Fwd: Restricted sums in BoxedRep

Thu Oct 15 20:23:44 UTC 2020

Ah yes.  In my research work I extended Mutable Constructor fields to allow
a mutable array field to help with that problem.  This allowed me to have a
Nil constructor in a sum type and a Node constructor with normal fields as
well as an array of mutable fields.  Pointer tagging worked as expected so
a case match on a dereference of a field would not need to follow the
pointer and no intermediate objects were between Node objects.

On Thu, Oct 15, 2020 at 4:08 PM David Feuer <david.feuer at gmail.com> wrote:

> This might be lost in the noise for MVar and TVar, but for arrays, there
> are tremendous advantages to cutting out the extra heap objects.
>
> On Thu, Oct 15, 2020, 1:10 PM Ryan Yates <fryguybob at gmail.com> wrote:
>
>> I haven't been following this discussion too closely, but in my research
>> work I found that some of the benefits that I wanted in this direction were
>> already there with pointer tagging.
>>
>> On Thu, Oct 15, 2020 at 12:59 PM David Feuer <david.feuer at gmail.com>
>> wrote:
>>
>>> Yes, that's something quite different. We'd need a whole different heap
>>> object type for such MVars and TVars. My approach covers the case where the
>>> unboxed thing can only take on a few values, for some value of "a few"
>>> which, depending on implementation, may or may not be very small. If the
>>> nulls point to actual heap objects in pre-allocated pinned memory (say),
>>> then up to 64 or so might be reasonable. If they point to "invalid" address
>>> space, then the numbers could go up a good bit.
>>>
>>> On Thu, Oct 15, 2020, 12:50 PM Carter Schonwald <
>>> carter.schonwald at gmail.com> wrote:
>>>
>>>> Indeed, I mean things that aren’t pointery, and could be represented by
>>>> a tvar paired with a mutable byte array or mvar with mutable byte array,
>>>> but which we’d want considered as a single heap object from the rts/gc
>>>> perspective.
>>>>
>>>> On Thu, Oct 15, 2020 at 11:58 AM David Feuer <david.feuer at gmail.com>
>>>> wrote:
>>>>
>>>>> Sorry, unlifted, not unboxed...
>>>>>
>>>>> On Thu, Oct 15, 2020, 11:57 AM David Feuer <david.feuer at gmail.com>
>>>>> wrote:
>>>>>
>>>>>> Putting unboxed things in TVar, MVar, etc., is part of Andrew
>>>>>> Martin's accepted BoxedRep proposal.
>>>>>>
>>>>>> On Thu, Oct 15, 2020, 11:44 AM Carter Schonwald <
>>>>>> carter.schonwald at gmail.com> wrote:
>>>>>>
>>>>>>> A related idea that came up recently and is perhaps simpler ties
>>>>>>> into this via the lens of having unboxed Mvars/tvars (even if it’s
>>>>>>> restricted to just things we can embed in a word64#)
>>>>>>>
>>>>>>> This came up in
>>>>>>> https://gitlab.haskell.org/ghc/ghc/-/issues/18798#note_307410,
>>>>>>> where viktor had millions of independent mvars holding what’s essentially a
>>>>>>> strict unit ()!
>>>>>>>
>>>>>>> The motivation in this later scenario is that in high concurrency
>>>>>>> settings, the less trivial stuff the gc needs to trace under updates, the
>>>>>>> better ghc scales.
>>>>>>>
>>>>>>> This may not be a use case david has in mind, but certainly seems
>>>>>>> related.
>>>>>>>
>>>>>>> Phrased more succinctly: gc perf dominates large heap / many core
>>>>>>> computation in Haskell via sensitivity to allocation volume / mutation
>>>>>>> volume (to ensure generational hypothesis stays valid), and providing tools
>>>>>>> to incrementally reduce the pressure with local changes would be good.
>>>>>>>
>>>>>>> So I’d propose / suggest that a baby step towards what david asks
>>>>>>> would be for us to work out some manner of unboxed tvar/mvar ref machinery
>>>>>>> that supports unboxed values.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> On Thu, Oct 15, 2020 at 5:32 AM Andreas Klebinger <
>>>>>>> klebinger.andreas at gmx.at> wrote:
>>>>>>>
>>>>>>>> From a implementors perspective my main questions would be:
>>>>>>>>
>>>>>>>> * How big is the benefit in practice? How many use cases are there?
>>>>>>>> * How bad are the costs? (Runtime overhead, rts complexity, ...)
>>>>>>>>
>>>>>>>> The details of how this would be exposed to a user would are
>>>>>>>> important.
>>>>>>>> But if the costs are too high for the drawbacks then it becomes a
>>>>>>>> moot point.
>>>>>>>>
>>>>>>>>
