Re: [OE-core][PATCH RFC v2 12/19] meta/lib/oe/tune.py: add riscv_pkgarch()

["Trevor Gamblin" <tgamblin@baylibre.com>]

On Thu Oct 8, 2026 at 11:26 AM EDT, Mark Hatle wrote:
>
>
> On 9/30/26 11:38 AM, Trevor Gamblin wrote:
>> Create a new function to help match RISC-V ISA extension combinations
>> against recognized profile strings (e.g. 'rva22u64'), so that we can use
>> simpler patterns for naming things like sstate/build paths.
>> 
>> Signed-off-by: Trevor Gamblin <tgamblin@baylibre.com>
>> ---
>>   meta/lib/oe/tune.py | 29 +++++++++++++++++++++++++++++
>>   1 file changed, 29 insertions(+)
>> 
>> diff --git a/meta/lib/oe/tune.py b/meta/lib/oe/tune.py
>> index 7296f39e66..0d8e4417e7 100644
>> --- a/meta/lib/oe/tune.py
>> +++ b/meta/lib/oe/tune.py
>> @@ -90,3 +90,32 @@ def riscv_isa_to_tune(isa):
>>       # Eliminate duplicates, but preserve the order
>>       feature = list(dict.fromkeys(feature))
>>       return ' '.join(feature)
>> +
>> +# riscv_pkgarch(features, exact_pkgarch)
>> +#
>> +# Prefer a RISCV_PROFILES name (plus any extensions enabled on top of it) as the
>> +# package arch fragment, since otherwise the full ISA extension list will exceed
>> +# what the sstate class can handle. When several profiles match, the one with
>> +# the most extensions is preferred, as it is the more specific/clearer match.
>> +# If there's no clear profile matching the extension list, then fall back to
>> +# exact_pkgarch.
>> +def riscv_pkgarch(features, exact_pkgarch):
>> +    if not features:
>> +        return exact_pkgarch
>> +
>> +    current = set(features.split())
>> +
>> +    best_match = None
>> +    best_features = None
>> +    for name in RISCV_PROFILES:
>
> Where is 'RISCV_PROFILES' defined?

It's at the top of this file, although it could make sense to separate it into
its own text file somewhere else.
>
> As long as someone is willing to keep this up to date with all of the new 
> profiles, I don't have a concern with it.

I had intended this for myself (it seems almost all that I do is RISC-V these
days :)).
>
> On my AMD platforms, we have dynamic (user defined) ISA for our parts, the users 
> don't really use 'profiles', but select individual isa components.  I would LIKE 
> (but only because it's easier for me) a way to see the ISA and not have it 
> converted back to a profile.  However, if profiles are the risc-v ISA standard, 
> I'm fine without a way to disable the translation.

That is a good point. I built the profile logic to help with shortening the
build paths, but also so that users who were selecting custom sets might end up
seeing that (for example) their custom extension list was "just" RVA20 plus
vector, or something like that. This would work a lot better if it could do a
comparison and also report subtractive matches, e.g. "your selection matches
RVA22 minus Zfhmin and Zicboz". However, I think there's limited value in adding
a capability like that, and it's out of scope for this series anyway. Maybe it'd
be worth exploring as an add-on for bitbake(-setup) as a NOTE when a build
starts, sometime in the future and if people request it.
>
> --Mark
>
>> +        profile_features = set(riscv_isa_to_tune(name).split())
>> +        if profile_features.issubset(current):
>> +            if best_features is None or len(profile_features) > len(best_features):
>> +                best_match = name
>> +                best_features = profile_features
>> +
>> +    if best_match is None:
>> +        return exact_pkgarch
>> +
>> +    extra = sorted(current - best_features)
>> +    return '_'.join([best_match] + extra)
>>