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;
}
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;
}
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_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;
}
}
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;
}
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;
}
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
}
}
}
}