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/Constants.h"
17 #include "llvm/DerivedTypes.h"
18 #include "llvm/GlobalVariable.h"
19 #include "llvm/IRBuilder.h"
20 #include "llvm/Instructions.h"
21 #include "llvm/IntrinsicInst.h"
22 #include "llvm/LLVMContext.h"
23 #include "llvm/MDBuilder.h"
24 #include "llvm/Metadata.h"
25 #include "llvm/Module.h"
26 #include "llvm/Operator.h"
27 #include "llvm/Type.h"
28 #include "llvm/ADT/DenseMap.h"
29 #include "llvm/ADT/STLExtras.h"
30 #include "llvm/ADT/SetVector.h"
31 #include "llvm/ADT/SmallPtrSet.h"
32 #include "llvm/ADT/SmallVector.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/Analysis/InstructionSimplify.h"
35 #include "llvm/Analysis/ValueTracking.h"
36 #include "llvm/Support/CFG.h"
37 #include "llvm/Support/CommandLine.h"
38 #include "llvm/Support/ConstantRange.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/NoFolder.h"
41 #include "llvm/Support/raw_ostream.h"
42 #include "llvm/DataLayout.h"
43 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
49 static cl::opt<unsigned>
50 PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1),
51 cl::desc("Control the amount of phi node folding to perform (default = 1)"));
54 DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false),
55 cl::desc("Duplicate return instructions into unconditional branches"));
58 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
59 cl::desc("Sink common instructions down to the end block"));
61 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
62 STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables");
63 STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block");
64 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
67 /// ValueEqualityComparisonCase - Represents a case of a switch.
68 struct ValueEqualityComparisonCase {
72 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
73 : Value(Value), Dest(Dest) {}
75 bool operator<(ValueEqualityComparisonCase RHS) const {
76 // Comparing pointers is ok as we only rely on the order for uniquing.
77 return Value < RHS.Value;
80 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
83 class SimplifyCFGOpt {
84 const DataLayout *const TD;
86 Value *isValueEqualityComparison(TerminatorInst *TI);
87 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
88 std::vector<ValueEqualityComparisonCase> &Cases);
89 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
91 IRBuilder<> &Builder);
92 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
93 IRBuilder<> &Builder);
95 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
96 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
97 bool SimplifyUnreachable(UnreachableInst *UI);
98 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
99 bool SimplifyIndirectBr(IndirectBrInst *IBI);
100 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder);
101 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder);
104 explicit SimplifyCFGOpt(const DataLayout *td) : TD(td) {}
105 bool run(BasicBlock *BB);
109 /// SafeToMergeTerminators - Return true if it is safe to merge these two
110 /// terminator instructions together.
112 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
113 if (SI1 == SI2) return false; // Can't merge with self!
115 // It is not safe to merge these two switch instructions if they have a common
116 // successor, and if that successor has a PHI node, and if *that* PHI node has
117 // conflicting incoming values from the two switch blocks.
118 BasicBlock *SI1BB = SI1->getParent();
119 BasicBlock *SI2BB = SI2->getParent();
120 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
122 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
123 if (SI1Succs.count(*I))
124 for (BasicBlock::iterator BBI = (*I)->begin();
125 isa<PHINode>(BBI); ++BBI) {
126 PHINode *PN = cast<PHINode>(BBI);
127 if (PN->getIncomingValueForBlock(SI1BB) !=
128 PN->getIncomingValueForBlock(SI2BB))
135 /// isProfitableToFoldUnconditional - Return true if it is safe and profitable
136 /// to merge these two terminator instructions together, where SI1 is an
137 /// unconditional branch. PhiNodes will store all PHI nodes in common
140 static bool isProfitableToFoldUnconditional(BranchInst *SI1,
143 SmallVectorImpl<PHINode*> &PhiNodes) {
144 if (SI1 == SI2) return false; // Can't merge with self!
145 assert(SI1->isUnconditional() && SI2->isConditional());
147 // We fold the unconditional branch if we can easily update all PHI nodes in
148 // common successors:
149 // 1> We have a constant incoming value for the conditional branch;
150 // 2> We have "Cond" as the incoming value for the unconditional branch;
151 // 3> SI2->getCondition() and Cond have same operands.
152 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
153 if (!Ci2) return false;
154 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
155 Cond->getOperand(1) == Ci2->getOperand(1)) &&
156 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
157 Cond->getOperand(1) == Ci2->getOperand(0)))
160 BasicBlock *SI1BB = SI1->getParent();
161 BasicBlock *SI2BB = SI2->getParent();
162 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
163 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
164 if (SI1Succs.count(*I))
165 for (BasicBlock::iterator BBI = (*I)->begin();
166 isa<PHINode>(BBI); ++BBI) {
167 PHINode *PN = cast<PHINode>(BBI);
168 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
169 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
171 PhiNodes.push_back(PN);
176 /// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will
177 /// now be entries in it from the 'NewPred' block. The values that will be
178 /// flowing into the PHI nodes will be the same as those coming in from
179 /// ExistPred, an existing predecessor of Succ.
180 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
181 BasicBlock *ExistPred) {
182 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
185 for (BasicBlock::iterator I = Succ->begin();
186 (PN = dyn_cast<PHINode>(I)); ++I)
187 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
191 /// GetIfCondition - Given a basic block (BB) with two predecessors (and at
192 /// least one PHI node in it), check to see if the merge at this block is due
193 /// to an "if condition". If so, return the boolean condition that determines
194 /// which entry into BB will be taken. Also, return by references the block
195 /// that will be entered from if the condition is true, and the block that will
196 /// be entered if the condition is false.
198 /// This does no checking to see if the true/false blocks have large or unsavory
199 /// instructions in them.
200 static Value *GetIfCondition(BasicBlock *BB, BasicBlock *&IfTrue,
201 BasicBlock *&IfFalse) {
202 PHINode *SomePHI = cast<PHINode>(BB->begin());
203 assert(SomePHI->getNumIncomingValues() == 2 &&
204 "Function can only handle blocks with 2 predecessors!");
205 BasicBlock *Pred1 = SomePHI->getIncomingBlock(0);
206 BasicBlock *Pred2 = SomePHI->getIncomingBlock(1);
208 // We can only handle branches. Other control flow will be lowered to
209 // branches if possible anyway.
210 BranchInst *Pred1Br = dyn_cast<BranchInst>(Pred1->getTerminator());
211 BranchInst *Pred2Br = dyn_cast<BranchInst>(Pred2->getTerminator());
212 if (Pred1Br == 0 || Pred2Br == 0)
215 // Eliminate code duplication by ensuring that Pred1Br is conditional if
217 if (Pred2Br->isConditional()) {
218 // If both branches are conditional, we don't have an "if statement". In
219 // reality, we could transform this case, but since the condition will be
220 // required anyway, we stand no chance of eliminating it, so the xform is
221 // probably not profitable.
222 if (Pred1Br->isConditional())
225 std::swap(Pred1, Pred2);
226 std::swap(Pred1Br, Pred2Br);
229 if (Pred1Br->isConditional()) {
230 // The only thing we have to watch out for here is to make sure that Pred2
231 // doesn't have incoming edges from other blocks. If it does, the condition
232 // doesn't dominate BB.
233 if (Pred2->getSinglePredecessor() == 0)
236 // If we found a conditional branch predecessor, make sure that it branches
237 // to BB and Pred2Br. If it doesn't, this isn't an "if statement".
238 if (Pred1Br->getSuccessor(0) == BB &&
239 Pred1Br->getSuccessor(1) == Pred2) {
242 } else if (Pred1Br->getSuccessor(0) == Pred2 &&
243 Pred1Br->getSuccessor(1) == BB) {
247 // We know that one arm of the conditional goes to BB, so the other must
248 // go somewhere unrelated, and this must not be an "if statement".
252 return Pred1Br->getCondition();
255 // Ok, if we got here, both predecessors end with an unconditional branch to
256 // BB. Don't panic! If both blocks only have a single (identical)
257 // predecessor, and THAT is a conditional branch, then we're all ok!
258 BasicBlock *CommonPred = Pred1->getSinglePredecessor();
259 if (CommonPred == 0 || CommonPred != Pred2->getSinglePredecessor())
262 // Otherwise, if this is a conditional branch, then we can use it!
263 BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator());
264 if (BI == 0) return 0;
266 assert(BI->isConditional() && "Two successors but not conditional?");
267 if (BI->getSuccessor(0) == Pred1) {
274 return BI->getCondition();
277 /// ComputeSpeculuationCost - Compute an abstract "cost" of speculating the
278 /// given instruction, which is assumed to be safe to speculate. 1 means
279 /// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive.
280 static unsigned ComputeSpeculationCost(const User *I) {
281 assert(isSafeToSpeculativelyExecute(I) &&
282 "Instruction is not safe to speculatively execute!");
283 switch (Operator::getOpcode(I)) {
285 // In doubt, be conservative.
287 case Instruction::GetElementPtr:
288 // GEPs are cheap if all indices are constant.
289 if (!cast<GEPOperator>(I)->hasAllConstantIndices())
292 case Instruction::Load:
293 case Instruction::Add:
294 case Instruction::Sub:
295 case Instruction::And:
296 case Instruction::Or:
297 case Instruction::Xor:
298 case Instruction::Shl:
299 case Instruction::LShr:
300 case Instruction::AShr:
301 case Instruction::ICmp:
302 case Instruction::Trunc:
303 case Instruction::ZExt:
304 case Instruction::SExt:
305 return 1; // These are all cheap.
307 case Instruction::Call:
308 case Instruction::Select:
313 /// DominatesMergePoint - If we have a merge point of an "if condition" as
314 /// accepted above, return true if the specified value dominates the block. We
315 /// don't handle the true generality of domination here, just a special case
316 /// which works well enough for us.
318 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
319 /// see if V (which must be an instruction) and its recursive operands
320 /// that do not dominate BB have a combined cost lower than CostRemaining and
321 /// are non-trapping. If both are true, the instruction is inserted into the
322 /// set and true is returned.
324 /// The cost for most non-trapping instructions is defined as 1 except for
325 /// Select whose cost is 2.
327 /// After this function returns, CostRemaining is decreased by the cost of
328 /// V plus its non-dominating operands. If that cost is greater than
329 /// CostRemaining, false is returned and CostRemaining is undefined.
330 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
331 SmallPtrSet<Instruction*, 4> *AggressiveInsts,
332 unsigned &CostRemaining) {
333 Instruction *I = dyn_cast<Instruction>(V);
335 // Non-instructions all dominate instructions, but not all constantexprs
336 // can be executed unconditionally.
337 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
342 BasicBlock *PBB = I->getParent();
344 // We don't want to allow weird loops that might have the "if condition" in
345 // the bottom of this block.
346 if (PBB == BB) return false;
348 // If this instruction is defined in a block that contains an unconditional
349 // branch to BB, then it must be in the 'conditional' part of the "if
350 // statement". If not, it definitely dominates the region.
351 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
352 if (BI == 0 || BI->isConditional() || BI->getSuccessor(0) != BB)
355 // If we aren't allowing aggressive promotion anymore, then don't consider
356 // instructions in the 'if region'.
357 if (AggressiveInsts == 0) return false;
359 // If we have seen this instruction before, don't count it again.
360 if (AggressiveInsts->count(I)) return true;
362 // Okay, it looks like the instruction IS in the "condition". Check to
363 // see if it's a cheap instruction to unconditionally compute, and if it
364 // only uses stuff defined outside of the condition. If so, hoist it out.
365 if (!isSafeToSpeculativelyExecute(I))
368 unsigned Cost = ComputeSpeculationCost(I);
370 if (Cost > CostRemaining)
373 CostRemaining -= Cost;
375 // Okay, we can only really hoist these out if their operands do
376 // not take us over the cost threshold.
377 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
378 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining))
380 // Okay, it's safe to do this! Remember this instruction.
381 AggressiveInsts->insert(I);
385 /// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr
386 /// and PointerNullValue. Return NULL if value is not a constant int.
387 static ConstantInt *GetConstantInt(Value *V, const DataLayout *TD) {
388 // Normal constant int.
389 ConstantInt *CI = dyn_cast<ConstantInt>(V);
390 if (CI || !TD || !isa<Constant>(V) || !V->getType()->isPointerTy())
393 // This is some kind of pointer constant. Turn it into a pointer-sized
394 // ConstantInt if possible.
395 IntegerType *PtrTy = TD->getIntPtrType(V->getContext());
397 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
398 if (isa<ConstantPointerNull>(V))
399 return ConstantInt::get(PtrTy, 0);
401 // IntToPtr const int.
402 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
403 if (CE->getOpcode() == Instruction::IntToPtr)
404 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
405 // The constant is very likely to have the right type already.
406 if (CI->getType() == PtrTy)
409 return cast<ConstantInt>
410 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
415 /// GatherConstantCompares - Given a potentially 'or'd or 'and'd together
416 /// collection of icmp eq/ne instructions that compare a value against a
417 /// constant, return the value being compared, and stick the constant into the
420 GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra,
421 const DataLayout *TD, bool isEQ, unsigned &UsedICmps) {
422 Instruction *I = dyn_cast<Instruction>(V);
423 if (I == 0) return 0;
425 // If this is an icmp against a constant, handle this as one of the cases.
426 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) {
427 if (ConstantInt *C = GetConstantInt(I->getOperand(1), TD)) {
428 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
431 return I->getOperand(0);
434 // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to
437 ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue());
439 // If this is an and/!= check then we want to optimize "x ugt 2" into
442 Span = Span.inverse();
444 // If there are a ton of values, we don't want to make a ginormous switch.
445 if (Span.getSetSize().ugt(8) || Span.isEmptySet())
448 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
449 Vals.push_back(ConstantInt::get(V->getContext(), Tmp));
451 return I->getOperand(0);
456 // Otherwise, we can only handle an | or &, depending on isEQ.
457 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And))
460 unsigned NumValsBeforeLHS = Vals.size();
461 unsigned UsedICmpsBeforeLHS = UsedICmps;
462 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, TD,
464 unsigned NumVals = Vals.size();
465 unsigned UsedICmpsBeforeRHS = UsedICmps;
466 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD,
470 Vals.resize(NumVals);
471 UsedICmps = UsedICmpsBeforeRHS;
474 // The RHS of the or/and can't be folded in and we haven't used "Extra" yet,
475 // set it and return success.
