C++ Logo


Advanced search

Re: [std-proposals] Relocation in C++

From: Sébastien Bini <sebastien.bini_at_[hidden]>
Date: Tue, 1 Feb 2022 18:02:23 +0100
Hi all,

@Marcin Jaczewski
> This has one big flaw, `T x` destructor is called by the caller
> function of `maybe_destroy` as the lifetime of `T x` needs to be
> longer than the call to this function.
> To be able to skip this destructor there should be ABI break to
> transfer info about destruction of variables in function.

Damn it, I didn't know that :/

@Maciej Cencora <m.cencora_at_[hidden]>
> Skipping destructor call for moved-from object could be only an
> optimization and not a requirement from standard PoV.
> This way implementers that do not wish to break ABI would keep
> caller-destroy convention, and others would switch to/use
> callee-destroy convention and have better code-gen.

This goes against the relocation constructor philosophy. But your point is
Considering that, that's +1 for T(T&&)=bitcopies;

@Giuseppe D'Angelo
> class IntBuffer {
> int *ptr;
> IntBuffer() : i(new int[42]) {}
> ~IntBuffer() { delete[] i; }
> };
> Then its correct move constructor is
> IntBuffer(IntBuffer&& other) noexcept
> : ptr(std::exchange(other.ptr, nullptr)) {}
> If you have a destroying move constructor / relocation constructor, and
> know that `other` is considered destroyed, then you can skip an operation:
> IntBuffer(IntBuffer~ other) noexcept
> : ptr(other.ptr) {} // correct?

> 1) you can't do that in your relocation constructor if `other` may still
> be explicitly destroyed after that constructor call (or you'll delete
> the buffer). Your relocation constructor must reset other.ptr, which
> means, it just falls back to being the move constructor. So what's the
> point of this whole approach?
other must not go through its destructor after being passed to the
relocation constructor. Otherwise, indeed, there would be trouble.
The reloc operator takes care of that, it calls the relocation constructor
and prevents the call to the destructor. Think of an explicit call to the
relocation constructor as an explicit call to the destructor. If you
explicitly call either, we assume you know what you are doing and will
somehow don't (re)call the destructor afterwards.

> 2) under the same assumption that destruction is a QoI, according to the
> paper (but I may have misread something) `other.ptr` in the relocation
> constructor is a relocation reference. This implies that
> a) you can't even implement the relocation as the move constructor;
> you're not allowed to write something into `other.ptr` after building
> `ptr`, as `other.ptr` is destroyed (?)
The relocation constructor and move constructor are separate things and do
different businesses, one likely does a single memcpy, the other needs to
transfer resource ownership.
You can still cast a relocation reference to an rvalue reference with a
static_cast or std::move.
You can write anything you want in other.ptr but it will have little value
since other is considered destructed at the end of the constructor call.

> b) if the compiler that built the relocation constructor decided to
> destroy `other.ptr`, and then `other` is then destroyed *again*, that's
> a double destruction of the `other.ptr` subobject.
Not sure I understand. Why would the compiler decide to destroy other.ptr
in the relocation constructor? And having other destructed again is a
programming error (like doing: `IntBuffer x; x.~IntPtr();` x is destructed
twice) and should not happen. Using the reloc operator guarantees things
will be safe.


