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 *ThenBB) {
1373 // Be conservative for now. FP select instruction can often be expensive.
1374 Value *BrCond = BI->getCondition();
1375 if (isa<FCmpInst>(BrCond))
1378 BasicBlock *BB = BI->getParent();
1379 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1381 // If ThenBB is actually on the false edge of the conditional branch, remember
1382 // to swap the select operands later.
1383 bool Invert = false;
1384 if (ThenBB != BI->getSuccessor(0)) {
1385 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1388 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1390 // Only speculatively execution a single instruction (not counting the
1391 // terminator) for now.
1392 Instruction *HInst = NULL;
1393 Instruction *Term = ThenBB->getTerminator();
1394 for (BasicBlock::iterator BBI = ThenBB->begin(), BBE = ThenBB->end();
1395 BBI != BBE; ++BBI) {
1396 Instruction *I = BBI;
1398 if (isa<DbgInfoIntrinsic>(I)) continue;
1399 if (I == Term) break;
1406 // Check the instruction to be hoisted, if there is one.
1408 // Don't hoist the instruction if it's unsafe or expensive.
1409 if (!isSafeToSpeculativelyExecute(HInst))
1411 if (ComputeSpeculationCost(HInst) > PHINodeFoldingThreshold)
1414 // Do not hoist the instruction if any of its operands are defined but not
1415 // used in this BB. The transformation will prevent the operand from
1416 // being sunk into the use block.
1417 for (User::op_iterator i = HInst->op_begin(), e = HInst->op_end();
1419 Instruction *OpI = dyn_cast<Instruction>(*i);
1420 if (OpI && OpI->getParent() == BB &&
1421 !OpI->mayHaveSideEffects() &&
1422 !OpI->isUsedInBasicBlock(BB))
1427 // Collect interesting PHIs, and scan for hazards.
1428 SmallSetVector<std::pair<Value *, Value *>, 4> PHIs;
1429 for (BasicBlock::iterator I = EndBB->begin();
1430 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1431 Value *OrigV = PN->getIncomingValueForBlock(BB);
1432 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1434 // Skip PHIs which are trivial.
1438 // Check for safety.
1439 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(ThenV)) {
1440 // An unfolded ConstantExpr could end up getting expanded into
1441 // Instructions. Don't speculate this and another instruction at
1445 if (!isSafeToSpeculativelyExecute(CE))
1447 if (ComputeSpeculationCost(CE) > PHINodeFoldingThreshold)
1451 // Ok, we may insert a select for this PHI.
1452 PHIs.insert(std::make_pair(ThenV, OrigV));
1455 // If there are no PHIs to process, bail early. This helps ensure idempotence
1460 // If we get here, we can hoist the instruction and if-convert.
1461 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1463 // Hoist the instruction.
1465 BB->getInstList().splice(BI, ThenBB->getInstList(), HInst);
1467 // Insert selects and rewrite the PHI operands.
1468 IRBuilder<true, NoFolder> Builder(BI);
1469 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
1470 Value *TrueV = PHIs[i].first;
1471 Value *FalseV = PHIs[i].second;
1473 // Create a select whose true value is the speculatively executed value and
1474 // false value is the previously determined FalseV.
1477 SI = cast<SelectInst>
1478 (Builder.CreateSelect(BrCond, FalseV, TrueV,
1479 FalseV->getName() + "." + TrueV->getName()));
1481 SI = cast<SelectInst>
1482 (Builder.CreateSelect(BrCond, TrueV, FalseV,
1483 TrueV->getName() + "." + FalseV->getName()));
1485 // Make the PHI node use the select for all incoming values for "then" and
1487 for (BasicBlock::iterator I = EndBB->begin();
1488 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1489 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1490 unsigned OrigI = PN->getBasicBlockIndex(BB);
1491 Value *ThenV = PN->getIncomingValue(ThenI);
1492 Value *OrigV = PN->getIncomingValue(OrigI);
1493 if (TrueV == ThenV && FalseV == OrigV) {
1494 PN->setIncomingValue(ThenI, SI);
1495 PN->setIncomingValue(OrigI, SI);
1504 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1505 /// across this block.
1506 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1507 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1510 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1511 if (isa<DbgInfoIntrinsic>(BBI))
1513 if (Size > 10) return false; // Don't clone large BB's.
1516 // We can only support instructions that do not define values that are
1517 // live outside of the current basic block.
1518 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end();
1520 Instruction *U = cast<Instruction>(*UI);
1521 if (U->getParent() != BB || isa<PHINode>(U)) return false;
1524 // Looks ok, continue checking.
1530 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1531 /// that is defined in the same block as the branch and if any PHI entries are
1532 /// constants, thread edges corresponding to that entry to be branches to their
1533 /// ultimate destination.
1534 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *TD) {
1535 BasicBlock *BB = BI->getParent();
1536 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1537 // NOTE: we currently cannot transform this case if the PHI node is used
1538 // outside of the block.
1539 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1542 // Degenerate case of a single entry PHI.
1543 if (PN->getNumIncomingValues() == 1) {
1544 FoldSingleEntryPHINodes(PN->getParent());
1548 // Now we know that this block has multiple preds and two succs.
1549 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1551 // Okay, this is a simple enough basic block. See if any phi values are
1553 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1554 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1555 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue;
1557 // Okay, we now know that all edges from PredBB should be revectored to
1558 // branch to RealDest.
1559 BasicBlock *PredBB = PN->getIncomingBlock(i);
1560 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1562 if (RealDest == BB) continue; // Skip self loops.
1563 // Skip if the predecessor's terminator is an indirect branch.
1564 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1566 // The dest block might have PHI nodes, other predecessors and other
1567 // difficult cases. Instead of being smart about this, just insert a new
1568 // block that jumps to the destination block, effectively splitting
1569 // the edge we are about to create.
1570 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1571 RealDest->getName()+".critedge",
1572 RealDest->getParent(), RealDest);
1573 BranchInst::Create(RealDest, EdgeBB);
1575 // Update PHI nodes.
1576 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1578 // BB may have instructions that are being threaded over. Clone these
1579 // instructions into EdgeBB. We know that there will be no uses of the
1580 // cloned instructions outside of EdgeBB.
1581 BasicBlock::iterator InsertPt = EdgeBB->begin();
1582 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1583 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1584 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1585 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1588 // Clone the instruction.
1589 Instruction *N = BBI->clone();
1590 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1592 // Update operands due to translation.
1593 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1595 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1596 if (PI != TranslateMap.end())
1600 // Check for trivial simplification.
1601 if (Value *V = SimplifyInstruction(N, TD)) {
1602 TranslateMap[BBI] = V;
1603 delete N; // Instruction folded away, don't need actual inst
1605 // Insert the new instruction into its new home.
1606 EdgeBB->getInstList().insert(InsertPt, N);
1607 if (!BBI->use_empty())
1608 TranslateMap[BBI] = N;
1612 // Loop over all of the edges from PredBB to BB, changing them to branch
1613 // to EdgeBB instead.
1614 TerminatorInst *PredBBTI = PredBB->getTerminator();
1615 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1616 if (PredBBTI->getSuccessor(i) == BB) {
1617 BB->removePredecessor(PredBB);
1618 PredBBTI->setSuccessor(i, EdgeBB);
1621 // Recurse, simplifying any other constants.
1622 return FoldCondBranchOnPHI(BI, TD) | true;
1628 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1629 /// PHI node, see if we can eliminate it.
1630 static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *TD) {
1631 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1632 // statement", which has a very simple dominance structure. Basically, we
1633 // are trying to find the condition that is being branched on, which
1634 // subsequently causes this merge to happen. We really want control
1635 // dependence information for this check, but simplifycfg can't keep it up
1636 // to date, and this catches most of the cases we care about anyway.
1637 BasicBlock *BB = PN->getParent();
1638 BasicBlock *IfTrue, *IfFalse;
1639 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1641 // Don't bother if the branch will be constant folded trivially.
1642 isa<ConstantInt>(IfCond))
1645 // Okay, we found that we can merge this two-entry phi node into a select.
1646 // Doing so would require us to fold *all* two entry phi nodes in this block.
1647 // At some point this becomes non-profitable (particularly if the target
1648 // doesn't support cmov's). Only do this transformation if there are two or
1649 // fewer PHI nodes in this block.
1650 unsigned NumPhis = 0;
1651 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1655 // Loop over the PHI's seeing if we can promote them all to select
1656 // instructions. While we are at it, keep track of the instructions
1657 // that need to be moved to the dominating block.
1658 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1659 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1660 MaxCostVal1 = PHINodeFoldingThreshold;
1662 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1663 PHINode *PN = cast<PHINode>(II++);
1664 if (Value *V = SimplifyInstruction(PN, TD)) {
1665 PN->replaceAllUsesWith(V);
1666 PN->eraseFromParent();
1670 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1672 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1677 // If we folded the first phi, PN dangles at this point. Refresh it. If
1678 // we ran out of PHIs then we simplified them all.
1679 PN = dyn_cast<PHINode>(BB->begin());
1680 if (PN == 0) return true;
1682 // Don't fold i1 branches on PHIs which contain binary operators. These can
1683 // often be turned into switches and other things.
1684 if (PN->getType()->isIntegerTy(1) &&
1685 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1686 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1687 isa<BinaryOperator>(IfCond)))
1690 // If we all PHI nodes are promotable, check to make sure that all
1691 // instructions in the predecessor blocks can be promoted as well. If
1692 // not, we won't be able to get rid of the control flow, so it's not
1693 // worth promoting to select instructions.
1694 BasicBlock *DomBlock = 0;
1695 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1696 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1697 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1700 DomBlock = *pred_begin(IfBlock1);
1701 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1702 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1703 // This is not an aggressive instruction that we can promote.
