clippy-driver has unnecessary couplings with cargo

[jsgf]

I'm trying to use clippy-driver in a non-cargo build system (Buck), by using it as a drop-in replacement for rustc. There are a few ways in which this could be improved:

1. clippy-driver looks to see if its doing a check build by explicitly looking for the string --emit=dep-info,metadata, which is the string that Cargo uses. This is brittle - for example, Buck's #check builds don't use dep-info. It should just check for the presence of metadata.
2. It also relies on CARGO_MANIFEST_DIR being set to find .?clippy.toml, and panics if it isn't. It should at least fall back to a default dir like ..
3. It relies on SYSROOT being set to find the compiler - it would be nice if there a way to derive SYSROOT from the path of the executable and inhibit the search for a rustup installation.

[oli-obk]

I think we should setup a somewhat realistic test scenario for non-cargo uses. Mostly this is covered by our compiletest suite but that one is very fine tuned to replicate some of cargo's quicks. We should just add a test script to travis that runs clippy-driver directly on a few files

[jsgf]

Thanks for the feedback - yeah I'm specifically looking for advice on how to test this. And I also want to actually deploy it locally to make sure these changes are necessary and sufficient.

[oli-obk]

For testing I'd create a script (to run locally and on CI) that runs cargo-clippy on a few files (e.g. in the tests/ui directory) and checks whether it gives the expected result.

To really test a non-rustup system we'd probably need to compile a cargo-clippy without any of the env vars set so it goes into the fallback. We probably need to create an entirely separate travis job to test that in a fully honest manner. As an intermediate solution you could touch src/driver.rs, clear the problematic env vars, and cargo build again. The resulting clippy-driver should then be fairly vanilla

[jsgf]

we'd probably need to compile a cargo-clippy without any of the env vars set

It did strike me as odd that setting compile-time env vars overrides all the runtime ways of setting things. The normal convention is to take these things in order of preference:

1. command-line arg
2. env var
3. built-in default (from env or whatever other build-time config)

(that is, more specific to a particular invocation to least)

Is there a specific reason for clippy-driver's order?

[oli-obk]

Is there a specific reason for clippy-driver's order?

Historically homegrown spaghetti code

I think it would be totally fine to invert the order.