Leaking implementation details during unification

[lcnr]

What is this

This is a design document for const generics. Any discussions about its content should be on zulip. The conclusions of these discussions should then be edited back into this issue. Please do not post any comments directly in this issue.

Content

When unifying generic constants, we have to avoid accidentally stabilizing
parts of our lowering.

fn foo<const B: bool>() -> [u8; if B { 3 } else { 7 }] {
    [0; match B {
        true => 3,
        false => 7,
    }]
}

If we allow foo to compile, we have to guarantee that we will continue supporting it
in the future. While this is fine if we intentionally allow it, it could also be allowed
by accident if we we use the same representation for if and match in the abstract representation for
generic constants.

Another issue here is the removal of noops, like redundant as-casts. This was an issue with a
previous implementation which relied on the MIR of generic constants for unification.

fn multiple_casts<const N: u8>() -> [u8; N as usize as usize] {
    [0; N as usize]
}

In general, the unification of generic constants has to add a new representation for expressions which can
be used in types. If we end up stabilizing feature(generic_const_exprs) with a suboptimal lifting from expressions
to that new representation, this will probably be quite difficult to fix while maintaining backwards compatability.