[ajross]

Isn't that exactly my point though? This is just a memcpy(). In C, you do some analysis to prove to yourself that the pointers are valid[1]. In this unsafe Rust code, the author did some analysis to prove the same thing. I mean, sure, the specific analyses use words and jargon that are different. I don't think that's particularly notable. This is C code, written in Rust.

[1] FWIW, memcpy() arguments are declared restrict post-C99, the strict aliasing thing doesn't apply, for exactly the reason you're imagining.

[steveklabnik]

> In C, you do some analysis to prove to yourself that the pointers are valid[1]

Right, and in Rust, you don't have to do it yourself: the language does it for you. If the signature were in C, you'd have to analyze the callers to make sure that this property is upheld when invoked. In Rust, the compiler does that for you.

> the strict aliasing thing doesn't apply

Yes, this is the case in this specific instance due to it being literally memcpy, but if it were any other function with the same signature, the problem would exist. Again, I picked some code at random, I'm not saying this one specific instance is even the best one. The broader point of "Rust has a type system that lets you encode more invariants than C's" is still broadly true.

[ajross]

> In Rust, the compiler does that for you.

No it doesn't? That comment is expressing a human analysis. The compiler would allow you to stuff any pointer in that you want, even ones that overlap. You're right that some side effects of the runtime can be exploited to do that analysis. But that's true of C too! (Like, "these are two separate heap blocks", or "these are owned by two separate objects", etc...). Still human analysis.

Frankly you're overselling hard here. A human author can absolutely mess that analysis up, which is the whole reason Rust calls it "unsafe" to begin with.

[steveklabnik]

I think you're misunderstanding of what I'm claiming is being checked. I don't mean the unsafe block directly. I mean that &mut Ts do not alias. That is checked by the compiler.

I'm saying that even in a codebase with a lot of unsafe, the checks that are still performed have value.

[ajross]

Sure, but C++ objects returned from operator new are likewise guaranteed not to alias. There's "value" there, but not a lot of value. And I repeat, you're overselling hard here. People who write rust like this are going to produce roughly the same amount of memory safety bugs, and pretending otherwise is frankly dangerous, IMHO.

[sophacles]

The difference is:

In c++ i could do something like:

x_ptr = new object y_ptr = x_ptr

copy(x_ptr, y_ptr)

In safe rust there is no way to call the function in question if that sort of aliasing has happened. This means that if you get a bug from your copy, its in the copy method - the possibility it's been used inappropriately has been eliminated.

It reduces the search space for problems from: everywhere that created a pointer that is ultimately used in the copy, to: the copy function itself.

It reduces the number of programmers who have to keep the memory semantics of that copy in their head from "potentially everyone" to just "those who directly implement and check copy".

Pretending that has no value is absurd.