//===-- SparcInstrSelection.cpp -------------------------------------------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file was developed by the LLVM research group and is distributed under
+// the University of Illinois Open Source License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
//
// BURS instruction selection for SPARC V9 architecture.
//
//===----------------------------------------------------------------------===//
-#include "SparcInternals.h"
#include "SparcInstrSelectionSupport.h"
+#include "SparcInternals.h"
#include "SparcRegClassInfo.h"
-#include "llvm/CodeGen/InstrSelectionSupport.h"
-#include "llvm/CodeGen/MachineInstrBuilder.h"
-#include "llvm/CodeGen/MachineInstrAnnot.h"
+#include "llvm/Constants.h"
+#include "llvm/ConstantHandling.h"
+#include "llvm/DerivedTypes.h"
+#include "llvm/Instructions.h"
+#include "llvm/Intrinsics.h"
+#include "llvm/Module.h"
#include "llvm/CodeGen/InstrForest.h"
#include "llvm/CodeGen/InstrSelection.h"
+#include "llvm/CodeGen/InstrSelectionSupport.h"
+#include "llvm/CodeGen/MachineCodeForInstruction.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFunctionInfo.h"
-#include "llvm/CodeGen/MachineCodeForInstruction.h"
-#include "llvm/DerivedTypes.h"
-#include "llvm/iTerminators.h"
-#include "llvm/iMemory.h"
-#include "llvm/iOther.h"
-#include "llvm/Function.h"
-#include "llvm/Constants.h"
-#include "llvm/ConstantHandling.h"
-#include "llvm/Intrinsics.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineInstrAnnot.h"
#include "Support/MathExtras.h"
-#include <math.h>
+#include <algorithm>
+#include <cmath>
static inline void Add3OperandInstr(unsigned Opcode, InstructionNode* Node,
std::vector<MachineInstr*>& mvec) {
}
-
-//---------------------------------------------------------------------------
-// Function: GetMemInstArgs
-//
-// Purpose:
-// Get the pointer value and the index vector for a memory operation
-// (GetElementPtr, Load, or Store). If all indices of the given memory
-// operation are constant, fold in constant indices in a chain of
-// preceding GetElementPtr instructions (if any), and return the
-// pointer value of the first instruction in the chain.
-// All folded instructions are marked so no code is generated for them.
-//
-// Return values:
-// Returns the pointer Value to use.
-// Returns the resulting IndexVector in idxVec.
-// Returns true/false in allConstantIndices if all indices are/aren't const.
-//---------------------------------------------------------------------------
-
-
//---------------------------------------------------------------------------
// Function: FoldGetElemChain
//
Value* ptrVal = gepI->getPointerOperand();
InstrTreeNode* ptrChild = gepNode->leftChild();
- // Extract the index vector of the GEP instructin.
+ // Extract the index vector of the GEP instruction.
// If all indices are constant and first index is zero, try to fold
// in preceding GEPs with all constant indices.
for (User::op_iterator OI=gepI->idx_begin(), OE=gepI->idx_end();
static inline MachineOpCode
ChooseBccInstruction(const InstructionNode* instrNode,
- bool& isFPBranch)
+ const Type*& setCCType)
{
InstructionNode* setCCNode = (InstructionNode*) instrNode->leftChild();
assert(setCCNode->getOpLabel() == SetCCOp);
BinaryOperator* setCCInstr =cast<BinaryOperator>(setCCNode->getInstruction());
- const Type* setCCType = setCCInstr->getOperand(0)->getType();
-
- isFPBranch = setCCType->isFloatingPoint(); // Return value: don't delete!
+ setCCType = setCCInstr->getOperand(0)->getType();
- if (isFPBranch)
+ if (setCCType->isFloatingPoint())
return ChooseBFpccInstruction(instrNode, setCCInstr);
else
return ChooseBpccInstruction(instrNode, setCCInstr);
}
-// Assumes that SUBcc v1, v2 -> v3 has been executed.
-// In most cases, we want to clear v3 and then follow it by instruction
-// MOVcc 1 -> v3.
-// Set mustClearReg=false if v3 need not be cleared before conditional move.
-// Set valueToMove=0 if we want to conditionally move 0 instead of 1
-// (i.e., we want to test inverse of a condition)
-// (The latter two cases do not seem to arise because SetNE needs nothing.)
+// ChooseMovpcciForSetCC -- Choose a conditional-move instruction
+// based on the type of SetCC operation.
//
// WARNING: since this function has only one caller, it always returns
// the opcode that expects an immediate and a register. If this function
//
// It will be necessary to expand convertOpcodeFromRegToImm() to handle the
// new cases of opcodes.
+//
static MachineOpCode
-ChooseMovpcciAfterSub(const InstructionNode* instrNode)
+ChooseMovpcciForSetCC(const InstructionNode* instrNode)
{
MachineOpCode opCode = V9::INVALID_OPCODE;
+
+ const Type* opType = instrNode->leftChild()->getValue()->getType();
+ assert(opType->isIntegral() || isa<PointerType>(opType));
+ bool noSign = opType->isUnsigned() || isa<PointerType>(opType);
switch(instrNode->getInstruction()->getOpcode())
{
- case Instruction::SetEQ: opCode = V9::MOVEi; break;
- case Instruction::SetLE: opCode = V9::MOVLEi; break;
- case Instruction::SetGE: opCode = V9::MOVGEi; break;
- case Instruction::SetLT: opCode = V9::MOVLi; break;
- case Instruction::SetGT: opCode = V9::MOVGi; break;
- case Instruction::SetNE: opCode = V9::MOVNEi; break;
+ case Instruction::SetEQ: opCode = V9::MOVEi; break;
+ case Instruction::SetLE: opCode = noSign? V9::MOVLEUi : V9::MOVLEi; break;
+ case Instruction::SetGE: opCode = noSign? V9::MOVCCi : V9::MOVGEi; break;
+ case Instruction::SetLT: opCode = noSign? V9::MOVCSi : V9::MOVLi; break;
+ case Instruction::SetGT: opCode = noSign? V9::MOVGUi : V9::MOVGi; break;
+ case Instruction::SetNE: opCode = V9::MOVNEi; break;
+ default: assert(0 && "Unrecognized LLVM instr!"); break;
+ }
+
+ return opCode;
+}
+
+
+// ChooseMovpregiForSetCC -- Choose a conditional-move-on-register-value
+// instruction based on the type of SetCC operation. These instructions
+// compare a register with 0 and perform the move is the comparison is true.
+//
+// WARNING: like the previous function, this function it always returns
+// the opcode that expects an immediate and a register. See above.
+//
+static MachineOpCode
+ChooseMovpregiForSetCC(const InstructionNode* instrNode)
+{
+ MachineOpCode opCode = V9::INVALID_OPCODE;
+
+ switch(instrNode->getInstruction()->getOpcode())
+ {
+ case Instruction::SetEQ: opCode = V9::MOVRZi; break;
+ case Instruction::SetLE: opCode = V9::MOVRLEZi; break;
+ case Instruction::SetGE: opCode = V9::MOVRGEZi; break;
+ case Instruction::SetLT: opCode = V9::MOVRLZi; break;
+ case Instruction::SetGT: opCode = V9::MOVRGZi; break;
+ case Instruction::SetNE: opCode = V9::MOVRNZi; break;
default: assert(0 && "Unrecognized VM instr!"); break;
}
return opCode;
}
+
static inline MachineOpCode
-ChooseConvertToFloatInstr(OpLabel vopCode, const Type* opType)
+ChooseConvertToFloatInstr(const TargetMachine& target,
+ OpLabel vopCode, const Type* opType)
{
assert((vopCode == ToFloatTy || vopCode == ToDoubleTy) &&
"Unrecognized convert-to-float opcode!");
+ assert((opType->isIntegral() || opType->isFloatingPoint() ||
+ isa<PointerType>(opType))
+ && "Trying to convert a non-scalar type to FLOAT/DOUBLE?");
MachineOpCode opCode = V9::INVALID_OPCODE;
-
- if (opType == Type::SByteTy || opType == Type::UByteTy ||
- opType == Type::ShortTy || opType == Type::UShortTy ||
- opType == Type::IntTy || opType == Type::UIntTy)
- opCode = (vopCode == ToFloatTy? V9::FITOS : V9::FITOD);
- else if (opType == Type::LongTy || opType == Type::ULongTy)
- opCode = (vopCode == ToFloatTy? V9::FXTOS : V9::FXTOD);
- else if (opType == Type::FloatTy)
- opCode = (vopCode == ToFloatTy? V9::INVALID_OPCODE : V9::FSTOD);
+
+ unsigned opSize = target.getTargetData().getTypeSize(opType);
+
+ if (opType == Type::FloatTy)
+ opCode = (vopCode == ToFloatTy? V9::NOP : V9::FSTOD);
else if (opType == Type::DoubleTy)
- opCode = (vopCode == ToFloatTy? V9::FDTOS : V9::INVALID_OPCODE);
- else
- assert(0 && "Cannot convert this type to DOUBLE on SPARC");
+ opCode = (vopCode == ToFloatTy? V9::FDTOS : V9::NOP);
+ else if (opSize <= 4)
+ opCode = (vopCode == ToFloatTy? V9::FITOS : V9::FITOD);
+ else {
+ assert(opSize == 8 && "Unrecognized type size > 4 and < 8!");
+ opCode = (vopCode == ToFloatTy? V9::FXTOS : V9::FXTOD);
+ }
return opCode;
}
static inline MachineOpCode
-ChooseConvertFPToIntInstr(Type::PrimitiveID tid, const Type* opType)
+ChooseConvertFPToIntInstr(const TargetMachine& target,
+ const Type* destType, const Type* opType)
{
- MachineOpCode opCode = V9::INVALID_OPCODE;;
-
assert((opType == Type::FloatTy || opType == Type::DoubleTy)
&& "This function should only be called for FLOAT or DOUBLE");
+ assert((destType->isIntegral() || isa<PointerType>(destType))
+ && "Trying to convert FLOAT/DOUBLE to a non-scalar type?");
- // SPARC does not have a float-to-uint conversion, only a float-to-int.
