TR::Register *IA32LinkageUtils::pushLongArg(
TR::Node *child,
TR::CodeGenerator *cg)
{
TR::Register *pushRegister;
if (child->getRegister() == NULL)
{
if (child->getOpCode().isLoadConst())
{
TR_X86OpCodes pushOp;
int32_t highValue = child->getLongIntHigh();
if (highValue >= -128 && highValue <= 127)
{
pushOp = PUSHImms;
}
else
{
pushOp = PUSHImm4;
}
generateImmInstruction(pushOp, child, highValue, cg);
int32_t lowValue = child->getLongIntLow();
if (lowValue >= -128 && lowValue <= 127)
{
pushOp = PUSHImms;
}
else
{
pushOp = PUSHImm4;
}
generateImmInstruction(pushOp, child, lowValue, cg);
cg->decReferenceCount(child);
return NULL;
}
else if (child->getOpCodeValue() == TR::dbits2l &&
!child->normalizeNanValues() &&
child->getReferenceCount() == 1)
{
pushRegister = pushDoubleArg(child->getFirstChild(), cg);
cg->decReferenceCount(child);
return pushRegister;
}
else if (child->getOpCode().isMemoryReference() &&
child->getReferenceCount() == 1)
{
TR::MemoryReference *lowMR = generateX86MemoryReference(child, cg);
generateMemInstruction(PUSHMem, child, generateX86MemoryReference(*lowMR,4, cg), cg);
generateMemInstruction(PUSHMem, child, lowMR, cg);
lowMR->decNodeReferenceCounts(cg);
return NULL;
}
}
pushRegister = cg->evaluate(child);
generateRegInstruction(PUSHReg, child, pushRegister->getHighOrder(), cg);
generateRegInstruction(PUSHReg, child, pushRegister->getLowOrder(), cg);
cg->decReferenceCount(child);
return pushRegister;
}
rcount_t
OMR::CodeGenerator::incReferenceCount(TR::Node *node)
{
TR::Register *reg = node->getRegister();
#ifdef J9_PROJECT_SPECIFIC
#if defined(TR_TARGET_S390)
if (reg && reg->getOpaquePseudoRegister())
{
TR_OpaquePseudoRegister * pseudoReg = reg->getOpaquePseudoRegister();
TR_StorageReference * pseudoRegStorageReference = pseudoReg->getStorageReference();
TR_ASSERT(pseudoRegStorageReference, "the pseudoReg should have a non-null storage reference\n");
pseudoRegStorageReference->incrementTemporaryReferenceCount();
}
#endif
#endif
rcount_t count = node->incReferenceCount();
if (self()->comp()->getOptions()->getTraceCGOption(TR_TraceCGEvaluation))
{
self()->getDebug()->printNodeEvaluation(node, "++ ", reg);
}
return count;
}
void TR::ARM64MemSrc1Instruction::assignRegisters(TR_RegisterKinds kindToBeAssigned)
{
TR::Machine *machine = cg()->machine();
TR::MemoryReference *mref = getMemoryReference();
TR::Register *sourceVirtual = getSource1Register();
if (getDependencyConditions())
getDependencyConditions()->assignPostConditionRegisters(this, kindToBeAssigned, cg());
sourceVirtual->block();
mref->assignRegisters(this, cg());
sourceVirtual->unblock();
mref->blockRegisters();
TR::RealRegister *assignedRegister = sourceVirtual->getAssignedRealRegister();
if (assignedRegister == NULL)
{
assignedRegister = machine->assignOneRegister(this, sourceVirtual);
}
mref->unblockRegisters();
setSource1Register(assignedRegister);
if (getDependencyConditions())
getDependencyConditions()->assignPreConditionRegisters(this->getPrev(), kindToBeAssigned, cg());
}
void decFutureUseCounts(uint32_t numberOfRegisters,
TR::CodeGenerator *cg)
{
for (uint32_t i = 0; i< numberOfRegisters; i++)
{
TR::Register *virtReg = _dependencies[i].getRegister();
virtReg->decFutureUseCount();
}
}
// also handles ilload
TR::Register *OMR::X86::AMD64::TreeEvaluator::lloadEvaluator(TR::Node *node, TR::CodeGenerator *cg)
{
TR::MemoryReference *sourceMR = generateX86MemoryReference(node, cg);
TR::Register *reg = TR::TreeEvaluator::loadMemory(node, sourceMR, TR_RematerializableLong, node->getOpCode().isIndirect(), cg);
reg->setMemRef(sourceMR);
node->setRegister(reg);
sourceMR->decNodeReferenceCounts(cg);
return reg;
}
TR::Register *IA32LinkageUtils::pushFloatArg(
TR::Node *child,
TR::CodeGenerator *cg)
{
TR::Register *pushRegister;
if (child->getRegister() == NULL)
{
if (child->getOpCodeValue() == TR::fconst)
{
int32_t value = child->getFloatBits();
TR_X86OpCodes pushOp;
if (value >= -128 && value <= 127)
{
pushOp = PUSHImms;
}
else
{
pushOp = PUSHImm4;
}
generateImmInstruction(pushOp, child, value, cg);
cg->decReferenceCount(child);
return NULL;
}
else if (child->getReferenceCount() == 1)
{
if (child->getOpCode().isLoad())
{
TR::MemoryReference *tempMR = generateX86MemoryReference(child, cg);
generateMemInstruction(PUSHMem, child, tempMR, cg);
tempMR->decNodeReferenceCounts(cg);
cg->decReferenceCount(child);
return NULL;
}
else if (child->getOpCodeValue() == TR::ibits2f)
{
pushRegister = pushIntegerWordArg(child->getFirstChild(), cg);
cg->decReferenceCount(child);
return pushRegister;
}
}
}
pushRegister = cg->evaluate(child);
TR::RealRegister *espReal = cg->machine()->getRealRegister(TR::RealRegister::esp);
generateRegImmInstruction(SUB4RegImms, child, espReal, 4, cg);
if (cg->useSSEForSinglePrecision() && pushRegister->getKind() == TR_FPR)
generateMemRegInstruction(MOVSSMemReg, child, generateX86MemoryReference(espReal, 0, cg), pushRegister, cg);
else
generateFPMemRegInstruction(FSTMemReg, child, generateX86MemoryReference(espReal, 0, cg), pushRegister, cg);
cg->decReferenceCount(child);
return pushRegister;
}
TR::Register* OMR::X86::TreeEvaluator::SIMDsplatsEvaluator(TR::Node* node, TR::CodeGenerator* cg)
{
TR::Node* childNode = node->getChild(0);
TR::Register* childReg = cg->evaluate(childNode);
TR::Register* resultReg = cg->allocateRegister(TR_VRF);
switch (node->getDataType())
{
case TR::VectorInt32:
generateRegRegInstruction(MOVDRegReg4, node, resultReg, childReg, cg);
generateRegRegImmInstruction(PSHUFDRegRegImm1, node, resultReg, resultReg, 0x00, cg); // 00 00 00 00 shuffle xxxA to AAAA
break;
case TR::VectorInt64:
if (TR::Compiler->target.is32Bit())
{
TR::Register* tempVectorReg = cg->allocateRegister(TR_VRF);
generateRegRegInstruction(MOVDRegReg4, node, tempVectorReg, childReg->getHighOrder(), cg);
generateRegImmInstruction(PSLLQRegImm1, node, tempVectorReg, 0x20, cg);
generateRegRegInstruction(MOVDRegReg4, node, resultReg, childReg->getLowOrder(), cg);
generateRegRegInstruction(PORRegReg, node, resultReg, tempVectorReg, cg);
cg->stopUsingRegister(tempVectorReg);
}
else
{
generateRegRegInstruction(MOVQRegReg8, node, resultReg, childReg, cg);
}
generateRegRegImmInstruction(PSHUFDRegRegImm1, node, resultReg, resultReg, 0x44, cg); // 01 00 01 00 shuffle xxBA to BABA
break;
case TR::VectorFloat:
generateRegRegImmInstruction(PSHUFDRegRegImm1, node, resultReg, childReg, 0x00, cg); // 00 00 00 00 shuffle xxxA to AAAA
break;
case TR::VectorDouble:
generateRegRegImmInstruction(PSHUFDRegRegImm1, node, resultReg, childReg, 0x44, cg); // 01 00 01 00 shuffle xxBA to BABA
break;
default:
if (cg->comp()->getOption(TR_TraceCG))
traceMsg(cg->comp(), "Unsupported data type, Node = %p\n", node);
TR_ASSERT(false, "Unsupported data type");
break;
}
node->setRegister(resultReg);
cg->decReferenceCount(childNode);
return resultReg;
}
// Create a NoReg dependency for each child of a call that has been evaluated into a register.
// Ignore children that do not have a register since their live range should not persist outside of
// the helper call stream.
//
TR::RegisterDependencyConditions *TR_OutlinedInstructions::formEvaluatedArgumentDepList()
{
int32_t i, c=0;
for (i=_callNode->getFirstArgumentIndex(); i<_callNode->getNumChildren(); i++)
{
TR::Register *reg = _callNode->getChild(i)->getRegister();
if (reg)
{
TR::RegisterPair *regPair = reg->getRegisterPair();
c += regPair? 2 : 1;
}
}
TR::RegisterDependencyConditions *depConds = NULL;
if (c)
{
TR::Machine *machine = _cg->machine();
depConds = generateRegisterDependencyConditions(0, c, _cg);
for (i=_callNode->getFirstArgumentIndex(); i<_callNode->getNumChildren(); i++)
{
TR::Register *reg = _callNode->getChild(i)->getRegister();
if (reg)
{
TR::RegisterPair *regPair = reg->getRegisterPair();
if (regPair)
{
depConds->addPostCondition(regPair->getLowOrder(), TR::RealRegister::NoReg, _cg);
depConds->addPostCondition(regPair->getHighOrder(), TR::RealRegister::NoReg, _cg);
}
else
{
depConds->addPostCondition(reg, TR::RealRegister::NoReg, _cg);
}
}
}
depConds->stopAddingConditions();
}
return depConds;
}
TR::Register *
OMR::X86::I386::CodeGenerator::longClobberEvaluate(TR::Node *node)
{
TR_ASSERT(node->getOpCode().is8Byte(), "assertion failure");
if (node->getReferenceCount() > 1)
{
TR::Register *temp = self()->evaluate(node);
TR::Register *lowReg = self()->allocateRegister();
TR::Register *highReg = self()->allocateRegister();
TR::RegisterPair *longReg = self()->allocateRegisterPair(lowReg, highReg);
generateRegRegInstruction(MOV4RegReg, node, lowReg, temp->getLowOrder(), self());
generateRegRegInstruction(MOV4RegReg, node, highReg, temp->getHighOrder(), self());
return longReg;
}
else
{
return self()->evaluate(node);
}
}
void TR::ARM64Trg1MemInstruction::assignRegisters(TR_RegisterKinds kindToBeAssigned)
{
TR::Machine *machine = cg()->machine();
TR::MemoryReference *mref = getMemoryReference();
TR::Register *targetVirtual = getTargetRegister();
if (getDependencyConditions())
getDependencyConditions()->assignPostConditionRegisters(this, kindToBeAssigned, cg());
mref->blockRegisters();
setTargetRegister(machine->assignOneRegister(this, targetVirtual));
mref->unblockRegisters();
targetVirtual->block();
mref->assignRegisters(this, cg());
targetVirtual->unblock();
if (getDependencyConditions())
getDependencyConditions()->assignPreConditionRegisters(this->getPrev(), kindToBeAssigned, cg());
}
void OMR::X86::Instruction::clobberRegsForRematerialisation()
{
// We assume most instructions modify all registers that appear in their
// postconditions, with a few exceptions.
//
if ( self()->cg()->enableRematerialisation()
&& self()->getDependencyConditions()
&& (self()->getOpCodeValue() != ASSOCREGS) // reg associations aren't really instructions, so they don't modify anything
&& (self()->getOpCodeValue() != LABEL) // labels must already be handled properly for a variety of reasons
&& (!self()->getOpCode().isShiftOp())
&& (!self()->getOpCode().isRotateOp()) // shifts and rotates often have a postcondition on ecx but don't clobber it
){
// Check the live discardable register list to see if this is the first
// instruction that kills the rematerialisable range of a register.
