void TR::ARMSystemLinkage::createEpilogue(TR::Instruction *cursor)
{
TR::CodeGenerator *codeGen = cg();
const TR::ARMLinkageProperties& properties = getProperties();
TR::Machine *machine = codeGen->machine();
TR::Node *lastNode = cursor->getNode();
TR::ResolvedMethodSymbol* bodySymbol = comp()->getJittedMethodSymbol();
TR::RealRegister *stackPtr = machine->getRealRegister(properties.getStackPointerRegister());
// restore link register (r14)
auto *stackSlot = new (trHeapMemory()) TR::MemoryReference(stackPtr, bodySymbol->getLocalMappingCursor(), codeGen);
cursor = generateMemSrc1Instruction(cg(), ARMOp_ldr, lastNode, stackSlot, machine->getRealRegister(TR::RealRegister::gr14), cursor);
// restore all preserved registers
for (int r = TR::RealRegister::gr4; r <= TR::RealRegister::gr11; ++r)
{
auto *stackSlot = new (trHeapMemory()) TR::MemoryReference(stackPtr, (TR::RealRegister::gr11 - r + 1)*4 + bodySymbol->getLocalMappingCursor(), codeGen);
cursor = generateMemSrc1Instruction(cg(), ARMOp_ldr, lastNode, stackSlot, machine->getRealRegister((TR::RealRegister::RegNum)r), cursor);
}
// remove space for preserved registers
auto frameSize = codeGen->getFrameSizeInBytes();
cursor = generateTrg1Src1ImmInstruction(codeGen, ARMOp_add, lastNode, stackPtr, stackPtr, frameSize, 0, cursor);
// return using `mov r15, r14`
TR::RealRegister *gr14 = machine->getRealRegister(TR::RealRegister::gr14);
TR::RealRegister *gr15 = machine->getRealRegister(TR::RealRegister::gr15);
cursor = generateTrg1Src1Instruction(codeGen, ARMOp_mov, lastNode, gr15, gr14, cursor);
}
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);
}
uint32_t
FrontEnd::calculateSizeOfStackAtlas(
bool encodeFourByteOffsets,
uint32_t numberOfSlotsMapped,
uint32_t bytesPerStackMap,
TR::Compilation *comp)
{
TR::CodeGenerator *cg = comp->cg();
TR::GCStackAtlas * stackAtlas = cg->getStackAtlas();
// Calculate the size of each individual map in the atlas. The fixed
// portion of the map contains:
//
// Low Code Offset (2 or 4)
// Stack map (depends on # of mapped parms/locals)
//
uint32_t sizeOfEncodedCodeOffsetInBytes = encodeFourByteOffsets ? 4 : 2;
uint32_t sizeOfSingleEncodedMapInBytes = sizeOfEncodedCodeOffsetInBytes;
sizeOfSingleEncodedMapInBytes += bytesPerStackMap;
// Calculate the atlas size
//
uint32_t atlasSize = sizeof(OMR::StackAtlasPOD);
ListIterator<TR_GCStackMap> mapIterator(&stackAtlas->getStackMapList());
TR_GCStackMap *mapCursor = mapIterator.getFirst();
while (mapCursor != NULL)
{
TR_GCStackMap *nextMapCursor = mapIterator.getNext();
if (!mapsAreIdentical(mapCursor, nextMapCursor, stackAtlas, comp))
{
atlasSize += sizeOfSingleEncodedMapInBytes;
}
mapCursor = nextMapCursor;
}
return atlasSize;
}
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::ARMSystemLinkage::createPrologue(TR::Instruction *cursor)
{
TR::CodeGenerator *codeGen = cg();
const TR::ARMLinkageProperties& properties = getProperties();
TR::Machine *machine = codeGen->machine();
TR::ResolvedMethodSymbol* bodySymbol = comp()->getJittedMethodSymbol();
TR::Node *firstNode = comp()->getStartTree()->getNode();
TR::RealRegister *stackPtr = machine->getRealRegister(properties.getStackPointerRegister());
// Entry breakpoint
//
if (comp()->getOption(TR_EntryBreakPoints))
{
cursor = new (trHeapMemory()) TR::Instruction(cursor, ARMOp_bad, firstNode, cg());
}
// allocate stack space
auto frameSize = codeGen->getFrameSizeInBytes();