1704 // Because of this, we won't be able to get rid of the control
1705 // flow, so the xform is not worth it.
1710 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1713 DomBlock = *pred_begin(IfBlock2);
1714 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1715 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1716 // This is not an aggressive instruction that we can promote.
1717 // Because of this, we won't be able to get rid of the control
1718 // flow, so the xform is not worth it.
1723 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1724 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1726 // If we can still promote the PHI nodes after this gauntlet of tests,
1727 // do all of the PHI's now.
1728 Instruction *InsertPt = DomBlock->getTerminator();
1729 IRBuilder<true, NoFolder> Builder(InsertPt);
1731 // Move all 'aggressive' instructions, which are defined in the
1732 // conditional parts of the if's up to the dominating block.
1734 DomBlock->getInstList().splice(InsertPt,
1735 IfBlock1->getInstList(), IfBlock1->begin(),
1736 IfBlock1->getTerminator());
1738 DomBlock->getInstList().splice(InsertPt,
1739 IfBlock2->getInstList(), IfBlock2->begin(),
1740 IfBlock2->getTerminator());
1742 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1743 // Change the PHI node into a select instruction.
1744 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1745 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1748 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1749 PN->replaceAllUsesWith(NV);
1751 PN->eraseFromParent();
1754 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1755 // has been flattened. Change DomBlock to jump directly to our new block to
1756 // avoid other simplifycfg's kicking in on the diamond.
1757 TerminatorInst *OldTI = DomBlock->getTerminator();
1758 Builder.SetInsertPoint(OldTI);
1759 Builder.CreateBr(BB);
1760 OldTI->eraseFromParent();
1764 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1765 /// to two returning blocks, try to merge them together into one return,
1766 /// introducing a select if the return values disagree.
1767 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1768 IRBuilder<> &Builder) {
1769 assert(BI->isConditional() && "Must be a conditional branch");
1770 BasicBlock *TrueSucc = BI->getSuccessor(0);
1771 BasicBlock *FalseSucc = BI->getSuccessor(1);
1772 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1773 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1775 // Check to ensure both blocks are empty (just a return) or optionally empty
1776 // with PHI nodes. If there are other instructions, merging would cause extra
1777 // computation on one path or the other.
1778 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1780 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1783 Builder.SetInsertPoint(BI);
1784 // Okay, we found a branch that is going to two return nodes. If
1785 // there is no return value for this function, just change the
1786 // branch into a return.
1787 if (FalseRet->getNumOperands() == 0) {
1788 TrueSucc->removePredecessor(BI->getParent());
1789 FalseSucc->removePredecessor(BI->getParent());
1790 Builder.CreateRetVoid();
1791 EraseTerminatorInstAndDCECond(BI);
1795 // Otherwise, figure out what the true and false return values are
1796 // so we can insert a new select instruction.
1797 Value *TrueValue = TrueRet->getReturnValue();
1798 Value *FalseValue = FalseRet->getReturnValue();
1800 // Unwrap any PHI nodes in the return blocks.
1801 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1802 if (TVPN->getParent() == TrueSucc)
1803 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1804 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1805 if (FVPN->getParent() == FalseSucc)
1806 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1808 // In order for this transformation to be safe, we must be able to
1809 // unconditionally execute both operands to the return. This is
1810 // normally the case, but we could have a potentially-trapping
1811 // constant expression that prevents this transformation from being
1813 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1816 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1820 // Okay, we collected all the mapped values and checked them for sanity, and
1821 // defined to really do this transformation. First, update the CFG.
1822 TrueSucc->removePredecessor(BI->getParent());
1823 FalseSucc->removePredecessor(BI->getParent());
1825 // Insert select instructions where needed.
1826 Value *BrCond = BI->getCondition();
1828 // Insert a select if the results differ.
1829 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1830 } else if (isa<UndefValue>(TrueValue)) {
1831 TrueValue = FalseValue;
1833 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1834 FalseValue, "retval");
1838 Value *RI = !TrueValue ?
1839 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
1843 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
1844 << "\n " << *BI << "NewRet = " << *RI
1845 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
1847 EraseTerminatorInstAndDCECond(BI);
1852 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
1853 /// probabilities of the branch taking each edge. Fills in the two APInt
1854 /// parameters and return true, or returns false if no or invalid metadata was
1856 static bool ExtractBranchMetadata(BranchInst *BI,
1857 uint64_t &ProbTrue, uint64_t &ProbFalse) {
1858 assert(BI->isConditional() &&
1859 "Looking for probabilities on unconditional branch?");
1860 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
1861 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
1862 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
1863 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
1864 if (!CITrue || !CIFalse) return false;
1865 ProbTrue = CITrue->getValue().getZExtValue();
1866 ProbFalse = CIFalse->getValue().getZExtValue();
1870 /// checkCSEInPredecessor - Return true if the given instruction is available
1871 /// in its predecessor block. If yes, the instruction will be removed.
1873 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
1874 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
1876 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
1877 Instruction *PBI = &*I;
1878 // Check whether Inst and PBI generate the same value.
1879 if (Inst->isIdenticalTo(PBI)) {
1880 Inst->replaceAllUsesWith(PBI);
1881 Inst->eraseFromParent();
1888 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
1889 /// predecessor branches to us and one of our successors, fold the block into
1890 /// the predecessor and use logical operations to pick the right destination.
1891 bool llvm::FoldBranchToCommonDest(BranchInst *BI) {
1892 BasicBlock *BB = BI->getParent();
1894 Instruction *Cond = 0;
1895 if (BI->isConditional())
1896 Cond = dyn_cast<Instruction>(BI->getCondition());
1898 // For unconditional branch, check for a simple CFG pattern, where
1899 // BB has a single predecessor and BB's successor is also its predecessor's
1900 // successor. If such pattern exisits, check for CSE between BB and its
1902 if (BasicBlock *PB = BB->getSinglePredecessor())
1903 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
1904 if (PBI->isConditional() &&
1905 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
1906 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
1907 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
1909 Instruction *Curr = I++;
1910 if (isa<CmpInst>(Curr)) {
1914 // Quit if we can't remove this instruction.
1915 if (!checkCSEInPredecessor(Curr, PB))
1924 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
1925 Cond->getParent() != BB || !Cond->hasOneUse())
1928 // Only allow this if the condition is a simple instruction that can be
1929 // executed unconditionally. It must be in the same block as the branch, and
1930 // must be at the front of the block.
1931 BasicBlock::iterator FrontIt = BB->front();
1933 // Ignore dbg intrinsics.
1934 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
1936 // Allow a single instruction to be hoisted in addition to the compare
1937 // that feeds the branch. We later ensure that any values that _it_ uses
1938 // were also live in the predecessor, so that we don't unnecessarily create
1939 // register pressure or inhibit out-of-order execution.
1940 Instruction *BonusInst = 0;
1941 if (&*FrontIt != Cond &&
1942 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond &&
1943 isSafeToSpeculativelyExecute(FrontIt)) {
1944 BonusInst = &*FrontIt;
1947 // Ignore dbg intrinsics.
1948 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
1951 // Only a single bonus inst is allowed.
1952 if (&*FrontIt != Cond)
1955 // Make sure the instruction after the condition is the cond branch.
1956 BasicBlock::iterator CondIt = Cond; ++CondIt;
1958 // Ingore dbg intrinsics.
1959 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
1964 // Cond is known to be a compare or binary operator. Check to make sure that
1965 // neither operand is a potentially-trapping constant expression.
1966 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
1969 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
1973 // Finally, don't infinitely unroll conditional loops.
1974 BasicBlock *TrueDest = BI->getSuccessor(0);
1975 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0;
1976 if (TrueDest == BB || FalseDest == BB)
1979 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
1980 BasicBlock *PredBlock = *PI;
1981 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
1983 // Check that we have two conditional branches. If there is a PHI node in
1984 // the common successor, verify that the same value flows in from both
1986 SmallVector<PHINode*, 4> PHIs;
1987 if (PBI == 0 || PBI->isUnconditional() ||
1988 (BI->isConditional() &&
1989 !SafeToMergeTerminators(BI, PBI)) ||
1990 (!BI->isConditional() &&
1991 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
1994 // Determine if the two branches share a common destination.
1995 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
1996 bool InvertPredCond = false;
1998 if (BI->isConditional()) {
1999 if (PBI->getSuccessor(0) == TrueDest)
2000 Opc = Instruction::Or;
2001 else if (PBI->getSuccessor(1) == FalseDest)
2002 Opc = Instruction::And;
2003 else if (PBI->getSuccessor(0) == FalseDest)
2004 Opc = Instruction::And, InvertPredCond = true;
2005 else if (PBI->getSuccessor(1) == TrueDest)
2006 Opc = Instruction::Or, InvertPredCond = true;
2010 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2014 // Ensure that any values used in the bonus instruction are also used
2015 // by the terminator of the predecessor. This means that those values
2016 // must already have been resolved, so we won't be inhibiting the
2017 // out-of-order core by speculating them earlier.
2019 // Collect the values used by the bonus inst
2020 SmallPtrSet<Value*, 4> UsedValues;
2021 for (Instruction::op_iterator OI = BonusInst->op_begin(),
2022 OE = BonusInst->op_end(); OI != OE; ++OI) {
2024 if (!isa<Constant>(V))
2025 UsedValues.insert(V);
2028 SmallVector<std::pair<Value*, unsigned>, 4> Worklist;
2029 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0));
2031 // Walk up to four levels back up the use-def chain of the predecessor's
2032 // terminator to see if all those values were used. The choice of four
2033 // levels is arbitrary, to provide a compile-time-cost bound.
