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 unsigned SpeculationCost = 0;
1391 for (BasicBlock::iterator BBI = ThenBB->begin(),
1392 BBE = llvm::prior(ThenBB->end());
1393 BBI != BBE; ++BBI) {
1394 Instruction *I = BBI;
1396 if (isa<DbgInfoIntrinsic>(I))
1399 // Only speculatively execution a single instruction (not counting the
1400 // terminator) for now.
1402 if (SpeculationCost > 1)
1405 // Don't hoist the instruction if it's unsafe or expensive.
1406 if (!isSafeToSpeculativelyExecute(I))
1408 if (ComputeSpeculationCost(I) > PHINodeFoldingThreshold)
1411 // Do not hoist the instruction if any of its operands are defined but not
1412 // used in this BB. The transformation will prevent the operand from
1413 // being sunk into the use block.
1414 for (User::op_iterator i = I->op_begin(), e = I->op_end();
1416 Instruction *OpI = dyn_cast<Instruction>(*i);
1417 if (OpI && OpI->getParent() == BB &&
1418 !OpI->mayHaveSideEffects() &&
1419 !OpI->isUsedInBasicBlock(BB))
1424 // Check that the PHI nodes can be converted to selects.
1425 bool HaveRewritablePHIs = false;
1426 for (BasicBlock::iterator I = EndBB->begin();
1427 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1428 Value *OrigV = PN->getIncomingValueForBlock(BB);
1429 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1431 // Skip PHIs which are trivial.
1435 HaveRewritablePHIs = true;
1436 ConstantExpr *CE = dyn_cast<ConstantExpr>(ThenV);
1438 continue; // Known safe and cheap.
1440 if (!isSafeToSpeculativelyExecute(CE))
1443 // Don't speculate into a select with a constant select expression operand.
1444 // FIXME: This should really be a cost metric, but our cost model doesn't
1445 // accurately model the expense of select.
1446 if (Operator::getOpcode(CE) == Instruction::Select)
1449 // Account for the cost of an unfolded ConstantExpr which could end up
1450 // getting expanded into Instructions.
1451 // FIXME: This doesn't account for how many operations are combined in the
1452 // constant expression.
1454 if (SpeculationCost > 1)
1458 // If there are no PHIs to process, bail early. This helps ensure idempotence
1460 if (!HaveRewritablePHIs)
1463 // If we get here, we can hoist the instruction and if-convert.
1464 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1466 // Hoist the instructions.
1467 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(),
1468 llvm::prior(ThenBB->end()));
1470 // Insert selects and rewrite the PHI operands.
1471 IRBuilder<true, NoFolder> Builder(BI);
1472 for (BasicBlock::iterator I = EndBB->begin();
1473 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1474 unsigned OrigI = PN->getBasicBlockIndex(BB);
1475 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1476 Value *OrigV = PN->getIncomingValue(OrigI);
1477 Value *ThenV = PN->getIncomingValue(ThenI);
1479 // Skip PHIs which are trivial.
1483 // Create a select whose true value is the speculatively executed value and
1484 // false value is the preexisting value. Swap them if the branch
1485 // destinations were inverted.
1486 Value *TrueV = ThenV, *FalseV = OrigV;
1488 std::swap(TrueV, FalseV);
1489 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV,
1490 TrueV->getName() + "." + FalseV->getName());
1491 PN->setIncomingValue(OrigI, V);
1492 PN->setIncomingValue(ThenI, V);
1499 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1500 /// across this block.
1501 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1502 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1505 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1506 if (isa<DbgInfoIntrinsic>(BBI))
1508 if (Size > 10) return false; // Don't clone large BB's.
1511 // We can only support instructions that do not define values that are
1512 // live outside of the current basic block.
1513 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end();
1515 Instruction *U = cast<Instruction>(*UI);
1516 if (U->getParent() != BB || isa<PHINode>(U)) return false;
1519 // Looks ok, continue checking.
1525 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1526 /// that is defined in the same block as the branch and if any PHI entries are
1527 /// constants, thread edges corresponding to that entry to be branches to their
1528 /// ultimate destination.
1529 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *TD) {
1530 BasicBlock *BB = BI->getParent();
1531 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1532 // NOTE: we currently cannot transform this case if the PHI node is used
1533 // outside of the block.
1534 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1537 // Degenerate case of a single entry PHI.
1538 if (PN->getNumIncomingValues() == 1) {
1539 FoldSingleEntryPHINodes(PN->getParent());
1543 // Now we know that this block has multiple preds and two succs.
1544 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1546 // Okay, this is a simple enough basic block. See if any phi values are
1548 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1549 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1550 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue;
1552 // Okay, we now know that all edges from PredBB should be revectored to
1553 // branch to RealDest.
1554 BasicBlock *PredBB = PN->getIncomingBlock(i);
1555 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1557 if (RealDest == BB) continue; // Skip self loops.
1558 // Skip if the predecessor's terminator is an indirect branch.
1559 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1561 // The dest block might have PHI nodes, other predecessors and other
1562 // difficult cases. Instead of being smart about this, just insert a new
1563 // block that jumps to the destination block, effectively splitting
1564 // the edge we are about to create.
1565 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1566 RealDest->getName()+".critedge",
1567 RealDest->getParent(), RealDest);
1568 BranchInst::Create(RealDest, EdgeBB);
1570 // Update PHI nodes.
1571 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1573 // BB may have instructions that are being threaded over. Clone these
1574 // instructions into EdgeBB. We know that there will be no uses of the
1575 // cloned instructions outside of EdgeBB.
1576 BasicBlock::iterator InsertPt = EdgeBB->begin();
1577 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1578 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1579 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1580 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1583 // Clone the instruction.
1584 Instruction *N = BBI->clone();
1585 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1587 // Update operands due to translation.
1588 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1590 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1591 if (PI != TranslateMap.end())
1595 // Check for trivial simplification.
1596 if (Value *V = SimplifyInstruction(N, TD)) {
1597 TranslateMap[BBI] = V;
1598 delete N; // Instruction folded away, don't need actual inst
1600 // Insert the new instruction into its new home.
1601 EdgeBB->getInstList().insert(InsertPt, N);
1602 if (!BBI->use_empty())
1603 TranslateMap[BBI] = N;
1607 // Loop over all of the edges from PredBB to BB, changing them to branch
1608 // to EdgeBB instead.
1609 TerminatorInst *PredBBTI = PredBB->getTerminator();
1610 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1611 if (PredBBTI->getSuccessor(i) == BB) {
1612 BB->removePredecessor(PredBB);
1613 PredBBTI->setSuccessor(i, EdgeBB);
1616 // Recurse, simplifying any other constants.
1617 return FoldCondBranchOnPHI(BI, TD) | true;
1623 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1624 /// PHI node, see if we can eliminate it.
1625 static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *TD) {
1626 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1627 // statement", which has a very simple dominance structure. Basically, we
1628 // are trying to find the condition that is being branched on, which
1629 // subsequently causes this merge to happen. We really want control
1630 // dependence information for this check, but simplifycfg can't keep it up
1631 // to date, and this catches most of the cases we care about anyway.
1632 BasicBlock *BB = PN->getParent();
1633 BasicBlock *IfTrue, *IfFalse;
1634 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1636 // Don't bother if the branch will be constant folded trivially.
1637 isa<ConstantInt>(IfCond))
1640 // Okay, we found that we can merge this two-entry phi node into a select.
1641 // Doing so would require us to fold *all* two entry phi nodes in this block.
1642 // At some point this becomes non-profitable (particularly if the target
1643 // doesn't support cmov's). Only do this transformation if there are two or
1644 // fewer PHI nodes in this block.
1645 unsigned NumPhis = 0;
1646 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1650 // Loop over the PHI's seeing if we can promote them all to select
1651 // instructions. While we are at it, keep track of the instructions
1652 // that need to be moved to the dominating block.
1653 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1654 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1655 MaxCostVal1 = PHINodeFoldingThreshold;
1657 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1658 PHINode *PN = cast<PHINode>(II++);
1659 if (Value *V = SimplifyInstruction(PN, TD)) {
1660 PN->replaceAllUsesWith(V);
1661 PN->eraseFromParent();
1665 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1667 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1672 // If we folded the first phi, PN dangles at this point. Refresh it. If
1673 // we ran out of PHIs then we simplified them all.
1674 PN = dyn_cast<PHINode>(BB->begin());
1675 if (PN == 0) return true;
1677 // Don't fold i1 branches on PHIs which contain binary operators. These can
1678 // often be turned into switches and other things.
1679 if (PN->getType()->isIntegerTy(1) &&
1680 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1681 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1682 isa<BinaryOperator>(IfCond)))
1685 // If we all PHI nodes are promotable, check to make sure that all
1686 // instructions in the predecessor blocks can be promoted as well. If
1687 // not, we won't be able to get rid of the control flow, so it's not
1688 // worth promoting to select instructions.
1689 BasicBlock *DomBlock = 0;
1690 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1691 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1692 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1695 DomBlock = *pred_begin(IfBlock1);
1696 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1697 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1698 // This is not an aggressive instruction that we can promote.
1699 // Because of this, we won't be able to get rid of the control
1700 // flow, so the xform is not worth it.
1705 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1708 DomBlock = *pred_begin(IfBlock2);
1709 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1710 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1711 // This is not an aggressive instruction that we can promote.
1712 // Because of this, we won't be able to get rid of the control
1713 // flow, so the xform is not worth it.
1718 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1719 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1721 // If we can still promote the PHI nodes after this gauntlet of tests,
1722 // do all of the PHI's now.
