Finish rewriting TLS

[alexcrichton]

This changes the interface to get, and it also changes the keys to be static slices instead of static functions.

This allows the removal of the unsafe interface because while functions can monomorphize from different types to the same actual function, static slices cannot do this.

From at least what I can tell, we don't need to worry about LLVM coalescing these addresses. If we ever use the unnamed_addr it looks like there's cause for worry, but there doesn't appear to be any coalescing atm.

[huonw]

The non-uppercase statics lint is on by default because of #7526, and it looks like you've removed the warnings about @mut borrows being extended by rewriting find_insert as find ... insert; I believe that those warnings are harmless, so making the code slower just to avoid them would seem to be bad.

[alexcrichton]

@huonw, I don't think that we should dictate the style of statics by default. Solving #7526 doesn't involve dictating what all programs should look like, but rather a more informative error message (if any).

Also you're right that it's "faster" by using find_or_insert rather than find and then insert, the lint is far more annoying than the few nanoseconds that this is going to add to the compilation time. If this does end up hitting perf, I'll revert it and just live with the lint for now.

[huonw]

Yes, warn by default is purely until that issue is properly resolved.

[alexcrichton]

r? @bblum on the last commit?

You initially had the idea of using a vector as the TLS key values, although apparently 0-sized globals all get allocated to the same location on linux (from a linker perspective, nothing to do with LLVM). I just wanted to make sure that this idea is ok with you as well.