Victor Stinner schrieb am 14.04.2016 um 21:56:

Which kind of usage do you see in Cython?

Mainly caching, I guess. We could avoid global/module name lookups in some cases, especially inside of loops.

Off-topic (PEP 510):

I really want to experiment automatic generation of Cython code from the Python using profiling to discover function parameters types. Then use the PEP 510 to attach the fast Cython code to a Python function, but fallback to bytecode if the types are different. See the example of builtin functions in the PEP: https://www.python.org/dev/peps/pep-0510/#using-builtin-function

Before having something fully automated, we can use some manual steps, like annotate manually function types, compile manually the code, etc.

Sounds like Cython's "Fused Types" could help here: http://docs.cython.org/src/userguide/fusedtypes.html It's essentially a generic functions implementation and you get a dispatch either at compile time or runtime, depending on where (Python/Cython) and how you call a function. Stefan

It would be nice to hear Barry Warsow who was opposed to the PEP in january. He wanted to wait until FAT Python was proven to really be faster, which is still not case right now. (I mean that I didnt't run seriously benchmarks, but early macro benchmarks are not really promising, only micro benchmarks. I expect better results when the implemenation will be more complete.) The main change since january is that Yury wrote a patch making method calls using the PEP. https://mail.python.org/pipermail/python-dev/2016-January/142772.html Victor Le jeudi 14 avril 2016, Guido van Rossum <guido@python.org> a écrit :

I'll wait a day before formally pronouncing to see if any objections are made, but it looks good to me.

On Thu, Apr 14, 2016 at 8:19 AM, Victor Stinner <victor.stinner@gmail.com <javascript:;>> wrote:

...

Hi,

I updated my PEP 509 to make the dictionary version globally unique. With *two* use cases of this PEP (Yury's method call patch and my FAT Python project), I think that the PEP is now ready to be accepted.

Globally unique identifier is a requirement for Yury's patch optimizing method calls ( https://bugs.python.org/issue26110 ). It allows to check for free if the dictionary was replaced.

I also renamed the ma_version field to ma_version_tag.

HTML version: https://www.python.org/dev/peps/pep-0509/

Victor

PEP: 509 Title: Add a private version to dict Version: $Revision$ Last-Modified: $Date$ Author: Victor Stinner <victor.stinner@gmail.com <javascript:;>> Status: Draft Type: Standards Track Content-Type: text/x-rst Created: 4-January-2016 Python-Version: 3.6

Abstract ========

Add a new private version to the builtin ``dict`` type, incremented at each dictionary creation and at each dictionary change, to implement fast guards on namespaces.

Rationale =========

In Python, the builtin ``dict`` type is used by many instructions. For example, the ``LOAD_GLOBAL`` instruction searchs for a variable in the global namespace, or in the builtins namespace (two dict lookups). Python uses ``dict`` for the builtins namespace, globals namespace, type namespaces, instance namespaces, etc. The local namespace (namespace of a function) is usually optimized to an array, but it can be a dict too.

Python is hard to optimize because almost everything is mutable: builtin functions, function code, global variables, local variables, ... can be modified at runtime. Implementing optimizations respecting the Python semantics requires to detect when "something changes": we will call these checks "guards".

The speedup of optimizations depends on the speed of guard checks. This PEP proposes to add a version to dictionaries to implement fast guards on namespaces.

Dictionary lookups can be skipped if the version does not change which is the common case for most namespaces. Since the version is globally unique, the version is also enough to check if the namespace dictionary was not replaced with a new dictionary. The performance of a guard does not depend on the number of watched dictionary entries, complexity of O(1), if the dictionary version does not change.

Example of optimization: copy the value of a global variable to function constants. This optimization requires a guard on the global variable to check if it was modified. If the variable is modified, the variable must be loaded at runtime when the function is called, instead of using the constant.

See the `PEP 510 -- Specialized functions with guards <https://www.python.org/dev/peps/pep-0510/>`_ for the concrete usage of guards to specialize functions and for the rationale on Python static optimizers.

Guard example =============

Pseudo-code of an fast guard to check if a dictionary entry was modified (created, updated or deleted) using an hypothetical ``dict_get_version(dict)`` function::

UNSET = object()

class GuardDictKey: def __init__(self, dict, key): self.dict = dict self.key = key self.value = dict.get(key, UNSET) self.version = dict_get_version(dict)

def check(self): """Return True if the dictionary entry did not changed and the dictionary was not replaced."""

