1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Peephole optimize the CFG.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "simplifycfg"
15 #include "llvm/Transforms/Utils/Local.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SetVector.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/Analysis/InstructionSimplify.h"
23 #include "llvm/Analysis/TargetTransformInfo.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/GlobalVariable.h"
29 #include "llvm/IR/IRBuilder.h"
30 #include "llvm/IR/Instructions.h"
31 #include "llvm/IR/IntrinsicInst.h"
32 #include "llvm/IR/LLVMContext.h"
33 #include "llvm/IR/MDBuilder.h"
34 #include "llvm/IR/Metadata.h"
35 #include "llvm/IR/Module.h"
36 #include "llvm/IR/Operator.h"
37 #include "llvm/IR/Type.h"
38 #include "llvm/Support/CFG.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Support/ConstantRange.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/NoFolder.h"
43 #include "llvm/Support/raw_ostream.h"
44 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
50 static cl::opt<unsigned>
51 PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1),
52 cl::desc("Control the amount of phi node folding to perform (default = 1)"));
55 DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false),
56 cl::desc("Duplicate return instructions into unconditional branches"));
59 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
60 cl::desc("Sink common instructions down to the end block"));
62 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
63 STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables");
64 STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block");
65 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
68 /// ValueEqualityComparisonCase - Represents a case of a switch.
69 struct ValueEqualityComparisonCase {
73 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
74 : Value(Value), Dest(Dest) {}
76 bool operator<(ValueEqualityComparisonCase RHS) const {
77 // Comparing pointers is ok as we only rely on the order for uniquing.
78 return Value < RHS.Value;
81 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
84 class SimplifyCFGOpt {
85 const TargetTransformInfo &TTI;
86 const DataLayout *const TD;
88 Value *isValueEqualityComparison(TerminatorInst *TI);
89 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
90 std::vector<ValueEqualityComparisonCase> &Cases);
91 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
93 IRBuilder<> &Builder);
94 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
95 IRBuilder<> &Builder);
97 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
98 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
99 bool SimplifyUnreachable(UnreachableInst *UI);
100 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
101 bool SimplifyIndirectBr(IndirectBrInst *IBI);
102 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder);
103 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder);
106 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout *TD)
107 : TTI(TTI), TD(TD) {}
108 bool run(BasicBlock *BB);
112 /// SafeToMergeTerminators - Return true if it is safe to merge these two
113 /// terminator instructions together.
115 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
116 if (SI1 == SI2) return false; // Can't merge with self!
118 // It is not safe to merge these two switch instructions if they have a common
119 // successor, and if that successor has a PHI node, and if *that* PHI node has
120 // conflicting incoming values from the two switch blocks.
121 BasicBlock *SI1BB = SI1->getParent();
122 BasicBlock *SI2BB = SI2->getParent();
123 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
125 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
126 if (SI1Succs.count(*I))
127 for (BasicBlock::iterator BBI = (*I)->begin();
128 isa<PHINode>(BBI); ++BBI) {
129 PHINode *PN = cast<PHINode>(BBI);
130 if (PN->getIncomingValueForBlock(SI1BB) !=
131 PN->getIncomingValueForBlock(SI2BB))
138 /// isProfitableToFoldUnconditional - Return true if it is safe and profitable
139 /// to merge these two terminator instructions together, where SI1 is an
140 /// unconditional branch. PhiNodes will store all PHI nodes in common
143 static bool isProfitableToFoldUnconditional(BranchInst *SI1,
146 SmallVectorImpl<PHINode*> &PhiNodes) {
147 if (SI1 == SI2) return false; // Can't merge with self!
148 assert(SI1->isUnconditional() && SI2->isConditional());
150 // We fold the unconditional branch if we can easily update all PHI nodes in
151 // common successors:
152 // 1> We have a constant incoming value for the conditional branch;
153 // 2> We have "Cond" as the incoming value for the unconditional branch;
154 // 3> SI2->getCondition() and Cond have same operands.
155 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
156 if (!Ci2) return false;
157 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
158 Cond->getOperand(1) == Ci2->getOperand(1)) &&
159 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
160 Cond->getOperand(1) == Ci2->getOperand(0)))
163 BasicBlock *SI1BB = SI1->getParent();
164 BasicBlock *SI2BB = SI2->getParent();
165 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
166 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
167 if (SI1Succs.count(*I))
168 for (BasicBlock::iterator BBI = (*I)->begin();
169 isa<PHINode>(BBI); ++BBI) {
170 PHINode *PN = cast<PHINode>(BBI);
171 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
172 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
174 PhiNodes.push_back(PN);
179 /// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will
180 /// now be entries in it from the 'NewPred' block. The values that will be
181 /// flowing into the PHI nodes will be the same as those coming in from
182 /// ExistPred, an existing predecessor of Succ.
183 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
184 BasicBlock *ExistPred) {
185 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
188 for (BasicBlock::iterator I = Succ->begin();
189 (PN = dyn_cast<PHINode>(I)); ++I)
190 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
194 /// GetIfCondition - Given a basic block (BB) with two predecessors (and at
195 /// least one PHI node in it), check to see if the merge at this block is due
196 /// to an "if condition". If so, return the boolean condition that determines
197 /// which entry into BB will be taken. Also, return by references the block
198 /// that will be entered from if the condition is true, and the block that will
199 /// be entered if the condition is false.
201 /// This does no checking to see if the true/false blocks have large or unsavory
202 /// instructions in them.
203 static Value *GetIfCondition(BasicBlock *BB, BasicBlock *&IfTrue,
204 BasicBlock *&IfFalse) {
205 PHINode *SomePHI = cast<PHINode>(BB->begin());
206 assert(SomePHI->getNumIncomingValues() == 2 &&
207 "Function can only handle blocks with 2 predecessors!");
208 BasicBlock *Pred1 = SomePHI->getIncomingBlock(0);
209 BasicBlock *Pred2 = SomePHI->getIncomingBlock(1);
211 // We can only handle branches. Other control flow will be lowered to
212 // branches if possible anyway.
213 BranchInst *Pred1Br = dyn_cast<BranchInst>(Pred1->getTerminator());
214 BranchInst *Pred2Br = dyn_cast<BranchInst>(Pred2->getTerminator());
215 if (Pred1Br == 0 || Pred2Br == 0)
218 // Eliminate code duplication by ensuring that Pred1Br is conditional if
220 if (Pred2Br->isConditional()) {
221 // If both branches are conditional, we don't have an "if statement". In
222 // reality, we could transform this case, but since the condition will be
223 // required anyway, we stand no chance of eliminating it, so the xform is
224 // probably not profitable.
225 if (Pred1Br->isConditional())
228 std::swap(Pred1, Pred2);
229 std::swap(Pred1Br, Pred2Br);
232 if (Pred1Br->isConditional()) {
233 // The only thing we have to watch out for here is to make sure that Pred2
234 // doesn't have incoming edges from other blocks. If it does, the condition
235 // doesn't dominate BB.
236 if (Pred2->getSinglePredecessor() == 0)
239 // If we found a conditional branch predecessor, make sure that it branches
240 // to BB and Pred2Br. If it doesn't, this isn't an "if statement".
241 if (Pred1Br->getSuccessor(0) == BB &&
242 Pred1Br->getSuccessor(1) == Pred2) {
245 } else if (Pred1Br->getSuccessor(0) == Pred2 &&
246 Pred1Br->getSuccessor(1) == BB) {
250 // We know that one arm of the conditional goes to BB, so the other must
251 // go somewhere unrelated, and this must not be an "if statement".
255 return Pred1Br->getCondition();
258 // Ok, if we got here, both predecessors end with an unconditional branch to
259 // BB. Don't panic! If both blocks only have a single (identical)
260 // predecessor, and THAT is a conditional branch, then we're all ok!
261 BasicBlock *CommonPred = Pred1->getSinglePredecessor();
262 if (CommonPred == 0 || CommonPred != Pred2->getSinglePredecessor())
265 // Otherwise, if this is a conditional branch, then we can use it!
266 BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator());
267 if (BI == 0) return 0;
269 assert(BI->isConditional() && "Two successors but not conditional?");
270 if (BI->getSuccessor(0) == Pred1) {
277 return BI->getCondition();
280 /// ComputeSpeculuationCost - Compute an abstract "cost" of speculating the
281 /// given instruction, which is assumed to be safe to speculate. 1 means
282 /// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive.
283 static unsigned ComputeSpeculationCost(const User *I) {
284 assert(isSafeToSpeculativelyExecute(I) &&
285 "Instruction is not safe to speculatively execute!");
286 switch (Operator::getOpcode(I)) {
288 // In doubt, be conservative.
290 case Instruction::GetElementPtr:
291 // GEPs are cheap if all indices are constant.
292 if (!cast<GEPOperator>(I)->hasAllConstantIndices())
295 case Instruction::Load:
296 case Instruction::Add:
297 case Instruction::Sub:
298 case Instruction::And:
299 case Instruction::Or:
300 case Instruction::Xor:
301 case Instruction::Shl:
302 case Instruction::LShr:
303 case Instruction::AShr:
304 case Instruction::ICmp:
305 case Instruction::Trunc:
306 case Instruction::ZExt:
307 case Instruction::SExt:
308 return 1; // These are all cheap.
310 case Instruction::Call:
311 case Instruction::Select:
316 /// DominatesMergePoint - If we have a merge point of an "if condition" as
317 /// accepted above, return true if the specified value dominates the block. We
318 /// don't handle the true generality of domination here, just a special case
319 /// which works well enough for us.
321 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
322 /// see if V (which must be an instruction) and its recursive operands
323 /// that do not dominate BB have a combined cost lower than CostRemaining and
324 /// are non-trapping. If both are true, the instruction is inserted into the
325 /// set and true is returned.
327 /// The cost for most non-trapping instructions is defined as 1 except for
328 /// Select whose cost is 2.
330 /// After this function returns, CostRemaining is decreased by the cost of
331 /// V plus its non-dominating operands. If that cost is greater than
332 /// CostRemaining, false is returned and CostRemaining is undefined.
333 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
334 SmallPtrSet<Instruction*, 4> *AggressiveInsts,
335 unsigned &CostRemaining) {
336 Instruction *I = dyn_cast<Instruction>(V);
338 // Non-instructions all dominate instructions, but not all constantexprs
339 // can be executed unconditionally.
340 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
345 BasicBlock *PBB = I->getParent();
347 // We don't want to allow weird loops that might have the "if condition" in
348 // the bottom of this block.
349 if (PBB == BB) return false;
351 // If this instruction is defined in a block that contains an unconditional
352 // branch to BB, then it must be in the 'conditional' part of the "if
353 // statement". If not, it definitely dominates the region.
354 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
355 if (BI == 0 || BI->isConditional() || BI->getSuccessor(0) != BB)
358 // If we aren't allowing aggressive promotion anymore, then don't consider
359 // instructions in the 'if region'.
360 if (AggressiveInsts == 0) return false;
362 // If we have seen this instruction before, don't count it again.
363 if (AggressiveInsts->count(I)) return true;
365 // Okay, it looks like the instruction IS in the "condition". Check to
366 // see if it's a cheap instruction to unconditionally compute, and if it
367 // only uses stuff defined outside of the condition. If so, hoist it out.
368 if (!isSafeToSpeculativelyExecute(I))
371 unsigned Cost = ComputeSpeculationCost(I);
373 if (Cost > CostRemaining)
376 CostRemaining -= Cost;
378 // Okay, we can only really hoist these out if their operands do
379 // not take us over the cost threshold.
380 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
381 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining))
383 // Okay, it's safe to do this! Remember this instruction.
384 AggressiveInsts->insert(I);
388 /// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr
389 /// and PointerNullValue. Return NULL if value is not a constant int.
390 static ConstantInt *GetConstantInt(Value *V, const DataLayout *TD) {
391 // Normal constant int.
392 ConstantInt *CI = dyn_cast<ConstantInt>(V);
393 if (CI || !TD || !isa<Constant>(V) || !V->getType()->isPointerTy())
396 // This is some kind of pointer constant. Turn it into a pointer-sized
397 // ConstantInt if possible.
398 IntegerType *PtrTy = cast<IntegerType>(TD->getIntPtrType(V->getType()));
400 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
401 if (isa<ConstantPointerNull>(V))
402 return ConstantInt::get(PtrTy, 0);
404 // IntToPtr const int.
405 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
406 if (CE->getOpcode() == Instruction::IntToPtr)
407 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
408 // The constant is very likely to have the right type already.
409 if (CI->getType() == PtrTy)
412 return cast<ConstantInt>
413 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
418 /// GatherConstantCompares - Given a potentially 'or'd or 'and'd together
419 /// collection of icmp eq/ne instructions that compare a value against a
420 /// constant, return the value being compared, and stick the constant into the
423 GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra,
424 const DataLayout *TD, bool isEQ, unsigned &UsedICmps) {
425 Instruction *I = dyn_cast<Instruction>(V);
426 if (I == 0) return 0;
428 // If this is an icmp against a constant, handle this as one of the cases.
429 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) {
430 if (ConstantInt *C = GetConstantInt(I->getOperand(1), TD)) {
431 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
434 return I->getOperand(0);
437 // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to
440 ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue());
442 // If this is an and/!= check then we want to optimize "x ugt 2" into
445 Span = Span.inverse();
447 // If there are a ton of values, we don't want to make a ginormous switch.
448 if (Span.getSetSize().ugt(8) || Span.isEmptySet())
451 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
452 Vals.push_back(ConstantInt::get(V->getContext(), Tmp));
454 return I->getOperand(0);
459 // Otherwise, we can only handle an | or &, depending on isEQ.
460 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And))
463 unsigned NumValsBeforeLHS = Vals.size();
464 unsigned UsedICmpsBeforeLHS = UsedICmps;
465 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, TD,
467 unsigned NumVals = Vals.size();
468 unsigned UsedICmpsBeforeRHS = UsedICmps;
469 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD,
473 Vals.resize(NumVals);
474 UsedICmps = UsedICmpsBeforeRHS;
477 // The RHS of the or/and can't be folded in and we haven't used "Extra" yet,
478 // set it and return success.
479 if (Extra == 0 || Extra == I->getOperand(1)) {
480 Extra = I->getOperand(1);
484 Vals.resize(NumValsBeforeLHS);
485 UsedICmps = UsedICmpsBeforeLHS;
489 // If the LHS can't be folded in, but Extra is available and RHS can, try to
491 if (Extra == 0 || Extra == I->getOperand(0)) {
492 Value *OldExtra = Extra;
493 Extra = I->getOperand(0);
494 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD,
497 assert(Vals.size() == NumValsBeforeLHS);
504 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
505 Instruction *Cond = 0;
506 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
507 Cond = dyn_cast<Instruction>(SI->getCondition());
508 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
509 if (BI->isConditional())
510 Cond = dyn_cast<Instruction>(BI->getCondition());
511 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
512 Cond = dyn_cast<Instruction>(IBI->getAddress());
515 TI->eraseFromParent();
516 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
519 /// isValueEqualityComparison - Return true if the specified terminator checks
520 /// to see if a value is equal to constant integer value.
