TR::RegisterDependencyConditions *OMR::Power::RegisterDependencyConditions::cloneAndFix(
TR::CodeGenerator * cg, TR::RegisterDependencyConditions *added)
{
TR::RegisterDependencyConditions *result;
TR::RegisterDependency *singlePair;
int32_t idx, postNum, addPost=0;
if (added != NULL)
{
addPost = added->getAddCursorForPost();
}
postNum = this->getAddCursorForPost();
result = new (cg->trHeapMemory()) TR::RegisterDependencyConditions(0, postNum+addPost, cg->trMemory());
for (idx=0; idx<postNum; idx++)
{
singlePair = this->getPostConditions()->getRegisterDependency(idx);
result->addPostCondition(singlePair->getRegister(), singlePair->getRealRegister(),
singlePair->getFlags());
}
for (idx=0; idx<addPost; idx++)
{
singlePair = added->getPostConditions()->getRegisterDependency(idx);
result->addPostCondition(singlePair->getRegister(), singlePair->getRealRegister(),
singlePair->getFlags());
}
return result;
}
TR::RegisterDependencyConditions *OMR::ARM::RegisterDependencyConditions::cloneAndFix(
TR::CodeGenerator * cg, TR::RegisterDependencyConditions *added)
{
TR::RegisterDependencyConditions *result;
TR_ARMRegisterDependency *singlePair;
int32_t idx, preNum, postNum, addPre=0, addPost=0;
TR::Register *postReg, *tempReg;
TR::RealRegister::RegNum rnum;
if (added != NULL)
{
addPre = added->getAddCursorForPre();
addPost = added->getAddCursorForPost();
}
preNum = this->getAddCursorForPre();
postNum = this->getAddCursorForPost();
result = new (cg->trHeapMemory()) TR::RegisterDependencyConditions(preNum + addPre, postNum + addPost, cg->trMemory());
for (idx=0; idx<postNum; idx++)
{
singlePair = this->getPostConditions()->getRegisterDependency(idx);
rnum = singlePair->getRealRegister();
tempReg = singlePair->getRegister();
result->addPostCondition(tempReg, rnum, singlePair->getFlags());
if (rnum == TR::RealRegister::gr0)
{
postReg = tempReg;
}
}
for (idx=0; idx<preNum; idx++)
{
singlePair = this->getPreConditions()->getRegisterDependency(idx);
rnum = singlePair->getRealRegister();
tempReg = singlePair->getRegister();
if (rnum == TR::RealRegister::gr0)
tempReg = postReg;
result->addPreCondition(tempReg, rnum, singlePair->getFlags());
}
for (idx=0; idx<addPost; idx++)
{
singlePair = added->getPostConditions()->getRegisterDependency(idx);
result->addPostCondition(singlePair->getRegister(), singlePair->getRealRegister(),
singlePair->getFlags());
}
for (idx=0; idx<addPre; idx++)
{
singlePair = added->getPreConditions()->getRegisterDependency(idx);
result->addPreCondition(singlePair->getRegister(), singlePair->getRealRegister(),
singlePair->getFlags());
}
return result;
}
TR::Register *
TR::AMD64SystemLinkage::buildIndirectDispatch(TR::Node *callNode)
{
TR::SymbolReference *methodSymRef = callNode->getSymbolReference();
TR_ASSERT(methodSymRef->getSymbol()->castToMethodSymbol()->isComputed(), "system linkage only supports computed indirect call for now %p\n", callNode);
// Evaluate VFT
//
TR::Register *vftRegister;
TR::Node *vftNode = callNode->getFirstChild();
if (vftNode->getRegister())
{
vftRegister = vftNode->getRegister();
}
else
{
vftRegister = cg()->evaluate(vftNode);
}
// Allocate adequate register dependencies.
//
// pre = number of argument registers + 1 for VFT register
// post = number of volatile + VMThread + return register
//
uint32_t pre = getProperties().getNumIntegerArgumentRegisters() + getProperties().getNumFloatArgumentRegisters() + 1;
uint32_t post = getProperties().getNumVolatileRegisters() + 1 + (callNode->getDataType() == TR::NoType ? 0 : 1);
#if defined (PYTHON) && 0
// Treat all preserved GP regs as volatile until register map support available.
//
post += getProperties().getNumberOfPreservedGPRegisters();
#endif
TR::RegisterDependencyConditions *callDeps = generateRegisterDependencyConditions(pre, 1, cg());
TR::RealRegister::RegNum scratchRegIndex = getProperties().getIntegerScratchRegister(1);
callDeps->addPostCondition(vftRegister, scratchRegIndex, cg());
callDeps->stopAddingPostConditions();
// Evaluate outgoing arguments on the system stack and build pre-conditions.
//
int32_t memoryArgSize = buildArgs(callNode, callDeps);
// Dispatch
//
generateRegInstruction(CALLReg, callNode, vftRegister, callDeps, cg());
cg()->resetIsLeafMethod();
// Build label post-conditions
//
TR::RegisterDependencyConditions *postDeps = generateRegisterDependencyConditions(0, post, cg());
TR::Register *returnReg = buildVolatileAndReturnDependencies(callNode, postDeps);
postDeps->stopAddingPostConditions();
TR::LabelSymbol *postDepLabel = generateLabelSymbol(cg());
generateLabelInstruction(LABEL, callNode, postDepLabel, postDeps, cg());
return returnReg;
}
// Create a NoReg dependency for each child of a call that has been evaluated into a register.
// Ignore children that do not have a register since their live range should not persist outside of
// the helper call stream.