2034 while (!Worklist.empty()) {
2035 std::pair<Value*, unsigned> Pair = Worklist.back();
2036 Worklist.pop_back();
2038 if (Pair.second >= 4) continue;
2039 UsedValues.erase(Pair.first);
2040 if (UsedValues.empty()) break;
2042 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) {
2043 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
2045 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1));
2049 if (!UsedValues.empty()) return false;
2052 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2053 IRBuilder<> Builder(PBI);
2055 // If we need to invert the condition in the pred block to match, do so now.
2056 if (InvertPredCond) {
2057 Value *NewCond = PBI->getCondition();
2059 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2060 CmpInst *CI = cast<CmpInst>(NewCond);
2061 CI->setPredicate(CI->getInversePredicate());
2063 NewCond = Builder.CreateNot(NewCond,
2064 PBI->getCondition()->getName()+".not");
2067 PBI->setCondition(NewCond);
2068 PBI->swapSuccessors();
2071 // If we have a bonus inst, clone it into the predecessor block.
2072 Instruction *NewBonus = 0;
2074 NewBonus = BonusInst->clone();
2075 PredBlock->getInstList().insert(PBI, NewBonus);
2076 NewBonus->takeName(BonusInst);
2077 BonusInst->setName(BonusInst->getName()+".old");
2080 // Clone Cond into the predecessor basic block, and or/and the
2081 // two conditions together.
2082 Instruction *New = Cond->clone();
2083 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus);
2084 PredBlock->getInstList().insert(PBI, New);
2085 New->takeName(Cond);
2086 Cond->setName(New->getName()+".old");
2088 if (BI->isConditional()) {
2089 Instruction *NewCond =
2090 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2092 PBI->setCondition(NewCond);
2094 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2095 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2097 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2099 SmallVector<uint64_t, 8> NewWeights;
2101 if (PBI->getSuccessor(0) == BB) {
2102 if (PredHasWeights && SuccHasWeights) {
2103 // PBI: br i1 %x, BB, FalseDest
2104 // BI: br i1 %y, TrueDest, FalseDest
2105 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2106 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2107 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2108 // TrueWeight for PBI * FalseWeight for BI.
2109 // We assume that total weights of a BranchInst can fit into 32 bits.
2110 // Therefore, we will not have overflow using 64-bit arithmetic.
2111 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2112 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2114 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2115 PBI->setSuccessor(0, TrueDest);
2117 if (PBI->getSuccessor(1) == BB) {
2118 if (PredHasWeights && SuccHasWeights) {
2119 // PBI: br i1 %x, TrueDest, BB
2120 // BI: br i1 %y, TrueDest, FalseDest
2121 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2122 // FalseWeight for PBI * TrueWeight for BI.
2123 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2124 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2125 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2126 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2128 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2129 PBI->setSuccessor(1, FalseDest);
2131 if (NewWeights.size() == 2) {
2132 // Halve the weights if any of them cannot fit in an uint32_t
2133 FitWeights(NewWeights);
2135 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2136 PBI->setMetadata(LLVMContext::MD_prof,
2137 MDBuilder(BI->getContext()).
2138 createBranchWeights(MDWeights));
2140 PBI->setMetadata(LLVMContext::MD_prof, NULL);
2142 // Update PHI nodes in the common successors.
2143 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2144 ConstantInt *PBI_C = cast<ConstantInt>(
2145 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2146 assert(PBI_C->getType()->isIntegerTy(1));
2147 Instruction *MergedCond = 0;
2148 if (PBI->getSuccessor(0) == TrueDest) {
2149 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2150 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2151 // is false: !PBI_Cond and BI_Value
2152 Instruction *NotCond =
2153 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2156 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2161 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2162 PBI->getCondition(), MergedCond,
2165 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2166 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2167 // is false: PBI_Cond and BI_Value
2169 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2170 PBI->getCondition(), New,
2172 if (PBI_C->isOne()) {
2173 Instruction *NotCond =
2174 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2177 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2178 NotCond, MergedCond,
2183 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2186 // Change PBI from Conditional to Unconditional.
2187 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2188 EraseTerminatorInstAndDCECond(PBI);
2192 // TODO: If BB is reachable from all paths through PredBlock, then we
2193 // could replace PBI's branch probabilities with BI's.
2195 // Copy any debug value intrinsics into the end of PredBlock.
2196 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2197 if (isa<DbgInfoIntrinsic>(*I))
2198 I->clone()->insertBefore(PBI);
2205 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2206 /// predecessor of another block, this function tries to simplify it. We know
2207 /// that PBI and BI are both conditional branches, and BI is in one of the
2208 /// successor blocks of PBI - PBI branches to BI.
2209 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2210 assert(PBI->isConditional() && BI->isConditional());
2211 BasicBlock *BB = BI->getParent();
2213 // If this block ends with a branch instruction, and if there is a
2214 // predecessor that ends on a branch of the same condition, make
2215 // this conditional branch redundant.
2216 if (PBI->getCondition() == BI->getCondition() &&
2217 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2218 // Okay, the outcome of this conditional branch is statically
2219 // knowable. If this block had a single pred, handle specially.
2220 if (BB->getSinglePredecessor()) {
2221 // Turn this into a branch on constant.
2222 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2223 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2225 return true; // Nuke the branch on constant.
2228 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2229 // in the constant and simplify the block result. Subsequent passes of
2230 // simplifycfg will thread the block.
2231 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2232 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2233 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2234 std::distance(PB, PE),
2235 BI->getCondition()->getName() + ".pr",
2237 // Okay, we're going to insert the PHI node. Since PBI is not the only
2238 // predecessor, compute the PHI'd conditional value for all of the preds.
2239 // Any predecessor where the condition is not computable we keep symbolic.
2240 for (pred_iterator PI = PB; PI != PE; ++PI) {
2241 BasicBlock *P = *PI;
2242 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2243 PBI != BI && PBI->isConditional() &&
2244 PBI->getCondition() == BI->getCondition() &&
2245 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2246 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2247 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2250 NewPN->addIncoming(BI->getCondition(), P);
2254 BI->setCondition(NewPN);
2259 // If this is a conditional branch in an empty block, and if any
2260 // predecessors is a conditional branch to one of our destinations,
2261 // fold the conditions into logical ops and one cond br.
2262 BasicBlock::iterator BBI = BB->begin();
2263 // Ignore dbg intrinsics.
2264 while (isa<DbgInfoIntrinsic>(BBI))
2270 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2275 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2277 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2278 PBIOp = 0, BIOp = 1;
2279 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2280 PBIOp = 1, BIOp = 0;
2281 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2286 // Check to make sure that the other destination of this branch
2287 // isn't BB itself. If so, this is an infinite loop that will
2288 // keep getting unwound.
2289 if (PBI->getSuccessor(PBIOp) == BB)
2292 // Do not perform this transformation if it would require
2293 // insertion of a large number of select instructions. For targets
2294 // without predication/cmovs, this is a big pessimization.
2295 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2297 unsigned NumPhis = 0;
2298 for (BasicBlock::iterator II = CommonDest->begin();
2299 isa<PHINode>(II); ++II, ++NumPhis)
2300 if (NumPhis > 2) // Disable this xform.
2303 // Finally, if everything is ok, fold the branches to logical ops.
2304 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2306 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2307 << "AND: " << *BI->getParent());
2310 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2311 // branch in it, where one edge (OtherDest) goes back to itself but the other
2312 // exits. We don't *know* that the program avoids the infinite loop
2313 // (even though that seems likely). If we do this xform naively, we'll end up
2314 // recursively unpeeling the loop. Since we know that (after the xform is
2315 // done) that the block *is* infinite if reached, we just make it an obviously
2316 // infinite loop with no cond branch.
2317 if (OtherDest == BB) {
2318 // Insert it at the end of the function, because it's either code,
2319 // or it won't matter if it's hot. :)
2320 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2321 "infloop", BB->getParent());
2322 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2323 OtherDest = InfLoopBlock;
2326 DEBUG(dbgs() << *PBI->getParent()->getParent());
2328 // BI may have other predecessors. Because of this, we leave
2329 // it alone, but modify PBI.
2331 // Make sure we get to CommonDest on True&True directions.
2332 Value *PBICond = PBI->getCondition();
2333 IRBuilder<true, NoFolder> Builder(PBI);
2335 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2337 Value *BICond = BI->getCondition();
2339 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2341 // Merge the conditions.
2342 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2344 // Modify PBI to branch on the new condition to the new dests.
2345 PBI->setCondition(Cond);
2346 PBI->setSuccessor(0, CommonDest);
2347 PBI->setSuccessor(1, OtherDest);
2349 // Update branch weight for PBI.
2350 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2351 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2353 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2355 if (PredHasWeights && SuccHasWeights) {
2356 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2357 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2358 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2359 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2360 // The weight to CommonDest should be PredCommon * SuccTotal +
2361 // PredOther * SuccCommon.
2362 // The weight to OtherDest should be PredOther * SuccOther.
2363 SmallVector<uint64_t, 2> NewWeights;
2364 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) +
2365 PredOther * SuccCommon);
2366 NewWeights.push_back(PredOther * SuccOther);
2367 // Halve the weights if any of them cannot fit in an uint32_t
2368 FitWeights(NewWeights);
2370 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end());
2371 PBI->setMetadata(LLVMContext::MD_prof,
2372 MDBuilder(BI->getContext()).
2373 createBranchWeights(MDWeights));
2376 // OtherDest may have phi nodes. If so, add an entry from PBI's
2377 // block that are identical to the entries for BI's block.