521 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
523 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
524 // Do not permit merging of large switch instructions into their
525 // predecessors unless there is only one predecessor.
526 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()),
527 pred_end(SI->getParent())) <= 128)
528 CV = SI->getCondition();
529 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
530 if (BI->isConditional() && BI->getCondition()->hasOneUse())
531 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition()))
532 if ((ICI->getPredicate() == ICmpInst::ICMP_EQ ||
533 ICI->getPredicate() == ICmpInst::ICMP_NE) &&
534 GetConstantInt(ICI->getOperand(1), TD))
535 CV = ICI->getOperand(0);
537 // Unwrap any lossless ptrtoint cast.
538 if (TD && CV && CV->getType() == TD->getIntPtrType(CV->getContext()))
539 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV))
540 CV = PTII->getOperand(0);
544 /// GetValueEqualityComparisonCases - Given a value comparison instruction,
545 /// decode all of the 'cases' that it represents and return the 'default' block.
546 BasicBlock *SimplifyCFGOpt::
547 GetValueEqualityComparisonCases(TerminatorInst *TI,
548 std::vector<ValueEqualityComparisonCase>
550 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
551 Cases.reserve(SI->getNumCases());
552 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
553 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
554 i.getCaseSuccessor()));
555 return SI->getDefaultDest();
558 BranchInst *BI = cast<BranchInst>(TI);
559 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
560 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
561 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
564 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
568 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries
569 /// in the list that match the specified block.
570 static void EliminateBlockCases(BasicBlock *BB,
571 std::vector<ValueEqualityComparisonCase> &Cases) {
572 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
575 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as
578 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
579 std::vector<ValueEqualityComparisonCase > &C2) {
580 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
582 // Make V1 be smaller than V2.
583 if (V1->size() > V2->size())
586 if (V1->size() == 0) return false;
587 if (V1->size() == 1) {
589 ConstantInt *TheVal = (*V1)[0].Value;
590 for (unsigned i = 0, e = V2->size(); i != e; ++i)
591 if (TheVal == (*V2)[i].Value)
595 // Otherwise, just sort both lists and compare element by element.
596 array_pod_sort(V1->begin(), V1->end());
597 array_pod_sort(V2->begin(), V2->end());
598 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
599 while (i1 != e1 && i2 != e2) {
600 if ((*V1)[i1].Value == (*V2)[i2].Value)
602 if ((*V1)[i1].Value < (*V2)[i2].Value)
610 /// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a
611 /// terminator instruction and its block is known to only have a single
612 /// predecessor block, check to see if that predecessor is also a value
613 /// comparison with the same value, and if that comparison determines the
614 /// outcome of this comparison. If so, simplify TI. This does a very limited
615 /// form of jump threading.
616 bool SimplifyCFGOpt::
617 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
619 IRBuilder<> &Builder) {
620 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
621 if (!PredVal) return false; // Not a value comparison in predecessor.
623 Value *ThisVal = isValueEqualityComparison(TI);
624 assert(ThisVal && "This isn't a value comparison!!");
625 if (ThisVal != PredVal) return false; // Different predicates.
627 // TODO: Preserve branch weight metadata, similarly to how
628 // FoldValueComparisonIntoPredecessors preserves it.
630 // Find out information about when control will move from Pred to TI's block.
631 std::vector<ValueEqualityComparisonCase> PredCases;
632 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
634 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
636 // Find information about how control leaves this block.
637 std::vector<ValueEqualityComparisonCase> ThisCases;
638 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
639 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
641 // If TI's block is the default block from Pred's comparison, potentially
642 // simplify TI based on this knowledge.
643 if (PredDef == TI->getParent()) {
644 // If we are here, we know that the value is none of those cases listed in
645 // PredCases. If there are any cases in ThisCases that are in PredCases, we
647 if (!ValuesOverlap(PredCases, ThisCases))
650 if (isa<BranchInst>(TI)) {
651 // Okay, one of the successors of this condbr is dead. Convert it to a
653 assert(ThisCases.size() == 1 && "Branch can only have one case!");
654 // Insert the new branch.
655 Instruction *NI = Builder.CreateBr(ThisDef);
658 // Remove PHI node entries for the dead edge.
659 ThisCases[0].Dest->removePredecessor(TI->getParent());
661 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
662 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
664 EraseTerminatorInstAndDCECond(TI);
668 SwitchInst *SI = cast<SwitchInst>(TI);
669 // Okay, TI has cases that are statically dead, prune them away.
670 SmallPtrSet<Constant*, 16> DeadCases;
671 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
672 DeadCases.insert(PredCases[i].Value);
674 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
675 << "Through successor TI: " << *TI);
677 // Collect branch weights into a vector.
678 SmallVector<uint32_t, 8> Weights;
679 MDNode* MD = SI->getMetadata(LLVMContext::MD_prof);
680 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
682 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
684 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i));
686 Weights.push_back(CI->getValue().getZExtValue());
688 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
690 if (DeadCases.count(i.getCaseValue())) {
692 std::swap(Weights[i.getCaseIndex()+1], Weights.back());
695 i.getCaseSuccessor()->removePredecessor(TI->getParent());
699 if (HasWeight && Weights.size() >= 2)
700 SI->setMetadata(LLVMContext::MD_prof,
701 MDBuilder(SI->getParent()->getContext()).
702 createBranchWeights(Weights));
704 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
708 // Otherwise, TI's block must correspond to some matched value. Find out
709 // which value (or set of values) this is.
710 ConstantInt *TIV = 0;
711 BasicBlock *TIBB = TI->getParent();
712 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
713 if (PredCases[i].Dest == TIBB) {
715 return false; // Cannot handle multiple values coming to this block.
716 TIV = PredCases[i].Value;
718 assert(TIV && "No edge from pred to succ?");
720 // Okay, we found the one constant that our value can be if we get into TI's
721 // BB. Find out which successor will unconditionally be branched to.
722 BasicBlock *TheRealDest = 0;
723 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
724 if (ThisCases[i].Value == TIV) {
725 TheRealDest = ThisCases[i].Dest;
729 // If not handled by any explicit cases, it is handled by the default case.
730 if (TheRealDest == 0) TheRealDest = ThisDef;
732 // Remove PHI node entries for dead edges.
733 BasicBlock *CheckEdge = TheRealDest;
734 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
735 if (*SI != CheckEdge)
736 (*SI)->removePredecessor(TIBB);
740 // Insert the new branch.
741 Instruction *NI = Builder.CreateBr(TheRealDest);
744 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
745 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
747 EraseTerminatorInstAndDCECond(TI);
752 /// ConstantIntOrdering - This class implements a stable ordering of constant
753 /// integers that does not depend on their address. This is important for
754 /// applications that sort ConstantInt's to ensure uniqueness.
755 struct ConstantIntOrdering {
756 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
757 return LHS->getValue().ult(RHS->getValue());
762 static int ConstantIntSortPredicate(const void *P1, const void *P2) {
763 const ConstantInt *LHS = *(const ConstantInt*const*)P1;
764 const ConstantInt *RHS = *(const ConstantInt*const*)P2;
765 if (LHS->getValue().ult(RHS->getValue()))
767 if (LHS->getValue() == RHS->getValue())
772 static inline bool HasBranchWeights(const Instruction* I) {
773 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof);
774 if (ProfMD && ProfMD->getOperand(0))
775 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
776 return MDS->getString().equals("branch_weights");
781 /// Get Weights of a given TerminatorInst, the default weight is at the front
782 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
784 static void GetBranchWeights(TerminatorInst *TI,
785 SmallVectorImpl<uint64_t> &Weights) {
786 MDNode* MD = TI->getMetadata(LLVMContext::MD_prof);
788 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
789 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(i));
791 Weights.push_back(CI->getValue().getZExtValue());
794 // If TI is a conditional eq, the default case is the false case,
795 // and the corresponding branch-weight data is at index 2. We swap the
796 // default weight to be the first entry.
797 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
798 assert(Weights.size() == 2);
799 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
800 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
801 std::swap(Weights.front(), Weights.back());
805 /// Sees if any of the weights are too big for a uint32_t, and halves all the
806 /// weights if any are.
807 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
809 for (unsigned i = 0; i < Weights.size(); ++i)
810 if (Weights[i] > UINT_MAX) {
818 for (unsigned i = 0; i < Weights.size(); ++i)
822 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value
823 /// equality comparison instruction (either a switch or a branch on "X == c").
824 /// See if any of the predecessors of the terminator block are value comparisons
825 /// on the same value. If so, and if safe to do so, fold them together.
826 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
827 IRBuilder<> &Builder) {
828 BasicBlock *BB = TI->getParent();
829 Value *CV = isValueEqualityComparison(TI); // CondVal
830 assert(CV && "Not a comparison?");
831 bool Changed = false;
833 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
834 while (!Preds.empty()) {
835 BasicBlock *Pred = Preds.pop_back_val();
837 // See if the predecessor is a comparison with the same value.
838 TerminatorInst *PTI = Pred->getTerminator();
839 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
841 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
842 // Figure out which 'cases' to copy from SI to PSI.
843 std::vector<ValueEqualityComparisonCase> BBCases;
844 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
846 std::vector<ValueEqualityComparisonCase> PredCases;
847 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
849 // Based on whether the default edge from PTI goes to BB or not, fill in
850 // PredCases and PredDefault with the new switch cases we would like to
852 SmallVector<BasicBlock*, 8> NewSuccessors;
854 // Update the branch weight metadata along the way
855 SmallVector<uint64_t, 8> Weights;
856 bool PredHasWeights = HasBranchWeights(PTI);
857 bool SuccHasWeights = HasBranchWeights(TI);
859 if (PredHasWeights) {
860 GetBranchWeights(PTI, Weights);
861 // branch-weight metadata is inconsistent here.
862 if (Weights.size() != 1 + PredCases.size())
863 PredHasWeights = SuccHasWeights = false;
864 } else if (SuccHasWeights)
865 // If there are no predecessor weights but there are successor weights,
866 // populate Weights with 1, which will later be scaled to the sum of
867 // successor's weights
868 Weights.assign(1 + PredCases.size(), 1);
870 SmallVector<uint64_t, 8> SuccWeights;
871 if (SuccHasWeights) {
872 GetBranchWeights(TI, SuccWeights);
873 // branch-weight metadata is inconsistent here.
874 if (SuccWeights.size() != 1 + BBCases.size())
875 PredHasWeights = SuccHasWeights = false;
876 } else if (PredHasWeights)
877 SuccWeights.assign(1 + BBCases.size(), 1);
879 if (PredDefault == BB) {
880 // If this is the default destination from PTI, only the edges in TI
881 // that don't occur in PTI, or that branch to BB will be activated.
882 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
883 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
884 if (PredCases[i].Dest != BB)
885 PTIHandled.insert(PredCases[i].Value);
887 // The default destination is BB, we don't need explicit targets.
888 std::swap(PredCases[i], PredCases.back());
890 if (PredHasWeights || SuccHasWeights) {
891 // Increase weight for the default case.
892 Weights[0] += Weights[i+1];
893 std::swap(Weights[i+1], Weights.back());
897 PredCases.pop_back();
901 // Reconstruct the new switch statement we will be building.
902 if (PredDefault != BBDefault) {
903 PredDefault->removePredecessor(Pred);
904 PredDefault = BBDefault;
905 NewSuccessors.push_back(BBDefault);
908 unsigned CasesFromPred = Weights.size();
909 uint64_t ValidTotalSuccWeight = 0;
910 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
911 if (!PTIHandled.count(BBCases[i].Value) &&
912 BBCases[i].Dest != BBDefault) {
913 PredCases.push_back(BBCases[i]);
914 NewSuccessors.push_back(BBCases[i].Dest);
915 if (SuccHasWeights || PredHasWeights) {
916 // The default weight is at index 0, so weight for the ith case
917 // should be at index i+1. Scale the cases from successor by
918 // PredDefaultWeight (Weights[0]).
919 Weights.push_back(Weights[0] * SuccWeights[i+1]);
920 ValidTotalSuccWeight += SuccWeights[i+1];
924 if (SuccHasWeights || PredHasWeights) {
925 ValidTotalSuccWeight += SuccWeights[0];
926 // Scale the cases from predecessor by ValidTotalSuccWeight.
927 for (unsigned i = 1; i < CasesFromPred; ++i)
928 Weights[i] *= ValidTotalSuccWeight;
929 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
930 Weights[0] *= SuccWeights[0];
933 // If this is not the default destination from PSI, only the edges
934 // in SI that occur in PSI with a destination of BB will be
936 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
937 std::map<ConstantInt*, uint64_t> WeightsForHandled;
938 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
939 if (PredCases[i].Dest == BB) {
940 PTIHandled.insert(PredCases[i].Value);
942 if (PredHasWeights || SuccHasWeights) {
943 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
944 std::swap(Weights[i+1], Weights.back());
948 std::swap(PredCases[i], PredCases.back());
949 PredCases.pop_back();
953 // Okay, now we know which constants were sent to BB from the
954 // predecessor. Figure out where they will all go now.
955 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
956 if (PTIHandled.count(BBCases[i].Value)) {
957 // If this is one we are capable of getting...
958 if (PredHasWeights || SuccHasWeights)
959 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
960 PredCases.push_back(BBCases[i]);
961 NewSuccessors.push_back(BBCases[i].Dest);
962 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
965 // If there are any constants vectored to BB that TI doesn't handle,
966 // they must go to the default destination of TI.
967 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
969 E = PTIHandled.end(); I != E; ++I) {
970 if (PredHasWeights || SuccHasWeights)
971 Weights.push_back(WeightsForHandled[*I]);
972 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
973 NewSuccessors.push_back(BBDefault);
977 // Okay, at this point, we know which new successor Pred will get. Make
978 // sure we update the number of entries in the PHI nodes for these
980 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i)
981 AddPredecessorToBlock(NewSuccessors[i], Pred, BB);
983 Builder.SetInsertPoint(PTI);
984 // Convert pointer to int before we switch.
985 if (CV->getType()->isPointerTy()) {
986 assert(TD && "Cannot switch on pointer without DataLayout");
987 CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getContext()),
991 // Now that the successors are updated, create the new Switch instruction.
