIs to process, bail early. This helps ensure idempotence
1509 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
1512 // If we get here, we can hoist the instruction and if-convert.
1513 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1515 // Insert a select of the value of the speculated store.
1516 if (SpeculatedStoreValue) {
1517 IRBuilder<true, NoFolder> Builder(BI);
1518 Value *TrueV = SpeculatedStore->getValueOperand();
1519 Value *FalseV = SpeculatedStoreValue;
1521 std::swap(TrueV, FalseV);
1522 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() +
1523 "." + FalseV->getName());
1524 SpeculatedStore->setOperand(0, S);
1527 // Hoist the instructions.
1528 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(),
1529 llvm::prior(ThenBB->end()));
1531 // Insert selects and rewrite the PHI operands.
1532 IRBuilder<true, NoFolder> Builder(BI);
1533 for (BasicBlock::iterator I = EndBB->begin();
1534 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1535 unsigned OrigI = PN->getBasicBlockIndex(BB);
1536 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1537 Value *OrigV = PN->getIncomingValue(OrigI);
1538 Value *ThenV = PN->getIncomingValue(ThenI);
1540 // Skip PHIs which are trivial.
1544 // Create a select whose true value is the speculatively executed value and
1545 // false value is the preexisting value. Swap them if the branch
1546 // destinations were inverted.
1547 Value *TrueV = ThenV, *FalseV = OrigV;
1549 std::swap(TrueV, FalseV);
1550 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV,
1551 TrueV->getName() + "." + FalseV->getName());
1552 PN->setIncomingValue(OrigI, V);
1553 PN->setIncomingValue(ThenI, V);
1560 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1561 /// across this block.
1562 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1563 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1566 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1567 if (isa<DbgInfoIntrinsic>(BBI))
1569 if (Size > 10) return false; // Don't clone large BB's.
1572 // We can only support instructions that do not define values that are
1573 // live outside of the current basic block.
1574 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end();
1576 Instruction *U = cast<Instruction>(*UI);
1577 if (U->getParent() != BB || isa<PHINode>(U)) return false;
1580 // Looks ok, continue checking.
1586 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1587 /// that is defined in the same block as the branch and if any PHI entries are
1588 /// constants, thread edges corresponding to that entry to be branches to their
1589 /// ultimate destination.
1590 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *TD) {
1591 BasicBlock *BB = BI->getParent();
1592 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1593 // NOTE: we currently cannot transform this case if the PHI node is used
1594 // outside of the block.
1595 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1598 // Degenerate case of a single entry PHI.
1599 if (PN->getNumIncomingValues() == 1) {
1600 FoldSingleEntryPHINodes(PN->getParent());
1604 // Now we know that this block has multiple preds and two succs.
1605 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1607 // Okay, this is a simple enough basic block. See if any phi values are
1609 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1610 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1611 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue;
1613 // Okay, we now know that all edges from PredBB should be revectored to
1614 // branch to RealDest.
1615 BasicBlock *PredBB = PN->getIncomingBlock(i);
1616 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1618 if (RealDest == BB) continue; // Skip self loops.
1619 // Skip if the predecessor's terminator is an indirect branch.
1620 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1622 // The dest block might have PHI nodes, other predecessors and other
1623 // difficult cases. Instead of being smart about this, just insert a new
1624 // block that jumps to the destination block, effectively splitting
1625 // the edge we are about to create.
1626 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1627 RealDest->getName()+".critedge",
1628 RealDest->getParent(), RealDest);
1629 BranchInst::Create(RealDest, EdgeBB);
1631 // Update PHI nodes.
1632 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1634 // BB may have instructions that are being threaded over. Clone these
1635 // instructions into EdgeBB. We know that there will be no uses of the
1636 // cloned instructions outside of EdgeBB.
1637 BasicBlock::iterator InsertPt = EdgeBB->begin();
1638 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1639 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1640 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1641 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1644 // Clone the instruction.
1645 Instruction *N = BBI->clone();
1646 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1648 // Update operands due to translation.
1649 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1651 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1652 if (PI != TranslateMap.end())
1656 // Check for trivial simplification.
1657 if (Value *V = SimplifyInstruction(N, TD)) {
1658 TranslateMap[BBI] = V;
1659 delete N; // Instruction folded away, don't need actual inst
1661 // Insert the new instruction into its new home.
1662 EdgeBB->getInstList().insert(InsertPt, N);
1663 if (!BBI->use_empty())
1664 TranslateMap[BBI] = N;
1668 // Loop over all of the edges from PredBB to BB, changing them to branch
1669 // to EdgeBB instead.
1670 TerminatorInst *PredBBTI = PredBB->getTerminator();
1671 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1672 if (PredBBTI->getSuccessor(i) == BB) {
1673 BB->removePredecessor(PredBB);
1674 PredBBTI->setSuccessor(i, EdgeBB);
1677 // Recurse, simplifying any other constants.
1678 return FoldCondBranchOnPHI(BI, TD) | true;
1684 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1685 /// PHI node, see if we can eliminate it.
1686 static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *TD) {
1687 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1688 // statement", which has a very simple dominance structure. Basically, we
1689 // are trying to find the condition that is being branched on, which
1690 // subsequently causes this merge to happen. We really want control
1691 // dependence information for this check, but simplifycfg can't keep it up
1692 // to date, and this catches most of the cases we care about anyway.
1693 BasicBlock *BB = PN->getParent();
1694 BasicBlock *IfTrue, *IfFalse;
1695 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1697 // Don't bother if the branch will be constant folded trivially.
1698 isa<ConstantInt>(IfCond))
1701 // Okay, we found that we can merge this two-entry phi node into a select.
1702 // Doing so would require us to fold *all* two entry phi nodes in this block.
1703 // At some point this becomes non-profitable (particularly if the target
1704 // doesn't support cmov's). Only do this transformation if there are two or
1705 // fewer PHI nodes in this block.
1706 unsigned NumPhis = 0;
1707 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1711 // Loop over the PHI's seeing if we can promote them all to select
1712 // instructions. While we are at it, keep track of the instructions
1713 // that need to be moved to the dominating block.
1714 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1715 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1716 MaxCostVal1 = PHINodeFoldingThreshold;
1718 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1719 PHINode *PN = cast<PHINode>(II++);
1720 if (Value *V = SimplifyInstruction(PN, TD)) {
1721 PN->replaceAllUsesWith(V);
1722 PN->eraseFromParent();
1726 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1728 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1733 // If we folded the first phi, PN dangles at this point. Refresh it. If
1734 // we ran out of PHIs then we simplified them all.
1735 PN = dyn_cast<PHINode>(BB->begin());
1736 if (PN == 0) return true;
1738 // Don't fold i1 branches on PHIs which contain binary operators. These can
1739 // often be turned into switches and other things.
1740 if (PN->getType()->isIntegerTy(1) &&
1741 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1742 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1743 isa<BinaryOperator>(IfCond)))
1746 // If we all PHI nodes are promotable, check to make sure that all
1747 // instructions in the predecessor blocks can be promoted as well. If
1748 // not, we won't be able to get rid of the control flow, so it's not
1749 // worth promoting to select instructions.
1750 BasicBlock *DomBlock = 0;
1751 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1752 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1753 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1756 DomBlock = *pred_begin(IfBlock1);
1757 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1758 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1759 // This is not an aggressive instruction that we can promote.
1760 // Because of this, we won't be able to get rid of the control
1761 // flow, so the xform is not worth it.
1766 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1769 DomBlock = *pred_begin(IfBlock2);
1770 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1771 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1772 // This is not an aggressive instruction that we can promote.
1773 // Because of this, we won't be able to get rid of the control
1774 // flow, so the xform is not worth it.
1779 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1780 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1782 // If we can still promote the PHI nodes after this gauntlet of tests,
1783 // do all of the PHI's now.
1784 Instruction *InsertPt = DomBlock->getTerminator();
1785 IRBuilder<true, NoFolder> Builder(InsertPt);
1787 // Move all 'aggressive' instructions, which are defined in the
1788 // conditional parts of the if's up to the dominating block.
1790 DomBlock->getInstList().splice(InsertPt,
1791 IfBlock1->getInstList(), IfBlock1->begin(),
1792 IfBlock1->getTerminator());
1794 DomBlock->getInstList().splice(InsertPt,
1795 IfBlock2->getInstList(), IfBlock2->begin(),
1796 IfBlock2->getTerminator());
1798 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1799 // Change the PHI node into a select instruction.
1800 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1801 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1804 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1805 PN->replaceAllUsesWith(NV);
1807 PN->eraseFromParent();
1810 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1811 // has been flattened. Change DomBlock to jump directly to our new block to
1812 // avoid other simplifycfg's kicking in on the diamond.
1813 TerminatorInst *OldTI = DomBlock->getTerminator();
1814 Builder.SetInsertPoint(OldTI);
1815 Builder.CreateBr(BB);
1816 OldTI->eraseFromParent();
1820 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1821 /// to two returning blocks, try to merge them together into one return,
1822 /// introducing a select if the return values disagree.
1823 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1824 IRBuilder<> &Builder) {
1825 assert(BI->isConditional() && "Must be a conditional branch");
1826 BasicBlock *TrueSucc = BI->getSuccessor(0);
1827 BasicBlock *FalseSucc = BI->getSuccessor(1);
1828 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1829 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1831 // Check to ensure both blocks are empty (just a return) or optionally empty
1832 // with PHI nodes. If there are other instructions, merging would cause extra
1833 // computation on one path or the other.
1834 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1836 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1839 Builder.SetInsertPoint(BI);
1840 // Okay, we found a branch that is going to two return nodes. If
1841 // there is no return value for this function, just change the
1842 // branch into a return.