- // For converting an FP value to uint32_t, we first need to convert to
- // uint64_t and then to uint32_t, or we may overflow the signed int
- // representation even for legal uint32_t values. This expansion is
- // done by the Preselection pass.
- //
- if (tid == Type::UIntTyID) {
- assert(tid != Type::UIntTyID && "FP-to-uint conversions must be expanded"
- " into FP->long->uint for SPARC v9: SO RUN PRESELECTION PASS!");
- } else if (tid == Type::SByteTyID || tid == Type::ShortTyID ||
- tid == Type::IntTyID || tid == Type::UByteTyID ||
- tid == Type::UShortTyID) {
+ MachineOpCode opCode = V9::INVALID_OPCODE;
+
+ unsigned destSize = target.getTargetData().getTypeSize(destType);
+
+ if (destType == Type::UIntTy)
+ assert(destType != Type::UIntTy && "Expand FP-to-uint beforehand.");
+ else if (destSize <= 4)
opCode = (opType == Type::FloatTy)? V9::FSTOI : V9::FDTOI;
- } else if (tid == Type::LongTyID || tid == Type::ULongTyID) {
- opCode = (opType == Type::FloatTy)? V9::FSTOX : V9::FDTOX;
- } else
- assert(0 && "Should not get here, Mo!");
+ else {
+ assert(destSize == 8 && "Unrecognized type size > 4 and < 8!");
+ opCode = (opType == Type::FloatTy)? V9::FSTOX : V9::FDTOX;
+ }
return opCode;
}
-MachineInstr*
-CreateConvertFPToIntInstr(Type::PrimitiveID destTID,
- Value* srcVal, Value* destVal)
+static MachineInstr*
+CreateConvertFPToIntInstr(const TargetMachine& target,
+ Value* srcVal,
+ Value* destVal,
+ const Type* destType)
{
- MachineOpCode opCode = ChooseConvertFPToIntInstr(destTID, srcVal->getType());
+ MachineOpCode opCode = ChooseConvertFPToIntInstr(target, destType,
+ srcVal->getType());
assert(opCode != V9::INVALID_OPCODE && "Expected to need conversion!");
return BuildMI(opCode, 2).addReg(srcVal).addRegDef(destVal);
}
// 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.
-//
+// SPARC does not have a float-to-uint conversion, only a float-to-int (fdtoi).
// 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. (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) V;
+// and MAXUNSIGNED (i.e., 2^31 <= V <= 2^32-1) would be converted incorrectly.
+// Therefore, for converting an FP value to uint32_t, we first need to convert
+// to uint64_t and then to uint32_t.
//
static void
CreateCodeToConvertFloatToInt(const TargetMachine& target,
std::vector<MachineInstr*>& mvec,
MachineCodeForInstruction& mcfi)
{
+ Function* F = destI->getParent()->getParent();
+
// 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.getTargetData().getTypeSize(destI->getType());
- const Type* destTypeToUse = (destSize > 4)? Type::DoubleTy : Type::FloatTy;
- TmpInstruction* destForCast = new TmpInstruction(mcfi, destTypeToUse, opVal);
- // Create the fp-to-int conversion code
- MachineInstr* M =CreateConvertFPToIntInstr(destI->getType()->getPrimitiveID(),
- opVal, destForCast);
- mvec.push_back(M);
+ const Type* castDestType = destI->getType(); // type for the cast instr result
+ const Type* castDestRegType; // type for cast instruction result reg
+ TmpInstruction* destForCast; // dest for cast instruction
+ Instruction* fpToIntCopyDest = destI; // dest for fp-reg-to-int-reg copy instr
+
+ // For converting an FP value to uint32_t, we first need to convert to
+ // uint64_t and then to uint32_t, as explained above.
+ if (destI->getType() == Type::UIntTy) {
+ castDestType = Type::ULongTy; // use this instead of type of destI
+ castDestRegType = Type::DoubleTy; // uint64_t needs 64-bit FP register.
+ destForCast = new TmpInstruction(mcfi, castDestRegType, opVal);
+ fpToIntCopyDest = new TmpInstruction(mcfi, castDestType, destForCast);
+ }
+ else {
+ castDestRegType = (destSize > 4)? Type::DoubleTy : Type::FloatTy;
+ destForCast = new TmpInstruction(mcfi, castDestRegType, opVal);
+ }
+
+ // Create the fp-to-int conversion instruction (src and dest regs are FP regs)
+ mvec.push_back(CreateConvertFPToIntInstr(target, opVal, destForCast,
+ castDestType));
// Create the fpreg-to-intreg copy code
- target.getInstrInfo().
- CreateCodeToCopyFloatToInt(target, destI->getParent()->getParent(),
- destForCast, destI, mvec, mcfi);
+ target.getInstrInfo().CreateCodeToCopyFloatToInt(target, F, destForCast,
+ fpToIntCopyDest, mvec, mcfi);
+
+ // Create the uint64_t to uint32_t conversion, if needed
+ if (destI->getType() == Type::UIntTy)
+ target.getInstrInfo().
+ CreateZeroExtensionInstructions(target, F, fpToIntCopyDest, destI,
+ /*numLowBits*/ 32, mvec, mcfi);
}
Value* shiftDest = destVal;
unsigned opSize = target.getTargetData().getTypeSize(argVal1->getType());
- if ((shiftOpCode == V9::SLLr6 || shiftOpCode == V9::SLLXr6) && opSize < 8) {
+ if ((shiftOpCode == V9::SLLr5 || shiftOpCode == V9::SLLXr6) && opSize < 8) {
// put SLL result into a temporary
shiftDest = new TmpInstruction(mcfi, argVal1, optArgVal2, "sllTmp");
}
if (resultType->isInteger() || isa<PointerType>(resultType)) {
bool isValidConst;
- int64_t C = GetConstantValueAsSignedInt(constOp, isValidConst);
+ int64_t C = (int64_t) target.getInstrInfo().ConvertConstantToIntType(target,
+ constOp, constOp->getType(), isValidConst);
if (isValidConst) {
unsigned pow;
bool needNeg = false;
mvec.push_back(M);
} else if (isPowerOf2(C, pow)) {
unsigned opSize = target.getTargetData().getTypeSize(resultType);
- MachineOpCode opCode = (opSize <= 32)? V9::SLLr6 : V9::SLLXr6;
+ MachineOpCode opCode = (opSize <= 32)? V9::SLLr5 : V9::SLLXr6;
CreateShiftInstructions(target, F, opCode, lval, NULL, pow,
destVal, mvec, mcfi);
}
if (resultType->isInteger()) {
unsigned pow;
bool isValidConst;
- int64_t C = GetConstantValueAsSignedInt(constOp, isValidConst);
+ int64_t C = (int64_t) target.getInstrInfo().ConvertConstantToIntType(target,
+ constOp, constOp->getType(), isValidConst);
if (isValidConst) {
bool needNeg = false;
if (C < 0) {
needNeg = true;
C = -C;
}
-
+
if (C == 1) {
mvec.push_back(BuildMI(V9::ADDr, 3).addReg(LHS).addMReg(ZeroReg)
.addRegDef(destVal));
} else if (isPowerOf2(C, pow)) {
unsigned opCode;
Value* shiftOperand;
+ unsigned opSize = target.getTargetData().getTypeSize(resultType);
if (resultType->isSigned()) {
- // The result may be negative and we need to add one before shifting
- // a negative value. Use:
- // srl i0, 31, x0; add x0, i0, i1 (if i0 is <= 32 bits)
- // or
- // srlx i0, 63, x0; add x0, i0, i1 (if i0 is 64 bits)
- // to compute i1=i0+1 if i0 < 0 and i1=i0 otherwise.
