Hello GHC HQ, I have been toying with phantom types in combination with "polymorphic" record-updates (which is a great feature imho), but stumbled over a limitation: GHC doesn't allow empty record updates (see toy example below), and I couldn't find a GHC language extension to relax this constraint. In the toy-example below it was easy to workaround by performing a dummy record update, but for more advanced uses workarounds becomes a bit more annoying. Is there a particular reason why empty record updates are disallowed by the Haskell Report? Would it be sensible, to allow empty record updates as a GHC language extension? hvr. ------------------------------------------------------- -- empty types for tagging data Clean data Dirty data Foo a = Foo { fa :: Int, fb :: String } data Bar a = Bar { ba :: Int, bb :: Foo a } markDirtyFoo :: Foo Clean -> Foo Dirty markDirtyFoo foo = foo { } -- rejected with "Empty record update" error markDirtyFoo foo = foo { fa = fa foo } -- workaround: dummy update markDirtyBar :: Bar Clean -> Bar Dirty markDirtyBar bar = bar { bb = markDirtyFoo (bb bar) } -- works
Trouble is, what type does this have? f x = x {} In your example the type annotations give the clue, but Haskell is mainly designed for type annotations to be optional. We require at least one field so we can figure out, from the field name, which type is being updated. Yes, something could doubtless be done by following the type annotations, much like higher-rank types, but it would be somewhat tricky to specify -- and the whole feature of type-changing update is (as you know) a bit obscure and not widely used, so it'd be adding complexity to an already-dark corner. See also the (inconclusive) discussion here, which would involve abolishing the entire type-changing update mechanism entirely! http://hackage.haskell.org/trac/ghc/wiki/Records Simon | -----Original Message----- | From: glasgow-haskell-users-bounces@haskell.org [mailto:glasgow-haskell-users- | bounces@haskell.org] On Behalf Of Herbert Valerio Riedel | Sent: 14 November 2011 14:31 | To: glasgow-haskell-users@haskell.org | Subject: Why not allow empty record updates? | | Hello GHC HQ, | | I have been toying with phantom types in combination with "polymorphic" | record-updates (which is a great feature imho), but stumbled over a | limitation: GHC doesn't allow empty record updates (see toy example | below), and I couldn't find a GHC language extension to relax this | constraint. In the toy-example below it was easy to workaround by | performing a dummy record update, but for more advanced uses workarounds | becomes a bit more annoying. | | Is there a particular reason why empty record updates are disallowed by | the Haskell Report? Would it be sensible, to allow empty record updates | as a GHC language extension? | | | hvr. | | ------------------------------------------------------- | -- empty types for tagging | data Clean | data Dirty | | data Foo a = Foo { fa :: Int, fb :: String } | data Bar a = Bar { ba :: Int, bb :: Foo a } | | markDirtyFoo :: Foo Clean -> Foo Dirty | markDirtyFoo foo = foo { } -- rejected with "Empty record update" error | markDirtyFoo foo = foo { fa = fa foo } -- workaround: dummy update | | markDirtyBar :: Bar Clean -> Bar Dirty | markDirtyBar bar = bar { bb = markDirtyFoo (bb bar) } -- works | | | | | | | _______________________________________________ | Glasgow-haskell-users mailing list | Glasgow-haskell-users@haskell.org | http://www.haskell.org/mailman/listinfo/glasgow-haskell-users
On 14 Nov 2011, at 22:09, Simon Peyton-Jones wrote:
Trouble is, what type does this have?
f x = x {}
f :: a -> a Empty record patterns {} are permitted, even for types that are not declared with named fields. So I don't see why an empty record update should require the type to be declared with named fields either. Regards, Malcolm
Simon Peyton-Jones wrote:
Trouble is, what type does this have? f x = x {}
Malcolm Wallace wrote:
Empty record patterns {} are permitted, even for types that are not declared with named fields. So I don't see why an empty record update should require the type to be declared with named fields either.
Yes. The translation of record updates given in the Report makes perfect sense for {}. It is only forbidden by "n >= 1", but no reason is given for that restriction. According to that translation, the type of x {} is the type of the case expression it translates to. Thanks, Yitz
Hmm yes. Fair enough. Does anyone care enough? I can see (now) that it wouldn't really be hard. | -----Original Message----- | From: glasgow-haskell-users-bounces@haskell.org [mailto:glasgow-haskell-users- | bounces@haskell.org] On Behalf Of Yitzchak Gale | Sent: 15 November 2011 11:16 | To: Malcolm Wallace | Cc: GHC-users List | Subject: Re: Why not allow empty record updates? | | Simon Peyton-Jones wrote: | >> Trouble is, what type does this have? | >> f x = x {} | | Malcolm Wallace wrote: | > Empty record patterns {} are permitted, even for types | > that are not declared with named fields. | > So I don't see why an empty record update should | > require the type to be declared with named fields either. | | Yes. The translation of record updates given in the Report | makes perfect sense for {}. It is only forbidden by | "n >= 1", but no reason is given for that restriction. | | According to that translation, the type of x {} is | the type of the case expression it translates to. | | Thanks, | Yitz | | _______________________________________________ | Glasgow-haskell-users mailing list | Glasgow-haskell-users@haskell.org | http://www.haskell.org/mailman/listinfo/glasgow-haskell-users
Quoting Yitzchak Gale
Yes. The translation of record updates given in the Report makes perfect sense for {}. It is only forbidden by "n >= 1", but no reason is given for that restriction.
