At opt-levels <= 1 the arithmetic operation methods do not get inlined, preventing other optimisations

[nagisa]

Consider code like this:

#![feature(core_intrinsics)]

pub fn method(a: usize, b: usize) -> usize {
    a.wrapping_sub(b)
}

pub fn intrinsic(a: usize, b: usize) -> usize {
    std::intrinsics::wrapping_sub(a, b)
}

compiler explorer

at -Copt-level=1 and lower, the generated assembly for versions with the method call will generate a function call, rather than direct operation. Once the inlining fails, other optimisations that could be done are inhibited, especially at -Copt-level=1`.

More generally speaking, I wonder if we might want to make these methods have a special annotation that would make the compiler generate the instructions directly much like it does for intrinsics right now. These seem like basic enough that #[inline(always)] might not be good enough (it being just a hint) and also possibly more expensive than necessary (something needs to do the inlining still).

[alecmocatta]

Relatedly, #74362 is an instance where not inlining a function that wraps an intrinsic inhibits DCE.

[bugadani]

#[inline(always)] might not be good enough

My naive assumption is that inline(always) would inline shorter functions more eagerly, since the relative cost of a function call is higher and also shorter functions don't contribute as much to binary bloat. Is this not the case?

[Lokathor]

Using the compiler explorer link, it seems like simply tagging the wrapping_ operations as #[inline(always)] will make them get inlined even at opt-level=0

[AngelicosPhosphoros]

These seem like basic enough that #[inline(always)] might not be good enough (it being just a hint) and also possibly more expensive than necessary (something needs to do the inlining still).

This is not true for opt-level=1, AFAIK. In this mode, LLVM runs always-inline which inlines functions without much thinking.
In higher optimization levels this pass doesn't run and inline(always) becomes hint.

[AngelicosPhosphoros]

Example of prevented optimization.
godbolt