X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTransforms%2FUtils%2FSimplifyCFG.cpp;h=bbeac4c35028e004eea9cc7000276490c6f8bfe0;hb=b824512b8d733ccdfb1b1a2e8a6950605acf1c52;hp=99eef52d1faa2afd36cb1023a170fbbab5370131;hpb=4d5f508318b2f666ed50a2ed5e49e4aa343a92d6;p=oota-llvm.git diff --git a/lib/Transforms/Utils/SimplifyCFG.cpp b/lib/Transforms/Utils/SimplifyCFG.cpp index 99eef52d1fa..bbeac4c3502 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. // //===----------------------------------------------------------------------===// // @@ -16,15 +16,22 @@ #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; +STATISTIC(NumSpeculations, "Number of speculative executed instructions"); + /// SafeToMergeTerminators - Return true if it is safe to merge these two /// terminator instructions together. /// @@ -36,7 +43,7 @@ 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)) @@ -76,63 +83,93 @@ static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred, 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.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); + + // 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; } @@ -141,11 +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; @@ -154,7 +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) { @@ -167,20 +201,23 @@ 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 @@ -191,19 +228,21 @@ static bool TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB, // 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. + // 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, - // 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. + // 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]); @@ -211,12 +250,9 @@ static bool TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB, } // Everything that jumped to BB now goes to Succ. - std::string OldName = BB->getName(); BB->replaceAllUsesWith(Succ); + if (!Succ->hasName()) Succ->takeName(BB); BB->eraseFromParent(); // Delete the old basic block. - - if (!OldName.empty() && !Succ->hasName()) // Transfer name if we can - Succ->setName(OldName); return true; } @@ -371,13 +407,15 @@ 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); @@ -390,7 +428,7 @@ static bool DominatesMergePoint(Value *V, BasicBlock *BB, // 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))) { @@ -406,6 +444,7 @@ static Value *GatherConstantSetEQs(Value *V, std::vector &Values){ if (LHS == RHS) return LHS; } + } return 0; } @@ -413,7 +452,7 @@ 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 (Instruction *Inst = dyn_cast(V)) { if (Inst->getOpcode() == Instruction::ICmp && cast(Inst)->getPredicate() == ICmpInst::ICMP_NE) { if (ConstantInt *C = dyn_cast(Inst->getOperand(1))) { @@ -429,6 +468,7 @@ static Value *GatherConstantSetNEs(Value *V, std::vector &Values){ if (LHS == RHS) return LHS; } + } return 0; } @@ -461,7 +501,7 @@ static bool GatherValueComparisons(Instruction *Cond, Value *&CompVal, static void ErasePossiblyDeadInstructionTree(Instruction *I) { if (!isInstructionTriviallyDead(I)) return; - std::vector InstrsToInspect; + SmallVector InstrsToInspect; InstrsToInspect.push_back(I); while (!InstrsToInspect.empty()) { @@ -478,8 +518,8 @@ static void ErasePossiblyDeadInstructionTree(Instruction *I) { } // 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))) + 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); // Remove dead instruction. @@ -531,7 +571,7 @@ GetValueEqualityComparisonCases(TerminatorInst *TI, } -// 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, std::vector > &Cases) { @@ -616,7 +656,7 @@ 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()); @@ -633,7 +673,7 @@ 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); @@ -657,11 +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 @@ -685,7 +726,7 @@ 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"; @@ -710,7 +751,7 @@ 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(); @@ -730,7 +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 @@ -798,7 +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); @@ -820,8 +862,8 @@ 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); } @@ -879,7 +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. @@ -898,8 +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) @@ -916,6 +958,129 @@ HoistTerminator: 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) { @@ -983,9 +1148,9 @@ 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 +1172,12 @@ 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. @@ -1152,21 +1318,363 @@ static bool FoldTwoEntryPHINode(PHINode *PN) { Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue); - std::string Name = PN->getName(); PN->setName(""); - PN->replaceAllUsesWith(new SelectInst(IfCond, TrueVal, FalseVal, - Name, AfterPHIIt)); + 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->getZExtValue() < RHS->getZExtValue(); + return LHS->getValue().ult(RHS->getValue()); } }; } @@ -1184,11 +1692,12 @@ 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)) { + 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 @@ -1218,6 +1727,12 @@ 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. // @@ -1228,15 +1743,16 @@ 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! @@ -1253,10 +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); @@ -1277,73 +1795,12 @@ 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())) { @@ -1352,7 +1809,7 @@ 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())) { @@ -1365,14 +1822,16 @@ 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... - 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; @@ -1405,14 +1864,13 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { } } 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); 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)) { @@ -1442,215 +1900,16 @@ 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(); - 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; - } - } - // 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 @@ -1668,7 +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(); @@ -1681,10 +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; } @@ -1737,14 +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; @@ -1798,28 +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; } @@ -1839,6 +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); + } } } @@ -1852,7 +2129,7 @@ bool llvm::SimplifyCFG(BasicBlock *BB) { Value *CompVal = 0; std::vector Values; bool TrueWhenEqual = GatherValueComparisons(Cond, CompVal, Values); - if (CompVal && CompVal->getType()->isIntegral()) { + if (CompVal && CompVal->getType()->isInteger()) { // There might be duplicate constants in the list, which the switch // instruction can't handle, remove them now. std::sort(Values.begin(), Values.end(), ConstantIntOrdering()); @@ -1864,7 +2141,8 @@ 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) @@ -1891,11 +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) - Changed |= FoldTwoEntryPHINode(PN); - return Changed; }