X-Git-Url: http://plrg.eecs.uci.edu/git/?p=oota-llvm.git;a=blobdiff_plain;f=lib%2FTarget%2FX86%2FX86ISelSimple.cpp;h=e1515672648d670c5cd61af5eaa7c9eb8c2cb3e9;hp=4cb6562211da599683df95e963dd890517cbb277;hb=b576c94c15af9a440f69d9d03c2afead7971118c;hpb=992447f91b720f3148d3d989a15f77fd60dd9ee7 diff --git a/lib/Target/X86/X86ISelSimple.cpp b/lib/Target/X86/X86ISelSimple.cpp index 4cb6562211d..e1515672648 100644 --- a/lib/Target/X86/X86ISelSimple.cpp +++ b/lib/Target/X86/X86ISelSimple.cpp @@ -1,6 +1,13 @@ //===-- InstSelectSimple.cpp - A simple instruction selector for x86 ------===// +// +// The LLVM Compiler Infrastructure // -// This file defines a simple peephole instruction selector for the x86 platform +// 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 defines a simple peephole instruction selector for the x86 target // //===----------------------------------------------------------------------===// @@ -8,35 +15,30 @@ #include "X86InstrInfo.h" #include "X86InstrBuilder.h" #include "llvm/Function.h" -#include "llvm/iTerminators.h" -#include "llvm/iOperators.h" -#include "llvm/iOther.h" -#include "llvm/iPHINode.h" -#include "llvm/iMemory.h" -#include "llvm/Type.h" +#include "llvm/Instructions.h" #include "llvm/DerivedTypes.h" #include "llvm/Constants.h" #include "llvm/Pass.h" +#include "llvm/Intrinsics.h" #include "llvm/CodeGen/MachineFunction.h" #include "llvm/CodeGen/MachineInstrBuilder.h" +#include "llvm/CodeGen/SSARegMap.h" +#include "llvm/CodeGen/MachineFrameInfo.h" +#include "llvm/CodeGen/MachineConstantPool.h" #include "llvm/Target/TargetMachine.h" -#include "llvm/Support/InstVisitor.h" #include "llvm/Target/MRegisterInfo.h" -#include - -using namespace MOTy; // Get Use, Def, UseAndDef - +#include "llvm/Support/InstVisitor.h" /// BMI - A special BuildMI variant that takes an iterator to insert the -/// instruction at as well as a basic block. -/// this is the version for when you have a destination register in mind. +/// instruction at as well as a basic block. This is the version for when you +/// have a destination register in mind. inline static MachineInstrBuilder BMI(MachineBasicBlock *MBB, MachineBasicBlock::iterator &I, - MachineOpCode Opcode, - unsigned NumOperands, + int Opcode, unsigned NumOperands, unsigned DestReg) { + assert(I >= MBB->begin() && I <= MBB->end() && "Bad iterator!"); MachineInstr *MI = new MachineInstr(Opcode, NumOperands+1, true, true); - I = ++MBB->insert(I, MI); + I = MBB->insert(I, MI)+1; return MachineInstrBuilder(MI).addReg(DestReg, MOTy::Def); } @@ -44,10 +46,10 @@ inline static MachineInstrBuilder BMI(MachineBasicBlock *MBB, /// instruction at as well as a basic block. inline static MachineInstrBuilder BMI(MachineBasicBlock *MBB, MachineBasicBlock::iterator &I, - MachineOpCode Opcode, - unsigned NumOperands) { + int Opcode, unsigned NumOperands) { + assert(I >= MBB->begin() && I <= MBB->end() && "Bad iterator!"); MachineInstr *MI = new MachineInstr(Opcode, NumOperands, true, true); - I = ++MBB->insert(I, MI); + I = MBB->insert(I, MI)+1; return MachineInstrBuilder(MI); } @@ -55,17 +57,16 @@ inline static MachineInstrBuilder BMI(MachineBasicBlock *MBB, namespace { struct ISel : public FunctionPass, InstVisitor { TargetMachine &TM; - MachineFunction *F; // The function we are compiling into - MachineBasicBlock *BB; // The current MBB we are compiling + MachineFunction *F; // The function we are compiling into + MachineBasicBlock *BB; // The current MBB we are compiling + int VarArgsFrameIndex; // FrameIndex for start of varargs area - unsigned CurReg; std::map RegMap; // Mapping between Val's and SSA Regs // MBBMap - Mapping between LLVM BB -> Machine BB std::map MBBMap; - ISel(TargetMachine &tm) - : TM(tm), F(0), BB(0), CurReg(MRegisterInfo::FirstVirtualRegister) {} + ISel(TargetMachine &tm) : TM(tm), F(0), BB(0) {} /// runOnFunction - Top level implementation of instruction selection for /// the entire function. @@ -73,9 +74,15 @@ namespace { bool runOnFunction(Function &Fn) { F = &MachineFunction::construct(&Fn, TM); + // Create all of the machine basic blocks for the function... for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) F->getBasicBlockList().push_back(MBBMap[I] = new MachineBasicBlock(I)); + BB = &F->front(); + + // Copy incoming arguments off of the stack... + LoadArgumentsToVirtualRegs(Fn); + // Instruction select everything except PHI nodes visit(Fn); @@ -84,9 +91,13 @@ namespace { RegMap.clear(); MBBMap.clear(); - CurReg = MRegisterInfo::FirstVirtualRegister; F = 0; - return false; // We never modify the LLVM itself. + // We always build a machine code representation for the function + return true; + } + + virtual const char *getPassName() const { + return "X86 Simple Instruction Selection"; } /// visitBasicBlock - This method is called when we are visiting a new basic @@ -98,6 +109,10 @@ namespace { BB = MBBMap[&LLVM_BB]; } + /// LoadArgumentsToVirtualRegs - Load all of the arguments to this function + /// from the stack into virtual registers. + /// + void LoadArgumentsToVirtualRegs(Function &F); /// SelectPHINodes - Insert machine code to generate phis. This is tricky /// because we have to generate our sources into the source basic blocks, @@ -112,16 +127,30 @@ namespace { // Control flow operators void visitReturnInst(ReturnInst &RI); void visitBranchInst(BranchInst &BI); + + struct ValueRecord { + Value *Val; + unsigned Reg; + const Type *Ty; + ValueRecord(unsigned R, const Type *T) : Val(0), Reg(R), Ty(T) {} + ValueRecord(Value *V) : Val(V), Reg(0), Ty(V->getType()) {} + }; + void doCall(const ValueRecord &Ret, MachineInstr *CallMI, + const std::vector &Args); void visitCallInst(CallInst &I); + void visitIntrinsicCall(LLVMIntrinsic::ID ID, CallInst &I); // Arithmetic operators void visitSimpleBinary(BinaryOperator &B, unsigned OpcodeClass); void visitAdd(BinaryOperator &B) { visitSimpleBinary(B, 0); } void visitSub(BinaryOperator &B) { visitSimpleBinary(B, 1); } - void doMultiply(unsigned destReg, const Type *resultType, - unsigned op0Reg, unsigned op1Reg, - MachineBasicBlock *MBB, - MachineBasicBlock::iterator &MBBI); + void doMultiply(MachineBasicBlock *MBB, MachineBasicBlock::iterator &MBBI, + unsigned DestReg, const Type *DestTy, + unsigned Op0Reg, unsigned Op1Reg); + void doMultiplyConst(MachineBasicBlock *MBB, + MachineBasicBlock::iterator &MBBI, + unsigned DestReg, const Type *DestTy, + unsigned Op0Reg, unsigned Op1Val); void visitMul(BinaryOperator &B); void visitDiv(BinaryOperator &B) { visitDivRem(B); } @@ -133,27 +162,26 @@ namespace { void visitOr (BinaryOperator &B) { visitSimpleBinary(B, 3); } void visitXor(BinaryOperator &B) { visitSimpleBinary(B, 4); } - // Binary comparison operators - void visitSetCCInst(SetCondInst &I, unsigned OpNum); - void visitSetEQ(SetCondInst &I) { visitSetCCInst(I, 0); } - void visitSetNE(SetCondInst &I) { visitSetCCInst(I, 1); } - void visitSetLT(SetCondInst &I) { visitSetCCInst(I, 2); } - void visitSetGT(SetCondInst &I) { visitSetCCInst(I, 3); } - void visitSetLE(SetCondInst &I) { visitSetCCInst(I, 4); } - void visitSetGE(SetCondInst &I) { visitSetCCInst(I, 5); } - + // Comparison operators... + void visitSetCondInst(SetCondInst &I); + unsigned EmitComparison(unsigned OpNum, Value *Op0, Value *Op1, + MachineBasicBlock *MBB, + MachineBasicBlock::iterator &MBBI); + // Memory Instructions void visitLoadInst(LoadInst &I); void visitStoreInst(StoreInst &I); void visitGetElementPtrInst(GetElementPtrInst &I); + void visitAllocaInst(AllocaInst &I); void visitMallocInst(MallocInst &I); void visitFreeInst(FreeInst &I); - void visitAllocaInst(AllocaInst &I); // Other operators void visitShiftInst(ShiftInst &I); void visitPHINode(PHINode &I) {} // PHI nodes handled by second pass void visitCastInst(CastInst &I); + void visitVANextInst(VANextInst &I); + void visitVAArgInst(VAArgInst &I); void visitInstruction(Instruction &I) { std::cerr << "Cannot instruction select: " << I; @@ -161,28 +189,71 @@ namespace { } /// promote32 - Make a value 32-bits wide, and put it somewhere. - void promote32 (const unsigned targetReg, Value *v); + /// + void promote32(unsigned targetReg, const ValueRecord &VR); + + /// EmitByteSwap - Byteswap SrcReg into DestReg. + /// + void EmitByteSwap(unsigned DestReg, unsigned SrcReg, unsigned Class); - // emitGEPOperation - Common code shared between visitGetElementPtrInst and - // constant expression GEP support. - // + /// emitGEPOperation - Common code shared between visitGetElementPtrInst and + /// constant expression GEP support. + /// void emitGEPOperation(MachineBasicBlock *BB, MachineBasicBlock::iterator&IP, Value *Src, User::op_iterator IdxBegin, User::op_iterator IdxEnd, unsigned TargetReg); + /// emitCastOperation - Common code shared between visitCastInst and + /// constant expression cast support. + void emitCastOperation(MachineBasicBlock *BB,MachineBasicBlock::iterator&IP, + Value *Src, const Type *DestTy, unsigned TargetReg); + + /// emitSimpleBinaryOperation - Common code shared between visitSimpleBinary + /// and constant expression support. + void emitSimpleBinaryOperation(MachineBasicBlock *BB, + MachineBasicBlock::iterator &IP, + Value *Op0, Value *Op1, + unsigned OperatorClass, unsigned TargetReg); + + /// emitSetCCOperation - Common code shared between visitSetCondInst and + /// constant expression support. + void emitSetCCOperation(MachineBasicBlock *BB, + MachineBasicBlock::iterator &IP, + Value *Op0, Value *Op1, unsigned Opcode, + unsigned TargetReg); + + /// copyConstantToRegister - Output the instructions required to put the /// specified constant into the specified register. /// - void copyConstantToRegister(Constant *C, unsigned Reg, - MachineBasicBlock *MBB, - MachineBasicBlock::iterator &MBBI); + void copyConstantToRegister(MachineBasicBlock *MBB, + MachineBasicBlock::iterator &MBBI, + Constant *C, unsigned Reg); - /// makeAnotherReg - This method returns the next register number - /// we haven't yet used. + /// makeAnotherReg - This method returns the next register number we haven't + /// yet used. + /// + /// Long values are handled somewhat specially. They are always allocated + /// as pairs of 32 bit integer values. The register number returned is the + /// lower 32 bits of the long value, and the regNum+1 is the upper 32 bits + /// of the long value. + /// unsigned makeAnotherReg(const Type *Ty) { + assert(dynamic_cast(TM.getRegisterInfo()) && + "Current target doesn't have X86 reg info??"); + const X86RegisterInfo *MRI = + static_cast(TM.getRegisterInfo()); + if (Ty == Type::LongTy || Ty == Type::ULongTy) { + const TargetRegisterClass *RC = MRI->getRegClassForType(Type::IntTy); + // Create the lower part + F->getSSARegMap()->createVirtualRegister(RC); + // Create the upper part. + return F->getSSARegMap()->createVirtualRegister(RC)-1; + } + // Add the mapping of regnumber => reg class to MachineFunction - F->addRegMap(CurReg, TM.getRegisterInfo()->getRegClassForType(Ty)); - return CurReg++; + const TargetRegisterClass *RC = MRI->getRegClassForType(Ty); + return F->getSSARegMap()->createVirtualRegister(RC); } /// getReg - This method turns an LLVM value into a register number. This @@ -207,31 +278,12 @@ namespace { // the register here... // if (Constant *C = dyn_cast(V)) { - copyConstantToRegister(C, Reg, MBB, IPt); + copyConstantToRegister(MBB, IPt, C, Reg); + RegMap.erase(V); // Assign a new name to this constant if ref'd again } else if (GlobalValue *GV = dyn_cast(V)) { // Move the address of the global into the register - BMI(MBB, IPt, X86::MOVir32, 1, Reg).addReg(GV); - } else if (Argument *A = dyn_cast(V)) { - // Find the position of the argument in the argument list. - const Function *f = F->getFunction (); - // The function's arguments look like this: - // [EBP] -- copy of old EBP - // [EBP + 4] -- return address - // [EBP + 8] -- first argument (leftmost lexically) - // So we want to start with counter = 2. - int counter = 2, argPos = -1; - for (Function::const_aiterator ai = f->abegin (), ae = f->aend (); - ai != ae; ++ai) { - if (&(*ai) == A) { - argPos = counter; - break; // Only need to find it once. ;-) - } - ++counter; - } - assert (argPos != -1 - && "Argument not found in current function's argument list"); - // Load it out of the stack frame at EBP + 4*argPos. - addRegOffset(BMI(MBB, IPt, X86::MOVmr32, 4, Reg), X86::EBP, 4*argPos); + BMI(MBB, IPt, X86::MOVir32, 1, Reg).addGlobalAddress(GV); + RegMap.erase(V); // Assign a new name to this address if ref'd again } return Reg; @@ -243,7 +295,7 @@ namespace { /// Representation. /// enum TypeClass { - cByte, cShort, cInt, cLong, cFloat, cDouble + cByte, cShort, cInt, cFP, cLong }; /// getClass - Turn a primitive type into a "class" number which is based on the @@ -259,56 +311,111 @@ static inline TypeClass getClass(const Type *Ty) { case Type::UIntTyID: case Type::PointerTyID: return cInt; // Int's and pointers are class #2 - case Type::LongTyID: - case Type::ULongTyID: //return cLong; // Longs are class #3 - return cInt; // FIXME: LONGS ARE TREATED AS INTS! + case Type::FloatTyID: + case Type::DoubleTyID: return cFP; // Floating Point is #3 - case Type::FloatTyID: return cFloat; // Float is class #4 - case Type::DoubleTyID: return cDouble; // Doubles are class #5 + case Type::LongTyID: + case Type::ULongTyID: return cLong; // Longs are class #4 default: assert(0 && "Invalid type to getClass!"); return cByte; // not reached } } +// getClassB - Just like getClass, but treat boolean values as bytes. +static inline TypeClass getClassB(const Type *Ty) { + if (Ty == Type::BoolTy) return cByte; + return getClass(Ty); +} + /// copyConstantToRegister - Output the instructions required to put the /// specified constant into the specified register. /// -void ISel::copyConstantToRegister(Constant *C, unsigned R, - MachineBasicBlock *MBB, - MachineBasicBlock::iterator &IP) { +void ISel::copyConstantToRegister(MachineBasicBlock *MBB, + MachineBasicBlock::iterator &IP, + Constant *C, unsigned R) { if (ConstantExpr *CE = dyn_cast(C)) { - if (CE->getOpcode() == Instruction::GetElementPtr) { - emitGEPOperation(BB, IP, CE->getOperand(0), + unsigned Class = 0; + switch (CE->getOpcode()) { + case Instruction::GetElementPtr: + emitGEPOperation(MBB, IP, CE->getOperand(0), CE->op_begin()+1, CE->op_end(), R); return; - } + case Instruction::Cast: + emitCastOperation(MBB, IP, CE->getOperand(0), CE->getType(), R); + return; + + case Instruction::Xor: ++Class; // FALL THROUGH + case Instruction::Or: ++Class; // FALL THROUGH + case Instruction::And: ++Class; // FALL THROUGH + case Instruction::Sub: ++Class; // FALL THROUGH + case Instruction::Add: + emitSimpleBinaryOperation(MBB, IP, CE->getOperand(0), CE->getOperand(1), + Class, R); + return; - std::cerr << "Offending expr: " << C << "\n"; - assert (0 && "Constant expressions not yet handled!\n"); + case Instruction::SetNE: + case Instruction::SetEQ: + case Instruction::SetLT: + case Instruction::SetGT: + case Instruction::SetLE: + case Instruction::SetGE: + emitSetCCOperation(MBB, IP, CE->getOperand(0), CE->getOperand(1), + CE->getOpcode(), R); + return; + + default: + std::cerr << "Offending expr: " << C << "\n"; + assert(0 && "Constant expression not yet handled!\n"); + } } if (C->getType()->isIntegral()) { - unsigned Class = getClass(C->getType()); - assert(Class != 3 && "Type not handled yet!"); + unsigned Class = getClassB(C->getType()); + + if (Class == cLong) { + // Copy the value into the register pair. + uint64_t Val = cast(C)->getRawValue(); + BMI(MBB, IP, X86::MOVir32, 1, R).addZImm(Val & 0xFFFFFFFF); + BMI(MBB, IP, X86::MOVir32, 1, R+1).addZImm(Val >> 32); + return; + } + + assert(Class <= cInt && "Type not handled yet!"); static const unsigned IntegralOpcodeTab[] = { X86::MOVir8, X86::MOVir16, X86::MOVir32 }; - if (C->getType()->isSigned()) { - ConstantSInt *CSI = cast(C); - BMI(MBB, IP, IntegralOpcodeTab[Class], 1, R).addSImm(CSI->getValue()); + if (C->getType() == Type::BoolTy) { + BMI(MBB, IP, X86::MOVir8, 1, R).addZImm(C == ConstantBool::True); } else { - ConstantUInt *CUI = cast(C); - BMI(MBB, IP, IntegralOpcodeTab[Class], 1, R).addZImm(CUI->getValue()); + ConstantInt *CI = cast(C); + BMI(MBB, IP, IntegralOpcodeTab[Class], 1, R).addZImm(CI->getRawValue()); + } + } else if (ConstantFP *CFP = dyn_cast(C)) { + double Value = CFP->getValue(); + if (Value == +0.0) + BMI(MBB, IP, X86::FLD0, 0, R); + else if (Value == +1.0) + BMI(MBB, IP, X86::FLD1, 0, R); + else { + // Otherwise we need to spill the constant to memory... + MachineConstantPool *CP = F->getConstantPool(); + unsigned CPI = CP->getConstantPoolIndex(CFP); + const Type *Ty = CFP->getType(); + + assert(Ty == Type::FloatTy || Ty == Type::DoubleTy && "Unknown FP type!"); + unsigned LoadOpcode = Ty == Type::FloatTy ? X86::FLDr32 : X86::FLDr64; + addConstantPoolReference(BMI(MBB, IP, LoadOpcode, 4, R), CPI); } + } else if (isa(C)) { // Copy zero (null pointer) to the register. BMI(MBB, IP, X86::MOVir32, 1, R).addZImm(0); } else if (ConstantPointerRef *CPR = dyn_cast(C)) { - unsigned SrcReg = getReg(CPR->getValue(), BB, IP); + unsigned SrcReg = getReg(CPR->getValue(), MBB, IP); BMI(MBB, IP, X86::MOVrr32, 1, R).addReg(SrcReg); } else { std::cerr << "Offending constant: " << C << "\n"; @@ -316,11 +423,76 @@ void ISel::copyConstantToRegister(Constant *C, unsigned R, } } +/// LoadArgumentsToVirtualRegs - Load all of the arguments to this function from +/// the stack into virtual registers. +/// +void ISel::LoadArgumentsToVirtualRegs(Function &Fn) { + // Emit instructions to load the arguments... On entry to a function on the + // X86, the stack frame looks like this: + // + // [ESP] -- return address + // [ESP + 4] -- first argument (leftmost lexically) + // [ESP + 8] -- second argument, if first argument is four bytes in size + // ... + // + unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot + MachineFrameInfo *MFI = F->getFrameInfo(); + + for (Function::aiterator I = Fn.abegin(), E = Fn.aend(); I != E; ++I) { + unsigned Reg = getReg(*I); + + int FI; // Frame object index + switch (getClassB(I->getType())) { + case cByte: + FI = MFI->CreateFixedObject(1, ArgOffset); + addFrameReference(BuildMI(BB, X86::MOVmr8, 4, Reg), FI); + break; + case cShort: + FI = MFI->CreateFixedObject(2, ArgOffset); + addFrameReference(BuildMI(BB, X86::MOVmr16, 4, Reg), FI); + break; + case cInt: + FI = MFI->CreateFixedObject(4, ArgOffset); + addFrameReference(BuildMI(BB, X86::MOVmr32, 4, Reg), FI); + break; + case cLong: + FI = MFI->CreateFixedObject(8, ArgOffset); + addFrameReference(BuildMI(BB, X86::MOVmr32, 4, Reg), FI); + addFrameReference(BuildMI(BB, X86::MOVmr32, 4, Reg+1), FI, 4); + ArgOffset += 4; // longs require 4 additional bytes + break; + case cFP: + unsigned Opcode; + if (I->getType() == Type::FloatTy) { + Opcode = X86::FLDr32; + FI = MFI->CreateFixedObject(4, ArgOffset); + } else { + Opcode = X86::FLDr64; + FI = MFI->CreateFixedObject(8, ArgOffset); + ArgOffset += 4; // doubles require 4 additional bytes + } + addFrameReference(BuildMI(BB, Opcode, 4, Reg), FI); + break; + default: + assert(0 && "Unhandled argument type!"); + } + ArgOffset += 4; // Each argument takes at least 4 bytes on the stack... + } + + // If the function takes variable number of arguments, add a frame offset for + // the start of the first vararg value... this is used to expand + // llvm.va_start. + if (Fn.getFunctionType()->isVarArg()) + VarArgsFrameIndex = MFI->CreateFixedObject(1, ArgOffset); +} + + /// SelectPHINodes - Insert machine code to generate phis. This is tricky /// because we have to generate our sources into the source basic blocks, not /// the current one. /// void ISel::SelectPHINodes() { + const TargetInstrInfo &TII = TM.getInstrInfo(); const Function &LF = *F->getFunction(); // The LLVM function... for (Function::const_iterator I = LF.begin(), E = LF.end(); I != E; ++I) { const BasicBlock *BB = I; @@ -329,135 +501,293 @@ void ISel::SelectPHINodes() { // Loop over all of the PHI nodes in the LLVM basic block... unsigned NumPHIs = 0; for (BasicBlock::const_iterator I = BB->begin(); - PHINode *PN = (PHINode*)dyn_cast(&*I); ++I) { + PHINode *PN = const_cast(dyn_cast(I)); ++I) { + // Create a new machine instr PHI node, and insert it. - MachineInstr *MI = BuildMI(X86::PHI, PN->getNumOperands(), getReg(*PN)); - MBB->insert(MBB->begin()+NumPHIs++, MI); // Insert it at the top of the BB + unsigned PHIReg = getReg(*PN); + MachineInstr *PhiMI = BuildMI(X86::PHI, PN->getNumOperands(), PHIReg); + MBB->insert(MBB->begin()+NumPHIs++, PhiMI); + + MachineInstr *LongPhiMI = 0; + if (PN->getType() == Type::LongTy || PN->getType() == Type::ULongTy) { + LongPhiMI = BuildMI(X86::PHI, PN->getNumOperands(), PHIReg+1); + MBB->insert(MBB->begin()+NumPHIs++, LongPhiMI); + } + + // PHIValues - Map of blocks to incoming virtual registers. We use this + // so that we only initialize one incoming value for a particular block, + // even if the block has multiple entries in the PHI node. + // + std::map PHIValues; for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { MachineBasicBlock *PredMBB = MBBMap[PN->getIncomingBlock(i)]; + unsigned ValReg; + std::map::iterator EntryIt = + PHIValues.lower_bound(PredMBB); + + if (EntryIt != PHIValues.end() && EntryIt->first == PredMBB) { + // We already inserted an initialization of the register for this + // predecessor. Recycle it. + ValReg = EntryIt->second; + + } else { + // Get the incoming value into a virtual register. + // + Value *Val = PN->getIncomingValue(i); + + // If this is a constant or GlobalValue, we may have to insert code + // into the basic block to compute it into a virtual register. + if (isa(Val) || isa(Val)) { + // Because we don't want to clobber any values which might be in + // physical registers with the computation of this constant (which + // might be arbitrarily complex if it is a constant expression), + // just insert the computation at the top of the basic block. + MachineBasicBlock::iterator PI = PredMBB->begin(); + + // Skip over any PHI nodes though! + while (PI != PredMBB->end() && (*PI)->getOpcode() == X86::PHI) + ++PI; + + ValReg = getReg(Val, PredMBB, PI); + } else { + ValReg = getReg(Val); + } + + // Remember that we inserted a value for this PHI for this predecessor + PHIValues.insert(EntryIt, std::make_pair(PredMBB, ValReg)); + } - // Get the incoming value into a virtual register. If it is not already - // available in a virtual register, insert the computation code into - // PredMBB - MachineBasicBlock::iterator PI = PredMBB->end()-1; - MI->addRegOperand(getReg(PN->getIncomingValue(i), PredMBB, PI)); - - - // FIXME: Pass in the MachineBasicBlocks instead of the basic blocks... - MI->addPCDispOperand(PN->getIncomingBlock(i)); // PredMBB + PhiMI->addRegOperand(ValReg); + PhiMI->addMachineBasicBlockOperand(PredMBB); + if (LongPhiMI) { + LongPhiMI->addRegOperand(ValReg+1); + LongPhiMI->addMachineBasicBlockOperand(PredMBB); + } } } } } +// canFoldSetCCIntoBranch - Return the setcc instruction if we can fold it into +// the conditional branch instruction which is the only user of the cc +// instruction. This is the case if the conditional branch is the only user of +// the setcc, and if the setcc is in the same basic block as the conditional +// branch. We also don't handle long arguments below, so we reject them here as +// well. +// +static SetCondInst *canFoldSetCCIntoBranch(Value *V) { + if (SetCondInst *SCI = dyn_cast(V)) + if (SCI->hasOneUse() && isa(SCI->use_back()) && + SCI->getParent() == cast(SCI->use_back())->getParent()) { + const Type *Ty = SCI->getOperand(0)->getType(); + if (Ty != Type::LongTy && Ty != Type::ULongTy) + return SCI; + } + return 0; +} + +// Return a fixed numbering for setcc instructions which does not depend on the +// order of the opcodes. +// +static unsigned getSetCCNumber(unsigned Opcode) { + switch(Opcode) { + default: assert(0 && "Unknown setcc instruction!"); + case Instruction::SetEQ: return 0; + case Instruction::SetNE: return 1; + case Instruction::SetLT: return 