+ // For N / 2^k, if the operand N is negative,
+ // we need to add (2^k - 1) before right-shifting by k, i.e.,
+ //
+ // (N / 2^k) = N >> k, if N >= 0;
+ // (N + 2^k - 1) >> k, if N < 0
+ //
+ // If N is <= 32 bits, use:
+ // sra N, 31, t1 // t1 = ~0, if N < 0, 0 else
+ // srl t1, 32-k, t2 // t2 = 2^k - 1, if N < 0, 0 else
+ // add t2, N, t3 // t3 = N + 2^k -1, if N < 0, N else
+ // sra t3, k, result // result = N / 2^k
//
- TmpInstruction *srlTmp, *addTmp;
+ // If N is 64 bits, use:
+ // srax N, k-1, t1 // t1 = sign bit in high k positions
+ // srlx t1, 64-k, t2 // t2 = 2^k - 1, if N < 0, 0 else
+ // add t2, N, t3 // t3 = N + 2^k -1, if N < 0, N else
+ // sra t3, k, result // result = N / 2^k
+ //
+ TmpInstruction *sraTmp, *srlTmp, *addTmp;
MachineCodeForInstruction& mcfi
= MachineCodeForInstruction::get(destVal);
- srlTmp = new TmpInstruction(mcfi, resultType, LHS, 0, "getSign");
+ sraTmp = new TmpInstruction(mcfi, resultType, LHS, 0, "getSign");
+ srlTmp = new TmpInstruction(mcfi, resultType, LHS, 0, "getPlus2km1");
addTmp = new TmpInstruction(mcfi, resultType, LHS, srlTmp,"incIfNeg");
- // Create the SRL or SRLX instruction to get the sign bit
- mvec.push_back(BuildMI((resultType==Type::LongTy) ?
- V9::SRLXi6 : V9::SRLi6, 3)
+ // Create the SRA or SRAX instruction to get the sign bit
+ mvec.push_back(BuildMI((opSize > 4)? V9::SRAXi6 : V9::SRAi5, 3)
.addReg(LHS)
- .addSImm((resultType==Type::LongTy)? 63 : 31)
+ .addSImm((resultType==Type::LongTy)? pow-1 : 31)
+ .addRegDef(sraTmp));
+
+ // Create the SRL or SRLX instruction to get the sign bit
+ mvec.push_back(BuildMI((opSize > 4)? V9::SRLXi6 : V9::SRLi5, 3)
+ .addReg(sraTmp)
+ .addSImm((resultType==Type::LongTy)? 64-pow : 32-pow)
.addRegDef(srlTmp));
- // Create the ADD instruction to add 1 for negative values
+ // Create the ADD instruction to add 2^pow-1 for negative values
mvec.push_back(BuildMI(V9::ADDr, 3).addReg(LHS).addReg(srlTmp)
.addRegDef(addTmp));
// Get the shift operand and "right-shift" opcode to do the divide
shiftOperand = addTmp;
- opCode = (resultType==Type::LongTy) ? V9::SRAXi6 : V9::SRAi6;
+ opCode = (opSize > 4)? V9::SRAXi6 : V9::SRAi5;
} else {
// Get the shift operand and "right-shift" opcode to do the divide
shiftOperand = LHS;
- opCode = (resultType==Type::LongTy) ? V9::SRLXi6 : V9::SRLi6;
+ opCode = (opSize > 4)? V9::SRLXi6 : V9::SRLi5;
}
// Now do the actual shift!
// compile time if the total size is a known constant.
if (isa<Constant>(numElementsVal)) {
bool isValid;
- int64_t numElem = GetConstantValueAsSignedInt(numElementsVal, isValid);
+ int64_t numElem = (int64_t) target.getInstrInfo().
+ ConvertConstantToIntType(target, numElementsVal,
+ numElementsVal->getType(), isValid);
assert(isValid && "Unexpectedly large array dimension in alloca!");
int64_t total = numElem * tsize;
if (int extra= total % target.getFrameInfo().getStackFrameSizeAlignment())
Value* tsizeVal = ConstantSInt::get(Type::IntTy, tsize);
// Create temporary values to hold the result of MUL, SLL, SRL
- // THIS CASE IS INCOMPLETE AND WILL BE FIXED SHORTLY.
+ // To pad `size' to next smallest multiple of 16:
+ // size = (size + 15) & (-16 = 0xfffffffffffffff0)
+ //
TmpInstruction* tmpProd = new TmpInstruction(mcfi,numElementsVal, tsizeVal);
- TmpInstruction* tmpSLL = new TmpInstruction(mcfi,numElementsVal, tmpProd);
- TmpInstruction* tmpSRL = new TmpInstruction(mcfi,numElementsVal, tmpSLL);
+ TmpInstruction* tmpAdd15= new TmpInstruction(mcfi,numElementsVal, tmpProd);
+ TmpInstruction* tmpAndf0= new TmpInstruction(mcfi,numElementsVal, tmpAdd15);
// Instruction 1: mul numElements, typeSize -> tmpProd
// This will optimize the MUL as far as possible.
- CreateMulInstruction(target, F, numElementsVal, tsizeVal, tmpProd,getMvec,
+ CreateMulInstruction(target, F, numElementsVal, tsizeVal, tmpProd, getMvec,
mcfi, INVALID_MACHINE_OPCODE);
- assert(0 && "Need to insert padding instructions here!");
+ // Instruction 2: andn tmpProd, 0x0f -> tmpAndn
+ getMvec.push_back(BuildMI(V9::ADDi, 3).addReg(tmpProd).addSImm(15)
+ .addReg(tmpAdd15, MOTy::Def));
- totalSizeVal = tmpProd;
+ // Instruction 3: add tmpAndn, 0x10 -> tmpAdd16
+ getMvec.push_back(BuildMI(V9::ANDi, 3).addReg(tmpAdd15).addSImm(-16)
+ .addReg(tmpAndf0, MOTy::Def));
+
+ totalSizeVal = tmpAndf0;
}
// Get the constant offset from SP for dynamically allocated storage
Function *F = result->getParent()->getParent();
MachineFunction &mcInfo = MachineFunction::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 paddedSizeIgnored;
+ // Put the variable in the dynamically sized area of the frame if either:
+ // (a) The offset is too large to use as an immediate in load/stores
+ // (check LDX because all load/stores have the same-size immed. field).
+ // (b) The object is "large", so it could cause many other locals,
+ // spills, and temporaries to have large offsets.
+ // NOTE: We use LARGE = 8 * argSlotSize = 64 bytes.
+ // You've gotta love having only 13 bits for constant offset values :-|.
+ //
+ unsigned paddedSize;
int offsetFromFP = mcInfo.getInfo()->computeOffsetforLocalVar(result,
- paddedSizeIgnored,
- tsize * numElements);
- if (! target.getInstrInfo().constantFitsInImmedField(V9::LDXi,offsetFromFP)) {
+ paddedSize,
+ tsize * numElements);
+
+ if (((int)paddedSize) > 8 * target.getFrameInfo().getSizeOfEachArgOnStack() ||
+ ! target.getInstrInfo().constantFitsInImmedField(V9::LDXi,offsetFromFP)) {
CreateCodeForVariableSizeAlloca(target, result, tsize,
ConstantSInt::get(Type::IntTy,numElements),
getMvec);
}
for (unsigned i=0,numOps=minstr->getNumImplicitRefs(); i<numOps; ++i)
- if (minstr->getImplicitRef(i) == unusedOp) {
- minstr->setImplicitRef(i, fwdOp,
- minstr->getImplicitOp(i).opIsDefOnly(),
- minstr->getImplicitOp(i).opIsDefAndUse());
- }
+ if (minstr->getImplicitRef(i) == unusedOp)
+ minstr->setImplicitRef(i, fwdOp);
}
}
}
{
switch (iid) {
case LLVMIntrinsic::va_start: {
- // Get the address of the first vararg value on stack and copy it to
- // the argument of va_start(va_list* ap).
+ // Get the address of the first incoming vararg argument on the stack
bool ignore;
Function* func = cast<Function>(callInstr.getParent()->getParent());
int numFixedArgs = func->getFunctionType()->getNumParams();
int firstVarArgOff = numFixedArgs * argSize + target.getFrameInfo().
getFirstIncomingArgOffset(MachineFunction::get(func), ignore);
mvec.push_back(BuildMI(V9::ADDi, 3).addMReg(fpReg).addSImm(firstVarArgOff).
- addReg(callInstr.getOperand(1)));
+ addRegDef(&callInstr));
return true;
}
return true; // no-op on Sparc
case LLVMIntrinsic::va_copy:
- // Simple copy of current va_list (arg2) to new va_list (arg1)
+ // Simple copy of current va_list (arg1) to new va_list (result)
mvec.push_back(BuildMI(V9::ORr, 3).
addMReg(target.getRegInfo().getZeroRegNum()).
- addReg(callInstr.getOperand(2)).
- addReg(callInstr.getOperand(1)));
+ addReg(callInstr.getOperand(1)).
