I've noticed a bunch of "main loop too long" warnings in console when
typing in Cargo.toml. Profiling showed that the culprit is `rustc
--print cfg` call.
I moved it to the background project loading phase, where it belongs.
This highlighted a problem: we generally use single `cfg`, while it
really should be per crate.
Rather than eagerly converting JSON, we losslessly keep it as is, and
change the shape of user-submitted data at the last moment.
This also allows us to remove a bunch of wrong Defaults
7071: Pass --all-targets to "cargo check" when discovering external resources r=matklad a=WasabiFan
There is a repro case and background in the linked issue.
In short, the goal of this MR is to allow rust-analyzer to discover proc-macros which come from your tests (including, most importantly, dev-dependencies).
By default, `cargo check` implies the equivalent of `--lib --bins`, meaning it'll check your libraries and binaries -- but not tests! This means proc-macros (or, I guess, build scripts as well) weren't discovered by rust-analyzer if they came from tests.
One solution would be to manually add `--lib --bins --tests` (i.e., just augment the effective options to include tests). However, in this MR, I threw in `--all-targets`, which [according to the docs](https://doc.rust-lang.org/cargo/commands/cargo-check.html#target-selection) implies `--benches --examples` too. I have absolutely no idea what repercussions that will have on rust-analyzer for other projects, nor do I know if it's a problem that build scripts will now be discovered for tests/examples/benches. But I am not aware of a reason you _wouldn't_ want to discover these things in your examples too.
I think the main drawback of this change is that it will likely slow down the `cargo check`. At a minimum, it'll now be checking your tests _and_ their dependencies. The `cargo check` docs also say that including `--tests` as I have here may cause your lib crate to be built _twice_, once for the normal target and again for unit tests. My reading of that caveat suggests that "building twice" means it's built once for the tests _inside_ your lib, with a test profile, and again for any consumers of your lib, now using a normal release profile or similar. This doesn't seem surprising.
Very minor caveat: `--tests` will not include tests within a binary if it has `test = false` set in `Cargo.toml`. (I discovered this manually by trial-and-error, but hey, it actually says that in the docs!) This is likely not an issue, but _does_ mean that if you are -- for whatever reason -- disabling tests like that and then manually specifying `cargo test --package <...> --bin <...>` to run them, rust-analyzer will remain unaware of proc-macros in your tests.
I have confirmed this fixes the original issue in my sandbox example linked in #7034 and in my own project in which I originally discovered this. I've left it configured as my default RA language server and will report back if I notice any unexpected side-effects.
Fixes#7034
Co-authored-by: Kaelin Laundry <wasabifan@outlook.com>
This modifies the logic for calling cargo metadata so that it will use
the host platform if no explicit target platform is given. This is
needed since cargo metadata defaults to outputting information for _all_
targets.
Fixes#6908.
Since we just tried running `rustup component add`, it doesn't make sense to me
to recommend trying that again. If we're reaching this case, it's probably more
likely that rustc was installed via package manager, in which case the source
should be installed the same way (e.g. if you install the rust-src package on
Ubuntu it will install a symlink in the right place to make our sysroot
detection work).
Folks report a ton of hard-to-diagnose issues, the solution for which
is "unset RUST_SRC_PATH". Let's just ignore RUST_SRC_PATH when it
won't work anyway!
