Add --dep-info flag to write Makefile compatible dependency files

[metajack]

This isn't super useful for libraries yet without #10593.

Fixes #7633.

[alexcrichton]

Out of curiosity, why are we reviving this issue? Or more specifically, why is the rustpkg solution not being used?

[metajack]

It doesn't look like rustpkg will be ready to handle all of Servo's needs in the near future and I want more reliable builds. With a few small changes (this + #10593) external build tooling becomes much easier to use.

[alexcrichton]

It sounds like this should maybe be a -Z flag becuase this won't be supported for forever? (nor is it the recommended way of doing things).

[metajack]

Why wouldn't we support it? There will always be reasons to use external build tooling, such as when non-Rust projects take Rust libraries as dependencies.

[pcwalton]

I agree with @metajack. This is a useful flag that we should support.

[alexcrichton]

This should also have a test, just to make sure that nothing dies along this path.

[metajack]

Rebased. Need to add a test still.

[metajack]

@alexcrichton r? test added. The rest was previously reviewed by @pcwalton although it had a minor interaction with your patches during rebase, so you might take a glace at that.

[metajack]

@alexcrichton I addressed your comments in two new commits. I'll squash it down after you review.

[metajack]

Fixed rustpkg's call to phase_6. Otherwise nothing changed.

[metajack]

Added some minor test fixes to account for TMPDIR.

[zargony]

This is cool. Makefiles are much easier now. Great work!

However, I found a slight problem. When calling rustc twice with the same crate source file, the second compile overwrites the dependency file generated by the first compile. This happens to me when compiling tests from the same source file as the library.
Because the deps file name is based on the source file name, rustc --dep-info --test lib.rs will overwrite the deps file from rustc --dep-info --dylib --rlib lib.rs.
Wouldn't it be more practical to use the target file name as the base for the deps file name? (Or maybe optionally passing a file name as command line parameter)

[metajack]

gcc does indeed have options for specifying the filename, but I didn't use that part, instead implementing the -MMD version which does similar inference as this.

A workaround (and the recommended way I think) is to have tests start from test.rs.

You might file a bug to add the ability for --dep-info to optionally take an output file name. That shouldn't be difficult to add.

[zargony]

What do you think about using the target filename with extension replaced by ".d" for the deps file? That sounds more practical to me since you could build more than one target from a source (e.g. with different --cfg settings, or library and test like I tried. I know the recommended way is to build tests from test.rs, but especially for small projects, using the same source file for tests can make things easier).

[metajack]

That sounds fine to me. Want to submit a PR? :)

[zargony]

Sure, I'll look into it. May take me a little time since I'm mostly familiar with libstd/libextra and not librustc yet, but changing the name shouldn't be too hard :)