//------------------------------------------------------------------------
// TreeNodeInfoInitPutArgStk: Set the NodeInfo for a GT_PUTARG_STK node
//
// Arguments:
// argNode - a GT_PUTARG_STK node
//
// Return Value:
// None.
//
// Notes:
// Set the child node(s) to be contained when we have a multireg arg
//
void Lowering::TreeNodeInfoInitPutArgStk(GenTreePutArgStk* argNode, fgArgTabEntryPtr info)
{
assert(argNode->gtOper == GT_PUTARG_STK);
GenTreePtr putArgChild = argNode->gtOp.gtOp1;
// Initialize 'argNode' as not contained, as this is both the default case
// and how MakeSrcContained expects to find things setup.
//
argNode->gtLsraInfo.srcCount = 1;
argNode->gtLsraInfo.dstCount = 0;
// Do we have a TYP_STRUCT argument (or a GT_FIELD_LIST), if so it must be a multireg pass-by-value struct
if ((putArgChild->TypeGet() == TYP_STRUCT) || (putArgChild->OperGet() == GT_FIELD_LIST))
{
// We will use store instructions that each write a register sized value
if (putArgChild->OperGet() == GT_FIELD_LIST)
{
// We consume all of the items in the GT_FIELD_LIST
argNode->gtLsraInfo.srcCount = info->numSlots;
putArgChild->SetContained();
}
else
{
#ifdef _TARGET_ARM64_
// We could use a ldp/stp sequence so we need two internal registers
argNode->gtLsraInfo.internalIntCount = 2;
#else // _TARGET_ARM_
// We could use a ldr/str sequence so we need a internal register
argNode->gtLsraInfo.internalIntCount = 1;
#endif // _TARGET_ARM_
if (putArgChild->OperGet() == GT_OBJ)
{
GenTreePtr objChild = putArgChild->gtOp.gtOp1;
if (objChild->OperGet() == GT_LCL_VAR_ADDR)
{
// We will generate all of the code for the GT_PUTARG_STK, the GT_OBJ and the GT_LCL_VAR_ADDR
// as one contained operation
//
MakeSrcContained(putArgChild, objChild);
}
}
// We will generate all of the code for the GT_PUTARG_STK and it's child node
// as one contained operation
//
MakeSrcContained(argNode, putArgChild);
}
}
else
{
// We must not have a multi-reg struct
assert(info->numSlots == 1);
}
}
//------------------------------------------------------------------------
// TreeNodeInfoInitShiftRotate: Set the NodeInfo for a shift or rotate.
//
// Arguments:
// tree - The node of interest
//
// Return Value:
// None.
//
void Lowering::TreeNodeInfoInitShiftRotate(GenTree* tree)
{
TreeNodeInfo* info = &(tree->gtLsraInfo);
LinearScan* l = m_lsra;
info->srcCount = 2;
info->dstCount = 1;
GenTreePtr shiftBy = tree->gtOp.gtOp2;
GenTreePtr source = tree->gtOp.gtOp1;
if (shiftBy->IsCnsIntOrI())
{
MakeSrcContained(tree, shiftBy);
}
#ifdef _TARGET_ARM_
// The first operand of a GT_LSH_HI and GT_RSH_LO oper is a GT_LONG so that
// we can have a three operand form. Increment the srcCount.
if (tree->OperGet() == GT_LSH_HI || tree->OperGet() == GT_RSH_LO)
{
assert(source->OperGet() == GT_LONG);
source->SetContained();
info->srcCount++;
if (tree->OperGet() == GT_LSH_HI)
{
GenTreePtr sourceLo = source->gtOp.gtOp1;
sourceLo->gtLsraInfo.isDelayFree = true;
}
else
{
GenTreePtr sourceHi = source->gtOp.gtOp2;
sourceHi->gtLsraInfo.isDelayFree = true;
}
source->gtLsraInfo.hasDelayFreeSrc = true;
info->hasDelayFreeSrc = true;
}
#endif // _TARGET_ARM_
}
//------------------------------------------------------------------------
// TreeNodeInfoInitBlockStore: Set the NodeInfo for a block store.