# read the version of the dict structure version = dict_get_version(self.dict) if version == self.version: # Fast-path: dictionary lookup avoided return True

# lookup in the dictionary value = self.dict.get(self.key, UNSET) if value is self.value: # another key was modified: # cache the new dictionary version self.version = version return True

# the key was modified return False

Usage of the dict version =========================

Speedup method calls 1.2x -------------------------

Yury Selivanov wrote a `patch to optimize method calls <https://bugs.python.org/issue26110>`_. The patch depends on the `implement per-opcode cache in ceval <https://bugs.python.org/issue26219>`_ patch which requires dictionary versions to invalidate the cache if the globals dictionary or the builtins dictionary has been modified.

The cache also requires that the dictionary version is globally unique. It is possible to define a function in a namespace and call it in a different namespace: using ``exec()`` with the *globals* parameter for example. In this case, the globals dictionary was changed and the cache must be invalidated.

Specialized functions using guards ----------------------------------

The `PEP 510 -- Specialized functions with guards <https://www.python.org/dev/peps/pep-0510/>`_ proposes an API to support specialized functions with guards. It allows to implement static optimizers for Python without breaking the Python semantics.

Example of a static Python optimizer: the `fatoptimizer <http://fatoptimizer.readthedocs.org/>`_ of the `FAT Python <http://faster-cpython.readthedocs.org/fat_python.html>`_ project implements many optimizations which require guards on namespaces. Examples:

* Call pure builtins: to replace ``len("abc")`` with ``3``, guards on ``builtins.__dict__['len']`` and ``globals()['len']`` are required * Loop unrolling: to unroll the loop ``for i in range(...): ...``, guards on ``builtins.__dict__['range']`` and ``globals()['range']`` are required

Pyjion ------

According of Brett Cannon, one of the two main developers of Pyjion, Pyjion can also benefit from dictionary version to implement optimizations.

Pyjion is a JIT compiler for Python based upon CoreCLR (Microsoft .NET Core runtime).

Unladen Swallow ---------------

Even if dictionary version was not explicitly mentioned, optimizing globals and builtins lookup was part of the Unladen Swallow plan: "Implement one of the several proposed schemes for speeding lookups of globals and builtins." Source: `Unladen Swallow ProjectPlan <https://code.google.com/p/unladen-swallow/wiki/ProjectPlan>`_.

Unladen Swallow is a fork of CPython 2.6.1 adding a JIT compiler implemented with LLVM. The project stopped in 2011: `Unladen Swallow Retrospective <http://qinsb.blogspot.com.au/2011/03/unladen-swallow-retrospective.html `_.

Changes =======

Add a ``ma_version_tag`` field to the ``PyDictObject`` structure with the C type ``PY_INT64_T``, 64-bit unsigned integer. Add also a global dictionary version. Each time a dictionary is created, the global version is incremented and the dictionary version is initialized to the global version. The global version is also incremented and copied to the dictionary version at each dictionary change:

* ``clear()`` if the dict was non-empty * ``pop(key)`` if the key exists * ``popitem()`` if the dict is non-empty * ``setdefault(key, value)`` if the `key` does not exist * ``__detitem__(key)`` if the key exists * ``__setitem__(key, value)`` if the `key` doesn't exist or if the value is not ``dict[key]`` * ``update(...)`` if new values are different than existing values: values are compared by identity, not by their content; the version can be incremented multiple times

The ``PyDictObject`` structure is not part of the stable ABI.

The field is called ``ma_version_tag`` rather than ``ma_version`` to suggest to compare it using ``version_tag == old_version_tag`` rather than ``version <= old_version`` which makes the integer overflow much likely.

Example using an hypothetical ``dict_get_version(dict)`` function::

>>> d = {} >>> dict_get_version(d) 100 >>> d['key'] = 'value' >>> dict_get_version(d) 101 >>> d['key'] = 'new value' >>> dict_get_version(d) 102 >>> del d['key'] >>> dict_get_version(d) 103

The version is not incremented if an existing key is set to the same value. For efficiency, values are compared by their identity: ``new_value is old_value``, not by their content: ``new_value == old_value``. Example::

>>> d = {} >>> value = object() >>> d['key'] = value >>> dict_get_version(d) 40 >>> d['key'] = value >>> dict_get_version(d) 40

.. note:: CPython uses some singleton like integers in the range [-5; 257], empty tuple, empty strings, Unicode strings of a single character in the range [U+0000; U+00FF], etc. When a key is set twice to the same singleton, the version is not modified.

