python-peps/pep-0701.rst

PEP: 701
Title: Syntactic formalization of f-strings
Author: Pablo Galindo <pablogsal@python.org>,
        Batuhan Taskaya <batuhan@python.org>,
        Lysandros Nikolaou <lisandrosnik@gmail.com>
Discussions-To: https://discuss.python.org/t/pep-701-syntactic-formalization-of-f-strings/22046
Status: Draft
Type: Standards Track
Content-Type: text/x-rst
Created: 15-Nov-2022
Python-Version: 3.12


Abstract
========

This document proposes to lift some of the restrictions originally formulated in
:pep:`498` and to provide a formalized grammar for f-strings that can be
integrated into the parser directly. The proposed syntactic formalization of
f-strings will have some small side-effects on how f-strings are parsed and
interpreted, allowing for a considerable number of advantages for end users and
library developers, while also dramatically reducing the maintenance cost of
the code dedicated to parsing f-strings.


Motivation
==========

When f-strings were originally introduced in :pep:`498`, the specification was
provided without providing a formal grammar for f-strings. Additionally, the
specification contains several restrictions that are imposed so the parsing of
f-strings could be implemented into CPython without modifying the existing
lexer. These limitations have been recognized previously and previous attempts
have been made to lift them in :pep:`536`, but `none of this work was ever implemented`_.
Some of these limitations (collected originally by :pep:`536`) are:

#. It is impossible to use the quote character delimiting the f-string
   within the expression portion::

    >>> f'Magic wand: { bag['wand'] }'
                                 ^
    SyntaxError: invalid syntax

#. A previously considered way around it would lead to escape sequences
   in executed code and is prohibited in f-strings::

    >>> f'Magic wand { bag[\'wand\'] } string'
    SyntaxError: f-string expression portion cannot include a backslash

#. Comments are forbidden even in multi-line f-strings::

    >>> f'''A complex trick: {
    ... bag['bag']  # recursive bags!
    ... }'''
    SyntaxError: f-string expression part cannot include '#'

#. Arbitrary nesting of expressions without expansion of escape sequences is
   available in many other languages that employ a string interpolation
   method that uses expressions instead of just variable names. Some examples:

   .. code-block:: text

      # Ruby
      "#{ "#{1+2}" }"

      # JavaScript
      `${`${1+2}`}`

      # Swift
      "\("\(1+2)")"

      # C#
      $"{$"{1+2}"}"

These limitations serve no purpose from a language user perspective and
can be lifted by giving f-string literals a regular grammar without exceptions
and implementing it using dedicated parse code.

The other issue that f-strings have is that the current implementation in
CPython relies on tokenising f-strings as ``STRING`` tokens and a post processing of
these tokens. This has the following problems:

#. It adds a considerable maintenance cost to the CPython parser. This is because
   the parsing code needs to be written by hand, which has historically led to a
   considerable number of inconsistencies and bugs. Writing and maintaining parsing
   code by hand in C has always been considered error prone and dangerous as it needs
   to deal with a lot of manual memory management over the original lexer buffers.

#. The f-string parsing code is not able to use the new improved error message mechanisms
   that the new PEG parser, originally introduced in :pep:`617`, has allowed. The
   improvements that these error messages brought has been greatly celebrated but
   unfortunately f-strings cannot benefit from them because they are parsed in a
   separate piece of the parsing machinery. This is especially unfortunate, since
   there are several syntactical features of f-strings that can be confusing due
   to the different implicit tokenization that happens inside the expression
   part (for instance ``f"{y:=3}"`` is not an assignment expression).

#. Other Python implementations have no way to know if they have implemented
   f-strings correctly because contrary to other language features, they are not
   part of the :ref:`official Python grammar <f-strings>`.
   This is important because several prominent
   alternative implementations are using CPython's PEG parser, `such as PyPy`_,
   and/or are basing their grammars on the official PEG grammar. The
   fact that f-strings use a separate parser prevents these alternative implementations
   from leveraging the official grammar and benefiting from improvements in error messages derived
   from the grammar.


A version of this proposal was originally `discussed on Python-Dev`_  and
`presented at the Python Language Summit 2022`_ where it was enthusiastically
received.

