Experiment with reusing monomorphized instances across crates

[marijnh]

Currently, crates generate their own versions of all generic functions they use. They could write info about the generated functions to the metadata, and crates compiling against them could check that first, and simply link the already existing instance when possible. This is bound to reduce code output size, and thus compilation time.

[graydon]

Yeah, this is worth doing. It'd also plausibly be worth pre-specializing a variety of generics in libstd and libcore and such. Maybe some kind of #[pre_specialize(int, str)] attribute or something.

[catamorphism]

Cleared assignee.

[thestinger]

@graydon: I think being able to pre-specialize is a good solution. C++11 supports this by letting you declare template instantiations as extern in the header and providing them in some translation unit and they came to that solution after many failed alternatives.

[thestinger]

LLVM supports this with externally_available combined with linkonce_odr.

[catamorphism]

Also see #9222 (though that doesn't solve the cross-crate issue).

[catamorphism]

P-low, not 1.0

[emberian]

Visiting for triage. Hasn't been done yet.

[steveklabnik]

Triage: no updates

[alexcrichton]

cc @brson, do you feel your work in this area is enough to close this? It sounded like experiments were done, conclusions were reached, and we probably don't want to land it in the compiler at this time?

Plus you get to close a pre-2k bug :)

[brson]

My report.

After a few weeks of on-and-off work I could not find a simple heuristic that would produce reliable wins. Switching large amounts of functions from internal to some form of external linkage causes the compilation profile to change in large but subtle ways, and I couldn't find a scenario where the changes were consistently in favor of better compile times.

I haven't given up hope though. I think this technique still has clear applications to profile-based (de)optimization and pre-instantiation of std library functions.

I do think it's not worth keeping this open. If somebody has a breakthrough someday it'll happen. If not, nbd.