Sprinkle #[inline] across scylla-cql crate to allow inlining into scylla crate

[wprzytula]

Motivation

Having read the guidelines on inlining (highly recommended for everyone!), I've understood that adding #[inline] explicitly is required to allow inlining across crates.

Even with recent rustc enhancements (related PR), the inlining policy is very conservative at the moment - requires that an (already optimized) function subject to cross-crate inlining does not contain any asserts nor calls.

Problem

scylla-cql crate's functions are largely short ones and, at the same time, are heavily called by scylla crate's routines.

Suggestion

Adding #[inline] to scylla-cql short functions (as already done in e.g. recently added RowSerializationContext) can bring considerable performance benefits on the ser/de code path.

Additionally, we can think about small functions in scylla that could make the driver's users benefit from their inlining.

Note

Do not forget that #[inline] requires transitivity. In the following code:

fn foo(x: i32) -> i32 {
  x * 2
}

#[inline]
pub fn bar(x: i32) -> i32 {
  foo(x) + 1
}

if bar is called from another crate, it won't be inlined (however, IIUC, it's probable that the call to foo() will be inlined into bar body).
In order to enable inlining bar(), it is mandatory* that both foo and bar functions are marked with #[inline].

*not necessarily after the mentioned recent changes to rustc.

[mykaul]

Do we compile / recommend to compile with LTO?

[Lorak-mmk]

Worth noting: @piodul found out that rustc already cross-crate inlines small functions without asserts / other calls: https://doc.rust-lang.org/stable/releases.html#version-1750-2023-12-28

Do we compile / recommend to compile with LTO?

We don't compile - we have no way to, profile settings for dependencies are ignored.
We don't recommend. Should we? I'm not sure. I think it highly depends on a specific application and it's not specific to the driver at all, so it's not our place to make such recommendations. Also, Rust already gets a lot of bad rep for compile times, so I get why people might be reluctant to enable LTO.

[wprzytula]

Worth noting: @piodul found out that rustc already cross-crate inlines small functions without asserts / other calls: https://doc.rust-lang.org/stable/releases.html#version-1750-2023-12-28

Updated the issue description to mention this.