Unlifted primop types

[David Feuer]

Don't pay *too* much attention to mkWeakIORef in particular; for some
reason it only makes a weak reference from an IORef to itself. It *should*
have been written

mkWeakIORef :: IORef a -> b -> IO () -> IO (Weak (IORef a))
mkWeakIORef r@(IORef (STRef r#)) b (IO finalizer) = IO $ \s ->
    case mkWeak# r# b finalizer s of (# s1, w #) -> (# s1, Weak w #)

and the current version should've been

mkSimpleWeakIORef r fin = mkWeakIORef r r fin

On Thu, Aug 23, 2018, 5:52 PM David Feuer <david.feuer at gmail.com> wrote:

> Weak pointers to unlifted heap objects are the opposite of dodgy! We have
> base library functions (e.g., Data.IORef.mkWeakIORef) for creating weak
> references to lifted objects that *actually* create references to the
> underlying unlifted objects instead. This is much more reliable, because
> the wrapper (e.g., the STRef constructor) could be stripped away by
> worker-wrapper, while the underlying object (e.g., the MutVar#) won't be.
>
> On Thu, Aug 23, 2018, 5:43 PM Simon Peyton Jones <simonpj at microsoft.com>
> wrote:
>
>> Gah.  We have no way to be polymorphic over all pointers (both lifted and
>> unlifted) but not over Int# etc.
>>
>>
>>
>> As you say, this is too much of a special case to make an invasive change.
>>
>>
>>
>> I’m quite dubious about making weak poitners to unlifted heap-allocated
>> objects.  I can’t say it’s wrong but it feels dodgy to me.
>>
>>
>>
>> Simon
>>
>>
>>
>> *From:* David Feuer <david.feuer at gmail.com>
>> *Sent:* 22 August 2018 14:59
>> *To:* Ben Gamari <ben at smart-cactus.org>
>> *Cc:* Simon Peyton Jones <simonpj at microsoft.com>; David Feuer <
>> david at well-typed.com>; ghc-devs <ghc-devs at haskell.org>
>> *Subject:* Re: Unlifted primop types
>>
>>
>>
>> The problem is that this also accepts things that aren't pointers at all!
>> We could fix that *for primops* by changing RuntimeRep to something like
>>
>>
>>
>> data RuntimeRep
>>
>>   = PtrRep Liftedness
>>
>>    | ...
>>
>>
>>
>> But that would only work for primops (at least for now) so it may not be
>> worth the breakage.
>>
>>
>>
>> On Wed, Aug 22, 2018, 7:45 AM Ben Gamari <ben at smart-cactus.org> wrote:
>>
>> Simon Peyton Jones <simonpj at microsoft.com> writes:
>>
>> > |  Huh! It looks like what we currently do for some primops is just use
>> a
>> > |  totally bogus kind. For example, mkWeak# will happily accept an Int#
>> as
>> > |  its first argument.
>> >
>> > Well, I see
>> >   primop  MkWeakOp "mkWeak#" GenPrimOp
>> >      o -> b -> (State# RealWorld -> (# State# RealWorld, c #))
>> >
>> > and I believe (from Ben's message) that the "o" means "open type
>> variable",
>> > which is the old terminology for what we now call levity-polymorphic.
>> >
>> Right; currently (largely for historical reasons) we use `o` to
>> accommodate cases that accept both lifted and unlifted pointers.
>>
>> Cheers,
>>
>> - Ben
>>
>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/ghc-devs/attachments/20180823/4aca06f0/attachment.html>