//
// Arguments:
// blkNode - The block store node of interest
//
// Return Value:
// None.
//
void Lowering::TreeNodeInfoInitBlockStore(GenTreeBlk* blkNode)
{
GenTree* dstAddr = blkNode->Addr();
unsigned size = blkNode->gtBlkSize;
GenTree* source = blkNode->Data();
LinearScan* l = m_lsra;
Compiler* compiler = comp;
// Sources are dest address and initVal or source.
// We may require an additional source or temp register for the size.
blkNode->gtLsraInfo.srcCount = 2;
blkNode->gtLsraInfo.dstCount = 0;
GenTreePtr srcAddrOrFill = nullptr;
bool isInitBlk = blkNode->OperIsInitBlkOp();
if (!isInitBlk)
{
// CopyObj or CopyBlk
if (source->gtOper == GT_IND)
{
srcAddrOrFill = blkNode->Data()->gtGetOp1();
// We're effectively setting source as contained, but can't call MakeSrcContained, because the
// "inheritance" of the srcCount is to a child not a parent - it would "just work" but could be misleading.
// If srcAddr is already non-contained, we don't need to change it.
if (srcAddrOrFill->gtLsraInfo.getDstCount() == 0)
{
srcAddrOrFill->gtLsraInfo.setDstCount(1);
srcAddrOrFill->gtLsraInfo.setSrcCount(source->gtLsraInfo.srcCount);
}
m_lsra->clearOperandCounts(source);
source->SetContained();
source->AsIndir()->Addr()->ClearContained();
}
else if (!source->IsMultiRegCall() && !source->OperIsSIMD())
{
assert(source->IsLocal());
MakeSrcContained(blkNode, source);
blkNode->gtLsraInfo.srcCount--;
}
}
if (isInitBlk)
{
GenTreePtr initVal = source;
if (initVal->OperIsInitVal())
{
initVal->SetContained();
initVal = initVal->gtGetOp1();
}
srcAddrOrFill = initVal;
if (blkNode->gtBlkOpKind == GenTreeBlk::BlkOpKindUnroll)
{
// TODO-ARM-CQ: Currently we generate a helper call for every
// initblk we encounter. Later on we should implement loop unrolling
// code sequences to improve CQ.
// For reference see the code in lsraxarch.cpp.
NYI_ARM("initblk loop unrolling is currently not implemented.");
#ifdef _TARGET_ARM64_
// No additional temporaries required
ssize_t fill = initVal->gtIntCon.gtIconVal & 0xFF;
if (fill == 0)
{
MakeSrcContained(blkNode, source);
blkNode->gtLsraInfo.srcCount--;
}
#endif // _TARGET_ARM64_
}
else
{
assert(blkNode->gtBlkOpKind == GenTreeBlk::BlkOpKindHelper);
// The helper follows the regular ABI.
dstAddr->gtLsraInfo.setSrcCandidates(l, RBM_ARG_0);
initVal->gtLsraInfo.setSrcCandidates(l, RBM_ARG_1);
if (size != 0)
{
// Reserve a temp register for the block size argument.
blkNode->gtLsraInfo.setInternalCandidates(l, RBM_ARG_2);
blkNode->gtLsraInfo.internalIntCount = 1;
}
else
{
// The block size argument is a third argument to GT_STORE_DYN_BLK
noway_assert(blkNode->gtOper == GT_STORE_DYN_BLK);
blkNode->gtLsraInfo.setSrcCount(3);
GenTree* sizeNode = blkNode->AsDynBlk()->gtDynamicSize;
sizeNode->gtLsraInfo.setSrcCandidates(l, RBM_ARG_2);
}
}
}
else
{
// CopyObj or CopyBlk
// Sources are src and dest and size if not constant.
if (blkNode->OperGet() == GT_STORE_OBJ)
{
// CopyObj
// We don't need to materialize the struct size but we still need
// a temporary register to perform the sequence of loads and stores.