1723 Instruction *InsertPt = DomBlock->getTerminator();
1724 IRBuilder<true, NoFolder> Builder(InsertPt);
1726 // Move all 'aggressive' instructions, which are defined in the
1727 // conditional parts of the if's up to the dominating block.
1729 DomBlock->getInstList().splice(InsertPt,
1730 IfBlock1->getInstList(), IfBlock1->begin(),
1731 IfBlock1->getTerminator());
1733 DomBlock->getInstList().splice(InsertPt,
1734 IfBlock2->getInstList(), IfBlock2->begin(),
1735 IfBlock2->getTerminator());
1737 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1738 // Change the PHI node into a select instruction.
1739 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1740 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1743 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1744 PN->replaceAllUsesWith(NV);
1746 PN->eraseFromParent();
1749 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1750 // has been flattened. Change DomBlock to jump directly to our new block to
1751 // avoid other simplifycfg's kicking in on the diamond.
1752 TerminatorInst *OldTI = DomBlock->getTerminator();
1753 Builder.SetInsertPoint(OldTI);
1754 Builder.CreateBr(BB);
1755 OldTI->eraseFromParent();
1759 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1760 /// to two returning blocks, try to merge them together into one return,
1761 /// introducing a select if the return values disagree.
1762 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1763 IRBuilder<> &Builder) {
1764 assert(BI->isConditional() && "Must be a conditional branch");
1765 BasicBlock *TrueSucc = BI->getSuccessor(0);
1766 BasicBlock *FalseSucc = BI->getSuccessor(1);
1767 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1768 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1770 // Check to ensure both blocks are empty (just a return) or optionally empty
1771 // with PHI nodes. If there are other instructions, merging would cause extra
1772 // computation on one path or the other.
1773 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1775 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1778 Builder.SetInsertPoint(BI);
1779 // Okay, we found a branch that is going to two return nodes. If
1780 // there is no return value for this function, just change the
1781 // branch into a return.
1782 if (FalseRet->getNumOperands() == 0) {
1783 TrueSucc->removePredecessor(BI->getParent());
1784 FalseSucc->removePredecessor(BI->getParent());
1785 Builder.CreateRetVoid();
1786 EraseTerminatorInstAndDCECond(BI);
1790 // Otherwise, figure out what the true and false return values are
1791 // so we can insert a new select instruction.
1792 Value *TrueValue = TrueRet->getReturnValue();
1793 Value *FalseValue = FalseRet->getReturnValue();
1795 // Unwrap any PHI nodes in the return blocks.
1796 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1797 if (TVPN->getParent() == TrueSucc)
1798 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1799 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1800 if (FVPN->getParent() == FalseSucc)
1801 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1803 // In order for this transformation to be safe, we must be able to
1804 // unconditionally execute both operands to the return. This is
1805 // normally the case, but we could have a potentially-trapping
1806 // constant expression that prevents this transformation from being
1808 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1811 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1815 // Okay, we collected all the mapped values and checked them for sanity, and
1816 // defined to really do this transformation. First, update the CFG.
1817 TrueSucc->removePredecessor(BI->getParent());
1818 FalseSucc->removePredecessor(BI->getParent());
1820 // Insert select instructions where needed.
1821 Value *BrCond = BI->getCondition();
1823 // Insert a select if the results differ.
1824 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1825 } else if (isa<UndefValue>(TrueValue)) {
1826 TrueValue = FalseValue;
1828 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1829 FalseValue, "retval");
1833 Value *RI = !TrueValue ?
1834 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
1838 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
1839 << "\n " << *BI << "NewRet = " << *RI
1840 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
1842 EraseTerminatorInstAndDCECond(BI);
1847 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
1848 /// probabilities of the branch taking each edge. Fills in the two APInt
1849 /// parameters and return true, or returns false if no or invalid metadata was
1851 static bool ExtractBranchMetadata(BranchInst *BI,
1852 uint64_t &ProbTrue, uint64_t &ProbFalse) {
1853 assert(BI->isConditional() &&
1854 "Looking for probabilities on unconditional branch?");
1855 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
1856 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
1857 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
1858 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
1859 if (!CITrue || !CIFalse) return false;
1860 ProbTrue = CITrue->getValue().getZExtValue();
1861 ProbFalse = CIFalse->getValue().getZExtValue();
1865 /// checkCSEInPredecessor - Return true if the given instruction is available
1866 /// in its predecessor block. If yes, the instruction will be removed.
1868 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
1869 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
1871 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
1872 Instruction *PBI = &*I;
1873 // Check whether Inst and PBI generate the same value.
1874 if (Inst->isIdenticalTo(PBI)) {
1875 Inst->replaceAllUsesWith(PBI);
1876 Inst->eraseFromParent();
1883 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
1884 /// predecessor branches to us and one of our successors, fold the block into
1885 /// the predecessor and use logical operations to pick the right destination.
1886 bool llvm::FoldBranchToCommonDest(BranchInst *BI) {
1887 BasicBlock *BB = BI->getParent();
1889 Instruction *Cond = 0;
1890 if (BI->isConditional())
1891 Cond = dyn_cast<Instruction>(BI->getCondition());
1893 // For unconditional branch, check for a simple CFG pattern, where
1894 // BB has a single predecessor and BB's successor is also its predecessor's
1895 // successor. If such pattern exisits, check for CSE between BB and its
1897 if (BasicBlock *PB = BB->getSinglePredecessor())
1898 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
1899 if (PBI->isConditional() &&
1900 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
1901 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
1902 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
1904 Instruction *Curr = I++;
1905 if (isa<CmpInst>(Curr)) {
1909 // Quit if we can't remove this instruction.
1910 if (!checkCSEInPredecessor(Curr, PB))
1919 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
1920 Cond->getParent() != BB || !Cond->hasOneUse())
1923 // Only allow this if the condition is a simple instruction that can be
1924 // executed unconditionally. It must be in the same block as the branch, and
1925 // must be at the front of the block.
1926 BasicBlock::iterator FrontIt = BB->front();
1928 // Ignore dbg intrinsics.
1929 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
1931 // Allow a single instruction to be hoisted in addition to the compare
1932 // that feeds the branch. We later ensure that any values that _it_ uses
1933 // were also live in the predecessor, so that we don't unnecessarily create
1934 // register pressure or inhibit out-of-order execution.
1935 Instruction *BonusInst = 0;
1936 if (&*FrontIt != Cond &&
1937 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond &&
1938 isSafeToSpeculativelyExecute(FrontIt)) {
1939 BonusInst = &*FrontIt;
1942 // Ignore dbg intrinsics.
1943 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
1946 // Only a single bonus inst is allowed.
1947 if (&*FrontIt != Cond)
1950 // Make sure the instruction after the condition is the cond branch.
1951 BasicBlock::iterator CondIt = Cond; ++CondIt;
1953 // Ingore dbg intrinsics.
1954 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
1959 // Cond is known to be a compare or binary operator. Check to make sure that
1960 // neither operand is a potentially-trapping constant expression.
1961 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
1964 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
1968 // Finally, don't infinitely unroll conditional loops.
1969 BasicBlock *TrueDest = BI->getSuccessor(0);
1970 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0;
1971 if (TrueDest == BB || FalseDest == BB)
1974 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
1975 BasicBlock *PredBlock = *PI;
1976 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
1978 // Check that we have two conditional branches. If there is a PHI node in
1979 // the common successor, verify that the same value flows in from both
1981 SmallVector<PHINode*, 4> PHIs;
1982 if (PBI == 0 || PBI->isUnconditional() ||
1983 (BI->isConditional() &&
1984 !SafeToMergeTerminators(BI, PBI)) ||
1985 (!BI->isConditional() &&
1986 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
1989 // Determine if the two branches share a common destination.
1990 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
1991 bool InvertPredCond = false;
1993 if (BI->isConditional()) {
1994 if (PBI->getSuccessor(0) == TrueDest)
1995 Opc = Instruction::Or;
1996 else if (PBI->getSuccessor(1) == FalseDest)
1997 Opc = Instruction::And;
1998 else if (PBI->getSuccessor(0) == FalseDest)
1999 Opc = Instruction::And, InvertPredCond = true;
2000 else if (PBI->getSuccessor(1) == TrueDest)
2001 Opc = Instruction::Or, InvertPredCond = true;
2005 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2009 // Ensure that any values used in the bonus instruction are also used
2010 // by the terminator of the predecessor. This means that those values
2011 // must already have been resolved, so we won't be inhibiting the
2012 // out-of-order core by speculating them earlier.
2014 // Collect the values used by the bonus inst
2015 SmallPtrSet<Value*, 4> UsedValues;
2016 for (Instruction::op_iterator OI = BonusInst->op_begin(),
2017 OE = BonusInst->op_end(); OI != OE; ++OI) {
2019 if (!isa<Constant>(V))
2020 UsedValues.insert(V);
2023 SmallVector<std::pair<Value*, unsigned>, 4> Worklist;
2024 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0));
2026 // Walk up to four levels back up the use-def chain of the predecessor's
2027 // terminator to see if all those values were used. The choice of four
2028 // levels is arbitrary, to provide a compile-time-cost bound.
2029 while (!Worklist.empty()) {
2030 std::pair<Value*, unsigned> Pair = Worklist.back();
2031 Worklist.pop_back();
2033 if (Pair.second >= 4) continue;
2034 UsedValues.erase(Pair.first);
2035 if (UsedValues.empty()) break;
2037 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) {
2038 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
2040 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1));
2044 if (!UsedValues.empty()) return false;
2047 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2048 IRBuilder<> Builder(PBI);
2050 // If we need to invert the condition in the pred block to match, do so now.
