Fingerprint out of root targets #15205
Conversation
|
r? @weihanglo rustbot has assigned @weihanglo. Use |
rustbot
added
the
S-waiting-on-review
|
This is the hack described in #12266 (comment) and hence has the same caveats that |
There was a problem hiding this comment.
Thanks for helping with this issue!
As you've said, this is hacky. Some high-level feedback:
PathSource::list_file(which then calls_list_files) is used in computing fingerprints for build scripts andcargo doc, as well as determining what is going to be packaged intocargo package. During packaging, any file outside the package root will not be packed into the resulting.cratefile. Cargo even emits a warning when the situation is detected. With this change,cargo-packagestarts including those files again. This is undesired, not even mention that where should their path be in the.cratetarball.- How do this hack know what really needs to track? It simply looks at the directory the target source file is in and assume it always at the root directory with all other relevant files. Yeah it cannot handle
#[path]we already knew, but it also doesn't respect git directory andinclude/excludeoptions. It may accidentally include more files they necessary.
To me I still think depinfo from rustdoc is a more sustainable approach. We might want to put more effort on that route than sorting out a hack.
Some other alternatives are on user side, like breaking them into sub packages for reasonable reuse.
There was a problem hiding this comment.
- allow me to link rustdoc: add support for --emit=dep-info rust#91982 which is the tracking issue for the "sustainable" approach mentioned
There was a problem hiding this comment.
Thanks! It is linked in #12266 (comment), which also calls out a relevant Cargo issue #9931.
|
Also note that the issue is not labeled as |
|
With rust-lang/rust#137684 underway, would you be interested in pivoting to integrating that into Cargo? |
### What does this PR try to resolve?
This leverages the unstable `--emit=depinfo` option from rustdoc,
so that rustdoc invocation rebuild can be better tracked
without traversing the entire directory.
Some design decisions made in the current implementation:
* Rustdoc's depinfo doesn't and shouldn't emit to `target/doc`,
as the directory is considered part of the final artifact directory.
In regard to that, we specify the dep-info output path to
the fingerprint directory of rustdoc invocation.
It looks like this
`target/debug/.fingerprint/serde-12d29d32b3b8b38f/doc-lib-serde.d`.
* We also start supporting `-Zchecksum-freshness` as a side effect.
Could make it a separate PR if desired.
* `-Zbinary-dep-depinfo` is not enabled along with this,
since doc generations don't really require any binary dependencies.
### How should we test and review this PR?
The tests added has covered these cases:
* target src outside package root, e.g., `lib.path = "../lib.rs"`
* `#[doc = include_str!("../outside/pkgroot")]`
* `#[path = "../outside/pkgroot"]`
* `env!`
### Additional information
Fixes #12266
Closes #15205
This PR partially fixes #15203 and #12266 but this is admitedly a bit hackish, let me know how you want to proceed with it.