diff options
Diffstat (limited to 'lib/Transforms/Utils/SimplifyCFG.cpp')
| -rw-r--r-- | lib/Transforms/Utils/SimplifyCFG.cpp | 276 |
1 files changed, 163 insertions, 113 deletions
diff --git a/lib/Transforms/Utils/SimplifyCFG.cpp b/lib/Transforms/Utils/SimplifyCFG.cpp index 681bf9c..c4c1423 100644 --- a/lib/Transforms/Utils/SimplifyCFG.cpp +++ b/lib/Transforms/Utils/SimplifyCFG.cpp @@ -40,12 +40,14 @@ #include "llvm/Support/ConstantRange.h" #include "llvm/Support/Debug.h" #include "llvm/Support/NoFolder.h" +#include "llvm/Support/PatternMatch.h" #include "llvm/Support/raw_ostream.h" #include "llvm/Transforms/Utils/BasicBlockUtils.h" #include <algorithm> #include <map> #include <set> using namespace llvm; +using namespace PatternMatch; static cl::opt<unsigned> PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1), @@ -59,6 +61,10 @@ static cl::opt<bool> SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true), cl::desc("Sink common instructions down to the end block")); +static cl::opt<bool> +HoistCondStores("simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true), + cl::desc("Hoist conditional stores if an unconditional store preceeds")); + STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps"); STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables"); STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block"); @@ -84,7 +90,6 @@ namespace { class SimplifyCFGOpt { const TargetTransformInfo &TTI; const DataLayout *const TD; - Value *isValueEqualityComparison(TerminatorInst *TI); BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases); @@ -190,94 +195,7 @@ static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred, PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred); } - -/// GetIfCondition - Given a basic block (BB) with two predecessors (and at -/// least one PHI node in it), check to see if the merge at this block is due -/// to an "if condition". If so, return the boolean condition that determines -/// which entry into BB will be taken. Also, return by references the block -/// that will be entered from if the condition is true, and the block that will -/// be entered if the condition is false. -/// -/// This does no checking to see if the true/false blocks have large or unsavory -/// instructions in them. -static Value *GetIfCondition(BasicBlock *BB, BasicBlock *&IfTrue, - BasicBlock *&IfFalse) { - PHINode *SomePHI = cast<PHINode>(BB->begin()); - assert(SomePHI->getNumIncomingValues() == 2 && - "Function can only handle blocks with 2 predecessors!"); - BasicBlock *Pred1 = SomePHI->getIncomingBlock(0); - BasicBlock *Pred2 = SomePHI->getIncomingBlock(1); - - // We can only handle branches. Other control flow will be lowered to - // branches if possible anyway. - BranchInst *Pred1Br = dyn_cast<BranchInst>(Pred1->getTerminator()); - BranchInst *Pred2Br = dyn_cast<BranchInst>(Pred2->getTerminator()); - if (Pred1Br == 0 || Pred2Br == 0) - return 0; - - // Eliminate code duplication by ensuring that Pred1Br is conditional if - // either are. - if (Pred2Br->isConditional()) { - // If both branches are conditional, we don't have an "if statement". In - // reality, we could transform this case, but since the condition will be - // required anyway, we stand no chance of eliminating it, so the xform is - // probably not profitable. - if (Pred1Br->isConditional()) - return 0; - - std::swap(Pred1, Pred2); - std::swap(Pred1Br, Pred2Br); - } - - if (Pred1Br->isConditional()) { - // The only thing we have to watch out for here is to make sure that Pred2 - // doesn't have incoming edges from other blocks. If it does, the condition - // doesn't dominate BB. - if (Pred2->getSinglePredecessor() == 0) - return 0; - - // If we found a conditional branch predecessor, make sure that it branches - // to BB and Pred2Br. If it doesn't, this isn't an "if statement". - if (Pred1Br->getSuccessor(0) == BB && - Pred1Br->getSuccessor(1) == Pred2) { - IfTrue = Pred1; - IfFalse = Pred2; - } else if (Pred1Br->getSuccessor(0) == Pred2 && - Pred1Br->getSuccessor(1) == BB) { - IfTrue = Pred2; - IfFalse = Pred1; - } else { - // We know that one arm of the conditional goes to BB, so the other must - // go somewhere unrelated, and this must not be an "if statement". - return 0; - } - - return