992 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
994 NewSI->setDebugLoc(PTI->getDebugLoc());
995 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
996 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest);
998 if (PredHasWeights || SuccHasWeights) {
999 // Halve the weights if any of them cannot fit in an uint32_t
1000 FitWeights(Weights);
1002 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1004 NewSI->setMetadata(LLVMContext::MD_prof,
1005 MDBuilder(BB->getContext()).
1006 createBranchWeights(MDWeights));
1009 EraseTerminatorInstAndDCECond(PTI);
1011 // Okay, last check. If BB is still a successor of PSI, then we must
1012 // have an infinite loop case. If so, add an infinitely looping block
1013 // to handle the case to preserve the behavior of the code.
1014 BasicBlock *InfLoopBlock = 0;
1015 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1016 if (NewSI->getSuccessor(i) == BB) {
1017 if (InfLoopBlock == 0) {
1018 // Insert it at the end of the function, because it's either code,
1019 // or it won't matter if it's hot. :)
1020 InfLoopBlock = BasicBlock::Create(BB->getContext(),
1021 "infloop", BB->getParent());
1022 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1024 NewSI->setSuccessor(i, InfLoopBlock);
1033 // isSafeToHoistInvoke - If we would need to insert a select that uses the
1034 // value of this invoke (comments in HoistThenElseCodeToIf explain why we
1035 // would need to do this), we can't hoist the invoke, as there is nowhere
1036 // to put the select in this case.
1037 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1038 Instruction *I1, Instruction *I2) {
1039 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1041 for (BasicBlock::iterator BBI = SI->begin();
1042 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1043 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1044 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1045 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
1053 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
1054 /// BB2, hoist any common code in the two blocks up into the branch block. The
1055 /// caller of this function guarantees that BI's block dominates BB1 and BB2.
1056 static bool HoistThenElseCodeToIf(BranchInst *BI) {
1057 // This does very trivial matching, with limited scanning, to find identical
1058 // instructions in the two blocks. In particular, we don't want to get into
1059 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1060 // such, we currently just scan for obviously identical instructions in an
1062 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1063 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1065 BasicBlock::iterator BB1_Itr = BB1->begin();
1066 BasicBlock::iterator BB2_Itr = BB2->begin();
1068 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++;
1069 // Skip debug info if it is not identical.
1070 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1071 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1072 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1073 while (isa<DbgInfoIntrinsic>(I1))
1075 while (isa<DbgInfoIntrinsic>(I2))
1078 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1079 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1082 // If we get here, we can hoist at least one instruction.
1083 BasicBlock *BIParent = BI->getParent();
1086 // If we are hoisting the terminator instruction, don't move one (making a
1087 // broken BB), instead clone it, and remove BI.
1088 if (isa<TerminatorInst>(I1))
1089 goto HoistTerminator;
1091 // For a normal instruction, we just move one to right before the branch,
1092 // then replace all uses of the other with the first. Finally, we remove
1093 // the now redundant second instruction.
1094 BIParent->getInstList().splice(BI, BB1->getInstList(), I1);
1095 if (!I2->use_empty())
1096 I2->replaceAllUsesWith(I1);
1097 I1->intersectOptionalDataWith(I2);
1098 I2->eraseFromParent();
1102 // Skip debug info if it is not identical.
1103 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1104 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1105 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1106 while (isa<DbgInfoIntrinsic>(I1))
1108 while (isa<DbgInfoIntrinsic>(I2))
1111 } while (I1->isIdenticalToWhenDefined(I2));
1116 // It may not be possible to hoist an invoke.
1117 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1120 // Okay, it is safe to hoist the terminator.
1121 Instruction *NT = I1->clone();
1122 BIParent->getInstList().insert(BI, NT);
1123 if (!NT->getType()->isVoidTy()) {
1124 I1->replaceAllUsesWith(NT);
1125 I2->replaceAllUsesWith(NT);
1129 IRBuilder<true, NoFolder> Builder(NT);
1130 // Hoisting one of the terminators from our successor is a great thing.
1131 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1132 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1133 // nodes, so we insert select instruction to compute the final result.
1134 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1135 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1137 for (BasicBlock::iterator BBI = SI->begin();
1138 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1139 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1140 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1141 if (BB1V == BB2V) continue;
1143 // These values do not agree. Insert a select instruction before NT
1144 // that determines the right value.
1145 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1147 SI = cast<SelectInst>
1148 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1149 BB1V->getName()+"."+BB2V->getName()));
1151 // Make the PHI node use the select for all incoming values for BB1/BB2
1152 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1153 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1154 PN->setIncomingValue(i, SI);
1158 // Update any PHI nodes in our new successors.
1159 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1160 AddPredecessorToBlock(*SI, BIParent, BB1);
1162 EraseTerminatorInstAndDCECond(BI);
1166 /// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd,
1167 /// check whether BBEnd has only two predecessors and the other predecessor
1168 /// ends with an unconditional branch. If it is true, sink any common code
1169 /// in the two predecessors to BBEnd.
1170 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1171 assert(BI1->isUnconditional());
1172 BasicBlock *BB1 = BI1->getParent();
1173 BasicBlock *BBEnd = BI1->getSuccessor(0);
1175 // Check that BBEnd has two predecessors and the other predecessor ends with
1176 // an unconditional branch.
1177 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1178 BasicBlock *Pred0 = *PI++;
1179 if (PI == PE) // Only one predecessor.
1181 BasicBlock *Pred1 = *PI++;
1182 if (PI != PE) // More than two predecessors.
1184 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1185 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1186 if (!BI2 || !BI2->isUnconditional())
1189 // Gather the PHI nodes in BBEnd.
1190 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2;
1191 Instruction *FirstNonPhiInBBEnd = 0;
1192 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end();
1194 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1195 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1196 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1197 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN);
1199 FirstNonPhiInBBEnd = &*I;
1203 if (!FirstNonPhiInBBEnd)
1207 // This does very trivial matching, with limited scanning, to find identical
1208 // instructions in the two blocks. We scan backward for obviously identical
1209 // instructions in an identical order.
1210 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1211 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(),
1212 RE2 = BB2->getInstList().rend();
1214 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1217 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1220 // Skip the unconditional branches.
1224 bool Changed = false;
1225 while (RI1 != RE1 && RI2 != RE2) {
1227 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1230 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1234 Instruction *I1 = &*RI1, *I2 = &*RI2;
1235 // I1 and I2 should have a single use in the same PHI node, and they
1236 // perform the same operation.
1237 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1238 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1239 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1240 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) ||
1241 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1242 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1243 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1244 !I1->hasOneUse() || !I2->hasOneUse() ||
1245 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() ||
1246 MapValueFromBB1ToBB2[I1].first != I2)
1249 // Check whether we should swap the operands of ICmpInst.
1250 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1251 bool SwapOpnds = false;
1252 if (ICmp1 && ICmp2 &&
1253 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1254 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1255 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1256 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1257 ICmp2->swapOperands();
1260 if (!I1->isSameOperationAs(I2)) {
1262 ICmp2->swapOperands();
1266 // The operands should be either the same or they need to be generated
1267 // with a PHI node after sinking. We only handle the case where there is
1268 // a single pair of different operands.
1269 Value *DifferentOp1 = 0, *DifferentOp2 = 0;
1270 unsigned Op1Idx = 0;
1271 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1272 if (I1->getOperand(I) == I2->getOperand(I))
1274 // Early exit if we have more-than one pair of different operands or
1275 // the different operand is already in MapValueFromBB1ToBB2.
1276 // Early exit if we need a PHI node to replace a constant.
1278 MapValueFromBB1ToBB2.find(I1->getOperand(I)) !=
1279 MapValueFromBB1ToBB2.end() ||
1280 isa<Constant>(I1->getOperand(I)) ||
1281 isa<Constant>(I2->getOperand(I))) {
1282 // If we can't sink the instructions, undo the swapping.
1284 ICmp2->swapOperands();
1287 DifferentOp1 = I1->getOperand(I);
1289 DifferentOp2 = I2->getOperand(I);
1292 // We insert the pair of different operands to MapValueFromBB1ToBB2 and
1293 // remove (I1, I2) from MapValueFromBB1ToBB2.
1295 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2,
1296 DifferentOp1->getName() + ".sink",
1298 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN);
1299 // I1 should use NewPN instead of DifferentOp1.
1300 I1->setOperand(Op1Idx, NewPN);
1301 NewPN->addIncoming(DifferentOp1, BB1);
1302 NewPN->addIncoming(DifferentOp2, BB2);
1303 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1305 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second;
1306 MapValueFromBB1ToBB2.erase(I1);
1308 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";);
1309 DEBUG(dbgs() << " " << *I2 << "\n";);
1310 // We need to update RE1 and RE2 if we are going to sink the first
1311 // instruction in the basic block down.
1312 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1313 // Sink the instruction.
1314 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1);
1315 if (!OldPN->use_empty())
1316 OldPN->replaceAllUsesWith(I1);
1317 OldPN->eraseFromParent();
1319 if (!I2->use_empty())
1320 I2->replaceAllUsesWith(I1);
1321 I1->intersectOptionalDataWith(I2);
1322 I2->eraseFromParent();
1325 RE1 = BB1->getInstList().rend();
1327 RE2 = BB2->getInstList().rend();
1328 FirstNonPhiInBBEnd = I1;
1335 /// \brief Speculate a conditional basic block flattening the CFG.
1337 /// Note that this is a very risky transform currently. Speculating
1338 /// instructions like this is most often not desirable. Instead, there is an MI
1339 /// pass which can do it with full awareness of the resource constraints.
1340 /// However, some cases are "obvious" and we should do directly. An example of
1341 /// this is speculating a single, reasonably cheap instruction.
1343 /// There is only one distinct advantage to flattening the CFG at the IR level:
1344 /// it makes very common but simplistic optimizations such as are common in
1345 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1346 /// modeling their effects with easier to reason about SSA value graphs.
1349 /// An illustration of this transform is turning this IR:
1352 /// %cmp = icmp ult %x, %y
1353 /// br i1 %cmp, label %EndBB, label %ThenBB
1355 /// %sub = sub %x, %y
1358 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1365 /// %cmp = icmp ult %x, %y
1366 /// %sub = sub %x, %y
1367 /// %cond = select i1 %cmp, 0, %sub
1371 /// \returns true if the conditional block is removed.
1372 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1) {
1373 // Be conservative for now. FP select instruction can often be expensive.
1374 Value *BrCond = BI->getCondition();
1375 if (isa<FCmpInst>(BrCond))
1378 // Only speculatively execution a single instruction (not counting the
1379 // terminator) for now.
1380 Instruction *HInst = NULL;
1381 Instruction *Term = BB1->getTerminator();
1382 for (BasicBlock::iterator BBI = BB1->begin(), BBE = BB1->end();
1383 BBI != BBE; ++BBI) {
1384 Instruction *I = BBI;
1386 if (isa<DbgInfoIntrinsic>(I)) continue;
1387 if (I == Term) break;
1394 BasicBlock *BIParent = BI->getParent();
1396 // Check the instruction to be hoisted, if there is one.
1398 // Don't hoist the instruction if it's unsafe or expensive.
1399 if (!isSafeToSpeculativelyExecute(HInst))
1401 if (ComputeSpeculationCost(HInst) > PHINodeFoldingThreshold)
1404 // Do not hoist the instruction if any of its operands are defined but not
1405 // used in this BB. The transformation will prevent the operand from
1406 // being sunk into the use block.
1407 for (User::op_iterator i = HInst->op_begin(), e = HInst->op_end();
1409 Instruction *OpI = dyn_cast<Instruction>(*i);
1410 if (OpI && OpI->getParent() == BIParent &&
1411 !OpI->mayHaveSideEffects() &&
1412 !OpI->isUsedInBasicBlock(BIParent))
1417 // If BB1 is actually on the false edge of the conditional branch, remember
1418 // to swap the select operands later.
1419 bool Invert = false;
1420 if (BB1 != BI->getSuccessor(0)) {
1421 assert(BB1 == BI->getSuccessor(1) && "No edge from 'if' block?");
1425 // Collect interesting PHIs, and scan for hazards.
1426 SmallSetVector<std::pair<Value *, Value *>, 4> PHIs;
1427 BasicBlock *BB2 = BB1->getTerminator()->getSuccessor(0);
1428 for (BasicBlock::iterator I = BB2->begin();
1429 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1430 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1431 Value *BIParentV = PN->getIncomingValueForBlock(BIParent);
1433 // Skip PHIs which are trivial.
1434 if (BB1V == BIParentV)
1437 // Check for safety.
1438 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BB1V)) {
1439 // An unfolded ConstantExpr could end up getting expanded into
1440 // Instructions. Don't speculate this and another instruction at
1444 if (!isSafeToSpeculativelyExecute(CE))
1446 if (ComputeSpeculationCost(CE) > PHINodeFoldingThreshold)
1450 // Ok, we may insert a select for this PHI.
1451 PHIs.insert(std::make_pair(BB1V, BIParentV));
1454 // If there are no PHIs to process, bail early. This helps ensure idempotence
1459 // If we get here, we can hoist the instruction and if-convert.
1460 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *BB1 << "\n";);
1462 // Hoist the instruction.
1464 BIParent->getInstList().splice(BI, BB1->getInstList(), HInst);
1466 // Insert selects and rewrite the PHI operands.
1467 IRBuilder<true, NoFolder> Builder(BI);
1468 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
1469 Value *TrueV = PHIs[i].first;
1470 Value *FalseV = PHIs[i].second;
1472 // Create a select whose true value is the speculatively executed value and
1473 // false value is the previously determined FalseV.
1476 SI = cast<SelectInst>
1477 (Builder.CreateSelect(BrCond, FalseV, TrueV,
1478 FalseV->getName() + "." + TrueV->getName()));
1480 SI = cast<SelectInst>
1481 (Builder.CreateSelect(BrCond, TrueV, FalseV,
1482 TrueV->getName() + "." + FalseV->getName()));
1484 // Make the PHI node use the select for all incoming values for "then" and
1486 for (BasicBlock::iterator I = BB2->begin();
1487 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1488 unsigned BB1I = PN->getBasicBlockIndex(BB1);
1489 unsigned BIParentI = PN->getBasicBlockIndex(BIParent);
1490 Value *BB1V = PN->getIncomingValue(BB1I);
1491 Value *BIParentV = PN->getIncomingValue(BIParentI);
1492 if (TrueV == BB1V && FalseV == BIParentV) {
1493 PN->setIncomingValue(BB1I, SI);
1494 PN->setIncomingValue(BIParentI, SI);
1503 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1504 /// across this block.
