PEP 653 -- Precise Semantics for Pattern Matching
PEP: | 653 |
---|---|
Title: | Precise Semantics for Pattern Matching |
Author: | Mark Shannon <mark at hotpy.org> |
Status: | Draft |
Type: | Standards Track |
Created: | 9-Feb-2021 |
Post-History: | 18-Feb-2021 |
Contents
Abstract
This PEP proposes a semantics for pattern matching that respects the general concept of PEP 634, but is more precise, easier to reason about, and should be faster.
The object model will be extended with three special (dunder) attributes to support pattern matching:
- A __match_kind__ attribute. Must be an integer.
- An __attributes__ attribute. Only needed for those classes wanting to customize matching the class pattern. If present, it must be a tuple of strings.
- A __deconstruct__() method. Only needed if __attributes__ is present. Returns an iterable over the components of the deconstructed object.
With this PEP:
- The semantics of pattern matching will be clearer, so that patterns are easier to reason about.
- It will be possible to implement pattern matching in a more efficient fashion.
- Pattern matching will be more usable for complex classes, by allowing classes more control over which patterns they match.
Motivation
Pattern matching in Python, as described in PEP 634, is to be added to Python 3.10. Unfortunately, PEP 634 is not as precise about the semantics as it could be, nor does it allow classes sufficient control over how they match patterns.
Precise semantics
PEP 634 explicitly includes a section on undefined behavior. Large amounts of undefined behavior may be acceptable in a language like C, but in Python it should be kept to a minimum. Pattern matching in Python can be defined more precisely without loosing expressiveness or performance.
Improved control over class matching
PEP 634 assumes that class instances are simply a collection of their attributes, and that deconstruction by attribute access is the dual of construction. That is not true, as many classes have a more complex relation between their constructor and internal attributes. Those classes need to be able to define their own deconstruction.
For example, using sympy, we might want to write:
# sin(x)**2 + cos(x)**2 == 1 case Add(Pow(sin(a), 2), Pow(cos(b), 2)) if a == b: return 1
For sympy to support this pattern for PEP 634 would be possible, but tricky and cumbersome. With this PEP it can be implemented easily [1].
PEP 634 also privileges some builtin classes with a special form of matching, the "self" match. For example the pattern list(x) matches a list and assigns the list to x. By allowing classes to choose which kinds of pattern they match, other classes can use this form as well.
Robustness
With this PEP, access to attributes during pattern matching becomes well defined and deterministic. This makes pattern matching less error prone when matching objects with hidden side effects, such as object-relational mappers. Objects will have control over their own deconstruction, which can help prevent unintended consequences should attribute access have side-effects.
PEP 634 relies on the collections.abc module when determining which patterns a value can match, implicitly importing it if necessary. This PEP will eliminate surprising import errors and misleading audit events from those imports.
Efficient implementation
The semantics proposed in this PEP will allow efficient implementation, partly as a result of having precise semantics and partly from using the object model.
With precise semantics, it is possible to reason about what code transformations are correct, and thus apply optimizations effectively.
Because the object model is a core part of Python, implementations already handle special attribute lookup efficiently. Looking up a special attribute is much faster than performing a subclass test on an abstract base class.
Rationale
The object model and special methods are at the core of the Python language. Consequently, implementations support them well. Using special attributes for pattern matching allows pattern matching to be implemented in a way that integrates well with the rest of the implementation, and is thus easier to maintain and is likely to perform better.
A match statement performs a sequence of pattern matches. In general, matching a pattern has three parts:
- Can the value match this kind of pattern?
- When deconstructed, does the value match this particular pattern?
- Is the guard true?
To determine whether a value can match a particular kind of pattern, we add the __match_kind__ attribute. This allows the kind of a value to be determined once and in a efficient fashion.
To deconstruct an object, pre-existing special methods can be used for sequence and mapping patterns, but something new is needed for class patterns. PEP 634 proposes using ad-hoc attribute access, disregarding the possibility of side-effects. This could be problematic should the attributes of the object be dynamically created or consume resources. By adding the __attributes__ attribute and __deconstruct__() method, objects can control how they are deconstructed, and patterns with a different set of attributes can be efficiently rejected. Should deconstruction of an object make no sense, then classes can define __match_kind__ to reject class patterns completely.
Specification
Additions to the object model
A __match_kind__ attribute will be added to object. It should be overridden by classes that want to match mapping or sequence patterns, or want change the default behavior when matching class patterns. It must be an integer and should be exactly one of these:
0 MATCH_SEQUENCE MATCH_MAPPING
bitwise ored with exactly one of these:
0 MATCH_DEFAULT MATCH_CLASS MATCH_SELF
Note
It does not matter what the actual values are. We will refer to them by name only. Symbolic constants will be provided both for Python and C, and once defined they will never be changed.
