Notice: While JavaScript is not essential for this website, your interaction with the content will be limited. Please turn JavaScript on for the full experience.

PEP 558 -- Defined semantics for locals()

PEP:558
Title:Defined semantics for locals()
Author:Nick Coghlan <ncoghlan at gmail.com>
BDFL-Delegate:Nathaniel J. Smith
Discussions-To:<python-dev at python.org>
Status:Draft
Type:Standards Track
Created:08-Sep-2017
Python-Version:3.11
Post-History:2017-09-08, 2019-05-22, 2019-05-30, 2019-12-30, 2021-07-18, 2021-08-26

Abstract

The semantics of the locals() builtin have historically been underspecified and hence implementation dependent.

This PEP proposes formally standardising on the behaviour of the CPython 3.10 reference implementation for most execution scopes, with some adjustments to the behaviour at function scope to make it more predictable and independent of the presence or absence of tracing functions.

In addition, it proposes that the following functions be added to the stable Python C API/ABI:

typedef enum {
  PyLocals_UNDEFINED = -1,
  PyLocals_DIRECT_REFERENCE = 0,
  PyLocals_SHALLOW_COPY = 1,
  _PyLocals_ENSURE_32BIT_ENUM = 2147483647
} PyLocals_Kind;

PyLocals_Kind PyLocals_GetKind();
PyObject * PyLocals_Get();
PyObject * PyLocals_GetCopy();
PyObject * PyLocals_GetView();

It also proposes the addition of several supporting functions and type definitions to the CPython C API.

Motivation

While the precise semantics of the locals() builtin are nominally undefined, in practice, many Python programs depend on it behaving exactly as it behaves in CPython (at least when no tracing functions are installed).

Other implementations such as PyPy are currently replicating that behaviour, up to and including replication of local variable mutation bugs that can arise when a trace hook is installed [1].

While this PEP considers CPython's current behaviour when no trace hooks are installed to be largely acceptable, it considers the current behaviour when trace hooks are installed to be problematic, as it causes bugs like [1] without even reliably enabling the desired functionality of allowing debuggers like pdb to mutate local variables [3].

Review of the initial PEP and the draft implementation then identified an opportunity for simplification of both the documentation and implementation of the function level locals() behaviour by updating it to return an independent snapshot of the function locals and closure variables on each call, rather than continuing to return the semi-dynamic intermittently updated shared copy that it has historically returned in CPython.

Proposal

The expected semantics of the locals() builtin change based on the current execution scope. For this purpose, the defined scopes of execution are:

  • module scope: top-level module code, as well as any other code executed using exec() or eval() with a single namespace
  • class scope: code in the body of a class statement, as well as any other code executed using exec() or eval() with separate local and global namespaces
  • function scope: code in the body of a def or async def statement, or any other construct that creates an optimized code block in CPython (e.g. comprehensions, lambda functions)

This PEP proposes elevating most of the current behaviour of the CPython reference implementation to become part of the language specification, except that each call to locals() at function scope will create a new dictionary object, rather than caching a common dict instance in the frame object that each invocation will update and return.

This PEP also proposes to largely eliminate the concept of a separate "tracing" mode from the CPython reference implementation. In releases up to and including Python 3.10, the CPython interpreter behaves differently when a trace hook has been registered in one or more threads via an implementation dependent mechanism like sys.settrace ([4]) in CPython's sys module or PyEval_SetTrace ([5]) in CPython's C API. If this PEP is accepted, then the only remaining behavioural difference when a trace hook is installed is that some optimisations in the interpreter eval loop are disabled when the tracing logic needs to run after each opcode.

This PEP proposes changes to CPython's behaviour at function scope that make the locals() builtin semantics when a trace hook is registered identical to those used when no trace hook is registered, while also making the related frame API semantics clearer and easier for interactive debuggers to rely on.

The proposed elimination of tracing mode affects the semantics of frame object references obtained through other means, such as via a traceback, or via the sys._getframe() API, as the write-through semantics needed for trace hook support are always provided by the f_locals attribute on frame objects, rather than being runtime state dependent.

New locals() documentation

The heart of this proposal is to revise the documentation for the locals() builtin to read as follows:

Return a mapping object representing the current local symbol table, with variable names as the keys, and their currently bound references as the values.

At module scope, as well as when using exec() or eval() with a single namespace, this function returns the same namespace as globals().

At class scope, it returns the namespace that will be passed to the metaclass constructor.

When using exec() or eval() with separate local and global namespaces, it returns the local namespace passed in to the function call.

In all of the above cases, each call to locals() in a given frame of execution will return the same mapping object. Changes made through the mapping object returned from locals() will be visible as bound, rebound, or deleted local variables, and binding, rebinding, or deleting local variables will immediately affect the contents of the returned mapping object.

At function scope (including for generators and coroutines), each call to locals() instead returns a fresh dictionary containing the current bindings of the function's local variables and any nonlocal cell references. In this case, name binding changes made via the returned dict are not written back to the corresponding local variables or nonlocal cell references, and binding, rebinding, or deleting local variables and nonlocal cell references does not affect the contents of previously returned dictionaries.

There would also be a versionchanged note for the release making this change:

In prior versions, the semantics of mutating the mapping object returned from locals() were formally undefined. In CPython specifically, the mapping returned at function scope could be implicitly refreshed by other operations, such as calling locals() again, or the interpreter implicitly invoking a Python level trace function. Obtaining the legacy CPython behaviour now requires explicit calls to update the initially returned dictionary with the results of subsequent calls to locals().

For reference, the current documentation of this builtin reads as follows:

Update and return a dictionary representing the current local symbol table. Free variables are returned by locals() when it is called in function blocks, but not in class blocks.

