Only propagate transforms entities with GlobalTransforms.

[StarArawn]

Objective

Fixes a performance issue when you have 1000s of entities in a bevy hierarchy without transforms.

This was prominently happening in bevy_ecs_tilemap.

Solution

Filter out entities that don't have a global transform.

Testing

CI
We should test some other way...

Migration Guide

• To avoid surprising performance pitfalls, Transform / GlobalTransform propagation is no longer performed down through hierarchies where intermediate parent are missing a GlobalTransform. To restore the previous behavior, add GlobalTransform::default to intermediate entities.

[alice-i-cecile]

I think this is the correct behavior. We shouldn't waste serious amounts of time in real applications trying to propagate transforms that don't exist.

[StarArawn]

Just to be clear in a scenario where you have this tree:
Transform entity> no transform entity > transform entity propagation would not flow up the tree.

[janhohenheim]

Approved, but I want to point out that this might be a surprising change for people who use organizational entities in their tree with only names, e.g.

• Player (has transform)
  • PhysicsStuff (has no transform)
    • Collider (has transform)
    • Joint (has transform)
    • Sensor (has transform)

This code would now suddenly break in a subtle way. Maybe a migration guide or some documentation would be good.

[alice-i-cecile]

I wrote up a nice little migration guide, and added a bit more context to the PR description :)

[superdump]

LGTM

[mockersf]

I think that it's showing something very wrong is happening.

Transform/GlobalTransform are supposed to be on all levels of the hierarchy. If you have holes, what is happening is pretty much undefined, and you're not using transforms in the way expected by the engine.

This should at minimum be a warning, at best be rejected

[mockersf]

This was prominently happening in bevy_ecs_tilemap.

What is bevy_ecs_tilemap doing to have that, and what does it expect?

[janhohenheim]

I assume they don't have holes but just big hierarchies without any transforms, which should be supported. I think the note about holes no longer being supported is just pointing out the logical consequences of this PR.

I would be careful with adding checks for warnings to transform propagation as it's quite a hot path. If they're added, I'd like to see a benchmark.

[aevyrie]

I assume they don't have holes but just big hierarchies without any transforms, which should be supported. I think the note about holes no longer being supported is just pointing out the logical consequences of this PR.

I would be careful with adding checks for warnings to transform propagation as it's quite a hot path. If they're added, I'd like to see a benchmark.

Bevy already has validation for components that should exist uninterrupted all the way up the tree, but they run in a separate system. There is no reason validation logic has to live inside the propagation code.