Classes inheriting from object will inherit __match_kind__ = MATCH_DEFAULT.
Classes which define __match_kind__ & MATCH_CLASS to be non-zero must implement one additional special attribute, and one special method:
- __attributes__: should hold a tuple of strings indicating the names of attributes that are to be considered for matching; it may be empty for postional-only matches.
- __deconstruct__(): should return a sequence which contains the parts of the deconstructed object.
Note
__attributes__ and __deconstruct__ will be automatically generated for dataclasses and named tuples.
The pattern matching implementation is not required to check that __attributes__ and __deconstruct__ behave as specified. If the value of __attributes__ or the result of __deconstruct__() is not as specified, then the implementation may raise any exception, or match the wrong pattern. Of course, implementations are free to check these properties and provide meaningful error messages if they can do so efficiently.
Semantics of the matching process
In the following, all variables of the form $var are temporary variables and are not visible to the Python program. They may be visible via introspection, but that is an implementation detail and should not be relied on. The psuedo-statement DONE is used to signify that matching is complete and that following patterns should be ignored. All the translations below include guards. If no guard is present, simply substitute the guard if True when translating.
Variables of the form $ALL_CAPS are meta-variables holding a syntactic element, they are not normal variables. So, $VARS = $items is not an assignment of $items to $VARS, but an unpacking of $items into the variables that $VARS holds. For example, with the abstract syntax case [$VARS]:, and the concrete syntax case[a, b]: then $VARS would hold the variables (a, b), not the values of those variables.
The psuedo-function QUOTE takes a variable and returns the name of that variable. For example, if the meta-variable $VAR held the variable foo then QUOTE($VAR) == "foo".
All additional code listed below that is not present in the original source will not trigger line events, conforming to PEP 626.
Preamble
Before any patterns are matched, the expression being matched is evaluated and its kind is determined:
match expr:
translates to:
$value = expr $kind = type($value).__match_kind__
In addition some helper variables are initialized:
$list = None $dict = None $attrs = None $items = None
Capture patterns
Capture patterns always match, so:
case capture_var if guard:
translates to:
capture_var = $value if guard: DONE
Wildcard patterns
Wildcard patterns always match, so:
case _ if guard:
translates to:
if guard: DONE
Literal Patterns
The literal pattern:
case LITERAL if guard:
translates to:
if $value == LITERAL and guard: DONE
except when the literal is one of None, True or False , when it translates to:
if $value is LITERAL and guard: DONE
Value Patterns
The value pattern:
case value.pattern if guard:
translates to:
if $value == value.pattern and guard: DONE
Sequence Patterns
Before matching the first sequence pattern, but after checking that $value is a sequence, $value is converted to a list.
A pattern not including a star pattern:
case [$VARS] if guard:
translates to:
if $kind & MATCH_SEQUENCE: if $list is None: $list = list($value) if len($list) == len($VARS): $VARS = $list if guard: DONE
Example: [2]
A pattern including a star pattern:
case [$VARS] if guard
translates to:
if $kind & MATCH_SEQUENCE: if $list is None: $list = list($value) if len($list) >= len($VARS): $VARS = $list # Note that $VARS includes a star expression. if guard: DONE
Example: [3]
Mapping Patterns
Before matching the first mapping pattern, but after checking that $value is a mapping, $value is converted to a dict.
A pattern not including a double-star pattern:
case {$KEYWORD_PATTERNS} if guard:
translates to:
if $kind & MATCH_MAPPING: if $dict is None: $dict = dict($value) if $dict.keys() == $KEYWORD_PATTERNS.keys(): # $KEYWORD_PATTERNS is a meta-variable mapping names to variables. for $KEYWORD in $KEYWORD_PATTERNS: $KEYWORD_PATTERNS[$KEYWORD] = $dict[QUOTE($KEYWORD)] if guard: DONE
Example: [4]
A pattern including a double-star pattern:
case {$KEYWORD_PATTERNS, **$DOUBLE_STARRED_PATTERN} if guard::
translates to:
if $kind & MATCH_MAPPING: if $dict is None: $dict = dict($value) if $dict.keys() >= $KEYWORD_PATTERNS.keys(): # $KEYWORD_PATTERNS is a meta-variable mapping names to variables. $tmp = dict($dict) for $KEYWORD in $KEYWORD_PATTERNS: $KEYWORD_PATTERNS[$KEYWORD] = $tmp.pop(QUOTE($KEYWORD)) $DOUBLE_STARRED_PATTERN = $tmp DONE
Example: [5]
Class Patterns
Class pattern with no arguments:
match ClsName() if guard:
translates to:
if $kind & MATCH_CLASS: if isinstance($value, ClsName): if guard: DONE
Class pattern with a single positional pattern:
match ClsName($PATTERN) if guard:
translates to:
if $kind & MATCH_SELF: if isinstance($value, ClsName): x = $value if guard: DONE else: As other positional-only class pattern
Positional-only class pattern:
match ClsName($VARS) if guard:
translates to:
if $kind & MATCH_CLASS: if isinstance($value, ClsName): if $items is None: $items = type($value).__deconstruct__($value) # $VARS is a meta-variable. if len($items) == len($VARS): $VARS = $items if guard: DONE
Note
__attributes__ is not checked when matching positional-only class patterns, this allows classes to match only positional-only patterns by setting __attributes__ to ().