Note: The contents of this dictionary should not be modified; changes may not affect the values of local and free variables used by the interpreter.

(In other words: the status quo is that the semantics and behaviour of locals() are formally implementation defined, whereas the proposed state after this PEP is that the only implementation defined behaviour will be that associated with whether or not the implementation emulates the CPython frame API, with the behaviour in all other cases being defined by the language and library references)

Module scope

At module scope, as well as when using exec() or eval() with a single namespace, locals() must return the same object as globals(), which must be the actual execution namespace (available as inspect.currentframe().f_locals in implementations that provide access to frame objects).

Variable assignments during subsequent code execution in the same scope must dynamically change the contents of the returned mapping, and changes to the returned mapping must change the values bound to local variable names in the execution environment.

To capture this expectation as part of the language specification, the following paragraph will be added to the documentation for locals():

At module scope, as well as when using exec() or eval() with a single namespace, this function returns the same namespace as globals().

This part of the proposal does not require any changes to the reference implementation - it is standardisation of the current behaviour.

Class scope

At class scope, as well as when using exec() or eval() with separate global and local namespaces, locals() must return the specified local namespace (which may be supplied by the metaclass __prepare__ method in the case of classes). As for module scope, this must be a direct reference to the actual execution namespace (available as inspect.currentframe().f_locals in implementations that provide access to frame objects).

Variable assignments during subsequent code execution in the same scope must change the contents of the returned mapping, and changes to the returned mapping must change the values bound to local variable names in the execution environment.

The mapping returned by locals() will not be used as the actual class namespace underlying the defined class (the class creation process will copy the contents to a fresh dictionary that is only accessible by going through the class machinery).

For nested classes defined inside a function, any nonlocal cells referenced from the class scope are not included in the locals() mapping.

To capture this expectation as part of the language specification, the following two paragraphs will be added to the documentation for locals():

When using exec() or eval() with separate local and global namespaces, [this function] returns the given local namespace.

At class scope, it returns the namespace that will be passed to the metaclass constructor.

This part of the proposal does not require any changes to the reference implementation - it is standardisation of the current behaviour.

Function scope

At function scope, interpreter implementations are granted significant freedom to optimise local variable access, and hence are NOT required to permit arbitrary modification of local and nonlocal variable bindings through the mapping returned from locals().

Historically, this leniency has been described in the language specification with the words "The contents of this dictionary should not be modified; changes may not affect the values of local and free variables used by the interpreter."

This PEP proposes to change that text to instead say:

At function scope (including for generators and coroutines), each call to locals() instead returns a fresh dictionary containing the current bindings of the function's local variables and any nonlocal cell references. In this case, name binding changes made via the returned dict are not written back to the corresponding local variables or nonlocal cell references, and binding, rebinding, or deleting local variables and nonlocal cell references does not affect the contents of previously returned dictionaries.

This part of the proposal does require changes to the CPython reference implementation, as CPython currently returns a shared mapping object that may be implicitly refreshed by additional calls to locals(), and the "write back" strategy currently used to support namespace changes from trace functions also doesn't comply with it (and causes the quirky behavioural problems mentioned in the Motivation above).

CPython Implementation Changes

Summary of proposed implementation-specific changes

  • Changes are made as necessary to provide the updated Python level semantics

  • Two new functions are added to the stable ABI to replicate the updated behaviour of the Python locals() builtin:

    PyObject * PyLocals_Get();
    PyLocals_Kind PyLocals_GetKind();
    
  • One new function is added to the stable ABI to efficiently get a snapshot of the local namespace in the running frame:

    PyObject * PyLocals_GetCopy();
    
  • One new function is added to the stable ABI to get a read-only view of the local namespace in the running frame:

    PyObject * PyLocals_GetView();
    
  • Corresponding frame accessor functions for these new public APIs are added to the CPython frame C API

  • On optimised frames, the Python level f_locals API will return dynamically created read/write proxy objects that directly access the frame's local and closure variable storage. To provide interoperability with the existing PyEval_GetLocals() API, the proxy objects will continue to use the C level frame locals data storage field to hold a value cache that also allows for storage of arbitrary additional keys. Additional details on the expected behaviour of these fast locals proxy objects are covered below.

  • No C API function is added to get access to a mutable mapping for the local namespace. Instead, PyObject_GetAttrString(frame, "f_locals") is used, the same API as is used in Python code.

  • PyEval_GetLocals() remains supported and does not emit a programmatic warning, but will be deprecated in the documentation in favour of the new APIs that don't rely on returning a borrowed reference

  • PyFrame_FastToLocals() and PyFrame_FastToLocalsWithError() remain supported and do not emit a programmatic warning, but will be deprecated in the documentation in favour of the new APIs that don't require direct access to the internal data storage layout of frame objects

  • PyFrame_LocalsToFast() always raises RuntimeError(), indicating that PyObject_GetAttrString(frame, "f_locals") should be used to obtain a mutable read/write mapping for the local variables.

  • The trace hook implementation will no longer call PyFrame_FastToLocals() implicitly. The version porting guide will recommend migrating to PyFrame_GetLocalsView() for read-only access and PyObject_GetAttrString(frame, "f_locals") for read/write access.

Providing the updated Python level semantics

The implementation of the locals() builtin is modified to return a distinct copy of the local namespace for optimised frames, rather than a direct reference to the internal frame value cache updated by the PyFrame_FastToLocals() C API and returned by the PyEval_GetLocals() C API.

Resolving the issues with tracing mode behaviour

The current cause of CPython's tracing mode quirks (both the side effects from simply installing a tracing function and the fact that writing values back to function locals only works for the specific function being traced) is the way that locals mutation support for trace hooks is currently implemented: the PyFrame_LocalsToFast function.

