Use of   by rustdoc makes APIs not copyable into Rust source code

[dtolnay]

Suppose we have this trait:

// lib.rs

pub trait Trait {
    fn method<T>()
    where
        T: Default;
}

Run cargo doc; it renders as:

Now let's begin writing an impl of this trait:

struct MyStruct;

impl Trait for MyStruct {
}

and copy in the method method signature directly from the documentation:

struct MyStruct;

impl Trait for MyStruct {
    fn method<T>()
    where
        T: Default {}
}

We get this distressing error from cargo check:

error: unknown start of token: \u{a0}
  --> src/main.rs:11:1
   |
11 |     where
   | ^
   |
help: Unicode character ' ' (No-Break Space) looks like ' ' (Space), but it is not
   |
11 |     where
   | +

error: unknown start of token: \u{a0}
  --> src/main.rs:11:2
   |
11 |     where
   |  ^
   |
help: Unicode character ' ' (No-Break Space) looks like ' ' (Space), but it is not
   |
11 |     where
   |  +

error: unknown start of token: \u{a0}
  --> src/main.rs:11:3
   |
11 |     where
   |   ^
   |
help: Unicode character ' ' (No-Break Space) looks like ' ' (Space), but it is not
   |
11 |     where
   |   +

error: unknown start of token: \u{a0}
  --> src/main.rs:11:4
   |
11 |     where
   |    ^
   |
help: Unicode character ' ' (No-Break Space) looks like ' ' (Space), but it is not
   |
11 |     where
   |    +

error: unknown start of token: \u{a0}
  --> src/main.rs:12:1
   |
12 |         T: Default {}
   | ^
   |
help: Unicode character ' ' (No-Break Space) looks like ' ' (Space), but it is not
   |
12 |         T: Default {}
   | +

error: unknown start of token: \u{a0}
  --> src/main.rs:12:2
   |
12 |         T: Default {}
   |  ^
   |
help: Unicode character ' ' (No-Break Space) looks like ' ' (Space), but it is not
   |
12 |         T: Default {}
   |  +

error: unknown start of token: \u{a0}
  --> src/main.rs:12:3
   |
12 |         T: Default {}
   |   ^
   |
help: Unicode character ' ' (No-Break Space) looks like ' ' (Space), but it is not
   |
12 |         T: Default {}
   |   +

error: unknown start of token: \u{a0}
  --> src/main.rs:12:4
   |
12 |         T: Default {}
   |    ^
   |
help: Unicode character ' ' (No-Break Space) looks like ' ' (Space), but it is not
   |
12 |         T: Default {}
   |    +

error: unknown start of token: \u{a0}
  --> src/main.rs:12:5
   |
12 |         T: Default {}
   |     ^
   |
help: Unicode character ' ' (No-Break Space) looks like ' ' (Space), but it is not
   |
12 |         T: Default {}
   |     +

error: unknown start of token: \u{a0}
  --> src/main.rs:12:6
   |
12 |         T: Default {}
   |      ^
   |
help: Unicode character ' ' (No-Break Space) looks like ' ' (Space), but it is not
   |
12 |         T: Default {}
   |      +

error: unknown start of token: \u{a0}
  --> src/main.rs:12:7
   |
12 |         T: Default {}
   |       ^
   |
help: Unicode character ' ' (No-Break Space) looks like ' ' (Space), but it is not
   |
12 |         T: Default {}
   |       +

error: unknown start of token: \u{a0}
  --> src/main.rs:12:8
   |
12 |         T: Default {}
   |        ^
   |
help: Unicode character ' ' (No-Break Space) looks like ' ' (Space), but it is not
   |
12 |         T: Default {}
   |        +

Hopefully there is some other way rustdoc would be able to display the same content in a way that doesn't suffer this limitation.

Is this just as simple as using white-space: pre; in the CSS and then writing ordinary spaces instead of  ? If so, that seems definitely worthwhile doing.

[dtolnay]

For now the best workaround I've found is to comment out the whole copied code using a block comment:

/*
impl Trait for MyStruct {
    fn method<T>()
    where
        T: Default {}
}
*/

Then run cargo fmt which will delete all the U+00A0 characters (unclear whether this behavior is a rustfmt bug):

/*
impl Trait for MyStruct {
    fn method<T>()
where
T: Default {}
}
*/

Then you can uncomment that code:

impl Trait for MyStruct {
    fn method<T>()
where
T: Default {}
}

and finally run rustfmt again to get normal spaces:

impl Trait for MyStruct {
    fn method<T>()
    where
        T: Default,
    {
    }
}

[est31]

The behaviour in rustdoc is pretty old, it occurs at least since 1.12 (haven't checked further back).

What is new though is the behaviour of Firefox to translate   to U+00A0. They changed this recently in version 103, mentioned in the release notes, and the bugzilla issue is here. There are years-old comments however saying that Chrome had similar behaviour.

[dtolnay]

That lines up. Yes I am using Firefox, and have been copying code from rustdoc rendered documentation for years, and don't remember it being a problem until recently.

Would it be worth triaging this as a regression, due to the recent regression in user experience, despite not being caused by a code change in rustdoc?

[est31]

Would it be worth triaging this as a regression, due to the recent regression in user experience, despite not being caused by a code change in rustdoc?

No idea really, IIRC the main purpose of those labels is to track regressions that rustc changes caused. Chromium users had the behaviour for 4+ years, maybe even since Chromium existed, so for them this is not a regression, but been the status quo.

This doesn't mean that this isn't a rustc side bug though, copying code from generated rustdoc should work. I've done it in the past myself. I usually only copy single lines however, and don't copy the indentation, so I didn't notice this change.

[est31]

More firefox issue threads:

• https://bugzilla.mozilla.org/show_bug.cgi?id=1769534
• https://bugzilla.mozilla.org/show_bug.cgi?id=1801186 is a similar bug to this one, where people complained that copy pasting code to an SQL terminal yielded in errors (as it was expecting 0x20 spaces). This should at least resolve the issues for indentation. Not sure if rustdoc emits nbsp in other situations too.