Hi Nilay,
Part of my point is that, when compiled with optimization, there should not
be any performance difference between the two versions. Thus I don't see a
need to maintain two versions, even in the makeRead/makeWrite parts of the
code (which I see you haven't updated). Others can weigh in if they
disagree.
As far as the second issue, I agree there may not be anything to do there
besides perhaps add some comments documenting the restriction. The const
idea is a good one though it may be unenforceable. Our options might be
clearer when we see how this feature is used.
All this said, I think in the long run we will still need to handle the x86
condition code specially and not treat them like GPRs, so I'm not yet
convinced that this patch is more than an academic exercise. Nevertheless,
I can see how it could be useful in the future.
Even if we handle cc registers in a special manner, this patch or
something similar would be required to remove some of the dependencies. We
still read and write most of the CC registers, even if we do not need to.
With this patch, these reads and writes would be avoided.
As far as I understand, special handling of the cc registers is aimed at
providing one physical register for several cc registers. It will remove
the register pressure, but not the artificial dependencies. I think you
have not gauged the need for the patch correctly.
--
Nilay
_______________________________________________
gem5-dev mailing list
[email protected]
http://m5sim.org/mailman/listinfo/gem5-dev
