invalid-shift-base UndefinedBehaviorSanitizer error

[omkardandekar]

trimesh/trimesh/grouping.py

Line 204 in 46b6aaf

np.bitwise_xor(hashable, column << (offset * precision), out=hashable)

We use trimesh and with Python 3.10. And, during testing, UndefinedBehaviorSanitizer (UBSAN) threw an error with this operation when the contents of the column are negative values. I see that this is an optimization put in place in case "array is 2D and smallish", but is there a way to accomplish this efficiently while not triggering UndefinedBehaviorSanitizer error? I would greatly appreciate and thoughts and suggestions. Thank you.

numpy/core/src/npymath/npy_math_internal.h.src:657:18: runtime error: left shift of negative value -2
#0 0x7f4188a47bca in npy_lshiftl numpy/core/src/npymath/npy_math_internal.h.src:657
#1 0x7f4188a7e11f in LONG_left_shift_avx2

SUMMARY: UndefinedBehaviorSanitizer: invalid-shift-base numpy/core/src/npymath/npy_math_internal.h.src:657:18 in

[omkardandekar]

Just following up on this. @mikedh or others: Do you have any suggestions or advice? Thank you!

[mikedh]

Seems like we can offset to the minimum there, which I just PR'd. Looking at that function there's probably a faster way to do the grouping but this is ok for now.

How were you running the tests? It doesn't look like there's an easy way to run that behavior checker via pytest.

[omkardandekar]

Thank you, so much for looking at this @mikedh! Much appreciated. Yes, it is a bit complicated to test in isolation -- we have this integrated in a larger codebase and running sanitizer checks on that. However, I tried your fix and invalid-shift-base error is now resolved. However, this has triggered another sanitizer error:

numpy/core/src/umath/loops.c.src:640:9: runtime error: signed integer overflow: 1790832976 - -365581955 cannot be represented in type 'int

So perhaps we are going over the 32-bit representation while applying the offset? Thanks.