added fuzzer cases #633
added fuzzer cases #633
Conversation
|
interesting, will check them tonight. |
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## master #633 +/- ##
=======================================
Coverage 91.40% 91.40%
=======================================
Files 35 35
Lines 4512 4514 +2
=======================================
+ Hits 4124 4126 +2
Misses 388 388 ☔ View full report in Codecov by Sentry. |
|
Wait, I wasn't running the tests right. I have a repro now. |
|
Progress, but not quite there yet. It's late, I'll pick this up tomorrow. |
| {-3.f, 3.f}, // | ||
| }); | ||
| polys.push_back({ | ||
| {1.40282347e+38f, 1.40282347e+38f}, // |
There was a problem hiding this comment.
I cut these down by a factor of ~2 because for the duplicated test case (two copies next to each other), one coordinate went to inf, so the calculated precision went to inf, and apparently CCW returns -1 when both the coordinate and the precision are inf. This doesn't seem to cause any problems in the algorithm, but our geometric check at the end is wrong. That doesn't feel like a bug particularly worth fixing, so I reduced the value a touch so that the test functions.
| @@ -286,7 +286,8 @@ class EarClip { | |||
| VertItr center = tail; | |||
| VertItr last = center; | |||
|
|
|||
| while (nextL != nextR && tail != nextR) { | |||
| while (nextL != nextR && tail != nextR && | |||
| nextL != (toLeft ? right : left)) { | |||
There was a problem hiding this comment.
This fixed the infinite loop
| @@ -550,6 +551,9 @@ class EarClip { | |||
| auto vert = poly.begin(); | |||
| polygon_.push_back({vert->idx, earsQueue_.end(), vert->pos}); | |||
| const VertItr first = std::prev(polygon_.end()); | |||
|
|
|||
| bound = glm::max( | |||
| bound, glm::max(glm::abs(first->pos.x), glm::abs(first->pos.y))); | |||
There was a problem hiding this comment.
Here I was missing a point in my precision calc, so if the first point was extreme it got the wrong precision.
|
Btw, I tried running the two fuzzer cases for several minutes and was unable to find crashing/infinite loop cases. I guess the robustness is much better now. Will try to write the fuzzer for valid/ε-valid inputs. |
* added fuzzer cases * not infinite looping anymore * fixed loop check * fixed bound calc
@pca006132 here I've added your two failing test cases from #620, but they aren't failing on my mac.