Re: OVERLAPPABLE/OVERLAPPING/OVERLAPS pragmas are confusing

Johan, to summarize: 1. If an instance is marked as OVERLAPPABLE, then clients may overlap it without having any pragmas 2. If an instance is NOT marked OVERLAPPABLE, then clients may still overlap it, but then they have to use an explicit OVERLAPPING pragma. So you should either add OVERLAPPABLE to your library, and then clients don't need to do anything, or you should remove it, and require that clients add OVERLAPPING. Note that using this mechanism across modules can be quite error prone. For example, you have to be very careful not to use an OVERLAPPABLE instance in your library, as if you do parts of the program might end up using one instance, and other parts may end up using another instance---GHC has no way of knowing about overlapping instance in client libraries, so it will simply use the best possible *local* instance. -Iavor On Tue, Aug 25, 2015 at 8:46 AM, Mikhail Glushenkov < the.dead.shall.rise@gmail.com> wrote:

Hi,

On 25 August 2015 at 14:18, Johan Tibell wrote:

...

The proposed change to my library is here: https://github.com/tibbe/cassava/pull/95/files

We remove the OverlappingInstances pragma and instead add an OVERLAPPABLE pragma like so:

instance {-# OVERLAPPABLE #-} FromField a => FromField (Maybe a) where

This causes clients of the library that previously compiled (e.g. the music-parts package) to no longer compile, due to a now lacking OVERLAPPING pragma in their code.

No, it's not quite like that. Client code can start to break when {-# LANGUAGE OverlappingInstances #-} is removed, as happened with the music-parts package. Adding an OVERLAPPABLE pragma to cassava's code made that error go away.

Client code can usually work around the problem of missing OVERLAPPABLE pragmas in the library by adding OVERLAPPING pragmas to their instances. The reason I suggested bumping cassava's version is that there may be some places in cassava that still need new pragmas that I've overlooked.

If GHC had an option for detecting overlapping instances at definition site, that'd help, I think, since then it'd be easier to find instances that need new pragmas. _______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs