// Process the structure recursively
//
int32_t
TR_ExpressionsSimplification::perform(TR_Structure * str)
{
if (trace())
traceMsg(comp(), "Analyzing root Structure : %p\n", str);
TR_RegionStructure *region;
// Only regions can be simplified
//
if (!(region = str->asRegion()))
return 0;
TR_RegionStructure::Cursor it(*region);
for (TR_StructureSubGraphNode *node = it.getCurrent();
node != 0;
node = it.getNext())
{
// Too strict
/*
if ((node->getPredecessors().size() == 1))
{
TR::CFGEdge *edge = node->getPredecessors().front();
TR_StructureSubGraphNode *pred = toStructureSubGraphNode(edge->getFrom());
TR_BlockStructure *b = pred->getStructure()->asBlock();
if (b && pred->getSuccessors().size() == 1))
perform(node->getStructure());
}
*/
perform(node->getStructure());
}
// debug only
//
/*
if (region->isNaturalLoop() &&
(region->getParent() &&
!region->getParent()->asRegion()->isCanonicalizedLoop()))
{
traceMsg(comp(), "Loop not canonicalized %x\n", region);
}
*/
TR::Block *entryBlock = region->getEntryBlock();
if (region->isNaturalLoop() && !entryBlock->isCold() &&
(region->getParent() /* &&
region->getParent()->asRegion()->isCanonicalizedLoop() */))
{
if (trace())
traceMsg(comp(), "Found candidate non cold loop %p for expression elimination\n", region);
findAndSimplifyInvariantLoopExpressions(region);
}
return 1; // Need to specify the cost
}
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
}