np.median doesn't support read-only arrays

[EmptySamurai]

np.median implementation in Numba doesn't support read-only arrays, while non-jitted version does. When I launch the snippet below

import numpy as np
from numba import jit

@jit(nopython=True)
def median(arr):
    return np.median(arr)

x = np.random.random(100)
x.setflags(write=False)

median(x)

I got an error

TypeError: Cannot modify value of type readonly array(float64, 1d, C)

My setup:

• Python 3.5.2 (Anaconda)
• numba==0.38.1
• numpy==1.14.2

[EmptySamurai]

It seems that problem is with .flatten() method, that is used by np.median.

First of all the following snippet produces the same error

import numpy as np
from numba import jit

@jit(nopython=True)
def flatten_and_change(arr):
    result = arr.flatten()
    result[0]=1
    return result

x = np.random.random(100)
x.setflags(write=False)
flatten_and_change(x)

numba.errors.InternalError: Cannot modify value of type readonly array(float64, 1d, C)

And second, I just can't return flattened array (maybe this should be considered as separate bug)

import numpy as np
from numba import jit

@jit(nopython=True)
def flatten(arr):
    return arr.flatten()

x = np.random.random(100)
x.setflags(write=False)
flatten(x)

.......
.......
~/anaconda3/lib/python3.6/site-packages/numba/targets/arrayobj.py in array_to_array(context, builder, fromty, toty, val)
   4435 def array_to_array(context, builder, fromty, toty, val):
   4436     # Type inference should have prevented illegal array casting.
-> 4437     assert toty.layout == 'A'
   4438     return val
   4439 

LoweringError: Failed at nopython (nopython mode backend)
Failed at nopython (nopython mode backend)

File "../../../../../../../../home/emptysamurai/anaconda3/lib/python3.6/site-packages/numba/targets/arrayobj.py", line 1593
[1] During: lowering "$0.8 = cast(value=$0.7)" at /home/emptysamurai/anaconda3/lib/python3.6/site-packages/numba/targets/arrayobj.py (1593)
[2] During: lowering "$0.3 = call $0.2(func=$0.2, args=[], kws=(), vararg=None)" at <ipython-input-40-8e9bf9ef1f40> (6)

[stuartarchibald]

Thanks for the report. I think #3070 (comment) is an oversight and a bug. .flatten() should return a mutable copy.

For the second example (returning flattened array), I can't reproduce on 0.39rc1, I think 24df695 (#2668) fixed it?