private static void EmitStoreGlobal(EmitterContext context)
{
OpCodeMemory op = (OpCodeMemory)context.CurrOp;
bool isSmallInt = op.Size < IntegerSize.B32;
int count = GetVectorCount(op.Size);
(Operand addrLow, Operand addrHigh) = Get40BitsAddress(context, op.Ra, op.Extended, op.Offset);
Operand bitOffset = GetBitOffset(context, addrLow);
for (int index = 0; index < count; index++)
{
Register rd = new Register(op.Rd.Index + index, RegisterType.Gpr);
Operand value = Register(rd);
if (isSmallInt)
{
Operand word = context.LoadGlobal(addrLow, addrHigh);
value = InsertSmallInt(context, op.Size, bitOffset, word, value);
}
context.StoreGlobal(context.IAdd(addrLow, Const(index * 4)), addrHigh, value);
if (rd.IsRZ)
{
break;
}
}
}
private static void EmitStore(EmitterContext context, MemoryRegion region)
{
OpCodeMemory op = (OpCodeMemory)context.CurrOp;
if (op.Size > IntegerSize.B128)
{
context.Config.GpuAccessor.Log($"Invalid store size: {op.Size}.");
}
bool isSmallInt = op.Size < IntegerSize.B32;
int count = 1;
switch (op.Size)
{
case IntegerSize.B64: count = 2; break;
case IntegerSize.B128: count = 4; break;
}
Operand baseOffset = context.IAdd(GetSrcA(context), Const(op.Offset));
Operand wordOffset = context.ShiftRightU32(baseOffset, Const(2));
Operand bitOffset = GetBitOffset(context, baseOffset);
for (int index = 0; index < count; index++)
{
bool isRz = op.Rd.IsRZ;
Register rd = new Register(isRz ? op.Rd.Index : op.Rd.Index + index, RegisterType.Gpr);
Operand value = Register(rd);
Operand offset = context.IAdd(wordOffset, Const(index));
if (isSmallInt)
{
Operand word = null;
switch (region)
{
case MemoryRegion.Local: word = context.LoadLocal(offset); break;
case MemoryRegion.Shared: word = context.LoadShared(offset); break;
}
value = InsertSmallInt(context, op.Size, bitOffset, word, value);
}
switch (region)
{
case MemoryRegion.Local: context.StoreLocal(offset, value); break;
case MemoryRegion.Shared: context.StoreShared(offset, value); break;
}
}
}
private static void EmitLoad(EmitterContext context, MemoryRegion region)
{
OpCodeMemory op = (OpCodeMemory)context.CurrOp;
if (op.Size > IntegerSize.B128)
{
context.Config.GpuAccessor.Log($"Invalid load size: {op.Size}.");
}
bool isSmallInt = op.Size < IntegerSize.B32;
int count = 1;
switch (op.Size)
{
case IntegerSize.B64: count = 2; break;
case IntegerSize.B128: count = 4; break;
}
Operand baseOffset = context.IAdd(GetSrcA(context), Const(op.Offset));
// Word offset = byte offset / 4 (one word = 4 bytes).
Operand wordOffset = context.ShiftRightU32(baseOffset, Const(2));
Operand bitOffset = GetBitOffset(context, baseOffset);
for (int index = 0; index < count; index++)
{
Register rd = new Register(op.Rd.Index + index, RegisterType.Gpr);
if (rd.IsRZ)
{
break;
}
Operand offset = context.IAdd(wordOffset, Const(index));
Operand value = null;
switch (region)
{
case MemoryRegion.Local: value = context.LoadLocal(offset); break;
case MemoryRegion.Shared: value = context.LoadShared(offset); break;
}
if (isSmallInt)
{
value = ExtractSmallInt(context, op.Size, bitOffset, value);
}
context.Copy(Register(rd), value);
}
}