blkNode->gtLsraInfo.internalIntCount = 1;
if (size >= 2 * REGSIZE_BYTES)
{
// We will use ldp/stp to reduce code size and improve performance
// so we need to reserve an extra internal register
blkNode->gtLsraInfo.internalIntCount++;
}
// We can't use the special Write Barrier registers, so exclude them from the mask
regMaskTP internalIntCandidates = RBM_ALLINT & ~(RBM_WRITE_BARRIER_DST_BYREF | RBM_WRITE_BARRIER_SRC_BYREF);
blkNode->gtLsraInfo.setInternalCandidates(l, internalIntCandidates);
// If we have a dest address we want it in RBM_WRITE_BARRIER_DST_BYREF.
dstAddr->gtLsraInfo.setSrcCandidates(l, RBM_WRITE_BARRIER_DST_BYREF);
// If we have a source address we want it in REG_WRITE_BARRIER_SRC_BYREF.
// Otherwise, if it is a local, codegen will put its address in REG_WRITE_BARRIER_SRC_BYREF,
// which is killed by a StoreObj (and thus needn't be reserved).
if (srcAddrOrFill != nullptr)
{
srcAddrOrFill->gtLsraInfo.setSrcCandidates(l, RBM_WRITE_BARRIER_SRC_BYREF);
}
}
else
{
// CopyBlk
short internalIntCount = 0;
regMaskTP internalIntCandidates = RBM_NONE;
if (blkNode->gtBlkOpKind == GenTreeBlk::BlkOpKindUnroll)
{
// TODO-ARM-CQ: cpblk loop unrolling is currently not implemented.
// In case of a CpBlk with a constant size and less than CPBLK_UNROLL_LIMIT size
// we should unroll the loop to improve CQ.
// For reference see the code in lsraxarch.cpp.
NYI_ARM("cpblk loop unrolling is currently not implemented.");
#ifdef _TARGET_ARM64_
internalIntCount = 1;
internalIntCandidates = RBM_ALLINT;
if (size >= 2 * REGSIZE_BYTES)
{
// We will use ldp/stp to reduce code size and improve performance
// so we need to reserve an extra internal register
internalIntCount++;
}
#endif // _TARGET_ARM64_
}
else
{
assert(blkNode->gtBlkOpKind == GenTreeBlk::BlkOpKindHelper);
dstAddr->gtLsraInfo.setSrcCandidates(l, RBM_ARG_0);
// The srcAddr goes in arg1.
if (srcAddrOrFill != nullptr)
{
srcAddrOrFill->gtLsraInfo.setSrcCandidates(l, RBM_ARG_1);
}
if (size != 0)
{
// Reserve a temp register for the block size argument.
internalIntCandidates |= RBM_ARG_2;
internalIntCount++;
}
else
{
// The block size argument is a third argument to GT_STORE_DYN_BLK
noway_assert(blkNode->gtOper == GT_STORE_DYN_BLK);
blkNode->gtLsraInfo.setSrcCount(3);
GenTree* blockSize = blkNode->AsDynBlk()->gtDynamicSize;
blockSize->gtLsraInfo.setSrcCandidates(l, RBM_ARG_2);
}
}
if (internalIntCount != 0)
{
blkNode->gtLsraInfo.internalIntCount = internalIntCount;
blkNode->gtLsraInfo.setInternalCandidates(l, internalIntCandidates);
}
}
}
}
//------------------------------------------------------------------------
// TreeNodeInfoInitPutArgSplit: Set the NodeInfo for a GT_PUTARG_SPLIT node
//
// Arguments:
// argNode - a GT_PUTARG_SPLIT node
//
// Return Value:
// None.
