Set transmute from fn item to type fn lint a hard error

[nikomatsakis]

Original title: Fn item types should be zero-sized

Per RFC 401, every fn in Rust should have its own unique, zero-sized type. This fits with the overall design of closures, which also have their own unique types (but those types are not zero-sized, as they carry the environment). None of these types have any explicit syntax at present. This has been a longstanding plan, but until recently it was not fully implemented. This is changing with PR #31710, which means that some existing code will need to be adjusted to account for the new behavior. This issue attempts to summarize how things work and describe what needs to be changed.

Zero-sized fn types

What all of this means is that, if you have a function declaration foo:

// for the purposes of this discussion, all of these different kinds of `fn` declarations are equivalent:
fn foo(x: i32) { ... }
extern "C" fn foo(x: i32);
impl i32 { fn foo(x: self) { ... } }

the type of foo is not fn(i32), as one might expect. Rather, it is a unique, zero-sized marker type that I will just write as typeof(foo). However, typeof(foo) can be coerced to a function pointer fn(i32), so you rarely notice this:

let x: fn(i32) = foo; // OK, coerces

The reason that this matter is that the type fn(i32) is not specify to any particular function: it's a function pointer. So calling x() results in a virtual call, whereas foo() is statically dispatched, because the type of foo tells us precisely what function is being called.

Impact on users

As noted above, coercions mean that most code will continue to work just fine before and after this issue is fully fixed. However, you can tell the difference in a few scenarios. Perhaps the most prominent is using transmute to convert a fn item into a fn pointer. This is often done as part of an FFI:

extern "C" fn foo(userdata: Box<i32>) {
   ...
}

let f: extern "C" fn(*mut i32) = transmute(foo);
callback(f);

Here, transmute is being used to convert the types of the fn arguments. This pattern is now incorrect because, because the type of foo is a function item (typeof(foo)), which is zero-sized, and the target type (fn()) is a function pointer, which is not zero-sized. For now, this pattern still works due to some special-cased code in the compiler, but that code is expected to be removed. For future compatibility, this pattern should be rewritten. There are a few possible ways to do this:

• change the original fn declaration to match the expected signature, and do the cast in the fn body (perhaps the best);
• cast the fn item fo a fn pointer before calling transmute, as shown here:
  • let f: extern "C" fn(*mut i32) = transmute(foo as extern "C" fn(_))
  • let f: extern "C" fn(*mut i32) = transmute(foo as usize) /* works too */

The same applies to transmutes to *mut fn(), which were observed frequently in practice. Note though that use of this type is generally incorrect. The intention is typically to describe a function pointer, but just fn() alone suffices for that. *mut fn() is a pointer to a fn pointer. (Since these values are typically just passed to C code, however, this rarely makes a difference in practice.)

Current status and implementation trail

• PR Implement unique types per fn item, rather than having all fn items have fn pointer type #19891 laid the groundwork with fresh, unique types per fn item, but left the back-end work undone
• PR Distinguish fn item types to allow reification from nothing to fn pointers. #31710 completes the work.

[nikomatsakis]

Nominating. Not sure how important this is.

[pnkfelix]

P-high, not 1.0 milestone.

[nikomatsakis]

On a related note, the subtyping relation is not quite right. This code should not type check:

fn foo<A>(_: A, _: A) { }

fn bar<B>() { }

fn main() {
    let x = bar::<u32>;
    let y = bar::<i32>;
    foo(x, y);
}

[nikomatsakis]

To elaborate, the type of x and y should be enough to precisely identify the code that's about to run, but here the types of two distinct versions of bar are considered compatible. Not good! This works out ok because we're still encoding the fn pointer anyway.

[nikomatsakis]

Put another way, the type of a fn item ought to have the values for all of its substs (sort of like bar<u32> vs bar<i32>). But the compiler is currently representing them as something like fn() {bar} -- that is, the fn signature plus a def-id.

[nikomatsakis]

Nominating. Not sure how seriously to take this (but probably not that hard to fix).

[pnkfelix]

Classifying even the type-system issue as a "bug that needs fixing" -- thus we can afford to leave this as P-high, not a 1.0-blocker.