Implementation and Performance ==============================

The `issue #26058: PEP 509: Add ma_version_tag to PyDictObject <https://bugs.python.org/issue26058>`_ contains a patch implementing this PEP.

On pybench and timeit microbenchmarks, the patch does not seem to add any overhead on dictionary operations.

When the version does not change, ``PyDict_GetItem()`` takes 14.8 ns for a dictionary lookup, whereas a guard check only takes 3.8 ns. Moreover, a guard can watch for multiple keys. For example, for an optimization using 10 global variables in a function, 10 dictionary lookups costs 148 ns, whereas the guard still only costs 3.8 ns when the version does not change (39x as fast).

The `fat module <http://fatoptimizer.readthedocs.org/en/latest/fat.html>`_ implements such guards: ``fat.GuardDict`` is based on the dictionary version.

Integer overflow ================

The implementation uses the C type ``PY_UINT64_T`` to store the version: a 64 bits unsigned integer. The C code uses ``version++``. On integer overflow, the version is wrapped to ``0`` (and then continue to be incremented) according to the C standard.

After an integer overflow, a guard can succeed whereas the watched dictionary key was modified. The bug only occurs at a guard check if there are exaclty ``2 ** 64`` dictionary creations or modifications since the previous guard check.

If a dictionary is modified every nanosecond, ``2 ** 64`` modifications takes longer than 584 years. Using a 32-bit version, it only takes 4 seconds. That's why a 64-bit unsigned type is also used on 32-bit systems. A dictionary lookup at the C level takes 14.8 ns.

A risk of a bug every 584 years is acceptable.

Alternatives ============

Expose the version at Python level as a read-only __version__ property ----------------------------------------------------------------------

The first version of the PEP proposed to expose the dictionary version as a read-only ``__version__`` property at Python level, and also to add the property to ``collections.UserDict`` (since this type must mimick the ``dict`` API).

There are multiple issues:

* To be consistent and avoid bad surprises, the version must be added to all mapping types. Implementing a new mapping type would require extra work for no benefit, since the version is only required on the ``dict`` type in practice. * All Python implementations must implement this new property, it gives more work to other implementations, whereas they may not use the dictionary version at all. * Exposing the dictionary version at Python level can lead the false assumption on performances. Checking ``dict.__version__`` at the Python level is not faster than a dictionary lookup. A dictionary lookup has a cost of 48.7 ns and checking a guard has a cost of 47.5 ns, the difference is only 1.2 ns (3%)::

$ ./python -m timeit -s 'd = {str(i):i for i in range(100)}' 'd["33"] == 33' 10000000 loops, best of 3: 0.0487 usec per loop $ ./python -m timeit -s 'd = {str(i):i for i in range(100)}' 'd.__version__ == 100' 10000000 loops, best of 3: 0.0475 usec per loop

* The ``__version__`` can be wrapped on integer overflow. It is error prone: using ``dict.__version__ <= guard_version`` is wrong, ``dict.__version__ == guard_version`` must be used instead to reduce the risk of bug on integer overflow (even if the integer overflow is unlikely in practice).

Mandatory bikeshedding on the property name:

* ``__cache_token__``: name proposed by Nick Coghlan, name coming from `abc.get_cache_token() <https://docs.python.org/3/library/abc.html#abc.get_cache_token>`_. * ``__version__`` * ``__timestamp__``

Add a version to each dict entry --------------------------------

A single version per dictionary requires to keep a strong reference to the value which can keep the value alive longer than expected. If we add also a version per dictionary entry, the guard can only store the entry version to avoid the strong reference to the value (only strong references to the dictionary and to the key are needed).

Changes: add a ``me_version`` field to the ``PyDictKeyEntry`` structure, the field has the C type ``PY_INT64_T``. When a key is created or modified, the entry version is set to the dictionary version which is incremented at any change (create, modify, delete).