476 if (Extra == 0 || Extra == I->getOperand(1)) {
477 Extra = I->getOperand(1);
481 Vals.resize(NumValsBeforeLHS);
482 UsedICmps = UsedICmpsBeforeLHS;
486 // If the LHS can't be folded in, but Extra is available and RHS can, try to
488 if (Extra == 0 || Extra == I->getOperand(0)) {
489 Value *OldExtra = Extra;
490 Extra = I->getOperand(0);
491 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD,
494 assert(Vals.size() == NumValsBeforeLHS);
501 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
502 Instruction *Cond = 0;
503 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
504 Cond = dyn_cast<Instruction>(SI->getCondition());
505 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
506 if (BI->isConditional())
507 Cond = dyn_cast<Instruction>(BI->getCondition());
508 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
509 Cond = dyn_cast<Instruction>(IBI->getAddress());
512 TI->eraseFromParent();
513 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
516 /// isValueEqualityComparison - Return true if the specified terminator checks
517 /// to see if a value is equal to constant integer value.
518 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
520 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
521 // Do not permit merging of large switch instructions into their
522 // predecessors unless there is only one predecessor.
523 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()),
524 pred_end(SI->getParent())) <= 128)
525 CV = SI->getCondition();
526 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
527 if (BI->isConditional() && BI->getCondition()->hasOneUse())
528 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition()))
529 if ((ICI->getPredicate() == ICmpInst::ICMP_EQ ||
530 ICI->getPredicate() == ICmpInst::ICMP_NE) &&
531 GetConstantInt(ICI->getOperand(1), TD))
532 CV = ICI->getOperand(0);
534 // Unwrap any lossless ptrtoint cast.
535 if (TD && CV && CV->getType() == TD->getIntPtrType(CV->getContext()))
536 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV))
537 CV = PTII->getOperand(0);
541 /// GetValueEqualityComparisonCases - Given a value comparison instruction,
542 /// decode all of the 'cases' that it represents and return the 'default' block.
543 BasicBlock *SimplifyCFGOpt::
544 GetValueEqualityComparisonCases(TerminatorInst *TI,
545 std::vector<ValueEqualityComparisonCase>
547 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
548 Cases.reserve(SI->getNumCases());
549 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
550 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
551 i.getCaseSuccessor()));
552 return SI->getDefaultDest();
555 BranchInst *BI = cast<BranchInst>(TI);
556 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
557 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
558 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
561 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
565 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries
566 /// in the list that match the specified block.
567 static void EliminateBlockCases(BasicBlock *BB,
568 std::vector<ValueEqualityComparisonCase> &Cases) {
569 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
572 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as
575 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
576 std::vector<ValueEqualityComparisonCase > &C2) {
577 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
579 // Make V1 be smaller than V2.
580 if (V1->size() > V2->size())
583 if (V1->size() == 0) return false;
584 if (V1->size() == 1) {
586 ConstantInt *TheVal = (*V1)[0].Value;
587 for (unsigned i = 0, e = V2->size(); i != e; ++i)
588 if (TheVal == (*V2)[i].Value)
592 // Otherwise, just sort both lists and compare element by element.
593 array_pod_sort(V1->begin(), V1->end());
594 array_pod_sort(V2->begin(), V2->end());
595 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
596 while (i1 != e1 && i2 != e2) {
597 if ((*V1)[i1].Value == (*V2)[i2].Value)
599 if ((*V1)[i1].Value < (*V2)[i2].Value)
607 /// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a
608 /// terminator instruction and its block is known to only have a single
609 /// predecessor block, check to see if that predecessor is also a value
610 /// comparison with the same value, and if that comparison determines the
611 /// outcome of this comparison. If so, simplify TI. This does a very limited
612 /// form of jump threading.
613 bool SimplifyCFGOpt::
614 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
616 IRBuilder<> &Builder) {
617 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
618 if (!PredVal) return false; // Not a value comparison in predecessor.
620 Value *ThisVal = isValueEqualityComparison(TI);
621 assert(ThisVal && "This isn't a value comparison!!");
622 if (ThisVal != PredVal) return false; // Different predicates.
624 // TODO: Preserve branch weight metadata, similarly to how
625 // FoldValueComparisonIntoPredecessors preserves it.
627 // Find out information about when control will move from Pred to TI's block.
628 std::vector<ValueEqualityComparisonCase> PredCases;
629 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
631 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
633 // Find information about how control leaves this block.
634 std::vector<ValueEqualityComparisonCase> ThisCases;
635 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
636 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
638 // If TI's block is the default block from Pred's comparison, potentially
639 // simplify TI based on this knowledge.
640 if (PredDef == TI->getParent()) {
641 // If we are here, we know that the value is none of those cases listed in
642 // PredCases. If there are any cases in ThisCases that are in PredCases, we
644 if (!ValuesOverlap(PredCases, ThisCases))
647 if (isa<BranchInst>(TI)) {
648 // Okay, one of the successors of this condbr is dead. Convert it to a
650 assert(ThisCases.size() == 1 && "Branch can only have one case!");
651 // Insert the new branch.
652 Instruction *NI = Builder.CreateBr(ThisDef);
655 // Remove PHI node entries for the dead edge.
656 ThisCases[0].Dest->removePredecessor(TI->getParent());
658 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
659 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
661 EraseTerminatorInstAndDCECond(TI);
665 SwitchInst *SI = cast<SwitchInst>(TI);
666 // Okay, TI has cases that are statically dead, prune them away.
667 SmallPtrSet<Constant*, 16> DeadCases;
668 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
669 DeadCases.insert(PredCases[i].Value);
671 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
672 << "Through successor TI: " << *TI);
674 // Collect branch weights into a vector.
675 SmallVector<uint32_t, 8> Weights;
676 MDNode* MD = SI->getMetadata(LLVMContext::MD_prof);
677 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
679 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
681 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i));
683 Weights.push_back(CI->getValue().getZExtValue());
685 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
687 if (DeadCases.count(i.getCaseValue())) {
689 std::swap(Weights[i.getCaseIndex()+1], Weights.back());
692 i.getCaseSuccessor()->removePredecessor(TI->getParent());
696 if (HasWeight && Weights.size() >= 2)
697 SI->setMetadata(LLVMContext::MD_prof,
698 MDBuilder(SI->getParent()->getContext()).
699 createBranchWeights(Weights));
701 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
705 // Otherwise, TI's block must correspond to some matched value. Find out
706 // which value (or set of values) this is.
707 ConstantInt *TIV = 0;
708 BasicBlock *TIBB = TI->getParent();
709 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
710 if (PredCases[i].Dest == TIBB) {
712 return false; // Cannot handle multiple values coming to this block.
713 TIV = PredCases[i].Value;
715 assert(TIV && "No edge from pred to succ?");
717 // Okay, we found the one constant that our value can be if we get into TI's
718 // BB. Find out which successor will unconditionally be branched to.
719 BasicBlock *TheRealDest = 0;
720 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
721 if (ThisCases[i].Value == TIV) {
722 TheRealDest = ThisCases[i].Dest;
726 // If not handled by any explicit cases, it is handled by the default case.
727 if (TheRealDest == 0) TheRealDest = ThisDef;
729 // Remove PHI node entries for dead edges.
730 BasicBlock *CheckEdge = TheRealDest;
731 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
732 if (*SI != CheckEdge)
733 (*SI)->removePredecessor(TIBB);
737 // Insert the new branch.
738 Instruction *NI = Builder.CreateBr(TheRealDest);
741 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
742 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
744 EraseTerminatorInstAndDCECond(TI);
749 /// ConstantIntOrdering - This class implements a stable ordering of constant
750 /// integers that does not depend on their address. This is important for
751 /// applications that sort ConstantInt's to ensure uniqueness.
752 struct ConstantIntOrdering {
753 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
754 return LHS->getValue().ult(RHS->getValue());
759 static int ConstantIntSortPredicate(const void *P1, const void *P2) {
760 const ConstantInt *LHS = *(const ConstantInt*const*)P1;
761 const ConstantInt *RHS = *(const ConstantInt*const*)P2;
762 if (LHS->getValue().ult(RHS->getValue()))
764 if (LHS->getValue() == RHS->getValue())
769 static inline bool HasBranchWeights(const Instruction* I) {
770 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof);
771 if (ProfMD && ProfMD->getOperand(0))
772 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
773 return MDS->getString().equals("branch_weights");
778 /// Get Weights of a given TerminatorInst, the default weight is at the front
779 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
781 static void GetBranchWeights(TerminatorInst *TI,
782 SmallVectorImpl<uint64_t> &Weights) {
783 MDNode* MD = TI->getMetadata(LLVMContext::MD_prof);
785 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
786 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(i));
788 Weights.push_back(CI->getValue().getZExtValue());
791 // If TI is a conditional eq, the default case is the false case,
792 // and the corresponding branch-weight data is at index 2. We swap the
793 // default weight to be the first entry.
794 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
795 assert(Weights.size() == 2);
796 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
797 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
798 std::swap(Weights.front(), Weights.back());
802 /// Sees if any of the weights are too big for a uint32_t, and halves all the
803 /// weights if any are.
804 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
806 for (unsigned i = 0; i < Weights.size(); ++i)
807 if (Weights[i] > UINT_MAX) {
815 for (unsigned i = 0; i < Weights.size(); ++i)
819 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value
820 /// equality comparison instruction (either a switch or a branch on "X == c").
821 /// See if any of the predecessors of the terminator block are value comparisons
822 /// on the same value. If so, and if safe to do so, fold them together.
823 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
824 IRBuilder<> &Builder) {
825 BasicBlock *BB = TI->getParent();
826 Value *CV = isValueEqualityComparison(TI); // CondVal
827 assert(CV && "Not a comparison?");
828 bool Changed = false;
830 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
831 while (!Preds.empty()) {
832 BasicBlock *Pred = Preds.pop_back_val();
834 // See if the predecessor is a comparison with the same value.
835 TerminatorInst *PTI = Pred->getTerminator();
836 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
838 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
839 // Figure out which 'cases' to copy from SI to PSI.
840 std::vector<ValueEqualityComparisonCase> BBCases;
841 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
843 std::vector<ValueEqualityComparisonCase> PredCases;
844 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
846 // Based on whether the default edge from PTI goes to BB or not, fill in
847 // PredCases and PredDefault with the new switch cases we would like to
849 SmallVector<BasicBlock*, 8> NewSuccessors;
851 // Update the branch weight metadata along the way
852 SmallVector<uint64_t, 8> Weights;
853 bool PredHasWeights = HasBranchWeights(PTI);
854 bool SuccHasWeights = HasBranchWeights(TI);
856 if (PredHasWeights) {
857 GetBranchWeights(PTI, Weights);
858 // branch-weight metadata is inconsistant here.
859 if (Weights.size() != 1 + PredCases.size())
860 PredHasWeights = SuccHasWeights = false;
861 } else if (SuccHasWeights)
862 // If there are no predecessor weights but there are successor weights,
863 // populate Weights with 1, which will later be scaled to the sum of
864 // successor's weights
865 Weights.assign(1 + PredCases.size(), 1);
867 SmallVector<uint64_t, 8> SuccWeights;
868 if (SuccHasWeights) {
869 GetBranchWeights(TI, SuccWeights);
870 // branch-weight metadata is inconsistant here.
871 if (SuccWeights.size() != 1 + BBCases.size())
872 PredHasWeights = SuccHasWeights = false;
873 } else if (PredHasWeights)
874 SuccWeights.assign(1 + BBCases.size(), 1);
876 if (PredDefault == BB) {
877 // If this is the default destination from PTI, only the edges in TI
878 // that don't occur in PTI, or that branch to BB will be activated.
879 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
880 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
881 if (PredCases[i].Dest != BB)
882 PTIHandled.insert(PredCases[i].Value);
884 // The default destination is BB, we don't need explicit targets.
885 std::swap(PredCases[i], PredCases.back());
887 if (PredHasWeights || SuccHasWeights) {
888 // Increase weight for the default case.
889 Weights[0] += Weights[i+1];
890 std::swap(Weights[i+1], Weights.back());
894 PredCases.pop_back();
898 // Reconstruct the new switch statement we will be building.
899 if (PredDefault != BBDefault) {
900 PredDefault->removePredecessor(Pred);
901 PredDefault = BBDefault;
902 NewSuccessors.push_back(BBDefault);
905 unsigned CasesFromPred = Weights.size();
906 uint64_t ValidTotalSuccWeight = 0;
907 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
908 if (!PTIHandled.count(BBCases[i].Value) &&
909 BBCases[i].Dest != BBDefault) {
910 PredCases.push_back(BBCases[i]);
911 NewSuccessors.push_back(BBCases[i].Dest);
912 if (SuccHasWeights || PredHasWeights) {
913 // The default weight is at index 0, so weight for the ith case
914 // should be at index i+1. Scale the cases from successor by
915 // PredDefaultWeight (Weights[0]).
916 Weights.push_back(Weights[0] * SuccWeights[i+1]);
917 ValidTotalSuccWeight += SuccWeights[i+1];
921 if (SuccHasWeights || PredHasWeights) {
922 ValidTotalSuccWeight += SuccWeights[0];
923 // Scale the cases from predecessor by ValidTotalSuccWeight.
924 for (unsigned i = 1; i < CasesFromPred; ++i)
925 Weights[i] *= ValidTotalSuccWeight;
926 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
927 Weights[0] *= SuccWeights[0];
930 // If this is not the default destination from PSI, only the edges
931 // in SI that occur in PSI with a destination of BB will be
933 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
934 std::map<ConstantInt*, uint64_t> WeightsForHandled;
935 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
936 if (PredCases[i].Dest == BB) {
937 PTIHandled.insert(PredCases[i].Value);
939 if (PredHasWeights || SuccHasWeights) {
940 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
941 std::swap(Weights[i+1], Weights.back());
945 std::swap(PredCases[i], PredCases.back());
946 PredCases.pop_back();
950 // Okay, now we know which constants were sent to BB from the
951 // predecessor. Figure out where they will all go now.