1843 if (FalseRet->getNumOperands() == 0) {
1844 TrueSucc->removePredecessor(BI->getParent());
1845 FalseSucc->removePredecessor(BI->getParent());
1846 Builder.CreateRetVoid();
1847 EraseTerminatorInstAndDCECond(BI);
1851 // Otherwise, figure out what the true and false return values are
1852 // so we can insert a new select instruction.
1853 Value *TrueValue = TrueRet->getReturnValue();
1854 Value *FalseValue = FalseRet->getReturnValue();
1856 // Unwrap any PHI nodes in the return blocks.
1857 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1858 if (TVPN->getParent() == TrueSucc)
1859 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1860 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1861 if (FVPN->getParent() == FalseSucc)
1862 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1864 // In order for this transformation to be safe, we must be able to
1865 // unconditionally execute both operands to the return. This is
1866 // normally the case, but we could have a potentially-trapping
1867 // constant expression that prevents this transformation from being
1869 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1872 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1876 // Okay, we collected all the mapped values and checked them for sanity, and
1877 // defined to really do this transformation. First, update the CFG.
1878 TrueSucc->removePredecessor(BI->getParent());
1879 FalseSucc->removePredecessor(BI->getParent());
1881 // Insert select instructions where needed.
1882 Value *BrCond = BI->getCondition();
1884 // Insert a select if the results differ.
1885 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1886 } else if (isa<UndefValue>(TrueValue)) {
1887 TrueValue = FalseValue;
1889 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1890 FalseValue, "retval");
1894 Value *RI = !TrueValue ?
1895 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
1899 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
1900 << "\n " << *BI << "NewRet = " << *RI
1901 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
1903 EraseTerminatorInstAndDCECond(BI);
1908 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
1909 /// probabilities of the branch taking each edge. Fills in the two APInt
1910 /// parameters and return true, or returns false if no or invalid metadata was
1912 static bool ExtractBranchMetadata(BranchInst *BI,
1913 uint64_t &ProbTrue, uint64_t &ProbFalse) {
1914 assert(BI->isConditional() &&
1915 "Looking for probabilities on unconditional branch?");
1916 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
1917 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
1918 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
1919 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
1920 if (!CITrue || !CIFalse) return false;
1921 ProbTrue = CITrue->getValue().getZExtValue();
1922 ProbFalse = CIFalse->getValue().getZExtValue();
1926 /// checkCSEInPredecessor - Return true if the given instruction is available
1927 /// in its predecessor block. If yes, the instruction will be removed.
1929 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
1930 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
1932 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
1933 Instruction *PBI = &*I;
1934 // Check whether Inst and PBI generate the same value.
1935 if (Inst->isIdenticalTo(PBI)) {
1936 Inst->replaceAllUsesWith(PBI);
1937 Inst->eraseFromParent();
1944 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
1945 /// predecessor branches to us and one of our successors, fold the block into
1946 /// the predecessor and use logical operations to pick the right destination.
1947 bool llvm::FoldBranchToCommonDest(BranchInst *BI) {
1948 BasicBlock *BB = BI->getParent();
1950 Instruction *Cond = 0;
1951 if (BI->isConditional())
1952 Cond = dyn_cast<Instruction>(BI->getCondition());
1954 // For unconditional branch, check for a simple CFG pattern, where
1955 // BB has a single predecessor and BB's successor is also its predecessor's
1956 // successor. If such pattern exisits, check for CSE between BB and its
1958 if (BasicBlock *PB = BB->getSinglePredecessor())
1959 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
1960 if (PBI->isConditional() &&
1961 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
1962 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
1963 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
1965 Instruction *Curr = I++;
1966 if (isa<CmpInst>(Curr)) {
1970 // Quit if we can't remove this instruction.
1971 if (!checkCSEInPredecessor(Curr, PB))
1980 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
1981 Cond->getParent() != BB || !Cond->hasOneUse())
1984 // Only allow this if the condition is a simple instruction that can be
1985 // executed unconditionally. It must be in the same block as the branch, and
1986 // must be at the front of the block.
1987 BasicBlock::iterator FrontIt = BB->front();
1989 // Ignore dbg intrinsics.
1990 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
1992 // Allow a single instruction to be hoisted in addition to the compare
1993 // that feeds the branch. We later ensure that any values that _it_ uses
1994 // were also live in the predecessor, so that we don't unnecessarily create
1995 // register pressure or inhibit out-of-order execution.
1996 Instruction *BonusInst = 0;
1997 if (&*FrontIt != Cond &&
1998 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond &&
1999 isSafeToSpeculativelyExecute(FrontIt)) {
2000 BonusInst = &*FrontIt;
2003 // Ignore dbg intrinsics.
2004 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
2007 // Only a single bonus inst is allowed.
2008 if (&*FrontIt != Cond)
2011 // Make sure the instruction after the condition is the cond branch.
2012 BasicBlock::iterator CondIt = Cond; ++CondIt;
2014 // Ingore dbg intrinsics.
2015 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
2020 // Cond is known to be a compare or binary operator. Check to make sure that
2021 // neither operand is a potentially-trapping constant expression.
2022 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2025 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2029 // Finally, don't infinitely unroll conditional loops.
2030 BasicBlock *TrueDest = BI->getSuccessor(0);
2031 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0;
2032 if (TrueDest == BB || FalseDest == BB)
2035 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2036 BasicBlock *PredBlock = *PI;
2037 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2039 // Check that we have two conditional branches. If there is a PHI node in
2040 // the common successor, verify that the same value flows in from both
2042 SmallVector<PHINode*, 4> PHIs;
2043 if (PBI == 0 || PBI->isUnconditional() ||
2044 (BI->isConditional() &&
2045 !SafeToMergeTerminators(BI, PBI)) ||
2046 (!BI->isConditional() &&
2047 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2050 // Determine if the two branches share a common destination.
2051 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2052 bool InvertPredCond = false;
2054 if (BI->isConditional()) {
2055 if (PBI->getSuccessor(0) == TrueDest)
2056 Opc = Instruction::Or;
2057 else if (PBI->getSuccessor(1) == FalseDest)
2058 Opc = Instruction::And;
2059 else if (PBI->getSuccessor(0) == FalseDest)
2060 Opc = Instruction::And, InvertPredCond = true;
2061 else if (PBI->getSuccessor(1) == TrueDest)
2062 Opc = Instruction::Or, InvertPredCond = true;
2066 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2070 // Ensure that any values used in the bonus instruction are also used
2071 // by the terminator of the predecessor. This means that those values
2072 // must already have been resolved, so we won't be inhibiting the
2073 // out-of-order core by speculating them earlier.
2075 // Collect the values used by the bonus inst
2076 SmallPtrSet<Value*, 4> UsedValues;
2077 for (Instruction::op_iterator OI = BonusInst->op_begin(),
2078 OE = BonusInst->op_end(); OI != OE; ++OI) {
2080 if (!isa<Constant>(V))
2081 UsedValues.insert(V);
2084 SmallVector<std::pair<Value*, unsigned>, 4> Worklist;
2085 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0));
2087 // Walk up to four levels back up the use-def chain of the predecessor's
2088 // terminator to see if all those values were used. The choice of four
2089 // levels is arbitrary, to provide a compile-time-cost bound.
2090 while (!Worklist.empty()) {
2091 std::pair<Value*, unsigned> Pair = Worklist.back();
2092 Worklist.pop_back();
2094 if (Pair.second >= 4) continue;
2095 UsedValues.erase(Pair.first);
2096 if (UsedValues.empty()) break;
2098 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) {
2099 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
2101 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1));
2105 if (!UsedValues.empty()) return false;
2108 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2109 IRBuilder<> Builder(PBI);
2111 // If we need to invert the condition in the pred block to match, do so now.
2112 if (InvertPredCond) {
2113 Value *NewCond = PBI->getCondition();
2115 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2116 CmpInst *CI = cast<CmpInst>(NewCond);
2117 CI->setPredicate(CI->getInversePredicate());
2119 NewCond = Builder.CreateNot(NewCond,
2120 PBI->getCondition()->getName()+".not");
2123 PBI->setCondition(NewCond);
2124 PBI->swapSuccessors();
2127 // If we have a bonus inst, clone it into the predecessor block.
2128 Instruction *NewBonus = 0;
2130 NewBonus = BonusInst->clone();
2131 PredBlock->getInstList().insert(PBI, NewBonus);
2132 NewBonus->takeName(BonusInst);
2133 BonusInst->setName(BonusInst->getName()+".old");
2136 // Clone Cond into the predecessor basic block, and or/and the
2137 // two conditions together.
2138 Instruction *New = Cond->clone();
2139 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus);
2140 PredBlock->getInstList().insert(PBI, New);
2141 New->takeName(Cond);
2142 Cond->setName(New->getName()+".old");
2144 if (BI->isConditional()) {
2145 Instruction *NewCond =
2146 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2148 PBI->setCondition(NewCond);
2150 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2151 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2153 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2155 SmallVector<uint64_t, 8> NewWeights;
2157 if (PBI->getSuccessor(0) == BB) {
2158 if (PredHasWeights && SuccHasWeights) {
2159 // PBI: br i1 %x, BB, FalseDest
2160 // BI: br i1 %y, TrueDest, FalseDest
2161 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2162 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2163 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2164 // TrueWeight for PBI * FalseWeight for BI.
2165 // We assume that total weights of a BranchInst can fit into 32 bits.
2166 // Therefore, we will not have overflow using 64-bit arithmetic.
