[Experiment] Export generic instances from libstd.

[michaelwoerister]

This should resolve issue #64140. However it is unclear if there are detrimental effects. Let's test if there are performance improvements to be had.

r? @ghost

[michaelwoerister]

@bors try @rust-timer queue

[rust-timer]

Awaiting bors try build completion

[bors]

⌛ Trying commit 73e77a7 with merge f8a2991...

[bors]

☀️ Try build successful - checks-azure
Build commit: f8a2991 (f8a2991ec6f9a64de373fd9ee5900bb0da828d42)

[rust-timer]

Queued f8a2991 with parent 212b2c7, future comparison URL.

[rust-timer]

Finished benchmarking try commit f8a2991, comparison URL.

[bjorn3]

This seems to negatively affect at least linker performance.

[panstromek]

Without actually understanding anything - it's interesting that all most regressed benchmarks are artificial little programs and stress tests, while the ones that improved are more "real" programs, like cargo, style-servo, ripgrep or inflate. Even though the page is full of red numbers, it's bit misleading because they all come from 50 LOC programs. Is that enough to even derive some conclusion?

[edit] - not entirely true, nvm

[bjorn3]

ripgrep-opt clean incremental regresses by 1.9% (linker and some other things). cranelift-codegen-opt baseline incremental regresses by 1.2% (mostly LLVM). Both are "real" programs/libraries.

[michaelwoerister]

Yes, it looks like this change causes quite a bit of overhead for loading the list of generic symbols available in the standard library. Small programs that then do not use any of these newly available symbols don't get a chance for amortizing that overhead.

ripgrep-opt clean incremental regresses by 1.9% (linker and some other things). cranelift-codegen-opt baseline incremental regresses by 1.2% (mostly LLVM). Both are "real" programs/libraries.

I wouldn't rely too much on wall-time numbers when it comes to changes below 3%. They are rather noisy.

I'm not quite sure how to proceed here. Those performance numbers clearly aren't a mandate to merge this :)

[bjorn3]

I was looking at the instructions count.