Rationale
=========

By building on top of the new Python PEG Parser (:pep:`617`), this PEP proposes
to redefine “f-strings”, especially emphasizing the clear separation of the
string component and the expression (or replacement, ``{...}``) component. :pep:`498`
summarizes the syntactical part of “f-strings” as the following:

    In Python source code, an f-string is a literal string, prefixed with ‘f’, which
    contains expressions inside braces. The expressions are replaced with their values.

However, :pep:`498` also contained a formal list of exclusions on what
can or cannot be contained inside the expression component (primarily due to the
limitations of the existing parser). By clearly establishing the formal grammar, we
now also have the ability to define the expression component of an f-string as truly "any
applicable Python expression" (in that particular context) without being bound
by the limitations imposed by the details of our implementation.

The formalization effort and the premise above also has a significant benefit for
Python programmers due to its ability to simplify and eliminate the obscure
limitations. This reduces the mental burden and the cognitive complexity of
f-string literals (as well as the Python language in general).

#. The expression component can include any string literal that a normal Python expression
   can include. This opens up the possibility of nesting string literals (formatted or
   not) inside the expression component of an f-string with the same quote type (and length)::

    >>> f"These are the things: {", ".join(things)}"

    >>> f"{source.removesuffix(".py")}.c: $(srcdir)/{source}"

    >>> f"{f"{f"infinite"}"}" + " " + f"{f"nesting!!!"}"

   This "feature" is not universally agreed to be desirable, and some users find this unreadable.
   For a discussion on the different views on this, see the :ref:`701-considerations-of-quote-reuse` section.

#. Another issue that has felt unintuitive to most is the lack of support for backslashes
   within the expression component of an f-string. One example that keeps coming up is including
   a newline character in the expression part for joining containers. For example::

    >>> a = ["hello", "world"]
    >>> f"{'\n'.join(a)}"
    File "<stdin>", line 1
        f"{'\n'.join(a)}"
                        ^
    SyntaxError: f-string expression part cannot include a backslash

   A common work-around for this was to either assign the newline to an intermediate variable or
   pre-create the whole string prior to creating the f-string::

    >>> a = ["hello", "world"]
    >>> joined = '\n'.join(a)
    >>> f"{joined}"
    'hello\nworld'

   It only feels natural to allow backslashes in the expression part now that the new PEG parser
   can easily support it.

    >>> a = ["hello", "world"]
    >>> f"{'\n'.join(a)}"
    'hello\nworld'

#. Before the changes proposed in this document, there was no explicit limit in
   how f-strings can be nested, but the fact that string quotes cannot be reused
   inside the expression component of f-strings made it impossible to nest
   f-strings arbitrarily. In fact, this is the most nested-fstring that can be
   written::

    >>> f"""{f'''{f'{f"{1+1}"}'}'''}"""
    '2'

   As this PEP allows placing **any** valid Python expression inside the
   expression component of the f-strings, it is now possible to reuse quotes and
   therefore is possible to nest f-strings arbitrarily::

    >>> f"{f"{f"{f"{f"{f"{1+1}"}"}"}"}"}"
    '2'
   
   Although this is just a consequence of allowing arbitrary expressions, the
   authors of this PEP do not believe that this is a fundamental benefit and we
   have decided that the language specification will not explicitly mandate that
   this nesting can be arbitrary. This is because allowing arbitrarily-deep
   nesting imposes a lot of extra complexity to the lexer implementation
   (particularly as lexer/parser pipelines need to allow "untokenizing" to
   support the 'f-string debugging expressions' and this is especially taxing when
   arbitrary nesting is allowed). Implementations are therefore free to impose a
   limit on the nesting depth if they need to. Note that this is not an uncommon
   situation, as the CPython implementation already imposes several limits all
   over the place, including a limit on the nesting depth of parentheses and
   brackets, a limit on the nesting of the blocks, a limit in the number of
   branches in ``if`` statements, a limit on the number of expressions in
   star-unpacking, etc.

Specification
=============

The formal proposed PEG grammar specification for f-strings is (see :pep:`617`
for details on the syntax):

.. code-block:: peg

    fstring
        | FSTRING_START fstring_middle* FSTRING_END
    fstring_middle
        | fstring_replacement_field
        | FSTRING_MIDDLE
    fstring_replacement_field
        | '{' (yield_expr | star_expressions) "="? [ "!" NAME ] [ ':' fstring_format_spec* ] '}'
    fstring_format_spec:
        | FSTRING_MIDDLE
        | fstring_replacement_field

The new tokens (``FSTRING_START``, ``FSTRING_MIDDLE``, ``FSTRING_END``) are defined
:ref:`later in this document <701-new-tokens>`.

This PEP leaves up to the implementation the level of f-string nesting allowed.
This means that limiting nesting is **not part of the language specification**
but also the language specification **doesn't mandate arbitrary nesting**. 

The new grammar will preserve the Abstract Syntax Tree (AST) of the current
implementation. This means that no semantic changes will be introduced by this
PEP on existing code that uses f-strings.

Handling of f-string debug expressions
--------------------------------------

Since Python 3.8, f-strings can be used to debug expressions by using the
``=`` operator. For example::

    >>> a = 1
    >>> f"{1+1=}"
    '1+1=2'

This semantics were not introduced formally in a PEP and they were implemented
in the current string parser as a special case in `bpo-36817
<https://bugs.python.org/issue?@action=redirect&bpo=36817>`_ and documented in
`the f-string lexical analysis section
<https://docs.python.org/3/reference/lexical_analysis.html#f-strings>`_.

This feature is not affected by the changes proposed in this PEP but is
important to specify that the formal handling of this feature requires the lexer
to be able to "untokenize" the expression part of the f-string. This is not a
problem for the current string parser as it can operate directly on the string
token contents. However, incorporating this feature into a given parser
implementation requires the lexer to keep track of the raw string contents of
the expression part of the f-string and make them available to the parser when
the parse tree is constructed for f-string nodes. A pure "untokenization" is not
enough because as specified currently, f-string debug expressions preserve whitespace in the expression,
including spaces after the ``{`` and the ``=`` characters. This means that the
raw string contents of the expression part of the f-string must be kept intact
and not just the associated tokens.

How parser/lexer implementations deal with this problem is of course up to the
implementation.

.. _701-new-tokens:

New tokens
----------

Three new tokens are introduced: ``FSTRING_START``, ``FSTRING_MIDDLE`` and
``FSTRING_END``. Different lexers may have different implementations that may be
more efficient than the ones proposed here given the context of the particular
implementation. However, the following definitions will be used as part of the
public APIs of CPython (such as the ``tokenize`` module) and are also provided
as a reference so that the reader can have a better understanding of the
proposed grammar changes and how the tokens are used:

* ``FSTRING_START``: This token includes the f-string prefix (``f``/``F``/``fr``) and the opening quote(s).
* ``FSTRING_MIDDLE``: This token includes a portion of text inside the string that's not part of the
  expression part and isn't an opening or closing brace. This can include the text between the opening quote
  and the first expression brace (``{``), the text between two expression braces (``}`` and ``{``) and the text
  between the last expression brace (``}``) and the closing quote.
* ``FSTRING_END``: This token includes the closing quote.

These tokens are always string parts and they are semantically equivalent to the
``STRING`` token with the restrictions specified. These tokens must be produced by the lexer
when lexing f-strings.  This means that **the tokenizer cannot produce a single token for f-strings anymore**.
How the lexer emits this token is **not specified** as this will heavily depend on every
implementation (even the Python version of the lexer in the standard library is implemented
differently to the one used by the PEG parser).

As an example::

    f'some words {a+b:.3f} more words {c+d=} final words'

will be tokenized as::

    FSTRING_START - "f'"
    FSTRING_MIDDLE - 'some words '
    LBRACE - '{'
    NAME - 'a'
    PLUS - '+'
    NAME - 'b'
    OP - ':'
    FSTRING_MIDDLE - '.3f'
    RBRACE - '}'
    FSTRING_MIDDLE - ' more words '
    LBRACE - '{'
    NAME - 'c'
    PLUS - '+'
    NAME - 'd'
    OP - '='
    RBRACE - '}'
    FSTRING_MIDDLE - ' final words'
    FSTRING_END - "'"

while ``f"""some words"""`` will be tokenized simply as::

    FSTRING_START - 'f"""'
    FSTRING_MIDDLE - 'some words'
    FSTRING_END - '"""'

.. _701-tokenize-changes:

Changes to the tokenize module
------------------------------

The :mod:`tokenize` module will be adapted to emit these tokens as described in the previous section
when parsing f-strings so tools can take advantage of this new tokenization schema and avoid having
to implement their own f-string tokenizer and parser.