Pseudo-code of an fast guard to check if a dictionary key was modified using hypothetical ``dict_get_version(dict)`` ``dict_get_entry_version(dict)`` functions::

UNSET = object()

class GuardDictKey: def __init__(self, dict, key): self.dict = dict self.key = key self.dict_version = dict_get_version(dict) self.entry_version = dict_get_entry_version(dict, key)

def check(self): """Return True if the dictionary entry did not changed and the dictionary was not replaced."""

# read the version of the dict structure dict_version = dict_get_version(self.dict) if dict_version == self.version: # Fast-path: dictionary lookup avoided return True

# lookup in the dictionary entry_version = get_dict_key_version(dict, key) if entry_version == self.entry_version: # another key was modified: # cache the new dictionary version self.dict_version = dict_version return True

# the key was modified return False

The main drawback of this option is the impact on the memory footprint. It increases the size of each dictionary entry, so the overhead depends on the number of buckets (dictionary entries, used or unused yet). For example, it increases the size of each dictionary entry by 8 bytes on 64-bit system.

In Python, the memory footprint matters and the trend is to reduce it. Examples:

* `PEP 393 -- Flexible String Representation <https://www.python.org/dev/peps/pep-0393/>`_ * `PEP 412 -- Key-Sharing Dictionary <https://www.python.org/dev/peps/pep-0412/>`_

Add a new dict subtype ----------------------

Add a new ``verdict`` type, subtype of ``dict``. When guards are needed, use the ``verdict`` for namespaces (module namespace, type namespace, instance namespace, etc.) instead of ``dict``.

Leave the ``dict`` type unchanged to not add any overhead (memory footprint) when guards are not needed.

Technical issue: a lot of C code in the wild, including CPython core, expecting the exact ``dict`` type. Issues:

* ``exec()`` requires a ``dict`` for globals and locals. A lot of code use ``globals={}``. It is not possible to cast the ``dict`` to a ``dict`` subtype because the caller expects the ``globals`` parameter to be modified (``dict`` is mutable). * Functions call directly ``PyDict_xxx()`` functions, instead of calling ``PyObject_xxx()`` if the object is a ``dict`` subtype * ``PyDict_CheckExact()`` check fails on ``dict`` subtype, whereas some functions require the exact ``dict`` type. * ``Python/ceval.c`` does not completely supports dict subtypes for namespaces

The ``exec()`` issue is a blocker issue.

Other issues:

* The garbage collector has a special code to "untrack" ``dict`` instances. If a ``dict`` subtype is used for namespaces, the garbage collector can be unable to break some reference cycles. * Some functions have a fast-path for ``dict`` which would not be taken for ``dict`` subtypes, and so it would make Python a little bit slower.

Prior Art =========

Method cache and type version tag ---------------------------------

In 2007, Armin Rigo wrote a patch to to implement a cache of methods. It was merged into Python 2.6. The patch adds a "type attribute cache version tag" (``tp_version_tag``) and a "valid version tag" flag to types (the ``PyTypeObject`` structure).

The type version tag is not available at the Python level.

The version tag has the C type ``unsigned int``. The cache is a global hash table of 4096 entries, shared by all types. The cache is global to "make it fast, have a deterministic and low memory footprint, and be easy to invalidate". Each cache entry has a version tag. A global version tag is used to create the next version tag, it also has the C type ``unsigned int``.

By default, a type has its "valid version tag" flag cleared to indicate that the version tag is invalid. When the first method of the type is cached, the version tag and the "valid version tag" flag are set. When a type is modified, the "valid version tag" flag of the type and its subclasses is cleared. Later, when a cache entry of these types is used, the entry is removed because its version tag is outdated.

On integer overflow, the whole cache is cleared and the global version tag is reset to ``0``.

See `Method cache (issue #1685986) <https://bugs.python.org/issue1685986>`_ and `Armin's method cache optimization updated for Python 2.6 (issue #1700288) <https://bugs.python.org/issue1700288>`_.

Globals / builtins cache ------------------------

In 2010, Antoine Pitrou proposed a `Globals / builtins cache (issue #10401) <http://bugs.python.org/issue10401>`_ which adds a private ``ma_version`` field to the ``PyDictObject`` structure (``dict`` type), the field has the C type ``Py_ssize_t``.

The patch adds a "global and builtin cache" to functions and frames, and changes ``LOAD_GLOBAL`` and ``STORE_GLOBAL`` instructions to use the cache.