When a trace function is installed, CPython currently does the following for function frames (those where the code object uses "fast locals" semantics):

  1. Calls PyFrame_FastToLocals to update the frame value cache
  2. Calls the trace hook (with tracing of the hook itself disabled)
  3. Calls PyFrame_LocalsToFast to capture any changes made to the frame value cache

This approach is problematic for a few different reasons:

  • Even if the trace function doesn't mutate the value cache, the final step resets any cell references back to the state they were in before the trace function was called (this is the root cause of the bug report in [1])
  • If the trace function does mutate the value cache, but then does something that causes the value cache to be refreshed from the frame, those changes are lost (this is one aspect of the bug report in [3])
  • If the trace function attempts to mutate the local variables of a frame other than the one being traced (e.g. frame.f_back.f_locals), those changes will almost certainly be lost (this is another aspect of the bug report in [3])
  • If a reference to the frame value cache (e.g. retrieved via locals()) is passed to another function, and that function mutates the value cache, then those changes may be written back to the execution frame if a trace hook is installed

The proposed resolution to this problem is to take advantage of the fact that whereas functions typically access their own namespace using the language defined locals() builtin, trace functions necessarily use the implementation dependent frame.f_locals interface, as a frame reference is what gets passed to hook implementations.

Instead of being a direct reference to the internal frame value cache historically returned by the locals() builtin, the Python level frame.f_locals will be updated to instead return instances of a dedicated fast locals proxy type that writes and reads values directly to and from the fast locals array on the underlying frame. Each access of the attribute produces a new instance of the proxy (so creating proxy instances is intentionally a cheap operation).

Despite the new proxy type becoming the preferred way to access local variables on optimised frames, the internal value cache stored on the frame is still retained for two key purposes:

  • maintaining backwards compatibility for and interoperability with the PyEval_GetLocals() C API
  • providing storage space for additional keys that don't have slots in the fast locals array (e.g. the __return__ and __exception__ keys set by pdb when tracing code execution for debugging purposes)

With the changes in this PEP, this internal frame value cache is no longer directly accessible from Python code (whereas historically it was both returned by the locals() builtin and available as the frame.f_locals attribute). Instead, the value cache is only accessible via the PyEval_GetLocals() C API and by directly accessing the internal storage of a frame object.

Fast locals proxy objects and the internal frame value cache returned by PyEval_GetLocals() offer the following behavioural guarantees:

  • changes made via a fast locals proxy will be immediately visible to the frame itself, to other fast locals proxy objects for the same frame, and in the internal value cache stored on the frame (it is this last point that provides PyEval_GetLocals() interoperability)
  • changes made directly to the internal frame value cache will never be visible to the frame itself, and will only be reliably visible via fast locals proxies for the same frame if the change relates to extra variables that don't have slots in the frame's fast locals array
  • changes made by executing code in the frame will be visible to newly created fast locals proxy objects, when directly accessing specific keys on existing fast locals proxy objects, and when performing intrinsically O(n) operations on existing fast locals proxy objects. Visibility in the internal frame value cache (and in fast locals proxy operations that rely on the frame) cache is subject to the cache update guidelines discussed in the next section

Due to the last point, the frame API documentation will recommend that a new frame.f_locals reference be retrieved whenever an optimised frame (or a related frame) might have been running code that binds or unbinds local variable or cell references, and the code iterates over the proxy, checks its length, or calls popitem(). This will be the most natural style of use in tracing function implementations, as those are passed references to frames rather than directly to frames.f_locals.

Fast locals proxy implementation details

Each fast locals proxy instance has two internal attributes that are not exposed as part of the Python runtime API:

  • frame: the underlying optimised frame that the proxy provides access to
  • frame_cache_updated: whether this proxy has already updated the frame's internal value cache at least once

In addition, proxy instances use and update the following attributes stored on the underlying frame:

  • fast_refs: a hidden mapping from variable names to either fast local storage offsets (for local variables) or to closure cells (for closure variables). This mapping is lazily initialized on the first frame read or write access through a fast locals proxy, rather than being eagerly populated as soon as the first fast locals proxy is created.
  • locals: the internal frame value cache returned by the PyEval_GetLocals() C API and updated by the PyFrame_FastToLocals() C API. This is the mapping that the locals() builtin returns in Python 3.10 and earlier.

__getitem__ operations on the proxy will populate the fast_refs mapping (if it is not already populated), and then either return the relevant value (if the key is found in either the fast_refs mapping or the internal frame value cache), or else raise KeyError. Variables that are defined on the frame but not currently bound raise KeyError (just as they're omitted from the result of locals()).

As the frame storage is always accessed directly, the proxy will automatically pick up name binding and unbinding operations that take place as the function executes. The internal value cache is implicitly updated when individual variables are read from the frame state (including for containment checks, which need to check if the name is currently bound or unbound).

Similarly, __setitem__ and __delitem__ operations on the proxy will directly affect the corresponding fast local or cell reference on the underlying frame, ensuring that changes are immediately visible to the running Python code, rather than needing to be written back to the runtime storage at some later time. Such changes are also immediately written to the internal frame value cache to reduce the opportunities for the cache to get out of sync with the frame state and to make them visible to users of the PyEval_GetLocals() C API.

Keys that are not defined as local or closure variables on the underlying frame are still written to the internal value cache on optimised frames. This allows utilities like pdb (which writes __return__ and __exception__ values into the frame's f_locals mapping) to continue working as they always have. These additional keys that do not correspond to a local or closure variable on the frame will be left alone by future cache sync operations.

Fast locals proxy objects offer a proxy-specific method that explicitly syncs the internal frame cache with the current state of the fast locals array: proxy.sync_frame_cache(). This method runs PyFrame_FastToLocalsWithError() to ensure the cache is consistent with the current frame state.