//
TR::RegisterDependencyConditions *TR_OutlinedInstructions::formEvaluatedArgumentDepList()
{
int32_t i, c=0;
for (i=_callNode->getFirstArgumentIndex(); i<_callNode->getNumChildren(); i++)
{
TR::Register *reg = _callNode->getChild(i)->getRegister();
if (reg)
{
TR::RegisterPair *regPair = reg->getRegisterPair();
c += regPair? 2 : 1;
}
}
TR::RegisterDependencyConditions *depConds = NULL;
if (c)
{
TR::Machine *machine = _cg->machine();
depConds = generateRegisterDependencyConditions(0, c, _cg);
for (i=_callNode->getFirstArgumentIndex(); i<_callNode->getNumChildren(); i++)
{
TR::Register *reg = _callNode->getChild(i)->getRegister();
if (reg)
{
TR::RegisterPair *regPair = reg->getRegisterPair();
if (regPair)
{
depConds->addPostCondition(regPair->getLowOrder(), TR::RealRegister::NoReg, _cg);
depConds->addPostCondition(regPair->getHighOrder(), TR::RealRegister::NoReg, _cg);
}
else
{
depConds->addPostCondition(reg, TR::RealRegister::NoReg, _cg);
}
}
}
depConds->stopAddingConditions();
}
return depConds;
}
void
TR_S390BinaryAnalyser::longSubtractAnalyser(TR::Node * root)
{
TR::Node * firstChild;
TR::Node * secondChild;
TR::Instruction * cursor = NULL;
TR::RegisterDependencyConditions * dependencies = NULL;
bool setsOrReadsCC = NEED_CC(root) || (root->getOpCodeValue() == TR::lusubb);
TR::InstOpCode::Mnemonic regToRegOpCode;
TR::InstOpCode::Mnemonic memToRegOpCode;
TR::Compilation *comp = TR::comp();
if (TR::Compiler->target.is64Bit() || cg()->use64BitRegsOn32Bit())
{
if (!setsOrReadsCC)
{
regToRegOpCode = TR::InstOpCode::SGR;
memToRegOpCode = TR::InstOpCode::SG;
}
else
{
regToRegOpCode = TR::InstOpCode::SLGR;
memToRegOpCode = TR::InstOpCode::SLG;
}
}
else
{
regToRegOpCode = TR::InstOpCode::SLR;
memToRegOpCode = TR::InstOpCode::SL;
}
firstChild = root->getFirstChild();
secondChild = root->getSecondChild();
TR::Register * firstRegister = firstChild->getRegister();
TR::Register * secondRegister = secondChild->getRegister();
setInputs(firstChild, firstRegister, secondChild, secondRegister,
false, false, comp);
/** Attempt to use SGH to subtract halfword (64 <- 16).
* The second child is a halfword from memory */
bool is16BitMemory2Operand = false;
if (TR::Compiler->target.cpu.getS390SupportsZ14() &&
secondChild->getOpCodeValue() == TR::s2l &&
secondChild->getFirstChild()->getOpCodeValue() == TR::sloadi &&
secondChild->isSingleRefUnevaluated() &&
secondChild->getFirstChild()->isSingleRefUnevaluated())
{
setMem2();
memToRegOpCode = TR::InstOpCode::SGH;
is16BitMemory2Operand = true;
}
if (getEvalChild1())
{
firstRegister = cg()->evaluate(firstChild);
}
if (getEvalChild2())
{
secondRegister = cg()->evaluate(secondChild);
}
remapInputs(firstChild, firstRegister, secondChild, secondRegister);
if ((root->getOpCodeValue() == TR::lusubb) &&
TR_S390ComputeCC::setCarryBorrow(root->getChild(2), false, cg()))
{
// use SLBGR rather than SLGR/SGR
// SLBG rather than SLG/SG
// or
// use SLBR rather than SLR
// SLB rather than SL
bool uses64bit = TR::Compiler->target.is64Bit() || cg()->use64BitRegsOn32Bit();
regToRegOpCode = uses64bit ? TR::InstOpCode::SLBGR : TR::InstOpCode::SLBR;
memToRegOpCode = uses64bit ? TR::InstOpCode::SLBG : TR::InstOpCode::SLB;
}
if (TR::Compiler->target.is64Bit() || cg()->use64BitRegsOn32Bit())
{
if (getCopyReg1())
{
TR::Register * thirdReg = cg()->allocate64bitRegister();
root->setRegister(thirdReg);
generateRRInstruction(cg(), TR::InstOpCode::LGR, root, thirdReg, firstRegister);
if (getBinaryReg3Reg2())
{
generateRRInstruction(cg(), regToRegOpCode, root, thirdReg, secondRegister);
}
else // assert getBinaryReg3Mem2() == true
{
TR::MemoryReference * longMR = generateS390MemoryReference(secondChild, cg());
generateRXInstruction(cg(), memToRegOpCode, root, thirdReg, longMR);
longMR->stopUsingMemRefRegister(cg());
}
}
else if (getBinaryReg1Reg2())
{
generateRRInstruction(cg(), regToRegOpCode, root, firstRegister, secondRegister);
root->setRegister(firstRegister);
}
else // assert getBinaryReg1Mem2() == true
{
TR_ASSERT( !getInvalid(), "TR_S390BinaryAnalyser::invalid case\n");
TR::Node* baseAddrNode = is16BitMemory2Operand ? secondChild->getFirstChild() : secondChild;
TR::MemoryReference * longMR = generateS390MemoryReference(baseAddrNode, cg());
generateRXInstruction(cg(), memToRegOpCode, root, firstRegister, longMR);
longMR->stopUsingMemRefRegister(cg());
root->setRegister(firstRegister);
if(is16BitMemory2Operand)
{
cg()->decReferenceCount(secondChild->getFirstChild());
}
}
}
else // if 32bit codegen...
