No aliasable, non-nullable, mutable pointers

[emberian]

In this situation:

// type jit_context_t = *mut Struct__jit_context;
pub struct Context {
    priv jcx: ffi::jit_context_t
}

impl Context {
    /// Create a new Context. Returns None on out-of-memory (as libjit does)
    pub fn new() -> Option<Context> {
        unsafe {
            ffi::jit_init();
            let x: ffi::jit_context_t = ffi::jit_context_create();
            if x == null() {
                None
            } else {
                Some(Context { jcx: x })
            }
        }
    }
}

Ideally I'd be storing jcx as some non-nullable pointer, such that Option<Context> could be a single word. Unfortunately there is no type usable today. Option<*mut T> is 16 bytes on x86_64, and will always be at least one byte more than just the pointer. Option<~T> will be 8 bytes, but needs handling about the destructor and also has aliasability guarantees, which will be invalidated by the underlying libjit code. Option<&'static mut T> also has aliasing concerns, on top of nasty lifetime hack.

There is no pointer type usable to get zero overhead in this. Given the decision on #10571 (which I agree with!) it would be nice to have something usable here. Perhaps even a #[not_null] attribute, per-field, could be usable?

[emberian]

cc @nikomatsakis @thestinger @Aatch @glaebhoerl @pnkfelix (vocal participants in #10571)

[emberian]

(In fact, ideally, that conditionally could just be written transmute(Context { jcx: x }), though some may see that as more evil than I do)

[huonw]

If we expose something like #9546 as an annotation, we can write this in a library (presumably).

[emberian]

This can't really use a non-raw pointer though because of the aliasing and mutability guarantees we have. The only type that semantically works here is *mut.

[emberian]

And specifically adt needs to care about this (it's the Option optimization), not just LLVM.

[flaper87]

cc @flaper87

[alexcrichton]

This is a general issue with smart pointers. Both Option<Rc<T>> and Option<Arc<T>> are a word larger than they need to be. Sadly, the compiler knows a lot more about ~T than it does about Rc<T>, as in the compiler can't guarantee that the pointer inside Rc<T> is never null. (food for thought)

[emberian]

If there's no opposition to a #[not_null] attribute here, I will write up an RFC.

[Aatch]

I'm not against a #[not_null] attribute as, like many attributes, it shouldn't noticeably change the semantics of the type, it should strictly be an optimization (similar to #[unsafe_no_drop_flag]).

[sfackler]

Maybe #[unsafe_not_null] in the spirit of #[unsafe_no_drop_flag] to give users a heads up that they need to make sure that they preserve the right invariants?

[emberian]

Sure, I don't really care about the name.

On Sat, Mar 29, 2014 at 7:26 PM, Steven Fackler notifications@github.comwrote:

Maybe #[unsafe_not_null] in the spirit of #[unsafe_no_drop_flag] to give
users a heads up that they need to make sure that they preserve the right
invariants?

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/13194#issuecomment-39012516
.

http://octayn.net/

[huonw]

(I feel like this issue is very similar to #7576, at least in its consequences.)

[nikomatsakis]

I think this is #7576, isn't it?

[emberian]

Well I don't specifically care about library smart pointers, but it's
similar. (My use-case is idiomatic zero-overhead FFI)

On Mon, Mar 31, 2014 at 10:55 AM, Niko Matsakis notifications@github.comwrote:

I think this is #7576 #7576,
isn't it?

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/13194#issuecomment-39097739
.

http://octayn.net/

[emberian]

rust-lang/rfcs#36