[MIR] Phantom unit write after return

[arielb1]

EDIT: I was confused about the issue

pub fn foo(mut value: u64) -> usize {
    return 0;
}

fn(arg0: u64) -> usize {
    let mut var0: u64; // value
    let mut tmp0: ();
    let mut tmp1: ();

    bb0: {
        var0 = arg0;
        return = const 0;
        goto -> bb1;
    }

    bb1: {
        return;
    }

    bb2: {
        drop(tmp0) -> bb3;
    }

    bb3: {
        return = ();
        goto -> bb1;
    }
}

[arielb1]

The write here is the empty write at the end of a block - which should not exist if the block is diverging.

Unclear whether we should not create the (ill-formed) write when building or remove it in post-processing.

[nagisa]

NB: This write gets removed by simplifycfg (you need to comment the pass out to reproduce the issue).

[arielb1]

This would still pose a problem to analysis passes that want to run before SimplifyCfg. I am adding an early pass for my MIR typeck PR to ameliorate this problem.

[nagisa]

Likely caused by this

[arielb1]

The root cause of the problem is that typeck allows "diverging" blocks to have any kind of return value instead of requiring it to be MIR. I think that clearing the diverging blocks in a post-pass is a better strategy than detecting diverging blocks in HAIR.

[solson]

Maybe we should only emit what @nagisa linked to when the actual type of the block is unit?

Generally, a missing final expression doesn't necessarily mean () could be inserted, e.g.:

fn foo() -> i32 { panic!(); } // OK
fn foo() -> i32 { panic!(); () } // expected `i32`, found `()` - http://is.gd/qVOTyf

EDIT: Do we want to omit the unit write even for a block of type unit if the block is diverging? Then what I said wouldn't help.

[varkor]

As far as I can tell, this was fixed by #47746, where the trailing unit is only added if the return type of the function is the unit type.