On Tue, Feb 1, 2022 at 5:10 PM Sébastien Bini <sebastien.bini_at_[hidden]>

> Hello again,
> Thank you both for your comments.
> @Gašper Ažman <gasper.azman_at_[hidden]> :
> > I'd very much like to see sections on the comparisons with the various
> papers that came in the past. It'll both show you're aware of all the
> historical context as an author, and save the readers the trouble of having
> to think through all of it when reviewing.
> I will work on it. Thank you for pointing out the proposals on the same
> topic, I find it hard to search for them. Is there some central place I'm
> missing? I only found http://www.open-std.org/jtc1/sc22/wg21/docs/papers/
> but it's not convenient to browse through all of them. Likewise, you
> mentioned this section should also "answer to the previously surfaced
> objections"; but do you know how I find them?
> I will also get in touch with Arthur O'Dwyer and Howard Hinnant. I see
> Arthur O'Dwyer is the author of one of the linked proposals but has Howard
> Hinnant worked on something similar?
> @Maciej Cencora:
> I take it your proposition is to add the reloc keyword to P1029R3. I
> honestly did not know of P1029R3, sorry about that. I think it's definitely
> something to consider.
> On one hand, I like the new T(T&&) = relocate/bitcopies; move constructor.
> I get your point, and it involves less changes in the core langage, which
> is a good thing.
> However, having a dedicated constructor for relocation allows for more
> freedom in the relocation implementation. In what I propose, the relocation
> constructor can be more than a memcpy, if needs be. I don't know if there
> are cases where relocation involves more than a memcpy, but it might
> (instance tracking?). It's something I need to ponder. But it's definitely
> a good point, thanks!
> Regards,
> Sébastien
> On Tue, Feb 1, 2022 at 12:41 PM Maciej Cencora <m.cencora_at_[hidden]>
> wrote:
>> Hi,
>> I don't think adding yet another reference type and yet another
>> special member function is a good way to solve this, as the complexity
>> in this area is already big. I think the solution should reuse as much
>> as possible of existing syntax.
>> What I think would be better, is that we kept move-constructors, but
>> we add a new syntax to mark if a class is relocatable:
>> MyClass(MyClass&& other) = relocate;
>> Marking such a constructor as relocate (no user-provided definition
>> allowed), would indicate to compiler two things: 1) moving is just a
>> trivial memcpy, 2) moved-from object is left in a state (e.g. default
>> constructed) where destructor call has no-side effect.
>> So current code:
>> MyClass newObject(std::move(other));
>> becomes (pseudo-code):
>> MyClass newObject = uninitialized;
>> new (&newObject) MyClass(other); // or memcpy
>> new (&other) MyClass();
>> // when 'other' goes out of scope, its destruction can be skipped
>> because it is a noop.
>> So far it does not change anything w.r.t. what we have now.
>> But if we add Sebastian's proposed operator reloc with such a
>> semantics that it will call move-constructor, and mark source object
>> as already destroyed we get the semantic checking that moved-from
>> object cannot be used anymore.
>> Operator reloc can be called on any type that is move-constructible,
>> it is just that for types marked as relocatable such an operation can
>> be better optimized.
>> MyClass newObject = reloc other;
>> // now other cannot be referenced any more
>> Since we do not introduce new types of references, or new types of
>> member functions, we can gradually migrate code from calling std::move
>> to using reloc operator (while preserving ABI and API compatibility).
>> This will also allow to finally have a optimal construction for types
>> with user-defined constructors:
>> struct Person
>> {
>> Person(std::string firstName, lastName)
>> : firstName(reloc firstName)
>> , lastName(reloc lastName)
>> {}
>> std::string firstName, lastName;
>> };
>> Person p1("John", "Doe"); // no temporaries, no move constructors
>> Person p2(p1.firstName, p2.lastName); // one copy, no temporaries, no
>> move constructors
>> Regards,
>> Maciej
>> wt., 1 lut 2022 o 11:04 Gašper Ažman via Std-Proposals
>> <std-proposals_at_[hidden]> napisał(a):
>> >
>> > Hi Sebastien,
>> >
>> > you sure made a pretty long write-up! What I'm missing on the first
>> skim-through is a thorough review of the currently published papers in the
>> space and answers to the previously surfaced objections.
>> >
>> > Some of the papers in this space:
>> > http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2020/p1144r5.html
>> > http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2014/n4158.pdf
>> > http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2020/p1029r3.pdf
>> > http://open-std.org/JTC1/SC22/WG21/docs/papers/2016/p0023r0.pdf
>> >
>> > I also couldn't find any way that your proposal handles maybe-destroy
