X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTransforms%2FUtils%2FSimplifyCFG.cpp;h=08c8a3f65e5183acea80021958d5cd67d293c43a;hb=9a7c743fc0d45e4f2331c4092d3f29bf18351e6f;hp=91664bf2d9d9d93f71c7a2ef3e3b9f682cdcdbc3;hpb=e1c99d4c69d20731579ae462db39c3c1393f50fd;p=oota-llvm.git diff --git a/lib/Transforms/Utils/SimplifyCFG.cpp b/lib/Transforms/Utils/SimplifyCFG.cpp index 91664bf2d9d..08c8a3f65e5 100644 --- a/lib/Transforms/Utils/SimplifyCFG.cpp +++ b/lib/Transforms/Utils/SimplifyCFG.cpp @@ -2,8 +2,8 @@ // // 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. // //===----------------------------------------------------------------------===// // @@ -15,6 +15,7 @@ #include "llvm/Transforms/Utils/Local.h" #include "llvm/Constants.h" #include "llvm/Instructions.h" +#include "llvm/IntrinsicInst.h" #include "llvm/Type.h" #include "llvm/DerivedTypes.h" #include "llvm/Support/CFG.h" @@ -22,12 +23,16 @@ #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; +STATISTIC(NumSpeculations, "Number of speculative executed instructions"); + /// SafeToMergeTerminators - Return true if it is safe to merge these two /// terminator instructions together. /// @@ -39,7 +44,7 @@ static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) { // 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)); + 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)) @@ -64,78 +69,107 @@ static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred, 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); - } + PHINode *PN; + for (BasicBlock::iterator I = Succ->begin(); + (PN = dyn_cast(I)); ++I) + PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), 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. -// +/// 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 CanPropagatePredecessorsForPHIs(BasicBlock *BB, BasicBlock *Succ) { assert(*succ_begin(BB) == Succ && "Succ is not successor of BB!"); - // 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! - // - if (isa(Succ->front())) { - std::set BBPreds(pred_begin(BB), pred_end(BB)); - 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). - if (PN->getIncomingValueForBlock(BB) != - PN->getIncomingValueForBlock(*PI)) - return false; // Values are not equal... + 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; + + typedef SmallPtrSet InstrSet; + InstrSet BBPHIs; + + // 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)); + + // 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.empty()) + 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); + + // 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 { + 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; } } - } - - // Finally, if BB has PHI nodes that are used by things other than the PHIs in - // Succ and Succ has predecessors that are not Succ and not Pred, we cannot - // fold these blocks, as we don't know whether BB dominates Succ or not to - // update the PHI nodes correctly. - if (!isa(BB->begin()) || Succ->getSinglePredecessor()) return true; - - // If the predecessors of Succ are only BB and Succ itself, we can handle this. - bool IsSafe = true; - for (pred_iterator PI = pred_begin(Succ), E = pred_end(Succ); PI != E; ++PI) - if (*PI != Succ && *PI != BB) { - IsSafe = false; - break; } - if (IsSafe) return true; - - // If the PHI nodes in BB are only used by instructions in Succ, we are ok if - // BB and Succ have no common predecessors. - for (BasicBlock::iterator I = BB->begin(); isa(I); ++I) { - PHINode *PN = cast(I); - for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end(); UI != E; - ++UI) - if (cast(*UI)->getParent() != Succ) + } + + // 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; + } } - - // Scan the predecessor sets of BB and Succ, making sure there are no common - // predecessors. Common predecessors would cause us to build a phi node with - // differing incoming values, which is not legal. - std::set BBPreds(pred_begin(BB), pred_end(BB)); - for (pred_iterator PI = pred_begin(Succ), E = pred_end(Succ); PI != E; ++PI) - if (BBPreds.count(*PI)) - return false; - + return true; } @@ -144,11 +178,8 @@ static bool CanPropagatePredecessorsForPHIs(BasicBlock *BB, BasicBlock *Succ) { /// branch. If possible, eliminate BB. static bool TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB, BasicBlock *Succ) { - // 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! - // + // 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; @@ -157,7 +188,7 @@ static bool TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB, // 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 std::vector BBPreds(pred_begin(BB), pred_end(BB)); + 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) { @@ -170,47 +201,54 @@ static bool TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB, 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 { - 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); - } + // 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())) { - std::vector + 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())) + 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 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. - 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, - // 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]); + continue; } + + // 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. @@ -310,15 +348,15 @@ static Value *GetIfCondition(BasicBlock *BB, } -// If we have a merge point of an "if condition" as accepted above, return true -// if the specified value dominates the block. We don't handle the true -// generality of domination here, just a special case which works well enough -// for us. -// -// If AggressiveInsts is non-null, and if V does not dominate BB, we check to -// see if V (which must be an instruction) is cheap to compute and is -// non-trapping. If both are true, the instruction is inserted into the set and -// true is returned. +/// DominatesMergePoint - If we have a merge point of an "if condition" as +/// accepted above, return true if the specified value dominates the block. We +/// don't handle the true generality of domination here, just a special case +/// which works well enough for us. +/// +/// If AggressiveInsts is non-null, and if V does not dominate BB, we check to +/// see if V (which must be an instruction) is cheap to compute and is +/// non-trapping. If both are true, the instruction is inserted into the set +/// and true is returned. static bool DominatesMergePoint(Value *V, BasicBlock *BB, std::set *AggressiveInsts) { Instruction *I = dyn_cast(V); @@ -351,6 +389,7 @@ static bool DominatesMergePoint(Value *V, BasicBlock *BB, // We can hoist loads that are non-volatile and obviously cannot trap. if (cast(I)->isVolatile()) return false; + // FIXME: A computation of a constant can trap! if (!isa(I->getOperand(0)) && !isa(I->getOperand(0))) return false; @@ -371,13 +410,15 @@ static bool DominatesMergePoint(Value *V, BasicBlock *BB, 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); @@ -386,11 +427,11 @@ static bool DominatesMergePoint(Value *V, BasicBlock *BB, return true; } -// 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. +/// 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 (Instruction *Inst = dyn_cast(V)) { if (Inst->getOpcode() == Instruction::ICmp && cast(Inst)->getPredicate() == ICmpInst::ICMP_EQ) { if (ConstantInt *C = dyn_cast(Inst->getOperand(1))) { @@ -406,14 +447,15 @@ static Value *GatherConstantSetEQs(Value *V, std::vector &Values){ if (LHS == RHS) return LHS; } + } return 0; } -// GatherConstantSetNEs - Given a potentially 'and'd together collection of -// setne instructions that compare a value against a constant, return the value -// being compared, and stick the constant into the Values vector. +/// GatherConstantSetNEs - Given a potentially 'and'd together collection of +/// 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 (Instruction *Inst = dyn_cast(V)) { if (Inst->getOpcode() == Instruction::ICmp && cast(Inst)->getPredicate() == ICmpInst::ICMP_NE) { if (ConstantInt *C = dyn_cast(Inst->getOperand(1))) { @@ -429,11 +471,10 @@ static Value *GatherConstantSetNEs(Value *V, std::vector &Values){ if (LHS == RHS) return LHS; } + } return 0; } - - /// GatherValueComparisons - If the specified Cond is an 'and' or 'or' of a /// bunch of comparisons of one value against constants, return the value and /// the constants being compared. @@ -455,40 +496,21 @@ static bool GatherValueComparisons(Instruction *Cond, Value *&CompVal, return false; } -/// ErasePossiblyDeadInstructionTree - If the specified instruction is dead and -/// 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)) return; - - std::vector InstrsToInspect; - InstrsToInspect.push_back(I); - - while (!InstrsToInspect.empty()) { - I = InstrsToInspect.back(); - InstrsToInspect.pop_back(); - - if (!isInstructionTriviallyDead(I)) continue; - - // 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; - } - - // Add operands of dead instruction to worklist. - for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) - if (Instruction *OpI = dyn_cast(I->getOperand(i))) - InstrsToInspect.push_back(OpI); - - // Remove dead instruction. - I->eraseFromParent(); +static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) { + Instruction* Cond = 0; + if (SwitchInst *SI = dyn_cast(TI)) { + Cond = dyn_cast(SI->getCondition()); + } else