Dependency on LIBCLANG_PATH is not propagated back to cargo

[nagisa]

I have upgraded clang from 7.0 to 10.0 and set LIBCLANG_PATH (the path changed) to a new clang, however to my surprise cargo build did not rebuild any of the crates that depend on bindgen to generate bindings.

Bindgen should be telling cargo to rerun the build.rs for the crate if following are true:

1. The used libclang shared library has changed;
2. LIBCLANG_PATH has changed…

and whatever other things can influence loading of the libclang.

[kulp]

I do not know all of the things that "can influence loading of the libclang", and my proposed #1809 may or may not sufficiently cover all the cases, but I think it meets the main need here.

[emilio]

Well, #1809 only addresses bindgen's tests. Ideally we should do something a bit more automatic here... Maybe CargoCallbacks should print this at some point, or Bindgen's generate if we detect we're running under Cargo? (what is the right way to detect that? Checking for OUT_DIR?)

[kulp]

@emilio, perhaps I misunderstand, but my demonstration in #1809 was meant to show that changing LIBCLANG_PATH when building a downstream crate does cause bindings to be regenerated, so this does not affect only bindgen's tests, I think.

[emilio]

Ah, indeed. We invoke testgen::main unconditionally, so it ends up working... I'd say it makes sense for that bit to be outside of the testgen module though, as it's not related to test generation.

But yeah, this should be closed, behavior here after your PR is correct, thanks @kulp!

[kulp]

Good point about testgen -- I had not looked very carefully at where I was adding the code, I admit (I just added it in the vicinity of some other cargo:rerun prints). I will look for a better place.

[kulp]

I see you are way ahead of me in #1811. Thanks !

[kulp]

I guess we should probably check BINDGEN_EXTRA_CLANG_ARGS too, from #1537, so I created #1813.