Support customization of fields on derive

[greyblake]

This PR addresses #33 and enables the following syntax for derive:

#[derive(Arbitrary)]
pub struct Rgb {
    // set `r` to Default::default()
    #[arbitrary(default)]
    pub r: u8,

    // set `g` to 255
    #[arbitrary(value = 255)]
    pub g: u8,

    // generate `b` with a custom function
    #[arbitrary(with = arbitrary_b)]
    pub b: u8,
}

fn arbitrary_b(u: &mut Unstructured) -> arbitrary::Result<u8> {
    u.int_in_range(64..=128)
}

Note

In case of custom function the lower bound for hint_size is defined through core::mem::size_of::<T>() which is correct for the types allocated on stack, but not for the heap allocated types.

[fitzgen]

Thanks! A few suggestions below before we merge this in.

[fitzgen]

Can we also add some doc tests that use compile_fail to check that we return compiler errors when

• #[arbitrary] is used multiple times on a single field
• #[arbitrary(unknown_attr)] and #[arbitrary(unknown_attr = unknown_val)] are used
• #[arbitrary(value)] and #[arbitrary(with)] are used without RHS values

Should be able to just do something like the following in this file:

/// Can only use `#[arbitrary]` once per field:
///
/// ```compile_fail
/// use arbitrary::*;
/// #[derive(Arbitrary)]
/// struct Foo {
///     #[arbitrary(with = foo)]
///     #[arbitrary(with = bar)]
///     field: u32,
/// }
/// fn foo(_: &mut Unstructured) -> u32 { todo!() }
/// fn bar(_: &mut Unstructured) -> u32 { todo!() }
/// ```
///
/// Etc...
pub struct CompileFailTests;

[greyblake]

Addressed in 8913ec0

I was puzzled where to put such doc tests (I don't think users really want to see them in the docs).
Eventually I look into serde again, they're using trybuild to test compile failures.
So I've decided to introduce trybuild to this project too, I think it's a nice tool for this purpose.

[fitzgen]

Was doing doc tests on a pub struct CompileFailTests in tests/derive.rs not getting picked up by rustdoc?

If not, then we could maybe avoid a new dependency by doing

// src/lib.rs

#[cfg(all(test, feature = "derive"))]
/// ...
pub struct CompileFailTests {}

So that it never shows up in the docs and only is built when we are running tests (with the derive enabled).

[greyblake]

@fitzgen I just tried doc tests within derive/src/lib.rs, they are not picked up.

If not, then we could maybe avoid a new dependency by doing

It's dev-dependency, shouldn't it be fine? I definitely like trybuild, it generates really helpful error messages and allows to structure tests in more easy to understand / maintainable way.

For example, this is what a reported failure looks like:

This way we actually test, the compiler fails in the way we expect it to fail. Doc tests may give false positives if something in the code gets mistyped.

I would see the cost of using trybuild is rather potential failures if rustc changes the way it prints errors.
But still I think pros of trybuild outweights its cons.

Please let me know, if you still insist on using doc tests.

[fitzgen]

I think if the test is in src/lib.rs rather than derive/src/lib.rs it should work? I'd really want to avoid failing tests on nightly but not stable unless as an absolute last resort.

[greyblake]

The remaining part is using syn::Result instead of panic. I will finish it today later.

[greyblake]

@fitzgen , @Manishearth I've addressed your comments guys. Thanks to you even learned a few new things today.

The easiest way to review the changes will be to review them commit by commit.

If you merge this PR, I would very very appreciate releasing a new version of the crate, so I can benefit from the changes in my main project.
Thanks.

[fitzgen]

Thanks! Looking great.

So trybuild is kinda nice because it lets you check the error messages, but if we can't have the tests passing on both stable and nightly at the same time, then I'd prefer exploring the doc tests work around instead.

[fitzgen]

Does this mean that the tests won't pass on both nightly and stable at the same time right now?

[greyblake]

That's correct. I got one tests broken on CI, because I was running nightly locally, but CI was running stable.

OK, i'll change it to doc tests then.

[greyblake]

@fitzgen Ok, here are the doc tests: 755b8e0

[fitzgen]

Thanks!!

[greyblake]

@fitzgen @Manishearth Thank you guys for accepting the PR. Would you mind publishing a new version?

[fitzgen]

Yeah I can cut a release in a little bit

[fitzgen]

published

[fitzgen]

Thanks again!

[greyblake]

@fitzgen Thank you!