uint8_t TR::ExternalOrderedPair32BitRelocation::collectModifier()
{
TR::Compilation *comp = TR::comp();
uint8_t * relocatableMethodCodeStart = (uint8_t *)comp->getRelocatableMethodCodeStart();
uint8_t * updateLocation;
uint8_t * updateLocation2;
TR_ExternalRelocationTargetKind kind = getTargetKind();
if (TR::Compiler->target.cpu.isPower() &&
(kind == TR_ArrayCopyHelper || kind == TR_ArrayCopyToc || kind == TR_RamMethod || kind == TR_GlobalValue || kind == TR_BodyInfoAddressLoad || kind == TR_DataAddress || kind == TR_DebugCounter))
{
TR::Instruction *instr = (TR::Instruction *)getUpdateLocation();
TR::Instruction *instr2 = (TR::Instruction *)getLocation2();
updateLocation = instr->getBinaryEncoding();
updateLocation2 = instr2->getBinaryEncoding();
}
else
{
updateLocation = getUpdateLocation();
updateLocation2 = getLocation2();
}
int32_t iLoc = updateLocation - relocatableMethodCodeStart;
int32_t iLoc2 = updateLocation2 - relocatableMethodCodeStart;
AOTcgDiag0(comp, "TR::ExternalOrderedPair32BitRelocation::collectModifier\n");
if ( (iLoc < MIN_SHORT_OFFSET || iLoc > MAX_SHORT_OFFSET ) || (iLoc2 < MIN_SHORT_OFFSET || iLoc2 > MAX_SHORT_OFFSET ) )
return RELOCATION_TYPE_WIDE_OFFSET | RELOCATION_TYPE_ORDERED_PAIR;
return RELOCATION_TYPE_ORDERED_PAIR;
}
void TR_OutlinedInstructions::assignRegistersOnOutlinedPath(TR_RegisterKinds kindsToBeAssigned, TR::X86VFPSaveInstruction *vfpSaveInstruction)
{
if (hasBeenRegisterAssigned())
{
TR_ASSERT(0, "these registers should not have been assigned already");
return;
}
// Register assign the outlined instructions.
//
_cg->doBackwardsRegisterAssignment(kindsToBeAssigned, _appendInstruction);
// Ensure correct VFP state at the start of the outlined instruction sequence.
//
generateVFPRestoreInstruction(comp()->getAppendInstruction(), vfpSaveInstruction, _cg);
// Link in the helper stream into the mainline code.
//
TR::Instruction *appendInstruction = comp()->getAppendInstruction();
appendInstruction->setNext(_firstInstruction);
_firstInstruction->setPrev(appendInstruction);
comp()->setAppendInstruction(_appendInstruction);
setHasBeenRegisterAssigned(true);
}
void TR_PPCOutOfLineCodeSection::assignRegisters(TR_RegisterKinds kindsToBeAssigned)
{
TR::Compilation* comp = _cg->comp();
if (hasBeenRegisterAssigned())
return;
// nested internal control flow assert:
_cg->setInternalControlFlowSafeNestingDepth(_cg->internalControlFlowNestingDepth());
// Create a dependency list on the first instruction in this stream that captures all current real register associations.
// This is necessary to get the register assigner back into its original state before the helper stream was processed.
_cg->incOutOfLineColdPathNestedDepth();
// This prevents the OOL entry label from resetting all register's startOfranges during RA
_cg->toggleIsInOOLSection();
TR::RegisterDependencyConditions *liveRealRegDeps = _cg->machine()->createCondForLiveAndSpilledGPRs(true, _cg->getSpilledRegisterList());
if (liveRealRegDeps)
_firstInstruction->setDependencyConditions(liveRealRegDeps);
_cg->toggleIsInOOLSection(); // toggle it back because swapInstructionListsWithCompilation() also calls toggle...
// Register assign the helper dispatch instructions.
swapInstructionListsWithCompilation();
_cg->doRegisterAssignment(kindsToBeAssigned);
swapInstructionListsWithCompilation();
_cg->decOutOfLineColdPathNestedDepth();
// Returning to mainline, reset this counter
_cg->setInternalControlFlowSafeNestingDepth(0);
// Link in the helper stream into the mainline code.