2167 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2168 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2170 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2171 PBI->setSuccessor(0, TrueDest);
2173 if (PBI->getSuccessor(1) == BB) {
2174 if (PredHasWeights && SuccHasWeights) {
2175 // PBI: br i1 %x, TrueDest, BB
2176 // BI: br i1 %y, TrueDest, FalseDest
2177 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2178 // FalseWeight for PBI * TrueWeight for BI.
2179 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2180 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2181 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2182 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2184 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2185 PBI->setSuccessor(1, FalseDest);
2187 if (NewWeights.size() == 2) {
2188 // Halve the weights if any of them cannot fit in an uint32_t
2189 FitWeights(NewWeights);
2191 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2192 PBI->setMetadata(LLVMContext::MD_prof,
2193 MDBuilder(BI->getContext()).
2194 createBranchWeights(MDWeights));
2196 PBI->setMetadata(LLVMContext::MD_prof, NULL);
2198 // Update PHI nodes in the common successors.
2199 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2200 ConstantInt *PBI_C = cast<ConstantInt>(
2201 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2202 assert(PBI_C->getType()->isIntegerTy(1));
2203 Instruction *MergedCond = 0;
2204 if (PBI->getSuccessor(0) == TrueDest) {
2205 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2206 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2207 // is false: !PBI_Cond and BI_Value
2208 Instruction *NotCond =
2209 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2212 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2217 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2218 PBI->getCondition(), MergedCond,
2221 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2222 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2223 // is false: PBI_Cond and BI_Value
2225 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2226 PBI->getCondition(), New,
2228 if (PBI_C->isOne()) {
2229 Instruction *NotCond =
2230 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2233 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2234 NotCond, MergedCond,
2239 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2242 // Change PBI from Conditional to Unconditional.
2243 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2244 EraseTerminatorInstAndDCECond(PBI);
2248 // TODO: If BB is reachable from all paths through PredBlock, then we
2249 // could replace PBI's branch probabilities with BI's.
2251 // Copy any debug value intrinsics into the end of PredBlock.
2252 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2253 if (isa<DbgInfoIntrinsic>(*I))
2254 I->clone()->insertBefore(PBI);
2261 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2262 /// predecessor of another block, this function tries to simplify it. We know
2263 /// that PBI and BI are both conditional branches, and BI is in one of the
2264 /// successor blocks of PBI - PBI branches to BI.
2265 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2266 assert(PBI->isConditional() && BI->isConditional());
2267 BasicBlock *BB = BI->getParent();
2269 // If this block ends with a branch instruction, and if there is a
2270 // predecessor that ends on a branch of the same condition, make
2271 // this conditional branch redundant.
2272 if (PBI->getCondition() == BI->getCondition() &&
2273 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2274 // Okay, the outcome of this conditional branch is statically
2275 // knowable. If this block had a single pred, handle specially.
2276 if (BB->getSinglePredecessor()) {
2277 // Turn this into a branch on constant.
2278 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2279 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2281 return true; // Nuke the branch on constant.
2284 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2285 // in the constant and simplify the block result. Subsequent passes of
2286 // simplifycfg will thread the block.
2287 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2288 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2289 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2290 std::distance(PB, PE),
2291 BI->getCondition()->getName() + ".pr",
2293 // Okay, we're going to insert the PHI node. Since PBI is not the only
2294 // predecessor, compute the PHI'd conditional value for all of the preds.
2295 // Any predecessor where the condition is not computable we keep symbolic.
2296 for (pred_iterator PI = PB; PI != PE; ++PI) {
2297 BasicBlock *P = *PI;
2298 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2299 PBI != BI && PBI->isConditional() &&
2300 PBI->getCondition() == BI->getCondition() &&
2301 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2302 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2303 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2306 NewPN->addIncoming(BI->getCondition(), P);
2310 BI->setCondition(NewPN);
2315 // If this is a conditional branch in an empty block, and if any
2316 // predecessors is a conditional branch to one of our destinations,
2317 // fold the conditions into logical ops and one cond br.
2318 BasicBlock::iterator BBI = BB->begin();
2319 // Ignore dbg intrinsics.
2320 while (isa<DbgInfoIntrinsic>(BBI))
2326 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2331 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2333 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2334 PBIOp = 0, BIOp = 1;
2335 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2336 PBIOp = 1, BIOp = 0;
2337 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2342 // Check to make sure that the other destination of this branch
2343 // isn't BB itself. If so, this is an infinite loop that will
2344 // keep getting unwound.
2345 if (PBI->getSuccessor(PBIOp) == BB)
2348 // Do not perform this transformation if it would require
2349 // insertion of a large number of select instructions. For targets
2350 // without predication/cmovs, this is a big pessimization.
2351 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2353 unsigned NumPhis = 0;
2354 for (BasicBlock::iterator II = CommonDest->begin();
2355 isa<PHINode>(II); ++II, ++NumPhis)
2356 if (NumPhis > 2) // Disable this xform.
2359 // Finally, if everything is ok, fold the branches to logical ops.
2360 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2362 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2363 << "AND: " << *BI->getParent());
2366 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2367 // branch in it, where one edge (OtherDest) goes back to itself but the other
2368 // exits. We don't *know* that the program avoids the infinite loop
2369 // (even though that seems likely). If we do this xform naively, we'll end up
2370 // recursively unpeeling the loop. Since we know that (after the xform is
2371 // done) that the block *is* infinite if reached, we just make it an obviously
2372 // infinite loop with no cond branch.
2373 if (OtherDest == BB) {
2374 // Insert it at the end of the function, because it's either code,
2375 // or it won't matter if it's hot. :)
2376 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2377 "infloop", BB->getParent());
2378 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2379 OtherDest = InfLoopBlock;
2382 DEBUG(dbgs() << *PBI->getParent()->getParent());
2384 // BI may have other predecessors. Because of this, we leave
2385 // it alone, but modify PBI.
2387 // Make sure we get to CommonDest on True&True directions.
2388 Value *PBICond = PBI->getCondition();
2389 IRBuilder<true, NoFolder> Builder(PBI);
2391 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2393 Value *BICond = BI->getCondition();
2395 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2397 // Merge the conditions.
2398 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2400 // Modify PBI to branch on the new condition to the new dests.
2401 PBI->setCondition(Cond);
2402 PBI->setSuccessor(0, CommonDest);
2403 PBI->setSuccessor(1, OtherDest);
2405 // Update branch weight for PBI.
2406 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2407 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2409 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2411 if (PredHasWeights && SuccHasWeights) {
2412 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2413 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2414 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2415 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2416 // The weight to CommonDest should be PredCommon * SuccTotal +
2417 // PredOther * SuccCommon.
2418 // The weight to OtherDest should be PredOther * SuccOther.
2419 SmallVector<uint64_t, 2> NewWeights;
2420 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) +
2421 PredOther * SuccCommon);
2422 NewWeights.push_back(PredOther * SuccOther);
2423 // Halve the weights if any of them cannot fit in an uint32_t
2424 FitWeights(NewWeights);
2426 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end());
2427 PBI->setMetadata(LLVMContext::MD_prof,
2428 MDBuilder(BI->getContext()).
2429 createBranchWeights(MDWeights));
2432 // OtherDest may have phi nodes. If so, add an entry from PBI's
2433 // block that are identical to the entries for BI's block.
2434 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2436 // We know that the CommonDest already had an edge from PBI to
2437 // it. If it has PHIs though, the PHIs may have different
2438 // entries for BB and PBI's BB. If so, insert a select to make
2441 for (BasicBlock::iterator II = CommonDest->begin();
2442 (PN = dyn_cast<PHINode>(II)); ++II) {
2443 Value *BIV = PN->getIncomingValueForBlock(BB);
2444 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2445 Value *PBIV = PN->getIncomingValue(PBBIdx);
2447 // Insert a select in PBI to pick the right value.
2448 Value *NV = cast<SelectInst>
2449 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2450 PN->setIncomingValue(PBBIdx, NV);
2454 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2455 DEBUG(dbgs() << *PBI->getParent()->getParent());
2457 // This basic block is probably dead. We know it has at least
2458 // one fewer predecessor.
2462 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2463 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2464 // Takes care of updating the successors and removing the old terminator.
2465 // Also makes sure not to introduce new successors by assuming that edges to
2466 // non-successor TrueBBs and FalseBBs aren't reachable.
2467 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2468 BasicBlock *TrueBB, BasicBlock *FalseBB,
2469 uint32_t TrueWeight,
2470 uint32_t FalseWeight){
2471 // Remove any superfluous successor edges from the CFG.
2472 // First, figure out which successors to preserve.
2473 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2475 BasicBlock *KeepEdge1 = TrueBB;
2476 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0;
2478 // Then remove the rest.
2479 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2480 BasicBlock *Succ = OldTerm->getSuccessor(I);
2481 // Make sure only to keep exactly one copy of each edge.
2482 if (Succ == KeepEdge1)
2484 else if (Succ == KeepEdge2)
2487 Succ->removePredecessor(OldTerm->getParent());
2490 IRBuilder<> Builder(OldTerm);
2491 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2493 // Insert an appropriate new terminator.
2494 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) {
2495 if (TrueBB == FalseBB)
2496 // We were only looking for one successor, and it was present.
2497 // Create an unconditional branch to it.
2498 Builder.CreateBr(TrueBB);
2500 // We found both of the successors we were looking for.
2501 // Create a conditional branch sharing the condition of the select.
2502 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2503 if (TrueWeight != FalseWeight)
2504 NewBI->setMetadata(LLVMContext::MD_prof,
2505 MDBuilder(OldTerm->getContext()).
2506 createBranchWeights(TrueWeight, FalseWeight));
2508 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2509 // Neither of the selected blocks were successors, so this
2510 // terminator must be unreachable.