952 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
953 if (PTIHandled.count(BBCases[i].Value)) {
954 // If this is one we are capable of getting...
955 if (PredHasWeights || SuccHasWeights)
956 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
957 PredCases.push_back(BBCases[i]);
958 NewSuccessors.push_back(BBCases[i].Dest);
959 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
962 // If there are any constants vectored to BB that TI doesn't handle,
963 // they must go to the default destination of TI.
964 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
966 E = PTIHandled.end(); I != E; ++I) {
967 if (PredHasWeights || SuccHasWeights)
968 Weights.push_back(WeightsForHandled[*I]);
969 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
970 NewSuccessors.push_back(BBDefault);
974 // Okay, at this point, we know which new successor Pred will get. Make
975 // sure we update the number of entries in the PHI nodes for these
977 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i)
978 AddPredecessorToBlock(NewSuccessors[i], Pred, BB);
980 Builder.SetInsertPoint(PTI);
981 // Convert pointer to int before we switch.
982 if (CV->getType()->isPointerTy()) {
983 assert(TD && "Cannot switch on pointer without DataLayout");
984 CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getContext()),
988 // Now that the successors are updated, create the new Switch instruction.
989 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
991 NewSI->setDebugLoc(PTI->getDebugLoc());
992 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
993 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest);
995 if (PredHasWeights || SuccHasWeights) {
996 // Halve the weights if any of them cannot fit in an uint32_t
999 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1001 NewSI->setMetadata(LLVMContext::MD_prof,
1002 MDBuilder(BB->getContext()).
1003 createBranchWeights(MDWeights));
1006 EraseTerminatorInstAndDCECond(PTI);
1008 // Okay, last check. If BB is still a successor of PSI, then we must
1009 // have an infinite loop case. If so, add an infinitely looping block
1010 // to handle the case to preserve the behavior of the code.
1011 BasicBlock *InfLoopBlock = 0;
1012 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1013 if (NewSI->getSuccessor(i) == BB) {
1014 if (InfLoopBlock == 0) {
1015 // Insert it at the end of the function, because it's either code,
1016 // or it won't matter if it's hot. :)
1017 InfLoopBlock = BasicBlock::Create(BB->getContext(),
1018 "infloop", BB->getParent());
1019 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1021 NewSI->setSuccessor(i, InfLoopBlock);
1030 // isSafeToHoistInvoke - If we would need to insert a select that uses the
1031 // value of this invoke (comments in HoistThenElseCodeToIf explain why we
1032 // would need to do this), we can't hoist the invoke, as there is nowhere
1033 // to put the select in this case.
1034 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1035 Instruction *I1, Instruction *I2) {
1036 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1038 for (BasicBlock::iterator BBI = SI->begin();
1039 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1040 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1041 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1042 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
1050 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
1051 /// BB2, hoist any common code in the two blocks up into the branch block. The
1052 /// caller of this function guarantees that BI's block dominates BB1 and BB2.
1053 static bool HoistThenElseCodeToIf(BranchInst *BI) {
1054 // This does very trivial matching, with limited scanning, to find identical
1055 // instructions in the two blocks. In particular, we don't want to get into
1056 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1057 // such, we currently just scan for obviously identical instructions in an
1059 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1060 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1062 BasicBlock::iterator BB1_Itr = BB1->begin();
1063 BasicBlock::iterator BB2_Itr = BB2->begin();
1065 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++;
1066 // Skip debug info if it is not identical.
1067 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1068 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1069 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1070 while (isa<DbgInfoIntrinsic>(I1))
1072 while (isa<DbgInfoIntrinsic>(I2))
1075 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1076 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1079 // If we get here, we can hoist at least one instruction.
1080 BasicBlock *BIParent = BI->getParent();
1083 // If we are hoisting the terminator instruction, don't move one (making a
1084 // broken BB), instead clone it, and remove BI.
1085 if (isa<TerminatorInst>(I1))
1086 goto HoistTerminator;
1088 // For a normal instruction, we just move one to right before the branch,
1089 // then replace all uses of the other with the first. Finally, we remove
1090 // the now redundant second instruction.
1091 BIParent->getInstList().splice(BI, BB1->getInstList(), I1);
1092 if (!I2->use_empty())
1093 I2->replaceAllUsesWith(I1);
1094 I1->intersectOptionalDataWith(I2);
1095 I2->eraseFromParent();
1099 // Skip debug info if it is not identical.
1100 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1101 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1102 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1103 while (isa<DbgInfoIntrinsic>(I1))
1105 while (isa<DbgInfoIntrinsic>(I2))
1108 } while (I1->isIdenticalToWhenDefined(I2));
1113 // It may not be possible to hoist an invoke.
1114 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1117 // Okay, it is safe to hoist the terminator.
1118 Instruction *NT = I1->clone();
1119 BIParent->getInstList().insert(BI, NT);
1120 if (!NT->getType()->isVoidTy()) {
1121 I1->replaceAllUsesWith(NT);
1122 I2->replaceAllUsesWith(NT);
1126 IRBuilder<true, NoFolder> Builder(NT);
1127 // Hoisting one of the terminators from our successor is a great thing.
1128 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1129 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1130 // nodes, so we insert select instruction to compute the final result.
1131 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1132 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1134 for (BasicBlock::iterator BBI = SI->begin();
1135 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1136 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1137 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1138 if (BB1V == BB2V) continue;
1140 // These values do not agree. Insert a select instruction before NT
1141 // that determines the right value.
1142 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1144 SI = cast<SelectInst>
1145 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1146 BB1V->getName()+"."+BB2V->getName()));
1148 // Make the PHI node use the select for all incoming values for BB1/BB2
1149 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1150 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1151 PN->setIncomingValue(i, SI);
1155 // Update any PHI nodes in our new successors.
1156 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1157 AddPredecessorToBlock(*SI, BIParent, BB1);
1159 EraseTerminatorInstAndDCECond(BI);
1163 /// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd,
1164 /// check whether BBEnd has only two predecessors and the other predecessor
1165 /// ends with an unconditional branch. If it is true, sink any common code
1166 /// in the two predecessors to BBEnd.
1167 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1168 assert(BI1->isUnconditional());
1169 BasicBlock *BB1 = BI1->getParent();
1170 BasicBlock *BBEnd = BI1->getSuccessor(0);
1172 // Check that BBEnd has two predecessors and the other predecessor ends with
1173 // an unconditional branch.
1174 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1175 BasicBlock *Pred0 = *PI++;
1176 if (PI == PE) // Only one predecessor.
1178 BasicBlock *Pred1 = *PI++;
1179 if (PI != PE) // More than two predecessors.
1181 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1182 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1183 if (!BI2 || !BI2->isUnconditional())
1186 // Gather the PHI nodes in BBEnd.
1187 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2;
1188 Instruction *FirstNonPhiInBBEnd = 0;
1189 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end();
1191 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1192 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1193 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1194 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN);
1196 FirstNonPhiInBBEnd = &*I;
1200 if (!FirstNonPhiInBBEnd)
1204 // This does very trivial matching, with limited scanning, to find identical
1205 // instructions in the two blocks. We scan backward for obviously identical
1206 // instructions in an identical order.
1207 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1208 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(),
1209 RE2 = BB2->getInstList().rend();
1211 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1214 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1217 // Skip the unconditional branches.
1221 bool Changed = false;
1222 while (RI1 != RE1 && RI2 != RE2) {
1224 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1227 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1231 Instruction *I1 = &*RI1, *I2 = &*RI2;
1232 // I1 and I2 should have a single use in the same PHI node, and they
1233 // perform the same operation.
1234 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1235 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1236 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1237 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) ||
1238 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1239 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1240 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1241 !I1->hasOneUse() || !I2->hasOneUse() ||
1242 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() ||
1243 MapValueFromBB1ToBB2[I1].first != I2)
1246 // Check whether we should swap the operands of ICmpInst.
1247 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1248 bool SwapOpnds = false;
1249 if (ICmp1 && ICmp2 &&
1250 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1251 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1252 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1253 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1254 ICmp2->swapOperands();
1257 if (!I1->isSameOperationAs(I2)) {
1259 ICmp2->swapOperands();
1263 // The operands should be either the same or they need to be generated
1264 // with a PHI node after sinking. We only handle the case where there is
1265 // a single pair of different operands.
1266 Value *DifferentOp1 = 0, *DifferentOp2 = 0;
1267 unsigned Op1Idx = 0;
1268 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1269 if (I1->getOperand(I) == I2->getOperand(I))
1271 // Early exit if we have more-than one pair of different operands or
1272 // the different operand is already in MapValueFromBB1ToBB2.
1273 // Early exit if we need a PHI node to replace a constant.
1275 MapValueFromBB1ToBB2.find(I1->getOperand(I)) !=
1276 MapValueFromBB1ToBB2.end() ||
1277 isa<Constant>(I1->getOperand(I)) ||
1278 isa<Constant>(I2->getOperand(I))) {
1279 // If we can't sink the instructions, undo the swapping.
1281 ICmp2->swapOperands();
1284 DifferentOp1 = I1->getOperand(I);
1286 DifferentOp2 = I2->getOperand(I);
1289 // We insert the pair of different operands to MapValueFromBB1ToBB2 and
1290 // remove (I1, I2) from MapValueFromBB1ToBB2.
1292 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2,
1293 DifferentOp1->getName() + ".sink",
1295 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN);
1296 // I1 should use NewPN instead of DifferentOp1.
1297 I1->setOperand(Op1Idx, NewPN);
1298 NewPN->addIncoming(DifferentOp1, BB1);
1299 NewPN->addIncoming(DifferentOp2, BB2);
1300 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1302 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second;
1303 MapValueFromBB1ToBB2.erase(I1);
1305 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";);
1306 DEBUG(dbgs() << " " << *I2 << "\n";);
1307 // We need to update RE1 and RE2 if we are going to sink the first
1308 // instruction in the basic block down.
1309 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1310 // Sink the instruction.
1311 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1);
1312 if (!OldPN->use_empty())
1313 OldPN->replaceAllUsesWith(I1);
1314 OldPN->eraseFromParent();
1316 if (!I2->use_empty())
1317 I2->replaceAllUsesWith(I1);
1318 I1->intersectOptionalDataWith(I2);
1319 I2->eraseFromParent();
1322 RE1 = BB1->getInstList().rend();
1324 RE2 = BB2->getInstList().rend();
1325 FirstNonPhiInBBEnd = I1;
1332 /// SpeculativelyExecuteBB - Given a conditional branch that goes to BB1
1333 /// and an BB2 and the only successor of BB1 is BB2, hoist simple code
1334 /// (for now, restricted to a single instruction that's side effect free) from
1335 /// the BB1 into the branch block to speculatively execute it.
1340 /// br i1 %t1, label %BB1, label %BB2
1342 /// %t3 = add %t2, c
1348 /// %t4 = add %t2, c
1349 /// %t3 = select i1 %t1, %t2, %t3
1350 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1) {
1351 // Only speculatively execution a single instruction (not counting the
1352 // terminator) for now.
1353 Instruction *HInst = NULL;
1354 Instruction *Term = BB1->getTerminator();
1355 for (BasicBlock::iterator BBI = BB1->begin(), BBE = BB1->end();
1356 BBI != BBE; ++BBI) {
1357 Instruction *I = BBI;
1359 if (isa<DbgInfoIntrinsic>(I)) continue;
1360 if (I == Term) break;
1367 BasicBlock *BIParent = BI->getParent();
1369 // Check the instruction to be hoisted, if there is one.
1371 // Don't hoist the instruction if it's unsafe or expensive.
1372 if (!isSafeToSpeculativelyExecute(HInst))
1374 if (ComputeSpeculationCost(HInst) > PHINodeFoldingThreshold)
1377 // Do not hoist the instruction if any of its operands are defined but not
1378 // used in this BB. The transformation will prevent the operand from
1379 // being sunk into the use block.
1380 for (User::op_iterator i = HInst->op_begin(), e = HInst->op_end();
1382 Instruction *OpI = dyn_cast<Instruction>(*i);
1383 if (OpI && OpI->getParent() == BIParent &&
1384 !OpI->mayHaveSideEffects() &&
1385 !OpI->isUsedInBasicBlock(BIParent))
1390 // Be conservative for now. FP select instruction can often be expensive.
1391 Value *BrCond = BI->getCondition();
1392 if (isa<FCmpInst>(BrCond))
1395 // If BB1 is actually on the false edge of the conditional branch, remember
1396 // to swap the select operands later.
1397 bool Invert = false;
1398 if (BB1 != BI->getSuccessor(0)) {
1399 assert(BB1 == BI->getSuccessor(1) && "No edge from 'if' block?");
1403 // Collect interesting PHIs, and scan for hazards.
1404 SmallSetVector<std::pair<Value *, Value *>, 4> PHIs;
1405 BasicBlock *BB2 = BB1->getTerminator()->getSuccessor(0);
1406 for (BasicBlock::iterator I = BB2->begin();
1407 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1408 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1409 Value *BIParentV = PN->getIncomingValueForBlock(BIParent);
1411 // Skip PHIs which are trivial.
1412 if (BB1V == BIParentV)
1415 // Check for saftey.
1416 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BB1V)) {
1417 // An unfolded ConstantExpr could end up getting expanded into
1418 // Instructions. Don't speculate this and another instruction at
1422 if (!isSafeToSpeculativelyExecute(CE))
1424 if (ComputeSpeculationCost(CE) > PHINodeFoldingThreshold)
1428 // Ok, we may insert a select for this PHI.
1429 PHIs.insert(std::make_pair(BB1V, BIParentV));
1432 // If there are no PHIs to process, bail early. This helps ensure idempotence
1437 // If we get here, we can hoist the instruction and if-convert.
1438 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *BB1 << "\n";);
1440 // Hoist the instruction.
1442 BIParent->getInstList().splice(BI, BB1->getInstList(), HInst);
1444 // Insert selects and rewrite the PHI operands.
1445 IRBuilder<true, NoFolder> Builder(BI);
1446 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
1447 Value *TrueV = PHIs[i].first;
1448 Value *FalseV = PHIs[i].second;
1450 // Create a select whose true value is the speculatively executed value and
1451 // false value is the previously determined FalseV.
