X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTarget%2FSparcV9%2FSparcV9InstrSelection.cpp;h=5195f4ac29faef839c3603f35fff817e89f62238;hb=edf3a727b7106cfa9f10aadd5e6f603bcc0b879f;hp=533d74c15f91bf3c31f2d003ebf8d8052ee26080;hpb=67a463ac04f439fa24ad9777e29ea487bb99b7c3;p=oota-llvm.git diff --git a/lib/Target/SparcV9/SparcV9InstrSelection.cpp b/lib/Target/SparcV9/SparcV9InstrSelection.cpp index 533d74c15f9..5195f4ac29f 100644 --- a/lib/Target/SparcV9/SparcV9InstrSelection.cpp +++ b/lib/Target/SparcV9/SparcV9InstrSelection.cpp @@ -1,44 +1,29 @@ -// $Id$ -//*************************************************************************** -// File: -// SparcInstrSelection.cpp -// -// Purpose: -// BURS instruction selection for SPARC V9 architecture. -// -// History: -// 7/02/01 - Vikram Adve - Created -//**************************************************************************/ +//===-- SparcInstrSelection.cpp -------------------------------------------===// +// +// BURS instruction selection for SPARC V9 architecture. +// +//===----------------------------------------------------------------------===// #include "SparcInternals.h" #include "SparcInstrSelectionSupport.h" +#include "SparcRegClassInfo.h" #include "llvm/CodeGen/InstrSelectionSupport.h" #include "llvm/CodeGen/MachineInstr.h" +#include "llvm/CodeGen/MachineInstrAnnot.h" #include "llvm/CodeGen/InstrForest.h" #include "llvm/CodeGen/InstrSelection.h" -#include "llvm/Support/MathExtras.h" +#include "llvm/CodeGen/MachineCodeForMethod.h" +#include "llvm/CodeGen/MachineCodeForInstruction.h" #include "llvm/DerivedTypes.h" #include "llvm/iTerminators.h" #include "llvm/iMemory.h" #include "llvm/iOther.h" -#include "llvm/BasicBlock.h" -#include "llvm/Method.h" -#include "llvm/ConstPoolVals.h" +#include "llvm/Function.h" +#include "llvm/Constants.h" +#include "llvm/ConstantHandling.h" +#include "Support/MathExtras.h" #include - -//******************** Internal Data Declarations ************************/ - - -//************************* Forward Declarations ***************************/ - - -static void SetMemOperands_Internal (MachineInstr* minstr, - const InstructionNode* vmInstrNode, - Value* ptrVal, - Value* arrayOffsetVal, - const vector& idxVec, - const TargetMachine& target); - +using std::vector; //************************ Internal Functions ******************************/ @@ -140,20 +125,20 @@ ChooseBFpccInstruction(const InstructionNode* instrNode, // Eventually the entire BURG instruction selection should be put // into a separate class that can hold such information. // The static cache is not too bad because the memory for these -// TmpInstructions will be freed along with the rest of the Method anyway. +// TmpInstructions will be freed along with the rest of the Function anyway. // static TmpInstruction* -GetTmpForCC(Value* boolVal, const Method* method, const Type* ccType) +GetTmpForCC(Value* boolVal, const Function *F, const Type* ccType) { - typedef hash_map BoolTmpCache; + typedef hash_map BoolTmpCache; static BoolTmpCache boolToTmpCache; // Map boolVal -> TmpInstruction* - static const Method* lastMethod = NULL; // Use to flush cache between methods + static const Function *lastFunction = 0;// Use to flush cache between funcs assert(boolVal->getType() == Type::BoolTy && "Weird but ok! Delete assert"); - if (lastMethod != method) + if (lastFunction != F) { - lastMethod = method; + lastFunction = F; boolToTmpCache.clear(); } @@ -161,7 +146,7 @@ GetTmpForCC(Value* boolVal, const Method* method, const Type* ccType) // directly written to map using the ref returned by operator[]. TmpInstruction*& tmpI = boolToTmpCache[boolVal]; if (tmpI == NULL) - tmpI = new TmpInstruction(TMP_INSTRUCTION_OPCODE, ccType, boolVal, NULL); + tmpI = new TmpInstruction(ccType, boolVal); return tmpI; } @@ -172,12 +157,13 @@ ChooseBccInstruction(const InstructionNode* instrNode, bool& isFPBranch) { InstructionNode* setCCNode = (InstructionNode*) instrNode->leftChild(); - BinaryOperator* setCCInstr = (BinaryOperator*) setCCNode->getInstruction(); + assert(setCCNode->getOpLabel() == SetCCOp); + BinaryOperator* setCCInstr =cast(setCCNode->getInstruction()); const Type* setCCType = setCCInstr->getOperand(0)->getType(); - isFPBranch = (setCCType == Type::FloatTy || setCCType == Type::DoubleTy); + isFPBranch = setCCType->isFloatingPoint(); // Return value: don't delete! - if (isFPBranch) + if (isFPBranch) return ChooseBFpccInstruction(instrNode, setCCInstr); else return ChooseBpccInstruction(instrNode, setCCInstr); @@ -238,12 +224,11 @@ ChooseMovpccAfterSub(const InstructionNode* instrNode, } static inline MachineOpCode -ChooseConvertToFloatInstr(const InstructionNode* instrNode, - const Type* opType) +ChooseConvertToFloatInstr(OpLabel vopCode, const Type* opType) { MachineOpCode opCode = INVALID_OPCODE; - switch(instrNode->getOpLabel()) + switch(vopCode) { case ToFloatTy: if (opType == Type::SByteTy || opType == Type::ShortTy || opType == Type::IntTy) @@ -259,12 +244,14 @@ ChooseConvertToFloatInstr(const InstructionNode* instrNode, break; case ToDoubleTy: - // Use FXTOD for all integer-to-double conversions. This has to be - // consistent with the code in CreateCodeToCopyIntToFloat() since - // that will be used to load the integer into an FP register. - // - if (opType == Type::SByteTy || opType == Type::ShortTy || - opType == Type::IntTy || opType == Type::LongTy) + // This is usually used in conjunction with CreateCodeToCopyIntToFloat(). + // Both functions should treat the integer as a 32-bit value for types + // of 4 bytes or less, and as a 64-bit value otherwise. + if (opType == Type::SByteTy || opType == Type::UByteTy || + opType == Type::ShortTy || opType == Type::UShortTy || + opType == Type::IntTy || opType == Type::UIntTy) + opCode = FITOD; + else if (opType == Type::LongTy || opType == Type::ULongTy) opCode = FXTOD; else if (opType == Type::FloatTy) opCode = FSTOD; @@ -282,64 +269,89 @@ ChooseConvertToFloatInstr(const InstructionNode* instrNode, } static inline MachineOpCode -ChooseConvertToIntInstr(const InstructionNode* instrNode, - const Type* opType) +ChooseConvertFPToIntInstr(Type::PrimitiveID tid, const Type* opType) { MachineOpCode opCode = INVALID_OPCODE;; - - int instrType = (int) instrNode->getOpLabel(); - - if (instrType == ToSByteTy || instrType == ToShortTy || instrType == ToIntTy) + + assert((opType == Type::FloatTy || opType == Type::DoubleTy) + && "This function should only be called for FLOAT or DOUBLE"); + + if (tid==Type::UIntTyID) { - switch (opType->getPrimitiveID()) - { - case Type::FloatTyID: opCode = FSTOI; break; - case Type::DoubleTyID: opCode = FDTOI; break; - default: - assert(0 && "Non-numeric non-bool type cannot be converted to Int"); - break; - } + assert(tid != Type::UIntTyID && "FP-to-uint conversions must be expanded" + " into FP->long->uint for SPARC v9: SO RUN PRESELECTION PASS!"); } - else if (instrType == ToLongTy) + else if (tid==Type::SByteTyID || tid==Type::ShortTyID || tid==Type::IntTyID || + tid==Type::UByteTyID || tid==Type::UShortTyID) { - switch (opType->getPrimitiveID()) - { - case Type::FloatTyID: opCode = FSTOX; break; - case Type::DoubleTyID: opCode = FDTOX; break; - default: - assert(0 && "Non-numeric non-bool type cannot be converted to Long"); - break; - } + opCode = (opType == Type::FloatTy)? FSTOI : FDTOI; + } + else if (tid==Type::LongTyID || tid==Type::ULongTyID) + { + opCode = (opType == Type::FloatTy)? FSTOX : FDTOX; } else assert(0 && "Should not get here, Mo!"); - + return opCode; } - -static inline MachineOpCode -ChooseAddInstructionByType(const Type* resultType) +MachineInstr* +CreateConvertFPToIntInstr(Type::PrimitiveID destTID, + Value* srcVal, Value* destVal) { - MachineOpCode opCode = INVALID_OPCODE; + MachineOpCode opCode = ChooseConvertFPToIntInstr(destTID, srcVal->getType()); + assert(opCode != INVALID_OPCODE && "Expected to need conversion!"); - if (resultType->isIntegral() || - resultType->isPointerType() || - resultType->isLabelType() || - isa(resultType) || - resultType == Type::BoolTy) - { - opCode = ADD; - } - else - switch(resultType->getPrimitiveID()) - { - case Type::FloatTyID: opCode = FADDS; break; - case Type::DoubleTyID: opCode = FADDD; break; - default: assert(0 && "Invalid type for ADD instruction"); break; - } - - return opCode; + MachineInstr* M = new MachineInstr(opCode); + M->SetMachineOperandVal(0, MachineOperand::MO_VirtualRegister, srcVal); + M->SetMachineOperandVal(1, MachineOperand::MO_VirtualRegister, destVal); + return M; +} + +// CreateCodeToConvertFloatToInt: Convert FP value to signed or unsigned integer +// The FP value must be converted to the dest type in an FP register, +// and the result is then copied from FP to int register via memory. +// +// Since fdtoi converts to signed integers, any FP value V between MAXINT+1 +// and MAXUNSIGNED (i.e., 2^31 <= V <= 2^32-1) would be converted incorrectly +// *only* when converting to an unsigned int. (Unsigned byte, short or long +// don't have this problem.) +// For unsigned int, we therefore have to generate the code sequence: +// +// if (V > (float) MAXINT) { +// unsigned result = (unsigned) (V - (float) MAXINT); +// result = result + (unsigned) MAXINT; +// } +// else +// result = (unsigned int) V; +// +static void +CreateCodeToConvertFloatToInt(const TargetMachine& target, + Value* opVal, + Instruction* destI, + std::vector& mvec, + MachineCodeForInstruction& mcfi) +{ + // Create a temporary to represent the FP register into which the + // int value will placed after conversion. The type of this temporary + // depends on the type of FP register to use: single-prec for a 32-bit + // int or smaller; double-prec for a 64-bit int. + // + size_t destSize = target.DataLayout.getTypeSize(destI->getType()); + const Type* destTypeToUse = (destSize > 4)? Type::DoubleTy : Type::FloatTy; + TmpInstruction* destForCast = new TmpInstruction(destTypeToUse, opVal); + mcfi.addTemp(destForCast); + + // Create the fp-to-int conversion code + MachineInstr* M =CreateConvertFPToIntInstr(destI->getType()->getPrimitiveID(), + opVal, destForCast); + mvec.push_back(M); + + // Create the fpreg-to-intreg copy code + target.getInstrInfo(). + CreateCodeToCopyFloatToInt(target, destI->getParent()->getParent(), + destForCast, destI, mvec, mcfi); } @@ -356,10 +368,10 @@ CreateMovFloatInstruction(const InstructionNode* instrNode, { MachineInstr* minstr = new MachineInstr((resultType == Type::FloatTy) ? FMOVS : FMOVD); - minstr->SetMachineOperand(0, MachineOperand::MO_VirtualRegister, - instrNode->leftChild()->getValue()); - minstr->SetMachineOperand(1, MachineOperand::MO_VirtualRegister, - instrNode->getValue()); + minstr->SetMachineOperandVal(0, MachineOperand::MO_VirtualRegister, + instrNode->leftChild()->getValue()); + minstr->SetMachineOperandVal(1, MachineOperand::MO_VirtualRegister, + instrNode->getValue()); return minstr; } @@ -369,20 +381,17 @@ CreateAddConstInstruction(const InstructionNode* instrNode) MachineInstr* minstr = NULL; Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue(); - assert(isa(constOp)); + assert(isa(constOp)); // Cases worth optimizing are: // (1) Add with 0 for float or double: use an FMOV of appropriate type, // instead of an FADD (1 vs 3 cycles). There is no integer MOV. // - const Type* resultType = instrNode->getInstruction()->getType(); - - if (resultType == Type::FloatTy || - resultType == Type::DoubleTy) - { - double dval = ((ConstPoolFP*) constOp)->getValue(); + if (ConstantFP *FPC = dyn_cast(constOp)) { + double dval = FPC->getValue(); if (dval == 0.0) - minstr = CreateMovFloatInstruction(instrNode, resultType); + minstr = CreateMovFloatInstruction(instrNode, + instrNode->getInstruction()->getType()); } return minstr; @@ -394,8 +403,7 @@ ChooseSubInstructionByType(const Type* resultType) { MachineOpCode opCode = INVALID_OPCODE; - if (resultType->isIntegral() || - resultType->isPointerType()) + if (resultType->isInteger() || isa(resultType)) { opCode = SUB; } @@ -417,21 +425,18 @@ CreateSubConstInstruction(const InstructionNode* instrNode) MachineInstr* minstr = NULL; Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue(); - assert(isa(constOp)); + assert(isa(constOp)); // Cases worth optimizing are: // (1) Sub with 0 for float or double: use an FMOV of appropriate type, // instead of an FSUB (1 vs 3 cycles). There is no integer MOV. // - const Type* resultType = instrNode->getInstruction()->getType(); - - if (resultType == Type::FloatTy || - resultType == Type::DoubleTy) - { - double dval = ((ConstPoolFP*) constOp)->getValue(); - if (dval == 0.0) - minstr = CreateMovFloatInstruction(instrNode, resultType); - } + if (ConstantFP *FPC = dyn_cast(constOp)) { + double dval = FPC->getValue(); + if (dval == 0.0) + minstr = CreateMovFloatInstruction(instrNode, + instrNode->getInstruction()->getType()); + } return minstr; } @@ -476,7 +481,7 @@ ChooseMulInstructionByType(const Type* resultType) { MachineOpCode opCode = INVALID_OPCODE; - if (resultType->isIntegral()) + if (resultType->isInteger()) opCode = MULX; else switch(resultType->getPrimitiveID()) @@ -490,55 +495,101 @@ ChooseMulInstructionByType(const Type* resultType) } -static inline MachineOpCode -ChooseMulInstruction(const InstructionNode* instrNode, - bool checkCasts) -{ - if (checkCasts && BothFloatToDouble(instrNode)) - return FSMULD; - - // else use the regular multiply instructions - return ChooseMulInstructionByType(instrNode->getInstruction()->getType()); -} - static inline MachineInstr* -CreateIntNegInstruction(TargetMachine& target, +CreateIntNegInstruction(const TargetMachine& target, Value* vreg) { MachineInstr* minstr = new MachineInstr(SUB); - minstr->SetMachineOperand(0, target.getRegInfo().getZeroRegNum()); - minstr->SetMachineOperand(1, MachineOperand::MO_VirtualRegister, vreg); - minstr->SetMachineOperand(2, MachineOperand::MO_VirtualRegister, vreg); + minstr->SetMachineOperandReg(0, target.getRegInfo().getZeroRegNum()); + minstr->SetMachineOperandVal(1, MachineOperand::MO_VirtualRegister, vreg); + minstr->SetMachineOperandVal(2, MachineOperand::MO_VirtualRegister, vreg); return minstr; } -static inline MachineInstr* -CreateMulConstInstruction(TargetMachine &target, - const InstructionNode* instrNode, - MachineInstr*& getMinstr2) +// Create instruction sequence for any shift operation. +// SLL or SLLX on an operand smaller than the integer reg. size (64bits) +// requires a second instruction for explicit sign-extension. +// Note that we only have to worry about a sign-bit appearing in the +// most significant bit of the operand after shifting (e.g., bit 32 of +// Int or bit 16 of Short), so we do not have to worry about results +// that are as large as a normal integer register. +// +static inline void +CreateShiftInstructions(const TargetMachine& target, + Function* F, + MachineOpCode shiftOpCode, + Value* argVal1, + Value* optArgVal2, /* Use optArgVal2 if not NULL */ + unsigned int optShiftNum, /* else use optShiftNum */ + Instruction* destVal, + vector& mvec, + MachineCodeForInstruction& mcfi) { - MachineInstr* minstr = NULL; // return NULL if we cannot exploit constant - getMinstr2 = NULL; // to create a cheaper instruction - bool needNeg = false; + assert((optArgVal2 != NULL || optShiftNum <= 64) && + "Large shift sizes unexpected, but can be handled below: " + "You need to check whether or not it fits in immed field below"); + + // If this is a logical left shift of a type smaller than the standard + // integer reg. size, we have to extend the sign-bit into upper bits + // of dest, so we need to put the result of the SLL into a temporary. + // + Value* shiftDest = destVal; + unsigned opSize = target.DataLayout.getTypeSize(argVal1->getType()); + if ((shiftOpCode == SLL || shiftOpCode == SLLX) + && opSize < target.DataLayout.getIntegerRegize()) + { // put SLL result into a temporary + shiftDest = new TmpInstruction(argVal1, optArgVal2, "sllTmp"); + mcfi.addTemp(shiftDest); + } + + MachineInstr* M = (optArgVal2 != NULL) + ? Create3OperandInstr(shiftOpCode, argVal1, optArgVal2, shiftDest) + : Create3OperandInstr_UImmed(shiftOpCode, argVal1, optShiftNum, shiftDest); + mvec.push_back(M); + + if (shiftDest != destVal) + { // extend the sign-bit of the result into all upper bits of dest + assert(8*opSize <= 32 && "Unexpected type size > 4 and < IntRegSize?"); + target.getInstrInfo(). + CreateSignExtensionInstructions(target, F, shiftDest, destVal, + 8*opSize, mvec, mcfi); + } +} - Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue(); - assert(isa(constOp)); + +// Does not create any instructions if we cannot exploit constant to +// create a cheaper instruction. +// This returns the approximate cost of the instructions generated, +// which is used to pick the cheapest when both operands are constant. +static inline unsigned int +CreateMulConstInstruction(const TargetMachine &target, Function* F, + Value* lval, Value* rval, Instruction* destVal, + vector& mvec, + MachineCodeForInstruction& mcfi) +{ + /* Use max. multiply cost, viz., cost of MULX */ + unsigned int cost = target.getInstrInfo().minLatency(MULX); + unsigned int firstNewInstr = mvec.size(); + + Value* constOp = rval; + if (! isa(constOp)) + return cost; // Cases worth optimizing are: // (1) Multiply by 0 or 1 for any type: replace with copy (ADD or FMOV) // (2) Multiply by 2^x for integer types: replace with Shift // - const Type* resultType = instrNode->getInstruction()->getType(); + const Type* resultType = destVal->getType(); - if (resultType->isIntegral() || resultType->isPointerType()) + if (resultType->isInteger() || isa(resultType)) { - unsigned pow; bool isValidConst; int64_t C = GetConstantValueAsSignedInt(constOp, isValidConst); if (isValidConst) { + unsigned pow; bool needNeg = false; if (C < 0) { @@ -548,60 +599,109 @@ CreateMulConstInstruction(TargetMachine &target, if (C == 0 || C == 1) { - minstr = new MachineInstr(ADD); - - if (C == 0) - minstr->SetMachineOperand(0, - target.getRegInfo().getZeroRegNum()); - else - minstr->SetMachineOperand(0,MachineOperand::MO_VirtualRegister, - instrNode->leftChild()->getValue()); - minstr->SetMachineOperand(1,target.getRegInfo().getZeroRegNum()); + cost = target.getInstrInfo().minLatency(ADD); + MachineInstr* M = (C == 0) + ? Create3OperandInstr_Reg(ADD, + target.getRegInfo().getZeroRegNum(), + target.getRegInfo().getZeroRegNum(), + destVal) + : Create3OperandInstr_Reg(ADD, lval, + target.getRegInfo().getZeroRegNum(), + destVal); + mvec.push_back(M); } - else if (IsPowerOf2(C, pow)) + else if (isPowerOf2(C, pow)) { - minstr = new MachineInstr((resultType == Type::LongTy) - ? SLLX : SLL); - minstr->SetMachineOperand(0, MachineOperand::MO_VirtualRegister, - instrNode->leftChild()->getValue()); - minstr->SetMachineOperand(1, MachineOperand::MO_UnextendedImmed, - pow); + unsigned int opSize = target.DataLayout.getTypeSize(resultType); + MachineOpCode opCode = (opSize <= 32)? SLL : SLLX; + CreateShiftInstructions(target, F, opCode, lval, NULL, pow, + destVal, mvec, mcfi); } - if (minstr && needNeg) + if (mvec.size() > 0 && needNeg) { // insert after the instr to flip the sign - getMinstr2 = CreateIntNegInstruction(target, - instrNode->getValue()); + MachineInstr* M = CreateIntNegInstruction(target, destVal); + mvec.push_back(M); } } } else { - if (resultType == Type::FloatTy || - resultType == Type::DoubleTy) + if (ConstantFP *FPC = dyn_cast(constOp)) { - double dval = ((ConstPoolFP*) constOp)->getValue(); + double dval = FPC->getValue(); if (fabs(dval) == 1) { - bool needNeg = (dval < 0); - - MachineOpCode opCode = needNeg + MachineOpCode opCode = (dval < 0) ? (resultType == Type::FloatTy? FNEGS : FNEGD) : (resultType == Type::FloatTy? FMOVS : FMOVD); - - minstr = new MachineInstr(opCode); - minstr->SetMachineOperand(0, - MachineOperand::MO_VirtualRegister, - instrNode->leftChild()->getValue()); + MachineInstr* M = Create2OperandInstr(opCode, lval, destVal); + mvec.push_back(M); } } } - if (minstr != NULL) - minstr->SetMachineOperand(2, MachineOperand::MO_VirtualRegister, - instrNode->getValue()); + if (firstNewInstr < mvec.size()) + { + cost = 0; + for (unsigned int i=firstNewInstr; i < mvec.size(); ++i) + cost += target.getInstrInfo().minLatency(mvec[i]->getOpCode()); + } - return minstr; + return cost; +} + + +// Does not create any instructions if we cannot exploit constant to +// create a cheaper instruction. +// +static inline void +CreateCheapestMulConstInstruction(const TargetMachine &target, + Function* F, + Value* lval, Value* rval, + Instruction* destVal, + vector& mvec, + MachineCodeForInstruction& mcfi) +{ + Value* constOp; + if (isa(lval) && isa(rval)) + { // both operands are constant: evaluate and "set" in dest + Constant* P = ConstantFoldBinaryInstruction(Instruction::Mul, + cast(lval), cast(rval)); + target.getInstrInfo().CreateCodeToLoadConst(target,F,P,destVal,mvec,mcfi); + } + else if (isa(rval)) // rval is constant, but not lval + CreateMulConstInstruction(target, F, lval, rval, destVal, mvec, mcfi); + else if (isa(lval)) // lval is constant, but not rval + CreateMulConstInstruction(target, F, lval, rval, destVal, mvec, mcfi); + + // else neither is constant + return; +} + +// Return NULL if we cannot exploit constant to create a cheaper instruction +static inline void +CreateMulInstruction(const TargetMachine &target, Function* F, + Value* lval, Value* rval, Instruction* destVal, + vector& mvec, + MachineCodeForInstruction& mcfi, + MachineOpCode forceMulOp = INVALID_MACHINE_OPCODE) +{ + unsigned int L = mvec.size(); + CreateCheapestMulConstInstruction(target,F, lval, rval, destVal, mvec, mcfi); + if (mvec.size() == L) + { // no instructions were added so create MUL reg, reg, reg. + // Use FSMULD if both operands are actually floats cast to doubles. + // Otherwise, use the default opcode for the appropriate type. + MachineOpCode mulOp = ((forceMulOp != INVALID_MACHINE_OPCODE) + ? forceMulOp + : ChooseMulInstructionByType(destVal->getType())); + MachineInstr* M = new MachineInstr(mulOp); + M->SetMachineOperandVal(0, MachineOperand::MO_VirtualRegister, lval); + M->SetMachineOperandVal(1, MachineOperand::MO_VirtualRegister, rval); + M->SetMachineOperandVal(2, MachineOperand::MO_VirtualRegister, destVal); + mvec.push_back(M); + } } @@ -617,7 +717,7 @@ ChooseDivInstruction(TargetMachine &target, const Type* resultType = instrNode->getInstruction()->getType(); - if (resultType->isIntegral()) + if (resultType->isInteger()) opCode = resultType->isSigned()? SDIVX : UDIVX; else switch(resultType->getPrimitiveID()) @@ -631,16 +731,18 @@ ChooseDivInstruction(TargetMachine &target, } -static inline MachineInstr* +// Return NULL if we cannot exploit constant to create a cheaper instruction +static inline void CreateDivConstInstruction(TargetMachine &target, const InstructionNode* instrNode, - MachineInstr*& getMinstr2) + vector& mvec) { - MachineInstr* minstr = NULL; - getMinstr2 = NULL; + MachineInstr* minstr1 = NULL; + MachineInstr* minstr2 = NULL; Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue(); - assert(isa(constOp)); + if (! isa(constOp)) + return; // Cases worth optimizing are: // (1) Divide by 1 for any type: replace with copy (ADD or FMOV) @@ -648,7 +750,7 @@ CreateDivConstInstruction(TargetMachine &target, // const Type* resultType = instrNode->getInstruction()->getType(); - if (resultType->isIntegral()) + if (resultType->isInteger()) { unsigned pow; bool isValidConst; @@ -664,36 +766,39 @@ CreateDivConstInstruction(TargetMachine &target, if (C == 1) { - minstr = new MachineInstr(ADD); - minstr->SetMachineOperand(0,MachineOperand::MO_VirtualRegister, - instrNode->leftChild()->getValue()); - minstr->SetMachineOperand(1,target.getRegInfo().getZeroRegNum()); + minstr1 = new MachineInstr(ADD); + minstr1->SetMachineOperandVal(0, + MachineOperand::MO_VirtualRegister, + instrNode->leftChild()->getValue()); + minstr1->SetMachineOperandReg(1, + target.getRegInfo().getZeroRegNum()); } - else if (IsPowerOf2(C, pow)) + else if (isPowerOf2(C, pow)) { MachineOpCode opCode= ((resultType->isSigned()) ? (resultType==Type::LongTy)? SRAX : SRA : (resultType==Type::LongTy)? SRLX : SRL); - minstr = new MachineInstr(opCode); - minstr->SetMachineOperand(0, MachineOperand::MO_VirtualRegister, + minstr1 = new MachineInstr(opCode); + minstr1->SetMachineOperandVal(0, + MachineOperand::MO_VirtualRegister, instrNode->leftChild()->getValue()); - minstr->SetMachineOperand(1, MachineOperand::MO_UnextendedImmed, - pow); + minstr1->SetMachineOperandConst(1, + MachineOperand::MO_UnextendedImmed, + pow); } - if (minstr && needNeg) + if (minstr1 && needNeg) { // insert after the instr to flip the sign - getMinstr2 = CreateIntNegInstruction(target, + minstr2 = CreateIntNegInstruction(target, instrNode->getValue()); } } } else { - if (resultType == Type::FloatTy || - resultType == Type::DoubleTy) + if (ConstantFP *FPC = dyn_cast(constOp)) { - double dval = ((ConstPoolFP*) constOp)->getValue(); + double dval = FPC->getValue(); if (fabs(dval) == 1) { bool needNeg = (dval < 0); @@ -702,18 +807,116 @@ CreateDivConstInstruction(TargetMachine &target, ? (resultType == Type::FloatTy? FNEGS : FNEGD) : (resultType == Type::FloatTy? FMOVS : FMOVD); - minstr = new MachineInstr(opCode); - minstr->SetMachineOperand(0, MachineOperand::MO_VirtualRegister, + minstr1 = new MachineInstr(opCode); + minstr1->SetMachineOperandVal(0, + MachineOperand::MO_VirtualRegister, instrNode->leftChild()->getValue()); } } } - if (minstr != NULL) - minstr->SetMachineOperand(2, MachineOperand::MO_VirtualRegister, - instrNode->getValue()); + if (minstr1 != NULL) + minstr1->SetMachineOperandVal(2, MachineOperand::MO_VirtualRegister, + instrNode->getValue()); - return minstr; + if (minstr1) + mvec.push_back(minstr1); + if (minstr2) + mvec.push_back(minstr2); +} + + +static void +CreateCodeForVariableSizeAlloca(const TargetMachine& target, + Instruction* result, + unsigned int tsize, + Value* numElementsVal, + vector& getMvec) +{ + MachineInstr* M; + MachineCodeForInstruction& mcfi = MachineCodeForInstruction::get(result); + + // Create a Value to hold the (constant) element size + Value* tsizeVal = ConstantSInt::get(Type::IntTy, tsize); + + // Get the constant offset from SP for dynamically allocated storage + // and create a temporary Value to hold it. + assert(result && result->getParent() && "Result value is not part of a fn?"); + Function *F = result->getParent()->getParent(); + MachineCodeForMethod& mcInfo = MachineCodeForMethod::get(F); + bool growUp; + ConstantSInt* dynamicAreaOffset = + ConstantSInt::get(Type::IntTy, + target.getFrameInfo().getDynamicAreaOffset(mcInfo,growUp)); + assert(! growUp && "Has SPARC v9 stack frame convention changed?"); + + // Create a temporary value to hold the result of MUL + TmpInstruction* tmpProd = new TmpInstruction(numElementsVal, tsizeVal); + mcfi.addTemp(tmpProd); + + // Instruction 1: mul numElements, typeSize -> tmpProd + CreateMulInstruction(target, F, numElementsVal, tsizeVal, tmpProd, getMvec, + mcfi, INVALID_MACHINE_OPCODE); + + // Instruction 2: sub %sp, tmpProd -> %sp + M = new MachineInstr(SUB); + M->SetMachineOperandReg(0, target.getRegInfo().getStackPointer()); + M->SetMachineOperandVal(1, MachineOperand::MO_VirtualRegister, tmpProd); + M->SetMachineOperandReg(2, target.getRegInfo().getStackPointer()); + getMvec.push_back(M); + + // Instruction 3: add %sp, frameSizeBelowDynamicArea -> result + M = new MachineInstr(ADD); + M->SetMachineOperandReg(0, target.getRegInfo().getStackPointer()); + M->SetMachineOperandVal(1, MachineOperand::MO_VirtualRegister, dynamicAreaOffset); + M->SetMachineOperandVal(2, MachineOperand::MO_VirtualRegister, result); + getMvec.push_back(M); +} + + +static void +CreateCodeForFixedSizeAlloca(const TargetMachine& target, + Instruction* result, + unsigned int tsize, + unsigned int numElements, + vector& getMvec) +{ + assert(tsize > 0 && "Illegal (zero) type size for alloca"); + assert(result && result->getParent() && + "Result value is not part of a function?"); + Function *F = result->getParent()->getParent(); + MachineCodeForMethod &mcInfo = MachineCodeForMethod::get(F); + + // Check if the offset would small enough to use as an immediate in + // load/stores (check LDX because all load/stores have the same-size immediate + // field). If not, put the variable in the dynamically sized area of the + // frame. + unsigned int paddedSizeIgnored; + int offsetFromFP = mcInfo.computeOffsetforLocalVar(target, result, + paddedSizeIgnored, + tsize * numElements); + if (! target.getInstrInfo().constantFitsInImmedField(LDX, offsetFromFP)) + { + CreateCodeForVariableSizeAlloca(target, result, tsize, + ConstantSInt::get(Type::IntTy,numElements), + getMvec); + return; + } + + // else offset fits in immediate field so go ahead and allocate it. + offsetFromFP = mcInfo.allocateLocalVar(target, result, tsize * numElements); + + // Create a temporary Value to hold the constant offset. + // This is needed because it may not fit in the immediate field. + ConstantSInt* offsetVal = ConstantSInt::get(Type::IntTy, offsetFromFP); + + // Instruction 1: add %fp, offsetFromFP -> result + MachineInstr* M = new MachineInstr(ADD); + M->SetMachineOperandReg(0, target.getRegInfo().getFramePointer()); + M->SetMachineOperandVal(1, MachineOperand::MO_VirtualRegister, offsetVal); + M->SetMachineOperandVal(2, MachineOperand::MO_VirtualRegister, result); + + getMvec.push_back(M); } @@ -734,135 +937,85 @@ CreateDivConstInstruction(TargetMachine &target, //------------------------------------------------------------------------ static void -SetOperandsForMemInstr(MachineInstr* minstr, - const InstructionNode* vmInstrNode, +SetOperandsForMemInstr(vector& mvec, + InstructionNode* vmInstrNode, const TargetMachine& target) { - MemAccessInst* memInst = (MemAccessInst*) vmInstrNode->getInstruction(); - - // Variables to hold the index vector, ptr value, and offset value. - // The major work here is to extract these for all 3 instruction types - // and then call the common function SetMemOperands_Internal(). - // - const vector* idxVec = &memInst->getIndices(); - vector* newIdxVec = NULL; - Value* ptrVal; - Value* arrayOffsetVal = NULL; - - // Test if a GetElemPtr instruction is being folded into this mem instrn. - // If so, it will be in the left child for Load and GetElemPtr, - // and in the right child for Store instructions. - // - InstrTreeNode* ptrChild = (vmInstrNode->getOpLabel() == Instruction::Store - ? vmInstrNode->rightChild() - : vmInstrNode->leftChild()); - - if (ptrChild->getOpLabel() == Instruction::GetElementPtr || - ptrChild->getOpLabel() == GetElemPtrIdx) - { - // There is a GetElemPtr instruction and there may be a chain of - // more than one. Use the pointer value of the last one in the chain. - // Fold the index vectors from the entire chain and from the mem - // instruction into one single index vector. - // Finally, we never fold for an array instruction so make that NULL. - - newIdxVec = new vector; - ptrVal = FoldGetElemChain((InstructionNode*) ptrChild, *newIdxVec); - - newIdxVec->insert(newIdxVec->end(), idxVec->begin(), idxVec->end()); - idxVec = newIdxVec; - - assert(! ((PointerType*)ptrVal->getType())->getValueType()->isArrayType() - && "GetElemPtr cannot be folded into array refs in selection"); - } - else - { - // There is no GetElemPtr instruction. - // Use the pointer value and the index vector from the Mem instruction. - // If it is an array reference, get the array offset value. - // - ptrVal = memInst->getPtrOperand(); - - const Type* opType = - ((const PointerType*) ptrVal->getType())->getValueType(); - if (opType->isArrayType()) - { - assert((memInst->getNumOperands() - == (unsigned) 1 + memInst->getFirstOffsetIdx()) - && "Array refs must be lowered before Instruction Selection"); - - arrayOffsetVal = memInst->getOperand(memInst->getFirstOffsetIdx()); - } - } - - SetMemOperands_Internal(minstr, vmInstrNode, ptrVal, arrayOffsetVal, - *idxVec, target); - - if (newIdxVec != NULL) - delete newIdxVec; -} + Instruction* memInst = vmInstrNode->getInstruction(); + vector::iterator mvecI = mvec.end() - 1; + // Index vector, ptr value, and flag if all indices are const. + vector idxVec; + bool allConstantIndices; + Value* ptrVal = GetMemInstArgs(vmInstrNode, idxVec, allConstantIndices); -static void -SetMemOperands_Internal(MachineInstr* minstr, - const InstructionNode* vmInstrNode, - Value* ptrVal, - Value* arrayOffsetVal, - const vector& idxVec, - const TargetMachine& target) -{ - MemAccessInst* memInst = (MemAccessInst*) vmInstrNode->getInstruction(); - - // Initialize so we default to storing the offset in a register. + // Now create the appropriate operands for the machine instruction. + // First, initialize so we default to storing the offset in a register. int64_t smallConstOffset = 0; Value* valueForRegOffset = NULL; - MachineOperand::MachineOperandType offsetOpType =MachineOperand::MO_VirtualRegister; + MachineOperand::MachineOperandType offsetOpType = + MachineOperand::MO_VirtualRegister; - // Check if there is an index vector and if so, if it translates to - // a small enough constant to fit in the immediate-offset field. + // Check if there is an index vector and if so, compute the + // right offset for structures and for arrays // - if (idxVec.size() > 0) + if (!idxVec.empty()) { - bool isConstantOffset = false; - unsigned offset = 0; + const PointerType* ptrType = cast(ptrVal->getType()); - const PointerType* ptrType = (PointerType*) ptrVal->getType(); - - if (ptrType->getValueType()->isStructType()) + // If all indices are constant, compute the combined offset directly. + if (allConstantIndices) { - // the offset is always constant for structs - isConstantOffset = true; - - // Compute the offset value using the index vector - offset = target.DataLayout.getIndexedOffset(ptrType, idxVec); + // Compute the offset value using the index vector. Create a + // virtual reg. for it since it may not fit in the immed field. + uint64_t offset = target.DataLayout.getIndexedOffset(ptrType,idxVec); + valueForRegOffset = ConstantSInt::get(Type::LongTy, offset); } else { - // It must be an array ref. Check if the offset is a constant, - // and that the indexing has been lowered to a single offset. + // There is at least one non-constant offset. Therefore, this must + // be an array ref, and must have been lowered to a single non-zero + // offset. (An extra leading zero offset, if any, can be ignored.) + // Generate code sequence to compute address from index. // - assert(ptrType->getValueType()->isArrayType()); - assert(arrayOffsetVal != NULL - && "Expect to be given Value* for array offsets"); - - if (ConstPoolVal *CPV = dyn_cast(arrayOffsetVal)) - { - isConstantOffset = true; // always constant for structs - assert(arrayOffsetVal->getType()->isIntegral()); - offset = (CPV->getType()->isSigned() - ? ((ConstPoolSInt*)CPV)->getValue() - : (int64_t) ((ConstPoolUInt*)CPV)->getValue()); - } - else - { - valueForRegOffset = arrayOffsetVal; - } - } - - if (isConstantOffset) - { - // create a virtual register for the constant - valueForRegOffset = ConstPoolSInt::get(Type::IntTy, offset); + bool firstIdxIsZero = + (idxVec[0] == Constant::getNullValue(idxVec[0]->getType())); + assert(idxVec.size() == 1U + firstIdxIsZero + && "Array refs must be lowered before Instruction Selection"); + + Value* idxVal = idxVec[firstIdxIsZero]; + + vector mulVec; + Instruction* addr = new TmpInstruction(Type::ULongTy, memInst); + MachineCodeForInstruction::get(memInst).addTemp(addr); + + // Get the array type indexed by idxVal, and compute its element size. + // The call to getTypeSize() will fail if size is not constant. + const Type* vecType = (firstIdxIsZero + ? GetElementPtrInst::getIndexedType(ptrType, + std::vector(1U, idxVec[0]), + /*AllowCompositeLeaf*/ true) + : ptrType); + const Type* eltType = cast(vecType)->getElementType(); + ConstantUInt* eltSizeVal = ConstantUInt::get(Type::ULongTy, + target.DataLayout.getTypeSize(eltType)); + + // CreateMulInstruction() folds constants intelligently enough. + CreateMulInstruction(target, memInst->getParent()->getParent(), + idxVal, /* lval, not likely to be const*/ + eltSizeVal, /* rval, likely to be constant */ + addr, /* result */ + mulVec, MachineCodeForInstruction::get(memInst), + INVALID_MACHINE_OPCODE); + + // Insert mulVec[] before *mvecI in mvec[] and update mvecI + // to point to the same instruction it pointed to before. + assert(mulVec.size() > 0 && "No multiply code created?"); + vector::iterator oldMvecI = mvecI; + for (unsigned i=0, N=mulVec.size(); i < N; ++i) + mvecI = mvec.insert(mvecI, mulVec[i]) + 1; // pts to mem instr + + valueForRegOffset = addr; } } else @@ -870,29 +1023,40 @@ SetMemOperands_Internal(MachineInstr* minstr, offsetOpType = MachineOperand::MO_SignExtendedImmed; smallConstOffset = 0; } + + // For STORE: + // Operand 0 is value, operand 1 is ptr, operand 2 is offset + // For LOAD or GET_ELEMENT_PTR, + // Operand 0 is ptr, operand 1 is offset, operand 2 is result. + // + unsigned offsetOpNum, ptrOpNum; + if (memInst->getOpcode() == Instruction::Store) + { + (*mvecI)->SetMachineOperandVal(0, MachineOperand::MO_VirtualRegister, + vmInstrNode->leftChild()->getValue()); + ptrOpNum = 1; + offsetOpNum = 2; + } + else + { + ptrOpNum = 0; + offsetOpNum = 1; + (*mvecI)->SetMachineOperandVal(2, MachineOperand::MO_VirtualRegister, + memInst); + } - // Operand 0 is value for STORE, ptr for LOAD or GET_ELEMENT_PTR - // It is the left child in the instruction tree in all cases. - Value* leftVal = vmInstrNode->leftChild()->getValue(); - minstr->SetMachineOperand(0, MachineOperand::MO_VirtualRegister, leftVal); + (*mvecI)->SetMachineOperandVal(ptrOpNum, MachineOperand::MO_VirtualRegister, + ptrVal); - // Operand 1 is ptr for STORE, offset for LOAD or GET_ELEMENT_PTR - // Operand 2 is offset for STORE, result reg for LOAD or GET_ELEMENT_PTR - // - unsigned offsetOpNum = (memInst->getOpcode() == Instruction::Store)? 