X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTransforms%2FUtils%2FSimplifyCFG.cpp;h=bbeac4c35028e004eea9cc7000276490c6f8bfe0;hb=b824512b8d733ccdfb1b1a2e8a6950605acf1c52;hp=986a53f11ffd35590cd8ae48fa959ff6f855de8c;hpb=bf5d4fb7d8ee4537955610ef9c48f7009418efc3;p=oota-llvm.git diff --git a/lib/Transforms/Utils/SimplifyCFG.cpp b/lib/Transforms/Utils/SimplifyCFG.cpp index 986a53f11ff..bbeac4c3502 100644 --- a/lib/Transforms/Utils/SimplifyCFG.cpp +++ b/lib/Transforms/Utils/SimplifyCFG.cpp @@ -1,10 +1,10 @@ //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===// -// +// // The LLVM Compiler Infrastructure // -// This file was developed by the LLVM research group and is distributed under -// the University of Illinois Open Source License. See LICENSE.TXT for details. -// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// //===----------------------------------------------------------------------===// // // Peephole optimize the CFG. @@ -16,79 +16,244 @@ #include "llvm/Constants.h" #include "llvm/Instructions.h" #include "llvm/Type.h" +#include "llvm/DerivedTypes.h" #include "llvm/Support/CFG.h" #include "llvm/Support/Debug.h" +#include "llvm/Analysis/ConstantFolding.h" +#include "llvm/Transforms/Utils/BasicBlockUtils.h" +#include "llvm/ADT/SmallVector.h" +#include "llvm/ADT/SmallPtrSet.h" +#include "llvm/ADT/Statistic.h" #include #include #include #include using namespace llvm; -// PropagatePredecessorsForPHIs - This gets "Succ" ready to have the -// predecessors from "BB". This is a little tricky because "Succ" has PHI -// nodes, which need to have extra slots added to them to hold the merge edges -// from BB's predecessors, and BB itself might have had PHI nodes in it. This -// function returns true (failure) if the Succ BB already has a predecessor that -// is a predecessor of BB and incoming PHI arguments would not be discernible. +STATISTIC(NumSpeculations, "Number of speculative executed instructions"); + +/// SafeToMergeTerminators - Return true if it is safe to merge these two +/// terminator instructions together. +/// +static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) { + if (SI1 == SI2) return false; // Can't merge with self! + + // It is not safe to merge these two switch instructions if they have a common + // successor, and if that successor has a PHI node, and if *that* PHI node has + // conflicting incoming values from the two switch blocks. + BasicBlock *SI1BB = SI1->getParent(); + BasicBlock *SI2BB = SI2->getParent(); + SmallPtrSet SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); + + for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) + if (SI1Succs.count(*I)) + for (BasicBlock::iterator BBI = (*I)->begin(); + isa(BBI); ++BBI) { + PHINode *PN = cast(BBI); + if (PN->getIncomingValueForBlock(SI1BB) != + PN->getIncomingValueForBlock(SI2BB)) + return false; + } + + return true; +} + +/// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will +/// now be entries in it from the 'NewPred' block. The values that will be +/// flowing into the PHI nodes will be the same as those coming in from +/// ExistPred, an existing predecessor of Succ. +static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred, + BasicBlock *ExistPred) { + assert(std::find(succ_begin(ExistPred), succ_end(ExistPred), Succ) != + succ_end(ExistPred) && "ExistPred is not a predecessor of Succ!"); + if (!isa(Succ->begin())) return; // Quick exit if nothing to do + + for (BasicBlock::iterator I = Succ->begin(); isa(I); ++I) { + PHINode *PN = cast(I); + Value *V = PN->getIncomingValueForBlock(ExistPred); + PN->addIncoming(V, NewPred); + } +} + +// CanPropagatePredecessorsForPHIs - Return true if we can fold BB, an +// almost-empty BB ending in an unconditional branch to Succ, into succ. // // Assumption: Succ is the single successor for BB. // -static bool PropagatePredecessorsForPHIs(BasicBlock *BB, BasicBlock *Succ) { +static bool CanPropagatePredecessorsForPHIs(BasicBlock *BB, BasicBlock *Succ) { assert(*succ_begin(BB) == Succ && "Succ is not successor of BB!"); - if (!isa(Succ->front())) - return false; // We can make the transformation, no problem. + DOUT << "Looking to fold " << BB->getNameStart() << " into " + << Succ->getNameStart() << "\n"; + // Shortcut, if there is only a single predecessor is must be BB and merging + // is always safe + if (Succ->getSinglePredecessor()) return true; - // If there is more than one predecessor, and there are PHI nodes in - // the successor, then we need to add incoming edges for the PHI nodes - // - const std::vector BBPreds(pred_begin(BB), pred_end(BB)); + typedef SmallPtrSet InstrSet; + InstrSet BBPHIs; - // Check to see if one of the predecessors of BB is already a predecessor of - // Succ. If so, we cannot do the transformation if there are any PHI nodes - // with incompatible values coming in from the two edges! - // - for (pred_iterator PI = pred_begin(Succ), PE = pred_end(Succ); PI != PE; ++PI) - if (std::find(BBPreds.begin(), BBPreds.end(), *PI) != BBPreds.end()) { - // Loop over all of the PHI nodes checking to see if there are - // incompatible values coming in. - for (BasicBlock::iterator I = Succ->begin(); isa(I); ++I) { - PHINode *PN = cast(I); - // Loop up the entries in the PHI node for BB and for *PI if the values - // coming in are non-equal, we cannot merge these two blocks (instead we - // should insert a conditional move or something, then merge the - // blocks). - int Idx1 = PN->getBasicBlockIndex(BB); - int Idx2 = PN->getBasicBlockIndex(*PI); - assert(Idx1 != -1 && Idx2 != -1 && - "Didn't have entries for my predecessors??"); - if (PN->getIncomingValue(Idx1) != PN->getIncomingValue(Idx2)) - return true; // Values are not equal... - } - } + // Make a list of all phi nodes in BB + BasicBlock::iterator BBI = BB->begin(); + while (isa(*BBI)) BBPHIs.insert(BBI++); + + // Make a list of the predecessors of BB + typedef SmallPtrSet BlockSet; + BlockSet BBPreds(pred_begin(BB), pred_end(BB)); - // Loop over all of the PHI nodes in the successor BB. + // Use that list to make another list of common predecessors of BB and Succ + BlockSet CommonPreds; + for (pred_iterator PI = pred_begin(Succ), PE = pred_end(Succ); + PI != PE; ++PI) + if (BBPreds.count(*PI)) + CommonPreds.insert(*PI); + + // Shortcut, if there are no common predecessors, merging is always safe + if (CommonPreds.begin() == CommonPreds.end()) + return true; + + // Look at all the phi nodes in Succ, to see if they present a conflict when + // merging these blocks for (BasicBlock::iterator I = Succ->begin(); isa(I); ++I) { PHINode *PN = cast(I); - Value *OldVal = PN->removeIncomingValue(BB, false); - assert(OldVal && "No entry in PHI for Pred BB!"); - - // If this incoming value is one of the PHI nodes in BB, the new entries in - // the PHI node are the entries from the old PHI. - if (isa(OldVal) && cast(OldVal)->getParent() == BB) { - PHINode *OldValPN = cast(OldVal); - for (unsigned i = 0, e = OldValPN->getNumIncomingValues(); i != e; ++i) - PN->addIncoming(OldValPN->getIncomingValue(i), - OldValPN->getIncomingBlock(i)); + + // If the incoming value from BB is again a PHINode in + // BB which has the same incoming value for *PI as PN does, we can + // merge the phi nodes and then the blocks can still be merged + PHINode *BBPN = dyn_cast(PN->getIncomingValueForBlock(BB)); + if (BBPN && BBPN->getParent() == BB) { + for (BlockSet::iterator PI = CommonPreds.begin(), PE = CommonPreds.end(); + PI != PE; PI++) { + if (BBPN->getIncomingValueForBlock(*PI) + != PN->getIncomingValueForBlock(*PI)) { + DOUT << "Can't fold, phi node " << *PN->getNameStart() << " in " + << Succ->getNameStart() << " is conflicting with " + << BBPN->getNameStart() << " with regard to common predecessor " + << (*PI)->getNameStart() << "\n"; + return false; + } + } + // Remove this phinode from the list of phis in BB, since it has been + // handled. + BBPHIs.erase(BBPN); } else { - for (std::vector::const_iterator PredI = BBPreds.begin(), - End = BBPreds.end(); PredI != End; ++PredI) { - // Add an incoming value for each of the new incoming values... - PN->addIncoming(OldVal, *PredI); + Value* Val = PN->getIncomingValueForBlock(BB); + for (BlockSet::iterator PI = CommonPreds.begin(), PE = CommonPreds.end(); + PI != PE; PI++) { + // See if the incoming value for the common predecessor is equal to the + // one for BB, in which case this phi node will not prevent the merging + // of the block. + if (Val != PN->getIncomingValueForBlock(*PI)) { + DOUT << "Can't fold, phi node " << *PN->getNameStart() << " in " + << Succ->getNameStart() << " is conflicting with regard to common " + << "predecessor " << (*PI)->getNameStart() << "\n"; + return false; + } } } } - return false; + + // If there are any other phi nodes in BB that don't have a phi node in Succ + // to merge with, they must be moved to Succ completely. However, for any + // predecessors of Succ, branches will be added to the phi node that just + // point to itself. So, for any common predecessors, this must not cause + // conflicts. + for (InstrSet::iterator I = BBPHIs.begin(), E = BBPHIs.end(); + I != E; I++) { + PHINode *PN = cast(*I); + for (BlockSet::iterator PI = CommonPreds.begin(), PE = CommonPreds.end(); + PI != PE; PI++) + if (PN->getIncomingValueForBlock(*PI) != PN) { + DOUT << "Can't fold, phi node " << *PN->getNameStart() << " in " + << BB->getNameStart() << " is conflicting with regard to common " + << "predecessor " << (*PI)->getNameStart() << "\n"; + return false; + } + } + + return true; +} + +/// TryToSimplifyUncondBranchFromEmptyBlock - BB contains an unconditional +/// branch to Succ, and contains no instructions other than PHI nodes and the +/// branch. If possible, eliminate BB. +static bool TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB, + BasicBlock *Succ) { + // Check to see if merging these blocks would cause conflicts for any of the + // phi nodes in BB or Succ. If not, we can safely merge. + if (!CanPropagatePredecessorsForPHIs(BB, Succ)) return false; + + DOUT << "Killing Trivial BB: \n" << *BB; + + if (isa(Succ->begin())) { + // If there is more than one pred of succ, and there are PHI nodes in + // the successor, then we need to add incoming edges for the PHI nodes + // + const SmallVector BBPreds(pred_begin(BB), pred_end(BB)); + + // Loop over all of the PHI nodes in the successor of BB. + for (BasicBlock::iterator I = Succ->begin(); isa(I); ++I) { + PHINode *PN = cast(I); + Value *OldVal = PN->removeIncomingValue(BB, false); + assert(OldVal && "No entry in PHI for Pred BB!"); + + // If this incoming value is one of the PHI nodes in BB, the new entries + // in the PHI node are the entries from the old PHI. + if (isa(OldVal) && cast(OldVal)->getParent() == BB) { + PHINode *OldValPN = cast(OldVal); + for (unsigned i = 0, e = OldValPN->getNumIncomingValues(); i != e; ++i) + // Note that, since we are merging phi nodes and BB and Succ might + // have common predecessors, we could end up with a phi node with + // identical incoming branches. This will be cleaned up later (and + // will trigger asserts if we try to clean it up now, without also + // simplifying the corresponding conditional branch). + PN->addIncoming(OldValPN->getIncomingValue(i), + OldValPN->getIncomingBlock(i)); + } else { + // Add an incoming value for each of the new incoming values. + for (unsigned i = 0, e = BBPreds.size(); i != e; ++i) + PN->addIncoming(OldVal, BBPreds[i]); + } + } + } + + if (isa(&BB->front())) { + SmallVector + OldSuccPreds(pred_begin(Succ), pred_end(Succ)); + + // Move all PHI nodes in BB to Succ if they are alive, otherwise + // delete them. + while (PHINode *PN = dyn_cast(&BB->front())) + if (PN->use_empty()) { + // Just remove the dead phi. This happens if Succ's PHIs were the only + // users of the PHI nodes. + PN->eraseFromParent(); + } else { + // The instruction is alive, so this means that BB must dominate all + // predecessors of Succ (Since all uses of the PN are after its + // definition, so in Succ or a block dominated by Succ. If a predecessor + // of Succ would not be dominated by BB, PN would violate the def before + // use SSA demand). Therefore, we can simply move the phi node to the + // next block. + Succ->getInstList().splice(Succ->begin(), + BB->getInstList(), BB->begin()); + + // We need to add new entries for the PHI node to account for + // predecessors of Succ that the PHI node does not take into + // account. At this point, since we know that BB dominated succ and all + // of its predecessors, this means that we should any newly added + // incoming edges should use the PHI node itself as the value for these + // edges, because they are loop back edges. + for (unsigned i = 0, e = OldSuccPreds.size(); i != e; ++i) + if (OldSuccPreds[i] != BB) + PN->addIncoming(PN, OldSuccPreds[i]); + } + } + + // Everything that jumped to BB now goes to Succ. + BB->replaceAllUsesWith(Succ); + if (!Succ->hasName()) Succ->takeName(BB); + BB->eraseFromParent(); // Delete the old basic block. + return true; } /// GetIfCondition - Given a basic block (BB) with two predecessors (and @@ -97,7 +262,7 @@ static bool PropagatePredecessorsForPHIs(BasicBlock *BB, BasicBlock *Succ) { /// 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. -/// +/// /// static Value *GetIfCondition(BasicBlock *BB, BasicBlock *&IfTrue, BasicBlock *&IfFalse) { @@ -193,7 +358,14 @@ static Value *GetIfCondition(BasicBlock *BB, static bool DominatesMergePoint(Value *V, BasicBlock *BB, std::set *AggressiveInsts) { Instruction *I = dyn_cast(V); - if (!I) return true; // Non-instructions all dominate instructions. + if (!I) { + // Non-instructions all dominate instructions, but not all constantexprs + // can be executed unconditionally. + if (ConstantExpr *C = dyn_cast(V)) + if (C->canTrap()) + return false; + return true; + } BasicBlock *PBB = I->getParent(); // We don't want to allow weird loops that might have the "if condition" in @@ -231,20 +403,19 @@ static bool DominatesMergePoint(Value *V, BasicBlock *BB, case Instruction::Or: case Instruction::Xor: case Instruction::Shl: - case Instruction::Shr: - case Instruction::SetEQ: - case Instruction::SetNE: - case Instruction::SetLT: - case Instruction::SetGT: - case Instruction::SetLE: - case Instruction::SetGE: + case Instruction::LShr: + case Instruction::AShr: + case Instruction::ICmp: + case Instruction::FCmp: + if (I->getOperand(0)->getType()->isFPOrFPVector()) + return false; // FP arithmetic might trap. break; // These are all cheap and non-trapping instructions. } - + // Okay, we can only really hoist these out if their operands are not // defined in the conditional region. - for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) - if (!DominatesMergePoint(I->getOperand(i), BB, 0)) + for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) + if (!DominatesMergePoint(*i, BB, 0)) return false; // Okay, it's safe to do this! Remember this instruction. AggressiveInsts->insert(I); @@ -253,12 +424,13 @@ static bool DominatesMergePoint(Value *V, BasicBlock *BB, return true; } -// GatherConstantSetEQs - Given a potentially 'or'd together collection of seteq -// instructions that compare a value against a constant, return the value being -// compared, and stick the constant into the Values vector. +// GatherConstantSetEQs - Given a potentially 'or'd together collection of +// icmp_eq instructions that compare a value against a constant, return the +// value being compared, and stick the constant into the Values vector. static Value *GatherConstantSetEQs(Value *V, std::vector &Values){ - if (Instruction *Inst = dyn_cast(V)) - if (Inst->getOpcode() == Instruction::SetEQ) { + if (Instruction *Inst = dyn_cast(V)) { + if (Inst->getOpcode() == Instruction::ICmp && + cast(Inst)->getPredicate() == ICmpInst::ICMP_EQ) { if (ConstantInt *C = dyn_cast(Inst->getOperand(1))) { Values.push_back(C); return Inst->getOperand(0); @@ -272,6 +444,7 @@ static Value *GatherConstantSetEQs(Value *V, std::vector &Values){ if (LHS == RHS) return LHS; } + } return 0; } @@ -279,8 +452,9 @@ static Value *GatherConstantSetEQs(Value *V, std::vector &Values){ // setne instructions that compare a value against a constant, return the value // being compared, and stick the constant into the Values vector. static Value *GatherConstantSetNEs(Value *V, std::vector &Values){ - if (Instruction *Inst = dyn_cast(V)) - if (Inst->getOpcode() == Instruction::SetNE) { + if (Instruction *Inst = dyn_cast(V)) { + if (Inst->getOpcode() == Instruction::ICmp && + cast(Inst)->getPredicate() == ICmpInst::ICMP_NE) { if (ConstantInt *C = dyn_cast(Inst->getOperand(1))) { Values.push_back(C); return Inst->getOperand(0); @@ -288,17 +462,13 @@ static Value *GatherConstantSetNEs(Value *V, std::vector &Values){ Values.push_back(C); return Inst->getOperand(1); } - } else if (Inst->getOpcode() == Instruction::Cast) { - // Cast of X to bool is really a comparison against zero. - assert(Inst->getType() == Type::BoolTy && "Can only handle bool values!"); - Values.push_back(ConstantInt::get(Inst->getOperand(0)->getType(), 0)); - return Inst->getOperand(0); } else if (Inst->getOpcode() == Instruction::And) { if (Value *LHS = GatherConstantSetNEs(Inst->getOperand(0), Values)) if (Value *RHS = GatherConstantSetNEs(Inst->getOperand(1), Values)) if (LHS == RHS) return LHS; } + } return 0; } @@ -317,7 +487,7 @@ static bool GatherValueComparisons(Instruction *Cond, Value *&CompVal, return true; } else if (Cond->getOpcode() == Instruction::And) { CompVal = GatherConstantSetNEs(Cond, Values); - + // Return false to indicate that the condition is false if the CompVal is // equal to one of the constants. return false; @@ -329,55 +499,31 @@ static bool GatherValueComparisons(Instruction *Cond, Value *&CompVal, /// has no side effects, nuke it. If it uses any instructions that become dead /// because the instruction is now gone, nuke them too. static void ErasePossiblyDeadInstructionTree(Instruction *I) { - if (isInstructionTriviallyDead(I)) { - std::vector Operands(I->op_begin(), I->op_end()); - I->getParent()->getInstList().erase(I); - for (unsigned i = 0, e = Operands.size(); i != e; ++i) - if (Instruction *OpI = dyn_cast(Operands[i])) - ErasePossiblyDeadInstructionTree(OpI); - } -} + if (!isInstructionTriviallyDead(I)) return; + + SmallVector InstrsToInspect; + InstrsToInspect.push_back(I); -/// SafeToMergeTerminators - Return true if it is safe to merge these two -/// terminator instructions together. -/// -static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) { - if (SI1 == SI2) return false; // Can't merge with self! + while (!InstrsToInspect.empty()) { + I = InstrsToInspect.back(); + InstrsToInspect.pop_back(); - // It is not safe to merge these two switch instructions if they have a common - // successor, and if that successor has a PHI node, and if *that* PHI node has - // conflicting incoming values from the two switch blocks. - BasicBlock *SI1BB = SI1->getParent(); - BasicBlock *SI2BB = SI2->getParent(); - std::set SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); + if (!isInstructionTriviallyDead(I)) continue; - for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) - if (SI1Succs.count(*I)) - for (BasicBlock::iterator BBI = (*I)->begin(); - isa(BBI); ++BBI) { - PHINode *PN = cast(BBI); - if (PN->getIncomingValueForBlock(SI1BB) != - PN->getIncomingValueForBlock(SI2BB)) - return false; + // If I is in the work list multiple times, remove previous instances. + for (unsigned i = 0, e = InstrsToInspect.size(); i != e; ++i) + if (InstrsToInspect[i] == I) { + InstrsToInspect.erase(InstrsToInspect.begin()+i); + --i, --e; } - - return true; -} -/// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will -/// now be entries in it from the 'NewPred' block. The values that will be -/// flowing into the PHI nodes will be the same as those coming in from -/// ExistPred, an existing predecessor of Succ. -static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred, - BasicBlock *ExistPred) { - assert(std::find(succ_begin(ExistPred), succ_end(ExistPred), Succ) != - succ_end(ExistPred) && "ExistPred is not a predecessor of Succ!"); - if (!isa(Succ->begin())) return; // Quick exit if nothing to do + // Add operands of dead instruction to worklist. + for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) + if (Instruction *OpI = dyn_cast(*i)) + InstrsToInspect.push_back(OpI); - for (BasicBlock::iterator I = Succ->begin(); isa(I); ++I) { - PHINode *PN = cast(I); - Value *V = PN->getIncomingValueForBlock(ExistPred); - PN->addIncoming(V, NewPred); + // Remove dead instruction. + I->eraseFromParent(); } } @@ -395,18 +541,18 @@ static Value *isValueEqualityComparison(TerminatorInst *TI) { } if (BranchInst *BI = dyn_cast(TI)) if (BI->isConditional() && BI->getCondition()->hasOneUse()) - if (SetCondInst *SCI = dyn_cast(BI->getCondition())) - if ((SCI->getOpcode() == Instruction::SetEQ || - SCI->getOpcode() == Instruction::SetNE) && - isa(SCI->getOperand(1))) - return SCI->getOperand(0); + if (ICmpInst *ICI = dyn_cast(BI->getCondition())) + if ((ICI->getPredicate() == ICmpInst::ICMP_EQ || + ICI->getPredicate() == ICmpInst::ICMP_NE) && + isa(ICI->getOperand(1))) + return ICI->getOperand(0); return 0; } // Given a value comparison instruction, decode all of the 'cases' that it // represents and return the 'default' block. static BasicBlock * -GetValueEqualityComparisonCases(TerminatorInst *TI, +GetValueEqualityComparisonCases(TerminatorInst *TI, std::vector > &Cases) { if (SwitchInst *SI = dyn_cast(TI)) { @@ -417,17 +563,17 @@ GetValueEqualityComparisonCases(TerminatorInst *TI, } BranchInst *BI = cast(TI); - SetCondInst *SCI = cast(BI->getCondition()); - Cases.push_back(std::make_pair(cast(SCI->getOperand(1)), - BI->getSuccessor(SCI->getOpcode() == - Instruction::SetNE))); - return BI->getSuccessor(SCI->getOpcode() == Instruction::SetEQ); + ICmpInst *ICI = cast(BI->getCondition()); + Cases.push_back(std::make_pair(cast(ICI->getOperand(1)), + BI->getSuccessor(ICI->getPredicate() == + ICmpInst::ICMP_NE))); + return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ); } -// EliminateBlockCases - Given an vector of bb/value pairs, remove any entries +// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries // in the list that match the specified block. -static void EliminateBlockCases(BasicBlock *BB, +static void EliminateBlockCases(BasicBlock *BB, std::vector > &Cases) { for (unsigned i = 0, e = Cases.size(); i != e; ++i) if (Cases[i].second == BB) { @@ -491,7 +637,7 @@ static bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases); EliminateBlockCases(PredDef, PredCases); // Remove default from cases. - + // Find information about how control leaves this block. std::vector > ThisCases; BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases); @@ -510,13 +656,13 @@ static bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, assert(ThisCases.size() == 1 && "Branch can only have one case!"); Value *Cond = BTI->getCondition(); // Insert the new branch. - Instruction *NI = new BranchInst(ThisDef, TI); + Instruction *NI = BranchInst::Create(ThisDef, TI); // Remove PHI node entries for the dead edge. ThisCases[0].second->removePredecessor(TI->getParent()); - DEBUG(std::cerr << "Threading pred instr: " << *Pred->getTerminator() - << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); + DOUT << "Threading pred instr: " << *Pred->getTerminator() + << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"; TI->eraseFromParent(); // Nuke the old one. // If condition is now dead, nuke it. @@ -527,12 +673,12 @@ static bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, } else { SwitchInst *SI = cast(TI); // Okay, TI has cases that are statically dead, prune them away. - std::set DeadCases; + SmallPtrSet DeadCases; for (unsigned i = 0, e = PredCases.size(); i != e; ++i) DeadCases.insert(PredCases[i].first); - DEBUG(std::cerr << "Threading pred instr: " << *Pred->getTerminator() - << "Through successor TI: " << *TI); + DOUT << "Threading pred instr: " << *Pred->getTerminator() + << "Through successor TI: " << *TI; for (unsigned i = SI->getNumCases()-1; i != 0; --i) if (DeadCases.count(SI->getCaseValue(i))) { @@ -540,7 +686,7 @@ static bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, SI->removeCase(i); } - DEBUG(std::cerr << "Leaving: " << *TI << "\n"); + DOUT << "Leaving: " << *TI << "\n"; return true; } } @@ -551,11 +697,12 @@ static bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, ConstantInt *TIV = 0; BasicBlock *TIBB = TI->getParent(); for (unsigned i = 0, e = PredCases.size(); i != e; ++i) - if (PredCases[i].second == TIBB) + if (PredCases[i].second == TIBB) { if (TIV == 0) TIV = PredCases[i].first; else return false; // Cannot handle multiple values coming to this block. + } assert(TIV && "No edge from pred to succ?"); // Okay, we found the one constant that our value can be if we get into TI's @@ -579,10 +726,10 @@ static bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, CheckEdge = 0; // Insert the new branch. - Instruction *NI = new BranchInst(TheRealDest, TI); + Instruction *NI = BranchInst::Create(TheRealDest, TI); - DEBUG(std::cerr << "Threading pred instr: " << *Pred->getTerminator() - << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"); + DOUT << "Threading pred instr: " << *Pred->getTerminator() + << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"; Instruction *Cond = 0; if (BranchInst *BI = dyn_cast(TI)) Cond = dyn_cast(BI->getCondition()); @@ -604,11 +751,11 @@ static bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI) { assert(CV && "Not a comparison?"); bool Changed = false; - std::vector Preds(pred_begin(BB), pred_end(BB)); + SmallVector Preds(pred_begin(BB), pred_end(BB)); while (!Preds.empty()) { BasicBlock *Pred = Preds.back(); Preds.pop_back(); - + // See if the predecessor is a comparison with the same value. TerminatorInst *PTI = Pred->getTerminator(); Value *PCV = isValueEqualityComparison(PTI); // PredCondVal @@ -624,7 +771,7 @@ static bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI) { // Based on whether the default edge from PTI goes to BB or not, fill in // PredCases and PredDefault with the new switch cases we would like to // build. - std::vector NewSuccessors; + SmallVector NewSuccessors; if (PredDefault == BB) { // If this is the default destination from PTI, only the edges in TI @@ -692,7 +839,8 @@ static bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI) { AddPredecessorToBlock(NewSuccessors[i], Pred, BB); // Now that the successors are updated, create the new Switch instruction. - SwitchInst *NewSI = new SwitchInst(CV, PredDefault, PredCases.size(),PTI); + SwitchInst *NewSI = SwitchInst::Create(CV, PredDefault, + PredCases.size(), PTI); for (unsigned i = 0, e = PredCases.size(); i != e; ++i) NewSI->addCase(PredCases[i].first, PredCases[i].second); @@ -714,19 +862,19 @@ static bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI) { if (InfLoopBlock == 0) { // Insert it at the end of the loop, because it's either code, // or it won't matter if it's hot. :) - InfLoopBlock = new BasicBlock("infloop", BB->getParent()); - new BranchInst(InfLoopBlock, InfLoopBlock); + InfLoopBlock = BasicBlock::Create("infloop", BB->getParent()); + BranchInst::Create(InfLoopBlock, InfLoopBlock); } NewSI->setSuccessor(i, InfLoopBlock); } - + Changed = true; } } return Changed; } -/// HoistThenElseCodeToIf - Given a conditional branch that codes to BB1 and +/// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and /// BB2, hoist any common code in the two blocks up into the branch block. The /// caller of this function guarantees that BI's block dominates BB1 and BB2. static bool HoistThenElseCodeToIf(BranchInst *BI) { @@ -739,7 +887,8 @@ static bool HoistThenElseCodeToIf(BranchInst *BI) { BasicBlock *BB2 = BI->getSuccessor(1); // The false destination Instruction *I1 = BB1->begin(), *I2 = BB2->begin(); - if (I1->getOpcode() != I2->getOpcode() || !I1->isIdenticalTo(I2)) + if (I1->getOpcode() != I2->getOpcode() || isa(I1) || + isa(I1) || !I1->isIdenticalTo(I2)) return false; // If we get here, we can hoist at least one instruction. @@ -750,7 +899,7 @@ static bool HoistThenElseCodeToIf(BranchInst *BI) { // broken BB), instead clone it, and remove BI. if (isa(I1)) goto HoistTerminator; - + // For a normal instruction, we just move one to right before the branch, // then replace all uses of the other with the first. Finally, we remove // the now redundant second instruction. @@ -758,7 +907,7 @@ static bool HoistThenElseCodeToIf(BranchInst *BI) { if (!I2->use_empty()) I2->replaceAllUsesWith(I1); BB2->getInstList().erase(I2); - + I1 = BB1->begin(); I2 = BB2->begin(); } while (I1->getOpcode() == I2->getOpcode() && I1->isIdenticalTo(I2)); @@ -772,7 +921,7 @@ HoistTerminator: if (NT->getType() != Type::VoidTy) { I1->replaceAllUsesWith(NT); I2->replaceAllUsesWith(NT); - NT->setName(I1->getName()); + NT->takeName(I1); } // Hoisting one of the terminators from our successor is a great thing. @@ -791,8 +940,8 @@ HoistTerminator: // that determines the right value. SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)]; if (SI == 0) - SI = new SelectInst(BI->getCondition(), BB1V, BB2V, - BB1V->getName()+"."+BB2V->getName(), NT); + SI = SelectInst::Create(BI->getCondition(), BB1V, BB2V, + BB1V->getName()+"."