perf: split BytecodeInterpreter.execute for default C2 compilation

未關閉
#2,093 0 則留言 0 個 reaction 已指派 0 人 在 GitHub 檢視

維護者通常 1 天內回覆

還沒有人認領這個 Issue。

評估

難度
5/5
預估耗時
一週以上
新手友好度
25/100
Issue 類型
重構
描述清晰度
描述清楚
活躍度
活躍
技術堆疊
java

研究方向

Start by reading BytecodeInterpreter.execute(SuspendedInterpreterFrame) and the profiling skill at .agents/skills/profile-perlonjava/SKILL.md; use dev/bench/benchmark_eval_string.pl to reproduce the eval STRING workload. Split the dispatch path while preserving opcode behavior and JVM/interpreter parity, then check HotSpot diagnostics, project tests, and the make gate. Compare randomized paired runs against the exact parent and check representative workloads and startup cost; done means default flags produce effective C2 compilation with a measured improvement and no material regression.

由索引模型根據 Issue 內容生成。

描述

area:backend enhancement

Summary

Split BytecodeInterpreter.execute(SuspendedInterpreterFrame) into smaller dispatch/handler methods so HotSpot's default C2 policy can optimize the eval STRING interpreter path without globally enabling compilation of huge methods.

This is follow-up work to eval performance investigations on master. PR #2059 contains allocation and frame-pooling changes, but its performance effect was small and did not address the dispatch method's JIT compilation limit.

Evidence

On master commit 3e725aa6e1 (Temurin 24.0.2, Apple Silicon), async-profiler sampling of dev/bench/benchmark_eval_string.pl showed BytecodeInterpreter.execute as the largest CPU leaf (199 of 1,694 samples, C1), followed by StatementResolver.parseStatement (59) and CompileAssignment.compileAssignmentOperator (48).

HotSpot compilation diagnostics report the private execute(SuspendedInterpreterFrame) method at 23,790 bytecodes. The default DontCompileHugeMethods policy prevents C2 from optimizing this hot dispatch path; the log also reports it as too large to inline.

As a diagnostic, disabling that policy with -XX:-DontCompileHugeMethods caused the 23,790-byte method to reach C2 level 4. Five randomized default/flag pairs on the same master JAR measured:

Default C2 enabled Speedup
52.4k/s 102.7k/s 1.96x
60.7k/s 98.1k/s 1.62x
65.7k/s 98.3k/s 1.50x
71.2k/s 106.9k/s 1.50x
73.2k/s 114.3k/s 1.56x

Each run executes one million iterations of dev/bench/benchmark_eval_string.pl. The geometric mean paired gain is about 1.60x. Host load caused variance in both modes; all five pairs favored C2.

Goal

Reduce the dispatch method's bytecode size enough that the normal JVM policy compiles its hot path effectively. Prefer extracting cold opcode families and keeping common eval opcodes on a direct, JIT-friendly path. Keep opcode semantics, exception behavior, interpreter state, and JVM/interpreter parity unchanged.

The global -XX:-DontCompileHugeMethods flag is a measured interim option, but this issue is to determine whether a source-level split can get the same or better result without changing HotSpot's huge-method policy for every method in the process.

Acceptance

  • HotSpot diagnostics on supported Java versions show the refactored hot interpreter path gets C2 compilation with default JVM flags.
  • Standard Perl behavior and project-owned tests cover affected opcode families; JVM and interpreter backends remain compatible.
  • Run the repository make gate on the immutable candidate.
  • Compare default-flag performance against the exact parent using randomized paired eval STRING benchmark runs with operation counts/checksums; retain the split only when the paired interval shows a real improvement.
  • Check representative non-eval workloads and startup/compilation cost so a faster eval benchmark does not hide a material regression elsewhere.

Related

主要語言
Perl
星號
67
分支
8
平均合併
5 小時 12 分鐘
30 天內合併 PR
131

環境準備

  • 提供 Dockerfile 或 Docker Compose 檔案
  • 沒有 Pull Request 範本
  • 閱讀貢獻指南

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。

fglock/PerlOnJava 的其他 Issue

查看 fglock/PerlOnJava 的全部 Issue

相似的 Issue

更多 Perl Issue

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。