cursor = generateTrg1Src1ImmInstruction(codeGen, ARMOp_sub, firstNode, stackPtr, stackPtr, frameSize, 0, cursor);
// spill argument registers
auto nextIntArgReg = 0;
auto nextFltArgReg = 0;
ListIterator<TR::ParameterSymbol> parameterIterator(&bodySymbol->getParameterList());
for (TR::ParameterSymbol *parameter = parameterIterator.getFirst();
parameter!=NULL && (nextIntArgReg < getProperties().getNumIntArgRegs() || nextFltArgReg < getProperties().getNumFloatArgRegs());
parameter=parameterIterator.getNext())
{
auto *stackSlot = new (trHeapMemory()) TR::MemoryReference(stackPtr, parameter->getParameterOffset(), codeGen);
switch (parameter->getDataType())
{
case TR::Int8:
case TR::Int16:
case TR::Int32:
case TR::Address:
if (nextIntArgReg < getProperties().getNumIntArgRegs())
{
cursor = generateMemSrc1Instruction(cg(), ARMOp_str, firstNode, stackSlot, machine->getRealRegister((TR::RealRegister::RegNum)(TR::RealRegister::gr0 + nextIntArgReg)), cursor);
nextIntArgReg++;
}
else
{
nextIntArgReg = getProperties().getNumIntArgRegs() + 1;
}
break;
case TR::Int64:
nextIntArgReg += nextIntArgReg & 0x1; // round to next even number
if (nextIntArgReg + 1 < getProperties().getNumIntArgRegs())
{
cursor = generateMemSrc1Instruction(cg(), ARMOp_str, firstNode, stackSlot, machine->getRealRegister((TR::RealRegister::RegNum)(TR::RealRegister::gr0 + nextIntArgReg)), cursor);
stackSlot = new (trHeapMemory()) TR::MemoryReference(stackPtr, parameter->getParameterOffset() + 4, codeGen);
cursor = generateMemSrc1Instruction(cg(), ARMOp_str, firstNode, stackSlot, machine->getRealRegister((TR::RealRegister::RegNum)(TR::RealRegister::gr0 + nextIntArgReg + 1)), cursor);
nextIntArgReg += 2;
}
else
{
nextIntArgReg = getProperties().getNumIntArgRegs() + 1;
}
break;
case TR::Float:
comp()->failCompilation<UnsupportedParameterType>("Compiling methods with a single precision floating point parameter is not supported");
break;
case TR::Double:
if (nextFltArgReg < getProperties().getNumFloatArgRegs())
{
cursor = generateMemSrc1Instruction(cg(), ARMOp_fstd, firstNode, stackSlot, machine->getRealRegister((TR::RealRegister::RegNum)(TR::RealRegister::fp0 + nextFltArgReg)), cursor);
nextFltArgReg += 1;
}
else
{
nextFltArgReg = getProperties().getNumFloatArgRegs() + 1;
}
break;
case TR::Aggregate:
TR_ASSERT(false, "Function parameters of aggregate types are not currently supported on ARM.");
}
}
// save all preserved registers
for (int r = TR::RealRegister::gr4; r <= TR::RealRegister::gr11; ++r)
{
auto *stackSlot = new (trHeapMemory()) TR::MemoryReference(stackPtr, (TR::RealRegister::gr11 - r + 1)*4 + bodySymbol->getLocalMappingCursor(), codeGen);
cursor = generateMemSrc1Instruction(cg(), ARMOp_str, firstNode, stackSlot, machine->getRealRegister((TR::RealRegister::RegNum)r), cursor);
}
// save link register (r14)
auto *stackSlot = new (trHeapMemory()) TR::MemoryReference(stackPtr, bodySymbol->getLocalMappingCursor(), codeGen);
cursor = generateMemSrc1Instruction(cg(), ARMOp_str, firstNode, stackSlot, machine->getRealRegister(TR::RealRegister::gr14), cursor);
}
TR::Node *
OMR::TransformUtil::scalarizeArrayCopy(
TR::Compilation *comp,
TR::Node *node,
TR::TreeTop *tt,
bool useElementType,
bool &didTransformArrayCopyNode,
TR::SymbolReference *sourceRef,
TR::SymbolReference *targetRef,
bool castToIntegral)
{
TR::CodeGenerator *cg = comp->cg();
didTransformArrayCopyNode = false;
if ((comp->getOptLevel() == noOpt) ||
