[partitioning] Don't create CGUs below a certain size

[michaelwoerister]

As a followup to #47316: Once we have a more accurate metric for CGU-size, we can also make sure to not sub-divide the program into smaller units than makes sense -- i.e. we don't want to have programs consisting of 16 one-function object files. We'd first have to find out concrete values for what is "too small" though. And I expect this to have little impact in practice. Consider it a mostly aesthetic improvement.

[michaelwoerister]

I have no data to back this up but this might allow to reduce our bootstrap time because it would affect most test cases? cc @rust-lang/infra

[aidanhs]

Just to re-state to make sure i understand: the intention would be to speed up bootstrap, while avoiding the test slowdown seen in #44675 (comment) by avoiding splitting to ensure multiple codegen units don't penalise many tiny compiles?

[michaelwoerister]

The hope would be that most test cases, which are usually small, would end up being compiled to just one codegen unit. As a result, compiling each test case would be faster because we are skipping unnecessary overhead that is introduced by splitting things into multiple CGUs.

This could be tested by making compiletest set -Ccodegen-units=1 for run-pass for example. If that makes the test suite execute faster then the optimization proposed here could be tuned to do so too.

[aidanhs]

Sounds like the previous PR might be worth looking at again as low hanging fruit then, with this being a more robust solution.

[Mark-Simulacrum]

This seems to have been implemented:

rust/compiler/rustc_monomorphize/src/partitioning.rs

Line 385 in 30dfb9e

const NON_INCR_MIN_CGU_SIZE: usize = 1800;

in #112448.