VisibleTypeApplication and AllowAmbiguousTypes

[Adam Gundry]

On 15/03/16 14:53, Richard Eisenberg wrote:
> I'm happy to change the setting, but my logic was this: With
> -XTypeApplications, there are no ambiguous types. Of course, there
> may be types that are unambiguous locally but might be ambiguous in
> downstream modules that don't use -XTypeApplications, and this is
> what Andrew is (quite validly) getting at.

Yes. I think the key point here is that TypeApplications is typically
required at use sites, whereas AllowAmbiguousTypes is needed at
definition sites.

AllowAmbiguousTypes is a heuristic: ambiguity can arise without it, and
even before TypeApplications there were cases in which
AllowAmbiguousTypes was necessary to permit polymorphic definitions that
could be used unambiguously in some contexts. With TypeApplications, any
definition can be used given a suitable context, but ambiguity is still
a useful rule for deciding whether a definition is likely to lead to
type inference trouble.

TL;DR I agree that we should drop the implication.

All the best,

Adam


> Richard
> 
> On Mar 15, 2016, at 9:16 AM, Ben Gamari <ben at smart-cactus.org>
> wrote:
> 
>> Andrew Martin <andrew.thaddeus at gmail.com> writes:
>> 
>>> I'm posting this because Richard said it would be the best place
>>> to raise this issue. I know it's a little bit late in the GHC8
>>> development process, but I would like to get the idea out there.
>>> 
>>> To my knowledge, in GHC8, turning on VisibleTypeApplication will
>>> also turn on AllowAmbiguousTypes. I think that a better behavior
>>> for most end users would be to not have this happen.
>>> 
>>> I need AllowAmbiguousTypes turned on in modules where I want to
>>> declare functions that are presently uncalled. At a call site
>>> though, I only need VisibleTypeApplication, not both extensions.
>>> The only reason I bring this up is because having
>>> AllowAmbiguousTypes on is usually bad. It makes it possible to
>>> write functions that don't work, but you don't learn that until 
>>> you get to the call site. If I write libraries that require 
>>> VisibleTypeAppliaction to call certain functions, I don't want
>>> users accidentally ending up turning on AllowAmbiguousTypes in
>>> their own code.
>>> 
>> It sounds reasonable to me to require that those defining functions
>> to be used with type application enable AllowAmbiguousTypes. In my
>> opinion we should avoid adding extension implications unless the
>> extensions in question are essentially always used together and
>> this doesn't appear to be true in the case of TypeApplications and 
>> AllowAmbiguousTypes.
>> 
>> Cheers, - Ben



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