Re: [PATCH v5 00/26] Generic `Allocator` support for Rust

[Benno Lossin <benno.lossin@proton.me>]

On 15.08.24 16:09, Danilo Krummrich wrote:
> On Thu, Aug 15, 2024 at 01:39:05PM +0000, Benno Lossin wrote:
>> On 15.08.24 15:33, Danilo Krummrich wrote:
>>> On Thu, Aug 15, 2024 at 02:34:50PM +0200, Alice Ryhl wrote:
>>>> On Thu, Aug 15, 2024 at 2:33 PM Danilo Krummrich <dakr@kernel.org> wrote:
>>>>>
>>>>> On Thu, Aug 15, 2024 at 11:20:32AM +0200, Alice Ryhl wrote:
>>>>>> On Thu, Aug 15, 2024 at 4:52 AM Danilo Krummrich <dakr@kernel.org> wrote:
>>>>>>>
>>>>>>> On Wed, Aug 14, 2024 at 12:32:15PM -0700, Boqun Feng wrote:
>>>>>>>> Hi Danilo,
>>>>>>>>
>>>>>>>> I'm trying to put your series on rust-dev, but I hit a few conflicts due
>>>>>>>> to the conflict with `Box::drop_contents`, which has been in rust-dev
>>>>>>>> for a while. And the conflict is not that trivial for me to resolve.
>>>>>>>> So just a head-up, that's a requirement for me to put it on rust-dev for
>>>>>>>> more tests from my end ;-)
>>>>>>>
>>>>>>> I rebased everything and you can fetch them from [1].
>>>>>>>
>>>>>>> I resolved the following conflicts:
>>>>>>>
>>>>>>>   - for `Box`, implement
>>>>>>>     - `drop_contents`
>>>>>>>     - `manually_drop_contents` [2]
>>>>>>
>>>>>> Not sure I like this name. It sounds like something that runs the
>>>>>> destructor, but it does the exact opposite.
>>>>>
>>>>> I thought it kinda makes sense, since it's analogous to `ManuallyDrop::new`.
>>>>>
>>>>> What about `Box::forget_contents` instead?
>>>>
>>>> One option is `into_manually_drop`. This uses the convention of using
>>>> the `into_*` prefix for conversions that take ownership of the
>>>> original value.
>>>
>>> The signature of the current `Box::manually_drop_contents` is the same as for
>>> `Box::drop_contents`, namely
>>> `fn manually_drop_contents(this: Self) -> Box<MaybeUninit<T>, A>`.
>>>
>>> `into_manually_drop` seems misleading for for returning a
>>> `Box<MaybeUninit<T>, A>`.
>>>
>>> I still think `forget_contents` hits it quite well. Just as `drop_contents`
>>> drops the value, `forget_contents` makes the `Box` forget the value.
>>
>> I think `forget_contents` sounds good. Can you please add some more docs
>> to that function though? Like an example and change "Manually drops the
>> contents, but keeps the allocation" to "Forgets the contents (does not
>> run the destructor), but keeps the allocation.".
> 
> I can't really think of a good real world example other than moving out of a
> `Box`, can you? Otherwise, maybe it just shouldn't be public?

Oh I thought you had a user for that function. In that case making it
private makes a lot of sense.

Also, `drop_contents` can be implemented with `forget_contents`, but
that might be a good follow up as a good-first-issue.

>> Another thing that I spotted while looking at the patch, `move_out`
>> doesn't need the `transmute_copy`, you should be able to just call
>> `read` on the pointer.
> 
> While technically it's the same I thought `transmute_copy` is considered better,
> since it has less stict safety requirements?

I would avoid `transmute_copy` as much as possible, transmuting
signifies to me that you somehow need to change the type and
`transmute_copy` means that the compiler isn't even able to figure out
that the two types have the same size (they are allowed to have
different size [only `Src` larger than `Dst` though], but I most often
have used this with generics where it was guaranteed to be the same
type, but the compiler failed to notice it.).

> For `transmute_copy` we only need to say, that dst has the same type as src,
> whereas for `read` the pointer must be valid for reads, properly aligned and
> point to an initialized value.

You can input a reference, so you get pointer validity for free, though
you still need to put that in the SAFETY comment.

---
Cheers,
Benno