1505 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1506 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1509 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1510 if (isa<DbgInfoIntrinsic>(BBI))
1512 if (Size > 10) return false; // Don't clone large BB's.
1515 // We can only support instructions that do not define values that are
1516 // live outside of the current basic block.
1517 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end();
1519 Instruction *U = cast<Instruction>(*UI);
1520 if (U->getParent() != BB || isa<PHINode>(U)) return false;
1523 // Looks ok, continue checking.
1529 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1530 /// that is defined in the same block as the branch and if any PHI entries are
1531 /// constants, thread edges corresponding to that entry to be branches to their
1532 /// ultimate destination.
1533 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *TD) {
1534 BasicBlock *BB = BI->getParent();
1535 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1536 // NOTE: we currently cannot transform this case if the PHI node is used
1537 // outside of the block.
1538 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1541 // Degenerate case of a single entry PHI.
1542 if (PN->getNumIncomingValues() == 1) {
1543 FoldSingleEntryPHINodes(PN->getParent());
1547 // Now we know that this block has multiple preds and two succs.
1548 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1550 // Okay, this is a simple enough basic block. See if any phi values are
1552 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1553 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1554 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue;
1556 // Okay, we now know that all edges from PredBB should be revectored to
1557 // branch to RealDest.
1558 BasicBlock *PredBB = PN->getIncomingBlock(i);
1559 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1561 if (RealDest == BB) continue; // Skip self loops.
1562 // Skip if the predecessor's terminator is an indirect branch.
1563 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1565 // The dest block might have PHI nodes, other predecessors and other
1566 // difficult cases. Instead of being smart about this, just insert a new
1567 // block that jumps to the destination block, effectively splitting
1568 // the edge we are about to create.
1569 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1570 RealDest->getName()+".critedge",
1571 RealDest->getParent(), RealDest);
1572 BranchInst::Create(RealDest, EdgeBB);
1574 // Update PHI nodes.
1575 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1577 // BB may have instructions that are being threaded over. Clone these
1578 // instructions into EdgeBB. We know that there will be no uses of the
1579 // cloned instructions outside of EdgeBB.
1580 BasicBlock::iterator InsertPt = EdgeBB->begin();
1581 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1582 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1583 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1584 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1587 // Clone the instruction.
1588 Instruction *N = BBI->clone();
1589 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1591 // Update operands due to translation.
1592 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1594 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1595 if (PI != TranslateMap.end())
1599 // Check for trivial simplification.
1600 if (Value *V = SimplifyInstruction(N, TD)) {
1601 TranslateMap[BBI] = V;
1602 delete N; // Instruction folded away, don't need actual inst
1604 // Insert the new instruction into its new home.
1605 EdgeBB->getInstList().insert(InsertPt, N);
1606 if (!BBI->use_empty())
1607 TranslateMap[BBI] = N;
1611 // Loop over all of the edges from PredBB to BB, changing them to branch
1612 // to EdgeBB instead.
1613 TerminatorInst *PredBBTI = PredBB->getTerminator();
1614 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1615 if (PredBBTI->getSuccessor(i) == BB) {
1616 BB->removePredecessor(PredBB);
1617 PredBBTI->setSuccessor(i, EdgeBB);
1620 // Recurse, simplifying any other constants.
1621 return FoldCondBranchOnPHI(BI, TD) | true;
1627 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1628 /// PHI node, see if we can eliminate it.
1629 static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *TD) {
1630 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1631 // statement", which has a very simple dominance structure. Basically, we
1632 // are trying to find the condition that is being branched on, which
1633 // subsequently causes this merge to happen. We really want control
1634 // dependence information for this check, but simplifycfg can't keep it up
1635 // to date, and this catches most of the cases we care about anyway.
1636 BasicBlock *BB = PN->getParent();
1637 BasicBlock *IfTrue, *IfFalse;
1638 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1640 // Don't bother if the branch will be constant folded trivially.
1641 isa<ConstantInt>(IfCond))
1644 // Okay, we found that we can merge this two-entry phi node into a select.
1645 // Doing so would require us to fold *all* two entry phi nodes in this block.
1646 // At some point this becomes non-profitable (particularly if the target
1647 // doesn't support cmov's). Only do this transformation if there are two or
1648 // fewer PHI nodes in this block.
1649 unsigned NumPhis = 0;
1650 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1654 // Loop over the PHI's seeing if we can promote them all to select
1655 // instructions. While we are at it, keep track of the instructions
1656 // that need to be moved to the dominating block.
1657 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1658 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1659 MaxCostVal1 = PHINodeFoldingThreshold;
1661 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1662 PHINode *PN = cast<PHINode>(II++);
1663 if (Value *V = SimplifyInstruction(PN, TD)) {
1664 PN->replaceAllUsesWith(V);
1665 PN->eraseFromParent();
1669 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1671 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1676 // If we folded the first phi, PN dangles at this point. Refresh it. If
1677 // we ran out of PHIs then we simplified them all.
1678 PN = dyn_cast<PHINode>(BB->begin());
1679 if (PN == 0) return true;
1681 // Don't fold i1 branches on PHIs which contain binary operators. These can
1682 // often be turned into switches and other things.
1683 if (PN->getType()->isIntegerTy(1) &&
1684 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1685 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1686 isa<BinaryOperator>(IfCond)))
1689 // If we all PHI nodes are promotable, check to make sure that all
1690 // instructions in the predecessor blocks can be promoted as well. If
1691 // not, we won't be able to get rid of the control flow, so it's not
1692 // worth promoting to select instructions.
1693 BasicBlock *DomBlock = 0;
1694 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1695 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1696 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1699 DomBlock = *pred_begin(IfBlock1);
1700 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1701 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1702 // This is not an aggressive instruction that we can promote.
1703 // Because of this, we won't be able to get rid of the control
1704 // flow, so the xform is not worth it.
1709 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1712 DomBlock = *pred_begin(IfBlock2);
1713 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1714 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1715 // This is not an aggressive instruction that we can promote.
1716 // Because of this, we won't be able to get rid of the control
1717 // flow, so the xform is not worth it.
1722 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1723 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1725 // If we can still promote the PHI nodes after this gauntlet of tests,
1726 // do all of the PHI's now.
1727 Instruction *InsertPt = DomBlock->getTerminator();
1728 IRBuilder<true, NoFolder> Builder(InsertPt);
1730 // Move all 'aggressive' instructions, which are defined in the
1731 // conditional parts of the if's up to the dominating block.
1733 DomBlock->getInstList().splice(InsertPt,
1734 IfBlock1->getInstList(), IfBlock1->begin(),
1735 IfBlock1->getTerminator());
1737 DomBlock->getInstList().splice(InsertPt,
1738 IfBlock2->getInstList(), IfBlock2->begin(),
1739 IfBlock2->getTerminator());
1741 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1742 // Change the PHI node into a select instruction.
1743 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1744 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1747 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1748 PN->replaceAllUsesWith(NV);
1750 PN->eraseFromParent();
1753 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1754 // has been flattened. Change DomBlock to jump directly to our new block to
1755 // avoid other simplifycfg's kicking in on the diamond.
1756 TerminatorInst *OldTI = DomBlock->getTerminator();
1757 Builder.SetInsertPoint(OldTI);
1758 Builder.CreateBr(BB);
1759 OldTI->eraseFromParent();
1763 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1764 /// to two returning blocks, try to merge them together into one return,
1765 /// introducing a select if the return values disagree.
1766 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1767 IRBuilder<> &Builder) {
1768 assert(BI->isConditional() && "Must be a conditional branch");
1769 BasicBlock *TrueSucc = BI->getSuccessor(0);
1770 BasicBlock *FalseSucc = BI->getSuccessor(1);
1771 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1772 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1774 // Check to ensure both blocks are empty (just a return) or optionally empty
1775 // with PHI nodes. If there are other instructions, merging would cause extra
1776 // computation on one path or the other.
1777 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1779 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1782 Builder.SetInsertPoint(BI);
1783 // Okay, we found a branch that is going to two return nodes. If
1784 // there is no return value for this function, just change the
1785 // branch into a return.
1786 if (FalseRet->getNumOperands() == 0) {
1787 TrueSucc->removePredecessor(BI->getParent());
1788 FalseSucc->removePredecessor(BI->getParent());
1789 Builder.CreateRetVoid();
1790 EraseTerminatorInstAndDCECond(BI);
1794 // Otherwise, figure out what the true and false return values are
1795 // so we can insert a new select instruction.
1796 Value *TrueValue = TrueRet->getReturnValue();
1797 Value *FalseValue = FalseRet->getReturnValue();
1799 // Unwrap any PHI nodes in the return blocks.
1800 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1801 if (TVPN->getParent() == TrueSucc)
1802 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1803 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1804 if (FVPN->getParent() == FalseSucc)
1805 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1807 // In order for this transformation to be safe, we must be able to
1808 // unconditionally execute both operands to the return. This is
1809 // normally the case, but we could have a potentially-trapping
1810 // constant expression that prevents this transformation from being
1812 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1815 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1819 // Okay, we collected all the mapped values and checked them for sanity, and
1820 // defined to really do this transformation. First, update the CFG.
1821 TrueSucc->removePredecessor(BI->getParent());
1822 FalseSucc->removePredecessor(BI->getParent());
1824 // Insert select instructions where needed.
1825 Value *BrCond = BI->getCondition();
1827 // Insert a select if the results differ.
1828 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1829 } else if (isa<UndefValue>(TrueValue)) {
1830 TrueValue = FalseValue;
1832 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1833 FalseValue, "retval");
1837 Value *RI = !TrueValue ?
1838 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
1842 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
1843 << "\n " << *BI << "NewRet = " << *RI
1844 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
1846 EraseTerminatorInstAndDCECond(BI);
1851 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
1852 /// probabilities of the branch taking each edge. Fills in the two APInt
1853 /// parameters and return true, or returns false if no or invalid metadata was
1855 static bool ExtractBranchMetadata(BranchInst *BI,
1856 uint64_t &ProbTrue, uint64_t &ProbFalse) {
1857 assert(BI->isConditional() &&
1858 "Looking for probabilities on unconditional branch?");
1859 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
1860 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
1861 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
1862 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
1863 if (!CITrue || !CIFalse) return false;
1864 ProbTrue = CITrue->getValue().getZExtValue();
1865 ProbFalse = CIFalse->getValue().getZExtValue();
1869 /// checkCSEInPredecessor - Return true if the given instruction is available
1870 /// in its predecessor block. If yes, the instruction will be removed.
1872 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
1873 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
1875 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
1876 Instruction *PBI = &*I;
1877 // Check whether Inst and PBI generate the same value.
1878 if (Inst->isIdenticalTo(PBI)) {
1879 Inst->replaceAllUsesWith(PBI);
1880 Inst->eraseFromParent();
1887 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
1888 /// predecessor branches to us and one of our successors, fold the block into
1889 /// the predecessor and use logical operations to pick the right destination.
1890 bool llvm::FoldBranchToCommonDest(BranchInst *BI) {
1891 BasicBlock *BB = BI->getParent();
1893 Instruction *Cond = 0;
1894 if (BI->isConditional())
1895 Cond = dyn_cast<Instruction>(BI->getCondition());
1897 // For unconditional branch, check for a simple CFG pattern, where
1898 // BB has a single predecessor and BB's successor is also its predecessor's
1899 // successor. If such pattern exisits, check for CSE between BB and its
1901 if (BasicBlock *PB = BB->getSinglePredecessor())
1902 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
1903 if (PBI->isConditional() &&
1904 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
1905 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
1906 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
1908 Instruction *Curr = I++;
1909 if (isa<CmpInst>(Curr)) {
1913 // Quit if we can't remove this instruction.
1914 if (!checkCSEInPredecessor(Curr, PB))
1923 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
1924 Cond->getParent() != BB || !Cond->hasOneUse())
1927 // Only allow this if the condition is a simple instruction that can be
1928 // executed unconditionally. It must be in the same block as the branch, and
1929 // must be at the front of the block.
1930 BasicBlock::iterator FrontIt = BB->front();
1932 // Ignore dbg intrinsics.
1933 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
1935 // Allow a single instruction to be hoisted in addition to the compare
1936 // that feeds the branch. We later ensure that any values that _it_ uses
1937 // were also live in the predecessor, so that we don't unnecessarily create
1938 // register pressure or inhibit out-of-order execution.
1939 Instruction *BonusInst = 0;
1940 if (&*FrontIt != Cond &&
1941 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond &&
1942 isSafeToSpeculativelyExecute(FrontIt)) {
1943 BonusInst = &*FrontIt;
1946 // Ignore dbg intrinsics.
1947 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
1950 // Only a single bonus inst is allowed.
1951 if (&*FrontIt != Cond)
1954 // Make sure the instruction after the condition is the cond branch.
1955 BasicBlock::iterator CondIt = Cond; ++CondIt;
1957 // Ingore dbg intrinsics.
1958 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
1963 // Cond is known to be a compare or binary operator. Check to make sure that
1964 // neither operand is a potentially-trapping constant expression.
1965 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
1968 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
1972 // Finally, don't infinitely unroll conditional loops.
1973 BasicBlock *TrueDest = BI->getSuccessor(0);
1974 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0;
1975 if (TrueDest == BB || FalseDest == BB)
1978 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
1979 BasicBlock *PredBlock = *PI;
1980 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
1982 // Check that we have two conditional branches. If there is a PHI node in
1983 // the common successor, verify that the same value flows in from both
1985 SmallVector<PHINode*, 4> PHIs;
1986 if (PBI == 0 || PBI->isUnconditional() ||
1987 (BI->isConditional() &&
1988 !SafeToMergeTerminators(BI, PBI)) ||
1989 (!BI->isConditional() &&
1990 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
1993 // Determine if the two branches share a common destination.
1994 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
1995 bool InvertPredCond = false;
1997 if (BI->isConditional()) {
1998 if (PBI->getSuccessor(0) == TrueDest)
1999 Opc = Instruction::Or;
2000 else if (PBI->getSuccessor(1) == FalseDest)
2001 Opc = Instruction::And;
2002 else if (PBI->getSuccessor(0) == FalseDest)
2003 Opc = Instruction::And, InvertPredCond = true;
2004 else if (PBI->getSuccessor(1) == TrueDest)
2005 Opc = Instruction::Or, InvertPredCond = true;
2009 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2013 // Ensure that any values used in the bonus instruction are also used
2014 // by the terminator of the predecessor. This means that those values
2015 // must already have been resolved, so we won't be inhibiting the
2016 // out-of-order core by speculating them earlier.