2378 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2380 // We know that the CommonDest already had an edge from PBI to
2381 // it. If it has PHIs though, the PHIs may have different
2382 // entries for BB and PBI's BB. If so, insert a select to make
2385 for (BasicBlock::iterator II = CommonDest->begin();
2386 (PN = dyn_cast<PHINode>(II)); ++II) {
2387 Value *BIV = PN->getIncomingValueForBlock(BB);
2388 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2389 Value *PBIV = PN->getIncomingValue(PBBIdx);
2391 // Insert a select in PBI to pick the right value.
2392 Value *NV = cast<SelectInst>
2393 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2394 PN->setIncomingValue(PBBIdx, NV);
2398 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2399 DEBUG(dbgs() << *PBI->getParent()->getParent());
2401 // This basic block is probably dead. We know it has at least
2402 // one fewer predecessor.
2406 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2407 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2408 // Takes care of updating the successors and removing the old terminator.
2409 // Also makes sure not to introduce new successors by assuming that edges to
2410 // non-successor TrueBBs and FalseBBs aren't reachable.
2411 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2412 BasicBlock *TrueBB, BasicBlock *FalseBB,
2413 uint32_t TrueWeight,
2414 uint32_t FalseWeight){
2415 // Remove any superfluous successor edges from the CFG.
2416 // First, figure out which successors to preserve.
2417 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2419 BasicBlock *KeepEdge1 = TrueBB;
2420 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0;
2422 // Then remove the rest.
2423 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2424 BasicBlock *Succ = OldTerm->getSuccessor(I);
2425 // Make sure only to keep exactly one copy of each edge.
2426 if (Succ == KeepEdge1)
2428 else if (Succ == KeepEdge2)
2431 Succ->removePredecessor(OldTerm->getParent());
2434 IRBuilder<> Builder(OldTerm);
2435 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2437 // Insert an appropriate new terminator.
2438 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) {
2439 if (TrueBB == FalseBB)
2440 // We were only looking for one successor, and it was present.
2441 // Create an unconditional branch to it.
2442 Builder.CreateBr(TrueBB);
2444 // We found both of the successors we were looking for.
2445 // Create a conditional branch sharing the condition of the select.
2446 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2447 if (TrueWeight != FalseWeight)
2448 NewBI->setMetadata(LLVMContext::MD_prof,
2449 MDBuilder(OldTerm->getContext()).
2450 createBranchWeights(TrueWeight, FalseWeight));
2452 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2453 // Neither of the selected blocks were successors, so this
2454 // terminator must be unreachable.
2455 new UnreachableInst(OldTerm->getContext(), OldTerm);
2457 // One of the selected values was a successor, but the other wasn't.
2458 // Insert an unconditional branch to the one that was found;
2459 // the edge to the one that wasn't must be unreachable.
2461 // Only TrueBB was found.
2462 Builder.CreateBr(TrueBB);
2464 // Only FalseBB was found.
2465 Builder.CreateBr(FalseBB);
2468 EraseTerminatorInstAndDCECond(OldTerm);
2472 // SimplifySwitchOnSelect - Replaces
2473 // (switch (select cond, X, Y)) on constant X, Y
2474 // with a branch - conditional if X and Y lead to distinct BBs,
2475 // unconditional otherwise.
2476 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2477 // Check for constant integer values in the select.
2478 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2479 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2480 if (!TrueVal || !FalseVal)
2483 // Find the relevant condition and destinations.
2484 Value *Condition = Select->getCondition();
2485 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2486 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2488 // Get weight for TrueBB and FalseBB.
2489 uint32_t TrueWeight = 0, FalseWeight = 0;
2490 SmallVector<uint64_t, 8> Weights;
2491 bool HasWeights = HasBranchWeights(SI);
2493 GetBranchWeights(SI, Weights);
2494 if (Weights.size() == 1 + SI->getNumCases()) {
2495 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2496 getSuccessorIndex()];
2497 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2498 getSuccessorIndex()];
2502 // Perform the actual simplification.
2503 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2504 TrueWeight, FalseWeight);
2507 // SimplifyIndirectBrOnSelect - Replaces
2508 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2509 // blockaddress(@fn, BlockB)))
2511 // (br cond, BlockA, BlockB).
2512 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2513 // Check that both operands of the select are block addresses.
2514 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2515 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2519 // Extract the actual blocks.
2520 BasicBlock *TrueBB = TBA->getBasicBlock();
2521 BasicBlock *FalseBB = FBA->getBasicBlock();
2523 // Perform the actual simplification.
2524 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2528 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2529 /// instruction (a seteq/setne with a constant) as the only instruction in a
2530 /// block that ends with an uncond branch. We are looking for a very specific
2531 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2532 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2533 /// default value goes to an uncond block with a seteq in it, we get something
2536 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2538 /// %tmp = icmp eq i8 %A, 92
2541 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2543 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2544 /// the PHI, merging the third icmp into the switch.
2545 static bool TryToSimplifyUncondBranchWithICmpInIt(
2546 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI,
2547 const DataLayout *TD) {
2548 BasicBlock *BB = ICI->getParent();
2550 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2552 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2554 Value *V = ICI->getOperand(0);
2555 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2557 // The pattern we're looking for is where our only predecessor is a switch on
2558 // 'V' and this block is the default case for the switch. In this case we can
2559 // fold the compared value into the switch to simplify things.
2560 BasicBlock *Pred = BB->getSinglePredecessor();
2561 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false;
2563 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2564 if (SI->getCondition() != V)
2567 // If BB is reachable on a non-default case, then we simply know the value of
2568 // V in this block. Substitute it and constant fold the icmp instruction
2570 if (SI->getDefaultDest() != BB) {
2571 ConstantInt *VVal = SI->findCaseDest(BB);
2572 assert(VVal && "Should have a unique destination value");
2573 ICI->setOperand(0, VVal);
2575 if (Value *V = SimplifyInstruction(ICI, TD)) {
2576 ICI->replaceAllUsesWith(V);
2577 ICI->eraseFromParent();
2579 // BB is now empty, so it is likely to simplify away.
2580 return SimplifyCFG(BB, TTI, TD) | true;
2583 // Ok, the block is reachable from the default dest. If the constant we're
2584 // comparing exists in one of the other edges, then we can constant fold ICI
2586 if (SI->findCaseValue(Cst) != SI->case_default()) {
2588 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2589 V = ConstantInt::getFalse(BB->getContext());
2591 V = ConstantInt::getTrue(BB->getContext());
2593 ICI->replaceAllUsesWith(V);
2594 ICI->eraseFromParent();
2595 // BB is now empty, so it is likely to simplify away.
2596 return SimplifyCFG(BB, TTI, TD) | true;
2599 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2601 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2602 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back());
2603 if (PHIUse == 0 || PHIUse != &SuccBlock->front() ||
2604 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2607 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2609 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2610 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2612 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2613 std::swap(DefaultCst, NewCst);
2615 // Replace ICI (which is used by the PHI for the default value) with true or
2616 // false depending on if it is EQ or NE.
2617 ICI->replaceAllUsesWith(DefaultCst);
2618 ICI->eraseFromParent();
2620 // Okay, the switch goes to this block on a default value. Add an edge from
2621 // the switch to the merge point on the compared value.
2622 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2623 BB->getParent(), BB);
2624 SmallVector<uint64_t, 8> Weights;
2625 bool HasWeights = HasBranchWeights(SI);
2627 GetBranchWeights(SI, Weights);
2628 if (Weights.size() == 1 + SI->getNumCases()) {
2629 // Split weight for default case to case for "Cst".
2630 Weights[0] = (Weights[0]+1) >> 1;
2631 Weights.push_back(Weights[0]);
2633 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2634 SI->setMetadata(LLVMContext::MD_prof,
2635 MDBuilder(SI->getContext()).
2636 createBranchWeights(MDWeights));
2639 SI->addCase(Cst, NewBB);
2641 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2642 Builder.SetInsertPoint(NewBB);
2643 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2644 Builder.CreateBr(SuccBlock);
2645 PHIUse->addIncoming(NewCst, NewBB);
2649 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2650 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2651 /// fold it into a switch instruction if so.
2652 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *TD,
2653 IRBuilder<> &Builder) {
2654 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2655 if (Cond == 0) return false;
2658 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2659 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2660 // 'setne's and'ed together, collect them.
2662 std::vector<ConstantInt*> Values;
2663 bool TrueWhenEqual = true;
2664 Value *ExtraCase = 0;
2665 unsigned UsedICmps = 0;
2667 if (Cond->getOpcode() == Instruction::Or) {
2668 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true,
2670 } else if (Cond->getOpcode() == Instruction::And) {
2671 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false,
2673 TrueWhenEqual = false;
2676 // If we didn't have a multiply compared value, fail.
2677 if (CompVal == 0) return false;
2679 // Avoid turning single icmps into a switch.
2683 // There might be duplicate constants in the list, which the switch
2684 // instruction can't handle, remove them now.
2685 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2686 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2688 // If Extra was used, we require at least two switch values to do the
2689 // transformation. A switch with one value is just an cond branch.
2690 if (ExtraCase && Values.size() < 2) return false;
2692 // TODO: Preserve branch weight metadata, similarly to how
2693 // FoldValueComparisonIntoPredecessors preserves it.
2695 // Figure out which block is which destination.
2696 BasicBlock *DefaultBB = BI->getSuccessor(1);
2697 BasicBlock *EdgeBB = BI->getSuccessor(0);
2698 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2700 BasicBlock *BB = BI->getParent();
2702 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2703 << " cases into SWITCH. BB is:\n" << *BB);
2705 // If there are any extra values that couldn't be folded into the switch
2706 // then we evaluate them with an explicit branch first. Split the block
2707 // right before the condbr to handle it.
2709 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2710 // Remove the uncond branch added to the old block.
2711 TerminatorInst *OldTI = BB->getTerminator();
2712 Builder.SetInsertPoint(OldTI);
2715 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2717 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2719 OldTI->eraseFromParent();
2721 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2722 // for the edge we just added.
