CALL_FUNCTION_EX argument passing routine can be improved for vectorcalls

未关闭
#95,757 5 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
35/100
Issue 类型
重构
描述清晰度
基本清楚
活跃度
停滞
技术栈
python
领域
compilers

调研方向

追踪参数流经 do_call_core、_PyVectorcall_Call、_PyStack_UnpackDict、_PyEval_Vector、_PyEvalFramePushAndInit 和 initialize_locals 的过程。首先了解 CALL_FUNCTION_EX 如何到达 vectorcall Python 函数,以及在每个边界处如何处理所有权。完成的标志是该路径避免使用中间 C 数组并重新创建 tuple/dict,同时保持正确的参数所有权和行为。

由索引模型根据 Issue 内容生成。

描述

interpreter-core performance type-feature

While working on https://github.com/python/cpython/issues/95126, I noticed that CALL_FUNCTION_EX for vectorcall Python functions is extremely wasteful. These are its steps:

  1. CALL_FUNCTION_EX calls do_call_core with a args tuple and kwargs dict.
  2. Eventually it's detected that the function supports vectorcall, and it calls _PyVectorcall_Call. This calls _PyStack_UnpackDict, which unpacks the args tuple and kwargs dict into a C array.
  3. This C array is passed to the vectorcall function. Eventually calling _PyEval_Vector.
  4. _PyEval_Vector calls _PyEvalFramePushAndInit, which sees that there's complex args, and recreates the tuple and args dict in initialize_locals from the C array.
  5. Frame is pushed and evaluated, then popped. End.

We shouldn't even be converting the arguments to a C array just to vectorcall. We should instead directly transfer ownership of args tuple and kwds dict because that would be significantly faster in this case (and save us the tuple and dict recreation dance).

CC @markshannon and @brandtbucher .

主要语言
Python
星标
77.2k
派生
36k
平均合并
1 天 9 小时
30 天内合并 PR
558

贡献指南

打开贡献指南

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

python/cpython 的其他 Issue

查看 python/cpython 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。