fn should coerce to unsafe fn

[cramertj]

Unsafe functions are a strict super-set of safe functions. It seems like fns should freely coerce to the unsafe fn of the appropriate type (i.e. fn() -> () to unsafe fn() -> ()). See this playground link for an example. Is there any particular reason not to allow this?

[petrochenkov]

fn() -> () coerces to unsafe fn() -> () right now. Modified example:

fn print_hello() -> () {
    println!("Hello!");
}

// Calls an unsafe function
fn call_unsafe(func: unsafe fn() -> ()) -> () {
    unsafe { func() }
}

fn main() {
    call_unsafe(print_hello as fn() -> ()); // OK
}

The problem is that print_hello is not a function pointer fn() -> (), but a unique unnameable function type fn() -> () {print_hello} that can be coerced to fn() -> (), so coercion to unsafe fn() -> () actually requires two coercions, and transitive coercions don't work yet, but they are supposed to work eventually.

[KalitaAlexey]

@petrochenkov Is it hard to add note to mention that as is required to make it?

[cramertj]

Yeah, I'd at least like to figure out if we can change the compiler error that occurs at the callsite. I never would have thought to cast to fn() -> () in order to get an unsafe fn() -> (), especially since a cast directly to unsafe fn() -> () errors as a non-scalar cast.

[eddyb]

@nikomatsakis Should we just allow reification from a safe function to an unsafe fn pointer?

[petrochenkov]

@KalitaAlexey

Is it hard to add note to mention that as is required to make it?

I don't know exactly, but I suspect that simply adding a direct coercion fn() {f} -> unsafe fn() is not harder than adding the note.

[KalitaAlexey]

@petrochenkov I think this is required some discussion between rust team. I'd like to have it because it doesn't do anything bad.

[nikomatsakis]

@eddyb I think so

[cramertj]

This is just a matter of combining coerce_from_fn_item and coerce_from_fn_pointer here, right?

[eddyb]

@cramertj Sort of, except the safe -> unsafe ptr case would have to also be safe -> safe in coerce_from_fn_item, so combining them might not work out that nicely.

[cramertj]

Right-- there are three possible end types for a safe fn item, compared to the two that exist currently.