Using a particular proxy instance to sync the frame cache sets the internal frame_cache_updated flag on that instance.

For most use cases, explicitly syncing the frame cache shouldn't be necessary, as the following intrinsically O(n) operations implicitly sync the frame cache whenever they're called on a proxy instance:

  • __str__
  • __or__ (dict union)
  • copy()

While the following operations will implicitly sync the frame cache if frame_cache_updated has not yet been set on that instance:

  • __len__
  • __iter__
  • __reversed__
  • keys()
  • values()
  • items()
  • popitem()
  • value comparison operations

Other Mapping and MutableMapping methods on the proxy will behave as expected for a mapping with these essential method semantics regardless of whether the internal frame value cache is up to date or not.

An additional benefit of storing only the variable value cache on the frame (rather than storing an instance of the proxy type), is that it avoids creating a reference cycle from the frame back to itself, so the frame will only be kept alive if another object retains a reference to a proxy instance.

Note: calling the proxy.clear() method has a similarly broad impact as calling PyFrame_LocalsToFast() on an empty frame value cache in earlier versions. Not only will the frame local variables be cleared, but also any cell variables accessible from the frame (whether those cells are owned by the frame itself or by an outer frame). This can clear a class's __class__ cell if called on the frame of a method that uses the zero-arg super() construct (or otherwise references __class__). This exceeds the scope of calling frame.clear(), as that only drop's the frame's references to cell variables, it doesn't clear the cells themselves. This PEP could be a potential opportunity to narrow the scope of attempts to clear the frame variables directly by leaving cells belonging to outer frames alone, and only clearing local variables and cells belonging directly to the frame underlying the proxy (this issue affects PEP 667 as well, as the question relates to the handling of cell variables, and is entirely independent of the internal frame value cache).

Changes to the stable C API/ABI

Unlike Python code, extension module functions that call in to the Python C API can be called from any kind of Python scope. This means it isn't obvious from the context whether locals() will return a snapshot or not, as it depends on the scope of the calling Python code, not the C code itself.

This means it is desirable to offer C APIs that give predictable, scope independent, behaviour. However, it is also desirable to allow C code to exactly mimic the behaviour of Python code at the same scope.

To enable mimicking the behaviour of Python code, the stable C ABI would gain the following new functions:

PyObject * PyLocals_Get();
PyLocals_Kind PyLocals_GetKind();

PyLocals_Get() is directly equivalent to the Python locals() builtin. It returns a new reference to the local namespace mapping for the active Python frame at module and class scope, and when using exec() or eval(). It returns a shallow copy of the active namespace at function/coroutine/generator scope.

PyLocals_GetKind() returns a value from the newly defined PyLocals_Kind enum, with the following options being available:

  • PyLocals_DIRECT_REFERENCE: PyLocals_Get() returns a direct reference to the local namespace for the running frame.
  • PyLocals_SHALLOW_COPY: PyLocals_Get() returns a shallow copy of the local namespace for the running frame.
  • PyLocals_UNDEFINED: an error occurred (e.g. no active Python thread state). A Python exception will be set if this value is returned.

Since the enum is used in the stable ABI, an additional 31-bit value is set to ensure that it is safe to cast arbitrary signed 32-bit signed integers to PyLocals_Kind values.

This query API allows extension module code to determine the potential impact of mutating the mapping returned by PyLocals_Get() without needing access to the details of the running frame object.

To allow extension module code to behave consistently regardless of the active Python scope, the stable C ABI would gain the following new functions:

PyObject * PyLocals_GetCopy();
PyObject * PyLocals_GetView();

PyLocals_GetCopy() returns a new dict instance populated from the current locals namespace. Roughly equivalent to dict(locals()) in Python code, but avoids the double-copy in the case where locals() already returns a shallow copy.

PyLocals_GetView() returns a new read-only mapping proxy instance for the current locals namespace. This view immediately reflects all local variable changes, independently of whether the running frame is optimised or not. However, some operations (e.g. length checking, iteration, mapping equality comparisons) may be subject to frame cache consistency issues on optimised frames (as noted above when describing the behaviour of the fast locals proxy).

The existing PyEval_GetLocals() API will retain its existing behaviour in CPython (mutable locals at class and module scope, shared dynamic snapshot otherwise). However, its documentation will be updated to note that the conditions under which the shared dynamic snapshot get updated have changed.

The PyEval_GetLocals() documentation will also be updated to recommend replacing usage of this API with whichever of the new APIs is most appropriate for the use case:

  • Use PyLocals_GetView() for read-only access to the current locals namespace.
  • Use PyLocals_GetCopy() for a regular mutable dict that contains a copy of the current locals namespace, but has no ongoing connection to the active frame.
  • Use PyLocals_Get() to exactly match the semantics of the Python level locals() builtin.
  • Query PyLocals_GetKind() explicitly to implement custom handling (e.g. raising a meaningful exception) for scopes where PyLocals_Get() would return a shallow copy rather than granting read/write access to the locals namespace.
  • Use implementation specific APIs (e.g. PyObject_GetAttrString(frame, "f_locals")) if read/write access to the frame is required and PyLocals_GetKind() returns something other than PyLocals_DIRECT_REFERENCE.

