[cppyy] Conversion of signed char pointer / vector into int8 numpy array

[AlkaidCheng]

Check duplicate issues.

• Checked for duplicates

Description

For common data types, e.g. int, double, we can convert a pointer / vector to numpy array with a direct construction:

import numpy as np
import cppyy
import cppyy.ll

vec = cppyy.gbl.std.vector([-1, 0, 1])
# works
arr = np.array(vec)
# also works
arr = np.array(vec.data())

Reproducer

However for int8 (or signed char in C++), the conversion only works for a pointer, but not for a vector:

import numpy as np
import cppyy
import cppyy.ll

arr = np.array([-1, 0, 1], dtype=np.int8)
view = cppyy.ll.cast[f"signed char*"](arr.ctypes.data)
view.reshape((3,))
vec = cppyy.gbl.std.vector['signed char'](view)
# does not work (OverflowError: character argument not in range(0x110000))
arr = np.array(vec)
# works
arr = np.array(vec.data())

ROOT version

cppyy==3.5.0

Installation method

conda (bundled with root)

Operating system

Linux

Additional context

No response

[siliataider]

Hi, thanks for your report!

When np.array() is called on a std::vector<signed char> numpy falls back to the iterator path wince __array__ was dropped in #18208.
The vector iterator picks a converter for each element based on value_type and for signed char the one selected is CharConverter::FromMemory which fails when converting -1.

The workaround with np.array(vec.data()) works because this goes through a raw memory buffer that already carries the correct dtype.

@aaronj0 was there a plan to reintegrate __array__?
Otherwise SCharAsIntConverter exists and could be enforced for this use case..