Currently, the following test case fails:

>>> import array
>>> a=array.array('u', 'foo')
>>> memoryview(a)==memoryview(a)
False

This is because the memoryview object doesn't support the u and w codes, as it should per PEP-3118. This patch fixes it.

Nick's comment in msg167963 got me thinking. Indeed, in Numpy the 'U'
specifier is similar to the struct module's 's' format code, only for
UCS4. So I'm questioning whether the current semantics of 'u' and 'w'
used by array.array were ever intended by the PEP authors:

import numpy

>>> nd = numpy.array(["A", "B"], dtype='U')
>>> nd
array(['A', 'B'],
      dtype='<U1')
>>> nd.tostring()
b'A\x00\x00\x00B\x00\x00\x00'
>>>
>>> nd = numpy.array(["ABC", "D"], dtype='U')
>>> nd
array(['ABC', 'D'],
      dtype='<U3')
>>> nd.tostring()
b'A\x00\x00\x00B\x00\x00\x00C\x00\x00\x00D\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00'
>>>

Internally, in NumPy 'U' is always UCS4, and the data type is a fixed
length string that has the length of the longest initializer element.

NumPy's use of 'U' seems vastly more useful for arrays than the behavior
of array.array:

>>> array.array('u', ['A', 'B'])
array('u', 'AB')
>>> array.array('u', ['ABC', 'D'])
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: array item must be unicode character

In Numpy, arrays of words are possible, with array.array they are not.

An additional thought: The convention in the struct module is to use
uppercase for unsigned types. So it would be a possibility to use
'C', 'U' and 'W', where '3C' would denote the same as '3s', except
for UCS1 instead of bytes.

Travis: can you please comment on what the intended semantics of the 'u' and 'w' specifiers is, in PEP-3118? More specifically:

• "an array/memoryview with format 'u' can support exactly one-character values (i.e. unicode strings of length 1)": true or false?
• "in a struct, an element of type 'u' will use up two bytes exactly (ignoring padding)": true or false?

I admit that the main thing that bothers me with the proposal in PEP-3118 is the inconsistency between c -> bytes, while u, w -> str

This was less of an issue in 2.x (which was the main frame of reference when the PEP was written), with implicit str/unicode interoperability, but seems quite jarring in the 3.x world.

Status quo:
struct module: 'c' = individual bytes, 's' = multi-byte sequence
array module: 'u' typecode may be either 2 bytes or 4 bytes (Py_UNICODE) (the addition of the 'w' typecode has been reverted)

My current inclination is still to apply Victor's patch from bpo-13072 (which changes array to export the appropriate integer typecodes for 'u' arrays) and otherwise punt on this for 3.3 and try to sort out the mess for 3.4.

For 3.4, I'm inclined to favour Stefan's proposal of C, U, W mapping to multi-point sequences of UCS-1, UCS-2, UCS-4 code points (with corresponding typecodes in the array module).

Support for lowercase 'u' would then never become an official part of the buffer API, existing only as an array typecode.

My current inclination is still to apply Victor's patch from bpo-13072
(which changes array to export the appropriate integer typecodes for
'u' arrays) and otherwise punt on this for 3.3 and try to sort out
the mess for 3.4.

I think this would be the worst choice. It would mean that we change
the format for exported array.arrays now for 3.3, and then change it
in 3.4 again. So anybody who cares about this would have to deal
with three different behaviors.

Note that the array module had been using 'u' and 'w' essentially
"forever" (i.e. since 3.0).

For 3.4, I'm inclined to favour Stefan's proposal of C, U, W mapping
to multi-point sequences of UCS-1, UCS-2, UCS-4 code points (with
corresponding typecodes in the array module).

Fine with me in principle, although I see a problem when NumPy uses
'U' for UCS-4, yet CPython declares it to be UCS-2. I also think that
Travis' explicit agreement must be sought.

I wouldn't change the export formats used for the 'u' typecode at all in 3.4 - I'd add new typecodes to array that match any new struct format characters and are exported accordingly. 'u' would *never* become a formally defined struct character, instead lingering in the array module as a legacy of the narrow/wide build distinction.

