[ghc-steering-committee] #536: Type-level literals as a separate language extension, rec: accept

[Adam Gundry]

This proposal is very helpful in exposing some of the divisions about 
the interpretation of extensions that exist within the committee! "Let a 
thousand flowers bloom" vs "more-extensions-make-me-wince", if you like.

For what it's worth I tend to be in the former camp. I think we should 
strive for orthogonality and avoid bundling together distinct features 
merely because they have related effects. Instead we should construct 
our language via composition of simple components. And I think it's 
reasonable to say that TypeLevelLiterals and (the rest of) DataKinds are 
clearly distinct and can be understood in isolation.

If the number of extensions is a problem, that points to the need to 
better structure our documentation and gather coherent subsets of 
extensions into manageable pieces (much as GHC2021 tries to do). Beyond 
that I'd like to understand the downsides better (e.g. what costs does 
it actually impose on HLS to have another extension?).

A conversation with one of the authors of Helium this week made me think 
that in the GHC proposals process we have tended to neglect the possible 
presence of other Haskell implementations. It's not realistic to expect 
that anyone will reimplement the whole of GHC Haskell, but perhaps there 
is a fragment large enough to be useful for a fair number of Haskell 
programs but also small enough to be independently implemented (e.g. by 
Helium). It would strike me as plausible that such an implementation 
might want to support TypeData+TypeLevelLiterals but not DataKinds.

Adam


On 09/03/2023 16:21, Richard Eisenberg wrote:
> Though I see the arguments in the other direction, I vote against this 
> proposal. As Joachim argues, adding an extension puts a tax on the 
> entire ecosystem by adding one more bit of context a reader has to have 
> to understand code. In addition, HLS will have to be aware of this 
> extension. And the benefits seem pretty marginal here.
> 
> Can we instead simply provide a warning? The warning here would be 
> something like the -Wpuns idea, in that it helps the author keep to a 
> particular subset of the language. (We might want to collect such 
> warnings somehow in our documentation.)
> 
> Actually, it occurs to me we already have the warning: 
> -Wunticked-promoted-constructors. We've put that warning in the 
> -Weverything bin (not even -Wall) because in the past we've decided that 
> we shouldn't recommend/require the ' when using a constructor in a type. 
> But actually, setting -Werror=unticked-promoted-constructors would seem 
> to be a perfect substitute for -XTypeLevelLiterals -XNoDataKinds. OK, 
> not perfect: you could still use a constructor in a type with a ' mark. 
> But I think that's great, actually, because perhaps an author wary of 
> constructors in types still might need to use one, in one place, in a 
> 2,000-line file. This ticks all the boxes (ha ha).
> 
> Having discovered this replacement, I'm now pretty strongly against this 
> proposal. I'll also post on GitHub.
> 
> Richard
> 
>> On Mar 9, 2023, at 4:26 AM, Arnaud Spiwack <arnaud.spiwack at tweag.io 
>> <mailto:arnaud.spiwack at tweag.io>> wrote:
>>
>> I can see where the author is coming from. But as Joachim, I find 
>> myself unconvinced that it is worth yet one extra extension (but then 
>> again, I'm the more-extensions-make-me-wince member of this committee).
>>
>> On Wed, 8 Mar 2023 at 13:56, Joachim Breitner 
>> <mail at joachim-breitner.de <mailto:mail at joachim-breitner.de>> wrote:
>>
>>     Hi,
>>
>>     Am Mittwoch, dem 08.03.2023 um 14:58 +0300 schrieb Vladislav Zavialov:
>>     > DataKinds is indeed a "bad" extension because it introduces
>>     ambiguity
>>     > to name resolution.
>>     >
>>     > There are two ways to avoid this issue:
>>     > 1. Keep namespaces separate and double down on using punned
>>     > constructor names.
>>     > 2. Merge terms and types, discouraging punned constructor names.
>>     >
>>     > Each approach could indeed lead to a "dialect" of sorts.
>>     >
>>     > * The (accepted and recently implemented) TypeData extension
>>     moves us
>>     > in the direction of option (1), and I imagine the proposed
>>     > TypeLevelLiterals would be often used in conjunction with TypeData.
>>     > This carves out a limited, conservative language subset for type-
>>     > level programming.
>>     > * At the same time, #270 "Support pun-free code" moves us in the
>>     > direction of option (2), and I expect it to be the better option
>>     when
>>     > it comes to fully fledged dependently-typed programming
>>     envisioned by
>>     > #378.
>>     >
>>     > If I had to choose, I'd go with option (2). I don't see myself using
>>     > TypeData, nor TypeLevelLiterals for that matter. But I am of a
>>     "let a
>>     > thousand flowers bloom" persuasion, hence my recommendation to
>>     > accept.
>>
>>
>>     Thanks for putting this into a clearer context.
>>
>>     Where does DataKind fit in? Is it a natural part of approach (2)
>>     (which
>>     is the general direction outlined by #378), or is yet another take on
>>     the issue?
>>     If DataKinds was not what we want according to #378, but the fragment
>>     carved out by TypeLevelLiterals was, then that would be a strong
>>     reason
>>     for me to acccept TypeLevelLiterals.
>>     But if DataKinds is the goal (or at least “the” goal), then I
>>     think I’d
>>     like to see stronger arguments for why it should be split.
>>
>>
>>     Cheers,
>>     Joachim
>>
>>     -- 
>>     Joachim Breitner
>>     mail at joachim-breitner.de <mailto:mail at joachim-breitner.de>
>>     http://www.joachim-breitner.de/ <http://www.joachim-breitner.de/>

-- 
Adam Gundry, Haskell Consultant
Well-Typed LLP, https://www.well-typed.com/

Registered in England & Wales, OC335890
27 Old Gloucester Street, London WC1N 3AX, England