if (BranchInst *BI = dyn_cast(TI)) { + if (BI->isConditional()) + Cond = dyn_cast(BI->getCondition()); } + + TI->eraseFromParent(); + if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond); } -// isValueEqualityComparison - Return true if the specified terminator checks to -// see if a value is equal to constant integer value. +/// isValueEqualityComparison - Return true if the specified terminator checks +/// to see if a value is equal to constant integer value. static Value *isValueEqualityComparison(TerminatorInst *TI) { if (SwitchInst *SI = dyn_cast(TI)) { // Do not permit merging of large switch instructions into their @@ -509,8 +531,8 @@ static Value *isValueEqualityComparison(TerminatorInst *TI) { return 0; } -// Given a value comparison instruction, decode all of the 'cases' that it -// represents and return the 'default' block. +/// GetValueEqualityComparisonCases - Given a value comparison instruction, +/// decode all of the 'cases' that it represents and return the 'default' block. static BasicBlock * GetValueEqualityComparisonCases(TerminatorInst *TI, std::vector > &Cases) { for (unsigned i = 0, e = Cases.size(); i != e; ++i) @@ -542,8 +564,8 @@ static void EliminateBlockCases(BasicBlock *BB, } } -// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as -// well. +/// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as +/// well. static bool ValuesOverlap(std::vector > &C1, std::vector > &C2) { @@ -577,12 +599,12 @@ ValuesOverlap(std::vector > &C1, return false; } -// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a -// terminator instruction and its block is known to only have a single -// predecessor block, check to see if that predecessor is also a value -// comparison with the same value, and if that comparison determines the outcome -// of this comparison. If so, simplify TI. This does a very limited form of -// jump threading. +/// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a +/// terminator instruction and its block is known to only have a single +/// predecessor block, check to see if that predecessor is also a value +/// comparison with the same value, and if that comparison determines the +/// outcome of this comparison. If so, simplify TI. This does a very limited +/// form of jump threading. static bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, BasicBlock *Pred) { Value *PredVal = isValueEqualityComparison(Pred->getTerminator()); @@ -610,13 +632,12 @@ static bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, // PredCases. If there are any cases in ThisCases that are in PredCases, we // can simplify TI. if (ValuesOverlap(PredCases, ThisCases)) { - if (BranchInst *BTI = dyn_cast(TI)) { + if (isa(TI)) { // Okay, one of the successors of this condbr is dead. Convert it to a // uncond br. 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()); @@ -624,16 +645,13 @@ static bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 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. - if (Instruction *CondI = dyn_cast(Cond)) - ErasePossiblyDeadInstructionTree(CondI); + EraseTerminatorInstAndDCECond(TI); return true; } 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); @@ -657,11 +675,12 @@ 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 @@ -685,32 +704,28 @@ static bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, CheckEdge = 0; // Insert the new branch. - Instruction *NI = new BranchInst(TheRealDest, TI); + Instruction *NI = BranchInst::Create(TheRealDest, TI); 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()); - TI->eraseFromParent(); // Nuke the old one. - if (Cond) ErasePossiblyDeadInstructionTree(Cond); + EraseTerminatorInstAndDCECond(TI); return true; } return false; } -// FoldValueComparisonIntoPredecessors - The specified terminator is a value -// equality comparison instruction (either a switch or a branch on "X == c"). -// See if any of the predecessors of the terminator block are value comparisons -// on the same value. If so, and if safe to do so, fold them together. +/// FoldValueComparisonIntoPredecessors - The specified terminator is a value +/// equality comparison instruction (either a switch or a branch on "X == c"). +/// See if any of the predecessors of the terminator block are value comparisons +/// on the same value. If so, and if safe to do so, fold them together. static bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI) { BasicBlock *BB = TI->getParent(); Value *CV = isValueEqualityComparison(TI); // CondVal 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(); @@ -730,7 +745,7 @@ 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 @@ -798,18 +813,12 @@ 