It doesn't make sense to me. The translation explodes a value into a case statement over its constructors; what constructors do you use when you don't know the type of the value? When n >= 1, you know the type of the value by looking where the field came from, and hence which constructors to use in the case statement. ~d
I wrote:
Yes. The translation of record updates given in the Report makes perfect sense for {}. It is only forbidden by "n >= 1", but no reason is given for that restriction.
d wagner wrote:
It doesn't make sense to me. The translation explodes a value into a case statement over its constructors; what constructors do you use when you don't know the type of the value? When n >= 1, you know the type of the value by looking where the field came from, and hence which constructors to use in the case statement.
Well, yes, you need to know the type. That's why I asked Simon if there is difficulty with implementation. I do have an algorithm in mind, but it seemed silly for me to write it out given that I know near zero about the implementation details of GHC's type checker. But if you insist, here is a rough sketch: Replace each subexpression of the form e {} by v' and add v = e to the let bindings, where v and v' are fresh variables. Resolve the type. For any v that resolves to an ADT (even if the types of its parameters are not resolved), replace e {} in the original expression by its case expansion. Repeat until all are resolved, rejecting the program if an iteration does not resolve any v. Resolve the final form of the expression one more time to obtain its type. Does this make any sense? Thanks, Yitz
On Tue, Nov 15, 2011 at 08:34:01AM +0000, Malcolm Wallace wrote:
On 14 Nov 2011, at 22:09, Simon Peyton-Jones wrote:
Trouble is, what type does this have?
f x = x {}
f :: a -> a
That wouldn't help the original poster, as it is incompatible with f :: Foo Clean -> Foo Dirty Thanks Ian
| > > Trouble is, what type does this have? | > > | > > f x = x {} | > | > f :: a -> a | | That wouldn't help the original poster, as it is incompatible with | f :: Foo Clean -> Foo Dirty Ah! *That* is why I said it was awkward. Thanks Ian. Simon
Simon Peyton-Jones wrote:
Trouble is, what type does this have? f x = x {}
Malcolm Wallace wrote:
f :: a -> a
Ian Lynagh wrote:
That wouldn't help the original poster, as it is incompatible with f :: Foo Clean -> Foo Dirty
Only because in that expression the type of x is not known.
...the whole feature of type-changing update is (as you know) a bit obscure and not widely used, so it'd be adding complexity to an already-dark corner.
To me, at least, that is surprising. The report implies that record updates are just sugar for the given case expression. Whether or not it changes a type parameter seems unimportant. In fact, I would even advocate adding a line of explanation in the Report that this is a convenient way of copying a value from an ADT to itself with a different type as its parameter. I agree with Malcolm that this is analogous to using empty record syntax in a pattern to avoid hard-coding the number of parameter to a constructor. I usually avoid using the combination of type parameters and record syntax altogether, mainly because this obvious syntax doesn't work. Perhaps that's the reason why type-changing update is not widely used. (Admittedly, I didn't think of Herbert's trick. But doesn't that seem like somewhat of an ugly hack?) Are you hesitant because of implementation difficulty, or only because you are worried about the semantics being confusing? In my opinion, it's more confusing the way it is now. Thanks, Yitz
On 11/15/11 12:33 PM, Yitzchak Gale wrote:
Simon Peyton-Jones wrote:
Trouble is, what type does this have? f x = x {}
Malcolm Wallace wrote:
f :: a -> a
Ian Lynagh wrote:
That wouldn't help the original poster, as it is incompatible with f :: Foo Clean -> Foo Dirty
Only because in that expression the type of x is not known.
...the whole feature of type-changing update is (as you know) a bit obscure and not widely used, so it'd be adding complexity to an already-dark corner.
To me, at least, that is surprising. The report implies that record updates are just sugar for the given case expression. Whether or not it changes a type parameter seems unimportant.
In fact, I would even advocate adding a line of explanation in the Report that this is a convenient way of copying a value from an ADT to itself with a different type as its parameter. I agree with Malcolm that this is analogous to using empty record syntax in a pattern to avoid hard-coding the number of parameter to a constructor.
I usually avoid using the combination of type parameters and record syntax altogether, mainly because this obvious syntax doesn't work. Perhaps that's the reason why type-changing update is not widely used.
(Admittedly, I didn't think of Herbert's trick. But doesn't that seem like somewhat of an ugly hack?)
Are you hesitant because of implementation difficulty, or only because you are worried about the semantics being confusing? In my opinion, it's more confusing the way it is now.