2723 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2725 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2726 << "\nEXTRABB = " << *BB);
2730 Builder.SetInsertPoint(BI);
2731 // Convert pointer to int before we switch.
2732 if (CompVal->getType()->isPointerTy()) {
2733 assert(TD && "Cannot switch on pointer without DataLayout");
2734 CompVal = Builder.CreatePtrToInt(CompVal,
2735 TD->getIntPtrType(CompVal->getContext()),
2739 // Create the new switch instruction now.
2740 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2742 // Add all of the 'cases' to the switch instruction.
2743 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2744 New->addCase(Values[i], EdgeBB);
2746 // We added edges from PI to the EdgeBB. As such, if there were any
2747 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2748 // the number of edges added.
2749 for (BasicBlock::iterator BBI = EdgeBB->begin();
2750 isa<PHINode>(BBI); ++BBI) {
2751 PHINode *PN = cast<PHINode>(BBI);
2752 Value *InVal = PN->getIncomingValueForBlock(BB);
2753 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2754 PN->addIncoming(InVal, BB);
2757 // Erase the old branch instruction.
2758 EraseTerminatorInstAndDCECond(BI);
2760 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2764 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2765 // If this is a trivial landing pad that just continues unwinding the caught
2766 // exception then zap the landing pad, turning its invokes into calls.
2767 BasicBlock *BB = RI->getParent();
2768 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2769 if (RI->getValue() != LPInst)
2770 // Not a landing pad, or the resume is not unwinding the exception that
2771 // caused control to branch here.
2774 // Check that there are no other instructions except for debug intrinsics.
2775 BasicBlock::iterator I = LPInst, E = RI;
2777 if (!isa<DbgInfoIntrinsic>(I))
2780 // Turn all invokes that unwind here into calls and delete the basic block.
2781 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2782 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2783 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2784 // Insert a call instruction before the invoke.
2785 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2787 Call->setCallingConv(II->getCallingConv());
2788 Call->setAttributes(II->getAttributes());
2789 Call->setDebugLoc(II->getDebugLoc());
2791 // Anything that used the value produced by the invoke instruction now uses
2792 // the value produced by the call instruction. Note that we do this even
2793 // for void functions and calls with no uses so that the callgraph edge is
2795 II->replaceAllUsesWith(Call);
2796 BB->removePredecessor(II->getParent());
2798 // Insert a branch to the normal destination right before the invoke.
2799 BranchInst::Create(II->getNormalDest(), II);
2801 // Finally, delete the invoke instruction!
2802 II->eraseFromParent();
2805 // The landingpad is now unreachable. Zap it.
2806 BB->eraseFromParent();
2810 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2811 BasicBlock *BB = RI->getParent();
2812 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2814 // Find predecessors that end with branches.
2815 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2816 SmallVector<BranchInst*, 8> CondBranchPreds;
2817 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2818 BasicBlock *P = *PI;
2819 TerminatorInst *PTI = P->getTerminator();
2820 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2821 if (BI->isUnconditional())
2822 UncondBranchPreds.push_back(P);
2824 CondBranchPreds.push_back(BI);
2828 // If we found some, do the transformation!
2829 if (!UncondBranchPreds.empty() && DupRet) {
2830 while (!UncondBranchPreds.empty()) {
2831 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
2832 DEBUG(dbgs() << "FOLDING: " << *BB
2833 << "INTO UNCOND BRANCH PRED: " << *Pred);
2834 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
2837 // If we eliminated all predecessors of the block, delete the block now.
2838 if (pred_begin(BB) == pred_end(BB))
2839 // We know there are no successors, so just nuke the block.
2840 BB->eraseFromParent();
2845 // Check out all of the conditional branches going to this return
2846 // instruction. If any of them just select between returns, change the
2847 // branch itself into a select/return pair.
2848 while (!CondBranchPreds.empty()) {
2849 BranchInst *BI = CondBranchPreds.pop_back_val();
2851 // Check to see if the non-BB successor is also a return block.
2852 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
2853 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
2854 SimplifyCondBranchToTwoReturns(BI, Builder))
2860 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
2861 BasicBlock *BB = UI->getParent();
2863 bool Changed = false;
2865 // If there are any instructions immediately before the unreachable that can
2866 // be removed, do so.
2867 while (UI != BB->begin()) {
2868 BasicBlock::iterator BBI = UI;
2870 // Do not delete instructions that can have side effects which might cause
2871 // the unreachable to not be reachable; specifically, calls and volatile
2872 // operations may have this effect.
2873 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
2875 if (BBI->mayHaveSideEffects()) {
2876 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
2877 if (SI->isVolatile())
2879 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
2880 if (LI->isVolatile())
2882 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
2883 if (RMWI->isVolatile())
2885 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
2886 if (CXI->isVolatile())
2888 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
2889 !isa<LandingPadInst>(BBI)) {
2892 // Note that deleting LandingPad's here is in fact okay, although it
2893 // involves a bit of subtle reasoning. If this inst is a LandingPad,
2894 // all the predecessors of this block will be the unwind edges of Invokes,
2895 // and we can therefore guarantee this block will be erased.
2898 // Delete this instruction (any uses are guaranteed to be dead)
2899 if (!BBI->use_empty())
2900 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
2901 BBI->eraseFromParent();
2905 // If the unreachable instruction is the first in the block, take a gander
2906 // at all of the predecessors of this instruction, and simplify them.
2907 if (&BB->front() != UI) return Changed;
2909 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
2910 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
2911 TerminatorInst *TI = Preds[i]->getTerminator();
2912 IRBuilder<> Builder(TI);
2913 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
2914 if (BI->isUnconditional()) {
2915 if (BI->getSuccessor(0) == BB) {
2916 new UnreachableInst(TI->getContext(), TI);
2917 TI->eraseFromParent();
2921 if (BI->getSuccessor(0) == BB) {
2922 Builder.CreateBr(BI->getSuccessor(1));
2923 EraseTerminatorInstAndDCECond(BI);
2924 } else if (BI->getSuccessor(1) == BB) {
2925 Builder.CreateBr(BI->getSuccessor(0));
2926 EraseTerminatorInstAndDCECond(BI);
2930 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
2931 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2933 if (i.getCaseSuccessor() == BB) {
2934 BB->removePredecessor(SI->getParent());
2939 // If the default value is unreachable, figure out the most popular
2940 // destination and make it the default.
2941 if (SI->getDefaultDest() == BB) {
2942 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
2943 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2945 std::pair<unsigned, unsigned> &entry =
2946 Popularity[i.getCaseSuccessor()];
2947 if (entry.first == 0) {
2949 entry.second = i.getCaseIndex();
2955 // Find the most popular block.
2956 unsigned MaxPop = 0;
2957 unsigned MaxIndex = 0;
2958 BasicBlock *MaxBlock = 0;
2959 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
2960 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
2961 if (I->second.first > MaxPop ||
2962 (I->second.first == MaxPop && MaxIndex > I->second.second)) {
2963 MaxPop = I->second.first;
2964 MaxIndex = I->second.second;
2965 MaxBlock = I->first;
2969 // Make this the new default, allowing us to delete any explicit
2971 SI->setDefaultDest(MaxBlock);
2974 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
2976 if (isa<PHINode>(MaxBlock->begin()))
2977 for (unsigned i = 0; i != MaxPop-1; ++i)
2978 MaxBlock->removePredecessor(SI->getParent());
2980 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2982 if (i.getCaseSuccessor() == MaxBlock) {
2988 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
2989 if (II->getUnwindDest() == BB) {
2990 // Convert the invoke to a call instruction. This would be a good
2991 // place to note that the call does not throw though.
2992 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
2993 II->removeFromParent(); // Take out of symbol table
2995 // Insert the call now...
2996 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
2997 Builder.SetInsertPoint(BI);
2998 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
2999 Args, II->getName());
3000 CI->setCallingConv(II->getCallingConv());
3001 CI->setAttributes(II->getAttributes());
3002 // If the invoke produced a value, the call does now instead.
3003 II->replaceAllUsesWith(CI);
3010 // If this block is now dead, remove it.
3011 if (pred_begin(BB) == pred_end(BB) &&
3012 BB != &BB->getParent()->getEntryBlock()) {
3013 // We know there are no successors, so just nuke the block.
3014 BB->eraseFromParent();
3021 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
3022 /// integer range comparison into a sub, an icmp and a branch.
3023 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3024 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3026 // Make sure all cases point to the same destination and gather the values.
3027 SmallVector<ConstantInt *, 16> Cases;
3028 SwitchInst::CaseIt I = SI->case_begin();
3029 Cases.push_back(I.getCaseValue());
3030 SwitchInst::CaseIt PrevI = I++;
3031 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
3032 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
3034 Cases.push_back(I.getCaseValue());
3036 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
3038 // Sort the case values, then check if they form a range we can transform.
3039 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3040 for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
3041 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
3045 Constant *Offset = ConstantExpr::getNeg(Cases.back());
3046 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
3048 Value *Sub = SI->getCondition();
3049 if (!Offset->isNullValue())
3050 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
3051 Value *Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3052 BranchInst *NewBI = Builder.CreateCondBr(
3053 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
3055 // Update weight for the newly-created conditional branch.
3056 SmallVector<uint64_t, 8> Weights;
3057 bool HasWeights = HasBranchWeights(SI);
3059 GetBranchWeights(SI, Weights);
3060 if (Weights.size() == 1 + SI->getNumCases()) {
3061 // Combine all weights for the cases to be the true weight of NewBI.