The change on the ``PyDictObject`` structure is very similar to this PEP.

Cached globals+builtins lookup ------------------------------

In 2006, Andrea Griffini proposed a patch implementing a `Cached globals+builtins lookup optimization <https://bugs.python.org/issue1616125>`_. The patch adds a private ``timestamp`` field to the ``PyDictObject`` structure (``dict`` type), the field has the C type ``size_t``.

Thread on python-dev: `About dictionary lookup caching <https://mail.python.org/pipermail/python-dev/2006-December/070348.html `_.

Guard against changing dict during iteration --------------------------------------------

In 2013, Serhiy Storchaka proposed `Guard against changing dict during iteration (issue #19332) <https://bugs.python.org/issue19332>`_ which adds a ``ma_count`` field to the ``PyDictObject`` structure (``dict`` type), the field has the C type ``size_t``. This field is incremented when the dictionary is modified, and so is very similar to the proposed dictionary version.

Sadly, the dictionary version proposed in this PEP doesn't help to detect dictionary mutation. The dictionary version changes when values are replaced, whereas modifying dictionary values while iterating on dictionary keys is legit in Python.

PySizer -------

`PySizer <http://pysizer.8325.org/>`_: a memory profiler for Python, Google Summer of Code 2005 project by Nick Smallbone.

This project has a patch for CPython 2.4 which adds ``key_time`` and ``value_time`` fields to dictionary entries. It uses a global process-wide counter for dictionaries, incremented each time that a dictionary is modified. The times are used to decide when child objects first appeared in their parent objects.

Discussion ==========

Thread on the mailing lists:

* python-dev: `PEP 509: Add a private version to dict <https://mail.python.org/pipermail/python-dev/2016-January/142685.html `_ (january 2016) * python-ideas: `RFC: PEP: Add dict.__version__ < https://mail.python.org/pipermail/python-ideas/2016-January/037702.html>`_ (january 2016)

Copyright =========

This document has been placed in the public domain. _______________________________________________ Python-Dev mailing list Python-Dev@python.org <javascript:;> https://mail.python.org/mailman/listinfo/python-dev Unsubscribe: https://mail.python.org/mailman/options/python-dev/guido%40python.org

-- --Guido van Rossum (python.org/~guido)

2016-04-14 22:50 GMT+02:00 Barry Warsaw <barry@python.org>:

Although I'm not totally convinced, I won't continue to object. You've provided some performance numbers in the PEP even without FAT, and you aren't exposing the API to Python, so it's not a burden being imposed on other implementations.

Cool! Ah right, the PEP evolved since its first version sent to python-ideas. I didn't recall the full context of the discussion. The PEP is now more complete and it has more known (future) use cases ;-) (now maybe also Cython?) Victor

From Armin Rigo Sent: Thursday, April 14, 2016 4:42 PM To: Victor Stinner Cc: Python Dev Subject: Re: [Python-Dev] RFC: PEP 509: Add a private version to dict

Hi Victor,

On 14 April 2016 at 17:19, Victor Stinner <victor.stinner@gmail.com> wrote:

...

Each time a dictionary is created, the global version is incremented and the dictionary version is initialized to the global version.

A detail, but why not set the version tag of new empty dictionaries to zero, always? Same after a clear(). This would satisfy the condition: equality of the version tag is supposed to mean "the dictionary content is precisely the same".

From Victor's original post:

"Globally unique identifier is a requirement for Yury's patch optimizing method calls ( https://bugs.python.org/issue26110 ). It allows to check for free if the dictionary was replaced." I think it's a good design idea, and there's no chance that this counter will ever overflow (I think Victor is using 64-bit unsigned integer). I don't think there's really any drawback to using a global vs per-dict counter (but Victor is better placed to answer that :)) -Emanuel ~Ducks lay where no programmer has ever been~

On Apr 14, 2016, at 11:17 PM, Victor Stinner wrote:

You're right that incrementing the global version is useless for these specific cases, and using the version 0 should work. It only matters that the version (version? version tag?) is different.

I will play with that. If I don't see any issue, I will update the PEP.

It's more an implementation detail, but it may help to mention it in the PEP.