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); - Instruction *DeadCond = 0; - if (BranchInst *BI = dyn_cast(PTI)) - // If PTI is a branch, remember the condition. - DeadCond = dyn_cast(BI->getCondition()); - Pred->getInstList().erase(PTI); - - // If the condition is dead now, remove the instruction tree. - if (DeadCond) ErasePossiblyDeadInstructionTree(DeadCond); + EraseTerminatorInstAndDCECond(PTI); // Okay, last check. If BB is still a successor of PSI, then we must // have an infinite loop case. If so, add an infinitely looping block @@ -818,10 +827,10 @@ static bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI) { for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i) if (NewSI->getSuccessor(i) == BB) { if (InfLoopBlock == 0) { - // Insert it at the end of the loop, because it's either code, + // Insert it at the end of the function, 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); } @@ -844,7 +853,14 @@ static bool HoistThenElseCodeToIf(BranchInst *BI) { BasicBlock *BB1 = BI->getSuccessor(0); // The true destination. BasicBlock *BB2 = BI->getSuccessor(1); // The false destination - Instruction *I1 = BB1->begin(), *I2 = BB2->begin(); + BasicBlock::iterator BB1_Itr = BB1->begin(); + BasicBlock::iterator BB2_Itr = BB2->begin(); + + Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++; + while (isa(I1)) + I1 = BB1_Itr++; + while (isa(I2)) + I2 = BB2_Itr++; if (I1->getOpcode() != I2->getOpcode() || isa(I1) || isa(I1) || !I1->isIdenticalTo(I2)) return false; @@ -866,8 +882,12 @@ static bool HoistThenElseCodeToIf(BranchInst *BI) { I2->replaceAllUsesWith(I1); BB2->getInstList().erase(I2); - I1 = BB1->begin(); - I2 = BB2->begin(); + I1 = BB1_Itr++; + while (isa(I1)) + I1 = BB1_Itr++; + I2 = BB2_Itr++; + while (isa(I2)) + I2 = BB2_Itr++; } while (I1->getOpcode() == I2->getOpcode() && I1->isIdenticalTo(I2)); return true; @@ -898,8 +918,8 @@ 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) @@ -912,7 +932,159 @@ HoistTerminator: for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) AddPredecessorToBlock(*SI, BIParent, BB1); - BI->eraseFromParent(); + EraseTerminatorInstAndDCECond(BI); + 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: + // FP arithmetic might trap. Not worth doing for vector ops. + if (I->getType()->isFloatingPoint() || isa(I->getType())) + return false; + break; + case Instruction::And: + case Instruction::Or: + case Instruction::Xor: + case Instruction::Shl: + case Instruction::LShr: + case Instruction::AShr: + // Don't mess with vector operations. + if (isa(I->getType())) + return false; + break; // These are all cheap and non-trapping instructions. + } + + // If the instruction is obviously dead, don't try to predicate it. + if (I->use_empty()) { + I->eraseFromParent(); + return true; + } + + // 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; + + BasicBlock *BB2 = BB1->getTerminator()->getSuccessor(0); + for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); + UI != E; ++UI) { + // Ignore any user that is not a PHI node in BB2. These can only occur in + // unreachable blocks, because they would not be dominated by the instr. + PHINode *PN = dyn_cast(UI); + if (!PN || PN->getParent() != BB2) + return false; + PHIUses.push_back(PN); + + Value *PHIV = PN->getIncomingValueForBlock(BIParent); + if (!FalseV) + FalseV = PHIV; + else if (FalseV != PHIV) + return false; // Inconsistent value when condition is false. + } + + assert(FalseV && "Must have at least one user, and it must be a PHI"); + + // 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)) { + SmallPtrSet BB1Insns; + for(BasicBlock::iterator BB1I = BB1->begin(), BB1E = BB1->end(); + BB1I != BB1E; ++BB1I) + BB1Insns.insert(BB1I); + for(Value::use_iterator UI = BrCond->use_begin(), UE = BrCond->use_end(); + UI != UE; ++UI) { + Instruction *Use = cast(*UI); + if (BB1Insns.count(Use)) { + // If BrCond uses the instruction that place it just before + // branch instruction. + InsertPos = BI; + break; + } + } + } else + InsertPos = BI; + BIParent->getInstList().splice(InsertPos, 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; } @@ -955,11 +1127,7 @@ static bool FoldCondBranchOnPHI(BranchInst *BI) { // 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(); + FoldSingleEntryPHINodes(PN->getParent()); return true; } @@ -983,9 +1151,9 @@ static bool FoldCondBranchOnPHI(BranchInst *BI) { // 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 = new BasicBlock(RealDest->getName()+".critedge", - RealDest->getParent(), RealDest); - new BranchInst(RealDest, EdgeBB); + 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) { @@ -1007,11 +1175,12 @@ static bool FoldCondBranchOnPHI(BranchInst *BI) { if (BBI->hasName()) N->setName(BBI->getName()+".c"); // Update operands due to translation. - for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { + for (User::op_iterator i = N->op_begin(), e = N->op_end(); + i != e; ++i) { std::map::iterator PI = - TranslateMap.find(N->getOperand(i)); + TranslateMap.find(*i); if (PI != TranslateMap.end()) - N->setOperand(i, PI->second); + *i = PI->second; } // Check for trivial simplification. @@ -1105,7 +1274,7 @@ static bool FoldTwoEntryPHINode(PHINode *PN) { DomBlock = *pred_begin(Pred); for (BasicBlock::iterator I = Pred->begin(); !isa(I); ++I) - if (!AggressiveInsts.count(I)) { + if (!AggressiveInsts.count(I) && !isa(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. @@ -1119,7 +1288,7 @@ static bool FoldTwoEntryPHINode(PHINode *PN) { DomBlock = *pred_begin(Pred); for (BasicBlock::iterator I = Pred->begin(); !isa(I); ++I) - if (!AggressiveInsts.count(I)) { + if (!AggressiveInsts.count(I) && !isa(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. @@ -1152,7 +1321,7 @@ static bool FoldTwoEntryPHINode(PHINode *PN) { Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue); - Value *NV = new SelectInst(IfCond, TrueVal, FalseVal, "", AfterPHIIt); + Value *NV = SelectInst::Create(IfCond, TrueVal, FalseVal, "", AfterPHIIt); PN->replaceAllUsesWith(NV); NV->takeName(PN); @@ -1161,6 +1330,387 @@ static bool FoldTwoEntryPHINode(PHINode *PN) { return true; } +/// isTerminatorFirstRelevantInsn - Return true if Term is very first +/// instruction ignoring Phi nodes and dbg intrinsics. +static bool isTerminatorFirstRelevantInsn(BasicBlock *BB, Instruction *Term) { + BasicBlock::iterator BBI = Term; + while (BBI != BB->begin()) { + --BBI; + if (!isa(BBI)) + break; + } + + if (isa(BBI) || &*BBI == Term || isa(BBI)) + return true; + return false; +} + +/// 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. + if (!isTerminatorFirstRelevantInsn(TrueSucc, TrueRet)) + return false; + if (!isTerminatorFirstRelevantInsn(FalseSucc, FalseRet)) + return false; + + // 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); + EraseTerminatorInstAndDCECond(BI); + return true; + } + + // Otherwise, figure out what the true and false return values are + // so we can insert a new select instruction. + Value *TrueValue = TrueRet->getReturnValue(); + Value *FalseValue = FalseRet->getReturnValue(); + + // Unwrap any PHI nodes in the return blocks. + if (PHINode *TVPN = dyn_cast_or_null(TrueValue)) + if (TVPN->getParent() == TrueSucc) + TrueValue = TVPN->getIncomingValueForBlock(BI->getParent()); + if (PHINode *FVPN = dyn_cast_or_null(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_or_null(TrueValue)) + if (TCV->canTrap()) + return false; + if (ConstantExpr *FCV = dyn_cast_or_null(FalseValue)) + if (FCV->canTrap()) + return false; + + // 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(); + if (TrueValue) { + // Insert a select if the results differ. + if (TrueValue == FalseValue || isa(FalseValue)) { + } else if (isa(TrueValue)) { + TrueValue = FalseValue; + } else { + TrueValue = SelectInst::Create(BrCond, TrueValue, + FalseValue, "retval", BI); + } + } + + Value *RI = !TrueValue ? + ReturnInst::Create(BI) : + ReturnInst::Create(TrueValue, BI); + + DOUT << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:" + << "\n " << *BI << "NewRet = " << *RI + << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc; + + EraseTerminatorInstAndDCECond(BI); + + 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) { + BasicBlock *BB = BI->getParent(); + Instruction *Cond = dyn_cast(BI->getCondition()); + if (Cond == 0) return false; + + + // 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. + BasicBlock::iterator FrontIt = BB->front(); + // Ignore dbg intrinsics. + while(isa(FrontIt)) + ++FrontIt; + if ((!isa(Cond) && !isa(Cond)) || + Cond->getParent() != BB || &*FrontIt != Cond || !Cond->hasOneUse()) { + return false; + } + + // Make sure the instruction after the condition is the cond branch. + BasicBlock::iterator CondIt = Cond; ++CondIt; + // Ingore dbg intrinsics. + while(isa(CondIt)) + ++CondIt; + if (&*CondIt != BI) { + assert (!isa(CondIt) && "Hey do not forget debug info!"); + return false; + } + + // Cond is known to be a compare or binary operator. Check to make sure that + // neither operand is a potentially-trapping constant expression. + if (ConstantExpr *CE = dyn_cast(Cond->getOperand(0))) + if (CE->canTrap()) + return false; + if (ConstantExpr *CE = dyn_cast(Cond->getOperand(1))) + if (CE->canTrap()) + 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()); + + // Check that we have two conditional branches. If there is a PHI node in + // the common successor, verify that the same value