{
bool zArchTrexsupported = performTransformation(comp, "O^O Use SL/SLB for long sub.");
TR::Register * highDiff = NULL;
TR::LabelSymbol * doneLSub = TR::LabelSymbol::create(cg()->trHeapMemory(),cg());
if (getCopyReg1())
{
TR::Register * lowThird = cg()->allocateRegister();
TR::Register * highThird = cg()->allocateRegister();
TR::RegisterPair * thirdReg = cg()->allocateConsecutiveRegisterPair(lowThird, highThird);
highDiff = highThird;
dependencies = new (cg()->trHeapMemory()) TR::RegisterDependencyConditions(0, 9, cg());
dependencies->addPostCondition(firstRegister, TR::RealRegister::EvenOddPair);
dependencies->addPostCondition(firstRegister->getHighOrder(), TR::RealRegister::LegalEvenOfPair);
dependencies->addPostCondition(firstRegister->getLowOrder(), TR::RealRegister::LegalOddOfPair);
// If 2nd operand has ref count of 1 and can be accessed by a memory reference,
// then second register will not be used.
if(secondRegister == firstRegister && !setsOrReadsCC)
{
TR_ASSERT( false, "lsub with identical children - fix Simplifier");
}
if (secondRegister != NULL && firstRegister != secondRegister)
{
dependencies->addPostCondition(secondRegister, TR::RealRegister::EvenOddPair);
dependencies->addPostCondition(secondRegister->getHighOrder(), TR::RealRegister::LegalEvenOfPair);
dependencies->addPostCondition(secondRegister->getLowOrder(), TR::RealRegister::LegalOddOfPair);
}
dependencies->addPostCondition(highThird, TR::RealRegister::AssignAny);
root->setRegister(thirdReg);
generateRRInstruction(cg(), TR::InstOpCode::LR, root, highThird, firstRegister->getHighOrder());
generateRRInstruction(cg(), TR::InstOpCode::LR, root, lowThird, firstRegister->getLowOrder());
if (getBinaryReg3Reg2())
{
if ((ENABLE_ZARCH_FOR_32 && zArchTrexsupported) || setsOrReadsCC)
{
generateRRInstruction(cg(), regToRegOpCode, root, lowThird, secondRegister->getLowOrder());
generateRRInstruction(cg(), TR::InstOpCode::SLBR, root, highThird, secondRegister->getHighOrder());
}
else
{
generateRRInstruction(cg(), TR::InstOpCode::SR, root, highThird, secondRegister->getHighOrder());
generateRRInstruction(cg(), TR::InstOpCode::SLR, root, lowThird, secondRegister->getLowOrder());
}
}
else // assert getBinaryReg3Mem2() == true
{
TR::MemoryReference * highMR = generateS390MemoryReference(secondChild, cg());
TR::MemoryReference * lowMR = generateS390MemoryReference(*highMR, 4, cg());
dependencies->addAssignAnyPostCondOnMemRef(highMR);
if ((ENABLE_ZARCH_FOR_32 && zArchTrexsupported) || setsOrReadsCC)
{
generateRXInstruction(cg(), memToRegOpCode, root, lowThird, lowMR);
generateRXInstruction(cg(), TR::InstOpCode::SLB, root, highThird, highMR);
}
else
{
generateRXInstruction(cg(), TR::InstOpCode::S, root, highThird, highMR);
generateRXInstruction(cg(), TR::InstOpCode::SL, root, lowThird, lowMR);
}
highMR->stopUsingMemRefRegister(cg());
lowMR->stopUsingMemRefRegister(cg());
}
}
else if (getBinaryReg1Reg2())
{
dependencies = new (cg()->trHeapMemory()) TR::RegisterDependencyConditions(0, 6, cg());
dependencies->addPostCondition(firstRegister, TR::RealRegister::EvenOddPair);
dependencies->addPostCondition(firstRegister->getHighOrder(), TR::RealRegister::LegalEvenOfPair);
dependencies->addPostCondition(firstRegister->getLowOrder(), TR::RealRegister::LegalOddOfPair);
if(secondRegister == firstRegister)
{
TR_ASSERT( false, "lsub with identical children - fix Simplifier");
}
if (secondRegister != firstRegister)
{
dependencies->addPostCondition(secondRegister, TR::RealRegister::EvenOddPair);
dependencies->addPostCondition(secondRegister->getHighOrder(), TR::RealRegister::LegalEvenOfPair);
dependencies->addPostCondition(secondRegister->getLowOrder(), TR::RealRegister::LegalOddOfPair);
}
if ((ENABLE_ZARCH_FOR_32 && zArchTrexsupported) || setsOrReadsCC)
{
generateRRInstruction(cg(), regToRegOpCode, root, firstRegister->getLowOrder(), secondRegister->getLowOrder());
generateRRInstruction(cg(), TR::InstOpCode::SLBR, root, firstRegister->getHighOrder(), secondRegister->getHighOrder());
}
else
{
generateRRInstruction(cg(), TR::InstOpCode::SR, root, firstRegister->getHighOrder(), secondRegister->getHighOrder());
generateRRInstruction(cg(), TR::InstOpCode::SLR, root, firstRegister->getLowOrder(), secondRegister->getLowOrder());
}
highDiff = firstRegister->getHighOrder();
root->setRegister(firstRegister);
}
else // assert getBinaryReg1Mem2() == true
{
TR_ASSERT( !getInvalid(),"TR_S390BinaryAnalyser::invalid case\n");
dependencies = new (cg()->trHeapMemory()) TR::RegisterDependencyConditions(0, 5, cg());
dependencies->addPostCondition(firstRegister, TR::RealRegister::EvenOddPair);
dependencies->addPostCondition(firstRegister->getHighOrder(), TR::RealRegister::LegalEvenOfPair);
dependencies->addPostCondition(firstRegister->getLowOrder(), TR::RealRegister::LegalOddOfPair);
TR::MemoryReference * highMR = generateS390MemoryReference(secondChild, cg());
TR::MemoryReference * lowMR = generateS390MemoryReference(*highMR, 4, cg());
dependencies->addAssignAnyPostCondOnMemRef(highMR);
if ((ENABLE_ZARCH_FOR_32 && zArchTrexsupported) || setsOrReadsCC)
{
generateRXInstruction(cg(), memToRegOpCode, root, firstRegister->getLowOrder(), lowMR);
generateRXInstruction(cg(), TR::InstOpCode::SLB, root, firstRegister->getHighOrder(), highMR);
}
else
{
generateRXInstruction(cg(), TR::InstOpCode::S, root, firstRegister->getHighOrder(), highMR);
generateRXInstruction(cg(), TR::InstOpCode::SL, root, firstRegister->getLowOrder(), lowMR);
}
highDiff = firstRegister->getHighOrder();
root->setRegister(firstRegister);
highMR->stopUsingMemRefRegister(cg());
lowMR->stopUsingMemRefRegister(cg());
}
if (!((ENABLE_ZARCH_FOR_32 && zArchTrexsupported) || setsOrReadsCC))
{
// Check for overflow in LS int. If overflow, we are done.
