diff --git a/lib/Transforms/Scalar/CodeGenPrepare.cpp b/lib/Transforms/Scalar/CodeGenPrepare.cpp index ccd44a22fe8..32226a02aa4 100644 --- a/lib/Transforms/Scalar/CodeGenPrepare.cpp +++ b/lib/Transforms/Scalar/CodeGenPrepare.cpp @@ -25,9 +25,12 @@ #include "llvm/Target/TargetLowering.h" #include "llvm/Target/TargetMachine.h" #include "llvm/Transforms/Utils/BasicBlockUtils.h" +#include "llvm/Transforms/Utils/Local.h" +#include "llvm/ADT/DenseMap.h" #include "llvm/ADT/SmallSet.h" #include "llvm/Support/Debug.h" #include "llvm/Support/Compiler.h" +#include "llvm/Support/GetElementPtrTypeIterator.h" using namespace llvm; namespace { @@ -44,7 +47,9 @@ namespace { bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const; void EliminateMostlyEmptyBlock(BasicBlock *BB); bool OptimizeBlock(BasicBlock &BB); - bool OptimizeGEPExpression(GetElementPtrInst *GEPI); + bool OptimizeLoadStoreInst(Instruction *I, Value *Addr, + const Type *AccessTy, + DenseMap &SunkAddrs); }; } static RegisterPass X("codegenprepare", @@ -291,312 +296,37 @@ static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum, Pass *P) { SplitCriticalEdge(TI, SuccNum, P, true); } - -/// InsertGEPComputeCode - Insert code into BB to compute Ptr+PtrOffset, -/// casting to the type of GEPI. -static Instruction *InsertGEPComputeCode(Instruction *&V, BasicBlock *BB, - Instruction *GEPI, Value *Ptr, - Value *PtrOffset) { - if (V) return V; // Already computed. - - // Figure out the insertion point - BasicBlock::iterator InsertPt; - if (BB == GEPI->getParent()) { - // If GEP is already inserted into BB, insert right after the GEP. - InsertPt = GEPI; - ++InsertPt; - } else { - // Otherwise, insert at the top of BB, after any PHI nodes - InsertPt = BB->begin(); - while (isa(InsertPt)) ++InsertPt; - } - - // If Ptr is itself a cast, but in some other BB, emit a copy of the cast into - // BB so that there is only one value live across basic blocks (the cast - // operand). - if (CastInst *CI = dyn_cast(Ptr)) - if (CI->getParent() != BB && isa(CI->getOperand(0)->getType())) - Ptr = CastInst::create(CI->getOpcode(), CI->getOperand(0), CI->getType(), - "", InsertPt); - - // Add the offset, cast it to the right type. - Ptr = BinaryOperator::createAdd(Ptr, PtrOffset, "", InsertPt); - // Ptr is an integer type, GEPI is pointer type ==> IntToPtr - return V = CastInst::create(Instruction::IntToPtr, Ptr, GEPI->getType(), - "", InsertPt); -} - -/// ReplaceUsesOfGEPInst - Replace all uses of RepPtr with inserted code to -/// compute its value. The RepPtr value can be computed with Ptr+PtrOffset. One -/// trivial way of doing this would be to evaluate Ptr+PtrOffset in RepPtr's -/// block, then ReplaceAllUsesWith'ing everything. However, we would prefer to -/// sink PtrOffset into user blocks where doing so will likely allow us to fold -/// the constant add into a load or store instruction. Additionally, if a user -/// is a pointer-pointer cast, we look through it to find its users. -static void ReplaceUsesOfGEPInst(Instruction *RepPtr, Value *Ptr, - Constant *PtrOffset, BasicBlock *DefBB, - GetElementPtrInst *GEPI, - std::map &InsertedExprs) { - while (!RepPtr->use_empty()) { - Instruction *User = cast(RepPtr->use_back()); - - // If the user is a Pointer-Pointer cast, recurse. Only BitCast can be - // used for a Pointer-Pointer cast. - if (isa(User)) { - ReplaceUsesOfGEPInst(User, Ptr, PtrOffset, DefBB, GEPI, InsertedExprs); - - // Drop the use of RepPtr. The cast is dead. Don't delete it now, else we - // could invalidate an iterator. - User->setOperand(0, UndefValue::get(RepPtr->getType())); - continue; - } - - // If this is a load of the pointer, or a store through the pointer, emit - // the increment into the load/store block. - Instruction *NewVal; - if (isa(User) || - (isa(User) && User->getOperand(0) != RepPtr)) { - NewVal = InsertGEPComputeCode(InsertedExprs[User->getParent()], - User->getParent(), GEPI, - Ptr, PtrOffset); - } else { - // If this use is not foldable into the addressing mode, use a version - // emitted