//------------------------------------------------------------------------
// 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);
}
}
}
}
//------------------------------------------------------------------------
// 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);
}
}
}