Compiler lets escape borrowed pointer to local variable.

[dim-an]

Following code can be successfully compiled using rust-0.7 and from master.

fn foo<'a>(a: &'a int) -> &'a Option<int> {
    let v: &'a Option<int> = &None;
    v
}

[thestinger]

cc @nikomatsakis

[dim-an]

As far as I understand type checker considers None to be static variable. But looks like in generated code None is on the stack.

~/workspace/rust (master) » cat ~/qux.rs 
fn foo<'a>(_: &'a int) -> &Option<int> {
    let v: &Option<int> = &None;
    v
}

fn main() {
    let a = 4;
    let b = foo(&a);
    let c = [1, ..1000]; // fill some stack space
    printfln!(b);
}

~/workspace/rust (master) » rust run ~/qux.rs 
/home/ermolovd/qux.rs:9:8: 9:9 warning: unused variable: `c` [-W unused-variable (default)]
/home/ermolovd/qux.rs:9     let c = [1, ..1000]; // fill some stack space
                                ^
rust: task failed at 'enum value matched no variant', /home/ermolovd/workspace/rust/src/libstd/repr.rs:521
rust: domain main @0x232a110 root task failed

[nikomatsakis]

So, the borrow checker considers None to be a static constant, but I guess that trans disagrees and creates a local temporary. I am not sure who is correct, to be honest, we should bring it up at a meeting I guess. I feel like the borrow checker is correct, but only for unary variant values.

[nikomatsakis]

Thinking more about it, there is also this special treatment of static constants that was needed for move checking (they can be moved, which seems surprising). But this treatment makes more sense if None is just a special kind of rvalue. So perhaps trans is correct and we should modify the memory categorizer / borrow checker to categorize nullary variants as a distinguished kind of rvalue (datasort refinements would be the more principled sol'n here).

[nikomatsakis]

(cc #1679, which is somewhat related)

[emberian]

Still a bug. This is pretty nasty, and I'm pretty sure it's a soundness hole. Nominating.

[pnkfelix]

accepted for first major release, P-backcompat-lang

[dmski]

No longer reproduces with master for me
cc #11681

[alexcrichton]

Flagging as needstest.

[dmski]

There is https://github.com/mozilla/rust/blob/master/src/test/compile-fail/regions-lifetime-of-struct-or-enum-variant.rs which looks ok. Should a test with the code from issue description be added?

[alexcrichton]

Those look sufficient to me! thanks