Change {Box,Arc,Rc,Weak}::into_raw to only work with A = Global

[Amanieu]

Also applies to Vec::into_raw_parts.

The expectation is that you can round-trip these methods with from_raw, but this is only true when using the global allocator. With custom allocators you should instead be using into_raw_with_allocator and from_raw_in.

The implementation of Box::leak is changed to use Box::into_raw_with_allocator and explicitly leak the allocator (which was already the existing behavior). This is because, for leak to be safe, the allocator must not free its underlying backing store. The Allocator trait only guarantees that allocated memory remains valid until the allocator is dropped.

[rustbot]

r? @joboet

rustbot has assigned @joboet.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

[Amanieu]

I don't know why this is failing on Miri... I only moved the methods and didn't change any implementation (except Box::leak).

[Amanieu]

@RalfJung Do you have any idea why these miri tests would fail? The only relevant change is moving Box::into_raw from Box<T, A> to Box<T, Global>.

[RalfJung]

Yeah it has to do with this comment. This is part of the general issue that adding allocators to Box was done without considering the fact that Box is a primitive type that has special treatment in the operational semantics (and all over the compiler).

[Amanieu]

Sure, but the implementation of Box::into_raw hasn't changed at all in this PR: https://github.com/Amanieu/rust/blob/ac178fc93c2e3ee1a7e53a14db69551a3207fb66/library/alloc/src/boxed.rs#L1152

I don't see why changing the bounds on that methods from Box<T, A> to Box<T, Global> makes miri tests now fail.

[RalfJung]

Miri has no choice but use some terrible hacks to work around the poor state that Box is in due to custom allocators at the moment. Those hacks are sensitive to whether code is generic or monomorphic. I'll try to take a look and see if we can update those hacks to be compatible with this PR.

[RalfJung]

I pushed a commit that should hopefully fix CI.

[joboet]

r=me pending approval from the rest of libs-api

[safinaskar]

@rustbot label +A-box +A-allocators +A-MIR +A-miri