2; + case Instruction::SetGE: return 3; + case Instruction::SetGT: return 4; + case Instruction::SetLE: return 5; + } +} +// LLVM -> X86 signed X86 unsigned +// ----- ---------- ------------ +// seteq -> sete sete +// setne -> setne setne +// setlt -> setl setb +// setge -> setge setae +// setgt -> setg seta +// setle -> setle setbe +// ---- +// sets // Used by comparison with 0 optimization +// setns +static const unsigned SetCCOpcodeTab[2][8] = { + { X86::SETEr, X86::SETNEr, X86::SETBr, X86::SETAEr, X86::SETAr, X86::SETBEr, + 0, 0 }, + { X86::SETEr, X86::SETNEr, X86::SETLr, X86::SETGEr, X86::SETGr, X86::SETLEr, + X86::SETSr, X86::SETNSr }, +}; -/// SetCC instructions - Here we just emit boilerplate code to set a byte-sized -/// register, then move it to wherever the result should be. -/// We handle FP setcc instructions by pushing them, doing a -/// compare-and-pop-twice, and then copying the concodes to the main -/// processor's concodes (I didn't make this up, it's in the Intel manual) -/// -void ISel::visitSetCCInst(SetCondInst &I, unsigned OpNum) { +// EmitComparison - This function emits a comparison of the two operands, +// returning the extended setcc code to use. +unsigned ISel::EmitComparison(unsigned OpNum, Value *Op0, Value *Op1, + MachineBasicBlock *MBB, + MachineBasicBlock::iterator &IP) { // The arguments are already supposed to be of the same type. - const Type *CompTy = I.getOperand(0)->getType(); - unsigned reg1 = getReg(I.getOperand(0)); - unsigned reg2 = getReg(I.getOperand(1)); + const Type *CompTy = Op0->getType(); + unsigned Class = getClassB(CompTy); + unsigned Op0r = getReg(Op0, MBB, IP); + + // Special case handling of: cmp R, i + if (Class == cByte || Class == cShort || Class == cInt) + if (ConstantInt *CI = dyn_cast(Op1)) { + uint64_t Op1v = cast(CI)->getRawValue(); + + // Mask off any upper bits of the constant, if there are any... + Op1v &= (1ULL << (8 << Class)) - 1; + + // If this is a comparison against zero, emit more efficient code. We + // can't handle unsigned comparisons against zero unless they are == or + // !=. These should have been strength reduced already anyway. + if (Op1v == 0 && (CompTy->isSigned() || OpNum < 2)) { + static const unsigned TESTTab[] = { + X86::TESTrr8, X86::TESTrr16, X86::TESTrr32 + }; + BMI(MBB, IP, TESTTab[Class], 2).addReg(Op0r).addReg(Op0r); + + if (OpNum == 2) return 6; // Map jl -> js + if (OpNum == 3) return 7; // Map jg -> jns + return OpNum; + } + + static const unsigned CMPTab[] = { + X86::CMPri8, X86::CMPri16, X86::CMPri32 + }; + + BMI(MBB, IP, CMPTab[Class], 2).addReg(Op0r).addZImm(Op1v); + return OpNum; + } - unsigned Class = getClass(CompTy); + unsigned Op1r = getReg(Op1, MBB, IP); switch (Class) { + default: assert(0 && "Unknown type class!"); // Emit: cmp , (do the comparison). We can // compare 8-bit with 8-bit, 16-bit with 16-bit, 32-bit with // 32-bit. case cByte: - BuildMI (BB, X86::CMPrr8, 2).addReg (reg1).addReg (reg2); + BMI(MBB, IP, X86::CMPrr8, 2).addReg(Op0r).addReg(Op1r); break; case cShort: - BuildMI (BB, X86::CMPrr16, 2).addReg (reg1).addReg (reg2); + BMI(MBB, IP, X86::CMPrr16, 2).addReg(Op0r).addReg(Op1r); break; case cInt: - BuildMI (BB, X86::CMPrr32, 2).addReg (reg1).addReg (reg2); + BMI(MBB, IP, X86::CMPrr32, 2).addReg(Op0r).addReg(Op1r); break; - - // Push the variables on the stack with fldl opcodes. - // FIXME: assuming var1, var2 are in memory, if not, spill to - // stack first - case cFloat: // Floats - BuildMI (BB, X86::FLDr32, 1).addReg (reg1); - BuildMI (BB, X86::FLDr32, 1).addReg (reg2); - break; - case cDouble: // Doubles - BuildMI (BB, X86::FLDr64, 1).addReg (reg1); - BuildMI (BB, X86::FLDr64, 1).addReg (reg2); + case cFP: + BMI(MBB, IP, X86::FpUCOM, 2).addReg(Op0r).addReg(Op1r); + BMI(MBB, IP, X86::FNSTSWr8, 0); + BMI(MBB, IP, X86::SAHF, 1); break; + case cLong: - default: - visitInstruction(I); - } + if (OpNum < 2) { // seteq, setne + unsigned LoTmp = makeAnotherReg(Type::IntTy); + unsigned HiTmp = makeAnotherReg(Type::IntTy); + unsigned FinalTmp = makeAnotherReg(Type::IntTy); + BMI(MBB, IP, X86::XORrr32, 2, LoTmp).addReg(Op0r).addReg(Op1r); + BMI(MBB, IP, X86::XORrr32, 2, HiTmp).addReg(Op0r+1).addReg(Op1r+1); + BMI(MBB, IP, X86::ORrr32, 2, FinalTmp).addReg(LoTmp).addReg(HiTmp); + break; // Allow the sete or setne to be generated from flags set by OR + } else { + // Emit a sequence of code which compares the high and low parts once + // each, then uses a conditional move to handle the overflow case. For + // example, a setlt for long would generate code like this: + // + // AL = lo(op1) < lo(op2) // Signedness depends on operands + // BL = hi(op1) < hi(op2) // Always unsigned comparison + // dest = hi(op1) == hi(op2) ? AL : BL; + // - if (CompTy->isFloatingPoint()) { - // (Non-trapping) compare and pop twice. - BuildMI (BB, X86::FUCOMPP, 0); - // Move fp status word (concodes) to ax. - BuildMI (BB, X86::FNSTSWr8, 1, X86::AX); - // Load real concodes from ax. - BuildMI (BB, X86::SAHF, 1).addReg(X86::AH); + // FIXME: This would be much better if we had hierarchical register + // classes! Until then, hardcode registers so that we can deal with their + // aliases (because we don't have conditional byte moves). + // + BMI(MBB, IP, X86::CMPrr32, 2).addReg(Op0r).addReg(Op1r); + BMI(MBB, IP, SetCCOpcodeTab[0][OpNum], 0, X86::AL); + BMI(MBB, IP, X86::CMPrr32, 2).addReg(Op0r+1).addReg(Op1r+1); + BMI(MBB, IP, SetCCOpcodeTab[CompTy->isSigned()][OpNum], 0, X86::BL); + BMI(MBB, IP, X86::IMPLICIT_DEF, 0, X86::BH); + BMI(MBB, IP, X86::IMPLICIT_DEF, 0, X86::AH); + BMI(MBB, IP, X86::CMOVErr16, 2, X86::BX).addReg(X86::BX).addReg(X86::AX); + // NOTE: visitSetCondInst knows that the value is dumped into the BL + // register at this point for long values... + return OpNum; + } } + return OpNum; +} - // Emit setOp instruction (extract concode; clobbers ax), - // using the following mapping: - // LLVM -> X86 signed X86 unsigned - // ----- ----- ----- - // seteq -> sete sete - // setne -> setne setne - // setlt -> setl setb - // setgt -> setg seta - // setle -> setle setbe - // setge -> setge setae - - static const unsigned OpcodeTab[2][6] = { - {X86::SETEr, X86::SETNEr, X86::SETBr, X86::SETAr, X86::SETBEr, X86::SETAEr}, - {X86::SETEr, X86::SETNEr, X86::SETLr, X86::SETGr, X86::SETLEr, X86::SETGEr}, - }; - BuildMI(BB, OpcodeTab[CompTy->isSigned()][OpNum], 0, X86::AL); - - // Put it in the result using a move. - BuildMI (BB, X86::MOVrr8, 1, getReg(I)).addReg(X86::AL); +/// SetCC instructions - Here we just emit boilerplate code to set a byte-sized +/// register, then move it to wherever the result should be. +/// +void ISel::visitSetCondInst(SetCondInst &I) { + if (canFoldSetCCIntoBranch(&I)) return; // Fold this into a branch... + + unsigned DestReg = getReg(I); + MachineBasicBlock::iterator MII = BB->end(); + emitSetCCOperation(BB, MII, I.getOperand(0), I.getOperand(1), I.getOpcode(), + DestReg); } +/// emitSetCCOperation - Common code shared between visitSetCondInst and +/// constant expression support. +void ISel::emitSetCCOperation(MachineBasicBlock *MBB, + MachineBasicBlock::iterator &IP, + Value *Op0, Value *Op1, unsigned Opcode, + unsigned TargetReg) { + unsigned OpNum = getSetCCNumber(Opcode); + OpNum = EmitComparison(OpNum, Op0, Op1, MBB, IP); + + const Type *CompTy = Op0->getType(); + unsigned CompClass = getClassB(CompTy); + bool isSigned = CompTy->isSigned() && CompClass != cFP; + + if (CompClass != cLong || OpNum < 2) { + // Handle normal comparisons with a setcc instruction... + BMI(MBB, IP, SetCCOpcodeTab[isSigned][OpNum], 0, TargetReg); + } else { + // Handle long comparisons by copying the value which is already in BL into + // the register we want... + BMI(MBB, IP, X86::MOVrr8, 1, TargetReg).addReg(X86::BL); + } +} + + + + /// promote32 - Emit instructions to turn a narrow operand into a 32-bit-wide /// operand, in the specified target register. -void -ISel::promote32 (unsigned targetReg, Value *v) -{ - unsigned vReg = getReg (v); - unsigned Class = getClass (v->getType ()); - bool isUnsigned = v->getType ()->isUnsigned (); - assert (((Class == cByte) || (Class == cShort) || (Class == cInt)) - && "Unpromotable operand class in promote32"); - switch (Class) - { - case cByte: - // Extend value into target register (8->32) - if (isUnsigned) - BuildMI (BB, X86::MOVZXr32r8, 1, targetReg).addReg (vReg); - else - BuildMI (BB, X86::MOVSXr32r8, 1, targetReg).addReg (vReg); - break; - case cShort: - // Extend value into target register (16->32) - if (isUnsigned) - BuildMI (BB, X86::MOVZXr32r16, 1, targetReg).addReg (vReg); - else - BuildMI (BB, X86::MOVSXr32r16, 1, targetReg).addReg (vReg); - break; - case cInt: - // Move value into target register (32->32) - BuildMI (BB, X86::MOVrr32, 1, targetReg).addReg (vReg); - break; - } +void ISel::promote32(unsigned targetReg, const ValueRecord &VR) { + bool isUnsigned = VR.Ty->isUnsigned(); + + // Make sure we have the register number for this value... + unsigned Reg = VR.Val ? getReg(VR.Val) : VR.Reg; + + switch (getClassB(VR.Ty)) { + case cByte: + // Extend value into target register (8->32) + if (isUnsigned) + BuildMI(BB, X86::MOVZXr32r8, 1, targetReg).addReg(Reg); + else + BuildMI(BB, X86::MOVSXr32r8, 1, targetReg).addReg(Reg); + break; + case cShort: + // Extend value into target register (16->32) + if (isUnsigned) + BuildMI(BB, X86::MOVZXr32r16, 1, targetReg).addReg(Reg); + else + BuildMI(BB, X86::MOVSXr32r16, 1, targetReg).addReg(Reg); + break; + case cInt: + // Move value into target register (32->32) + BuildMI(BB, X86::MOVrr32, 1, targetReg).addReg(Reg); + break; + default: + assert(0 && "Unpromotable operand class in promote32"); + } } /// 'ret' instruction - Here we are interested in meeting the x86 ABI. As such, @@ -471,120 +801,202 @@ ISel::promote32 (unsigned targetReg, Value *v) /// ret long, ulong : Move value into EAX/EDX and return /// ret float/double : Top of FP stack /// -void -ISel::visitReturnInst (ReturnInst &I) -{ - if (I.getNumOperands () == 0) - { - // Emit a 'ret' instruction - BuildMI (BB, X86::RET, 0); - return; - } - Value *rv = I.getOperand (0); - unsigned Class = getClass (rv->getType ()); - switch (Class) - { - // integral return values: extend or move into EAX and return. - case cByte: - case cShort: - case cInt: - promote32 (X86::EAX, rv); - break; - // ret float/double: top of FP stack - // FLD - case cFloat: // Floats - BuildMI (BB, X86::FLDr32, 1).addReg (getReg (rv)); - break; - case cDouble: // Doubles - BuildMI (BB, X86::FLDr64, 1).addReg (getReg (rv)); - break; - case cLong: - // ret long: use EAX(least significant 32 bits)/EDX (most - // significant 32)...uh, I think so Brain, but how do i call - // up the two parts of the value from inside this mouse - // cage? *zort* - default: - visitInstruction (I); - } +void ISel::visitReturnInst(ReturnInst &I) { + if (I.getNumOperands() == 0) { + BuildMI(BB, X86::RET, 0); // Just emit a 'ret' instruction + return; + } + + Value *RetVal = I.getOperand(0); + unsigned RetReg = getReg(RetVal); + switch (getClassB(RetVal->getType())) { + case cByte: // integral return values: extend or move into EAX and return + case cShort: + case cInt: + promote32(X86::EAX, ValueRecord(RetReg, RetVal->getType())); + // Declare that EAX is live on exit + BuildMI(BB, X86::IMPLICIT_USE, 2).addReg(X86::EAX).addReg(X86::ESP); + break; + case cFP: // Floats & Doubles: Return in ST(0) + BuildMI(BB, X86::FpSETRESULT, 1).addReg(RetReg); + // Declare that top-of-stack is live on exit + BuildMI(BB, X86::IMPLICIT_USE, 2).addReg(X86::ST0).addReg(X86::ESP); + break; + case cLong: + BuildMI(BB, X86::MOVrr32, 1, X86::EAX).addReg(RetReg); + BuildMI(BB, X86::MOVrr32, 1, X86::EDX).addReg(RetReg+1); + // Declare that EAX & EDX are live on exit + BuildMI(BB, X86::IMPLICIT_USE, 3).addReg(X86::EAX).addReg(X86::EDX).addReg(X86::ESP); + break; + default: + visitInstruction(I); + } // Emit a 'ret' instruction - BuildMI (BB, X86::RET, 0); + BuildMI(BB, X86::RET, 0); +} + +// getBlockAfter - Return the basic block