static bool safeToMoveGuard(TR::Block *destination, TR::TreeTop *guardCandidate,
TR::TreeTop *branchDest, TR_BitVector &privArgSymRefs)
{
static char *disablePrivArgMovement = feGetEnv("TR_DisableRuntimeGuardPrivArgMovement");
TR::TreeTop *start = destination ? destination->getExit() : TR::comp()->getStartTree();
if (guardCandidate->getNode()->isHCRGuard())
{
for (TR::TreeTop *tt = start; tt && tt != guardCandidate; tt = tt->getNextTreeTop())
{
if (tt->getNode()->canGCandReturn())
return false;
}
}
else if (guardCandidate->getNode()->isOSRGuard())
{
for (TR::TreeTop *tt = start; tt && tt != guardCandidate; tt = tt->getNextTreeTop())
{
if (TR::comp()->isPotentialOSRPoint(tt->getNode(), NULL, true))
return false;
}
}
else
{
privArgSymRefs.empty();
for (TR::TreeTop *tt = start; tt && tt != guardCandidate; tt = tt->getNextTreeTop())
{
// It's safe to move the guard if there are only priv arg stores and live monitor stores
// ahead of the guard
if (tt->getNode()->getOpCodeValue() != TR::BBStart
&& tt->getNode()->getOpCodeValue() != TR::BBEnd
&& !tt->getNode()->chkIsPrivatizedInlinerArg()
&& !(tt->getNode()->getOpCode().hasSymbolReference() && tt->getNode()->getSymbol()->holdsMonitoredObject())
&& !tt->getNode()->isNopableInlineGuard())
return false;
if (tt->getNode()->chkIsPrivatizedInlinerArg() && (disablePrivArgMovement ||
// If the priv arg is not for this guard
(guardCandidate->getNode()->getInlinedSiteIndex() > -1 &&
// if priv arg store does not have the same inlined site index as the guard's caller, that means it is not a priv arg for this guard,
// then we cannot move the guard and its priv args up across other calls' priv args
tt->getNode()->getInlinedSiteIndex() != TR::comp()->getInlinedCallSite(guardCandidate->getNode()->getInlinedSiteIndex())._byteCodeInfo.getCallerIndex())))
return false;
if (tt->getNode()->chkIsPrivatizedInlinerArg())
privArgSymRefs.set(tt->getNode()->getSymbolReference()->getReferenceNumber());
if (tt->getNode()->isNopableInlineGuard()
&& tt->getNode()->getBranchDestination() != branchDest)
return false;
}
}
return true;
}
TR::Instruction *
TR_X86SystemLinkage::savePreservedRegisters(TR::Instruction *cursor)
{
// For IA32, if disableShrinkWrapping, usePushForPreservedRegs will be true; otherwise false;
// For X64, shrinkWraping is always on, and usePushForPreservedRegs always false;
// TR_ASSERT(!getProperties().getUsesPushesForPreservedRegs(), "assertion failure");
TR::ResolvedMethodSymbol *bodySymbol = comp()->getJittedMethodSymbol();
const int32_t localSize = getProperties().getOffsetToFirstLocal() - bodySymbol->getLocalMappingCursor();
const int32_t pointerSize = getProperties().getPointerSize();
int32_t offsetCursor = -localSize + getProperties().getOffsetToFirstLocal() - pointerSize;
if (_properties.getUsesPushesForPreservedRegs())
{
for (int32_t pindex = _properties.getMaxRegistersPreservedInPrologue()-1;
pindex >= 0;
pindex--)
{
TR::RealRegister::RegNum idx = _properties.getPreservedRegister((uint32_t)pindex);
TR::RealRegister *reg = machine()->getX86RealRegister(idx);
if (reg->getHasBeenAssignedInMethod() && reg->getState() != TR::RealRegister::Locked)
{
cursor = new (trHeapMemory()) TR::X86RegInstruction(cursor, PUSHReg, reg, cg());
}
}
}
else
{
TR_BitVector *p = cg()->getPreservedRegsInPrologue();
for (int32_t pindex = getProperties().getMaxRegistersPreservedInPrologue()-1;
pindex >= 0;
pindex--)
{
TR::RealRegister::RegNum idx = _properties.getPreservedRegister((uint32_t)pindex);
TR::RealRegister *reg = machine()->getX86RealRegister(getProperties().getPreservedRegister((uint32_t)pindex));
if(reg->getHasBeenAssignedInMethod() && reg->getState() != TR::RealRegister::Locked)
{
if (!p || p->get(idx))
{
cursor = generateMemRegInstruction(
cursor,
movOpcodes[MemReg][fullRegisterMovType(reg)],
generateX86MemoryReference(machine()->getX86RealRegister(TR::RealRegister::vfp), offsetCursor, cg()),
reg,
cg()
);
}
offsetCursor -= pointerSize;
}
}
}
return cursor;
}
bool TR_DominatorVerifier::isExpensiveAlgorithmCorrect(TR_DominatorsChk &expensiveAlgorithm)
{
int32_t i,j;
_nodesSeenOnEveryPath = new (trStackMemory()) TR_BitVector(_numBlocks,trMemory(), stackAlloc);
_nodesSeenOnCurrentPath = new (trStackMemory()) TR_BitVector(_numBlocks,trMemory(), stackAlloc);
_dominatorsChkInfo = expensiveAlgorithm.getDominatorsChkInfo();
for (i = 2; i < _numBlocks-1; i++)
{
TR_BitVector *bucket = _dominatorsChkInfo[i]._tmpbucket;
for (j = 0; j < _numBlocks-1; j++)
{
if (bucket->get(j)) // dominator according to algorithm
{
// Initializing these BitVectors before checking
// dominators for the next block.
// The last bit is not changed for either bit vector - is that what
// was intended?
//
_nodesSeenOnEveryPath->setAll(_numBlocks-1);
int32_t lastBit = _nodesSeenOnCurrentPath->get(_numBlocks-1);
_nodesSeenOnCurrentPath->empty();
if (lastBit)
_nodesSeenOnCurrentPath->set(_numBlocks-1);
if ( ! dominates(_dominatorsChkInfo[j+1]._block,_dominatorsChkInfo[i]._block) ) // dominator according to the CFG
{
if (debug("traceVER"))
{
dumpOptDetails(comp(), " Dominator info for expensive algorithm is incorrect \n");
dumpOptDetails(comp(), " Dominator of [%p] is [%p] as per the algorithm\n", _dominatorsChkInfo[i]._block, _dominatorsChkInfo[j+1]._block);
dumpOptDetails(comp(), " But [%p] is not an the dominator of [%p] as per the Control Flow Graph", _dominatorsChkInfo[j+1]._block, _dominatorsChkInfo[i]._block);
}
return false;
}
}
}
}
return true;
}
/**
* Collect direct loads in a node and its children, adding them to the provided BitVector.
*
* @param node The node to consider.
* @param loadSymRefs BitVector of symbol reference numbers seen.
* @param checklist Checklist of visited nodes.