//
TR::ClobberingInstruction *clob = NULL;
TR_X86RegisterDependencyGroup *post = self()->getDependencyConditions()->getPostConditions();
for (uint32_t i = 0; i < self()->getDependencyConditions()->getNumPostConditions(); i++)
{
TR::Register *reg = post->getRegisterDependency(i)->getRegister();
if (reg->isDiscardable())
{
if (!clob)
{
clob = new (self()->cg()->trHeapMemory()) TR::ClobberingInstruction(self(), self()->cg()->trMemory());
self()->cg()->addClobberingInstruction(clob);
}
clob->addClobberedRegister(reg);
self()->cg()->removeLiveDiscardableRegister(reg);
self()->cg()->clobberLiveDependentDiscardableRegisters(clob, reg);
if (debug("dumpRemat"))
diagnostic("---> Clobbering %s discardable postcondition register %s at instruction %s\n",
self()->cg()->getDebug()->toString(reg->getRematerializationInfo()),
self()->cg()->getDebug()->getName(reg),
self()->cg()->getDebug()->getName(self()));
}
}
}
}
rcount_t
OMR::CodeGenerator::decReferenceCount(TR::Node * node)
{
TR::Register *reg = node->getRegister();
// restricted registers go dead when ref count==2 because
// their ref count was inced in prepareNodeForInstructionSelection
if ((node->getReferenceCount() == 1) &&
reg && self()->getLiveRegisters(reg->getKind()))
{
TR_ASSERT(reg->isLive() ||
(diagnostic("\n*** Error: Register %s for node "
"[%s] died prematurely\n",
reg->getRegisterName(self()->comp()),
node->getName(self()->comp()->getDebug())),
0),
"Node %s register should be live",self()->getDebug()->getName(node));
TR_LiveRegisterInfo *liveRegister = reg->getLiveRegisterInfo();
TR::Register *pair = reg->getRegisterPair();
if (pair)
{
pair->getHighOrder()->getLiveRegisterInfo()->decNodeCount();
pair->getLowOrder()->getLiveRegisterInfo()->decNodeCount();
}
if (liveRegister && liveRegister->decNodeCount() == 0)
{
// The register is now dead
//
self()->getLiveRegisters(reg->getKind())->registerIsDead(reg);
}
}
#ifdef J9_PROJECT_SPECIFIC
#if defined(TR_TARGET_S390)
if (reg && reg->getOpaquePseudoRegister())
{
TR_OpaquePseudoRegister *pseudoReg = reg->getOpaquePseudoRegister();
TR_StorageReference *storageReference = pseudoReg->getStorageReference();
TR_ASSERT(storageReference,"the pseudoReg should have a non-null storage reference\n");
storageReference->decrementTemporaryReferenceCount();
if (node->getReferenceCount() == 1)
{
storageReference->decOwningRegisterCount();
if (self()->traceBCDCodeGen())
traceMsg(self()->comp(),"\tdecrement owningRegisterCount %d->%d on ref #%d (%s) for reg %s as %s (%p) refCount == 1 (going to 0)\n",
storageReference->getOwningRegisterCount()+1,
storageReference->getOwningRegisterCount(),
storageReference->getReferenceNumber(),
self()->getDebug()->getName(storageReference->getSymbol()),
self()->getDebug()->getName(reg),
node->getOpCode().getName(),
node);
}
}
else if (node->getOpCode().hasSymbolReference() && node->getSymbolReference() && node->getSymbolReference()->isTempVariableSizeSymRef())
{
TR_ASSERT(false,"tempMemSlots should only be attached to pseudoRegisters and not node %p\n",node);
}
#endif
#endif
rcount_t count = node->decReferenceCount();
if (self()->comp()->getOptions()->getTraceCGOption(TR_TraceCGEvaluation))
{
self()->getDebug()->printNodeEvaluation(node, "-- ", reg);
}
return count;
}
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);
}
// Build arguments for system linkage dispatch.
//
int32_t TR::AMD64SystemLinkage::buildArgs(
TR::Node *callNode,
TR::RegisterDependencyConditions *deps)
{
TR::SymbolReference *methodSymRef = callNode->getSymbolReference();
TR::MethodSymbol *methodSymbol = methodSymRef->getSymbol()->castToMethodSymbol();
TR::RealRegister::RegNum noReg = TR::RealRegister::NoReg;
TR::RealRegister *espReal = machine()->getX86RealRegister(TR::RealRegister::esp);
int32_t firstNodeArgument = callNode->getFirstArgumentIndex();
int32_t lastNodeArgument = callNode->getNumChildren() - 1;
int32_t offset = 0;
int32_t sizeOfOutGoingArgs= 0;
uint16_t numIntArgs = 0,
numFloatArgs = 0;
int32_t first, last, direction;
int32_t numCopiedRegs = 0;
TR::Register *copiedRegs[TR::X86LinkageProperties::MaxArgumentRegisters];
if (getProperties().passArgsRightToLeft())
{
first = lastNodeArgument;
last = firstNodeArgument - 1;
direction = -1;
}
else
{
first = firstNodeArgument;
last = lastNodeArgument + 1;
direction = 1;
}
// If the dispatch is indirect we must add the VFT register to the preconditions
// so that it gets register assigned with the other preconditions to the call.
//
if (callNode->getOpCode().isIndirect())
{
TR::Node *vftChild = callNode->getFirstChild();
TR_ASSERT(vftChild->getRegister(), "expecting VFT child to be evaluated");
TR::RealRegister::RegNum scratchRegIndex = getProperties().getIntegerScratchRegister(1);
deps->addPreCondition(vftChild->getRegister(), scratchRegIndex, cg());
}
int32_t i;
for (i = first; i != last; i += direction)
{
TR::parmLayoutResult layoutResult;
TR::RealRegister::RegNum rregIndex = noReg;
TR::Node *child = callNode->getChild(i);
layoutParm(child, sizeOfOutGoingArgs, numIntArgs, numFloatArgs, layoutResult);
if (layoutResult.abstract & TR::parmLayoutResult::IN_LINKAGE_REG_PAIR)
{
// TODO: AMD64 SysV ABI might put a struct into a pair of linkage registerr
TR_ASSERT(false, "haven't support linkage_reg_pair yet.\n");
}
else if (layoutResult.abstract & TR::parmLayoutResult::IN_LINKAGE_REG)
{
TR_RegisterKinds regKind = layoutResult.regs[0].regKind;
uint32_t regIndex = layoutResult.regs[0].regIndex;
TR_ASSERT(regKind == TR_GPR || regKind == TR_FPR, "linkage registers includes TR_GPR and TR_FPR\n");
rregIndex = (regKind == TR_FPR) ? getProperties().getFloatArgumentRegister(regIndex): getProperties().getIntegerArgumentRegister(regIndex);
}
else
{
offset = layoutResult.offset;
}
TR::Register *vreg;
vreg = cg()->evaluate(child);
bool needsStackOffsetUpdate = false;
if (rregIndex != noReg)
{
// For NULL JNI reference parameters, it is possible that the NULL value will be evaluated into
// a different register than the child. In that case it is not necessary to copy the temporary scratch
// register across the call.
//
if ((child->getReferenceCount() > 1) &&
(vreg == child->getRegister()))
{
TR::Register *argReg = cg()->allocateRegister();
if (vreg->containsCollectedReference())
argReg->setContainsCollectedReference();
generateRegRegInstruction(TR::Linkage::movOpcodes(RegReg, movType(child->getDataType())), child, argReg, vreg, cg());
vreg = argReg;
copiedRegs[numCopiedRegs++] = vreg;
}
deps->addPreCondition(vreg, rregIndex, cg());
}
else
{
// Ideally, we would like to push rather than move
generateMemRegInstruction(TR::Linkage::movOpcodes(MemReg, fullRegisterMovType(vreg)),
child,
generateX86MemoryReference(espReal, offset, cg()),
vreg,
cg());
}
cg()->decReferenceCount(child);
}
// Now that we're finished making the preconditions, all the interferences
// are established and we can kill these regs.
//
for (i = 0; i < numCopiedRegs; i++)
cg()->stopUsingRegister(copiedRegs[i]);
deps->stopAddingPreConditions();
return sizeOfOutGoingArgs;
}
TR::Register *
TR::AMD64SystemLinkage::buildVolatileAndReturnDependencies(
TR::Node *callNode,
TR::RegisterDependencyConditions *deps)
{
if (callNode->getOpCode().isIndirect())
{
TR::Node *vftChild = callNode->getFirstChild();
if (vftChild->getRegister() && (vftChild->getReferenceCount() > 1))
{
}
else
{
// VFT child dies here; decrement it early so it doesn't interfere with dummy regs.
cg()->recursivelyDecReferenceCount(vftChild);
}
}
TR_ASSERT(deps != NULL, "expected register dependencies");
// Figure out which is the return register.
//
TR::RealRegister::RegNum returnRegIndex;
TR_RegisterKinds returnKind;
switch (callNode->getDataType())
{
case TR::NoType:
returnRegIndex = TR::RealRegister::NoReg;
returnKind = TR_NoRegister;
break;
case TR::Int8:
case TR::Int16:
case TR::Int32:
case TR::Int64:
case TR::Address:
returnRegIndex = getProperties().getIntegerReturnRegister();
returnKind = TR_GPR;
break;
case TR::Float:
case TR::Double:
returnRegIndex = getProperties().getFloatReturnRegister();
returnKind = TR_FPR;
break;
case TR::Aggregate:
default:
TR_ASSERT(false, "Unrecognized call node data type: #%d", (int)callNode->getDataType());
break;
}
// Kill all non-preserved int and float regs besides the return register.
//
int32_t i;
TR::RealRegister::RegNum scratchIndex = getProperties().getIntegerScratchRegister(1);
for (i=0; i<getProperties().getNumVolatileRegisters(); i++)
{
TR::RealRegister::RegNum regIndex = getProperties()._volatileRegisters[i];
if (regIndex != returnRegIndex)
{
TR_RegisterKinds rk = (i < getProperties()._numberOfVolatileGPRegisters) ? TR_GPR : TR_FPR;
TR::Register *dummy = cg()->allocateRegister(rk);
deps->addPostCondition(dummy, regIndex, cg());
// Note that we don't setPlaceholderReg here. If this volatile reg is also volatile
// in the caller's linkage, then that flag doesn't matter much anyway. If it's preserved
// in the caller's linkage, then we don't want to set that flag because we want this
// use of the register to count as a "real" use, thereby motivating the prologue to
// preserve the register.
// A scratch register is necessary to call the native without a trampoline.
//
if (callNode->getOpCode().isIndirect() || (regIndex != scratchIndex))
cg()->stopUsingRegister(dummy);
}
}
#if defined (PYTHON) && 0
// Evict the preserved registers across the call
//
for (i=0; i<getProperties().getNumberOfPreservedGPRegisters(); i++)
{
TR::RealRegister::RegNum regIndex = getProperties()._preservedRegisters[i];
TR::Register *dummy = cg()->allocateRegister(TR_GPR);
deps->addPostCondition(dummy, regIndex, cg());
// Note that we don't setPlaceholderReg here. If this volatile reg is also volatile
// in the caller's linkage, then that flag doesn't matter much anyway. If it's preserved
// in the caller's linkage, then we don't want to set that flag because we want this
// use of the register to count as a "real" use, thereby motivating the prologue to
// preserve the register.
// A scratch register is necessary to call the native without a trampoline.
//
if (callNode->getOpCode().isIndirect() || (regIndex != scratchIndex))
cg()->stopUsingRegister(dummy);
}
#endif
if (callNode->getOpCode().isIndirect())
{
TR::Node *vftChild = callNode->getFirstChild();
if (vftChild->getRegister() && (vftChild->getReferenceCount() > 1))
{
// VFT child survives the call, so we must include it in the postconditions.
deps->addPostCondition(vftChild->getRegister(), TR::RealRegister::NoReg, cg());
cg()->recursivelyDecReferenceCount(vftChild);
}
}
// Now that everything is dead, we can allocate the return register without
// interference
//
TR::Register *returnRegister;
if (returnRegIndex)
{
TR_ASSERT(returnKind != TR_NoRegister, "assertion failure");
if (callNode->getDataType() == TR::Address)
returnRegister = cg()->allocateCollectedReferenceRegister();
else
{
returnRegister = cg()->allocateRegister(returnKind);
if (callNode->getDataType() == TR::Float)
returnRegister->setIsSinglePrecision();
}
deps->addPostCondition(returnRegister, returnRegIndex, cg());
}
else
returnRegister = NULL;
// The reg dependency is left open intentionally, and need to be closed by
// the caller. The reason is because, child class might call this method, while
// adding more register dependecies; if we close the reg dependency here,
// the child class won't be able to add more register dependencies.