2511 new UnreachableInst(OldTerm->getContext(), OldTerm);
2513 // One of the selected values was a successor, but the other wasn't.
2514 // Insert an unconditional branch to the one that was found;
2515 // the edge to the one that wasn't must be unreachable.
2517 // Only TrueBB was found.
2518 Builder.CreateBr(TrueBB);
2520 // Only FalseBB was found.
2521 Builder.CreateBr(FalseBB);
2524 EraseTerminatorInstAndDCECond(OldTerm);
2528 // SimplifySwitchOnSelect - Replaces
2529 // (switch (select cond, X, Y)) on constant X, Y
2530 // with a branch - conditional if X and Y lead to distinct BBs,
2531 // unconditional otherwise.
2532 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2533 // Check for constant integer values in the select.
2534 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2535 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2536 if (!TrueVal || !FalseVal)
2539 // Find the relevant condition and destinations.
2540 Value *Condition = Select->getCondition();
2541 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2542 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2544 // Get weight for TrueBB and FalseBB.
2545 uint32_t TrueWeight = 0, FalseWeight = 0;
2546 SmallVector<uint64_t, 8> Weights;
2547 bool HasWeights = HasBranchWeights(SI);
2549 GetBranchWeights(SI, Weights);
2550 if (Weights.size() == 1 + SI->getNumCases()) {
2551 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2552 getSuccessorIndex()];
2553 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2554 getSuccessorIndex()];
2558 // Perform the actual simplification.
2559 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2560 TrueWeight, FalseWeight);
2563 // SimplifyIndirectBrOnSelect - Replaces
2564 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2565 // blockaddress(@fn, BlockB)))
2567 // (br cond, BlockA, BlockB).
2568 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2569 // Check that both operands of the select are block addresses.
2570 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2571 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2575 // Extract the actual blocks.
2576 BasicBlock *TrueBB = TBA->getBasicBlock();
2577 BasicBlock *FalseBB = FBA->getBasicBlock();
2579 // Perform the actual simplification.
2580 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2584 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2585 /// instruction (a seteq/setne with a constant) as the only instruction in a
2586 /// block that ends with an uncond branch. We are looking for a very specific
2587 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2588 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2589 /// default value goes to an uncond block with a seteq in it, we get something
2592 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2594 /// %tmp = icmp eq i8 %A, 92
2597 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2599 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2600 /// the PHI, merging the third icmp into the switch.
2601 static bool TryToSimplifyUncondBranchWithICmpInIt(
2602 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI,
2603 const DataLayout *TD) {
2604 BasicBlock *BB = ICI->getParent();
2606 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2608 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2610 Value *V = ICI->getOperand(0);
2611 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2613 // The pattern we're looking for is where our only predecessor is a switch on
2614 // 'V' and this block is the default case for the switch. In this case we can
2615 // fold the compared value into the switch to simplify things.
2616 BasicBlock *Pred = BB->getSinglePredecessor();
2617 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false;
2619 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2620 if (SI->getCondition() != V)
2623 // If BB is reachable on a non-default case, then we simply know the value of
2624 // V in this block. Substitute it and constant fold the icmp instruction
2626 if (SI->getDefaultDest() != BB) {
2627 ConstantInt *VVal = SI->findCaseDest(BB);
2628 assert(VVal && "Should have a unique destination value");
2629 ICI->setOperand(0, VVal);
2631 if (Value *V = SimplifyInstruction(ICI, TD)) {
2632 ICI->replaceAllUsesWith(V);
2633 ICI->eraseFromParent();
2635 // BB is now empty, so it is likely to simplify away.
2636 return SimplifyCFG(BB, TTI, TD) | true;
2639 // Ok, the block is reachable from the default dest. If the constant we're
2640 // comparing exists in one of the other edges, then we can constant fold ICI
2642 if (SI->findCaseValue(Cst) != SI->case_default()) {
2644 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2645 V = ConstantInt::getFalse(BB->getContext());
2647 V = ConstantInt::getTrue(BB->getContext());
2649 ICI->replaceAllUsesWith(V);
2650 ICI->eraseFromParent();
2651 // BB is now empty, so it is likely to simplify away.
2652 return SimplifyCFG(BB, TTI, TD) | true;
2655 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2657 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2658 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back());
2659 if (PHIUse == 0 || PHIUse != &SuccBlock->front() ||
2660 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2663 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2665 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2666 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2668 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2669 std::swap(DefaultCst, NewCst);
2671 // Replace ICI (which is used by the PHI for the default value) with true or
2672 // false depending on if it is EQ or NE.
2673 ICI->replaceAllUsesWith(DefaultCst);
2674 ICI->eraseFromParent();
2676 // Okay, the switch goes to this block on a default value. Add an edge from
2677 // the switch to the merge point on the compared value.
2678 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2679 BB->getParent(), BB);
2680 SmallVector<uint64_t, 8> Weights;
2681 bool HasWeights = HasBranchWeights(SI);
2683 GetBranchWeights(SI, Weights);
2684 if (Weights.size() == 1 + SI->getNumCases()) {
2685 // Split weight for default case to case for "Cst".
2686 Weights[0] = (Weights[0]+1) >> 1;
2687 Weights.push_back(Weights[0]);
2689 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2690 SI->setMetadata(LLVMContext::MD_prof,
2691 MDBuilder(SI->getContext()).
2692 createBranchWeights(MDWeights));
2695 SI->addCase(Cst, NewBB);
2697 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2698 Builder.SetInsertPoint(NewBB);
2699 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2700 Builder.CreateBr(SuccBlock);
2701 PHIUse->addIncoming(NewCst, NewBB);
2705 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2706 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2707 /// fold it into a switch instruction if so.
2708 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *TD,
2709 IRBuilder<> &Builder) {
2710 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2711 if (Cond == 0) return false;
2714 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2715 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2716 // 'setne's and'ed together, collect them.
2718 std::vector<ConstantInt*> Values;
2719 bool TrueWhenEqual = true;
2720 Value *ExtraCase = 0;
2721 unsigned UsedICmps = 0;
2723 if (Cond->getOpcode() == Instruction::Or) {
2724 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true,
2726 } else if (Cond->getOpcode() == Instruction::And) {
2727 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false,
2729 TrueWhenEqual = false;
2732 // If we didn't have a multiply compared value, fail.
2733 if (CompVal == 0) return false;
2735 // Avoid turning single icmps into a switch.
2739 // There might be duplicate constants in the list, which the switch
2740 // instruction can't handle, remove them now.
2741 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2742 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2744 // If Extra was used, we require at least two switch values to do the
2745 // transformation. A switch with one value is just an cond branch.
2746 if (ExtraCase && Values.size() < 2) return false;
2748 // TODO: Preserve branch weight metadata, similarly to how
2749 // FoldValueComparisonIntoPredecessors preserves it.
2751 // Figure out which block is which destination.
2752 BasicBlock *DefaultBB = BI->getSuccessor(1);
2753 BasicBlock *EdgeBB = BI->getSuccessor(0);
2754 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2756 BasicBlock *BB = BI->getParent();
2758 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2759 << " cases into SWITCH. BB is:\n" << *BB);
2761 // If there are any extra values that couldn't be folded into the switch
2762 // then we evaluate them with an explicit branch first. Split the block
2763 // right before the condbr to handle it.
2765 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2766 // Remove the uncond branch added to the old block.
2767 TerminatorInst *OldTI = BB->getTerminator();
2768 Builder.SetInsertPoint(OldTI);
2771 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2773 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2775 OldTI->eraseFromParent();
2777 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2778 // for the edge we just added.
2779 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2781 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2782 << "\nEXTRABB = " << *BB);
2786 Builder.SetInsertPoint(BI);
2787 // Convert pointer to int before we switch.
2788 if (CompVal->getType()->isPointerTy()) {
2789 assert(TD && "Cannot switch on pointer without DataLayout");
2790 CompVal = Builder.CreatePtrToInt(CompVal,
2791 TD->getIntPtrType(CompVal->getContext()),
2795 // Create the new switch instruction now.
2796 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2798 // Add all of the 'cases' to the switch instruction.
2799 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2800 New->addCase(Values[i], EdgeBB);
2802 // We added edges from PI to the EdgeBB. As such, if there were any
2803 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2804 // the number of edges added.
2805 for (BasicBlock::iterator BBI = EdgeBB->begin();
2806 isa<PHINode>(BBI); ++BBI) {
2807 PHINode *PN = cast<PHINode>(BBI);
2808 Value *InVal = PN->getIncomingValueForBlock(BB);
2809 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2810 PN->addIncoming(InVal, BB);
2813 // Erase the old branch instruction.
2814 EraseTerminatorInstAndDCECond(BI);
2816 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2820 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2821 // If this is a trivial landing pad that just continues unwinding the caught
2822 // exception then zap the landing pad, turning its invokes into calls.
2823 BasicBlock *BB = RI->getParent();
2824 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2825 if (RI->getValue() != LPInst)
2826 // Not a landing pad, or the resume is not unwinding the exception that
2827 // caused control to branch here.
2830 // Check that there are no other instructions except for debug intrinsics.
2831 BasicBlock::iterator I = LPInst, E = RI;
2833 if (!isa<DbgInfoIntrinsic>(I))
2836 // Turn all invokes that unwind here into calls and delete the basic block.
2837 bool InvokeRequiresTableEntry = false;
2838 bool Changed = false;
2839 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2840 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2842 if (II->hasFnAttr(Attribute::UWTable)) {
2843 // Don't remove an `invoke' instruction if the ABI requires an entry into
2845 InvokeRequiresTableEntry = true;
2849 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2851 // Insert a call instruction before the invoke.
