void
TR::ARM64SystemLinkage::createEpilogue(TR::Instruction *cursor)
{
TR::CodeGenerator *codeGen = cg();
const TR::ARM64LinkageProperties& properties = getProperties();
TR::Machine *machine = codeGen->machine();
TR::Node *lastNode = cursor->getNode();
TR::ResolvedMethodSymbol *bodySymbol = comp()->getJittedMethodSymbol();
TR::RealRegister *sp = machine->getRealRegister(properties.getStackPointerRegister());
// restore callee-saved registers
uint32_t offset = bodySymbol->getLocalMappingCursor();
for (int r = TR::RealRegister::x19; r <= TR::RealRegister::x28; r++)
{
TR::RealRegister *rr = machine->getRealRegister((TR::RealRegister::RegNum)r);
if (rr->getHasBeenAssignedInMethod())
{
TR::MemoryReference *stackSlot = new (trHeapMemory()) TR::MemoryReference(sp, offset, codeGen);
cursor = generateTrg1MemInstruction(cg(), TR::InstOpCode::ldrimmx, lastNode, rr, stackSlot, cursor);
offset += 8;
}
}
for (int r = TR::RealRegister::v8; r <= TR::RealRegister::v15; r++)
{
TR::RealRegister *rr = machine->getRealRegister((TR::RealRegister::RegNum)r);
if (rr->getHasBeenAssignedInMethod())
{
TR::MemoryReference *stackSlot = new (trHeapMemory()) TR::MemoryReference(sp, offset, codeGen);
cursor = generateTrg1MemInstruction(cg(), TR::InstOpCode::vldrimmd, lastNode, rr, stackSlot, cursor);
offset += 8;
}
}
// restore link register (x30)
TR::RealRegister *lr = machine->getRealRegister(TR::RealRegister::lr);
if (machine->getLinkRegisterKilled())
{
TR::MemoryReference *stackSlot = new (trHeapMemory()) TR::MemoryReference(sp, 0, codeGen);
cursor = generateTrg1MemInstruction(cg(), TR::InstOpCode::ldrimmx, lastNode, lr, stackSlot, cursor);
}
// remove space for preserved registers
uint32_t frameSize = codeGen->getFrameSizeInBytes();
if (constantIsUnsignedImm12(frameSize))
{
cursor = generateTrg1Src1ImmInstruction(codeGen, TR::InstOpCode::addimmx, lastNode, sp, sp, frameSize, cursor);
}
else
{
TR_UNIMPLEMENTED();
}
// return
cursor = generateRegBranchInstruction(codeGen, TR::InstOpCode::ret, lastNode, lr, cursor);
}
TR::Instruction *
TR_X86SystemLinkage::savePreservedRegisters(TR::Instruction *cursor)
{
// For IA32, if disableShrinkWrapping, usePushForPreservedRegs will be true; otherwise false;
// For X64, shrinkWraping is always on, and usePushForPreservedRegs always false;
// TR_ASSERT(!getProperties().getUsesPushesForPreservedRegs(), "assertion failure");
TR::ResolvedMethodSymbol *bodySymbol = comp()->getJittedMethodSymbol();
const int32_t localSize = getProperties().getOffsetToFirstLocal() - bodySymbol->getLocalMappingCursor();
const int32_t pointerSize = getProperties().getPointerSize();
int32_t offsetCursor = -localSize + getProperties().getOffsetToFirstLocal() - pointerSize;
if (_properties.getUsesPushesForPreservedRegs())
{
for (int32_t pindex = _properties.getMaxRegistersPreservedInPrologue()-1;
pindex >= 0;
pindex--)
{
TR::RealRegister::RegNum idx = _properties.getPreservedRegister((uint32_t)pindex);
TR::RealRegister *reg = machine()->getX86RealRegister(idx);
if (reg->getHasBeenAssignedInMethod() && reg->getState() != TR::RealRegister::Locked)
{
cursor = new (trHeapMemory()) TR::X86RegInstruction(cursor, PUSHReg, reg, cg());
}
}
}
else
{
TR_BitVector *p = cg()->getPreservedRegsInPrologue();
for (int32_t pindex = getProperties().getMaxRegistersPreservedInPrologue()-1;
pindex >= 0;
pindex--)
{