*/
static void collectDirectLoads(TR::Node *node, TR_BitVector &loadSymRefs, TR::NodeChecklist &checklist)
{
if (checklist.contains(node))
return;
checklist.add(node);
if (node->getOpCode().isLoadVarDirect())
loadSymRefs.set(node->getSymbolReference()->getReferenceNumber());
for (int i = 0; i < node->getNumChildren(); i++)
collectDirectLoads(node->getChild(i), loadSymRefs, checklist);
}
/**
* Search for direct loads in the taken side of a guard
*
* @param firstBlock The guard's branch destination
* @param coldPathLoads BitVector of symbol reference numbers for any direct loads seen until the merge back to mainline
*/
static void collectColdPathLoads(TR::Block* firstBlock, TR_BitVector &coldPathLoads)
{
TR_Stack<TR::Block*> blocksToCheck(TR::comp()->trMemory(), 8, false, stackAlloc);
blocksToCheck.push(firstBlock);
TR::NodeChecklist checklist(TR::comp());
coldPathLoads.empty();
while (!blocksToCheck.isEmpty())
{
TR::Block *block = blocksToCheck.pop();
for (TR::TreeTop *tt = block->getFirstRealTreeTop(); tt->getNode()->getOpCodeValue() != TR::BBEnd; tt = tt->getNextTreeTop())
collectDirectLoads(tt->getNode(), coldPathLoads, checklist);
// Search for any successors that have not merged with the mainline
for (auto itr = block->getSuccessors().begin(), end = block->getSuccessors().end(); itr != end; ++itr)
{
TR::Block *dest = (*itr)->getTo()->asBlock();
if (dest != TR::comp()->getFlowGraph()->getEnd() && dest->getPredecessors().size() == 1)
blocksToCheck.push(dest);
}
}
}
TR_BitVector *
OMR::SymbolReference::getUseDefAliasesBV(bool isDirectCall, bool includeGCSafePoint)
{
TR::Compilation *comp = TR::comp();
TR::Region &aliasRegion = comp->aliasRegion();
int32_t bvInitialSize = comp->getSymRefCount();
TR_BitVectorGrowable growability = growable;
// allow more than one shadow for an array type. Used by LoopAliasRefiner
const bool supportArrayRefinement=true;
int32_t kind = _symbol->getKind();
TR::SymbolReferenceTable * symRefTab = comp->getSymRefTab();
// !!! NOTE !!!
// THERE IS A COPY OF THIS LOGIC IN sharesSymbol
//
if (!self()->reallySharesSymbol(comp))
{
switch (kind)
{
case TR::Symbol::IsShadow:
case TR::Symbol::IsStatic:
{
// For unresolved constant dynamic, we need to invoke a Java bootstrap method,
// which can have arbitrary side effects, so the aliasing should be conservative here.
// isConstObjectRef now returns true for condy, so we add an explicit condition,
// more like a short-circuit, to say if we are unresolved and not isConstObjectRef
// (this is the same as before), or if we are unresolved and condy
// (this is the extra condition added), we would return conservative aliases.
if ((self()->isUnresolved() && (_symbol->isConstantDynamic() || !_symbol->isConstObjectRef())) ||
_symbol->isVolatile() || self()->isLiteralPoolAddress() ||
self()->isFromLiteralPool() || _symbol->isUnsafeShadowSymbol() ||
(_symbol->isArrayShadowSymbol() && comp->getMethodSymbol()->hasVeryRefinedAliasSets()))
{
// getUseDefAliases might not return NULL
}
else if (!symRefTab->aliasBuilder.mutableGenericIntShadowHasBeenCreated())
{
// getUseDefAliases must return NULL
return NULL;
}
else if (kind == TR::Symbol::IsStatic && !symRefTab->aliasBuilder.litPoolGenericIntShadowHasBeenCreated())
{
// getUseDefAliases must return NULL
return NULL;
}
break;
}
}
}
// now do stuff for various kinds of symbols
//
switch (kind)
{
case TR::Symbol::IsMethod:
{
TR::MethodSymbol * methodSymbol = _symbol->castToMethodSymbol();
if (!methodSymbol->isHelper())
return symRefTab->aliasBuilder.methodAliases(self());
if (symRefTab->isNonHelper(self(), TR::SymbolReferenceTable::arraySetSymbol) ||
symRefTab->isNonHelper(self(), TR::SymbolReferenceTable::osrFearPointHelperSymbol) ||
symRefTab->isNonHelper(self(), TR::SymbolReferenceTable::potentialOSRPointHelperSymbol))
{
return &symRefTab->aliasBuilder.defaultMethodDefAliases();
}
if (symRefTab->isNonHelper(self(), TR::SymbolReferenceTable::arrayCmpSymbol))
return 0;
switch (self()->getReferenceNumber())
{
case TR_methodTypeCheck:
case TR_nullCheck:
return &symRefTab->aliasBuilder.defaultMethodDefAliasesWithoutImmutable();
case TR_arrayBoundsCheck:
case TR_checkCast:
case TR_divCheck:
case TR_typeCheckArrayStore:
case TR_arrayStoreException:
case TR_incompatibleReceiver:
case TR_IncompatibleClassChangeError:
case TR_reportFinalFieldModified:
case TR_reportMethodEnter:
case TR_reportStaticMethodEnter:
case TR_reportMethodExit:
case TR_acquireVMAccess:
case TR_instanceOf:
case TR_checkAssignable:
case TR_throwCurrentException:
case TR_releaseVMAccess:
case TR_stackOverflow:
case TR_writeBarrierStore:
case TR_writeBarrierBatchStore:
case TR_jitProfileAddress:
case TR_jitProfileWarmCompilePICAddress:
case TR_jitProfileValue:
case TR_jitProfileLongValue:
case TR_jitProfileBigDecimalValue:
case TR_jitProfileParseBuffer:
return 0;
case TR_asyncCheck:
case TR_writeBarrierClassStoreRealTimeGC:
case TR_writeBarrierStoreRealTimeGC:
case TR_aNewArray:
case TR_newObject:
case TR_newObjectNoZeroInit:
case TR_newArray:
case TR_multiANewArray:
if ((comp->generateArraylets() || comp->isDLT()) && includeGCSafePoint)
return &symRefTab->aliasBuilder.gcSafePointSymRefNumbers();
else
return 0;
case TR_aThrow:
return 0;
// The monitor exit symbol needs to be aliased with all fields in the
// current class to ensure that all references to fields are evaluated
// before the monitor exit
case TR_monitorExit:
case TR_monitorEntry:
case TR_transactionExit:
case TR_transactionEntry:
default:
// The following is the place to check for
// a use of killsAllMethodSymbolRef... However,
// it looks like the default action is sufficient.