return returnRegister;
}
int32_t TR::ARM64SystemLinkage::buildArgs(TR::Node *callNode,
TR::RegisterDependencyConditions *dependencies)
{
const TR::ARM64LinkageProperties &properties = getProperties();
TR::ARM64MemoryArgument *pushToMemory = NULL;
TR::Register *argMemReg;
TR::Register *tempReg;
int32_t argIndex = 0;
int32_t numMemArgs = 0;
int32_t argSize = 0;
int32_t numIntegerArgs = 0;
int32_t numFloatArgs = 0;
int32_t totalSize;
int32_t i;
TR::Node *child;
TR::DataType childType;
TR::DataType resType = callNode->getType();
uint32_t firstArgumentChild = callNode->getFirstArgumentIndex();
/* Step 1 - figure out how many arguments are going to be spilled to memory i.e. not in registers */
for (i = firstArgumentChild; i < callNode->getNumChildren(); i++)
{
child = callNode->getChild(i);
childType = child->getDataType();
switch (childType)
{
case TR::Int8:
case TR::Int16:
case TR::Int32:
case TR::Int64:
case TR::Address:
if (numIntegerArgs >= properties.getNumIntArgRegs())
numMemArgs++;
numIntegerArgs++;
break;
case TR::Float:
case TR::Double:
if (numFloatArgs >= properties.getNumFloatArgRegs())
numMemArgs++;
numFloatArgs++;
break;
default:
TR_ASSERT(false, "Argument type %s is not supported\n", childType.toString());
}
}
// From here, down, any new stack allocations will expire / die when the function returns
TR::StackMemoryRegion stackMemoryRegion(*trMemory());
/* End result of Step 1 - determined number of memory arguments! */
if (numMemArgs > 0)
{
pushToMemory = new (trStackMemory()) TR::ARM64MemoryArgument[numMemArgs];
argMemReg = cg()->allocateRegister();
}
totalSize = numMemArgs * 8;
// align to 16-byte boundary
totalSize = (totalSize + 15) & (~15);
numIntegerArgs = 0;
numFloatArgs = 0;
for (i = firstArgumentChild; i < callNode->getNumChildren(); i++)
{
TR::MemoryReference *mref = NULL;
TR::Register *argRegister;
TR::InstOpCode::Mnemonic op;
child = callNode->getChild(i);
childType = child->getDataType();
switch (childType)
{
case TR::Int8:
case TR::Int16:
case TR::Int32:
case TR::Int64:
case TR::Address:
if (childType == TR::Address)
argRegister = pushAddressArg(child);
else if (childType == TR::Int64)
argRegister = pushLongArg(child);
else
argRegister = pushIntegerWordArg(child);
if (numIntegerArgs < properties.getNumIntArgRegs())
{
if (!cg()->canClobberNodesRegister(child, 0))
{
if (argRegister->containsCollectedReference())
tempReg = cg()->allocateCollectedReferenceRegister();
else
tempReg = cg()->allocateRegister();
generateMovInstruction(cg(), callNode, tempReg, argRegister);
argRegister = tempReg;
}
if (numIntegerArgs == 0 &&
(resType.isAddress() || resType.isInt32() || resType.isInt64()))
{
TR::Register *resultReg;
if (resType.isAddress())
resultReg = cg()->allocateCollectedReferenceRegister();
else
resultReg = cg()->allocateRegister();
dependencies->addPreCondition(argRegister, TR::RealRegister::x0);
dependencies->addPostCondition(resultReg, TR::RealRegister::x0);
}
else
{
addDependency(dependencies, argRegister, properties.getIntegerArgumentRegister(numIntegerArgs), TR_GPR, cg());
}
}
else
{
// numIntegerArgs >= properties.getNumIntArgRegs()
if (childType == TR::Address || childType == TR::Int64)
{
op = TR::InstOpCode::strpostx;
}
else
{
op = TR::InstOpCode::strpostw;
}
mref = getOutgoingArgumentMemRef(argMemReg, argRegister, op, pushToMemory[argIndex++]);
argSize += 8; // always 8-byte aligned
}
numIntegerArgs++;
break;
case TR::Float:
case TR::Double:
if (childType == TR::Float)
argRegister = pushFloatArg(child);
else
argRegister = pushDoubleArg(child);
if (numFloatArgs < properties.getNumFloatArgRegs())
{
if (!cg()->canClobberNodesRegister(child, 0))
{
tempReg = cg()->allocateRegister(TR_FPR);
op = (childType == TR::Float) ? TR::InstOpCode::fmovs : TR::InstOpCode::fmovd;
generateTrg1Src1Instruction(cg(), op, callNode, tempReg, argRegister);
argRegister = tempReg;
}
if ((numFloatArgs == 0 && resType.isFloatingPoint()))
{
TR::Register *resultReg;
if (resType.getDataType() == TR::Float)
resultReg = cg()->allocateSinglePrecisionRegister();
else
resultReg = cg()->allocateRegister(TR_FPR);
dependencies->addPreCondition(argRegister, TR::RealRegister::v0);
dependencies->addPostCondition(resultReg, TR::RealRegister::v0);
}
else
{
addDependency(dependencies, argRegister, properties.getFloatArgumentRegister(numFloatArgs), TR_FPR, cg());
}
}
else
{
// numFloatArgs >= properties.getNumFloatArgRegs()
if (childType == TR::Double)
{
op = TR::InstOpCode::vstrpostd;
}
else
{
op = TR::InstOpCode::vstrposts;
}
mref = getOutgoingArgumentMemRef(argMemReg, argRegister, op, pushToMemory[argIndex++]);
argSize += 8; // always 8-byte aligned
}
numFloatArgs++;
break;
} // end of switch
} // end of for
// NULL deps for non-preserved and non-system regs
while (numIntegerArgs < properties.getNumIntArgRegs())
{
if (numIntegerArgs == 0 && resType.isAddress())
{
dependencies->addPreCondition(cg()->allocateRegister(), properties.getIntegerArgumentRegister(0));
dependencies->addPostCondition(cg()->allocateCollectedReferenceRegister(), properties.getIntegerArgumentRegister(0));
}
else
{
addDependency(dependencies, NULL, properties.getIntegerArgumentRegister(numIntegerArgs), TR_GPR, cg());
}
numIntegerArgs++;
}
int32_t floatRegsUsed = (numFloatArgs > properties.getNumFloatArgRegs()) ? properties.getNumFloatArgRegs() : numFloatArgs;
for (i = (TR::RealRegister::RegNum)((uint32_t)TR::RealRegister::v0 + floatRegsUsed); i <= TR::RealRegister::LastFPR; i++)
{
if (!properties.getPreserved((TR::RealRegister::RegNum)i))
{
// NULL dependency for non-preserved regs
addDependency(dependencies, NULL, (TR::RealRegister::RegNum)i, TR_FPR, cg());
}
}
if (numMemArgs > 0)
{
TR::RealRegister *sp = cg()->machine()->getRealRegister(properties.getStackPointerRegister());
generateTrg1Src1ImmInstruction(cg(), TR::InstOpCode::subimmx, callNode, argMemReg, sp, totalSize);
for (argIndex = 0; argIndex < numMemArgs; argIndex++)
{
TR::Register *aReg = pushToMemory[argIndex].argRegister;
generateMemSrc1Instruction(cg(), pushToMemory[argIndex].opCode, callNode, pushToMemory[argIndex].argMemory, aReg);
cg()->stopUsingRegister(aReg);
}
cg()->stopUsingRegister(argMemReg);
}
return totalSize;
}
void TR_PPCRegisterDependencyGroup::assignRegisters(TR::Instruction *currentInstruction,
TR_RegisterKinds kindToBeAssigned,
uint32_t numberOfRegisters,
TR::CodeGenerator *cg)
{
// *this swipeable for debugging purposes
TR::Machine *machine = cg->machine();
TR::Register *virtReg;
TR::RealRegister::RegNum dependentRegNum;
TR::RealRegister *dependentRealReg, *assignedRegister, *realReg;
int i, j;
TR::Compilation *comp = cg->comp();
int num_gprs = 0;
int num_fprs = 0;
int num_vrfs = 0;
// Use to do lookups using real register numbers
TR_PPCRegisterDependencyMap map(_dependencies, numberOfRegisters);
if (!comp->getOption(TR_DisableOOL))
{
for (i = 0; i< numberOfRegisters; i++)
{
virtReg = _dependencies[i].getRegister();
dependentRegNum = _dependencies[i].getRealRegister();
if (dependentRegNum == TR::RealRegister::SpilledReg)
{
TR_ASSERT(virtReg->getBackingStorage(),"should have a backing store if dependentRegNum == spillRegIndex()\n");
if (virtReg->getAssignedRealRegister())
{
// this happens when the register was first spilled in main line path then was reverse spilled
// and assigned to a real register in OOL path. We protected the backing store when doing
// the reverse spill so we could re-spill to the same slot now
traceMsg (comp,"\nOOL: Found register spilled in main line and re-assigned inside OOL");
TR::Node *currentNode = currentInstruction->getNode();
TR::RealRegister *assignedReg = toRealRegister(virtReg->getAssignedRegister());
TR::MemoryReference *tempMR = new (cg->trHeapMemory()) TR::MemoryReference(currentNode, (TR::SymbolReference*)virtReg->getBackingStorage()->getSymbolReference(), sizeof(uintptr_t), cg);
TR::InstOpCode::Mnemonic opCode;
TR_RegisterKinds rk = virtReg->getKind();
switch (rk)
{
case TR_GPR:
opCode =TR::InstOpCode::Op_load;
break;
case TR_FPR:
opCode = virtReg->isSinglePrecision() ? TR::InstOpCode::lfs : TR::InstOpCode::lfd;
break;
default:
TR_ASSERT(0, "\nRegister kind not supported in OOL spill\n");
break;
}
TR::Instruction *inst = generateTrg1MemInstruction(cg, opCode, currentNode, assignedReg, tempMR, currentInstruction);
assignedReg->setAssignedRegister(NULL);
virtReg->setAssignedRegister(NULL);
assignedReg->setState(TR::RealRegister::Free);
if (comp->getDebug())
cg->traceRegisterAssignment("Generate reload of virt %s due to spillRegIndex dep at inst %p\n",comp->getDebug()->getName(virtReg),currentInstruction);
cg->traceRAInstruction(inst);
}
if (!(std::find(cg->getSpilledRegisterList()->begin(), cg->getSpilledRegisterList()->end(), virtReg) != cg->getSpilledRegisterList()->end()))
cg->getSpilledRegisterList()->push_front(virtReg);
}
// we also need to free up all locked backing storage if we are exiting the OOL during backwards RA assignment
else if (currentInstruction->isLabel() && virtReg->getAssignedRealRegister())
{
TR::PPCLabelInstruction *labelInstr = (TR::PPCLabelInstruction *)currentInstruction;
TR_BackingStore * location = virtReg->getBackingStorage();
TR_RegisterKinds rk = virtReg->getKind();
int32_t dataSize;
if (labelInstr->getLabelSymbol()->isStartOfColdInstructionStream() && location)
{
traceMsg (comp,"\nOOL: Releasing backing storage (%p)\n", location);
if (rk == TR_GPR)
dataSize = TR::Compiler->om.sizeofReferenceAddress();
else
dataSize = 8;
location->setMaxSpillDepth(0);
cg->freeSpill(location,dataSize,0);
virtReg->setBackingStorage(NULL);
}
}
}
}
for (i = 0; i < numberOfRegisters; i++)
{
map.addDependency(_dependencies[i], i);
virtReg = _dependencies[i].getRegister();
dependentRegNum = _dependencies[i].getRealRegister();
if (dependentRegNum != TR::RealRegister::SpilledReg)
{
if (virtReg->getKind() == TR_GPR)
num_gprs++;
else if (virtReg->getKind() == TR_FPR)
num_fprs++;
else if (virtReg->getKind() == TR_VRF)
num_vrfs++;
}
}
#ifdef DEBUG
int locked_gprs = 0;
int locked_fprs = 0;
int locked_vrfs = 0;
// count up how many registers are locked for each type
for(i = TR::RealRegister::FirstGPR; i <= TR::RealRegister::LastGPR; i++)
{
realReg = machine->getPPCRealRegister((TR::RealRegister::RegNum)i);
if (realReg->getState() == TR::RealRegister::Locked)
locked_gprs++;
}
for(i = TR::RealRegister::FirstFPR; i <= TR::RealRegister::LastFPR; i++)
{
realReg = machine->getPPCRealRegister((TR::RealRegister::RegNum)i);
if (realReg->getState() == TR::RealRegister::Locked)
locked_fprs++;
}
for(i = TR::RealRegister::FirstVRF; i <= TR::RealRegister::LastVRF; i++)
{
realReg = machine->getPPCRealRegister((TR::RealRegister::RegNum)i);
if (realReg->getState() == TR::RealRegister::Locked)
locked_vrfs++;
}
TR_ASSERT( locked_gprs == machine->getNumberOfLockedRegisters(TR_GPR),"Inconsistent number of locked GPRs");
TR_ASSERT( locked_fprs == machine->getNumberOfLockedRegisters(TR_FPR),"Inconsistent number of locked FPRs");
TR_ASSERT( locked_vrfs == machine->getNumberOfLockedRegisters(TR_VRF), "Inconsistent number of locked VRFs");
#endif
// To handle circular dependencies, we block a real register if (1) it is already assigned to a correct
// virtual register and (2) if it is assigned to one register in the list but is required by another.
// However, if all available registers are requested, we do not block in case (2) to avoid all registers
// being blocked.