2051 if (InvertPredCond) {
2052 Value *NewCond = PBI->getCondition();
2054 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2055 CmpInst *CI = cast<CmpInst>(NewCond);
2056 CI->setPredicate(CI->getInversePredicate());
2058 NewCond = Builder.CreateNot(NewCond,
2059 PBI->getCondition()->getName()+".not");
2062 PBI->setCondition(NewCond);
2063 PBI->swapSuccessors();
2066 // If we have a bonus inst, clone it into the predecessor block.
2067 Instruction *NewBonus = 0;
2069 NewBonus = BonusInst->clone();
2070 PredBlock->getInstList().insert(PBI, NewBonus);
2071 NewBonus->takeName(BonusInst);
2072 BonusInst->setName(BonusInst->getName()+".old");
2075 // Clone Cond into the predecessor basic block, and or/and the
2076 // two conditions together.
2077 Instruction *New = Cond->clone();
2078 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus);
2079 PredBlock->getInstList().insert(PBI, New);
2080 New->takeName(Cond);
2081 Cond->setName(New->getName()+".old");
2083 if (BI->isConditional()) {
2084 Instruction *NewCond =
2085 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2087 PBI->setCondition(NewCond);
2089 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2090 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2092 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2094 SmallVector<uint64_t, 8> NewWeights;
2096 if (PBI->getSuccessor(0) == BB) {
2097 if (PredHasWeights && SuccHasWeights) {
2098 // PBI: br i1 %x, BB, FalseDest
2099 // BI: br i1 %y, TrueDest, FalseDest
2100 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2101 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2102 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2103 // TrueWeight for PBI * FalseWeight for BI.
2104 // We assume that total weights of a BranchInst can fit into 32 bits.
2105 // Therefore, we will not have overflow using 64-bit arithmetic.
2106 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2107 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2109 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2110 PBI->setSuccessor(0, TrueDest);
2112 if (PBI->getSuccessor(1) == BB) {
2113 if (PredHasWeights && SuccHasWeights) {
2114 // PBI: br i1 %x, TrueDest, BB
2115 // BI: br i1 %y, TrueDest, FalseDest
2116 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2117 // FalseWeight for PBI * TrueWeight for BI.
2118 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2119 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2120 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2121 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2123 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2124 PBI->setSuccessor(1, FalseDest);
2126 if (NewWeights.size() == 2) {
2127 // Halve the weights if any of them cannot fit in an uint32_t
2128 FitWeights(NewWeights);
2130 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2131 PBI->setMetadata(LLVMContext::MD_prof,
2132 MDBuilder(BI->getContext()).
2133 createBranchWeights(MDWeights));
2135 PBI->setMetadata(LLVMContext::MD_prof, NULL);
2137 // Update PHI nodes in the common successors.
2138 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2139 ConstantInt *PBI_C = cast<ConstantInt>(
2140 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2141 assert(PBI_C->getType()->isIntegerTy(1));
2142 Instruction *MergedCond = 0;
2143 if (PBI->getSuccessor(0) == TrueDest) {
2144 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2145 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2146 // is false: !PBI_Cond and BI_Value
2147 Instruction *NotCond =
2148 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2151 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2156 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2157 PBI->getCondition(), MergedCond,
2160 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2161 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2162 // is false: PBI_Cond and BI_Value
2164 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2165 PBI->getCondition(), New,
2167 if (PBI_C->isOne()) {
2168 Instruction *NotCond =
2169 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2172 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2173 NotCond, MergedCond,
2178 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2181 // Change PBI from Conditional to Unconditional.
2182 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2183 EraseTerminatorInstAndDCECond(PBI);
2187 // TODO: If BB is reachable from all paths through PredBlock, then we
2188 // could replace PBI's branch probabilities with BI's.
2190 // Copy any debug value intrinsics into the end of PredBlock.
2191 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2192 if (isa<DbgInfoIntrinsic>(*I))
2193 I->clone()->insertBefore(PBI);
2200 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2201 /// predecessor of another block, this function tries to simplify it. We know
2202 /// that PBI and BI are both conditional branches, and BI is in one of the
2203 /// successor blocks of PBI - PBI branches to BI.
2204 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2205 assert(PBI->isConditional() && BI->isConditional());
2206 BasicBlock *BB = BI->getParent();
2208 // If this block ends with a branch instruction, and if there is a
2209 // predecessor that ends on a branch of the same condition, make
2210 // this conditional branch redundant.
2211 if (PBI->getCondition() == BI->getCondition() &&
2212 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2213 // Okay, the outcome of this conditional branch is statically
2214 // knowable. If this block had a single pred, handle specially.
2215 if (BB->getSinglePredecessor()) {
2216 // Turn this into a branch on constant.
2217 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2218 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2220 return true; // Nuke the branch on constant.
2223 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2224 // in the constant and simplify the block result. Subsequent passes of
2225 // simplifycfg will thread the block.
2226 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2227 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2228 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2229 std::distance(PB, PE),
2230 BI->getCondition()->getName() + ".pr",
2232 // Okay, we're going to insert the PHI node. Since PBI is not the only
2233 // predecessor, compute the PHI'd conditional value for all of the preds.
2234 // Any predecessor where the condition is not computable we keep symbolic.
2235 for (pred_iterator PI = PB; PI != PE; ++PI) {
2236 BasicBlock *P = *PI;
2237 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2238 PBI != BI && PBI->isConditional() &&
2239 PBI->getCondition() == BI->getCondition() &&
2240 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2241 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2242 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2245 NewPN->addIncoming(BI->getCondition(), P);
2249 BI->setCondition(NewPN);
2254 // If this is a conditional branch in an empty block, and if any
2255 // predecessors is a conditional branch to one of our destinations,
2256 // fold the conditions into logical ops and one cond br.
2257 BasicBlock::iterator BBI = BB->begin();
2258 // Ignore dbg intrinsics.
2259 while (isa<DbgInfoIntrinsic>(BBI))
2265 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2270 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2272 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2273 PBIOp = 0, BIOp = 1;
2274 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2275 PBIOp = 1, BIOp = 0;
2276 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2281 // Check to make sure that the other destination of this branch
2282 // isn't BB itself. If so, this is an infinite loop that will
2283 // keep getting unwound.
2284 if (PBI->getSuccessor(PBIOp) == BB)
2287 // Do not perform this transformation if it would require
2288 // insertion of a large number of select instructions. For targets
2289 // without predication/cmovs, this is a big pessimization.
2290 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2292 unsigned NumPhis = 0;
2293 for (BasicBlock::iterator II = CommonDest->begin();
2294 isa<PHINode>(II); ++II, ++NumPhis)
2295 if (NumPhis > 2) // Disable this xform.
2298 // Finally, if everything is ok, fold the branches to logical ops.
2299 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2301 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2302 << "AND: " << *BI->getParent());
2305 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2306 // branch in it, where one edge (OtherDest) goes back to itself but the other
2307 // exits. We don't *know* that the program avoids the infinite loop
2308 // (even though that seems likely). If we do this xform naively, we'll end up
2309 // recursively unpeeling the loop. Since we know that (after the xform is
2310 // done) that the block *is* infinite if reached, we just make it an obviously
2311 // infinite loop with no cond branch.
2312 if (OtherDest == BB) {
2313 // Insert it at the end of the function, because it's either code,
2314 // or it won't matter if it's hot. :)
2315 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2316 "infloop", BB->getParent());
2317 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2318 OtherDest = InfLoopBlock;
2321 DEBUG(dbgs() << *PBI->getParent()->getParent());
2323 // BI may have other predecessors. Because of this, we leave
2324 // it alone, but modify PBI.
2326 // Make sure we get to CommonDest on True&True directions.
2327 Value *PBICond = PBI->getCondition();
2328 IRBuilder<true, NoFolder> Builder(PBI);
2330 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2332 Value *BICond = BI->getCondition();
2334 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2336 // Merge the conditions.
2337 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2339 // Modify PBI to branch on the new condition to the new dests.
2340 PBI->setCondition(Cond);
2341 PBI->setSuccessor(0, CommonDest);
2342 PBI->setSuccessor(1, OtherDest);
2344 // Update branch weight for PBI.
2345 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2346 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2348 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2350 if (PredHasWeights && SuccHasWeights) {
2351 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2352 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2353 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2354 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2355 // The weight to CommonDest should be PredCommon * SuccTotal +
2356 // PredOther * SuccCommon.
2357 // The weight to OtherDest should be PredOther * SuccOther.
2358 SmallVector<uint64_t, 2> NewWeights;
2359 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) +
2360 PredOther * SuccCommon);
2361 NewWeights.push_back(PredOther * SuccOther);
2362 // Halve the weights if any of them cannot fit in an uint32_t
2363 FitWeights(NewWeights);
2365 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end());
2366 PBI->setMetadata(LLVMContext::MD_prof,
2367 MDBuilder(BI->getContext()).
2368 createBranchWeights(MDWeights));
2371 // OtherDest may have phi nodes. If so, add an entry from PBI's
2372 // block that are identical to the entries for BI's block.
2373 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2375 // We know that the CommonDest already had an edge from PBI to
2376 // it. If it has PHIs though, the PHIs may have different
2377 // entries for BB and PBI's BB. If so, insert a select to make
2380 for (BasicBlock::iterator II = CommonDest->begin();
2381 (PN = dyn_cast<PHINode>(II)); ++II) {
2382 Value *BIV = PN->getIncomingValueForBlock(BB);
2383 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2384 Value *PBIV = PN->getIncomingValue(PBBIdx);
2386 // Insert a select in PBI to pick the right value.