[nikomatsakis]

Tagging as P-backcompat-lang since compiler incorrectly accepts things it shouldn't accept.

[eddyb]

Closing it as in "this has been fully implemented", by merging #34198.

[glaebhoerl]

I have to admit that I I continue to wonder if we should just define the behavior of transmute in this case to be that it reifies into a fn ptr. It's a very useful kind of conversion.

The latter point may be true, but transmute doing anything besides actual bit-for-bit transmuting feels very wrong to me.

[nikomatsakis]

The latter point may be true, but transmute doing anything besides actual bit-for-bit transmuting feels very wrong to me.

Yeah...I know.

[nikomatsakis]

cc @pnkfelix and @withoutboats -- your thoughts here?

[withoutboats]

@nikomatsakis can't you always as cast? Is making transmute just about convenience? I wouldn't like to add special cases to something as unsafe as transmute just for convenience.

[nikomatsakis]

@withoutboats you can, indeed. In any case, the proposal is to remove the special case from transmute (and hence you will get a hard error).

What I am saying is merely that I would like a more concise syntax here, since when you actually have to write (e.g.) x as fn(_, _) -> _ that is pretty long and annoying. My particular problem arose in a macro, where I also didn't have the type readily on hand.

---

@eddyb -- one thing, though it has more to do w/ the PR than this, I think we should make sure to give a very clear error message, along with a suggestion of what to write instead.

[withoutboats]

@rfcbot reviewed

you can, indeed. ... What I am saying is merely that I would like a more concise syntax here

Sounds reasonable, just wanted to make sure I understand that this special case is just about concision & not expressibility.

[eddyb]

@nikomatsakis In that case, we have to keep the special-casing code where we're not emitting the lint.
At least it'd still be an error and trans would stay clean.

[nikomatsakis]

@eddyb

In that case, we have to keep the special-casing code where we're not emitting the lint.

Yes, such is life. I left a comment in the PR.

[nikomatsakis]

Basically nice errors == special-case code. =)

[pnkfelix]

@rfcbot reviewed

[nikomatsakis]

It seems that @rfcbot went on vacation. I will put this into Final Comment Period manually:

🔔 Hear ye, hear ye! This issue is now entering a 3-week final comment period, which expires February 7, 2017. 🔔

The intention is to convert this forward compatibility warning into a hard error.

[rfcbot]

🔔 This is now entering its final comment period, as per the review above. 🔔

psst @eddyb, I wasn't able to add the final-comment-period label, please do so.

[rfcbot]

The final comment period is now complete.

[nikomatsakis]

Although the FCP is complete, the change has not been finalized, and I want to float one possible tweak to the rules that might go a long way to eliminating the annoyance that these rules can cause. Each function has a unique type, but we could make the unique types for extern "C" functions have he size of a pointer, and only make extern "Rust" functions be zero-sized. After all, a common time to send pointers is in FFI.

[wycats]

I raised this with @nikomatsakis privately because this broke a significant amount of code I wrote that I just came back to after a number of months.

In my view extern "C" fn is more or less equivalent to repr(C), and means something like: "a function intended to be called using the C ABI." In general, C code is calling functions using the C ABI, and C code also requires an actual pointer in order to work.

While there may be some cases of extern "C" functions that are only ever called from Rust, this seems exceedingly rare compared to extern "C" functions meant to be called from C. Since pointers are part of the C representation in such an exceedingly large % of cases, the extra type-cast we're imposing here seems inappropriate.

[eddyb]

@wycats You should never need the transmute in the first place, except when dealing with variadics (and we could even add coercions for those!).

[wycats]

@eddyb you don't need the transmute because you can do a manual cast?

[eddyb]

transmute is not something you should ever touch, unless there is no other way.

I suppose we should've been stricter and not allow many type combinations, where a better option exists.

[nikomatsakis]

My take on this proposal that I floated earlier:

I still think it's worth considering making extern "C" fn types not be zero-sized, but it seems like it ought to go through an RFC, and there is no reason for that decision to affect the FCP of this particular issue.

[wagenet]

The main thing that makes me sad about this is that it seems like a regression. Code that used to work now breaks without a major version change. At least there is a workaround.