Performance regression since resource as components

[mockersf]

Bevy version and features

main since #20934

What you did

cargo run --release --example bevymark -- --waves 60 --per-wave 500 --benchmark --mode mesh2d

What went wrong

Confirmed also on windows:

7950x, 4070ti, windows
cargo run --release --example bevymark -- --waves 60 --per-wave 500 --benchmark --mode mesh2d
13.7ms Before Resource as Components
15.3ms After Resource as Components

[alice-i-cecile]

Previously discussed in #22915. We're collectively leaning against reverting, so I'm going to close that and move the discussion to this issue.

[alice-i-cecile]

The most useful thing to do next is to gather and post a before / after trace on the commits immediately surrounding the offending PR.

Existing hypotheses as to what might have caused this:

• much slower resource_scope due to table moves
• something's no longer being in-lined
• slower system parallelism due to larger Access data structures
• overhead from more archetypes

If you have other theories, please let me know! I'm sure I've forgotten or overlooked some.

[hymm]

Ran a trace and seeing a lot of things being slower for example

    pub(crate) fn asset_events_condition(assets: Res<Self>) -> bool {
        !assets.queued_events.is_empty()
    }

this system is much slower. For Assets<Font> this is 450 ns vs 190 ns. I wouldn't be surprised if this is death by a thousand cuts because our resource access being slower. I'll keep looking into things more.