rvalue lifetime too short when equivalent `match` works

[alexcrichton]

Sometimes the lifetime of an expression is considered too short when it's an rvalue, but the equivalent code using match works just fine:

use std::hashmap::HashMap;
struct A(int, int);
fn main() {
    let m: HashMap<int, A> = HashMap::new();

    let A(ref _a, ref _b) = *m.get(&1); // compiler error!
    let (a, b) = match *m.get(&1) { A(ref _a, ref _b) => (_a, _b) }; // ok
}

This may be related to #3511, but this seems like it may just be a bug in lifetimes somewhere?

ping @nikomatsakis

[alexcrichton]

Nominating for production-ready: after niko's patches about ref bindings in irrefut patterns landed, this form of destructuring assignment is almost there! It's really painful to use match all over the place when the majority of the surrounding code uses owned pointers.

[alexcrichton]

This is also a continuation of the bug in #4653 (the llvm assertions of which were very recently fixed).

[nikomatsakis]

I'll take a look. This does indeed look like regionck is failing to propagate the necessary constraints.

[thestinger]

This is still an issue.

[catamorphism]

Accepted for production-ready

[flaper87]

cc @flaper87

[nikomatsakis]

I...expect this may work now but I'm not entirely sure.

[flaper87]

I does compile now, I'll submit a test case for it