in the GEP block. - NewVal = InsertGEPComputeCode(InsertedExprs[DefBB], DefBB, GEPI, - Ptr, PtrOffset); - } - - if (GEPI->getType() != RepPtr->getType()) { - BasicBlock::iterator IP = NewVal; - ++IP; - // NewVal must be a GEP which must be pointer type, so BitCast - NewVal = new BitCastInst(NewVal, RepPtr->getType(), "", IP); - } - User->replaceUsesOfWith(RepPtr, NewVal); - } -} - -/// OptimizeGEPExpression - Since we are doing basic-block-at-a-time instruction -/// selection, we want to be a bit careful about some things. In particular, if -/// we have a GEP instruction that is used in a different block than it is -/// defined, the addressing expression of the GEP cannot be folded into loads or -/// stores that use it. In this case, decompose the GEP and move constant -/// indices into blocks that use it. -bool CodeGenPrepare::OptimizeGEPExpression(GetElementPtrInst *GEPI) { - // If this GEP is only used inside the block it is defined in, there is no - // need to rewrite it. - bool isUsedOutsideDefBB = false; - BasicBlock *DefBB = GEPI->getParent(); - for (Value::use_iterator UI = GEPI->use_begin(), E = GEPI->use_end(); - UI != E; ++UI) { - if (cast(*UI)->getParent() != DefBB) { - isUsedOutsideDefBB = true; - break; - } - } - if (!isUsedOutsideDefBB) return false; - - // If this GEP has no non-zero constant indices, there is nothing we can do, - // ignore it. - bool hasConstantIndex = false; - bool hasVariableIndex = false; - for (GetElementPtrInst::op_iterator OI = GEPI->op_begin()+1, - E = GEPI->op_end(); OI != E; ++OI) { - if (ConstantInt *CI = dyn_cast(*OI)) { - if (!CI->isZero()) { - hasConstantIndex = true; - break; - } - } else { - hasVariableIndex = true; - } - } - - // If this is a "GEP X, 0, 0, 0", turn this into a cast. - if (!hasConstantIndex && !hasVariableIndex) { - /// The GEP operand must be a pointer, so must its result -> BitCast - Value *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(), - GEPI->getName(), GEPI); - GEPI->replaceAllUsesWith(NC); - GEPI->eraseFromParent(); - return true; - } - - // If this is a GEP &Alloca, 0, 0, forward subst the frame index into uses. - if (!hasConstantIndex && !isa(GEPI->getOperand(0))) - return false; - - // If we don't have target lowering info, we can't lower the GEP. - if (!TLI) return false; - const TargetData *TD = TLI->getTargetData(); - - // Otherwise, decompose the GEP instruction into multiplies and adds. Sum the - // constant offset (which we now know is non-zero) and deal with it later. - uint64_t ConstantOffset = 0; - const Type *UIntPtrTy = TD->getIntPtrType(); - Value *Ptr = new PtrToIntInst(GEPI->getOperand(0), UIntPtrTy, "", GEPI); - const Type *Ty = GEPI->getOperand(0)->getType(); - - for (GetElementPtrInst::op_iterator OI = GEPI->op_begin()+1, - E = GEPI->op_end(); OI != E; ++OI) { - Value *Idx = *OI; - if (const StructType *StTy = dyn_cast(Ty)) { - unsigned Field = cast(Idx)->getZExtValue(); - if (Field) - ConstantOffset += TD->getStructLayout(StTy)->getElementOffset(Field); - Ty = StTy->getElementType(Field); - } else { - Ty = cast(Ty)->getElementType(); - - // Handle constant subscripts. - if (ConstantInt *CI = dyn_cast(Idx)) { - if (CI->getZExtValue() == 0) continue; - ConstantOffset += (int64_t)TD->getTypeSize(Ty)*CI->getSExtValue(); - continue; - } - - // Ptr = Ptr + Idx * ElementSize; - - // Cast Idx to UIntPtrTy if needed. - Idx = CastInst::createIntegerCast(Idx, UIntPtrTy, true/*SExt*/, "", GEPI); - - uint64_t ElementSize = TD->getTypeSize(Ty); - // Mask off bits that should not be set. - ElementSize &= ~0ULL >> (64-UIntPtrTy->getPrimitiveSizeInBits()); - Constant *SizeCst = ConstantInt::get(UIntPtrTy, ElementSize); - - // Multiply by the element size and add to the base. - Idx = BinaryOperator::createMul(Idx, SizeCst, "", GEPI); - Ptr = BinaryOperator::createAdd(Ptr, Idx, "", GEPI); - } - } - - // Make sure that the offset fits in uintptr_t. - ConstantOffset &= ~0ULL >> (64-UIntPtrTy->getPrimitiveSizeInBits()); - Constant *PtrOffset = ConstantInt::get(UIntPtrTy, ConstantOffset); - - // Okay, we