generateS390BranchInstruction(cg(), TR::InstOpCode::BRC,TR::InstOpCode::COND_MASK3, root, doneLSub);
// Increment MS int due to overflow in LS int
generateRIInstruction(cg(), TR::InstOpCode::AHI, root, highDiff, -1);
generateS390LabelInstruction(cg(), TR::InstOpCode::LABEL, root, doneLSub, dependencies);
}
}
cg()->decReferenceCount(firstChild);
cg()->decReferenceCount(secondChild);
return;
}
TR::Register *OMR::X86::AMD64::TreeEvaluator::dbits2lEvaluator(TR::Node *node, TR::CodeGenerator *cg)
{
// TODO:AMD64: Peepholing
TR::Node *child = node->getFirstChild();
TR::Register *sreg = cg->evaluate(child);
TR::Register *treg = cg->allocateRegister(TR_GPR);
generateRegRegInstruction(MOVQReg8Reg, node, treg, sreg, cg);
if (node->normalizeNanValues())
{
static char *disableFastNormalizeNaNs = feGetEnv("TR_disableFastNormalizeNaNs");
if (disableFastNormalizeNaNs)
{
// This one is not clever, but it is simple, and it's based directly
// on the IA32 version which is known to work, so is safer.
//
TR::RegisterDependencyConditions *deps = generateRegisterDependencyConditions((uint8_t)0, (uint8_t)1, cg);
deps->addPostCondition(treg, TR::RealRegister::NoReg, cg);
TR::IA32ConstantDataSnippet *nan1Snippet = cg->findOrCreate8ByteConstant(node, DOUBLE_NAN_1_LOW);
TR::IA32ConstantDataSnippet *nan2Snippet = cg->findOrCreate8ByteConstant(node, DOUBLE_NAN_2_LOW);
TR::MemoryReference *nan1MR = generateX86MemoryReference(nan1Snippet, cg);
TR::MemoryReference *nan2MR = generateX86MemoryReference(nan2Snippet, cg);
TR::LabelSymbol *startLabel = TR::LabelSymbol::create(cg->trHeapMemory(),cg);
TR::LabelSymbol *normalizeLabel = TR::LabelSymbol::create(cg->trHeapMemory(),cg);
TR::LabelSymbol *endLabel = TR::LabelSymbol::create(cg->trHeapMemory(),cg);
startLabel->setStartInternalControlFlow();
endLabel ->setEndInternalControlFlow();
generateLabelInstruction( LABEL, node, startLabel, cg);
generateRegMemInstruction( CMP8RegMem, node, treg, nan1MR, cg);
generateLabelInstruction( JGE4, node, normalizeLabel, cg);
generateRegMemInstruction( CMP8RegMem, node, treg, nan2MR, cg);
generateLabelInstruction( JB4, node, endLabel, cg);
generateLabelInstruction( LABEL, node, normalizeLabel, cg);
generateRegImm64Instruction( MOV8RegImm64, node, treg, DOUBLE_NAN, cg);
generateLabelInstruction( LABEL, node, endLabel, deps, cg);
}
else
{
// A bunch of bookkeeping
//
uint64_t nanDetector = DOUBLE_NAN_2_LOW;
TR::RegisterDependencyConditions *internalControlFlowDeps = generateRegisterDependencyConditions((uint8_t)0, (uint8_t)1, cg);
internalControlFlowDeps->addPostCondition(treg, TR::RealRegister::NoReg, cg);
TR::RegisterDependencyConditions *helperDeps = generateRegisterDependencyConditions((uint8_t)1, (uint8_t)1, cg);
helperDeps->addPreCondition( treg, TR::RealRegister::eax, cg);
helperDeps->addPostCondition(treg, TR::RealRegister::eax, cg);
TR::IA32ConstantDataSnippet *nanDetectorSnippet = cg->findOrCreate8ByteConstant(node, nanDetector);
TR::MemoryReference *nanDetectorMR = generateX86MemoryReference(nanDetectorSnippet, cg);
TR::LabelSymbol *startLabel = TR::LabelSymbol::create(cg->trHeapMemory(),cg);
TR::LabelSymbol *slowPathLabel = TR::LabelSymbol::create(cg->trHeapMemory(),cg);
TR::LabelSymbol *normalizeLabel = TR::LabelSymbol::create(cg->trHeapMemory(),cg);
TR::LabelSymbol *endLabel = TR::LabelSymbol::create(cg->trHeapMemory(),cg);
startLabel->setStartInternalControlFlow();
endLabel ->setEndInternalControlFlow();
// Fast path: if subtracting nanDetector leaves CF=0 or OF=1, then it
// must be a NaN.
//
generateLabelInstruction( LABEL, node, startLabel, cg);
generateRegMemInstruction( CMP8RegMem, node, treg, nanDetectorMR, cg);
generateLabelInstruction( JAE4, node, slowPathLabel, cg);
generateLabelInstruction( JO4, node, slowPathLabel, cg);
// Slow path
//
TR_OutlinedInstructions *slowPath = new (cg->trHeapMemory()) TR_OutlinedInstructions(slowPathLabel, cg);
cg->getOutlinedInstructionsList().push_front(slowPath);
slowPath->swapInstructionListsWithCompilation();
generateLabelInstruction(NULL, LABEL, slowPathLabel, cg)->setNode(node);
generateRegImm64Instruction(MOV8RegImm64, node, treg, DOUBLE_NAN, cg);
generateLabelInstruction( JMP4, node, endLabel, cg);
slowPath->swapInstructionListsWithCompilation();
// Merge point
//
generateLabelInstruction(LABEL, node, endLabel, internalControlFlowDeps, cg);
}
}
node->setRegister(treg);
cg->decReferenceCount(child);
return treg;
}
TR::Register *TR::AMD64SystemLinkage::buildDirectDispatch(
TR::Node *callNode,
bool spillFPRegs)
{
TR::SymbolReference *methodSymRef = callNode->getSymbolReference();
TR::MethodSymbol *methodSymbol = methodSymRef->getSymbol()->castToMethodSymbol();
TR::Register *returnReg;
// Allocate adequate register dependencies.