2018 // Collect the values used by the bonus inst
2019 SmallPtrSet<Value*, 4> UsedValues;
2020 for (Instruction::op_iterator OI = BonusInst->op_begin(),
2021 OE = BonusInst->op_end(); OI != OE; ++OI) {
2023 if (!isa<Constant>(V))
2024 UsedValues.insert(V);
2027 SmallVector<std::pair<Value*, unsigned>, 4> Worklist;
2028 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0));
2030 // Walk up to four levels back up the use-def chain of the predecessor's
2031 // terminator to see if all those values were used. The choice of four
2032 // levels is arbitrary, to provide a compile-time-cost bound.
2033 while (!Worklist.empty()) {
2034 std::pair<Value*, unsigned> Pair = Worklist.back();
2035 Worklist.pop_back();
2037 if (Pair.second >= 4) continue;
2038 UsedValues.erase(Pair.first);
2039 if (UsedValues.empty()) break;
2041 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) {
2042 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
2044 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1));
2048 if (!UsedValues.empty()) return false;
2051 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2052 IRBuilder<> Builder(PBI);
2054 // If we need to invert the condition in the pred block to match, do so now.
2055 if (InvertPredCond) {
2056 Value *NewCond = PBI->getCondition();
2058 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2059 CmpInst *CI = cast<CmpInst>(NewCond);
2060 CI->setPredicate(CI->getInversePredicate());
2062 NewCond = Builder.CreateNot(NewCond,
2063 PBI->getCondition()->getName()+".not");
2066 PBI->setCondition(NewCond);
2067 PBI->swapSuccessors();
2070 // If we have a bonus inst, clone it into the predecessor block.
2071 Instruction *NewBonus = 0;
2073 NewBonus = BonusInst->clone();
2074 PredBlock->getInstList().insert(PBI, NewBonus);
2075 NewBonus->takeName(BonusInst);
2076 BonusInst->setName(BonusInst->getName()+".old");
2079 // Clone Cond into the predecessor basic block, and or/and the
2080 // two conditions together.
2081 Instruction *New = Cond->clone();
2082 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus);
2083 PredBlock->getInstList().insert(PBI, New);
2084 New->takeName(Cond);
2085 Cond->setName(New->getName()+".old");
2087 if (BI->isConditional()) {
2088 Instruction *NewCond =
2089 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2091 PBI->setCondition(NewCond);
2093 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2094 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2096 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2098 SmallVector<uint64_t, 8> NewWeights;
2100 if (PBI->getSuccessor(0) == BB) {
2101 if (PredHasWeights && SuccHasWeights) {
2102 // PBI: br i1 %x, BB, FalseDest
2103 // BI: br i1 %y, TrueDest, FalseDest
2104 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2105 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2106 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2107 // TrueWeight for PBI * FalseWeight for BI.
2108 // We assume that total weights of a BranchInst can fit into 32 bits.
2109 // Therefore, we will not have overflow using 64-bit arithmetic.
2110 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2111 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2113 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2114 PBI->setSuccessor(0, TrueDest);
2116 if (PBI->getSuccessor(1) == BB) {
2117 if (PredHasWeights && SuccHasWeights) {
2118 // PBI: br i1 %x, TrueDest, BB
2119 // BI: br i1 %y, TrueDest, FalseDest
2120 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2121 // FalseWeight for PBI * TrueWeight for BI.
2122 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2123 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2124 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2125 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2127 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2128 PBI->setSuccessor(1, FalseDest);
2130 if (NewWeights.size() == 2) {
2131 // Halve the weights if any of them cannot fit in an uint32_t
2132 FitWeights(NewWeights);
2134 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2135 PBI->setMetadata(LLVMContext::MD_prof,
2136 MDBuilder(BI->getContext()).
2137 createBranchWeights(MDWeights));
2139 PBI->setMetadata(LLVMContext::MD_prof, NULL);
2141 // Update PHI nodes in the common successors.
2142 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2143 ConstantInt *PBI_C = cast<ConstantInt>(
2144 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2145 assert(PBI_C->getType()->isIntegerTy(1));
2146 Instruction *MergedCond = 0;
2147 if (PBI->getSuccessor(0) == TrueDest) {
2148 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2149 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2150 // is false: !PBI_Cond and BI_Value
2151 Instruction *NotCond =
2152 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2155 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2160 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2161 PBI->getCondition(), MergedCond,
2164 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2165 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2166 // is false: PBI_Cond and BI_Value
2168 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2169 PBI->getCondition(), New,
2171 if (PBI_C->isOne()) {
2172 Instruction *NotCond =
2173 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2176 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2177 NotCond, MergedCond,
2182 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2185 // Change PBI from Conditional to Unconditional.
2186 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2187 EraseTerminatorInstAndDCECond(PBI);
2191 // TODO: If BB is reachable from all paths through PredBlock, then we
2192 // could replace PBI's branch probabilities with BI's.
2194 // Copy any debug value intrinsics into the end of PredBlock.
2195 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2196 if (isa<DbgInfoIntrinsic>(*I))
2197 I->clone()->insertBefore(PBI);
2204 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2205 /// predecessor of another block, this function tries to simplify it. We know
2206 /// that PBI and BI are both conditional branches, and BI is in one of the
2207 /// successor blocks of PBI - PBI branches to BI.
2208 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2209 assert(PBI->isConditional() && BI->isConditional());
2210 BasicBlock *BB = BI->getParent();
2212 // If this block ends with a branch instruction, and if there is a
2213 // predecessor that ends on a branch of the same condition, make
2214 // this conditional branch redundant.
2215 if (PBI->getCondition() == BI->getCondition() &&
2216 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2217 // Okay, the outcome of this conditional branch is statically
2218 // knowable. If this block had a single pred, handle specially.
2219 if (BB->getSinglePredecessor()) {
2220 // Turn this into a branch on constant.
2221 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2222 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2224 return true; // Nuke the branch on constant.
2227 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2228 // in the constant and simplify the block result. Subsequent passes of
2229 // simplifycfg will thread the block.
2230 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2231 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2232 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2233 std::distance(PB, PE),
2234 BI->getCondition()->getName() + ".pr",
2236 // Okay, we're going to insert the PHI node. Since PBI is not the only
2237 // predecessor, compute the PHI'd conditional value for all of the preds.
2238 // Any predecessor where the condition is not computable we keep symbolic.
2239 for (pred_iterator PI = PB; PI != PE; ++PI) {
2240 BasicBlock *P = *PI;
2241 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2242 PBI != BI && PBI->isConditional() &&
2243 PBI->getCondition() == BI->getCondition() &&
2244 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2245 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2246 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2249 NewPN->addIncoming(BI->getCondition(), P);
2253 BI->setCondition(NewPN);
2258 // If this is a conditional branch in an empty block, and if any
2259 // predecessors is a conditional branch to one of our destinations,
2260 // fold the conditions into logical ops and one cond br.
2261 BasicBlock::iterator BBI = BB->begin();
2262 // Ignore dbg intrinsics.
2263 while (isa<DbgInfoIntrinsic>(BBI))
2269 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2274 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2276 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2277 PBIOp = 0, BIOp = 1;
2278 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2279 PBIOp = 1, BIOp = 0;
2280 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2285 // Check to make sure that the other destination of this branch
2286 // isn't BB itself. If so, this is an infinite loop that will
2287 // keep getting unwound.
2288 if (PBI->getSuccessor(PBIOp) == BB)
2291 // Do not perform this transformation if it would require
2292 // insertion of a large number of select instructions. For targets
2293 // without predication/cmovs, this is a big pessimization.
2294 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2296 unsigned NumPhis = 0;
2297 for (BasicBlock::iterator II = CommonDest->begin();
2298 isa<PHINode>(II); ++II, ++NumPhis)
2299 if (NumPhis > 2) // Disable this xform.
2302 // Finally, if everything is ok, fold the branches to logical ops.
2303 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2305 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2306 << "AND: " << *BI->getParent());
2309 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2310 // branch in it, where one edge (OtherDest) goes back to itself but the other
2311 // exits. We don't *know* that the program avoids the infinite loop
2312 // (even though that seems likely). If we do this xform naively, we'll end up
2313 // recursively unpeeling the loop. Since we know that (after the xform is
2314 // done) that the block *is* infinite if reached, we just make it an obviously
2315 // infinite loop with no cond branch.
2316 if (OtherDest == BB) {
2317 // Insert it at the end of the function, because it's either code,
2318 // or it won't matter if it's hot. :)
2319 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2320 "infloop", BB->getParent());
2321 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2322 OtherDest = InfLoopBlock;
2325 DEBUG(dbgs() << *PBI->getParent()->getParent());
2327 // BI may have other predecessors. Because of this, we leave
2328 // it alone, but modify PBI.
2330 // Make sure we get to CommonDest on True&True directions.
2331 Value *PBICond = PBI->getCondition();
2332 IRBuilder<true, NoFolder> Builder(PBI);
2334 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2336 Value *BICond = BI->getCondition();
2338 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2340 // Merge the conditions.
2341 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2343 // Modify PBI to branch on the new condition to the new dests.
2344 PBI->setCondition(Cond);
2345 PBI->setSuccessor(0, CommonDest);
2346 PBI->setSuccessor(1, OtherDest);
2348 // Update branch weight for PBI.
2349 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2350 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2352 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2354 if (PredHasWeights && SuccHasWeights) {
2355 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2356 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2357 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2358 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2359 // The weight to CommonDest should be PredCommon * SuccTotal +
2360 // PredOther * SuccCommon.
2361 // The weight to OtherDest should be PredOther * SuccOther.
2362 SmallVector<uint64_t, 2> NewWeights;
2363 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) +
2364 PredOther * SuccCommon);
2365 NewWeights.push_back(PredOther * SuccOther);
2366 // Halve the weights if any of them cannot fit in an uint32_t
2367 FitWeights(NewWeights);
2369 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end());
2370 PBI->setMetadata(LLVMContext::MD_prof,
2371 MDBuilder(BI->getContext()).
2372 createBranchWeights(MDWeights));
2375 // OtherDest may have phi nodes. If so, add an entry from PBI's
2376 // block that are identical to the entries for BI's block.
2377 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2379 // We know that the CommonDest already had an edge from PBI to
2380 // it. If it has PHIs though, the PHIs may have different
2381 // entries for BB and PBI's BB. If so, insert a select to make
2384 for (BasicBlock::iterator II = CommonDest->begin();
2385 (PN = dyn_cast<PHINode>(II)); ++II) {
2386 Value *BIV = PN->getIncomingValueForBlock(BB);
2387 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2388 Value *PBIV = PN->getIncomingValue(PBBIdx);
2390 // Insert a select in PBI to pick the right value.
2391 Value *NV = cast<SelectInst>
2392 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2393 PN->setIncomingValue(PBBIdx, NV);
2397 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2398 DEBUG(dbgs() << *PBI->getParent()->getParent());
2400 // This basic block is probably dead. We know it has at least
2401 // one fewer predecessor.
2405 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2406 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2407 // Takes care of updating the successors and removing the old terminator.
2408 // Also makes sure not to introduce new successors by assuming that edges to
2409 // non-successor TrueBBs and FalseBBs aren't reachable.
2410 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2411 BasicBlock *TrueBB, BasicBlock *FalseBB,
2412 uint32_t TrueWeight,
2413 uint32_t FalseWeight){
2414 // Remove any superfluous successor edges from the CFG.
2415 // First, figure out which successors to preserve.
2416 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2418 BasicBlock *KeepEdge1 = TrueBB;
2419 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0;
2421 // Then remove the rest.
2422 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2423 BasicBlock *Succ = OldTerm->getSuccessor(I);
2424 // Make sure only to keep exactly one copy of each edge.
2425 if (Succ == KeepEdge1)
2427 else if (Succ == KeepEdge2)
2430 Succ->removePredecessor(OldTerm->getParent());
2433 IRBuilder<> Builder(OldTerm);
2434 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2436 // Insert an appropriate new terminator.
2437 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) {
2438 if (TrueBB == FalseBB)
2439 // We were only looking for one successor, and it was present.
2440 // Create an unconditional branch to it.
2441 Builder.CreateBr(TrueBB);
2443 // We found both of the successors we were looking for.
2444 // Create a conditional branch sharing the condition of the select.
2445 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2446 if (TrueWeight != FalseWeight)
2447 NewBI->setMetadata(LLVMContext::MD_prof,
2448 MDBuilder(OldTerm->getContext()).
2449 createBranchWeights(TrueWeight, FalseWeight));
2451 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2452 // Neither of the selected blocks were successors, so this
2453 // terminator must be unreachable.
2454 new UnreachableInst(OldTerm->getContext(), OldTerm);
2456 // One of the selected values was a successor, but the other wasn't.
2457 // Insert an unconditional branch to the one that was found;
2458 // the edge to the one that wasn't must be unreachable.
2460 // Only TrueBB was found.
2461 Builder.CreateBr(TrueBB);
2463 // Only FalseBB was found.
2464 Builder.CreateBr(FalseBB);
2467 EraseTerminatorInstAndDCECond(OldTerm);
2471 // SimplifySwitchOnSelect - Replaces
2472 // (switch (select cond, X, Y)) on constant X, Y
2473 // with a branch - conditional if X and Y lead to distinct BBs,
2474 // unconditional otherwise.
2475 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2476 // Check for constant integer values in the select.
2477 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2478 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2479 if (!TrueVal || !FalseVal)
2482 // Find the relevant condition and destinations.
2483 Value *Condition = Select->getCondition();
2484 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2485 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2487 // Get weight for TrueBB and FalseBB.
2488 uint32_t TrueWeight = 0, FalseWeight = 0;
2489 SmallVector<uint64_t, 8> Weights;
2490 bool HasWeights = HasBranchWeights(SI);
2492 GetBranchWeights(SI, Weights);
2493 if (Weights.size() == 1 + SI->getNumCases()) {
2494 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2495 getSuccessorIndex()];
2496 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2497 getSuccessorIndex()];
2501 // Perform the actual simplification.
2502 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2503 TrueWeight, FalseWeight);
2506 // SimplifyIndirectBrOnSelect - Replaces
2507 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2508 // blockaddress(@fn, BlockB)))
2510 // (br cond, BlockA, BlockB).
2511 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2512 // Check that both operands of the select are block addresses.
2513 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2514 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2518 // Extract the actual blocks.
2519 BasicBlock *TrueBB = TBA->getBasicBlock();
2520 BasicBlock *FalseBB = FBA->getBasicBlock();
2522 // Perform the actual simplification.