+ addRegDef(&callInstr));
+ return true;
+
+ case LLVMIntrinsic::sigsetjmp:
+ case LLVMIntrinsic::setjmp: {
+ // act as if we return 0
+ unsigned g0 = target.getRegInfo().getZeroRegNum();
+ mvec.push_back(BuildMI(V9::ORr,3).addMReg(g0).addMReg(g0)
+ .addReg(&callInstr, MOTy::Def));
+ return true;
+ }
+
+ case LLVMIntrinsic::siglongjmp:
+ case LLVMIntrinsic::longjmp: {
+ // call abort()
+ Module* M = callInstr.getParent()->getParent()->getParent();
+ const FunctionType *voidvoidFuncTy =
+ FunctionType::get(Type::VoidTy, std::vector<const Type*>(), false);
+ Function *F = M->getOrInsertFunction("abort", voidvoidFuncTy);
+ assert(F && "Unable to get or create `abort' function declaration");
+
+ // Create hidden virtual register for return address with type void*
+ TmpInstruction* retAddrReg =
+ new TmpInstruction(MachineCodeForInstruction::get(&callInstr),
+ PointerType::get(Type::VoidTy), &callInstr);
+
+ // Use a descriptor to pass information about call arguments
+ // to the register allocator. This descriptor will be "owned"
+ // and freed automatically when the MachineCodeForInstruction
+ // object for the callInstr goes away.
+ CallArgsDescriptor* argDesc =
+ new CallArgsDescriptor(&callInstr, retAddrReg, false, false);
+
+ MachineInstr* callMI = BuildMI(V9::CALL, 1).addPCDisp(F);
+ callMI->addImplicitRef(retAddrReg, /*isDef*/ true);
+
+ mvec.push_back(callMI);
+ mvec.push_back(BuildMI(V9::NOP, 0));
return true;
+ }
default:
return false;
unsigned allocaSize = 0;
MachineInstr* M, *M2;
unsigned L;
+ bool foldCase = false;
mvec.clear();
// Let's check for chain rules outside the switch so that we don't have
// to duplicate the list of chain rule production numbers here again
//
- if (ThisIsAChainRule(ruleForNode))
- {
- // Chain rules have a single nonterminal on the RHS.
- // Get the rule that matches the RHS non-terminal and use that instead.
- //
- assert(nts[0] && ! nts[1]
- && "A chain rule should have only one RHS non-terminal!");
- nextRule = burm_rule(subtreeRoot->state, nts[0]);
- nts = burm_nts[nextRule];
- GetInstructionsByRule(subtreeRoot, nextRule, nts, target, mvec);
- }
- else
- {
- switch(ruleForNode) {
- case 1: // stmt: Ret
- case 2: // stmt: RetValue(reg)
+ if (ThisIsAChainRule(ruleForNode)) {
+ // Chain rules have a single nonterminal on the RHS.
+ // Get the rule that matches the RHS non-terminal and use that instead.
+ //
+ assert(nts[0] && ! nts[1]
+ && "A chain rule should have only one RHS non-terminal!");
+ nextRule = burm_rule(subtreeRoot->state, nts[0]);
+ nts = burm_nts[nextRule];
+ GetInstructionsByRule(subtreeRoot, nextRule, nts, target, mvec);
+ } else {
+ switch(ruleForNode) {
+ case 1: // stmt: Ret
+ case 2: // stmt: RetValue(reg)
{ // NOTE: Prepass of register allocation is responsible
// for moving return value to appropriate register.
// Copy the return value to the required return register.
BuildMI(V9::JMPLRETi, 3).addReg(returnAddrTmp).addSImm(8)
.addMReg(target.getRegInfo().getZeroRegNum(), MOTy::Def);
- // Insert a copy to copy the return value to the appropriate register
- // -- For FP values, create a FMOVS or FMOVD instruction
- // -- For non-FP values, create an add-with-0 instruction
+ // If there is a value to return, we need to:
+ // (a) Sign-extend the value if it is smaller than 8 bytes (reg size)
+ // (b) Insert a copy to copy the return value to the appropriate reg.
+ // -- For FP values, create a FMOVS or FMOVD instruction
+ // -- For non-FP values, create an add-with-0 instruction
//
if (retVal != NULL) {
const UltraSparcRegInfo& regInfo =
: (unsigned) SparcIntRegClass::i0);
retRegNum = regInfo.getUnifiedRegNum(regClassID, retRegNum);
- // Create a virtual register to represent it and mark
- // this vreg as being an implicit operand of the ret MI
+ // () Insert sign-extension instructions for small signed values.
+ //
+ Value* retValToUse = retVal;
+ if (retType->isIntegral() && retType->isSigned()) {
+ unsigned retSize = target.getTargetData().getTypeSize(retType);
+ if (retSize <= 4) {
+ // create a temporary virtual reg. to hold the sign-extension
+ retValToUse = new TmpInstruction(mcfi, retVal);
+
+ // sign-extend retVal and put the result in the temporary reg.
+ target.getInstrInfo().CreateSignExtensionInstructions
+ (target, returnInstr->getParent()->getParent(),
+ retVal, retValToUse, 8*retSize, mvec, mcfi);
+ }
+ }
+
+ // (b) Now, insert a copy to to the appropriate register:
+ // -- For FP values, create a FMOVS or FMOVD instruction
+ // -- For non-FP values, create an add-with-0 instruction
+ //
+ // First, create a virtual register to represent the register and
+ // mark this vreg as being an implicit operand of the ret MI.
TmpInstruction* retVReg =
- new TmpInstruction(mcfi, retVal, NULL, "argReg");
-
+ new TmpInstruction(mcfi, retValToUse, NULL, "argReg");
+
retMI->addImplicitRef(retVReg);
-
+
if (retType->isFloatingPoint())
M = (BuildMI(retType==Type::FloatTy? V9::FMOVS : V9::FMOVD, 2)
- .addReg(retVal).addReg(retVReg, MOTy::Def));
+ .addReg(retValToUse).addReg(retVReg, MOTy::Def));
else
M = (BuildMI(ChooseAddInstructionByType(retType), 3)
- .addReg(retVal).addSImm((int64_t) 0)
+ .addReg(retValToUse).addSImm((int64_t) 0)
.addReg(retVReg, MOTy::Def));
// Mark the operand with the register it should be assigned
if ((constVal->getType()->isInteger()
|| isa<PointerType>(constVal->getType()))
- && GetConstantValueAsSignedInt(constVal, isValidConst) == 0
+ && target.getInstrInfo().ConvertConstantToIntType(target,
+ constVal, constVal->getType(), isValidConst) == 0
&& isValidConst)
{
// That constant is a zero after all...
// TmpInstruction representing that CC.
//
BranchInst* brInst = cast<BranchInst>(subtreeRoot->getInstruction());
- bool isFPBranch;
- unsigned Opcode = ChooseBccInstruction(subtreeRoot, isFPBranch);
+ const Type* setCCType;
+ unsigned Opcode = ChooseBccInstruction(subtreeRoot, setCCType);
Value* ccValue = GetTmpForCC(subtreeRoot->leftChild()->getValue(),
brInst->getParent()->getParent(),
- isFPBranch? Type::FloatTy : Type::IntTy,
+ setCCType,
MachineCodeForInstruction::get(brInst));
M = BuildMI(Opcode, 2).addCCReg(ccValue)
.addPCDisp(brInst->getSuccessor(0));
assert(0 && "VRegList should never be the topmost non-chain rule");
break;
- case 21: // bool: Not(bool,reg): Both these are implemented as:
- case 421: // reg: BNot(reg,reg): reg = reg XOR-NOT 0
+ case 21: // bool: Not(bool,reg): Compute with a conditional-move-on-reg
+ { // First find the unary operand. It may be left or right, usually right.
+ Instruction* notI = subtreeRoot->getInstruction();
+ Value* notArg = BinaryOperator::getNotArgument(
+ cast<BinaryOperator>(subtreeRoot->getInstruction()));
+ unsigned ZeroReg = target.getRegInfo().getZeroRegNum();
+
+ // Unconditionally set register to 0
+ mvec.push_back(BuildMI(V9::SETHI, 2).addZImm(0).addRegDef(notI));
+
+ // Now conditionally move 1 into the register.
+ // Mark the register as a use (as well as a def) because the old
+ // value will be retained if the condition is false.
+ mvec.push_back(BuildMI(V9::MOVRZi, 3).addReg(notArg).addZImm(1)
+ .addReg(notI, MOTy::UseAndDef));
+
+ break;
+ }
+
+ case 421: // reg: BNot(reg,reg): Compute as reg = reg XOR-NOT 0
{ // First find the unary operand. It may be left or right, usually right.
Value* notArg = BinaryOperator::getNotArgument(
cast<BinaryOperator>(subtreeRoot->getInstruction()));
break;
}
+ case 322: // reg: Not(tobool, reg):
+ // Fold CAST-TO-BOOL with NOT by inverting the sense of cast-to-bool
+ foldCase = true;
+ // Just fall through!