which occurs lexically after the +// specified one. +static inline BasicBlock *getBlockAfter(BasicBlock *BB) { + Function::iterator I = BB; ++I; // Get iterator to next block + return I != BB->getParent()->end() ? &*I : 0; } /// visitBranchInst - Handle conditional and unconditional branches here. Note /// that since code layout is frozen at this point, that if we are trying to /// jump to a block that is the immediate successor of the current block, we can -/// just make a fall-through. (but we don't currently). +/// just make a fall-through (but we don't currently). /// -void -ISel::visitBranchInst (BranchInst & BI) -{ - if (BI.isConditional ()) - { - BasicBlock *ifTrue = BI.getSuccessor (0); - BasicBlock *ifFalse = BI.getSuccessor (1); // this is really unobvious - - // simplest thing I can think of: compare condition with zero, - // followed by jump-if-equal to ifFalse, and jump-if-nonequal to - // ifTrue - unsigned int condReg = getReg (BI.getCondition ()); - BuildMI (BB, X86::CMPri8, 2).addReg (condReg).addZImm (0); - BuildMI (BB, X86::JNE, 1).addPCDisp (BI.getSuccessor (0)); - BuildMI (BB, X86::JE, 1).addPCDisp (BI.getSuccessor (1)); +void ISel::visitBranchInst(BranchInst &BI) { + BasicBlock *NextBB = getBlockAfter(BI.getParent()); // BB after current one + + if (!BI.isConditional()) { // Unconditional branch? + if (BI.getSuccessor(0) != NextBB) + BuildMI(BB, X86::JMP, 1).addPCDisp(BI.getSuccessor(0)); + return; + } + + // See if we can fold the setcc into the branch itself... + SetCondInst *SCI = canFoldSetCCIntoBranch(BI.getCondition()); + if (SCI == 0) { + // Nope, cannot fold setcc into this branch. Emit a branch on a condition + // computed some other way... + unsigned condReg = getReg(BI.getCondition()); + BuildMI(BB, X86::CMPri8, 2).addReg(condReg).addZImm(0); + if (BI.getSuccessor(1) == NextBB) { + if (BI.getSuccessor(0) != NextBB) + BuildMI(BB, X86::JNE, 1).addPCDisp(BI.getSuccessor(0)); + } else { + BuildMI(BB, X86::JE, 1).addPCDisp(BI.getSuccessor(1)); + + if (BI.getSuccessor(0) != NextBB) + BuildMI(BB, X86::JMP, 1).addPCDisp(BI.getSuccessor(0)); } - else // unconditional branch - { - BuildMI (BB, X86::JMP, 1).addPCDisp (BI.getSuccessor (0)); + return; + } + + unsigned OpNum = getSetCCNumber(SCI->getOpcode()); + MachineBasicBlock::iterator MII = BB->end(); + OpNum = EmitComparison(OpNum, SCI->getOperand(0), SCI->getOperand(1), BB, MII); + + const Type *CompTy = SCI->getOperand(0)->getType(); + bool isSigned = CompTy->isSigned() && getClassB(CompTy) != cFP; + + + // LLVM -> X86 signed X86 unsigned + // ----- ---------- ------------ + // seteq -> je je + // setne -> jne jne + // setlt -> jl jb + // setge -> jge jae + // setgt -> jg ja + // setle -> jle jbe + // ---- + // js // Used by comparison with 0 optimization + // jns + + static const unsigned OpcodeTab[2][8] = { + { X86::JE, X86::JNE, X86::JB, X86::JAE, X86::JA, X86::JBE, 0, 0 }, + { X86::JE, X86::JNE, X86::JL, X86::JGE, X86::JG, X86::JLE, + X86::JS, X86::JNS }, + }; + + if (BI.getSuccessor(0) != NextBB) { + BuildMI(BB, OpcodeTab[isSigned][OpNum], 1).addPCDisp(BI.getSuccessor(0)); + if (BI.getSuccessor(1) != NextBB) + BuildMI(BB, X86::JMP, 1).addPCDisp(BI.getSuccessor(1)); + } else { + // Change to the inverse condition... + if (BI.getSuccessor(1) != NextBB) { + OpNum ^= 1; + BuildMI(BB, OpcodeTab[isSigned][OpNum], 1).addPCDisp(BI.getSuccessor(1)); } + } } -/// visitCallInst - Push args on stack and do a procedure call instruction. -void -ISel::visitCallInst (CallInst & CI) -{ - // keep a counter of how many bytes we pushed on the stack - unsigned bytesPushed = 0; - - // Push the arguments on the stack in reverse order, as specified by - // the ABI. - for (unsigned i = CI.getNumOperands()-1; i >= 1; --i) - { - Value *v = CI.getOperand (i); - switch (getClass (v->getType ())) - { - case cByte: - case cShort: - // Promote V to 32 bits wide, and move the result into EAX, - // then push EAX. - promote32 (X86::EAX, v); - BuildMI (BB, X86::PUSHr32, 1).addReg (X86::EAX); - bytesPushed += 4; - break; - case cInt: - case cFloat: { - unsigned Reg = getReg(v); - BuildMI (BB, X86::PUSHr32, 1).addReg(Reg); - bytesPushed += 4; - break; - } - default: - // FIXME: long/ulong/double args not handled. - visitInstruction (CI); - break; + +/// doCall - This emits an abstract call instruction, setting up the arguments +/// and the return value as appropriate. For the actual function call itself, +/// it inserts the specified CallMI instruction into the stream. +/// +void ISel::doCall(const ValueRecord &Ret, MachineInstr *CallMI, + const std::vector &Args) { + + // Count how many bytes are to be pushed on the stack... + unsigned NumBytes = 0; + + if (!Args.empty()) { + for (unsigned i = 0, e = Args.size(); i != e; ++i) + switch (getClassB(Args[i].Ty)) { + case cByte: case cShort: case cInt: + NumBytes += 4; break; + case cLong: + NumBytes += 8; break; + case cFP: + NumBytes += Args[i].Ty == Type::FloatTy ? 4 : 8; + break; + default: assert(0 && "Unknown class!"); + } + + // Adjust the stack pointer for the new arguments... + BuildMI(BB, X86::ADJCALLSTACKDOWN, 1).addZImm(NumBytes); + + // Arguments go on the stack in reverse order, as specified by the ABI. + unsigned ArgOffset = 0; + for (unsigned i = 0, e = Args.size(); i != e; ++i) { + unsigned ArgReg = Args[i].Val ? getReg(Args[i].Val) : Args[i].Reg; + switch (getClassB(Args[i].Ty)) { + case cByte: + case cShort: { + // Promote arg to 32 bits wide into a temporary register... + unsigned R = makeAnotherReg(Type::UIntTy); + promote32(R, Args[i]); + addRegOffset(BuildMI(BB, X86::MOVrm32, 5), + X86::ESP, ArgOffset).addReg(R); + break; + } + case cInt: + addRegOffset(BuildMI(BB, X86::MOVrm32, 5), + X86::ESP, ArgOffset).addReg(ArgReg); + break; + case cLong: + addRegOffset(BuildMI(BB, X86::MOVrm32, 5), + X86::ESP, ArgOffset).addReg(ArgReg); + addRegOffset(BuildMI(BB, X86::MOVrm32, 5), + X86::ESP, ArgOffset+4).addReg(ArgReg+1); + ArgOffset += 4; // 8 byte entry, not 4. + break; + + case cFP: + if (Args[i].Ty == Type::FloatTy) { + addRegOffset(BuildMI(BB, X86::FSTr32, 5), + X86::ESP, ArgOffset).addReg(ArgReg); + } else { + assert(Args[i].Ty == Type::DoubleTy && "Unknown FP type!"); + addRegOffset(BuildMI(BB, X86::FSTr64, 5), + X86::ESP, ArgOffset).addReg(ArgReg); + ArgOffset += 4; // 8 byte entry, not 4. } + break; + + default: assert(0 && "Unknown class!"); + } + ArgOffset += 4; } - // Emit a CALL instruction with PC-relative displacement. - BuildMI (BB, X86::CALLpcrel32, 1).addPCDisp (CI.getCalledValue ()); + } else { + BuildMI(BB, X86::ADJCALLSTACKDOWN, 1).addZImm(0); + } + + BB->push_back(CallMI); - // Adjust the stack by `bytesPushed' amount if non-zero - if (bytesPushed > 0) - BuildMI (BB, X86::ADDri32, 2).addReg(X86::ESP).addZImm(bytesPushed); + BuildMI(BB, X86::ADJCALLSTACKUP, 1).addZImm(NumBytes); // If there is a return value, scavenge the result from the location the call // leaves it in... // - if (CI.getType() != Type::VoidTy) { - unsigned resultTypeClass = getClass (CI.getType ()); - switch (resultTypeClass) { + if (Ret.Ty != Type::VoidTy) { + unsigned DestClass = getClassB(Ret.Ty); + switch (DestClass) { case cByte: case cShort: case cInt: { @@ -594,95 +1006,329 @@ ISel::visitCallInst (CallInst & CI) X86::MOVrr8, X86::MOVrr16, X86::MOVrr32 }; static const unsigned AReg[] = { X86::AL, X86::AX, X86::EAX }; - BuildMI (BB, regRegMove[resultTypeClass], 1, - getReg (CI)).addReg (AReg[resultTypeClass]); + BuildMI(BB, regRegMove[DestClass], 1, Ret.Reg).addReg(AReg[DestClass]); break; } - case cFloat: - // Floating-point return values live in %st(0) (i.e., the top of - // the FP stack.) The general way to approach this is to do a - // FSTP to save the top of the FP stack on the real stack, then - // do a MOV to load the top of the real stack into the target - // register. - visitInstruction (CI); // FIXME: add the right args for the calls below - // BuildMI (BB, X86::FSTPm32, 0); - // BuildMI (BB, X86::MOVmr32, 0); + case cFP: // Floating-point return values live in %ST(0) + BuildMI(BB, X86::FpGETRESULT, 1, Ret.Reg); break; - default: - std::cerr << "Cannot get return value for call of type '" - << *CI.getType() << "'\n"; - visitInstruction(CI); + case cLong: // Long values are left in EDX:EAX + BuildMI(BB, X86::MOVrr32, 1, Ret.Reg).addReg(X86::EAX); + BuildMI(BB, X86::MOVrr32, 1, Ret.Reg+1).addReg(X86::EDX); + break; + default: assert(0 && "Unknown class!"); + } + } +} + + +/// visitCallInst - Push args on stack and do a procedure call instruction. +void ISel::visitCallInst(CallInst &CI) { + MachineInstr *TheCall; + if (Function *F = CI.getCalledFunction()) { + // Is it an intrinsic function call? + if (LLVMIntrinsic::ID ID = (LLVMIntrinsic::ID)F->getIntrinsicID()) { + visitIntrinsicCall(ID, CI); // Special intrinsics are not handled here + return; } + + // Emit a CALL instruction with PC-relative displacement. + TheCall = BuildMI(X86::CALLpcrel32, 1).addGlobalAddress(F, true); + } else { // Emit an indirect call... + unsigned Reg = getReg(CI.getCalledValue()); + TheCall = BuildMI(X86::CALLr32, 1).addReg(Reg); + } + + std::vector Args; + for (unsigned i = 1, e = CI.getNumOperands(); i != e; ++i) + Args.push_back(ValueRecord(CI.getOperand(i))); + + unsigned DestReg = CI.getType() != Type::VoidTy ? getReg(CI) : 0; + doCall(ValueRecord(DestReg, CI.getType()), TheCall, Args); +} + + +void ISel::visitIntrinsicCall(LLVMIntrinsic::ID ID, CallInst &CI) { + unsigned TmpReg1, TmpReg2; + switch (ID) { + case LLVMIntrinsic::va_start: + // Get the address of the first vararg value... + TmpReg1 = getReg(CI); + addFrameReference(BuildMI(BB, X86::LEAr32, 5, TmpReg1), VarArgsFrameIndex); + return; + + case LLVMIntrinsic::va_copy: + TmpReg1 = getReg(CI); + TmpReg2 = getReg(CI.getOperand(1)); + BuildMI(BB, X86::MOVrr32, 1, TmpReg1).addReg(TmpReg2); + return; + case LLVMIntrinsic::va_end: return; // Noop on X86 + + case LLVMIntrinsic::longjmp: + case LLVMIntrinsic::siglongjmp: + BuildMI(BB, X86::CALLpcrel32, 1).addExternalSymbol("abort", true); + return; + + case LLVMIntrinsic::setjmp: + case LLVMIntrinsic::sigsetjmp: + // Setjmp always returns zero... + BuildMI(BB, X86::MOVir32, 1, getReg(CI)).addZImm(0); + return; + default: assert(0 && "Unknown intrinsic for X86!"); } } + /// visitSimpleBinary - Implement simple binary operators for integral types... -/// OperatorClass is one of: 0 for Add, 1 for Sub, 2 for And, 3 for Or, -/// 4 for Xor. -/// +/// OperatorClass is one of: 0 for Add, 1 for Sub, 2 for And, 3 for Or, 4 for +/// Xor. void ISel::visitSimpleBinary(BinaryOperator &B, unsigned OperatorClass) { - if (B.getType() == Type::BoolTy) // FIXME: Handle bools for logicals - visitInstruction(B); + unsigned DestReg = getReg(B); + MachineBasicBlock::iterator MI = BB->end(); + emitSimpleBinaryOperation(BB, MI, B.getOperand(0), B.getOperand(1), + OperatorClass, DestReg); +} - unsigned Class = getClass(B.getType()); - if (Class > 2) // FIXME: Handle longs - visitInstruction(B); +/// emitSimpleBinaryOperation - Implement simple binary operators for integral +/// types... OperatorClass is one of: 0 for Add, 1 for Sub, 2 for And, 3 for +/// Or, 4 for Xor. +/// +/// emitSimpleBinaryOperation - Common code shared between visitSimpleBinary +/// and constant expression support. +/// +void ISel::emitSimpleBinaryOperation(MachineBasicBlock *MBB, + MachineBasicBlock::iterator &IP, + Value *Op0, Value *Op1, + unsigned OperatorClass, unsigned DestReg) { + unsigned Class = getClassB(Op0->getType()); + + // sub 0, X -> neg X + if (OperatorClass == 1 && Class != cLong) + if (ConstantInt *CI = dyn_cast(Op0)) + if (CI->isNullValue()) { + unsigned op1Reg = getReg(Op1, MBB, IP); + switch (Class) { + default: assert(0 && "Unknown class for this function!"); + case cByte: + BMI(MBB, IP, X86::NEGr8, 1, DestReg).addReg(op1Reg); + return; + case cShort: + BMI(MBB, IP, X86::NEGr16, 1, DestReg).addReg(op1Reg); + return; + case cInt: + BMI(MBB, IP, X86::NEGr32, 1, DestReg).addReg(op1Reg); + return; + } + } - static const unsigned