1454 SI = cast<SelectInst>
1455 (Builder.CreateSelect(BrCond, FalseV, TrueV,
1456 FalseV->getName() + "." + TrueV->getName()));
1458 SI = cast<SelectInst>
1459 (Builder.CreateSelect(BrCond, TrueV, FalseV,
1460 TrueV->getName() + "." + FalseV->getName()));
1462 // Make the PHI node use the select for all incoming values for "then" and
1464 for (BasicBlock::iterator I = BB2->begin();
1465 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1466 unsigned BB1I = PN->getBasicBlockIndex(BB1);
1467 unsigned BIParentI = PN->getBasicBlockIndex(BIParent);
1468 Value *BB1V = PN->getIncomingValue(BB1I);
1469 Value *BIParentV = PN->getIncomingValue(BIParentI);
1470 if (TrueV == BB1V && FalseV == BIParentV) {
1471 PN->setIncomingValue(BB1I, SI);
1472 PN->setIncomingValue(BIParentI, SI);
1481 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1482 /// across this block.
1483 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1484 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1487 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1488 if (isa<DbgInfoIntrinsic>(BBI))
1490 if (Size > 10) return false; // Don't clone large BB's.
1493 // We can only support instructions that do not define values that are
1494 // live outside of the current basic block.
1495 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end();
1497 Instruction *U = cast<Instruction>(*UI);
1498 if (U->getParent() != BB || isa<PHINode>(U)) return false;
1501 // Looks ok, continue checking.
1507 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1508 /// that is defined in the same block as the branch and if any PHI entries are
1509 /// constants, thread edges corresponding to that entry to be branches to their
1510 /// ultimate destination.
1511 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *TD) {
1512 BasicBlock *BB = BI->getParent();
1513 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1514 // NOTE: we currently cannot transform this case if the PHI node is used
1515 // outside of the block.
1516 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1519 // Degenerate case of a single entry PHI.
1520 if (PN->getNumIncomingValues() == 1) {
1521 FoldSingleEntryPHINodes(PN->getParent());
1525 // Now we know that this block has multiple preds and two succs.
1526 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1528 // Okay, this is a simple enough basic block. See if any phi values are
1530 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1531 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1532 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue;
1534 // Okay, we now know that all edges from PredBB should be revectored to
1535 // branch to RealDest.
1536 BasicBlock *PredBB = PN->getIncomingBlock(i);
1537 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1539 if (RealDest == BB) continue; // Skip self loops.
1540 // Skip if the predecessor's terminator is an indirect branch.
1541 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1543 // The dest block might have PHI nodes, other predecessors and other
1544 // difficult cases. Instead of being smart about this, just insert a new
1545 // block that jumps to the destination block, effectively splitting
1546 // the edge we are about to create.
1547 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1548 RealDest->getName()+".critedge",
1549 RealDest->getParent(), RealDest);
1550 BranchInst::Create(RealDest, EdgeBB);
1552 // Update PHI nodes.
1553 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1555 // BB may have instructions that are being threaded over. Clone these
1556 // instructions into EdgeBB. We know that there will be no uses of the
1557 // cloned instructions outside of EdgeBB.
1558 BasicBlock::iterator InsertPt = EdgeBB->begin();
1559 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1560 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1561 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1562 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1565 // Clone the instruction.
1566 Instruction *N = BBI->clone();
1567 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1569 // Update operands due to translation.
1570 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1572 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1573 if (PI != TranslateMap.end())
1577 // Check for trivial simplification.
1578 if (Value *V = SimplifyInstruction(N, TD)) {
1579 TranslateMap[BBI] = V;
1580 delete N; // Instruction folded away, don't need actual inst
1582 // Insert the new instruction into its new home.
1583 EdgeBB->getInstList().insert(InsertPt, N);
1584 if (!BBI->use_empty())
1585 TranslateMap[BBI] = N;
1589 // Loop over all of the edges from PredBB to BB, changing them to branch
1590 // to EdgeBB instead.
1591 TerminatorInst *PredBBTI = PredBB->getTerminator();
1592 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1593 if (PredBBTI->getSuccessor(i) == BB) {
1594 BB->removePredecessor(PredBB);
1595 PredBBTI->setSuccessor(i, EdgeBB);
1598 // Recurse, simplifying any other constants.
1599 return FoldCondBranchOnPHI(BI, TD) | true;
1605 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1606 /// PHI node, see if we can eliminate it.
1607 static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *TD) {
1608 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1609 // statement", which has a very simple dominance structure. Basically, we
1610 // are trying to find the condition that is being branched on, which
1611 // subsequently causes this merge to happen. We really want control
1612 // dependence information for this check, but simplifycfg can't keep it up
1613 // to date, and this catches most of the cases we care about anyway.
1614 BasicBlock *BB = PN->getParent();
1615 BasicBlock *IfTrue, *IfFalse;
1616 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1618 // Don't bother if the branch will be constant folded trivially.
1619 isa<ConstantInt>(IfCond))
1622 // Okay, we found that we can merge this two-entry phi node into a select.
1623 // Doing so would require us to fold *all* two entry phi nodes in this block.
1624 // At some point this becomes non-profitable (particularly if the target
1625 // doesn't support cmov's). Only do this transformation if there are two or
1626 // fewer PHI nodes in this block.
1627 unsigned NumPhis = 0;
1628 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1632 // Loop over the PHI's seeing if we can promote them all to select
1633 // instructions. While we are at it, keep track of the instructions
1634 // that need to be moved to the dominating block.
1635 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1636 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1637 MaxCostVal1 = PHINodeFoldingThreshold;
1639 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1640 PHINode *PN = cast<PHINode>(II++);
1641 if (Value *V = SimplifyInstruction(PN, TD)) {
1642 PN->replaceAllUsesWith(V);
1643 PN->eraseFromParent();
1647 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1649 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1654 // If we folded the first phi, PN dangles at this point. Refresh it. If
1655 // we ran out of PHIs then we simplified them all.
1656 PN = dyn_cast<PHINode>(BB->begin());
1657 if (PN == 0) return true;
1659 // Don't fold i1 branches on PHIs which contain binary operators. These can
1660 // often be turned into switches and other things.
1661 if (PN->getType()->isIntegerTy(1) &&
1662 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1663 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1664 isa<BinaryOperator>(IfCond)))
1667 // If we all PHI nodes are promotable, check to make sure that all
1668 // instructions in the predecessor blocks can be promoted as well. If
1669 // not, we won't be able to get rid of the control flow, so it's not
1670 // worth promoting to select instructions.
1671 BasicBlock *DomBlock = 0;
1672 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1673 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1674 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1677 DomBlock = *pred_begin(IfBlock1);
1678 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1679 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1680 // This is not an aggressive instruction that we can promote.
1681 // Because of this, we won't be able to get rid of the control
1682 // flow, so the xform is not worth it.
1687 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1690 DomBlock = *pred_begin(IfBlock2);
1691 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1692 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1693 // This is not an aggressive instruction that we can promote.
1694 // Because of this, we won't be able to get rid of the control
1695 // flow, so the xform is not worth it.
1700 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1701 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1703 // If we can still promote the PHI nodes after this gauntlet of tests,
1704 // do all of the PHI's now.
1705 Instruction *InsertPt = DomBlock->getTerminator();
1706 IRBuilder<true, NoFolder> Builder(InsertPt);
1708 // Move all 'aggressive' instructions, which are defined in the
1709 // conditional parts of the if's up to the dominating block.
1711 DomBlock->getInstList().splice(InsertPt,
1712 IfBlock1->getInstList(), IfBlock1->begin(),
1713 IfBlock1->getTerminator());
1715 DomBlock->getInstList().splice(InsertPt,
1716 IfBlock2->getInstList(), IfBlock2->begin(),
1717 IfBlock2->getTerminator());
1719 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1720 // Change the PHI node into a select instruction.
1721 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1722 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1725 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1726 PN->replaceAllUsesWith(NV);
1728 PN->eraseFromParent();
1731 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1732 // has been flattened. Change DomBlock to jump directly to our new block to
1733 // avoid other simplifycfg's kicking in on the diamond.
1734 TerminatorInst *OldTI = DomBlock->getTerminator();
1735 Builder.SetInsertPoint(OldTI);
1736 Builder.CreateBr(BB);
1737 OldTI->eraseFromParent();
1741 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1742 /// to two returning blocks, try to merge them together into one return,
1743 /// introducing a select if the return values disagree.
1744 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1745 IRBuilder<> &Builder) {
1746 assert(BI->isConditional() && "Must be a conditional branch");
1747 BasicBlock *TrueSucc = BI->getSuccessor(0);
1748 BasicBlock *FalseSucc = BI->getSuccessor(1);
1749 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1750 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1752 // Check to ensure both blocks are empty (just a return) or optionally empty
1753 // with PHI nodes. If there are other instructions, merging would cause extra
1754 // computation on one path or the other.
1755 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1757 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1760 Builder.SetInsertPoint(BI);
1761 // Okay, we found a branch that is going to two return nodes. If
1762 // there is no return value for this function, just change the
1763 // branch into a return.
1764 if (FalseRet->getNumOperands() == 0) {
1765 TrueSucc->removePredecessor(BI->getParent());
1766 FalseSucc->removePredecessor(BI->getParent());
1767 Builder.CreateRetVoid();
1768 EraseTerminatorInstAndDCECond(BI);
1772 // Otherwise, figure out what the true and false return values are
1773 // so we can insert a new select instruction.
1774 Value *TrueValue = TrueRet->getReturnValue();
1775 Value *FalseValue = FalseRet->getReturnValue();
1777 // Unwrap any PHI nodes in the return blocks.
1778 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1779 if (TVPN->getParent() == TrueSucc)
1780 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1781 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1782 if (FVPN->getParent() == FalseSucc)
1783 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1785 // In order for this transformation to be safe, we must be able to
1786 // unconditionally execute both operands to the return. This is
1787 // normally the case, but we could have a potentially-trapping
1788 // constant expression that prevents this transformation from being
1790 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1793 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1797 // Okay, we collected all the mapped values and checked them for sanity, and
1798 // defined to really do this transformation. First, update the CFG.
1799 TrueSucc->removePredecessor(BI->getParent());
1800 FalseSucc->removePredecessor(BI->getParent());
1802 // Insert select instructions where needed.
1803 Value *BrCond = BI->getCondition();
1805 // Insert a select if the results differ.
1806 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1807 } else if (isa<UndefValue>(TrueValue)) {
1808 TrueValue = FalseValue;
1810 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1811 FalseValue, "retval");
1815 Value *RI = !TrueValue ?
1816 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
1820 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
1821 << "\n " << *BI << "NewRet = " << *RI
1822 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
1824 EraseTerminatorInstAndDCECond(BI);
1829 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
1830 /// probabilities of the branch taking each edge. Fills in the two APInt
1831 /// parameters and return true, or returns false if no or invalid metadata was
1833 static bool ExtractBranchMetadata(BranchInst *BI,
1834 uint64_t &ProbTrue, uint64_t &ProbFalse) {
1835 assert(BI->isConditional() &&
1836 "Looking for probabilities on unconditional branch?");
1837 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
1838 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
1839 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
1840 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
1841 if (!CITrue || !CIFalse) return false;
1842 ProbTrue = CITrue->getValue().getZExtValue();
1843 ProbFalse = CIFalse->getValue().getZExtValue();
1847 /// checkCSEInPredecessor - Return true if the given instruction is available
1848 /// in its predecessor block. If yes, the instruction will be removed.
1850 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
1851 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
1853 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
1854 Instruction *PBI = &*I;
1855 // Check whether Inst and PBI generate the same value.
1856 if (Inst->isIdenticalTo(PBI)) {
1857 Inst->replaceAllUsesWith(PBI);
1858 Inst->eraseFromParent();
1865 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
1866 /// predecessor branches to us and one of our successors, fold the block into
1867 /// the predecessor and use logical operations to pick the right destination.
1868 bool llvm::FoldBranchToCommonDest(BranchInst *BI) {
1869 BasicBlock *BB = BI->getParent();
1871 Instruction *Cond = 0;
1872 if (BI->isConditional())
1873 Cond = dyn_cast<Instruction>(BI->getCondition());
1875 // For unconditional branch, check for a simple CFG pattern, where
1876 // BB has a single predecessor and BB's successor is also its predecessor's
1877 // successor. If such pattern exisits, check for CSE between BB and its
1879 if (BasicBlock *PB = BB->getSinglePredecessor())
1880 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
1881 if (PBI->isConditional() &&
1882 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
1883 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
1884 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
1886 Instruction *Curr = I++;
1887 if (isa<CmpInst>(Curr)) {
1891 // Quit if we can't remove this instruction.
1892 if (!checkCSEInPredecessor(Curr, PB))
1901 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
1902 Cond->getParent() != BB || !Cond->hasOneUse())
1905 // Only allow this if the condition is a simple instruction that can be
1906 // executed unconditionally. It must be in the same block as the branch, and
1907 // must be at the front of the block.
1908 BasicBlock::iterator FrontIt = BB->front();
1910 // Ignore dbg intrinsics.
1911 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
1913 // Allow a single instruction to be hoisted in addition to the compare
1914 // that feeds the branch. We later ensure that any values that _it_ uses
1915 // were also live in the predecessor, so that we don't unnecessarily create
1916 // register pressure or inhibit out-of-order execution.
1917 Instruction *BonusInst = 0;
1918 if (&*FrontIt != Cond &&
1919 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond &&
1920 isSafeToSpeculativelyExecute(FrontIt)) {
1921 BonusInst = &*FrontIt;
1924 // Ignore dbg intrinsics.
1925 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
1928 // Only a single bonus inst is allowed.
1929 if (&*FrontIt != Cond)
1932 // Make sure the instruction after the condition is the cond branch.
1933 BasicBlock::iterator CondIt = Cond; ++CondIt;
1935 // Ingore dbg intrinsics.
1936 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
1941 // Cond is known to be a compare or binary operator. Check to make sure that
1942 // neither operand is a potentially-trapping constant expression.
1943 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
1946 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
1950 // Finally, don't infinitely unroll conditional loops.
