python / python/cpython

Specialize calls to inherited `list.append` on list subclasses

オープン
#155,456 コメント 4 件 リアクション 0 件 担当者 0 名 GitHub で見る

まだ誰も着手していません。

interpreter-core pending performance type-feature
主要言語
Python
スター
77.2k
フォーク
35.9k
PR マージ指標
PR 指標を取得中

説明

Feature or enhancement

Proposal:

I was reading the code around CALL_LIST_APPEND and noticed a possible optimization for list subclasses that inherit list.append without overriding it.

I am wondering whether inherited list.append calls could be supported deliberately, with matching specialization time and runtime checks.

For example:

class MyList(list):
    pass

def append_one(values):
    values.append(1)

For an exact list, this call can specialize to CALL_LIST_APPEND. For MyList, it currently specializes to CALL_METHOD_DESCRIPTOR_O, even though attribute lookup still resolves to the built-in list.append.


The specialization is made in specialize_method_descriptor().

It checks that the receiver is an exact list:

if ((PyObject *)descr == list_append && oparg == 1) {
    assert(self_or_null != NULL);
    if (PyList_CheckExact(self_or_null)) {
        specialize(instr, CALL_LIST_APPEND);
        return 0;
    }
}

and CALL_LIST_APPEND macro is defined as follows:

macro(CALL_LIST_APPEND) =
    unused/1 +
    unused/2 +
    _GUARD_CALLABLE_LIST_APPEND +
    _GUARD_NOS_NOT_NULL +
    _GUARD_NOS_LIST +
    _CALL_LIST_APPEND +
    POP_TOP +
    POP_TOP;

As far as I understand, this should exclude subclasses that shadow append.

But, _CALL_LIST_APPEND ultimately uses _PyList_AppendTakeRef, which accepts list subclasses through PyList_Check, just like the generic list.append path. So the underlying append operation already appears to support subclasses.

The remaining problem is ensuring that the specialization and runtime guards admit them safely.

Possible approach

The safer structure seems to be leaving the existing instruction unchanged and adding a separate specialization for list subclasses such as CALL_LIST_APPEND_SUBTYPE

The specializer could select the existing instruction for exact lists and the new instruction for subclasses:

if (PyList_CheckExact(self_or_null)) {
    specialize(instr, CALL_LIST_APPEND);
}
else if (PyList_Check(self_or_null)) {
    specialize(instr, CALL_LIST_APPEND_SUBTYPE);
}

Both instructions could then share _CALL_LIST_APPEND while keeping their receiver assumptions separate:

macro(CALL_LIST_APPEND) =
    unused/1 +
    unused/2 +
    _GUARD_CALLABLE_LIST_APPEND +
    _GUARD_NOS_NOT_NULL +
    _GUARD_NOS_LIST +
    _CALL_LIST_APPEND +
    POP_TOP +
    POP_TOP;

macro(CALL_LIST_APPEND_SUBTYPE) =
    unused/1 +
    unused/2 +
    _GUARD_CALLABLE_LIST_APPEND +
    _GUARD_NOS_NOT_NULL +
    _GUARD_NOS_LIST_SUBTYPE +
    _CALL_LIST_APPEND +
    POP_TOP +
    POP_TOP;
op(_GUARD_NOS_LIST_SUBTYPE, (self -- self)) {
    EXIT_IF(!PyList_Check(self));
}

This keeps the meaning of CALL_LIST_APPEND unchanged, while isolating the new subtype behavior.

Alternative approach

A smaller alternative would be to let the existing CALL_LIST_APPEND instruction accept list subclasses:

- if (PyList_CheckExact(self_or_null)) {
+ if (PyList_Check(self_or_null)) {
      specialize(instr, CALL_LIST_APPEND);
  }

and replace its exact list guard:

 macro(CALL_LIST_APPEND) =
     unused/1 +
     unused/2 +
     _GUARD_CALLABLE_LIST_APPEND +
     _GUARD_NOS_NOT_NULL +
-    _GUARD_NOS_LIST +
+    _GUARD_NOS_LIST_SUBTYPE +
     _CALL_LIST_APPEND +
     POP_TOP +
     POP_TOP;

It is a smaller code change, but it broadens the meaning of an existing instruction. I have not verified all assumptions outside, so I am not sure that changing the existing instruction is the right approach.


Is there any semantic or implementation reason why a list subclass inheriting the built-in list.append should not reuse the _CALL_LIST_APPEND fast path?

If the optimization is worthwhile, would it be preferable to extend the existing instruction, or to keep the new behavior isolated in a separate specialization?

Has this already been discussed elsewhere?

No response given

Links to previous discussion of this feature:

https://github.com/python/cpython/issues/141367

Linked PRs
  • gh-155474

コントリビューションガイド

コントリビューションガイドを開く

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

調査の方向性

Python/specialize.c の specialize_method_descriptor() から始め、CALL_LIST_APPEND、その guards、_CALL_LIST_APPEND を追跡します。これまでの議論について、リンクされた issue 141367 と PR gh-155474 を読んでください。list のサブクラスにおける継承された list.append の特殊化について、安全性と適用範囲が解決され、検証されれば完了です。

索引モデルが issue の本文から書いたものです。

評価

技術スタック
c, python
領域
compilers, performance
issue の種類
機能追加
難易度
5/5
見積もり時間
1週間以上
活発さ
停滞
明瞭さ
おおむね明確
初心者へのやさしさ
30/100

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。