2852 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2854 Call->setCallingConv(II->getCallingConv());
2855 Call->setAttributes(II->getAttributes());
2856 Call->setDebugLoc(II->getDebugLoc());
2858 // Anything that used the value produced by the invoke instruction now uses
2859 // the value produced by the call instruction. Note that we do this even
2860 // for void functions and calls with no uses so that the callgraph edge is
2862 II->replaceAllUsesWith(Call);
2863 BB->removePredecessor(II->getParent());
2865 // Insert a branch to the normal destination right before the invoke.
2866 BranchInst::Create(II->getNormalDest(), II);
2868 // Finally, delete the invoke instruction!
2869 II->eraseFromParent();
2873 if (!InvokeRequiresTableEntry)
2874 // The landingpad is now unreachable. Zap it.
2875 BB->eraseFromParent();
2880 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2881 BasicBlock *BB = RI->getParent();
2882 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2884 // Find predecessors that end with branches.
2885 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2886 SmallVector<BranchInst*, 8> CondBranchPreds;
2887 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2888 BasicBlock *P = *PI;
2889 TerminatorInst *PTI = P->getTerminator();
2890 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2891 if (BI->isUnconditional())
2892 UncondBranchPreds.push_back(P);
2894 CondBranchPreds.push_back(BI);
2898 // If we found some, do the transformation!
2899 if (!UncondBranchPreds.empty() && DupRet) {
2900 while (!UncondBranchPreds.empty()) {
2901 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
2902 DEBUG(dbgs() << "FOLDING: " << *BB
2903 << "INTO UNCOND BRANCH PRED: " << *Pred);
2904 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
2907 // If we eliminated all predecessors of the block, delete the block now.
2908 if (pred_begin(BB) == pred_end(BB))
2909 // We know there are no successors, so just nuke the block.
2910 BB->eraseFromParent();
2915 // Check out all of the conditional branches going to this return
2916 // instruction. If any of them just select between returns, change the
2917 // branch itself into a select/return pair.
2918 while (!CondBranchPreds.empty()) {
2919 BranchInst *BI = CondBranchPreds.pop_back_val();
2921 // Check to see if the non-BB successor is also a return block.
2922 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
2923 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
2924 SimplifyCondBranchToTwoReturns(BI, Builder))
2930 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
2931 BasicBlock *BB = UI->getParent();
2933 bool Changed = false;
2935 // If there are any instructions immediately before the unreachable that can
2936 // be removed, do so.
2937 while (UI != BB->begin()) {
2938 BasicBlock::iterator BBI = UI;
2940 // Do not delete instructions that can have side effects which might cause
2941 // the unreachable to not be reachable; specifically, calls and volatile
2942 // operations may have this effect.
2943 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
2945 if (BBI->mayHaveSideEffects()) {
2946 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
2947 if (SI->isVolatile())
2949 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
2950 if (LI->isVolatile())
2952 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
2953 if (RMWI->isVolatile())
2955 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
2956 if (CXI->isVolatile())
2958 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
2959 !isa<LandingPadInst>(BBI)) {
2962 // Note that deleting LandingPad's here is in fact okay, although it
2963 // involves a bit of subtle reasoning. If this inst is a LandingPad,
2964 // all the predecessors of this block will be the unwind edges of Invokes,
2965 // and we can therefore guarantee this block will be erased.
2968 // Delete this instruction (any uses are guaranteed to be dead)
2969 if (!BBI->use_empty())
2970 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
2971 BBI->eraseFromParent();
2975 // If the unreachable instruction is the first in the block, take a gander
2976 // at all of the predecessors of this instruction, and simplify them.
2977 if (&BB->front() != UI) return Changed;
2979 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
2980 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
2981 TerminatorInst *TI = Preds[i]->getTerminator();
2982 IRBuilder<> Builder(TI);
2983 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
2984 if (BI->isUnconditional()) {
2985 if (BI->getSuccessor(0) == BB) {
2986 new UnreachableInst(TI->getContext(), TI);
2987 TI->eraseFromParent();
2991 if (BI->getSuccessor(0) == BB) {
2992 Builder.CreateBr(BI->getSuccessor(1));
2993 EraseTerminatorInstAndDCECond(BI);
2994 } else if (BI->getSuccessor(1) == BB) {
2995 Builder.CreateBr(BI->getSuccessor(0));
2996 EraseTerminatorInstAndDCECond(BI);
3000 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3001 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3003 if (i.getCaseSuccessor() == BB) {
3004 BB->removePredecessor(SI->getParent());
3009 // If the default value is unreachable, figure out the most popular
3010 // destination and make it the default.
3011 if (SI->getDefaultDest() == BB) {
3012 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
3013 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3015 std::pair<unsigned, unsigned> &entry =
3016 Popularity[i.getCaseSuccessor()];
3017 if (entry.first == 0) {
3019 entry.second = i.getCaseIndex();
3025 // Find the most popular block.
3026 unsigned MaxPop = 0;
3027 unsigned MaxIndex = 0;
3028 BasicBlock *MaxBlock = 0;
3029 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
3030 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
3031 if (I->second.first > MaxPop ||
3032 (I->second.first == MaxPop && MaxIndex > I->second.second)) {
3033 MaxPop = I->second.first;
3034 MaxIndex = I->second.second;
3035 MaxBlock = I->first;
3039 // Make this the new default, allowing us to delete any explicit
3041 SI->setDefaultDest(MaxBlock);
3044 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
3046 if (isa<PHINode>(MaxBlock->begin()))
3047 for (unsigned i = 0; i != MaxPop-1; ++i)
3048 MaxBlock->removePredecessor(SI->getParent());
3050 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3052 if (i.getCaseSuccessor() == MaxBlock) {
3058 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
3059 if (II->getUnwindDest() == BB) {
3060 // Convert the invoke to a call instruction. This would be a good
3061 // place to note that the call does not throw though.
3062 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
3063 II->removeFromParent(); // Take out of symbol table
3065 // Insert the call now...
3066 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
3067 Builder.SetInsertPoint(BI);
3068 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
3069 Args, II->getName());
3070 CI->setCallingConv(II->getCallingConv());
3071 CI->setAttributes(II->getAttributes());
3072 // If the invoke produced a value, the call does now instead.
3073 II->replaceAllUsesWith(CI);
3080 // If this block is now dead, remove it.
3081 if (pred_begin(BB) == pred_end(BB) &&
3082 BB != &BB->getParent()->getEntryBlock()) {
3083 // We know there are no successors, so just nuke the block.
3084 BB->eraseFromParent();
3091 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
3092 /// integer range comparison into a sub, an icmp and a branch.
3093 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3094 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3096 // Make sure all cases point to the same destination and gather the values.
3097 SmallVector<ConstantInt *, 16> Cases;
3098 SwitchInst::CaseIt I = SI->case_begin();
3099 Cases.push_back(I.getCaseValue());
3100 SwitchInst::CaseIt PrevI = I++;
3101 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
3102 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
3104 Cases.push_back(I.getCaseValue());
3106 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
3108 // Sort the case values, then check if they form a range we can transform.
3109 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3110 for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
3111 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
3115 Constant *Offset = ConstantExpr::getNeg(Cases.back());
3116 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
3118 Value *Sub = SI->getCondition();
3119 if (!Offset->isNullValue())
3120 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
3122 // If NumCases overflowed, then all possible values jump to the successor.
3123 if (NumCases->isNullValue() && SI->getNumCases() != 0)
3124 Cmp = ConstantInt::getTrue(SI->getContext());
3126 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3127 BranchInst *NewBI = Builder.CreateCondBr(
3128 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
3130 // Update weight for the newly-created conditional branch.
3131 SmallVector<uint64_t, 8> Weights;
3132 bool HasWeights = HasBranchWeights(SI);
3134 GetBranchWeights(SI, Weights);
3135 if (Weights.size() == 1 + SI->getNumCases()) {
3136 // Combine all weights for the cases to be the true weight of NewBI.
3137 // We assume that the sum of all weights for a Terminator can fit into 32
3139 uint32_t NewTrueWeight = 0;
3140 for (unsigned I = 1, E = Weights.size(); I != E; ++I)
3141 NewTrueWeight += (uint32_t)Weights[I];
3142 NewBI->setMetadata(LLVMContext::MD_prof,
3143 MDBuilder(SI->getContext()).
3144 createBranchWeights(NewTrueWeight,
3145 (uint32_t)Weights[0]));
3149 // Prune obsolete incoming values off the successor's PHI nodes.
3150 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
3151 isa<PHINode>(BBI); ++BBI) {
3152 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
3153 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3155 SI->eraseFromParent();
3160 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
3161 /// and use it to remove dead cases.
3162 static bool EliminateDeadSwitchCases(SwitchInst *SI) {
3163 Value *Cond = SI->getCondition();
3164 unsigned Bits = Cond->getType()->getIntegerBitWidth();
3165 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3166 ComputeMaskedBits(Cond, KnownZero, KnownOne);
3168 // Gather dead cases.
3169 SmallVector<ConstantInt*, 8> DeadCases;
3170 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3171 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3172 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3173 DeadCases.push_back(I.getCaseValue());
3174 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3175 << I.getCaseValue() << "' is dead.\n");
3179 SmallVector<uint64_t, 8> Weights;
3180 bool HasWeight = HasBranchWeights(SI);
3182 GetBranchWeights(SI, Weights);
3183 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3186 // Remove dead cases from the switch.