Changes to the public CPython C API

The existing PyEval_GetLocals() API returns a borrowed reference, which means it cannot be updated to return the new shallow copies at function scope. Instead, it will continue to return a borrowed reference to an internal dynamic snapshot stored on the frame object. This shared mapping will behave similarly to the existing shared mapping in Python 3.10 and earlier, but the exact conditions under which it gets refreshed will be different. Specifically, it will be updated only in the following circumstance:

  • any call to PyEval_GetLocals(), PyLocals_Get(), PyLocals_GetCopy(), or the Python locals() builtin while the frame is running
  • any call to PyFrame_GetLocals(), PyFrame_GetLocalsCopy(), _PyFrame_BorrowLocals(), PyFrame_FastToLocals(), or PyFrame_FastToLocalsWithError() for the frame
  • retrieving the f_locals attribute from a Python level frame object
  • any call to the sync_frame_cache() method on a fast locals proxy referencing that frame
  • any operation on a fast locals proxy object that requires the shared mapping to be up to date on the underlying frame. In the initial reference implementation, those operations are those that are intrinsically O(n) operations (flp.copy() and rendering as a string), as well as those that refresh the cache entries for individual keys.

Accessing the frame "view" APIs will not implicitly update the shared dynamic snapshot, and the CPython trace hook handling will no longer implicitly update it either.

(Note: even though PyEval_GetLocals() is part of the stable C API/ABI, the specifics of when the namespace it returns gets refreshed are still an interpreter implementation detail)

The additions to the public CPython C API are the frame level enhancements needed to support the stable C API/ABI updates:

PyLocals_Kind PyFrame_GetLocalsKind(frame);
PyObject * PyFrame_GetLocals(frame);
PyObject * PyFrame_GetLocalsCopy(frame);
PyObject * PyFrame_GetLocalsView(frame);
PyObject * _PyFrame_BorrowLocals(frame);

PyFrame_GetLocalsKind(frame) is the underlying API for PyLocals_GetKind().

PyFrame_GetLocals(frame) is the underlying API for PyLocals_Get().

PyFrame_GetLocalsCopy(frame) is the underlying API for PyLocals_GetCopy().

PyFrame_GetLocalsView(frame) is the underlying API for PyLocals_GetView().

_PyFrame_BorrowLocals(frame) is the underlying API for PyEval_GetLocals(). The underscore prefix is intended to discourage use and to indicate that code using it is unlikely to be portable across implementations. However, it is documented and visible to the linker in order to avoid having to access the internals of the frame struct from the PyEval_GetLocals() implementation.

The PyFrame_LocalsToFast() function will be changed to always emit RuntimeError, explaining that it is no longer a supported operation, and affected code should be updated to use PyObject_GetAttrString(frame, "f_locals") to obtain a read/write proxy instead.

In addition to the above documented interfaces, the draft reference implementation also exposes the following undocumented interfaces:

PyTypeObject _PyFastLocalsProxy_Type;
#define _PyFastLocalsProxy_CheckExact(self) Py_IS_TYPE(op, &_PyFastLocalsProxy_Type)

This type is what the reference implementation actually returns from PyObject_GetAttrString(frame, "f_locals") for optimized frames (i.e. when PyFrame_GetLocalsKind() returns PyLocals_SHALLOW_COPY).

Reducing the runtime overhead of trace hooks

As noted in [9], the implicit call to PyFrame_FastToLocals() in the Python trace hook support isn't free, and could be rendered unnecessary if the frame proxy read values directly from the frame instead of getting them from the mapping.

As the new frame locals proxy type doesn't require separate data refresh steps, this PEP incorporates Victor Stinner's proposal to no longer implicitly call PyFrame_FastToLocalsWithError() before calling trace hooks implemented in Python.

Code using the new frame view APIs will have the dynamic locals snapshot implicitly refreshed when accessing methods that need it, while code using the PyEval_GetLocals() API will implicitly refresh it when making that call.

The PEP necessarily also drops the implicit call to PyFrame_LocalsToFast() when returning from a trace hook, as that API now always raises an exception.

Rationale and Design Discussion

Changing locals() to return independent snapshots at function scope

The locals() builtin is a required part of the language, and in the reference implementation it has historically returned a mutable mapping with the following characteristics:

  • each call to locals() returns the same mapping object
  • for namespaces where locals() returns a reference to something other than the actual local execution namespace, each call to locals() updates the mapping object with the current state of the local variables and any referenced nonlocal cells
  • changes to the returned mapping usually aren't written back to the local variable bindings or the nonlocal cell references, but write backs can be triggered by doing one of the following:
    • installing a Python level trace hook (write backs then happen whenever the trace hook is called)
    • running a function level wildcard import (requires bytecode injection in Py3)
    • running an exec statement in the function's scope (Py2 only, since exec became an ordinary builtin in Python 3)

Originally this PEP proposed to retain the first two of these properties, while changing the third in order to address the outright behaviour bugs that it can cause.

In [7] Nathaniel Smith made a persuasive case that we could make the behaviour of locals() at function scope substantially less confusing by retaining only the second property and having each call to locals() at function scope return an independent snapshot of the local variables and closure references rather than updating an implicitly shared snapshot.

As this revised design also made the implementation markedly easier to follow, the PEP was updated to propose this change in behaviour, rather than retaining the historical shared snapshot.

Keeping locals() as a snapshot at function scope

As discussed in [7], it would theoretically be possible to change the semantics of the locals() builtin to return the write-through proxy at function scope, rather than switching it to return independent snapshots.

This PEP doesn't (and won't) propose this as it's a backwards incompatible change in practice, even though code that relies on the current behaviour is technically operating in an undefined area of the language specification.

Consider the following code snippet:

def example():
    x = 1
    locals()["x"] = 2
    print(x)

Even with a trace hook installed, that function will consistently print 1 on the current reference interpreter implementation:

>>> example()
1
>>> import sys
>>> def basic_hook(*args):
...     return basic_hook
...
>>> sys.settrace(basic_hook)
>>> example()
1

Similarly, locals() can be passed to the exec() and eval() builtins at function scope (either explicitly or implicitly) without risking unexpected rebinding of local variables or closure references.

Provoking the reference interpreter into incorrectly mutating the local variable state requires a more complex setup where a nested function closes over a variable being rebound in the outer function, and due to the use of either threads, generators, or coroutines, it's possible for a trace function to start running for the nested function before the rebinding operation in the outer function, but finish running after the rebinding operation has taken place (in which case the rebinding will be reverted, which is the bug reported in [1]).

In addition to preserving the de facto semantics which have been in place since PEP 227 introduced nested scopes in Python 2.1, the other benefit of restricting the write-through proxy support to the implementation-defined frame object API is that it means that only interpreter implementations which emulate the full frame API need to offer the write-through capability at all, and that JIT-compiled implementations only need to enable it when a frame introspection API is invoked, or a trace hook is installed, not whenever locals() is accessed at function scope.

Returning snapshots from locals() at function scope also means that static analysis for function level code will be more reliable, as only access to the frame machinery will allow rebinding of local and nonlocal variable references in a way that is hidden from static analysis.

Retaining the internal frame value cache

Retaining the internal frame value cache results in some visible quirks when frame proxy instances are kept around and re-used after name binding and unbinding operations have been executed on the frame.

The primary reason for retaining the frame value cache is to maintain backwards compatibility with the PyEval_GetLocals() API. That API returns a borrowed reference, so it must refer to persistent state stored on the frame object. Storing a fast locals proxy object on the frame creates a problematic reference cycle, so the cleanest option is to instead continue to return a frame value cache, just as this function has done since optimised frames were first introduced.

With the frame value cache being kept around anyway, it then further made sense to rely on it to simplify the fast locals proxy mapping implementation.

Delaying implicit frame value cache updates

Earlier iterations of this PEP proposed updating the internal frame value cache whenever a new fast locals proxy instance was created for that frame. They also proposed storing a separate copy of the fast_refs lookup mapping on each

What happens with the default args for eval() and exec()?

These are formally defined as inheriting globals() and locals() from the calling scope by default.

There isn't any need for the PEP to change these defaults, so it doesn't, and exec() and eval() will start running in a shallow copy of the local namespace when that is what locals() returns.

This behaviour will have potential performance implications, especially for functions with large numbers of local variables (e.g. if these functions are called in a loop, calling globals() and locals() once before the loop and then passing the namespace into the function explicitly will give the same semantics and performance characteristics as the status quo, whereas relying on the implicit default would create a new shallow copy of the local namespace on each iteration).

(Note: the reference implementation draft PR has updated the locals() and vars(), eval(), and exec() builtins to use PyLocals_Get(). The dir() builtin still uses PyEval_GetLocals(), since it's only using it to make a list from the keys).

Changing the frame API semantics in regular operation

Earlier versions of this PEP proposed having the semantics of the frame f_locals attribute depend on whether or not a tracing hook was currently installed - only providing the write-through proxy behaviour when a tracing hook was active, and otherwise behaving the same as the historical locals() builtin.

That was adopted as the original design proposal for a couple of key reasons, one pragmatic and one more philosophical:

  • Object allocations and method wrappers aren't free, and tracing functions aren't the only operations that access frame locals from outside the function. Restricting the changes to tracing mode meant that the additional memory and execution time overhead of these changes would be as close to zero in regular operation as we can possibly make them.
  • "Don't change what isn't broken": the current tracing mode problems are caused by a requirement that's specific to tracing mode (support for external rebinding of function local variable references), so it made sense to also restrict any related fixes to tracing mode

However, actually attempting to implement and document that dynamic approach highlighted the fact that it makes for a really subtle runtime state dependent behaviour distinction in how frame.f_locals works, and creates several new edge cases around how f_locals behaves as trace functions are added and removed.

Accordingly, the design was switched to the current one, where frame.f_locals is always a write-through proxy, and locals() is always a snapshot, which is both simpler to implement and easier to explain.

Regardless of how the CPython reference implementation chooses to handle this, optimising compilers and interpreters also remain free to impose additional restrictions on debuggers, such as making local variable mutation through frame objects an opt-in behaviour that may disable some optimisations (just as the emulation of CPython's frame API is already an opt-in flag in some Python implementations).

Continuing to support storing additional data on optimised frames

One of the draft iterations of this PEP proposed removing the ability to store additional data on optimised frames by writing to frame.f_locals keys that didn't correspond to local or closure variable names on the underlying frame.

While this idea offered some attractive simplification of the fast locals proxy implementation, pdb stores __return__ and __exception__ values on arbitrary frames, so the standard library test suite fails if that functionality no longer works.

Accordingly, the ability to store arbitrary keys was retained, at the expense of certain operations on proxy objects currently either being slower than desired (as they need to update the dynamic snapshot in order to provide correct behaviour), or else assuming that the cache is currently up to date (and hence potentially giving an incorrect answer if the frame state has changed in a way that doesn't automatically update the cache contents).

It is expected that the exact details of the interaction between the fast locals proxy and the f_locals value cache on the underlying frame will evolve over time as opportunities for improvement are identified.

Historical semantics at function scope

The current semantics of mutating locals() and frame.f_locals in CPython are rather quirky due to historical implementation details:

  • actual execution uses the fast locals array for local variable bindings and cell references for nonlocal variables
  • there's a PyFrame_FastToLocals operation that populates the frame's f_locals attribute based on the current state of the fast locals array and any referenced cells. This exists for three reasons:
    • allowing trace functions to read the state of local variables
    • allowing traceback processors to read the state of local variables
    • allowing locals() to read the state of local variables
  • a direct reference to frame.f_locals is returned from locals(), so if you hand out multiple concurrent references, then all those references will be to the exact same dictionary
  • the two common calls to the reverse operation, PyFrame_LocalsToFast, were removed in the migration to Python 3: exec is no longer a statement (and hence can no longer affect function local namespaces), and the compiler now disallows the use of from module import * operations at function scope
  • however, two obscure calling paths remain: PyFrame_LocalsToFast is called as part of returning from a trace function (which allows debuggers to make changes to the local variable state), and you can also still inject the IMPORT_STAR opcode when creating a function directly from a code object rather than via the compiler

This proposal deliberately doesn't formalise these semantics as is, since they only make sense in terms of the historical evolution of the language and the reference implementation, rather than being deliberately designed.

