>>> 'xyz'.center(20, '\U00100140')
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: The fill character must be exactly one character long

str.ljust and str.rjust are similarly affected.

The question is, what should it do with such an input? Pretend it's a single char (but other chars in the source string won't get the same treatment)? Treat it as a two-char string (but then center() and friends should logically be extended to accept strings of arbitrary lengths)?

str.__format__ and friends (int, float, complex) also have this same problem. For example, when they're computing the "fill" character:

>>> format('', 'x^')
''

>>> format('', '\U00100140^')
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ValueError: Invalid conversion specification

On Wed, Nov 24, 2010 at 10:33 AM, Antoine Pitrou <report@bugs.python.org> wrote:
..
> The question is, what should it do with such an input?

I think the rule for such functions should be that if
input.encode('utf-8') is the same on wide and narrow builds, then the
output.encode('utf-8') should be the same.

> Pretend it's a single char (but other chars in the source string won't get the same treatment)?

Yes, *and* surrogate pairs in the source string should count for one
char as well.

> Treat it as a two-char string (but then center() and friends should logically be
> extended to accept strings of arbitrary lengths)?

No.  For better or worse, on wide builds these methods effectively
operate on code points.  They don't interpret multi-code-point-
graphemes or take grapheme width into account:

--------------------
​​​​​​​​​​​​​​​​​123
--------------------

Application code has to ascertain that it is dealing with with fixed
width characters in the target font before using these methods for
text alignment.

Alexander Belopolsky wrote:
> 
> New submission from Alexander Belopolsky <belopolsky@users.sourceforge.net>:
> 
>>>> 'xyz'.center(20, '\U00100140')
> Traceback (most recent call last):
>   File "<stdin>", line 1, in <module>
> TypeError: The fill character must be exactly one character long
> 
> str.ljust and str.rjust are similarly affected.

I don't think we should change that for the formatting methods.

See my reply on python-dev:

str.center(n) centers the string in a padded string that
is composed of n code units. Whether that operation will result
in a text that's centered visually on output is a completely
different story. The original string could contain surrogates,
it could also contain combing code points, so the visual
presentation of the result may very well not be centered at
all; it may not even appear as having the length n to the user.

Since we're not going change the semantics of those APIs,
it is OK to not support padding with non-BMP code points on
UCS-2 builds.

Supporting such cases would only cause problems:

* if the methods would pad with surrogates, the resulting
  string would no longer have length n; breaking the
  assumption that len(str.center(n)) == n

* if the methods would pad with half the number of surroagtes
  to make sure that len(str.center(n)) == n, the resulting
  output to e.g. a terminal would be further off, than what
  you already have with surrogates and combining code points
  in the original string.

On Wed, Nov 24, 2010 at 3:37 PM, Marc-Andre Lemburg
<report@bugs.python.org> wrote:
..
> I don't think we should change that for the formatting methods.

That's a reasonable position.  What about

>>> unicodedata.category('\N{OLD ITALIC LETTER A}')
'Lo'
>>> '\N{OLD ITALIC LETTER A}'.isalpha()
False

the str.isalpha() method is underspecified in the reference manual,
but a comment in unicodectype.c describes Py_UNICODE_ISALPHA as
follows:

/* Returns 1 for Unicode characters having the category 'Ll', 'Lu',
'Lt',
  'Lo' or 'Lm',  0 otherwise. */

I don't have a wide build handy, but I am fairly sure  '\N{OLD ITALIC
LETTER A}'.isalpha() would produce True there.  The result above is
simply consequence of surrogates considered to be non-letters:

>>> [c.isalpha() for c in '\N{OLD ITALIC LETTER A}']
[False, False]

Here is another str method not ready for non-BMP chars:


>>> u = '\U00010140'
>>> u.translate({ord(u):ord('A')})
'𐅀'

(expected 'A')

>>> u = 'B'
>>> u.translate({ord(u):ord('A')})
'A'

I think that methods like str.isalpha can and should be fixed. Since _PyUnicode_IsAlpha now accepts a Py_UCS4, the body of unicode_isalpha can be changed to convert normal chars and surrogates pairs to a Py_UCS4 before calling Py_UNICODE_ISALPHA.
The attached patch is a proof of concept of this approach and returns True for '\N{OLD ITALIC LETTER A}'.isalpha() on a narrow build.
It still has a number of issues that should be addressed (check for narrow builds, check for lone surrogates, check for high surrogate at the end of a string, fix compiler warnings ...) but it should be good enough as a PoC.