//
// pre = number of argument registers
// post = number of volatile + return register
//
uint32_t pre = getProperties().getNumIntegerArgumentRegisters() + getProperties().getNumFloatArgumentRegisters();
uint32_t post = getProperties().getNumVolatileRegisters() + (callNode->getDataType() == TR::NoType ? 0 : 1);
#if defined (PYTHON) && 0
// Treat all preserved GP regs as volatile until register map support available.
//
post += getProperties().getNumberOfPreservedGPRegisters();
#endif
TR::RegisterDependencyConditions *preDeps = generateRegisterDependencyConditions(pre, 0, cg());
TR::RegisterDependencyConditions *postDeps = generateRegisterDependencyConditions(0, post, cg());
// Evaluate outgoing arguments on the system stack and build pre-conditions.
//
int32_t memoryArgSize = buildArgs(callNode, preDeps);
// Build post-conditions.
//
returnReg = buildVolatileAndReturnDependencies(callNode, postDeps);
postDeps->stopAddingPostConditions();
// Find the second scratch register in the post dependency list.
//
TR::Register *scratchReg = NULL;
TR::RealRegister::RegNum scratchRegIndex = getProperties().getIntegerScratchRegister(1);
for (int32_t i=0; i<post; i++)
{
if (postDeps->getPostConditions()->getRegisterDependency(i)->getRealRegister() == scratchRegIndex)
{
scratchReg = postDeps->getPostConditions()->getRegisterDependency(i)->getRegister();
break;
}
}
#if defined(PYTHON) && 0
// For Python, store the instruction that contains the GC map at this site into
// the frame object.
//
TR::SymbolReference *frameObjectSymRef =
comp()->getSymRefTab()->findOrCreateAutoSymbol(comp()->getMethodSymbol(), 0, TR::Address, true, false, true);
TR::Register *frameObjectRegister = cg()->allocateRegister();
generateRegMemInstruction(
L8RegMem,
callNode,
frameObjectRegister,
generateX86MemoryReference(frameObjectSymRef, cg()),
cg());
TR::RealRegister *espReal = cg()->machine()->getX86RealRegister(TR::RealRegister::esp);
TR::Register *gcMapPCRegister = cg()->allocateRegister();
generateRegMemInstruction(
LEA8RegMem,
callNode,
gcMapPCRegister,
generateX86MemoryReference(espReal, -8, cg()),
cg());
// Use "volatile" registers across the call. Once proper register map support
// is implemented, r14 and r15 will no longer be volatile and a different pair
// should be chosen.
//
TR::RegisterDependencyConditions *gcMapDeps = generateRegisterDependencyConditions(0, 2, cg());
gcMapDeps->addPostCondition(frameObjectRegister, TR::RealRegister::r14, cg());
gcMapDeps->addPostCondition(gcMapPCRegister, TR::RealRegister::r15, cg());
gcMapDeps->stopAddingPostConditions();
generateMemRegInstruction(
S8MemReg,
callNode,
generateX86MemoryReference(frameObjectRegister, fe()->getPythonGCMapPCOffsetInFrame(), cg()),
gcMapPCRegister,
gcMapDeps,
cg());
cg()->stopUsingRegister(frameObjectRegister);
cg()->stopUsingRegister(gcMapPCRegister);
#endif
TR::Instruction *instr;
if (methodSymbol->getMethodAddress())
{
TR_ASSERT(scratchReg, "could not find second scratch register");
auto LoadRegisterInstruction = generateRegImm64SymInstruction(
MOV8RegImm64,
callNode,
scratchReg,
(uintptr_t)methodSymbol->getMethodAddress(),
methodSymRef,
cg());
if (TR::Options::getCmdLineOptions()->getOption(TR_EmitRelocatableELFFile))
{
LoadRegisterInstruction->setReloKind(TR_NativeMethodAbsolute);
}
instr = generateRegInstruction(CALLReg, callNode, scratchReg, preDeps, cg());
}
else
{
instr = generateImmSymInstruction(CALLImm4, callNode, (uintptrj_t)methodSymbol->getMethodAddress(), methodSymRef, preDeps, cg());
}
cg()->resetIsLeafMethod();
instr->setNeedsGCMap(getProperties().getPreservedRegisterMapForGC());
cg()->stopUsingRegister(scratchReg);
TR::LabelSymbol *postDepLabel = generateLabelSymbol(cg());
generateLabelInstruction(LABEL, callNode, postDepLabel, postDeps, cg());
return returnReg;
}
TR::RegisterDependencyConditions *
OMR::Power::RegisterDependencyConditions::clone(
TR::CodeGenerator *cg,
TR::RegisterDependencyConditions *added)
{
TR::RegisterDependencyConditions *result;
TR::RegisterDependency *singlePair;
int32_t idx, preNum, postNum, addPre=0, addPost=0;
#if defined(DEBUG)
int32_t preGPR=0, postGPR=0;
#endif
if (added != NULL)
{
addPre = added->getAddCursorForPre();
addPost = added->getAddCursorForPost();
}
preNum = this->getAddCursorForPre();
postNum = this->getAddCursorForPost();
result = new (cg->trHeapMemory()) TR::RegisterDependencyConditions(getNumPreConditions()+addPre,
getNumPostConditions()+addPost, cg->trMemory());
for (idx=0; idx<postNum; idx++)
{
singlePair = this->getPostConditions()->getRegisterDependency(idx);
//eliminate duplicate virtual->NoReg dependencies in 'this' conditions
if (!((TR::RealRegister::NoReg == singlePair->getRealRegister()) && (added->postConditionContainsVirtual(singlePair->getRegister()))))
result->addPostCondition(singlePair->getRegister(), singlePair->getRealRegister(),
singlePair->getFlags());
#if defined(DEBUG)
if (singlePair->getRegister()->getKind() == TR_GPR)
postGPR++;
#endif
}
for (idx=0; idx<preNum; idx++)
{
singlePair = this->getPreConditions()->getRegisterDependency(idx);
//eliminate duplicate virtual->NoReg dependencies in 'this' conditions
if (!((TR::RealRegister::NoReg == singlePair->getRealRegister()) && (added->preConditionContainsVirtual(singlePair->getRegister()))))