flows in from both + // blocks. + 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; + + DOUT << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB; + + // 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; + + + if (ConstantExpr *CE = dyn_cast(BI->getCondition())) + if (CE->canTrap()) + 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); + + DOUT << "FOLDING BRs:" << *PBI->getParent() + << "AND: " << *BI->getParent(); + + + // If OtherDest *is* BB, then BB is a basic block with a single conditional + // branch in it, where one edge (OtherDest) goes back to itself but the other + // exits. We don't *know* that the program avoids the infinite loop + // (even though that seems likely). If we do this xform naively, we'll end up + // recursively unpeeling the loop. Since we know that (after the xform is + // done) that the block *is* infinite if reached, we just make it an obviously + // infinite loop with no cond branch. + if (OtherDest == BB) { + // Insert it at the end of the function, because it's either code, + // or it won't matter if it's hot. :) + BasicBlock *InfLoopBlock = BasicBlock::Create("infloop", BB->getParent()); + BranchInst::Create(InfLoopBlock, InfLoopBlock); + OtherDest = InfLoopBlock; + } + + 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 @@ -1172,46 +1722,27 @@ namespace { }; } -// 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 -// of the CFG. It returns true if a modification was made. -// -// WARNING: The entry node of a function may not be simplified. -// +/// 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 +/// of the CFG. It returns true if a modification was made. +/// +/// WARNING: The entry node of a function may not be simplified. +/// bool llvm::SimplifyCFG(BasicBlock *BB) { bool Changed = false; Function *M = BB->getParent(); 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)) { + // Remove basic blocks that have no predecessors... or that just have themself + // as a predecessor. These are unreachable. + if (pred_begin(BB) == pred_end(BB) || BB->getSinglePredecessor() == 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. - for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) - SI->removePredecessor(BB); - - while (!BB->empty()) { - Instruction &I = BB->back(); - // If this instruction is used, replace uses with an arbitrary - // 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 use undef instead - I.replaceAllUsesWith(UndefValue::get(I.getType())); - - // Remove the instruction from the basic block - BB->getInstList().pop_back(); - } - M->getBasicBlockList().erase(BB); + DeleteDeadBlock(BB); return true; } @@ -1219,6 +1750,12 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { // away... Changed |= ConstantFoldTerminator(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) + Changed |= FoldTwoEntryPHINode(PN); + // If this is a returning block with only PHI nodes in it, fold the return // instruction into any unconditional branch predecessors. // @@ -1226,18 +1763,18 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { // different return values, fold the replace the branch/return with a select // and return. if (ReturnInst *RI = dyn_cast(BB->getTerminator())) { - BasicBlock::iterator BBI = BB->getTerminator(); - if (BBI == BB->begin() || isa(--BBI)) { + if (isTerminatorFirstRelevantInsn(BB, BB->getTerminator())) { // 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! @@ -1252,12 +1789,21 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { Instruction *NewRet = RI->clone(); Pred->getInstList().push_back(NewRet); + BasicBlock::iterator BBI = RI; + if (BBI != BB->begin()) { + // Move region end info into the predecessor. + if (DbgRegionEndInst *DREI = dyn_cast(--BBI)) + DREI->moveBefore(NewRet); + } + // 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); @@ -1278,73 +1824,12 @@ 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()); - - // 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 ((!isa(TrueValue) || - !cast(TrueValue)->canTrap()) && - (!isa(TrueValue) || - !cast(TrueValue)->canTrap())) { - 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; - - DOUT << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:" - << "\n " << *BI << "Select = " << *NewRetVal - << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc; - - new ReturnInst(NewRetVal, BI); - BI->eraseFromParent(); - if (Instruction *BrCondI = dyn_cast(BrCond)) - if (isInstructionTriviallyDead(BrCondI)) - BrCondI->eraseFromParent(); - return true; - } - } - } + if (isa(BI->getSuccessor(0)->getTerminator()) && + isa(BI->getSuccessor(1)->getTerminator()) && + SimplifyCondBranchToTwoReturns(BI)) + return true; } } } else if (isa(BB->begin())) { @@ -1353,7 +1838,7 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { // 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())) { @@ -1366,14 +1851,16 @@ 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... SmallVector Args(II->op_begin()+3, II->op_end()); - CallInst *CI = new CallInst(II->getCalledValue(), - &Args[0], Args.size(), II->getName(), BI); + CallInst *CI = CallInst::Create(II->getCalledValue(), + Args.begin(), Args.end(), + II->getName(), BI); CI->setCallingConv(II->getCallingConv()); + CI->setAttributes(II->getAttributes()); // If the invoke produced a value, the Call now does instead II->replaceAllUsesWith(CI); delete II; @@ -1400,20 +1887,26 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { // If the block only contains the switch, see if we can fold the block // away into any preds. - if (SI == &BB->front()) + BasicBlock::iterator BBI = BB->begin(); + // Ignore dbg intrinsics. + while (isa(BBI)) + ++BBI; + if (SI == &*BBI) if (FoldValueComparisonIntoPredecessors(SI)) return SimplifyCFG(BB) || 1; } } else if (BranchInst *BI = dyn_cast(BB->getTerminator())) { if (BI->isUnconditional()) { - BasicBlock::iterator BBI = BB->begin(); // Skip over phi nodes... - while (isa(*BBI)) ++BBI; + BasicBlock::iterator BBI = BB->getFirstNonPHI(); BasicBlock *Succ = BI->getSuccessor(0); + // Ignore dbg intrinsics. + while (isa(BBI)) + ++BBI; if (BBI->isTerminator() && // Terminator is the only non-phi instruction! Succ != BB) // Don't hurt infinite loops! if (TryToSimplifyUncondBranchFromEmptyBlock(BB, Succ)) - return 1; + return true; } else { // Conditional branch if (isValueEqualityComparison(BI)) { @@ -1425,14 +1918,26 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { return SimplifyCFG(BB) || 1; // This block must be empty, except for the setcond inst, if it exists. + // Ignore dbg intrinsics. BasicBlock::iterator I = BB->begin(); - if (&*I == BI || - (&*I == cast(BI->getCondition()) && - &*++I == BI)) + // Ignore dbg intrinsics. + while (isa(I)) + ++I; + if (&*I == BI) { if (FoldValueComparisonIntoPredecessors(BI)) return SimplifyCFG(BB) | true; + } else if (&*I == cast(BI->getCondition())){ + ++I; + // Ignore dbg intrinsics. + while (isa(I)) + ++I; + if(&*I == BI) { + 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())) @@ -1443,215 +1948,16 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { // 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 (Instruction *Cond = dyn_cast(BI->getCondition())) - if ((isa(Cond) || isa(Cond)) && - 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 && FalseDest != BB) || - (PBI->getSuccessor(1) == FalseDest && TrueDest != BB)) { - // 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"); - 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; - } - } - // Scan predessor blocks for conditional branchs. + // 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 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 SimplifyCFG(BB); // Nuke the branch on constant. - } - - // Otherwise, if there are multiple predecessors, insert a PHI - // that merges in the constant and simplify the block result. - if (BlockIsSimpleEnoughToThreadThrough(BB)) { - PHINode *NewPN = new PHINode(Type::Int1Ty, - BI->getCondition()->getName()+".pr", - BB->begin()); - for (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); - // This will thread the branch. - return SimplifyCFG(BB) | 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) { - 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 { - PBIOp = BIOp = -1; - } - - // 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 (PBIOp != -1 && PBI->getSuccessor(PBIOp) == BB) - PBIOp = BIOp = -1; - - // 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. - if (PBIOp != -1) { - BasicBlock *CommonDest = PBI->getSuccessor(PBIOp); - - unsigned NumPhis = 0; - for (BasicBlock::iterator II = CommonDest->begin(); - isa(II); ++II, ++NumPhis) { - if (NumPhis > 2) { - // Disable this xform. - PBIOp = -1; - break; - } - } - } - - // Finally, if everything is ok, fold the branches to logical ops. - if (PBIOp != -1) { - BasicBlock *CommonDest = PBI->getSuccessor(PBIOp); - 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(); - - // 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 = new SelectInst(PBICond, PBIV, BIV, - PBIV->getName()+".mux", PBI); - PN->setIncomingValue(PBBIdx, NV); - } - } - - DOUT << "INTO: " << *PBI->getParent(); - - // This basic block is probably dead. We know it has at least - // one fewer predecessor. - return SimplifyCFG(BB) | true; - } - } - } + 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 @@ -1660,7 +1966,18 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { while (Unreachable != BB->begin()) { BasicBlock::iterator BBI = Unreachable; --BBI; + // Do not delete instructions that can have side effects, like calls + // (which may never return) and volatile loads and stores. if (isa(BBI)) break; + + if (StoreInst *SI = dyn_cast(BBI)) + if (SI->isVolatile()) + break; + + if (LoadInst *LI = dyn_cast(BBI)) + if (LI->isVolatile()) + break; + // Delete this instruction BB->getInstList().erase(BBI); Changed = true; @@ -1669,7 +1986,7 @@ 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(); @@ -1682,11 +1999,11 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { } } else { if (BI->getSuccessor(0) == BB) { - new BranchInst(BI->getSuccessor(1), BI); - BI->eraseFromParent(); + BranchInst::Create(BI->getSuccessor(1), BI); + EraseTerminatorInstAndDCECond(BI); } else if (BI->getSuccessor(1) == BB) { - new BranchInst(BI->getSuccessor(0), BI); - BI->eraseFromParent(); + BranchInst::Create(BI->getSuccessor(0), BI); + EraseTerminatorInstAndDCECond(BI); Changed = true; } } @@ -1738,15 +2055,16 @@ 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... SmallVector Args(II->op_begin()+3, II->op_end()); - CallInst *CI = new CallInst(II->getCalledValue(), - &Args[0], Args.size(), - II->getName(), BI); + CallInst *CI = CallInst::Create(II->getCalledValue(), + Args.begin(), Args.end(), + II->getName(), BI); CI->setCallingConv(II->getCallingConv()); + CI->setAttributes(II->getAttributes()); // If the invoke produced a value, the Call does now instead. II->replaceAllUsesWith(CI); delete II; @@ -1768,6 +2086,12 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { // pred, and if there is only one distinct successor of the predecessor, and // if there are no PHI nodes. // + if (MergeBlockIntoPredecessor(BB)) + return true; + + // Otherwise, if this block only has a single predecessor, and if that block + // is a conditional branch, see if we can hoist any code from this block up + // into our predecessor. 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 @@ -1775,57 +2099,7 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { OnlyPred = 0; // There are multiple different predecessors... break; } - - BasicBlock *OnlySucc = 0; - if (OnlyPred && OnlyPred != BB && // Don't break self loops - OnlyPred->getTerminator()->getOpcode() != Instruction::Invoke) { - // Check to see if there is only one distinct successor... - succ_iterator SI(succ_begin(OnlyPred)), SE(succ_end(OnlyPred)); - OnlySucc = BB; - for (; SI != SE; ++SI) - if (*SI != OnlySucc) { - OnlySucc = 0; // There are multiple distinct successors! - break; - } - } - - if (OnlySucc) { - 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 - // multiple duplicate (but guaranteed to be equal) entries for the - // incoming edges. This occurs when there are multiple edges from - // OnlyPred to OnlySucc. - // - while (PHINode *PN = dyn_cast(&BB->front())) { - PN->replaceAllUsesWith(PN->getIncomingValue(0)); - BB->getInstList().pop_front(); // Delete the phi node. - } - - // Delete the unconditional branch from the predecessor. - OnlyPred->getInstList().pop_back(); - - // 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. - BB->replaceAllUsesWith(OnlyPred); - - // Inherit predecessors name if it exists. - if (!OnlyPred->hasName()) - OnlyPred->takeName(BB); - - // Erase basic block from the function. - M->getBasicBlockList().erase(BB); - - return true; - } - - // Otherwise, if this block only has a single predecessor, and if that block - // is a conditional branch, see if we can hoist any code from this block up - // into our predecessor. + if (OnlyPred) if (BranchInst *BI = dyn_cast(OnlyPred->getTerminator())) if (BI->isConditional()) { @@ -1833,12 +2107,31 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { BasicBlock *OtherBB = BI->getSuccessor(BI->getSuccessor(0) == BB); PI = pred_begin(OtherBB); ++PI; + if (PI == pred_end(OtherBB)) { // We have a conditional branch to two blocks that are only reachable // from the condbr. We know that the condbr dominates the two blocks, // 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 { + BasicBlock* 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); + } } } @@ -1864,7 +2157,8 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { 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) @@ -1882,20 +2176,10 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { } // Erase the old branch instruction. - (*PI)->getInstList().erase(BI); - - // Erase the potentially condition tree that was used to computed the - // branch condition. - ErasePossiblyDeadInstructionTree(Cond); + EraseTerminatorInstAndDCECond(BI); 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); - return Changed; }