TR::RealRegister::RegNum idx = _properties.getPreservedRegister((uint32_t)pindex);
TR::RealRegister *reg = machine()->getX86RealRegister(getProperties().getPreservedRegister((uint32_t)pindex));
if(reg->getHasBeenAssignedInMethod() && reg->getState() != TR::RealRegister::Locked)
{
if (!p || p->get(idx))
{
cursor = generateMemRegInstruction(
cursor,
movOpcodes[MemReg][fullRegisterMovType(reg)],
generateX86MemoryReference(machine()->getX86RealRegister(TR::RealRegister::vfp), offsetCursor, cg()),
reg,
cg()
);
}
offsetCursor -= pointerSize;
}
}
}
return cursor;
}
TR::Instruction *
TR::X86SystemLinkage::savePreservedRegisters(TR::Instruction *cursor)
{
// For IA32 usePushForPreservedRegs will be true;
// For X64, usePushForPreservedRegs always false;
TR::ResolvedMethodSymbol *bodySymbol = comp()->getJittedMethodSymbol();
const int32_t localSize = getProperties().getOffsetToFirstLocal() - bodySymbol->getLocalMappingCursor();
const int32_t pointerSize = getProperties().getPointerSize();
int32_t offsetCursor = -localSize + getProperties().getOffsetToFirstLocal() - pointerSize;
if (_properties.getUsesPushesForPreservedRegs())
{
for (int32_t pindex = _properties.getMaxRegistersPreservedInPrologue()-1;
pindex >= 0;
pindex--)
{
TR::RealRegister::RegNum idx = _properties.getPreservedRegister((uint32_t)pindex);
TR::RealRegister *reg = machine()->getX86RealRegister(idx);
if (reg->getHasBeenAssignedInMethod() && reg->getState() != TR::RealRegister::Locked)
{
cursor = new (trHeapMemory()) TR::X86RegInstruction(cursor, PUSHReg, reg, cg());
}
}
}
else
{
for (int32_t pindex = getProperties().getMaxRegistersPreservedInPrologue()-1;
pindex >= 0;
pindex--)
{
TR::RealRegister::RegNum idx = _properties.getPreservedRegister((uint32_t)pindex);
TR::RealRegister *reg = machine()->getX86RealRegister(getProperties().getPreservedRegister((uint32_t)pindex));
if(reg->getHasBeenAssignedInMethod() && reg->getState() != TR::RealRegister::Locked)
{
cursor = generateMemRegInstruction(
cursor,
TR::Linkage::movOpcodes(MemReg, fullRegisterMovType(reg)),
generateX86MemoryReference(machine()->getX86RealRegister(TR::RealRegister::vfp), offsetCursor, cg()),
reg,
cg()
);
offsetCursor -= pointerSize;
}
}
}
return cursor;
}
void
TR::ARM64SystemLinkage::createPrologue(TR::Instruction *cursor, List<TR::ParameterSymbol> &parmList)
{
TR::CodeGenerator *codeGen = cg();
TR::Machine *machine = codeGen->machine();
TR::ResolvedMethodSymbol *bodySymbol = comp()->getJittedMethodSymbol();
const TR::ARM64LinkageProperties& properties = getProperties();
TR::RealRegister *sp = machine->getRealRegister(properties.getStackPointerRegister());
TR::Node *firstNode = comp()->getStartTree()->getNode();
// allocate stack space
uint32_t frameSize = (uint32_t)codeGen->getFrameSizeInBytes();
if (constantIsUnsignedImm12(frameSize))
{
cursor = generateTrg1Src1ImmInstruction(codeGen, TR::InstOpCode::subimmx, firstNode, sp, sp, frameSize, cursor);
}
else
{
TR_UNIMPLEMENTED();
}
// save link register (x30)
if (machine->getLinkRegisterKilled())
{
TR::MemoryReference *stackSlot = new (trHeapMemory()) TR::MemoryReference(sp, 0, codeGen);
cursor = generateMemSrc1Instruction(cg(), TR::InstOpCode::strimmx, firstNode, stackSlot, machine->getRealRegister(TR::RealRegister::x30), cursor);
}
// spill argument registers
int32_t nextIntArgReg = 0;
int32_t nextFltArgReg = 0;