3062 // We assume that the sum of all weights for a Terminator can fit into 32
3064 uint32_t NewTrueWeight = 0;
3065 for (unsigned I = 1, E = Weights.size(); I != E; ++I)
3066 NewTrueWeight += (uint32_t)Weights[I];
3067 NewBI->setMetadata(LLVMContext::MD_prof,
3068 MDBuilder(SI->getContext()).
3069 createBranchWeights(NewTrueWeight,
3070 (uint32_t)Weights[0]));
3074 // Prune obsolete incoming values off the successor's PHI nodes.
3075 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
3076 isa<PHINode>(BBI); ++BBI) {
3077 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
3078 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3080 SI->eraseFromParent();
3085 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
3086 /// and use it to remove dead cases.
3087 static bool EliminateDeadSwitchCases(SwitchInst *SI) {
3088 Value *Cond = SI->getCondition();
3089 unsigned Bits = cast<IntegerType>(Cond->getType())->getBitWidth();
3090 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3091 ComputeMaskedBits(Cond, KnownZero, KnownOne);
3093 // Gather dead cases.
3094 SmallVector<ConstantInt*, 8> DeadCases;
3095 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3096 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3097 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3098 DeadCases.push_back(I.getCaseValue());
3099 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3100 << I.getCaseValue() << "' is dead.\n");
3104 SmallVector<uint64_t, 8> Weights;
3105 bool HasWeight = HasBranchWeights(SI);
3107 GetBranchWeights(SI, Weights);
3108 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3111 // Remove dead cases from the switch.
3112 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3113 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3114 assert(Case != SI->case_default() &&
3115 "Case was not found. Probably mistake in DeadCases forming.");
3117 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3121 // Prune unused values from PHI nodes.
3122 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3123 SI->removeCase(Case);
3126 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3127 SI->setMetadata(LLVMContext::MD_prof,
3128 MDBuilder(SI->getParent()->getContext()).
3129 createBranchWeights(MDWeights));
3132 return !DeadCases.empty();
3135 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
3136 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3137 /// by an unconditional branch), look at the phi node for BB in the successor
3138 /// block and see if the incoming value is equal to CaseValue. If so, return
3139 /// the phi node, and set PhiIndex to BB's index in the phi node.
3140 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3143 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3144 return NULL; // BB must be empty to be a candidate for simplification.
3145 if (!BB->getSinglePredecessor())
3146 return NULL; // BB must be dominated by the switch.
3148 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3149 if (!Branch || !Branch->isUnconditional())
3150 return NULL; // Terminator must be unconditional branch.
3152 BasicBlock *Succ = Branch->getSuccessor(0);
3154 BasicBlock::iterator I = Succ->begin();
3155 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3156 int Idx = PHI->getBasicBlockIndex(BB);
3157 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3159 Value *InValue = PHI->getIncomingValue(Idx);
3160 if (InValue != CaseValue) continue;
3169 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
3170 /// instruction to a phi node dominated by the switch, if that would mean that
3171 /// some of the destination blocks of the switch can be folded away.
3172 /// Returns true if a change is made.
3173 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3174 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3175 ForwardingNodesMap ForwardingNodes;
3177 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3178 ConstantInt *CaseValue = I.getCaseValue();
3179 BasicBlock *CaseDest = I.getCaseSuccessor();
3182 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3186 ForwardingNodes[PHI].push_back(PhiIndex);
3189 bool Changed = false;
3191 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3192 E = ForwardingNodes.end(); I != E; ++I) {
3193 PHINode *Phi = I->first;
3194 SmallVector<int,4> &Indexes = I->second;
3196 if (Indexes.size() < 2) continue;
3198 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3199 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3206 /// ValidLookupTableConstant - Return true if the backend will be able to handle
3207 /// initializing an array of constants like C.
3208 static bool ValidLookupTableConstant(Constant *C) {
3209 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3210 return CE->isGEPWithNoNotionalOverIndexing();
3212 return isa<ConstantFP>(C) ||
3213 isa<ConstantInt>(C) ||
3214 isa<ConstantPointerNull>(C) ||
3215 isa<GlobalValue>(C) ||
3219 /// LookupConstant - If V is a Constant, return it. Otherwise, try to look up
3220 /// its constant value in ConstantPool, returning 0 if it's not there.
3221 static Constant *LookupConstant(Value *V,
3222 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3223 if (Constant *C = dyn_cast<Constant>(V))
3225 return ConstantPool.lookup(V);
3228 /// ConstantFold - Try to fold instruction I into a constant. This works for
3229 /// simple instructions such as binary operations where both operands are
3230 /// constant or can be replaced by constants from the ConstantPool. Returns the
3231 /// resulting constant on success, 0 otherwise.
3232 static Constant *ConstantFold(Instruction *I,
3233 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3234 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
3235 Constant *A = LookupConstant(BO->getOperand(0), ConstantPool);
3238 Constant *B = LookupConstant(BO->getOperand(1), ConstantPool);
3241 return ConstantExpr::get(BO->getOpcode(), A, B);
3244 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
3245 Constant *A = LookupConstant(I->getOperand(0), ConstantPool);
3248 Constant *B = LookupConstant(I->getOperand(1), ConstantPool);
3251 return ConstantExpr::getCompare(Cmp->getPredicate(), A, B);
3254 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3255 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3258 if (A->isAllOnesValue())
3259 return LookupConstant(Select->getTrueValue(), ConstantPool);
3260 if (A->isNullValue())
3261 return LookupConstant(Select->getFalseValue(), ConstantPool);
3265 if (CastInst *Cast = dyn_cast<CastInst>(I)) {
3266 Constant *A = LookupConstant(I->getOperand(0), ConstantPool);
3269 return ConstantExpr::getCast(Cast->getOpcode(), A, Cast->getDestTy());
3275 /// GetCaseResults - Try to determine the resulting constant values in phi nodes
3276 /// at the common destination basic block, *CommonDest, for one of the case
3277 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3278 /// case), of a switch instruction SI.
3279 static bool GetCaseResults(SwitchInst *SI,
3280 ConstantInt *CaseVal,
3281 BasicBlock *CaseDest,
3282 BasicBlock **CommonDest,
3283 SmallVector<std::pair<PHINode*,Constant*>, 4> &Res) {
3284 // The block from which we enter the common destination.
3285 BasicBlock *Pred = SI->getParent();
3287 // If CaseDest is empty except for some side-effect free instructions through
3288 // which we can constant-propagate the CaseVal, continue to its successor.
3289 SmallDenseMap<Value*, Constant*> ConstantPool;
3290 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3291 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3293 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3294 // If the terminator is a simple branch, continue to the next block.
3295 if (T->getNumSuccessors() != 1)
3298 CaseDest = T->getSuccessor(0);
3299 } else if (isa<DbgInfoIntrinsic>(I)) {
3300 // Skip debug intrinsic.
3302 } else if (Constant *C = ConstantFold(I, ConstantPool)) {
3303 // Instruction is side-effect free and constant.
3304 ConstantPool.insert(std::make_pair(I, C));
3310 // If we did not have a CommonDest before, use the current one.
3312 *CommonDest = CaseDest;
3313 // If the destination isn't the common one, abort.
3314 if (CaseDest != *CommonDest)
3317 // Get the values for this case from phi nodes in the destination block.
3318 BasicBlock::iterator I = (*CommonDest)->begin();
3319 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3320 int Idx = PHI->getBasicBlockIndex(Pred);
3324 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3329 // Note: If the constant comes from constant-propagating the case value
3330 // through the CaseDest basic block, it will be safe to remove the
3331 // instructions in that block. They cannot be used (except in the phi nodes
3332 // we visit) outside CaseDest, because that block does not dominate its
3333 // successor. If it did, we would not be in this phi node.
3335 // Be conservative about which kinds of constants we support.
3336 if (!ValidLookupTableConstant(ConstVal))
3339 Res.push_back(std::make_pair(PHI, ConstVal));
3346 /// SwitchLookupTable - This class represents a lookup table that can be used
3347 /// to replace a switch.
3348 class SwitchLookupTable {
3350 /// SwitchLookupTable - Create a lookup table to use as a switch replacement
3351 /// with the contents of Values, using DefaultValue to fill any holes in the
3353 SwitchLookupTable(Module &M,
3355 ConstantInt *Offset,
3356 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values,
3357 Constant *DefaultValue,
3358 const DataLayout *TD);
3360 /// BuildLookup - Build instructions with Builder to retrieve the value at
3361 /// the position given by Index in the lookup table.
3362 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3364 /// WouldFitInRegister - Return true if a table with TableSize elements of
3365 /// type ElementType would fit in a target-legal register.
3366 static bool WouldFitInRegister(const DataLayout *TD,
3368 const Type *ElementType);
3371 // Depending on the contents of the table, it can be represented in
3374 // For tables where each element contains the same value, we just have to
3375 // store that single value and return it for each lookup.
3378 // For small tables with integer elements, we can pack them into a bitmap
3379 // that fits into a target-legal register. Values are retrieved by
3380 // shift and mask operations.
3383 // The table is stored as an array of values. Values are retrieved by load
3384 // instructions from the table.
3388 // For SingleValueKind, this is the single value.
3389 Constant *SingleValue;
3391 // For BitMapKind, this is the bitmap.
3392 ConstantInt *BitMap;
3393 IntegerType *BitMapElementTy;
3395 // For ArrayKind, this is the array.
3396 GlobalVariable *Array;
3400 SwitchLookupTable::SwitchLookupTable(Module &M,
3402 ConstantInt *Offset,
3403 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values,
3404 Constant *DefaultValue,
3405 const DataLayout *TD)
3406 : SingleValue(0), BitMap(0), BitMapElementTy(0), Array(0) {
3407 assert(Values.size() && "Can't build lookup table without values!");
3408 assert(TableSize >= Values.size() && "Can't fit values in table!");
3410 // If all values in the table are equal, this is that value.
