commit 6a813346b846acd0394187d3e58745b11d052a62
author Devin Jeanpierre <jeanpierreda@google.com> Mon Aug 05 17:15:18 2024 -0700
committer Copybara-Service <copybara-worker@google.com> Mon Aug 05 17:16:23 2024 -0700
tree 4e69a7d62cb3b572203af5a436786646c16022aa
parent 03247396adfb3c7ed280c942e7a7d5c78431d3c0

Zip the HIR and MIR together, and use it to expose the `ffi` aliases.

See[]

(I would suggest not approving without at least in principle approving that doc.)

## What?

The core of this CL is that, while recursing through the `ty::Ty`, we also recurse through the corresponding `hir::Ty` if we have one.
We then get to see aliases, as long as they haven't been "lost" because we lost them in the HIR. (This can happen for at least two reasons: because the HIR is in another crate (not sure about the boundaries of this?), and because the HIR had an opaque barrier, such as an inferred type, a generic type, or an associated type).

## Why?

This has the following benefits:

 *  The direct impetus is that it allows us to map `c_char` to `char`, even though it's actually a `u8` or `i8`. Rust has two 8-bit types, but the Rust standard library uses three. This provides a _relatively_ intuitive behavior. See[]

 *  In a future CL, zipping this will allow us to create bindings for aliases to non-public types. For instance, consider the following:

    https://play.rust-lang.org/?version=stable&mode=debug&edition=2021&gist=725fa6d6140ab9894b6796c71d84abf0

    ```rust
    pub trait MyTrait {
        type Type;
    }
    const _ : () = {
        pub struct Hidden;
        impl MyTrait for i32 {
            type Type = Hidden;
        }
    };
    pub type X = <i32 as MyTrait>::Type;
    ```

    We may need to see aliases to find the public instance of this, otherwise we are lost. The `rustc_middle::ty::Ty`, however, has no useful name attached to it so far as I know.

---

This is not perfect. For `c_char`, ideally we'd use a newtype, not an alias. The HIR will get lost in many predictable places:

1. When the entity is in another crate (?). For example, a trait defined in another crate. This also might be why `pub type MyChar = core::ffi::c_char` doesn't work, see TODO -- have not yet investigated, and we'd need to fully outline the limitations of this approach before the bug is marked fixed.
2. When this is an _instantiation_. For a function `fn foo<T>()`, `foo::<i8>` and `foo::<c_char>` are the _same function_, and HIR is lost here.

IMO, we will want to do this regardless of whether we can broadly embrace the use of newtypes, simply because users would _expect_ c_char etc. to work. Walking the HIR like this forces it to work in a predictable set of places: it works wherever it _can_ work, and doesn't work wherever it can't. Easy enough.

(See[]

PiperOrigin-RevId: 659732898
Change-Id: I700b771caa57ca792ceb656ca3a4a4be9499a1d9