ListIterator<TR::ParameterSymbol> parameterIterator(&parmList);
for (TR::ParameterSymbol *parameter = parameterIterator.getFirst();
parameter != NULL && (nextIntArgReg < getProperties().getNumIntArgRegs() || nextFltArgReg < getProperties().getNumFloatArgRegs());
parameter = parameterIterator.getNext())
{
TR::MemoryReference *stackSlot = new (trHeapMemory()) TR::MemoryReference(sp, parameter->getParameterOffset(), codeGen);
TR::InstOpCode::Mnemonic op;
switch (parameter->getDataType())
{
case TR::Int8:
case TR::Int16:
case TR::Int32:
case TR::Int64:
case TR::Address:
if (nextIntArgReg < getProperties().getNumIntArgRegs())
{
op = (parameter->getSize() == 8) ? TR::InstOpCode::strimmx : TR::InstOpCode::strimmw;
cursor = generateMemSrc1Instruction(cg(), op, firstNode, stackSlot, machine->getRealRegister((TR::RealRegister::RegNum)(TR::RealRegister::x0 + nextIntArgReg)), cursor);
nextIntArgReg++;
}
else
{
nextIntArgReg = getProperties().getNumIntArgRegs() + 1;
}
break;
case TR::Float:
case TR::Double:
if (nextFltArgReg < getProperties().getNumFloatArgRegs())
{
op = (parameter->getSize() == 8) ? TR::InstOpCode::vstrimmd : TR::InstOpCode::vstrimms;
cursor = generateMemSrc1Instruction(cg(), op, firstNode, stackSlot, machine->getRealRegister((TR::RealRegister::RegNum)(TR::RealRegister::v0 + nextFltArgReg)), cursor);
nextFltArgReg++;
}
else
{
nextFltArgReg = getProperties().getNumFloatArgRegs() + 1;
}
break;
case TR::Aggregate:
TR_ASSERT(false, "Function parameters of aggregate types are not currently supported on AArch64.");
break;
default:
TR_ASSERT(false, "Unknown parameter type.");
}
}
// save callee-saved registers
uint32_t offset = bodySymbol->getLocalMappingCursor();
for (int r = TR::RealRegister::x19; r <= TR::RealRegister::x28; r++)
{
TR::RealRegister *rr = machine->getRealRegister((TR::RealRegister::RegNum)r);
if (rr->getHasBeenAssignedInMethod())
{
TR::MemoryReference *stackSlot = new (trHeapMemory()) TR::MemoryReference(sp, offset, codeGen);
cursor = generateMemSrc1Instruction(cg(), TR::InstOpCode::strimmx, firstNode, stackSlot, rr, cursor);
offset += 8;
}
}
for (int r = TR::RealRegister::v8; r <= TR::RealRegister::v15; r++)
{
TR::RealRegister *rr = machine->getRealRegister((TR::RealRegister::RegNum)r);
if (rr->getHasBeenAssignedInMethod())
{
TR::MemoryReference *stackSlot = new (trHeapMemory()) TR::MemoryReference(sp, offset, codeGen);
cursor = generateMemSrc1Instruction(cg(), TR::InstOpCode::vstrimmd, firstNode, stackSlot, rr, cursor);
offset += 8;
}
}
}
void
TR::ARM64SystemLinkage::mapStack(TR::ResolvedMethodSymbol *method)
{
TR::Machine *machine = cg()->machine();
uint32_t stackIndex = 0;
ListIterator<TR::AutomaticSymbol> automaticIterator(&method->getAutomaticList());
TR::AutomaticSymbol *localCursor = automaticIterator.getFirst();
stackIndex = 8; // [sp+0] is for link register
// map non-long/double automatics
while (localCursor != NULL)
{
if (localCursor->getGCMapIndex() < 0
&& localCursor->getDataType() != TR::Int64
&& localCursor->getDataType() != TR::Double)
{
localCursor->setOffset(stackIndex);
stackIndex += (localCursor->getSize() + 3) & (~3);
}
localCursor = automaticIterator.getNext();
}
stackIndex += (stackIndex & 0x4) ? 4 : 0; // align to 8 bytes
automaticIterator.reset();
localCursor = automaticIterator.getFirst();
// map long/double automatics
while (localCursor != NULL)
{
if (localCursor->getDataType() == TR::Int64