//if (symRefTab->findKillsAllMethodSymbolRef() == self())
// {
// }
return &symRefTab->aliasBuilder.defaultMethodDefAliases();
}
}
case TR::Symbol::IsResolvedMethod:
{
TR::ResolvedMethodSymbol * resolvedMethodSymbol = _symbol->castToResolvedMethodSymbol();
if (!comp->getOption(TR_EnableHCR))
{
switch (resolvedMethodSymbol->getRecognizedMethod())
{
#ifdef J9_PROJECT_SPECIFIC
case TR::java_lang_System_arraycopy:
{
TR_BitVector * aliases = new (aliasRegion) TR_BitVector(bvInitialSize, aliasRegion, growability);
*aliases |= symRefTab->aliasBuilder.arrayElementSymRefs();
if (comp->generateArraylets())
*aliases |= symRefTab->aliasBuilder.arrayletElementSymRefs();
return aliases;
}
if (resolvedMethodSymbol->isPureFunction())
return NULL;
case TR::java_lang_Double_longBitsToDouble:
case TR::java_lang_Double_doubleToLongBits:
case TR::java_lang_Float_intBitsToFloat:
case TR::java_lang_Float_floatToIntBits:
case TR::java_lang_Double_doubleToRawLongBits:
case TR::java_lang_Float_floatToRawIntBits:
case TR::java_lang_Math_sqrt:
case TR::java_lang_StrictMath_sqrt:
case TR::java_lang_Math_sin:
case TR::java_lang_StrictMath_sin:
case TR::java_lang_Math_cos:
case TR::java_lang_StrictMath_cos:
case TR::java_lang_Math_max_I:
case TR::java_lang_Math_min_I:
case TR::java_lang_Math_max_L:
case TR::java_lang_Math_min_L:
case TR::java_lang_Math_abs_I:
case TR::java_lang_Math_abs_L:
case TR::java_lang_Math_abs_F:
case TR::java_lang_Math_abs_D:
case TR::java_lang_Math_pow:
case TR::java_lang_StrictMath_pow:
case TR::java_lang_Math_exp:
case TR::java_lang_StrictMath_exp:
case TR::java_lang_Math_log:
case TR::java_lang_StrictMath_log:
case TR::java_lang_Math_floor:
case TR::java_lang_Math_ceil:
case TR::java_lang_Math_copySign_F:
case TR::java_lang_Math_copySign_D:
case TR::java_lang_StrictMath_floor:
case TR::java_lang_StrictMath_ceil:
case TR::java_lang_StrictMath_copySign_F:
case TR::java_lang_StrictMath_copySign_D:
case TR::com_ibm_Compiler_Internal__TR_Prefetch:
case TR::java_nio_Bits_keepAlive:
if ((comp->generateArraylets() || comp->isDLT()) && includeGCSafePoint)
return &symRefTab->aliasBuilder.gcSafePointSymRefNumbers();
else
return 0;
// no aliasing on DFP dummy stubs
case TR::java_math_BigDecimal_DFPPerformHysteresis:
case TR::java_math_BigDecimal_DFPUseDFP:
case TR::java_math_BigDecimal_DFPHWAvailable:
case TR::java_math_BigDecimal_DFPCompareTo:
case TR::java_math_BigDecimal_DFPUnscaledValue:
case TR::com_ibm_dataaccess_DecimalData_DFPFacilityAvailable:
case TR::com_ibm_dataaccess_DecimalData_DFPUseDFP:
case TR::com_ibm_dataaccess_DecimalData_DFPConvertPackedToDFP:
case TR::com_ibm_dataaccess_DecimalData_DFPConvertDFPToPacked:
case TR::com_ibm_dataaccess_DecimalData_createZeroBigDecimal:
case TR::com_ibm_dataaccess_DecimalData_getlaside:
case TR::com_ibm_dataaccess_DecimalData_setlaside:
case TR::com_ibm_dataaccess_DecimalData_getflags:
case TR::com_ibm_dataaccess_DecimalData_setflags:
if (!(
#ifdef TR_TARGET_S390
TR::Compiler->target.cpu.getS390SupportsDFP() ||
#endif
TR::Compiler->target.cpu.supportsDecimalFloatingPoint()) ||
comp->getOption(TR_DisableDFP))
return NULL;
#endif //J9_PROJECT_SPECIFIC
default:
break;
}
}
#ifdef J9_PROJECT_SPECIFIC
TR_ResolvedMethod * method = resolvedMethodSymbol->getResolvedMethod();
TR_PersistentMethodInfo * methodInfo = TR_PersistentMethodInfo::get(method);
if (methodInfo && (methodInfo->hasRefinedAliasSets() ||
comp->getMethodHotness() >= veryHot ||
resolvedMethodSymbol->hasVeryRefinedAliasSets()) &&
(method->isStatic() || method->isFinal() || isDirectCall))
{
TR_BitVector * aliases = new (aliasRegion) TR_BitVector(bvInitialSize, aliasRegion, growability);
if ((comp->generateArraylets() || comp->isDLT()) && includeGCSafePoint)
*aliases |= symRefTab->aliasBuilder.gcSafePointSymRefNumbers();
if (methodInfo->doesntKillAnything() && !comp->getOption(TR_DisableRefinedAliases))
return aliases;
if ((resolvedMethodSymbol->hasVeryRefinedAliasSets() || comp->getMethodHotness() >= hot) &&
!debug("disableVeryRefinedCallAliasSets"))
{
TR_BitVector * exactAliases = 0;
if (resolvedMethodSymbol->hasVeryRefinedAliasSets())
exactAliases = symRefTab->aliasBuilder.getVeryRefinedCallAliasSets(resolvedMethodSymbol);
else
{
resolvedMethodSymbol->setHasVeryRefinedAliasSets(true);
List<void> methodsPeeked(comp->trMemory());
exactAliases = addVeryRefinedCallAliasSets(resolvedMethodSymbol, aliases, &methodsPeeked);
symRefTab->aliasBuilder.setVeryRefinedCallAliasSets(resolvedMethodSymbol, exactAliases);
}
if (exactAliases)
{
return exactAliases;
}
}
// From here on, we're just checking refined alias info.
// If refined aliases are disabled, return the conservative answer
// we would have returned had we never attempted to use refined
// aliases at all.
//
if (comp->getOption(TR_DisableRefinedAliases))
return symRefTab->aliasBuilder.methodAliases(self());
if (!methodInfo->doesntKillAddressArrayShadows())
{
symRefTab->aliasBuilder.addAddressArrayShadows(aliases);
if (comp->generateArraylets())
aliases->set(symRefTab->getArrayletShadowIndex(TR::Address));
}
if (!methodInfo->doesntKillIntArrayShadows())
{
symRefTab->aliasBuilder.addIntArrayShadows(aliases);
if (comp->generateArraylets())
{
aliases->set(symRefTab->getArrayletShadowIndex(TR::Int32));
}
}
if (!methodInfo->doesntKillNonIntPrimitiveArrayShadows())
{
symRefTab->aliasBuilder.addNonIntPrimitiveArrayShadows(aliases);
if (comp->generateArraylets())
{
aliases->set(symRefTab->getArrayletShadowIndex(TR::Int8));
aliases->set(symRefTab->getArrayletShadowIndex(TR::Int16));
aliases->set(symRefTab->getArrayletShadowIndex(TR::Int32));
aliases->set(symRefTab->getArrayletShadowIndex(TR::Int64));
aliases->set(symRefTab->getArrayletShadowIndex(TR::Float));
aliases->set(symRefTab->getArrayletShadowIndex(TR::Double));
}
}
if (!methodInfo->doesntKillAddressFields())
*aliases |= symRefTab->aliasBuilder.addressShadowSymRefs();
if (!methodInfo->doesntKillIntFields())
*aliases |= symRefTab->aliasBuilder.intShadowSymRefs();
if (!methodInfo->doesntKillNonIntPrimitiveFields())
*aliases |= symRefTab->aliasBuilder.nonIntPrimitiveShadowSymRefs();
if (!methodInfo->doesntKillAddressStatics())
*aliases |= symRefTab->aliasBuilder.addressStaticSymRefs();
if (!methodInfo->doesntKillIntStatics())
*aliases |= symRefTab->aliasBuilder.intStaticSymRefs();