1951 BasicBlock *TrueDest = BI->getSuccessor(0);
1952 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0;
1953 if (TrueDest == BB || FalseDest == BB)
1956 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
1957 BasicBlock *PredBlock = *PI;
1958 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
1960 // Check that we have two conditional branches. If there is a PHI node in
1961 // the common successor, verify that the same value flows in from both
1963 SmallVector<PHINode*, 4> PHIs;
1964 if (PBI == 0 || PBI->isUnconditional() ||
1965 (BI->isConditional() &&
1966 !SafeToMergeTerminators(BI, PBI)) ||
1967 (!BI->isConditional() &&
1968 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
1971 // Determine if the two branches share a common destination.
1972 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
1973 bool InvertPredCond = false;
1975 if (BI->isConditional()) {
1976 if (PBI->getSuccessor(0) == TrueDest)
1977 Opc = Instruction::Or;
1978 else if (PBI->getSuccessor(1) == FalseDest)
1979 Opc = Instruction::And;
1980 else if (PBI->getSuccessor(0) == FalseDest)
1981 Opc = Instruction::And, InvertPredCond = true;
1982 else if (PBI->getSuccessor(1) == TrueDest)
1983 Opc = Instruction::Or, InvertPredCond = true;
1987 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
1991 // Ensure that any values used in the bonus instruction are also used
1992 // by the terminator of the predecessor. This means that those values
1993 // must already have been resolved, so we won't be inhibiting the
1994 // out-of-order core by speculating them earlier.
1996 // Collect the values used by the bonus inst
1997 SmallPtrSet<Value*, 4> UsedValues;
1998 for (Instruction::op_iterator OI = BonusInst->op_begin(),
1999 OE = BonusInst->op_end(); OI != OE; ++OI) {
2001 if (!isa<Constant>(V))
2002 UsedValues.insert(V);
2005 SmallVector<std::pair<Value*, unsigned>, 4> Worklist;
2006 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0));
2008 // Walk up to four levels back up the use-def chain of the predecessor's
2009 // terminator to see if all those values were used. The choice of four
2010 // levels is arbitrary, to provide a compile-time-cost bound.
2011 while (!Worklist.empty()) {
2012 std::pair<Value*, unsigned> Pair = Worklist.back();
2013 Worklist.pop_back();
2015 if (Pair.second >= 4) continue;
2016 UsedValues.erase(Pair.first);
2017 if (UsedValues.empty()) break;
2019 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) {
2020 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
2022 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1));
2026 if (!UsedValues.empty()) return false;
2029 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2030 IRBuilder<> Builder(PBI);
2032 // If we need to invert the condition in the pred block to match, do so now.
2033 if (InvertPredCond) {
2034 Value *NewCond = PBI->getCondition();
2036 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2037 CmpInst *CI = cast<CmpInst>(NewCond);
2038 CI->setPredicate(CI->getInversePredicate());
2040 NewCond = Builder.CreateNot(NewCond,
2041 PBI->getCondition()->getName()+".not");
2044 PBI->setCondition(NewCond);
2045 PBI->swapSuccessors();
2048 // If we have a bonus inst, clone it into the predecessor block.
2049 Instruction *NewBonus = 0;
2051 NewBonus = BonusInst->clone();
2052 PredBlock->getInstList().insert(PBI, NewBonus);
2053 NewBonus->takeName(BonusInst);
2054 BonusInst->setName(BonusInst->getName()+".old");
2057 // Clone Cond into the predecessor basic block, and or/and the
2058 // two conditions together.
2059 Instruction *New = Cond->clone();
2060 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus);
2061 PredBlock->getInstList().insert(PBI, New);
2062 New->takeName(Cond);
2063 Cond->setName(New->getName()+".old");
2065 if (BI->isConditional()) {
2066 Instruction *NewCond =
2067 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2069 PBI->setCondition(NewCond);
2071 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2072 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2074 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2076 SmallVector<uint64_t, 8> NewWeights;
2078 if (PBI->getSuccessor(0) == BB) {
2079 if (PredHasWeights && SuccHasWeights) {
2080 // PBI: br i1 %x, BB, FalseDest
2081 // BI: br i1 %y, TrueDest, FalseDest
2082 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2083 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2084 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2085 // TrueWeight for PBI * FalseWeight for BI.
2086 // We assume that total weights of a BranchInst can fit into 32 bits.
2087 // Therefore, we will not have overflow using 64-bit arithmetic.
2088 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2089 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2091 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2092 PBI->setSuccessor(0, TrueDest);
2094 if (PBI->getSuccessor(1) == BB) {
2095 if (PredHasWeights && SuccHasWeights) {
2096 // PBI: br i1 %x, TrueDest, BB
2097 // BI: br i1 %y, TrueDest, FalseDest
2098 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2099 // FalseWeight for PBI * TrueWeight for BI.
2100 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2101 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2102 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2103 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2105 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2106 PBI->setSuccessor(1, FalseDest);
2108 if (NewWeights.size() == 2) {
2109 // Halve the weights if any of them cannot fit in an uint32_t
2110 FitWeights(NewWeights);
2112 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2113 PBI->setMetadata(LLVMContext::MD_prof,
2114 MDBuilder(BI->getContext()).
2115 createBranchWeights(MDWeights));
2117 PBI->setMetadata(LLVMContext::MD_prof, NULL);
2119 // Update PHI nodes in the common successors.
2120 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2121 ConstantInt *PBI_C = cast<ConstantInt>(
2122 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2123 assert(PBI_C->getType()->isIntegerTy(1));
2124 Instruction *MergedCond = 0;
2125 if (PBI->getSuccessor(0) == TrueDest) {
2126 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2127 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2128 // is false: !PBI_Cond and BI_Value
2129 Instruction *NotCond =
2130 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2133 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2138 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2139 PBI->getCondition(), MergedCond,
2142 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2143 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2144 // is false: PBI_Cond and BI_Value
2146 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2147 PBI->getCondition(), New,
2149 if (PBI_C->isOne()) {
2150 Instruction *NotCond =
2151 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2154 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2155 NotCond, MergedCond,
2160 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2163 // Change PBI from Conditional to Unconditional.
2164 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2165 EraseTerminatorInstAndDCECond(PBI);
2169 // TODO: If BB is reachable from all paths through PredBlock, then we
2170 // could replace PBI's branch probabilities with BI's.
2172 // Copy any debug value intrinsics into the end of PredBlock.
2173 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2174 if (isa<DbgInfoIntrinsic>(*I))
2175 I->clone()->insertBefore(PBI);
2182 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2183 /// predecessor of another block, this function tries to simplify it. We know
2184 /// that PBI and BI are both conditional branches, and BI is in one of the
2185 /// successor blocks of PBI - PBI branches to BI.
2186 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2187 assert(PBI->isConditional() && BI->isConditional());
2188 BasicBlock *BB = BI->getParent();
2190 // If this block ends with a branch instruction, and if there is a
2191 // predecessor that ends on a branch of the same condition, make
2192 // this conditional branch redundant.
2193 if (PBI->getCondition() == BI->getCondition() &&
2194 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2195 // Okay, the outcome of this conditional branch is statically
2196 // knowable. If this block had a single pred, handle specially.
2197 if (BB->getSinglePredecessor()) {
2198 // Turn this into a branch on constant.
2199 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2200 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2202 return true; // Nuke the branch on constant.
2205 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2206 // in the constant and simplify the block result. Subsequent passes of
2207 // simplifycfg will thread the block.
2208 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2209 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2210 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2211 std::distance(PB, PE),
2212 BI->getCondition()->getName() + ".pr",
2214 // Okay, we're going to insert the PHI node. Since PBI is not the only
2215 // predecessor, compute the PHI'd conditional value for all of the preds.
2216 // Any predecessor where the condition is not computable we keep symbolic.
2217 for (pred_iterator PI = PB; PI != PE; ++PI) {
2218 BasicBlock *P = *PI;
2219 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2220 PBI != BI && PBI->isConditional() &&
2221 PBI->getCondition() == BI->getCondition() &&
2222 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2223 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2224 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2227 NewPN->addIncoming(BI->getCondition(), P);
2231 BI->setCondition(NewPN);
2236 // If this is a conditional branch in an empty block, and if any
2237 // predecessors is a conditional branch to one of our destinations,
2238 // fold the conditions into logical ops and one cond br.
2239 BasicBlock::iterator BBI = BB->begin();
2240 // Ignore dbg intrinsics.
2241 while (isa<DbgInfoIntrinsic>(BBI))
2247 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2252 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2254 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2255 PBIOp = 0, BIOp = 1;
2256 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2257 PBIOp = 1, BIOp = 0;
2258 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2263 // Check to make sure that the other destination of this branch
2264 // isn't BB itself. If so, this is an infinite loop that will
2265 // keep getting unwound.
2266 if (PBI->getSuccessor(PBIOp) == BB)
2269 // Do not perform this transformation if it would require
2270 // insertion of a large number of select instructions. For targets
2271 // without predication/cmovs, this is a big pessimization.
2272 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2274 unsigned NumPhis = 0;
2275 for (BasicBlock::iterator II = CommonDest->begin();
2276 isa<PHINode>(II); ++II, ++NumPhis)
2277 if (NumPhis > 2) // Disable this xform.
2280 // Finally, if everything is ok, fold the branches to logical ops.
2281 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2283 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2284 << "AND: " << *BI->getParent());
2287 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2288 // branch in it, where one edge (OtherDest) goes back to itself but the other
2289 // exits. We don't *know* that the program avoids the infinite loop
2290 // (even though that seems likely). If we do this xform naively, we'll end up
2291 // recursively unpeeling the loop. Since we know that (after the xform is
2292 // done) that the block *is* infinite if reached, we just make it an obviously
2293 // infinite loop with no cond branch.
2294 if (OtherDest == BB) {
2295 // Insert it at the end of the function, because it's either code,
2296 // or it won't matter if it's hot. :)
2297 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2298 "infloop", BB->getParent());
2299 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2300 OtherDest = InfLoopBlock;
2303 DEBUG(dbgs() << *PBI->getParent()->getParent());
2305 // BI may have other predecessors. Because of this, we leave
2306 // it alone, but modify PBI.
2308 // Make sure we get to CommonDest on True&True directions.
2309 Value *PBICond = PBI->getCondition();
2310 IRBuilder<true, NoFolder> Builder(PBI);
2312 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2314 Value *BICond = BI->getCondition();
2316 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2318 // Merge the conditions.
2319 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2321 // Modify PBI to branch on the new condition to the new dests.
2322 PBI->setCondition(Cond);
2323 PBI->setSuccessor(0, CommonDest);
2324 PBI->setSuccessor(1, OtherDest);
2326 // Update branch weight for PBI.
2327 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2328 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2330 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2332 if (PredHasWeights && SuccHasWeights) {
2333 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2334 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2335 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2336 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2337 // The weight to CommonDest should be PredCommon * SuccTotal +
2338 // PredOther * SuccCommon.
2339 // The weight to OtherDest should be PredOther * SuccOther.
2340 SmallVector<uint64_t, 2> NewWeights;
2341 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) +
2342 PredOther * SuccCommon);
2343 NewWeights.push_back(PredOther * SuccOther);
2344 // Halve the weights if any of them cannot fit in an uint32_t
2345 FitWeights(NewWeights);
2347 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end());
2348 PBI->setMetadata(LLVMContext::MD_prof,
2349 MDBuilder(BI->getContext()).
2350 createBranchWeights(MDWeights));
2353 // OtherDest may have phi nodes. If so, add an entry from PBI's
2354 // block that are identical to the entries for BI's block.
2355 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2357 // We know that the CommonDest already had an edge from PBI to
2358 // it. If it has PHIs though, the PHIs may have different
2359 // entries for BB and PBI's BB. If so, insert a select to make
2362 for (BasicBlock::iterator II = CommonDest->begin();
2363 (PN = dyn_cast<PHINode>(II)); ++II) {
2364 Value *BIV = PN->getIncomingValueForBlock(BB);
2365 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2366 Value *PBIV = PN->getIncomingValue(PBBIdx);
2368 // Insert a select in PBI to pick the right value.
2369 Value *NV = cast<SelectInst>
2370 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2371 PN->setIncomingValue(PBBIdx, NV);
2375 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2376 DEBUG(dbgs() << *PBI->getParent()->getParent());
2378 // This basic block is probably dead. We know it has at least
2379 // one fewer predecessor.
2383 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2384 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2385 // Takes care of updating the successors and removing the old terminator.
2386 // Also makes sure not to introduce new successors by assuming that edges to
2387 // non-successor TrueBBs and FalseBBs aren't reachable.
2388 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2389 BasicBlock *TrueBB, BasicBlock *FalseBB,
2390 uint32_t TrueWeight,
2391 uint32_t FalseWeight){
2392 // Remove any superfluous successor edges from the CFG.
2393 // First, figure out which successors to preserve.
2394 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2396 BasicBlock *KeepEdge1 = TrueBB;
2397 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0;
2399 // Then remove the rest.
2400 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2401 BasicBlock *Succ = OldTerm->getSuccessor(I);
2402 // Make sure only to keep exactly one copy of each edge.
2403 if (Succ == KeepEdge1)
2405 else if (Succ == KeepEdge2)
2408 Succ->removePredecessor(OldTerm->getParent());
2411 IRBuilder<> Builder(OldTerm);
2412 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2414 // Insert an appropriate new terminator.
2415 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) {
2416 if (TrueBB == FalseBB)
2417 // We were only looking for one successor, and it was present.
2418 // Create an unconditional branch to it.
2419 Builder.CreateBr(TrueBB);
2421 // We found both of the successors we were looking for.
2422 // Create a conditional branch sharing the condition of the select.
2423 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2424 if (TrueWeight != FalseWeight)
2425 NewBI->setMetadata(LLVMContext::MD_prof,
2426 MDBuilder(OldTerm->getContext()).
2427 createBranchWeights(TrueWeight, FalseWeight));
2429 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2430 // Neither of the selected blocks were successors, so this
2431 // terminator must be unreachable.
2432 new UnreachableInst(OldTerm->getContext(), OldTerm);
2434 // One of the selected values was a successor, but the other wasn't.
2435 // Insert an unconditional branch to the one that was found;
2436 // the edge to the one that wasn't must be unreachable.
2438 // Only TrueBB was found.