+
case 22: // reg: ToBoolTy(reg):
{
- const Type* opType = subtreeRoot->leftChild()->getValue()->getType();
- assert(opType->isIntegral() || isa<PointerType>(opType));
- forwardOperandNum = 0; // forward first operand to user
+ Instruction* castI = subtreeRoot->getInstruction();
+ Value* opVal = subtreeRoot->leftChild()->getValue();
+ assert(opVal->getType()->isIntegral() ||
+ isa<PointerType>(opVal->getType()));
+
+ // Unconditionally set register to 0
+ mvec.push_back(BuildMI(V9::SETHI, 2).addZImm(0).addRegDef(castI));
+
+ // Now conditionally move 1 into the register.
+ // Mark the register as a use (as well as a def) because the old
+ // value will be retained if the condition is false.
+ MachineOpCode opCode = foldCase? V9::MOVRZi : V9::MOVRNZi;
+ mvec.push_back(BuildMI(opCode, 3).addReg(opVal).addZImm(1)
+ .addReg(castI, MOTy::UseAndDef));
+
break;
}
case 26: // reg: ToShortTy(reg)
case 27: // reg: ToUIntTy(reg)
case 28: // reg: ToIntTy(reg)
+ case 29: // reg: ToULongTy(reg)
+ case 30: // reg: ToLongTy(reg)
{
//======================================================================
// Rules for integer conversions:
//
// 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 smaller than destination, zero-extend or sign-extend
+ // according to the signedness of the *operand*: source decides:
+ // (1) If operand is signed, sign-extend it.
+ // If dest is unsigned, zero-ext the result!
+ // (2) If operand is unsigned, our current invariant is that
+ // it's high bits are correct, so zero-extension is not needed.
//
- // -- 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.
+ // -- If operand is same size as or larger than destination,
+ // zero-extend or sign-extend according to the signedness of
+ // the *destination*: destination decides:
+ // (1) If destination is signed, sign-extend (truncating if needed)
+ // This choice is implementation defined. We sign-extend the
+ // operand, which matches both Sun's cc and gcc3.2.
+ // (2) If destination is unsigned, zero-extend (truncating if needed)
//======================================================================
Instruction* destI = subtreeRoot->getInstruction();
+ Function* currentFunc = destI->getParent()->getParent();
+ MachineCodeForInstruction& mcfi=MachineCodeForInstruction::get(destI);
+
Value* opVal = subtreeRoot->leftChild()->getValue();
const Type* opType = opVal->getType();
- if (opType->isIntegral() || isa<PointerType>(opType)) {
- unsigned opSize = target.getTargetData().getTypeSize(opType);
- unsigned destSize =
- target.getTargetData().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
+ const Type* destType = destI->getType();
+ unsigned opSize = target.getTargetData().getTypeSize(opType);
+ unsigned destSize = target.getTargetData().getTypeSize(destType);
+
+ bool isIntegral = opType->isIntegral() || isa<PointerType>(opType);
+
+ if (opType == Type::BoolTy ||
+ opType == destType ||
+ isIntegral && opSize == destSize && opSize == 8) {
+ // nothing to do in all these cases
+ forwardOperandNum = 0; // forward first operand to user
+
} else if (opType->isFloatingPoint()) {
- CreateCodeToConvertFloatToInt(target, opVal, destI, mvec,
- MachineCodeForInstruction::get(destI));
- if (destI->getType()->isUnsigned())
+
+ CreateCodeToConvertFloatToInt(target, opVal, destI, mvec, mcfi);
+ if (destI->getType()->isUnsigned() && destI->getType() !=Type::UIntTy)
maskUnsignedResult = true; // not handled by fp->int code
- } else
- assert(0 && "Unrecognized operand type for convert-to-unsigned");
- break;
- }
+ } else if (isIntegral) {
+
+ bool opSigned = opType->isSigned();
+ bool destSigned = destType->isSigned();
+ unsigned extSourceInBits = 8 * std::min<unsigned>(opSize, destSize);
+
+ assert(! (opSize == destSize && opSigned == destSigned) &&
+ "How can different int types have same size and signedness?");
+
+ bool signExtend = (opSize < destSize && opSigned ||
+ opSize >= destSize && destSigned);
+
+ bool signAndZeroExtend = (opSize < destSize && destSize < 8u &&
+ opSigned && !destSigned);
+ assert(!signAndZeroExtend || signExtend);
+
+ bool zeroExtendOnly = opSize >= destSize && !destSigned;
+ assert(!zeroExtendOnly || !signExtend);
+
+ if (signExtend) {
+ Value* signExtDest = (signAndZeroExtend
+ ? new TmpInstruction(mcfi, destType, opVal)
+ : destI);
+
+ target.getInstrInfo().CreateSignExtensionInstructions
+ (target, currentFunc,opVal,signExtDest,extSourceInBits,mvec,mcfi);
+
+ if (signAndZeroExtend)
+ target.getInstrInfo().CreateZeroExtensionInstructions
+ (target, currentFunc, signExtDest, destI, 8*destSize, mvec, mcfi);
+ }
+ else if (zeroExtendOnly) {
+ target.getInstrInfo().CreateZeroExtensionInstructions
+ (target, currentFunc, opVal, destI, extSourceInBits, mvec, mcfi);
+ }
+ else
+ forwardOperandNum = 0; // forward first operand to user
- case 29: // reg: ToULongTy(reg)
- case 30: // reg: ToLongTy(reg)
- {
- Value* opVal = subtreeRoot->leftChild()->getValue();
- const Type* opType = opVal->getType();
- if (opType->isIntegral() || isa<PointerType>(opType))
- forwardOperandNum = 0; // forward first operand to user
- else if (opType->isFloatingPoint()) {
- Instruction* destI = subtreeRoot->getInstruction();
- CreateCodeToConvertFloatToInt(target, opVal, destI, mvec,
- MachineCodeForInstruction::get(destI));
} else
- assert(0 && "Unrecognized operand type for convert-to-signed");
+ assert(0 && "Unrecognized operand type for convert-to-integer");
+
break;
}
if (forwardOperandNum != 0) { // we do need the cast
Value* leftVal = subtreeRoot->leftChild()->getValue();
const Type* opType = leftVal->getType();
- MachineOpCode opCode=ChooseConvertToFloatInstr(
+ MachineOpCode opCode=ChooseConvertToFloatInstr(target,
subtreeRoot->getOpLabel(), opType);
- if (opCode == V9::INVALID_OPCODE) { // no conversion needed
+ if (opCode == V9::NOP) { // no conversion needed
forwardOperandNum = 0; // forward first operand to user
} else {
// If the source operand is a non-FP type it must be
// ELSE FALL THROUGH
case 36: // reg: Div(reg, reg)
+ {
maskUnsignedResult = true;
- Add3OperandInstr(ChooseDivInstruction(target, subtreeRoot),
- subtreeRoot, mvec);
+
+ // If either operand of divide is smaller than 64 bits, we have
+ // to make sure the unused top bits are correct because they affect
+ // the result. These bits are already correct for unsigned values.
+ // They may be incorrect for signed values, so sign extend to fill in.
+ Instruction* divI = subtreeRoot->getInstruction();
+ Value* divOp1 = subtreeRoot->leftChild()->getValue();
+ Value* divOp2 = subtreeRoot->rightChild()->getValue();
+ Value* divOp1ToUse = divOp1;
+ Value* divOp2ToUse = divOp2;
+ if (divI->getType()->isSigned()) {
+ unsigned opSize=target.getTargetData().getTypeSize(divI->getType());
+ if (opSize < 8) {
+ MachineCodeForInstruction& mcfi=MachineCodeForInstruction::get(divI);
+ divOp1ToUse = new TmpInstruction(mcfi, divOp1);
+ divOp2ToUse = new TmpInstruction(mcfi, divOp2);
+ target.getInstrInfo().
+ CreateSignExtensionInstructions(target,
+ divI->getParent()->getParent(),
+ divOp1, divOp1ToUse,
+ 8*opSize, mvec, mcfi);
+ target.getInstrInfo().
+ CreateSignExtensionInstructions(target,
+ divI->getParent()->getParent(),
+ divOp2, divOp2ToUse,
+ 8*opSize, mvec, mcfi);
+ }
+ }
+
+ mvec.push_back(BuildMI(ChooseDivInstruction(target, subtreeRoot), 3)
+ .addReg(divOp1ToUse)
+ .addReg(divOp2ToUse)
+ .addRegDef(divI));
+
break;
+ }
case 37: // reg: Rem(reg, reg)
case 237: // reg: Rem(reg, Constant)
{
maskUnsignedResult = true;
- Instruction* remInstr = subtreeRoot->getInstruction();
-
- MachineCodeForInstruction& mcfi=MachineCodeForInstruction::get(remInstr);
- TmpInstruction* quot = new TmpInstruction(mcfi,
- subtreeRoot->leftChild()->getValue(),
- subtreeRoot->rightChild()->getValue());
- TmpInstruction* prod = new TmpInstruction(mcfi,
- quot,
- subtreeRoot->rightChild()->getValue());
+
+ Instruction* remI = subtreeRoot->getInstruction();
+ Value* divOp1 = subtreeRoot->leftChild()->getValue();
+ Value* divOp2 = subtreeRoot->rightChild()->getValue();
+
+ MachineCodeForInstruction& mcfi = MachineCodeForInstruction::get(remI);
- M = BuildMI(ChooseDivInstruction(target, subtreeRoot), 3)
- .addReg(subtreeRoot->leftChild()->getValue())
- .addReg(subtreeRoot->rightChild()->getValue())
- .addRegDef(quot);
- mvec.push_back(M);
+ // If second operand of divide is smaller than 64 bits, we have
+ // to make sure the unused top bits are correct because they affect
+ // the result. These bits are already correct for unsigned values.