OpcodeTab[][4] = { - // Arithmetic operators - { X86::ADDrr8, X86::ADDrr16, X86::ADDrr32, 0 }, // ADD - { X86::SUBrr8, X86::SUBrr16, X86::SUBrr32, 0 }, // SUB + if (!isa(Op1) || Class == cLong) { + static const unsigned OpcodeTab[][4] = { + // Arithmetic operators + { X86::ADDrr8, X86::ADDrr16, X86::ADDrr32, X86::FpADD }, // ADD + { X86::SUBrr8, X86::SUBrr16, X86::SUBrr32, X86::FpSUB }, // SUB + + // Bitwise operators + { X86::ANDrr8, X86::ANDrr16, X86::ANDrr32, 0 }, // AND + { X86:: ORrr8, X86:: ORrr16, X86:: ORrr32, 0 }, // OR + { X86::XORrr8, X86::XORrr16, X86::XORrr32, 0 }, // XOR + }; + + bool isLong = false; + if (Class == cLong) { + isLong = true; + Class = cInt; // Bottom 32 bits are handled just like ints + } + + unsigned Opcode = OpcodeTab[OperatorClass][Class]; + assert(Opcode && "Floating point arguments to logical inst?"); + unsigned Op0r = getReg(Op0, MBB, IP); + unsigned Op1r = getReg(Op1, MBB, IP); + BMI(MBB, IP, Opcode, 2, DestReg).addReg(Op0r).addReg(Op1r); + + if (isLong) { // Handle the upper 32 bits of long values... + static const unsigned TopTab[] = { + X86::ADCrr32, X86::SBBrr32, X86::ANDrr32, X86::ORrr32, X86::XORrr32 + }; + BMI(MBB, IP, TopTab[OperatorClass], 2, + DestReg+1).addReg(Op0r+1).addReg(Op1r+1); + } + return; + } + // Special case: op Reg, + ConstantInt *Op1C = cast(Op1); + unsigned Op0r = getReg(Op0, MBB, IP); + + // xor X, -1 -> not X + if (OperatorClass == 4 && Op1C->isAllOnesValue()) { + static unsigned const NOTTab[] = { X86::NOTr8, X86::NOTr16, X86::NOTr32 }; + BMI(MBB, IP, NOTTab[Class], 1, DestReg).addReg(Op0r); + return; + } + + // add X, -1 -> dec X + if (OperatorClass == 0 && Op1C->isAllOnesValue()) { + static unsigned const DECTab[] = { X86::DECr8, X86::DECr16, X86::DECr32 }; + BMI(MBB, IP, DECTab[Class], 1, DestReg).addReg(Op0r); + return; + } + + // add X, 1 -> inc X + if (OperatorClass == 0 && Op1C->equalsInt(1)) { + static unsigned const DECTab[] = { X86::INCr8, X86::INCr16, X86::INCr32 }; + BMI(MBB, IP, DECTab[Class], 1, DestReg).addReg(Op0r); + return; + } + + static const unsigned OpcodeTab[][3] = { + // Arithmetic operators + { X86::ADDri8, X86::ADDri16, X86::ADDri32 }, // ADD + { X86::SUBri8, X86::SUBri16, X86::SUBri32 }, // SUB + // Bitwise operators - { X86::ANDrr8, X86::ANDrr16, X86::ANDrr32, 0 }, // AND - { X86:: ORrr8, X86:: ORrr16, X86:: ORrr32, 0 }, // OR - { X86::XORrr8, X86::XORrr16, X86::XORrr32, 0 }, // XOR + { X86::ANDri8, X86::ANDri16, X86::ANDri32 }, // AND + { X86:: ORri8, X86:: ORri16, X86:: ORri32 }, // OR + { X86::XORri8, X86::XORri16, X86::XORri32 }, // XOR }; + assert(Class < 3 && "General code handles 64-bit integer types!"); unsigned Opcode = OpcodeTab[OperatorClass][Class]; - unsigned Op0r = getReg(B.getOperand(0)); - unsigned Op1r = getReg(B.getOperand(1)); - BuildMI(BB, Opcode, 2, getReg(B)).addReg(Op0r).addReg(Op1r); + uint64_t Op1v = cast(Op1C)->getRawValue(); + + // Mask off any upper bits of the constant, if there are any... + Op1v &= (1ULL << (8 << Class)) - 1; + BMI(MBB, IP, Opcode, 2, DestReg).addReg(Op0r).addZImm(Op1v); } -/// doMultiply - Emit appropriate instructions to multiply together -/// the registers op0Reg and op1Reg, and put the result in destReg. -/// The type of the result should be given as resultType. -void -ISel::doMultiply(unsigned destReg, const Type *resultType, - unsigned op0Reg, unsigned op1Reg, - MachineBasicBlock *MBB, MachineBasicBlock::iterator &MBBI) -{ - unsigned Class = getClass (resultType); - - // FIXME: - assert (Class <= 2 && "Someday, we will learn how to multiply" - "longs and floating-point numbers. This is not that day."); - - static const unsigned Regs[] ={ X86::AL , X86::AX , X86::EAX }; - static const unsigned MulOpcode[]={ X86::MULrr8, X86::MULrr16, X86::MULrr32 }; - static const unsigned MovOpcode[]={ X86::MOVrr8, X86::MOVrr16, X86::MOVrr32 }; - unsigned Reg = Regs[Class]; +/// doMultiply - Emit appropriate instructions to multiply together the +/// registers op0Reg and op1Reg, and put the result in DestReg. The type of the +/// result should be given as DestTy. +/// +void ISel::doMultiply(MachineBasicBlock *MBB, MachineBasicBlock::iterator &MBBI, + unsigned DestReg, const Type *DestTy, + unsigned op0Reg, unsigned op1Reg) { + unsigned Class = getClass(DestTy); + switch (Class) { + case cFP: // Floating point multiply + BMI(BB, MBBI, X86::FpMUL, 2, DestReg).addReg(op0Reg).addReg(op1Reg); + return; + case cInt: + case cShort: + BMI(BB, MBBI, Class == cInt ? X86::IMULrr32 : X86::IMULrr16, 2, DestReg) + .addReg(op0Reg).addReg(op1Reg); + return; + case cByte: + // Must use the MUL instruction, which forces use of AL... + BMI(MBB, MBBI, X86::MOVrr8, 1, X86::AL).addReg(op0Reg); + BMI(MBB, MBBI, X86::MULr8, 1).addReg(op1Reg); + BMI(MBB, MBBI, X86::MOVrr8, 1, DestReg).addReg(X86::AL); + return; + default: + case cLong: assert(0 && "doMultiply cannot operate on LONG values!"); + } +} - // Emit a MOV to put the first operand into the appropriately-sized - // subreg of EAX. - BMI(MBB, MBBI, MovOpcode[Class], 1, Reg).addReg (op0Reg); +// ExactLog2 - This function solves for (Val == 1 << (N-1)) and returns N. It +// returns zero when the input is not exactly a power of two. +static unsigned ExactLog2(unsigned Val) { + if (Val == 0) return 0; + unsigned Count = 0; + while (Val != 1) { + if (Val & 1) return 0; + Val >>= 1; + ++Count; + } + return Count+1; +} + +void ISel::doMultiplyConst(MachineBasicBlock *MBB, + MachineBasicBlock::iterator &IP, + unsigned DestReg, const Type *DestTy, + unsigned op0Reg, unsigned ConstRHS) { + unsigned Class = getClass(DestTy); + + // If the element size is exactly a power of 2, use a shift to get it. + if (unsigned Shift = ExactLog2(ConstRHS)) { + switch (Class) { + default: assert(0 && "Unknown class for this function!"); + case cByte: + BMI(MBB, IP, X86::SHLir32, 2, DestReg).addReg(op0Reg).addZImm(Shift-1); + return; + case cShort: + BMI(MBB, IP, X86::SHLir32, 2, DestReg).addReg(op0Reg).addZImm(Shift-1); + return; + case cInt: + BMI(MBB, IP, X86::SHLir32, 2, DestReg).addReg(op0Reg).addZImm(Shift-1); + return; + } + } - // Emit the appropriate multiply instruction. - BMI(MBB, MBBI, MulOpcode[Class], 1).addReg (op1Reg); + if (Class == cShort) { + BMI(MBB, IP, X86::IMULri16, 2, DestReg).addReg(op0Reg).addZImm(ConstRHS); + return; + } else if (Class == cInt) { + BMI(MBB, IP, X86::IMULri32, 2, DestReg).addReg(op0Reg).addZImm(ConstRHS); + return; + } + + // Most general case, emit a normal multiply... + static const unsigned MOVirTab[] = { + X86::MOVir8, X86::MOVir16, X86::MOVir32 + }; - // Emit another MOV to put the result into the destination register. - BMI(MBB, MBBI, MovOpcode[Class], 1, destReg).addReg (Reg); + unsigned TmpReg = makeAnotherReg(DestTy); + BMI(MBB, IP, MOVirTab[Class], 1, TmpReg).addZImm(ConstRHS); + + // Emit a MUL to multiply the register holding the index by + // elementSize, putting the result in OffsetReg. + doMultiply(MBB, IP, DestReg, DestTy, op0Reg, TmpReg); } /// visitMul - Multiplies are not simple binary operators because they must deal /// with the EAX register explicitly. /// void ISel::visitMul(BinaryOperator &I) { - MachineBasicBlock::iterator MBBI = BB->end(); - doMultiply (getReg (I), I.getType (), - getReg (I.getOperand (0)), getReg (I.getOperand (1)), - BB, MBBI); + unsigned Op0Reg = getReg(I.getOperand(0)); + unsigned DestReg = getReg(I); + + // Simple scalar multiply? + if (I.getType() != Type::LongTy && I.getType() != Type::ULongTy) { + if (ConstantInt *CI = dyn_cast(I.getOperand(1))) { + unsigned Val = (unsigned)CI->getRawValue(); // Cannot be 64-bit constant + MachineBasicBlock::iterator MBBI = BB->end(); + doMultiplyConst(BB, MBBI, DestReg, I.getType(), Op0Reg, Val); + } else { + unsigned Op1Reg = getReg(I.getOperand(1)); + MachineBasicBlock::iterator MBBI = BB->end(); + doMultiply(BB, MBBI, DestReg, I.getType(), Op0Reg, Op1Reg); + } + } else { + unsigned Op1Reg = getReg(I.getOperand(1)); + + // Long value. We have to do things the hard way... + // Multiply the two low parts... capturing carry into EDX + BuildMI(BB, X86::MOVrr32, 1, X86::EAX).addReg(Op0Reg); + BuildMI(BB, X86::MULr32, 1).addReg(Op1Reg); // AL*BL + + unsigned OverflowReg = makeAnotherReg(Type::UIntTy); + BuildMI(BB, X86::MOVrr32, 1, DestReg).addReg(X86::EAX); // AL*BL + BuildMI(BB, X86::MOVrr32, 1, OverflowReg).addReg(X86::EDX); // AL*BL >> 32 + + MachineBasicBlock::iterator MBBI = BB->end(); + unsigned AHBLReg = makeAnotherReg(Type::UIntTy); // AH*BL + BMI(BB, MBBI, X86::IMULrr32, 2, AHBLReg).addReg(Op0Reg+1).addReg(Op1Reg); + + unsigned AHBLplusOverflowReg = makeAnotherReg(Type::UIntTy); + BuildMI(BB, X86::ADDrr32, 2, // AH*BL+(AL*BL >> 32) + AHBLplusOverflowReg).addReg(AHBLReg).addReg(OverflowReg); + + MBBI = BB->end(); + unsigned ALBHReg = makeAnotherReg(Type::UIntTy); // AL*BH + BMI(BB, MBBI, X86::IMULrr32, 2, ALBHReg).addReg(Op0Reg).addReg(Op1Reg+1); + + BuildMI(BB, X86::ADDrr32, 2, // AL*BH + AH*BL + (AL*BL >> 32) + DestReg+1).addReg(AHBLplusOverflowReg).addReg(ALBHReg); + } } @@ -693,45 +1339,81 @@ void ISel::visitMul(BinaryOperator &I) { /// void ISel::visitDivRem(BinaryOperator &I) { unsigned Class = getClass(I.getType()); - if (Class > 2) // FIXME: Handle longs - visitInstruction(I); + unsigned Op0Reg, Op1Reg, ResultReg = getReg(I); + + switch (Class) { + case cFP: // Floating point divide + if (I.getOpcode() == Instruction::Div) { + Op0Reg = getReg(I.getOperand(0)); + Op1Reg = getReg(I.getOperand(1)); + BuildMI(BB, X86::FpDIV, 2, ResultReg).addReg(Op0Reg).addReg(Op1Reg); + } else { // Floating point remainder... + MachineInstr *TheCall = + BuildMI(X86::CALLpcrel32, 1).addExternalSymbol("fmod", true); + std::vector Args; + Args.push_back(ValueRecord(I.getOperand(0))); + Args.push_back(ValueRecord(I.getOperand(1))); + doCall(ValueRecord(ResultReg, Type::DoubleTy), TheCall, Args); + } + return; + case cLong: { + static const char *FnName[] = + { "__moddi3", "__divdi3", "__umoddi3", "__udivdi3" }; + + unsigned NameIdx = I.getType()->isUnsigned()*2; + NameIdx += I.getOpcode() == Instruction::Div; + MachineInstr *TheCall = + BuildMI(X86::CALLpcrel32, 1).addExternalSymbol(FnName[NameIdx], true); + + std::vector Args; + Args.push_back(ValueRecord(I.getOperand(0))); + Args.push_back(ValueRecord(I.getOperand(1))); + doCall(ValueRecord(ResultReg, Type::LongTy), TheCall, Args); + return; + } + case cByte: case cShort: case cInt: + break; // Small integrals, handled below... + default: assert(0 && "Unknown class!"); + } static const unsigned Regs[] ={ X86::AL , X86::AX , X86::EAX }; static const unsigned MovOpcode[]={ X86::MOVrr8, X86::MOVrr16, X86::MOVrr32 }; - static const unsigned ExtOpcode[]={ X86::CBW , X86::CWD , X86::CDQ }; + static const unsigned SarOpcode[]={ X86::SARir8, X86::SARir16, X86::SARir32 }; static const unsigned ClrOpcode[]={ X86::XORrr8, X86::XORrr16, X86::XORrr32 }; static const unsigned ExtRegs[] ={ X86::AH , X86::DX , X86::EDX }; static const unsigned DivOpcode[][4] = { - { X86::DIVrr8 , X86::DIVrr16 , X86::DIVrr32 , 0 }, // Unsigned division - { X86::IDIVrr8, X86::IDIVrr16, X86::IDIVrr32, 0 }, // Signed division + { X86::DIVr8 , X86::DIVr16 , X86::DIVr32 , 0 }, // Unsigned division + { X86::IDIVr8, X86::IDIVr16, X86::IDIVr32, 0 }, // Signed division }; bool isSigned = I.getType()->isSigned(); unsigned Reg = Regs[Class]; unsigned ExtReg = ExtRegs[Class]; - unsigned Op0Reg = getReg(I.getOperand(0)); - unsigned Op1Reg = getReg(I.getOperand(1)); // Put the first operand into one of the A registers... + Op0Reg = getReg(I.getOperand(0)); BuildMI(BB, MovOpcode[Class], 1, Reg).addReg(Op0Reg); if (isSigned) { // Emit a sign extension instruction... - BuildMI(BB, ExtOpcode[Class], 0); + unsigned ShiftResult = makeAnotherReg(I.getType()); + BuildMI(BB, SarOpcode[Class], 2, ShiftResult).addReg(Op0Reg).addZImm(31); + BuildMI(BB, MovOpcode[Class], 1, ExtReg).addReg(ShiftResult); } else { // If unsigned, emit a zeroing instruction... (reg = xor reg, reg) BuildMI(BB, ClrOpcode[Class], 2, ExtReg).addReg(ExtReg).addReg(ExtReg); } // Emit