if (!methodInfo->doesntKillNonIntPrimitiveStatics())
*aliases |= symRefTab->aliasBuilder.nonIntPrimitiveStaticSymRefs();
TR_BitVector *methodAliases = symRefTab->aliasBuilder.methodAliases(self());
*aliases &= *methodAliases;
return aliases;
}
#endif
return symRefTab->aliasBuilder.methodAliases(self());
}
case TR::Symbol::IsShadow:
{
if ((self()->isUnresolved() && !_symbol->isConstObjectRef()) || _symbol->isVolatile() || self()->isLiteralPoolAddress() || self()->isFromLiteralPool() ||
(_symbol->isUnsafeShadowSymbol() && !self()->reallySharesSymbol()))
{
if (symRefTab->aliasBuilder.unsafeArrayElementSymRefs().get(self()->getReferenceNumber()))
{
TR_BitVector *aliases = new (aliasRegion) TR_BitVector(bvInitialSize, aliasRegion, growability);
*aliases |= comp->getSymRefTab()->aliasBuilder.defaultMethodDefAliasesWithoutImmutable();
*aliases -= symRefTab->aliasBuilder.cpSymRefs();
return aliases;
}
else
return &comp->getSymRefTab()->aliasBuilder.defaultMethodDefAliasesWithoutImmutable();
}
TR_BitVector *aliases = NULL;
if (_symbol == symRefTab->findGenericIntShadowSymbol())
{
aliases = new (aliasRegion) TR_BitVector(bvInitialSize, aliasRegion, growability);
*aliases |= symRefTab->aliasBuilder.arrayElementSymRefs();
if (comp->generateArraylets())
*aliases |= symRefTab->aliasBuilder.arrayletElementSymRefs();
*aliases |= symRefTab->aliasBuilder.genericIntShadowSymRefs();
*aliases |= symRefTab->aliasBuilder.genericIntArrayShadowSymRefs();
*aliases |= symRefTab->aliasBuilder.genericIntNonArrayShadowSymRefs();
*aliases |= symRefTab->aliasBuilder.unsafeSymRefNumbers();
#ifdef J9_PROJECT_SPECIFIC
*aliases |= symRefTab->aliasBuilder.unresolvedShadowSymRefs();
#endif
if (symRefTab->aliasBuilder.conservativeGenericIntShadowAliasing())
{
*aliases |= symRefTab->aliasBuilder.addressShadowSymRefs();
*aliases |= symRefTab->aliasBuilder.intShadowSymRefs();
*aliases |= symRefTab->aliasBuilder.nonIntPrimitiveShadowSymRefs();
}
aliases->set(self()->getReferenceNumber());
return aliases;
}
if (self()->reallySharesSymbol(comp))
{
aliases = new (aliasRegion) TR_BitVector(bvInitialSize, aliasRegion, growability);
self()->setSharedShadowAliases(aliases, symRefTab);
}
if (symRefTab->findGenericIntShadowSymbol())
{
if (!aliases)
aliases = new (aliasRegion) TR_BitVector(bvInitialSize, aliasRegion, growability);
self()->setLiteralPoolAliases(aliases, symRefTab);
if (symRefTab->aliasBuilder.conservativeGenericIntShadowAliasing() || self()->isUnresolved())
{
*aliases |= symRefTab->aliasBuilder.genericIntShadowSymRefs();
*aliases |= symRefTab->aliasBuilder.genericIntArrayShadowSymRefs();
*aliases |= symRefTab->aliasBuilder.genericIntNonArrayShadowSymRefs();
}
}
if (_symbol->isArrayShadowSymbol() &&
symRefTab->findGenericIntShadowSymbol())
{
if (!aliases)
aliases = new (aliasRegion) TR_BitVector(bvInitialSize, aliasRegion, growability);
*aliases |= symRefTab->aliasBuilder.genericIntShadowSymRefs();
*aliases |= symRefTab->aliasBuilder.genericIntArrayShadowSymRefs();
if (supportArrayRefinement && self()->getIndependentSymRefs())
*aliases -= *self()->getIndependentSymRefs();
}
#ifdef J9_PROJECT_SPECIFIC
// make TR::PackedDecimal aliased with TR::Int8(byte)
if (_symbol->isArrayShadowSymbol() && _symbol->getDataType() == TR::PackedDecimal)
{
if (!aliases)
aliases = new (aliasRegion) TR_BitVector(bvInitialSize, aliasRegion, growability);
aliases->set(symRefTab->getArrayShadowIndex(TR::Int8));
}
//the other way around.
if (_symbol->isArrayShadowSymbol() && _symbol->getDataType() == TR::Int8)
{
if (!aliases)
aliases = new (aliasRegion) TR_BitVector(bvInitialSize, aliasRegion, growability);
aliases->set(symRefTab->getArrayShadowIndex(TR::PackedDecimal));
}
#endif
// alias vector arrays shadows with corresponding scalar array shadows
if (_symbol->isArrayShadowSymbol() && _symbol->getDataType().isVector())
{
if (!aliases)
aliases = new (aliasRegion) TR_BitVector(bvInitialSize, aliasRegion, growability);
aliases->set(symRefTab->getArrayShadowIndex(_symbol->getDataType().vectorToScalar()));
}
// the other way around
if (_symbol->isArrayShadowSymbol() && !_symbol->getDataType().isVector())
{
if (!aliases)
aliases = new (aliasRegion) TR_BitVector(bvInitialSize, aliasRegion, growability);
aliases->set(symRefTab->getArrayShadowIndex(_symbol->getDataType().scalarToVector()));
}
if (_symbol->isArrayShadowSymbol() &&
!symRefTab->aliasBuilder.immutableArrayElementSymRefs().isEmpty())
{
if (!aliases)
aliases = new (aliasRegion) TR_BitVector(bvInitialSize, aliasRegion, growability);
TR::DataType type = _symbol->getDataType();
TR_BitVectorIterator bvi(symRefTab->aliasBuilder.arrayElementSymRefs());
int32_t symRefNum;
while (bvi.hasMoreElements())
{
symRefNum = bvi.getNextElement();
if (symRefTab->getSymRef(symRefNum)->getSymbol()->getDataType() == type)
aliases->set(symRefNum);
}
}
if (_symbol->isArrayShadowSymbol() &&
supportArrayRefinement &&
comp->getMethodSymbol()->hasVeryRefinedAliasSets())
{
if (!aliases)
aliases = new (aliasRegion) TR_BitVector(bvInitialSize, aliasRegion, growability);
TR::DataType type = _symbol->getDataType();
TR_BitVectorIterator bvi(symRefTab->aliasBuilder.arrayElementSymRefs());
int32_t symRefNum;
while (bvi.hasMoreElements())
{
symRefNum = bvi.getNextElement();
if (symRefTab->getSymRef(symRefNum)->getSymbol()->getDataType() == type)
aliases->set(symRefNum);
}
if (self()->getIndependentSymRefs())
*aliases -= *self()->getIndependentSymRefs();
return aliases;
}
if (aliases)
aliases->set(self()->getReferenceNumber());
if (symRefTab->aliasBuilder.unsafeArrayElementSymRefs().get(self()->getReferenceNumber()))
*aliases -= symRefTab->aliasBuilder.cpSymRefs();
else if (symRefTab->aliasBuilder.cpSymRefs().get(self()->getReferenceNumber()))
*aliases -= symRefTab->aliasBuilder.unsafeArrayElementSymRefs();
return aliases;
}
case TR::Symbol::IsStatic:
{
// For unresolved constant dynamic, we need to invoke a Java bootstrap method,
// which can have arbitrary side effects, so the aliasing should be conservative here.
// isConstObjectRef now returns true for condy, so we add an explicit condition,
// more like a short-circuit, to say if we are unresolved and not isConstObjectRef
// (this is the same as before), or if we are unresolved and condy
// (this is the extra condition added), we would return conservative aliases.
