[ghc-steering-committee] Proposal #631: Set program exit code by main return type, recommendation: accept something

[Adam Gundry]

I vote:

2
3a
That's it

on the basis articulated in 
https://github.com/ghc-proposals/ghc-proposals/pull/631#issuecomment-1985236743. 
Adding a warning solves the problem of GHC surprising users by ignoring 
`main :: IO ExitCode`. Adding an extension for a tiny bit of convenience 
seems undermotivated IMHO, and I agree that making the change without 
extension is unjustified.

I'm very sceptical of 4, because it introduces an extension that 
approximately nobody will think to use (and if it gets turned on by 
default at some point, we risk significant confusion with `main :: IO 
ExitCode` quietly doing different things on different language editions).

Adam


On 22/04/2024 15:15, Arnaud Spiwack wrote:
> Ok, so, after an extended period of time (which, obviously, is my 
> fault), we essentially have one option per opinion (well, I guess 3a and 
> 4 have two, but that's not really what consensus is like). Simon, the 
> other day, advised me to reduce the number of choices based on Shea's 
> preferences. I decided to not include Shea's original proposal of doing 
> the backward incompatible change without extension as I believe that it 
> is contrary to the spirit of the language editions and all that.
> 
> *I'm calling for a vote* on the three following options. As per our 
> customs, this is preference voting please order the following options. 
> If you want to vote against an option, rank it after “That's it” (or 
> omit it altogether). Explanation of the summary below
> 
> I'm leaving the vote open until *Wednesday 1st May*. After which, I'll 
> tally, and synthesise the committee's final position.
> 
> 2. [Summary: 00WWWN] No change in behaviour, just add a warning when 
> `main` has a type that isn't `main :: IO ()` or `main :: IO Void`, very 
> much including `main :: IO ExitCode` (Shae's second favourite alternative)
> 3a. [Summary: -XNoWombat: 00WWWW / -XWombat: TTETET] A warning is added 
> as in 2, but, additionally, an extension is introduced. When the 
> extension is turned on, we always call the proposed `ExitStatus` type 
> class on the returned value to determine the program's exit code. 
> (Shae's favourite alternative)
> 4. [Summary: -XNoWombat: 00000N / -XWombat 00I00N] No warning is 
> introduced, but an extension is. When the extension is turned on, 
> everything is as today, except when `main :: IO ExitCode`, in which case 
> we program's exit code is the exit code returned by `main` (Simon PJ's 
> favourite)
> That's it
> 
> ------
> 
> The summaries are based on the following examples. Each get a letter 
> representing the behaviour under the proposal, as described in the legend
> 
> The examples:
> module Ex1 where {  ..; main :: IO Void }
> module Ex2 where {  ..; main :: IO () }
> module Ex3 where {  ..; main :: IO Int }
> module Ex4 where { ...; main :: IO ExitCode }   -- ExitCode exists already
> module Ex5 where { ...; main :: IO Bool }         -- No ExitStatus 
> instance for Bool
> module Ex6 where { ...; data T = ..; main :: IO T; instance ExitStatus T 
> where ... }
> 
> The legend:
> - 0: main exits with exit code 0 (success)
> - E: type error
> - W: a warning is emitted
> - T: the `ExitStatus` type class is used to select the exit code.
> - I: The exit code is the returned value (only apply to `main :: IO 
> ExitCode`).
> - N: not available under this alternative
> 
> On Thu, 28 Mar 2024 at 16:18, Arnaud Spiwack <arnaud.spiwack at tweag.io 
> <mailto:arnaud.spiwack at tweag.io>> wrote:
> 
>     @SImonPJ I didn't include these two options because I hadn't
>     understood that they had any backing.
> 
>     Personally: I don't like (1c) and (4)
>        - (1c) doesn't really address Shea's concern that the behaviour
>     is currently surprising, as you need to actively turn an extension
>     on to have the new behaviour, so you need to already know that the
>     default behaviour is counterintuitive.
>        - (4) is weird without type classes. Like what happens if I `type
>     T = ExitCode; main :: IO T`? Certainly `main` must not return with
>     exit code 0.
> 
>     We are having an issue here, the typical bikeshedding issue I
>     imagine, that there's about 1 proposal per member of the committee.
>     I'm not sure how to solve this efficiently, but I don't think it'll
>     be easy to drive consensus.
> 
>     I did ask Shea for his favoured options. He told me that if he can't
>     have 1a, he prefers 3a (I promise I didn't influence him!).
> 
>     On Thu, 28 Mar 2024 at 09:18, Simon Marlow <marlowsd at gmail.com
>     <mailto:marlowsd at gmail.com>> wrote:
> 
>         I think we can discount 1a because it doesn't satisfy the
>         stability principles, right?
> 
>         Out of the others, I would probably go with 1b or 3a as the most
>         predictable behaviours. I also like Simon's (4) (gated by an
>         extension, that we hope to enable in GHC2027).
> 
>         Cheers
>         Simon
> 
>         On Tue, 26 Mar 2024 at 09:35, Arnaud Spiwack
>         <arnaud.spiwack at tweag.io <mailto:arnaud.spiwack at tweag.io>> wrote:
> 
>             Alright, so here are the plausible alternatives
> 
>             1a. New type-class-based behaviour without extension
>             1b. New type-class-based behaviour gated by an extension
>             2. Just a warning (when main isn't at type IO () or IO Void)
>             3a. A warning + the new type-class-based behaviour gated by
>             an extension. With the extension, types that don't implement
>             the type class raise an error.
>             3b. A warning + the new type-class-based behaviour gated by
>             an extension. With the extension, types that don't implement
>             the type class raise a warning (which could have a different
>             phrasing than without the extension).
> 
>             Let's vote!
> 
>             On Fri, 22 Mar 2024 at 15:30, Malte Ott
>             <malte.ott at maralorn.de <mailto:malte.ott at maralorn.de>> wrote:
> 
>                 On 2024-03-22 08:58, Arnaud Spiwack wrote:
>                  > @Malte, in my opinion, with the extension on, types
>                 which are not covered
>                  > by the type class should error out.
> 
>                 Ah, I see. Well, I am fine either way.
> 
>                 I just don’t see much value in deciding for the user
>                 which code problems are
>                 unacceptable. Especially since this will make the
>                 corresponding language
>                 extension more breaking and thus harder to make the default.
>                 Others have voiced similar opinions in the GitHub thread.
> 
>                 Best
>                 Malte
>                 _______________________________________________


-- 
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