You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
solidcore: a 2D or 3D vector with a component above 1.34e154 has an infinite length, and unitize gives a zero vector instead of an error #16041
solidcore's 2D and 3D vectors measure their length by summing the squares of their components. For a component above about 1.34e154 (the square root of f64::MAX), the square overflows to infinity. Then length() returns infinity, and unitize() returns a zero vector as the "unit", with no error.
Where.solidcore/crates/solidcore/src/geom/vector2d.rs (Vec2::length_squared, length and unitize, lines 127-160 on main) and geom/vector3d.rs (Vec3::length_squared, length, unitize and unitize_or_z, from line 258). Both are SMLib's Length and Unitize, ported unchanged.
Reproduction (the same IEEE double arithmetic as the code):
let v = Vec2::new(1e155,0.0);// length_squared: 1e155 * 1e155 = infassert_eq!(v.length(),f64::INFINITY);// the length is finite: 1e155let u = v.unitize().unwrap();// Ok, not an errorassert_eq!((u.x, u.y),(0.0,0.0));// the "unit" is zero: 1e155 / inf
Expected: length() gives 1e155, and unitize() gives (1, 0). The direction is finite and well defined.
Actual: length() gives infinity, and unitize() gives (0, 0), which every caller takes as a valid unit.
The same happens for Vec3. The threshold is about sqrt(f64::MAX) = 1.3408e154.
Who reaches it. I found no real design or rig case that does: model coordinates are millimetres. A design or STEP file with a huge coordinate could reach it, if nothing on the way checks the size. I haven't confirmed which readers do.
How it was found.#16038 (a drawing's points along a line) needed a unit that never overflows. It scales the vector by its largest component before squaring (scaled3 in prism/profile.rs). That avoids the overflow for #16038's own callers only. The general fix belongs here, in geom. One open question for the fix: ported code keeps SMLib's results bit for bit (AGENTS.md §4b), so a scaled length should probably apply only where the plain one overflows. That's for triage.
Nothing is claimed. I'm leaving this for triage, as solidcore's AGENTS.md §0 asks.
Activity