"invalid value encountered in left_shift" for broadcasted int16 array

[asmeurer]

Reproducing code example:

For some reason, np.left_shift(np.array([0], dtype=np.int16), np.array([[31, 31, 31], [31, 31, 31], [31, 31, 31]], dtype=np.int16)) gives a floating point warning, but similar function calls do not.

>>> np.seterr(all='raise')
{'divide': 'warn', 'over': 'warn', 'under': 'ignore', 'invalid': 'warn'}
>>> np.left_shift(np.array([0], dtype=np.int8), np.array([31], dtype=np.int16))
array([0], dtype=int16)
>>> np.left_shift(np.array([0], dtype=np.int16), np.array([31], dtype=np.int16))
array([0], dtype=int16)
>>> np.left_shift(np.array([0], dtype=np.int16), np.array([[31, 31, 31]], dtype=np.int16))
array([[0, 0, 0]], dtype=int16)
>>> np.left_shift(np.array([0], dtype=np.int16), np.array([[31, 31, 31], [31, 31, 31]], dtype=np.int16))
array([[0, 0, 0],
       [0, 0, 0]], dtype=int16)
>>> np.left_shift(np.array([0], dtype=np.int16), np.array([[31, 31, 31], [31, 31, 31], [31, 31, 31]], dtype=np.int16))
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
FloatingPointError: invalid value encountered in left_shift
>>> np.left_shift(np.array([0], dtype=np.int16), np.array([[30, 30, 30], [30, 30, 30], [30, 30, 30]], dtype=np.int16))
array([[0, 0, 0],
       [0, 0, 0],
       [0, 0, 0]], dtype=int16)

Error message:

NumPy/Python version information:

>>> np.__version__
'1.21.0.dev0+1549.g7c4b2482a'
>>> import sys
>>> sys.version
'3.9.2 | packaged by conda-forge | (default, Feb 21 2021, 05:02:20) \n[Clang 11.0.1 ]'

[seberg]

This sounds either like a clang bug, or a duplicate of #18005.

EDIT: Unless we have SIMD here and that is related, but I doubt it?

[asmeurer]

I just tested on Linux (3.7.8 | packaged by conda-forge | (default, Jul 31 2020, 02:25:08) \n[GCC 7.5.0]) and didn't get the error, so that seems plausible.

[asmeurer]

#19049 doesn't seem to have fixed this, so I guess it was a different issue.

[r-devulap]

I confirmed that clang-10 also sets the invalid flag with this code. It is coming from a cvttps2dq %xmm2,%xmm5 instruction within the SHORT_left_shift_avx2 function. I have no idea why clang uses a floating point instruction for integers (the generated assembly is insanely long to understand), but the code comments indicate this isn't the first time this is happening, see:

numpy/numpy/core/src/umath/loops.c.src

Line 672 in 36cc9be

// For some reason, our macOS CI sets an "invalid" flag here, but only

The invalid flag is currently suppressed for INT and UINT. Would it be appropriate to extend this to other integer types as well?