@mihe, can you please look at this carefully. It solves the MRP that was provided, but I might have broken other things.

It solves the MRP that was provided, but I might have broken other things.

@jrouwe It does indeed seem to break the re-adding of a SoftBody3D that has previously been removed from the scene tree, where it'll appear way off in the distance when added back into the scene tree. You can reproduce it in this modified version of the MRP found in #112348: soft-body-reentering-issue.zip

It also seems to trigger a Jolt assertion when the body is added back again:

ERROR: Jolt Physics assertion 'mat_eigvec.IsClose(eigval_eigvec, max(mat_eigvec.LengthSq(), eigval_eigvec.LengthSq()) * 1.0e-6f)' failed with message '' at 'thirdparty/jolt_physics/Jolt/Math/EigenValueSymmetric.h:88'

EDIT: It looks like the Jolt assertion happens even without this PR, so I guess that's something to address separately.

Apart from that, I'm not seeing any issues in the few projects I've tested with. Functionality that involve the vertex positions, like moving pinned vertices, seems to work fine.

I have not triggered any assert on my side.

The way the transforms are handled with soft bodies is 'interesting'. The problem is that in JoltSoftBody3D::_ref_shared_data we're getting the vertices from the mesh. When it is first inserted, these vertices are in local space. When the body is inserted the 2nd time, they are in world space, presumably because of:

I'm not sure if that means that JoltSoftBody3D::_ref_shared_data has been incorrect all along as it uses a global map to map from mesh RID to shared mesh settings. This may mean that if soft bodies A and B share the same mesh:

• insert A, remove A (this removes the shared mesh as it is was last ref), insert A -> mesh is in world space.
• insert A, insert B, remove B, insert B -> mesh from original A insert is used which is in local space -> failure (both with the old code and the new).

I have not triggered any assert on my side.

Are you using dev_build=yes? You'll need that for Jolt assertions to be enabled. I'm still seeing the assertion in the project I linked above.

The way the transforms are handled with soft bodies is 'interesting'.

Yeah, SoftBody3D as a node has never been in a good state. It needs a serious overhaul in general.

I'm not sure if that means that JoltSoftBody3D::_ref_shared_data has been incorrect all along

Well, I'm not sure how I missed this in my initial implementation, but it turns out this whole sharing of JPH::SoftBodySharedSettings is moot during actual runtime/simulation anyway, since SoftBody3D will always copy its mesh in non-editor runs in order to add/remove some flags on the mesh:

It's only in the editor (for some reason) that the meshes are left as-is and thus shared, which I guess I must have extrapolated to think that it applied to runtime as well, without actually double-checking to see if that was the case.

I guess there's reasonable justification for just removing all that stuff then, which would not only simplify the code a bit, but also mean that JoltSoftBody3D::set_stiffness_coefficient could actually update its JPH::SoftBodySharedSettings::VertexAttributes::mCompliance when changed.

Are you using dev_build=yes? You'll need that for Jolt assertions to be enabled. I'm still seeing the assertion in the project I linked above.

I was somehow assuming that an assert would trigger a breakpoint in the debugger, but I see it isn't. I'm getting the assert too.

The issue is that on reinsert, the vertices are all in world space. This causes some precision issues in the inertia tensor calculation and triggers this assert. I don't think it's a big deal as the error is relatively small and the inertia isn't used for anything anyway:

I would say we merge this change to fix the bug and then we remove the sharing of the mesh in another change. When we do that, we can get rid of:

jolt_settings->mPosition = JPH::RVec3::sZero(); // We'll be getting the vertices in world space when we re-insert the body so we need to reset the position here.

and when extracting the mesh vertices, subtract jolt_settings->mPosition so that the inertia calculation is local again. This also has the benefit that the sim won't see world space vertices on reinsert during the first update (before it recenters due to jolt_settings->mUpdatePosition = true;).

Do you want to remove the sharing code or should I do it?

Do you want to remove the sharing code or should I do it?

You're more than welcome to.

Thanks!

Cherry-picked for 4.5.2.