bitwise shift of booleans? #997

Activity

1. rgommers commented on Mar 3, 2026

   

   rgommers

   on Mar 3, 2026

   Member

   More actions

   Good catch, that looks like an inconsistency indeed. In gh-54, all bitwise_* functions say "integer or bool dtype" except for bitwise_left_shift and bitwise_right_shift. So that was deliberate; there were no review comments on it.

   If you look at the NumPy docs, it does also mandate integer dtype for the input arrays. Hence, the acceptance of bool arrays and upcasting them to int8 seems like an accident of implementation. It seems to just follow from the operator implementation, and as soon as the second argument is integer it doesn't matter anymore:

   >>> x = np.asarray([True, False])
   >>> x << x
   array([2, 0], dtype=int8)
   >>> x << 1
   array([2, 0])
   >>> (x << 1).dtype
   dtype('int64')
   >>> x << np.asarray([2, 3], dtype=np.int16)
   array([4, 0], dtype=int16)

   However, does x << x really make any sense to begin with? I don't think so.

   The edit that can be made to the standard is for the bitwise operator section to say this I think:

   Bitwise operators should be defined for arrays having integer and boolean data types, except << and >> which should be defined only for integer data types.

   👍React with 👍2Reacted by Matt Haberland and Athan
2. kgryte

   modified the milestones: v2025, v2026 on Mar 4, 2026
3. kgryte

   added

   Narrative ContentNarrative documentation content.Narrative documentation content.

   on Mar 4, 2026