// We will end up with the OOL items attached at the bottom of the instruction stream
TR::Instruction *appendInstruction = _cg->getAppendInstruction();
appendInstruction->setNext(_firstInstruction);
_firstInstruction->setPrev(appendInstruction);
_cg->setAppendInstruction(_appendInstruction);
setHasBeenRegisterAssigned(true);
}
void TR::ExternalOrderedPair32BitRelocation::apply(TR::CodeGenerator *codeGen)
{
TR::Compilation *comp = codeGen->comp();
AOTcgDiag0(comp, "TR::ExternalOrderedPair32BitRelocation::apply\n");
TR::IteratedExternalRelocation *rec = getRelocationRecord();
uint8_t *codeStart = (uint8_t *)comp->getRelocatableMethodCodeStart();
TR_ExternalRelocationTargetKind kind = getRelocationRecord()->getTargetKind();
if (TR::Compiler->target.cpu.isPower() &&
(kind == TR_ArrayCopyHelper || kind == TR_ArrayCopyToc || kind == TR_RamMethodSequence || kind == TR_GlobalValue || kind == TR_BodyInfoAddressLoad || kind == TR_DataAddress || kind == TR_DebugCounter))
{
TR::Instruction *instr = (TR::Instruction *)getUpdateLocation();
TR::Instruction *instr2 = (TR::Instruction *)getLocation2();
rec->addRelocationEntry((uint32_t)(instr->getBinaryEncoding() - codeStart));
rec->addRelocationEntry((uint32_t)(instr2->getBinaryEncoding() - codeStart));
}
else
{
rec->addRelocationEntry(getUpdateLocation() - codeStart);
rec->addRelocationEntry(getLocation2() - codeStart);
}
}
TR::Register *TR::AMD64SystemLinkage::buildDirectDispatch(
TR::Node *callNode,
bool spillFPRegs)
{
TR::SymbolReference *methodSymRef = callNode->getSymbolReference();
TR::MethodSymbol *methodSymbol = methodSymRef->getSymbol()->castToMethodSymbol();
TR::Register *returnReg;
// Allocate adequate register dependencies.
//
// pre = number of argument registers
// post = number of volatile + return register
//
uint32_t pre = getProperties().getNumIntegerArgumentRegisters() + getProperties().getNumFloatArgumentRegisters();
uint32_t post = getProperties().getNumVolatileRegisters() + (callNode->getDataType() == TR::NoType ? 0 : 1);
#if defined (PYTHON) && 0
// Treat all preserved GP regs as volatile until register map support available.
//
post += getProperties().getNumberOfPreservedGPRegisters();
#endif
TR::RegisterDependencyConditions *preDeps = generateRegisterDependencyConditions(pre, 0, cg());
TR::RegisterDependencyConditions *postDeps = generateRegisterDependencyConditions(0, post, cg());
// Evaluate outgoing arguments on the system stack and build pre-conditions.
//
int32_t memoryArgSize = buildArgs(callNode, preDeps);
// Build post-conditions.
//
returnReg = buildVolatileAndReturnDependencies(callNode, postDeps);
postDeps->stopAddingPostConditions();
// Find the second scratch register in the post dependency list.
//
TR::Register *scratchReg = NULL;
TR::RealRegister::RegNum scratchRegIndex = getProperties().getIntegerScratchRegister(1);
for (int32_t i=0; i<post; i++)
{
if (postDeps->getPostConditions()->getRegisterDependency(i)->getRealRegister() == scratchRegIndex)
{
scratchReg = postDeps->getPostConditions()->getRegisterDependency(i)->getRegister();
break;
}
}
#if defined(PYTHON) && 0
// For Python, store the instruction that contains the GC map at this site into
// the frame object.