result->addPreCondition(singlePair->getRegister(), singlePair->getRealRegister(),
singlePair->getFlags());
#if defined(DEBUG)
if (singlePair->getRegister()->getKind() == TR_GPR)
preGPR++;
#endif
}
for (idx=0; idx<addPost; idx++)
{
singlePair = added->getPostConditions()->getRegisterDependency(idx);
result->addPostCondition(singlePair->getRegister(), singlePair->getRealRegister(),
singlePair->getFlags());
#if defined(DEBUG)
if (singlePair->getRegister()->getKind() == TR_GPR)
postGPR++;
#endif
}
for (idx=0; idx<addPre; idx++)
{
singlePair = added->getPreConditions()->getRegisterDependency(idx);
result->addPreCondition(singlePair->getRegister(), singlePair->getRealRegister(),
singlePair->getFlags());
#if defined(DEBUG)
if (singlePair->getRegister()->getKind() == TR_GPR)
preGPR++;
#endif
}
#if defined(DEBUG)
int32_t max_gpr = cg->getProperties().getNumAllocatableIntegerRegisters();
TR_ASSERT(preGPR<=max_gpr && postGPR<=max_gpr, "Over the limit of available global regsiters.");
#endif
return result;
}
TR::RegisterDependencyConditions* TR_PPCScratchRegisterDependencyConditions::createDependencyConditions(TR::CodeGenerator *cg,
TR_PPCScratchRegisterDependencyConditions *pre,
TR_PPCScratchRegisterDependencyConditions *post)
{
int32_t preCount = pre ? pre->getNumberOfDependencies() : 0;
int32_t postCount = post ? post->getNumberOfDependencies() : 0;
TR_LiveRegisters *lrVector = cg->getLiveRegisters(TR_VSX_VECTOR);
bool liveVSXVectorReg = (!lrVector || (lrVector->getNumberOfLiveRegisters() > 0));
TR_LiveRegisters *lrScalar = cg->getLiveRegisters(TR_VSX_SCALAR);
bool liveVSXScalarReg = (!lrScalar || (lrScalar->getNumberOfLiveRegisters() > 0));
if (liveVSXVectorReg)
{
preCount += 64;
postCount += 64;
}
else if (liveVSXScalarReg)
{
preCount += 32;
postCount += 32;
}
TR::RegisterDependencyConditions *dependencies = new (cg->trHeapMemory()) TR::RegisterDependencyConditions(preCount,
postCount,
cg->trMemory());
for (int i = 0; i < (pre ? pre->_numGPRDeps : 0); ++i)
{
dependencies->addPreCondition(pre->_gprDeps[i].getRegister(), pre->_gprDeps[i].getRealRegister(), pre->_gprDeps[i].getFlags());
if (pre->_excludeGPR0 & (1 << i))
dependencies->getPreConditions()->getRegisterDependency(i)->setExcludeGPR0();
}
for (int i = 0; i < (post ? post->_numGPRDeps : 0); ++i)
{
dependencies->addPostCondition(post->_gprDeps[i].getRegister(), post->_gprDeps[i].getRealRegister(), post->_gprDeps[i].getFlags());
if (post->_excludeGPR0 & (1 << i))
dependencies->getPostConditions()->getRegisterDependency(i)->setExcludeGPR0();
}
for (int i = 0; i < (pre ? pre->_numCCRDeps : 0); ++i)
{
dependencies->addPreCondition(pre->_ccrDeps[i].getRegister(), pre->_ccrDeps[i].getRealRegister(), pre->_ccrDeps[i].getFlags());
}
for (int i = 0; i < (post ? post->_numCCRDeps : 0); ++i)
{
dependencies->addPostCondition(post->_ccrDeps[i].getRegister(), post->_ccrDeps[i].getRealRegister(), post->_ccrDeps[i].getFlags());
}
const TR_PPCLinkageProperties& properties = cg->getLinkage()->getProperties();
if (liveVSXVectorReg)
{
for (int32_t i=TR::RealRegister::FirstVSR; i<=TR::RealRegister::LastVSR; i++)
{
if (!properties.getPreserved((TR::RealRegister::RegNum)i))
{
TR::Register *vreg = cg->allocateRegister(TR_FPR);
vreg->setPlaceholderReg();
dependencies->addPreCondition(vreg, (TR::RealRegister::RegNum)i);
dependencies->addPostCondition(vreg, (TR::RealRegister::RegNum)i);
}
}
}
else
{
if (liveVSXScalarReg)
{
for (int32_t i=TR::RealRegister::vsr32; i<=TR::RealRegister::LastVSR; i++)
{
if (!properties.getPreserved((TR::RealRegister::RegNum)i))
{
TR::Register *vreg = cg->allocateRegister(TR_FPR);
vreg->setPlaceholderReg();
dependencies->addPreCondition(vreg, (TR::RealRegister::RegNum)i);
dependencies->addPostCondition(vreg, (TR::RealRegister::RegNum)i);
}
}
}
}
return dependencies;
}
void OMR::Power::RegisterDependencyConditions::bookKeepingRegisterUses(TR::Instruction *instr, TR::CodeGenerator *cg)
{
// *this swipeable for debugging purposes
TR::Register *virtReg;
TR::RealRegister::RegNum regNum;
TR::RegisterDependencyConditions *assoc;
int numAssoc = 0;
if (instr->getOpCodeValue() == TR::InstOpCode::assocreg)
return;
// Don't track associations or emit assocregs in outlined code
// Register assigner can save/restore associations across outlined sections properly, however no such mechanism exists for instruction selection
// so we don't want these associations to clobber the associations that were set in main line code, which are more important
// TODO: Fix this by saving/restoring the associations in swapInstructionListsWithCompilation()
bool isOOL = cg->getIsInOOLSection();
TR::Machine *machine = cg->machine();
assoc = !isOOL ? new (cg->trHeapMemory()) TR::RegisterDependencyConditions(0,_addCursorForPre, cg->trMemory()) : 0;
for (int i = 0; i < _addCursorForPre; i++)
{
virtReg = _preConditions->getRegisterDependency(i)->getRegister();
regNum = _preConditions->getRegisterDependency(i)->getRealRegister();
if (!isOOL)
{
// Add to the association condition when the association is changed
// from one virtual register to another
if (machine->getVirtualAssociatedWithReal(regNum) != virtReg && machine->getVirtualAssociatedWithReal(regNum) != 0)
{
assoc->addPostCondition(machine->getVirtualAssociatedWithReal(regNum), regNum);
numAssoc++;
}
// Keep track of the virtual register map to real registers!
