mirror of
https://github.com/recp/cglm.git
synced 2026-08-14 16:30:43 -05:00
94xhn
1c260ceb7d
glm_aabb_sphere() and glm_aabb2d_circle() pick which box face to measure axis distance to via `!(a - 1)`, where `a = (s[i] < box[0][i]) + (s[i] > box[1][i])` is 1 for either out-of-range direction. `!(a - 1)` always resolves to index 1 (the max face) whenever a == 1, so when the query point is below the box's min bound on an axis, distance is measured to the wrong (max) face instead of the min face, inflating dmin and producing false-negative intersection results. The above-max direction happens to pick the correct face already, which is why this went unnoticed since the index-selection logic was introduced in #180 (fix for #179). Fix: index directly with the "above max" boolean (s[i] > box[1][i]) instead of the derived !(a - 1) expression -- it's 0 for below-min (selects box[0], correct) and 1 for above-max (selects box[1], correct), and is already computed as part of `a`/`b`/`c` so no extra branching is introduced. Same fix applied to both the 3D (box.h) and 2D (aabb2d.h) variants since they share the identical pattern. Verified against a hand-written Ericson/Graphics-Gems reference point-AABB distance test: 200k randomized box/sphere fuzz cases show 3418/200000 false negatives before this fix and 0/200000 after, with no change to the already-correct above-max and fully-enclosed cases.