Matching strings in beta and nightly

[hank-der-hafenarbeiter]

when matching &str in nightly or beta the following code always returns the first match case:

#[derive(Debug, PartialEq)]
enum Reg {
    EAX,
    EBX,
    ECX,
    EDX,
    ESP,
    EBP,
    ISP,
}

fn string_to_reg(_s:&str) -> Reg {
    match _s.as_ref() {
        "EAX" => Reg::EAX,
        "EBX" => Reg::EBX,
        "ECX" => Reg::ECX,
        "EDX" => Reg::EDX,
        "EBP" => Reg::EBP,
        "ESP" => Reg::ESP,
        "ISP" => Reg::ISP,
        &_    => panic!("bla bla bla"),  //removing this "&" fixes it
    }
}

fn main() {
    let vec = vec!["EAX", "EBX", "ECX", "EDX", "ESP", "EBP", "ISP"];
    let mut iter = vec.iter();
    println!("{:?}", string_to_reg(""));
    println!("{:?}", string_to_reg(iter.next().unwrap()));
    println!("{:?}", string_to_reg(iter.next().unwrap()));
    println!("{:?}", string_to_reg(iter.next().unwrap()));
    println!("{:?}", string_to_reg(iter.next().unwrap()));
    println!("{:?}",  string_to_reg(iter.next().unwrap()));     
    println!("{:?}", string_to_reg(iter.next().unwrap()));
    println!("{:?}",  string_to_reg(iter.next().unwrap()));
}

Expected behaviour is returning the fitting case. This can be fixed by removing the "&" in front of the last case.

Meta

tested on my machine with: rustc 1.11.0-nightly (bb4a79b 2016-06-15)
tested on here with beta and nightly showing the bug and stable giving a runtime error.

[oli-obk]

The MIR looks fine, and it works fine with -Zorbit enabled

[arielb1]

I think I broke old trans _match with my slice-patterns PR (#32202).

[DemiMarie]

Can we enforce MIR trans whenever slice patterns are in use?

[nikomatsakis]

Discussed in @rust-lang/compiler meeting -- @arielb1 is going to try and revert PR #32202 on the beta branch only to fix this or else find a fix. We don't have to do anything on nightly because it's correct in MIR -- though it means that old trans is kind of broken.

Also, @arielb1 says it's all @eddyb's fault. ;)

[nikomatsakis]

Update from @rust-lang/compiler meeting. @eddyb took a stab at fixing it but did not succeed. The problem is that, in the old code, literals (like "foo") and &_ are considered incompatible, and hence the match code just kind of does the wrong thing.

@eddyb points out that, on stable, this code does not currently compile. So to fix this for beta we could add some custom check that denies using & pattern to match an &str type. Or just using the old check-match.

[nikomatsakis]

This is actually fixed on beta.