Class patterns with keyword patterns:
match ClsName($VARS, $KEYWORD_PATTERNS) if guard:
translates to:
if $kind & MATCH_CLASS: if isinstance($value, ClsName): if $attrs is None: $attrs = type($value).__attributes__ if $items is None: $items = type($value).__deconstruct__($value) $right_attrs = attrs[len($VARS):] if set($right_attrs) >= set($KEYWORD_PATTERNS): $VARS = items[:len($VARS)] for $KEYWORD in $KEYWORD_PATTERNS: $index = $attrs.index(QUOTE($KEYWORD)) $KEYWORD_PATTERNS[$KEYWORD] = $items[$index] if guard: DONE
Example: [6]
Class patterns with all keyword patterns:
match ClsName($KEYWORD_PATTERNS) if guard:
translates to:
if $kind & MATCH_CLASS: As above with $VARS == () elif $kind & MATCH_DEFAULT: if isinstance($value, ClsName) and hasattr($value, "__dict__"): if $value.__dict__.keys() >= set($KEYWORD_PATTERNS): for $KEYWORD in $KEYWORD_PATTERNS: $KEYWORD_PATTERNS[$KEYWORD] = $value.__dict__[QUOTE($KEYWORD)] if guard: DONE
Example: [7]
Non-conforming __match_kind__
All classes should ensure that the the value of __match_kind__ follows the specification. Therefore, implementations can assume, without checking, that all the following are true:
(__match_kind__ & (MATCH_SEQUENCE | MATCH_MAPPING)) != (MATCH_SEQUENCE | MATCH_MAPPING) (__match_kind__ & (MATCH_SELF | MATCH_CLASS)) != (MATCH_SELF | MATCH_CLASS) (__match_kind__ & (MATCH_SELF | MATCH_DEFAULT)) != (MATCH_SELF | MATCH_DEFAULT) (__match_kind__ & (MATCH_DEFAULT | MATCH_CLASS)) != (MATCH_DEFAULT | MATCH_CLASS)
Thus, implementations can assume that __match_kind__ & MATCH_SEQUENCE implies (__match_kind__ & MATCH_MAPPING) == 0, and vice-versa. Likewise for MATCH_SELF, MATCH_CLASS and MATCH_DEFAULT.
If __match_kind__ does not follow the specification, then implementations may treat any of the expressions of the form $kind & MATCH_... above as having any value.
Implementation of __match_kind__ in the standard library
object.__match_kind__ will be MATCH_DEFAULT.
For common builtin classes __match_kind__ will be:
- bool: MATCH_SELF
- bytearray: MATCH_SELF
- bytes: MATCH_SELF
- float: MATCH_SELF
- frozenset: MATCH_SELF
- int: MATCH_SELF
- set: MATCH_SELF
- str: MATCH_SELF
- list: MATCH_SEQUENCE | MATCH_SELF
- tuple: MATCH_SEQUENCE | MATCH_SELF
- dict: MATCH_MAPPING | MATCH_SELF
Named tuples will have __match_kind__ set to MATCH_SEQUENCE | MATCH_CLASS.
- All other standard library classes for which issubclass(cls, collections.abc.Mapping) is true will have __match_kind__ set to MATCH_MAPPING.
- All other standard library classes for which issubclass(cls, collections.abc.Sequence) is true will have __match_kind__ set to MATCH_SEQUENCE.
Legal optimizations
The above semantics implies a lot of redundant effort and copying in the implementation. However, it is possible to implement the above semantics efficiently by employing semantic preserving transformations on the naive implementation.
When performing matching, implementations are allowed to treat the following functions and methods as pure:
- cls.__len__() for any class supporting MATCH_SEQUENCE
- dict.keys()
- dict.__contains__()
- dict.__getitem__()
Implementations are also allowed to freely replace isinstance(obj, cls) with issubclass(type(obj), cls) and vice-versa.
Security Implications
None.
Implementation
The naive implementation that follows from the specification will not be very efficient. Fortunately, there are some reasonably straightforward transformations that can be used to improve performance. Performance should be comparable to the implementation of PEP 634 (at time of writing) by the release of 3.10. Further performance improvements may have to wait for the 3.11 release.
Possible optimizations
The following is not part of the specification, but guidelines to help developers create an efficient implementation.
