Make rustc's own lib directory configurable and change the default to ru...

[jhasse]

...stlib. Fixes #3319

[alexcrichton]

This seems like a fairly roundabout way to fix the issue. I much prefer @brson's suggestion which is basically what you were doing in #11045. The host DLLs should all be in bin on windows, but otherwise everything should be the same as on unixes where we have lib/$target/....

Did you have a particular use case for configuring the name of this directory?

[jhasse]

In the last comment from brson he adviced to rename the rustc dir to
rustlib (and as I interpreted it ditch his last idea). It's a pitty that no
one answered me in the issue.

IMHO this should also be merged. rustlib seems like a better name for the
dir to me.

I made it configurable because there might be some restrictions in distro
how directories in /usr/lib should be named. E.g. a distro wants the names
there to resamble the package name than this could be changed back to rustc.

[brson]

I like this solution, and never really liked the rustc name for that directory.

[alexcrichton]

I don't think that this is quite right when dealing with the snapshot. If the snapshot has one directory configured and the local installation has another, there will be a mismatch. We can guarantee that all snapshots are built with rustlib as the subdirectory.

[jhasse]

I'm not sure if I understand this correctly. Doesn't CFG_ENABLE_LOCAL_RUST mean that the build process shouldn't use a snapshot?

[alexcrichton]

Ah good point, I didn't look closely enough!

[jhasse]

Any news about merging this? Tests fail because of snapshots using rustc :/

[alexcrichton]

If there's trouble with the snapshots, you can always add #[cfg(stage0)] gates to the code, but otherwise it's up to you to fix the test errors. We're around to help out (on IRC/github) for pointers, but we don't normally take over PRs (unless ownership is agreed upon to be transferred).

[jhasse]

The problem is that the snapshot expects the old directory structure. Therefore #[cfg(stage0)] won't help me.

The solution is to rebuild the snapshots (with my trick described here: #3319 (comment) ). I can't do that.

[alexcrichton]

This is going to need some temporary logic for stage0 because we don't do special snapshots. The idea of the snapshots are to be a history of the repository, and every snapshot is based off of some commit found in the history of the master branch. If we made new custom snapshots they would not be based on a commit in master, or the snapshots would only work to merge one patch (blocking all others).

The stage0 logic may need some patches in the makefiles for now, but it'll all get removed after the next snapshot.

[brson]

The makefiles can have a bunch of lines like ifeq ($(0),0) where $(0) is the variable representing the stage to alter the stage 0 logic.

[jhasse]

:'(

working on it ...

[jhasse]

Okay new patch copies the files in both directories for stage0. Should I create a PR/patch which removes these lines which can be merged after the next snapshot?

[jhasse]

Okay I was using too many *, sorry. It did work fine on my machine (Windows and Linux) but bors has some other directory laying around apparently. New patch uses exact directory names.

[jhasse]

I'm very sorry for the failed test, I didn't notice it. This new commit should work (added use metadata::filesearch; in rpath.rs).