|| localCursor->getDataType() == TR::Double)
{
localCursor->setOffset(stackIndex);
stackIndex += (localCursor->getSize() + 7) & (~7);
}
localCursor = automaticIterator.getNext();
}
method->setLocalMappingCursor(stackIndex);
// allocate space for preserved registers (x19-x28, v8-v15)
for (int r = TR::RealRegister::x19; r <= TR::RealRegister::x28; r++)
{
TR::RealRegister *rr = machine->getRealRegister((TR::RealRegister::RegNum)r);
if (rr->getHasBeenAssignedInMethod())
{
stackIndex += 8;
}
}
for (int r = TR::RealRegister::v8; r <= TR::RealRegister::v15; r++)
{
TR::RealRegister *rr = machine->getRealRegister((TR::RealRegister::RegNum)r);
if (rr->getHasBeenAssignedInMethod())
{
stackIndex += 8;
}
}
/*
* Because the rest of the code generator currently expects **all** arguments
* to be passed on the stack, arguments passed in registers must be spilled
* in the callee frame. To map the arguments correctly, we use two loops. The
* first maps the arguments that will come in registers onto the callee stack.
* At the end of this loop, the `stackIndex` is the the size of the frame.
* The second loop then maps the remaining arguments onto the caller frame.
*/
int32_t nextIntArgReg = 0;
int32_t nextFltArgReg = 0;
ListIterator<TR::ParameterSymbol> parameterIterator(&method->getParameterList());
for (TR::ParameterSymbol *parameter = parameterIterator.getFirst();
parameter != NULL && (nextIntArgReg < getProperties().getNumIntArgRegs() || nextFltArgReg < getProperties().getNumFloatArgRegs());
parameter = parameterIterator.getNext())
{
switch (parameter->getDataType())
{
case TR::Int8:
case TR::Int16:
case TR::Int32:
case TR::Int64:
case TR::Address:
if (nextIntArgReg < getProperties().getNumIntArgRegs())
{
nextIntArgReg++;
mapSingleParameter(parameter, stackIndex, true);
}
else
{
nextIntArgReg = getProperties().getNumIntArgRegs() + 1;
}
break;
case TR::Float:
case TR::Double:
if (nextFltArgReg < getProperties().getNumFloatArgRegs())
{
nextFltArgReg++;
mapSingleParameter(parameter, stackIndex, true);
}
else
{
nextFltArgReg = getProperties().getNumFloatArgRegs() + 1;
}
break;
case TR::Aggregate:
TR_ASSERT(false, "Function parameters of aggregate types are not currently supported on AArch64.");
break;
default:
TR_ASSERT(false, "Unknown parameter type.");
}
}
// save the stack frame size, aligned to 16 bytes
stackIndex = (stackIndex + 15) & (~15);
cg()->setFrameSizeInBytes(stackIndex);
nextIntArgReg = 0;
nextFltArgReg = 0;
parameterIterator.reset();
for (TR::ParameterSymbol *parameter = parameterIterator.getFirst();
parameter != NULL && (nextIntArgReg < getProperties().getNumIntArgRegs() || nextFltArgReg < getProperties().getNumFloatArgRegs());
parameter = parameterIterator.getNext())
{
switch (parameter->getDataType())
{
case TR::Int8:
case TR::Int16:
case TR::Int32:
case TR::Int64:
case TR::Address:
if (nextIntArgReg < getProperties().getNumIntArgRegs())
{
nextIntArgReg++;
}
else
{
mapSingleParameter(parameter, stackIndex, false);
}
break;
case TR::Float:
case TR::Double:
if (nextFltArgReg < getProperties().getNumFloatArgRegs())
{
nextFltArgReg++;
}
else
{
mapSingleParameter(parameter, stackIndex, false);
}
break;
case TR::Aggregate:
TR_ASSERT(false, "Function parameters of aggregate types are not currently supported on AArch64.");
break;
default:
TR_ASSERT(false, "Unknown parameter type.");
}
}
}