Proposing several additions to the stable C API/ABI

Historically, the CPython C API (and subsequently, the stable ABI) has exposed only a single API function related to the Python locals builtin: PyEval_GetLocals(). However, as it returns a borrowed reference, it is not possible to adapt that interface directly to supporting the new locals() semantics proposed in this PEP.

An earlier iteration of this PEP proposed a minimalist adaptation to the new semantics: one C API function that behaved like the Python locals() builtin, and another that behaved like the frame.f_locals descriptor (creating and returning the write-through proxy if necessary).

The feedback [8] on that version of the C API was that it was too heavily based on how the Python level semantics were implemented, and didn't account for the behaviours that authors of C extensions were likely to need.

The broader API now being proposed came from grouping the potential reasons for wanting to access the Python locals() namespace from an extension module into the following cases:

  • needing to exactly replicate the semantics of the Python level locals() operation. This is the PyLocals_Get() API.
  • needing to behave differently depending on whether writes to the result of PyLocals_Get() will be visible to Python code or not. This is handled by the PyLocals_GetKind() query API.
  • always wanting a mutable namespace that has been pre-populated from the current Python locals() namespace, but not wanting any changes to be visible to Python code. This is the PyLocals_GetCopy() API.
  • always wanting a read-only view of the current locals namespace, without incurring the runtime overhead of making a full copy each time. This is the PyLocals_GetView() API.

Historically, these kinds of checks and operations would only have been possible if a Python implementation emulated the full CPython frame API. With the proposed API, extension modules can instead ask more clearly for the semantics that they actually need, giving Python implementations more flexibility in how they provide those capabilities.

Comparison with PEP 667

PEP 667 offers a partially competing proposal for this PEP that suggests it would be reasonable to eliminate the internal frame value cache on optimised frames entirely.

These changes were originally offered as amendments to PEP 558, and the PEP author rejected them for three main reasons:

  • the claim that PyEval_GetLocals() is unfixable because it returns a borrowed reference is simply false, as it is still working in the PEP 558 reference implementation. All that is required to keep it working is to retain the internal frame value cache and design the fast locals proxy in such a way that it is reasonably straightforward to keep the cache up to date with changes in the frame state without incurring significant runtime overhead when the cache isn't needed. Given that this claim is false, the proposal to require that all code using the PyEval_GetLocals() API be rewritten to use a new API with different refcounting semantics fails PEP 387's requirement that API compatibility breaks should have a large benefit to breakage ratio (since there's no significant benefit gained from dropping the cache, no code breakage can be justified). The only genuinely unfixable public API is PyFrame_LocalsToFast() (which is why both PEPs propose breaking that).
  • without some form of internal value cache, the API performance characteristics of the fast locals proxy mapping become quite unintuitive. len(proxy), for example, becomes consistently O(n) in the number of variables defined on the frame, as the proxy has to iterate over the entire fast locals array to see which names are currently bound to values before it can determine the answer. By contrast, maintaining an internal frame value cache allows proxies to largely be treated as normal dictionaries from an algorithmic complexity point of view, with allowances only needing to be made for the initial implicit O(n) cache refresh that runs the first time an operation that relies on the cache being up to date is executed.
  • the claim that a cache-free implementation would be simpler is highly suspect, as PEP 667 includes only a pure Python sketch of a subset of a mutable mapping implementation, rather than a full-fledged C implementation of a new mapping type integrated with the underlying data storage for optimised frames. PEP 558's fast locals proxy implementation delegates heavily to the frame value cache for the operations needed to fully implement the mutable mapping API, allowing it to re-use the existing dict implementations of the following operations:
    • __len__
    • __str__
    • __or__ (dict union)
    • __iter__ (allowing the dict_keyiterator type to be reused)
    • __reversed__ (allowing the dict_reversekeyiterator type to be reused)
    • keys() (allowing the dict_keys type to be reused)
    • values() (allowing the dict_values type to be reused)
    • items() (allowing the dict_items type to be reused)
    • copy()
    • popitem()
    • value comparison operations

Of the three reasons, the first is the most important (since we need compelling reasons to break API backwards compatibility, and we don't have them).

The other two points relate to why the author of this PEP doesn't believe PEP 667's proposal would actually offer any significant benefits to either API consumers (while the author of this PEP concedes that PEP 558's internal frame cache sync management is more complex to deal with than PEP 667's API algorithmic complexity quirks, it's still markedly less complex than the tracing mode semantics in current Python versions) or to CPython core developers (the author of this PEP certainly didn't want to write C implementations of five new fast locals proxy specific mutable mapping helper types when he could instead just write a single cache refresh helper method and then reuse the existing builtin dict method implementations).