//
// Notes:
// Set the child node(s) to be contained
//
void Lowering::TreeNodeInfoInitPutArgSplit(GenTreePutArgSplit* argNode, TreeNodeInfo& info, fgArgTabEntryPtr argInfo)
{
assert(argNode->gtOper == GT_PUTARG_SPLIT);
GenTreePtr putArgChild = argNode->gtOp.gtOp1;
// Registers for split argument corresponds to source
argNode->gtLsraInfo.dstCount = argInfo->numRegs;
info.srcCount += argInfo->numRegs;
regNumber argReg = argInfo->regNum;
regMaskTP argMask = RBM_NONE;
for (unsigned i = 0; i < argInfo->numRegs; i++)
{
argMask |= genRegMask((regNumber)((unsigned)argReg + i));
}
argNode->gtLsraInfo.setDstCandidates(m_lsra, argMask);
if (putArgChild->OperGet() == GT_FIELD_LIST)
{
// Generated code:
// 1. Consume all of the items in the GT_FIELD_LIST (source)
// 2. Store to target slot and move to target registers (destination) from source
//
argNode->gtLsraInfo.srcCount = argInfo->numRegs + argInfo->numSlots;
// To avoid redundant moves, have the argument operand computed in the
// register in which the argument is passed to the call.
GenTreeFieldList* fieldListPtr = putArgChild->AsFieldList();
for (unsigned idx = 0; fieldListPtr != nullptr; fieldListPtr = fieldListPtr->Rest(), idx++)
{
if (idx < argInfo->numRegs)
{
GenTreePtr node = fieldListPtr->gtGetOp1();
node->gtLsraInfo.setSrcCandidates(m_lsra, genRegMask((regNumber)((unsigned)argReg + idx)));
}
}
putArgChild->SetContained();
}
else
{
assert(putArgChild->TypeGet() == TYP_STRUCT);
assert(putArgChild->OperGet() == GT_OBJ);
// We could use a ldr/str sequence so we need a internal register
argNode->gtLsraInfo.srcCount = 1;
argNode->gtLsraInfo.internalIntCount = 1;
regMaskTP internalMask = RBM_ALLINT & ~argMask;
argNode->gtLsraInfo.setInternalCandidates(m_lsra, internalMask);
GenTreePtr objChild = putArgChild->gtOp.gtOp1;
if (objChild->OperGet() == GT_LCL_VAR_ADDR)
{
// We will generate all of the code for the GT_PUTARG_SPLIT, the GT_OBJ and the GT_LCL_VAR_ADDR
// as one contained operation
//
MakeSrcContained(putArgChild, objChild);
putArgChild->gtLsraInfo.srcCount--;
}
argNode->gtLsraInfo.srcCount = putArgChild->gtLsraInfo.srcCount;
MakeSrcContained(argNode, putArgChild);
}
}
//------------------------------------------------------------------------
// TreeNodeInfoInitCall: Set the NodeInfo for a call.
//
// Arguments:
// call - The call node of interest
//
// Return Value:
// None.
//
void Lowering::TreeNodeInfoInitCall(GenTreeCall* call)
{
TreeNodeInfo* info = &(call->gtLsraInfo);
LinearScan* l = m_lsra;
Compiler* compiler = comp;
bool hasMultiRegRetVal = false;
ReturnTypeDesc* retTypeDesc = nullptr;
info->srcCount = 0;
if (call->TypeGet() != TYP_VOID)
{
hasMultiRegRetVal = call->HasMultiRegRetVal();
if (hasMultiRegRetVal)
{
// dst count = number of registers in which the value is returned by call
retTypeDesc = call->GetReturnTypeDesc();
info->dstCount = retTypeDesc->GetReturnRegCount();
}
else
{
info->dstCount = 1;
}
}
else
{
info->dstCount = 0;
}
GenTree* ctrlExpr = call->gtControlExpr;
if (call->gtCallType == CT_INDIRECT)
{
// either gtControlExpr != null or gtCallAddr != null.
// Both cannot be non-null at the same time.
assert(ctrlExpr == nullptr);
assert(call->gtCallAddr != nullptr);
ctrlExpr = call->gtCallAddr;
}
// set reg requirements on call target represented as control sequence.
if (ctrlExpr != nullptr)
{
// we should never see a gtControlExpr whose type is void.
assert(ctrlExpr->TypeGet() != TYP_VOID);
info->srcCount++;
// In case of fast tail implemented as jmp, make sure that gtControlExpr is
// computed into a register.
if (call->IsFastTailCall())
{
NYI_ARM("tail call");
#ifdef _TARGET_ARM64_
// Fast tail call - make sure that call target is always computed in IP0
// so that epilog sequence can generate "br xip0" to achieve fast tail call.
ctrlExpr->gtLsraInfo.setSrcCandidates(l, genRegMask(REG_IP0));
#endif // _TARGET_ARM64_
}
}
#ifdef _TARGET_ARM_
else
{
info->internalIntCount = 1;
}
#endif // _TARGET_ARM_
RegisterType registerType = call->TypeGet();
// Set destination candidates for return value of the call.