And good point that U would need to match UCS-4 to be consistent with NumPy. Perhaps we can just add 'U' in 3.4 and forget about UCS-2 entirely?

I wouldn't change the export formats used for the 'u' typecode at
all in 3.4 - I'd add new typecodes to array that match any new
struct format characters and are exported accordingly. 'u' would
*never* become a formally defined struct character, instead
lingering in the array module as a legacy of the narrow/wide build
distinction.

I think it is a desirable property that for an array A and an index
I, that A[I] == memoryview(A)[I]. Exporting the elements of an 'u'
array as integers would break that property.

So if we do want to support Unicode arrays (which some people apparently
want to see - I haven't heard anybody saying they actually *need* such
a type), the buffer type of it should be "unicode", in some form, not
"number".

I would be fine with deprecating the 'u' type arrays, acknowledging
that the Py_UNICODE element type is even more useless now than before.
If that is done, there is no point in fixing anything about it. If
it exports using the 'u' and 'w' codes - fine. If then memoryview
doesn't work properly - fine; this is a deprecated feature.

It should be fixed only if we want to support it "properly" (which I
believe this patch would do).

I guess the main alternative to deprecation that preserves the invariant you describe would be to propagate the "u == Py_UNICODE" definition to memoryview. Since we're trying to phase out Py_UNICODE, deprecation seems the more sensible course.

Perhaps just a documented deprecation for now, like the rest of the Py_UNICODE based APIs?

Martin v. Loewis <report@bugs.python.org> wrote:

I would be fine with deprecating the 'u' type arrays, acknowledging
that the Py_UNICODE element type is even more useless now than before.
If that is done, there is no point in fixing anything about it. If
it exports using the 'u' and 'w' codes - fine. If then memoryview
doesn't work properly - fine; this is a deprecated feature.

From the perspective of memoryview backwards compatibility, deprecation is fine.
In 3.2, memoryview could really only handle one-dimensional buffers of unsigned
bytes:

>>> import array
>>> a = array.array('u', "ABC")
>>> x = memoryview(a)
>>> a[0] == x[0]
False
>>> a[0]
'A'

# Indexing returns bytes instead of str:
>>> x[0]
b'A\x00'
>>> 

# Index assignment attempts to do slice assignment:
>>> x[0] = 'Z'
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: 'str' does not support the buffer interface
>>> 

I'm +1 for deprecating 'u' and 'w' in the array module, accept that memoryview
cannot handle 'u' and 'w' and fix the situation properly in 3.4. I agree that
the latter would require people to come up with actual use cases.

Well, apparently people do use 'u', see bpo-15035.

bpo-15035 indicates that there is a need for UCS-2 arrays, using 'u' arrays was technically incorrect, since it is based on Py_UNICODE, whereas the API in question uses UniChar (which apparently is a two-byte type).

Martin v. L??wis <report@bugs.python.org> wrote:

bpo-15035 indicates that there is a need for UCS-2 arrays, using 'u' arrays was technically incorrect, since it is based on Py_UNICODE, whereas the API in question uses UniChar (which apparently is a two-byte type).

Right, thanks for clearing that up. Then bpo-15035 would indeed support deprecating
'u' and 'w' and moving on to UCS2 and UCS4 arrays.

The documentation already specifies that 'u' is deprecated and doesn't mention the 'w' code. I think we can close this issue.

Closing sounds good to me

Trigger the same bug....

I want to effectively slice big unicode string. So I decide to use memoryview for that in order to eliminate memory copying.

In [33]: a = array.array('u', 'превед')
In [34]: m = memoryview(a)
In [35]: m[2:]
Out[35]: <memory at 0x7efc98fcc048>
In [36]: m[0]
...
NotImplementedError: memoryview: format w not supported

1. Why format 'w' error if I asked 'u' ?
2. Format 'w' is not listed in https://docs.python.org/3.5/library/array.html
3. What is alternative for fast slicing, like memoryview(bytearray(b'test')), but for unicode ?