bool block_gprs = true;
bool block_fprs = true;
bool block_vrfs = true;
TR_ASSERT(num_gprs <= (TR::RealRegister::LastGPR - TR::RealRegister::FirstGPR + 1 - machine->getNumberOfLockedRegisters(TR_GPR)), "Too many GPR dependencies, unable to assign" );
TR_ASSERT(num_fprs <= (TR::RealRegister::LastFPR - TR::RealRegister::FirstFPR + 1 - machine->getNumberOfLockedRegisters(TR_FPR)), "Too many FPR dependencies, unable to assign" );
TR_ASSERT(num_vrfs <= (TR::RealRegister::LastVRF - TR::RealRegister::FirstVRF + 1 - machine->getNumberOfLockedRegisters(TR_VRF)), "Too many VRF dependencies, unable to assign" );
if (num_gprs == (TR::RealRegister::LastGPR - TR::RealRegister::FirstGPR + 1 - machine->getNumberOfLockedRegisters(TR_GPR)))
block_gprs = false;
if (num_fprs == (TR::RealRegister::LastFPR - TR::RealRegister::FirstFPR + 1 - machine->getNumberOfLockedRegisters(TR_FPR)))
block_fprs = false;
if (num_vrfs == (TR::RealRegister::LastVRF - TR::RealRegister::FirstVRF + 1 - machine->getNumberOfLockedRegisters(TR_VRF)))
block_vrfs = false;
for (i = 0; i < numberOfRegisters; i++)
{
virtReg = _dependencies[i].getRegister();
if (virtReg->getAssignedRealRegister()!=NULL)
{
if (_dependencies[i].getRealRegister() == TR::RealRegister::NoReg)
{
virtReg->block();
}
else
{
TR::RealRegister::RegNum assignedRegNum;
assignedRegNum = toRealRegister(virtReg->getAssignedRealRegister())->getRegisterNumber();
// always block if required register and assigned register match;
// block if assigned register is required by other dependency but only if
// any spare registers are left to avoid blocking all existing registers
if (_dependencies[i].getRealRegister() == assignedRegNum ||
(map.getDependencyWithTarget(assignedRegNum) &&
((virtReg->getKind() != TR_GPR || block_gprs) &&
(virtReg->getKind() != TR_FPR || block_fprs) &&
(virtReg->getKind() != TR_VRF || block_vrfs))))
{
virtReg->block();
}
}
}
}
// Assign all virtual regs that depend on a specific real reg that is free
for (i = 0; i < numberOfRegisters; i++)
{
virtReg = _dependencies[i].getRegister();
dependentRegNum = _dependencies[i].getRealRegister();
dependentRealReg = machine->getPPCRealRegister(dependentRegNum);
if (dependentRegNum != TR::RealRegister::NoReg &&
dependentRegNum != TR::RealRegister::SpilledReg &&
dependentRealReg->getState() == TR::RealRegister::Free)
{
assignFreeRegisters(currentInstruction, &_dependencies[i], map, cg);
}
}
// Assign all virtual regs that depend on a specfic real reg that is not free
for (i = 0; i < numberOfRegisters; i++)
{
virtReg = _dependencies[i].getRegister();
assignedRegister = NULL;
if (virtReg->getAssignedRealRegister() != NULL)
{
assignedRegister = toRealRegister(virtReg->getAssignedRealRegister());
}
dependentRegNum = _dependencies[i].getRealRegister();
dependentRealReg = machine->getPPCRealRegister(dependentRegNum);
if (dependentRegNum != TR::RealRegister::NoReg &&
dependentRegNum != TR::RealRegister::SpilledReg &&
dependentRealReg != assignedRegister)
{
bool depsBlocked = false;
switch (_dependencies[i].getRegister()->getKind())
{
case TR_GPR:
depsBlocked = block_gprs;
break;
case TR_FPR:
depsBlocked = block_fprs;
break;
case TR_VRF:
depsBlocked = block_vrfs;
break;
}
assignContendedRegisters(currentInstruction, &_dependencies[i], map, depsBlocked, cg);
}
}
// Assign all virtual regs that depend on NoReg but exclude gr0
for (i=0; i<numberOfRegisters; i++)
{
if (_dependencies[i].getRealRegister() == TR::RealRegister::NoReg && _dependencies[i].getExcludeGPR0())
{
TR::RealRegister *realOne;
virtReg = _dependencies[i].getRegister();
realOne = virtReg->getAssignedRealRegister();
if (realOne!=NULL && toRealRegister(realOne)->getRegisterNumber()==TR::RealRegister::gr0)
{
if ((assignedRegister = machine->findBestFreeRegister(currentInstruction, virtReg->getKind(), true, false, virtReg)) == NULL)
{
assignedRegister = machine->freeBestRegister(currentInstruction, virtReg, NULL, true);
}
machine->coerceRegisterAssignment(currentInstruction, virtReg, assignedRegister->getRegisterNumber());
}
else if (realOne == NULL)
{
machine->assignOneRegister(currentInstruction, virtReg, true);
}
virtReg->block();
}
}
// Assign all virtual regs that depend on NoReg
for (i=0; i<numberOfRegisters; i++)
{
if (_dependencies[i].getRealRegister() == TR::RealRegister::NoReg && !_dependencies[i].getExcludeGPR0())
{
TR::RealRegister *realOne;
virtReg = _dependencies[i].getRegister();
realOne = virtReg->getAssignedRealRegister();
if (!realOne)
{
machine->assignOneRegister(currentInstruction, virtReg, false);
}
virtReg->block();
}
}
unblockRegisters(numberOfRegisters);
for (i = 0; i < numberOfRegisters; i++)
{
TR::Register *dependentRegister = getRegisterDependency(i)->getRegister();
// dependentRegister->getAssignedRegister() is NULL if the reg has already been spilled due to a spilledReg dep
if (comp->getOption(TR_DisableOOL) || (!(cg->isOutOfLineColdPath()) && !(cg->isOutOfLineHotPath())))
{
TR_ASSERT(dependentRegister->getAssignedRegister(),
"assignedRegister can not be NULL");
}
if (dependentRegister->getAssignedRegister())
{
TR::RealRegister *assignedRegister = dependentRegister->getAssignedRegister()->getRealRegister();
if (getRegisterDependency(i)->getRealRegister() == TR::RealRegister::NoReg)
getRegisterDependency(i)->setRealRegister(toRealRegister(assignedRegister)->getRegisterNumber());
machine->decFutureUseCountAndUnlatch(dependentRegister);
}
}
}
static void assignContendedRegisters(TR::Instruction *currentInstruction,
TR::RegisterDependency *dep,
TR_PPCRegisterDependencyMap& map,
bool depsBlocked,
TR::CodeGenerator *cg)
{
// *this swipeable for debugging purposes
TR::Machine *machine = cg->machine();
dep = findDependencyChainHead(dep, map);
TR::Register *virtReg = dep->getRegister();
TR::RealRegister::RegNum targetRegNum = dep->getRealRegister();
TR::RealRegister *targetReg = machine->getPPCRealRegister(targetRegNum);
TR::RealRegister *assignedReg = virtReg->getAssignedRealRegister() ?
toRealRegister(virtReg->getAssignedRealRegister()) : NULL;
// Chain of length 1
if (!assignedReg || !map.getDependencyWithTarget(assignedReg->getRegisterNumber()))
{
machine->coerceRegisterAssignment(currentInstruction, virtReg, targetRegNum);
virtReg->block();
return;
}
// Chain of length 2, handled here instead of below to get 3*xor exchange on GPRs
if (map.getDependencyWithTarget(assignedReg->getRegisterNumber()) == map.getDependencyWithAssigned(targetRegNum))
{
TR::Register *targetVirtReg = targetReg->getAssignedRegister();
machine->coerceRegisterAssignment(currentInstruction, virtReg, targetRegNum);
virtReg->block();
targetVirtReg->block();
return;
}
// Grab a spare reg in order to free the target of the first dep
// At this point the first dep's target could be blocked, assigned, or NoReg
// If it's blocked or assigned we allocate a spare and assign the target's virtual to it
// If it's NoReg, the spare reg will be used as the first dep's actual target
TR::RealRegister *spareReg = machine->findBestFreeRegister(currentInstruction, virtReg->getKind(),
targetRegNum == TR::RealRegister::NoReg ? dep->getExcludeGPR0() : false, false,
targetRegNum == TR::RealRegister::NoReg ? virtReg : targetReg->getAssignedRegister());
bool haveFreeSpare = spareReg != NULL;
if (!spareReg)
{
// If the regs in this dep group are not blocked we need to make sure we don't spill a reg that's in the middle of the chain
if (!depsBlocked)
{
if (targetRegNum == TR::RealRegister::NoReg)
spareReg = machine->freeBestRegister(currentInstruction,
map.getDependencyWithTarget(assignedReg->getRegisterNumber())->getRegister(),
assignedReg, false);
else
spareReg = machine->freeBestRegister(currentInstruction, virtReg, targetReg, false);
}
else
{
if (targetRegNum == TR::RealRegister::NoReg)
spareReg = machine->freeBestRegister(currentInstruction, virtReg, NULL, dep->getExcludeGPR0());
else
spareReg = machine->freeBestRegister(currentInstruction, targetReg->getAssignedRegister(), NULL, false);
}
}
if (targetRegNum != TR::RealRegister::NoReg && spareReg != targetReg)
{
machine->coerceRegisterAssignment(currentInstruction, targetReg->getAssignedRegister(), spareReg->getRegisterNumber());
}
TR_ASSERT(targetRegNum == TR::RealRegister::NoReg ||
targetReg->getState() == TR::RealRegister::Free, "Expecting free target register");
if (depsBlocked || targetRegNum != TR::RealRegister::NoReg || haveFreeSpare)
{
machine->coerceRegisterAssignment(currentInstruction, virtReg,
targetRegNum == TR::RealRegister::NoReg ?
spareReg->getRegisterNumber() : targetRegNum);
virtReg->block();
}
dep = map.getDependencyWithTarget(assignedReg->getRegisterNumber());
while (dep)
{
virtReg = dep->getRegister();
targetRegNum = dep->getRealRegister();
targetReg = machine->getPPCRealRegister(targetRegNum);
assignedReg = virtReg->getAssignedRealRegister() ?
toRealRegister(virtReg->getAssignedRealRegister()) : NULL;
TR_ASSERT(targetReg->getState() == TR::RealRegister::Free || targetReg == spareReg,
"Expecting free target register or target to have been filled to free spare register");
machine->coerceRegisterAssignment(currentInstruction, virtReg, targetRegNum);
virtReg->block();
dep = assignedReg ?
map.getDependencyWithTarget(assignedReg->getRegisterNumber()) : NULL;
}
}
void OMR::X86::Instruction::assignRegisters(TR_RegisterKinds kindsToBeAssigned)
{
if (!self()->getDependencyConditions())
{
// Fast path when there are no dependency conditions.
//
return;
}
if (self()->getOpCodeValue() != ASSOCREGS)
{
self()->aboutToAssignRegDeps();
if ((self()->cg()->getAssignmentDirection() == self()->cg()->Backward))
{
self()->getDependencyConditions()->assignPostConditionRegisters(self(), kindsToBeAssigned, self()->cg());
self()->getDependencyConditions()->assignPreConditionRegisters(self(), kindsToBeAssigned, self()->cg());
}
else
{
self()->getDependencyConditions()->assignPreConditionRegisters(self()->getPrev(), kindsToBeAssigned, self()->cg());
self()->getDependencyConditions()->assignPostConditionRegisters(self(), kindsToBeAssigned, self()->cg());
}
}
else if ((self()->getOpCodeValue() == ASSOCREGS) && self()->cg()->enableRegisterAssociations())
{
if (kindsToBeAssigned & TR_GPR_Mask)
{
TR::Machine *machine = self()->cg()->machine();
// First traverse the existing associations and remove them
// so that they don't interfere with the new ones
//
for (int i = TR::RealRegister::FirstGPR;
i <= TR::RealRegister::LastAssignableGPR;
++i)
{
// Skip non-assignable registers
//
if (machine->getX86RealRegister((TR::RealRegister::RegNum)i)->getState() == TR::RealRegister::Locked)
continue;
TR::Register *virtReg = machine->getVirtualAssociatedWithReal((TR::RealRegister::RegNum)i);
if (virtReg)
{
virtReg->setAssociation(TR::RealRegister::NoReg);
}
}
// Next loop through and set up the new associations (both on the machine
// and by associating the virtual registers with their real dependencies)
//
TR_X86RegisterDependencyGroup *depGroup = self()->getDependencyConditions()->getPostConditions();
for (int j = 0; j < self()->getDependencyConditions()->getNumPostConditions(); ++j)
{
TR::RegisterDependency *dep = depGroup->getRegisterDependency(j);
machine->setVirtualAssociatedWithReal(dep->getRealRegister(), dep->getRegister());
}
machine->setGPRWeightsFromAssociations();
}
}
}
void
TR_S390BinaryAnalyser::genericAnalyser(TR::Node * root,
TR::InstOpCode::Mnemonic regToRegOpCode,
TR::InstOpCode::Mnemonic memToRegOpCode,
TR::InstOpCode::Mnemonic copyOpCode)
{
TR::Node * firstChild;
TR::Node * secondChild;
firstChild = root->getFirstChild();
secondChild = root->getSecondChild();
TR::Register * firstRegister = firstChild->getRegister();
TR::Register * secondRegister = secondChild->getRegister();
TR::Compilation *comp = TR::comp();
TR::SymbolReference * firstReference = firstChild->getOpCode().hasSymbolReference() ? firstChild->getSymbolReference() : NULL;
TR::SymbolReference * secondReference = secondChild->getOpCode().hasSymbolReference() ? secondChild->getSymbolReference() : NULL;
setInputs(firstChild, firstRegister, secondChild, secondRegister,
false, false, comp,
(cg()->isAddressOfStaticSymRefWithLockedReg(firstReference) ||
cg()->isAddressOfPrivateStaticSymRefWithLockedReg(firstReference)),
(cg()->isAddressOfStaticSymRefWithLockedReg(secondReference) ||
cg()->isAddressOfPrivateStaticSymRefWithLockedReg(secondReference)));
/*
* Check if SH or CH can be used to evaluate this integer subtract/compare node.
* The second operand of SH/CH is a 16-bit number from memory. And using
* these directly can save a load instruction.