machine->setVirtualAssociatedWithReal(regNum, virtReg);
}
instr->useRegister(virtReg);
if (!isOOL)
{
cg->setRealRegisterAssociation(virtReg, regNum);
if (_preConditions->getRegisterDependency(i)->getExcludeGPR0())
cg->addRealRegisterInterference(virtReg, TR::RealRegister::gr0);
}
}
if (numAssoc > 0)
{
// Emit an AssocRegs instruction to track the previous association
assoc->setNumPostConditions(numAssoc, cg->trMemory());
generateDepInstruction(cg, TR::InstOpCode::assocreg, instr->getNode(), assoc, instr->getPrev());
}
for (int j = 0; j < _addCursorForPost; j++)
{
virtReg = _postConditions->getRegisterDependency(j)->getRegister();
regNum = _postConditions->getRegisterDependency(j)->getRealRegister();
instr->useRegister(virtReg);
if (!isOOL)
{
cg->setRealRegisterAssociation(virtReg, regNum);
if (_postConditions->getRegisterDependency(j)->getExcludeGPR0())
cg->addRealRegisterInterference(virtReg, TR::RealRegister::gr0);
}
}
}
TR::Register *TR_X86FPCompareAnalyser::fpCompareAnalyser(TR::Node *root,
TR_X86OpCodes cmpRegRegOpCode,
TR_X86OpCodes cmpRegMemOpCode,
TR_X86OpCodes cmpiRegRegOpCode,
bool useFCOMIInstructions)
{
TR::Node *firstChild,
*secondChild;
TR::ILOpCodes cmpOp = root->getOpCodeValue();
bool reverseMemOp = false;
bool reverseCmpOp = false;
TR::Compilation* comp = _cg->comp();
TR_X86OpCodes cmpInstr = useFCOMIInstructions ? cmpiRegRegOpCode : cmpRegRegOpCode;
// Some operators must have their operands swapped to improve the generated
// code needed to evaluate the result of the comparison.
//
bool mustSwapOperands = (cmpOp == TR::iffcmple ||
cmpOp == TR::ifdcmple ||
cmpOp == TR::iffcmpgtu ||
cmpOp == TR::ifdcmpgtu ||
cmpOp == TR::fcmple ||
cmpOp == TR::dcmple ||
cmpOp == TR::fcmpgtu ||
cmpOp == TR::dcmpgtu ||
(useFCOMIInstructions &&
(cmpOp == TR::iffcmplt ||
cmpOp == TR::ifdcmplt ||
cmpOp == TR::iffcmpgeu ||
cmpOp == TR::ifdcmpgeu ||
cmpOp == TR::fcmplt ||
cmpOp == TR::dcmplt ||
cmpOp == TR::fcmpgeu ||
cmpOp == TR::dcmpgeu))) ? true : false;
// Some operators should not have their operands swapped to improve the generated
// code needed to evaluate the result of the comparison.
//
bool preventOperandSwapping = (cmpOp == TR::iffcmpltu ||
cmpOp == TR::ifdcmpltu ||
cmpOp == TR::iffcmpge ||
cmpOp == TR::ifdcmpge ||
cmpOp == TR::fcmpltu ||
cmpOp == TR::dcmpltu ||
cmpOp == TR::fcmpge ||
cmpOp == TR::dcmpge ||
(useFCOMIInstructions &&
(cmpOp == TR::iffcmpgt ||
cmpOp == TR::ifdcmpgt ||
cmpOp == TR::iffcmpleu ||
cmpOp == TR::ifdcmpleu ||
cmpOp == TR::fcmpgt ||
cmpOp == TR::dcmpgt ||
cmpOp == TR::fcmpleu ||
cmpOp == TR::dcmpleu))) ? true : false;
// For correctness, don't swap operands of these operators.
//
if (cmpOp == TR::fcmpg || cmpOp == TR::fcmpl ||
cmpOp == TR::dcmpg || cmpOp == TR::dcmpl)
{
preventOperandSwapping = true;
}
// Initial operand evaluation ordering.
//
if (preventOperandSwapping || (!mustSwapOperands && _cg->whichChildToEvaluate(root) == 0))
{
firstChild = root->getFirstChild();
secondChild = root->getSecondChild();
setReversedOperands(false);
}
else
{
firstChild = root->getSecondChild();
secondChild = root->getFirstChild();
setReversedOperands(true);
}
TR::Register *firstRegister = firstChild->getRegister();
TR::Register *secondRegister = secondChild->getRegister();
setInputs(firstChild,
firstRegister,
secondChild,
secondRegister,
useFCOMIInstructions,
// If either 'preventOperandSwapping' or 'mustSwapOperands' is set then the
// initial operand ordering set above must be maintained.
//
preventOperandSwapping || mustSwapOperands);
// Make sure any required operand ordering is respected.