2387 Value *NV = cast<SelectInst>
2388 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2389 PN->setIncomingValue(PBBIdx, NV);
2393 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2394 DEBUG(dbgs() << *PBI->getParent()->getParent());
2396 // This basic block is probably dead. We know it has at least
2397 // one fewer predecessor.
2401 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2402 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2403 // Takes care of updating the successors and removing the old terminator.
2404 // Also makes sure not to introduce new successors by assuming that edges to
2405 // non-successor TrueBBs and FalseBBs aren't reachable.
2406 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2407 BasicBlock *TrueBB, BasicBlock *FalseBB,
2408 uint32_t TrueWeight,
2409 uint32_t FalseWeight){
2410 // Remove any superfluous successor edges from the CFG.
2411 // First, figure out which successors to preserve.
2412 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2414 BasicBlock *KeepEdge1 = TrueBB;
2415 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0;
2417 // Then remove the rest.
2418 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2419 BasicBlock *Succ = OldTerm->getSuccessor(I);
2420 // Make sure only to keep exactly one copy of each edge.
2421 if (Succ == KeepEdge1)
2423 else if (Succ == KeepEdge2)
2426 Succ->removePredecessor(OldTerm->getParent());
2429 IRBuilder<> Builder(OldTerm);
2430 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2432 // Insert an appropriate new terminator.
2433 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) {
2434 if (TrueBB == FalseBB)
2435 // We were only looking for one successor, and it was present.
2436 // Create an unconditional branch to it.
2437 Builder.CreateBr(TrueBB);
2439 // We found both of the successors we were looking for.
2440 // Create a conditional branch sharing the condition of the select.
2441 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2442 if (TrueWeight != FalseWeight)
2443 NewBI->setMetadata(LLVMContext::MD_prof,
2444 MDBuilder(OldTerm->getContext()).
2445 createBranchWeights(TrueWeight, FalseWeight));
2447 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2448 // Neither of the selected blocks were successors, so this
2449 // terminator must be unreachable.
2450 new UnreachableInst(OldTerm->getContext(), OldTerm);
2452 // One of the selected values was a successor, but the other wasn't.
2453 // Insert an unconditional branch to the one that was found;
2454 // the edge to the one that wasn't must be unreachable.
2456 // Only TrueBB was found.
2457 Builder.CreateBr(TrueBB);
2459 // Only FalseBB was found.
2460 Builder.CreateBr(FalseBB);
2463 EraseTerminatorInstAndDCECond(OldTerm);
2467 // SimplifySwitchOnSelect - Replaces
2468 // (switch (select cond, X, Y)) on constant X, Y
2469 // with a branch - conditional if X and Y lead to distinct BBs,
2470 // unconditional otherwise.
2471 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2472 // Check for constant integer values in the select.
2473 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2474 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2475 if (!TrueVal || !FalseVal)
2478 // Find the relevant condition and destinations.
2479 Value *Condition = Select->getCondition();
2480 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2481 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2483 // Get weight for TrueBB and FalseBB.
2484 uint32_t TrueWeight = 0, FalseWeight = 0;
2485 SmallVector<uint64_t, 8> Weights;
2486 bool HasWeights = HasBranchWeights(SI);
2488 GetBranchWeights(SI, Weights);
2489 if (Weights.size() == 1 + SI->getNumCases()) {
2490 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2491 getSuccessorIndex()];
2492 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2493 getSuccessorIndex()];
2497 // Perform the actual simplification.
2498 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2499 TrueWeight, FalseWeight);
2502 // SimplifyIndirectBrOnSelect - Replaces
2503 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2504 // blockaddress(@fn, BlockB)))
2506 // (br cond, BlockA, BlockB).
2507 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2508 // Check that both operands of the select are block addresses.
2509 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2510 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2514 // Extract the actual blocks.
2515 BasicBlock *TrueBB = TBA->getBasicBlock();
2516 BasicBlock *FalseBB = FBA->getBasicBlock();
2518 // Perform the actual simplification.
2519 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2523 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2524 /// instruction (a seteq/setne with a constant) as the only instruction in a
2525 /// block that ends with an uncond branch. We are looking for a very specific
2526 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2527 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2528 /// default value goes to an uncond block with a seteq in it, we get something
2531 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2533 /// %tmp = icmp eq i8 %A, 92
2536 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2538 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2539 /// the PHI, merging the third icmp into the switch.
2540 static bool TryToSimplifyUncondBranchWithICmpInIt(
2541 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI,
2542 const DataLayout *TD) {
2543 BasicBlock *BB = ICI->getParent();
2545 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2547 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2549 Value *V = ICI->getOperand(0);
2550 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2552 // The pattern we're looking for is where our only predecessor is a switch on
2553 // 'V' and this block is the default case for the switch. In this case we can
2554 // fold the compared value into the switch to simplify things.
2555 BasicBlock *Pred = BB->getSinglePredecessor();
2556 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false;
2558 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2559 if (SI->getCondition() != V)
2562 // If BB is reachable on a non-default case, then we simply know the value of
2563 // V in this block. Substitute it and constant fold the icmp instruction
2565 if (SI->getDefaultDest() != BB) {
2566 ConstantInt *VVal = SI->findCaseDest(BB);
2567 assert(VVal && "Should have a unique destination value");
2568 ICI->setOperand(0, VVal);
2570 if (Value *V = SimplifyInstruction(ICI, TD)) {
2571 ICI->replaceAllUsesWith(V);
2572 ICI->eraseFromParent();
2574 // BB is now empty, so it is likely to simplify away.
2575 return SimplifyCFG(BB, TTI, TD) | true;
2578 // Ok, the block is reachable from the default dest. If the constant we're
2579 // comparing exists in one of the other edges, then we can constant fold ICI
2581 if (SI->findCaseValue(Cst) != SI->case_default()) {
2583 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2584 V = ConstantInt::getFalse(BB->getContext());
2586 V = ConstantInt::getTrue(BB->getContext());
2588 ICI->replaceAllUsesWith(V);
2589 ICI->eraseFromParent();
2590 // BB is now empty, so it is likely to simplify away.
2591 return SimplifyCFG(BB, TTI, TD) | true;
2594 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2596 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2597 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back());
2598 if (PHIUse == 0 || PHIUse != &SuccBlock->front() ||
2599 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2602 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2604 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2605 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2607 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2608 std::swap(DefaultCst, NewCst);
2610 // Replace ICI (which is used by the PHI for the default value) with true or
2611 // false depending on if it is EQ or NE.
2612 ICI->replaceAllUsesWith(DefaultCst);
2613 ICI->eraseFromParent();
2615 // Okay, the switch goes to this block on a default value. Add an edge from
2616 // the switch to the merge point on the compared value.
2617 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2618 BB->getParent(), BB);
2619 SmallVector<uint64_t, 8> Weights;
2620 bool HasWeights = HasBranchWeights(SI);
2622 GetBranchWeights(SI, Weights);
2623 if (Weights.size() == 1 + SI->getNumCases()) {
2624 // Split weight for default case to case for "Cst".
2625 Weights[0] = (Weights[0]+1) >> 1;
2626 Weights.push_back(Weights[0]);
2628 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2629 SI->setMetadata(LLVMContext::MD_prof,
2630 MDBuilder(SI->getContext()).
2631 createBranchWeights(MDWeights));
2634 SI->addCase(Cst, NewBB);
2636 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2637 Builder.SetInsertPoint(NewBB);
2638 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2639 Builder.CreateBr(SuccBlock);
2640 PHIUse->addIncoming(NewCst, NewBB);
2644 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2645 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2646 /// fold it into a switch instruction if so.
2647 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *TD,
2648 IRBuilder<> &Builder) {
2649 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2650 if (Cond == 0) return false;
2653 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2654 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2655 // 'setne's and'ed together, collect them.
2657 std::vector<ConstantInt*> Values;
2658 bool TrueWhenEqual = true;
2659 Value *ExtraCase = 0;
2660 unsigned UsedICmps = 0;
2662 if (Cond->getOpcode() == Instruction::Or) {
2663 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true,
2665 } else if (Cond->getOpcode() == Instruction::And) {
2666 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false,
2668 TrueWhenEqual = false;
2671 // If we didn't have a multiply compared value, fail.
2672 if (CompVal == 0) return false;
2674 // Avoid turning single icmps into a switch.
2678 // There might be duplicate constants in the list, which the switch
2679 // instruction can't handle, remove them now.
2680 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2681 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2683 // If Extra was used, we require at least two switch values to do the
2684 // transformation. A switch with one value is just an cond branch.
2685 if (ExtraCase && Values.size() < 2) return false;
2687 // TODO: Preserve branch weight metadata, similarly to how
2688 // FoldValueComparisonIntoPredecessors preserves it.
2690 // Figure out which block is which destination.
2691 BasicBlock *DefaultBB = BI->getSuccessor(1);
2692 BasicBlock *EdgeBB = BI->getSuccessor(0);
2693 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2695 BasicBlock *BB = BI->getParent();
2697 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2698 << " cases into SWITCH. BB is:\n" << *BB);
2700 // If there are any extra values that couldn't be folded into the switch
2701 // then we evaluate them with an explicit branch first. Split the block
2702 // right before the condbr to handle it.
2704 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2705 // Remove the uncond branch added to the old block.
2706 TerminatorInst *OldTI = BB->getTerminator();
2707 Builder.SetInsertPoint(OldTI);
2710 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2712 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2714 OldTI->eraseFromParent();
2716 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2717 // for the edge we just added.
2718 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2720 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2721 << "\nEXTRABB = " << *BB);
2725 Builder.SetInsertPoint(BI);
2726 // Convert pointer to int before we switch.