2439 Builder.CreateBr(TrueBB);
2441 // Only FalseBB was found.
2442 Builder.CreateBr(FalseBB);
2445 EraseTerminatorInstAndDCECond(OldTerm);
2449 // SimplifySwitchOnSelect - Replaces
2450 // (switch (select cond, X, Y)) on constant X, Y
2451 // with a branch - conditional if X and Y lead to distinct BBs,
2452 // unconditional otherwise.
2453 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2454 // Check for constant integer values in the select.
2455 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2456 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2457 if (!TrueVal || !FalseVal)
2460 // Find the relevant condition and destinations.
2461 Value *Condition = Select->getCondition();
2462 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2463 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2465 // Get weight for TrueBB and FalseBB.
2466 uint32_t TrueWeight = 0, FalseWeight = 0;
2467 SmallVector<uint64_t, 8> Weights;
2468 bool HasWeights = HasBranchWeights(SI);
2470 GetBranchWeights(SI, Weights);
2471 if (Weights.size() == 1 + SI->getNumCases()) {
2472 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2473 getSuccessorIndex()];
2474 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2475 getSuccessorIndex()];
2479 // Perform the actual simplification.
2480 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2481 TrueWeight, FalseWeight);
2484 // SimplifyIndirectBrOnSelect - Replaces
2485 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2486 // blockaddress(@fn, BlockB)))
2488 // (br cond, BlockA, BlockB).
2489 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2490 // Check that both operands of the select are block addresses.
2491 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2492 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2496 // Extract the actual blocks.
2497 BasicBlock *TrueBB = TBA->getBasicBlock();
2498 BasicBlock *FalseBB = FBA->getBasicBlock();
2500 // Perform the actual simplification.
2501 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2505 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2506 /// instruction (a seteq/setne with a constant) as the only instruction in a
2507 /// block that ends with an uncond branch. We are looking for a very specific
2508 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2509 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2510 /// default value goes to an uncond block with a seteq in it, we get something
2513 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2515 /// %tmp = icmp eq i8 %A, 92
2518 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2520 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2521 /// the PHI, merging the third icmp into the switch.
2522 static bool TryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI,
2523 const DataLayout *TD,
2524 IRBuilder<> &Builder) {
2525 BasicBlock *BB = ICI->getParent();
2527 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2529 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2531 Value *V = ICI->getOperand(0);
2532 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2534 // The pattern we're looking for is where our only predecessor is a switch on
2535 // 'V' and this block is the default case for the switch. In this case we can
2536 // fold the compared value into the switch to simplify things.
2537 BasicBlock *Pred = BB->getSinglePredecessor();
2538 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false;
2540 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2541 if (SI->getCondition() != V)
2544 // If BB is reachable on a non-default case, then we simply know the value of
2545 // V in this block. Substitute it and constant fold the icmp instruction
2547 if (SI->getDefaultDest() != BB) {
2548 ConstantInt *VVal = SI->findCaseDest(BB);
2549 assert(VVal && "Should have a unique destination value");
2550 ICI->setOperand(0, VVal);
2552 if (Value *V = SimplifyInstruction(ICI, TD)) {
2553 ICI->replaceAllUsesWith(V);
2554 ICI->eraseFromParent();
2556 // BB is now empty, so it is likely to simplify away.
2557 return SimplifyCFG(BB) | true;
2560 // Ok, the block is reachable from the default dest. If the constant we're
2561 // comparing exists in one of the other edges, then we can constant fold ICI
2563 if (SI->findCaseValue(Cst) != SI->case_default()) {
2565 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2566 V = ConstantInt::getFalse(BB->getContext());
2568 V = ConstantInt::getTrue(BB->getContext());
2570 ICI->replaceAllUsesWith(V);
2571 ICI->eraseFromParent();
2572 // BB is now empty, so it is likely to simplify away.
2573 return SimplifyCFG(BB) | true;
2576 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2578 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2579 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back());
2580 if (PHIUse == 0 || PHIUse != &SuccBlock->front() ||
2581 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2584 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2586 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2587 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2589 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2590 std::swap(DefaultCst, NewCst);
2592 // Replace ICI (which is used by the PHI for the default value) with true or
2593 // false depending on if it is EQ or NE.
2594 ICI->replaceAllUsesWith(DefaultCst);
2595 ICI->eraseFromParent();
2597 // Okay, the switch goes to this block on a default value. Add an edge from
2598 // the switch to the merge point on the compared value.
2599 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2600 BB->getParent(), BB);
2601 SmallVector<uint64_t, 8> Weights;
2602 bool HasWeights = HasBranchWeights(SI);
2604 GetBranchWeights(SI, Weights);
2605 if (Weights.size() == 1 + SI->getNumCases()) {
2606 // Split weight for default case to case for "Cst".
2607 Weights[0] = (Weights[0]+1) >> 1;
2608 Weights.push_back(Weights[0]);
2610 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2611 SI->setMetadata(LLVMContext::MD_prof,
2612 MDBuilder(SI->getContext()).
2613 createBranchWeights(MDWeights));
2616 SI->addCase(Cst, NewBB);
2618 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2619 Builder.SetInsertPoint(NewBB);
2620 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2621 Builder.CreateBr(SuccBlock);
2622 PHIUse->addIncoming(NewCst, NewBB);
2626 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2627 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2628 /// fold it into a switch instruction if so.
2629 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *TD,
2630 IRBuilder<> &Builder) {
2631 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2632 if (Cond == 0) return false;
2635 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2636 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2637 // 'setne's and'ed together, collect them.
2639 std::vector<ConstantInt*> Values;
2640 bool TrueWhenEqual = true;
2641 Value *ExtraCase = 0;
2642 unsigned UsedICmps = 0;
2644 if (Cond->getOpcode() == Instruction::Or) {
2645 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true,
2647 } else if (Cond->getOpcode() == Instruction::And) {
2648 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false,
2650 TrueWhenEqual = false;
2653 // If we didn't have a multiply compared value, fail.
2654 if (CompVal == 0) return false;
2656 // Avoid turning single icmps into a switch.
2660 // There might be duplicate constants in the list, which the switch
2661 // instruction can't handle, remove them now.
2662 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2663 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2665 // If Extra was used, we require at least two switch values to do the
2666 // transformation. A switch with one value is just an cond branch.
2667 if (ExtraCase && Values.size() < 2) return false;
2669 // TODO: Preserve branch weight metadata, similarly to how
2670 // FoldValueComparisonIntoPredecessors preserves it.
2672 // Figure out which block is which destination.
2673 BasicBlock *DefaultBB = BI->getSuccessor(1);
2674 BasicBlock *EdgeBB = BI->getSuccessor(0);
2675 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2677 BasicBlock *BB = BI->getParent();
2679 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2680 << " cases into SWITCH. BB is:\n" << *BB);
2682 // If there are any extra values that couldn't be folded into the switch
2683 // then we evaluate them with an explicit branch first. Split the block
2684 // right before the condbr to handle it.
2686 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2687 // Remove the uncond branch added to the old block.
2688 TerminatorInst *OldTI = BB->getTerminator();
2689 Builder.SetInsertPoint(OldTI);
2692 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2694 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2696 OldTI->eraseFromParent();
2698 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2699 // for the edge we just added.
2700 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2702 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2703 << "\nEXTRABB = " << *BB);
2707 Builder.SetInsertPoint(BI);
2708 // Convert pointer to int before we switch.
2709 if (CompVal->getType()->isPointerTy()) {
2710 assert(TD && "Cannot switch on pointer without DataLayout");
2711 CompVal = Builder.CreatePtrToInt(CompVal,
2712 TD->getIntPtrType(CompVal->getContext()),
2716 // Create the new switch instruction now.
2717 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2719 // Add all of the 'cases' to the switch instruction.
2720 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2721 New->addCase(Values[i], EdgeBB);
2723 // We added edges from PI to the EdgeBB. As such, if there were any
2724 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2725 // the number of edges added.
2726 for (BasicBlock::iterator BBI = EdgeBB->begin();
2727 isa<PHINode>(BBI); ++BBI) {
2728 PHINode *PN = cast<PHINode>(BBI);
2729 Value *InVal = PN->getIncomingValueForBlock(BB);
2730 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2731 PN->addIncoming(InVal, BB);
2734 // Erase the old branch instruction.
2735 EraseTerminatorInstAndDCECond(BI);
2737 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2741 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2742 // If this is a trivial landing pad that just continues unwinding the caught
2743 // exception then zap the landing pad, turning its invokes into calls.
2744 BasicBlock *BB = RI->getParent();
2745 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2746 if (RI->getValue() != LPInst)
2747 // Not a landing pad, or the resume is not unwinding the exception that
2748 // caused control to branch here.
2751 // Check that there are no other instructions except for debug intrinsics.
2752 BasicBlock::iterator I = LPInst, E = RI;
2754 if (!isa<DbgInfoIntrinsic>(I))
2757 // Turn all invokes that unwind here into calls and delete the basic block.
2758 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2759 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2760 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2761 // Insert a call instruction before the invoke.
2762 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2764 Call->setCallingConv(II->getCallingConv());
2765 Call->setAttributes(II->getAttributes());
2766 Call->setDebugLoc(II->getDebugLoc());
2768 // Anything that used the value produced by the invoke instruction now uses
2769 // the value produced by the call instruction. Note that we do this even
2770 // for void functions and calls with no uses so that the callgraph edge is
2772 II->replaceAllUsesWith(Call);
2773 BB->removePredecessor(II->getParent());
2775 // Insert a branch to the normal destination right before the invoke.
2776 BranchInst::Create(II->getNormalDest(), II);
2778 // Finally, delete the invoke instruction!
2779 II->eraseFromParent();
2782 // The landingpad is now unreachable. Zap it.
2783 BB->eraseFromParent();
2787 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2788 BasicBlock *BB = RI->getParent();
2789 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2791 // Find predecessors that end with branches.
2792 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2793 SmallVector<BranchInst*, 8> CondBranchPreds;
2794 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2795 BasicBlock *P = *PI;
2796 TerminatorInst *PTI = P->getTerminator();
2797 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2798 if (BI->isUnconditional())
2799 UncondBranchPreds.push_back(P);
2801 CondBranchPreds.push_back(BI);
2805 // If we found some, do the transformation!
2806 if (!UncondBranchPreds.empty() && DupRet) {
2807 while (!UncondBranchPreds.empty()) {
2808 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
2809 DEBUG(dbgs() << "FOLDING: " << *BB
2810 << "INTO UNCOND BRANCH PRED: " << *Pred);
2811 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
2814 // If we eliminated all predecessors of the block, delete the block now.
2815 if (pred_begin(BB) == pred_end(BB))
2816 // We know there are no successors, so just nuke the block.
2817 BB->eraseFromParent();
2822 // Check out all of the conditional branches going to this return
2823 // instruction. If any of them just select between returns, change the
2824 // branch itself into a select/return pair.
2825 while (!CondBranchPreds.empty()) {
2826 BranchInst *BI = CondBranchPreds.pop_back_val();
2828 // Check to see if the non-BB successor is also a return block.
2829 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
2830 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
2831 SimplifyCondBranchToTwoReturns(BI, Builder))
2837 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
2838 BasicBlock *BB = UI->getParent();
2840 bool Changed = false;
2842 // If there are any instructions immediately before the unreachable that can
2843 // be removed, do so.
2844 while (UI != BB->begin()) {
2845 BasicBlock::iterator BBI = UI;
2847 // Do not delete instructions that can have side effects which might cause
2848 // the unreachable to not be reachable; specifically, calls and volatile
2849 // operations may have this effect.
2850 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
2852 if (BBI->mayHaveSideEffects()) {
2853 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
2854 if (SI->isVolatile())
2856 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
2857 if (LI->isVolatile())
2859 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
2860 if (RMWI->isVolatile())
2862 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
2863 if (CXI->isVolatile())
2865 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
2866 !isa<LandingPadInst>(BBI)) {
2869 // Note that deleting LandingPad's here is in fact okay, although it
2870 // involves a bit of subtle reasoning. If this inst is a LandingPad,
2871 // all the predecessors of this block will be the unwind edges of Invokes,
2872 // and we can therefore guarantee this block will be erased.
2875 // Delete this instruction (any uses are guaranteed to be dead)
2876 if (!BBI->use_empty())
2877 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
2878 BBI->eraseFromParent();
2882 // If the unreachable instruction is the first in the block, take a gander
2883 // at all of the predecessors of this instruction, and simplify them.
2884 if (&BB->front() != UI) return Changed;
2886 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
2887 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
2888 TerminatorInst *TI = Preds[i]->getTerminator();
2889 IRBuilder<> Builder(TI);
2890 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
2891 if (BI->isUnconditional()) {
2892 if (BI->getSuccessor(0) == BB) {
2893 new UnreachableInst(TI->getContext(), TI);
2894 TI->eraseFromParent();
2898 if (BI->getSuccessor(0) == BB) {
2899 Builder.CreateBr(BI->getSuccessor(1));
2900 EraseTerminatorInstAndDCECond(BI);
2901 } else if (BI->getSuccessor(1) == BB) {
2902 Builder.CreateBr(BI->getSuccessor(0));
2903 EraseTerminatorInstAndDCECond(BI);
2907 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
2908 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2910 if (i.getCaseSuccessor() == BB) {
2911 BB->removePredecessor(SI->getParent());
2916 // If the default value is unreachable, figure out the most popular
2917 // destination and make it the default.
2918 if (SI->getDefaultDest() == BB) {
2919 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
2920 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2922 std::pair<unsigned, unsigned> &entry =
2923 Popularity[i.getCaseSuccessor()];
2924 if (entry.first == 0) {
2926 entry.second = i.getCaseIndex();
2932 // Find the most popular block.
