::drop-with-owned-fields
Safe and sound owned access to a struct
's fields in Drop
: no more unsafe
usage of
ManuallyDrop
!
The #[drop_with_owned_fields]
attribute of this crate automates and
encapsulates the process of wrapping the fields of a struct in ManuallyDrop
, which is typically
needed when having the intention to drop()
-in-place a certain field
before others are, or when actually needing to take
full, owned, access
to that field. These two operations are unsafe
, despite the notorious soundness of the whole
pattern, which is quite unfortunate. The objective of this crate is to properly identify and
automate this "notoriously sound" pattern, so as to expose a safe and sound API for users to take
advantage of, with all the power of the type system supporting them and astly nudging them away
from bugs.
Examples
Example: Defer<impl FnOnce()>
Take, for instance, the following, rather typical, example:
;
Alas, our usage of the correct FnOnce()
bound (since we only need to call it once) has made this
snippet fail!
-
Full error message:
# /* error[E0507]: cannot move out of `self.f` which is behind a mutable reference --> src/_lib.rs:37:9 | 7 | (self.f)() | ^^^^^^^^-- | | | `self.f` moved due to this call | move occurs because `self.f` has type `F`, which does not implement the `Copy` trait | note: this value implements `FnOnce`, which causes it to be moved when called --> src/_lib.rs:37:9 | 7 | (self.f)() | ^^^^^^^^ # */
Enter #[drop_with_owned_fields]
:
#
#
use drop_with_owned_fields;
Note that the second usage of #[drop_with_owned_fields]
on that
impl Drop
block is only supported by enabling the "drop-sugar"
feature of the crate, which
shall, in turn, enable the "full"
features of ::syn
(resulting in a slightly higher
from-scratch compile-time, should no other crate in the dependency tree have enabled it
already).
Without it, the Drop
block and logic would have had to be spelled out a bit more explicitly, like
so:
#
#
use *;
or if DestructuredFieldsOf::<Self>
is deemed unæsthetic:
#
#
use *;
// 👇 👇
or, you can skip the destructuring in the fn
arg position as well if you fancy:
#
#
use *;
- (the advantage of destructuring is that you can be sure not to be forgetting to properly handle
some field; with that being said, the "forgotten" fields are still owned in that
fn
body, just anonymously (or bythis
), so they get dropped, normally, at the end of thefn
.)
If you forget to impl DropWithOwnedFields (with or without sugar), like so:
# use drop_with_owned_fields;
#
#
#
you will then get the following compiler error message:
# /*
error[E0277]: the trait bound `Example: DropWithOwnedFields` is not satisfied
--> src/_lib.rs:130:1
|
6 | #[drop_with_owned_fields(as _)]
| ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ the trait `DropWithOwnedFields` is not implemented for `Example`
|
= note: The `#[drop_with_owned_fields]` annotation expects 🫵 you to provide a companion `impl` of `DropWithOwnedFields` (the whole point!).
If you have enabled the `"drop-sugar"` Cargo feature, you can even write a direct `impl` of `Drop` instead, but with a `#[drop_with_owned_fields]` annotation on top of it.
= note: this error originates in the attribute macro `drop_with_owned_fields` (...)
# */
Example: .transaction.commit()
in Drop
Another case where one may need owned access to a field in drop is when the field is doing its own
linear/affine-ish types thing, and having different "destructors" requiring and consuming an owned
self
.
Typically, transaction
handles do this for their .commit()
and .roll_back()
functions:
use drop_with_owned_fields;
use Transaction;
// where:
#
#
Unsugaring
Taking the Defer<F>
example, for instance, but renamed as Foo
:
#
#
unsugars to code along the following lines (papering over robust namespacing and privacy):
# use ;
# use *;
#
// == Unsugaring of the `Drop` impl: ==
// == Unsugaring of the `struct Foo` definition: ==
/// This is what defines `DestructuredFieldsOf<Foo<F>>` to be `FooඞFields<F>`.
#
// The `ManuallyDrop` unsafe-but-sound pattern!