3187 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3188 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3189 assert(Case != SI->case_default() &&
3190 "Case was not found. Probably mistake in DeadCases forming.");
3192 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3196 // Prune unused values from PHI nodes.
3197 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3198 SI->removeCase(Case);
3201 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3202 SI->setMetadata(LLVMContext::MD_prof,
3203 MDBuilder(SI->getParent()->getContext()).
3204 createBranchWeights(MDWeights));
3207 return !DeadCases.empty();
3210 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
3211 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3212 /// by an unconditional branch), look at the phi node for BB in the successor
3213 /// block and see if the incoming value is equal to CaseValue. If so, return
3214 /// the phi node, and set PhiIndex to BB's index in the phi node.
3215 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3218 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3219 return NULL; // BB must be empty to be a candidate for simplification.
3220 if (!BB->getSinglePredecessor())
3221 return NULL; // BB must be dominated by the switch.
3223 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3224 if (!Branch || !Branch->isUnconditional())
3225 return NULL; // Terminator must be unconditional branch.
3227 BasicBlock *Succ = Branch->getSuccessor(0);
3229 BasicBlock::iterator I = Succ->begin();
3230 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3231 int Idx = PHI->getBasicBlockIndex(BB);
3232 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3234 Value *InValue = PHI->getIncomingValue(Idx);
3235 if (InValue != CaseValue) continue;
3244 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
3245 /// instruction to a phi node dominated by the switch, if that would mean that
3246 /// some of the destination blocks of the switch can be folded away.
3247 /// Returns true if a change is made.
3248 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3249 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3250 ForwardingNodesMap ForwardingNodes;
3252 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3253 ConstantInt *CaseValue = I.getCaseValue();
3254 BasicBlock *CaseDest = I.getCaseSuccessor();
3257 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3261 ForwardingNodes[PHI].push_back(PhiIndex);
3264 bool Changed = false;
3266 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3267 E = ForwardingNodes.end(); I != E; ++I) {
3268 PHINode *Phi = I->first;
3269 SmallVectorImpl<int> &Indexes = I->second;
3271 if (Indexes.size() < 2) continue;
3273 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3274 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3281 /// ValidLookupTableConstant - Return true if the backend will be able to handle
3282 /// initializing an array of constants like C.
3283 static bool ValidLookupTableConstant(Constant *C) {
3284 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3285 return CE->isGEPWithNoNotionalOverIndexing();
3287 return isa<ConstantFP>(C) ||
3288 isa<ConstantInt>(C) ||
3289 isa<ConstantPointerNull>(C) ||
3290 isa<GlobalValue>(C) ||
3294 /// LookupConstant - If V is a Constant, return it. Otherwise, try to look up
3295 /// its constant value in ConstantPool, returning 0 if it's not there.
3296 static Constant *LookupConstant(Value *V,
3297 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3298 if (Constant *C = dyn_cast<Constant>(V))
3300 return ConstantPool.lookup(V);
3303 /// ConstantFold - Try to fold instruction I into a constant. This works for
3304 /// simple instructions such as binary operations where both operands are
3305 /// constant or can be replaced by constants from the ConstantPool. Returns the
3306 /// resulting constant on success, 0 otherwise.
3307 static Constant *ConstantFold(Instruction *I,
3308 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3309 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
3310 Constant *A = LookupConstant(BO->getOperand(0), ConstantPool);
3313 Constant *B = LookupConstant(BO->getOperand(1), ConstantPool);
3316 return ConstantExpr::get(BO->getOpcode(), A, B);
3319 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
3320 Constant *A = LookupConstant(I->getOperand(0), ConstantPool);
3323 Constant *B = LookupConstant(I->getOperand(1), ConstantPool);
3326 return ConstantExpr::getCompare(Cmp->getPredicate(), A, B);
3329 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3330 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3333 if (A->isAllOnesValue())
3334 return LookupConstant(Select->getTrueValue(), ConstantPool);
3335 if (A->isNullValue())
3336 return LookupConstant(Select->getFalseValue(), ConstantPool);
3340 if (CastInst *Cast = dyn_cast<CastInst>(I)) {
3341 Constant *A = LookupConstant(I->getOperand(0), ConstantPool);
3344 return ConstantExpr::getCast(Cast->getOpcode(), A, Cast->getDestTy());
3350 /// GetCaseResults - Try to determine the resulting constant values in phi nodes
3351 /// at the common destination basic block, *CommonDest, for one of the case
3352 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3353 /// case), of a switch instruction SI.
3355 GetCaseResults(SwitchInst *SI,
3356 ConstantInt *CaseVal,
3357 BasicBlock *CaseDest,
3358 BasicBlock **CommonDest,
3359 SmallVectorImpl<std::pair<PHINode*,Constant*> > &Res) {
3360 // The block from which we enter the common destination.
3361 BasicBlock *Pred = SI->getParent();
3363 // If CaseDest is empty except for some side-effect free instructions through
3364 // which we can constant-propagate the CaseVal, continue to its successor.
3365 SmallDenseMap<Value*, Constant*> ConstantPool;
3366 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3367 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3369 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3370 // If the terminator is a simple branch, continue to the next block.
3371 if (T->getNumSuccessors() != 1)
3374 CaseDest = T->getSuccessor(0);
3375 } else if (isa<DbgInfoIntrinsic>(I)) {
3376 // Skip debug intrinsic.
3378 } else if (Constant *C = ConstantFold(I, ConstantPool)) {
3379 // Instruction is side-effect free and constant.
3380 ConstantPool.insert(std::make_pair(I, C));
3386 // If we did not have a CommonDest before, use the current one.
3388 *CommonDest = CaseDest;
3389 // If the destination isn't the common one, abort.
3390 if (CaseDest != *CommonDest)
3393 // Get the values for this case from phi nodes in the destination block.
3394 BasicBlock::iterator I = (*CommonDest)->begin();
3395 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3396 int Idx = PHI->getBasicBlockIndex(Pred);
3400 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3405 // Note: If the constant comes from constant-propagating the case value
3406 // through the CaseDest basic block, it will be safe to remove the
3407 // instructions in that block. They cannot be used (except in the phi nodes
3408 // we visit) outside CaseDest, because that block does not dominate its
3409 // successor. If it did, we would not be in this phi node.
3411 // Be conservative about which kinds of constants we support.
3412 if (!ValidLookupTableConstant(ConstVal))
3415 Res.push_back(std::make_pair(PHI, ConstVal));
3422 /// SwitchLookupTable - This class represents a lookup table that can be used
3423 /// to replace a switch.
3424 class SwitchLookupTable {
3426 /// SwitchLookupTable - Create a lookup table to use as a switch replacement
3427 /// with the contents of Values, using DefaultValue to fill any holes in the
3429 SwitchLookupTable(Module &M,
3431 ConstantInt *Offset,
3432 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3433 Constant *DefaultValue,
3434 const DataLayout *TD);
3436 /// BuildLookup - Build instructions with Builder to retrieve the value at
3437 /// the position given by Index in the lookup table.
3438 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3440 /// WouldFitInRegister - Return true if a table with TableSize elements of
3441 /// type ElementType would fit in a target-legal register.
3442 static bool WouldFitInRegister(const DataLayout *TD,
3444 const Type *ElementType);
3447 // Depending on the contents of the table, it can be represented in
3450 // For tables where each element contains the same value, we just have to
3451 // store that single value and return it for each lookup.
3454 // For small tables with integer elements, we can pack them into a bitmap
3455 // that fits into a target-legal register. Values are retrieved by
3456 // shift and mask operations.
3459 // The table is stored as an array of values. Values are retrieved by load
3460 // instructions from the table.
3464 // For SingleValueKind, this is the single value.
3465 Constant *SingleValue;
3467 // For BitMapKind, this is the bitmap.
3468 ConstantInt *BitMap;
3469 IntegerType *BitMapElementTy;
3471 // For ArrayKind, this is the array.
3472 GlobalVariable *Array;
3476 SwitchLookupTable::SwitchLookupTable(Module &M,
3478 ConstantInt *Offset,
3479 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3480 Constant *DefaultValue,
3481 const DataLayout *TD)
3482 : SingleValue(0), BitMap(0), BitMapElementTy(0), Array(0) {
3483 assert(Values.size() && "Can't build lookup table without values!");
3484 assert(TableSize >= Values.size() && "Can't fit values in table!");
3486 // If all values in the table are equal, this is that value.
3487 SingleValue = Values.begin()->second;
3489 // Build up the table contents.
3490 SmallVector<Constant*, 64> TableContents(TableSize);
3491 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3492 ConstantInt *CaseVal = Values[I].first;
3493 Constant *CaseRes = Values[I].second;
3494 assert(CaseRes->getType() == DefaultValue->getType());
3496 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3498 TableContents[Idx] = CaseRes;
3500 if (CaseRes != SingleValue)
3504 // Fill in any holes in the table with the default result.
3505 if (Values.size() < TableSize) {
3506 for (uint64_t I = 0; I < TableSize; ++I) {
3507 if (!TableContents[I])
3508 TableContents[I] = DefaultValue;
3511 if (DefaultValue != SingleValue)
3515 // If each element in the table contains the same value, we only need to store
3516 // that single value.
3518 Kind = SingleValueKind;
3522 // If the type is integer and the table fits in a register, build a bitmap.
3523 if (WouldFitInRegister(TD, TableSize, DefaultValue->getType())) {
3524 IntegerType *IT = cast<IntegerType>(DefaultValue->getType());
3525 APInt TableInt(TableSize * IT->getBitWidth(), 0);
3526 for (uint64_t I = TableSize; I > 0; --I) {
3527 TableInt <<= IT->getBitWidth();
3528 // Insert values into the bitmap. Undef values are set to zero.