2727 if (CompVal->getType()->isPointerTy()) {
2728 assert(TD && "Cannot switch on pointer without DataLayout");
2729 CompVal = Builder.CreatePtrToInt(CompVal,
2730 TD->getIntPtrType(CompVal->getContext()),
2734 // Create the new switch instruction now.
2735 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2737 // Add all of the 'cases' to the switch instruction.
2738 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2739 New->addCase(Values[i], EdgeBB);
2741 // We added edges from PI to the EdgeBB. As such, if there were any
2742 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2743 // the number of edges added.
2744 for (BasicBlock::iterator BBI = EdgeBB->begin();
2745 isa<PHINode>(BBI); ++BBI) {
2746 PHINode *PN = cast<PHINode>(BBI);
2747 Value *InVal = PN->getIncomingValueForBlock(BB);
2748 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2749 PN->addIncoming(InVal, BB);
2752 // Erase the old branch instruction.
2753 EraseTerminatorInstAndDCECond(BI);
2755 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2759 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2760 // If this is a trivial landing pad that just continues unwinding the caught
2761 // exception then zap the landing pad, turning its invokes into calls.
2762 BasicBlock *BB = RI->getParent();
2763 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2764 if (RI->getValue() != LPInst)
2765 // Not a landing pad, or the resume is not unwinding the exception that
2766 // caused control to branch here.
2769 // Check that there are no other instructions except for debug intrinsics.
2770 BasicBlock::iterator I = LPInst, E = RI;
2772 if (!isa<DbgInfoIntrinsic>(I))
2775 // Turn all invokes that unwind here into calls and delete the basic block.
2776 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2777 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2778 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2779 // Insert a call instruction before the invoke.
2780 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2782 Call->setCallingConv(II->getCallingConv());
2783 Call->setAttributes(II->getAttributes());
2784 Call->setDebugLoc(II->getDebugLoc());
2786 // Anything that used the value produced by the invoke instruction now uses
2787 // the value produced by the call instruction. Note that we do this even
2788 // for void functions and calls with no uses so that the callgraph edge is
2790 II->replaceAllUsesWith(Call);
2791 BB->removePredecessor(II->getParent());
2793 // Insert a branch to the normal destination right before the invoke.
2794 BranchInst::Create(II->getNormalDest(), II);
2796 // Finally, delete the invoke instruction!
2797 II->eraseFromParent();
2800 // The landingpad is now unreachable. Zap it.
2801 BB->eraseFromParent();
2805 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2806 BasicBlock *BB = RI->getParent();
2807 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2809 // Find predecessors that end with branches.
2810 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2811 SmallVector<BranchInst*, 8> CondBranchPreds;
2812 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2813 BasicBlock *P = *PI;
2814 TerminatorInst *PTI = P->getTerminator();
2815 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2816 if (BI->isUnconditional())
2817 UncondBranchPreds.push_back(P);
2819 CondBranchPreds.push_back(BI);
2823 // If we found some, do the transformation!
2824 if (!UncondBranchPreds.empty() && DupRet) {
2825 while (!UncondBranchPreds.empty()) {
2826 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
2827 DEBUG(dbgs() << "FOLDING: " << *BB
2828 << "INTO UNCOND BRANCH PRED: " << *Pred);
2829 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
2832 // If we eliminated all predecessors of the block, delete the block now.
2833 if (pred_begin(BB) == pred_end(BB))
2834 // We know there are no successors, so just nuke the block.
2835 BB->eraseFromParent();
2840 // Check out all of the conditional branches going to this return
2841 // instruction. If any of them just select between returns, change the
2842 // branch itself into a select/return pair.
2843 while (!CondBranchPreds.empty()) {
2844 BranchInst *BI = CondBranchPreds.pop_back_val();
2846 // Check to see if the non-BB successor is also a return block.
2847 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
2848 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
2849 SimplifyCondBranchToTwoReturns(BI, Builder))
2855 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
2856 BasicBlock *BB = UI->getParent();
2858 bool Changed = false;
2860 // If there are any instructions immediately before the unreachable that can
2861 // be removed, do so.
2862 while (UI != BB->begin()) {
2863 BasicBlock::iterator BBI = UI;
2865 // Do not delete instructions that can have side effects which might cause
2866 // the unreachable to not be reachable; specifically, calls and volatile
2867 // operations may have this effect.
2868 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
2870 if (BBI->mayHaveSideEffects()) {
2871 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
2872 if (SI->isVolatile())
2874 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
2875 if (LI->isVolatile())
2877 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
2878 if (RMWI->isVolatile())
2880 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
2881 if (CXI->isVolatile())
2883 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
2884 !isa<LandingPadInst>(BBI)) {
2887 // Note that deleting LandingPad's here is in fact okay, although it
2888 // involves a bit of subtle reasoning. If this inst is a LandingPad,
2889 // all the predecessors of this block will be the unwind edges of Invokes,
2890 // and we can therefore guarantee this block will be erased.
2893 // Delete this instruction (any uses are guaranteed to be dead)
2894 if (!BBI->use_empty())
2895 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
2896 BBI->eraseFromParent();
2900 // If the unreachable instruction is the first in the block, take a gander
2901 // at all of the predecessors of this instruction, and simplify them.
2902 if (&BB->front() != UI) return Changed;
2904 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
2905 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
2906 TerminatorInst *TI = Preds[i]->getTerminator();
2907 IRBuilder<> Builder(TI);
2908 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
2909 if (BI->isUnconditional()) {
2910 if (BI->getSuccessor(0) == BB) {
2911 new UnreachableInst(TI->getContext(), TI);
2912 TI->eraseFromParent();
2916 if (BI->getSuccessor(0) == BB) {
2917 Builder.CreateBr(BI->getSuccessor(1));
2918 EraseTerminatorInstAndDCECond(BI);
2919 } else if (BI->getSuccessor(1) == BB) {
2920 Builder.CreateBr(BI->getSuccessor(0));
2921 EraseTerminatorInstAndDCECond(BI);
2925 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
2926 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2928 if (i.getCaseSuccessor() == BB) {
2929 BB->removePredecessor(SI->getParent());
2934 // If the default value is unreachable, figure out the most popular
2935 // destination and make it the default.
2936 if (SI->getDefaultDest() == BB) {
2937 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
2938 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2940 std::pair<unsigned, unsigned> &entry =
2941 Popularity[i.getCaseSuccessor()];
2942 if (entry.first == 0) {
2944 entry.second = i.getCaseIndex();
2950 // Find the most popular block.
2951 unsigned MaxPop = 0;
2952 unsigned MaxIndex = 0;
2953 BasicBlock *MaxBlock = 0;
2954 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
2955 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
2956 if (I->second.first > MaxPop ||
2957 (I->second.first == MaxPop && MaxIndex > I->second.second)) {
2958 MaxPop = I->second.first;
2959 MaxIndex = I->second.second;
2960 MaxBlock = I->first;
2964 // Make this the new default, allowing us to delete any explicit
2966 SI->setDefaultDest(MaxBlock);
2969 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
2971 if (isa<PHINode>(MaxBlock->begin()))
2972 for (unsigned i = 0; i != MaxPop-1; ++i)
2973 MaxBlock->removePredecessor(SI->getParent());
2975 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2977 if (i.getCaseSuccessor() == MaxBlock) {
2983 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
2984 if (II->getUnwindDest() == BB) {
2985 // Convert the invoke to a call instruction. This would be a good
2986 // place to note that the call does not throw though.
2987 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
2988 II->removeFromParent(); // Take out of symbol table
2990 // Insert the call now...
2991 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
2992 Builder.SetInsertPoint(BI);
2993 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
2994 Args, II->getName());
2995 CI->setCallingConv(II->getCallingConv());
2996 CI->setAttributes(II->getAttributes());
2997 // If the invoke produced a value, the call does now instead.
2998 II->replaceAllUsesWith(CI);
3005 // If this block is now dead, remove it.
3006 if (pred_begin(BB) == pred_end(BB) &&
3007 BB != &BB->getParent()->getEntryBlock()) {
3008 // We know there are no successors, so just nuke the block.
3009 BB->eraseFromParent();
3016 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
3017 /// integer range comparison into a sub, an icmp and a branch.
3018 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3019 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3021 // Make sure all cases point to the same destination and gather the values.
3022 SmallVector<ConstantInt *, 16> Cases;
3023 SwitchInst::CaseIt I = SI->case_begin();
3024 Cases.push_back(I.getCaseValue());
3025 SwitchInst::CaseIt PrevI = I++;
3026 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
3027 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
3029 Cases.push_back(I.getCaseValue());
3031 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
3033 // Sort the case values, then check if they form a range we can transform.
3034 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3035 for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
3036 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
3040 Constant *Offset = ConstantExpr::getNeg(Cases.back());
3041 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
3043 Value *Sub = SI->getCondition();
3044 if (!Offset->isNullValue())
3045 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
3046 Value *Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3047 BranchInst *NewBI = Builder.CreateCondBr(
3048 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
3050 // Update weight for the newly-created conditional branch.
3051 SmallVector<uint64_t, 8> Weights;
3052 bool HasWeights = HasBranchWeights(SI);
3054 GetBranchWeights(SI, Weights);
3055 if (Weights.size() == 1 + SI->getNumCases()) {
3056 // Combine all weights for the cases to be the true weight of NewBI.
3057 // We assume that the sum of all weights for a Terminator can fit into 32
3059 uint32_t NewTrueWeight = 0;
3060 for (unsigned I = 1, E = Weights.size(); I != E; ++I)
3061 NewTrueWeight += (uint32_t)Weights[I];
3062 NewBI->setMetadata(LLVMContext::MD_prof,
3063 MDBuilder(SI->getContext()).
3064 createBranchWeights(NewTrueWeight,
3065 (uint32_t)Weights[0]));
3069 // Prune obsolete incoming values off the successor's PHI nodes.