*/
bool is16BitMemory2Operand = false;
if (secondChild->getOpCodeValue() == TR::s2i &&
secondChild->getFirstChild()->getOpCodeValue() == TR::sloadi &&
secondChild->isSingleRefUnevaluated() &&
secondChild->getFirstChild()->isSingleRefUnevaluated())
{
bool supported = true;
if (memToRegOpCode == TR::InstOpCode::S)
{
memToRegOpCode = TR::InstOpCode::SH;
}
else if (memToRegOpCode == TR::InstOpCode::C)
{
memToRegOpCode = TR::InstOpCode::CH;
}
else
{
supported = false;
}
if (supported)
{
setMem2();
is16BitMemory2Operand = true;
}
}
if (getEvalChild1())
{
firstRegister = cg()->evaluate(firstChild);
}
if (getEvalChild2())
{
secondRegister = cg()->evaluate(secondChild);
}
remapInputs(firstChild, firstRegister, secondChild, secondRegister);
if (getCopyReg1())
{
TR::Register * thirdReg;
bool done = false;
if (firstRegister->getKind() == TR_GPR64)
{
thirdReg = cg()->allocate64bitRegister();
}
else if (firstRegister->getKind() == TR_VRF)
{
TR_ASSERT(false,"VRF: genericAnalyser unimplemented");
}
else if (firstRegister->getKind() != TR_FPR && firstRegister->getKind() != TR_VRF)
{
thirdReg = cg()->allocateRegister();
}
else
{
thirdReg = cg()->allocateRegister(TR_FPR);
}
if (cg()->getS390ProcessorInfo()->supportsArch(TR_S390ProcessorInfo::TR_z196))
{
if (getBinaryReg3Reg2() || secondRegister != NULL)
{
if (regToRegOpCode == TR::InstOpCode::SR)
{
generateRRRInstruction(cg(), TR::InstOpCode::SRK, root, thirdReg, firstRegister, secondRegister);
done = true;
}
else if (regToRegOpCode == TR::InstOpCode::SLR)
{
generateRRRInstruction(cg(), TR::InstOpCode::SLRK, root, thirdReg, firstRegister, secondRegister);
done = true;
}
else if (regToRegOpCode == TR::InstOpCode::SGR)
{
generateRRRInstruction(cg(), TR::InstOpCode::SGRK, root, thirdReg, firstRegister, secondRegister);
done = true;
}
else if (regToRegOpCode == TR::InstOpCode::SLGR)
{
generateRRRInstruction(cg(), TR::InstOpCode::SLGRK, root, thirdReg, firstRegister, secondRegister);
done = true;
}
}
}
if (!done)
{
generateRRInstruction(cg(), copyOpCode, root, thirdReg, firstRegister);
if (getBinaryReg3Reg2() || (secondRegister != NULL))
{
generateRRInstruction(cg(), regToRegOpCode, root, thirdReg, secondRegister);
}
else
{
TR::Node* loadBaseAddr = is16BitMemory2Operand ? secondChild->getFirstChild() : secondChild;
TR::MemoryReference * tempMR = generateS390MemoryReference(loadBaseAddr, cg());
//floating-point arithmatics don't have RXY format instructions, so no long displacement
if (secondChild->getOpCode().isFloatingPoint())
{
tempMR->enforce4KDisplacementLimit(secondChild, cg(), NULL);
}
generateRXInstruction(cg(), memToRegOpCode, root, thirdReg, tempMR);
tempMR->stopUsingMemRefRegister(cg());
if (is16BitMemory2Operand)
{
cg()->decReferenceCount(secondChild->getFirstChild());
}
}
}
root->setRegister(thirdReg);
}
else if (getBinaryReg1Reg2())
{
generateRRInstruction(cg(), regToRegOpCode, root, firstRegister, secondRegister);
root->setRegister(firstRegister);
}
else // assert getBinaryReg1Mem2() == true
{
TR_ASSERT( !getInvalid(), "TR_S390BinaryAnalyser::invalid case\n");
TR::MemoryReference * tempMR = generateS390MemoryReference(is16BitMemory2Operand ? secondChild->getFirstChild() : secondChild, cg());
//floating-point arithmatics don't have RXY format instructions, so no long displacement
if (secondChild->getOpCode().isFloatingPoint())
{
tempMR->enforce4KDisplacementLimit(secondChild, cg(), NULL);
}
generateRXInstruction(cg(), memToRegOpCode, root, firstRegister, tempMR);
tempMR->stopUsingMemRefRegister(cg());
if (is16BitMemory2Operand)
cg()->decReferenceCount(secondChild->getFirstChild());
root->setRegister(firstRegister);
}
cg()->decReferenceCount(firstChild);
cg()->decReferenceCount(secondChild);
return;
}
TR::Register *IA32LinkageUtils::pushDoubleArg(
TR::Node *child,
TR::CodeGenerator *cg)
{
TR::Register *pushRegister;
if (child->getRegister() == NULL)
{
if (child->getOpCodeValue() == TR::dconst)
{
TR_X86OpCodes pushOp;
int32_t highValue = child->getLongIntHigh();
if (highValue >= -128 && highValue <= 127)
{
pushOp = PUSHImms;
}
else
{
pushOp = PUSHImm4;
}
generateImmInstruction(pushOp, child, highValue, cg);
int32_t lowValue = child->getLongIntLow();
if (lowValue >= -128 && lowValue <= 127)
{
pushOp = PUSHImms;
}
else
{
pushOp = PUSHImm4;
}
generateImmInstruction(pushOp, child, lowValue, cg);
cg->decReferenceCount(child);
return NULL;
}
else if (child->getReferenceCount() == 1)
{
if (child->getOpCode().isLoad())
{
TR::MemoryReference *lowMR = generateX86MemoryReference(child, cg);
generateMemInstruction(PUSHMem, child, generateX86MemoryReference(*lowMR, 4, cg), cg);
generateMemInstruction(PUSHMem, child, lowMR, cg);
lowMR->decNodeReferenceCounts(cg);
cg->decReferenceCount(child);
return NULL;
}
else if (child->getOpCodeValue() == TR::lbits2d)
{
pushRegister = pushLongArg(child->getFirstChild(), cg);
cg->decReferenceCount(child);
return pushRegister;
}
}
}
pushRegister = cg->evaluate(child);
TR::RealRegister *espReal = cg->machine()->getRealRegister(TR::RealRegister::esp);
generateRegImmInstruction(SUB4RegImms, child, espReal, 8, cg);
if (cg->useSSEForSinglePrecision() && pushRegister->getKind() == TR_FPR)
generateMemRegInstruction(MOVSDMemReg, child, generateX86MemoryReference(espReal, 0, cg), pushRegister, cg);
else
generateFPMemRegInstruction(DSTMemReg, child, generateX86MemoryReference(espReal, 0, cg), pushRegister, cg);
cg->decReferenceCount(child);
return pushRegister;
}
TR::Register *TR_X86FPCompareAnalyser::fpCompareAnalyser(TR::Node *root,
TR_X86OpCodes cmpRegRegOpCode,
TR_X86OpCodes cmpRegMemOpCode,
TR_X86OpCodes cmpiRegRegOpCode,
bool useFCOMIInstructions)
{
TR::Node *firstChild,
*secondChild;
TR::ILOpCodes cmpOp = root->getOpCodeValue();
bool reverseMemOp = false;
bool reverseCmpOp = false;
TR::Compilation* comp = _cg->comp();
TR_X86OpCodes cmpInstr = useFCOMIInstructions ? cmpiRegRegOpCode : cmpRegRegOpCode;
// Some operators must have their operands swapped to improve the generated
// code needed to evaluate the result of the comparison.
//
bool mustSwapOperands = (cmpOp == TR::iffcmple ||
cmpOp == TR::ifdcmple ||
cmpOp == TR::iffcmpgtu ||
cmpOp == TR::ifdcmpgtu ||
cmpOp == TR::fcmple ||
cmpOp == TR::dcmple ||
cmpOp == TR::fcmpgtu ||
cmpOp == TR::dcmpgtu ||
(useFCOMIInstructions &&
(cmpOp == TR::iffcmplt ||
cmpOp == TR::ifdcmplt ||
cmpOp == TR::iffcmpgeu ||
cmpOp == TR::ifdcmpgeu ||
cmpOp == TR::fcmplt ||
cmpOp == TR::dcmplt ||
cmpOp == TR::fcmpgeu ||
cmpOp == TR::dcmpgeu))) ? true : false;
// Some operators should not have their operands swapped to improve the generated
// code needed to evaluate the result of the comparison.
//
bool preventOperandSwapping = (cmpOp == TR::iffcmpltu ||
cmpOp == TR::ifdcmpltu ||
cmpOp == TR::iffcmpge ||
cmpOp == TR::ifdcmpge ||
cmpOp == TR::fcmpltu ||
cmpOp == TR::dcmpltu ||
cmpOp == TR::fcmpge ||
cmpOp == TR::dcmpge ||
(useFCOMIInstructions &&
(cmpOp == TR::iffcmpgt ||
cmpOp == TR::ifdcmpgt ||
cmpOp == TR::iffcmpleu ||
cmpOp == TR::ifdcmpleu ||
cmpOp == TR::fcmpgt ||
cmpOp == TR::dcmpgt ||
cmpOp == TR::fcmpleu ||
cmpOp == TR::dcmpleu))) ? true : false;
// For correctness, don't swap operands of these operators.
//
if (cmpOp == TR::fcmpg || cmpOp == TR::fcmpl ||
cmpOp == TR::dcmpg || cmpOp == TR::dcmpl)
{
preventOperandSwapping = true;
}
// Initial operand evaluation ordering.
//
if (preventOperandSwapping || (!mustSwapOperands && _cg->whichChildToEvaluate(root) == 0))
{
firstChild = root->getFirstChild();
secondChild = root->getSecondChild();
setReversedOperands(false);
}
else
{
firstChild = root->getSecondChild();
secondChild = root->getFirstChild();
setReversedOperands(true);
}
TR::Register *firstRegister = firstChild->getRegister();
TR::Register *secondRegister = secondChild->getRegister();
setInputs(firstChild,
firstRegister,
secondChild,
secondRegister,
useFCOMIInstructions,
// If either 'preventOperandSwapping' or 'mustSwapOperands' is set then the
// initial operand ordering set above must be maintained.
//
preventOperandSwapping || mustSwapOperands);
// Make sure any required operand ordering is respected.
//
if ((getCmpReg2Reg1() || getCmpReg2Mem1()) &&
(mustSwapOperands || preventOperandSwapping))
{
reverseCmpOp = getCmpReg2Reg1() ? true : false;
reverseMemOp = getCmpReg2Mem1() ? true : false;
}
// If we are not comparing with a memory operand, one of them evaluates
// to a zero, and the zero is not already on the stack, then we can use
// FTST to save a register.
//
// (With a memory operand, either the constant zero needs to be loaded
// to use FCOM, or the memory operand needs to be loaded to use FTST,
// so there is no gain in using FTST.)
//
// If the constant zero is in the target register, using FTST means the
// comparison will be reversed. We cannot do this if the initial ordering
// of the operands must be maintained.
//
// Finally, if FTST is used and this is the last use of the target, the
// target register may need to be explicitly popped.
//
TR::Register *targetRegisterForFTST = NULL;
TR::Node *targetChildForFTST = NULL;
if (getEvalChild1() && isUnevaluatedZero(firstChild)) // do we need getEvalChild1() here?
{
if ( ((getCmpReg1Reg2() || reverseCmpOp) && !(preventOperandSwapping || mustSwapOperands)) ||
(getCmpReg2Reg1() && !reverseCmpOp))
{
if (getEvalChild2())
{
secondRegister = _cg->evaluate(secondChild);
}
targetRegisterForFTST = secondRegister;
targetChildForFTST = secondChild;
notReversedOperands();
}
}
else if (getEvalChild2() && isUnevaluatedZero(secondChild)) // do we need getEvalChild2() here?
{
if ( (getCmpReg1Reg2() || reverseCmpOp) ||
(getCmpReg2Reg1() && !reverseCmpOp && !(preventOperandSwapping || mustSwapOperands)) )
{
if (getEvalChild1())
{
firstRegister = _cg->evaluate(firstChild);
}
targetRegisterForFTST = firstRegister;
targetChildForFTST = firstChild;
}
}
if (!targetRegisterForFTST)
{
// If we have a choice, evaluate the target operand last. By doing so, we
// help out the register assigner because the target must be TOS. This
// avoids an unneccessary FXCH for the target.
//
if (getEvalChild1() && getEvalChild2())
{
if (getCmpReg1Reg2() || getCmpReg1Mem2())
{
secondRegister = _cg->evaluate(secondChild);
firstRegister = _cg->evaluate(firstChild);
}
else
{
firstRegister = _cg->evaluate(firstChild);
secondRegister = _cg->evaluate(secondChild);
}
}
else
{
if (getEvalChild1())
{
firstRegister = _cg->evaluate(firstChild);
}
if (getEvalChild2())
{
secondRegister = _cg->evaluate(secondChild);
}
}
}
// Adjust the FP precision of feeding operands.
//
if (firstRegister &&
(firstRegister->needsPrecisionAdjustment() ||
comp->getOption(TR_StrictFPCompares) ||
(firstRegister->mayNeedPrecisionAdjustment() && secondChild->getOpCode().isLoadConst()) ||
(firstRegister->mayNeedPrecisionAdjustment() && !secondRegister)))
{
TR::TreeEvaluator::insertPrecisionAdjustment(firstRegister, root, _cg);
}
if (secondRegister &&
(secondRegister->needsPrecisionAdjustment() ||
comp->getOption(TR_StrictFPCompares) ||
(secondRegister->mayNeedPrecisionAdjustment() && firstChild->getOpCode().isLoadConst()) ||
(secondRegister->mayNeedPrecisionAdjustment() && !firstRegister)))
{
TR::TreeEvaluator::insertPrecisionAdjustment(secondRegister, root, _cg);
}
// Generate the compare instruction.
//
if (targetRegisterForFTST)
{
generateFPRegInstruction(FTSTReg, root, targetRegisterForFTST, _cg);
}
else if (!useFCOMIInstructions && (getCmpReg1Mem2() || reverseMemOp))
{
TR::MemoryReference *tempMR = generateX86MemoryReference(secondChild, _cg);
generateFPRegMemInstruction(cmpRegMemOpCode, root, firstRegister, tempMR, _cg);
tempMR->decNodeReferenceCounts(_cg);
}
else if (!useFCOMIInstructions && getCmpReg2Mem1())
{
TR::MemoryReference *tempMR = generateX86MemoryReference(firstChild, _cg);
generateFPRegMemInstruction(cmpRegMemOpCode, root, secondRegister, tempMR, _cg);
notReversedOperands();
tempMR->decNodeReferenceCounts(_cg);
}
else if (getCmpReg1Reg2() || reverseCmpOp)
{
generateFPCompareRegRegInstruction(cmpInstr, root, firstRegister, secondRegister, _cg);
}
else if (getCmpReg2Reg1())
{
generateFPCompareRegRegInstruction(cmpInstr, root, secondRegister, firstRegister, _cg);
notReversedOperands();
}
_cg->decReferenceCount(firstChild);
_cg->decReferenceCount(secondChild);
// Evaluate the comparison.