2933 unsigned MaxPop = 0;
2934 unsigned MaxIndex = 0;
2935 BasicBlock *MaxBlock = 0;
2936 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
2937 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
2938 if (I->second.first > MaxPop ||
2939 (I->second.first == MaxPop && MaxIndex > I->second.second)) {
2940 MaxPop = I->second.first;
2941 MaxIndex = I->second.second;
2942 MaxBlock = I->first;
2946 // Make this the new default, allowing us to delete any explicit
2948 SI->setDefaultDest(MaxBlock);
2951 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
2953 if (isa<PHINode>(MaxBlock->begin()))
2954 for (unsigned i = 0; i != MaxPop-1; ++i)
2955 MaxBlock->removePredecessor(SI->getParent());
2957 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2959 if (i.getCaseSuccessor() == MaxBlock) {
2965 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
2966 if (II->getUnwindDest() == BB) {
2967 // Convert the invoke to a call instruction. This would be a good
2968 // place to note that the call does not throw though.
2969 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
2970 II->removeFromParent(); // Take out of symbol table
2972 // Insert the call now...
2973 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
2974 Builder.SetInsertPoint(BI);
2975 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
2976 Args, II->getName());
2977 CI->setCallingConv(II->getCallingConv());
2978 CI->setAttributes(II->getAttributes());
2979 // If the invoke produced a value, the call does now instead.
2980 II->replaceAllUsesWith(CI);
2987 // If this block is now dead, remove it.
2988 if (pred_begin(BB) == pred_end(BB) &&
2989 BB != &BB->getParent()->getEntryBlock()) {
2990 // We know there are no successors, so just nuke the block.
2991 BB->eraseFromParent();
2998 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
2999 /// integer range comparison into a sub, an icmp and a branch.
3000 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3001 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3003 // Make sure all cases point to the same destination and gather the values.
3004 SmallVector<ConstantInt *, 16> Cases;
3005 SwitchInst::CaseIt I = SI->case_begin();
3006 Cases.push_back(I.getCaseValue());
3007 SwitchInst::CaseIt PrevI = I++;
3008 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
3009 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
3011 Cases.push_back(I.getCaseValue());
3013 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
3015 // Sort the case values, then check if they form a range we can transform.
3016 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3017 for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
3018 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
3022 Constant *Offset = ConstantExpr::getNeg(Cases.back());
3023 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
3025 Value *Sub = SI->getCondition();
3026 if (!Offset->isNullValue())
3027 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
3028 Value *Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3029 BranchInst *NewBI = Builder.CreateCondBr(
3030 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
3032 // Update weight for the newly-created conditional branch.
3033 SmallVector<uint64_t, 8> Weights;
3034 bool HasWeights = HasBranchWeights(SI);
3036 GetBranchWeights(SI, Weights);
3037 if (Weights.size() == 1 + SI->getNumCases()) {
3038 // Combine all weights for the cases to be the true weight of NewBI.
3039 // We assume that the sum of all weights for a Terminator can fit into 32
3041 uint32_t NewTrueWeight = 0;
3042 for (unsigned I = 1, E = Weights.size(); I != E; ++I)
3043 NewTrueWeight += (uint32_t)Weights[I];
3044 NewBI->setMetadata(LLVMContext::MD_prof,
3045 MDBuilder(SI->getContext()).
3046 createBranchWeights(NewTrueWeight,
3047 (uint32_t)Weights[0]));
3051 // Prune obsolete incoming values off the successor's PHI nodes.
3052 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
3053 isa<PHINode>(BBI); ++BBI) {
3054 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
3055 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3057 SI->eraseFromParent();
3062 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
3063 /// and use it to remove dead cases.
3064 static bool EliminateDeadSwitchCases(SwitchInst *SI) {
3065 Value *Cond = SI->getCondition();
3066 unsigned Bits = cast<IntegerType>(Cond->getType())->getBitWidth();
3067 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3068 ComputeMaskedBits(Cond, KnownZero, KnownOne);
3070 // Gather dead cases.
3071 SmallVector<ConstantInt*, 8> DeadCases;
3072 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3073 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3074 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3075 DeadCases.push_back(I.getCaseValue());
3076 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3077 << I.getCaseValue() << "' is dead.\n");
3081 SmallVector<uint64_t, 8> Weights;
3082 bool HasWeight = HasBranchWeights(SI);
3084 GetBranchWeights(SI, Weights);
3085 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3088 // Remove dead cases from the switch.
3089 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3090 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3091 assert(Case != SI->case_default() &&
3092 "Case was not found. Probably mistake in DeadCases forming.");
3094 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3098 // Prune unused values from PHI nodes.
3099 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3100 SI->removeCase(Case);
3103 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3104 SI->setMetadata(LLVMContext::MD_prof,
3105 MDBuilder(SI->getParent()->getContext()).
3106 createBranchWeights(MDWeights));
3109 return !DeadCases.empty();
3112 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
3113 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3114 /// by an unconditional branch), look at the phi node for BB in the successor
3115 /// block and see if the incoming value is equal to CaseValue. If so, return
3116 /// the phi node, and set PhiIndex to BB's index in the phi node.
3117 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3120 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3121 return NULL; // BB must be empty to be a candidate for simplification.
3122 if (!BB->getSinglePredecessor())
3123 return NULL; // BB must be dominated by the switch.
3125 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3126 if (!Branch || !Branch->isUnconditional())
3127 return NULL; // Terminator must be unconditional branch.
3129 BasicBlock *Succ = Branch->getSuccessor(0);
3131 BasicBlock::iterator I = Succ->begin();
3132 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3133 int Idx = PHI->getBasicBlockIndex(BB);
3134 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3136 Value *InValue = PHI->getIncomingValue(Idx);
3137 if (InValue != CaseValue) continue;
3146 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
3147 /// instruction to a phi node dominated by the switch, if that would mean that
3148 /// some of the destination blocks of the switch can be folded away.
3149 /// Returns true if a change is made.
3150 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3151 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3152 ForwardingNodesMap ForwardingNodes;
3154 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3155 ConstantInt *CaseValue = I.getCaseValue();
3156 BasicBlock *CaseDest = I.getCaseSuccessor();
3159 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3163 ForwardingNodes[PHI].push_back(PhiIndex);
3166 bool Changed = false;
3168 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3169 E = ForwardingNodes.end(); I != E; ++I) {
3170 PHINode *Phi = I->first;
3171 SmallVector<int,4> &Indexes = I->second;
3173 if (Indexes.size() < 2) continue;
3175 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3176 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3183 /// ValidLookupTableConstant - Return true if the backend will be able to handle
3184 /// initializing an array of constants like C.
3185 static bool ValidLookupTableConstant(Constant *C) {
3186 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3187 return CE->isGEPWithNoNotionalOverIndexing();
3189 return isa<ConstantFP>(C) ||
3190 isa<ConstantInt>(C) ||
3191 isa<ConstantPointerNull>(C) ||
3192 isa<GlobalValue>(C) ||
3196 /// GetCaseResulsts - Try to determine the resulting constant values in phi
3197 /// nodes at the common destination basic block for one of the case
3198 /// destinations of a switch instruction.
3199 static bool GetCaseResults(SwitchInst *SI,
3200 BasicBlock *CaseDest,
3201 BasicBlock **CommonDest,
3202 SmallVector<std::pair<PHINode*,Constant*>, 4> &Res) {
3203 // The block from which we enter the common destination.
3204 BasicBlock *Pred = SI->getParent();
3206 // If CaseDest is empty, continue to its successor.
3207 if (CaseDest->getFirstNonPHIOrDbg() == CaseDest->getTerminator() &&
3208 !isa<PHINode>(CaseDest->begin())) {
3210 TerminatorInst *Terminator = CaseDest->getTerminator();
3211 if (Terminator->getNumSuccessors() != 1)
3215 CaseDest = Terminator->getSuccessor(0);
3218 // If we did not have a CommonDest before, use the current one.
3220 *CommonDest = CaseDest;
3221 // If the destination isn't the common one, abort.
3222 if (CaseDest != *CommonDest)
3225 // Get the values for this case from phi nodes in the destination block.
3226 BasicBlock::iterator I = (*CommonDest)->begin();
3227 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3228 int Idx = PHI->getBasicBlockIndex(Pred);
3232 Constant *ConstVal = dyn_cast<Constant>(PHI->getIncomingValue(Idx));
3236 // Be conservative about which kinds of constants we support.
3237 if (!ValidLookupTableConstant(ConstVal))
3240 Res.push_back(std::make_pair(PHI, ConstVal));
3247 /// SwitchLookupTable - This class represents a lookup table that can be used
3248 /// to replace a switch.
3249 class SwitchLookupTable {
3251 /// SwitchLookupTable - Create a lookup table to use as a switch replacement
3252 /// with the contents of Values, using DefaultValue to fill any holes in the
3254 SwitchLookupTable(Module &M,
3256 ConstantInt *Offset,
3257 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values,
3258 Constant *DefaultValue,
3259 const DataLayout *TD);
3261 /// BuildLookup - Build instructions with Builder to retrieve the value at
3262 /// the position given by Index in the lookup table.
3263 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3265 /// WouldFitInRegister - Return true if a table with TableSize elements of
3266 /// type ElementType would fit in a target-legal register.
3267 static bool WouldFitInRegister(const DataLayout *TD,
3269 const Type *ElementType);
3272 // Depending on the contents of the table, it can be represented in
3275 // For tables where each element contains the same value, we just have to
3276 // store that single value and return it for each lookup.
3279 // For small tables with integer elements, we can pack them into a bitmap
3280 // that fits into a target-legal register. Values are retrieved by
3281 // shift and mask operations.
3284 // The table is stored as an array of values. Values are retrieved by load
3285 // instructions from the table.
3289 // For SingleValueKind, this is the single value.
3290 Constant *SingleValue;
3292 // For BitMapKind, this is the bitmap.
3293 ConstantInt *BitMap;
3294 IntegerType *BitMapElementTy;
3296 // For ArrayKind, this is the array.
3297 GlobalVariable *Array;
3301 SwitchLookupTable::SwitchLookupTable(Module &M,
3303 ConstantInt *Offset,
3304 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values,
3305 Constant *DefaultValue,
3306 const DataLayout *TD) {
3307 assert(Values.size() && "Can't build lookup table without values!");
3308 assert(TableSize >= Values.size() && "Can't fit values in table!");
3310 // If all values in the table are equal, this is that value.
3311 SingleValue = Values.begin()->second;
3313 // Build up the table contents.
3314 SmallVector<Constant*, 64> TableContents(TableSize);
3315 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3316 ConstantInt *CaseVal = Values[I].first;
3317 Constant *CaseRes = Values[I].second;
3318 assert(CaseRes->getType() == DefaultValue->getType());
3320 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3322 TableContents[Idx] = CaseRes;
3324 if (CaseRes != SingleValue)
3328 // Fill in any holes in the table with the default result.
3329 if (Values.size() < TableSize) {
3330 for (uint64_t I = 0; I < TableSize; ++I) {
3331 if (!TableContents[I])
3332 TableContents[I] = DefaultValue;
3335 if (DefaultValue != SingleValue)
3339 // If each element in the table contains the same value, we only need to store
3340 // that single value.
3342 Kind = SingleValueKind;
3346 // If the type is integer and the table fits in a register, build a bitmap.
3347 if (WouldFitInRegister(TD, TableSize, DefaultValue->getType())) {
3348 IntegerType *IT = cast<IntegerType>(DefaultValue->getType());
3349 APInt TableInt(TableSize * IT->getBitWidth(), 0);
3350 for (uint64_t I = TableSize; I > 0; --I) {
3351 TableInt <<= IT->getBitWidth();
3352 // Insert values into the bitmap. Undef values are set to zero.
3353 if (!isa<UndefValue>(TableContents[I - 1])) {
3354 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
3355 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
3358 BitMap = ConstantInt::get(M.getContext(), TableInt);
3359 BitMapElementTy = IT;
3365 // Store the table in an array.
3366 ArrayType *ArrayTy = ArrayType::get(DefaultValue->getType(), TableSize);
3367 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3369 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3370 GlobalVariable::PrivateLinkage,
3373 Array->setUnnamedAddr(true);
3377 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
3379 case SingleValueKind:
3382 // Type of the bitmap (e.g. i59).
3383 IntegerType *MapTy = BitMap->getType();
3385 // Cast Index to the same type as the bitmap.
3386 // Note: The Index is <= the number of elements in the table, so
3387 // truncating it to the width of the bitmask is safe.
3388 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
3390 // Multiply the shift amount by the element width.
3391 ShiftAmt = Builder.CreateMul(ShiftAmt,
3392 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
3396 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
3397 "switch.downshift");
3399 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
3403 Value *GEPIndices[] = { Builder.getInt32(0), Index };
3404 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices,
3406 return Builder.CreateLoad(GEP, "switch.load");
3409 llvm_unreachable("Unknown lookup table kind!");
3412 bool SwitchLookupTable::WouldFitInRegister(const DataLayout *TD,
3414 const Type *ElementType) {
3417 const IntegerType *IT = dyn_cast<IntegerType>(ElementType);
3420 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
3421 // are <= 15, we could try to narrow the type.
3423 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
3424 if (TableSize >= UINT_MAX/IT->getBitWidth())
3426 return TD->fitsInLegalInteger(TableSize * IT->getBitWidth());
3429 /// ShouldBuildLookupTable - Determine whether a lookup table should be built
3430 /// for this switch, based on the number of caes, size of the table and the
3431 /// types of the results.
3432 static bool ShouldBuildLookupTable(SwitchInst *SI,
3434 const DataLayout *TD,
3435 const SmallDenseMap<PHINode*, Type*>& ResultTypes) {
3436 // The table density should be at least 40%. This is the same criterion as for
3437 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3438 // FIXME: Find the best cut-off.
3439 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
3440 return false; // TableSize overflowed, or mul below might overflow.
3441 if (SI->getNumCases() * 10 >= TableSize * 4)
3444 // If each table would fit in a register, we should build it anyway.
3445 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(),
3446 E = ResultTypes.end(); I != E; ++I) {
3447 if (!SwitchLookupTable::WouldFitInRegister(TD, TableSize, I->second))
3453 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3454 /// phi nodes in a common successor block with different constant values,
3455 /// replace the switch with lookup tables.
3456 static bool SwitchToLookupTable(SwitchInst *SI,
3457 IRBuilder<> &Builder,
3458 const DataLayout* TD) {
3459 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3460 // FIXME: Handle unreachable cases.
3462 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3463 // split off a dense part and build a lookup table for that.