#ifdef _TARGET_ARM_
if (call->IsHelperCall(compiler, CORINFO_HELP_INIT_PINVOKE_FRAME))
{
// The ARM CORINFO_HELP_INIT_PINVOKE_FRAME helper uses a custom calling convention that returns with
// TCB in REG_PINVOKE_TCB. fgMorphCall() sets the correct argument registers.
info->setDstCandidates(l, RBM_PINVOKE_TCB);
}
else
#endif // _TARGET_ARM_
if (hasMultiRegRetVal)
{
assert(retTypeDesc != nullptr);
info->setDstCandidates(l, retTypeDesc->GetABIReturnRegs());
}
else if (varTypeIsFloating(registerType))
{
info->setDstCandidates(l, RBM_FLOATRET);
}
else if (registerType == TYP_LONG)
{
info->setDstCandidates(l, RBM_LNGRET);
}
else
{
info->setDstCandidates(l, RBM_INTRET);
}
// If there is an explicit this pointer, we don't want that node to produce anything
// as it is redundant
if (call->gtCallObjp != nullptr)
{
GenTreePtr thisPtrNode = call->gtCallObjp;
if (thisPtrNode->gtOper == GT_PUTARG_REG)
{
l->clearOperandCounts(thisPtrNode);
thisPtrNode->SetContained();
l->clearDstCount(thisPtrNode->gtOp.gtOp1);
}
else
{
l->clearDstCount(thisPtrNode);
}
}
// First, count reg args
bool callHasFloatRegArgs = false;
for (GenTreePtr list = call->gtCallLateArgs; list; list = list->MoveNext())
{
assert(list->OperIsList());
GenTreePtr argNode = list->Current();
fgArgTabEntryPtr curArgTabEntry = compiler->gtArgEntryByNode(call, argNode);
assert(curArgTabEntry);
if (curArgTabEntry->regNum == REG_STK)
{
// late arg that is not passed in a register
assert(argNode->gtOper == GT_PUTARG_STK);
TreeNodeInfoInitPutArgStk(argNode->AsPutArgStk(), curArgTabEntry);
continue;
}
// A GT_FIELD_LIST has a TYP_VOID, but is used to represent a multireg struct
if (argNode->OperGet() == GT_FIELD_LIST)
{
argNode->SetContained();
// There could be up to 2-4 PUTARG_REGs in the list (3 or 4 can only occur for HFAs)
regNumber argReg = curArgTabEntry->regNum;
for (GenTreeFieldList* entry = argNode->AsFieldList(); entry != nullptr; entry = entry->Rest())
{
TreeNodeInfoInitPutArgReg(entry->Current()->AsUnOp(), argReg, *info, false, &callHasFloatRegArgs);
// Update argReg for the next putarg_reg (if any)
argReg = genRegArgNext(argReg);
#if defined(_TARGET_ARM_)
// A double register is modelled as an even-numbered single one
if (entry->Current()->TypeGet() == TYP_DOUBLE)
{
argReg = genRegArgNext(argReg);
}
#endif // _TARGET_ARM_
}
}
#ifdef _TARGET_ARM_
else if (argNode->OperGet() == GT_PUTARG_SPLIT)
{
fgArgTabEntryPtr curArgTabEntry = compiler->gtArgEntryByNode(call, argNode);
TreeNodeInfoInitPutArgSplit(argNode->AsPutArgSplit(), *info, curArgTabEntry);
}
#endif
else
{
TreeNodeInfoInitPutArgReg(argNode->AsUnOp(), curArgTabEntry->regNum, *info, false, &callHasFloatRegArgs);
}
}
// Now, count stack args
// Note that these need to be computed into a register, but then
// they're just stored to the stack - so the reg doesn't
// need to remain live until the call. In fact, it must not
// because the code generator doesn't actually consider it live,
// so it can't be spilled.
