I want to re-raise this issue. This continues to be a place where we can easily make an “easy win” in terms of the compile-time performance experienced by our end-users.
In practice, I think that the number of regression reports that hit LLVM assertions “in the wild” is small. I would appreciate feedback from @compiler_subteam on whether this is true.
The idea would be roughly this:
- Disable LLVM assertions on nightly builds.
- Continue to use LLVM assertions on the CI builds prepared by bors and used by crater.
- Offer some way to easily make use of those builds for end-users.
The last bullet point is a question mark, since we have no means to make this easy right now. But I personally feel like it’s probably enough if, when we see some suspicious crash, we can easily test it with LLVM assertions enabled to see if that diagnoses the problem.
We discussed this in the core team meeting recently and I believe the general feeling was that the above plan was a good idea, modulo resolving the final bullet point. How do people feel about this? I think it’s time we make a final decision.