I can see why you might want a global version number, but not doing so would eliminate an implicit reliance on the GIL, or in a GIL-less implementation <wink> a lock around incrementing the global version number. -Barry

On Thu, 14 Apr 2016 at 15:14 Victor Stinner <victor.stinner@gmail.com> wrote:

2016-04-14 23:29 GMT+02:00 Barry Warsaw <barry@python.org>:

...

I can see why you might want a global version number, but not doing so would eliminate an implicit reliance on the GIL, or in a GIL-less implementation <wink> a lock around incrementing the global version number.

It's not like the builtin dict type is going to become GIL-free... So I think that it's ok to use a global version.

A very few know that, but the GIL has some advantages sometimes...

And even if it was GIL-free you do run the risk of two dicts ending up at the same version # by simply mutating the same number of times if the counters were per-dict instead of process-wide.

On 4/14/2016 3:33 PM, Victor Stinner wrote:

When we will be able to get ride of the GIL for the dict type, we will probably be able to get an atomic "global_version++" for 64-bit integer. Right now, I don't think that an atomic int64++ is available on 32-bit archs. By the time we get an atomic increment for 64-bit integer, we'll be wanting it for 128-bit...

Le vendredi 15 avril 2016, Stefan Behnel <stefan_ml@behnel.de> a écrit :

How can that be achieved? If the tag is just a sequentially growing number, creating two dicts and applying one operation to the first one should give both the same version tag, right?

Armin didn't propose to get ride of the global version. a = dict() # version = 0 b = dict() # version = 0 a['key'] = 'value' # version = 300 b['key'] = 'value' # version = 301 Victor PS: It looks like the iPad Gmail app foces me to use HTML, I don't know how to use plain text :-/

Victor Stinner <victor.stinner <at> gmail.com> writes:

Hi,

2016-04-14 22:42 GMT+02:00 Armin Rigo <arigo <at> tunes.org>:

...

Hi Victor,

On 14 April 2016 at 17:19, Victor Stinner <victor.stinner <at>

gmail.com> wrote:

...

...

Each time a dictionary is created, the global version is incremented and the dictionary version is initialized to the global version.

A detail, but why not set the version tag of new empty dictionaries to zero, always? Same after a clear(). This would satisfy the condition: equality of the version tag is supposed to mean "the dictionary content is precisely the same".

You're right that incrementing the global version is useless for these specific cases, and using the version 0 should work. It only matters that the version (version? version tag?) is different.

Why do this? It's a nice property that two dicts always have different version tags, and now you're killing this property for... no obvious reason? Do you really think dict.clear() is in need of micro-optimizing a couple CPU cycles away? Regards Antoine.

Hi, FYI I updated the implementation of the PEP 509: https://bugs.python.org/issue26058 2016-04-15 11:01 GMT+02:00 Antoine Pitrou <antoine@python.org>:

Why do this? It's a nice property that two dicts always have different version tags, and now you're killing this property for... no obvious reason?

Do you really think dict.clear() is in need of micro-optimizing a couple CPU cycles away?

So, I played with Armin's idea. I confirm that it works for my use case, guards on dict keys. It should also work on Yury's use case. Antoine is right, it's really a micro-optimization. It shouldn't help much for the integer overflow (which is not a real issue in practice). I propose to leave the PEP unchanged to keep the nice property of unique identifier for empty dictionaries. It can help for future use cases. Victor

On 2016-04-14 4:42 PM, Armin Rigo wrote:

Hi Victor,

On 14 April 2016 at 17:19, Victor Stinner <victor.stinner@gmail.com> wrote:

...

Each time a dictionary is created, the global version is incremented and the dictionary version is initialized to the global version. A detail, but why not set the version tag of new empty dictionaries to zero, always? Same after a clear(). This would satisfy the condition: equality of the version tag is supposed to mean "the dictionary content is precisely the same".

So {}.version_tag == {}.version_tag == 0 {'a':1}.version_tag != {'a':1}.version_tag right? For my patches I need globally unique version tags (making an exception for empty dicts is OK). Yury

On Thu, Apr 14, 2016, 17:14 MRAB <python@mrabarnett.plus.com> wrote:

On 2016-04-14 21:42, Armin Rigo wrote:

...

Hi Victor,

On 14 April 2016 at 17:19, Victor Stinner <victor.stinner@gmail.com> wrote:

...

Each time a dictionary is created, the global version is incremented and the dictionary version is initialized to the global version.