GenTreePtr args = call->gtCallArgs;
while (args)
{
GenTreePtr arg = args->gtOp.gtOp1;
// Skip arguments that have been moved to the Late Arg list
if (!(args->gtFlags & GTF_LATE_ARG))
{
if (arg->gtOper == GT_PUTARG_STK)
{
fgArgTabEntryPtr curArgTabEntry = compiler->gtArgEntryByNode(call, arg);
assert(curArgTabEntry);
assert(curArgTabEntry->regNum == REG_STK);
TreeNodeInfoInitPutArgStk(arg->AsPutArgStk(), curArgTabEntry);
}
#ifdef _TARGET_ARM_
else if (arg->OperGet() == GT_PUTARG_SPLIT)
{
fgArgTabEntryPtr curArgTabEntry = compiler->gtArgEntryByNode(call, arg);
TreeNodeInfoInitPutArgSplit(arg->AsPutArgSplit(), *info, curArgTabEntry);
}
#endif
else
{
TreeNodeInfo* argInfo = &(arg->gtLsraInfo);
if (argInfo->dstCount != 0)
{
argInfo->isLocalDefUse = true;
}
argInfo->dstCount = 0;
}
}
args = args->gtOp.gtOp2;
}
// If it is a fast tail call, it is already preferenced to use IP0.
// Therefore, no need set src candidates on call tgt again.
if (call->IsVarargs() && callHasFloatRegArgs && !call->IsFastTailCall() && (ctrlExpr != nullptr))
{
NYI_ARM("float reg varargs");
// Don't assign the call target to any of the argument registers because
// we will use them to also pass floating point arguments as required
// by Arm64 ABI.
ctrlExpr->gtLsraInfo.setSrcCandidates(l, l->allRegs(TYP_INT) & ~(RBM_ARG_REGS));
}
#ifdef _TARGET_ARM_
if (call->NeedsNullCheck())
{
info->internalIntCount++;
}
#endif // _TARGET_ARM_
}
//------------------------------------------------------------------------
// TreeNodeInfoInitIndir: Specify register requirements for address expression
// of an indirection operation.
//
// Arguments:
// indirTree - GT_IND, GT_STOREIND, block node or GT_NULLCHECK gentree node
//
void Lowering::TreeNodeInfoInitIndir(GenTreePtr indirTree)
{
assert(indirTree->OperIsIndir());
// If this is the rhs of a block copy (i.e. non-enregisterable struct),
// it has no register requirements.
if (indirTree->TypeGet() == TYP_STRUCT)
{
return;
}
GenTreePtr addr = indirTree->gtGetOp1();
TreeNodeInfo* info = &(indirTree->gtLsraInfo);
GenTreePtr base = nullptr;
GenTreePtr index = nullptr;
unsigned cns = 0;
unsigned mul;
bool rev;
bool modifiedSources = false;
bool makeContained = true;
if ((addr->OperGet() == GT_LEA) && IsSafeToContainMem(indirTree, addr))
{
GenTreeAddrMode* lea = addr->AsAddrMode();
base = lea->Base();
index = lea->Index();
cns = lea->gtOffset;
#ifdef _TARGET_ARM_
// ARM floating-point load/store doesn't support a form similar to integer
// ldr Rdst, [Rbase + Roffset] with offset in a register. The only supported
// form is vldr Rdst, [Rbase + imm] with a more limited constraint on the imm.
if (lea->HasIndex() || !emitter::emitIns_valid_imm_for_vldst_offset(cns))
{
if (indirTree->OperGet() == GT_STOREIND)
{
if (varTypeIsFloating(indirTree->AsStoreInd()->Data()))
{
makeContained = false;
}
}
else if (indirTree->OperGet() == GT_IND)
{
if (varTypeIsFloating(indirTree))
{
makeContained = false;
}
}
}
#endif
if (makeContained)
{
m_lsra->clearOperandCounts(addr);
addr->SetContained();
// The srcCount is decremented because addr is now "contained",
// then we account for the base and index below, if they are non-null.