>>>>>>>> David Feuer schrieb am 14.10.2020 um 22:21:
>>>>>>>>
>>>>>>>> Forwarded from Andrew Martin below. I think we want more than just
>>>>>>>> Maybe (more than one null), but the nesting I described is certainly more
>>>>>>>> convenience than necessity.
>>>>>>>>
>>>>>>>> ---------- Forwarded message ---------
>>>>>>>> From: Andrew Martin <andrew.thaddeus at gmail.com>
>>>>>>>> Date: Wed, Oct 14, 2020, 4:14 PM
>>>>>>>> Subject: Re: Restricted sums in BoxedRep
>>>>>>>> To: David Feuer <david.feuer at gmail.com>
>>>>>>>>
>>>>>>>>
>>>>>>>> You'll have to forward this to the ghc-devs list to share it with
>>>>>>>> others since I'm not currently subscribed to it, but I've had this same
>>>>>>>> thought before. It is discussed at
>>>>>>>> https://github.com/andrewthad/impure-containers/issues/12. Here's
>>>>>>>> the relevant excerpt:
>>>>>>>>
>>>>>>>>> Relatedly, I was thinking the other day that after finishing
>>>>>>>>> implementing
>>>>>>>>> https://github.com/ghc-proposals/ghc-proposals/blob/master/proposals/0203-pointer-rep.rst,
>>>>>>>>> I should really look at seeing if it's possible to add this
>>>>>>>>> maybe-of-a-lifted value trick straight to GHC. I think that with:
>>>>>>>>>
>>>>>>>>> data RuntimpRep
>>>>>>>>>   = BoxedRep Levity
>>>>>>>>>   | MaybeBoxedRep Levity
>>>>>>>>>   | IntRep
>>>>>>>>>   | ...
>>>>>>>>>
>>>>>>>>> data BuiltinMaybe :: forall (v :: Levity). TYPE v -> TYPE ('MaybeBoxedRep v)
>>>>>>>>>
>>>>>>>>> This doesn't have the nesting issues because the kind system
>>>>>>>>> prevents nesting. But anyway, back to the original question. I would
>>>>>>>>> recommend not using Maybe.Unsafe and using unpacked-maybe
>>>>>>>>> instead. The latter is definitely safe, and it only costs an extra machine
>>>>>>>>> word of space in each data constructor it gets used in, and it doesn't
>>>>>>>>> introduce more indirections.
>>>>>>>>>
>>>>>>>>
>>>>>>>> On Tue, Oct 13, 2020 at 5:47 PM David Feuer <david.feuer at gmail.com>
>>>>>>>> wrote:
>>>>>>>>
>>>>>>>>> Null pointers are widely known to be a lousy language feature in
>>>>>>>>> general, but there are certain situations where they're *really* useful for
>>>>>>>>> compact representation. For example, we define
>>>>>>>>>
>>>>>>>>>     newtype TMVar a = TMVar (TVar (Maybe a))
>>>>>>>>>
>>>>>>>>> We don't, however, actually use the fact that (Maybe a) is lifted.
>>>>>>>>> So we could represent this much more efficiently using something like
>>>>>>>>>
>>>>>>>>>     newtype TMVar a = TMVar (TVar a)
>>>>>>>>>
>>>>>>>>> where Nothing is represented by a distinguished "null" pointer.
>>>>>>>>>
>>>>>>>>> While it's possible to implement this sort of thing in user code
>>>>>>>>> (with lots of fuss and care), it's not very nice at all. What I'd really
>>>>>>>>> like to be able to do is represent certain kinds of sums like this natively.
>>>>>>>>>
>>>>>>>>> Now that we're getting BoxedRep, I think we can probably make it
>>>>>>>>> happen. The trick is to add a special Levity constructor representing sums
>>>>>>>>> of particular shapes. Specifically, we can represent a type like this if it
>>>>>>>>> is a possibly-nested sum which, when flattened into a single sum, consists
>>>>>>>>> of some number of nullary tuples and at most one Lifted or Unlifted type.
>>>>>>>>> Then we can have (inline) primops to convert between the BoxedRep and the
>>>>>>>>> sum-of-sums representations.
>>>>>>>>>
>>>>>>>>> Anyone have thoughts on details for what the Levity constructor
>>>>>>>>> arguments might look like?
>>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> --
>>>>>>>> -Andrew Thaddeus Martin
>>>>>>>>
>>>>>>>>
>>>>>>>> _______________________________________________
>>>>>>>> ghc-devs mailing listghc-devs at haskell.orghttp://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs
>>>>>>>>
>>>>>>>>
>>>>>>>> _______________________________________________
>>>>>>>> ghc-devs mailing list
>>>>>>>> ghc-devs at haskell.org
>>>>>>>> http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs
>>>>>>>>
>>>>>>> _______________________________________________
>>> ghc-devs mailing list
>>> ghc-devs at haskell.org
>>> http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs
>>>
>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/ghc-devs/attachments/20201015/6e75425a/attachment.html>