A detail, but why not set the version tag of new empty dictionaries to zero, always? Same after a clear(). This would satisfy the condition: equality of the version tag is supposed to mean "the dictionary content is precisely the same".

If you did that, wouldn't it then be possible to replace an empty dict with another empty dict with you noticing?

If you meant to say "without" then yes. Would that matter?

Nope because this is about versioining content, so having identical/empty content compare equal is fine. -brett

_______________________________________________ Python-Dev mailing list Python-Dev@python.org https://mail.python.org/mailman/listinfo/python-dev Unsubscribe: https://mail.python.org/mailman/options/python-dev/brett%40python.org

On 15 April 2016 at 18:54, Jim J. Jewett <jimjjewett@gmail.com> wrote:

[2A] Do you want to promise that replacing a value with a non-identical object *will* trigger a version_tag update *even* if the objects are equal?

I would vote no, but I realize backwards-compatibility may create such a promise implicitly.

It needs to trigger a version update. Equality doesn't guarantee any kind of equivalence in Python. It's not even guaranteed that a==b will come to the same value if evaluated twice in a row. An example:

...

...

from fractions import Fraction as F F(1) == 1 True d = globals() d['a'] = F(1) a.limit_denominator() Fraction(1, 1) d['a'] = 1 a.limit_denominator() Traceback (most recent call last): File "<stdin>", line 1, in <module> AttributeError: 'int' object has no attribute 'limit_denominator'

-- Oscar

On Fri, Apr 15, 2016, at 16:41, Victor Stinner wrote:

If the dictionary values are modified during the loop, the dict version is increased. But it's allowed to modify values when you iterate on *keys*.

Why is iterating over items different from iterating over keys? in other words, why do I have to write: for k in dict: v = dict[k] ...do some stuff... dict[k] = something rather than for k, v in dict.items(): ...do some stuff... dict[k] = something It's not clear why the latter is something you want to prevent.

On 04/15/2016 01:41 PM, Victor Stinner wrote:

2016-04-15 19:54 GMT+02:00 Jim J. Jewett:

...

(2) Why *promise* not to update the version_tag when replacing a value with itself?

It's an useful property. For example, let's say that you have a guard on globals()['value']. The guard is created with value=3. An unit test replaces the value with 50, but then restore the value to its previous value (3). Later, the guard is checked to decide if an optimization can be used.

I don't understand -- shouldn't the version be incremented with the value was replaced with 50, and again when re-replaced with 3?

...

(6) I'm also not sure why version_tag *doesn't* solve the problem of dicts that fool the iteration guards by mutating without changing size ( https://bugs.python.org/issue19332 ) ... are you just saying that the iterator views aren't allowed to rely on the version-tag remaining stable, because replacing a value (as opposed to a key-value pair) is allowed?

If the dictionary values are modified during the loop, the dict version is increased. But it's allowed to modify values when you iterate on *keys*.

I don't understand. Could you provide a small example? -- ~Ethan~

On Fri, Apr 15, 2016 at 4:41 PM, Victor Stinner <victor.stinner@gmail.com> wrote:

2016-04-15 19:54 GMT+02:00 Jim J. Jewett <jimjjewett@gmail.com>:

...

(2) Why *promise* not to update the version_tag when replacing a value with itself?

It's an useful property. For example, let's say that you have a guard on globals()['value']. The guard is created with value=3. An unit test replaces the value with 50, but then restore the value to its previous value (3). Later, the guard is checked to decide if an optimization can be used.

If the dict version is increased, you need a lookup. If the dict version is not increased, the guard is cheap.

I would expect the version to be increased twice, and therefore to require a lookup. Are you suggesting that unittest should provide an example of resetting the version back to the original value when it cleans up after itself?

In C, it's very cheap to implement the test "new_value == old_value", it just compares two pointers.

Yeah, I understand that it is likely a win in terms of performance, and a good way to start off (given that you're willing to do the work). I just worry that you may end up closing off even better optimizations later, if you make too many promises about exactly how you will do which ones. Today, dict only cares about ==, and you (reasonably) think that full == isn't always worth running ... but when it comes to which tests *are* worth running, I'm not confident that the answers won't change over the years.

...

[2A] Do you want to promise that replacing a value with a non-identical object *will* trigger a version_tag update *even* if the objects are equal?

