We're currently attempting [1] to upgrade to the tip of LLVM in Rust and it looks like all our tests are green but it looks like we've got a compile-time performance regression that may prevent us from upgrading. We've tested LLVM 7 against what we're currently running against, LLVM 6, and it looks like across the board we're seeing a bit of a performance regression [2].

Taking a look at the biggest regression on that list it looks like LLVM's Jump Threading pass may be the culprit here. I've extracted some IR from the Rust compiler and verified that this command:

   $ opt -O2 lib.ll -o /dev/null

takes ~18s locally on LLVM 6 and ~40s on LLVM's master branch right now (or at least pretty close to it as of this writing). Passing `-time-passes` shows that on LLVM 6 the Jump Threading pass takes 13.7s and on LLVM 7 it takes 34.9s  where other passes remain roughly the same.

Next up I'll try to bisect this to see if I can hunt down a particular commit, but in the meantime someone else may have an idea as well!

[1]: https://github.com/rust-lang/rust/pull/51966
[2]: https://perf.rust-lang.org/compare.html?start=81d5c3e78374df15c40f463cef27918a80f3b234&end=921ebab62aa9541715ebde1db0ecf3b0d0891e92&stat=instructions%3Au