>> references. For instance, how do you handle
>> >
>> > bool maybe_destroy(T& x) {
>> > if (rand() % 2) { T(static_cast<T~>(x)); return true; }
>> > return false;
>> > }
>> >
>> > The compiler will still emit the destructor call for x in the caller of
>> maybe_destroy - there's nothing in the interface that communicates whether
>> an object has been destroyed or not, *to the compiler*. Being able to
>> handle this case has been the major stumbling block for pretty much every
>> proposal in this space.
>> >
>> > G
>> >
>> > On Tue, Feb 1, 2022 at 9:26 AM Sébastien Bini via Std-Proposals <
>> std-proposals_at_[hidden]> wrote:
>> >>
>> >> Hello everyone,
>> >>
>> >> I've worked on a proposal in the last months to introduce relocation
>> (a destructive move) in C++. I wrote the proposal (you will find it
>> enclosed with this email) but it's quite large so I'll write a small
>> overview here.
>> >>
>> >> Relocation allows to "move" one object into another one. It is similar
>> to move constructors, except that relocation guarantees that moved objects
>> are destructed right after and cannot be reused.
>> >>
>> >> This proposal is motivated by:
>> >> - the confusing and often not properly implemented "moved-from" state,
>> introduced by move constructor.
>> >> - relocation is simpler to implement than move semantics as the
>> relocated object can be left in a dirty invalid state.
>> >> - const variables cannot be moved with C++ move semantics. But they
>> can "relocated" with this proposal.
>> >>
>> >> This introduces three main additions to the language:
>> >>
>> >> 1. Relocation reference
>> >>
>> >> A new type of reference, called "relocation reference". A relocation
>> reference on T is denoted by T~. This reference is mainly introduced
>> because of the new constructor.
>> >>
>> >> 2. Relocation constructor
>> >>
>> >> We introduce a new constructor, the relocation constructor:
>> >> class T {
>> >> T(T~ other) noexcept;
>> >> };
>> >>
>> >> This constructor has a new feature no other C++ constructor has: it
>> acts as a destructor with regards to its parameter ("other" in the code
>> sample). Hence when this constructor is called, a new instance is built as
>> always, but the instance it was built from is destructed. This means that
>> the destructor of the moved instance must not be called (otherwise the
>> instance would be destructed twice).
>> >>
>> >> This constructor is usually quite straightforward to implement, you
>> need simply to copy all data-members from other into the new instance.
>> >>
>> >> For instance, implementing the relocation constructor for
>> std::unique_ptr is simple. You simply need to copy the internal pointer to
>> the new instance. The moved instance can be left untouched, and the memory
>> it still owns won't be deleted as the moved instance destructor will not be
>> called.
>> >> In fact there is so little to do that the default implementation does
>> the job:
>> >> unique_ptr(unique_ptr~) noexcept = default;
>> >>
>> >> 3. The reloc operator
>> >>
>> >> Lastly, we introduce a new unary operator: reloc. It will usually be
>> used like this: auto y = reloc x;
>> >> This relocates x into y, leaving x in a destructed state.
>> >> This new operator (a) handles the construction of the new instance
>> (will use the relocation constructor, the move constructor or the copy
>> constructor, picked in that order), (b) ensures the destruction of the
>> relocated instance and (c) prevents any further use in the code of the
>> relocated instance.
>> >>
>> >> Consider the following scenario:
>> >> const T x;
>> >> auto y = reloc x;
>> >> // std::cout << x << std::endl;
>> >>
>> >> The second line builds y from x:
>> >> - Case 1: the type of x provides a relocation constructor: The
>> relocation constructor is called. At the end of the expression x is
>> considered destructed because of the relocation constructor. The destructor
>> of x will not be called when its end of scope is reached.
>> >> - Case 2: the type of x does not provide a relocation constructor, but
>> a move or copy constructor: The move constructor is called if it exists (if
>> not the copy constructor is called) to construct y from x. reloc must
>> ensure the destruction of x, so the destructor of x is called at the end of
>> the evaluation of the second line.
>> >>
>> >> The third line attempts to reuse a variable that was relocated.
>> Uncommenting this line will raise a compile error. reloc forbids further
>> mention of a name that resolves to a relocated object.
>> >>
>> >> This is it, thank you for reading :)
>> >>
>> >> Best regards,
>> >> Sébastien Bini
>> >> --
>> >> Std-Proposals mailing list
>> >> Std-Proposals_at_[hidden]
>> >> https://lists.isocpp.org/mailman/listinfo.cgi/std-proposals
>> >
>> > --
>> > Std-Proposals mailing list
>> > Std-Proposals_at_[hidden]
>> > https://lists.isocpp.org/mailman/listinfo.cgi/std-proposals

Received on 2022-02-01 17:02:35