+ // They may be incorrect for signed values, so sign extend to fill in.
+ //
+ Value* divOpToUse = divOp2;
+ if (divOp2->getType()->isSigned()) {
+ unsigned opSize=target.getTargetData().getTypeSize(divOp2->getType());
+ if (opSize < 8) {
+ divOpToUse = new TmpInstruction(mcfi, divOp2);
+ target.getInstrInfo().
+ CreateSignExtensionInstructions(target,
+ remI->getParent()->getParent(),
+ divOp2, divOpToUse,
+ 8*opSize, mvec, mcfi);
+ }
+ }
+
+ // Now compute: result = rem V1, V2 as:
+ // result = V1 - (V1 / signExtend(V2)) * signExtend(V2)
+ //
+ TmpInstruction* quot = new TmpInstruction(mcfi, divOp1, divOpToUse);
+ TmpInstruction* prod = new TmpInstruction(mcfi, quot, divOpToUse);
+
+ mvec.push_back(BuildMI(ChooseDivInstruction(target, subtreeRoot), 3)
+ .addReg(divOp1).addReg(divOpToUse).addRegDef(quot));
- unsigned MulOpcode =
- ChooseMulInstructionByType(subtreeRoot->getInstruction()->getType());
- Value *MulRHS = subtreeRoot->rightChild()->getValue();
- M = BuildMI(MulOpcode, 3).addReg(quot).addReg(MulRHS).addReg(prod,
- MOTy::Def);
- mvec.push_back(M);
+ mvec.push_back(BuildMI(ChooseMulInstructionByType(remI->getType()), 3)
+ .addReg(quot).addReg(divOpToUse).addRegDef(prod));
+
+ mvec.push_back(BuildMI(ChooseSubInstructionByType(remI->getType()), 3)
+ .addReg(divOp1).addReg(prod).addRegDef(remI));
- unsigned Opcode = ChooseSubInstructionByType(
- subtreeRoot->getInstruction()->getType());
- M = BuildMI(Opcode, 3).addReg(subtreeRoot->leftChild()->getValue())
- .addReg(prod).addRegDef(subtreeRoot->getValue());
- mvec.push_back(M);
break;
}
case 38: // bool: And(bool, bool)
+ case 138: // bool: And(bool, not)
case 238: // bool: And(bool, boolconst)
case 338: // reg : BAnd(reg, reg)
case 538: // reg : BAnd(reg, Constant)
Add3OperandInstr(V9::ANDr, subtreeRoot, mvec);
break;
- case 138: // bool: And(bool, not)
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.
+ // If the type is boolean, set 1 or 0 in the result register.
InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
Value* notArg = BinaryOperator::getNotArgument(
cast<BinaryOperator>(notNode->getInstruction()));
notNode->markFoldedIntoParent();
- Value *LHS = subtreeRoot->leftChild()->getValue();
- Value *Dest = subtreeRoot->getValue();
- mvec.push_back(BuildMI(V9::ANDNr, 3).addReg(LHS).addReg(notArg)
- .addReg(Dest, MOTy::Def));
+ Value *lhs = subtreeRoot->leftChild()->getValue();
+ Value *dest = subtreeRoot->getValue();
+ mvec.push_back(BuildMI(V9::ANDNr, 3).addReg(lhs).addReg(notArg)
+ .addReg(dest, MOTy::Def));
+
+ if (notArg->getType() == Type::BoolTy) {
+ // set 1 in result register if result of above is non-zero
+ mvec.push_back(BuildMI(V9::MOVRNZi, 3).addReg(dest).addZImm(1)
+ .addReg(dest, MOTy::UseAndDef));
+ }
+
break;
}
case 39: // bool: Or(bool, bool)
+ case 139: // bool: Or(bool, not)
case 239: // bool: Or(bool, boolconst)
case 339: // reg : BOr(reg, reg)
case 539: // reg : BOr(reg, Constant)
Add3OperandInstr(V9::ORr, subtreeRoot, mvec);
break;
- case 139: // bool: Or(bool, not)
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.
+ // If the type is boolean, set 1 or 0 in the result register.
InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
Value* notArg = BinaryOperator::getNotArgument(
cast<BinaryOperator>(notNode->getInstruction()));
notNode->markFoldedIntoParent();
- Value *LHS = subtreeRoot->leftChild()->getValue();
- Value *Dest = subtreeRoot->getValue();
- mvec.push_back(BuildMI(V9::ORNr, 3).addReg(LHS).addReg(notArg)
- .addReg(Dest, MOTy::Def));
+ Value *lhs = subtreeRoot->leftChild()->getValue();
+ Value *dest = subtreeRoot->getValue();
+
+ mvec.push_back(BuildMI(V9::ORNr, 3).addReg(lhs).addReg(notArg)
+ .addReg(dest, MOTy::Def));
+
+ if (notArg->getType() == Type::BoolTy) {
+ // set 1 in result register if result of above is non-zero
+ mvec.push_back(BuildMI(V9::MOVRNZi, 3).addReg(dest).addZImm(1)
+ .addReg(dest, MOTy::UseAndDef));
+ }
+
break;
}
case 40: // bool: Xor(bool, bool)
+ case 140: // bool: Xor(bool, not)
case 240: // bool: Xor(bool, boolconst)
case 340: // reg : BXor(reg, reg)
case 540: // reg : BXor(reg, Constant)
Add3OperandInstr(V9::XORr, subtreeRoot, mvec);
break;
- case 140: // bool: Xor(bool, not)
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.
+ // If the type is boolean, set 1 or 0 in the result register.
InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
Value* notArg = BinaryOperator::getNotArgument(
cast<BinaryOperator>(notNode->getInstruction()));
notNode->markFoldedIntoParent();
- Value *LHS = subtreeRoot->leftChild()->getValue();
- Value *Dest = subtreeRoot->getValue();
- mvec.push_back(BuildMI(V9::XNORr, 3).addReg(LHS).addReg(notArg)
- .addReg(Dest, MOTy::Def));
+ Value *lhs = subtreeRoot->leftChild()->getValue();
+ Value *dest = subtreeRoot->getValue();
+ mvec.push_back(BuildMI(V9::XNORr, 3).addReg(lhs).addReg(notArg)
+ .addReg(dest, MOTy::Def));
+
+ if (notArg->getType() == Type::BoolTy) {
+ // set 1 in result register if result of above is non-zero
+ mvec.push_back(BuildMI(V9::MOVRNZi, 3).addReg(dest).addZImm(1)
+ .addReg(dest, MOTy::UseAndDef));
+ }
break;
}
- case 41: // boolconst: SetCC(reg, Constant)
+ case 41: // setCCconst: SetCC(reg, Constant)
+ { // Comparison is with a constant:
//
- // 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 the bool result must be computed into a register (see below),
+ // and the constant is int ZERO, we can use the MOVR[op] instructions
+ // and avoid the SUBcc instruction entirely.
+ // Otherwise this is just the same as case 42, so just fall through.
//
+ // The result of the SetCC must be computed and stored in a register if
+ // it is used outside the current basic block (so it must be computed
+ // as a boolreg) or it is used by anything other than a branch.
+ // We will use a conditional move to do this.
+ //
+ Instruction* setCCInstr = subtreeRoot->getInstruction();
+ bool computeBoolVal = (subtreeRoot->parent() == NULL ||
+ ! AllUsesAreBranches(setCCInstr));
+
+ if (computeBoolVal) {
+ InstrTreeNode* constNode = subtreeRoot->rightChild();
+ assert(constNode &&
+ constNode->getNodeType() ==InstrTreeNode::NTConstNode);
+ Constant *constVal = cast<Constant>(constNode->getValue());
+ bool isValidConst;
+
+ if ((constVal->getType()->isInteger()
+ || isa<PointerType>(constVal->getType()))
+ && target.getInstrInfo().ConvertConstantToIntType(target,
+ constVal, constVal->getType(), isValidConst) == 0
+ && isValidConst)
+ {
+ // That constant is an integer zero after all...
+ // Use a MOVR[op] to compute the boolean result
+ // Unconditionally set register to 0
+ mvec.push_back(BuildMI(V9::SETHI, 2).addZImm(0)
+ .addRegDef(setCCInstr));
+
+ // Now conditionally move 1 into the register.
+ // Mark the register as a use (as well as a def) because the old
+ // value will be retained if the condition is false.