//
if (getReversedOperands())
{
cmpOp = TR::ILOpCode(cmpOp).getOpCodeForSwapChildren();
TR::Node::recreate(root, cmpOp);
}
if (useFCOMIInstructions && !targetRegisterForFTST)
{
return NULL;
}
// We must manually move the FP condition flags to the EFLAGS register if we don't
// use the FCOMI instructions.
//
TR::Register *accRegister = _cg->allocateRegister();
TR::RegisterDependencyConditions *dependencies = generateRegisterDependencyConditions((uint8_t)1, 1, _cg);
dependencies->addPreCondition(accRegister, TR::RealRegister::eax, _cg);
dependencies->addPostCondition(accRegister, TR::RealRegister::eax, _cg);
generateRegInstruction(STSWAcc, root, accRegister, dependencies, _cg);
// Pop the FTST target register if it is not used any more.
//
if (targetRegisterForFTST &&
targetChildForFTST && targetChildForFTST->getReferenceCount() == 0)
{
generateFPSTiST0RegRegInstruction(FSTRegReg, root, targetRegisterForFTST, targetRegisterForFTST, _cg);
}
return accRegister;
}
void
TR_S390BinaryAnalyser::longSubtractAnalyser(TR::Node * root)
{
TR::Node * firstChild;
TR::Node * secondChild;
TR::Instruction * cursor = NULL;
TR::RegisterDependencyConditions * dependencies = NULL;
bool setsOrReadsCC = NEED_CC(root) || (root->getOpCodeValue() == TR::lusubb);
TR::InstOpCode::Mnemonic regToRegOpCode;
TR::InstOpCode::Mnemonic memToRegOpCode;
TR::Compilation *comp = TR::comp();
if (TR::Compiler->target.is64Bit() || cg()->use64BitRegsOn32Bit())
{
if (!setsOrReadsCC)
{
regToRegOpCode = TR::InstOpCode::SGR;
memToRegOpCode = TR::InstOpCode::SG;
}
else
{
regToRegOpCode = TR::InstOpCode::SLGR;
memToRegOpCode = TR::InstOpCode::SLG;
}
}
else
{
regToRegOpCode = TR::InstOpCode::SLR;
memToRegOpCode = TR::InstOpCode::SL;
}
firstChild = root->getFirstChild();
secondChild = root->getSecondChild();
TR::Register * firstRegister = firstChild->getRegister();
TR::Register * secondRegister = secondChild->getRegister();
setInputs(firstChild, firstRegister, secondChild, secondRegister,
false, false, comp);
/** Attempt to use SGH to subtract halfword (64 <- 16).
* The second child is a halfword from memory */
bool is16BitMemory2Operand = false;
if (TR::Compiler->target.cpu.getS390SupportsZ14() &&
secondChild->getOpCodeValue() == TR::s2l &&
secondChild->getFirstChild()->getOpCodeValue() == TR::sloadi &&
secondChild->isSingleRefUnevaluated() &&
secondChild->getFirstChild()->isSingleRefUnevaluated())
{
setMem2();
memToRegOpCode = TR::InstOpCode::SGH;
is16BitMemory2Operand = true;
}
if (getEvalChild1())
{
firstRegister = cg()->evaluate(firstChild);
}
if (getEvalChild2())
{
secondRegister = cg()->evaluate(secondChild);
}
remapInputs(firstChild, firstRegister, secondChild, secondRegister);
if ((root->getOpCodeValue() == TR::lusubb) &&
TR_S390ComputeCC::setCarryBorrow(root->getChild(2), false, cg()))
{
// use SLBGR rather than SLGR/SGR
// SLBG rather than SLG/SG
// or
// use SLBR rather than SLR
// SLB rather than SL
bool uses64bit = TR::Compiler->target.is64Bit() || cg()->use64BitRegsOn32Bit();
regToRegOpCode = uses64bit ? TR::InstOpCode::SLBGR : TR::InstOpCode::SLBR;
memToRegOpCode = uses64bit ? TR::InstOpCode::SLBG : TR::InstOpCode::SLB;
}
if (TR::Compiler->target.is64Bit() || cg()->use64BitRegsOn32Bit())
{
if (getCopyReg1())
{
TR::Register * thirdReg = cg()->allocate64bitRegister();
root->setRegister(thirdReg);
generateRRInstruction(cg(), TR::InstOpCode::LGR, root, thirdReg, firstRegister);
if (getBinaryReg3Reg2())
{
generateRRInstruction(cg(), regToRegOpCode, root, thirdReg, secondRegister);
}
else // assert getBinaryReg3Mem2() == true
{
TR::MemoryReference * longMR = generateS390MemoryReference(secondChild, cg());
generateRXInstruction(cg(), memToRegOpCode, root, thirdReg, longMR);
longMR->stopUsingMemRefRegister(cg());
}
}
else if (getBinaryReg1Reg2())
{
generateRRInstruction(cg(), regToRegOpCode, root, firstRegister, secondRegister);
root->setRegister(firstRegister);
}
else // assert getBinaryReg1Mem2() == true
{
TR_ASSERT( !getInvalid(), "TR_S390BinaryAnalyser::invalid case\n");
TR::Node* baseAddrNode = is16BitMemory2Operand ? secondChild->getFirstChild() : secondChild;
TR::MemoryReference * longMR = generateS390MemoryReference(baseAddrNode, cg());
generateRXInstruction(cg(), memToRegOpCode, root, firstRegister, longMR);
longMR->stopUsingMemRefRegister(cg());
root->setRegister(firstRegister);
if(is16BitMemory2Operand)
{
cg()->decReferenceCount(secondChild->getFirstChild());
}
}
}
else // if 32bit codegen...
{
bool zArchTrexsupported = performTransformation(comp, "O^O Use SL/SLB for long sub.");
TR::Register * highDiff = NULL;
TR::LabelSymbol * doneLSub = TR::LabelSymbol::create(cg()->trHeapMemory(),cg());
if (getCopyReg1())
{
TR::Register * lowThird = cg()->allocateRegister();
TR::Register * highThird = cg()->allocateRegister();
TR::RegisterPair * thirdReg = cg()->allocateConsecutiveRegisterPair(lowThird, highThird);
highDiff = highThird;
dependencies = new (cg()->trHeapMemory()) TR::RegisterDependencyConditions(0, 9, cg());
dependencies->addPostCondition(firstRegister, TR::RealRegister::EvenOddPair);
dependencies->addPostCondition(firstRegister->getHighOrder(), TR::RealRegister::LegalEvenOfPair);
dependencies->addPostCondition(firstRegister->getLowOrder(), TR::RealRegister::LegalOddOfPair);
// If 2nd operand has ref count of 1 and can be accessed by a memory reference,
// then second register will not be used.
if(secondRegister == firstRegister && !setsOrReadsCC)
{
TR_ASSERT( false, "lsub with identical children - fix Simplifier");
}
if (secondRegister != NULL && firstRegister != secondRegister)
{
dependencies->addPostCondition(secondRegister, TR::RealRegister::EvenOddPair);
dependencies->addPostCondition(secondRegister->getHighOrder(), TR::RealRegister::LegalEvenOfPair);
dependencies->addPostCondition(secondRegister->getLowOrder(), TR::RealRegister::LegalOddOfPair);
}
dependencies->addPostCondition(highThird, TR::RealRegister::AssignAny);
root->setRegister(thirdReg);
generateRRInstruction(cg(), TR::InstOpCode::LR, root, highThird, firstRegister->getHighOrder());
generateRRInstruction(cg(), TR::InstOpCode::LR, root, lowThird, firstRegister->getLowOrder());
if (getBinaryReg3Reg2())
{
if ((ENABLE_ZARCH_FOR_32 && zArchTrexsupported) || setsOrReadsCC)
{
generateRRInstruction(cg(), regToRegOpCode, root, lowThird, secondRegister->getLowOrder());
generateRRInstruction(cg(), TR::InstOpCode::SLBR, root, highThird, secondRegister->getHighOrder());
}
else
{
generateRRInstruction(cg(), TR::InstOpCode::SR, root, highThird, secondRegister->getHighOrder());
generateRRInstruction(cg(), TR::InstOpCode::SLR, root, lowThird, secondRegister->getLowOrder());
}
}
else // assert getBinaryReg3Mem2() == true
{
TR::MemoryReference * highMR = generateS390MemoryReference(secondChild, cg());
TR::MemoryReference * lowMR = generateS390MemoryReference(*highMR, 4, cg());
dependencies->addAssignAnyPostCondOnMemRef(highMR);
if ((ENABLE_ZARCH_FOR_32 && zArchTrexsupported) || setsOrReadsCC)
{
generateRXInstruction(cg(), memToRegOpCode, root, lowThird, lowMR);
generateRXInstruction(cg(), TR::InstOpCode::SLB, root, highThird, highMR);
}
else
{
generateRXInstruction(cg(), TR::InstOpCode::S, root, highThird, highMR);
generateRXInstruction(cg(), TR::InstOpCode::SL, root, lowThird, lowMR);
}
highMR->stopUsingMemRefRegister(cg());
lowMR->stopUsingMemRefRegister(cg());
}
}
else if (getBinaryReg1Reg2())
{
dependencies = new (cg()->trHeapMemory()) TR::RegisterDependencyConditions(0, 6, cg());
dependencies->addPostCondition(firstRegister, TR::RealRegister::EvenOddPair);
dependencies->addPostCondition(firstRegister->getHighOrder(), TR::RealRegister::LegalEvenOfPair);
dependencies->addPostCondition(firstRegister->getLowOrder(), TR::RealRegister::LegalOddOfPair);
if(secondRegister == firstRegister)
{
TR_ASSERT( false, "lsub with identical children - fix Simplifier");
}
if (secondRegister != firstRegister)
{
dependencies->addPostCondition(secondRegister, TR::RealRegister::EvenOddPair);
dependencies->addPostCondition(secondRegister->getHighOrder(), TR::RealRegister::LegalEvenOfPair);
dependencies->addPostCondition(secondRegister->getLowOrder(), TR::RealRegister::LegalOddOfPair);
}
if ((ENABLE_ZARCH_FOR_32 && zArchTrexsupported) || setsOrReadsCC)
{
generateRRInstruction(cg(), regToRegOpCode, root, firstRegister->getLowOrder(), secondRegister->getLowOrder());
generateRRInstruction(cg(), TR::InstOpCode::SLBR, root, firstRegister->getHighOrder(), secondRegister->getHighOrder());
}
else
{
generateRRInstruction(cg(), TR::InstOpCode::SR, root, firstRegister->getHighOrder(), secondRegister->getHighOrder());
generateRRInstruction(cg(), TR::InstOpCode::SLR, root, firstRegister->getLowOrder(), secondRegister->getLowOrder());
}
highDiff = firstRegister->getHighOrder();
root->setRegister(firstRegister);
}
else // assert getBinaryReg1Mem2() == true
{
TR_ASSERT( !getInvalid(),"TR_S390BinaryAnalyser::invalid case\n");
dependencies = new (cg()->trHeapMemory()) TR::RegisterDependencyConditions(0, 5, cg());
dependencies->addPostCondition(firstRegister, TR::RealRegister::EvenOddPair);
dependencies->addPostCondition(firstRegister->getHighOrder(), TR::RealRegister::LegalEvenOfPair);
dependencies->addPostCondition(firstRegister->getLowOrder(), TR::RealRegister::LegalOddOfPair);
TR::MemoryReference * highMR = generateS390MemoryReference(secondChild, cg());
TR::MemoryReference * lowMR = generateS390MemoryReference(*highMR, 4, cg());
dependencies->addAssignAnyPostCondOnMemRef(highMR);
if ((ENABLE_ZARCH_FOR_32 && zArchTrexsupported) || setsOrReadsCC)
{
generateRXInstruction(cg(), memToRegOpCode, root, firstRegister->getLowOrder(), lowMR);
generateRXInstruction(cg(), TR::InstOpCode::SLB, root, firstRegister->getHighOrder(), highMR);
}
else
{
generateRXInstruction(cg(), TR::InstOpCode::S, root, firstRegister->getHighOrder(), highMR);
generateRXInstruction(cg(), TR::InstOpCode::SL, root, firstRegister->getLowOrder(), lowMR);
}
highDiff = firstRegister->getHighOrder();
root->setRegister(firstRegister);
highMR->stopUsingMemRefRegister(cg());
lowMR->stopUsingMemRefRegister(cg());
}
if (!((ENABLE_ZARCH_FOR_32 && zArchTrexsupported) || setsOrReadsCC))
{
// Check for overflow in LS int. If overflow, we are done.