3070 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
3071 isa<PHINode>(BBI); ++BBI) {
3072 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
3073 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3075 SI->eraseFromParent();
3080 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
3081 /// and use it to remove dead cases.
3082 static bool EliminateDeadSwitchCases(SwitchInst *SI) {
3083 Value *Cond = SI->getCondition();
3084 unsigned Bits = cast<IntegerType>(Cond->getType())->getBitWidth();
3085 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3086 ComputeMaskedBits(Cond, KnownZero, KnownOne);
3088 // Gather dead cases.
3089 SmallVector<ConstantInt*, 8> DeadCases;
3090 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3091 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3092 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3093 DeadCases.push_back(I.getCaseValue());
3094 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3095 << I.getCaseValue() << "' is dead.\n");
3099 SmallVector<uint64_t, 8> Weights;
3100 bool HasWeight = HasBranchWeights(SI);
3102 GetBranchWeights(SI, Weights);
3103 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3106 // Remove dead cases from the switch.
3107 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3108 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3109 assert(Case != SI->case_default() &&
3110 "Case was not found. Probably mistake in DeadCases forming.");
3112 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3116 // Prune unused values from PHI nodes.
3117 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3118 SI->removeCase(Case);
3121 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3122 SI->setMetadata(LLVMContext::MD_prof,
3123 MDBuilder(SI->getParent()->getContext()).
3124 createBranchWeights(MDWeights));
3127 return !DeadCases.empty();
3130 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
3131 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3132 /// by an unconditional branch), look at the phi node for BB in the successor
3133 /// block and see if the incoming value is equal to CaseValue. If so, return
3134 /// the phi node, and set PhiIndex to BB's index in the phi node.
3135 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3138 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3139 return NULL; // BB must be empty to be a candidate for simplification.
3140 if (!BB->getSinglePredecessor())
3141 return NULL; // BB must be dominated by the switch.
3143 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3144 if (!Branch || !Branch->isUnconditional())
3145 return NULL; // Terminator must be unconditional branch.
3147 BasicBlock *Succ = Branch->getSuccessor(0);
3149 BasicBlock::iterator I = Succ->begin();
3150 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3151 int Idx = PHI->getBasicBlockIndex(BB);
3152 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3154 Value *InValue = PHI->getIncomingValue(Idx);
3155 if (InValue != CaseValue) continue;
3164 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
3165 /// instruction to a phi node dominated by the switch, if that would mean that
3166 /// some of the destination blocks of the switch can be folded away.
3167 /// Returns true if a change is made.
3168 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3169 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3170 ForwardingNodesMap ForwardingNodes;
3172 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3173 ConstantInt *CaseValue = I.getCaseValue();
3174 BasicBlock *CaseDest = I.getCaseSuccessor();
3177 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3181 ForwardingNodes[PHI].push_back(PhiIndex);
3184 bool Changed = false;
3186 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3187 E = ForwardingNodes.end(); I != E; ++I) {
3188 PHINode *Phi = I->first;
3189 SmallVector<int,4> &Indexes = I->second;
3191 if (Indexes.size() < 2) continue;
3193 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3194 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3201 /// ValidLookupTableConstant - Return true if the backend will be able to handle
3202 /// initializing an array of constants like C.
3203 static bool ValidLookupTableConstant(Constant *C) {
3204 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3205 return CE->isGEPWithNoNotionalOverIndexing();
3207 return isa<ConstantFP>(C) ||
3208 isa<ConstantInt>(C) ||
3209 isa<ConstantPointerNull>(C) ||
3210 isa<GlobalValue>(C) ||
3214 /// LookupConstant - If V is a Constant, return it. Otherwise, try to look up
3215 /// its constant value in ConstantPool, returning 0 if it's not there.
3216 static Constant *LookupConstant(Value *V,
3217 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3218 if (Constant *C = dyn_cast<Constant>(V))
3220 return ConstantPool.lookup(V);
3223 /// ConstantFold - Try to fold instruction I into a constant. This works for
3224 /// simple instructions such as binary operations where both operands are
3225 /// constant or can be replaced by constants from the ConstantPool. Returns the
3226 /// resulting constant on success, 0 otherwise.
3227 static Constant *ConstantFold(Instruction *I,
3228 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3229 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
3230 Constant *A = LookupConstant(BO->getOperand(0), ConstantPool);
3233 Constant *B = LookupConstant(BO->getOperand(1), ConstantPool);
3236 return ConstantExpr::get(BO->getOpcode(), A, B);
3239 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
3240 Constant *A = LookupConstant(I->getOperand(0), ConstantPool);
3243 Constant *B = LookupConstant(I->getOperand(1), ConstantPool);
3246 return ConstantExpr::getCompare(Cmp->getPredicate(), A, B);
3249 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3250 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3253 if (A->isAllOnesValue())
3254 return LookupConstant(Select->getTrueValue(), ConstantPool);
3255 if (A->isNullValue())
3256 return LookupConstant(Select->getFalseValue(), ConstantPool);
3260 if (CastInst *Cast = dyn_cast<CastInst>(I)) {
3261 Constant *A = LookupConstant(I->getOperand(0), ConstantPool);
3264 return ConstantExpr::getCast(Cast->getOpcode(), A, Cast->getDestTy());
3270 /// GetCaseResults - Try to determine the resulting constant values in phi nodes
3271 /// at the common destination basic block, *CommonDest, for one of the case
3272 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3273 /// case), of a switch instruction SI.
3274 static bool GetCaseResults(SwitchInst *SI,
3275 ConstantInt *CaseVal,
3276 BasicBlock *CaseDest,
3277 BasicBlock **CommonDest,
3278 SmallVector<std::pair<PHINode*,Constant*>, 4> &Res) {
3279 // The block from which we enter the common destination.
3280 BasicBlock *Pred = SI->getParent();
3282 // If CaseDest is empty except for some side-effect free instructions through
3283 // which we can constant-propagate the CaseVal, continue to its successor.
3284 SmallDenseMap<Value*, Constant*> ConstantPool;
3285 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3286 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3288 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3289 // If the terminator is a simple branch, continue to the next block.
3290 if (T->getNumSuccessors() != 1)
3293 CaseDest = T->getSuccessor(0);
3294 } else if (isa<DbgInfoIntrinsic>(I)) {
3295 // Skip debug intrinsic.
3297 } else if (Constant *C = ConstantFold(I, ConstantPool)) {
3298 // Instruction is side-effect free and constant.
3299 ConstantPool.insert(std::make_pair(I, C));
3305 // If we did not have a CommonDest before, use the current one.
3307 *CommonDest = CaseDest;
3308 // If the destination isn't the common one, abort.
3309 if (CaseDest != *CommonDest)
3312 // Get the values for this case from phi nodes in the destination block.
3313 BasicBlock::iterator I = (*CommonDest)->begin();
3314 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3315 int Idx = PHI->getBasicBlockIndex(Pred);
3319 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3324 // Note: If the constant comes from constant-propagating the case value
3325 // through the CaseDest basic block, it will be safe to remove the
3326 // instructions in that block. They cannot be used (except in the phi nodes
3327 // we visit) outside CaseDest, because that block does not dominate its
3328 // successor. If it did, we would not be in this phi node.
3330 // Be conservative about which kinds of constants we support.
3331 if (!ValidLookupTableConstant(ConstVal))
3334 Res.push_back(std::make_pair(PHI, ConstVal));
3341 /// SwitchLookupTable - This class represents a lookup table that can be used
3342 /// to replace a switch.
3343 class SwitchLookupTable {
3345 /// SwitchLookupTable - Create a lookup table to use as a switch replacement
3346 /// with the contents of Values, using DefaultValue to fill any holes in the
3348 SwitchLookupTable(Module &M,
3350 ConstantInt *Offset,
3351 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values,
3352 Constant *DefaultValue,
3353 const DataLayout *TD);
3355 /// BuildLookup - Build instructions with Builder to retrieve the value at
3356 /// the position given by Index in the lookup table.
3357 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3359 /// WouldFitInRegister - Return true if a table with TableSize elements of
3360 /// type ElementType would fit in a target-legal register.
3361 static bool WouldFitInRegister(const DataLayout *TD,
3363 const Type *ElementType);
3366 // Depending on the contents of the table, it can be represented in
3369 // For tables where each element contains the same value, we just have to
3370 // store that single value and return it for each lookup.
3373 // For small tables with integer elements, we can pack them into a bitmap
3374 // that fits into a target-legal register. Values are retrieved by
3375 // shift and mask operations.
3378 // The table is stored as an array of values. Values are retrieved by load
3379 // instructions from the table.
3383 // For SingleValueKind, this is the single value.
3384 Constant *SingleValue;
3386 // For BitMapKind, this is the bitmap.
3387 ConstantInt *BitMap;
3388 IntegerType *BitMapElementTy;
3390 // For ArrayKind, this is the array.
3391 GlobalVariable *Array;
3395 SwitchLookupTable::SwitchLookupTable(Module &M,
3397 ConstantInt *Offset,
3398 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values,
3399 Constant *DefaultValue,
3400 const DataLayout *TD)
3401 : SingleValue(0), BitMap(0), BitMapElementTy(0), Array(0) {
3402 assert(Values.size() && "Can't build lookup table without values!");
3403 assert(TableSize >= Values.size() && "Can't fit values in table!");
3405 // If all values in the table are equal, this is that value.
3406 SingleValue = Values.begin()->second;
3408 // Build up the table contents.