3465 // FIXME: This creates arrays of GEPs to constant strings, which means each
3466 // GEP needs a runtime relocation in PIC code. We should just build one big
3467 // string and lookup indices into that.
3469 // Ignore the switch if the number of cases is too small.
3470 // This is similar to the check when building jump tables in
3471 // SelectionDAGBuilder::handleJTSwitchCase.
3472 // FIXME: Determine the best cut-off.
3473 if (SI->getNumCases() < 4)
3476 // Figure out the corresponding result for each case value and phi node in the
3477 // common destination, as well as the the min and max case values.
3478 assert(SI->case_begin() != SI->case_end());
3479 SwitchInst::CaseIt CI = SI->case_begin();
3480 ConstantInt *MinCaseVal = CI.getCaseValue();
3481 ConstantInt *MaxCaseVal = CI.getCaseValue();
3483 BasicBlock *CommonDest = NULL;
3484 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
3485 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
3486 SmallDenseMap<PHINode*, Constant*> DefaultResults;
3487 SmallDenseMap<PHINode*, Type*> ResultTypes;
3488 SmallVector<PHINode*, 4> PHIs;
3490 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
3491 ConstantInt *CaseVal = CI.getCaseValue();
3492 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
3493 MinCaseVal = CaseVal;
3494 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
3495 MaxCaseVal = CaseVal;
3497 // Resulting value at phi nodes for this case value.
3498 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
3500 if (!GetCaseResults(SI, CI.getCaseSuccessor(), &CommonDest, Results))
3503 // Append the result from this case to the list for each phi.
3504 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) {
3505 if (!ResultLists.count(I->first))
3506 PHIs.push_back(I->first);
3507 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second));
3511 // Get the resulting values for the default case.
3512 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
3513 if (!GetCaseResults(SI, SI->getDefaultDest(), &CommonDest, DefaultResultsList))
3515 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) {
3516 PHINode *PHI = DefaultResultsList[I].first;
3517 Constant *Result = DefaultResultsList[I].second;
3518 DefaultResults[PHI] = Result;
3519 ResultTypes[PHI] = Result->getType();
3522 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
3523 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
3524 if (!ShouldBuildLookupTable(SI, TableSize, TD, ResultTypes))
3527 // Create the BB that does the lookups.
3528 Module &Mod = *CommonDest->getParent()->getParent();
3529 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
3531 CommonDest->getParent(),
3534 // Check whether the condition value is within the case range, and branch to
3536 Builder.SetInsertPoint(SI);
3537 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
3539 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
3540 MinCaseVal->getType(), TableSize));
3541 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
3543 // Populate the BB that does the lookups.
3544 Builder.SetInsertPoint(LookupBB);
3545 bool ReturnedEarly = false;
3546 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
3547 PHINode *PHI = PHIs[I];
3549 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI],
3550 DefaultResults[PHI], TD);
3552 Value *Result = Table.BuildLookup(TableIndex, Builder);
3554 // If the result is used to return immediately from the function, we want to
3555 // do that right here.
3556 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin()) &&
3557 *PHI->use_begin() == CommonDest->getFirstNonPHIOrDbg()) {
3558 Builder.CreateRet(Result);
3559 ReturnedEarly = true;
3563 PHI->addIncoming(Result, LookupBB);
3567 Builder.CreateBr(CommonDest);
3569 // Remove the switch.
3570 for (unsigned i = 0; i < SI->getNumSuccessors(); ++i) {
3571 BasicBlock *Succ = SI->getSuccessor(i);
3572 if (Succ == SI->getDefaultDest()) continue;
3573 Succ->removePredecessor(SI->getParent());
3575 SI->eraseFromParent();
3581 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
3582 // If this switch is too complex to want to look at, ignore it.
3583 if (!isValueEqualityComparison(SI))
3586 BasicBlock *BB = SI->getParent();
3588 // If we only have one predecessor, and if it is a branch on this value,
3589 // see if that predecessor totally determines the outcome of this switch.
3590 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3591 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
3592 return SimplifyCFG(BB) | true;
3594 Value *Cond = SI->getCondition();
3595 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
3596 if (SimplifySwitchOnSelect(SI, Select))
3597 return SimplifyCFG(BB) | true;
3599 // If the block only contains the switch, see if we can fold the block
3600 // away into any preds.
3601 BasicBlock::iterator BBI = BB->begin();
3602 // Ignore dbg intrinsics.
3603 while (isa<DbgInfoIntrinsic>(BBI))
3606 if (FoldValueComparisonIntoPredecessors(SI, Builder))
3607 return SimplifyCFG(BB) | true;
3609 // Try to transform the switch into an icmp and a branch.
3610 if (TurnSwitchRangeIntoICmp(SI, Builder))
3611 return SimplifyCFG(BB) | true;
3613 // Remove unreachable cases.
3614 if (EliminateDeadSwitchCases(SI))
3615 return SimplifyCFG(BB) | true;
3617 if (ForwardSwitchConditionToPHI(SI))
3618 return SimplifyCFG(BB) | true;
3620 if (SwitchToLookupTable(SI, Builder, TD))
3621 return SimplifyCFG(BB) | true;
3626 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
3627 BasicBlock *BB = IBI->getParent();
3628 bool Changed = false;
3630 // Eliminate redundant destinations.
3631 SmallPtrSet<Value *, 8> Succs;
3632 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
3633 BasicBlock *Dest = IBI->getDestination(i);
3634 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) {
3635 Dest->removePredecessor(BB);
3636 IBI->removeDestination(i);
3642 if (IBI->getNumDestinations() == 0) {
3643 // If the indirectbr has no successors, change it to unreachable.
3644 new UnreachableInst(IBI->getContext(), IBI);
3645 EraseTerminatorInstAndDCECond(IBI);
3649 if (IBI->getNumDestinations() == 1) {
3650 // If the indirectbr has one successor, change it to a direct branch.
3651 BranchInst::Create(IBI->getDestination(0), IBI);
3652 EraseTerminatorInstAndDCECond(IBI);
3656 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
3657 if (SimplifyIndirectBrOnSelect(IBI, SI))
3658 return SimplifyCFG(BB) | true;
3663 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
3664 BasicBlock *BB = BI->getParent();
3666 if (SinkCommon && SinkThenElseCodeToEnd(BI))
3669 // If the Terminator is the only non-phi instruction, simplify the block.
3670 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime();
3671 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
3672 TryToSimplifyUncondBranchFromEmptyBlock(BB))
3675 // If the only instruction in the block is a seteq/setne comparison
3676 // against a constant, try to simplify the block.
3677 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
3678 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
3679 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
3681 if (I->isTerminator() &&
3682 TryToSimplifyUncondBranchWithICmpInIt(ICI, TD, Builder))
3686 // If this basic block is ONLY a compare and a branch, and if a predecessor
3687 // branches to us and our successor, fold the comparison into the
3688 // predecessor and use logical operations to update the incoming value
3689 // for PHI nodes in common successor.
3690 if (FoldBranchToCommonDest(BI))
3691 return SimplifyCFG(BB) | true;
3696 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
3697 BasicBlock *BB = BI->getParent();
3699 // Conditional branch
3700 if (isValueEqualityComparison(BI)) {
3701 // If we only have one predecessor, and if it is a branch on this value,
3702 // see if that predecessor totally determines the outcome of this
3704 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3705 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
3706 return SimplifyCFG(BB) | true;
3708 // This block must be empty, except for the setcond inst, if it exists.
3709 // Ignore dbg intrinsics.
3710 BasicBlock::iterator I = BB->begin();
3711 // Ignore dbg intrinsics.
3712 while (isa<DbgInfoIntrinsic>(I))
3715 if (FoldValueComparisonIntoPredecessors(BI, Builder))
3716 return SimplifyCFG(BB) | true;
3717 } else if (&*I == cast<Instruction>(BI->getCondition())){
3719 // Ignore dbg intrinsics.
3720 while (isa<DbgInfoIntrinsic>(I))
3722 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
3723 return SimplifyCFG(BB) | true;
3727 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
3728 if (SimplifyBranchOnICmpChain(BI, TD, Builder))
3731 // If this basic block is ONLY a compare and a branch, and if a predecessor
3732 // branches to us and one of our successors, fold the comparison into the
3733 // predecessor and use logical operations to pick the right destination.
3734 if (FoldBranchToCommonDest(BI))
3735 return SimplifyCFG(BB) | true;
3737 // We have a conditional branch to two blocks that are only reachable
3738 // from BI. We know that the condbr dominates the two blocks, so see if
3739 // there is any identical code in the "then" and "else" blocks. If so, we
3740 // can hoist it up to the branching block.
3741 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) {
3742 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3743 if (HoistThenElseCodeToIf(BI))
3744 return SimplifyCFG(BB) | true;
3746 // If Successor #1 has multiple preds, we may be able to conditionally
3747 // execute Successor #0 if it branches to successor #1.
3748 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
3749 if (Succ0TI->getNumSuccessors() == 1 &&
3750 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
3751 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0)))
3752 return SimplifyCFG(BB) | true;
3754 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3755 // If Successor #0 has multiple preds, we may be able to conditionally
3756 // execute Successor #1 if it branches to successor #0.
3757 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
3758 if (Succ1TI->getNumSuccessors() == 1 &&
3759 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
3760 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1)))
3761 return SimplifyCFG(BB) | true;
3764 // If this is a branch on a phi node in the current block, thread control
3765 // through this block if any PHI node entries are constants.
3766 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
3767 if (PN->getParent() == BI->getParent())
3768 if (FoldCondBranchOnPHI(BI, TD))
3769 return SimplifyCFG(BB) | true;
3771 // Scan predecessor blocks for conditional branches.
3772 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
3773 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
3774 if (PBI != BI && PBI->isConditional())
3775 if (SimplifyCondBranchToCondBranch(PBI, BI))
3776 return SimplifyCFG(BB) | true;
3781 /// Check if passing a value to an instruction will cause undefined behavior.
3782 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
3783 Constant *C = dyn_cast<Constant>(V);
3790 if (C->isNullValue()) {
3791 // Only look at the first use, avoid hurting compile time with long uselists
3792 User *Use = *I->use_begin();
3794 // Now make sure that there are no instructions in between that can alter
3795 // control flow (eg. calls)
3796 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
3797 if (i == I->getParent()->end() || i->mayHaveSideEffects())
3800 // Look through GEPs. A load from a GEP derived from NULL is still undefined
3801 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
3802 if (GEP->getPointerOperand() == I)
3803 return passingValueIsAlwaysUndefined(V, GEP);
3805 // Look through bitcasts.
3806 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
3807 return passingValueIsAlwaysUndefined(V, BC);
3809 // Load from null is undefined.
3810 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
3811 return LI->getPointerAddressSpace() == 0;
3813 // Store to null is undefined.
3814 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
3815 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
3820 /// If BB has an incoming value that will always trigger undefined behavior
3821 /// (eg. null pointer dereference), remove the branch leading here.
3822 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
3823 for (BasicBlock::iterator i = BB->begin();
3824 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
3825 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
3826 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
3827 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
3828 IRBuilder<> Builder(T);
3829 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
3830 BB->removePredecessor(PHI->getIncomingBlock(i));
3831 // Turn uncoditional branches into unreachables and remove the dead
3832 // destination from conditional branches.
3833 if (BI->isUnconditional())
3834 Builder.CreateUnreachable();
3836 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
3837 BI->getSuccessor(0));
3838 BI->eraseFromParent();
3841 // TODO: SwitchInst.
3847 bool SimplifyCFGOpt::run(BasicBlock *BB) {
3848 bool Changed = false;
3850 assert(BB && BB->getParent() && "Block not embedded in function!");
3851 assert(BB->getTerminator() && "Degenerate basic block encountered!");
3853 // Remove basic blocks that have no predecessors (except the entry block)...
3854 // or that just have themself as a predecessor. These are unreachable.
3855 if ((pred_begin(BB) == pred_end(BB) &&
3856 BB != &BB->getParent()->getEntryBlock()) ||
3857 BB->getSinglePredecessor() == BB) {
3858 DEBUG(dbgs() << "Removing BB: \n" << *BB);
3859 DeleteDeadBlock(BB);
3863 // Check to see if we can constant propagate this terminator instruction
3865 Changed |= ConstantFoldTerminator(BB, true);
3867 // Check for and eliminate duplicate PHI nodes in this block.
3868 Changed |= EliminateDuplicatePHINodes(BB);
3870 // Check for and remove branches that will always cause undefined behavior.
3871 Changed |= removeUndefIntroducingPredecessor(BB);
3873 // Merge basic blocks into their predecessor if there is only one distinct
3874 // pred, and if there is only one distinct successor of the predecessor, and
3875 // if there are no PHI nodes.
3877 if (MergeBlockIntoPredecessor(BB))
3880 IRBuilder<> Builder(BB);
3882 // If there is a trivial two-entry PHI node in this basic block, and we can
3883 // eliminate it, do so now.
3884 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
3885 if (PN->getNumIncomingValues() == 2)
3886 Changed |= FoldTwoEntryPHINode(PN, TD);
3888 Builder.SetInsertPoint(BB->getTerminator());
3889 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
3890 if (BI->isUnconditional()) {
3891 if (SimplifyUncondBranch(BI, Builder)) return true;
3893 if (SimplifyCondBranch(BI, Builder)) return true;
3895 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
3896 if (SimplifyReturn(RI, Builder)) return true;
3897 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
3898 if (SimplifyResume(RI, Builder)) return true;
3899 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
3900 if (SimplifySwitch(SI, Builder)) return true;
3901 } else if (UnreachableInst *UI =
3902 dyn_cast<UnreachableInst>(BB->getTerminator())) {
3903 if (SimplifyUnreachable(UI)) return true;
3904 } else if (IndirectBrInst *IBI =
3905 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
3906 if (SimplifyIndirectBr(IBI)) return true;
3912 /// SimplifyCFG - This function is used to do simplification of a CFG. For
3913 /// example, it adjusts branches to branches to eliminate the extra hop, it
3914 /// eliminates unreachable basic blocks, and does other "peephole" optimization
3915 /// of the CFG. It returns true if a modification was made.
3917 bool llvm::SimplifyCFG(BasicBlock *BB, const DataLayout *TD) {
3918 return SimplifyCFGOpt(TD).run(BB);