Try out using ccache on Travis and build LLVM from source

[alexcrichton]

We haven't tried this out in awhile, so let's see what the build times look like
if we use ccache.

[rust-highfive]

r? @brson

(rust_highfive has picked a reviewer for you, use r? to override)

[alexcrichton]

probably want to hold off on merging until we get a good idea on what the travis times look like

[larsbergstrom]

Yeah, and will probably have to get a first full build through to fill the cache and then do a retry to see how much is shaved off.

[brson]

r=me

[alexcrichton]

OK looks like an uncached build was 30m longer clocking in a 1h47m, gonna retrigger and see what the cache does now.

[alexcrichton]

Ok this clocked in at 1hr 20m (yay!) so looks like caching does indeed work and shave about a half hour. It's still about 10m slower than our builds which install via apt-get, and I also remember now that we're using that to ensure that we build against the literal 3.7 release, so I don't think we want to do this just yet. If we move more CI to Travis, however, seems like this would indeed be a real win!

[larsbergstrom]

You could also build in both configs (via the matrix/include stuff, e.g., https://github.com/servo/servo/blob/master/.travis.yml#L3-L5), to ensure that you build against literal 3.7 but check for regressions against LLVM head.

[alexcrichton]

Oh right that's true, forgot to mention that. @brson how do you feel about expanding our Travis matrix? Right now all builds are clocking in at just over an hour, and it looks like we don't have a large backlog, so expanding the matrix shouldn't hopefully flood us too much with a backlog of builds.