3409 SmallVector<Constant*, 64> TableContents(TableSize);
3410 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3411 ConstantInt *CaseVal = Values[I].first;
3412 Constant *CaseRes = Values[I].second;
3413 assert(CaseRes->getType() == DefaultValue->getType());
3415 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3417 TableContents[Idx] = CaseRes;
3419 if (CaseRes != SingleValue)
3423 // Fill in any holes in the table with the default result.
3424 if (Values.size() < TableSize) {
3425 for (uint64_t I = 0; I < TableSize; ++I) {
3426 if (!TableContents[I])
3427 TableContents[I] = DefaultValue;
3430 if (DefaultValue != SingleValue)
3434 // If each element in the table contains the same value, we only need to store
3435 // that single value.
3437 Kind = SingleValueKind;
3441 // If the type is integer and the table fits in a register, build a bitmap.
3442 if (WouldFitInRegister(TD, TableSize, DefaultValue->getType())) {
3443 IntegerType *IT = cast<IntegerType>(DefaultValue->getType());
3444 APInt TableInt(TableSize * IT->getBitWidth(), 0);
3445 for (uint64_t I = TableSize; I > 0; --I) {
3446 TableInt <<= IT->getBitWidth();
3447 // Insert values into the bitmap. Undef values are set to zero.
3448 if (!isa<UndefValue>(TableContents[I - 1])) {
3449 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
3450 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
3453 BitMap = ConstantInt::get(M.getContext(), TableInt);
3454 BitMapElementTy = IT;
3460 // Store the table in an array.
3461 ArrayType *ArrayTy = ArrayType::get(DefaultValue->getType(), TableSize);
3462 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3464 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3465 GlobalVariable::PrivateLinkage,
3468 Array->setUnnamedAddr(true);
3472 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
3474 case SingleValueKind:
3477 // Type of the bitmap (e.g. i59).
3478 IntegerType *MapTy = BitMap->getType();
3480 // Cast Index to the same type as the bitmap.
3481 // Note: The Index is <= the number of elements in the table, so
3482 // truncating it to the width of the bitmask is safe.
3483 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
3485 // Multiply the shift amount by the element width.
3486 ShiftAmt = Builder.CreateMul(ShiftAmt,
3487 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
3491 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
3492 "switch.downshift");
3494 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
3498 Value *GEPIndices[] = { Builder.getInt32(0), Index };
3499 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices,
3501 return Builder.CreateLoad(GEP, "switch.load");
3504 llvm_unreachable("Unknown lookup table kind!");
3507 bool SwitchLookupTable::WouldFitInRegister(const DataLayout *TD,
3509 const Type *ElementType) {
3512 const IntegerType *IT = dyn_cast<IntegerType>(ElementType);
3515 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
3516 // are <= 15, we could try to narrow the type.
3518 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
3519 if (TableSize >= UINT_MAX/IT->getBitWidth())
3521 return TD->fitsInLegalInteger(TableSize * IT->getBitWidth());
3524 /// ShouldBuildLookupTable - Determine whether a lookup table should be built
3525 /// for this switch, based on the number of caes, size of the table and the
3526 /// types of the results.
3527 static bool ShouldBuildLookupTable(SwitchInst *SI,
3529 const TargetTransformInfo &TTI,
3530 const DataLayout *TD,
3531 const SmallDenseMap<PHINode*, Type*>& ResultTypes) {
3532 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
3533 return false; // TableSize overflowed, or mul below might overflow.
3535 bool AllTablesFitInRegister = true;
3536 bool HasIllegalType = false;
3537 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(),
3538 E = ResultTypes.end(); I != E; ++I) {
3539 Type *Ty = I->second;
3541 // Saturate this flag to true.
3542 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
3544 // Saturate this flag to false.
3545 AllTablesFitInRegister = AllTablesFitInRegister &&
3546 SwitchLookupTable::WouldFitInRegister(TD, TableSize, Ty);
3548 // If both flags saturate, we're done. NOTE: This *only* works with
3549 // saturating flags, and all flags have to saturate first due to the
3550 // non-deterministic behavior of iterating over a dense map.
3551 if (HasIllegalType && !AllTablesFitInRegister)
3555 // If each table would fit in a register, we should build it anyway.
3556 if (AllTablesFitInRegister)
3559 // Don't build a table that doesn't fit in-register if it has illegal types.
3563 // The table density should be at least 40%. This is the same criterion as for
3564 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3565 // FIXME: Find the best cut-off.
3566 return SI->getNumCases() * 10 >= TableSize * 4;
3569 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3570 /// phi nodes in a common successor block with different constant values,
3571 /// replace the switch with lookup tables.
3572 static bool SwitchToLookupTable(SwitchInst *SI,
3573 IRBuilder<> &Builder,
3574 const TargetTransformInfo &TTI,
3575 const DataLayout* TD) {
3576 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3578 // Only build lookup table when we have a target that supports it.
3579 if (!TTI.shouldBuildLookupTables())
3582 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3583 // split off a dense part and build a lookup table for that.
3585 // FIXME: This creates arrays of GEPs to constant strings, which means each
3586 // GEP needs a runtime relocation in PIC code. We should just build one big
3587 // string and lookup indices into that.
3589 // Ignore the switch if the number of cases is too small.
3590 // This is similar to the check when building jump tables in
3591 // SelectionDAGBuilder::handleJTSwitchCase.
3592 // FIXME: Determine the best cut-off.
3593 if (SI->getNumCases() < 4)
3596 // Figure out the corresponding result for each case value and phi node in the
3597 // common destination, as well as the the min and max case values.
3598 assert(SI->case_begin() != SI->case_end());
3599 SwitchInst::CaseIt CI = SI->case_begin();
3600 ConstantInt *MinCaseVal = CI.getCaseValue();
3601 ConstantInt *MaxCaseVal = CI.getCaseValue();
3603 BasicBlock *CommonDest = 0;
3604 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
3605 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
3606 SmallDenseMap<PHINode*, Constant*> DefaultResults;
3607 SmallDenseMap<PHINode*, Type*> ResultTypes;
3608 SmallVector<PHINode*, 4> PHIs;
3610 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
3611 ConstantInt *CaseVal = CI.getCaseValue();
3612 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
3613 MinCaseVal = CaseVal;
3614 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
3615 MaxCaseVal = CaseVal;
3617 // Resulting value at phi nodes for this case value.
3618 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
3620 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
3624 // Append the result from this case to the list for each phi.
3625 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) {
3626 if (!ResultLists.count(I->first))
3627 PHIs.push_back(I->first);
3628 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second));
3632 // Get the resulting values for the default case.
3633 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
3634 if (!GetCaseResults(SI, 0, SI->getDefaultDest(), &CommonDest,
3635 DefaultResultsList))
3637 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) {
3638 PHINode *PHI = DefaultResultsList[I].first;
3639 Constant *Result = DefaultResultsList[I].second;
3640 DefaultResults[PHI] = Result;
3641 ResultTypes[PHI] = Result->getType();
3644 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
3645 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
3646 if (!ShouldBuildLookupTable(SI, TableSize, TTI, TD, ResultTypes))
3649 // Create the BB that does the lookups.
3650 Module &Mod = *CommonDest->getParent()->getParent();
3651 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
3653 CommonDest->getParent(),
3656 // Check whether the condition value is within the case range, and branch to
3658 Builder.SetInsertPoint(SI);
3659 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
3661 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
3662 MinCaseVal->getType(), TableSize));
3663 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
3665 // Populate the BB that does the lookups.
3666 Builder.SetInsertPoint(LookupBB);
3667 bool ReturnedEarly = false;
3668 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
3669 PHINode *PHI = PHIs[I];
3671 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI],
3672 DefaultResults[PHI], TD);
3674 Value *Result = Table.BuildLookup(TableIndex, Builder);
3676 // If the result is used to return immediately from the function, we want to
3677 // do that right here.
3678 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin()) &&
3679 *PHI->use_begin() == CommonDest->getFirstNonPHIOrDbg()) {
3680 Builder.CreateRet(Result);
3681 ReturnedEarly = true;
3685 PHI->addIncoming(Result, LookupBB);
3689 Builder.CreateBr(CommonDest);
3691 // Remove the switch.
3692 for (unsigned i = 0; i < SI->getNumSuccessors(); ++i) {
3693 BasicBlock *Succ = SI->getSuccessor(i);
3694 if (Succ == SI->getDefaultDest()) continue;
3695 Succ->removePredecessor(SI->getParent());
3697 SI->eraseFromParent();
3703 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
3704 BasicBlock *BB = SI->getParent();
3706 if (isValueEqualityComparison(SI)) {
3707 // If we only have one predecessor, and if it is a branch on this value,
3708 // see if that predecessor totally determines the outcome of this switch.
3709 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3710 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
3711 return SimplifyCFG(BB, TTI, TD) | true;
3713 Value *Cond = SI->getCondition();
3714 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
3715 if (SimplifySwitchOnSelect(SI, Select))
3716 return SimplifyCFG(BB, TTI, TD) | true;
3718 // If the block only contains the switch, see if we can fold the block
3719 // away into any preds.
3720 BasicBlock::iterator BBI = BB->begin();
3721 // Ignore dbg intrinsics.
3722 while (isa<DbgInfoIntrinsic>(BBI))
3725 if (FoldValueComparisonIntoPredecessors(SI, Builder))
3726 return SimplifyCFG(BB, TTI, TD) | true;
3729 // Try to transform the switch into an icmp and a branch.
3730 if (TurnSwitchRangeIntoICmp(SI, Builder))
3731 return SimplifyCFG(BB, TTI, TD) | true;
3733 // Remove unreachable cases.