+ MachineOpCode movOpCode = ChooseMovpregiForSetCC(subtreeRoot);
+ mvec.push_back(BuildMI(movOpCode, 3)
+ .addReg(subtreeRoot->leftChild()->getValue())
+ .addZImm(1).addReg(setCCInstr, MOTy::UseAndDef));
+
+ break;
+ }
+ }
+ // ELSE FALL THROUGH
+ }
+
case 42: // bool: SetCC(reg, reg):
{
// This generates a SUBCC instruction, putting the difference in a
// result reg. if needed, and/or setting a condition code if needed.
//
Instruction* setCCInstr = subtreeRoot->getInstruction();
- Value* leftVal = subtreeRoot->leftChild()->getValue();
- bool isFPCompare = leftVal->getType()->isFloatingPoint();
+ Value* leftVal = subtreeRoot->leftChild()->getValue();
+ Value* rightVal = subtreeRoot->rightChild()->getValue();
+ const Type* opType = leftVal->getType();
+ bool isFPCompare = opType->isFloatingPoint();
// If the boolean result of the SetCC is used outside the current basic
// block (so it must be computed as a boolreg) or is used by anything
//
TmpInstruction* tmpForCC = GetTmpForCC(setCCInstr,
setCCInstr->getParent()->getParent(),
- isFPCompare ? Type::FloatTy : Type::IntTy,
+ leftVal->getType(),
MachineCodeForInstruction::get(setCCInstr));
+
+ // If the operands are signed values smaller than 4 bytes, then they
+ // must be sign-extended in order to do a valid 32-bit comparison
+ // and get the right result in the 32-bit CC register (%icc).
+ //
+ Value* leftOpToUse = leftVal;
+ Value* rightOpToUse = rightVal;
+ if (opType->isIntegral() && opType->isSigned()) {
+ unsigned opSize = target.getTargetData().getTypeSize(opType);
+ if (opSize < 4) {
+ MachineCodeForInstruction& mcfi =
+ MachineCodeForInstruction::get(setCCInstr);
+
+ // create temporary virtual regs. to hold the sign-extensions
+ leftOpToUse = new TmpInstruction(mcfi, leftVal);
+ rightOpToUse = new TmpInstruction(mcfi, rightVal);
+
+ // sign-extend each operand and put the result in the temporary reg.
+ target.getInstrInfo().CreateSignExtensionInstructions
+ (target, setCCInstr->getParent()->getParent(),
+ leftVal, leftOpToUse, 8*opSize, mvec, mcfi);
+ target.getInstrInfo().CreateSignExtensionInstructions
+ (target, setCCInstr->getParent()->getParent(),
+ rightVal, rightOpToUse, 8*opSize, mvec, mcfi);
+ }
+ }
+
if (! isFPCompare) {
// Integer condition: set CC and discard result.
- M = BuildMI(V9::SUBccr, 4)
- .addReg(subtreeRoot->leftChild()->getValue())
- .addReg(subtreeRoot->rightChild()->getValue())
- .addMReg(target.getRegInfo().getZeroRegNum(), MOTy::Def)
- .addCCReg(tmpForCC, MOTy::Def);
+ mvec.push_back(BuildMI(V9::SUBccr, 4)
+ .addReg(leftOpToUse)
+ .addReg(rightOpToUse)
+ .addMReg(target.getRegInfo().getZeroRegNum(),MOTy::Def)
+ .addCCReg(tmpForCC, MOTy::Def));
} else {
// FP condition: dest of FCMP should be some FCCn register
- M = BuildMI(ChooseFcmpInstruction(subtreeRoot), 3)
- .addCCReg(tmpForCC, MOTy::Def)
- .addReg(subtreeRoot->leftChild()->getValue())
- .addReg(subtreeRoot->rightChild()->getValue());
+ mvec.push_back(BuildMI(ChooseFcmpInstruction(subtreeRoot), 3)
+ .addCCReg(tmpForCC, MOTy::Def)
+ .addReg(leftOpToUse)
+ .addReg(rightOpToUse));
}
- mvec.push_back(M);
if (computeBoolVal) {
MachineOpCode movOpCode = (isFPCompare
? ChooseMovFpcciInstruction(subtreeRoot)
- : ChooseMovpcciAfterSub(subtreeRoot));
+ : ChooseMovpcciForSetCC(subtreeRoot));
// Unconditionally set register to 0
M = BuildMI(V9::SETHI, 2).addZImm(0).addRegDef(setCCInstr);
// This can also handle any intrinsics that are just function calls.
//
if (! specialIntrinsic) {
- MachineFunction& MF =
- MachineFunction::get(callInstr->getParent()->getParent());
+ Function* currentFunc = callInstr->getParent()->getParent();
+ MachineFunction& MF = MachineFunction::get(currentFunc);
MachineCodeForInstruction& mcfi =
MachineCodeForInstruction::get(callInstr);
const UltraSparcRegInfo& regInfo =
new CallArgsDescriptor(callInstr, retAddrReg,isVarArgs,noPrototype);
assert(callInstr->getOperand(0) == callee
&& "This is assumed in the loop below!");
-
+
+ // Insert sign-extension instructions for small signed values,
+ // if this is an unknown function (i.e., called via a funcptr)
+ // or an external one (i.e., which may not be compiled by llc).
+ //
+ if (calledFunc == NULL || calledFunc->isExternal()) {
+ for (unsigned i=1, N=callInstr->getNumOperands(); i < N; ++i) {
+ Value* argVal = callInstr->getOperand(i);
+ const Type* argType = argVal->getType();
+ if (argType->isIntegral() && argType->isSigned()) {
+ unsigned argSize = target.getTargetData().getTypeSize(argType);
+ if (argSize <= 4) {
+ // create a temporary virtual reg. to hold the sign-extension
+ TmpInstruction* argExtend = new TmpInstruction(mcfi, argVal);
+
+ // sign-extend argVal and put the result in the temporary reg.
+ target.getInstrInfo().CreateSignExtensionInstructions
+ (target, currentFunc, argVal, argExtend,
+ 8*argSize, mvec, mcfi);
+
+ // replace argVal with argExtend in CallArgsDescriptor
+ argDesc->getArgInfo(i-1).replaceArgVal(argExtend);
+ }
+ }
+ }
+ }
+
// Insert copy instructions to get all the arguments into
// all the places that they need to be.
//
for (unsigned i=1, N=callInstr->getNumOperands(); i < N; ++i) {
int argNo = i-1;
- Value* argVal = callInstr->getOperand(i);
+ CallArgInfo& argInfo = argDesc->getArgInfo(argNo);
+ Value* argVal = argInfo.getArgVal(); // don't use callInstr arg here
const Type* argType = argVal->getType();
- unsigned regType = regInfo.getRegType(argType);
+ unsigned regType = regInfo.getRegTypeForDataType(argType);
unsigned argSize = target.getTargetData().getTypeSize(argType);
int regNumForArg = TargetRegInfo::getInvalidRegNum();
unsigned regClassIDOfArgReg;
- CallArgInfo& argInfo = argDesc->getArgInfo(argNo);
// Check for FP arguments to varargs functions.
// Any such argument in the first $K$ args must be passed in an
// K = #integer argument registers.
bool isFPArg = argVal->getType()->isFloatingPoint();
if (isVarArgs && isFPArg) {
- // If it is a function with no prototype, pass value
- // as an FP value as well as a varargs value
- if (noPrototype)
- argInfo.setUseFPArgReg();
-
- // If this arg. is in the first $K$ regs, add copy-
+
+ if (noPrototype) {
+ // It is a function with no prototype: pass value
+ // as an FP value as well as a varargs value. The FP value
+ // may go in a register or on the stack. The copy instruction
+ // to the outgoing reg/stack is created by the normal argument
+ // handling code since this is the "normal" passing mode.
+ //
+ regNumForArg = regInfo.regNumForFPArg(regType,
+ false, false, argNo,
+ regClassIDOfArgReg);
+ if (regNumForArg == regInfo.getInvalidRegNum())
+ argInfo.setUseStackSlot();
+ else
+ argInfo.setUseFPArgReg();
+ }
+
+ // If this arg. is in the first $K$ regs, add special copy-
// float-to-int instructions to pass the value as an int.
- // To check if it is in teh first $K$, get the register
- // number for the arg #i.
- int copyRegNum = regInfo.regNumForIntArg(false, false,
- argNo, regClassIDOfArgReg);
+ // To check if it is in the first $K$, get the register
+ // number for the arg #i. These copy instructions are
+ // generated here because they are extra cases and not needed
+ // for the normal argument handling (some code reuse is
+ // possible though -- later).
+ //
+ int copyRegNum = regInfo.regNumForIntArg(false, false, argNo,
+ regClassIDOfArgReg);
if (copyRegNum != regInfo.getInvalidRegNum()) {
// Create a virtual register to represent copyReg. Mark
// this vreg as being an implicit operand of the call MI
const Type* loadTy = (argType == Type::FloatTy
? Type::IntTy : Type::LongTy);
- TmpInstruction* argVReg= new TmpInstruction(mcfi,loadTy,
- argVal, NULL, "argRegCopy");
+ TmpInstruction* argVReg = new TmpInstruction(mcfi, loadTy,
+ argVal, NULL,
+ "argRegCopy");
callMI->addImplicitRef(argVReg);
-
+
// Get a temp stack location to use to copy
// float-to-int via the stack.
