On 9/13/21 5:56 PM, Peter Maydell wrote:
> On Mon, 13 Sept 2021 at 16:22, Philippe Mathieu-Daudé <f4...@amsat.org> wrote:
>> On 9/13/21 3:50 PM, Peter Maydell wrote:
>>> On Mon, 13 Sept 2021 at 14:10, <marcandre.lur...@redhat.com> wrote:
>>> I still wonder if this would be clearer if we listed 'tcg' as
>>> an accelerator like all the others, rather than describing it as
>>> a kind of special case in the text above the table. After all, you
>>> can select it with '-accel tcg' the same as any other.
>>>
>>> I think the information in the table is correct now.
>>
>> What is TCI at this point?
>
> It's a different backend to TCG. (Almost always you do not want it.)
It might become a TCG sub-type.
> It's not a separate accelerator type.
Indeed.
>>>> +Other architectures are not actively maintained. They may be removed in
>>>> future
>>>> +releases.
>>
>> What are the other archs which could be removed?
>
> It's not so much that we have explicit code to handle any other host
> architectures which we might remove. It's more that we do not specifically
> say "we don't know what this host arch is and we don't have a TCG backend
> for it, sorry" for unknown host architectures; instead we suggest that
> the user might try using TCI.
>
> In future in theory we might:
> * just say "unsupported" there instead (making unknown
> host CPU arch a fatal error, in the same way unknown host OS
> is a fatal error)
> * drop TCI entirely
>
> I think both of those would be controversial (but I'm not against
> them myself ;-)).
So we had this discussion 5/6 months ago, and 2 users said they use
it for research, so Richard did a tremendous work to update it to
remove any limitation on TCG, which mostly ended in a rewrite:
$ for i in tcg/tci*;do git blame $i;done|cut -c25-47|sort|uniq -c|sort -nr
1106 Richard Henderson
274 Stefan Weil
6 Daniele Buono
4 Pranith Kumar
4 Emilio G. Cota
2 Paolo Bonzini
2 Aurelien Jarno
1 Philippe Mathieu-Daudé
1 Peter Maydell
Then it took me various week-ends to review/test it. At this point
I consider it stable and removing it isn't justified anymore.
> We could perhaps rephrase the sentence to something like:
>
> Other host architectures are not supported. It is possible to
> build QEMU on an unsupported host architecture using the configure
> ``--enable-tcg-interpreter`` option to enable the experimental
> TCI support, but note that this is very slow and is not recommended.
Sounds good.
> If we want to actually deprecate building on non-supported host
> architectures we should say that via the usual deprecation process.
Agreed.
Regards,
Phil.
