Re: [PATCH RFC 2/2] RISC-V: add T-Head vector errata handling

["Stefan O'Rear" <sorear@fastmail.com>]

On Fri, Jun 23, 2023, at 7:26 PM, Heiko Stübner wrote:
> Am Freitag, 23. Juni 2023, 12:22:35 CEST schrieb Heiko Stübner:
>> Am Freitag, 23. Juni 2023, 05:06:44 CEST schrieb Stefan O'Rear:
>> > On Thu, Jun 22, 2023, at 4:35 PM, Heiko Stübner wrote:
>> > > Am Donnerstag, 22. Juni 2023, 20:58:37 CEST schrieb Stefan O'Rear:
>> > >> Are you aware of "3.7. Vector Fixed-Point Fields in fcsr" in
>> > >> riscv-v-spec-0.7.1.pdf?
>> > >
>> > > oh wow, thanks a lot for that pointer, now I understand your concern.
>> > >
>> > > So in vector-0.7.1 fcsr[10:9] mirrors vxrm and fcsr[8] mirrors vxsat.
>> > >
>> > >
>> > > On a positive note, the T-Head cores seem to not implement the full
>> > > vector 0.7.1 specification after all, in the documentation I have [0]
>> > > fcsr[31:8] are declared as "0" and uppermost bits are [7:5] for the "frm"
>> > > field.
>> > 
>> > Given that the pdf you linked does not mention any vector CSRs, I am not
>> > confident that it provides a complete and accurate description of vector
>> > functionality in other registers for the C906 with vector extension.
>> > 
>> > Assuming that you have access to such a chip, I would be much happier with
>> > the proposed "just a comment" approach if our understanding of the behavior
>> > were confirmed on hardware (specifically: csr_write(CSR_FCSR, 0x700) should
>> > not affect csr_read(CSR_VXRM) or csr_read(CSR_VXSAT)).
>> 
>> For one, you're right that I should definitly try to confirm this on hardware :-) .
>
> ok, so now I know the documentation is wrong.
>
> 	before, vxrm 0x0, vxsat 0x0
> writing the 0x700 to fcsr
> 	after, vxrm 0x3, vxsat 0x1
>
> Essentially the link between the CSRs really is there - oh fun.
> So we're back at your original concern - sadly.
>
> I guess I need to figure out how to not have this stuff break
> because relying on the fpu parts to handle feels not correct
> at first glance.

I don't see a conceptual problem here.

There are a large number of vector instructions that access the floating point
state (frm, fflags, f{0..31}).  These instructions require a valid floating
point context, sstatus.FS>0, trap if sstatus.FS=0, and set sstatus.FS=3 if they
change anything.  vfadd, vfsub, vfmul, vfdiv, vfmv, etc, etc in both 0.7.1 and 1.0.

0.7.1 extends the floating point state to include vxrm and vxsat and adds vaaddu,
vnclip, vsmul, etc to the list of instructions which access both floating point and
vector state.  This breaks floating-point emulation (if a core provides integer
vectors in hardware, it provides a fcsr register with three writeable bits, which
means that accesses to fcsr won't trap and can't be emulated to provide access to
the software frm and fflags), which is probably why the behavior was changed in 1.0,
but is otherwise internally consistent.

You can continue to treat "fpu parts" and "vector parts" as independent as long as
you recognize vxrm and vxsat as _exclusively_ owned by the "fpu parts".

Access to vxrm and vxsat should be exclusively controlled by sstatus.FS since they
are aliases for fields in fcsr, while access to vstart/vtype/vlen should be
exclusively controlled by sstatus.VS.  It would be good to test this, since it's not
actually in the spec (risc-v has a rather systematic underspecification problem),
and T-Head's idea of obviously correct behavior may differ from mine.

1.0 puts vxrm and vxsat into the vector state, controlled by sstatus.VS; vsmul now
works on the vector state only, so as far as state management is concerned it now
acts like vadd and not like vfadd.  This is also internally consistent, but vxcsr
must be owned by whoever manages sstatus.VS.

It's a little bit weird to say "vxrm and vxsat live in different parts of the kernel
state depending on V extension version" but it might be less weird to say "fcsr and
all its parts lives somewhere, vxcsr and all its parts lives somewhere else,
vxcsr doesn't exist in 0.7.1".  Is this adequate?

-s

>
> Heiko