Use string interpolation instead of token streams to generate Baked mod.rs

[Manishearth]

Baked data is set up to run rustfmt on generated files, if available.

In Google's build system, we don't do that, because the hermetic build won't have a rustfmt binary available. This might be fixable with some work.

Either way, this isn't usually a big deal: baked data isn't meant to be legible anyway.

However, the mod.rs file is useful: you might need to edit it when adding a more markers. It's nice for it to always be formatted to make this easy.

Also, mod.rs is not that interesting a file from a codegen point of view anyway: the benefit of using token streams is slight to nonexistant.

[robertbastian]

However, the mod.rs file is useful: you might need to edit it when adding a more markers. It's nice for it to always be formatted to make this easy.

Baked data output is not meant to be editable like this. You can easily extend baked data output by include!-ing it, e.g. include!("bake1/mod.rs"); include!("bake2/mod.rs"); to merge two baked outputs.

I'd like to hear more about your use case for editing this file.

[Manishearth]

You can easily extend baked data output by include!-ing it, e.g. include!("bake1/mod.rs"); include!("bake2/mod.rs"); to merge two baked outputs.

mod.rs does includes but it also defines some macros, so no, this does not work. You need to edit the provider macro as well.

I'd like to hear more about your use case for editing this file.

Google has a single unified build for rust crates, so you cannot easily "build without compiled_data" to do datagen (the way ICU4X make bakeddata does). This leads to a philisophical cyclic dependency when updating baked data: we need the new versions of the crates to compile, but we need baked data for that, but we need the crates to compile to be able to run datagen, ....

To help with this, we have "stubdata", empty data that can be copied in to serve as a temporary stand-in for real data while we run datagen. We also sometimes just copy real ICU4X data.

It is easy to copy in data when we modify an existing data marker: we just replace the file. It is harder to do this for new data keys, because we need to manually edit mod.rs.

I've framed this from Google's needs, but single-feature-set build systems aren't uncommon.

[Manishearth]

FWIW at least in part I want to do this change because in situations like this I consider templated strings to be far better than token streams.