Format attribute name with repr() when formatting AttributeError error message

[vstinner]

Feature or enhancement

Some code paths format the attribute name with str() to format the AttributeError error message. Example:

>>> getattr(object(), "a'b")
AttributeError: 'object' object has no attribute 'a'b'

The quote character ' should be escaped in the error message ('a\'b'), or the attribute name should be quoted with " character ("a'b").

The expected error is:

AttributeError: 'object' object has no attribute "a'b"

I propose replacing str() with repr() to format the attribute name.

This issue is a follow-up of #153786 (comment) discussion (another AttributeError change).

Linked PRs

• gh-154610: Use repr() to format the attribute name in AttributeError #154614

[serhiy-storchaka]

How does it work with StrEnum?

[vstinner]

How does it work with StrEnum?

My PR changes how StrEnum instances and str subclasses are formatted.

Currently, PyErr_Format(PyExc_AttributeError, "...%U", ..., name) calls PyUnicode_FromFormatV() which calls _PyUnicodeWriter_WriteStr() to render %U. The _PyUnicodeWriter_WriteStr(writer, obj) call expects a str instance and writes str.__str__(obj) even if the __str__() method is overridden in a str subclass.

---

StrEnum example:

import enum

class Build(enum.StrEnum):
    DEBUG = enum.auto()
    OPTIMIZED = enum.auto()

getattr(object(), Build.DEBUG)

Before (Python 3.15):

AttributeError: 'object' object has no attribute 'debug'

After:

AttributeError: 'object' object has no attribute <Build.DEBUG: 'debug'>

---

Example of a str subclass:

class MyStr(str):
    def __repr__(self):
        return '<custom repr>'

getattr(object(), MyStr('abc'))

Before (Python 3.15):

AttributeError: 'object' object has no attribute 'abc'

After:

AttributeError: 'object' object has no attribute <custom repr>

[serhiy-storchaka]

__str__() should not be called for %U. We have %S for this.

[serhiy-storchaka]

My concerns:

• Such errors are often used in EAFP code. Error message is generated, but not used. The faster it executes and the less third-party code that can fail is executed, the better.
• Such errors can happen in stressful cases, for example in process of the interpreter shutdown or when it is failing from memory starvation. The less third-party code that can fail is executed, the better.

%U rarely fails (especially if we forbid execution of __str__()). %R is more complex, it usually involves several memory allocations.

[vstinner]

Sometimes, I wish that we could more easily extend PyUnicode_FromFormat() formats. For example, it would be nice to have a format similar to %U but which would quote the string: similar to repr(str) but don't call __repr__() method on str subclasses. Adding a suffix such as %Uq format is not possible, since currently, it's parsed as %U format + q literal string. There is %#U where # usually means "alternative format". But I'm not sure if # relationship to "repr/quote" is obvious enough.

Note: PEP 737 added %T, %#T, %N and %#N where # only changes the separator: use colon (:) instead of dot (.).

[serhiy-storchaka]

I thought about adding variants of %c and %s which escape non-printable characters.

[serhiy-storchaka]

Implemented in #154669.