For what it's worth, I do the exact same thing in the project I've been working on. The phantom type is a clean/dirty bit even :) It's an incredibly helpful thing to have for records. Especially for the context I'm in: I'm generating summary data over gobs of input, but the input can come incrementally. So long as the core of the summary is correct, then I don't care about maintaining the cache fields while I'm just shoveling data in; but I do want to make sure the caches are valid before I try to get any information out. This is exactly the sort of type-level hackery which makes Haskell a joy to work in and other languages such a pain. So far I've just defined helper functions to adjust the phantom type[1], each of which is implemented by (\x -> x { foo = foo x }). It's a horrible hack, but at least it's hidden away in library functions instead of something I have to look at. The annoying part is that when I adjust the members of the records, if I remove or rename foo then I have to fix all those coercion functions too. [1] set bit to Clean, set bit to Dirty, and unsafe set bit to 'a'. -- Live well, ~wren
Sent from my iPad
On Nov 15, 2011, at 7:18 PM, wren ng thornton
On 11/15/11 12:33 PM, Yitzchak Gale wrote:
Simon Peyton-Jones wrote:
Trouble is, what type does this have? f x = x {}
Malcolm Wallace wrote:
f :: a -> a
Ian Lynagh wrote:
That wouldn't help the original poster, as it is incompatible with f :: Foo Clean -> Foo Dirty
Only because in that expression the type of x is not known.
...the whole feature of type-changing update is (as you know) a bit obscure and not widely used, so it'd be adding complexity to an already-dark corner.
To me, at least, that is surprising. The report implies that record updates are just sugar for the given case expression. Whether or not it changes a type parameter seems unimportant.
In fact, I would even advocate adding a line of explanation in the Report that this is a convenient way of copying a value from an ADT to itself with a different type as its parameter. I agree with Malcolm that this is analogous to using empty record syntax in a pattern to avoid hard-coding the number of parameter to a constructor.
I usually avoid using the combination of type parameters and record syntax altogether, mainly because this obvious syntax doesn't work. Perhaps that's the reason why type-changing update is not widely used.
(Admittedly, I didn't think of Herbert's trick. But doesn't that seem like somewhat of an ugly hack?)
Are you hesitant because of implementation difficulty, or only because you are worried about the semantics being confusing? In my opinion, it's more confusing the way it is now.
For what it's worth, I do the exact same thing in the project I've been working on. The phantom type is a clean/dirty bit even :)
It's an incredibly helpful thing to have for records. Especially for the context I'm in: I'm generating summary data over gobs of input, but the input can come incrementally. So long as the core of the summary is correct, then I don't care about maintaining the cache fields while I'm just shoveling data in; but I do want to make sure the caches are valid before I try to get any information out. This is exactly the sort of type-level hackery which makes Haskell a joy to work in and other languages such a pain.
So far I've just defined helper functions to adjust the phantom type[1], each of which is implemented by (\x -> x { foo = foo x }). It's a horrible hack, but at least it's hidden away in library functions instead of something I have to look at. The annoying part is that when I adjust the members of the records, if I remove or rename foo then I have to fix all those coercion functions too.
My biggest issue is loss of sharing, but you could always use castFoo = asTypeOf unsafeCoerce $ \x -> x { foo = foo x } to maximize sharing, but that doesn't help with the code rewriting, Or less horrifically just carry the phantom in a newtype wrapper wrapped around your record, and cast by putting it on and taking it off, which also maximizes sharing in exchange for newtype noise on access.
| >>>>>> Trouble is, what type does this have? | >>>>>> f x = x {} | >> | >> Malcolm Wallace wrote: | >>>>> f :: a -> a | >> | >> Ian Lynagh wrote: | >>>> That wouldn't help the original poster, as it is incompatible with | >>>> f :: Foo Clean -> Foo Dirty There are several different things going on in this thread. 1. What does f x = x {} mean? The report http://www.haskell.org/onlinereport/haskell2010/haskellch3.html#x8-490003.15 says we should treat it like f x = case x of C1 a b -> C1 a b C2 v -> C2 v but if there are no fields how do we know what C1 and C2 are? The whole section only makes sense if you know x's type. So Malcolm's suggestion of "f :: forall a. a -> a" would be non-uniform with the non-empty cases. When we *do* know the type then the above translation makes sense, and even allows the f :: Foo Clean -> Foo Dirty type-change. Now two further issues arise: 2. When do we "know the type"? If the type is supposed to come from an enclosing type signature, to specify the type system one would need to specify the way that type annotations propagate. This isn't impossible (we do it for higher-rank types), but it seems like a big hammer for this particular nut. 3. Edward wants to maintain sharing, meaning presumably that no fresh record is allocated. That makes sense, but sadly System FC (GHC's intermediate language) has no way to express it. We'd need some new axioms claiming that forall ab. Foo a ~ Foo b. But that's a question for another time. Moreover, it affects non-record types just as much. Simon
participants (8)
-
Edward Kmett -
Herbert Valerio Riedel -
Ian Lynagh -
Malcolm Wallace -
Simon Peyton-Jones -
wagnerdm＠seas.upenn.edu -
wren ng thornton -
Yitzchak Gale