the appropriate divide or remainder instruction... + Op1Reg = getReg(I.getOperand(1)); BuildMI(BB, DivOpcode[isSigned][Class], 1).addReg(Op1Reg); // Figure out which register we want to pick the result out of... unsigned DestReg = (I.getOpcode() == Instruction::Div) ? Reg : ExtReg; // Put the result into the destination register... - BuildMI(BB, MovOpcode[Class], 1, getReg(I)).addReg(DestReg); + BuildMI(BB, MovOpcode[Class], 1, ResultReg).addReg(DestReg); } @@ -740,190 +1422,536 @@ void ISel::visitDivRem(BinaryOperator &I) { /// shift values equal to 1. Even the general case is sort of special, /// because the shift amount has to be in CL, not just any old register. /// -void ISel::visitShiftInst (ShiftInst &I) { - unsigned Op0r = getReg (I.getOperand(0)); +void ISel::visitShiftInst(ShiftInst &I) { + unsigned SrcReg = getReg(I.getOperand(0)); unsigned DestReg = getReg(I); bool isLeftShift = I.getOpcode() == Instruction::Shl; - bool isOperandSigned = I.getType()->isUnsigned(); - unsigned OperandClass = getClass(I.getType()); - - if (OperandClass > 2) - visitInstruction(I); // Can't handle longs yet! - - if (ConstantUInt *CUI = dyn_cast (I.getOperand (1))) - { - // The shift amount is constant, guaranteed to be a ubyte. Get its value. - assert(CUI->getType() == Type::UByteTy && "Shift amount not a ubyte?"); - unsigned char shAmt = CUI->getValue(); - - static const unsigned ConstantOperand[][4] = { - { X86::SHRir8, X86::SHRir16, X86::SHRir32, 0 }, // SHR - { X86::SARir8, X86::SARir16, X86::SARir32, 0 }, // SAR - { X86::SHLir8, X86::SHLir16, X86::SHLir32, 0 }, // SHL - { X86::SHLir8, X86::SHLir16, X86::SHLir32, 0 }, // SAL = SHL - }; + bool isSigned = I.getType()->isSigned(); + unsigned Class = getClass(I.getType()); + + static const unsigned ConstantOperand[][4] = { + { X86::SHRir8, X86::SHRir16, X86::SHRir32, X86::SHRDir32 }, // SHR + { X86::SARir8, X86::SARir16, X86::SARir32, X86::SHRDir32 }, // SAR + { X86::SHLir8, X86::SHLir16, X86::SHLir32, X86::SHLDir32 }, // SHL + { X86::SHLir8, X86::SHLir16, X86::SHLir32, X86::SHLDir32 }, // SAL = SHL + }; - const unsigned *OpTab = // Figure out the operand table to use - ConstantOperand[isLeftShift*2+isOperandSigned]; + static const unsigned NonConstantOperand[][4] = { + { X86::SHRrr8, X86::SHRrr16, X86::SHRrr32 }, // SHR + { X86::SARrr8, X86::SARrr16, X86::SARrr32 }, // SAR + { X86::SHLrr8, X86::SHLrr16, X86::SHLrr32 }, // SHL + { X86::SHLrr8, X86::SHLrr16, X86::SHLrr32 }, // SAL = SHL + }; + + // Longs, as usual, are handled specially... + if (Class == cLong) { + // If we have a constant shift, we can generate much more efficient code + // than otherwise... + // + if (ConstantUInt *CUI = dyn_cast(I.getOperand(1))) { + unsigned Amount = CUI->getValue(); + if (Amount < 32) { + const unsigned *Opc = ConstantOperand[isLeftShift*2+isSigned]; + if (isLeftShift) { + BuildMI(BB, Opc[3], 3, + DestReg+1).addReg(SrcReg+1).addReg(SrcReg).addZImm(Amount); + BuildMI(BB, Opc[2], 2, DestReg).addReg(SrcReg).addZImm(Amount); + } else { + BuildMI(BB, Opc[3], 3, + DestReg).addReg(SrcReg ).addReg(SrcReg+1).addZImm(Amount); + BuildMI(BB, Opc[2], 2, DestReg+1).addReg(SrcReg+1).addZImm(Amount); + } + } else { // Shifting more than 32 bits + Amount -= 32; + if (isLeftShift) { + BuildMI(BB, X86::SHLir32, 2,DestReg+1).addReg(SrcReg).addZImm(Amount); + BuildMI(BB, X86::MOVir32, 1,DestReg ).addZImm(0); + } else { + unsigned Opcode = isSigned ? X86::SARir32 : X86::SHRir32; + BuildMI(BB, Opcode, 2, DestReg).addReg(SrcReg+1).addZImm(Amount); + BuildMI(BB, X86::MOVir32, 1, DestReg+1).addZImm(0); + } + } + } else { + unsigned TmpReg = makeAnotherReg(Type::IntTy); + + if (!isLeftShift && isSigned) { + // If this is a SHR of a Long, then we need to do funny sign extension + // stuff. TmpReg gets the value to use as the high-part if we are + // shifting more than 32 bits. + BuildMI(BB, X86::SARir32, 2, TmpReg).addReg(SrcReg).addZImm(31); + } else { + // Other shifts use a fixed zero value if the shift is more than 32 + // bits. + BuildMI(BB, X86::MOVir32, 1, TmpReg).addZImm(0); + } - // Emit: reg, shamt (shift-by-immediate opcode "ir" form.) - BuildMI(BB, OpTab[OperandClass], 2, DestReg).addReg(Op0r).addZImm(shAmt); + // Initialize CL with the shift amount... + unsigned ShiftAmount = getReg(I.getOperand(1)); + BuildMI(BB, X86::MOVrr8, 1, X86::CL).addReg(ShiftAmount); + + unsigned TmpReg2 = makeAnotherReg(Type::IntTy); + unsigned TmpReg3 = makeAnotherReg(Type::IntTy); + if (isLeftShift) { + // TmpReg2 = shld inHi, inLo + BuildMI(BB, X86::SHLDrr32, 2, TmpReg2).addReg(SrcReg+1).addReg(SrcReg); + // TmpReg3 = shl inLo, CL + BuildMI(BB, X86::SHLrr32, 1, TmpReg3).addReg(SrcReg); + + // Set the flags to indicate whether the shift was by more than 32 bits. + BuildMI(BB, X86::TESTri8, 2).addReg(X86::CL).addZImm(32); + + // DestHi = (>32) ? TmpReg3 : TmpReg2; + BuildMI(BB, X86::CMOVNErr32, 2, + DestReg+1).addReg(TmpReg2).addReg(TmpReg3); + // DestLo = (>32) ? TmpReg : TmpReg3; + BuildMI(BB, X86::CMOVNErr32, 2, DestReg).addReg(TmpReg3).addReg(TmpReg); + } else { + // TmpReg2 = shrd inLo, inHi + BuildMI(BB, X86::SHRDrr32, 2, TmpReg2).addReg(SrcReg).addReg(SrcReg+1); + // TmpReg3 = s[ah]r inHi, CL + BuildMI(BB, isSigned ? X86::SARrr32 : X86::SHRrr32, 1, TmpReg3) + .addReg(SrcReg+1); + + // Set the flags to indicate whether the shift was by more than 32 bits. + BuildMI(BB, X86::TESTri8, 2).addReg(X86::CL).addZImm(32); + + // DestLo = (>32) ? TmpReg3 : TmpReg2; + BuildMI(BB, X86::CMOVNErr32, 2, + DestReg).addReg(TmpReg2).addReg(TmpReg3); + + // DestHi = (>32) ? TmpReg : TmpReg3; + BuildMI(BB, X86::CMOVNErr32, 2, + DestReg+1).addReg(TmpReg3).addReg(TmpReg); + } } - else - { - // The shift amount is non-constant. - // - // In fact, you can only shift with a variable shift amount if - // that amount is already in the CL register, so we have to put it - // there first. - // + return; + } - // Emit: move cl, shiftAmount (put the shift amount in CL.) - BuildMI(BB, X86::MOVrr8, 1, X86::CL).addReg(getReg(I.getOperand(1))); + if (ConstantUInt *CUI = dyn_cast(I.getOperand(1))) { + // The shift amount is constant, guaranteed to be a ubyte. Get its value. + assert(CUI->getType() == Type::UByteTy && "Shift amount not a ubyte?"); - // This is a shift right (SHR). - static const unsigned NonConstantOperand[][4] = { - { X86::SHRrr8, X86::SHRrr16, X86::SHRrr32, 0 }, // SHR - { X86::SARrr8, X86::SARrr16, X86::SARrr32, 0 }, // SAR - { X86::SHLrr8, X86::SHLrr16, X86::SHLrr32, 0 }, // SHL - { X86::SHLrr8, X86::SHLrr16, X86::SHLrr32, 0 }, // SAL = SHL - }; + const unsigned *Opc = ConstantOperand[isLeftShift*2+isSigned]; + BuildMI(BB, Opc[Class], 2, DestReg).addReg(SrcReg).addZImm(CUI->getValue()); + } else { // The shift amount is non-constant. + BuildMI(BB, X86::MOVrr8, 1, X86::CL).addReg(getReg(I.getOperand(1))); - const unsigned *OpTab = // Figure out the operand table to use - NonConstantOperand[isLeftShift*2+isOperandSigned]; + const unsigned *Opc = NonConstantOperand[isLeftShift*2+isSigned]; + BuildMI(BB, Opc[Class], 1, DestReg).addReg(SrcReg); + } +} - BuildMI(BB, OpTab[OperandClass], 1, DestReg).addReg(Op0r); - } + +/// EmitByteSwap - Byteswap SrcReg into DestReg. +/// +void ISel::EmitByteSwap(unsigned DestReg, unsigned SrcReg, unsigned Class) { + // Emit the byte swap instruction... + switch (Class) { + case cByte: + // No byteswap necessary for 8 bit value... + BuildMI(BB, X86::MOVrr8, 1, DestReg).addReg(SrcReg); + break; + case cInt: + // Use the 32 bit bswap instruction to do a 32 bit swap... + BuildMI(BB, X86::BSWAPr32, 1, DestReg).addReg(SrcReg); + break; + + case cShort: + // For 16 bit we have to use an xchg instruction, because there is no + // 16-bit bswap. XCHG is necessarily not in SSA form, so we force things + // into AX to do the xchg. + // + BuildMI(BB, X86::MOVrr16, 1, X86::AX).addReg(SrcReg); + BuildMI(BB, X86::XCHGrr8, 2).addReg(X86::AL, MOTy::UseAndDef) + .addReg(X86::AH, MOTy::UseAndDef); + BuildMI(BB, X86::MOVrr16, 1, DestReg).addReg(X86::AX); + break; + default: assert(0 && "Cannot byteswap this class!"); + } } /// visitLoadInst - Implement LLVM load instructions in terms of the x86 'mov' -/// instruction. +/// instruction. The load and store instructions are the only place where we +/// need to worry about the memory layout of the target machine. /// void ISel::visitLoadInst(LoadInst &I) { - unsigned Class = getClass(I.getType()); - if (Class > 2) // FIXME: Handle longs and others... - visitInstruction(I); + unsigned SrcAddrReg = getReg(I.getOperand(0)); + unsigned DestReg = getReg(I); - static const unsigned Opcode[] = { X86::MOVmr8, X86::MOVmr16, X86::MOVmr32 }; + unsigned Class = getClassB(I.getType()); - unsigned AddressReg = getReg(I.getOperand(0)); - addDirectMem(BuildMI(BB, Opcode[Class], 4, getReg(I)), AddressReg); -} + if (Class == cLong) { + addDirectMem(BuildMI(BB, X86::MOVmr32, 4, DestReg), SrcAddrReg); + addRegOffset(BuildMI(BB, X86::MOVmr32, 4, DestReg+1), SrcAddrReg, 4); + return; + } + static const unsigned Opcodes[] = { + X86::MOVmr8, X86::MOVmr16, X86::MOVmr32, X86::FLDr32 + }; + unsigned Opcode = Opcodes[Class]; + if (I.getType() == Type::DoubleTy) Opcode = X86::FLDr64; + addDirectMem(BuildMI(BB, Opcode, 4, DestReg), SrcAddrReg); +} /// visitStoreInst - Implement LLVM store instructions in terms of the x86 'mov' /// instruction. /// void ISel::visitStoreInst(StoreInst &I) { - unsigned Class = getClass(I.getOperand(0)->getType()); - if (Class > 2) // FIXME: Handle longs and others... - visitInstruction(I); + unsigned ValReg = getReg(I.getOperand(0)); + unsigned AddressReg = getReg(I.getOperand(1)); + + const Type *ValTy = I.getOperand(0)->getType(); + unsigned Class = getClassB(ValTy); - static const unsigned Opcode[] = { X86::MOVrm8, X86::MOVrm16, X86::MOVrm32 }; + if (Class == cLong) { + addDirectMem(BuildMI(BB, X86::MOVrm32, 1+4), AddressReg).addReg(ValReg); + addRegOffset(BuildMI(BB, X86::MOVrm32, 1+4), AddressReg,4).addReg(ValReg+1); + return; + } - unsigned ValReg = getReg(I.getOperand(0)); - unsigned AddressReg = getReg(I.getOperand(1)); - addDirectMem(BuildMI(BB, Opcode[Class], 1+4), AddressReg).addReg(ValReg); + static const unsigned Opcodes[] = { + X86::MOVrm8, X86::MOVrm16, X86::MOVrm32, X86::FSTr32 + }; + unsigned Opcode = Opcodes[Class]; + if (ValTy == Type::DoubleTy) Opcode = X86::FSTr64; + addDirectMem(BuildMI(BB, Opcode, 1+4), AddressReg).addReg(ValReg); } /// visitCastInst - Here we have various kinds of copying with or without /// sign extension going on. -void -ISel::visitCastInst (CastInst &CI) -{ - const Type *targetType = CI.getType (); - Value *operand = CI.getOperand (0); - unsigned int operandReg = getReg (operand); - const Type *sourceType = operand->getType (); - unsigned int destReg = getReg (CI); - // - // Currently we handle: - // - // 1) cast * to bool - // - // 2) cast {sbyte, ubyte} to {sbyte, ubyte} - // cast {short, ushort} to {ushort, short} - // cast {int, uint, ptr} to {int, uint, ptr} - // - // 3) cast {sbyte, ubyte} to {ushort, short} - // cast {sbyte, ubyte} to {int, uint, ptr} - // cast {short, ushort} to {int, uint, ptr} - // - // 4) cast {int, uint, ptr} to {short, ushort} - // cast {int, uint, ptr} to {sbyte, ubyte} - // cast {short, ushort} to {sbyte, ubyte} - - // 1) Implement casts to bool by using compare on the operand followed - // by set if not zero on the result. - if (targetType == Type::BoolTy) - { - BuildMI (BB, X86::CMPri8, 2).addReg (operandReg).addZImm (0); - BuildMI (BB, X86::SETNEr, 1, destReg); - return; +void ISel::visitCastInst(CastInst &CI) { + Value *Op = CI.getOperand(0); + // If this is a cast from a 32-bit integer to a Long type, and the only uses + // of the case are GEP instructions, then the cast does not need to be + // generated explicitly, it will be folded into the GEP. + if (CI.getType() == Type::LongTy && + (Op->getType() == Type::IntTy || Op->getType() == Type::UIntTy)) { + bool AllUsesAreGEPs = true; + for (Value::use_iterator I = CI.use_begin(), E = CI.use_end(); I != E; ++I) + if (!isa(*I)) { + AllUsesAreGEPs = false; + break; + } + + // No need to codegen this cast if all users are getelementptr instrs... + if (AllUsesAreGEPs) return; + } + + unsigned DestReg = getReg(CI); + MachineBasicBlock::iterator MI = BB->end(); + emitCastOperation(BB, MI, Op, CI.getType(), DestReg); +} + +/// emitCastOperation - Common code shared between visitCastInst and +/// constant expression cast support. +void ISel::emitCastOperation(MachineBasicBlock *BB, + MachineBasicBlock::iterator &IP, + Value *Src, const Type *DestTy, + unsigned DestReg) { + unsigned SrcReg = getReg(Src, BB, IP); + const Type *SrcTy = Src->getType(); + unsigned SrcClass = getClassB(SrcTy); + unsigned DestClass = getClassB(DestTy); + + // Implement casts to bool by using compare on the operand followed by set if + // not zero on the result. + if (DestTy == Type::BoolTy) { + switch (SrcClass) { + case cByte: + BMI(BB, IP, X86::TESTrr8, 2).addReg(SrcReg).addReg(SrcReg); + break; + case cShort: + BMI(BB, IP, X86::TESTrr16, 2).addReg(SrcReg).addReg(SrcReg); + break; + case cInt: + BMI(BB, IP, X86::TESTrr32, 2).addReg(SrcReg).addReg(SrcReg); + break; + case cLong: { + unsigned TmpReg = makeAnotherReg(Type::IntTy); + BMI(BB, IP, X86::ORrr32, 2, TmpReg).addReg(SrcReg).addReg(SrcReg+1); + break; + } + case cFP: + assert(0 && "FIXME: implement cast FP to bool"); + abort(); } - // 2) Implement casts between values of the same type class (as determined - // by getClass) by using a register-to-register move. - unsigned srcClass = sourceType == Type::BoolTy ? cByte : getClass(sourceType); - unsigned targClass = getClass (targetType); - static const unsigned regRegMove[] = { - X86::MOVrr8, X86::MOVrr16, X86::MOVrr32 + // If the zero flag is not set, then the value is true, set the byte to + // true. + BMI(BB, IP, X86::SETNEr, 1, DestReg); + return; + } + + static const unsigned RegRegMove[] = { + X86::MOVrr8, X86::MOVrr16, X86::MOVrr32, X86::FpMOV, X86::MOVrr32 }; - if ((srcClass < 3) && (targClass < 3) && (srcClass == targClass)) - { - BuildMI (BB, regRegMove[srcClass], 1, destReg).addReg (operandReg); - return; + + // Implement casts between values of the same type class (as determined by + // getClass) by using a register-to-register move. + if (SrcClass == DestClass) { + if (SrcClass <= cInt || (SrcClass == cFP && SrcTy == DestTy)) { + BMI(BB, IP, RegRegMove[SrcClass], 1, DestReg).addReg(SrcReg); + } else if (SrcClass == cFP) { + if (SrcTy == Type::FloatTy) { // double -> float + assert(DestTy == Type::DoubleTy && "Unknown cFP member!"); + BMI(BB, IP, X86::FpMOV, 1, DestReg).addReg(SrcReg); + } else { // float -> double + assert(SrcTy == Type::DoubleTy && DestTy == Type::FloatTy && + "Unknown cFP member!"); + // Truncate from double to float by storing to memory as short, then + // reading it back. + unsigned FltAlign = TM.getTargetData().getFloatAlignment(); + int FrameIdx = F->getFrameInfo()->CreateStackObject(4, FltAlign); + addFrameReference(BMI(BB, IP, X86::FSTr32, 5), FrameIdx).addReg(SrcReg); + addFrameReference(BMI(BB, IP, X86::FLDr32, 5, DestReg), FrameIdx); + } + } else if (SrcClass == cLong) { + BMI(BB, IP, X86::MOVrr32, 1, DestReg).addReg(SrcReg); + BMI(BB, IP, X86::MOVrr32, 1, DestReg+1).addReg(SrcReg+1); + } else { + assert(0 && "Cannot handle this type of cast instruction!"); + abort(); } - // 3) Handle cast of SMALLER int to LARGER int using a move with sign - // extension or zero extension, depending on whether the source type - // was signed. - if ((srcClass < 3) && (targClass < 3) && (srcClass < targClass)) - { - static const unsigned ops[] = { - X86::MOVSXr16r8, X86::MOVSXr32r8, X86::MOVSXr32r16, - X86::MOVZXr16r8, X86::MOVZXr32r8, X86::MOVZXr32r16 - }; - unsigned srcSigned = sourceType->isSigned (); - BuildMI (BB, ops[3 * srcSigned + srcClass + targClass - 1], 1, - destReg).addReg (operandReg); - return; + return; + } + + // Handle cast of SMALLER int to LARGER int using a move with sign extension + // or zero extension, depending on whether the source type was signed. + if (SrcClass <= cInt && (DestClass <= cInt || DestClass == cLong) && + SrcClass < DestClass) { + bool isLong = DestClass == cLong; + if (isLong) DestClass = cInt; + + static const unsigned Opc[][4] = { + { X86::MOVSXr16r8, X86::MOVSXr32r8, X86::MOVSXr32r16, X86::MOVrr32 }, // s + { X86::MOVZXr16r8, X86::MOVZXr32r8, X86::MOVZXr32r16, X86::MOVrr32 } // u + }; + + bool isUnsigned = SrcTy->isUnsigned(); + BMI(BB, IP, Opc[isUnsigned][SrcClass + DestClass - 1], 1, + DestReg).addReg(SrcReg); + + if (isLong) { // Handle upper 32 bits as appropriate... + if (isUnsigned) // Zero out top bits... + BMI(BB, IP, X86::MOVir32, 1, DestReg+1).addZImm(0); + else // Sign extend bottom half... + BMI(BB, IP, X86::SARir32, 2, DestReg+1).addReg(DestReg).addZImm(31); } - // 4) Handle cast of LARGER int to SMALLER int using a move to EAX - // followed by a move out of AX or AL. - if ((srcClass < 3) && (targClass < 3) && (srcClass > targClass)) - { - static const unsigned AReg[] = { X86::AL, X86::AX, X86::EAX }; - BuildMI (BB, regRegMove[srcClass], 1, - AReg[srcClass]).addReg (operandReg); - BuildMI (BB, regRegMove[targClass], 1, destReg).addReg (AReg[srcClass]); - return; + return; + } + + // Special case long -> int ... + if (SrcClass == cLong && DestClass == cInt) { + BMI(BB, IP, X86::MOVrr32, 1, DestReg).addReg(SrcReg); + return; + } + + // Handle cast of LARGER int to SMALLER int using a move to EAX followed by a + // move out of AX or AL. + if ((SrcClass <= cInt || SrcClass == cLong) && DestClass <= cInt + && SrcClass > DestClass) { + static const unsigned AReg[] = { X86::AL, X86::AX, X86::EAX, 0, X86::EAX }; + BMI(BB, IP, RegRegMove[SrcClass], 1, AReg[SrcClass]).addReg(SrcReg); + BMI(BB, IP, RegRegMove[DestClass], 1, DestReg).addReg(AReg[DestClass]); + return; + } + + // Handle casts from integer to floating point now... + if (DestClass == cFP) { + // Promote the integer to a type supported by FLD. We do this because there + // are no unsigned FLD instructions, so we must promote an unsigned value to + // a larger signed value, then use FLD on the larger value. + // + const Type *PromoteType = 0; + unsigned PromoteOpcode; + switch (SrcTy->getPrimitiveID()) { + case Type::BoolTyID: + case Type::SByteTyID: + // We don't have the facilities for directly loading byte sized data from + // memory (even signed). Promote it to 16 bits. + PromoteType = Type::ShortTy; + PromoteOpcode = X86::MOVSXr16r8; + break; + case Type::UByteTyID: + PromoteType = Type::ShortTy; + PromoteOpcode = X86::MOVZXr16r8; + break; + case Type::UShortTyID: + PromoteType = Type::IntTy; + PromoteOpcode = X86::MOVZXr32r16; + break; + case Type::UIntTyID: { + // Make a 64 bit temporary... and zero out the top of it... + unsigned TmpReg = makeAnotherReg(Type::LongTy); + BMI(BB, IP, X86::MOVrr32, 1, TmpReg).addReg(SrcReg); + BMI(BB, IP, X86::MOVir32, 1, TmpReg+1).addZImm(0); + SrcTy = Type::LongTy; + SrcClass = cLong; + SrcReg = TmpReg; + break; } + case Type::ULongTyID: + assert("FIXME: not implemented: cast ulong X to fp type!"); + default: // No promotion needed... + break; + } + + if (PromoteType) { + unsigned TmpReg = makeAnotherReg(PromoteType); + BMI(BB, IP, SrcTy->isSigned() ? X86::MOVSXr16r8 : X86::MOVZXr16r8, + 1, TmpReg).addReg(SrcReg); + SrcTy = PromoteType; + SrcClass = getClass(PromoteType); + SrcReg = TmpReg; + } + + // Spill the integer to memory and reload it from there... + int FrameIdx = + F->getFrameInfo()->CreateStackObject(SrcTy, TM.getTargetData()); + + if (SrcClass == cLong) { + addFrameReference(BMI(BB, IP, X86::MOVrm32, 5), FrameIdx).addReg(SrcReg); + addFrameReference(BMI(BB, IP, X86::MOVrm32, 5), + FrameIdx, 4).addReg(SrcReg+1); + } else { + static const unsigned Op1[] = { X86::MOVrm8, X86::MOVrm16, X86::MOVrm32 }; + addFrameReference(BMI(BB, IP, Op1[SrcClass], 5), FrameIdx).addReg(SrcReg); + } + + static const unsigned Op2[] = + { 0/*byte*/, X86::FILDr16, X86::FILDr32, 0/*FP*/, X86::FILDr64 }; + addFrameReference(BMI(BB, IP, Op2[SrcClass], 5, DestReg), FrameIdx); + return; + } + + // Handle casts from floating point to integer now... + if (SrcClass == cFP) { + // Change the floating point control register to use "round towards zero" + // mode when truncating to an integer value. + // + int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2); + addFrameReference(BMI(BB, IP, X86::FNSTCWm16, 4), CWFrameIdx); + + // Load the old value of the high byte of the control word... + unsigned HighPartOfCW = makeAnotherReg(Type::UByteTy); + addFrameReference(BMI(BB, IP, X86::MOVmr8, 4, HighPartOfCW), CWFrameIdx, 1); + + // Set the high part to be round to zero... + addFrameReference(BMI(BB, IP, X86::MOVim8, 5), CWFrameIdx, 1).addZImm(12); + + // Reload the modified control word now... + addFrameReference(BMI(BB, IP, X86::FLDCWm16, 4), CWFrameIdx); + + // Restore the memory image of control word to original value + addFrameReference(BMI(BB, IP, X86::MOVrm8, 5), + CWFrameIdx, 1).addReg(HighPartOfCW); + + // We don't have the facilities for directly storing byte sized data to + // memory. Promote it to 16 bits. We also must promote unsigned values to + // larger classes because we only have signed FP stores. + unsigned StoreClass = DestClass; + const Type *StoreTy = DestTy; + if (StoreClass == cByte || DestTy->isUnsigned()) + switch (StoreClass) { + case cByte: StoreTy = Type::ShortTy; StoreClass = cShort; break; + case cShort: StoreTy = Type::IntTy; StoreClass = cInt; break; + case cInt: StoreTy = Type::LongTy; StoreClass = cLong; break; + // The following treatment of cLong may not be perfectly right, + // but it survives chains of casts of the form + // double->ulong->double. + case cLong: StoreTy = Type::LongTy; StoreClass = cLong; break; + default: assert(0 && "Unknown store class!"); + } + + // Spill the integer to memory and reload it from there... + int FrameIdx = + F->getFrameInfo()->CreateStackObject(StoreTy, TM.getTargetData()); + + static const unsigned Op1[] = + { 0, X86::FISTr16, X86::FISTr32, 0, X86::FISTPr64 }; + addFrameReference(BMI(BB, IP, Op1[StoreClass], 5), FrameIdx).addReg(SrcReg); + + if (DestClass == cLong) { + addFrameReference(BMI(BB, IP, X86::MOVmr32, 4, DestReg), FrameIdx); + addFrameReference(BMI(BB, IP, X86::MOVmr32, 4, DestReg+1), FrameIdx, 4); + } else { + static const unsigned Op2[] = { X86::MOVmr8, X86::MOVmr16, X86::MOVmr32 }; + addFrameReference(BMI(BB, IP, Op2[DestClass], 4, DestReg), FrameIdx); + } + + // Reload the original control word now... + addFrameReference(BMI(BB, IP, X86::FLDCWm16, 4), CWFrameIdx); + return; + } + // Anything we haven't handled already, we can't (yet) handle at all. - // - // FP to integral casts can be handled with FISTP to store onto the - // stack while converting to integer, followed by a MOV to load from - // the stack into the result register. Integral to FP casts can be - // handled with MOV to store onto the stack, followed by a FILD to - // load from the stack while converting to FP. For the moment, I - // can't quite get straight in my head how to borrow myself some - // stack space and write on it. Otherwise, this would be trivial. - visitInstruction (CI); + assert(0 && "Unhandled cast instruction!"); + abort(); +} + +/// visitVANextInst - Implement the va_next instruction... +/// +void ISel::visitVANextInst(VANextInst &I) { + unsigned VAList = getReg(I.getOperand(0)); + unsigned DestReg = getReg(I); + + unsigned Size; + switch (I.getArgType()->getPrimitiveID()) { + default: + std::cerr << I; + assert(0 && "Error: bad type for va_next instruction!"); + return; + case Type::PointerTyID: + case Type::UIntTyID: + case Type::IntTyID: + Size = 4; + break; + case Type::ULongTyID: + case Type::LongTyID: + case Type::DoubleTyID: + Size = 8; + break; + } + + // Increment the VAList pointer... + BuildMI(BB, X86::ADDri32, 2, DestReg).addReg(VAList).addZImm(Size); +} + +void ISel::visitVAArgInst(VAArgInst &I) { + unsigned VAList = getReg(I.getOperand(0)); + unsigned DestReg = getReg(I); + + switch (I.getType()->getPrimitiveID()) { + default: + std::cerr << I; + assert(0 && "Error: bad type for va_next instruction!"); + return; + case Type::PointerTyID: + case Type::UIntTyID: + case Type::IntTyID: + addDirectMem(BuildMI(BB, X86::MOVmr32, 4, DestReg), VAList); + break; + case Type::ULongTyID: + case Type::LongTyID: + addDirectMem(BuildMI(BB, X86::MOVmr32, 4, DestReg), VAList); + addRegOffset(BuildMI(BB, X86::MOVmr32, 4, DestReg+1), VAList, 4); + break; + case Type::DoubleTyID: + addDirectMem(BuildMI(BB, X86::FLDr64, 4, DestReg), VAList); + break; + } } -/// visitGetElementPtrInst - I don't know, most programs don't have -/// getelementptr instructions, right? That means we can put off -/// implementing this, right? Right. This method emits machine -/// instructions to perform type-safe pointer arithmetic. I am -/// guessing this could be cleaned up somewhat to use fewer temporary -/// registers. -void -ISel::visitGetElementPtrInst (GetElementPtrInst &I) -{ + +void ISel::visitGetElementPtrInst(GetElementPtrInst &I) { + unsigned outputReg = getReg(I); MachineBasicBlock::iterator MI = BB->end(); emitGEPOperation(BB, MI, I.getOperand(0), - I.op_begin()+1, I.op_end(), getReg(I)); + I.op_begin()+1, I.op_end(), outputReg); } void ISel::emitGEPOperation(MachineBasicBlock *MBB, @@ -932,122 +1960,183 @@ void ISel::emitGEPOperation(MachineBasicBlock *MBB, User::op_iterator IdxEnd, unsigned TargetReg) { const TargetData &TD = TM.getTargetData(); const Type *Ty = Src->getType(); - unsigned basePtrReg = getReg(Src, BB, IP); + unsigned BaseReg = getReg(Src, MBB, IP); // GEPs have zero or more indices; we must perform a struct access // or array access for each one. for (GetElementPtrInst::op_iterator oi = IdxBegin, oe = IdxEnd; oi != oe; ++oi) { Value *idx = *oi; - unsigned nextBasePtrReg = makeAnotherReg(Type::UIntTy); - if (const StructType *StTy = dyn_cast (Ty)) { + unsigned NextReg = BaseReg; + if (const StructType *StTy = dyn_cast(Ty)) { // It's a struct access. idx is the index into the structure, // which names the field. This index must have ubyte type. - const ConstantUInt *CUI = cast (idx); - assert (CUI->getType () == Type::UByteTy + const ConstantUInt *CUI = cast(idx); + assert(CUI->getType() == Type::UByteTy && "Funny-looking structure index in GEP"); // Use the TargetData structure to pick out what the layout of // the structure is in memory. Since the structure index must // be constant, we can get its value and use it to find the // right byte offset from the StructLayout class's list of // structure member offsets. - unsigned idxValue = CUI->getValue (); - unsigned memberOffset = - TD.getStructLayout (StTy)->MemberOffsets[idxValue]; - // Emit an ADD to add memberOffset to the basePtr. - BMI(MBB, IP, X86::ADDri32, 2, - nextBasePtrReg).addReg (basePtrReg).addZImm (memberOffset); + unsigned idxValue = CUI->getValue(); + unsigned FieldOff = TD.getStructLayout(StTy)->MemberOffsets[idxValue]; + if (FieldOff) { + NextReg = makeAnotherReg(Type::UIntTy); + // Emit an ADD to add FieldOff to the basePtr. + BMI(MBB, IP, X86::ADDri32, 2,NextReg).addReg(BaseReg).addZImm(FieldOff); + } // The next type is the member of the structure selected by the // index. - Ty = StTy->getElementTypes ()[idxValue]; - } else if (const SequentialType *SqTy = cast (Ty)) { + Ty = StTy->getElementTypes()[idxValue]; + } else if (const SequentialType *SqTy = cast(Ty)) { // It's an array or pointer access: [ArraySize x ElementType]. - const Type *typeOfSequentialTypeIndex = SqTy->getIndexType (); + // idx is the index into the array. Unlike with structure // indices, we may not know its actual value at code-generation // time. - assert (idx->getType () == typeOfSequentialTypeIndex - && "Funny-looking array index in GEP"); - // We want to add basePtrReg to (idxReg * sizeof - // ElementType). First, we must find the size of the pointed-to - // type. (Not coincidentally, the next type is the type of the - // elements in the array.) - Ty = SqTy->getElementType (); - unsigned elementSize = TD.getTypeSize (Ty); - unsigned elementSizeReg = makeAnotherReg(typeOfSequentialTypeIndex); - copyConstantToRegister(ConstantSInt::get(typeOfSequentialTypeIndex, - elementSize), elementSizeReg, - BB, IP); - - unsigned idxReg = getReg(idx, BB, IP); - // Emit a MUL to multiply the register holding the index by - // elementSize, putting the result in memberOffsetReg. - unsigned memberOffsetReg = makeAnotherReg(Type::UIntTy); - doMultiply (memberOffsetReg, typeOfSequentialTypeIndex, - elementSizeReg, idxReg, BB, IP); - // Emit an ADD to add memberOffsetReg to the basePtr. - BMI(MBB, IP, X86::ADDrr32, 2, - nextBasePtrReg).addReg (basePtrReg).addReg (memberOffsetReg); + assert(idx->getType() == Type::LongTy && "Bad GEP array index!"); + + // Most GEP instructions use a [cast (int/uint) to LongTy] as their + // operand on X86. Handle this case directly now... + if (CastInst *CI = dyn_cast(idx)) + if (CI->getOperand(0)->getType() == Type::IntTy || + CI->getOperand(0)->getType() == Type::UIntTy) + idx = CI->getOperand(0); + + // We want to add BaseReg to(idxReg * sizeof ElementType). First, we + // must find the size of the pointed-to type (Not coincidentally, the next + // type is the type of the elements in the array). + Ty = SqTy->getElementType(); + unsigned elementSize = TD.getTypeSize(Ty); + + // If idxReg is a constant, we don't need to perform the multiply! + if (ConstantSInt *CSI = dyn_cast(idx)) { + if (!CSI->isNullValue()) { + unsigned Offset = elementSize*CSI->getValue(); + NextReg = makeAnotherReg(Type::UIntTy); + BMI(MBB, IP, X86::ADDri32, 2,NextReg).addReg(BaseReg).addZImm(Offset); + } + } else if (elementSize == 1) { + // If the element size is 1, we don't have to multiply, just add + unsigned idxReg = getReg(idx, MBB, IP); + NextReg = makeAnotherReg(Type::UIntTy); + BMI(MBB, IP, X86::ADDrr32, 2, NextReg).addReg(BaseReg).addReg(idxReg); + } else { + unsigned idxReg = getReg(idx, MBB, IP); + unsigned OffsetReg = makeAnotherReg(Type::UIntTy); + + doMultiplyConst(MBB, IP, OffsetReg, Type::IntTy, idxReg, elementSize); + + // Emit an ADD to add OffsetReg to the basePtr. + NextReg = makeAnotherReg(Type::UIntTy); + BMI(MBB, IP, X86::ADDrr32, 2,NextReg).addReg(BaseReg).addReg(OffsetReg); + } } // Now that we are here, further indices refer to subtypes of this - // one, so we don't need to worry about basePtrReg itself, anymore. - basePtrReg = nextBasePtrReg; + // one, so we don't need to worry about BaseReg itself, anymore. + BaseReg = NextReg; } // After we have processed all the indices, the result is left in - // basePtrReg. Move it to the register where we were expected to + // BaseReg. Move it to the register where we were expected to // put the answer. A 32-bit move should do it, because we are in // ILP32 land. - BMI(MBB, IP, X86::MOVrr32, 1, TargetReg).addReg (basePtrReg); + BMI(MBB, IP, X86::MOVrr32, 1, TargetReg).addReg(BaseReg); } -/// visitMallocInst - I know that personally, whenever I want to remember -/// something, I have to clear off some space in my brain. -void -ISel::visitMallocInst (MallocInst &I) -{ - // We assume that by this point, malloc instructions have been - // lowered to calls, and dlsym will magically find malloc for us. - // So we do not want to see malloc instructions here. - visitInstruction (I); -} - +/// visitAllocaInst - If this is a fixed size alloca, allocate space from the +/// frame manager, otherwise do it the hard way. +/// +void ISel::visitAllocaInst(AllocaInst &I) { + // Find the data size of the alloca inst's getAllocatedType. + const Type *Ty = I.getAllocatedType(); + unsigned TySize = TM.getTargetData().getTypeSize(Ty); -/// visitFreeInst - same story as MallocInst -void -ISel::visitFreeInst (FreeInst &I) -{ - // We assume that by this point, free instructions have been - // lowered to calls, and dlsym will magically find free for us. - // So we do not want to see free instructions here. - visitInstruction (I); -} + // If this is a fixed size alloca in the entry block for the function, + // statically stack allocate the space. + // + if (ConstantUInt *CUI = dyn_cast(I.getArraySize())) { + if (I.getParent() == I.getParent()->getParent()->begin()) { + TySize *= CUI->getValue(); // Get total allocated size... + unsigned Alignment = TM.getTargetData().getTypeAlignment(Ty); + + // Create a new stack object using the frame manager... + int FrameIdx = F->getFrameInfo()->CreateStackObject(TySize, Alignment); + addFrameReference(BuildMI(BB, X86::LEAr32, 5, getReg(I)), FrameIdx); + return; + } + } + + // Create a register to hold the temporary result of multiplying the type size + // constant by the variable amount. + unsigned TotalSizeReg = makeAnotherReg(Type::UIntTy); + unsigned SrcReg1 = getReg(I.getArraySize()); + + // TotalSizeReg = mul , + MachineBasicBlock::iterator MBBI = BB->end(); + doMultiplyConst(BB, MBBI, TotalSizeReg, Type::UIntTy, SrcReg1, TySize); + // AddedSize = add , 15 + unsigned AddedSizeReg = makeAnotherReg(Type::UIntTy); + BuildMI(BB, X86::ADDri32, 2, AddedSizeReg).addReg(TotalSizeReg).addZImm(15); -/// visitAllocaInst - I want some stack space. Come on, man, I said I -/// want some freakin' stack space. -void -ISel::visitAllocaInst (AllocaInst &I) -{ - // Find the data size of the alloca inst's getAllocatedType. - const Type *allocatedType = I.getAllocatedType (); - const TargetData &TD = TM.DataLayout; - unsigned allocatedTypeSize = TD.getTypeSize (allocatedType); - // Keep stack 32-bit aligned. - unsigned int allocatedTypeWords = allocatedTypeSize / 4; - if (allocatedTypeSize % 4 != 0) { allocatedTypeWords++; } + // AlignedSize = and , ~15 + unsigned AlignedSize = makeAnotherReg(Type::UIntTy); + BuildMI(BB, X86::ANDri32, 2, AlignedSize).addReg(AddedSizeReg).addZImm(~15); + // Subtract size from stack pointer, thereby allocating some space. - BuildMI (BB, X86::SUBri32, 1, X86::ESP).addZImm (allocatedTypeWords * 4); + BuildMI(BB, X86::SUBrr32, 2, X86::ESP).addReg(X86::ESP).addReg(AlignedSize); + // Put a pointer to the space into the result register, by copying // the stack pointer. - BuildMI (BB, X86::MOVrr32, 1, getReg (I)).addReg (X86::ESP); + BuildMI(BB, X86::MOVrr32, 1, getReg(I)).addReg(X86::ESP); + + // Inform the Frame Information that we have just allocated a variable-sized + // object. + F->getFrameInfo()->CreateVariableSizedObject(); } - -/// createSimpleX86InstructionSelector - This pass converts an LLVM function +/// visitMallocInst - Malloc instructions are code generated into direct calls +/// to the library malloc. +/// +void ISel::visitMallocInst(MallocInst &I) { + unsigned AllocSize = TM.getTargetData().getTypeSize(I.getAllocatedType()); + unsigned Arg; + + if (ConstantUInt *C = dyn_cast(I.getOperand(0))) { + Arg = getReg(ConstantUInt::get(Type::UIntTy, C->getValue() * AllocSize)); + } else { + Arg = makeAnotherReg(Type::UIntTy); + unsigned Op0Reg = getReg(I.getOperand(0)); + MachineBasicBlock::iterator MBBI = BB->end(); + doMultiplyConst(BB, MBBI, Arg, Type::UIntTy, Op0Reg, AllocSize); + } + + std::vector Args; + Args.push_back(ValueRecord(Arg, Type::UIntTy)); + MachineInstr *TheCall = BuildMI(X86::CALLpcrel32, + 1).addExternalSymbol("malloc", true); + doCall(ValueRecord(getReg(I), I.getType()), TheCall, Args); +} + + +/// visitFreeInst - Free instructions are code gen'd to call the free libc +/// function. +/// +void ISel::visitFreeInst(FreeInst &I) { + std::vector Args; + Args.push_back(ValueRecord(I.getOperand(0))); + MachineInstr *TheCall = BuildMI(X86::CALLpcrel32, + 1).addExternalSymbol("free", true); + doCall(ValueRecord(0, Type::VoidTy), TheCall, Args); +} + + +/// createX86SimpleInstructionSelector - This pass converts an LLVM function /// into a machine code representation is a very simple peep-hole fashion. The /// generated code sucks but the implementation is nice and simple. /// -Pass *createSimpleX86InstructionSelector(TargetMachine &TM) { +FunctionPass *createX86SimpleInstructionSelector(TargetMachine &TM) { return new ISel(TM); }