info->srcCount--;
}
}
else if (comp->codeGen->genCreateAddrMode(addr, -1, true, 0, &rev, &base, &index, &mul, &cns, true /*nogen*/) &&
!(modifiedSources = AreSourcesPossiblyModifiedLocals(indirTree, base, index)))
{
// An addressing mode will be constructed that may cause some
// nodes to not need a register, and cause others' lifetimes to be extended
// to the GT_IND or even its parent if it's an assignment
assert(base != addr);
m_lsra->clearOperandCounts(addr);
addr->SetContained();
// Traverse the computation below GT_IND to find the operands
// for the addressing mode, marking the various constants and
// intermediate results as not consuming/producing.
// If the traversal were more complex, we might consider using
// a traversal function, but the addressing mode is only made
// up of simple arithmetic operators, and the code generator
// only traverses one leg of each node.
bool foundBase = (base == nullptr);
bool foundIndex = (index == nullptr);
GenTreePtr nextChild = nullptr;
for (GenTreePtr child = addr; child != nullptr && !child->OperIsLeaf(); child = nextChild)
{
nextChild = nullptr;
GenTreePtr op1 = child->gtOp.gtOp1;
GenTreePtr op2 = (child->OperIsBinary()) ? child->gtOp.gtOp2 : nullptr;
if (op1 == base)
{
foundBase = true;
}
else if (op1 == index)
{
foundIndex = true;
}
else
{
m_lsra->clearOperandCounts(op1);
op1->SetContained();
if (!op1->OperIsLeaf())
{
nextChild = op1;
}
}
if (op2 != nullptr)
{
if (op2 == base)
{
foundBase = true;
}
else if (op2 == index)
{
foundIndex = true;
}
else
{
m_lsra->clearOperandCounts(op2);
op2->SetContained();
if (!op2->OperIsLeaf())
{
assert(nextChild == nullptr);
nextChild = op2;
}
}
}
}
assert(foundBase && foundIndex);
info->srcCount--; // it gets incremented below.
}
else if (addr->gtOper == GT_ARR_ELEM)
{
// The GT_ARR_ELEM consumes all the indices and produces the offset.
// The array object lives until the mem access.
// We also consume the target register to which the address is
// computed
info->srcCount++;
assert(addr->gtLsraInfo.srcCount >= 2);
addr->gtLsraInfo.srcCount -= 1;
}
else
{
// it is nothing but a plain indir
info->srcCount--; // base gets added in below
base = addr;
}
if (!makeContained)
{
return;
}
if (base != nullptr)
{
info->srcCount++;
}
if (index != nullptr && !modifiedSources)
{
info->srcCount++;
}
// On ARM we may need a single internal register
// (when both conditions are true then we still only need a single internal register)
if ((index != nullptr) && (cns != 0))
{
// ARM does not support both Index and offset so we need an internal register
info->internalIntCount = 1;
}
else if (!emitter::emitIns_valid_imm_for_ldst_offset(cns, emitTypeSize(indirTree)))
{
// This offset can't be contained in the ldr/str instruction, so we need an internal register
info->internalIntCount = 1;
}
}
//------------------------------------------------------------------------
// LowerBlockStore: Set block store type
//
// Arguments:
// blkNode - The block store node of interest
//
// Return Value:
// None.
//
void Lowering::LowerBlockStore(GenTreeBlk* blkNode)
{
GenTree* dstAddr = blkNode->Addr();
unsigned size = blkNode->gtBlkSize;
GenTree* source = blkNode->Data();
Compiler* compiler = comp;
// Sources are dest address and initVal or source.