generateS390BranchInstruction(cg(), TR::InstOpCode::BRC,TR::InstOpCode::COND_MASK3, root, doneLSub);
// Increment MS int due to overflow in LS int
generateRIInstruction(cg(), TR::InstOpCode::AHI, root, highDiff, -1);
generateS390LabelInstruction(cg(), TR::InstOpCode::LABEL, root, doneLSub, dependencies);
}
}
cg()->decReferenceCount(firstChild);
cg()->decReferenceCount(secondChild);
return;
}
OMR::Power::RegisterDependencyConditions::RegisterDependencyConditions(
TR::CodeGenerator *cg,
TR::Node *node,
uint32_t extranum,
TR::Instruction **cursorPtr)
{
List<TR::Register> regList(cg->trMemory());
TR::Instruction *iCursor = (cursorPtr==NULL)?NULL:*cursorPtr;
int32_t totalNum = node->getNumChildren() + extranum;
int32_t i;
cg->comp()->incVisitCount();
int32_t numLongs = 0;
//
// Pre-compute how many longs are global register candidates
//
for (i = 0; i < node->getNumChildren(); ++i)
{
TR::Node *child = node->getChild(i);
TR::Register *reg = child->getRegister();
if (reg!=NULL /* && reg->getKind()==TR_GPR */)
{
if (child->getHighGlobalRegisterNumber() > -1)
numLongs++;
}
}
totalNum = totalNum + numLongs;
_preConditions = new (totalNum, cg->trMemory()) TR_PPCRegisterDependencyGroup;
_postConditions = new (totalNum, cg->trMemory()) TR_PPCRegisterDependencyGroup;
_numPreConditions = totalNum;
_addCursorForPre = 0;
_numPostConditions = totalNum;
_addCursorForPost = 0;
// First, handle dependencies that match current association
for (i=0; i<node->getNumChildren(); i++)
{
TR::Node *child = node->getChild(i);
TR::Register *reg = child->getRegister();
TR::Register *highReg = NULL;
TR::RealRegister::RegNum regNum = (TR::RealRegister::RegNum)cg->getGlobalRegister(child->getGlobalRegisterNumber());
TR::RealRegister::RegNum highRegNum;
if (child->getHighGlobalRegisterNumber() > -1)
{
highRegNum = (TR::RealRegister::RegNum)cg->getGlobalRegister(child->getHighGlobalRegisterNumber());
TR::RegisterPair *regPair = reg->getRegisterPair();
TR_ASSERT(regPair, "assertion failure");
highReg = regPair->getHighOrder();
reg = regPair->getLowOrder();
if (highReg->getAssociation() != highRegNum ||
reg->getAssociation() != regNum)
continue;
}
else if (reg->getAssociation() != regNum)
continue;
TR_ASSERT(!regList.find(reg) && (!highReg || !regList.find(highReg)), "Should not happen\n");
addPreCondition(reg, regNum);
addPostCondition(reg, regNum);
regList.add(reg);
if (highReg)
{
addPreCondition(highReg, highRegNum);
addPostCondition(highReg, highRegNum);
regList.add(highReg);
}
}
// Second pass to handle dependencies for which association does not exist
// or does not match
for (i=0; i<node->getNumChildren(); i++)
{
TR::Node *child = node->getChild(i);
TR::Register *reg = child->getRegister();
TR::Register *highReg = NULL;
TR::Register *copyReg = NULL;
TR::Register *highCopyReg = NULL;
TR::RealRegister::RegNum regNum = (TR::RealRegister::RegNum)cg->getGlobalRegister(child->getGlobalRegisterNumber());
TR::RealRegister::RegNum highRegNum;
if (child->getHighGlobalRegisterNumber() > -1)
{
highRegNum = (TR::RealRegister::RegNum)cg->getGlobalRegister(child->getHighGlobalRegisterNumber());
TR::RegisterPair *regPair = reg->getRegisterPair();
TR_ASSERT(regPair, "assertion failure");
highReg = regPair->getHighOrder();
reg = regPair->getLowOrder();
if (highReg->getAssociation() == highRegNum &&
reg->getAssociation() == regNum)
continue;
}
else if (reg->getAssociation() == regNum)
continue;
if (regList.find(reg) || (highReg && regList.find(highReg)))
{
TR::InstOpCode::Mnemonic opCode;
TR_RegisterKinds kind = reg->getKind();
switch (kind)
{
case TR_GPR:
opCode = TR::InstOpCode::mr;
break;
case TR_FPR:
opCode = TR::InstOpCode::fmr;
break;
case TR_VRF:
opCode = TR::InstOpCode::vor;
//TR_ASSERT(0, "VMX not fully supported.");
break;
case TR_VSX_VECTOR:
opCode = TR::InstOpCode::xxlor;
break;
case TR_CCR:
opCode = TR::InstOpCode::mcrf;
break;
default:
TR_ASSERT(0, "Invalid register kind.");
}
if (regList.find(reg))
{
bool containsInternalPointer = false;
if (reg->getPinningArrayPointer())
containsInternalPointer = true;
copyReg = (reg->containsCollectedReference() && !containsInternalPointer) ?
cg->allocateCollectedReferenceRegister() : cg->allocateRegister(kind);
if (containsInternalPointer)
{
copyReg->setContainsInternalPointer();
copyReg->setPinningArrayPointer(reg->getPinningArrayPointer());
}
if (opCode == TR::InstOpCode::vor || opCode == TR::InstOpCode::xxlor)
iCursor = generateTrg1Src2Instruction(cg, opCode, node, copyReg, reg, reg, iCursor);
else
iCursor = generateTrg1Src1Instruction(cg, opCode, node, copyReg, reg, iCursor);
reg = copyReg;
}
if (highReg && regList.find(highReg))
{
bool containsInternalPointer = false;
if (highReg->getPinningArrayPointer())
containsInternalPointer = true;
highCopyReg = (highReg->containsCollectedReference() && !containsInternalPointer) ?
cg->allocateCollectedReferenceRegister() : cg->allocateRegister(kind);
if (containsInternalPointer)
{
highCopyReg->setContainsInternalPointer();
highCopyReg->setPinningArrayPointer(highReg->getPinningArrayPointer());
}
if (opCode == TR::InstOpCode::vor || opCode == TR::InstOpCode::xxlor)
iCursor = generateTrg1Src2Instruction(cg, opCode, node, highCopyReg, highReg, highReg, iCursor);
else
iCursor = generateTrg1Src1Instruction(cg, opCode, node, highCopyReg, highReg, iCursor);
highReg = highCopyReg;
}
}
addPreCondition(reg, regNum);
addPostCondition(reg, regNum);
if (copyReg != NULL)
cg->stopUsingRegister(copyReg);
else
regList.add(reg);
if (highReg)
{
addPreCondition(highReg, highRegNum);
addPostCondition(highReg, highRegNum);
if (highCopyReg != NULL)
cg->stopUsingRegister(highCopyReg);
else
regList.add(highReg);
}
}
if (iCursor!=NULL && cursorPtr!=NULL)
*cursorPtr = iCursor;
}
/*
* users should call the longSubtractAnalyser or longSubtractAnalyserWithExplicitOperands APIs instead of calling this one directly
*/
TR::Register* TR_X86SubtractAnalyser::longSubtractAnalyserImpl(TR::Node *root, TR::Node *&firstChild, TR::Node *&secondChild)
{
TR::Register *firstRegister = firstChild->getRegister();
TR::Register *secondRegister = secondChild->getRegister();
TR::Register *targetRegister = NULL;
bool firstHighZero = false;
bool secondHighZero = false; bool useSecondHighOrder = false;
TR_X86OpCodes regRegOpCode = SUB4RegReg;
TR_X86OpCodes regMemOpCode = SUB4RegMem;
bool needsEflags = NEED_CC(root) || (root->getOpCodeValue() == TR::lusubb);
// Can generate better code for long adds when one or more children have a high order zero word
// can avoid the evaluation when we don't need the result of such nodes for another parent.
//
if (firstChild->isHighWordZero() && !needsEflags)
{
firstHighZero = true;
}
if (secondChild->isHighWordZero() && !needsEflags)
{
secondHighZero = true;
TR::ILOpCodes secondOp = secondChild->getOpCodeValue();
if (secondChild->getReferenceCount() == 1 && secondRegister == 0)
{
if (secondOp == TR::iu2l || secondOp == TR::su2l ||
secondOp == TR::bu2l ||
(secondOp == TR::lushr &&
secondChild->getSecondChild()->getOpCodeValue() == TR::iconst &&
(secondChild->getSecondChild()->getInt() & TR::TreeEvaluator::shiftMask(true)) == 32))
{
secondChild = secondChild->getFirstChild();
secondRegister = secondChild->getRegister();
if (secondOp == TR::lushr)
{
useSecondHighOrder = true;
}
}
}
}
setInputs(firstChild, firstRegister, secondChild, secondRegister);
if (isVolatileMemoryOperand(firstChild))
resetMem1();
if (isVolatileMemoryOperand(secondChild))
resetMem2();
if (getEvalChild1())
{
firstRegister = _cg->evaluate(firstChild);
}
if (getEvalChild2())
{
secondRegister = _cg->evaluate(secondChild);
}
if (secondHighZero && secondRegister && secondRegister->getRegisterPair())
{
if (!useSecondHighOrder)
{
secondRegister = secondRegister->getLowOrder();
}
else
{
secondRegister = secondRegister->getHighOrder();
}
}
if (root->getOpCodeValue() == TR::lusubb &&
TR_X86ComputeCC::setCarryBorrow(root->getChild(2), true, _cg))
{
// use SBB rather than SUB
//
regRegOpCode = SBB4RegReg;
regMemOpCode = SBB4RegMem;
}
if (getCopyReg1())
{
TR::Register *lowThird = _cg->allocateRegister();
TR::Register *highThird = _cg->allocateRegister();
TR::RegisterPair *thirdReg = _cg->allocateRegisterPair(lowThird, highThird);
targetRegister = thirdReg;
generateRegRegInstruction(MOV4RegReg, root, lowThird, firstRegister->getLowOrder(), _cg);
if (firstHighZero)
{
generateRegRegInstruction(XOR4RegReg, root, highThird, highThird, _cg);
}
else
{
generateRegRegInstruction(MOV4RegReg, root, highThird, firstRegister->getHighOrder(), _cg);
}
if (getSubReg3Reg2())
{
if (secondHighZero)
{
generateRegRegInstruction(regRegOpCode, root, lowThird, secondRegister, _cg);
generateRegImmInstruction(SBB4RegImms, root, highThird, 0, _cg);
}
else
{
generateRegRegInstruction(regRegOpCode, root, lowThird, secondRegister->getLowOrder(), _cg);
generateRegRegInstruction(SBB4RegReg, root, highThird, secondRegister->getHighOrder(), _cg);
}
}
else // assert getSubReg3Mem2() == true
{
TR::MemoryReference *lowMR = generateX86MemoryReference(secondChild, _cg);
/**
* The below code is needed to ensure correct behaviour when the subtract analyser encounters a lushr bytecode that shifts
* by 32 bits. This is the only case where the useSecondHighOrder bit is set.
* When the first child of the lushr is in a register, code above handles the shift. When the first child of the lushr is in
* memory, the below ensures that the upper part of the first child of the lushr is used as lowMR.
*/
if (useSecondHighOrder)
{
TR_ASSERT(secondHighZero, "useSecondHighOrder should be consistent with secondHighZero. useSecondHighOrder subsumes secondHighZero");
lowMR = generateX86MemoryReference(*lowMR, 4, _cg);
}
generateRegMemInstruction(regMemOpCode, root, lowThird, lowMR, _cg);
if (secondHighZero)
{
generateRegImmInstruction(SBB4RegImms, root, highThird, 0, _cg);
}
else
{
TR::MemoryReference *highMR = generateX86MemoryReference(*lowMR, 4, _cg);
generateRegMemInstruction(SBB4RegMem, root, highThird, highMR, _cg);
}
lowMR->decNodeReferenceCounts(_cg);
}
}
else if (getSubReg1Reg2())
{
if (secondHighZero)
{
generateRegRegInstruction(regRegOpCode, root, firstRegister->getLowOrder(), secondRegister, _cg);
generateRegImmInstruction(SBB4RegImms, root, firstRegister->getHighOrder(), 0, _cg);
}
else
{
generateRegRegInstruction(regRegOpCode, root, firstRegister->getLowOrder(), secondRegister->getLowOrder(), _cg);
generateRegRegInstruction(SBB4RegReg, root, firstRegister->getHighOrder(), secondRegister->getHighOrder(), _cg);
}
targetRegister = firstRegister;
}
else // assert getSubReg1Mem2() == true
{
TR::MemoryReference *lowMR = generateX86MemoryReference(secondChild, _cg);
/**
* The below code is needed to ensure correct behaviour when the subtract analyser encounters a lushr bytecode that shifts
* by 32 bits. This is the only case where the useSecondHighOrder bit is set.
* When the first child of the lushr is in a register, code above handles the shift. When the first child of the lushr is in
* memory, the below ensures that the upper part of the first child of the lushr is used as lowMR.
