python / python/cpython

pypy-style constant promotion to pool

未關閉
#140,928 1 則留言 5 個 reaction 已指派 0 人 在 GitHub 檢視

還沒有人認領這個 Issue。

interpreter-core topic-JIT type-feature
主要語言
Python
星號
77.2k
分支
35.9k
PR 合併指標
PR 指標待擷取

描述

Proposal:

We should have a JIT constant pool and a JIT directive in the optimizer_bytecodes.c to promote certain values to constants. This would unlock significantly more future optimizations.

Steps needed:

  1. Add a constant pool to the executor object, so we can keep promoted constant alive. We already need this to unlock more constant propagation opportunities. The pool would be a PyListObject living in the vmdata field here https://github.com/python/cpython/blob/main/Include/internal/pycore_optimizer.h#L43 .
  2. Introduce a new optimizer macro: PROMOTE_TO_CONST_POOL and associated function promote_to_const_pool(ctx, obj) in optimizer_symbols.c. This macro will store any value it receives as a constant in the constant pool. This allows us to promote contants we see to the constant pool. This takes inspiration from RPython's hint(obj, promote=True)

What optimizations does this unlock?

  1. More constant propagation: we can now constant propagate not just immortal stuff but non-immortal things as well.
  2. Refcount removal. Anything in the the pool can use the cheaper _LOAD_CONST_INLINE_BORROW . Which will automatically benefit this work https://github.com/python/cpython/issues/134584 and vice versa. Ie, we automatically make it cheaper to load and store these values!
  3. On GIL/FT, we can promote method objects to the pool as well, which will allow us to perform more optimizations on them, such as LOAD_ATTR_METHOD folding.

We should also perhaps explore promoting globals to pool constants. Though that's a little risky, since we don't want to hold large objects around longer than needed.

The first PR should implement the first two steps and optimization 1. Optimization 2 should be a separate PR.

Linked PRs
  • gh-140968

貢獻指南

開啟貢獻指南

從這裡開始

  1. 先讀完整個 Issue,再讀專案的貢獻指南。
  2. 在 Issue 下留言說明你要接手 —— 這能避免兩個人做同樣的事。
  3. Fork 儲存庫,在一個分支上完成修改。
  4. 送出 Pull Request,並在描述裡引用這個 Issue 編號。

研究方向

先從 Include/internal/pycore_optimizer.h 和 executor 的 vmdata 欄位開始,接著檢查 optimizer_bytecodes.c 和 optimizer_symbols.c。第一個 PR 在 executor 具有 PyListObject 常數池、PROMOTE_TO_CONST_POOL 和 promote_to_const_pool 存在,且非 immortal 值獲得常數傳播時完成;移除 refcount 明確是獨立的工作。

由索引模型根據 Issue 內容生成。

評估

技術堆疊
c
領域
compilers
Issue 類型
功能
難度
5/5
預估耗時
一週以上
活躍度
停滯
描述清晰度
基本清楚
新手友好度
25/100

把新 issue 寄到你的電子郵件信箱

精選適合新手參與的 GitHub issue 摘要。