3411 SingleValue = Values.begin()->second;
3413 // Build up the table contents.
3414 SmallVector<Constant*, 64> TableContents(TableSize);
3415 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3416 ConstantInt *CaseVal = Values[I].first;
3417 Constant *CaseRes = Values[I].second;
3418 assert(CaseRes->getType() == DefaultValue->getType());
3420 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3422 TableContents[Idx] = CaseRes;
3424 if (CaseRes != SingleValue)
3428 // Fill in any holes in the table with the default result.
3429 if (Values.size() < TableSize) {
3430 for (uint64_t I = 0; I < TableSize; ++I) {
3431 if (!TableContents[I])
3432 TableContents[I] = DefaultValue;
3435 if (DefaultValue != SingleValue)
3439 // If each element in the table contains the same value, we only need to store
3440 // that single value.
3442 Kind = SingleValueKind;
3446 // If the type is integer and the table fits in a register, build a bitmap.
3447 if (WouldFitInRegister(TD, TableSize, DefaultValue->getType())) {
3448 IntegerType *IT = cast<IntegerType>(DefaultValue->getType());
3449 APInt TableInt(TableSize * IT->getBitWidth(), 0);
3450 for (uint64_t I = TableSize; I > 0; --I) {
3451 TableInt <<= IT->getBitWidth();
3452 // Insert values into the bitmap. Undef values are set to zero.
3453 if (!isa<UndefValue>(TableContents[I - 1])) {
3454 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
3455 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
3458 BitMap = ConstantInt::get(M.getContext(), TableInt);
3459 BitMapElementTy = IT;
3465 // Store the table in an array.
3466 ArrayType *ArrayTy = ArrayType::get(DefaultValue->getType(), TableSize);
3467 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3469 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3470 GlobalVariable::PrivateLinkage,
3473 Array->setUnnamedAddr(true);
3477 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
3479 case SingleValueKind:
3482 // Type of the bitmap (e.g. i59).
3483 IntegerType *MapTy = BitMap->getType();
3485 // Cast Index to the same type as the bitmap.
3486 // Note: The Index is <= the number of elements in the table, so
3487 // truncating it to the width of the bitmask is safe.
3488 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
3490 // Multiply the shift amount by the element width.
3491 ShiftAmt = Builder.CreateMul(ShiftAmt,
3492 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
3496 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
3497 "switch.downshift");
3499 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
3503 Value *GEPIndices[] = { Builder.getInt32(0), Index };
3504 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices,
3506 return Builder.CreateLoad(GEP, "switch.load");
3509 llvm_unreachable("Unknown lookup table kind!");
3512 bool SwitchLookupTable::WouldFitInRegister(const DataLayout *TD,
3514 const Type *ElementType) {
3517 const IntegerType *IT = dyn_cast<IntegerType>(ElementType);
3520 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
3521 // are <= 15, we could try to narrow the type.
3523 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
3524 if (TableSize >= UINT_MAX/IT->getBitWidth())
3526 return TD->fitsInLegalInteger(TableSize * IT->getBitWidth());
3529 /// ShouldBuildLookupTable - Determine whether a lookup table should be built
3530 /// for this switch, based on the number of caes, size of the table and the
3531 /// types of the results.
3532 static bool ShouldBuildLookupTable(SwitchInst *SI,
3534 const TargetTransformInfo &TTI,
3535 const DataLayout *TD,
3536 const SmallDenseMap<PHINode*, Type*>& ResultTypes) {
3537 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
3538 return false; // TableSize overflowed, or mul below might overflow.
3540 bool AllTablesFitInRegister = true;
3541 bool HasIllegalType = false;
3542 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(),
3543 E = ResultTypes.end(); I != E; ++I) {
3544 Type *Ty = I->second;
3546 // Saturate this flag to true.
3547 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
3549 // Saturate this flag to false.
3550 AllTablesFitInRegister = AllTablesFitInRegister &&
3551 SwitchLookupTable::WouldFitInRegister(TD, TableSize, Ty);
3553 // If both flags saturate, we're done. NOTE: This *only* works with
3554 // saturating flags, and all flags have to saturate first due to the
3555 // non-deterministic behavior of iterating over a dense map.
3556 if (HasIllegalType && !AllTablesFitInRegister)
3560 // If each table would fit in a register, we should build it anyway.
3561 if (AllTablesFitInRegister)
3564 // Don't build a table that doesn't fit in-register if it has illegal types.
3568 // The table density should be at least 40%. This is the same criterion as for
3569 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3570 // FIXME: Find the best cut-off.
3571 return SI->getNumCases() * 10 >= TableSize * 4;
3574 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3575 /// phi nodes in a common successor block with different constant values,
3576 /// replace the switch with lookup tables.
3577 static bool SwitchToLookupTable(SwitchInst *SI,
3578 IRBuilder<> &Builder,
3579 const TargetTransformInfo &TTI,
3580 const DataLayout* TD) {
3581 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3583 // Only build lookup table when we have a target that supports it.
3584 if (!TTI.shouldBuildLookupTables())
3587 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3588 // split off a dense part and build a lookup table for that.
3590 // FIXME: This creates arrays of GEPs to constant strings, which means each
3591 // GEP needs a runtime relocation in PIC code. We should just build one big
3592 // string and lookup indices into that.
3594 // Ignore the switch if the number of cases is too small.
3595 // This is similar to the check when building jump tables in
3596 // SelectionDAGBuilder::handleJTSwitchCase.
3597 // FIXME: Determine the best cut-off.
3598 if (SI->getNumCases() < 4)
3601 // Figure out the corresponding result for each case value and phi node in the
3602 // common destination, as well as the the min and max case values.
3603 assert(SI->case_begin() != SI->case_end());
3604 SwitchInst::CaseIt CI = SI->case_begin();
3605 ConstantInt *MinCaseVal = CI.getCaseValue();
3606 ConstantInt *MaxCaseVal = CI.getCaseValue();
3608 BasicBlock *CommonDest = 0;
3609 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
3610 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
3611 SmallDenseMap<PHINode*, Constant*> DefaultResults;
3612 SmallDenseMap<PHINode*, Type*> ResultTypes;
3613 SmallVector<PHINode*, 4> PHIs;
3615 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
3616 ConstantInt *CaseVal = CI.getCaseValue();
3617 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
3618 MinCaseVal = CaseVal;
3619 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
3620 MaxCaseVal = CaseVal;
3622 // Resulting value at phi nodes for this case value.
3623 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
3625 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
3629 // Append the result from this case to the list for each phi.
3630 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) {
3631 if (!ResultLists.count(I->first))
3632 PHIs.push_back(I->first);
3633 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second));
3637 // Get the resulting values for the default case.
3638 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
3639 if (!GetCaseResults(SI, 0, SI->getDefaultDest(), &CommonDest,
3640 DefaultResultsList))
3642 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) {
3643 PHINode *PHI = DefaultResultsList[I].first;
3644 Constant *Result = DefaultResultsList[I].second;
3645 DefaultResults[PHI] = Result;
3646 ResultTypes[PHI] = Result->getType();
3649 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
3650 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
3651 if (!ShouldBuildLookupTable(SI, TableSize, TTI, TD, ResultTypes))
3654 // Create the BB that does the lookups.
3655 Module &Mod = *CommonDest->getParent()->getParent();
3656 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
3658 CommonDest->getParent(),
3661 // Check whether the condition value is within the case range, and branch to
3663 Builder.SetInsertPoint(SI);
3664 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
3666 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
3667 MinCaseVal->getType(), TableSize));
3668 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
3670 // Populate the BB that does the lookups.
3671 Builder.SetInsertPoint(LookupBB);
3672 bool ReturnedEarly = false;
3673 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
3674 PHINode *PHI = PHIs[I];
3676 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI],
3677 DefaultResults[PHI], TD);
3679 Value *Result = Table.BuildLookup(TableIndex, Builder);
3681 // If the result is used to return immediately from the function, we want to
3682 // do that right here.
3683 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin()) &&
3684 *PHI->use_begin() == CommonDest->getFirstNonPHIOrDbg()) {
3685 Builder.CreateRet(Result);
3686 ReturnedEarly = true;
3690 PHI->addIncoming(Result, LookupBB);
3694 Builder.CreateBr(CommonDest);
3696 // Remove the switch.
3697 for (unsigned i = 0; i < SI->getNumSuccessors(); ++i) {
3698 BasicBlock *Succ = SI->getSuccessor(i);
3699 if (Succ == SI->getDefaultDest()) continue;
3700 Succ->removePredecessor(SI->getParent());
3702 SI->eraseFromParent();
3708 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
3709 BasicBlock *BB = SI->getParent();
3711 if (isValueEqualityComparison(SI)) {
3712 // If we only have one predecessor, and if it is a branch on this value,
3713 // see if that predecessor totally determines the outcome of this switch.
3714 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3715 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
3716 return SimplifyCFG(BB, TTI, TD) | true;
3718 Value *Cond = SI->getCondition();
3719 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
3720 if (SimplifySwitchOnSelect(SI, Select))
3721 return SimplifyCFG(BB, TTI, TD) | true;
3723 // If the block only contains the switch, see if we can fold the block
3724 // away into any preds.
3725 BasicBlock::iterator BBI = BB->begin();
3726 // Ignore dbg intrinsics.
3727 while (isa<DbgInfoIntrinsic>(BBI))
3730 if (FoldValueComparisonIntoPredecessors(SI, Builder))
3731 return SimplifyCFG(BB, TTI, TD) | true;
3734 // Try to transform the switch into an icmp and a branch.
3735 if (TurnSwitchRangeIntoICmp(SI, Builder))
3736 return SimplifyCFG(BB, TTI, TD) | true;
3738 // Remove unreachable cases.
