Crashes when changing view properties with certain data about libfive

[omgitsraven]

OK, I've found a couple of new situations that will crash Ao, one of which might not be related to this thread but I'll include it anyway since I have a feeling it'll still be useful -- and the other of which is DEFINITELY related, and is the simplest test case yet, and doesn't do anything to "cheat" the distance function, which should rule that out as a possible cause...

CRASH 3:

idol_CRASH3.zip

In the line:
(scale x y z 1 0.5 1 (lambda (x y z)
change the second 1 to a 2, to make it
(scale x y z 1 0.5 2 (lambda (x y z)
and Ao will instantly crash.

CRASH 4:

idol_CRASH4.zip

Change the negative end of the Z axis of the bounding box from -6 to -8, and Ao will instantly crash.

[omgitsraven]

fwiw, turning down the resolution then scaling up the model in code (after increasing the bounds to make room) also causes Crash 2, so it does seem like it has to do with trying to render that function at a certain amount of detail, rather than just something about the literal resolution value itself.

[mkeeter]

I added CRASH4 as a unit tests, where it reproduces the problem.

The issue here is that during meshing, two evaluations of the same point give slightly different values (?!). The first evaluation gives 0, which is inside the shape; the second gives a tiny positive value (1e-9 or so), which is outside.

There's something strange going on with Eigen: if I disable vectorization with -D EIGEN_DONT_VECTORIZE -D EIGEN_DISABLE_UNALIGNED_ARRAY_ASSERT), the test passes, but other unrelated tests start failing (?!).

Anyways, 5f89d8a should work around the issue by not doing the second evaluation (so there's no room to disagree). This fixes all the tests cases above on my machine; can you pull and test it out?

[omgitsraven]

That fixes all four for me :) Thanks so much!