TR::Block *
TR_ExpressionsSimplification::findPredecessorBlock(TR::CFGNode *entryNode)
{
if (!(entryNode->getPredecessors().size() == 2))
return 0;
TR::Block *block = 0;
for (auto edge = entryNode->getPredecessors().begin(); edge != entryNode->getPredecessors().end(); ++edge)
{
if ((*edge)->getFrom()->getSuccessors().size() == 1)
{
block = toBlock((*edge)->getFrom());
if (block->getStructureOf()->isLoopInvariantBlock())
break;
else
block = 0;
}
}
//traceMsg(comp(), "Commoned code will be put in block_%d\n", block->getNumber());
return block;
}
TR::TreeTop *
OMR::Simplifier::simplifyExtendedBlock(TR::TreeTop * treeTop)
{
TR::Block * block = 0;
_containingStructure = NULL;
_blockRemoved = false;
for (; treeTop; treeTop = treeTop->getNextTreeTop())
{
TR::Node * node = treeTop->getNode();
TR_ASSERT(node->getOpCodeValue() == TR::BBStart, "Simplification, expected BBStart treetop");
TR::Block * b = node->getBlock();
if (block && !b->isExtensionOfPreviousBlock())
break;
if (b->isOSRCodeBlock() || b->isOSRCatchBlock())
{
b->setHasBeenVisited();
treeTop = b->getExit();
continue;
}
#ifdef DEBUG
if (block != b)
b->setHasBeenVisited();
#endif
if (!block && _reassociate &&
comp()->getFlowGraph()->getStructure() != NULL // [99391] getStructureOf() only valid if structure isn't invalidated
)
{ // b is first block in the extended block
TR_BlockStructure *blockStructure = b->getStructureOf();
if(blockStructure)
{
TR_Structure *parent = blockStructure->getParent();
while (parent)
{
TR_RegionStructure *region = parent->asRegion();
if (region->isNaturalLoop() /* || region->containsInternalCycles() */)
{
_containingStructure = region;
break;
}
parent = parent->getParent();
}
}
}
block = b;
if (trace())
traceMsg(comp(), "simplifying block_%d\n", block->getNumber());
_performLowerTreeSimplifier=NULL;
_performLowerTreeNode=NULL;
simplify(block);
if(_performLowerTreeSimplifier)
{
_performLowerTreeNode = postWalkLowerTreeSimplifier(_performLowerTreeSimplifier, _performLowerTreeNode, block, (TR::Simplifier *) this);
_performLowerTreeSimplifier->setNode(_performLowerTreeNode);
}
// If the block itself was removed from the CFG during simplification, find
// the next 'legitimate' block to be simplified
//
//if (comp()->getFlowGraph()->getRemovedNodes().find(block))
if(block->nodeIsRemoved())
{
TR::TreeTop * tt = findNextLegalTreeTop(comp(), block);
// in certain cases the removed block might be the last one we haven't
// visited and therefore we won't be able to find a treetop to continue
// in such cases we exit the loop
//
treeTop = tt ? tt->getPrevTreeTop() : 0;
if (!treeTop)
break;
}
else
{
treeTop = block->getExit();
}
}
// now remove any unreachable blocks
//
if (_blockRemoved)
{
// if the next block to be processed has been removed,
// find the next valid block to process
//
if (treeTop)
{
TR::Block *b = treeTop->getNode()->getBlock();
//if (comp()->getFlowGraph()->getRemovedNodes().find(b))
if(b->nodeIsRemoved())
treeTop = findNextLegalTreeTop(comp(), b);
}
}
return treeTop;
}
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
}