if ((self()->isUnresolved() && (_symbol->isConstantDynamic() || !_symbol->isConstObjectRef())) ||
self()->isLiteralPoolAddress() || self()->isFromLiteralPool() || _symbol->isVolatile())
{
return &comp->getSymRefTab()->aliasBuilder.defaultMethodDefAliases();
}
TR_BitVector *aliases = NULL;
if (self()->reallySharesSymbol(comp))
{
aliases = new (aliasRegion) TR_BitVector(bvInitialSize, aliasRegion, growability);
self()->setSharedStaticAliases(aliases, symRefTab);
}
if (symRefTab->findGenericIntShadowSymbol())
{
if (!aliases)
aliases = new (aliasRegion) TR_BitVector(bvInitialSize, aliasRegion, growability);
self()->setLiteralPoolAliases(aliases, symRefTab);
}
if (aliases)
aliases->set(self()->getReferenceNumber());
return aliases;
}
case TR::Symbol::IsMethodMetaData:
{
TR_BitVector *aliases = NULL;
return aliases;
}
default:
//TR_ASSERT(0, "getUseDefAliasing called for non method");
if (comp->generateArraylets() && comp->getSymRefTab()->aliasBuilder.gcSafePointSymRefNumbers().get(self()->getReferenceNumber()) && includeGCSafePoint)
return &comp->getSymRefTab()->aliasBuilder.gcSafePointSymRefNumbers();
else
return 0;
}
}
void TR_ReachingDefinitions::initializeGenAndKillSetInfoForNode(TR::Node *node, TR_BitVector &defsKilled, bool seenException, int32_t blockNum, TR::Node *parent)
{
// Update gen and kill info for nodes in this subtree
//
int32_t i;
if (node->getVisitCount() == comp()->getVisitCount())
return;
node->setVisitCount(comp()->getVisitCount());
// Process the children first
//
for (i = node->getNumChildren()-1; i >= 0; --i)
{
initializeGenAndKillSetInfoForNode(node->getChild(i), defsKilled, seenException, blockNum, node);
}
bool irrelevantStore = false;
scount_t nodeIndex = node->getLocalIndex();
if (nodeIndex <= 0)
{
if (node->getOpCode().isStore() &&
node->getSymbol()->isAutoOrParm() &&
node->storedValueIsIrrelevant())
{
irrelevantStore = true;
}
else
return;
}
bool foundDefsToKill = false;
int32_t numDefNodes = 0;
defsKilled.empty();
TR::ILOpCode &opCode = node->getOpCode();
TR::SymbolReference *symRef;
TR::Symbol *sym;
uint16_t symIndex;
uint32_t num_aliases;
if (_useDefInfo->_useDefForRegs &&
(opCode.isLoadReg() ||
opCode.isStoreReg()))
{
sym = NULL;
symRef = NULL;
symIndex = _useDefInfo->getNumSymbols() + node->getGlobalRegisterNumber();
num_aliases = 1;
}
else
{
symRef = node->getSymbolReference();
sym = symRef->getSymbol();
symIndex = symRef->getSymbol()->getLocalIndex();
num_aliases = _useDefInfo->getNumAliases(symRef, _aux);
}
if (symIndex == NULL_USEDEF_SYMBOL_INDEX || node->getOpCode().isCall() || node->getOpCode().isFence() ||
(parent && parent->getOpCode().isResolveCheck() && num_aliases > 1))
{
// A call or unresolved reference is a definition of all
// symbols it is aliased with
//
numDefNodes = num_aliases;
//for all symbols that are a mustdef of a call, kill defs of those symbols
if (node->getOpCode().isCall())
foundDefsToKill = false;
}
else if (irrelevantStore || _useDefInfo->isExpandedDefIndex(nodeIndex))
{
// DefOnly node defines all symbols it is aliased with
// UseDef node(load) defines only the symbol itself
//
if (!irrelevantStore)
{
numDefNodes = num_aliases;
numDefNodes = _useDefInfo->isExpandedUseDefIndex(nodeIndex) ? 1 : numDefNodes;
if (!_useDefInfo->getDefsForSymbolIsZero(symIndex, _aux) &&
(!sym ||
(!sym->isShadow() &&
!sym->isMethod())))
{
foundDefsToKill = true;
// defsKilled ORed with defsForSymbol(symIndex);
_useDefInfo->getDefsForSymbol(defsKilled, symIndex, _aux);
}
if (node->getOpCode().isStoreIndirect())
{
int32_t memSymIndex = _useDefInfo->getMemorySymbolIndex(node);
if (memSymIndex != -1 &&
!_useDefInfo->getDefsForSymbolIsZero(memSymIndex, _aux))
{
foundDefsToKill = true;
// defsKilled ORed with defsForSymbol(symIndex);
_useDefInfo->getDefsForSymbol(defsKilled, memSymIndex, _aux);
}
}
}
else if (!_useDefInfo->getDefsForSymbolIsZero(symIndex, _aux))
{
numDefNodes = 1;
foundDefsToKill = true;
// defsKilled ORed with defsForSymbol(symIndex);
_useDefInfo->getDefsForSymbol(defsKilled, symIndex, _aux);
}
}
else
{
numDefNodes = 0;
}
if (foundDefsToKill)
{
if (_regularKillSetInfo[blockNum] == NULL)
allocateContainer(&_regularKillSetInfo[blockNum]);
*_regularKillSetInfo[blockNum] |= defsKilled;
if (!seenException)
{
if (_exceptionKillSetInfo[blockNum] == NULL)
allocateContainer(&_exceptionKillSetInfo[blockNum]);
*_exceptionKillSetInfo[blockNum] |= defsKilled;
}
}
if (_regularGenSetInfo[blockNum] == NULL)
allocateContainer(&_regularGenSetInfo[blockNum]);
else if (foundDefsToKill)
*_regularGenSetInfo[blockNum] -= defsKilled;
if (_exceptionGenSetInfo[blockNum] == NULL)
allocateContainer(&_exceptionGenSetInfo[blockNum]);
else if (foundDefsToKill && !seenException)
*_exceptionGenSetInfo[blockNum] -= defsKilled;
if (!irrelevantStore)
{
for (i = 0; i < numDefNodes; ++i)
{
_regularGenSetInfo[blockNum]->set(nodeIndex+i);
_exceptionGenSetInfo[blockNum]->set(nodeIndex+i);
}
}
else // fake up the method entry def as the def index to "gen" to avoid a use without a def completely
{
_regularGenSetInfo[blockNum]->set(sym->getLocalIndex());
_exceptionGenSetInfo[blockNum]->set(sym->getLocalIndex());
}
}
TR::Instruction *OMR::Power::Linkage::saveArguments(TR::Instruction *cursor, bool fsd, bool saveOnly,
List<TR::ParameterSymbol> &parmList)
{
#define REAL_REGISTER(ri) machine->getRealRegister(ri)
#define REGNUM(ri) ((TR::RealRegister::RegNum)(ri))
const TR::PPCLinkageProperties& properties = self()->getProperties();
TR::Machine *machine = self()->machine();
TR::RealRegister *stackPtr = self()->cg()->getStackPointerRegister();
TR::ResolvedMethodSymbol *bodySymbol = self()->comp()->getJittedMethodSymbol();
ListIterator<TR::ParameterSymbol> paramIterator(&parmList);
TR::ParameterSymbol *paramCursor;
TR::Node *firstNode = self()->comp()->getStartTree()->getNode();
TR_BitVector freeScratchable;
int32_t busyMoves[3][64];
int32_t busyIndex = 0, i1;
bool all_saved = false;
// the freeScratchable structure will not be used when saveOnly == true
// no additional conditions were added with the intention of keeping the code easier to read
// and not full of if conditions
freeScratchable.init(TR::RealRegister::LastFPR + 1, self()->trMemory());
// first, consider all argument registers free
for (i1=TR::RealRegister::FirstGPR; i1<=TR::RealRegister::LastFPR; i1++)
{
if (!properties.getReserved(REGNUM(i1)))
{
freeScratchable.set(i1);
}
}
// second, go through all parameters and reset registers that are actually used
for (paramCursor=paramIterator.getFirst(); paramCursor!=NULL; paramCursor=paramIterator.getNext())
{
int32_t lri = paramCursor->getLinkageRegisterIndex();
TR::DataType type = paramCursor->getType();
if (lri >= 0)
{
TR::RealRegister::RegNum regNum;
bool twoRegs = (TR::Compiler->target.is32Bit() && type.isInt64() && lri < properties.getNumIntArgRegs()-1);
if (!type.isFloatingPoint())
{
regNum = properties.getIntegerArgumentRegister(lri);
if (paramCursor->isReferencedParameter()) freeScratchable.reset(regNum);
if (twoRegs)
if (paramCursor->isReferencedParameter()) freeScratchable.reset(regNum+1);
}
else
{
regNum = properties.getFloatArgumentRegister(lri);
if (paramCursor->isReferencedParameter()) freeScratchable.reset(regNum);
if (twoRegs)
if (paramCursor->isReferencedParameter()) freeScratchable.reset(regNum+1);
}
}
}
for (paramCursor=paramIterator.getFirst(); paramCursor!=NULL; paramCursor=paramIterator.getNext())
{
int32_t lri = paramCursor->getLinkageRegisterIndex();
int32_t ai = paramCursor->getAllocatedIndex();
int32_t offset = self()->calculateParameterRegisterOffset(paramCursor->getParameterOffset(), *paramCursor);
TR::DataType type = paramCursor->getType();
int32_t dtype = type.getDataType();
// TODO: Is there an accurate assume to insert here ?