It's already written in the PEP:

I read that as a description of what the code does, rather than a spec for what it should do... so it isn't clear whether I could count on that remaining true. For example, if I know that my dict values are all 4-digit integers, can I write: d[k] = d[k] + 0 and be assured that the version_tag will bump? Or is that something that a future optimizer might optimize out?

...

(3) It is worth being explicit on whether empty dicts can share a version_tag of 0. If this PEP is about dict content, then that seems fine, and it may well be worth optimizing dict creation.

This is not part of the PEP yet. I'm not sure that I will modify the PEP to use the version 0 for empty dictionaries. Antoine doesn't seem to be convinced :-)

True. But do note that "not hitting the global counter an extra time for every dict creation" is a more compelling reason than "we could speed up dict.clear(), sometimes".

...

(4) Please be explicit about the locking around version++; it is enough to say that the relevant methods already need to hold the GIL (assuming that is true).

I don't think that it's important to mention it in the PEP. It's more an implementation detail. The version can be protected by atomic operations.

Now I'm the one arguing from a specific implementation. :D My thought was that any sort of locking (including atomic operations) is slow, but if the GIL is already held, then there is no *extra* locking cost. (Well, a slightly longer hold on the lock, but...)

...

(5) I'm not sure I understand the arguments around a per-entry version.

...

On the one hand, you never need a strong reference to the value; if it has been collected, then it has obviously been removed from the dict and should trigger a change even with per-dict.

Let's say that you watch the key1 of a dict. The key2 is modified, it increases the version. Later, you test the guard: to check if the key1 was modified, you need to lookup the key and compare the value. You need the value to compare it.

And the value for key1 is still there, so you can. The only reason you would notice that the key2 value had gone away is if you also care about key2 -- in which case the cached value is out of date, regardless of what specific value it used to hold.

...

(6) I'm also not sure why version_tag *doesn't* solve the problem of dicts that fool the iteration guards by mutating without changing size ( https://bugs.python.org/issue19332 ) ... are you just saying that the iterator views aren't allowed to rely on the version-tag remaining stable, because replacing a value (as opposed to a key-value pair) is allowed?

If the dictionary values are modified during the loop, the dict version is increased. But it's allowed to modify values when you iterate on *keys*.

Sure. So? I see three cases: (A) I don't care that the collection changed. The python implementation might, but I don't. (So no bug even today.) (B) I want to process exactly the collection that I started with. If some of the values get replaced, then I want to complain, even if python doesn't. version_tag is what I want. (C) I want to process exactly the original keys, but go ahead and use updated values. The bug still bites, but ... I don't think this case is any more common than B. -jJ

On Fri, Apr 15, 2016 at 7:31 PM, Victor Stinner <victor.stinner@gmail.com> wrote:

.2016-04-15 23:45 GMT+02:00 Jim J. Jewett <jimjjewett@gmail.com>: ...

...

I just worry that you may end up closing off even better optimizations later, if you make too many promises about exactly how you will do which ones.

...

Today, dict only cares about ==, and you (reasonably) think that full == isn't always worth running ... but when it comes to which tests *are* worth running, I'm not confident that the answers won't change over the years.

I checked, currently there is no unit test for a==b, only for a is b. I will add add a test for a==b but a is not b, and ensure that the version is increased.

Again, why? Why not just say "If an object is replaced by something equal to itself, the version_tag may not be changed. While the initial heuristics are simply to check for identity but not full equality, this may change in future releases."

...

For example, if I know that my dict values are all 4-digit integers, can I write:

d[k] = d[k] + 0

and be assured that the version_tag will bump? Or is that something that a future optimizer might optimize out?

Hum, I will try to clarify that.

I would prefer that you clarify it to say that while the initial patch doesn't optimize that out, a future optimizer might.

The problem with storing an identifier (a pointer in C) with no strong reference is when the object is destroyed, a new object can likely get the same identifier. So it's likely that "dict[key] is old_value_id" can be true even if dict[key] is now a new object.

Yes, but it shouldn't actually be destroyed until it is removed from the dict, which should change version_tag, so that there will be no need to compare it.

Do you want to modify the PEP 509 to fix this issue? Or you don't understand why the PEP 509 cannot be used to fix the issue? I'm lost...

I believe it *does* fix the issue in some (but not all) cases. -jJ