python / python/cpython

Encode immediate def-use in the JIT specializer

未关闭
#148,080 4 条评论 1 个 reaction 已指派 0 人 在 GitHub 查看

还没有人认领这个 Issue。

interpreter-core topic-JIT type-feature
主要语言
Python
星标
77.2k
派生
35.9k
PR 合并指标
PR 指标待抓取

描述

Feature or enhancement

Proposal:

Our JIT IR naturally encodes liveness in it by virtue of being a stack IR. However, we're missing a crucial piece of info to allow for stronger optimizations: def-use.

Background: Normal SSA (static single assignment form) IR already naturally encodes def-use, and LLVM's SSA flavor also includes use-def as a property.

The idea of def-use is as the name suggests, to include information of where something is defined and where it is used (hence the name def-use).

We need def-use for the following optimizations:

  1. Eliminating redundant ops after we've constant folded combined with refcount elimination. The problem is that the current refcount elimination complicates the stack effect. So naturally the current remove_unneeded_uops pass cannot do its job properly for all of those, as the ref shuffling form of uops blocks it. This issue is becoming a bigger one as we are starting to constant fold more in the JIT as we record more info, which leaves a bunch of redundant ops lying around that we could really do away with. E.g we can constant fold most LOAD_ATTR, but we leave their operand loads lying around even when no guards need it.
  2. To feed info to the future partial evaluator (PE). In my past experiments, partial evaluation is not worth it if the unboxed operand goes straight into a boxed consumer (e.g. a C API call). This mostly wrecks the perf of PE. We need a good heuristic to determine if something is worth to unbox or whether we should keep it boxed. A reminder that the PE pass is planned to take info from the current specializer pass.

Design (def-use):

typedef union {
    uintptr_t bits;
} JitOptRef;

becomes

typedef struct {
    uintptr_t bits;
    _PyUOpInstruction *originator;
} JitOptRef;

At selected uops, we then assign orginator to be the originating instruction this symbol comes from. E.g.

[0xfefefefe] LOAD_FAST x

symbol x's originator will point to 0xfefefefe (LOAD_FAST).

Not that this encodes only the most recent def-use as a property of the reference, which should be safe. We do not need full def-use chains further back, as an interesting property of a stack IR I've observed, but not yet proven, is that local rewrites can generalize to a whole-trace optimization if done correctly. E.g. see how remove_unneeded_uops just by performing local rewrites is able to remove dead code for a larger-than-local chunk of the trace.

Has this already been discussed elsewhere?

No response given

Links to previous discussion of this feature:

No response

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。

调研方向

首先阅读 JIT 专门器的 JitOptRef 设计,以及 issue 中提到的 remove_unneeded_uops pass。跟踪选定的 uops 在哪里创建或更新引用,然后确定 originator 数据如何支持冗余操作消除以及未来的部分求值输入。完成的实现应在指定的 uops 处包含 def-use 信息,同时不破坏现有的栈效果处理。

由索引模型根据 Issue 内容生成。

评估

技术栈
c, python
领域
compilers, performance
Issue 类型
功能
难度
5/5
预计耗时
一周以上
活跃度
冷清
描述清晰度
基本清楚
新手友好度
35/100

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。