!comp->getOption(TR_ScalarizeSSOps) ||
node->getOpCodeValue() != TR::arraycopy ||
node->getNumChildren() != 3 ||
comp->requiresSpineChecks() ||
!node->getChild(2)->getOpCode().isLoadConst() ||
cg->getOptimizationPhaseIsComplete())
return node;
int64_t byteLen = node->getChild(2)->get64bitIntegralValue();
if (byteLen == 0)
{
if (tt)
{
// Anchor the first two children
if (!node->getFirstChild()->safeToDoRecursiveDecrement())
TR::TreeTop::create(comp, tt->getPrevTreeTop(),
TR::Node::create(TR::treetop, 1, node->getFirstChild()));
if (!node->getSecondChild()->safeToDoRecursiveDecrement())
TR::TreeTop::create(comp, tt->getPrevTreeTop(),
TR::Node::create(TR::treetop, 1, node->getSecondChild()));
tt->getPrevTreeTop()->join(tt->getNextTreeTop());
tt->getNode()->recursivelyDecReferenceCount();
didTransformArrayCopyNode = true;
}
return node;
}
else if (byteLen < 0)
{
return node;
}
else if (byteLen > TR_MAX_OTYPE_SIZE)
{
return node;
}
TR::DataType dataType = TR::Aggregate;
// Get the element datatype from the (hidden) 4th child
TR::DataType elementType = node->getArrayCopyElementType();
int32_t elementSize = TR::Symbol::convertTypeToSize(elementType);
if (byteLen == elementSize)
{
dataType = elementType;
}
else if (!useElementType)
{
switch (byteLen)
{
case 1: dataType = TR::Int8; break;
case 2: dataType = TR::Int16; break;
case 4: dataType = TR::Int32; break;
case 8: dataType = TR::Int64; break;
}
}
else
{
return node;
}
// load/store double on 64-bit PPC requires offset to be word aligned
// abort if this requirement is not met.
// TODO: also need to check if the first two children are aload nodes
bool cannot_use_load_store_long = false;
if (TR::Compiler->target.cpu.isPower())
if (dataType == TR::Int64 && TR::Compiler->target.is64Bit())
{
TR::Node * firstChild = node->getFirstChild();
if (firstChild->getNumChildren() == 2)
{
TR::Node *offsetChild = firstChild->getSecondChild();
TR_ASSERT(offsetChild->getOpCodeValue() != TR::iconst, "iconst shouldn't be used for 64-bit array indexing");
if (offsetChild->getOpCodeValue() == TR::lconst)
{
if ((offsetChild->getLongInt() & 0x3) != 0)
cannot_use_load_store_long = true;
}
}
TR::Node *secondChild = node->getSecondChild();
if (secondChild->getNumChildren() == 2)
{
TR::Node *offsetChild = secondChild->getSecondChild();
TR_ASSERT(offsetChild->getOpCodeValue() != TR::iconst, "iconst shouldn't be used for 64-bit array indexing");
if (offsetChild->getOpCodeValue() == TR::lconst)
{
if ((offsetChild->getLongInt() & 0x3) != 0)
cannot_use_load_store_long = true;
}
}
}
if (cannot_use_load_store_long) return node;
TR::SymbolReference *nodeRef;
targetRef = comp->getSymRefTab()->findOrCreateGenericIntShadowSymbolReference(0);
sourceRef = targetRef;
bool trace = comp->getOption(TR_TraceScalarizeSSOps);
if (trace)
traceMsg(comp,"scalarizeArrayCopy: node %p got targetRef (#%d) and sourceRef (#%d)\n",
node,targetRef?targetRef->getReferenceNumber():-1,sourceRef?sourceRef->getReferenceNumber():-1);
if (targetRef == NULL || sourceRef == NULL)
{
if (trace)
traceMsg(comp,"do not scalarizeArrayCopy node %p : targetRef is NULL (%s) or sourceRef is NULL (%s)\n",node,targetRef?"no":"yes",sourceRef?"no":"yes");
return node;
}
#ifdef J9_PROJECT_SPECIFIC
if (targetRef->getSymbol()->getDataType().isBCD() ||
sourceRef->getSymbol()->getDataType().isBCD())
{
return node;
}
#endif
if (performTransformation(comp, "%sScalarize arraycopy 0x%p\n", OPT_DETAILS, node))
{