3734 if (EliminateDeadSwitchCases(SI))
3735 return SimplifyCFG(BB, TTI, TD) | true;
3737 if (ForwardSwitchConditionToPHI(SI))
3738 return SimplifyCFG(BB, TTI, TD) | true;
3740 if (SwitchToLookupTable(SI, Builder, TTI, TD))
3741 return SimplifyCFG(BB, TTI, TD) | true;
3746 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
3747 BasicBlock *BB = IBI->getParent();
3748 bool Changed = false;
3750 // Eliminate redundant destinations.
3751 SmallPtrSet<Value *, 8> Succs;
3752 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
3753 BasicBlock *Dest = IBI->getDestination(i);
3754 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) {
3755 Dest->removePredecessor(BB);
3756 IBI->removeDestination(i);
3762 if (IBI->getNumDestinations() == 0) {
3763 // If the indirectbr has no successors, change it to unreachable.
3764 new UnreachableInst(IBI->getContext(), IBI);
3765 EraseTerminatorInstAndDCECond(IBI);
3769 if (IBI->getNumDestinations() == 1) {
3770 // If the indirectbr has one successor, change it to a direct branch.
3771 BranchInst::Create(IBI->getDestination(0), IBI);
3772 EraseTerminatorInstAndDCECond(IBI);
3776 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
3777 if (SimplifyIndirectBrOnSelect(IBI, SI))
3778 return SimplifyCFG(BB, TTI, TD) | true;
3783 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
3784 BasicBlock *BB = BI->getParent();
3786 if (SinkCommon && SinkThenElseCodeToEnd(BI))
3789 // If the Terminator is the only non-phi instruction, simplify the block.
3790 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime();
3791 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
3792 TryToSimplifyUncondBranchFromEmptyBlock(BB))
3795 // If the only instruction in the block is a seteq/setne comparison
3796 // against a constant, try to simplify the block.
3797 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
3798 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
3799 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
3801 if (I->isTerminator() &&
3802 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI, TD))
3806 // If this basic block is ONLY a compare and a branch, and if a predecessor
3807 // branches to us and our successor, fold the comparison into the
3808 // predecessor and use logical operations to update the incoming value
3809 // for PHI nodes in common successor.
3810 if (FoldBranchToCommonDest(BI))
3811 return SimplifyCFG(BB, TTI, TD) | true;
3816 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
3817 BasicBlock *BB = BI->getParent();
3819 // Conditional branch
3820 if (isValueEqualityComparison(BI)) {
3821 // If we only have one predecessor, and if it is a branch on this value,
3822 // see if that predecessor totally determines the outcome of this
3824 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3825 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
3826 return SimplifyCFG(BB, TTI, TD) | true;
3828 // This block must be empty, except for the setcond inst, if it exists.
3829 // Ignore dbg intrinsics.
3830 BasicBlock::iterator I = BB->begin();
3831 // Ignore dbg intrinsics.
3832 while (isa<DbgInfoIntrinsic>(I))
3835 if (FoldValueComparisonIntoPredecessors(BI, Builder))
3836 return SimplifyCFG(BB, TTI, TD) | true;
3837 } else if (&*I == cast<Instruction>(BI->getCondition())){
3839 // Ignore dbg intrinsics.
3840 while (isa<DbgInfoIntrinsic>(I))
3842 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
3843 return SimplifyCFG(BB, TTI, TD) | true;
3847 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
3848 if (SimplifyBranchOnICmpChain(BI, TD, Builder))
3851 // If this basic block is ONLY a compare and a branch, and if a predecessor
3852 // branches to us and one of our successors, fold the comparison into the
3853 // predecessor and use logical operations to pick the right destination.
3854 if (FoldBranchToCommonDest(BI))
3855 return SimplifyCFG(BB, TTI, TD) | true;
3857 // We have a conditional branch to two blocks that are only reachable
3858 // from BI. We know that the condbr dominates the two blocks, so see if
3859 // there is any identical code in the "then" and "else" blocks. If so, we
3860 // can hoist it up to the branching block.
3861 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) {
3862 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3863 if (HoistThenElseCodeToIf(BI))
3864 return SimplifyCFG(BB, TTI, TD) | true;
3866 // If Successor #1 has multiple preds, we may be able to conditionally
3867 // execute Successor #0 if it branches to successor #1.
3868 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
3869 if (Succ0TI->getNumSuccessors() == 1 &&
3870 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
3871 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0)))
3872 return SimplifyCFG(BB, TTI, TD) | true;
3874 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3875 // If Successor #0 has multiple preds, we may be able to conditionally
3876 // execute Successor #1 if it branches to successor #0.
3877 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
3878 if (Succ1TI->getNumSuccessors() == 1 &&
3879 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
3880 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1)))
3881 return SimplifyCFG(BB, TTI, TD) | true;
3884 // If this is a branch on a phi node in the current block, thread control
3885 // through this block if any PHI node entries are constants.
3886 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
3887 if (PN->getParent() == BI->getParent())
3888 if (FoldCondBranchOnPHI(BI, TD))
3889 return SimplifyCFG(BB, TTI, TD) | true;
3891 // Scan predecessor blocks for conditional branches.
3892 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
3893 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
3894 if (PBI != BI && PBI->isConditional())
3895 if (SimplifyCondBranchToCondBranch(PBI, BI))
3896 return SimplifyCFG(BB, TTI, TD) | true;
3901 /// Check if passing a value to an instruction will cause undefined behavior.
3902 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
3903 Constant *C = dyn_cast<Constant>(V);
3910 if (C->isNullValue()) {
3911 // Only look at the first use, avoid hurting compile time with long uselists
3912 User *Use = *I->use_begin();
3914 // Now make sure that there are no instructions in between that can alter
3915 // control flow (eg. calls)
3916 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
3917 if (i == I->getParent()->end() || i->mayHaveSideEffects())
3920 // Look through GEPs. A load from a GEP derived from NULL is still undefined
3921 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
3922 if (GEP->getPointerOperand() == I)
3923 return passingValueIsAlwaysUndefined(V, GEP);
3925 // Look through bitcasts.
3926 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
3927 return passingValueIsAlwaysUndefined(V, BC);
3929 // Load from null is undefined.
3930 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
3931 return LI->getPointerAddressSpace() == 0;
3933 // Store to null is undefined.
3934 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
3935 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
3940 /// If BB has an incoming value that will always trigger undefined behavior
3941 /// (eg. null pointer dereference), remove the branch leading here.
3942 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
3943 for (BasicBlock::iterator i = BB->begin();
3944 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
3945 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
3946 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
3947 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
3948 IRBuilder<> Builder(T);
3949 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
3950 BB->removePredecessor(PHI->getIncomingBlock(i));
3951 // Turn uncoditional branches into unreachables and remove the dead
3952 // destination from conditional branches.
3953 if (BI->isUnconditional())
3954 Builder.CreateUnreachable();
3956 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
3957 BI->getSuccessor(0));
3958 BI->eraseFromParent();
3961 // TODO: SwitchInst.
3967 bool SimplifyCFGOpt::run(BasicBlock *BB) {
3968 bool Changed = false;
3970 assert(BB && BB->getParent() && "Block not embedded in function!");
3971 assert(BB->getTerminator() && "Degenerate basic block encountered!");
3973 // Remove basic blocks that have no predecessors (except the entry block)...
3974 // or that just have themself as a predecessor. These are unreachable.
3975 if ((pred_begin(BB) == pred_end(BB) &&
3976 BB != &BB->getParent()->getEntryBlock()) ||
3977 BB->getSinglePredecessor() == BB) {
3978 DEBUG(dbgs() << "Removing BB: \n" << *BB);
3979 DeleteDeadBlock(BB);
3983 // Check to see if we can constant propagate this terminator instruction
3985 Changed |= ConstantFoldTerminator(BB, true);
3987 // Check for and eliminate duplicate PHI nodes in this block.
3988 Changed |= EliminateDuplicatePHINodes(BB);
3990 // Check for and remove branches that will always cause undefined behavior.
3991 Changed |= removeUndefIntroducingPredecessor(BB);
3993 // Merge basic blocks into their predecessor if there is only one distinct
3994 // pred, and if there is only one distinct successor of the predecessor, and
3995 // if there are no PHI nodes.
3997 if (MergeBlockIntoPredecessor(BB))
4000 IRBuilder<> Builder(BB);
4002 // If there is a trivial two-entry PHI node in this basic block, and we can
4003 // eliminate it, do so now.
4004 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
4005 if (PN->getNumIncomingValues() == 2)
4006 Changed |= FoldTwoEntryPHINode(PN, TD);
4008 Builder.SetInsertPoint(BB->getTerminator());
4009 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
4010 if (BI->isUnconditional()) {
4011 if (SimplifyUncondBranch(BI, Builder)) return true;
4013 if (SimplifyCondBranch(BI, Builder)) return true;
4015 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
4016 if (SimplifyReturn(RI, Builder)) return true;
4017 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
4018 if (SimplifyResume(RI, Builder)) return true;
4019 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
4020 if (SimplifySwitch(SI, Builder)) return true;
4021 } else if (UnreachableInst *UI =
4022 dyn_cast<UnreachableInst>(BB->getTerminator())) {
4023 if (SimplifyUnreachable(UI)) return true;
4024 } else if (IndirectBrInst *IBI =
4025 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
4026 if (SimplifyIndirectBr(IBI)) return true;
4032 /// SimplifyCFG - This function is used to do simplification of a CFG. For
4033 /// example, it adjusts branches to branches to eliminate the extra hop, it
4034 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4035 /// of the CFG. It returns true if a modification was made.
4037 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
4038 const DataLayout *TD) {
4039 return SimplifyCFGOpt(TTI, TD).run(BB);