2523 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2527 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2528 /// instruction (a seteq/setne with a constant) as the only instruction in a
2529 /// block that ends with an uncond branch. We are looking for a very specific
2530 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2531 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2532 /// default value goes to an uncond block with a seteq in it, we get something
2535 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2537 /// %tmp = icmp eq i8 %A, 92
2540 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2542 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2543 /// the PHI, merging the third icmp into the switch.
2544 static bool TryToSimplifyUncondBranchWithICmpInIt(
2545 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI,
2546 const DataLayout *TD) {
2547 BasicBlock *BB = ICI->getParent();
2549 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2551 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2553 Value *V = ICI->getOperand(0);
2554 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2556 // The pattern we're looking for is where our only predecessor is a switch on
2557 // 'V' and this block is the default case for the switch. In this case we can
2558 // fold the compared value into the switch to simplify things.
2559 BasicBlock *Pred = BB->getSinglePredecessor();
2560 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false;
2562 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2563 if (SI->getCondition() != V)
2566 // If BB is reachable on a non-default case, then we simply know the value of
2567 // V in this block. Substitute it and constant fold the icmp instruction
2569 if (SI->getDefaultDest() != BB) {
2570 ConstantInt *VVal = SI->findCaseDest(BB);
2571 assert(VVal && "Should have a unique destination value");
2572 ICI->setOperand(0, VVal);
2574 if (Value *V = SimplifyInstruction(ICI, TD)) {
2575 ICI->replaceAllUsesWith(V);
2576 ICI->eraseFromParent();
2578 // BB is now empty, so it is likely to simplify away.
2579 return SimplifyCFG(BB, TTI, TD) | true;
2582 // Ok, the block is reachable from the default dest. If the constant we're
2583 // comparing exists in one of the other edges, then we can constant fold ICI
2585 if (SI->findCaseValue(Cst) != SI->case_default()) {
2587 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2588 V = ConstantInt::getFalse(BB->getContext());
2590 V = ConstantInt::getTrue(BB->getContext());
2592 ICI->replaceAllUsesWith(V);
2593 ICI->eraseFromParent();
2594 // BB is now empty, so it is likely to simplify away.
2595 return SimplifyCFG(BB, TTI, TD) | true;
2598 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2600 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2601 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back());
2602 if (PHIUse == 0 || PHIUse != &SuccBlock->front() ||
2603 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2606 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2608 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2609 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2611 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2612 std::swap(DefaultCst, NewCst);
2614 // Replace ICI (which is used by the PHI for the default value) with true or
2615 // false depending on if it is EQ or NE.
2616 ICI->replaceAllUsesWith(DefaultCst);
2617 ICI->eraseFromParent();
2619 // Okay, the switch goes to this block on a default value. Add an edge from
2620 // the switch to the merge point on the compared value.
2621 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2622 BB->getParent(), BB);
2623 SmallVector<uint64_t, 8> Weights;
2624 bool HasWeights = HasBranchWeights(SI);
2626 GetBranchWeights(SI, Weights);
2627 if (Weights.size() == 1 + SI->getNumCases()) {
2628 // Split weight for default case to case for "Cst".
2629 Weights[0] = (Weights[0]+1) >> 1;
2630 Weights.push_back(Weights[0]);
2632 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2633 SI->setMetadata(LLVMContext::MD_prof,
2634 MDBuilder(SI->getContext()).
2635 createBranchWeights(MDWeights));
2638 SI->addCase(Cst, NewBB);
2640 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2641 Builder.SetInsertPoint(NewBB);
2642 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2643 Builder.CreateBr(SuccBlock);
2644 PHIUse->addIncoming(NewCst, NewBB);
2648 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2649 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2650 /// fold it into a switch instruction if so.
2651 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *TD,
2652 IRBuilder<> &Builder) {
2653 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2654 if (Cond == 0) return false;
2657 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2658 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2659 // 'setne's and'ed together, collect them.
2661 std::vector<ConstantInt*> Values;
2662 bool TrueWhenEqual = true;
2663 Value *ExtraCase = 0;
2664 unsigned UsedICmps = 0;
2666 if (Cond->getOpcode() == Instruction::Or) {
2667 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true,
2669 } else if (Cond->getOpcode() == Instruction::And) {
2670 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false,
2672 TrueWhenEqual = false;
2675 // If we didn't have a multiply compared value, fail.
2676 if (CompVal == 0) return false;
2678 // Avoid turning single icmps into a switch.
2682 // There might be duplicate constants in the list, which the switch
2683 // instruction can't handle, remove them now.
2684 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2685 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2687 // If Extra was used, we require at least two switch values to do the
2688 // transformation. A switch with one value is just an cond branch.
2689 if (ExtraCase && Values.size() < 2) return false;
2691 // TODO: Preserve branch weight metadata, similarly to how
2692 // FoldValueComparisonIntoPredecessors preserves it.
2694 // Figure out which block is which destination.
2695 BasicBlock *DefaultBB = BI->getSuccessor(1);
2696 BasicBlock *EdgeBB = BI->getSuccessor(0);
2697 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2699 BasicBlock *BB = BI->getParent();
2701 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2702 << " cases into SWITCH. BB is:\n" << *BB);
2704 // If there are any extra values that couldn't be folded into the switch
2705 // then we evaluate them with an explicit branch first. Split the block
2706 // right before the condbr to handle it.
2708 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2709 // Remove the uncond branch added to the old block.
2710 TerminatorInst *OldTI = BB->getTerminator();
2711 Builder.SetInsertPoint(OldTI);
2714 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2716 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2718 OldTI->eraseFromParent();
2720 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2721 // for the edge we just added.
2722 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2724 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2725 << "\nEXTRABB = " << *BB);
2729 Builder.SetInsertPoint(BI);
2730 // Convert pointer to int before we switch.
2731 if (CompVal->getType()->isPointerTy()) {
2732 assert(TD && "Cannot switch on pointer without DataLayout");
2733 CompVal = Builder.CreatePtrToInt(CompVal,
2734 TD->getIntPtrType(CompVal->getContext()),
2738 // Create the new switch instruction now.
2739 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2741 // Add all of the 'cases' to the switch instruction.
2742 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2743 New->addCase(Values[i], EdgeBB);
2745 // We added edges from PI to the EdgeBB. As such, if there were any
2746 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2747 // the number of edges added.
2748 for (BasicBlock::iterator BBI = EdgeBB->begin();
2749 isa<PHINode>(BBI); ++BBI) {
2750 PHINode *PN = cast<PHINode>(BBI);
2751 Value *InVal = PN->getIncomingValueForBlock(BB);
2752 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2753 PN->addIncoming(InVal, BB);
2756 // Erase the old branch instruction.
2757 EraseTerminatorInstAndDCECond(BI);
2759 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2763 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2764 // If this is a trivial landing pad that just continues unwinding the caught
2765 // exception then zap the landing pad, turning its invokes into calls.
2766 BasicBlock *BB = RI->getParent();
2767 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2768 if (RI->getValue() != LPInst)
2769 // Not a landing pad, or the resume is not unwinding the exception that
2770 // caused control to branch here.
2773 // Check that there are no other instructions except for debug intrinsics.
2774 BasicBlock::iterator I = LPInst, E = RI;
2776 if (!isa<DbgInfoIntrinsic>(I))
2779 // Turn all invokes that unwind here into calls and delete the basic block.
2780 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2781 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2782 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2783 // Insert a call instruction before the invoke.
2784 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2786 Call->setCallingConv(II->getCallingConv());
2787 Call->setAttributes(II->getAttributes());
2788 Call->setDebugLoc(II->getDebugLoc());
2790 // Anything that used the value produced by the invoke instruction now uses
2791 // the value produced by the call instruction. Note that we do this even
2792 // for void functions and calls with no uses so that the callgraph edge is
2794 II->replaceAllUsesWith(Call);
2795 BB->removePredecessor(II->getParent());
2797 // Insert a branch to the normal destination right before the invoke.
2798 BranchInst::Create(II->getNormalDest(), II);
2800 // Finally, delete the invoke instruction!
2801 II->eraseFromParent();
2804 // The landingpad is now unreachable. Zap it.
2805 BB->eraseFromParent();
2809 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2810 BasicBlock *BB = RI->getParent();
2811 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2813 // Find predecessors that end with branches.
2814 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2815 SmallVector<BranchInst*, 8> CondBranchPreds;
2816 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2817 BasicBlock *P = *PI;
2818 TerminatorInst *PTI = P->getTerminator();
2819 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2820 if (BI->isUnconditional())
2821 UncondBranchPreds.push_back(P);
2823 CondBranchPreds.push_back(BI);
2827 // If we found some, do the transformation!
2828 if (!UncondBranchPreds.empty() && DupRet) {
2829 while (!UncondBranchPreds.empty()) {
2830 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
2831 DEBUG(dbgs() << "FOLDING: " << *BB
2832 << "INTO UNCOND BRANCH PRED: " << *Pred);
2833 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
2836 // If we eliminated all predecessors of the block, delete the block now.
2837 if (pred_begin(BB) == pred_end(BB))
2838 // We know there are no successors, so just nuke the block.
2839 BB->eraseFromParent();
2844 // Check out all of the conditional branches going to this return
2845 // instruction. If any of them just select between returns, change the
2846 // branch itself into a select/return pair.
2847 while (!CondBranchPreds.empty()) {
2848 BranchInst *BI = CondBranchPreds.pop_back_val();
2850 // Check to see if the non-BB successor is also a return block.
2851 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
2852 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
2853 SimplifyCondBranchToTwoReturns(BI, Builder))
2859 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
2860 BasicBlock *BB = UI->getParent();
2862 bool Changed = false;
2864 // If there are any instructions immediately before the unreachable that can
2865 // be removed, do so.
2866 while (UI != BB->begin()) {
2867 BasicBlock::iterator BBI = UI;
2869 // Do not delete instructions that can have side effects which might cause
2870 // the unreachable to not be reachable; specifically, calls and volatile
2871 // operations may have this effect.
2872 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
2874 if (BBI->mayHaveSideEffects()) {
2875 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
2876 if (SI->isVolatile())
2878 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
2879 if (LI->isVolatile())
2881 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
2882 if (RMWI->isVolatile())
2884 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
2885 if (CXI->isVolatile())
2887 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
2888 !isa<LandingPadInst>(BBI)) {
2891 // Note that deleting LandingPad's here is in fact okay, although it
2892 // involves a bit of subtle reasoning. If this inst is a LandingPad,
2893 // all the predecessors of this block will be the unwind edges of Invokes,
2894 // and we can therefore guarantee this block will be erased.
2897 // Delete this instruction (any uses are guaranteed to be dead)
2898 if (!BBI->use_empty())
2899 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
2900 BBI->eraseFromParent();
2904 // If the unreachable instruction is the first in the block, take a gander
2905 // at all of the predecessors of this instruction, and simplify them.
2906 if (&BB->front() != UI) return Changed;
2908 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
2909 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
2910 TerminatorInst *TI = Preds[i]->getTerminator();
2911 IRBuilder<> Builder(TI);
2912 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
2913 if (BI->isUnconditional()) {
2914 if (BI->getSuccessor(0) == BB) {
2915 new UnreachableInst(TI->getContext(), TI);
2916 TI->eraseFromParent();
2920 if (BI->getSuccessor(0) == BB) {
2921 Builder.CreateBr(BI->getSuccessor(1));
2922 EraseTerminatorInstAndDCECond(BI);
2923 } else if (BI->getSuccessor(1) == BB) {
2924 Builder.CreateBr(BI->getSuccessor(0));
2925 EraseTerminatorInstAndDCECond(BI);
2929 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
2930 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2932 if (i.getCaseSuccessor() == BB) {
2933 BB->removePredecessor(SI->getParent());
2938 // If the default value is unreachable, figure out the most popular
2939 // destination and make it the default.
2940 if (SI->getDefaultDest() == BB) {
2941 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
2942 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2944 std::pair<unsigned, unsigned> &entry =
2945 Popularity[i.getCaseSuccessor()];
2946 if (entry.first == 0) {
2948 entry.second = i.getCaseIndex();
2954 // Find the most popular block.
2955 unsigned MaxPop = 0;
2956 unsigned MaxIndex = 0;
2957 BasicBlock *MaxBlock = 0;
2958 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
2959 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
2960 if (I->second.first > MaxPop ||
2961 (I->second.first == MaxPop && MaxIndex > I->second.second)) {
2962 MaxPop = I->second.first;
2963 MaxIndex = I->second.second;
2964 MaxBlock = I->first;
2968 // Make this the new default, allowing us to delete any explicit
2970 SI->setDefaultDest(MaxBlock);
2973 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
2975 if (isa<PHINode>(MaxBlock->begin()))
2976 for (unsigned i = 0; i != MaxPop-1; ++i)
2977 MaxBlock->removePredecessor(SI->getParent());
2979 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2981 if (i.getCaseSuccessor() == MaxBlock) {
2987 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
2988 if (II->getUnwindDest() == BB) {
2989 // Convert the invoke to a call instruction. This would be a good
2990 // place to note that the call does not throw though.
2991 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
2992 II->removeFromParent(); // Take out of symbol table
2994 // Insert the call now...
2995 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
2996 Builder.SetInsertPoint(BI);
2997 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
2998 Args, II->getName());
2999 CI->setCallingConv(II->getCallingConv());
3000 CI->setAttributes(II->getAttributes());
3001 // If the invoke produced a value, the call does now instead.
3002 II->replaceAllUsesWith(CI);
3009 // If this block is now dead, remove it.
3010 if (pred_begin(BB) == pred_end(BB) &&
3011 BB != &BB->getParent()->getEntryBlock()) {
3012 // We know there are no successors, so just nuke the block.
3013 BB->eraseFromParent();
3020 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
3021 /// integer range comparison into a sub, an icmp and a branch.
3022 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3023 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3025 // Make sure all cases point to the same destination and gather the values.
3026 SmallVector<ConstantInt *, 16> Cases;
3027 SwitchInst::CaseIt I = SI->case_begin();
3028 Cases.push_back(I.getCaseValue());
3029 SwitchInst::CaseIt PrevI = I++;
3030 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
3031 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
3033 Cases.push_back(I.getCaseValue());
3035 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
3037 // Sort the case values, then check if they form a range we can transform.
3038 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3039 for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
3040 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
3044 Constant *Offset = ConstantExpr::getNeg(Cases.back());
3045 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
3047 Value *Sub = SI->getCondition();
3048 if (!Offset->isNullValue())
3049 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
3050 Value *Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3051 BranchInst *NewBI = Builder.CreateCondBr(
3052 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
3054 // Update weight for the newly-created conditional branch.
3055 SmallVector<uint64_t, 8> Weights;
3056 bool HasWeights = HasBranchWeights(SI);
3058 GetBranchWeights(SI, Weights);
3059 if (Weights.size() == 1 + SI->getNumCases()) {
3060 // Combine all weights for the cases to be the true weight of NewBI.