*/
if (useSecondHighOrder)
lowMR = generateX86MemoryReference(*lowMR, 4, _cg);
generateRegMemInstruction(regMemOpCode, root, firstRegister->getLowOrder(), lowMR, _cg);
if (secondHighZero)
{
generateRegImmInstruction(SBB4RegImms, root, firstRegister->getHighOrder(), 0, _cg);
}
else
{
TR::MemoryReference *highMR = generateX86MemoryReference(*lowMR, 4, _cg);
generateRegMemInstruction(SBB4RegMem, root, firstRegister->getHighOrder(), highMR, _cg);
}
targetRegister = firstRegister;
lowMR->decNodeReferenceCounts(_cg);
}
return targetRegister;
}
TR::RegisterDependencyConditions* TR_PPCScratchRegisterDependencyConditions::createDependencyConditions(TR::CodeGenerator *cg,
TR_PPCScratchRegisterDependencyConditions *pre,
TR_PPCScratchRegisterDependencyConditions *post)
{
int32_t preCount = pre ? pre->getNumberOfDependencies() : 0;
int32_t postCount = post ? post->getNumberOfDependencies() : 0;
TR_LiveRegisters *lrVector = cg->getLiveRegisters(TR_VSX_VECTOR);
bool liveVSXVectorReg = (!lrVector || (lrVector->getNumberOfLiveRegisters() > 0));
TR_LiveRegisters *lrScalar = cg->getLiveRegisters(TR_VSX_SCALAR);
bool liveVSXScalarReg = (!lrScalar || (lrScalar->getNumberOfLiveRegisters() > 0));
if (liveVSXVectorReg)
{
preCount += 64;
postCount += 64;
}
else if (liveVSXScalarReg)
{
preCount += 32;
postCount += 32;
}
TR::RegisterDependencyConditions *dependencies = new (cg->trHeapMemory()) TR::RegisterDependencyConditions(preCount,
postCount,
cg->trMemory());
for (int i = 0; i < (pre ? pre->_numGPRDeps : 0); ++i)
{
dependencies->addPreCondition(pre->_gprDeps[i].getRegister(), pre->_gprDeps[i].getRealRegister(), pre->_gprDeps[i].getFlags());
if (pre->_excludeGPR0 & (1 << i))
dependencies->getPreConditions()->getRegisterDependency(i)->setExcludeGPR0();
}
for (int i = 0; i < (post ? post->_numGPRDeps : 0); ++i)
{
dependencies->addPostCondition(post->_gprDeps[i].getRegister(), post->_gprDeps[i].getRealRegister(), post->_gprDeps[i].getFlags());
if (post->_excludeGPR0 & (1 << i))
dependencies->getPostConditions()->getRegisterDependency(i)->setExcludeGPR0();
}
for (int i = 0; i < (pre ? pre->_numCCRDeps : 0); ++i)
{
dependencies->addPreCondition(pre->_ccrDeps[i].getRegister(), pre->_ccrDeps[i].getRealRegister(), pre->_ccrDeps[i].getFlags());
}
for (int i = 0; i < (post ? post->_numCCRDeps : 0); ++i)
{
dependencies->addPostCondition(post->_ccrDeps[i].getRegister(), post->_ccrDeps[i].getRealRegister(), post->_ccrDeps[i].getFlags());
}
const TR_PPCLinkageProperties& properties = cg->getLinkage()->getProperties();
if (liveVSXVectorReg)
{
for (int32_t i=TR::RealRegister::FirstVSR; i<=TR::RealRegister::LastVSR; i++)
{
if (!properties.getPreserved((TR::RealRegister::RegNum)i))
{
TR::Register *vreg = cg->allocateRegister(TR_FPR);
vreg->setPlaceholderReg();
dependencies->addPreCondition(vreg, (TR::RealRegister::RegNum)i);
dependencies->addPostCondition(vreg, (TR::RealRegister::RegNum)i);
}
}
}
else
{
if (liveVSXScalarReg)
{
for (int32_t i=TR::RealRegister::vsr32; i<=TR::RealRegister::LastVSR; i++)
{
if (!properties.getPreserved((TR::RealRegister::RegNum)i))
{
TR::Register *vreg = cg->allocateRegister(TR_FPR);
vreg->setPlaceholderReg();
dependencies->addPreCondition(vreg, (TR::RealRegister::RegNum)i);
dependencies->addPostCondition(vreg, (TR::RealRegister::RegNum)i);
}
}
}
}
return dependencies;
}
TR::Register *
TR::IA32SystemLinkage::buildVolatileAndReturnDependencies(
TR::Node *callNode,
TR::RegisterDependencyConditions *deps)
{
TR_ASSERT(deps != NULL, "expected register dependencies");
// Allocate virtual register for return value
//
TR::Register *integerReturnReg = NULL;
TR::Register *longReturnReg = NULL;
TR::Register *fpReturnReg = NULL;
TR::Register *returnReg = NULL; // An alias for one of the above
switch (callNode->getDataType())
{
case TR::NoType:
break;
case TR::Int8:
case TR::Int16:
case TR::Int32:
returnReg = integerReturnReg = cg()->allocateRegister();
break;
case TR::Address:
returnReg = integerReturnReg = cg()->allocateCollectedReferenceRegister();
break;
case TR::Float:
returnReg = fpReturnReg = cg()->allocateSinglePrecisionRegister(TR_X87);
break;
case TR::Double:
returnReg = fpReturnReg = cg()->allocateRegister(TR_X87);
break;
case TR::Int64:
returnReg = longReturnReg = (TR::Register*)cg()->allocateRegisterPair(cg()->allocateRegister(), cg()->allocateRegister());
break;
case TR::Aggregate:
default:
TR_ASSERT(false, "return type still not supported");
}
// Deps for volatile regs
//
// TODO: This should be less dependent on the real registers, but the way
// _properties is set up makes that very hard.
TR_ASSERT(_properties.getIntegerReturnRegister() == TR::RealRegister::eax, "assertion failure");
TR_ASSERT(_properties.getLongLowReturnRegister() == TR::RealRegister::eax, "assertion failure");
TR_ASSERT(_properties.getLongHighReturnRegister() == TR::RealRegister::edx, "assertion failure");
TR_ASSERT(_properties.getFloatReturnRegister() == TR::RealRegister::st0, "assertion failure");
if (longReturnReg)
{
deps->addPostCondition(returnReg->getLowOrder(), TR::RealRegister::eax, cg());
deps->addPostCondition(returnReg->getHighOrder(), TR::RealRegister::edx, cg());
}
else if (integerReturnReg)
{
deps->addPostCondition(returnReg, TR::RealRegister::eax, cg());
deps->addPostCondition(cg()->allocateRegister(), TR::RealRegister::edx, cg());
}
else
{
deps->addPostCondition(cg()->allocateRegister(), TR::RealRegister::eax, cg());
deps->addPostCondition(cg()->allocateRegister(), TR::RealRegister::edx, cg());
}
deps->addPostCondition(cg()->allocateRegister(), TR::RealRegister::ecx, cg());
// st0
if (fpReturnReg)
{
deps->addPostCondition(returnReg, _properties.getFloatReturnRegister(), cg());
}
else
{
// No need for a dummy dep here because FPREGSPILL instruction takes care of it
}
// The reg dependency is left open intentionally, and need to be closed by
// the caller. The reason is because, child class might call this method, while
// adding more register dependecies; if we close the reg dependency here,
// the child class could add NO more register dependencies.
return returnReg;
}
void TR_ARMRegisterDependencyGroup::assignRegisters(TR::Instruction *currentInstruction,
TR_RegisterKinds kindToBeAssigned,
uint32_t numberOfRegisters,
TR::CodeGenerator *cg)
{
TR::Compilation *comp = cg->comp();
TR::Machine *machine = cg->machine();
TR::Register *virtReg;
TR::RealRegister::RegNum dependentRegNum;
TR::RealRegister *dependentRealReg, *assignedRegister;
uint32_t i, j;
bool changed;
if (!comp->getOption(TR_DisableOOL))
{
for (i = 0; i< numberOfRegisters; i++)
{
virtReg = dependencies[i].getRegister();
dependentRegNum = dependencies[i].getRealRegister();
if (dependentRegNum == TR::RealRegister::SpilledReg)
{
TR_ASSERT(virtReg->getBackingStorage(),"should have a backing store if dependentRegNum == spillRegIndex()\n");
if (virtReg->getAssignedRealRegister())
{
// this happens when the register was first spilled in main line path then was reverse spilled
// and assigned to a real register in OOL path. We protected the backing store when doing
// the reverse spill so we could re-spill to the same slot now
traceMsg (comp,"\nOOL: Found register spilled in main line and re-assigned inside OOL");
TR::Node *currentNode = currentInstruction->getNode();
TR::RealRegister *assignedReg = toRealRegister(virtReg->getAssignedRegister());
TR::MemoryReference *tempMR = new (cg->trHeapMemory()) TR::MemoryReference(currentNode, (TR::SymbolReference*)virtReg->getBackingStorage()->getSymbolReference(), sizeof(uintptr_t), cg);
TR_ARMOpCodes opCode;
TR_RegisterKinds rk = virtReg->getKind();
switch (rk)
{
case TR_GPR:
opCode = ARMOp_ldr;
break;
case TR_FPR:
opCode = virtReg->isSinglePrecision() ? ARMOp_ldfs : ARMOp_ldfd;
break;
default:
TR_ASSERT(0, "\nRegister kind not supported in OOL spill\n");
break;
}
TR::Instruction *inst = generateTrg1MemInstruction(cg, opCode, currentNode, assignedReg, tempMR, currentInstruction);
assignedReg->setAssignedRegister(NULL);
virtReg->setAssignedRegister(NULL);
assignedReg->setState(TR::RealRegister::Free);
if (comp->getDebug())
cg->traceRegisterAssignment("Generate reload of virt %s due to spillRegIndex dep at inst %p\n", cg->comp()->getDebug()->getName(virtReg),currentInstruction);
cg->traceRAInstruction(inst);
}
if (!(std::find(cg->getSpilledRegisterList()->begin(), cg->getSpilledRegisterList()->end(), virtReg) != cg->getSpilledRegisterList()->end()))
cg->getSpilledRegisterList()->push_front(virtReg);
}
// we also need to free up all locked backing storage if we are exiting the OOL during backwards RA assignment
else if (currentInstruction->isLabel() && virtReg->getAssignedRealRegister())
{
TR::ARMLabelInstruction *labelInstr = (TR::ARMLabelInstruction *)currentInstruction;
TR_BackingStore *location = virtReg->getBackingStorage();
TR_RegisterKinds rk = virtReg->getKind();
int32_t dataSize;
if (labelInstr->getLabelSymbol()->isStartOfColdInstructionStream() && location)
{
traceMsg (comp,"\nOOL: Releasing backing storage (%p)\n", location);
if (rk == TR_GPR)
dataSize = TR::Compiler->om.sizeofReferenceAddress();
else
dataSize = 8;
location->setMaxSpillDepth(0);
cg->freeSpill(location,dataSize,0);
virtReg->setBackingStorage(NULL);
}
}
}
}
for (i = 0; i < numberOfRegisters; i++)
{
virtReg = dependencies[i].getRegister();
if (virtReg->getAssignedRealRegister()!=NULL)
{
if (dependencies[i].getRealRegister() == TR::RealRegister::NoReg)
{
virtReg->block();
}
else
{
dependentRegNum = toRealRegister(virtReg->getAssignedRealRegister())->getRegisterNumber();
for (j=0; j<numberOfRegisters; j++)
{
if (dependentRegNum == dependencies[j].getRealRegister())
{
virtReg->block();
break;
}
}
}
}
}
do
{
changed = false;
for (i = 0; i < numberOfRegisters; i++)
{
virtReg = dependencies[i].getRegister();
dependentRegNum = dependencies[i].getRealRegister();
dependentRealReg = machine->getRealRegister(dependentRegNum);
if (dependentRegNum != TR::RealRegister::NoReg &&
dependentRegNum != TR::RealRegister::SpilledReg &&
dependentRealReg->getState() == TR::RealRegister::Free)
{
machine->coerceRegisterAssignment(currentInstruction, virtReg, dependentRegNum);
virtReg->block();
changed = true;
}
}
} while (changed == true);
do
{
changed = false;
for (i = 0; i < numberOfRegisters; i++)
{
virtReg = dependencies[i].getRegister();
assignedRegister = NULL;
if (virtReg->getAssignedRealRegister() != NULL)
{
assignedRegister = toRealRegister(virtReg->getAssignedRealRegister());
}
dependentRegNum = dependencies[i].getRealRegister();
dependentRealReg = machine->getRealRegister(dependentRegNum);
if (dependentRegNum != TR::RealRegister::NoReg &&
dependentRegNum != TR::RealRegister::SpilledReg &&
dependentRealReg != assignedRegister)
{
machine->coerceRegisterAssignment(currentInstruction, virtReg, dependentRegNum);
virtReg->block();
changed = true;
}
}
} while (changed == true);
for (i=0; i<numberOfRegisters; i++)
{
if (dependencies[i].getRealRegister() == TR::RealRegister::NoReg)
{
bool excludeGPR0 = dependencies[i].getExcludeGPR0()?true:false;
TR::RealRegister *realOne;
virtReg = dependencies[i].getRegister();
realOne = virtReg->getAssignedRealRegister();
if (realOne!=NULL && excludeGPR0 && toRealRegister(realOne)->getRegisterNumber()==TR::RealRegister::gr0)
{
if ((assignedRegister = machine->findBestFreeRegister(virtReg->getKind(), true)) == NULL)
{
assignedRegister = machine->freeBestRegister(currentInstruction, virtReg->getKind(), NULL, true);
}
machine->coerceRegisterAssignment(currentInstruction, virtReg, assignedRegister->getRegisterNumber());
}
else if (realOne == NULL)
{
if (virtReg->getTotalUseCount() == virtReg->getFutureUseCount())
{
if ((assignedRegister = machine->findBestFreeRegister(virtReg->getKind(), excludeGPR0, true)) == NULL)
{
assignedRegister = machine->freeBestRegister(currentInstruction, virtReg->getKind(), NULL, excludeGPR0);
}
}
else
{
assignedRegister = machine->reverseSpillState(currentInstruction, virtReg, NULL, excludeGPR0);
}
virtReg->setAssignedRegister(assignedRegister);
assignedRegister->setAssignedRegister(virtReg);
assignedRegister->setState(TR::RealRegister::Assigned);
virtReg->block();
}
}
}
unblockRegisters(numberOfRegisters);
for (i = 0; i < numberOfRegisters; i++)
{
TR::Register *dependentRegister = getRegisterDependency(i)->getRegister();
if (dependentRegister->getAssignedRegister())
{
TR::RealRegister *assignedRegister = dependentRegister->getAssignedRegister()->getRealRegister();
if (getRegisterDependency(i)->getRealRegister() == TR::RealRegister::NoReg)
getRegisterDependency(i)->setRealRegister(toRealRegister(assignedRegister)->getRegisterNumber());
if (dependentRegister->decFutureUseCount() == 0)
{
dependentRegister->setAssignedRegister(NULL);
assignedRegister->setAssignedRegister(NULL);
assignedRegister->setState(TR::RealRegister::Unlatched); // Was setting to Free
}
}
}
}