On Thu, Jun 13, 2024 at 07:22:54PM +0200, Miguel Ojeda wrote:
> On Thu, Jun 13, 2024 at 6:31 PM Boqun Feng <boqun.feng@xxxxxxxxx> wrote:
> >
> > So let's start with some basic and simple until we really have a need
> > for generic `Atomic<T>`. Thoughts?
>
> I don't want to delay this, but using generics would be more flexible,
> right? e.g. it could allow us to have atomics of, say, newtypes, if
> that were to be useful.
>
> Is there a particular disadvantage of using the generics? The two
> cases you mentioned would just be written explicitly, right?
>
> One disadvantage would be that they are different from the Rust
> standard library ones, e.g. in case we wanted third-party code to use
> them, but could be provided if needed later on.
>
Well, the other thing is AtomicI32 -> atomic_t and AtomicI64 ->
atomic64_t are perfect mappings, and we can treat AtomicI32 and
AtomicI64 as a separate layer that wires C atomics into Rust. As I said,
we can build `Atomic<T>` on top of this layer, like:
Atomic<T>
|
V
AtomicI{32,64}
|
V
atomic{,64}_t
and if we drop this layer, the dependencies become:
Atomic<i32,i64> <- Atomic<u32,u64>
|
V
atomic{,64}_t
i.e. in the same layer of Atomic<T>, some of them directly depend on
some other Atomic<T> types, which doesn't look very clean to me. And it
might be difficult for architecture maintainers to track the exact
dependency for Rust code.
This is also the reason why I didn't use Rust macros to generate
AtomicI32 and AtomicI64 implementation: I use a script to generate .rs
file. This ensures AtomicI32 and AtomicI64 stay with the exact same set
of APIs as atomic{,64}_t (described by scripts/atomic/atomics.tbl. Put
it in another way, I guess you can think AtomicI32 and AtomicI64 as some
sort of intrinsic layer provided by C. And should we need it, we can
build an Atomic<T> layer on top of it.
Does this make sense?
Regards,
Boqun
> Cheers,
> Miguel