3739 if (EliminateDeadSwitchCases(SI))
3740 return SimplifyCFG(BB, TTI, TD) | true;
3742 if (ForwardSwitchConditionToPHI(SI))
3743 return SimplifyCFG(BB, TTI, TD) | true;
3745 if (SwitchToLookupTable(SI, Builder, TTI, TD))
3746 return SimplifyCFG(BB, TTI, TD) | true;
3751 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
3752 BasicBlock *BB = IBI->getParent();
3753 bool Changed = false;
3755 // Eliminate redundant destinations.
3756 SmallPtrSet<Value *, 8> Succs;
3757 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
3758 BasicBlock *Dest = IBI->getDestination(i);
3759 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) {
3760 Dest->removePredecessor(BB);
3761 IBI->removeDestination(i);
3767 if (IBI->getNumDestinations() == 0) {
3768 // If the indirectbr has no successors, change it to unreachable.
3769 new UnreachableInst(IBI->getContext(), IBI);
3770 EraseTerminatorInstAndDCECond(IBI);
3774 if (IBI->getNumDestinations() == 1) {
3775 // If the indirectbr has one successor, change it to a direct branch.
3776 BranchInst::Create(IBI->getDestination(0), IBI);
3777 EraseTerminatorInstAndDCECond(IBI);
3781 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
3782 if (SimplifyIndirectBrOnSelect(IBI, SI))
3783 return SimplifyCFG(BB, TTI, TD) | true;
3788 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
3789 BasicBlock *BB = BI->getParent();
3791 if (SinkCommon && SinkThenElseCodeToEnd(BI))
3794 // If the Terminator is the only non-phi instruction, simplify the block.
3795 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime();
3796 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
3797 TryToSimplifyUncondBranchFromEmptyBlock(BB))
3800 // If the only instruction in the block is a seteq/setne comparison
3801 // against a constant, try to simplify the block.
3802 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
3803 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
3804 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
3806 if (I->isTerminator() &&
3807 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI, TD))
3811 // If this basic block is ONLY a compare and a branch, and if a predecessor
3812 // branches to us and our successor, fold the comparison into the
3813 // predecessor and use logical operations to update the incoming value
3814 // for PHI nodes in common successor.
3815 if (FoldBranchToCommonDest(BI))
3816 return SimplifyCFG(BB, TTI, TD) | true;
3821 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
3822 BasicBlock *BB = BI->getParent();
3824 // Conditional branch
3825 if (isValueEqualityComparison(BI)) {
3826 // If we only have one predecessor, and if it is a branch on this value,
3827 // see if that predecessor totally determines the outcome of this
3829 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3830 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
3831 return SimplifyCFG(BB, TTI, TD) | true;
3833 // This block must be empty, except for the setcond inst, if it exists.
3834 // Ignore dbg intrinsics.
3835 BasicBlock::iterator I = BB->begin();
3836 // Ignore dbg intrinsics.
3837 while (isa<DbgInfoIntrinsic>(I))
3840 if (FoldValueComparisonIntoPredecessors(BI, Builder))
3841 return SimplifyCFG(BB, TTI, TD) | true;
3842 } else if (&*I == cast<Instruction>(BI->getCondition())){
3844 // Ignore dbg intrinsics.
3845 while (isa<DbgInfoIntrinsic>(I))
3847 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
3848 return SimplifyCFG(BB, TTI, TD) | true;
3852 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
3853 if (SimplifyBranchOnICmpChain(BI, TD, Builder))
3856 // If this basic block is ONLY a compare and a branch, and if a predecessor
3857 // branches to us and one of our successors, fold the comparison into the
3858 // predecessor and use logical operations to pick the right destination.
3859 if (FoldBranchToCommonDest(BI))
3860 return SimplifyCFG(BB, TTI, TD) | true;
3862 // We have a conditional branch to two blocks that are only reachable
3863 // from BI. We know that the condbr dominates the two blocks, so see if
3864 // there is any identical code in the "then" and "else" blocks. If so, we
3865 // can hoist it up to the branching block.
3866 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) {
3867 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3868 if (HoistThenElseCodeToIf(BI))
3869 return SimplifyCFG(BB, TTI, TD) | true;
3871 // If Successor #1 has multiple preds, we may be able to conditionally
3872 // execute Successor #0 if it branches to successor #1.
3873 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
3874 if (Succ0TI->getNumSuccessors() == 1 &&
3875 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
3876 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0)))
3877 return SimplifyCFG(BB, TTI, TD) | true;
3879 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3880 // If Successor #0 has multiple preds, we may be able to conditionally
3881 // execute Successor #1 if it branches to successor #0.
3882 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
3883 if (Succ1TI->getNumSuccessors() == 1 &&
3884 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
3885 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1)))
3886 return SimplifyCFG(BB, TTI, TD) | true;
3889 // If this is a branch on a phi node in the current block, thread control
3890 // through this block if any PHI node entries are constants.
3891 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
3892 if (PN->getParent() == BI->getParent())
3893 if (FoldCondBranchOnPHI(BI, TD))
3894 return SimplifyCFG(BB, TTI, TD) | true;
3896 // Scan predecessor blocks for conditional branches.
3897 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
3898 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
3899 if (PBI != BI && PBI->isConditional())
3900 if (SimplifyCondBranchToCondBranch(PBI, BI))
3901 return SimplifyCFG(BB, TTI, TD) | true;
3906 /// Check if passing a value to an instruction will cause undefined behavior.
3907 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
3908 Constant *C = dyn_cast<Constant>(V);
3915 if (C->isNullValue()) {
3916 // Only look at the first use, avoid hurting compile time with long uselists
3917 User *Use = *I->use_begin();
3919 // Now make sure that there are no instructions in between that can alter
3920 // control flow (eg. calls)
3921 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
3922 if (i == I->getParent()->end() || i->mayHaveSideEffects())
3925 // Look through GEPs. A load from a GEP derived from NULL is still undefined
3926 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
3927 if (GEP->getPointerOperand() == I)
3928 return passingValueIsAlwaysUndefined(V, GEP);
3930 // Look through bitcasts.
3931 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
3932 return passingValueIsAlwaysUndefined(V, BC);
3934 // Load from null is undefined.
3935 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
3936 return LI->getPointerAddressSpace() == 0;
3938 // Store to null is undefined.
3939 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
3940 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
3945 /// If BB has an incoming value that will always trigger undefined behavior
3946 /// (eg. null pointer dereference), remove the branch leading here.
3947 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
3948 for (BasicBlock::iterator i = BB->begin();
3949 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
3950 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
3951 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
3952 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
3953 IRBuilder<> Builder(T);
3954 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
3955 BB->removePredecessor(PHI->getIncomingBlock(i));
3956 // Turn uncoditional branches into unreachables and remove the dead
3957 // destination from conditional branches.
3958 if (BI->isUnconditional())
3959 Builder.CreateUnreachable();
3961 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
3962 BI->getSuccessor(0));
3963 BI->eraseFromParent();
3966 // TODO: SwitchInst.
3972 bool SimplifyCFGOpt::run(BasicBlock *BB) {
3973 bool Changed = false;
3975 assert(BB && BB->getParent() && "Block not embedded in function!");
3976 assert(BB->getTerminator() && "Degenerate basic block encountered!");
3978 // Remove basic blocks that have no predecessors (except the entry block)...
3979 // or that just have themself as a predecessor. These are unreachable.
3980 if ((pred_begin(BB) == pred_end(BB) &&
3981 BB != &BB->getParent()->getEntryBlock()) ||
3982 BB->getSinglePredecessor() == BB) {
3983 DEBUG(dbgs() << "Removing BB: \n" << *BB);
3984 DeleteDeadBlock(BB);
3988 // Check to see if we can constant propagate this terminator instruction
3990 Changed |= ConstantFoldTerminator(BB, true);
3992 // Check for and eliminate duplicate PHI nodes in this block.
3993 Changed |= EliminateDuplicatePHINodes(BB);
3995 // Check for and remove branches that will always cause undefined behavior.
3996 Changed |= removeUndefIntroducingPredecessor(BB);
3998 // Merge basic blocks into their predecessor if there is only one distinct
3999 // pred, and if there is only one distinct successor of the predecessor, and
4000 // if there are no PHI nodes.
4002 if (MergeBlockIntoPredecessor(BB))
4005 IRBuilder<> Builder(BB);
4007 // If there is a trivial two-entry PHI node in this basic block, and we can
4008 // eliminate it, do so now.
4009 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
4010 if (PN->getNumIncomingValues() == 2)
4011 Changed |= FoldTwoEntryPHINode(PN, TD);
4013 Builder.SetInsertPoint(BB->getTerminator());
4014 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
4015 if (BI->isUnconditional()) {
4016 if (SimplifyUncondBranch(BI, Builder)) return true;
4018 if (SimplifyCondBranch(BI, Builder)) return true;
4020 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
4021 if (SimplifyReturn(RI, Builder)) return true;
4022 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
4023 if (SimplifyResume(RI, Builder)) return true;
4024 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
4025 if (SimplifySwitch(SI, Builder)) return true;
4026 } else if (UnreachableInst *UI =
4027 dyn_cast<UnreachableInst>(BB->getTerminator())) {
4028 if (SimplifyUnreachable(UI)) return true;
4029 } else if (IndirectBrInst *IBI =
4030 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
4031 if (SimplifyIndirectBr(IBI)) return true;
4037 /// SimplifyCFG - This function is used to do simplification of a CFG. For
4038 /// example, it adjusts branches to branches to eliminate the extra hop, it
4039 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4040 /// of the CFG. It returns true if a modification was made.
4042 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
4043 const DataLayout *TD) {
4044 return SimplifyCFGOpt(TTI, TD).run(BB);