How to produce these new tokens
-------------------------------

One way existing lexers can be adapted to emit these tokens is to incorporate a
stack of "lexer modes" or to use a stack of different lexers. This is because
the lexer needs to switch from "regular Python lexing" to "f-string lexing" when
it encounters an f-string start token and as f-strings can be nested, the
context needs to be preserved until the f-string closes. Also, the "lexer mode"
inside an f-string expression part needs to behave as a "super-set" of the
regular Python lexer (as it needs to be able to switch back to f-string lexing
when it encounters the ``}`` terminator for the expression part as well as
handling f-string formatting and debug expressions). For reference, here is a
draft of the algorithm to modify a CPython-like tokenizer to emit these new
tokens:

1. If the lexer detects that an f-string is starting (by detecting the letter
   'f/F' and one of the possible quotes) keep advancing until a valid quote is
   detected (one of ``"``, ``"""``, ``'`` or ``'''``) and emit a
   ``FSTRING_START`` token with the contents captured (the 'f/F' and the
   starting quote). Push a new tokenizer mode to the tokenizer mode stack for
   "F-string tokenization". Go to step 2.
2. Keep consuming tokens until a one of the following is encountered:

   * A closing quote equal to the opening quote.
   * An opening brace (``{``) or a closing brace (``}``) that is not immediately
     followed by another opening/closing brace.

   In all cases, if the character buffer is not empty, emit a ``FSTRING_MIDDLE``
   token with the contents captured so far but transform any double
   opening/closing braces into single opening/closing braces.  Now, proceed as
   follows depending on the character encountered:

   * If a closing quote matching the opening quite is encountered go to step 4.
   * If an opening bracket (not immediately followed by another opening bracket)
     is encountered, go to step 3.
   * If a closing bracket (not immediately followed by another closing bracket)
     is encountered, emit a token for the closing bracket and go to step 2.

3. Push a new tokenizer mode to the tokenizer mode stack for "Regular Python
   tokenization withing f-string" and proceed to tokenize with it. This mode
   tokenizes as the "Regular Python tokenization" until a ``!``, ``:``, ``=``
   character is encountered or if a ``}`` character is encountered with the same
   level of nesting as the opening bracket token that was pushed when we enter the
   f-string part. Using this mode, emit tokens until one of the stop points are
   reached. When this happens, emit the corresponding token for the stopping
   character encountered and, pop the current tokenizer mode from the tokenizer mode
   stack and go to step 2.
4. Emit a ``FSTRING_END`` token with the contents captured and pop the current
   tokenizer mode (corresponding to "F-string tokenization") and go back to
   "Regular Python mode".

Of course, as mentioned before, it is not possible to provide a precise
specification of how this should be done for an arbitrary tokenizer as it will
depend on the specific implementation and nature of the lexer to be changed.

Consequences of the new grammar
-------------------------------

All restrictions mentioned in the PEP are lifted from f-string literals, as explained below:

* Expression portions may now contain strings delimited with the same kind of
  quote that is used to delimit the f-string literal.
* Backslashes may now appear within expressions just like anywhere else in
  Python code. In case of strings nested within f-string literals, escape sequences are
  expanded when the innermost string is evaluated.
* New lines are now allowed within expression brackets. This means that these are now allowed::

    >>> x = 1
    >>> f"___{
    ...     x
    ... }___"
    '___1___'

    >>> f"___{(
    ...     x
    ... )}___"
    '___1___'

* Comments, using the ``#`` character, are allowed within the expression part of an f-string.
  Note that comments require that the closing bracket (``}``) of the expression part to be present in
  a different line as the one the comment is in or otherwise it will be ignored as part of the comment.

.. _701-considerations-of-quote-reuse:

Considerations regarding quote reuse
------------------------------------

One of the consequences of the grammar proposed here is that, as mentioned above,
f-string expressions can now contain strings delimited with the same kind of quote
that is used to delimit the external f-string literal. For example:

    >>> f" something { my_dict["key"] } something else "

In the `discussion thread for this PEP <https://discuss.python.org/t/pep-701-syntactic-formalization-of-f-strings/22046>`_,
several concerns have been raised regarding this aspect and we want to collect them here,
as these should be taken into consideration when accepting or rejecting this PEP.