if (lri >= 0)
{
if (!paramCursor->isReferencedParameter() && !paramCursor->isParmHasToBeOnStack()) continue;
TR::RealRegister::RegNum regNum;
bool twoRegs = (TR::Compiler->target.is32Bit() && type.isInt64() && lri < properties.getNumIntArgRegs()-1);
if (type.isFloatingPoint())
regNum = properties.getFloatArgumentRegister(lri);
else
regNum = properties.getIntegerArgumentRegister(lri);
// Do not save arguments to the stack if in Full Speed Debug and saveOnly is not set.
// If not in Full Speed Debug, the arguments will be saved.
if (((ai<0 || self()->hasToBeOnStack(paramCursor)) && !fsd) || (fsd && saveOnly))
{
switch (dtype)
{
case TR::Int8:
case TR::Int16:
case TR::Int32:
{
TR::InstOpCode::Mnemonic op = TR::InstOpCode::stw;
if (!all_saved) cursor = generateMemSrc1Instruction(self()->cg(), op, firstNode,
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, 4, self()->cg()), REAL_REGISTER(regNum), cursor);
}
break;
case TR::Address:
if (!all_saved) cursor = generateMemSrc1Instruction(self()->cg(),TR::InstOpCode::Op_st, firstNode,
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, TR::Compiler->om.sizeofReferenceAddress(), self()->cg()), REAL_REGISTER(regNum), cursor);
break;
case TR::Int64:
if (!all_saved) cursor = generateMemSrc1Instruction(self()->cg(),TR::InstOpCode::Op_st, firstNode,
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, TR::Compiler->om.sizeofReferenceAddress(), self()->cg()), REAL_REGISTER(regNum), cursor);
if (twoRegs)
{
if (!all_saved) cursor = generateMemSrc1Instruction(self()->cg(), TR::InstOpCode::stw, firstNode,
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset+4, 4, self()->cg()),
REAL_REGISTER(REGNUM(regNum+1)), cursor);
if (ai<0)
freeScratchable.set(regNum+1);
}
break;
case TR::Float:
cursor = generateMemSrc1Instruction(self()->cg(), TR::InstOpCode::stfs, firstNode,
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, 4, self()->cg()),
REAL_REGISTER(regNum), cursor);
break;
case TR::Double:
cursor = generateMemSrc1Instruction(self()->cg(), TR::InstOpCode::stfd, firstNode,
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, 8, self()->cg()),
REAL_REGISTER(regNum), cursor);
break;
default:
TR_ASSERT(false, "assertion failure");
break;
}
if (ai<0)
freeScratchable.set(regNum);
}
// Global register is allocated to this argument.
// Don't process if in Full Speed Debug and saveOnly is set
if (ai>=0 && (!fsd || !saveOnly))
{
if (regNum != ai) // Equal assignment: do nothing
{
if (freeScratchable.isSet(ai))
{
cursor = generateTrg1Src1Instruction(self()->cg(),
(type.isFloatingPoint()) ? TR::InstOpCode::fmr:TR::InstOpCode::mr,
firstNode, REAL_REGISTER(REGNUM(ai)), REAL_REGISTER(regNum), cursor);
freeScratchable.reset(ai);
freeScratchable.set(regNum);
}
else // The status of target global register is unclear (i.e. it is a arg reg)
{
busyMoves[0][busyIndex] = regNum;
busyMoves[1][busyIndex] = ai;
busyMoves[2][busyIndex] = 0;
busyIndex++;
}
}
if (TR::Compiler->target.is32Bit() && type.isInt64())
{
int32_t aiLow = paramCursor->getAllocatedLow();
if (!twoRegs) // Low part needs to come from memory
{
offset += 4; // We are dealing with the low part
if (freeScratchable.isSet(aiLow))
{
cursor = generateTrg1MemInstruction(self()->cg(), TR::InstOpCode::lwz, firstNode, REAL_REGISTER(REGNUM(aiLow)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, 4, self()->cg()), cursor);
freeScratchable.reset(aiLow);
}
else
{
busyMoves[0][busyIndex] = offset;
busyMoves[1][busyIndex] = aiLow;
busyMoves[2][busyIndex] = 1;
busyIndex++;
}
}
else if (regNum+1 != aiLow) // Low part needs to be moved
{
if (freeScratchable.isSet(aiLow))
{
cursor = generateTrg1Src1Instruction(self()->cg(), TR::InstOpCode::mr,
firstNode, REAL_REGISTER(REGNUM(aiLow)),
REAL_REGISTER(REGNUM(regNum+1)), cursor);
freeScratchable.reset(aiLow);
freeScratchable.set(regNum+1);
}
else
{
busyMoves[0][busyIndex] = regNum+1;
busyMoves[1][busyIndex] = aiLow;
busyMoves[2][busyIndex] = 0;
busyIndex++;
}
}
}
}
}
// Don't process if in Full Speed Debug and saveOnly is set
else if (ai >= 0 && (!fsd || !saveOnly)) // lri<0: arg needs to come from memory
{
switch (dtype)
{
case TR::Int8:
case TR::Int16:
case TR::Int32:
if (freeScratchable.isSet(ai))
{
cursor = generateTrg1MemInstruction(self()->cg(), TR::InstOpCode::lwz, firstNode, REAL_REGISTER(REGNUM(ai)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, 4, self()->cg()), cursor);
freeScratchable.reset(ai);
}
else
{
busyMoves[0][busyIndex] = offset;
busyMoves[1][busyIndex] = ai;
busyMoves[2][busyIndex] = 1;
busyIndex++;
}
break;
case TR::Address:
if (freeScratchable.isSet(ai))
{
cursor = generateTrg1MemInstruction(self()->cg(),TR::InstOpCode::Op_load, firstNode, REAL_REGISTER(REGNUM(ai)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, TR::Compiler->om.sizeofReferenceAddress(), self()->cg()), cursor);
freeScratchable.reset(ai);
}
else
{
busyMoves[0][busyIndex] = offset;
busyMoves[1][busyIndex] = ai;
if (TR::Compiler->target.is64Bit())
busyMoves[2][busyIndex] = 2;
else
busyMoves[2][busyIndex] = 1;
busyIndex++;
}
break;
case TR::Int64:
if (TR::Compiler->target.is64Bit())
{
if (freeScratchable.isSet(ai))
{
cursor = generateTrg1MemInstruction(self()->cg(), TR::InstOpCode::ld, firstNode, REAL_REGISTER(REGNUM(ai)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, 8, self()->cg()), cursor);
freeScratchable.reset(ai);
}
else
{
busyMoves[0][busyIndex] = offset;
busyMoves[1][busyIndex] = ai;
busyMoves[2][busyIndex] = 2;
busyIndex++;
}
}
else // 32-bit
{
if (freeScratchable.isSet(ai))
{
cursor = generateTrg1MemInstruction(self()->cg(), TR::InstOpCode::lwz, firstNode, REAL_REGISTER(REGNUM(ai)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, 4, self()->cg()), cursor);
freeScratchable.reset(ai);
}
else
{
busyMoves[0][busyIndex] = offset;
busyMoves[1][busyIndex] = ai;
busyMoves[2][busyIndex] = 1;
busyIndex++;
}
ai = paramCursor->getAllocatedLow();
if (freeScratchable.isSet(ai))
{
cursor = generateTrg1MemInstruction(self()->cg(), TR::InstOpCode::lwz, firstNode, REAL_REGISTER(REGNUM(ai)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset+4, 4, self()->cg()), cursor);
freeScratchable.reset(ai);
}
else
{
busyMoves[0][busyIndex] = offset+4;
busyMoves[1][busyIndex] = ai;
busyMoves[2][busyIndex] = 1;
busyIndex++;
}
}
break;
case TR::Float:
if (freeScratchable.isSet(ai))
{
cursor = generateTrg1MemInstruction(self()->cg(), TR::InstOpCode::lfs, firstNode, REAL_REGISTER(REGNUM(ai)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, 4, self()->cg()), cursor);
freeScratchable.reset(ai);
}
else
{
busyMoves[0][busyIndex] = offset;
busyMoves[1][busyIndex] = ai;
busyMoves[2][busyIndex] = 3;
busyIndex++;
}
break;
case TR::Double:
if (freeScratchable.isSet(ai))
{
cursor = generateTrg1MemInstruction(self()->cg(), TR::InstOpCode::lfd, firstNode, REAL_REGISTER(REGNUM(ai)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, offset, 8, self()->cg()), cursor);
freeScratchable.reset(ai);
}
else
{