TR::Node *store = TR::TransformUtil::scalarizeAddressParameter(comp, node->getSecondChild(), byteLen, dataType, targetRef, true);
TR::Node *load = TR::TransformUtil::scalarizeAddressParameter(comp, node->getFirstChild(), byteLen, dataType, sourceRef, false);
if (tt)
{
// Transforming
// treetop
// arrayCopy <-- node
// into
// *store
//
node->recursivelyDecReferenceCount();
tt->setNode(node);
}
else
{
for (int16_t c = node->getNumChildren() - 1; c >= 0; c--)
cg->recursivelyDecReferenceCount(node->getChild(c));
}
TR::Node::recreate(node, store->getOpCodeValue());
node->setSymbolReference(store->getSymbolReference());
if (store->getOpCode().isStoreIndirect())
{
node->setChild(0, store->getFirstChild());
node->setAndIncChild(1, load);
node->setNumChildren(2);
}
else
{
node->setAndIncChild(0, load);
node->setNumChildren(1);
}
didTransformArrayCopyNode = true;
}
return node;
}
uint8_t *
FrontEnd::createStackAtlas(
bool encodeFourByteOffsets,
uint32_t numberOfSlotsMapped,
uint32_t bytesPerStackMap,
uint8_t *encodedAtlasBaseAddress,
uint32_t atlasSizeInBytes,
TR::Compilation *comp)
{
TR::CodeGenerator *cg = comp->cg();
TR::GCStackAtlas *stackAtlas = cg->getStackAtlas();
stackAtlas->setAtlasBits(encodedAtlasBaseAddress);
// Calculate the size of each individual map in the atlas. The fixed
// portion of the map contains:
//
// Low Code Offset (2 or 4)
// Stack map (depends on # of mapped parms/locals)
//
uint32_t sizeOfEncodedCodeOffsetInBytes = encodeFourByteOffsets ? 4 : 2;
uint32_t sizeOfSingleEncodedMapInBytes = sizeOfEncodedCodeOffsetInBytes;
sizeOfSingleEncodedMapInBytes += bytesPerStackMap;
// Encode the atlas
//
OMR::StackAtlasPOD *pyAtlas = (OMR::StackAtlasPOD *)encodedAtlasBaseAddress;
pyAtlas->numberOfMaps = stackAtlas->getNumberOfMaps();
pyAtlas->bytesPerStackMap = bytesPerStackMap;
// Offset to the MAPPED pyFrameObject parameter
//
pyAtlas->frameObjectParmOffset = 0;
// Lowest stack offset where MAPPED locals begin.
//
pyAtlas->localBaseOffset = stackAtlas->getLocalBaseOffset();
// Abort if we have overflowed the fields in pyAtlas.
//
if (bytesPerStackMap > USHRT_MAX ||
stackAtlas->getNumberOfMaps() > USHRT_MAX ||
stackAtlas->getNumberOfParmSlotsMapped() > USHRT_MAX ||
stackAtlas->getParmBaseOffset() < SHRT_MIN || stackAtlas->getParmBaseOffset() > SHRT_MAX ||
stackAtlas->getLocalBaseOffset() < SHRT_MIN || stackAtlas->getLocalBaseOffset() > SHRT_MAX)
{
comp->failCompilation<TR::CompilationException>("Overflowed the fields in pyAtlas");
}
// Maps are in reverse order in list from what we want in the atlas
// so advance to the address where the last map should go and start
// building the maps moving back toward the beginning of the atlas.
//
uint8_t *cursorInEncodedAtlas = encodedAtlasBaseAddress + atlasSizeInBytes;
ListIterator<TR_GCStackMap> mapIterator(&stackAtlas->getStackMapList());
TR_GCStackMap *mapCursor = mapIterator.getFirst();
while (mapCursor != NULL)
{
// Move back from the end of the atlas till the current map can be fit in,
// then pass the cursor to the routine that actually creates and fills in
// the stack map
//
TR_GCStackMap *nextMapCursor = mapIterator.getNext();
if (!mapsAreIdentical(mapCursor, nextMapCursor, stackAtlas, comp))
{
cursorInEncodedAtlas -= sizeOfSingleEncodedMapInBytes;
encodeStackMap(mapCursor, cursorInEncodedAtlas, encodeFourByteOffsets, bytesPerStackMap, comp);
}
mapCursor = nextMapCursor;
}
return encodedAtlasBaseAddress;
}