have now emitted all of the variable index parts to the BB that - // the GEP is defined in. Loop over all of the using instructions, inserting - // an "add Ptr, ConstantOffset" into each block that uses it and update the - // instruction to use the newly computed value, making GEPI dead. When the - // user is a load or store instruction address, we emit the add into the user - // block, otherwise we use a canonical version right next to the gep (these - // won't be foldable as addresses, so we might as well share the computation). - - std::map InsertedExprs; - ReplaceUsesOfGEPInst(GEPI, Ptr, PtrOffset, DefBB, GEPI, InsertedExprs); - - // Finally, the GEP is dead, remove it. - GEPI->eraseFromParent(); - - return true; -} - -/// SinkInvariantGEPIndex - If a GEP instruction has a variable index that has -/// been hoisted out of the loop by LICM pass, sink it back into the use BB -/// if it can be determined that the index computation can be folded into the -/// addressing mode of the load / store uses. -static bool SinkInvariantGEPIndex(BinaryOperator *BinOp, - const TargetLowering &TLI) { - // Only look at Add. - if (BinOp->getOpcode() != Instruction::Add) - return false; - - // DestBBs - These are the blocks where a copy of BinOp will be inserted. - SmallSet DestBBs; - BasicBlock *DefBB = BinOp->getParent(); - bool MadeChange = false; - for (Value::use_iterator UI = BinOp->use_begin(), E = BinOp->use_end(); - UI != E; ++UI) { - Instruction *GEPI = cast(*UI); - // Only look for GEP use in another block. - if (GEPI->getParent() == DefBB) continue; - - if (isa(GEPI)) { - // If the GEP has another variable index, abondon. - bool hasVariableIndex = false; - for (GetElementPtrInst::op_iterator OI = GEPI->op_begin()+1, - OE = GEPI->op_end(); OI != OE; ++OI) - if (*OI != BinOp && !isa(*OI)) { - hasVariableIndex = true; - break; - } - if (hasVariableIndex) - break; - - BasicBlock *GEPIBB = GEPI->getParent(); - for (Value::use_iterator UUI = GEPI->use_begin(), UE = GEPI->use_end(); - UUI != UE; ++UUI) { - Instruction *GEPIUser = cast(*UUI); - const Type *UseTy = NULL; - if (LoadInst *Load = dyn_cast(GEPIUser)) - UseTy = Load->getType(); - else if (StoreInst *Store = dyn_cast(GEPIUser)) - UseTy = Store->getOperand(0)->getType(); - - // Check if it is possible to fold the expression to address mode. - if (UseTy && isa(BinOp->getOperand(1))) { - int64_t Cst = cast(BinOp->getOperand(1))->getSExtValue(); - // e.g. load (gep i32 * %P, (X+42)) => load (%P + X*4 + 168). - TargetLowering::AddrMode AM; - // FIXME: This computation isn't right, scale is incorrect. - AM.Scale = TLI.getTargetData()->getTypeSize(UseTy); - // FIXME: Should should also include other fixed offsets. - AM.BaseOffs = Cst*AM.Scale; - - if (TLI.isLegalAddressingMode(AM, UseTy)) { - DestBBs.insert(GEPIBB); - MadeChange = true; - break; - } - } - } - } - } - - // Nothing to do. - if (!MadeChange) - return false; - - /// InsertedOps - Only insert a duplicate in each block once. - std::map InsertedOps; - for (Value::use_iterator UI = BinOp->use_begin(), E = BinOp->use_end(); - UI != E; ) { - Instruction *User = cast(*UI); - BasicBlock *UserBB = User->getParent(); - - // Preincrement use iterator so we don't invalidate it. - ++UI; - - // If any user in this BB wants it, replace all the uses in the BB. - if (DestBBs.count(UserBB)) { - // Sink it into user block. - BinaryOperator *&InsertedOp = InsertedOps[UserBB]; - if (!InsertedOp) { - BasicBlock::iterator InsertPt = UserBB->begin(); - while (isa(InsertPt)) ++InsertPt; - - InsertedOp = - BinaryOperator::create(BinOp->getOpcode(), BinOp->getOperand(0), - BinOp->getOperand(1), "", InsertPt); - } - - User->replaceUsesOfWith(BinOp, InsertedOp); - } - } - - if (BinOp->use_empty()) - BinOp->eraseFromParent(); - - return true; -} - -/// OptimizeNoopCopyExpression - We have determined that the specified cast -/// instruction is a noop copy (e.g. it's casting from one pointer type to -/// another, int->uint, or int->sbyte on PPC. +/// OptimizeNoopCopyExpression - If the specified cast instruction is a noop +/// copy (e.g. it's casting from one pointer type to another, int->uint, or +/// int->sbyte on PPC), sink it into user blocks to reduce the number of virtual +/// registers that must be created and coallesced. /// /// Return true if any changes are made. -static bool OptimizeNoopCopyExpression(CastInst *CI) { +static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){ + // If this is a noop copy, + MVT::ValueType SrcVT = TLI.getValueType(CI->getOperand(0)->getType()); + MVT::ValueType DstVT = TLI.getValueType(CI->getType()); + + // This is an fp<->int conversion? + if (MVT::isInteger(SrcVT) != MVT::isInteger(DstVT)) + return false; + + // If this is an extension, it will be a zero or sign extension, which + // isn't a noop. + if (SrcVT < DstVT) return false; + + // If these values will be promoted, find out what they will be promoted + // to. This helps us consider truncates on PPC as noop copies when they + // are. + if (TLI.getTypeAction(SrcVT) == TargetLowering::Promote) + SrcVT = TLI.getTypeToTransformTo(SrcVT); + if (TLI.getTypeAction(DstVT) == TargetLowering::Promote) + DstVT = TLI.getTypeToTransformTo(DstVT); + + // If, after promotion, these are the same types, this is a noop copy. + if (SrcVT != DstVT) + return false; + BasicBlock *DefBB = CI->getParent(); /// InsertedCasts - Only insert a cast in each block once. @@ -646,8 +376,469 @@ static bool OptimizeNoopCopyExpression(CastInst *CI) { return MadeChange; } +/// EraseDeadInstructions - Erase any dead instructions +static void EraseDeadInstructions(Value *V) { + Instruction *I = dyn_cast(V); + if (!I || !I->use_empty()) return; + + SmallPtrSet Insts; + Insts.insert(I); + + while (!Insts.empty()) { + I = *Insts.begin(); + Insts.erase(I); + if (isInstructionTriviallyDead(I)) { + for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) + if (Instruction *U = dyn_cast(I->getOperand(i))) + Insts.insert(U); + I->eraseFromParent(); + } + } +} +/// ExtAddrMode - This is an extended version of TargetLowering::AddrMode which +/// holds actual Value*'s for register values. +struct ExtAddrMode : public TargetLowering::AddrMode { + Value *BaseReg; + Value *ScaledReg; + ExtAddrMode() : BaseReg(0), ScaledReg(0) {} + void dump() const; +}; + +static std::ostream &operator<<(std::ostream &OS, const ExtAddrMode &AM) { + bool NeedPlus = false; + OS << "["; + if (AM.BaseGV) + OS << (NeedPlus ? " + " : "") + << "GV:%" << AM.BaseGV->getName(), NeedPlus = true; + + if (AM.BaseOffs) + OS << (NeedPlus ? " + " : "") << AM.BaseOffs, NeedPlus = true; + + if (AM.BaseReg) + OS << (NeedPlus ? " + " : "") + << "Base:%" << AM.BaseReg->getName(), NeedPlus = true; + if (AM.Scale) + OS << (NeedPlus ? " + " : "") + << AM.Scale << "*%" << AM.ScaledReg->getName(), NeedPlus = true; + + return OS << "]"; +} + +void ExtAddrMode::dump() const { + cerr << *this << "\n"; +} + +static bool TryMatchingScaledValue(Value *ScaleReg, int64_t Scale, + const Type *AccessTy, ExtAddrMode &AddrMode, + SmallVector &AddrModeInsts, + const TargetLowering &TLI, unsigned Depth); + +/// FindMaximalLegalAddressingMode - If we can, try to merge the computation of +/// Addr into the specified addressing mode. If Addr can't be added to AddrMode +/// this returns false. This assumes that Addr is either a pointer type or +/// intptr_t for the target. +static bool FindMaximalLegalAddressingMode(Value *Addr, const Type *AccessTy, + ExtAddrMode &AddrMode, + SmallVector &AddrModeInsts, + const TargetLowering &TLI, + unsigned Depth) { + + // If this is a global variable, fold it into the addressing mode if possible. + if (GlobalValue *GV = dyn_cast(Addr)) { + if (AddrMode.BaseGV == 0) { + AddrMode.BaseGV = GV; + if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) + return true; + AddrMode.BaseGV = 0; + } + } else if (ConstantInt *CI = dyn_cast(Addr)) { + AddrMode.BaseOffs += CI->getSExtValue(); + if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) + return true; + AddrMode.BaseOffs -= CI->getSExtValue(); + } else