GenTreePtr srcAddrOrFill = nullptr;
bool isInitBlk = blkNode->OperIsInitBlkOp();
if (!isInitBlk)
{
// CopyObj or CopyBlk
if ((blkNode->OperGet() == GT_STORE_OBJ) && ((blkNode->AsObj()->gtGcPtrCount == 0) || blkNode->gtBlkOpGcUnsafe))
{
blkNode->SetOper(GT_STORE_BLK);
}
if (source->gtOper == GT_IND)
{
srcAddrOrFill = blkNode->Data()->gtGetOp1();
}
}
if (isInitBlk)
{
GenTreePtr initVal = source;
if (initVal->OperIsInitVal())
{
initVal->SetContained();
initVal = initVal->gtGetOp1();
}
srcAddrOrFill = initVal;
#ifdef _TARGET_ARM64_
if ((size != 0) && (size <= INITBLK_UNROLL_LIMIT) && initVal->IsCnsIntOrI())
{
// TODO-ARM-CQ: Currently we generate a helper call for every
// initblk we encounter. Later on we should implement loop unrolling
// code sequences to improve CQ.
// For reference see the code in LowerXArch.cpp.
NYI_ARM("initblk loop unrolling is currently not implemented.");
// The fill value of an initblk is interpreted to hold a
// value of (unsigned int8) however a constant of any size
// may practically reside on the evaluation stack. So extract
// the lower byte out of the initVal constant and replicate
// it to a larger constant whose size is sufficient to support
// the largest width store of the desired inline expansion.
ssize_t fill = initVal->gtIntCon.gtIconVal & 0xFF;
if (fill == 0)
{
MakeSrcContained(blkNode, source);
}
else if (size < REGSIZE_BYTES)
{
initVal->gtIntCon.gtIconVal = 0x01010101 * fill;
}
else
{
initVal->gtIntCon.gtIconVal = 0x0101010101010101LL * fill;
initVal->gtType = TYP_LONG;
}
blkNode->gtBlkOpKind = GenTreeBlk::BlkOpKindUnroll;
}
else
#endif // _TARGET_ARM64_
{
blkNode->gtBlkOpKind = GenTreeBlk::BlkOpKindHelper;
}
}
else
{
// CopyObj or CopyBlk
// Sources are src and dest and size if not constant.
if (blkNode->OperGet() == GT_STORE_OBJ)
{
// CopyObj
GenTreeObj* objNode = blkNode->AsObj();
unsigned slots = objNode->gtSlots;
#ifdef DEBUG
// CpObj must always have at least one GC-Pointer as a member.
assert(objNode->gtGcPtrCount > 0);
assert(dstAddr->gtType == TYP_BYREF || dstAddr->gtType == TYP_I_IMPL);
CORINFO_CLASS_HANDLE clsHnd = objNode->gtClass;
size_t classSize = compiler->info.compCompHnd->getClassSize(clsHnd);
size_t blkSize = roundUp(classSize, TARGET_POINTER_SIZE);
// Currently, the EE always round up a class data structure so
// we are not handling the case where we have a non multiple of pointer sized
// struct. This behavior may change in the future so in order to keeps things correct
// let's assert it just to be safe. Going forward we should simply
// handle this case.
assert(classSize == blkSize);
assert((blkSize / TARGET_POINTER_SIZE) == slots);
assert(objNode->HasGCPtr());
#endif
blkNode->gtBlkOpKind = GenTreeBlk::BlkOpKindUnroll;
}
else // CopyBlk
{
// In case of a CpBlk with a constant size and less than CPBLK_UNROLL_LIMIT size
// we should unroll the loop to improve CQ.
// For reference see the code in lowerxarch.cpp.
if ((size != 0) && (size <= INITBLK_UNROLL_LIMIT))
{
blkNode->gtBlkOpKind = GenTreeBlk::BlkOpKindUnroll;
}
else
{
// In case we have a constant integer this means we went beyond
// CPBLK_UNROLL_LIMIT bytes of size, still we should never have the case of
// any GC-Pointers in the src struct.
blkNode->gtBlkOpKind = GenTreeBlk::BlkOpKindHelper;
}
}
// CopyObj or CopyBlk
if (source->gtOper == GT_IND)
{
MakeSrcContained(blkNode, source);
}
else if (!source->IsMultiRegCall() && !source->OperIsSIMD())
{
assert(source->IsLocal());
MakeSrcContained(blkNode, source);
}
}
}