Pred1Br->getCondition(); - } - - // Ok, if we got here, both predecessors end with an unconditional branch to - // BB. Don't panic! If both blocks only have a single (identical) - // predecessor, and THAT is a conditional branch, then we're all ok! - BasicBlock *CommonPred = Pred1->getSinglePredecessor(); - if (CommonPred == 0 || CommonPred != Pred2->getSinglePredecessor()) - return 0; - - // Otherwise, if this is a conditional branch, then we can use it! - BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator()); - if (BI == 0) return 0; - - assert(BI->isConditional() && "Two successors but not conditional?"); - if (BI->getSuccessor(0) == Pred1) { - IfTrue = Pred1; - IfFalse = Pred2; - } else { - IfTrue = Pred2; - IfFalse = Pred1; - } - return BI->getCondition(); -} - -/// ComputeSpeculuationCost - Compute an abstract "cost" of speculating the +/// ComputeSpeculationCost - Compute an abstract "cost" of speculating the /// given instruction, which is assumed to be safe to speculate. 1 means /// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive. static unsigned ComputeSpeculationCost(const User *I) { @@ -428,7 +346,24 @@ GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra, // If this is an icmp against a constant, handle this as one of the cases. if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) { if (ConstantInt *C = GetConstantInt(I->getOperand(1), TD)) { + Value *RHSVal; + ConstantInt *RHSC; + if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) { + // (x & ~2^x) == y --> x == y || x == y|2^x + // This undoes a transformation done by instcombine to fuse 2 compares. + if (match(ICI->getOperand(0), + m_And(m_Value(RHSVal), m_ConstantInt(RHSC)))) { + APInt Not = ~RHSC->getValue(); + if (Not.isPowerOf2()) { + Vals.push_back(C); + Vals.push_back( + ConstantInt::get(C->getContext(), C->getValue() | Not)); + UsedICmps++; + return RHSVal; + } + } + UsedICmps++; Vals.push_back(C); return I->getOperand(0); @@ -439,6 +374,13 @@ GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra, ConstantRange Span = ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue()); + // Shift the range if the compare is fed by an add. This is the range + // compare idiom as emitted by instcombine. + bool hasAdd = + match(I->getOperand(0), m_Add(m_Value(RHSVal), m_ConstantInt(RHSC))); + if (hasAdd) + Span = Span.subtract(RHSC->getValue()); + // If this is an and/!= check then we want to optimize "x ugt 2" into // x != 0 && x != 1. if (!isEQ) @@ -451,7 +393,7 @@ GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra, for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp) Vals.push_back(ConstantInt::get(V->getContext(), Tmp)); UsedICmps++; - return I->getOperand(0); + return hasAdd ? RHSVal : I->getOperand(0); } return 0; } @@ -529,9 +471,7 @@ Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) { } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) if (BI->isConditional() && BI->getCondition()->hasOneUse()) if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) - if ((ICI->getPredicate() == ICmpInst::ICMP_EQ || - ICI->getPredicate() == ICmpInst::ICMP_NE) && - GetConstantInt(ICI->getOperand(1), TD)) + if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), TD)) CV = ICI->getOperand(0); // Unwrap any lossless ptrtoint cast. @@ -1079,9 +1019,9 @@ static bool HoistThenElseCodeToIf(BranchInst *BI) { (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))) return false; - // If we get here, we can hoist at least one instruction. BasicBlock *BIParent = BI->getParent(); + bool Changed = false; do { // If we are hoisting the terminator instruction, don't move one (making a // broken BB), instead clone it, and remove BI. @@ -1096,6 +1036,7 @@ static bool HoistThenElseCodeToIf(BranchInst *BI) { I2->replaceAllUsesWith(I1); I1->intersectOptionalDataWith(I2); I2->eraseFromParent(); + Changed = true; I1 = BB1_Itr++; I2 = BB2_Itr++; @@ -1115,7 +1056,23 @@ static bool HoistThenElseCodeToIf(BranchInst *BI) { HoistTerminator: // It may not be possible to hoist an invoke. if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) - return true; + return Changed; + + for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { + PHINode *PN; + for (BasicBlock::iterator BBI = SI->begin(); + (PN = dyn_cast<PHINode>(BBI)); ++BBI) { + Value *BB1V = PN->getIncomingValueForBlock(BB1); + Value *BB2V = PN->getIncomingValueForBlock(BB2); + if (BB1V == BB2V) + continue; + + if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V)) + return Changed; + if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V)) + return Changed; + } + } // Okay, it is safe to hoist the terminator. Instruction *NT = I1->clone(); @@ -1332,6 +1289,66 @@ static bool SinkThenElseCodeToEnd(BranchInst *BI1) { return Changed; } +/// \brief Determine if we can hoist sink a sole store instruction out of a +/// conditional block. +/// +/// We are looking for code like the following: +/// BrBB: +/// store i32 %add, i32* %arrayidx2 +/// ... // No other stores or function calls (we could be calling a memory +/// ... // function). +/// %cmp = icmp ult %x, %y +/// br i1 %cmp, label %EndBB, label %ThenBB +/// ThenBB: +/// store i32 %add5, i32* %arrayidx2 +/// br label EndBB +/// EndBB: +/// ... +/// We are going to transform this into: +/// BrBB: +/// store i32 %add, i32* %arrayidx2 +/// ... // +/// %cmp = icmp ult %x, %y +/// %add.add5 = select i1 %cmp, i32 %add, %add5 +/// store i32 %add.add5, i32* %arrayidx2 +/// ... +/// +/// \return The pointer to the value of the previous store if the store can be +/// hoisted into the predecessor block. 0 otherwise. +static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB, + BasicBlock *StoreBB, BasicBlock *EndBB) { + StoreInst *StoreToHoist = dyn_cast<StoreInst>(I); + if (!StoreToHoist) + return 0; + + // Volatile or atomic. + if (!StoreToHoist->isSimple()) + return 0; + + Value *StorePtr = StoreToHoist->getPointerOperand(); + + // Look for a store to the same pointer in BrBB. + unsigned MaxNumInstToLookAt = 10; + for (BasicBlock::reverse_iterator RI = BrBB->rbegin(), + RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) { + Instruction *CurI = &*RI; + + // Could be calling an instruction that effects memory like free(). + if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI)) + return 0; + + StoreInst *SI = dyn_cast<StoreInst>(CurI); + // Found the previous store make sure it stores to the same location. + if (SI && SI->getPointerOperand() == StorePtr) + // Found the previous store, return its value operand. + return SI->getValueOperand(); + else if (SI) + return 0; // Unknown store. + } + + return 0; +} + /// \brief Speculate a conditional basic block flattening the CFG. /// /// Note that this is a very risky transform currently. Speculating @@ -1395,6 +1412,8 @@ static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB) { SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts; unsigned SpeculationCost = 0; + Value *SpeculatedStoreValue = 0; + StoreInst *SpeculatedStore = 0; for (BasicBlock::iterator BBI = ThenBB->begin(), BBE = llvm::prior(ThenBB->end()); BBI != BBE; ++BBI) { @@ -1410,13 +1429,21 @@ static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB) { return false; // Don't hoist the instruction if it's unsafe or expensive. - if (!isSafeToSpeculativelyExecute(I)) + if (!isSafeToSpeculativelyExecute(I) && + !(HoistCondStores && + (SpeculatedStoreValue = isSafeToSpeculateStore(I, BB, ThenBB, + EndBB)))) return false; - if (ComputeSpeculationCost(I) > PHINodeFoldingThreshold) + if (!SpeculatedStoreValue && + ComputeSpeculationCost(I) > PHINodeFoldingThreshold) return false; + // Store the store speculation candidate. + if (SpeculatedStoreValue) + SpeculatedStore = cast<StoreInst>(I); + // Do not hoist the instruction if any of its operands are defined but not - // used in this BB. The transformation will prevent the operand from + // used in BB. The transformation will prevent the operand from // being sunk into the use block. for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) { @@ -1448,18 +1475,23 @@ static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB) { Value *OrigV = PN->getIncomingValueForBlock(BB); Value *ThenV = PN->getIncomingValueForBlock(ThenBB); + // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf. // Skip PHIs which are trivial. if (ThenV == OrigV) continue; HaveRewritablePHIs = true; - ConstantExpr *CE = dyn_cast<ConstantExpr>(ThenV); - if (!CE) + ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV); + ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV); + if (!OrigCE && !ThenCE) continue; // Known safe and cheap. - if (!isSafeToSpeculativelyExecute(CE)) + if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) || + (OrigCE && !isSafeToSpeculativelyExecute(OrigCE))) return false; - if (ComputeSpeculationCost(CE) > PHINodeFoldingThreshold) + unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE) : 0; + unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE) : 0; + if (OrigCost + ThenCost > 2 * PHINodeFoldingThreshold) return false; // Account for the cost of an unfolded ConstantExpr which could end up @@ -1473,12 +1505,24 @@ static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB) { // If there are no PHIs to process, bail early. This helps ensure idempotence // as well. - if (!HaveRewritablePHIs) + if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue)) return false; // If we get here, we can hoist the instruction and if-convert. DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";); + // Insert a select of the value of the speculated store. + if (SpeculatedStoreValue) { + IRBuilder<true, NoFolder> Builder(BI); + Value *TrueV = SpeculatedStore->getValueOperand(); + Value *FalseV = SpeculatedStoreValue; + if (Invert) + std::swap(TrueV, FalseV); + Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() + + "." + FalseV->getName()); + SpeculatedStore->setOperand(0, S); + } + // Hoist the instructions. BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(), llvm::prior(ThenBB->end())); @@ -3073,7 +3117,12 @@ static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) { Value *Sub = SI->getCondition(); if (!Offset->isNullValue()) Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off"); - Value *Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch"); + Value *Cmp; + // If NumCases overflowed, then all possible values jump to the successor. + if (NumCases->isNullValue() && SI->getNumCases() != 0) + Cmp = ConstantInt::getTrue(SI->getContext()); + else + Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch"); BranchInst *NewBI = Builder.CreateCondBr( Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest()); @@ -3216,7 +3265,7 @@ static bool ForwardSwitchConditionToPHI(SwitchInst *SI) { for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(), E = ForwardingNodes.end(); I != E; ++I) { PHINode *Phi = I->first; - SmallVector<int,4> &Indexes = I->second; + SmallVectorImpl<int> &Indexes = I->second; if (Indexes.size() < 2) continue; @@ -3301,11 +3350,12 @@ static Constant *ConstantFold(Instruction *I, /// at the common destination basic block, *CommonDest, for one of the case /// destionations CaseDest corresponding to value CaseVal (0 for the default /// case), of a switch instruction SI. -static bool GetCaseResults(SwitchInst *SI, - ConstantInt *CaseVal, - BasicBlock *CaseDest, - BasicBlock **CommonDest, - SmallVector<std::pair<PHINode*,Constant*>, 4> &Res) { +static bool +GetCaseResults(SwitchInst *SI, + ConstantInt *CaseVal, + BasicBlock *CaseDest, + BasicBlock **CommonDest, + SmallVectorImpl<std::pair<PHINode*,Constant*> > &Res) { // The block from which we enter the common destination. BasicBlock *Pred = SI->getParent(); @@ -3378,7 +3428,7 @@ namespace { SwitchLookupTable(Module &M, uint64_t TableSize, ConstantInt *Offset, - const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values, + const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values, Constant *DefaultValue, const DataLayout *TD); @@ -3425,7 +3475,7 @@ namespace { SwitchLookupTable::SwitchLookupTable(Module &M, uint64_t TableSize, ConstantInt *Offset, - const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values, + const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values, Constant *DefaultValue, const DataLayout *TD) : SingleValue(0), BitMap(0), BitMapElementTy(0), Array(0) { @@ -3552,7 +3602,7 @@ bool SwitchLookupTable::WouldFitInRegister(const DataLayout *TD, } /// ShouldBuildLookupTable - Determine whether a lookup table should be built -/// for this switch, based on the number of caes, size of the table and the +/// for this switch, based on the number of cases, size of the table and the /// types of the results. static bool ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize, |