//
if ((getCmpReg2Reg1() || getCmpReg2Mem1()) &&
(mustSwapOperands || preventOperandSwapping))
{
reverseCmpOp = getCmpReg2Reg1() ? true : false;
reverseMemOp = getCmpReg2Mem1() ? true : false;
}
// If we are not comparing with a memory operand, one of them evaluates
// to a zero, and the zero is not already on the stack, then we can use
// FTST to save a register.
//
// (With a memory operand, either the constant zero needs to be loaded
// to use FCOM, or the memory operand needs to be loaded to use FTST,
// so there is no gain in using FTST.)
//
// If the constant zero is in the target register, using FTST means the
// comparison will be reversed. We cannot do this if the initial ordering
// of the operands must be maintained.
//
// Finally, if FTST is used and this is the last use of the target, the
// target register may need to be explicitly popped.
//
TR::Register *targetRegisterForFTST = NULL;
TR::Node *targetChildForFTST = NULL;
if (getEvalChild1() && isUnevaluatedZero(firstChild)) // do we need getEvalChild1() here?
{
if ( ((getCmpReg1Reg2() || reverseCmpOp) && !(preventOperandSwapping || mustSwapOperands)) ||
(getCmpReg2Reg1() && !reverseCmpOp))
{
if (getEvalChild2())
{
secondRegister = _cg->evaluate(secondChild);
}
targetRegisterForFTST = secondRegister;
targetChildForFTST = secondChild;
notReversedOperands();
}
}
else if (getEvalChild2() && isUnevaluatedZero(secondChild)) // do we need getEvalChild2() here?
{
if ( (getCmpReg1Reg2() || reverseCmpOp) ||
(getCmpReg2Reg1() && !reverseCmpOp && !(preventOperandSwapping || mustSwapOperands)) )
{
if (getEvalChild1())
{
firstRegister = _cg->evaluate(firstChild);
}
targetRegisterForFTST = firstRegister;
targetChildForFTST = firstChild;
}
}
if (!targetRegisterForFTST)
{
// If we have a choice, evaluate the target operand last. By doing so, we
// help out the register assigner because the target must be TOS. This
// avoids an unneccessary FXCH for the target.
//
if (getEvalChild1() && getEvalChild2())
{
if (getCmpReg1Reg2() || getCmpReg1Mem2())
{
secondRegister = _cg->evaluate(secondChild);
firstRegister = _cg->evaluate(firstChild);
}
else
{
firstRegister = _cg->evaluate(firstChild);
secondRegister = _cg->evaluate(secondChild);
}
}
else
{
if (getEvalChild1())
{
firstRegister = _cg->evaluate(firstChild);
}
if (getEvalChild2())
{
secondRegister = _cg->evaluate(secondChild);
}
}
}
// Adjust the FP precision of feeding operands.
//
if (firstRegister &&
(firstRegister->needsPrecisionAdjustment() ||
comp->getOption(TR_StrictFPCompares) ||
(firstRegister->mayNeedPrecisionAdjustment() && secondChild->getOpCode().isLoadConst()) ||
(firstRegister->mayNeedPrecisionAdjustment() && !secondRegister)))
{
TR::TreeEvaluator::insertPrecisionAdjustment(firstRegister, root, _cg);
}
if (secondRegister &&
(secondRegister->needsPrecisionAdjustment() ||
comp->getOption(TR_StrictFPCompares) ||
(secondRegister->mayNeedPrecisionAdjustment() && firstChild->getOpCode().isLoadConst()) ||
(secondRegister->mayNeedPrecisionAdjustment() && !firstRegister)))
{
TR::TreeEvaluator::insertPrecisionAdjustment(secondRegister, root, _cg);
}
// Generate the compare instruction.
//
if (targetRegisterForFTST)
{
generateFPRegInstruction(FTSTReg, root, targetRegisterForFTST, _cg);
}
else if (!useFCOMIInstructions && (getCmpReg1Mem2() || reverseMemOp))
{
TR::MemoryReference *tempMR = generateX86MemoryReference(secondChild, _cg);
generateFPRegMemInstruction(cmpRegMemOpCode, root, firstRegister, tempMR, _cg);
tempMR->decNodeReferenceCounts(_cg);
}
else if (!useFCOMIInstructions && getCmpReg2Mem1())
{
TR::MemoryReference *tempMR = generateX86MemoryReference(firstChild, _cg);
generateFPRegMemInstruction(cmpRegMemOpCode, root, secondRegister, tempMR, _cg);
notReversedOperands();
tempMR->decNodeReferenceCounts(_cg);
}
else if (getCmpReg1Reg2() || reverseCmpOp)
{
generateFPCompareRegRegInstruction(cmpInstr, root, firstRegister, secondRegister, _cg);
}
else if (getCmpReg2Reg1())
{
generateFPCompareRegRegInstruction(cmpInstr, root, secondRegister, firstRegister, _cg);
notReversedOperands();
}
_cg->decReferenceCount(firstChild);
_cg->decReferenceCount(secondChild);
// Evaluate the comparison.
//
if (getReversedOperands())
{
cmpOp = TR::ILOpCode(cmpOp).getOpCodeForSwapChildren();
TR::Node::recreate(root, cmpOp);
}
if (useFCOMIInstructions && !targetRegisterForFTST)
{
return NULL;
}
// We must manually move the FP condition flags to the EFLAGS register if we don't
// use the FCOMI instructions.
//
TR::Register *accRegister = _cg->allocateRegister();
TR::RegisterDependencyConditions *dependencies = generateRegisterDependencyConditions((uint8_t)1, 1, _cg);
dependencies->addPreCondition(accRegister, TR::RealRegister::eax, _cg);
dependencies->addPostCondition(accRegister, TR::RealRegister::eax, _cg);
generateRegInstruction(STSWAcc, root, accRegister, dependencies, _cg);
// Pop the FTST target register if it is not used any more.
//
if (targetRegisterForFTST &&
targetChildForFTST && targetChildForFTST->getReferenceCount() == 0)
{
generateFPSTiST0RegRegInstruction(FSTRegReg, root, targetRegisterForFTST, targetRegisterForFTST, _cg);
}
return accRegister;
}