On Wed, Jun 22, 2011 at 12:18 PM, nathan binkert <[email protected]> wrote:
>> ... but just to be clear, I'm saying let's keep readBytes/writeBytes
>> as the official ExecContext interface, but provide global templates
>> that map basic data types like uint64_t to those functions as a
>> convenience for when they can be used (but these templates can be
>> bypassed when appropriate). Then we can hide the byte swapping and
>> tracing goop inside those templates in the typical case.
>
> The write templates in AtomicSimpleCPU and base_dyn_inst are
> different. I don't see how they become global. Also, I think that
> the the need to have many specializations of read and write stems from
> the improper implementation of those templates. (base_dyn_inst)
> doesn't need specializations. It should be possible for us to write
> the templates for read/write in such a way that they only need to know
> the ISA (endianness).
You're comparing this:
BaseDynInst<Impl>::write(T data, Addr addr, unsigned flags, uint64_t *res)
{
if (traceData) {
traceData->setData(data);
}
data = TheISA::htog(data);
return writeBytes((uint8_t *)&data, sizeof(T), addr, flags, res);
}
with this:
AtomicSimpleCPU::write(T data, Addr addr, unsigned flags, uint64_t *res)
{
uint8_t *dataPtr = (uint8_t *)&data;
if (traceData)
traceData->setData(data);
data = htog(data);
Fault fault = writeBytes(dataPtr, sizeof(data), addr, flags, res);
if (fault == NoFault && data_write_req.isSwap()) {
*res = gtoh((T)*res);
}
return fault;
}
and you don't think they look very similar? We'd have to figure out where
*res gets updated in O3 and InOrder, but other than that they look the same
to me. Now that I see how similar they are I'm surprised you aren't anxious
to eliminate the redundancy...
Steve
_______________________________________________
gem5-dev mailing list
[email protected]
http://m5sim.org/mailman/listinfo/gem5-dev