busyMoves[0][busyIndex] = offset;
busyMoves[1][busyIndex] = ai;
busyMoves[2][busyIndex] = 4;
busyIndex++;
}
break;
default:
break;
}
}
}
if (!fsd || !saveOnly)
{
bool freeMore = true;
int32_t numMoves = busyIndex;
while (freeMore && numMoves>0)
{
freeMore = false;
for (i1=0; i1<busyIndex; i1++)
{
int32_t source = busyMoves[0][i1];
int32_t target = busyMoves[1][i1];
if (!(target<0) && freeScratchable.isSet(target))
{
switch(busyMoves[2][i1])
{
case 0:
cursor = generateTrg1Src1Instruction(self()->cg(), (source<=TR::RealRegister::LastGPR)?TR::InstOpCode::mr:TR::InstOpCode::fmr,
firstNode, REAL_REGISTER(REGNUM(target)), REAL_REGISTER(REGNUM(source)), cursor);
freeScratchable.set(source);
break;
case 1:
cursor = generateTrg1MemInstruction(self()->cg(), TR::InstOpCode::lwz, firstNode, REAL_REGISTER(REGNUM(target)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, source, 4, self()->cg()), cursor);
break;
case 2:
cursor = generateTrg1MemInstruction(self()->cg(), TR::InstOpCode::ld, firstNode, REAL_REGISTER(REGNUM(target)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, source, 8, self()->cg()), cursor);
break;
case 3:
cursor = generateTrg1MemInstruction(self()->cg(), TR::InstOpCode::lfs, firstNode, REAL_REGISTER(REGNUM(target)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, source, 4, self()->cg()), cursor);
break;
case 4:
cursor = generateTrg1MemInstruction(self()->cg(), TR::InstOpCode::lfd, firstNode, REAL_REGISTER(REGNUM(target)),
new (self()->trHeapMemory()) TR::MemoryReference(stackPtr, source, 8, self()->cg()), cursor);
break;
}
freeScratchable.reset(target);
freeMore = true;
busyMoves[0][i1] = busyMoves[1][i1] = -1;
numMoves--;
}
}
}
TR_ASSERT(numMoves<=0, "Circular argument register dependency can and should be avoided.");
}
return(cursor);
}
void
OMR::CodeGenPhase::performSetupForInstructionSelectionPhase(TR::CodeGenerator * cg, TR::CodeGenPhase * phase)
{
TR::Compilation *comp = cg->comp();
if (TR::Compiler->target.cpu.isZ() && TR::Compiler->om.shouldGenerateReadBarriersForFieldLoads())
{
// TODO (GuardedStorage): We need to come up with a better solution than anchoring aloadi's
// to enforce certain evaluation order
traceMsg(comp, "GuardedStorage: in performSetupForInstructionSelectionPhase\n");
auto mapAllocator = getTypedAllocator<std::pair<TR::TreeTop*, TR::TreeTop*> >(comp->allocator());
std::map<TR::TreeTop*, TR::TreeTop*, std::less<TR::TreeTop*>, TR::typed_allocator<std::pair<TR::TreeTop* const, TR::TreeTop*>, TR::Allocator> >
currentTreeTopToappendTreeTop(std::less<TR::TreeTop*> (), mapAllocator);
TR_BitVector *unAnchorableAloadiNodes = comp->getBitVectorPool().get();
for (TR::PreorderNodeIterator iter(comp->getStartTree(), comp); iter != NULL; ++iter)
{
TR::Node *node = iter.currentNode();
traceMsg(comp, "GuardedStorage: Examining node = %p\n", node);
// isNullCheck handles both TR::NULLCHK and TR::ResolveAndNULLCHK
// both of which do not operate on their child but their
// grandchild (or greatgrandchild).
if (node->getOpCode().isNullCheck())
{
// An aloadi cannot be anchored if there is a Null Check on
// its child. There are two situations where this occurs.
// The first is when doing an aloadi off some node that is
// being NULLCHK'd (see Ex1). The second is when doing an
// icalli in which case the aloadi loads the VFT of an
// object that must be NULLCHK'd (see Ex2).
//
// Ex1:
// n1n NULLCHK on n3n
// n2n aloadi f <-- First Child And Parent of Null Chk'd Node
// n3n aload O
//
// Ex2:
// n1n NULLCHK on n4n
// n2n icall foo <-- First Child
// n3n aloadi <vft> <-- Parent of Null Chk'd Node
// n4n aload O
// n4n ==> aload O
TR::Node *nodeBeingNullChkd = node->getNullCheckReference();
if (nodeBeingNullChkd)
{
TR::Node *firstChild = node->getFirstChild();
TR::Node *parentOfNullChkdNode = NULL;
if (firstChild->getOpCode().isCall() &&
firstChild->getOpCode().isIndirect())
{
parentOfNullChkdNode = firstChild->getFirstChild();
}
else
{
parentOfNullChkdNode = firstChild;
}
if (parentOfNullChkdNode &&
parentOfNullChkdNode->getOpCodeValue() == TR::aloadi &&
parentOfNullChkdNode->getNumChildren() > 0 &&
parentOfNullChkdNode->getFirstChild() == nodeBeingNullChkd)
{
unAnchorableAloadiNodes->set(parentOfNullChkdNode->getGlobalIndex());
traceMsg(comp, "GuardedStorage: Cannot anchor %p\n", firstChild);
}
}
}
else
{
bool shouldAnchorNode = false;
if (node->getOpCodeValue() == TR::aloadi &&
!unAnchorableAloadiNodes->isSet(node->getGlobalIndex()))
{
shouldAnchorNode = true;
}
else if (node->getOpCodeValue() == TR::aload &&
node->getSymbol()->isStatic() &&
node->getSymbol()->isCollectedReference())
{
shouldAnchorNode = true;
}
if (shouldAnchorNode)
{
TR::TreeTop* anchorTreeTop = TR::TreeTop::create(comp, TR::Node::create(TR::treetop, 1, node));
TR::TreeTop* appendTreeTop = iter.currentTree();
if (currentTreeTopToappendTreeTop.count(appendTreeTop) > 0)
{
appendTreeTop = currentTreeTopToappendTreeTop[appendTreeTop];
}
// Anchor the aload/aloadi before the current treetop
appendTreeTop->insertBefore(anchorTreeTop);
currentTreeTopToappendTreeTop[iter.currentTree()] = anchorTreeTop;
traceMsg(comp, "GuardedStorage: Anchored %p to treetop = %p\n", node, anchorTreeTop);
}
}
}
comp->getBitVectorPool().release(unAnchorableAloadiNodes);
}
if (cg->shouldBuildStructure() &&
(comp->getFlowGraph()->getStructure() != NULL))
{
TR_Structure *rootStructure = TR_RegionAnalysis::getRegions(comp);
comp->getFlowGraph()->setStructure(rootStructure);
}
phase->reportPhase(SetupForInstructionSelectionPhase);
// Dump preIR
if (comp->getOption(TR_TraceRegisterPressureDetails) && !comp->getOption(TR_DisableRegisterPressureSimulation))
{
traceMsg(comp, " { Post optimization register pressure simulation\n");
TR_BitVector emptyBitVector;
vcount_t vc = comp->incVisitCount();
cg->initializeRegisterPressureSimulator();
for (TR::Block *block = comp->getStartBlock(); block; block = block->getNextExtendedBlock())
{
TR_LinkHead<TR_RegisterCandidate> emptyCandidateList;
TR::CodeGenerator::TR_RegisterPressureState state(NULL, 0, emptyBitVector, emptyBitVector, &emptyCandidateList, cg->getNumberOfGlobalGPRs(), cg->getNumberOfGlobalFPRs(), cg->getNumberOfGlobalVRFs(), vc);
TR::CodeGenerator::TR_RegisterPressureSummary summary(state._gprPressure, state._fprPressure, state._vrfPressure);
cg->simulateBlockEvaluation(block, &state, &summary);
}
traceMsg(comp, " }\n");
}
TR::LexicalMemProfiler mp(phase->getName(), comp->phaseMemProfiler());
LexicalTimer pt(phase->getName(), comp->phaseTimer());
cg->setUpForInstructionSelection();
}
TR_GlobalRegisterNumber
OMR::X86::I386::CodeGenerator::pickRegister(
TR_RegisterCandidate *rc,
TR::Block **allBlocks,
TR_BitVector &availableRegisters,
TR_GlobalRegisterNumber &highRegisterNumber,
TR_LinkHead<TR_RegisterCandidate> *candidates)
{
if (!self()->comp()->getOption(TR_DisableRegisterPressureSimulation))
{
if (self()->comp()->getOption(TR_AssignEveryGlobalRegister))
{
// This is not really necessary except for testing purposes.