Splitting evaluation into lanes
Since the first step in matching each pattern is check to against the kind, it is possible to combine all the checks against kind into a single multi-way branch at the beginning of the match. The list of cases can then be duplicated into several "lanes" each corresponding to one kind. It is then trivial to remove unmatchable cases from each lane. Depending on the kind, different optimization strategies are possible for each lane. Note that the body of the match clause does not need to be duplicated, just the pattern.
Sequence patterns
This is probably the most complex to optimize and the most profitable in terms of performance. Since each pattern can only match a range of lengths, often only a single length, the sequence of tests can be rewitten in as an explicit iteration over the sequence, attempting to match only those patterns that apply to that sequence length.
For example:
case []: A case [x]: B case [x, y]: C case other: D
Can be compiled roughly as:
# Choose lane $i = iter($value) for $0 in $i: break else: A goto done for $1 in $i: break else: x = $0 B goto done for $2 in $i: del $0, $1, $2 break else: x = $0 y = $1 C goto done other = $value D done:
Mapping patterns
The best stategy here is probably to form a decision tree based on the size of the mapping and which keys are present. There is no point repeatedly testing for the presence of a key. For example:
match obj: case {a:x, b:y}: W case {a:x, c:y}: X case {a:x, b:_, c:y}: Y case other: Z
If the key "a" is not present when checking for case X, there is no need to check it again for Y.
The mapping lane can be implemented, roughly as:
# Choose lane if len($dict) == 2: if "a" in $dict: if "b" in $dict: x = $dict["a"] y = $dict["b"] goto W if "c" in $dict: x = $dict["a"] y = $dict["c"] goto X elif len(dict) == 3: if "a" in $dict and "b" in $dict: x = $dict["a"] y = $dict["c"] goto Y other = $value goto Z
Summary of differences between this PEP and PEP 634
The changes to the semantics can be summarized as:
- Selecting the kind of pattern uses cls.__match_kind__ instead of issubclass(cls, collections.abc.Mapping) and issubclass(cls, collections.abc.Sequence) and allows classes control over which kinds of pattern they match.
- Class matching is via the __attributes__ attribute and __deconstruct__ method, rather than the __match_args__ method, and allows classes more control over how they are deconstructed.
- The default behavior when matching a class pattern with keyword patterns is changed. Only the instance dictionary is used. This is to avoid unintended capture of bound-methods.
There are no changes to syntax.
Rejected Ideas
None, as yet.
Open Issues
None, as yet.
Code examples
[1] |
class Basic: __match_kind__ = MATCH_CLASS __attributes__ = () def __deconstruct__(self): return self._args
[2] |
This:
case [a, b] if a is b:
translates to:
if $kind & MATCH_SEQUENCE: if $list is None: $list = list($value) if len($list) == 2: a, b = $list if a is b: DONE
[3] |
This:
case [a, *b, c]:
translates to:
if $kind & MATCH_SEQUENCE: if $list is None: $list = list($value) if len($list) >= 2: a, *b, c = $list DONE
[4] |
This:
case {"x": x, "y": y} if x > 2:
translates to:
if $kind & MATCH_MAPPING: if $dict is None: $dict = dict($value) if $dict.keys() == {"x", "y"}: x = $dict["x"] y = $dict["y"] if x > 2: DONE
[5] |
This:
case {"x": x, "y": y, **: z}:
translates to:
if $kind & MATCH_MAPPING: if $dict is None: $dict = dict($value) if $dict.keys() >= {"x", "y"}: $tmp = dict($dict) x = $tmp.pop("x") y = $tmp.pop("y") z = $tmp DONE
[6] |
This:
match ClsName(x, a=y):
translates to:
if $kind & MATCH_CLASS: if isinstance($value, ClsName): if $attrs is None: $attrs = type($value).__attributes__ if $items is None: $items = type($value).__deconstruct__($value) $right_attrs = $attrs[1:] if "a" in $right_attrs: $y_index = $attrs.index("a") x = $items[0] y = $items[$y_index] DONE
[7] |
This:
match ClsName(a=x, b=y):
translates to:
if $kind & MATCH_CLASS: if isinstance($value, ClsName): if $attrs is None: $attrs = type($value).__attributes__ if $items is None: $items = type($value).__deconstruct__($value) if "a" in $attrs and "b" in $attrs: $x_index = $attrs.index("a") x = $items[$x_index] $y_index = $attrs.index("b") y = $items[$y_index] DONE elif $kind & MATCH_DEFAULT: if isinstance($value, ClsName) and hasattr($value, "__dict__"): $obj_dict = $value.__dict__ if "a" in $obj_dict and "b" in $obj_dict: x = $obj_dict["a"] y = $obj_dict["b"] DONE
Copyright
This document is placed in the public domain or under the CC0-1.0-Universal license, whichever is more permissive.