3529 if (!isa<UndefValue>(TableContents[I - 1])) {
3530 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
3531 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
3534 BitMap = ConstantInt::get(M.getContext(), TableInt);
3535 BitMapElementTy = IT;
3541 // Store the table in an array.
3542 ArrayType *ArrayTy = ArrayType::get(DefaultValue->getType(), TableSize);
3543 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3545 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3546 GlobalVariable::PrivateLinkage,
3549 Array->setUnnamedAddr(true);
3553 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
3555 case SingleValueKind:
3558 // Type of the bitmap (e.g. i59).
3559 IntegerType *MapTy = BitMap->getType();
3561 // Cast Index to the same type as the bitmap.
3562 // Note: The Index is <= the number of elements in the table, so
3563 // truncating it to the width of the bitmask is safe.
3564 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
3566 // Multiply the shift amount by the element width.
3567 ShiftAmt = Builder.CreateMul(ShiftAmt,
3568 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
3572 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
3573 "switch.downshift");
3575 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
3579 Value *GEPIndices[] = { Builder.getInt32(0), Index };
3580 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices,
3582 return Builder.CreateLoad(GEP, "switch.load");
3585 llvm_unreachable("Unknown lookup table kind!");
3588 bool SwitchLookupTable::WouldFitInRegister(const DataLayout *TD,
3590 const Type *ElementType) {
3593 const IntegerType *IT = dyn_cast<IntegerType>(ElementType);
3596 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
3597 // are <= 15, we could try to narrow the type.
3599 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
3600 if (TableSize >= UINT_MAX/IT->getBitWidth())
3602 return TD->fitsInLegalInteger(TableSize * IT->getBitWidth());
3605 /// ShouldBuildLookupTable - Determine whether a lookup table should be built
3606 /// for this switch, based on the number of cases, size of the table and the
3607 /// types of the results.
3608 static bool ShouldBuildLookupTable(SwitchInst *SI,
3610 const TargetTransformInfo &TTI,
3611 const DataLayout *TD,
3612 const SmallDenseMap<PHINode*, Type*>& ResultTypes) {
3613 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
3614 return false; // TableSize overflowed, or mul below might overflow.
3616 bool AllTablesFitInRegister = true;
3617 bool HasIllegalType = false;
3618 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(),
3619 E = ResultTypes.end(); I != E; ++I) {
3620 Type *Ty = I->second;
3622 // Saturate this flag to true.
3623 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
3625 // Saturate this flag to false.
3626 AllTablesFitInRegister = AllTablesFitInRegister &&
3627 SwitchLookupTable::WouldFitInRegister(TD, TableSize, Ty);
3629 // If both flags saturate, we're done. NOTE: This *only* works with
3630 // saturating flags, and all flags have to saturate first due to the
3631 // non-deterministic behavior of iterating over a dense map.
3632 if (HasIllegalType && !AllTablesFitInRegister)
3636 // If each table would fit in a register, we should build it anyway.
3637 if (AllTablesFitInRegister)
3640 // Don't build a table that doesn't fit in-register if it has illegal types.
3644 // The table density should be at least 40%. This is the same criterion as for
3645 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3646 // FIXME: Find the best cut-off.
3647 return SI->getNumCases() * 10 >= TableSize * 4;
3650 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3651 /// phi nodes in a common successor block with different constant values,
3652 /// replace the switch with lookup tables.
3653 static bool SwitchToLookupTable(SwitchInst *SI,
3654 IRBuilder<> &Builder,
3655 const TargetTransformInfo &TTI,
3656 const DataLayout* TD) {
3657 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3659 // Only build lookup table when we have a target that supports it.
3660 if (!TTI.shouldBuildLookupTables())
3663 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3664 // split off a dense part and build a lookup table for that.
3666 // FIXME: This creates arrays of GEPs to constant strings, which means each
3667 // GEP needs a runtime relocation in PIC code. We should just build one big
3668 // string and lookup indices into that.
3670 // Ignore the switch if the number of cases is too small.
3671 // This is similar to the check when building jump tables in
3672 // SelectionDAGBuilder::handleJTSwitchCase.
3673 // FIXME: Determine the best cut-off.
3674 if (SI->getNumCases() < 4)
3677 // Figure out the corresponding result for each case value and phi node in the
3678 // common destination, as well as the the min and max case values.
3679 assert(SI->case_begin() != SI->case_end());
3680 SwitchInst::CaseIt CI = SI->case_begin();
3681 ConstantInt *MinCaseVal = CI.getCaseValue();
3682 ConstantInt *MaxCaseVal = CI.getCaseValue();
3684 BasicBlock *CommonDest = 0;
3685 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
3686 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
3687 SmallDenseMap<PHINode*, Constant*> DefaultResults;
3688 SmallDenseMap<PHINode*, Type*> ResultTypes;
3689 SmallVector<PHINode*, 4> PHIs;
3691 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
3692 ConstantInt *CaseVal = CI.getCaseValue();
3693 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
3694 MinCaseVal = CaseVal;
3695 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
3696 MaxCaseVal = CaseVal;
3698 // Resulting value at phi nodes for this case value.
3699 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
3701 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
3705 // Append the result from this case to the list for each phi.
3706 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) {
3707 if (!ResultLists.count(I->first))
3708 PHIs.push_back(I->first);
3709 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second));
3713 // Get the resulting values for the default case.
3714 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
3715 if (!GetCaseResults(SI, 0, SI->getDefaultDest(), &CommonDest,
3716 DefaultResultsList))
3718 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) {
3719 PHINode *PHI = DefaultResultsList[I].first;
3720 Constant *Result = DefaultResultsList[I].second;
3721 DefaultResults[PHI] = Result;
3722 ResultTypes[PHI] = Result->getType();
3725 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
3726 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
3727 if (!ShouldBuildLookupTable(SI, TableSize, TTI, TD, ResultTypes))
3730 // Create the BB that does the lookups.
3731 Module &Mod = *CommonDest->getParent()->getParent();
3732 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
3734 CommonDest->getParent(),
3737 // Compute the table index value.
3738 Builder.SetInsertPoint(SI);
3739 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
3742 // Compute the maximum table size representable by the integer type we are
3744 const unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
3745 const uint64_t MaxTableSize = CaseSize > 63? UINT64_MAX : 1ULL << CaseSize;
3746 assert(MaxTableSize >= TableSize &&
3747 "It is impossible for a switch to have more entries than the max "
3748 "representable value of its input integer type's size.");
3750 // If we have a covered lookup table, unconditionally branch to the lookup table
3751 // BB. Otherwise, check if the condition value is within the case range. If it
3752 // is so, branch to the new BB. Otherwise branch to SI's default destination.
3753 const bool GeneratingCoveredLookupTable = MaxTableSize == TableSize;
3754 if (GeneratingCoveredLookupTable) {
3755 Builder.CreateBr(LookupBB);
3757 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
3758 MinCaseVal->getType(), TableSize));
3759 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
3762 // Populate the BB that does the lookups.
3763 Builder.SetInsertPoint(LookupBB);
3764 bool ReturnedEarly = false;
3765 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
3766 PHINode *PHI = PHIs[I];
3768 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI],
3769 DefaultResults[PHI], TD);
3771 Value *Result = Table.BuildLookup(TableIndex, Builder);
3773 // If the result is used to return immediately from the function, we want to
3774 // do that right here.
3775 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin()) &&
3776 *PHI->use_begin() == CommonDest->getFirstNonPHIOrDbg()) {
3777 Builder.CreateRet(Result);
3778 ReturnedEarly = true;
3782 PHI->addIncoming(Result, LookupBB);
3786 Builder.CreateBr(CommonDest);
3788 // Remove the switch.
3789 for (unsigned i = 0; i < SI->getNumSuccessors(); ++i) {
3790 BasicBlock *Succ = SI->getSuccessor(i);
3792 // If we are not generating a covered lookup table, we will have a
3793 // conditional branch from SI's parent BB to SI's default destination if our
3794 // input value lies outside of our case range. Thus in that case leave the
3795 // default destination BB as a predecessor of SI's parent BB.
3796 if (Succ == SI->getDefaultDest() && !GeneratingCoveredLookupTable)
3798 Succ->removePredecessor(SI->getParent());
3800 SI->eraseFromParent();
3806 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
3807 BasicBlock *BB = SI->getParent();
3809 if (isValueEqualityComparison(SI)) {
3810 // If we only have one predecessor, and if it is a branch on this value,
3811 // see if that predecessor totally determines the outcome of this switch.
3812 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3813 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
3814 return SimplifyCFG(BB, TTI, TD) | true;
3816 Value *Cond = SI->getCondition();
3817 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
3818 if (SimplifySwitchOnSelect(SI, Select))
3819 return SimplifyCFG(BB, TTI, TD) | true;
3821 // If the block only contains the switch, see if we can fold the block
3822 // away into any preds.
3823 BasicBlock::iterator BBI = BB->begin();
3824 // Ignore dbg intrinsics.
3825 while (isa<DbgInfoIntrinsic>(BBI))
3828 if (FoldValueComparisonIntoPredecessors(SI, Builder))
3829 return SimplifyCFG(BB, TTI, TD) | true;
3832 // Try to transform the switch into an icmp and a branch.
3833 if (TurnSwitchRangeIntoICmp(SI, Builder))
3834 return SimplifyCFG(BB, TTI, TD) | true;
3836 // Remove unreachable cases.
