Re: [Caml-list] destructive local opens

[Goswin von Brederlow <goswin-v-b@web.de>]

On Wed, Aug 05, 2015 at 08:40:21AM +0200, Petter A. Urkedal wrote:
> 2015-08-04 11:33 GMT+02:00 Goswin von Brederlow <goswin-v-b@web.de>:
> > On Tue, Aug 04, 2015 at 08:51:34AM +0200, Petter Urkedal wrote:
> >> On 2015-08-03, Daniel Bünzli wrote:
> >> > Le lundi, 3 août 2015 à 15:08, Nils Becker a écrit :
> >> > > It's possible that people actually want M.() to mean let open! more
> >> > > often than let open. For me I think that's the case.
> >> >
> >> > If you are in the vector case, I don't think that's the case. With Gg [1] I often had the following kind of subtly wrong code (can't remember the exact details but something similar):
> >> >
> >> > let ox = V2.((dot v ox) * ox) in
> >> > V2.(3 * ox + oy)
> >> >
> >> > The reality is that M.() is inherently dangerous, especially from an API evolution perspective where new identifiers with matching type may get introduced later leading to silent semantic changes in your code. So we should not encourage people to use M.!() as it's going to make the problem even more acute. Besides we should have that 44 warning by default so that we see the problems, but for now it's impossible to live with 44 and a Gg like library.
> >>
> >> This suggests another option.  If type information is available at the
> >> point this warning is emitted, then the warning could be issued only in
> >> the case when the type of the shadowing identifier matched that of the
> >> shadowed identifier.
> >>
> >> This assumes the common case for shadowing is to redefine operators or
> >> common functions at a custom type, the use of M.() being an alternative
> >> to overloading.  Loosely the warning should be emitted if the chosen
> >> identifier is not the one which would have been chosen by some sensible
> >> overloading scheme, but instead we make a simple estimate.
> >>
> >> This could still go wrong, since the type required by the context may be
> >> general than the type of both the shadowed and shadowing terms, so a
> >> better rule might be to issue the warning if both are admissible in the
> >> given context, though my guess is that's harder to implement.
> >
> > I like the idea but how feasable is it? Most of the time I figure the
> > type being infered from the operator being used. The use of M.(*)
> > makes it the custom type while simple (*) would make it int. The type
> > system would have to track the ambiguity until some other use of the
> > arguments or result decide the proper type. And if it doesn't resolve
> > then emit the warning.
> 
> Yes, it's probably not that easy to implement.  Not familiar with the
> code base.  At least it requires postponing the warning till type
> information is available, and maybe cluttering the parse tree with
> sub-nodes containing the types and locations of shadowed identifiers.
> The types at the point the check is made will have been analysed under
> the assumption that the right resolution was made.

Ahh, I was thinking wrong. In

    M.(x + y)

the type of x and y would usualy be unknown at first. The type of x
and y don't matter though. Only the existance and type of M.(+) should
be considered and would usualy be fixed by the module. Worst case you
have to add type annotations or a signature to the module to make its
API properly typed. Then when the compiler hits '+' it sees the M.(+).
It then has to keep looking for the previous (+) and compare the two
types. It could then not simply compare them but check if they can be
unified one direction or the other. So overloading int -> int -> int
with 'a -> 'a -> 'a would give the warning while int -> int -> int
with Vec.t -> Vec.t -> Vec.t would not. In case of 'a adding type
annotations to the operator would fix the warning.

> I originally assumed we had closed types, only type variables stemming
> from let-polymorphism.  That may be the case for M-bound identifiers
> in M.(...), but the context may contain local identifiers or
> file-level identifiers with unresolved existentials.  Nevertheless, I
> think the common case when we don't want to emit a warning is when
> M.(...) shadows a global binding whose type was already fixed.

MfG
	Goswin