I would also suggest to introduce a set of macros to handle surrogates (e.g. detect, combine) and use it in all the functions that need to work with them.

Here is another proof of concept patch for the isalpha issue that introduces a higher level abstraction macro - Py_UNICODE_NEXT.  It should be possible to reuse this macro in all isxyz methods and other places where surrogates are currently processed.  I should be possible to come up with a pure macro definition of Py_UNICODE_NEXT.

issue9200 already proposes a similar change to str.is* methods.

As a practical matter, I think that for at least the next decade, people are at least as likely to want to fill with a composed, multi-BMP-codepoint 'char' (grapheme) as with a non-BMP char. So to me, failure with the latter is no worse than failure with the former.

The underlying problem is that centering k chars within n spaces with fill i is based on one-char per code encodings *and* fixed pitch fonts with one-char per space. That model is not universally applicable, so I do not consider it a bug that functions based on that model are also not universally applicable. Perhaps docs should be clearer about the limitations of many of the string methods in the new context.

A full general solution to the general problem of centering requires a shift to physical units (points or mm) and detailed font information, including kerning. This is beyond the scope of a string method.

So I consider this a feature request for a partial generalization of unclear utility and unclear definition.

On Fri, Nov 26, 2010 at 6:37 PM, Terry J. Reedy <report@bugs.python.org> wrote:
>
> Terry J. Reedy <tjreedy@udel.edu> added the comment:
>
> As a practical matter, I think that for at least the next decade, people are at least as likely to
> want to fill with a composed, multi-BMP-codepoint 'char' (grapheme) as with a non-BMP char.
> So to me, failure with the latter is no worse than failure with the former.
>

I disagree. '\N{AEGEAN WORD SEPARATOR DOT}'  ('𐄁') looks like a
reasonably shaped fill character, while say 'Z\N{COMBINING ACUTE
ACCENT}\N{COMBINING GRAVE ACCENT}' ('Ź̀') does not.  Yet this is not
the point of this bug report.  The point is that Python user should
not care (much) about how many bytes per character Python uses under
the hood or what is the numeric value of the character that she can
enter in her program.

> The underlying problem is that centering k chars within n spaces with fill i is based
> on one-char per code encodings *and* fixed pitch fonts with one-char per space.

No. ' Section Title '.center(40, '*') will look good regardless of
font width and even more so when combined with <center> tag or its
equivalent in a given application.

I think these macros would be a reasonable approach. I think str.center, etc. should support non-BMP chars, because to not do so can raise an exception. Supporting composed graphemes seems like another problem altogether. And while we could fix that, it's clearly a larger step.

I agree that s.center(char, n).encode('utf-8') should be the same on both the builds -- even if their len() will be different -- for the following reasons:

1) the string will eventually be encoded, and if they the result is the same on both builds, it will look the same too;
2) trying to keep the same len() will generate different results and it won't work in case of odd width like 'foo'.center(surrogate_pair, 5) because you can't put half surrogate.

After reading the additional messages here and on a similar issue Alexander opened after this, I seem the point of wanting to make the difference between the two types of builds as transparent as sensibly possible. From that viewpoint, rejection of composed chars is not as bad because both types of builds act the same.

This issue has been fixed in Python 3.3 thanks to the PEP 393.

It can still be fixed on 2.7/3.2 though.

I'm just going to close this and say "use 3.3".