+BB2V->getName(), NT); // Make the PHI node use the select for all incoming values for BB1/BB2 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2) @@ -804,23 +953,732 @@ HoistTerminator: // Update any PHI nodes in our new successors. for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) AddPredecessorToBlock(*SI, BIParent, BB1); + + BI->eraseFromParent(); + return true; +} + +/// SpeculativelyExecuteBB - Given a conditional branch that goes to BB1 +/// and an BB2 and the only successor of BB1 is BB2, hoist simple code +/// (for now, restricted to a single instruction that's side effect free) from +/// the BB1 into the branch block to speculatively execute it. +static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1) { + // Only speculatively execution a single instruction (not counting the + // terminator) for now. + BasicBlock::iterator BBI = BB1->begin(); + ++BBI; // must have at least a terminator + if (BBI == BB1->end()) return false; // only one inst + ++BBI; + if (BBI != BB1->end()) return false; // more than 2 insts. + + // Be conservative for now. FP select instruction can often be expensive. + Value *BrCond = BI->getCondition(); + if (isa(BrCond) && + cast(BrCond)->getOpcode() == Instruction::FCmp) + return false; + + // If BB1 is actually on the false edge of the conditional branch, remember + // to swap the select operands later. + bool Invert = false; + if (BB1 != BI->getSuccessor(0)) { + assert(BB1 == BI->getSuccessor(1) && "No edge from 'if' block?"); + Invert = true; + } + + // Turn + // BB: + // %t1 = icmp + // br i1 %t1, label %BB1, label %BB2 + // BB1: + // %t3 = add %t2, c + // br label BB2 + // BB2: + // => + // BB: + // %t1 = icmp + // %t4 = add %t2, c + // %t3 = select i1 %t1, %t2, %t3 + Instruction *I = BB1->begin(); + switch (I->getOpcode()) { + default: return false; // Not safe / profitable to hoist. + case Instruction::Add: + case Instruction::Sub: + case Instruction::And: + case Instruction::Or: + case Instruction::Xor: + case Instruction::Shl: + case Instruction::LShr: + case Instruction::AShr: + if (!I->getOperand(0)->getType()->isInteger()) + // FP arithmetic might trap. Not worth doing for vector ops. + return false; + break; // These are all cheap and non-trapping instructions. + } + + // Can we speculatively execute the instruction? And what is the value + // if the condition is false? Consider the phi uses, if the incoming value + // from the "if" block are all the same V, then V is the value of the + // select if the condition is false. + BasicBlock *BIParent = BI->getParent(); + SmallVector PHIUses; + Value *FalseV = NULL; + for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); + UI != E; ++UI) { + PHINode *PN = dyn_cast(UI); + if (!PN) + continue; + PHIUses.push_back(PN); + Value *PHIV = PN->getIncomingValueForBlock(BIParent); + if (!FalseV) + FalseV = PHIV; + else if (FalseV != PHIV) + return false; // Don't know the value when condition is false. + } + if (!FalseV) // Can this happen? + return false; + + // 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 + // being sunk into the use block. + for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) { + Instruction *OpI = dyn_cast(*i); + if (OpI && OpI->getParent() == BIParent && + !OpI->isUsedInBasicBlock(BIParent)) + return false; + } + + // If we get here, we can hoist the instruction. Try to place it before the + // icmp instruction preceeding the conditional branch. + BasicBlock::iterator InsertPos = BI; + if (InsertPos != BIParent->begin()) + --InsertPos; + if (InsertPos == BrCond && !isa(BrCond)) + BIParent->getInstList().splice(InsertPos, BB1->getInstList(), I); + else + BIParent->getInstList().splice(BI, BB1->getInstList(), I); + + // Create a select whose true value is the speculatively executed value and + // false value is the previously determined FalseV. + SelectInst *SI; + if (Invert) + SI = SelectInst::Create(BrCond, FalseV, I, + FalseV->getName() + "." + I->getName(), BI); + else + SI = SelectInst::Create(BrCond, I, FalseV, + I->getName() + "." + FalseV->getName(), BI); + + // Make the PHI node use the select for all incoming values for "then" and + // "if" blocks. + for (unsigned i = 0, e = PHIUses.size(); i != e; ++i) { + PHINode *PN = PHIUses[i]; + for (unsigned j = 0, ee = PN->getNumIncomingValues(); j != ee; ++j) + if (PN->getIncomingBlock(j) == BB1 || + PN->getIncomingBlock(j) == BIParent) + PN->setIncomingValue(j, SI); + } + + ++NumSpeculations; + return true; +} + +/// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch +/// across this block. +static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) { + BranchInst *BI = cast(BB->getTerminator()); + unsigned Size = 0; + + // If this basic block contains anything other than a PHI (which controls the + // branch) and branch itself, bail out. FIXME: improve this in the future. + for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI, ++Size) { + if (Size > 10) return false; // Don't clone large BB's. + + // We can only support instructions that are do not define values that are + // live outside of the current basic block. + for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end(); + UI != E; ++UI) { + Instruction *U = cast(*UI); + if (U->getParent() != BB || isa(U)) return false; + } + + // Looks ok, continue checking. + } + + return true; +} + +/// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value +/// that is defined in the same block as the branch and if any PHI entries are +/// constants, thread edges corresponding to that entry to be branches to their +/// ultimate destination. +static bool FoldCondBranchOnPHI(BranchInst *BI) { + BasicBlock *BB = BI->getParent(); + PHINode *PN = dyn_cast(BI->getCondition()); + // NOTE: we currently cannot transform this case if the PHI node is used + // outside of the block. + if (!PN || PN->getParent() != BB || !PN->hasOneUse()) + return false; + + // Degenerate case of a single entry PHI. + if (PN->getNumIncomingValues() == 1) { + if (PN->getIncomingValue(0) != PN) + PN->replaceAllUsesWith(PN->getIncomingValue(0)); + else + PN->replaceAllUsesWith(UndefValue::get(PN->getType())); + PN->eraseFromParent(); + return true; + } + + // Now we know that this block has multiple preds and two succs. + if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false; + + // Okay, this is a simple enough basic block. See if any phi values are + // constants. + for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { + ConstantInt *CB; + if ((CB = dyn_cast(PN->getIncomingValue(i))) && + CB->getType() == Type::Int1Ty) { + // Okay, we now know that all edges from PredBB should be revectored to + // branch to RealDest. + BasicBlock *PredBB = PN->getIncomingBlock(i); + BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue()); + + if (RealDest == BB) continue; // Skip self loops. + + // The dest block might have PHI nodes, other predecessors and other + // difficult cases. Instead of being smart about this, just insert a new + // block that jumps to the destination block, effectively splitting + // the edge we are about to create. + BasicBlock *EdgeBB = BasicBlock::Create(RealDest->getName()+".critedge", + RealDest->getParent(), RealDest); + BranchInst::Create(RealDest, EdgeBB); + PHINode *PN; + for (BasicBlock::iterator BBI = RealDest->begin(); + (PN = dyn_cast(BBI)); ++BBI) { + Value *V = PN->getIncomingValueForBlock(BB); + PN->addIncoming(V, EdgeBB); + } + + // BB may have instructions that are being threaded over. Clone these + // instructions into EdgeBB. We know that there will be no uses of the + // cloned instructions outside of EdgeBB. + BasicBlock::iterator InsertPt = EdgeBB->begin(); + std::map TranslateMap; // Track translated values. + for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { + if (PHINode *PN = dyn_cast(BBI)) { + TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB); + } else { + // Clone the instruction. + Instruction *N = BBI->clone(); + if (BBI->hasName()) N->setName(BBI->getName()+".c"); + + // Update operands due to translation. + for (User::op_iterator i = N->op_begin(), e = N->op_end(); + i != e; ++i) { + std::map::iterator PI = + TranslateMap.find(*i); + if (PI != TranslateMap.end()) + *i = PI->second; + } + + // Check for trivial simplification. + if (Constant *C = ConstantFoldInstruction(N)) { + TranslateMap[BBI] = C; + delete N; // Constant folded away, don't need actual inst + } else { + // Insert the new instruction into its new home. + EdgeBB->getInstList().insert(InsertPt, N); + if (!BBI->use_empty()) + TranslateMap[BBI] = N; + } + } + } + + // Loop over all of the edges from PredBB to BB, changing them to branch + // to EdgeBB instead. + TerminatorInst *PredBBTI = PredBB->getTerminator(); + for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i) + if (PredBBTI->getSuccessor(i) == BB) { + BB->removePredecessor(PredBB); + PredBBTI->setSuccessor(i, EdgeBB); + } + + // Recurse, simplifying any other constants. + return FoldCondBranchOnPHI(BI) | true; + } + } + + return false; +} + +/// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry +/// PHI node, see if we can eliminate it. +static bool FoldTwoEntryPHINode(PHINode *PN) { + // Ok, this is a two entry PHI node. Check to see if this is a simple "if + // statement", which has a very simple dominance structure. Basically, we + // are trying to find the condition that is being branched on, which + // subsequently causes this merge to happen. We really want control + // dependence information for this check, but simplifycfg can't keep it up + // to date, and this catches most of the cases we care about anyway. + // + BasicBlock *BB = PN->getParent(); + BasicBlock *IfTrue, *IfFalse; + Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse); + if (!IfCond) return false; + + // Okay, we found that we can merge this two-entry phi node into a select. + // Doing so would require us to fold *all* two entry phi nodes in this block. + // At some point this becomes non-profitable (particularly if the target + // doesn't support cmov's). Only do this transformation if there are two or + // fewer PHI nodes in this block. + unsigned NumPhis = 0; + for (BasicBlock::iterator I = BB->begin(); isa(I); ++NumPhis, ++I) + if (NumPhis > 2) + return false; + DOUT << "FOUND IF CONDITION! " << *IfCond << " T: " + << IfTrue->getName() << " F: " << IfFalse->getName() << "\n"; + + // Loop over the PHI's seeing if we can promote them all to select + // instructions. While we are at it, keep track of the instructions + // that need to be moved to the dominating block. + std::set AggressiveInsts; + + BasicBlock::iterator AfterPHIIt = BB->begin(); + while (isa(AfterPHIIt)) { + PHINode *PN = cast(AfterPHIIt++); + if (PN->getIncomingValue(0) == PN->getIncomingValue(1)) { + if (PN->getIncomingValue(0) != PN) + PN->replaceAllUsesWith(PN->getIncomingValue(0)); + else + PN->replaceAllUsesWith(UndefValue::get(PN->getType())); + } else if (!DominatesMergePoint(PN->getIncomingValue(0), BB, + &AggressiveInsts) || + !DominatesMergePoint(PN->getIncomingValue(1), BB, + &AggressiveInsts)) { + return false; + } + } + + // If we all PHI nodes are promotable, check to make sure that all + // instructions in the predecessor blocks can be promoted as well. If + // not, we won't be able to get rid of the control flow, so it's not + // worth promoting to select instructions. + BasicBlock *DomBlock = 0, *IfBlock1 = 0, *IfBlock2 = 0; + PN = cast(BB->begin()); + BasicBlock *Pred = PN->getIncomingBlock(0); + if (cast(Pred->getTerminator())->isUnconditional()) { + IfBlock1 = Pred; + DomBlock = *pred_begin(Pred); + for (BasicBlock::iterator I = Pred->begin(); + !isa(I); ++I) + if (!AggressiveInsts.count(I)) { + // This is not an aggressive instruction that we can promote. + // Because of this, we won't be able to get rid of the control + // flow, so the xform is not worth it. + return false; + } + } + + Pred = PN->getIncomingBlock(1); + if (cast(Pred->getTerminator())->isUnconditional()) { + IfBlock2 = Pred; + DomBlock = *pred_begin(Pred); + for (BasicBlock::iterator I = Pred->begin(); + !isa(I); ++I) + if (!AggressiveInsts.count(I)) { + // This is not an aggressive instruction that we can promote. + // Because of this, we won't be able to get rid of the control + // flow, so the xform is not worth it. + return false; + } + } + + // If we can still promote the PHI nodes after this gauntlet of tests, + // do all of the PHI's now. + + // Move all 'aggressive' instructions, which are defined in the + // conditional parts of the if's up to the dominating block. + if (IfBlock1) { + DomBlock->getInstList().splice(DomBlock->getTerminator(), + IfBlock1->getInstList(), + IfBlock1->begin(), + IfBlock1->getTerminator()); + } + if (IfBlock2) { + DomBlock->getInstList().splice(DomBlock->getTerminator(), + IfBlock2->getInstList(), + IfBlock2->begin(), + IfBlock2->getTerminator()); + } + + while (PHINode *PN = dyn_cast(BB->begin())) { + // Change the PHI node into a select instruction. + Value *TrueVal = + PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse); + Value *FalseVal = + PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue); + + Value *NV = SelectInst::Create(IfCond, TrueVal, FalseVal, "", AfterPHIIt); + PN->replaceAllUsesWith(NV); + NV->takeName(PN); + + BB->getInstList().erase(PN); + } + return true; +} + +/// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes +/// to two returning blocks, try to merge them together into one return, +/// introducing a select if the return values disagree. +static bool SimplifyCondBranchToTwoReturns(BranchInst *BI) { + assert(BI->isConditional() && "Must be a conditional branch"); + BasicBlock *TrueSucc = BI->getSuccessor(0); + BasicBlock *FalseSucc = BI->getSuccessor(1); + ReturnInst *TrueRet = cast(TrueSucc->getTerminator()); + ReturnInst *FalseRet = cast(FalseSucc->getTerminator()); + + // Check to ensure both blocks are empty (just a return) or optionally empty + // with PHI nodes. If there are other instructions, merging would cause extra + // computation on one path or the other. + BasicBlock::iterator BBI = TrueRet; + if (BBI != TrueSucc->begin() && !isa(--BBI)) + return false; // Not empty with optional phi nodes. + BBI = FalseRet; + if (BBI != FalseSucc->begin() && !isa(--BBI)) + return false; // Not empty with optional phi nodes. + + // Okay, we found a branch that is going to two return nodes. If + // there is no return value for this function, just change the + // branch into a return. + if (FalseRet->getNumOperands() == 0) { + TrueSucc->removePredecessor(BI->getParent()); + FalseSucc->removePredecessor(BI->getParent()); + ReturnInst::Create(0, BI); + BI->eraseFromParent(); + return true; + } + + // Otherwise, build up the result values for the new return. + SmallVector TrueResult; + SmallVector FalseResult; + + for (unsigned i = 0, e = TrueRet->getNumOperands(); i != e; ++i) { + // Otherwise, figure out what the true and false return values are + // so we can insert a new select instruction. + Value *TrueValue = TrueRet->getOperand(i); + Value *FalseValue = FalseRet->getOperand(i); + + // Unwrap any PHI nodes in the return blocks. + if (PHINode *TVPN = dyn_cast(TrueValue)) + if (TVPN->getParent() == TrueSucc) + TrueValue = TVPN->getIncomingValueForBlock(BI->getParent()); + if (PHINode *FVPN = dyn_cast(FalseValue)) + if (FVPN->getParent() == FalseSucc) + FalseValue = FVPN->getIncomingValueForBlock(BI->getParent()); + + // In order for this transformation to be safe, we must be able to + // unconditionally execute both operands to the return. This is + // normally the case, but we could have a potentially-trapping + // constant expression that prevents this transformation from being + // safe. + if (ConstantExpr *TCV = dyn_cast(TrueValue)) + if (TCV->canTrap()) + return false; + if (ConstantExpr *FCV = dyn_cast(FalseValue)) + if (FCV->canTrap()) + return false; + + TrueResult.push_back(TrueValue); + FalseResult.push_back(FalseValue); + } + + // Okay, we collected all the mapped values and checked them for sanity, and + // defined to really do this transformation. First, update the CFG. + TrueSucc->removePredecessor(BI->getParent()); + FalseSucc->removePredecessor(BI->getParent()); + + // Insert select instructions where needed. + Value *BrCond = BI->getCondition(); + for (unsigned i = 0, e = TrueRet->getNumOperands(); i != e; ++i) { + // Insert a select if the results differ. + if (TrueResult[i] == FalseResult[i] || isa(FalseResult[i])) + continue; + if (isa(TrueResult[i])) { + TrueResult[i] = FalseResult[i]; + continue; + } + + TrueResult[i] = SelectInst::Create(BrCond, TrueResult[i], + FalseResult[i], "retval", BI); + } + + Value *RI = ReturnInst::Create(&TrueResult[0], TrueResult.size(), BI); + + DOUT << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:" + << "\n " << *BI << "NewRet = " << *RI + << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc; + BI->eraseFromParent(); + + if (Instruction *BrCondI = dyn_cast(BrCond)) + ErasePossiblyDeadInstructionTree(BrCondI); + return true; +} + +/// FoldBranchToCommonDest - If this basic block is ONLY a setcc and a branch, +/// and if a predecessor branches to us and one of our successors, fold the +/// setcc into the predecessor and use logical operations to pick the right +/// destination. +static bool FoldBranchToCommonDest(BranchInst *BI) { + Instruction *Cond = dyn_cast(BI->getCondition()); + if (Cond == 0) return false; + + BasicBlock *BB = BI->getParent(); + + // Only allow this if the condition is a simple instruction that can be + // executed unconditionally. It must be in the same block as the branch, and + // must be at the front of the block. + if ((!isa(Cond) && !isa(Cond)) || + Cond->getParent() != BB || &BB->front() != Cond || !Cond->hasOneUse()) + return false; + + // Make sure the instruction after the condition is the cond branch. + BasicBlock::iterator CondIt = Cond; ++CondIt; + if (&*CondIt != BI) + return false; + + // Finally, don't infinitely unroll conditional loops. + BasicBlock *TrueDest = BI->getSuccessor(0); + BasicBlock *FalseDest = BI->getSuccessor(1); + if (TrueDest == BB || FalseDest == BB) + return false; + + for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { + BasicBlock *PredBlock = *PI; + BranchInst *PBI = dyn_cast(PredBlock->getTerminator()); + if (PBI == 0 || PBI->isUnconditional() || + !SafeToMergeTerminators(BI, PBI)) + continue; + + Instruction::BinaryOps Opc; + bool InvertPredCond = false; + + if (PBI->getSuccessor(0) == TrueDest) + Opc = Instruction::Or; + else if (PBI->getSuccessor(1) == FalseDest) + Opc = Instruction::And; + else if (PBI->getSuccessor(0) == FalseDest) + Opc = Instruction::And, InvertPredCond = true; + else if (PBI->getSuccessor(1) == TrueDest) + Opc = Instruction::Or, InvertPredCond = true; + else + continue; + + // If we need to invert the condition in the pred block to match, do so now. + if (InvertPredCond) { + Value *NewCond = + BinaryOperator::CreateNot(PBI->getCondition(), + PBI->getCondition()->getName()+".not", PBI); + PBI->setCondition(NewCond); + BasicBlock *OldTrue = PBI->getSuccessor(0); + BasicBlock *OldFalse = PBI->getSuccessor(1); + PBI->setSuccessor(0, OldFalse); + PBI->setSuccessor(1, OldTrue); + } + + // Clone Cond into the predecessor basic block, and or/and the + // two conditions together. + Instruction *New = Cond->clone(); + PredBlock->getInstList().insert(PBI, New); + New->takeName(Cond); + Cond->setName(New->getName()+".old"); + + Value *NewCond = BinaryOperator::Create(Opc, PBI->getCondition(), + New, "or.cond", PBI); + PBI->setCondition(NewCond); + if (PBI->getSuccessor(0) == BB) { + AddPredecessorToBlock(TrueDest, PredBlock, BB); + PBI->setSuccessor(0, TrueDest); + } + if (PBI->getSuccessor(1) == BB) { + AddPredecessorToBlock(FalseDest, PredBlock, BB); + PBI->setSuccessor(1, FalseDest); + } + return true; + } + return false; +} + +/// SimplifyCondBranchToCondBranch - If we have a conditional branch as a +/// predecessor of another block, this function tries to simplify it. We know +/// that PBI and BI are both conditional branches, and BI is in one of the +/// successor blocks of PBI - PBI branches to BI. +static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { + assert(PBI->isConditional() && BI->isConditional()); + BasicBlock *BB = BI->getParent(); + + // If this block ends with a branch instruction, and if there is a + // predecessor that ends on a branch of the same condition, make + // this conditional branch redundant. + if (PBI->getCondition() == BI->getCondition() && + PBI->getSuccessor(0) != PBI->getSuccessor(1)) { + // Okay, the outcome of this conditional branch is statically + // knowable. If this block had a single pred, handle specially. + if (BB->getSinglePredecessor()) { + // Turn this into a branch on constant. + bool CondIsTrue = PBI->getSuccessor(0) == BB; + BI->setCondition(ConstantInt::get(Type::Int1Ty, CondIsTrue)); + return true; // Nuke the branch on constant. + } + + // Otherwise, if there are multiple predecessors, insert a PHI that merges + // in the constant and simplify the block result. Subsequent passes of + // simplifycfg will thread the block. + if (BlockIsSimpleEnoughToThreadThrough(BB)) { + PHINode *NewPN = PHINode::Create(Type::Int1Ty, + BI->getCondition()->getName() + ".pr", + BB->begin()); + // Okay, we're going to insert the PHI node. Since PBI is not the only + // predecessor, compute the PHI'd conditional value for all of the preds. + // Any predecessor where the condition is not computable we keep symbolic. + for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) + if ((PBI = dyn_cast((*PI)->getTerminator())) && + PBI != BI && PBI->isConditional() && + PBI->getCondition() == BI->getCondition() && + PBI->getSuccessor(0) != PBI->getSuccessor(1)) { + bool CondIsTrue = PBI->getSuccessor(0) == BB; + NewPN->addIncoming(ConstantInt::get(Type::Int1Ty, + CondIsTrue), *PI); + } else { + NewPN->addIncoming(BI->getCondition(), *PI); + } + + BI->setCondition(NewPN); + return true; + } + } + + // If this is a conditional branch in an empty block, and if any + // predecessors is a conditional branch to one of our destinations, + // fold the conditions into logical ops and one cond br. + if (&BB->front() != BI) + return false; + + int PBIOp, BIOp; + if (PBI->getSuccessor(0) == BI->getSuccessor(0)) + PBIOp = BIOp = 0; + else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) + PBIOp = 0, BIOp = 1; + else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) + PBIOp = 1, BIOp = 0; + else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) + PBIOp = BIOp = 1; + else + return false; + + // Check to make sure that the other destination of this branch + // isn't BB itself. If so, this is an infinite loop that will + // keep getting unwound. + if (PBI->getSuccessor(PBIOp) == BB) + return false; + + // Do not perform this transformation if it would require + // insertion of a large number of select instructions. For targets + // without predication/cmovs, this is a big pessimization. + BasicBlock *CommonDest = PBI->getSuccessor(PBIOp); + + unsigned NumPhis = 0; + for (BasicBlock::iterator II = CommonDest->begin(); + isa(II); ++II, ++NumPhis) + if (NumPhis > 2) // Disable this xform. + return false; + + // Finally, if everything is ok, fold the branches to logical ops. + BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1); + + // If OtherDest *is* BB, then this is a basic block with just + // a conditional branch in it, where one edge (OtherDesg) goes + // back to the block. We know that the program doesn't get + // stuck in the infinite loop, so the condition must be such + // that OtherDest isn't branched through. Forward to CommonDest, + // and avoid an infinite loop at optimizer time. + if (OtherDest == BB) + OtherDest = CommonDest; + + DOUT << "FOLDING BRs:" << *PBI->getParent() + << "AND: " << *BI->getParent(); + + DOUT << *PBI->getParent()->getParent(); + + // BI may have other predecessors. Because of this, we leave + // it alone, but modify PBI. + + // Make sure we get to CommonDest on True&True directions. + Value *PBICond = PBI->getCondition(); + if (PBIOp) + PBICond = BinaryOperator::CreateNot(PBICond, + PBICond->getName()+".not", + PBI); + Value *BICond = BI->getCondition(); + if (BIOp) + BICond = BinaryOperator::CreateNot(BICond, + BICond->getName()+".not", + PBI); + // Merge the conditions. + Value *Cond = BinaryOperator::CreateOr(PBICond, BICond, "brmerge", PBI); + + // Modify PBI to branch on the new condition to the new dests. + PBI->setCondition(Cond); + PBI->setSuccessor(0, CommonDest); + PBI->setSuccessor(1, OtherDest); + + // OtherDest may have phi nodes. If so, add an entry from PBI's + // block that are identical to the entries for BI's block. + PHINode *PN; + for (BasicBlock::iterator II = OtherDest->begin(); + (PN = dyn_cast(II)); ++II) { + Value *V = PN->getIncomingValueForBlock(BB); + PN->addIncoming(V, PBI->getParent()); + } + + // We know that the CommonDest already had an edge from PBI to + // it. If it has PHIs though, the PHIs may have different + // entries for BB and PBI's BB. If so, insert a select to make + // them agree. + for (BasicBlock::iterator II = CommonDest->begin(); + (PN = dyn_cast(II)); ++II) { + Value *BIV = PN->getIncomingValueForBlock(BB); + unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent()); + Value *PBIV = PN->getIncomingValue(PBBIdx); + if (BIV != PBIV) { + // Insert a select in PBI to pick the right value. + Value *NV = SelectInst::Create(PBICond, PBIV, BIV, + PBIV->getName()+".mux", PBI); + PN->setIncomingValue(PBBIdx, NV); + } + } + + DOUT << "INTO: " << *PBI->getParent(); + + DOUT << *PBI->getParent()->getParent(); + + // This basic block is probably dead. We know it has at least + // one fewer predecessor. return true; } + namespace { /// ConstantIntOrdering - This class implements a stable ordering of constant /// integers that does not depend on their address. This is important for /// applications that sort ConstantInt's to ensure uniqueness. struct ConstantIntOrdering { bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const { - return LHS->getRawValue() < RHS->getRawValue(); + return LHS->getValue().ult(RHS->getValue()); } }; } - // SimplifyCFG - This function is used to do simplification of a CFG. For // example, it adjusts branches to branches to eliminate the extra hop, it // eliminates unreachable basic blocks, and does other "peephole" optimization @@ -834,12 +1692,13 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { assert(BB && BB->getParent() && "Block not embedded in function!"); assert(BB->getTerminator() && "Degenerate basic block encountered!"); - assert(&BB->getParent()->front() != BB && "Can't Simplify entry block!"); + assert(&BB->getParent()->getEntryBlock() != BB && + "Can't Simplify entry block!"); // Remove basic blocks that have no predecessors... which are unreachable. - if (pred_begin(BB) == pred_end(BB) || - *pred_begin(BB) == BB && ++pred_begin(BB) == pred_end(BB)) { - DEBUG(std::cerr << "Removing BB: \n" << *BB); + if ((pred_begin(BB) == pred_end(BB)) || + (*pred_begin(BB) == BB && ++pred_begin(BB) == pred_end(BB))) { + DOUT << "Removing BB: \n" << *BB; // Loop through all of our successors and make sure they know that one // of their predecessors is going away. @@ -849,14 +1708,14 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { while (!BB->empty()) { Instruction &I = BB->back(); // If this instruction is used, replace uses with an arbitrary - // constant value. Because control flow can't get here, we don't care - // what we replace the value with. Note that since this block is + // value. Because control flow can't get here, we don't care + // what we replace the value with. Note that since this block is // unreachable, and all values contained within it must dominate their // uses, that all uses will eventually be removed. - if (!I.use_empty()) - // Make all users of this instruction reference the constant instead - I.replaceAllUsesWith(Constant::getNullValue(I.getType())); - + if (!I.use_empty()) + // Make all users of this instruction use undef instead + I.replaceAllUsesWith(UndefValue::get(I.getType())); + // Remove the instruction from the basic block BB->getInstList().pop_back(); } @@ -868,65 +1727,11 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { // away... Changed |= ConstantFoldTerminator(BB); - // Check to see if this block has no non-phi instructions and only a single - // successor. If so, replace references to this basic block with references - // to the successor. - succ_iterator SI(succ_begin(BB)); - if (SI != succ_end(BB) && ++SI == succ_end(BB)) { // One succ? - BasicBlock::iterator BBI = BB->begin(); // Skip over phi nodes... - while (isa(*BBI)) ++BBI; - - BasicBlock *Succ = *succ_begin(BB); // There is exactly one successor. - if (BBI->isTerminator() && // Terminator is the only non-phi instruction! - Succ != BB) { // Don't hurt infinite loops! - // If our successor has PHI nodes, then we need to update them to include - // entries for BB's predecessors, not for BB itself. Be careful though, - // if this transformation fails (returns true) then we cannot do this - // transformation! - // - if (!PropagatePredecessorsForPHIs(BB, Succ)) { - DEBUG(std::cerr << "Killing Trivial BB: \n" << *BB); - - if (isa(&BB->front())) { - std::vector - OldSuccPreds(pred_begin(Succ), pred_end(Succ)); - - // Move all PHI nodes in BB to Succ if they are alive, otherwise - // delete them. - while (PHINode *PN = dyn_cast(&BB->front())) - if (PN->use_empty()) - BB->getInstList().erase(BB->begin()); // Nuke instruction. - else { - // The instruction is alive, so this means that Succ must have - // *ONLY* had BB as a predecessor, and the PHI node is still valid - // now. Simply move it into Succ, because we know that BB - // strictly dominated Succ. - BB->getInstList().remove(BB->begin()); - Succ->getInstList().push_front(PN); - - // We need to add new entries for the PHI node to account for - // predecessors of Succ that the PHI node does not take into - // account. At this point, since we know that BB dominated succ, - // this means that we should any newly added incoming edges should - // use the PHI node as the value for these edges, because they are - // loop back edges. - for (unsigned i = 0, e = OldSuccPreds.size(); i != e; ++i) - if (OldSuccPreds[i] != BB) - PN->addIncoming(PN, OldSuccPreds[i]); - } - } - - // Everything that jumped to BB now goes to Succ. - std::string OldName = BB->getName(); - BB->replaceAllUsesWith(Succ); - BB->eraseFromParent(); // Delete the old basic block. - - if (!OldName.empty() && !Succ->hasName()) // Transfer name if we can - Succ->setName(OldName); - return true; - } - } - } + // If there is a trivial two-entry PHI node in this basic block, and we can + // eliminate it, do so now. + if (PHINode *PN = dyn_cast(BB->begin())) + if (PN->getNumIncomingValues() == 2) + Changed |= FoldTwoEntryPHINode(PN); // If this is a returning block with only PHI nodes in it, fold the return // instruction into any unconditional branch predecessors. @@ -938,21 +1743,24 