Make small functions implicitly `#[inline]`

[jonas-schievink]

It is very common to clutter libraries with #[inline] annotations on small/trivial functions to ensure that they optimize well. It shouldn't really be necessary to do this by hand so often. Instead, the compiler could use its inlining heuristic (which already exists for the MIR inliner) to decide when it is likely to be beneficial to make a function inlineable.

This was previously attempted in #70550

[thomcc]

I'd appreciate if there were a way to get the current behavior then — something between #[inline(never)] and #[inline]... Avoiding #[inline] can substantially improve compile times.

That said supporting this complicates this issue.

[thomcc]

I stand by what I said here — but at the same time I did just go through the (tedious) process of manually adding this to a crate that had very few #[inline] annotations in rusqlite/rusqlite#834

[jonas-schievink]

I wonder if doing this plus MIR inlining might not end up improving compile times. This is only about small functions, after all.

[mati865]

What is the current status on this (or how to push it forward)?

Looking at #70550 the MIR inliner seems to already have what it needs but it's not enabled yet.

[bjorn3]

We do not currentlt store the MIR for non-#[inline] and non-generic functions in the crate metadata as necessary for inlining cross-crate.

[tmiasko]

Implemented in #116505.