3837 if (EliminateDeadSwitchCases(SI))
3838 return SimplifyCFG(BB, TTI, TD) | true;
3840 if (ForwardSwitchConditionToPHI(SI))
3841 return SimplifyCFG(BB, TTI, TD) | true;
3843 if (SwitchToLookupTable(SI, Builder, TTI, TD))
3844 return SimplifyCFG(BB, TTI, TD) | true;
3849 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
3850 BasicBlock *BB = IBI->getParent();
3851 bool Changed = false;
3853 // Eliminate redundant destinations.
3854 SmallPtrSet<Value *, 8> Succs;
3855 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
3856 BasicBlock *Dest = IBI->getDestination(i);
3857 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) {
3858 Dest->removePredecessor(BB);
3859 IBI->removeDestination(i);
3865 if (IBI->getNumDestinations() == 0) {
3866 // If the indirectbr has no successors, change it to unreachable.
3867 new UnreachableInst(IBI->getContext(), IBI);
3868 EraseTerminatorInstAndDCECond(IBI);
3872 if (IBI->getNumDestinations() == 1) {
3873 // If the indirectbr has one successor, change it to a direct branch.
3874 BranchInst::Create(IBI->getDestination(0), IBI);
3875 EraseTerminatorInstAndDCECond(IBI);
3879 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
3880 if (SimplifyIndirectBrOnSelect(IBI, SI))
3881 return SimplifyCFG(BB, TTI, TD) | true;
3886 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
3887 BasicBlock *BB = BI->getParent();
3889 if (SinkCommon && SinkThenElseCodeToEnd(BI))
3892 // If the Terminator is the only non-phi instruction, simplify the block.
3893 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime();
3894 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
3895 TryToSimplifyUncondBranchFromEmptyBlock(BB))
3898 // If the only instruction in the block is a seteq/setne comparison
3899 // against a constant, try to simplify the block.
3900 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
3901 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
3902 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
3904 if (I->isTerminator() &&
3905 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI, TD))
3909 // If this basic block is ONLY a compare and a branch, and if a predecessor
3910 // branches to us and our successor, fold the comparison into the
3911 // predecessor and use logical operations to update the incoming value
3912 // for PHI nodes in common successor.
3913 if (FoldBranchToCommonDest(BI))
3914 return SimplifyCFG(BB, TTI, TD) | true;
3919 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
3920 BasicBlock *BB = BI->getParent();
3922 // Conditional branch
3923 if (isValueEqualityComparison(BI)) {
3924 // If we only have one predecessor, and if it is a branch on this value,
3925 // see if that predecessor totally determines the outcome of this
3927 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3928 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
3929 return SimplifyCFG(BB, TTI, TD) | true;
3931 // This block must be empty, except for the setcond inst, if it exists.
3932 // Ignore dbg intrinsics.
3933 BasicBlock::iterator I = BB->begin();
3934 // Ignore dbg intrinsics.
3935 while (isa<DbgInfoIntrinsic>(I))
3938 if (FoldValueComparisonIntoPredecessors(BI, Builder))
3939 return SimplifyCFG(BB, TTI, TD) | true;
3940 } else if (&*I == cast<Instruction>(BI->getCondition())){
3942 // Ignore dbg intrinsics.
3943 while (isa<DbgInfoIntrinsic>(I))
3945 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
3946 return SimplifyCFG(BB, TTI, TD) | true;
3950 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
3951 if (SimplifyBranchOnICmpChain(BI, TD, Builder))
3954 // If this basic block is ONLY a compare and a branch, and if a predecessor
3955 // branches to us and one of our successors, fold the comparison into the
3956 // predecessor and use logical operations to pick the right destination.
3957 if (FoldBranchToCommonDest(BI))
3958 return SimplifyCFG(BB, TTI, TD) | true;
3960 // We have a conditional branch to two blocks that are only reachable
3961 // from BI. We know that the condbr dominates the two blocks, so see if
3962 // there is any identical code in the "then" and "else" blocks. If so, we
3963 // can hoist it up to the branching block.
3964 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) {
3965 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3966 if (HoistThenElseCodeToIf(BI))
3967 return SimplifyCFG(BB, TTI, TD) | true;
3969 // If Successor #1 has multiple preds, we may be able to conditionally
3970 // execute Successor #0 if it branches to successor #1.
3971 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
3972 if (Succ0TI->getNumSuccessors() == 1 &&
3973 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
3974 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0)))
3975 return SimplifyCFG(BB, TTI, TD) | true;
3977 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3978 // If Successor #0 has multiple preds, we may be able to conditionally
3979 // execute Successor #1 if it branches to successor #0.
3980 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
3981 if (Succ1TI->getNumSuccessors() == 1 &&
3982 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
3983 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1)))
3984 return SimplifyCFG(BB, TTI, TD) | true;
3987 // If this is a branch on a phi node in the current block, thread control
3988 // through this block if any PHI node entries are constants.
3989 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
3990 if (PN->getParent() == BI->getParent())
3991 if (FoldCondBranchOnPHI(BI, TD))
3992 return SimplifyCFG(BB, TTI, TD) | true;
3994 // Scan predecessor blocks for conditional branches.
3995 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
3996 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
3997 if (PBI != BI && PBI->isConditional())
3998 if (SimplifyCondBranchToCondBranch(PBI, BI))
3999 return SimplifyCFG(BB, TTI, TD) | true;
4004 /// Check if passing a value to an instruction will cause undefined behavior.
4005 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
4006 Constant *C = dyn_cast<Constant>(V);
4013 if (C->isNullValue()) {
4014 // Only look at the first use, avoid hurting compile time with long uselists
4015 User *Use = *I->use_begin();
4017 // Now make sure that there are no instructions in between that can alter
4018 // control flow (eg. calls)
4019 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
4020 if (i == I->getParent()->end() || i->mayHaveSideEffects())
4023 // Look through GEPs. A load from a GEP derived from NULL is still undefined
4024 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
4025 if (GEP->getPointerOperand() == I)
4026 return passingValueIsAlwaysUndefined(V, GEP);
4028 // Look through bitcasts.
4029 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
4030 return passingValueIsAlwaysUndefined(V, BC);
4032 // Load from null is undefined.
4033 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
4034 if (!LI->isVolatile())
4035 return LI->getPointerAddressSpace() == 0;
4037 // Store to null is undefined.
4038 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
4039 if (!SI->isVolatile())
4040 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
4045 /// If BB has an incoming value that will always trigger undefined behavior
4046 /// (eg. null pointer dereference), remove the branch leading here.
4047 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
4048 for (BasicBlock::iterator i = BB->begin();
4049 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
4050 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
4051 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
4052 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
4053 IRBuilder<> Builder(T);
4054 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
4055 BB->removePredecessor(PHI->getIncomingBlock(i));
4056 // Turn uncoditional branches into unreachables and remove the dead
4057 // destination from conditional branches.
4058 if (BI->isUnconditional())
4059 Builder.CreateUnreachable();
4061 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
4062 BI->getSuccessor(0));
4063 BI->eraseFromParent();
4066 // TODO: SwitchInst.
4072 bool SimplifyCFGOpt::run(BasicBlock *BB) {
4073 bool Changed = false;
4075 assert(BB && BB->getParent() && "Block not embedded in function!");
4076 assert(BB->getTerminator() && "Degenerate basic block encountered!");
4078 // Remove basic blocks that have no predecessors (except the entry block)...
4079 // or that just have themself as a predecessor. These are unreachable.
4080 if ((pred_begin(BB) == pred_end(BB) &&
4081 BB != &BB->getParent()->getEntryBlock()) ||
4082 BB->getSinglePredecessor() == BB) {
4083 DEBUG(dbgs() << "Removing BB: \n" << *BB);
4084 DeleteDeadBlock(BB);
4088 // Check to see if we can constant propagate this terminator instruction
4090 Changed |= ConstantFoldTerminator(BB, true);
4092 // Check for and eliminate duplicate PHI nodes in this block.
4093 Changed |= EliminateDuplicatePHINodes(BB);
4095 // Check for and remove branches that will always cause undefined behavior.
4096 Changed |= removeUndefIntroducingPredecessor(BB);
4098 // Merge basic blocks into their predecessor if there is only one distinct
4099 // pred, and if there is only one distinct successor of the predecessor, and
4100 // if there are no PHI nodes.
4102 if (MergeBlockIntoPredecessor(BB))
4105 IRBuilder<> Builder(BB);
4107 // If there is a trivial two-entry PHI node in this basic block, and we can
4108 // eliminate it, do so now.
4109 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
4110 if (PN->getNumIncomingValues() == 2)
4111 Changed |= FoldTwoEntryPHINode(PN, TD);
4113 Builder.SetInsertPoint(BB->getTerminator());
4114 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
4115 if (BI->isUnconditional()) {
4116 if (SimplifyUncondBranch(BI, Builder)) return true;
4118 if (SimplifyCondBranch(BI, Builder)) return true;
4120 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
4121 if (SimplifyReturn(RI, Builder)) return true;
4122 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
4123 if (SimplifyResume(RI, Builder)) return true;
4124 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
4125 if (SimplifySwitch(SI, Builder)) return true;
4126 } else if (UnreachableInst *UI =
4127 dyn_cast<UnreachableInst>(BB->getTerminator())) {
4128 if (SimplifyUnreachable(UI)) return true;
4129 } else if (IndirectBrInst *IBI =
4130 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
4131 if (SimplifyIndirectBr(IBI)) return true;
4137 /// SimplifyCFG - This function is used to do simplification of a CFG. For
4138 /// example, it adjusts branches to branches to eliminate the extra hop, it
4139 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4140 /// of the CFG. It returns true if a modification was made.
4142 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
4143 const DataLayout *TD) {
4144 return SimplifyCFGOpt(TTI, TD).run(BB);