Some of these objections include:

* Many people find quote reuse withing the same string confusing and hard to read. This is because
  allowing quote reuse will violate a current property of Python as it stands today: the fact that
  strings are fully delimited by two consecutive pairs of the same kind of quote, which by itself is a very simple rule.
  One of the reasons quote reuse may be harder for humans to parse, leading to less readable
  code, is that the quote character is the same for both start and
  end (as opposed to other delimiters). 

* Some users have raised concerns that quote reuse may break some lexer and syntax highlighting tools that rely
  on simple mechanisms to detect strings and f-strings, such as regular expressions or simple delimiter
  matching tools. Introducing quote reuse in f-strings will either make it trickier to keep these tools
  working or will break the tools altogether (as, for instance, regular expressions cannot parse arbitrary nested
  structures with delimiters). The IDLE editor, included in the standard library, is an example of a
  tool which may need some work to correctly apply syntax highlighting to f-strings.

Here are some of the arguments in favour:

* Many languages that allow similar syntactic constructs (normally called "string interpolation") allow quote
  reuse and arbitrary nesting. These languages include JavaScript, Ruby, C#, Bash, Swift and many others.
  The fact that many languages allow quote reuse can be a compelling argument in favour of allowing it in Python. This
  is because it will make the language more familiar to users coming from other languages. 

* As many other popular languages allow quote reuse in string interpolation constructs, this means that editors
  that support syntax highlighting for these languages will already have the necessary tools to support syntax
  highlighting for f-strings with quote reuse in Python. This means that although the files that handle syntax
  highlighting for Python will need to be updated to support this new feature, is not expected to be impossible
  or very hard to do.

* One advantage of allowing quote reuse is that it composes cleanly with other syntax. Sometimes this is referred to
  as "referential transparency". An example of this is that if we have ``f(x+1)``, assuming ``a`` is a brand new variable, it
  should behave the same as ``a = x+1; f(a)``. And vice versa.  So if we have::

    def py2c(source):
        prefix = source.removesuffix(".py")
        return f"{prefix}.c"

  It should be expected that if we replace the variable ``prefix`` with its definition, the answer should be the same::

    def py2c(source):
        return f"{source.removesuffix(".py")}.c"

* Code generators (like `ast.unparse <https://docs.python.org/3/library/ast.html#ast.unparse>`_ from standard library) in their
  current form rely on complicated algorithms to ensure expressions within an f-string are properly suited for the context in
  which they are being used. These non-trivial algorithms come with challanges such as finding an unused quote type (by tracking
  the outer quotes), and generating string representations which would not include backslashes if possible. Allowing quote reuse
  and backslashes would simplify the code generators which deals with f-strings considerably, as the regular Python expression logic
  can be used inside and outside of f-strings without any special treatment.

* Limiting quote reuse will considerably increase the complexity of the implementation of the proposed changes. This is because
  it will force the parser to have the context that is parsing an expression part of an f-string with a given quote in order
  to know if it needs to reject an expression that reuses the quote. Carrying this context around is not trivial in parsers that
  can backtrack arbitrarily (such as the PEG parser). The issue becomes even more complex if we consider that f-strings can be
  arbitrarily nested and therefore several quote types may need to be rejected.

  To gather feedback from the community,
  `a poll <https://discuss.python.org/t/pep-701-syntactic-formalization-of-f-strings/22046/24>`__
  has been initiated to get a sense of how the community feels about this aspect of the PEP.

Backwards Compatibility
=======================

This PEP does not introduce any backwards incompatible syntactic or semantic changes
to the Python language. However, the :mod:`tokenize` module (a quasi-public part of the standard
library) will need to be updated to support the new f-string tokens (to allow tool authors
to correctly tokenize f-strings). See :ref:`701-tokenize-changes` for more details regarding
how the public API of ``tokenize`` will be affected.

How to Teach This
=================

As the concept of f-strings is already ubiquitous in the Python community, there is
no fundamental need for users to learn anything new. However, as the formalized grammar
allows some new possibilities, it is important that the formal grammar is added to the
documentation and explained in detail, explicitly mentioning what constructs are possible
since this PEP is aiming to avoid confusion.

