btAxisSweep3Internal<BP_FP_INT_TYPE>::testAabbOverlap ?

[Graviu]

Why does bullet check the pHandle->m_maxEdges[axis] rather than pHandle->m_aabbMax.get<axis>?

while i have yet to see any false negatives (and i haven't checked too many cases), i have seen a few different cases where objects would fail the 2nd method, yet return true for the first. Isn't there more performance to be gained skipping a needless nearphase, than is gained from the memory address over m_pos check?

[Erwin Coumans]

It would be interesting to do some benchmarks on this. I would expect the additional pairs to be negligible and broadphase performance gain to outperform this extra cost.
Do you have time do some timings?

By the way, by default we are moving towards using the dynamic aabb tree-based broadphase, implemented in btDbvtBroadphase.
Thanks,
Erwin