// Conceptually, the common pickRegister code should be free to make
// its choices based only on performance considerations, and shouldn't
// need to worry about correctness. When SupportsVMThreadGRA is not set,
// it is incorrect to choose the VMThread register. Therefore we mask
// it out here.
//
// Having said that, the common code *does* already mask out the
// VMThread register for convenience, so under normal circumstances,
// this code is redundant. It is only necessary when
// TR_AssignEveryGlobalRegister is set.
//
availableRegisters -= *self()->getGlobalRegisters(TR_vmThreadSpill, self()->comp()->getMethodSymbol()->getLinkageConvention());
}
return OMR::CodeGenerator::pickRegister(rc, allBlocks, availableRegisters, highRegisterNumber, candidates);
}
if ((rc->getSymbol()->getDataType() == TR::Float) ||
(rc->getSymbol()->getDataType() == TR::Double))
{
if (availableRegisters.get(7))
return 7;
if (availableRegisters.get(8))
return 8;
if (availableRegisters.get(9))
return 9;
if (availableRegisters.get(10))
return 10;
if (availableRegisters.get(11))
return 11;
if (availableRegisters.get(12))
return 12;
return -1;
}
if (!_assignedGlobalRegisters)
_assignedGlobalRegisters = new (self()->trStackMemory()) TR_BitVector(self()->comp()->getSymRefCount(), self()->trMemory(), stackAlloc, growable);
if (availableRegisters.get(5))
return 5; // esi
if (availableRegisters.get(2))
return 2; // ecx
static char *dontUseEBXasGPR = feGetEnv("dontUseEBXasGPR");
if (!dontUseEBXasGPR && availableRegisters.get(1))
return 1;
#ifdef J9_PROJECT_SPECIFIC
TR::RecognizedMethod rm = self()->comp()->getMethodSymbol()->getRecognizedMethod();
if (rm == TR::java_util_HashtableHashEnumerator_hasMoreElements)
{
if (availableRegisters.get(4))
return 4; // edi
if (availableRegisters.get(3))
return 3; // edx
}
else
#endif
{
int32_t numExtraRegs = 0;
int32_t maxRegisterPressure = 0;
vcount_t visitCount = self()->comp()->incVisitCount();
TR_BitVectorIterator bvi(rc->getBlocksLiveOnEntry());
int32_t maxFrequency = 0;
while (bvi.hasMoreElements())
{
int32_t liveBlockNum = bvi.getNextElement();
TR::Block *block = allBlocks[liveBlockNum];
if (block->getFrequency() > maxFrequency)
maxFrequency = block->getFrequency();
}
int32_t maxStaticFrequency = 0;
if (maxFrequency == 0)
{
bvi.setBitVector(rc->getBlocksLiveOnEntry());
while (bvi.hasMoreElements())
{
int32_t liveBlockNum = bvi.getNextElement();
TR::Block *block = allBlocks[liveBlockNum];
TR_BlockStructure *blockStructure = block->getStructureOf();
int32_t blockWeight = 1;
if (blockStructure &&
!block->isCold())
{
blockStructure->calculateFrequencyOfExecution(&blockWeight);
if (blockWeight > maxStaticFrequency)
maxStaticFrequency = blockWeight;
}
}
}
bool assigningEDX = false;
if (!availableRegisters.get(4) &&
availableRegisters.get(3))
assigningEDX = true;
bool vmThreadUsed = false;
bvi.setBitVector(rc->getBlocksLiveOnEntry());
while (bvi.hasMoreElements())
{
int32_t liveBlockNum = bvi.getNextElement();
TR::Block *block = allBlocks[liveBlockNum];
_assignedGlobalRegisters->empty();
int32_t numAssignedGlobalRegs = 0;
TR_RegisterCandidate *prev;
for (prev = candidates->getFirst(); prev; prev = prev->getNext())
{
bool gprCandidate = true;
if ((prev->getSymbol()->getDataType() == TR::Float) ||
(prev->getSymbol()->getDataType() == TR::Double))
gprCandidate = false;
if (gprCandidate && prev->getBlocksLiveOnEntry().get(liveBlockNum))
{
numAssignedGlobalRegs++;
if (prev->getDataType() == TR::Int64)
numAssignedGlobalRegs++;
_assignedGlobalRegisters->set(prev->getSymbolReference()->getReferenceNumber());
}
}
maxRegisterPressure = self()->estimateRegisterPressure(block, visitCount, maxStaticFrequency, maxFrequency, vmThreadUsed, numAssignedGlobalRegs, _assignedGlobalRegisters, rc->getSymbolReference(), assigningEDX);
if (maxRegisterPressure >= self()->getMaximumNumbersOfAssignableGPRs())
break;
}
// Determine if we can spare any extra registers for this candidate without spilling
// in any hot (critical) blocks
//
if (maxRegisterPressure < self()->getMaximumNumbersOfAssignableGPRs())
numExtraRegs = self()->getMaximumNumbersOfAssignableGPRs() - maxRegisterPressure;
//dumpOptDetails("For global register candidate %d reg pressure is %d maxRegs %d numExtraRegs %d\n", rc->getSymbolReference()->getReferenceNumber(), maxRegisterPressure, comp()->cg()->getMaximumNumbersOfAssignableGPRs(), numExtraRegs);
if (numExtraRegs > 0)
{
if (availableRegisters.get(4))
return 4; // edi
if (availableRegisters.get(3))
return 3; // edx
}
}
return -1; // -1 ==> don't use a global register
}
