BUG: core: Fix *_like strides for str and bytes dtype.

When a user passed a noncontiguous array to empty_like (or any other _like
function) with dtype=str or dtype=bytes, the returned array would have
strides that were all 0, even though the data type was 'U1' or 'S1'.

The problem was in the C code in the function PyArray_NewLikeArrayWithShape.
Because the default order is "keep", the branch with the comment
/* KEEPORDER needs some analysis of the strides */ was taken. That branch
uses the itemsize (i.e. dtype->elsize) of the requested dtype, but for str
or bytes, that itemsize is 0, so the computed strides end up being 0. It
is in the subsequent call to PyArray_NewFromDescr_int that the itemsize is
updated to be nonzero.

The fix is to use the same check for str or bytes that is used in
PyArray_NewFromDescr_int in PyArray_NewLikeArrayWithShape to figure out
the itemsize when computing the strides in the KEEPORDER  branch.

Closes gh-19860.

0 files changed, 0 insertions, 0 deletions