3061 // We assume that the sum of all weights for a Terminator can fit into 32
3063 uint32_t NewTrueWeight = 0;
3064 for (unsigned I = 1, E = Weights.size(); I != E; ++I)
3065 NewTrueWeight += (uint32_t)Weights[I];
3066 NewBI->setMetadata(LLVMContext::MD_prof,
3067 MDBuilder(SI->getContext()).
3068 createBranchWeights(NewTrueWeight,
3069 (uint32_t)Weights[0]));
3073 // Prune obsolete incoming values off the successor's PHI nodes.
3074 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
3075 isa<PHINode>(BBI); ++BBI) {
3076 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
3077 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3079 SI->eraseFromParent();
3084 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
3085 /// and use it to remove dead cases.
3086 static bool EliminateDeadSwitchCases(SwitchInst *SI) {
3087 Value *Cond = SI->getCondition();
3088 unsigned Bits = cast<IntegerType>(Cond->getType())->getBitWidth();
3089 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3090 ComputeMaskedBits(Cond, KnownZero, KnownOne);
3092 // Gather dead cases.
3093 SmallVector<ConstantInt*, 8> DeadCases;
3094 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3095 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3096 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3097 DeadCases.push_back(I.getCaseValue());
3098 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3099 << I.getCaseValue() << "' is dead.\n");
3103 SmallVector<uint64_t, 8> Weights;
3104 bool HasWeight = HasBranchWeights(SI);
3106 GetBranchWeights(SI, Weights);
3107 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3110 // Remove dead cases from the switch.
3111 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3112 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3113 assert(Case != SI->case_default() &&
3114 "Case was not found. Probably mistake in DeadCases forming.");
3116 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3120 // Prune unused values from PHI nodes.
3121 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3122 SI->removeCase(Case);
3125 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3126 SI->setMetadata(LLVMContext::MD_prof,
3127 MDBuilder(SI->getParent()->getContext()).
3128 createBranchWeights(MDWeights));
3131 return !DeadCases.empty();
3134 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
3135 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3136 /// by an unconditional branch), look at the phi node for BB in the successor
3137 /// block and see if the incoming value is equal to CaseValue. If so, return
3138 /// the phi node, and set PhiIndex to BB's index in the phi node.
3139 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3142 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3143 return NULL; // BB must be empty to be a candidate for simplification.
3144 if (!BB->getSinglePredecessor())
3145 return NULL; // BB must be dominated by the switch.
3147 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3148 if (!Branch || !Branch->isUnconditional())
3149 return NULL; // Terminator must be unconditional branch.
3151 BasicBlock *Succ = Branch->getSuccessor(0);
3153 BasicBlock::iterator I = Succ->begin();
3154 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3155 int Idx = PHI->getBasicBlockIndex(BB);
3156 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3158 Value *InValue = PHI->getIncomingValue(Idx);
3159 if (InValue != CaseValue) continue;
3168 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
3169 /// instruction to a phi node dominated by the switch, if that would mean that
3170 /// some of the destination blocks of the switch can be folded away.
3171 /// Returns true if a change is made.
3172 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3173 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3174 ForwardingNodesMap ForwardingNodes;
3176 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3177 ConstantInt *CaseValue = I.getCaseValue();
3178 BasicBlock *CaseDest = I.getCaseSuccessor();
3181 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3185 ForwardingNodes[PHI].push_back(PhiIndex);
3188 bool Changed = false;
3190 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3191 E = ForwardingNodes.end(); I != E; ++I) {
3192 PHINode *Phi = I->first;
3193 SmallVector<int,4> &Indexes = I->second;
3195 if (Indexes.size() < 2) continue;
3197 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3198 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3205 /// ValidLookupTableConstant - Return true if the backend will be able to handle
3206 /// initializing an array of constants like C.
3207 static bool ValidLookupTableConstant(Constant *C) {
3208 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3209 return CE->isGEPWithNoNotionalOverIndexing();
3211 return isa<ConstantFP>(C) ||
3212 isa<ConstantInt>(C) ||
3213 isa<ConstantPointerNull>(C) ||
3214 isa<GlobalValue>(C) ||
3218 /// LookupConstant - If V is a Constant, return it. Otherwise, try to look up
3219 /// its constant value in ConstantPool, returning 0 if it's not there.
3220 static Constant *LookupConstant(Value *V,
3221 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3222 if (Constant *C = dyn_cast<Constant>(V))
3224 return ConstantPool.lookup(V);
3227 /// ConstantFold - Try to fold instruction I into a constant. This works for
3228 /// simple instructions such as binary operations where both operands are
3229 /// constant or can be replaced by constants from the ConstantPool. Returns the
3230 /// resulting constant on success, 0 otherwise.
3231 static Constant *ConstantFold(Instruction *I,
3232 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3233 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
3234 Constant *A = LookupConstant(BO->getOperand(0), ConstantPool);
3237 Constant *B = LookupConstant(BO->getOperand(1), ConstantPool);
3240 return ConstantExpr::get(BO->getOpcode(), A, B);
3243 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
3244 Constant *A = LookupConstant(I->getOperand(0), ConstantPool);
3247 Constant *B = LookupConstant(I->getOperand(1), ConstantPool);
3250 return ConstantExpr::getCompare(Cmp->getPredicate(), A, B);
3253 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3254 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3257 if (A->isAllOnesValue())
3258 return LookupConstant(Select->getTrueValue(), ConstantPool);
3259 if (A->isNullValue())
3260 return LookupConstant(Select->getFalseValue(), ConstantPool);
3264 if (CastInst *Cast = dyn_cast<CastInst>(I)) {
3265 Constant *A = LookupConstant(I->getOperand(0), ConstantPool);
3268 return ConstantExpr::getCast(Cast->getOpcode(), A, Cast->getDestTy());
3274 /// GetCaseResults - Try to determine the resulting constant values in phi nodes
3275 /// at the common destination basic block, *CommonDest, for one of the case
3276 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3277 /// case), of a switch instruction SI.
3278 static bool GetCaseResults(SwitchInst *SI,
3279 ConstantInt *CaseVal,
3280 BasicBlock *CaseDest,
3281 BasicBlock **CommonDest,
3282 SmallVector<std::pair<PHINode*,Constant*>, 4> &Res) {
3283 // The block from which we enter the common destination.
3284 BasicBlock *Pred = SI->getParent();
3286 // If CaseDest is empty except for some side-effect free instructions through
3287 // which we can constant-propagate the CaseVal, continue to its successor.
3288 SmallDenseMap<Value*, Constant*> ConstantPool;
3289 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3290 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3292 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3293 // If the terminator is a simple branch, continue to the next block.
3294 if (T->getNumSuccessors() != 1)
3297 CaseDest = T->getSuccessor(0);
3298 } else if (isa<DbgInfoIntrinsic>(I)) {
3299 // Skip debug intrinsic.
3301 } else if (Constant *C = ConstantFold(I, ConstantPool)) {
3302 // Instruction is side-effect free and constant.
3303 ConstantPool.insert(std::make_pair(I, C));
3309 // If we did not have a CommonDest before, use the current one.
3311 *CommonDest = CaseDest;
3312 // If the destination isn't the common one, abort.
3313 if (CaseDest != *CommonDest)
3316 // Get the values for this case from phi nodes in the destination block.
3317 BasicBlock::iterator I = (*CommonDest)->begin();
3318 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3319 int Idx = PHI->getBasicBlockIndex(Pred);
3323 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3328 // Note: If the constant comes from constant-propagating the case value
3329 // through the CaseDest basic block, it will be safe to remove the
3330 // instructions in that block. They cannot be used (except in the phi nodes
3331 // we visit) outside CaseDest, because that block does not dominate its
3332 // successor. If it did, we would not be in this phi node.
3334 // Be conservative about which kinds of constants we support.
3335 if (!ValidLookupTableConstant(ConstVal))
3338 Res.push_back(std::make_pair(PHI, ConstVal));
3345 /// SwitchLookupTable - This class represents a lookup table that can be used
3346 /// to replace a switch.
3347 class SwitchLookupTable {
3349 /// SwitchLookupTable - Create a lookup table to use as a switch replacement
3350 /// with the contents of Values, using DefaultValue to fill any holes in the
3352 SwitchLookupTable(Module &M,
3354 ConstantInt *Offset,
3355 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values,
3356 Constant *DefaultValue,
3357 const DataLayout *TD);
3359 /// BuildLookup - Build instructions with Builder to retrieve the value at
3360 /// the position given by Index in the lookup table.
3361 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3363 /// WouldFitInRegister - Return true if a table with TableSize elements of
3364 /// type ElementType would fit in a target-legal register.
3365 static bool WouldFitInRegister(const DataLayout *TD,
3367 const Type *ElementType);
3370 // Depending on the contents of the table, it can be represented in
3373 // For tables where each element contains the same value, we just have to
3374 // store that single value and return it for each lookup.
3377 // For small tables with integer elements, we can pack them into a bitmap
3378 // that fits into a target-legal register. Values are retrieved by
3379 // shift and mask operations.
3382 // The table is stored as an array of values. Values are retrieved by load
3383 // instructions from the table.
3387 // For SingleValueKind, this is the single value.
3388 Constant *SingleValue;
3390 // For BitMapKind, this is the bitmap.
3391 ConstantInt *BitMap;
3392 IntegerType *BitMapElementTy;
3394 // For ArrayKind, this is the array.
3395 GlobalVariable *Array;
3399 SwitchLookupTable::SwitchLookupTable(Module &M,
3401 ConstantInt *Offset,
3402 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values,
3403 Constant *DefaultValue,
3404 const DataLayout *TD)
3405 : SingleValue(0), BitMap(0), BitMapElementTy(0), Array(0) {
3406 assert(Values.size() && "Can't build lookup table without values!");
3407 assert(TableSize >= Values.size() && "Can't fit values in table!");
3409 // If all values in the table are equal, this is that value.
3410 SingleValue = Values.begin()->second;
3412 // Build up the table contents.
3413 SmallVector<Constant*, 64> TableContents(TableSize);
3414 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3415 ConstantInt *CaseVal = Values[I].first;
3416 Constant *CaseRes = Values[I].second;
3417 assert(CaseRes->getType() == DefaultValue->getType());
3419 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3421 TableContents[Idx] = CaseRes;
3423 if (CaseRes != SingleValue)
3427 // Fill in any holes in the table with the default result.
3428 if (Values.size() < TableSize) {
3429 for (uint64_t I = 0; I < TableSize; ++I) {
3430 if (!TableContents[I])
3431 TableContents[I] = DefaultValue;
3434 if (DefaultValue != SingleValue)
3438 // If each element in the table contains the same value, we only need to store
3439 // that single value.
3441 Kind = SingleValueKind;
3445 // If the type is integer and the table fits in a register, build a bitmap.
3446 if (WouldFitInRegister(TD, TableSize, DefaultValue->getType())) {
3447 IntegerType *IT = cast<IntegerType>(DefaultValue->getType());
3448 APInt TableInt(TableSize * IT->getBitWidth(), 0);
3449 for (uint64_t I = TableSize; I > 0; --I) {
3450 TableInt <<= IT->getBitWidth();
3451 // Insert values into the bitmap. Undef values are set to zero.
3452 if (!isa<UndefValue>(TableContents[I - 1])) {
3453 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
3454 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
3457 BitMap = ConstantInt::get(M.getContext(), TableInt);
3458 BitMapElementTy = IT;
3464 // Store the table in an array.
3465 ArrayType *ArrayTy = ArrayType::get(DefaultValue->getType(), TableSize);
3466 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3468 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3469 GlobalVariable::PrivateLinkage,
3472 Array->setUnnamedAddr(true);
3476 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
3478 case SingleValueKind:
3481 // Type of the bitmap (e.g. i59).
3482 IntegerType *MapTy = BitMap->getType();
3484 // Cast Index to the same type as the bitmap.
3485 // Note: The Index is <= the number of elements in the table, so
3486 // truncating it to the width of the bitmask is safe.
3487 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
3489 // Multiply the shift amount by the element width.
3490 ShiftAmt = Builder.CreateMul(ShiftAmt,
3491 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
3495 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
3496 "switch.downshift");
3498 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
3502 Value *GEPIndices[] = { Builder.getInt32(0), Index };
3503 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices,
3505 return Builder.CreateLoad(GEP, "switch.load");
3508 llvm_unreachable("Unknown lookup table kind!");
3511 bool SwitchLookupTable::WouldFitInRegister(const DataLayout *TD,
3513 const Type *ElementType) {
3516 const IntegerType *IT = dyn_cast<IntegerType>(ElementType);
3519 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
3520 // are <= 15, we could try to narrow the type.
3522 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
3523 if (TableSize >= UINT_MAX/IT->getBitWidth())
3525 return TD->fitsInLegalInteger(TableSize * IT->getBitWidth());
3528 /// ShouldBuildLookupTable - Determine whether a lookup table should be built
3529 /// for this switch, based on the number of caes, size of the table and the
3530 /// types of the results.
3531 static bool ShouldBuildLookupTable(SwitchInst *SI,
3533 const TargetTransformInfo &TTI,
3534 const DataLayout *TD,
3535 const SmallDenseMap<PHINode*, Type*>& ResultTypes) {
3536 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
3537 return false; // TableSize overflowed, or mul below might overflow.
3539 bool AllTablesFitInRegister = true;
3540 bool HasIllegalType = false;
3541 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(),
3542 E = ResultTypes.end(); I != E; ++I) {
3543 Type *Ty = I->second;
3545 // Saturate this flag to true.
3546 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
3548 // Saturate this flag to false.
3549 AllTablesFitInRegister = AllTablesFitInRegister &&
3550 SwitchLookupTable::WouldFitInRegister(TD, TableSize, Ty);
3552 // If both flags saturate, we're done. NOTE: This *only* works with
3553 // saturating flags, and all flags have to saturate first due to the
3554 // non-deterministic behavior of iterating over a dense map.
3555 if (HasIllegalType && !AllTablesFitInRegister)
3559 // If each table would fit in a register, we should build it anyway.
3560 if (AllTablesFitInRegister)
3563 // Don't build a table that doesn't fit in-register if it has illegal types.
3567 // The table density should be at least 40%. This is the same criterion as for
3568 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3569 // FIXME: Find the best cut-off.
3570 return SI->getNumCases() * 10 >= TableSize * 4;
3573 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3574 /// phi nodes in a common successor block with different constant values,
3575 /// replace the switch with lookup tables.