//
int tmpOffset = MF.getInfo()->allocateLocalVar(argVReg);
// Generate the store from FP reg to stack
- M = BuildMI(ChooseStoreInstruction(argType), 3)
+ unsigned StoreOpcode = ChooseStoreInstruction(argType);
+ M = BuildMI(convertOpcodeFromRegToImm(StoreOpcode), 3)
.addReg(argVal).addMReg(regInfo.getFramePointer())
.addSImm(tmpOffset);
mvec.push_back(M);
// Generate the load from stack to int arg reg
- M = BuildMI(ChooseLoadInstruction(loadTy), 3)
+ unsigned LoadOpcode = ChooseLoadInstruction(loadTy);
+ M = BuildMI(convertOpcodeFromRegToImm(LoadOpcode), 3)
.addMReg(regInfo.getFramePointer()).addSImm(tmpOffset)
.addReg(argVReg, MOTy::Def);
// Mark operand with register it should be assigned
// both for copy and for the callMI
M->SetRegForOperand(M->getNumOperands()-1, copyRegNum);
- callMI->SetRegForImplicitRef(
- callMI->getNumImplicitRefs()-1, copyRegNum);
-
+ callMI->SetRegForImplicitRef(callMI->getNumImplicitRefs()-1,
+ copyRegNum);
mvec.push_back(M);
// Add info about the argument to the CallArgsDescriptor
M = BuildMI(storeOpCode, 3).addReg(argVal)
.addMReg(regInfo.getStackPointer()).addSImm(argOffset);
mvec.push_back(M);
- } else {
+ }
+ else if (regNumForArg != regInfo.getInvalidRegNum()) {
+
// Create a virtual register to represent the arg reg. Mark
// this vreg as being an implicit operand of the call MI.
TmpInstruction* argVReg =
new TmpInstruction(mcfi, argVal, NULL, "argReg");
callMI->addImplicitRef(argVReg);
-
+
// Generate the reg-to-reg copy into the outgoing arg reg.
// -- For FP values, create a FMOVS or FMOVD instruction
// -- For non-FP values, create an add-with-0 instruction
M = (BuildMI(ChooseAddInstructionByType(argType), 3)
.addReg(argVal).addSImm((int64_t) 0)
.addReg(argVReg, MOTy::Def));
-
+
// Mark the operand with the register it should be assigned
M->SetRegForOperand(M->getNumOperands()-1, regNumForArg);
callMI->SetRegForImplicitRef(callMI->getNumImplicitRefs()-1,
mvec.push_back(M);
}
+ else
+ assert(argInfo.getArgCopy() != regInfo.getInvalidRegNum() &&
+ "Arg. not in stack slot, primary or secondary register?");
}
// add call instruction and delay slot before copying return value
const Type* opType = argVal1->getType();
assert((opType->isInteger() || isa<PointerType>(opType)) &&
"Shl unsupported for other types");
+ unsigned opSize = target.getTargetData().getTypeSize(opType);
CreateShiftInstructions(target, shlInstr->getParent()->getParent(),
- (opType == Type::LongTy)? V9::SLLXr6:V9::SLLr6,
+ (opSize > 4)? V9::SLLXr6:V9::SLLr5,
argVal1, argVal2, 0, shlInstr, mvec,
MachineCodeForInstruction::get(shlInstr));
break;
const Type* opType = subtreeRoot->leftChild()->getValue()->getType();
assert((opType->isInteger() || isa<PointerType>(opType)) &&
"Shr unsupported for other types");
+ unsigned opSize = target.getTargetData().getTypeSize(opType);
Add3OperandInstr(opType->isSigned()
- ? (opType == Type::LongTy ? V9::SRAXr6 : V9::SRAr6)
- : (opType == Type::LongTy ? V9::SRLXr6 : V9::SRLr6),
+ ? (opSize > 4? V9::SRAXr6 : V9::SRAr5)
+ : (opSize > 4? V9::SRLXr6 : V9::SRLr5),
subtreeRoot, mvec);
break;
}
case 64: // reg: Phi(reg,reg)
break; // don't forward the value
- case 65: // reg: VaArg(reg)
- {
- // Use value initialized by va_start as pointer to args on the stack.
- // Load argument via current pointer value, then increment pointer.
+ case 65: // reg: VANext(reg): the va_next(va_list, type) instruction
+ { // Increment the va_list pointer register according to the type.
+ // All LLVM argument types are <= 64 bits, so use one doubleword.
+ Instruction* vaNextI = subtreeRoot->getInstruction();
+ assert(target.getTargetData().getTypeSize(vaNextI->getType()) <= 8 &&
+ "We assumed that all LLVM parameter types <= 8 bytes!");
int argSize = target.getFrameInfo().getSizeOfEachArgOnStack();
+ mvec.push_back(BuildMI(V9::ADDi, 3).addReg(vaNextI->getOperand(0)).
+ addSImm(argSize).addRegDef(vaNextI));
+ break;
+ }
+
+ case 66: // reg: VAArg (reg): the va_arg instruction
+ { // Load argument from stack using current va_list pointer value.
+ // Use 64-bit load for all non-FP args, and LDDF or double for FP.
Instruction* vaArgI = subtreeRoot->getInstruction();
- mvec.push_back(BuildMI(V9::LDXi, 3).addReg(vaArgI->getOperand(0)).
+ MachineOpCode loadOp = (vaArgI->getType()->isFloatingPoint()
+ ? (vaArgI->getType() == Type::FloatTy
+ ? V9::LDFi : V9::LDDFi)
+ : V9::LDXi);
+ mvec.push_back(BuildMI(loadOp, 3).addReg(vaArgI->getOperand(0)).
addSImm(0).addRegDef(vaArgI));
- mvec.push_back(BuildMI(V9::ADDi, 3).addReg(vaArgI->getOperand(0)).
- addSImm(argSize).addRegDef(vaArgI->getOperand(0)));
break;
}
if (dest->getType()->isUnsigned()) {
unsigned destSize=target.getTargetData().getTypeSize(dest->getType());
if (destSize <= 4) {
- // Mask high bits. Use a TmpInstruction to represent the
+ // Mask high 64 - N bits, where N = 4*destSize.
+
+ // 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(MachineCodeForInstruction::get(dest),
- dest->getType(), dest, NULL, "maskHi");
+ //
+ MachineCodeForInstruction& mcfi = MachineCodeForInstruction::get(dest);
+ TmpInstruction *tmpI = new TmpInstruction(mcfi, dest->getType(),
+ dest, NULL, "maskHi");
+ Value* srlArgToUse = tmpI;
+
+ unsigned numSubst = 0;
+ for (unsigned i=0, N=mvec.size(); i < N; ++i) {
+
+ // Make sure we substitute all occurrences of dest in these instrs.
+ // Otherwise, we will have bogus code.
+ bool someArgsWereIgnored = false;
+
+ // Make sure not to substitute an upwards-exposed use -- that would
+ // introduce a use of `tmpI' with no preceding def. Therefore,
+ // substitute a use or def-and-use operand only if a previous def
+ // operand has already been substituted (i.e., numSusbt > 0).
+ //
+ numSubst += mvec[i]->substituteValue(dest, tmpI,
+ /*defsOnly*/ numSubst == 0,
+ /*notDefsAndUses*/ numSubst > 0,
+ someArgsWereIgnored);
+ assert(!someArgsWereIgnored &&
+ "Operand `dest' exists but not replaced: probably bogus!");
+ }
+ assert(numSubst > 0 && "Operand `dest' not replaced: probably bogus!");
+
+ // Left shift 32-N if size (N) is less than 32 bits.
+ // Use another tmp. virtual register to represent this result.
+ if (destSize < 4) {
+ srlArgToUse = new TmpInstruction(mcfi, dest->getType(),
+ tmpI, NULL, "maskHi2");
+ mvec.push_back(BuildMI(V9::SLLXi6, 3).addReg(tmpI)
+ .addZImm(8*(4-destSize))
+ .addReg(srlArgToUse, MOTy::Def));
+ }
- for (unsigned i=0, N=mvec.size(); i < N; ++i)
- mvec[i]->substituteValue(dest, tmpI);
+ // Logical right shift 32-N to get zero extension in top 64-N bits.
+ mvec.push_back(BuildMI(V9::SRLi5, 3).addReg(srlArgToUse)
+ .addZImm(8*(4-destSize)).addReg(dest, MOTy::Def));
- M = BuildMI(V9::SRLi6, 3).addReg(tmpI).addZImm(8*(4-destSize))
- .addReg(dest, MOTy::Def);
- mvec.push_back(M);
} else if (destSize < 8) {
assert(0 && "Unsupported type size: 32 < size < 64 bits");
}