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { BasicBlock::iterator BBI = BB->getTerminator(); if (BBI == BB->begin() || isa(--BBI)) { // Find predecessors that end with branches. - std::vector UncondBranchPreds; - std::vector CondBranchPreds; + SmallVector UncondBranchPreds; + SmallVector CondBranchPreds; for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { TerminatorInst *PTI = (*PI)->getTerminator(); - if (BranchInst *BI = dyn_cast(PTI)) + if (BranchInst *BI = dyn_cast(PTI)) { if (BI->isUnconditional()) UncondBranchPreds.push_back(*PI); else CondBranchPreds.push_back(BI); + } } - + // If we found some, do the transformation! if (!UncondBranchPreds.empty()) { while (!UncondBranchPreds.empty()) { BasicBlock *Pred = UncondBranchPreds.back(); + DOUT << "FOLDING: " << *BB + << "INTO UNCOND BRANCH PRED: " << *Pred; UncondBranchPreds.pop_back(); Instruction *UncondBranch = Pred->getTerminator(); // Clone the return and add it to the end of the predecessor. @@ -961,10 +1769,12 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { // If the return instruction returns a value, and if the value was a // PHI node in "BB", propagate the right value into the return. - if (NewRet->getNumOperands() == 1) - if (PHINode *PN = dyn_cast(NewRet->getOperand(0))) + for (User::op_iterator i = NewRet->op_begin(), e = NewRet->op_end(); + i != e; ++i) + if (PHINode *PN = dyn_cast(*i)) if (PN->getParent() == BB) - NewRet->setOperand(0, PN->getIncomingValueForBlock(Pred)); + *i = PN->getIncomingValueForBlock(Pred); + // Update any PHI nodes in the returning block to realize that we no // longer branch to them. BB->removePredecessor(Pred); @@ -985,68 +1795,21 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { while (!CondBranchPreds.empty()) { BranchInst *BI = CondBranchPreds.back(); CondBranchPreds.pop_back(); - BasicBlock *TrueSucc = BI->getSuccessor(0); - BasicBlock *FalseSucc = BI->getSuccessor(1); - BasicBlock *OtherSucc = TrueSucc == BB ? FalseSucc : TrueSucc; // Check to see if the non-BB successor is also a return block. - if (isa(OtherSucc->getTerminator())) { - // Check to see if there are only PHI instructions in this block. - BasicBlock::iterator OSI = OtherSucc->getTerminator(); - if (OSI == OtherSucc->begin() || isa(--OSI)) { - // Okay, we found a branch that is going to two return nodes. If - // there is no return value for this function, just change the - // branch into a return. - if (RI->getNumOperands() == 0) { - TrueSucc->removePredecessor(BI->getParent()); - FalseSucc->removePredecessor(BI->getParent()); - new ReturnInst(0, BI); - BI->getParent()->getInstList().erase(BI); - return true; - } - - // Otherwise, figure out what the true and false return values are - // so we can insert a new select instruction. - Value *TrueValue = TrueSucc->getTerminator()->getOperand(0); - Value *FalseValue = FalseSucc->getTerminator()->getOperand(0); - - // Unwrap any PHI nodes in the return blocks. - if (PHINode *TVPN = dyn_cast(TrueValue)) - if (TVPN->getParent() == TrueSucc) - TrueValue = TVPN->getIncomingValueForBlock(BI->getParent()); - if (PHINode *FVPN = dyn_cast(FalseValue)) - if (FVPN->getParent() == FalseSucc) - FalseValue = FVPN->getIncomingValueForBlock(BI->getParent()); - - TrueSucc->removePredecessor(BI->getParent()); - FalseSucc->removePredecessor(BI->getParent()); - - // Insert a new select instruction. - Value *NewRetVal; - Value *BrCond = BI->getCondition(); - if (TrueValue != FalseValue) - NewRetVal = new SelectInst(BrCond, TrueValue, - FalseValue, "retval", BI); - else - NewRetVal = TrueValue; - - new ReturnInst(NewRetVal, BI); - BI->getParent()->getInstList().erase(BI); - if (BrCond->use_empty()) - if (Instruction *BrCondI = dyn_cast(BrCond)) - BrCondI->getParent()->getInstList().erase(BrCondI); - return true; - } - } + if (isa(BI->getSuccessor(0)->getTerminator()) && + isa(BI->getSuccessor(1)->getTerminator()) && + SimplifyCondBranchToTwoReturns(BI)) + return true; } } - } else if (UnwindInst *UI = dyn_cast(BB->begin())) { + } else if (isa(BB->begin())) { // Check to see if the first instruction in this block is just an unwind. // If so, replace any invoke instructions which use this as an exception // destination with call instructions, and any unconditional branch // predecessor with an unwind. // - std::vector Preds(pred_begin(BB), pred_end(BB)); + SmallVector Preds(pred_begin(BB), pred_end(BB)); while (!Preds.empty()) { BasicBlock *Pred = Preds.back(); if (BranchInst *BI = dyn_cast(Pred->getTerminator())) { @@ -1059,19 +1822,22 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { if (II->getUnwindDest() == BB) { // Insert a new branch instruction before the invoke, because this // is now a fall through... - BranchInst *BI = new BranchInst(II->getNormalDest(), II); + BranchInst *BI = BranchInst::Create(II->getNormalDest(), II); Pred->getInstList().remove(II); // Take out of symbol table - + // Insert the call now... - std::vector Args(II->op_begin()+3, II->op_end()); - CallInst *CI = new CallInst(II->getCalledValue(), Args, - II->getName(), BI); + SmallVector Args(II->op_begin()+3, II->op_end()); + CallInst *CI = CallInst::Create(II->getCalledValue(), + Args.begin(), Args.end(), + II->getName(), BI); + CI->setCallingConv(II->getCallingConv()); + CI->setParamAttrs(II->getParamAttrs()); // If the invoke produced a value, the Call now does instead II->replaceAllUsesWith(CI); delete II; Changed = true; } - + Preds.pop_back(); } @@ -1097,8 +1863,17 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { return SimplifyCFG(BB) || 1; } } else if (BranchInst *BI = dyn_cast(BB->getTerminator())) { - if (BI->isConditional()) { - if (Value *CompVal = isValueEqualityComparison(BI)) { + if (BI->isUnconditional()) { + BasicBlock::iterator BBI = BB->getFirstNonPHI(); + + BasicBlock *Succ = BI->getSuccessor(0); + if (BBI->isTerminator() && // Terminator is the only non-phi instruction! + Succ != BB) // Don't hurt infinite loops! + if (TryToSimplifyUncondBranchFromEmptyBlock(BB, Succ)) + return true; + + } else { // Conditional branch + if (isValueEqualityComparison(BI)) { // If we only have one predecessor, and if it is a branch on this value, // see if that predecessor totally determines the outcome of this // switch. @@ -1114,86 +1889,27 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { if (FoldValueComparisonIntoPredecessors(BI)) return SimplifyCFG(BB) | true; } + + // If this is a branch on a phi node in the current block, thread control + // through this block if any PHI node entries are constants. + if (PHINode *PN = dyn_cast(BI->getCondition())) + if (PN->getParent() == BI->getParent()) + if (FoldCondBranchOnPHI(BI)) + return SimplifyCFG(BB) | true; // If this basic block is ONLY a setcc and a branch, and if a predecessor // branches to us and one of our successors, fold the setcc into the // predecessor and use logical operations to pick the right destination. - BasicBlock *TrueDest = BI->getSuccessor(0); - BasicBlock *FalseDest = BI->getSuccessor(1); - if (BinaryOperator *Cond = dyn_cast(BI->getCondition())) - if (Cond->getParent() == BB && &BB->front() == Cond && - Cond->getNext() == BI && Cond->hasOneUse() && - TrueDest != BB && FalseDest != BB) - for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI!=E; ++PI) - if (BranchInst *PBI = dyn_cast((*PI)->getTerminator())) - if (PBI->isConditional() && SafeToMergeTerminators(BI, PBI)) { - BasicBlock *PredBlock = *PI; - if (PBI->getSuccessor(0) == FalseDest || - PBI->getSuccessor(1) == TrueDest) { - // Invert the predecessors condition test (xor it with true), - // which allows us to write this code once. - Value *NewCond = - BinaryOperator::createNot(PBI->getCondition(), - PBI->getCondition()->getName()+".not", PBI); - PBI->setCondition(NewCond); - BasicBlock *OldTrue = PBI->getSuccessor(0); - BasicBlock *OldFalse = PBI->getSuccessor(1); - PBI->setSuccessor(0, OldFalse); - PBI->setSuccessor(1, OldTrue); - } + if (FoldBranchToCommonDest(BI)) + return SimplifyCFG(BB) | 1; - if (PBI->getSuccessor(0) == TrueDest || - PBI->getSuccessor(1) == FalseDest) { - // Clone Cond into the predecessor basic block, and or/and the - // two conditions together. - Instruction *New = Cond->clone(); - New->setName(Cond->getName()); - Cond->setName(Cond->getName()+".old"); - PredBlock->getInstList().insert(PBI, New); - Instruction::BinaryOps Opcode = - PBI->getSuccessor(0) == TrueDest ? - Instruction::Or : Instruction::And; - Value *NewCond = - BinaryOperator::create(Opcode, PBI->getCondition(), - New, "bothcond", PBI); - PBI->setCondition(NewCond); - if (PBI->getSuccessor(0) == BB) { - AddPredecessorToBlock(TrueDest, PredBlock, BB); - PBI->setSuccessor(0, TrueDest); - } - if (PBI->getSuccessor(1) == BB) { - AddPredecessorToBlock(FalseDest, PredBlock, BB); - PBI->setSuccessor(1, FalseDest); - } - return SimplifyCFG(BB) | 1; - } - } - // If this block ends with a branch instruction, and if there is one - // predecessor, see if the previous block ended with a branch on the same - // condition, which makes this conditional branch redundant. - pred_iterator PI(pred_begin(BB)), PE(pred_end(BB)); - BasicBlock *OnlyPred = *PI++; - for (; PI != PE; ++PI)// Search all predecessors, see if they are all same - if (*PI != OnlyPred) { - OnlyPred = 0; // There are multiple different predecessors... - break; - } - - if (OnlyPred) - if (BranchInst *PBI = dyn_cast(OnlyPred->getTerminator())) - if (PBI->isConditional() && - PBI->getCondition() == BI->getCondition() && - (PBI->getSuccessor(0) != BB || PBI->getSuccessor(1) != BB)) { - // Okay, the outcome of this conditional branch is statically - // knowable. Delete the outgoing CFG edge that is impossible to - // execute. - bool CondIsTrue = PBI->getSuccessor(0) == BB; - BI->getSuccessor(CondIsTrue)->removePredecessor(BB); - new BranchInst(BI->getSuccessor(!CondIsTrue), BB); - BB->getInstList().erase(BI); - return SimplifyCFG(BB) | true; - } + // Scan predecessor blocks for conditional branches. + for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) + if (BranchInst *PBI = dyn_cast((*PI)->getTerminator())) + if (PBI != BI && PBI->isConditional()) + if (SimplifyCondBranchToCondBranch(PBI, BI)) + return SimplifyCFG(BB) | true; } } else if (isa(BB->getTerminator())) { // If there are any instructions immediately before the unreachable that can @@ -1211,7 +1927,7 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { // If the unreachable instruction is the first in the block, take a gander // at all of the predecessors of this instruction, and simplify them. if (&BB->front() == Unreachable) { - std::vector Preds(pred_begin(BB), pred_end(BB)); + SmallVector Preds(pred_begin(BB), pred_end(BB)); for (unsigned i = 0, e = Preds.size(); i != e; ++i) { TerminatorInst *TI = Preds[i]->getTerminator(); @@ -1224,10 +1940,10 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { } } else { if (BI->getSuccessor(0) == BB) { - new BranchInst(BI->getSuccessor(1), BI); + BranchInst::Create(BI->getSuccessor(1), BI); BI->eraseFromParent(); } else if (BI->getSuccessor(1) == BB) { - new BranchInst(BI->getSuccessor(0), BI); + BranchInst::Create(BI->getSuccessor(0), BI); BI->eraseFromParent(); Changed = true; } @@ -1235,6 +1951,7 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { } else if (SwitchInst *SI = dyn_cast(TI)) { for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) if (SI->getSuccessor(i) == BB) { + BB->removePredecessor(SI->getParent()); SI->removeCase(i); --i; --e; Changed = true; @@ -1262,6 +1979,12 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { SI->setSuccessor(0, MaxBlock); Changed = true; + // If MaxBlock has phinodes in it, remove MaxPop-1 entries from + // it. + if (isa(MaxBlock->begin())) + for (unsigned i = 0; i != MaxPop-1; ++i) + MaxBlock->removePredecessor(SI->getParent()); + for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) if (SI->getSuccessor(i) == MaxBlock) { SI->removeCase(i); @@ -1273,13 +1996,16 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { if (II->getUnwindDest() == BB) { // Convert the invoke to a call instruction. This would be a good // place to note that the call does not throw though. - BranchInst *BI = new BranchInst(II->getNormalDest(), II); + BranchInst *BI = BranchInst::Create(II->getNormalDest(), II); II->removeFromParent(); // Take out of symbol table - + // Insert the call now... - std::vector Args(II->op_begin()+3, II->op_end()); - CallInst *CI = new CallInst(II->getCalledValue(), Args, - II->getName(), BI); + SmallVector Args(II->op_begin()+3, II->op_end()); + CallInst *CI = CallInst::Create(II->getCalledValue(), + Args.begin(), Args.end(), + II->getName(), BI); + CI->setCallingConv(II->getCallingConv()); + CI->setParamAttrs(II->getParamAttrs()); // If the invoke produced a value, the Call does now instead. II->replaceAllUsesWith(CI); delete II; @@ -1323,8 +2049,7 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { } if (OnlySucc) { - DEBUG(std::cerr << "Merging: " << *BB << "into: " << *OnlyPred); - TerminatorInst *Term = OnlyPred->getTerminator(); + DOUT << "Merging: " << *BB << "into: " << *OnlyPred; // Resolve any PHI nodes at the start of the block. They are all // guaranteed to have exactly one entry if they exist, unless there are @@ -1334,28 +2059,26 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { // while (PHINode *PN = dyn_cast(&BB->front())) { PN->replaceAllUsesWith(PN->getIncomingValue(0)); - BB->getInstList().pop_front(); // Delete the phi node... + BB->getInstList().pop_front(); // Delete the phi node. } - // Delete the unconditional branch from the predecessor... + // Delete the unconditional branch from the predecessor. OnlyPred->getInstList().pop_back(); - - // Move all definitions in the successor to the predecessor... + + // Move all definitions in the successor to the predecessor. OnlyPred->getInstList().splice(OnlyPred->end(), BB->getInstList()); - + // Make all PHI nodes that referred to BB now refer to Pred as their - // source... + // source. BB->replaceAllUsesWith(OnlyPred); - std::string OldName = BB->getName(); - - // Erase basic block from the function... + // Inherit predecessors name if it exists. + if (!OnlyPred->hasName()) + OnlyPred->takeName(BB); + + // Erase basic block from the function. M->getBasicBlockList().erase(BB); - // Inherit predecessors name if it exists... - if (!OldName.empty() && !OnlyPred->hasName()) - OnlyPred->setName(OldName); - return true; } @@ -1375,6 +2098,24 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { // so see if there is any identical code in the "then" and "else" // blocks. If so, we can hoist it up to the branching block. Changed |= HoistThenElseCodeToIf(BI); + } else { + OnlySucc = NULL; + for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); + SI != SE; ++SI) { + if (!OnlySucc) + OnlySucc = *SI; + else if (*SI != OnlySucc) { + OnlySucc = 0; // There are multiple distinct successors! + break; + } + } + + if (OnlySucc == OtherBB) { + // If BB's only successor is the other successor of the predecessor, + // i.e. a triangle, see if we can hoist any code from this block up + // to the "if" block. + Changed |= SpeculativelyExecuteBB(BI, BB); + } } } @@ -1393,19 +2134,20 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { // instruction can't handle, remove them now. std::sort(Values.begin(), Values.end(), ConstantIntOrdering()); Values.erase(std::unique(Values.begin(), Values.end()), Values.end()); - + // Figure out which block is which destination. BasicBlock *DefaultBB = BI->getSuccessor(1); BasicBlock *EdgeBB = BI->getSuccessor(0); if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB); - + // Create the new switch instruction now. - SwitchInst *New = new SwitchInst(CompVal, DefaultBB,Values.size(),BI); - + SwitchInst *New = SwitchInst::Create(CompVal, DefaultBB, + Values.size(), BI); + // Add all of the 'cases' to the switch instruction. for (unsigned i = 0, e = Values.size(); i != e; ++i) New->addCase(Values[i], EdgeBB); - + // We added edges from PI to the EdgeBB. As such, if there were any // PHI nodes in EdgeBB, they need entries to be added corresponding to // the number of edges added. @@ -1427,118 +2169,5 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { } } - // If there is a trivial two-entry PHI node in this basic block, and we can - // eliminate it, do so now. - if (PHINode *PN = dyn_cast(BB->begin())) - if (PN->getNumIncomingValues() == 2) { - // Ok, this is a two entry PHI node. Check to see if this is a simple "if - // statement", which has a very simple dominance structure. Basically, we - // are trying to find the condition that is being branched on, which - // subsequently causes this merge to happen. We really want control - // dependence information for this check, but simplifycfg can't keep it up - // to date, and this catches most of the cases we care about anyway. - // - BasicBlock *IfTrue, *IfFalse; - if (Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse)) { - DEBUG(std::cerr << "FOUND IF CONDITION! " << *IfCond << " T: " - << IfTrue->getName() << " F: " << IfFalse->getName() << "\n"); - - // Loop over the PHI's seeing if we can promote them all to select - // instructions. While we are at it, keep track of the instructions - // that need to be moved to the dominating block. - std::set AggressiveInsts; - bool CanPromote = true; - - BasicBlock::iterator AfterPHIIt = BB->begin(); - while (isa(AfterPHIIt)) { - PHINode *PN = cast(AfterPHIIt++); - if (PN->getIncomingValue(0) == PN->getIncomingValue(1)) - PN->replaceAllUsesWith(PN->getIncomingValue(0)); - else if (!DominatesMergePoint(PN->getIncomingValue(0), BB, - &AggressiveInsts) || - !DominatesMergePoint(PN->getIncomingValue(1), BB, - &AggressiveInsts)) { - CanPromote = false; - break; - } - } - - // Did we eliminate all PHI's? - CanPromote |= AfterPHIIt == BB->begin(); - - // If we all PHI nodes are promotable, check to make sure that all - // instructions in the predecessor blocks can be promoted as well. If - // not, we won't be able to get rid of the control flow, so it's not - // worth promoting to select instructions. - BasicBlock *DomBlock = 0, *IfBlock1 = 0, *IfBlock2 = 0; - if (CanPromote) { - PN = cast(BB->begin()); - BasicBlock *Pred = PN->getIncomingBlock(0); - if (cast(Pred->getTerminator())->isUnconditional()) { - IfBlock1 = Pred; - DomBlock = *pred_begin(Pred); - for (BasicBlock::iterator I = Pred->begin(); - !isa(I); ++I) - if (!AggressiveInsts.count(I)) { - // This is not an aggressive instruction that we can promote. - // Because of this, we won't be able to get rid of the control - // flow, so the xform is not worth it. - CanPromote = false; - break; - } - } - - Pred = PN->getIncomingBlock(1); - if (CanPromote && - cast(Pred->getTerminator())->isUnconditional()) { - IfBlock2 = Pred; - DomBlock = *pred_begin(Pred); - for (BasicBlock::iterator I = Pred->begin(); - !isa(I); ++I) - if (!AggressiveInsts.count(I)) { - // This is not an aggressive instruction that we can promote. - // Because of this, we won't be able to get rid of the control - // flow, so the xform is not worth it. - CanPromote = false; - break; - } - } - } - - // If we can still promote the PHI nodes after this gauntlet of tests, - // do all of the PHI's now. - if (CanPromote) { - // Move all 'aggressive' instructions, which are defined in the - // conditional parts of the if's up to the dominating block. - if (IfBlock1) { - DomBlock->getInstList().splice(DomBlock->getTerminator(), - IfBlock1->getInstList(), - IfBlock1->begin(), - IfBlock1->getTerminator()); - } - if (IfBlock2) { - DomBlock->getInstList().splice(DomBlock->getTerminator(), - IfBlock2->getInstList(), - IfBlock2->begin(), - IfBlock2->getTerminator()); - } - - while (PHINode *PN = dyn_cast(BB->begin())) { - // Change the PHI node into a select instruction. - Value *TrueVal = - PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse); - Value *FalseVal = - PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue); - - std::string Name = PN->getName(); PN->setName(""); - PN->replaceAllUsesWith(new SelectInst(IfCond, TrueVal, FalseVal, - Name, AfterPHIIt)); - BB->getInstList().erase(PN); - } - Changed = true; - } - } - } - return Changed; }