On 2022/8/17 21:26, Richard Henderson wrote:
On 8/17/22 04:10, WANG Xuerui wrote:
From my own experiences, different use cases care about different
aspects of the CPU, and that IMO is an argument in favor of providing
both (high-fidelity models named after actual product model names,
and virtual models named after ISA levels). But before we have truly
high-fidelity models I think we should start with the virtual ones
first. And don't pretend the currently implemented model is LA464 --
the kernel change I've linked to [1] implies the opposite.
No, it simply pointed to a bug in qemu that could have been fixed.
The trouble with inventing virtual models is that no one knows what
they mean. Targeting real hardware is better, because we have a
documented standard.
Hmm, I've looked up more context and it is indeed reasonable to
generally name the QEMU models after real existing models. But in this
case we could face a problem with Loongson's nomenclature: all of
Loongson 3A5000, 3C5000 and 3C5000L are LA464, yet they should be
distinguishable software-side by checking the model name CSR. But with
only one CPU model that is LA464, currently this CSR is hard-coded to
read "3A5000", and this can hurt IMO. And when we finally add LA264 and
LA364 they would be identical ISA-level-wise, again the only
differentiator is the model name CSR.
And by "not high-fidelity", I mean some of the features present on real
HW might never get implemented, or actually implementable, like the DVFS
mechanism needed by cpufreq. And I believe Bibo wouldn't have to change
the kernel if it's not needed after all to run the unmodified upstream
kernel on top of qemu-system-loongarch64. (I would of course accept, and
learn something along the way, if this turns out not to be the case though.)
Lastly, the "ISA level" I proposed is not arbitrarily made up; it's
direct reference to the ISA manual revision. Each time the ISA gets some
addition/revision the ISA manual has to be updated, and currently the
manual's revision is the only reliable source of said information.
(Loongson has a history of naming cores badly, like with the MIPS 3B1500
and 3A4000, both were "GS464V"; and 3A5000 was originally GS464V too,
even though the insn encodings and some semantics have been entirely
different.)
In conclusion, I'd accept the micro-architecture naming if the model CSR
behavior could be sorted out, otherwise I'd personally prefer real model
names if ISA level naming is not going to fly. This is not a strong
objection to the current way of doing things though, more like some
minor but anyway needed discussion that happened a bit late. Sorry for
not chiming in earlier during the review process.