2 : 1; if (offsetOpType == MachineOperand::MO_VirtualRegister) { assert(valueForRegOffset != NULL); - minstr->SetMachineOperand(offsetOpNum, offsetOpType, valueForRegOffset); + (*mvecI)->SetMachineOperandVal(offsetOpNum, offsetOpType, + valueForRegOffset); } else - minstr->SetMachineOperand(offsetOpNum, offsetOpType, smallConstOffset); - - if (memInst->getOpcode() == Instruction::Store) - minstr->SetMachineOperand(1, MachineOperand::MO_VirtualRegister, ptrVal); - else - minstr->SetMachineOperand(2, MachineOperand::MO_VirtualRegister, - vmInstrNode->getValue()); + (*mvecI)->SetMachineOperandConst(offsetOpNum, offsetOpType, + smallConstOffset); } @@ -900,6 +1064,9 @@ SetMemOperands_Internal(MachineInstr* minstr, // Substitute operand `operandNum' of the instruction in node `treeNode' // in place of the use(s) of that instruction in node `parent'. // Check both explicit and implicit operands! +// Also make sure to skip over a parent who: +// (1) is a list node in the Burg tree, or +// (2) itself had its results forwarded to its parent // static void ForwardOperand(InstructionNode* treeNode, @@ -920,156 +1087,54 @@ ForwardOperand(InstructionNode* treeNode, InstructionNode* parentInstrNode = (InstructionNode*) parent; Instruction* userInstr = parentInstrNode->getInstruction(); - MachineCodeForVMInstr& mvec = userInstr->getMachineInstrVec(); - for (unsigned i=0, N=mvec.size(); i < N; i++) + MachineCodeForInstruction &mvec = MachineCodeForInstruction::get(userInstr); + + // The parent's mvec would be empty if it was itself forwarded. + // Recursively call ForwardOperand in that case... + // + if (mvec.size() == 0) { - MachineInstr* minstr = mvec[i]; - - for (unsigned i=0, numOps=minstr->getNumOperands(); i < numOps; ++i) + assert(parent->parent() != NULL && + "Parent could not have been forwarded, yet has no instructions?"); + ForwardOperand(treeNode, parent->parent(), operandNum); + } + else + { + for (unsigned i=0, N=mvec.size(); i < N; i++) { - const MachineOperand& mop = minstr->getOperand(i); - if (mop.getOperandType() == MachineOperand::MO_VirtualRegister && - mop.getVRegValue() == unusedOp) + MachineInstr* minstr = mvec[i]; + for (unsigned i=0, numOps=minstr->getNumOperands(); i < numOps; ++i) { - minstr->SetMachineOperand(i, MachineOperand::MO_VirtualRegister, - fwdOp); + const MachineOperand& mop = minstr->getOperand(i); + if (mop.getOperandType() == MachineOperand::MO_VirtualRegister && + mop.getVRegValue() == unusedOp) + minstr->SetMachineOperandVal(i, + MachineOperand::MO_VirtualRegister, fwdOp); } + + for (unsigned i=0,numOps=minstr->getNumImplicitRefs(); igetImplicitRef(i) == unusedOp) + minstr->setImplicitRef(i, fwdOp, + minstr->implicitRefIsDefined(i), + minstr->implicitRefIsDefinedAndUsed(i)); } - - for (unsigned i=0, numOps=minstr->getNumImplicitRefs(); i < numOps; ++i) - if (minstr->getImplicitRef(i) == unusedOp) - minstr->setImplicitRef(i, fwdOp, minstr->implicitRefIsDefined(i)); } } - -void UltraSparcInstrInfo:: -CreateCopyInstructionsByType(const TargetMachine& target, - Value* src, - Instruction* dest, - vector& minstrVec) const +inline bool +AllUsesAreBranches(const Instruction* setccI) { - bool loadConstantToReg = false; - - const Type* resultType = dest->getType(); - - MachineOpCode opCode = ChooseAddInstructionByType(resultType); - if (opCode == INVALID_OPCODE) - { - assert(0 && "Unsupported result type in CreateCopyInstructionsByType()"); - return; - } - - // if `src' is a constant that doesn't fit in the immed field or if it is - // a global variable (i.e., a constant address), generate a load - // instruction instead of an add - // - if (isa(src)) - { - unsigned int machineRegNum; - int64_t immedValue; - MachineOperand::MachineOperandType opType = - ChooseRegOrImmed(src, opCode, target, /*canUseImmed*/ true, - machineRegNum, immedValue); - - if (opType == MachineOperand::MO_VirtualRegister) - loadConstantToReg = true; - } - else if (isa(src)) - loadConstantToReg = true; - - if (loadConstantToReg) - { // `src' is constant and cannot fit in immed field for the ADD - // Insert instructions to "load" the constant into a register - vector tempVec; - target.getInstrInfo().CreateCodeToLoadConst(src,dest,minstrVec,tempVec); - for (unsigned i=0; i < tempVec.size(); i++) - dest->getMachineInstrVec().addTempValue(tempVec[i]); - } - else - { // Create the appropriate add instruction. - // Make `src' the second operand, in case it is a constant - // Use (unsigned long) 0 for a NULL pointer value. - // - const Type* nullValueType = - (resultType->getPrimitiveID() == Type::PointerTyID)? Type::ULongTy - : resultType; - MachineInstr* minstr = new MachineInstr(opCode); - minstr->SetMachineOperand(0, MachineOperand::MO_VirtualRegister, - ConstPoolVal::getNullConstant(nullValueType)); - minstr->SetMachineOperand(1, MachineOperand::MO_VirtualRegister, src); - minstr->SetMachineOperand(2, MachineOperand::MO_VirtualRegister, dest); - minstrVec.push_back(minstr); - } + for (Value::use_const_iterator UI=setccI->use_begin(), UE=setccI->use_end(); + UI != UE; ++UI) + if (! isa(*UI) // ignore tmp instructions here + && cast(*UI)->getOpcode() != Instruction::Br) + return false; + return true; } - - //******************* Externally Visible Functions *************************/ - -//------------------------------------------------------------------------ -// External Function: GetInstructionsForProlog -// External Function: GetInstructionsForEpilog -// -// Purpose: -// Create prolog and epilog code for procedure entry and exit -//------------------------------------------------------------------------ - -extern unsigned -GetInstructionsForProlog(BasicBlock* entryBB, - TargetMachine &target, - MachineInstr** mvec) -{ - int64_t s0=0; // used to avoid overloading ambiguity below - - const MachineFrameInfo& frameInfo = target.getFrameInfo(); - - // The second operand is the stack size. If it does not fit in the - // immediate field, we either have to find an unused register in the - // caller's window or move some elements to the dynamically allocated - // area of the stack frame (just above save area and method args). - Method* method = entryBB->getParent(); - MachineCodeForMethod& mcInfo = MachineCodeForMethod::get(method); - unsigned int staticStackSize = mcInfo.getStaticStackSize(); - - if (staticStackSize < (unsigned) frameInfo.getMinStackFrameSize()) - staticStackSize = (unsigned) frameInfo.getMinStackFrameSize(); - - if (unsigned padsz = (staticStackSize % - (unsigned) frameInfo.getStackFrameSizeAlignment())) - staticStackSize += frameInfo.getStackFrameSizeAlignment() - padsz; - - assert(target.getInstrInfo().constantFitsInImmedField(SAVE, staticStackSize) - && "Stack size too large for immediate field of SAVE instruction. Need additional work as described in the comment above"); - - mvec[0] = new MachineInstr(SAVE); - mvec[0]->SetMachineOperand(0, target.getRegInfo().getStackPointer()); - mvec[0]->SetMachineOperand(1, MachineOperand::MO_SignExtendedImmed, - - (int) staticStackSize); - mvec[0]->SetMachineOperand(2, target.getRegInfo().getStackPointer()); - - return 1; -} - - -extern unsigned -GetInstructionsForEpilog(BasicBlock* anExitBB, - TargetMachine &target, - MachineInstr** mvec) -{ - int64_t s0=0; // used to avoid overloading ambiguity below - - mvec[0] = new MachineInstr(RESTORE); - mvec[0]->SetMachineOperand(0, target.getRegInfo().getZeroRegNum()); - mvec[0]->SetMachineOperand(1, MachineOperand::MO_SignExtendedImmed, s0); - mvec[0]->SetMachineOperand(2, target.getRegInfo().getZeroRegNum()); - - return 1; -} - - //------------------------------------------------------------------------ // External Function: ThisIsAChainRule // @@ -1083,7 +1148,6 @@ ThisIsAChainRule(int eruleno) switch(eruleno) { case 111: // stmt: reg - case 113: // stmt: bool case 123: case 124: case 125: @@ -1102,8 +1166,10 @@ ThisIsAChainRule(int eruleno) case 242: case 243: case 244: + case 245: + case 321: return true; break; - + default: return false; break; } @@ -1118,22 +1184,27 @@ ThisIsAChainRule(int eruleno) // patterns chosen by the BURG-generated parser. //------------------------------------------------------------------------ -unsigned +void GetInstructionsByRule(InstructionNode* subtreeRoot, int ruleForNode, short* nts, TargetMachine &target, - MachineInstr** mvec) + vector& mvec) { - int numInstr = 1; // initialize for common case bool checkCast = false; // initialize here to use fall-through + bool maskUnsignedResult = false; int nextRule; int forwardOperandNum = -1; - int64_t s0=0, s8=8; // variables holding constants to avoid - uint64_t u0=0; // overloading ambiguities below + unsigned int allocaSize = 0; + MachineInstr* M, *M2; + unsigned int L; + + mvec.clear(); - for (unsigned i=0; i < MAX_INSTR_PER_VMINSTR; i++) - mvec[i] = NULL; + // If the code for this instruction was folded into the parent (user), + // then do nothing! + if (subtreeRoot->isFoldedIntoParent()) + return; // // Let's check for chain rules outside the switch so that we don't have @@ -1148,7 +1219,7 @@ GetInstructionsByRule(InstructionNode* subtreeRoot, && "A chain rule should have only one RHS non-terminal!"); nextRule = burm_rule(subtreeRoot->state, nts[0]); nts = burm_nts[nextRule]; - numInstr = GetInstructionsByRule(subtreeRoot, nextRule, nts,target,mvec); + GetInstructionsByRule(subtreeRoot, nextRule, nts, target, mvec); } else { @@ -1161,46 +1232,45 @@ GetInstructionsByRule(InstructionNode* subtreeRoot, // Mark the return value register as an implicit ref of // the machine instruction. // Finally put a NOP in the delay slot. - ReturnInst* returnInstr = (ReturnInst*) subtreeRoot->getInstruction(); + ReturnInst *returnInstr = + cast(subtreeRoot->getInstruction()); assert(returnInstr->getOpcode() == Instruction::Ret); - Method* method = returnInstr->getParent()->getParent(); - Instruction* returnReg = new TmpInstruction(TMP_INSTRUCTION_OPCODE, - returnInstr, NULL); - returnInstr->getMachineInstrVec().addTempValue(returnReg); + Instruction* returnReg = new TmpInstruction(returnInstr); + MachineCodeForInstruction::get(returnInstr).addTemp(returnReg); - mvec[0] = new MachineInstr(JMPLRET); - mvec[0]->SetMachineOperand(0, MachineOperand::MO_VirtualRegister, + M = new MachineInstr(JMPLRET); + M->SetMachineOperandReg(0, MachineOperand::MO_VirtualRegister, returnReg); - mvec[0]->SetMachineOperand(1, MachineOperand::MO_SignExtendedImmed,s8); - mvec[0]->SetMachineOperand(2, target.getRegInfo().getZeroRegNum()); + M->SetMachineOperandConst(1,MachineOperand::MO_SignExtendedImmed, + (int64_t)8); + M->SetMachineOperandReg(2, target.getRegInfo().getZeroRegNum()); if (returnInstr->getReturnValue() != NULL) - mvec[0]->addImplicitRef(returnInstr->getReturnValue()); + M->addImplicitRef(returnInstr->getReturnValue()); - unsigned n = numInstr++; // delay slot - mvec[n] = new MachineInstr(NOP); + mvec.push_back(M); + mvec.push_back(new MachineInstr(NOP)); break; } case 3: // stmt: Store(reg,reg) case 4: // stmt: Store(reg,ptrreg) - mvec[0] = new MachineInstr( - ChooseStoreInstruction( - subtreeRoot->leftChild()->getValue()->getType())); - SetOperandsForMemInstr(mvec[0], subtreeRoot, target); + mvec.push_back(new MachineInstr( + ChooseStoreInstruction( + subtreeRoot->leftChild()->getValue()->getType()))); + SetOperandsForMemInstr(mvec, subtreeRoot, target); break; case 5: // stmt: BrUncond - mvec[0] = new MachineInstr(BA); - mvec[0]->SetMachineOperand(0, MachineOperand::MO_CCRegister, - (Value*)NULL); - mvec[0]->SetMachineOperand(1, MachineOperand::MO_PCRelativeDisp, - ((BranchInst*) subtreeRoot->getInstruction())->getSuccessor(0)); + M = new MachineInstr(BA); + M->SetMachineOperandVal(0, MachineOperand::MO_PCRelativeDisp, + cast(subtreeRoot->getInstruction())->getSuccessor(0)); + mvec.push_back(M); // delay slot - mvec[numInstr++] = new MachineInstr(NOP); + mvec.push_back(new MachineInstr(NOP)); break; case 206: // stmt: BrCond(setCCconst) @@ -1212,95 +1282,95 @@ GetInstructionsByRule(InstructionNode* subtreeRoot, InstrTreeNode* constNode = subtreeRoot->leftChild()->rightChild(); assert(constNode && constNode->getNodeType() ==InstrTreeNode::NTConstNode); - ConstPoolVal* constVal = (ConstPoolVal*) constNode->getValue(); + Constant *constVal = cast(constNode->getValue()); bool isValidConst; - - if ((constVal->getType()->isIntegral() - || constVal->getType()->isPointerType()) + + if ((constVal->getType()->isInteger() + || isa(constVal->getType())) && GetConstantValueAsSignedInt(constVal, isValidConst) == 0 && isValidConst) { - BranchInst* brInst=cast(subtreeRoot->getInstruction()); - // That constant is a zero after all... // Use the left child of setCC as the first argument! - mvec[0] = new MachineInstr(ChooseBprInstruction(subtreeRoot)); - mvec[0]->SetMachineOperand(0, MachineOperand::MO_VirtualRegister, - subtreeRoot->leftChild()->leftChild()->getValue()); - mvec[0]->SetMachineOperand(1, MachineOperand::MO_PCRelativeDisp, - brInst->getSuccessor(0)); - + // Mark the setCC node so that no code is generated for it. + InstructionNode* setCCNode = (InstructionNode*) + subtreeRoot->leftChild(); + assert(setCCNode->getOpLabel() == SetCCOp); + setCCNode->markFoldedIntoParent(); + + BranchInst* brInst=cast(subtreeRoot->getInstruction()); + + M = new MachineInstr(ChooseBprInstruction(subtreeRoot)); + M->SetMachineOperandVal(0, MachineOperand::MO_VirtualRegister, + setCCNode->leftChild()->getValue()); + M->SetMachineOperandVal(1, MachineOperand::MO_PCRelativeDisp, + brInst->getSuccessor(0)); + mvec.push_back(M); + // delay slot - mvec[numInstr++] = new MachineInstr(NOP); + mvec.push_back(new MachineInstr(NOP)); // false branch - int n = numInstr++; - mvec[n] = new MachineInstr(BA); - mvec[n]->SetMachineOperand(0, MachineOperand::MO_CCRegister, - (Value*) NULL); - mvec[n]->SetMachineOperand(1, MachineOperand::MO_PCRelativeDisp, - brInst->getSuccessor(1)); + M = new MachineInstr(BA); + M->SetMachineOperandVal(0, MachineOperand::MO_PCRelativeDisp, + brInst->getSuccessor(1)); + mvec.push_back(M); // delay slot - mvec[numInstr++] = new MachineInstr(NOP); + mvec.push_back(new MachineInstr(NOP)); break; } // ELSE FALL THROUGH } - case 6: // stmt: BrCond(bool) - { // bool => boolean was computed with some boolean operator - // (SetCC, Not, ...). We need to check whether the type was a FP, - // signed int or unsigned int, and check the branching condition in - // order to choose the branch to use. + case 6: // stmt: BrCond(setCC) + { // bool => boolean was computed with SetCC. + // The branch to use depends on whether it is FP, signed, or unsigned. // If it is an integer CC, we also need to find the unique // TmpInstruction representing that CC. // BranchInst* brInst = cast(subtreeRoot->getInstruction()); bool isFPBranch; - mvec[0] = new MachineInstr(ChooseBccInstruction(subtreeRoot, - isFPBranch)); - + M = new MachineInstr(ChooseBccInstruction(subtreeRoot, isFPBranch)); + Value* ccValue = GetTmpForCC(subtreeRoot->leftChild()->getValue(), brInst->getParent()->getParent(), isFPBranch? Type::FloatTy : Type::IntTy); - mvec[0]->SetMachineOperand(0, MachineOperand::MO_CCRegister, ccValue); - mvec[0]->SetMachineOperand(1, MachineOperand::MO_PCRelativeDisp, - brInst->getSuccessor(0)); - + M->SetMachineOperandVal(0, MachineOperand::MO_CCRegister, ccValue); + M->SetMachineOperandVal(1, MachineOperand::MO_PCRelativeDisp, + brInst->getSuccessor(0)); + mvec.push_back(M); + // delay slot - mvec[numInstr++] = new MachineInstr(NOP); - + mvec.push_back(new MachineInstr(NOP)); + // false branch - int n = numInstr++; - mvec[n] = new MachineInstr(BA); - mvec[n]->SetMachineOperand(0, MachineOperand::MO_CCRegister, - (Value*) NULL); - mvec[n]->SetMachineOperand(1, MachineOperand::MO_PCRelativeDisp, - brInst->getSuccessor(1)); - + M = new MachineInstr(BA); + M->SetMachineOperandVal(0, MachineOperand::MO_PCRelativeDisp, + brInst->getSuccessor(1)); + mvec.push_back(M); + // delay slot - mvec[numInstr++] = new MachineInstr(NOP); + mvec.push_back(new MachineInstr(NOP)); break; } case 208: // stmt: BrCond(boolconst) { // boolconst => boolean is a constant; use BA to first or second label - ConstPoolVal* constVal = - cast(subtreeRoot->leftChild()->getValue()); - unsigned dest = ((ConstPoolBool*) constVal)->getValue()? 0 : 1; + Constant* constVal = + cast(subtreeRoot->leftChild()->getValue()); + unsigned dest = cast(constVal)->getValue()? 0 : 1; - mvec[0] = new MachineInstr(BA); - mvec[0]->SetMachineOperand(0, MachineOperand::MO_CCRegister, - (Value*) NULL); - mvec[0]->SetMachineOperand(1, MachineOperand::MO_PCRelativeDisp, - ((BranchInst*) subtreeRoot->getInstruction())->getSuccessor(dest)); + M = new MachineInstr(BA); + M->SetMachineOperandVal(0, MachineOperand::MO_PCRelativeDisp, + cast(subtreeRoot->getInstruction())->getSuccessor(dest)); + mvec.push_back(M); // delay slot - mvec[numInstr++] = new MachineInstr(NOP); + mvec.push_back(new MachineInstr(NOP)); break; } @@ -1308,162 +1378,176 @@ GetInstructionsByRule(InstructionNode* subtreeRoot, { // boolreg => boolean is stored in an existing register. // Just use the branch-on-integer-register instruction! // - mvec[0] = new MachineInstr(BRNZ); - mvec[0]->SetMachineOperand(0, MachineOperand::MO_VirtualRegister, + M = new MachineInstr(BRNZ); + M->SetMachineOperandVal(0, MachineOperand::MO_VirtualRegister, subtreeRoot->leftChild()->getValue()); - mvec[0]->SetMachineOperand(1, MachineOperand::MO_PCRelativeDisp, - ((BranchInst*) subtreeRoot->getInstruction())->getSuccessor(0)); + M->SetMachineOperandVal(1, MachineOperand::MO_PCRelativeDisp, + cast(subtreeRoot->getInstruction())->getSuccessor(0)); + mvec.push_back(M); // delay slot - mvec[numInstr++] = new MachineInstr(NOP); // delay slot + mvec.push_back(new MachineInstr(NOP)); // false branch - int n = numInstr++; - mvec[n] = new MachineInstr(BA); - mvec[n]->SetMachineOperand(0, MachineOperand::MO_CCRegister, - (Value*) NULL); - mvec[n]->SetMachineOperand(1, MachineOperand::MO_PCRelativeDisp, - ((BranchInst*) subtreeRoot->getInstruction())->getSuccessor(1)); + M = new MachineInstr(BA); + M->SetMachineOperandVal(0, MachineOperand::MO_PCRelativeDisp, + cast(subtreeRoot->getInstruction())->getSuccessor(1)); + mvec.push_back(M); // delay slot - mvec[numInstr++] = new MachineInstr(NOP); + mvec.push_back(new MachineInstr(NOP)); break; } case 9: // stmt: Switch(reg) assert(0 && "*** SWITCH instruction is not implemented yet."); - numInstr = 0; break; case 10: // reg: VRegList(reg, reg) assert(0 && "VRegList should never be the topmost non-chain rule"); break; - case 21: // bool: Not(bool): Both these are implemented as: - case 321: // reg: BNot(reg) : reg = reg XOR-NOT 0 - mvec[0] = new MachineInstr(XNOR); - mvec[0]->SetMachineOperand(0, MachineOperand::MO_VirtualRegister, - subtreeRoot->leftChild()->getValue()); - mvec[0]->SetMachineOperand(1, target.getRegInfo().getZeroRegNum()); - mvec[0]->SetMachineOperand(2, MachineOperand::MO_VirtualRegister, - subtreeRoot->getValue()); + case 21: // bool: Not(bool,reg): Both these are implemented as: + case 421: // reg: BNot(reg,reg): reg = reg XOR-NOT 0 + { // First find the unary operand. It may be left or right, usually right. + Value* notArg = BinaryOperator::getNotArgument( + cast(subtreeRoot->getInstruction())); + mvec.push_back(Create3OperandInstr_Reg(XNOR, notArg, + target.getRegInfo().getZeroRegNum(), + subtreeRoot->getValue())); break; + } - case 322: // reg: ToBoolTy(bool): case 22: // reg: ToBoolTy(reg): { const Type* opType = subtreeRoot->leftChild()->getValue()->getType(); - assert(opType->isIntegral() || opType->isPointerType() - || opType == Type::BoolTy); - numInstr = 0; - forwardOperandNum = 0; + assert(opType->isIntegral() || isa(opType)); + forwardOperandNum = 0; // forward first operand to user break; } case 23: // reg: ToUByteTy(reg) + case 24: // reg: ToSByteTy(reg) case 25: // reg: ToUShortTy(reg) + case 26: // reg: ToShortTy(reg) case 27: // reg: ToUIntTy(reg) - case 29: // reg: ToULongTy(reg) + case 28: // reg: ToIntTy(reg) { - const Type* opType = subtreeRoot->leftChild()->getValue()->getType(); - assert(opType->isIntegral() || - opType->isPointerType() || - opType == Type::BoolTy && "Cast is illegal for other types"); - numInstr = 0; - forwardOperandNum = 0; + //====================================================================== + // Rules for integer conversions: + // + //-------- + // From ISO 1998 C++ Standard, Sec. 4.7: + // + // 2. If the destination type is unsigned, the resulting value is + // the least unsigned integer congruent to the source integer + // (modulo 2n where n is the number of bits used to represent the + // unsigned type). [Note: In a two s complement representation, + // this conversion is conceptual and there is no change in the + // bit pattern (if there is no truncation). ] + // + // 3. If the destination type is signed, the value is unchanged if + // it can be represented in the destination type (and bitfield width); + // otherwise, the value is implementation-defined. + //-------- + // + // Since we assume 2s complement representations, this implies: + // + // -- if operand is smaller than destination, zero-extend or sign-extend + // according to the signedness of the *operand*: source decides. + // ==> we have to do nothing here! + // + // -- if operand is same size as or larger than destination, and the + // destination is *unsigned*, zero-extend the operand: dest. decides + // + // -- if operand is same size as or larger than destination, and the + // destination is *signed*, the choice is implementation defined: + // we sign-extend the operand: i.e., again dest. decides. + // Note: this matches both Sun's cc and gcc3.2. + //====================================================================== + + Instruction* destI = subtreeRoot->getInstruction(); + Value* opVal = subtreeRoot->leftChild()->getValue(); + const Type* opType = opVal->getType(); + if (opType->isIntegral() || isa(opType)) + { + unsigned opSize = target.DataLayout.getTypeSize(opType); + unsigned destSize = target.DataLayout.getTypeSize(destI->getType()); + if (opSize >= destSize) + { // Operand is same size as or larger than dest: + // zero- or sign-extend, according to the signeddness of + // the destination (see above). + if (destI->getType()->isSigned()) + target.getInstrInfo().CreateSignExtensionInstructions(target, + destI->getParent()->getParent(), opVal, destI, 8*destSize, + mvec, MachineCodeForInstruction::get(destI)); + else + target.getInstrInfo().CreateZeroExtensionInstructions(target, + destI->getParent()->getParent(), opVal, destI, 8*destSize, + mvec, MachineCodeForInstruction::get(destI)); + } + else + forwardOperandNum = 0; // forward first operand to user + } + else if (opType->isFloatingPoint()) + { + CreateCodeToConvertFloatToInt(target, opVal, destI, mvec, + MachineCodeForInstruction::get(destI)); + if (destI->getType()->isUnsigned()) + maskUnsignedResult = true; // not handled by fp->int code + } + else + assert(0 && "Unrecognized operand type for convert-to-unsigned"); + break; } - - case 24: // reg: ToSByteTy(reg) - case 26: // reg: ToShortTy(reg) - case 28: // reg: ToIntTy(reg) + + case 29: // reg: ToULongTy(reg) case 30: // reg: ToLongTy(reg) { - const Type* opType = subtreeRoot->leftChild()->getValue()->getType(); - if (opType->isIntegral() - || opType->isPointerType() - || opType == Type::BoolTy) + Value* opVal = subtreeRoot->leftChild()->getValue(); + const Type* opType = opVal->getType(); + if (opType->isIntegral() || isa(opType)) + forwardOperandNum = 0; // forward first operand to user + else if (opType->isFloatingPoint()) { - numInstr = 0; - forwardOperandNum = 0; + Instruction* destI = subtreeRoot->getInstruction(); + CreateCodeToConvertFloatToInt(target, opVal, destI, mvec, + MachineCodeForInstruction::get(destI)); } else - { - // If the source operand is an FP type, the int result must be - // copied from float to int register via memory! - Instruction *dest = subtreeRoot->getInstruction(); - Value* leftVal = subtreeRoot->leftChild()->getValue(); - Value* destForCast; - vector minstrVec; - - if (opType == Type::FloatTy || opType == Type::DoubleTy) - { - // Create a temporary to represent the INT register - // into which the FP value will be copied via memory. - // The type of this temporary will determine the FP - // register used: single-prec for a 32-bit int or smaller, - // double-prec for a 64-bit int. - // - const Type* destTypeToUse = - (dest->getType() == Type::LongTy)? Type::DoubleTy - : Type::FloatTy; - destForCast = new TmpInstruction(TMP_INSTRUCTION_OPCODE, - destTypeToUse, leftVal, NULL); - dest->getMachineInstrVec().addTempValue(destForCast); - - vector tempVec; - target.getInstrInfo().CreateCodeToCopyFloatToInt( - dest->getParent()->getParent(), - (TmpInstruction*) destForCast, dest, - minstrVec, tempVec, target); - - for (unsigned i=0; i < tempVec.size(); ++i) - dest->getMachineInstrVec().addTempValue(tempVec[i]); - } - else - destForCast = leftVal; - - MachineOpCode opCode=ChooseConvertToIntInstr(subtreeRoot, opType); - assert(opCode != INVALID_OPCODE && "Expected to need conversion!"); - - mvec[0] = new MachineInstr(opCode); - mvec[0]->SetMachineOperand(0, MachineOperand::MO_VirtualRegister, - leftVal); - mvec[0]->SetMachineOperand(1, MachineOperand::MO_VirtualRegister, - destForCast); - - assert(numInstr == 1 && "Should be initialized to 1 at the top"); - for (unsigned i=0; i < minstrVec.size(); ++i) - mvec[numInstr++] = minstrVec[i]; - } + assert(0 && "Unrecognized operand type for convert-to-signed"); break; - } + } case 31: // reg: ToFloatTy(reg): case 32: // reg: ToDoubleTy(reg): case 232: // reg: ToDoubleTy(Constant): - + // If this instruction has a parent (a user) in the tree // and the user is translated as an FsMULd instruction, // then the cast is unnecessary. So check that first. // In the future, we'll want to do the same for the FdMULq instruction, // so do the check here instead of only for ToFloatTy(reg). // - if (subtreeRoot->parent() != NULL && - ((InstructionNode*) subtreeRoot->parent())->getInstruction()->getMachineInstrVec()[0]->getOpCode() == FSMULD) + if (subtreeRoot->parent() != NULL) { - numInstr = 0; - forwardOperandNum = 0; + const MachineCodeForInstruction& mcfi = + MachineCodeForInstruction::get( + cast(subtreeRoot->parent())->getInstruction()); + if (mcfi.size() == 0 || mcfi.front()->getOpCode() == FSMULD) + forwardOperandNum = 0; // forward first operand to user } - else + + if (forwardOperandNum != 0) // we do need the cast { Value* leftVal = subtreeRoot->leftChild()->getValue(); const Type* opType = leftVal->getType(); - MachineOpCode opCode=ChooseConvertToFloatInstr(subtreeRoot,opType); + MachineOpCode opCode=ChooseConvertToFloatInstr( + subtreeRoot->getOpLabel(), opType); if (opCode == INVALID_OPCODE) // no conversion needed { - numInstr = 0; - forwardOperandNum = 0; + forwardOperandNum = 0; // forward first operand to user } else { @@ -1472,7 +1556,7 @@ GetInstructionsByRule(InstructionNode* subtreeRoot, Instruction *dest = subtreeRoot->getInstruction(); Value* srcForCast; int n = 0; - if (opType != Type::FloatTy && opType != Type::DoubleTy) + if (! opType->isFloatingPoint()) { // Create a temporary to represent the FP register // into which the integer will be copied via memory. @@ -1480,68 +1564,69 @@ GetInstructionsByRule(InstructionNode* subtreeRoot, // register used: single-prec for a 32-bit int or smaller, // double-prec for a 64-bit int. // - const Type* srcTypeToUse = - (leftVal->getType() == Type::LongTy)? Type::DoubleTy - : Type::FloatTy; - - srcForCast = new TmpInstruction(TMP_INSTRUCTION_OPCODE, - srcTypeToUse, dest, NULL); - dest->getMachineInstrVec().addTempValue(srcForCast); - - vector minstrVec; - vector tempVec; - target.getInstrInfo().CreateCodeToCopyIntToFloat( + uint64_t srcSize = + target.DataLayout.getTypeSize(leftVal->getType()); + Type* tmpTypeToUse = + (srcSize <= 4)? Type::FloatTy : Type::DoubleTy; + srcForCast = new TmpInstruction(tmpTypeToUse, dest); + MachineCodeForInstruction &destMCFI = + MachineCodeForInstruction::get(dest); + destMCFI.addTemp(srcForCast); + + target.getInstrInfo().CreateCodeToCopyIntToFloat(target, dest->getParent()->getParent(), - leftVal, (TmpInstruction*) srcForCast, - minstrVec, tempVec, target); - - for (unsigned i=0; i < minstrVec.size(); ++i) - mvec[n++] = minstrVec[i]; - - for (unsigned i=0; i < tempVec.size(); ++i) - dest->getMachineInstrVec().addTempValue(tempVec[i]); + leftVal, cast(srcForCast), + mvec, destMCFI); } else srcForCast = leftVal; - - MachineInstr* castI = new MachineInstr(opCode); - castI->SetMachineOperand(0, MachineOperand::MO_VirtualRegister, - srcForCast); - castI->SetMachineOperand(1, MachineOperand::MO_VirtualRegister, - dest); - mvec[n++] = castI; - numInstr = n; + + M = new MachineInstr(opCode); + M->SetMachineOperandVal(0, MachineOperand::MO_VirtualRegister, + srcForCast); + M->SetMachineOperandVal(1, MachineOperand::MO_VirtualRegister, + dest); + mvec.push_back(M); } } break; case 19: // reg: ToArrayTy(reg): case 20: // reg: ToPointerTy(reg): - numInstr = 0; - forwardOperandNum = 0; + forwardOperandNum = 0; // forward first operand to user break; case 233: // reg: Add(reg, Constant) - mvec[0] = CreateAddConstInstruction(subtreeRoot); - if (mvec[0] != NULL) - break; + maskUnsignedResult = true; + M = CreateAddConstInstruction(subtreeRoot); + if (M != NULL) + { + mvec.push_back(M); + break; + } // ELSE FALL THROUGH - + case 33: // reg: Add(reg, reg) - mvec[0] = new MachineInstr(ChooseAddInstruction(subtreeRoot)); - Set3OperandsFromInstr(mvec[0], subtreeRoot, target); + maskUnsignedResult = true; + mvec.push_back(new MachineInstr(ChooseAddInstruction(subtreeRoot))); + Set3OperandsFromInstr(mvec.back(), subtreeRoot, target); break; case 234: // reg: Sub(reg, Constant) - mvec[0] = CreateSubConstInstruction(subtreeRoot); - if (mvec[0] != NULL) - break; + maskUnsignedResult = true; + M = CreateSubConstInstruction(subtreeRoot); + if (M != NULL) + { + mvec.push_back(M); + break; + } // ELSE FALL THROUGH - + case 34: // reg: Sub(reg, reg) - mvec[0] = new MachineInstr(ChooseSubInstructionByType( - subtreeRoot->getInstruction()->getType())); - Set3OperandsFromInstr(mvec[0], subtreeRoot, target); + maskUnsignedResult = true; + mvec.push_back(new MachineInstr(ChooseSubInstructionByType( + subtreeRoot->getInstruction()->getType()))); + Set3OperandsFromInstr(mvec.back(), subtreeRoot, target); break; case 135: // reg: Mul(todouble, todouble) @@ -1549,73 +1634,82 @@ GetInstructionsByRule(InstructionNode* subtreeRoot, // FALL THROUGH case 35: // reg: Mul(reg, reg) - mvec[0] =new MachineInstr(ChooseMulInstruction(subtreeRoot,checkCast)); - Set3OperandsFromInstr(mvec[0], subtreeRoot, target); + { + maskUnsignedResult = true; + MachineOpCode forceOp = ((checkCast && BothFloatToDouble(subtreeRoot)) + ? FSMULD + : INVALID_MACHINE_OPCODE); + Instruction* mulInstr = subtreeRoot->getInstruction(); + CreateMulInstruction(target, mulInstr->getParent()->getParent(), + subtreeRoot->leftChild()->getValue(), + subtreeRoot->rightChild()->getValue(), + mulInstr, mvec, + MachineCodeForInstruction::get(mulInstr),forceOp); break; - + } case 335: // reg: Mul(todouble, todoubleConst) checkCast = true; // FALL THROUGH case 235: // reg: Mul(reg, Constant) - mvec[0] = CreateMulConstInstruction(target, subtreeRoot, mvec[1]); - if (mvec[0] == NULL) - { - mvec[0] = new MachineInstr(ChooseMulInstruction(subtreeRoot, - checkCast)); - Set3OperandsFromInstr(mvec[0], subtreeRoot, target); - } - else - if (mvec[1] != NULL) - ++numInstr; + { + maskUnsignedResult = true; + MachineOpCode forceOp = ((checkCast && BothFloatToDouble(subtreeRoot)) + ? FSMULD + : INVALID_MACHINE_OPCODE); + Instruction* mulInstr = subtreeRoot->getInstruction(); + CreateMulInstruction(target, mulInstr->getParent()->getParent(), + subtreeRoot->leftChild()->getValue(), + subtreeRoot->rightChild()->getValue(), + mulInstr, mvec, + MachineCodeForInstruction::get(mulInstr), + forceOp); break; - + } case 236: // reg: Div(reg, Constant) - mvec[0] = CreateDivConstInstruction(target, subtreeRoot, mvec[1]); - if (mvec[0] != NULL) - { - if (mvec[1] != NULL) - ++numInstr; - } - else + maskUnsignedResult = true; + L = mvec.size(); + CreateDivConstInstruction(target, subtreeRoot, mvec); + if (mvec.size() > L) + break; // ELSE FALL THROUGH - + case 36: // reg: Div(reg, reg) - mvec[0] = new MachineInstr(ChooseDivInstruction(target, subtreeRoot)); - Set3OperandsFromInstr(mvec[0], subtreeRoot, target); + maskUnsignedResult = true; + mvec.push_back(new MachineInstr(ChooseDivInstruction(target, subtreeRoot))); + Set3OperandsFromInstr(mvec.back(), subtreeRoot, target); break; case 37: // reg: Rem(reg, reg) case 237: // reg: Rem(reg, Constant) { + maskUnsignedResult = true; Instruction* remInstr = subtreeRoot->getInstruction(); - TmpInstruction* quot = new TmpInstruction(TMP_INSTRUCTION_OPCODE, + TmpInstruction* quot = new TmpInstruction( subtreeRoot->leftChild()->getValue(), subtreeRoot->rightChild()->getValue()); - TmpInstruction* prod = new TmpInstruction(TMP_INSTRUCTION_OPCODE, + TmpInstruction* prod = new TmpInstruction( quot, subtreeRoot->rightChild()->getValue()); - remInstr->getMachineInstrVec().addTempValue(quot); - remInstr->getMachineInstrVec().addTempValue(prod); + MachineCodeForInstruction::get(remInstr).addTemp(quot).addTemp(prod); - mvec[0] = new MachineInstr(ChooseDivInstruction(target, subtreeRoot)); - Set3OperandsFromInstr(mvec[0], subtreeRoot, target); - mvec[0]->SetMachineOperand(2, MachineOperand::MO_VirtualRegister,quot); + M = new MachineInstr(ChooseDivInstruction(target, subtreeRoot)); + Set3OperandsFromInstr(M, subtreeRoot, target); + M->SetMachineOperandVal(2, MachineOperand::MO_VirtualRegister,quot); + mvec.push_back(M); - int n = numInstr++; - mvec[n] = new MachineInstr(ChooseMulInstructionByType( - subtreeRoot->getInstruction()->getType())); - mvec[n]->SetMachineOperand(0, MachineOperand::MO_VirtualRegister,quot); - mvec[n]->SetMachineOperand(1, MachineOperand::MO_VirtualRegister, - subtreeRoot->rightChild()->getValue()); - mvec[n]->SetMachineOperand(2, MachineOperand::MO_VirtualRegister,prod); + M = Create3OperandInstr(ChooseMulInstructionByType( + subtreeRoot->getInstruction()->getType()), + quot, subtreeRoot->rightChild()->getValue(), + prod); + mvec.push_back(M); - n = numInstr++; - mvec[n] = new MachineInstr(ChooseSubInstructionByType( + M = new MachineInstr(ChooseSubInstructionByType( subtreeRoot->getInstruction()->getType())); - Set3OperandsFromInstr(mvec[n], subtreeRoot, target); - mvec[n]->SetMachineOperand(1, MachineOperand::MO_VirtualRegister,prod); + Set3OperandsFromInstr(M, subtreeRoot, target); + M->SetMachineOperandVal(1, MachineOperand::MO_VirtualRegister,prod); + mvec.push_back(M); break; } @@ -1624,81 +1718,82 @@ GetInstructionsByRule(InstructionNode* subtreeRoot, case 238: // bool: And(bool, boolconst) case 338: // reg : BAnd(reg, reg) case 538: // reg : BAnd(reg, Constant) - mvec[0] = new MachineInstr(AND); - Set3OperandsFromInstr(mvec[0], subtreeRoot, target); + mvec.push_back(new MachineInstr(AND)); + Set3OperandsFromInstr(mvec.back(), subtreeRoot, target); break; case 138: // bool: And(bool, not) - case 438: // bool: BAnd(bool, not) - mvec[0] = new MachineInstr(ANDN); - Set3OperandsFromInstr(mvec[0], subtreeRoot, target); + case 438: // bool: BAnd(bool, bnot) + { // Use the argument of NOT as the second argument! + // Mark the NOT node so that no code is generated for it. + InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild(); + Value* notArg = BinaryOperator::getNotArgument( + cast(notNode->getInstruction())); + notNode->markFoldedIntoParent(); + mvec.push_back(Create3OperandInstr(ANDN, + subtreeRoot->leftChild()->getValue(), + notArg, subtreeRoot->getValue())); break; + } case 39: // bool: Or(bool, bool) case 239: // bool: Or(bool, boolconst) case 339: // reg : BOr(reg, reg) case 539: // reg : BOr(reg, Constant) - mvec[0] = new MachineInstr(ORN); - Set3OperandsFromInstr(mvec[0], subtreeRoot, target); + mvec.push_back(new MachineInstr(OR)); + Set3OperandsFromInstr(mvec.back(), subtreeRoot, target); break; case 139: // bool: Or(bool, not) - case 439: // bool: BOr(bool, not) - mvec[0] = new MachineInstr(ORN); - Set3OperandsFromInstr(mvec[0], subtreeRoot, target); + case 439: // bool: BOr(bool, bnot) + { // Use the argument of NOT as the second argument! + // Mark the NOT node so that no code is generated for it. + InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild(); + Value* notArg = BinaryOperator::getNotArgument( + cast(notNode->getInstruction())); + notNode->markFoldedIntoParent(); + mvec.push_back(Create3OperandInstr(ORN, + subtreeRoot->leftChild()->getValue(), + notArg, subtreeRoot->getValue())); break; + } case 40: // bool: Xor(bool, bool) case 240: // bool: Xor(bool, boolconst) case 340: // reg : BXor(reg, reg) case 540: // reg : BXor(reg, Constant) - mvec[0] = new MachineInstr(XOR); - Set3OperandsFromInstr(mvec[0], subtreeRoot, target); + mvec.push_back(new MachineInstr(XOR)); + Set3OperandsFromInstr(mvec.back(), subtreeRoot, target); break; case 140: // bool: Xor(bool, not) - case 440: // bool: BXor(bool, not) - mvec[0] = new MachineInstr(XNOR); - Set3OperandsFromInstr(mvec[0], subtreeRoot, target); + case 440: // bool: BXor(bool, bnot) + { // Use the argument of NOT as the second argument! + // Mark the NOT node so that no code is generated for it. + InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild(); + Value* notArg = BinaryOperator::getNotArgument( + cast(notNode->getInstruction())); + notNode->markFoldedIntoParent(); + mvec.push_back(Create3OperandInstr(XNOR, + subtreeRoot->leftChild()->getValue(), + notArg, subtreeRoot->getValue())); break; + } case 41: // boolconst: SetCC(reg, Constant) - // Check if this is an integer comparison, and - // there is a parent, and the parent decided to use - // a branch-on-integer-register instead of branch-on-condition-code. - // If so, the SUBcc instruction is not required. - // (However, we must still check for constants to be loaded from - // the constant pool so that such a load can be associated with - // this instruction.) // - // Otherwise this is just the same as case 42, so just fall through. + // If the SetCC was folded into the user (parent), it will be + // caught above. All other cases are the same as case 42, + // so just fall through. // - if ((subtreeRoot->leftChild()->getValue()->getType()->isIntegral() || - subtreeRoot->leftChild()->getValue()->getType()->isPointerType()) - && subtreeRoot->parent() != NULL) - { - InstructionNode* parent = (InstructionNode*) subtreeRoot->parent(); - assert(parent->getNodeType() == InstrTreeNode::NTInstructionNode); - const vector& - minstrVec = parent->getInstruction()->getMachineInstrVec(); - MachineOpCode parentOpCode; - if (parent->getInstruction()->getOpcode() == Instruction::Br && - (parentOpCode = minstrVec[0]->getOpCode()) >= BRZ && - parentOpCode <= BRGEZ) - { - numInstr = 0; // don't forward the operand! - break; - } - } - // ELSE FALL THROUGH - case 42: // bool: SetCC(reg, reg): { // This generates a SUBCC instruction, putting the difference in // a result register, and setting a condition code. // // If the boolean result of the SetCC is used by anything other - // than a single branch instruction, the boolean must be + // than a branch instruction, or if it is used outside the current + // basic block, the boolean must be // computed and stored in the result register. Otherwise, discard // the difference (by using %g0) and keep only the condition code. // @@ -1709,9 +1804,9 @@ GetInstructionsByRule(InstructionNode* subtreeRoot, // InstructionNode* parentNode = (InstructionNode*) subtreeRoot->parent(); Instruction* setCCInstr = subtreeRoot->getInstruction(); - bool keepBoolVal = (parentNode == NULL || - parentNode->getInstruction()->getOpcode() - != Instruction::Br); + + bool keepBoolVal = parentNode == NULL || + ! AllUsesAreBranches(setCCInstr); bool subValIsBoolVal = setCCInstr->getOpcode() == Instruction::SetNE; bool keepSubVal = keepBoolVal && subValIsBoolVal; bool computeBoolVal = keepBoolVal && ! subValIsBoolVal; @@ -1719,7 +1814,7 @@ GetInstructionsByRule(InstructionNode* subtreeRoot, bool mustClearReg; int valueToMove; MachineOpCode movOpCode = 0; - + // Mark the 4th operand as being a CC register, and as a def // A TmpInstruction is created to represent the CC "result". // Unlike other instances of TmpInstruction, this one is used @@ -1732,13 +1827,12 @@ GetInstructionsByRule(InstructionNode* subtreeRoot, // a FP condition code register. // Value* leftVal = subtreeRoot->leftChild()->getValue(); - bool isFPCompare = (leftVal->getType() == Type::FloatTy || - leftVal->getType() == Type::DoubleTy); + bool isFPCompare = leftVal->getType()->isFloatingPoint(); TmpInstruction* tmpForCC = GetTmpForCC(setCCInstr, setCCInstr->getParent()->getParent(), - isFPCompare? Type::FloatTy : Type::IntTy); - setCCInstr->getMachineInstrVec().addTempValue(tmpForCC); + isFPCompare ? Type::FloatTy : Type::IntTy); + MachineCodeForInstruction::get(setCCInstr).addTemp(tmpForCC); if (! isFPCompare) { @@ -1747,11 +1841,11 @@ GetInstructionsByRule(InstructionNode* subtreeRoot, // a separate instruction to compute the bool result, so discard // result of SUBcc instruction anyway. // - mvec[0] = new MachineInstr(SUBcc); - Set3OperandsFromInstr(mvec[0], subtreeRoot, target, ! keepSubVal); - - mvec[0]->SetMachineOperand(3, MachineOperand::MO_CCRegister, - tmpForCC, /*def*/true); + M = new MachineInstr(SUBcc); + Set3OperandsFromInstr(M, subtreeRoot, target, ! keepSubVal); + M->SetMachineOperandVal(3, MachineOperand::MO_CCRegister, + tmpForCC, /*def*/true); + mvec.push_back(M); if (computeBoolVal) { // recompute bool using the integer condition codes @@ -1762,13 +1856,14 @@ GetInstructionsByRule(InstructionNode* subtreeRoot, else { // FP condition: dest of FCMP should be some FCCn register - mvec[0] = new MachineInstr(ChooseFcmpInstruction(subtreeRoot)); - mvec[0]->SetMachineOperand(0, MachineOperand::MO_CCRegister, + M = new MachineInstr(ChooseFcmpInstruction(subtreeRoot)); + M->SetMachineOperandVal(0, MachineOperand::MO_CCRegister, tmpForCC); - mvec[0]->SetMachineOperand(1,MachineOperand::MO_VirtualRegister, + M->SetMachineOperandVal(1,MachineOperand::MO_VirtualRegister, subtreeRoot->leftChild()->getValue()); - mvec[0]->SetMachineOperand(2,MachineOperand::MO_VirtualRegister, + M->SetMachineOperandVal(2,MachineOperand::MO_VirtualRegister, subtreeRoot->rightChild()->getValue()); + mvec.push_back(M); if (computeBoolVal) {// recompute bool using the FP condition codes @@ -1782,264 +1877,229 @@ GetInstructionsByRule(InstructionNode* subtreeRoot, { if (mustClearReg) {// Unconditionally set register to 0 - int n = numInstr++; - mvec[n] = new MachineInstr(SETHI); - mvec[n]->SetMachineOperand(0,MachineOperand::MO_UnextendedImmed, - s0); - mvec[n]->SetMachineOperand(1,MachineOperand::MO_VirtualRegister, - setCCInstr); + M = new MachineInstr(SETHI); + M->SetMachineOperandConst(0,MachineOperand::MO_UnextendedImmed, + (int64_t)0); + M->SetMachineOperandVal(1, MachineOperand::MO_VirtualRegister, + setCCInstr); + mvec.push_back(M); } // Now conditionally move `valueToMove' (0 or 1) into the register - int n = numInstr++; - mvec[n] = new MachineInstr(movOpCode); - mvec[n]->SetMachineOperand(0, MachineOperand::MO_CCRegister, - tmpForCC); - mvec[n]->SetMachineOperand(1, MachineOperand::MO_UnextendedImmed, - valueToMove); - mvec[n]->SetMachineOperand(2, MachineOperand::MO_VirtualRegister, - setCCInstr); + // Mark the register as a use (as well as a def) because the old + // value should be retained if the condition is false. + M = new MachineInstr(movOpCode); + M->SetMachineOperandVal(0, MachineOperand::MO_CCRegister, + tmpForCC); + M->SetMachineOperandConst(1, MachineOperand::MO_UnextendedImmed, + valueToMove); + M->SetMachineOperandVal(2, MachineOperand::MO_VirtualRegister, + setCCInstr, /*isDef*/ true, + /*isDefAndUse*/ true); + mvec.push_back(M); } break; } - case 43: // boolreg: VReg - case 44: // boolreg: Constant - numInstr = 0; - break; - case 51: // reg: Load(reg) case 52: // reg: Load(ptrreg) - case 53: // reg: LoadIdx(reg,reg) - case 54: // reg: LoadIdx(ptrreg,reg) - mvec[0] = new MachineInstr(ChooseLoadInstruction( - subtreeRoot->getValue()->getType())); - SetOperandsForMemInstr(mvec[0], subtreeRoot, target); + mvec.push_back(new MachineInstr(ChooseLoadInstruction( + subtreeRoot->getValue()->getType()))); + SetOperandsForMemInstr(mvec, subtreeRoot, target); break; case 55: // reg: GetElemPtr(reg) case 56: // reg: GetElemPtrIdx(reg,reg) - if (subtreeRoot->parent() != NULL) - { - // If the parent was a memory operation and not an array access, - // the parent will fold this instruction in so generate nothing. - // - Instruction* parent = - cast(subtreeRoot->parent()->getValue()); - if (parent->getOpcode() == Instruction::Load || - parent->getOpcode() == Instruction::Store || - parent->getOpcode() == Instruction::GetElementPtr) - { - // Check if the parent is an array access, - // If so, we still need to generate this instruction. - GetElementPtrInst* getElemInst = - cast(subtreeRoot->getInstruction()); - const PointerType* ptrType = - (const PointerType*) getElemInst->getPtrOperand()->getType(); - if (! ptrType->getValueType()->isArrayType()) - {// we don't need a separate instr - numInstr = 0; // don't forward operand! - break; - } - } - } - // else in all other cases we need to a separate ADD instruction - mvec[0] = new MachineInstr(ADD); - SetOperandsForMemInstr(mvec[0], subtreeRoot, target); + // If the GetElemPtr was folded into the user (parent), it will be + // caught above. For other cases, we have to compute the address. + mvec.push_back(new MachineInstr(ADD)); + SetOperandsForMemInstr(mvec, subtreeRoot, target); break; case 57: // reg: Alloca: Implement as 1 instruction: { // add %fp, offsetFromFP -> result - Instruction* instr = subtreeRoot->getInstruction(); - const PointerType* instrType = (const PointerType*) instr->getType(); - assert(instrType->isPointerType()); - int tsize = (int) - target.findOptimalStorageSize(instrType->getValueType()); - assert(tsize != 0 && "Just to check when this can happen"); - - Method* method = instr->getParent()->getParent(); - MachineCodeForMethod& mcInfo = MachineCodeForMethod::get(method); - int offsetFromFP = mcInfo.allocateLocalVar(target, instr); - - // Create a temporary Value to hold the constant offset. - // This is needed because it may not fit in the immediate field. - ConstPoolSInt* offsetVal=ConstPoolSInt::get(Type::IntTy, offsetFromFP); - - // Instruction 1: add %fp, offsetFromFP -> result - mvec[0] = new MachineInstr(ADD); - mvec[0]->SetMachineOperand(0, target.getRegInfo().getFramePointer()); - mvec[0]->SetMachineOperand(1, MachineOperand::MO_VirtualRegister, - offsetVal); - mvec[0]->SetMachineOperand(2, MachineOperand::MO_VirtualRegister, - instr); + AllocationInst* instr = + cast(subtreeRoot->getInstruction()); + unsigned int tsize = + target.DataLayout.getTypeSize(instr->getAllocatedType()); + assert(tsize != 0); + CreateCodeForFixedSizeAlloca(target, instr, tsize, 1, mvec); break; } - + case 58: // reg: Alloca(reg): Implement as 3 instructions: // mul num, typeSz -> tmp // sub %sp, tmp -> %sp { // add %sp, frameSizeBelowDynamicArea -> result - Instruction* instr = subtreeRoot->getInstruction(); - const PointerType* instrType = (const PointerType*) instr->getType(); - assert(instrType->isPointerType() && - instrType->getValueType()->isArrayType()); - const Type* eltType = - ((ArrayType*) instrType->getValueType())->getElementType(); - int tsize = (int) target.findOptimalStorageSize(eltType); - - assert(tsize != 0 && "Just to