3576 static bool SwitchToLookupTable(SwitchInst *SI,
3577 IRBuilder<> &Builder,
3578 const TargetTransformInfo &TTI,
3579 const DataLayout* TD) {
3580 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3582 // Only build lookup table when we have a target that supports it.
3583 if (!TTI.shouldBuildLookupTables())
3586 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3587 // split off a dense part and build a lookup table for that.
3589 // FIXME: This creates arrays of GEPs to constant strings, which means each
3590 // GEP needs a runtime relocation in PIC code. We should just build one big
3591 // string and lookup indices into that.
3593 // Ignore the switch if the number of cases is too small.
3594 // This is similar to the check when building jump tables in
3595 // SelectionDAGBuilder::handleJTSwitchCase.
3596 // FIXME: Determine the best cut-off.
3597 if (SI->getNumCases() < 4)
3600 // Figure out the corresponding result for each case value and phi node in the
3601 // common destination, as well as the the min and max case values.
3602 assert(SI->case_begin() != SI->case_end());
3603 SwitchInst::CaseIt CI = SI->case_begin();
3604 ConstantInt *MinCaseVal = CI.getCaseValue();
3605 ConstantInt *MaxCaseVal = CI.getCaseValue();
3607 BasicBlock *CommonDest = 0;
3608 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
3609 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
3610 SmallDenseMap<PHINode*, Constant*> DefaultResults;
3611 SmallDenseMap<PHINode*, Type*> ResultTypes;
3612 SmallVector<PHINode*, 4> PHIs;
3614 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
3615 ConstantInt *CaseVal = CI.getCaseValue();
3616 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
3617 MinCaseVal = CaseVal;
3618 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
3619 MaxCaseVal = CaseVal;
3621 // Resulting value at phi nodes for this case value.
3622 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
3624 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
3628 // Append the result from this case to the list for each phi.
3629 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) {
3630 if (!ResultLists.count(I->first))
3631 PHIs.push_back(I->first);
3632 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second));
3636 // Get the resulting values for the default case.
3637 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
3638 if (!GetCaseResults(SI, 0, SI->getDefaultDest(), &CommonDest,
3639 DefaultResultsList))
3641 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) {
3642 PHINode *PHI = DefaultResultsList[I].first;
3643 Constant *Result = DefaultResultsList[I].second;
3644 DefaultResults[PHI] = Result;
3645 ResultTypes[PHI] = Result->getType();
3648 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
3649 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
3650 if (!ShouldBuildLookupTable(SI, TableSize, TTI, TD, ResultTypes))
3653 // Create the BB that does the lookups.
3654 Module &Mod = *CommonDest->getParent()->getParent();
3655 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
3657 CommonDest->getParent(),
3660 // Check whether the condition value is within the case range, and branch to
3662 Builder.SetInsertPoint(SI);
3663 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
3665 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
3666 MinCaseVal->getType(), TableSize));
3667 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
3669 // Populate the BB that does the lookups.
3670 Builder.SetInsertPoint(LookupBB);
3671 bool ReturnedEarly = false;
3672 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
3673 PHINode *PHI = PHIs[I];
3675 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI],
3676 DefaultResults[PHI], TD);
3678 Value *Result = Table.BuildLookup(TableIndex, Builder);
3680 // If the result is used to return immediately from the function, we want to
3681 // do that right here.
3682 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin()) &&
3683 *PHI->use_begin() == CommonDest->getFirstNonPHIOrDbg()) {
3684 Builder.CreateRet(Result);
3685 ReturnedEarly = true;
3689 PHI->addIncoming(Result, LookupBB);
3693 Builder.CreateBr(CommonDest);
3695 // Remove the switch.
3696 for (unsigned i = 0; i < SI->getNumSuccessors(); ++i) {
3697 BasicBlock *Succ = SI->getSuccessor(i);
3698 if (Succ == SI->getDefaultDest()) continue;
3699 Succ->removePredecessor(SI->getParent());
3701 SI->eraseFromParent();
3707 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
3708 BasicBlock *BB = SI->getParent();
3710 if (isValueEqualityComparison(SI)) {
3711 // If we only have one predecessor, and if it is a branch on this value,
3712 // see if that predecessor totally determines the outcome of this switch.
3713 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3714 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
3715 return SimplifyCFG(BB, TTI, TD) | true;
3717 Value *Cond = SI->getCondition();
3718 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
3719 if (SimplifySwitchOnSelect(SI, Select))
3720 return SimplifyCFG(BB, TTI, TD) | true;
3722 // If the block only contains the switch, see if we can fold the block
3723 // away into any preds.
3724 BasicBlock::iterator BBI = BB->begin();
3725 // Ignore dbg intrinsics.
3726 while (isa<DbgInfoIntrinsic>(BBI))
3729 if (FoldValueComparisonIntoPredecessors(SI, Builder))
3730 return SimplifyCFG(BB, TTI, TD) | true;
3733 // Try to transform the switch into an icmp and a branch.
3734 if (TurnSwitchRangeIntoICmp(SI, Builder))
3735 return SimplifyCFG(BB, TTI, TD) | true;
3737 // Remove unreachable cases.
3738 if (EliminateDeadSwitchCases(SI))
3739 return SimplifyCFG(BB, TTI, TD) | true;
3741 if (ForwardSwitchConditionToPHI(SI))
3742 return SimplifyCFG(BB, TTI, TD) | true;
3744 if (SwitchToLookupTable(SI, Builder, TTI, TD))
3745 return SimplifyCFG(BB, TTI, TD) | true;
3750 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
3751 BasicBlock *BB = IBI->getParent();
3752 bool Changed = false;
3754 // Eliminate redundant destinations.
3755 SmallPtrSet<Value *, 8> Succs;
3756 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
3757 BasicBlock *Dest = IBI->getDestination(i);
3758 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) {
3759 Dest->removePredecessor(BB);
3760 IBI->removeDestination(i);
3766 if (IBI->getNumDestinations() == 0) {
3767 // If the indirectbr has no successors, change it to unreachable.
3768 new UnreachableInst(IBI->getContext(), IBI);
3769 EraseTerminatorInstAndDCECond(IBI);
3773 if (IBI->getNumDestinations() == 1) {
3774 // If the indirectbr has one successor, change it to a direct branch.
3775 BranchInst::Create(IBI->getDestination(0), IBI);
3776 EraseTerminatorInstAndDCECond(IBI);
3780 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
3781 if (SimplifyIndirectBrOnSelect(IBI, SI))
3782 return SimplifyCFG(BB, TTI, TD) | true;
3787 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
3788 BasicBlock *BB = BI->getParent();
3790 if (SinkCommon && SinkThenElseCodeToEnd(BI))
3793 // If the Terminator is the only non-phi instruction, simplify the block.
3794 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime();
3795 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
3796 TryToSimplifyUncondBranchFromEmptyBlock(BB))
3799 // If the only instruction in the block is a seteq/setne comparison
3800 // against a constant, try to simplify the block.
3801 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
3802 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
3803 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
3805 if (I->isTerminator() &&
3806 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI, TD))
3810 // If this basic block is ONLY a compare and a branch, and if a predecessor
3811 // branches to us and our successor, fold the comparison into the
3812 // predecessor and use logical operations to update the incoming value
3813 // for PHI nodes in common successor.
3814 if (FoldBranchToCommonDest(BI))
3815 return SimplifyCFG(BB, TTI, TD) | true;
3820 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
3821 BasicBlock *BB = BI->getParent();
3823 // Conditional branch
3824 if (isValueEqualityComparison(BI)) {
3825 // If we only have one predecessor, and if it is a branch on this value,
3826 // see if that predecessor totally determines the outcome of this
3828 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3829 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
3830 return SimplifyCFG(BB, TTI, TD) | true;
3832 // This block must be empty, except for the setcond inst, if it exists.
3833 // Ignore dbg intrinsics.
3834 BasicBlock::iterator I = BB->begin();
3835 // Ignore dbg intrinsics.
3836 while (isa<DbgInfoIntrinsic>(I))
3839 if (FoldValueComparisonIntoPredecessors(BI, Builder))
3840 return SimplifyCFG(BB, TTI, TD) | true;
3841 } else if (&*I == cast<Instruction>(BI->getCondition())){
3843 // Ignore dbg intrinsics.
3844 while (isa<DbgInfoIntrinsic>(I))
3846 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
3847 return SimplifyCFG(BB, TTI, TD) | true;
3851 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
3852 if (SimplifyBranchOnICmpChain(BI, TD, Builder))
3855 // If this basic block is ONLY a compare and a branch, and if a predecessor
3856 // branches to us and one of our successors, fold the comparison into the
3857 // predecessor and use logical operations to pick the right destination.
3858 if (FoldBranchToCommonDest(BI))
3859 return SimplifyCFG(BB, TTI, TD) | true;
3861 // We have a conditional branch to two blocks that are only reachable
3862 // from BI. We know that the condbr dominates the two blocks, so see if
3863 // there is any identical code in the "then" and "else" blocks. If so, we
3864 // can hoist it up to the branching block.
3865 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) {
3866 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3867 if (HoistThenElseCodeToIf(BI))
3868 return SimplifyCFG(BB, TTI, TD) | true;
3870 // If Successor #1 has multiple preds, we may be able to conditionally
3871 // execute Successor #0 if it branches to successor #1.
3872 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
3873 if (Succ0TI->getNumSuccessors() == 1 &&
3874 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
3875 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0)))
3876 return SimplifyCFG(BB, TTI, TD) | true;
3878 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3879 // If Successor #0 has multiple preds, we may be able to conditionally
3880 // execute Successor #1 if it branches to successor #0.
3881 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
3882 if (Succ1TI->getNumSuccessors() == 1 &&
3883 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
3884 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1)))
3885 return SimplifyCFG(BB, TTI, TD) | true;
3888 // If this is a branch on a phi node in the current block, thread control
3889 // through this block if any PHI node entries are constants.
3890 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
3891 if (PN->getParent() == BI->getParent())
3892 if (FoldCondBranchOnPHI(BI, TD))
3893 return SimplifyCFG(BB, TTI, TD) | true;
3895 // Scan predecessor blocks for conditional branches.
3896 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
3897 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
3898 if (PBI != BI && PBI->isConditional())
3899 if (SimplifyCondBranchToCondBranch(PBI, BI))
3900 return SimplifyCFG(BB, TTI, TD) | true;
3905 /// Check if passing a value to an instruction will cause undefined behavior.
3906 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
3907 Constant *C = dyn_cast<Constant>(V);
3914 if (C->isNullValue()) {
3915 // Only look at the first use, avoid hurting compile time with long uselists
3916 User *Use = *I->use_begin();
3918 // Now make sure that there are no instructions in between that can alter
3919 // control flow (eg. calls)
3920 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
3921 if (i == I->getParent()->end() || i->mayHaveSideEffects())
3924 // Look through GEPs. A load from a GEP derived from NULL is still undefined
3925 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
3926 if (GEP->getPointerOperand() == I)
3927 return passingValueIsAlwaysUndefined(V, GEP);
3929 // Look through bitcasts.
3930 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
3931 return passingValueIsAlwaysUndefined(V, BC);
3933 // Load from null is undefined.
3934 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
3935 return LI->getPointerAddressSpace() == 0;
3937 // Store to null is undefined.
3938 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
3939 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
3944 /// If BB has an incoming value that will always trigger undefined behavior
3945 /// (eg. null pointer dereference), remove the branch leading here.
3946 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
3947 for (BasicBlock::iterator i = BB->begin();
3948 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
3949 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
3950 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
3951 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
3952 IRBuilder<> Builder(T);
3953 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
3954 BB->removePredecessor(PHI->getIncomingBlock(i));
3955 // Turn uncoditional branches into unreachables and remove the dead
3956 // destination from conditional branches.
3957 if (BI->isUnconditional())
3958 Builder.CreateUnreachable();
3960 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
3961 BI->getSuccessor(0));
3962 BI->eraseFromParent();
3965 // TODO: SwitchInst.
3971 bool SimplifyCFGOpt::run(BasicBlock *BB) {
3972 bool Changed = false;
3974 assert(BB && BB->getParent() && "Block not embedded in function!");
3975 assert(BB->getTerminator() && "Degenerate basic block encountered!");
3977 // Remove basic blocks that have no predecessors (except the entry block)...
3978 // or that just have themself as a predecessor. These are unreachable.
3979 if ((pred_begin(BB) == pred_end(BB) &&
3980 BB != &BB->getParent()->getEntryBlock()) ||
3981 BB->getSinglePredecessor() == BB) {
3982 DEBUG(dbgs() << "Removing BB: \n" << *BB);
3983 DeleteDeadBlock(BB);
3987 // Check to see if we can constant propagate this terminator instruction
3989 Changed |= ConstantFoldTerminator(BB, true);
3991 // Check for and eliminate duplicate PHI nodes in this block.
3992 Changed |= EliminateDuplicatePHINodes(BB);
3994 // Check for and remove branches that will always cause undefined behavior.
3995 Changed |= removeUndefIntroducingPredecessor(BB);
3997 // Merge basic blocks into their predecessor if there is only one distinct
3998 // pred, and if there is only one distinct successor of the predecessor, and
3999 // if there are no PHI nodes.
4001 if (MergeBlockIntoPredecessor(BB))
4004 IRBuilder<> Builder(BB);
4006 // If there is a trivial two-entry PHI node in this basic block, and we can
4007 // eliminate it, do so now.
4008 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
4009 if (PN->getNumIncomingValues() == 2)
4010 Changed |= FoldTwoEntryPHINode(PN, TD);
4012 Builder.SetInsertPoint(BB->getTerminator());
4013 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
4014 if (BI->isUnconditional()) {
4015 if (SimplifyUncondBranch(BI, Builder)) return true;
4017 if (SimplifyCondBranch(BI, Builder)) return true;
4019 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
4020 if (SimplifyReturn(RI, Builder)) return true;
4021 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
4022 if (SimplifyResume(RI, Builder)) return true;
4023 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
4024 if (SimplifySwitch(SI, Builder)) return true;
4025 } else if (UnreachableInst *UI =
4026 dyn_cast<UnreachableInst>(BB->getTerminator())) {
4027 if (SimplifyUnreachable(UI)) return true;
4028 } else if (IndirectBrInst *IBI =
4029 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
4030 if (SimplifyIndirectBr(IBI)) return true;
4036 /// SimplifyCFG - This function is used to do simplification of a CFG. For
4037 /// example, it adjusts branches to branches to eliminate the extra hop, it
4038 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4039 /// of the CFG. It returns true if a modification was made.
4041 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
4042 const DataLayout *TD) {
4043 return SimplifyCFGOpt(TTI, TD).run(BB);