if (isa(Addr)) { + return true; + } + + // Look through constant exprs and instructions. + unsigned Opcode = ~0U; + User *AddrInst = 0; + if (Instruction *I = dyn_cast(Addr)) { + Opcode = I->getOpcode(); + AddrInst = I; + } else if (ConstantExpr *CE = dyn_cast(Addr)) { + Opcode = CE->getOpcode(); + AddrInst = CE; + } + + // Limit recursion to avoid exponential behavior. + if (Depth == 5) { AddrInst = 0; Opcode = ~0U; } + + // If this is really an instruction, add it to our list of related + // instructions. + if (Instruction *I = dyn_cast_or_null(AddrInst)) + AddrModeInsts.push_back(I); + + switch (Opcode) { + case Instruction::PtrToInt: + // PtrToInt is always a noop, as we know that the int type is pointer sized. + if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, + AddrMode, AddrModeInsts, TLI, Depth)) + return true; + break; + case Instruction::IntToPtr: + // This inttoptr is a no-op if the integer type is pointer sized. + if (TLI.getValueType(AddrInst->getOperand(0)->getType()) == + TLI.getPointerTy()) { + if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, + AddrMode, AddrModeInsts, TLI, Depth)) + return true; + } + break; + case Instruction::Add: { + // Check to see if we can merge in the RHS then the LHS. If so, we win. + ExtAddrMode BackupAddrMode = AddrMode; + unsigned OldSize = AddrModeInsts.size(); + if (FindMaximalLegalAddressingMode(AddrInst->getOperand(1), AccessTy, + AddrMode, AddrModeInsts, TLI, Depth+1) && + FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, + AddrMode, AddrModeInsts, TLI, Depth+1)) + return true; + + // Restore the old addr mode info. + AddrMode = BackupAddrMode; + AddrModeInsts.resize(OldSize); + + // Otherwise this was over-aggressive. Try merging in the LHS then the RHS. + if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, + AddrMode, AddrModeInsts, TLI, Depth+1) && + FindMaximalLegalAddressingMode(AddrInst->getOperand(1), AccessTy, + AddrMode, AddrModeInsts, TLI, Depth+1)) + return true; + + // Otherwise we definitely can't merge the ADD in. + AddrMode = BackupAddrMode; + AddrModeInsts.resize(OldSize); + break; + } + case Instruction::Or: { + ConstantInt *RHS = dyn_cast(AddrInst->getOperand(1)); + if (!RHS) break; + // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD. + break; + } + case Instruction::Mul: + case Instruction::Shl: { + // Can only handle X*C and X << C, and can only handle this when the scale + // field is available. + ConstantInt *RHS = dyn_cast(AddrInst->getOperand(1)); + if (!RHS) break; + int64_t Scale = RHS->getSExtValue(); + if (Opcode == Instruction::Shl) + Scale = 1 << Scale; + + if (TryMatchingScaledValue(AddrInst->getOperand(0), Scale, AccessTy, + AddrMode, AddrModeInsts, TLI, Depth)) + return true; + break; + } + case Instruction::GetElementPtr: { + // Scan the GEP. We check it if it contains constant offsets and at most + // one variable offset. + int VariableOperand = -1; + unsigned VariableScale = 0; + + int64_t ConstantOffset = 0; + const TargetData *TD = TLI.getTargetData(); + gep_type_iterator GTI = gep_type_begin(AddrInst); + for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) { + if (const StructType *STy = dyn_cast(*GTI)) { + const StructLayout *SL = TD->getStructLayout(STy); + unsigned Idx = + cast(AddrInst->getOperand(i))->getZExtValue(); + ConstantOffset += SL->getElementOffset(Idx); + } else { + uint64_t TypeSize = TD->getTypeSize(GTI.getIndexedType()); + if (ConstantInt *CI = dyn_cast(AddrInst->getOperand(i))) { + ConstantOffset += CI->getSExtValue()*TypeSize; + } else if (TypeSize) { // Scales of zero don't do anything. + // We only allow one variable index at the moment. + if (VariableOperand != -1) { + VariableOperand = -2; + break; + } + + // Remember the variable index. + VariableOperand = i; + VariableScale = TypeSize; + } + } + } + + // If the GEP had multiple variable indices, punt. + if (VariableOperand == -2) + break; + + // A common case is for the GEP to only do a constant offset. In this case, + // just add it to the disp field