// -- Niceties --
// -- 1. `.field_name` access sugar:
// Ditto for `DerefMut`
#
// -- 2. Constructor builder/helper
#
#
Mainly, notice the very important addition of a "companion struct
", FooඞFields<F>
:
The companion struct FooඞFields<…>
This is the struct
containing all of the fields laid out as they initially were for the original
Foo
definition. The trick having been to split the original Foo
definition (as input to
the macro) into two struct
definitions:
-
the
Foo<…>
original type, which does have the desired extra/customizedDrop impl
, but in exchange of that it had to forsake carrying the fields directly, using aManuallyDrop<FooඞFields<…>>
layer instead. -
a companion
FooඞFields<…>
definition, a "verbatim copy" of the original input but for its name (e.g., it has the original field definitions), but which has no extra/customizedimpl Drop
whatsoever.It is guaranteed to have the same fields as the original
Foo
definition, in terms of:-
accessing these fields, implicitly, within the
Deref{,Mut}
ofFoo
; -
deconstructing it when
impl
ementingDropWithOwnedFields
forFoo
; -
constructing this
FooFields { … }
instance, which, as we are about to see, shall be paramount for the instantiation of aFoo { … }
value.
-
-
if the
ඞ
in the name scares you, don't worry, this only happens if you have forfeited interest in naming it yourself by using theas _
attribute arg. Otherwise it can easily be renamed and madepub
lic by usingas struct YourName
instead. -
otherwise, the default name is currently left unspecified, and probably even private[^path].
In that case, the
DestructureFields
trait
can be used, especially itsFields
associated type, to still be able to refer to this type.Hence the convenience
DestructuredFieldsOf<_>
type
alias:= Fields =
This "proxy type" yields a properly specified and usable way to refer to the
…Fields
, no matter what the actual name of…Fields
ends up being 🙂.
[^path]: or rather, sealed, as in, trapped in a private module which ought to result in an
unnameable containing type path. The type itself, in the sense of actual type privacy, needs to
be pub
(or rather, at least as pub
as the original struct
definition), in order for the
DestructureFields::Fields associated type to be well-formed. And this difference
can actually be witnessed in practice, since DestructuredFieldsOf<Foo> is just as
pub
lic and reachable as Foo
is, no matter how much FooඞFields
might have been sealed / how
much private its containing mod
ule might be.
Renaming the companion struct
Since having to type DestructuredFieldsOf::<Foo<F>> all the time can be deemed
cumbersome and noisy, the #[drop_with_owned_fields]
attribute takes
this as …
attribute arg, which can optionally be of the form
as $($pub:vis)? struct $StructName:ident
:
- to both override the name of that companion struct,
- and adjust its visibility so that it be allowed to be fully
pub
lic should the author wish so:
use *;
#
#
If, on the other hand, you are fine using DestructuredFieldsOf::<Foo<F>>, or
just don't really need to, thanks to the "drop-sugar"
feature, then you are free to dismiss this
and use as _
instead.
Do note that this "dismissal" at the call-site is interpreted, by the macro, as a license to use a
private[^path] name for the companion struct
, hiding it from the docs. If you wish the companion
struct
to be pub
, then do put in the effort to say so, and name it, rather than using _
.
Braced literal construction of Foo { … }
Alas, this ceases to be available once the #[drop_with_owned_fields]
pass has happened onto Foo
's definition. This is the one and main "regression" which using this
attribute entails. Such is the price to pay for the safe-and-sound ManuallyDrop
pattern, I guess.
Indeed, instead, we have something along the lines of:
- (With
.manually_drop_fields
being a field name left private.)
This, obviously, prevents the "typical" braced-struct
-literal construction of a Foo { … }
.
# use ManuallyDrop;
#
let _foo = Foo ;
# /*
error[E0560]: struct `Foo<_>` has no field named `f`
--> src/_lib.rs:392:5
|
12 | f: || (),
| ^ `Foo<_>` does not have this field
|
# */
Instead, the workaround is to involve the perfectly-available braced-struct
-literal construction
of the FooFields { … } struct
and its eponymous fields, and then simply call .into()
to
convert it "back" into a Foo { … }
:
use drop_with_owned_fields;
// 👇
or, if you are really hang up on not naming DeferFields
:
use *;
#
#