replace type_use with LLVM's mergefunc pass

[thestinger]

The type_use implementation is buggy, leading to nightmarish code generation errors. The mergefunc pass will be strictly better in terms of code size reduction since it compares equality of the function bytecode after normalizing it. I think it will catch nearly every valid case we currently do with type_use, if not all.

We should drop the maintenance burden as soon as we get mergefunc working (or sooner) and revisit it as a potential compile-time optimization in the future.

[thestinger]

We also have to figure out where exactly to run mergefunc in the list of passes and #8957 will determine whether it leaves behind stubs.

If we run it early, we'll save on compile-time, and if we run it after optimizations we'll save more on code size. I'm sure how much mergefunc itself is going to cost yet. The big decision is whether to run it before or after the function passes. Of course, this could also vary based on optimization level.

[huonw]

Can we run it twice?

[thestinger]

Yes, but it might be too expensive.

[thestinger]

mergefunc is receiving attention upstream, so it will likely be robust enough to enable soon.

It now works well enough for libextra, but not libstd. Sadly, bugpoint fails to narrow down a test case as it ends up chasing down another bug instead...