//
TR::SymbolReference *frameObjectSymRef =
comp()->getSymRefTab()->findOrCreateAutoSymbol(comp()->getMethodSymbol(), 0, TR::Address, true, false, true);
TR::Register *frameObjectRegister = cg()->allocateRegister();
generateRegMemInstruction(
L8RegMem,
callNode,
frameObjectRegister,
generateX86MemoryReference(frameObjectSymRef, cg()),
cg());
TR::RealRegister *espReal = cg()->machine()->getX86RealRegister(TR::RealRegister::esp);
TR::Register *gcMapPCRegister = cg()->allocateRegister();
generateRegMemInstruction(
LEA8RegMem,
callNode,
gcMapPCRegister,
generateX86MemoryReference(espReal, -8, cg()),
cg());
// Use "volatile" registers across the call. Once proper register map support
// is implemented, r14 and r15 will no longer be volatile and a different pair
// should be chosen.
//
TR::RegisterDependencyConditions *gcMapDeps = generateRegisterDependencyConditions(0, 2, cg());
gcMapDeps->addPostCondition(frameObjectRegister, TR::RealRegister::r14, cg());
gcMapDeps->addPostCondition(gcMapPCRegister, TR::RealRegister::r15, cg());
gcMapDeps->stopAddingPostConditions();
generateMemRegInstruction(
S8MemReg,
callNode,
generateX86MemoryReference(frameObjectRegister, fe()->getPythonGCMapPCOffsetInFrame(), cg()),
gcMapPCRegister,
gcMapDeps,
cg());
cg()->stopUsingRegister(frameObjectRegister);
cg()->stopUsingRegister(gcMapPCRegister);
#endif
TR::Instruction *instr;
if (methodSymbol->getMethodAddress())
{
TR_ASSERT(scratchReg, "could not find second scratch register");
auto LoadRegisterInstruction = generateRegImm64SymInstruction(
MOV8RegImm64,
callNode,
scratchReg,
(uintptr_t)methodSymbol->getMethodAddress(),
methodSymRef,
cg());
if (TR::Options::getCmdLineOptions()->getOption(TR_EmitRelocatableELFFile))
{
LoadRegisterInstruction->setReloKind(TR_NativeMethodAbsolute);
}
instr = generateRegInstruction(CALLReg, callNode, scratchReg, preDeps, cg());
}
else
{
instr = generateImmSymInstruction(CALLImm4, callNode, (uintptrj_t)methodSymbol->getMethodAddress(), methodSymRef, preDeps, cg());
}
cg()->resetIsLeafMethod();
instr->setNeedsGCMap(getProperties().getPreservedRegisterMapForGC());
cg()->stopUsingRegister(scratchReg);
TR::LabelSymbol *postDepLabel = generateLabelSymbol(cg());
generateLabelInstruction(LABEL, callNode, postDepLabel, postDeps, cg());
return returnReg;
}
TR::Register *TR::AMD64SystemLinkage::buildDirectDispatch(
TR::Node *callNode,
bool spillFPRegs)
{
TR::SymbolReference *methodSymRef = callNode->getSymbolReference();
TR::MethodSymbol *methodSymbol = methodSymRef->getSymbol()->castToMethodSymbol();
TR::Register *returnReg;
// Allocate adequate register dependencies.
//
// pre = number of argument registers
// post = number of volatile + return register
//
uint32_t pre = getProperties().getNumIntegerArgumentRegisters() + getProperties().getNumFloatArgumentRegisters();
uint32_t post = getProperties().getNumVolatileRegisters() + (callNode->getDataType() == TR::NoType ? 0 : 1);
TR::RegisterDependencyConditions *preDeps = generateRegisterDependencyConditions(pre, 0, cg());
TR::RegisterDependencyConditions *postDeps = generateRegisterDependencyConditions(0, post, cg());
// Evaluate outgoing arguments on the system stack and build pre-conditions.
//
int32_t memoryArgSize = buildArgs(callNode, preDeps);
// Build post-conditions.
//
returnReg = buildVolatileAndReturnDependencies(callNode, postDeps);
postDeps->stopAddingPostConditions();
// Find the second scratch register in the post dependency list.