check when this can happen"); - - // Create a temporary Value to hold the constant type-size - ConstPoolSInt* tsizeVal = ConstPoolSInt::get(Type::IntTy, tsize); - - // Create a temporary Value to hold the constant offset from SP - Method* method = instr->getParent()->getParent(); - bool ignore; // we don't need this - ConstPoolSInt* dynamicAreaOffset = ConstPoolSInt::get(Type::IntTy, - target.getFrameInfo().getDynamicAreaOffset(MachineCodeForMethod::get(method), - ignore)); - - // Create a temporary value to hold `tmp' - Instruction* tmpInstr = new TmpInstruction(TMP_INSTRUCTION_OPCODE, - subtreeRoot->leftChild()->getValue(), - NULL /*could insert tsize here*/); - subtreeRoot->getInstruction()->getMachineInstrVec().addTempValue(tmpInstr); - - // Instruction 1: mul numElements, typeSize -> tmp - mvec[0] = new MachineInstr(MULX); - mvec[0]->SetMachineOperand(0, MachineOperand::MO_VirtualRegister, - subtreeRoot->leftChild()->getValue()); - mvec[0]->SetMachineOperand(1, MachineOperand::MO_VirtualRegister, - tsizeVal); - mvec[0]->SetMachineOperand(2, MachineOperand::MO_VirtualRegister, - tmpInstr); + AllocationInst* instr = + cast(subtreeRoot->getInstruction()); + const Type* eltType = instr->getAllocatedType(); - // Instruction 2: sub %sp, tmp -> %sp - numInstr++; - mvec[1] = new MachineInstr(SUB); - mvec[1]->SetMachineOperand(0, target.getRegInfo().getStackPointer()); - mvec[1]->SetMachineOperand(1, MachineOperand::MO_VirtualRegister, - tmpInstr); - mvec[1]->SetMachineOperand(2, target.getRegInfo().getStackPointer()); + // If #elements is constant, use simpler code for fixed-size allocas + int tsize = (int) target.DataLayout.getTypeSize(eltType); + Value* numElementsVal = NULL; + bool isArray = instr->isArrayAllocation(); - // Instruction 3: add %sp, frameSizeBelowDynamicArea -> result - numInstr++; - mvec[2] = new MachineInstr(ADD); - mvec[2]->SetMachineOperand(0, target.getRegInfo().getStackPointer()); - mvec[2]->SetMachineOperand(1, MachineOperand::MO_VirtualRegister, - dynamicAreaOffset); - mvec[2]->SetMachineOperand(2,MachineOperand::MO_VirtualRegister,instr); + if (!isArray || + isa(numElementsVal = instr->getArraySize())) + { // total size is constant: generate code for fixed-size alloca + unsigned int numElements = isArray? + cast(numElementsVal)->getValue() : 1; + CreateCodeForFixedSizeAlloca(target, instr, tsize, + numElements, mvec); + } + else // total size is not constant. + CreateCodeForVariableSizeAlloca(target, instr, tsize, + numElementsVal, mvec); break; } case 61: // reg: Call - { // Generate a call-indirect (i.e., jmpl) for now to expose - // the potential need for registers. If an absolute address - // is available, replace this with a CALL instruction. - // Mark both the indirection register and the return-address - // register as hidden virtual registers. - // Also, mark the operands of the Call and return value (if - // any) as implicit operands of the CALL machine instruction. + { // Generate a direct (CALL) or indirect (JMPL) call. + // Mark the return-address register, the indirection + // register (for indirect calls), the operands of the Call, + // and the return value (if any) as implicit operands + // of the machine instruction. + // + // If this is a varargs function, floating point arguments + // have to passed in integer registers so insert + // copy-float-to-int instructions for each float operand. // CallInst *callInstr = cast(subtreeRoot->getInstruction()); Value *callee = callInstr->getCalledValue(); - - Instruction* retAddrReg = new TmpInstruction(TMP_INSTRUCTION_OPCODE, - callInstr, NULL); - - // Note temporary values in the machineInstrVec for the VM instr. - // - // WARNING: Operands 0..N-1 must go in slots 0..N-1 of implicitUses. - // The result value must go in slot N. This is assumed - // in register allocation. - // - callInstr->getMachineInstrVec().addTempValue(retAddrReg); - - + + // Create hidden virtual register for return address with type void* + TmpInstruction* retAddrReg = + new TmpInstruction(PointerType::get(Type::VoidTy), callInstr); + MachineCodeForInstruction::get(callInstr).addTemp(retAddrReg); + // Generate the machine instruction and its operands. // Use CALL for direct function calls; this optimistically assumes // the PC-relative address fits in the CALL address field (22 bits). // Use JMPL for indirect calls. // - if (callee->getValueType() == Value::MethodVal) - { // direct function call - mvec[0] = new MachineInstr(CALL); - mvec[0]->SetMachineOperand(0, MachineOperand::MO_PCRelativeDisp, - callee); - } - else - { // indirect function call - mvec[0] = new MachineInstr(JMPLCALL); - mvec[0]->SetMachineOperand(0, MachineOperand::MO_VirtualRegister, - callee); - mvec[0]->SetMachineOperand(1, MachineOperand::MO_SignExtendedImmed, - (int64_t) 0); - mvec[0]->SetMachineOperand(2, MachineOperand::MO_VirtualRegister, - retAddrReg); - } + if (isa(callee)) // direct function call + M = Create1OperandInstr_Addr(CALL, callee); + else // indirect function call + M = Create3OperandInstr_SImmed(JMPLCALL, callee, + (int64_t) 0, retAddrReg); + mvec.push_back(M); + + const FunctionType* funcType = + cast(cast(callee->getType()) + ->getElementType()); + bool isVarArgs = funcType->isVarArg(); + bool noPrototype = isVarArgs && funcType->getNumParams() == 0; - // Add the call operands and return value as implicit refs - for (unsigned i=0, N=callInstr->getNumOperands(); i < N; ++i) - if (callInstr->getOperand(i) != callee) - mvec[0]->addImplicitRef(callInstr->getOperand(i)); + // Use an annotation to pass information about call arguments + // to the register allocator. + CallArgsDescriptor* argDesc = new CallArgsDescriptor(callInstr, + retAddrReg, isVarArgs, noPrototype); + M->addAnnotation(argDesc); + + assert(callInstr->getOperand(0) == callee + && "This is assumed in the loop below!"); + + for (unsigned i=1, N=callInstr->getNumOperands(); i < N; ++i) + { + Value* argVal = callInstr->getOperand(i); + Instruction* intArgReg = NULL; + + // Check for FP arguments to varargs functions. + // Any such argument in the first $K$ args must be passed in an + // integer register, where K = #integer argument registers. + if (isVarArgs && argVal->getType()->isFloatingPoint()) + { + // If it is a function with no prototype, pass value + // as an FP value as well as a varargs value + if (noPrototype) + argDesc->getArgInfo(i-1).setUseFPArgReg(); + + // If this arg. is in the first $K$ regs, add a copy + // float-to-int instruction to pass the value as an integer. + if (i <= target.getRegInfo().GetNumOfIntArgRegs()) + { + MachineCodeForInstruction &destMCFI = + MachineCodeForInstruction::get(callInstr); + intArgReg = new TmpInstruction(Type::IntTy, argVal); + destMCFI.addTemp(intArgReg); + + vector copyMvec; + target.getInstrInfo().CreateCodeToCopyFloatToInt(target, + callInstr->getParent()->getParent(), + argVal, (TmpInstruction*) intArgReg, + copyMvec, destMCFI); + mvec.insert(mvec.begin(),copyMvec.begin(),copyMvec.end()); + + argDesc->getArgInfo(i-1).setUseIntArgReg(); + argDesc->getArgInfo(i-1).setArgCopy(intArgReg); + } + else + // Cannot fit in first $K$ regs so pass the arg on the stack + argDesc->getArgInfo(i-1).setUseStackSlot(); + } + + if (intArgReg) + mvec.back()->addImplicitRef(intArgReg); + + mvec.back()->addImplicitRef(argVal); + } + // Add the return value as an implicit ref. The call operands + // were added above. if (callInstr->getType() != Type::VoidTy) - mvec[0]->addImplicitRef(callInstr, /*isDef*/ true); + mvec.back()->addImplicitRef(callInstr, /*isDef*/ true); // For the CALL instruction, the ret. addr. reg. is also implicit - if (callee->getValueType() == Value::MethodVal) - mvec[0]->addImplicitRef(retAddrReg, /*isDef*/ true); + if (isa(callee)) + mvec.back()->addImplicitRef(retAddrReg, /*isDef*/ true); - mvec[numInstr++] = new MachineInstr(NOP); // delay slot + // delay slot + mvec.push_back(new MachineInstr(NOP)); break; } - + case 62: // reg: Shl(reg, reg) - { const Type* opType = subtreeRoot->leftChild()->getValue()->getType(); - assert(opType->isIntegral() - || opType == Type::BoolTy - || opType->isPointerType()&& "Shl unsupported for other types"); - mvec[0] = new MachineInstr((opType == Type::LongTy)? SLLX : SLL); - Set3OperandsFromInstr(mvec[0], subtreeRoot, target); + { + Value* argVal1 = subtreeRoot->leftChild()->getValue(); + Value* argVal2 = subtreeRoot->rightChild()->getValue(); + Instruction* shlInstr = subtreeRoot->getInstruction(); + + const Type* opType = argVal1->getType(); + assert((opType->isInteger() || isa(opType)) && + "Shl unsupported for other types"); + + CreateShiftInstructions(target, shlInstr->getParent()->getParent(), + (opType == Type::LongTy)? SLLX : SLL, + argVal1, argVal2, 0, shlInstr, mvec, + MachineCodeForInstruction::get(shlInstr)); break; } case 63: // reg: Shr(reg, reg) { const Type* opType = subtreeRoot->leftChild()->getValue()->getType(); - assert(opType->isIntegral() - || opType == Type::BoolTy - || opType->isPointerType() &&"Shr unsupported for other types"); - mvec[0] = new MachineInstr((opType->isSigned() - ? ((opType == Type::LongTy)? SRAX : SRA) - : ((opType == Type::LongTy)? SRLX : SRL))); - Set3OperandsFromInstr(mvec[0], subtreeRoot, target); + assert((opType->isInteger() || isa(opType)) && + "Shr unsupported for other types"); + mvec.push_back(new MachineInstr((opType->isSigned() + ? ((opType == Type::LongTy)? SRAX : SRA) + : ((opType == Type::LongTy)? SRLX : SRL)))); + Set3OperandsFromInstr(mvec.back(), subtreeRoot, target); break; } case 64: // reg: Phi(reg,reg) - { // This instruction has variable #operands, so resultPos is 0. - Instruction* phi = subtreeRoot->getInstruction(); - mvec[0] = new MachineInstr(PHI, 1 + phi->getNumOperands()); - mvec[0]->SetMachineOperand(0, MachineOperand::MO_VirtualRegister, - subtreeRoot->getValue()); - for (unsigned i=0, N=phi->getNumOperands(); i < N; i++) - mvec[0]->SetMachineOperand(i+1, MachineOperand::MO_VirtualRegister, - phi->getOperand(i)); - break; - } - + break; // don't forward the value + case 71: // reg: VReg case 72: // reg: Constant - numInstr = 0; // don't forward the value - break; + break; // don't forward the value default: assert(0 && "Unrecognized BURG rule"); - numInstr = 0; break; } } - + if (forwardOperandNum >= 0) { // We did not generate a machine instruction but need to use operand. // If user is in the same tree, replace Value in its machine operand. @@ -2050,16 +2110,43 @@ GetInstructionsByRule(InstructionNode* subtreeRoot, else { vector minstrVec; - target.getInstrInfo().CreateCopyInstructionsByType(target, - subtreeRoot->getInstruction()->getOperand(forwardOperandNum), - subtreeRoot->getInstruction(), minstrVec); + Instruction* instr = subtreeRoot->getInstruction(); + target.getInstrInfo(). + CreateCopyInstructionsByType(target, + instr->getParent()->getParent(), + instr->getOperand(forwardOperandNum), + instr, minstrVec, + MachineCodeForInstruction::get(instr)); assert(minstrVec.size() > 0); - for (unsigned i=0; i < minstrVec.size(); ++i) - mvec[numInstr++] = minstrVec[i]; + mvec.insert(mvec.end(), minstrVec.begin(), minstrVec.end()); } } - - return numInstr; -} - + if (maskUnsignedResult) + { // If result is unsigned and smaller than int reg size, + // we need to clear high bits of result value. + assert(forwardOperandNum < 0 && "Need mask but no instruction generated"); + Instruction* dest = subtreeRoot->getInstruction(); + if (dest->getType()->isUnsigned()) + { + unsigned destSize = target.DataLayout.getTypeSize(dest->getType()); + if (destSize <= 4) + { // Mask high bits. Use a TmpInstruction to represent the + // intermediate result before masking. Since those instructions + // have already been generated, go back and substitute tmpI + // for dest in the result position of each one of them. + TmpInstruction *tmpI = new TmpInstruction(dest->getType(), dest, + NULL, "maskHi"); + MachineCodeForInstruction::get(dest).addTemp(tmpI); + + for (unsigned i=0, N=mvec.size(); i < N; ++i) + mvec[i]->substituteValue(dest, tmpI); + + M = Create3OperandInstr_UImmed(SRL, tmpI, 8*(4-destSize), dest); + mvec.push_back(M); + } + else if (destSize < target.DataLayout.getIntegerRegize()) + assert(0 && "Unsupported type size: 32 < size < 64 bits"); + } + } +}