Taking the specific frame access example cited in PEP 667:

def foo():
    x = sys._getframe().f_locals
    y = locals()
    print(tuple(x))
    print(tuple(y))

Following the implementation improvements prompted by the suggestions in PEP 667, PEP 558 prints the same result as PEP 667 does:

('x', 'y')
('x',)

That said, it's certainly possible to desynchronise the cache quite easily when keeping proxy references around while letting code run in the frame. This isn't a new problem, as it's similar to the way that sys._getframe().f_locals behaves in existing versions when no trace hooks are installed. The following example:

def foo():
    x = sys._getframe().f_locals
    print(tuple(x))
    y = locals()
    print(tuple(x))
    print(tuple(y))

will print the following under PEP 558, as the first tuple(x) call consumes the single implicit cache update performed by the proxy instance, and y hasn't been bound yet when the locals() call refreshes it again:

('x',)
('x',)
('x',)

However, this is the origin of the coding style guideline in the body of the PEP: don't keep fast locals proxy references around if code might have been executed in that frame since the proxy instance was created. With the code updated to follow that guideline:

def foo():
    x = sys._getframe().f_locals
    print(tuple(x))
    y = locals()
    x = sys._getframe().f_locals
    print(tuple(x))
    print(tuple(y))

The output once again becomes the same as it would be under PEP 667:

('x',)
('x', 'y',)
('x',)

Tracing function implementations, which are expected to be the main consumer of the fast locals proxy API, generally won't run into the above problem, since they get passed a reference to the frame object (and retrieve a fresh fast locals proxy instance from that), while the frame itself isn't running code while the trace function is running. If the trace function does allow code to be run on the frame (e.g. it's a debugger), then it should also follow the coding guideline and retrieve a new proxy instance each time it allows code to run in the frame.

Most trace functions are going to be reading or writing individual keys, or running intrinsically O(n) operations like iterating over all currently bound variables, so they also shouldn't be impacted too badly by the performance quirks in the PEP 667 proposal. The most likely source of annoyance would be the O(n) len(proxy) implementation.

Note: the simplest way to convert the PEP 558 reference implementation into a PEP 667 implementation that doesn't break PyEval_GetLocals() would be to remove the frame_cache_updated checks in affected operations, and instead always sync the frame cache in those methods. Adopting that approach would change the algorithmic complexity of the following operations as shown (where n is the number of local and cell variables defined on the frame):

  • __len__: O(1) -> O(n)
  • __iter__: O(1) -> O(n)
  • __reversed__: O(1) -> O(n)
  • keys(): O(1) -> O(n)
  • values(): O(1) -> O(n)
  • items(): O(1) -> O(n)
  • popitem(): O(1) -> O(n)
  • value comparison operations: no longer benefit from O(1) length check shortcut

Keeping the iterator/iterable retrieval methods as O(1) would involve writing custom replacements for the corresponding builtin dict helper types. popitem() could be improved from "always O(n)" to "O(n) worst case" by creating a custom implementation that iterates over the fast locals array directly. The length check and value comparison operations have very limited opportunities for improvement: without a cache, the only way to know how many variables are currently bound is to iterate over all of them and check, and if the implementation is going to be spending that much time on an operation anyway, it may as well spend it updating the frame value cache and then consuming the result.

This feels worse than PEP 558 as written, where folks that don't want to think too hard about the cache management details, and don't care about potential performance issues with large frames, are free to add as many proxy.sync_frame_cache() (or other internal frame cache updating) calls to their code as they like.

Implementation

The reference implementation update is in development as a draft pull request on GitHub ([6]).

Acknowledgements

Thanks to Nathaniel J. Smith for proposing the write-through proxy idea in [1] and pointing out some critical design flaws in earlier iterations of the PEP that attempted to avoid introducing such a proxy.

Thanks to Steve Dower and Petr Viktorin for asking that more attention be paid to the developer experience of the proposed C API additions [8,13]_.

Thanks to Larry Hastings for the suggestion on how to use enums in the stable ABI while ensuring that they safely support typecasting from arbitrary integers.

Thanks to Mark Shannon for pushing for further simplification of the C level API and semantics, as well as significant clarification of the PEP text (and for restarting discussion on the PEP in early 2021 after a further year of inactivity) [10,11,12]_. Mark's comments that were ultimately published as PEP 667 also directly resulted in several implementation efficiency improvements that avoid incurring the cost of redundant O(n) mapping refresh operations when the relevant mappings aren't used.

References

[1](1, 2, 3, 4, 5) Broken local variable assignment given threads + trace hook + closure (https://bugs.python.org/issue30744)
[2]Clarify the required behaviour of locals() (https://bugs.python.org/issue17960)
[3](1, 2, 3) Updating function local variables from pdb is unreliable (https://bugs.python.org/issue9633)
[4]CPython's Python API for installing trace hooks (https://docs.python.org/dev/library/sys.html#sys.settrace)
[5]CPython's C API for installing trace hooks (https://docs.python.org/3/c-api/init.html#c.PyEval_SetTrace)
[6]PEP 558 reference implementation (https://github.com/python/cpython/pull/3640/files)
[7](1, 2) Nathaniel's review of possible function level semantics for locals() (https://mail.python.org/pipermail/python-dev/2019-May/157738.html)
[8]Discussion of more intentionally designed C API enhancements (https://discuss.python.org/t/pep-558-defined-semantics-for-locals/2936/3)
[9]Disable automatic update of frame locals during tracing (https://bugs.python.org/issue42197)
[10]python-dev thread: Resurrecting PEP 558 (Defined semantics for locals()) (https://mail.python.org/archives/list/python-dev@python.org/thread/TUQOEWQSCQZPUDV2UFFKQ3C3I4WGFPAJ/)
[11]python-dev thread: Comments on PEP 558 (https://mail.python.org/archives/list/python-dev@python.org/thread/A3UN4DGBCOB45STE6AQBITJFW6UZE43O/)
[12]python-dev thread: More comments on PEP 558 (https://mail.python.org/archives/list/python-dev@python.org/thread/7TKPMD5LHCBXGFUIMKDAUZELRH6EX76S/)
[13]Petr Viktorin's suggestion to use an enum for PyLocals_Get's behaviour (https://mail.python.org/archives/list/python-dev@python.org/message/BTQUBHIVE766RPIWLORC5ZYRCRC4CEBL/)
Source: https://github.com/python/peps/blob/master/pep-0558.rst