and check validity. + if (VariableOperand == -1) { + AddrMode.BaseOffs += ConstantOffset; + if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){ + // Check to see if we can fold the base pointer in too. + if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, + AddrMode, AddrModeInsts, TLI, + Depth+1)) + return true; + } + AddrMode.BaseOffs -= ConstantOffset; + } else { + // Check that this has no base reg yet. If so, we won't have a place to + // put the base of the GEP (assuming it is not a null ptr). + bool SetBaseReg = false; + if (AddrMode.HasBaseReg) { + if (!isa(AddrInst->getOperand(0))) + break; + } else { + AddrMode.HasBaseReg = true; + AddrMode.BaseReg = AddrInst->getOperand(0); + SetBaseReg = true; + } + + // See if the scale amount is valid for this target. + AddrMode.BaseOffs += ConstantOffset; + if (TryMatchingScaledValue(AddrInst->getOperand(VariableOperand), + VariableScale, AccessTy, AddrMode, + AddrModeInsts, TLI, Depth)) { + if (!SetBaseReg) return true; + + // If this match succeeded, we know that we can form an address with the + // GepBase as the basereg. See if we can match *more*. + AddrMode.HasBaseReg = false; + AddrMode.BaseReg = 0; + if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, + AddrMode, AddrModeInsts, TLI, + Depth+1)) + return true; + // Strange, shouldn't happen. Restore the base reg and succeed the easy + // way. + AddrMode.HasBaseReg = true; + AddrMode.BaseReg = AddrInst->getOperand(0); + return true; + } + + AddrMode.BaseOffs -= ConstantOffset; + if (SetBaseReg) { + AddrMode.HasBaseReg = false; + AddrMode.BaseReg = 0; + } + } + break; + } + } + + if (Instruction *I = dyn_cast_or_null(AddrInst)) { + assert(AddrModeInsts.back() == I && "Stack imbalance"); + AddrModeInsts.pop_back(); + } + + // Worse case, the target should support [reg] addressing modes. :) + if (!AddrMode.HasBaseReg) { + AddrMode.HasBaseReg = true; + // Still check for legality in case the target supports [imm] but not [i+r]. + if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) { + AddrMode.BaseReg = Addr; + return true; + } + AddrMode.HasBaseReg = false; + } + + // If the base register is already taken, see if we can do [r+r]. + if (AddrMode.Scale == 0) { + AddrMode.Scale = 1; + if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) { + AddrMode.ScaledReg = Addr; + return true; + } + AddrMode.Scale = 0; + } + // Couldn't match. + return false; +} + +/// TryMatchingScaledValue - Try adding ScaleReg*Scale to the specified +/// addressing mode. Return true if this addr mode is legal for the target, +/// false if not. +static bool TryMatchingScaledValue(Value *ScaleReg, int64_t Scale, + const Type *AccessTy, ExtAddrMode &AddrMode, + SmallVector &AddrModeInsts, + const TargetLowering &TLI, unsigned Depth) { + // If we already have a scale of this value, we can add to it, otherwise, we + // need an available scale field. + if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg) + return false; + + ExtAddrMode InputAddrMode = AddrMode; + + // Add scale to turn X*4+X*3 -> X*7. This could also do things like + // [A+B + A*7] -> [B+A*8]. + AddrMode.Scale += Scale; + AddrMode.ScaledReg = ScaleReg; + + if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) { + // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now + // to see if ScaleReg is actually X+C. If so, we can turn this into adding + // X*Scale + C*Scale to addr mode. + BinaryOperator *BinOp = dyn_cast(ScaleReg); + if (BinOp && BinOp->getOpcode() == Instruction::Add && + isa(BinOp->getOperand(1)) && InputAddrMode.ScaledReg ==0) { + + InputAddrMode.Scale = Scale; + InputAddrMode.ScaledReg = BinOp->getOperand(0); + InputAddrMode.BaseOffs += + cast(BinOp->getOperand(1))->getSExtValue()*Scale; + if (TLI.isLegalAddressingMode(InputAddrMode, AccessTy)) { + AddrModeInsts.push_back(BinOp); + AddrMode = InputAddrMode; + return true; + } + } + + // Otherwise, not (x+c)*scale, just return what we have. + return true; + } + + // Otherwise, back this attempt out. + AddrMode.Scale -= Scale; + if (AddrMode.Scale == 0) AddrMode.ScaledReg = 0; + + return false; +} + + +/// IsNonLocalValue - Return true if the specified values are defined in a +/// different basic block than BB. +static bool IsNonLocalValue(Value *V, BasicBlock *BB) { + if (Instruction *I = dyn_cast(V)) + return I->getParent() != BB; + return false; +} + +/// OptimizeLoadStoreInst - Load and Store Instructions have often have +/// addressing modes that can do significant amounts of computation. As such, +/// instruction selection will try to get the load or store to do as much +/// computation as possible for the program. The problem is that isel can only +/// see within a single block. As such, we sink as much legal addressing mode +/// stuff into the block as possible. +bool CodeGenPrepare::OptimizeLoadStoreInst(Instruction *LdStInst, Value *Addr, + const Type *AccessTy, + DenseMap &SunkAddrs) { + // Figure out what addressing mode will be built up for this operation. + SmallVector AddrModeInsts; + ExtAddrMode AddrMode; + bool Success = FindMaximalLegalAddressingMode(Addr, AccessTy, AddrMode, + AddrModeInsts, *TLI, 0); + Success = Success; assert(Success && "Couldn't select *anything*?"); + + // Check to see if any of the instructions supersumed by this addr mode are + // non-local to I's BB. + bool AnyNonLocal = false; + for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) { + if (IsNonLocalValue(AddrModeInsts[i], LdStInst->getParent())) { + AnyNonLocal = true; + break; + } + } + + // If all the instructions matched are already in this BB, don't do anything. + if (!AnyNonLocal) { + DEBUG(cerr << "CGP: Found local addrmode: " << AddrMode << "\n"); + return false; + } + + // Insert this computation right after this user. Since our caller is + // scanning from the top of the BB to the bottom, reuse of the expr are + // guaranteed to happen later. + BasicBlock::iterator InsertPt = LdStInst; + + // Now that we determined the addressing expression we want to use and know + // that we have to sink it into this block. Check to see if we have already + // done this for some other load/store instr in this block. If so, reuse the + // computation. + Value *&SunkAddr = SunkAddrs[Addr]; + if (SunkAddr) { + DEBUG(cerr << "CGP: Reusing nonlocal addrmode: " << AddrMode << "\n"); + if (SunkAddr->getType() != Addr->getType()) + SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt); + } else { + DEBUG(cerr << "CGP: SINKING nonlocal addrmode: " << AddrMode << "\n"); + const Type *IntPtrTy = TLI->getTargetData()->getIntPtrType(); + + Value *Result = 0; + // Start with the scale value. + if (AddrMode.Scale) { + Value *V = AddrMode.ScaledReg; + if (V->getType() == IntPtrTy) { + // done. + } else if (isa(V->getType())) { + V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt); + } else if (cast(IntPtrTy)->getBitWidth() < + cast(V->getType())->getBitWidth()) { + V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt); + } else { + V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt); + } + if (AddrMode.Scale != 1) + V = BinaryOperator::createMul(V, ConstantInt::get(IntPtrTy, + AddrMode.Scale), + "sunkaddr", InsertPt); + Result = V; + } + + // Add in the base register. + if (AddrMode.BaseReg) { + Value *V = AddrMode.BaseReg; + if (V->getType() != IntPtrTy) + V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt); + if (Result) + Result = BinaryOperator::createAdd(Result, V, "sunkaddr", InsertPt); + else + Result = V; + } + + // Add in the BaseGV if present. + if (AddrMode.BaseGV) { + Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr", + InsertPt); + if (Result) + Result = BinaryOperator::createAdd(Result, V, "sunkaddr", InsertPt); + else + Result = V; + } + + // Add in the Base Offset if present. + if (AddrMode.BaseOffs) { + Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs); + if (Result) + Result = BinaryOperator::createAdd(Result, V, "sunkaddr", InsertPt); + else + Result = V; + } + + if (Result == 0) + SunkAddr = Constant::getNullValue(Addr->getType()); + else + SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt); + } + + LdStInst->replaceUsesOfWith(Addr, SunkAddr); + + if (Addr->use_empty()) + EraseDeadInstructions(Addr); + return true; +} + // In this pass we look for GEP and cast instructions that are used // across basic blocks and rewrite them to improve basic-block-at-a-time // selection. @@ -665,21 +856,15 @@ bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) { } + // Keep track of non-local addresses that have been sunk into this block. + // This allows us to avoid inserting duplicate code for blocks with multiple + // load/stores of the same address. + DenseMap SunkAddrs; + for (BasicBlock::iterator BBI = BB.begin(), E = BB.end(); BBI != E; ) { Instruction *I = BBI++; - if (CallInst *CI = dyn_cast(I)) { - // If we found an inline asm expession, and if the target knows how to - // lower it to normal LLVM code, do so now. - if (TLI && isa(CI->getCalledValue())) - if (const TargetAsmInfo *TAI = - TLI->getTargetMachine().getTargetAsmInfo()) { - if (TAI->ExpandInlineAsm(CI)) - BBI = BB.begin(); - } - } else if (GetElementPtrInst *GEPI = dyn_cast(I)) { - MadeChange |= OptimizeGEPExpression(GEPI); - } else if (CastInst *CI = dyn_cast(I)) { + if (CastInst *CI = dyn_cast(I)) { // If the source of the cast is a constant, then this should have // already been constant folded. The only reason NOT to constant fold // it is if something (e.g. LSR) was careful to place the constant @@ -689,37 +874,53 @@ bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) { if (isa(CI->getOperand(0))) continue; - if (!TLI) continue; - - // If this is a noop copy, sink it into user blocks to reduce the number - // of virtual registers that must be created and coallesced. - MVT::ValueType SrcVT = TLI->getValueType(CI->getOperand(0)->getType()); - MVT::ValueType DstVT = TLI->getValueType(CI->getType()); - - // This is an fp<->int conversion? - if (MVT::isInteger(SrcVT) != MVT::isInteger(DstVT)) - continue; - - // If this is an extension, it will be a zero or sign extension, which - // isn't a noop. - if (SrcVT < DstVT) continue; - - // If these values will be promoted, find out what they will be promoted - // to. This helps us consider truncates on PPC as noop copies when they - // are. - if (TLI->getTypeAction(SrcVT) == TargetLowering::Promote) - SrcVT = TLI->getTypeToTransformTo(SrcVT); - if (TLI->getTypeAction(DstVT) == TargetLowering::Promote) - DstVT = TLI->getTypeToTransformTo(DstVT); - - // If, after promotion, these are the same types, this is a noop copy. - if (SrcVT == DstVT) - MadeChange |= OptimizeNoopCopyExpression(CI); - } else if (BinaryOperator *BinOp = dyn_cast(I)) { if (TLI) - MadeChange |= SinkInvariantGEPIndex(BinOp, *TLI); + MadeChange |= OptimizeNoopCopyExpression(CI, *TLI); + } else if (LoadInst *LI = dyn_cast(I)) { + if (TLI) + MadeChange |= OptimizeLoadStoreInst(I, I->getOperand(0), LI->getType(), + SunkAddrs); + } else if (StoreInst *SI = dyn_cast(I)) { + if (TLI) + MadeChange |= OptimizeLoadStoreInst(I, SI->getOperand(1), + SI->getOperand(0)->getType(), + SunkAddrs); + } else if (GetElementPtrInst *GEPI = dyn_cast(I)) { + bool HasNonZeroIdx = false; + for (GetElementPtrInst::op_iterator OI = GEPI->op_begin()+1, + E = GEPI->op_end(); OI != E; ++OI) { + if (ConstantInt *CI = dyn_cast(*OI)) { + if (!CI->isZero()) { + HasNonZeroIdx = true; + break; + } + } else { + HasNonZeroIdx = true; + break; + } + } + + if (!HasNonZeroIdx) { + /// The GEP operand must be a pointer, so must its result -> BitCast + Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(), + GEPI->getName(), GEPI); + GEPI->replaceAllUsesWith(NC); + GEPI->eraseFromParent(); + MadeChange = true; + BBI = NC; + } + } else if (CallInst *CI = dyn_cast(I)) { + // If we found an inline asm expession, and if the target knows how to + // lower it to normal LLVM code, do so now. + if (TLI && isa(CI->getCalledValue())) + if (const TargetAsmInfo *TAI = + TLI->getTargetMachine().getTargetAsmInfo()) { + if (TAI->ExpandInlineAsm(CI)) + BBI = BB.begin(); + } } } + return MadeChange; }