It is also beneficial to provide users with a simple framework for understanding what can
be placed inside an f-string expression. In this case the authors think that this work will
make it even simpler to explain this aspect of the language, since it can be summarized as:

    You can place any valid Python expression inside an f-string expression.

With the changes in this PEP, there is no need to clarify that string quotes are
limited to be different from the quotes of the enclosing string, because this is
now allowed: as an arbitrary Python string can contain any possible choice of
quotes, so can any f-string expression. Additionally there is no need to clarify
that certain things are not allowed in the expression part because of
implementation restrictions such as comments, new line characters or
backslashes. 

The only "surprising" difference is that as f-strings allow specifying a
format, expressions that allow a ``:`` character at the top level still need to be
enclosed in parenthesis. This is not new to this work, but it is important to
emphasize that this restriction is still in place. This allows for an easier
modification of the summary:

    You can place any valid Python expression inside
    an f-string expression, and everything after a ``:`` character at the top level will
    be identified as a format specification.


Reference Implementation
========================

A reference implementation can be found in the implementation_ fork.

Rejected Ideas
==============

#. Although we think the readability arguments that have been raised against
   allowing quote reuse in f-string expressions are valid and very important,
   we have decided to propose not rejecting quote reuse in f-strings at the parser
   level. The reason is that one of the cornerstones of this PEP is to reduce the
   complexity and maintenance of parsing f-strings in CPython and this will not
   only work against that goal, but it may even make the implementation even more
   complex than the current one. We believe that forbidding quote reuse should be
   done in linters and code style tools and not in the parser, the same way other
   confusing or hard-to-read constructs in the language are handled today.

#. We have decided not to lift the restriction that some expression portions
   need to wrap ``':'`` and ``'!'`` in parentheses at the top level, e.g.::

    >>> f'Useless use of lambdas: { lambda x: x*2 }'
    SyntaxError: unexpected EOF while parsing
   
   The reason is that this would this will introduce a considerable amount of
   complexity for no real benefit. This is due to the fact that the ``:`` character
   normally separates the f-string format specification. This format specification
   is currently tokenized as a string. As the tokenizer MUST tokenize what's on the
   right of the ``:`` as either a string or a stream of tokens, this won't allow the
   parser to differentiate between the different semantics as that would require the
   tokenizer to backtrack and produce a different set of tokens (this is, first try
   as a stream of tokens, and if it fails, try as a string for a format specifier).

   As there is no fundamental advantage in being able to allow lambdas and similar
   expressions at the top level, we have decided to keep the restriction that these must
   be parenthesized if needed::

    >>> f'Useless use of lambdas: { (lambda x: x*2) }'

#. We have decided to disallow (for the time being) using escaped braces (``\{`` and ``\}``)
   in addition to the ``{{`` and ``}}`` syntax. Although the authors of the PEP believe that
   allowing escaped braces is a good idea, we have decided to not include it in this PEP, as it is not strictly
   necessary for the formalization of f-strings proposed here, and it can be
   added independently in a regular CPython issue.

Open Issues
===========

None yet


Footnotes
=========


.. _official Python grammar: https://docs.python.org/3/reference/lexical_analysis.html#formatted-string-literals

.. _none of this work was ever implemented: https://mail.python.org/archives/list/python-dev@python.org/thread/N43O4KNLZW4U7YZC4NVPCETZIVRDUVU2/#NM2A37THVIXXEYR4J5ZPTNLXGGUNFRLZ

.. _such as PyPy: https://foss.heptapod.net/pypy/pypy/-/commit/fe120f89bf07e64a41de62b224e4a3d80e0fe0d4/pipelines?ref=branch%2Fpy3.9

.. _discussed on Python-Dev: https://mail.python.org/archives/list/python-dev@python.org/thread/54N3MOYVBDSJQZTU6MTCPLUPIFSDN5IS/#SAYU6SMP4KT7G7AQ6WVQYUDOSZPKHJMS

.. _presented at the Python Language Summit 2022: https://pyfound.blogspot.com/2022/05/the-2022-python-language-summit-f.html

.. _implementation: https://github.com/we-like-parsers/cpython/tree/fstring-grammar


Copyright
=========

This document is placed in the public domain or under the
CC0-1.0-Universal license, whichever is more permissive.