//
TR::Register *scratchReg = NULL;
TR::RealRegister::RegNum scratchRegIndex = getProperties().getIntegerScratchRegister(1);
for (int32_t i=0; i<post; i++)
{
if (postDeps->getPostConditions()->getRegisterDependency(i)->getRealRegister() == scratchRegIndex)
{
scratchReg = postDeps->getPostConditions()->getRegisterDependency(i)->getRegister();
break;
}
}
TR::Instruction *instr;
if (methodSymbol->getMethodAddress())
{
TR_ASSERT(scratchReg, "could not find second scratch register");
auto LoadRegisterInstruction = generateRegImm64SymInstruction(
MOV8RegImm64,
callNode,
scratchReg,
(uintptr_t)methodSymbol->getMethodAddress(),
methodSymRef,
cg());
if (comp()->getOption(TR_EmitRelocatableELFFile))
{
LoadRegisterInstruction->setReloKind(TR_NativeMethodAbsolute);
}
instr = generateRegInstruction(CALLReg, callNode, scratchReg, preDeps, cg());
}
else
{
instr = generateImmSymInstruction(CALLImm4, callNode, (uintptrj_t)methodSymbol->getMethodAddress(), methodSymRef, preDeps, cg());
}
cg()->resetIsLeafMethod();
instr->setNeedsGCMap(getProperties().getPreservedRegisterMapForGC());
cg()->stopUsingRegister(scratchReg);
TR::LabelSymbol *postDepLabel = generateLabelSymbol(cg());
generateLabelInstruction(LABEL, callNode, postDepLabel, postDeps, cg());
return returnReg;
}
void TR_OutlinedInstructions::assignRegisters(TR_RegisterKinds kindsToBeAssigned, TR::X86VFPSaveInstruction *vfpSaveInstruction)
{
if (hasBeenRegisterAssigned())
return;
// nested internal control flow assert:
_cg->setInternalControlFlowSafeNestingDepth(_cg->internalControlFlowNestingDepth());
// Create a dependency list on the first instruction in this stream that captures all
// current real register associations. This is necessary to get the register assigner
// back into its original state before the helper stream was processed.
//
TR::RegisterDependencyConditions *liveRealRegDeps = _cg->machine()->createDepCondForLiveGPRs();
_firstInstruction->setDependencyConditions(liveRealRegDeps);
#if 0
// If the outlined section jumps back to a section that's expecting a certain register
// state then add register dependencies on the exit branch to set that state.
//
if (_postDependencyMergeList)
{
TR::RegisterDependencyConditions *mergeDeps = _postDependencyMergeList->clone(_cg);
TR_ASSERT(_appendInstruction->getDependencyConditions() == NULL, "unexpected reg deps on OOL append instruction");
_appendInstruction->setDependencyConditions(mergeDeps);
TR_X86RegisterDependencyGroup *depGroup = mergeDeps->getPostConditions();
for (int32_t i=0; i<mergeDeps->getNumPostConditions(); i++)
{
TR::RegisterDependency *dependency = depGroup->getRegisterDependency(i);
TR::Register *virtReg = dependency->getRegister();
virtReg->incTotalUseCount();
virtReg->incFutureUseCount();
#ifdef DEBUG
// Ensure all register dependencies have been assigned.
//
TR_ASSERT(dependency->getRealRegister() != TR::RealRegister::NoReg, "unassigned merge dep register");
TR_ASSERT(virtReg->getAssignedRealRegister() == _cg->machine()->getX86RealRegister(dependency->getRealRegister()), "unexpected(?) register assignment");
#endif
}
}
#endif
// TODO:AMD64: Fix excessive register assignment exchanges in outlined instruction dispatch.
// Ensure correct VFP state at the start of the outlined instruction sequence.
//
generateVFPRestoreInstruction(comp()->getAppendInstruction(), vfpSaveInstruction, _cg);
// Link in the helper stream into the mainline code.
//
TR::Instruction *appendInstruction = comp()->getAppendInstruction();
appendInstruction->setNext(_firstInstruction);
_firstInstruction->setPrev(appendInstruction);
comp()->setAppendInstruction(_appendInstruction);
// Register assign the helper dispatch instructions.
//
_cg->doBackwardsRegisterAssignment(kindsToBeAssigned, _appendInstruction, appendInstruction);
// Returning to mainline, reset this counter
_cg->setInternalControlFlowSafeNestingDepth(0);
setHasBeenRegisterAssigned(true);
}
