Fix Clone::clone graph aliases and DESTROY timing

未关闭
#2,159 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

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
28/100
Issue 类型
缺陷
描述清晰度
基本清楚
活跃度
活跃
技术栈
java, perl
领域
backend

调研方向

Start with the REFERENCE, HASHREFERENCE and ARRAYREFERENCE branches of deepClone in the Java Clone implementation, and check that src/main/perl/lib/Clone.pm loads it. Run the graph probe (t/dump.pl) and the unblessed-hash DESTROY probe on both backends to reproduce. Done means the shared-referent and cycle probe matches system Perl, cloned objects are destroyed before END, and t/dclone.t and t/06refcnt.t pass. Coordinate owner-tracking changes with #2043.

由索引模型根据 Issue 内容生成。

描述

area:cpan-port area:memory area:runtime bug

Summary

Clone::Any 1.01 fails under PerlOnJava because its preferred Clone::clone provider does not preserve a scalar-reference graph and does not release a cloned object at lexical scope exit. Both failures reproduce on the JVM and bytecode interpreter; system Perl passes the distribution suite and the focused reproductions.

CPAN random tester run: 20261120-195805-13340 (Clone-Any--from--Clone-Any). The archived run failed t/06refcnt.t and t/dclone.t. The distribution is pure Perl and has no XS/native prerequisite. Its required Devel::UseAnyFunc dependency is pure Perl.

Reproduction

With Clone-Any 1.01 and its dependency installed, run the graph probe from the distribution root:

use Clone::Any qw(clone);
require './t/dump.pl';

my $value = 'x';
my $shared = \$value;
my $input = [$shared, \$value];
my $copy = clone($input);

print main::dump($input) eq main::dump($copy) ? "same\n" : "different\n";

Expected output is same. Perl 5.42.2 prints same; PerlOnJava prints different on both backends. This is a reduced form of t/dclone.t test 4.

The lifetime failure reduces to cloning an unblessed hash, then blessing both values into a class with DESTROY before leaving scope:

use Clone::Any qw(clone);
package Probe;
sub DESTROY { my $self = shift; print "DESTROY $self->{name}\n" }
package main;
{
    my $source = { name => 'source' };
    my $copy = clone($source);
    $copy->{name} = 'copy';
    bless $source, 'Probe';
    bless $copy, 'Probe';
}
print "END\n";

Expected system-Perl ordering is both DESTROY lines before END. PerlOnJava destroys source before END but delays copy until global destruction. The same pattern accounts for t/06refcnt.t tests 6–9, which fail on both PerlOnJava backends.

Cause

Clone::Any tries Clone::clone before its other providers. PerlOnJava's src/main/perl/lib/Clone.pm sees Clone::clone as defined after XSLoader::load, so it does not load Clone::PP; the active provider is the Java org.perlonjava.runtime.perlmodule.Clone implementation. On the reproducing launcher, Clone/PP.pm is absent from %INC.

Two concrete defects are visible in Clone.java:

  1. In the REFERENCE branch, deepClone stores and looks up a clone using scalar.value, which is the referent node, and maps it to a reference wrapper. Repeated references to that referent can therefore return the same wrapper for distinct input reference nodes. The reduced graph probe demonstrates the resulting structural mismatch. The cache needs to preserve Perl reference graph edges while still reconnecting cycles and shared referents.
  2. The HASHREFERENCE and ARRAYREFERENCE branches create fresh clones with newHash.createReference() and newArray.createReference(). Those methods set localBindingExists for named aggregate references. Fresh clone aggregates have no named backing variable; the corresponding createAnonymousReference() methods deliberately leave that flag clear so DESTROY can run when the last owner is released. A focused scope probe confirms the cloned hash's DESTROY is delayed until global destruction.

The ARRAYREFERENCE branch also appends cloned elements directly to newArray.elements, unlike the HASHREFERENCE branch's newHash.put(...). Audit this path for owner-slot acquisition and release so cloned arrays preserve element lifetimes as well as graph structure.

Evidence

  • System Perl 5.42.2: full suite passes, 7 files / 40 tests. PERL5OPT=-I. is required for this 2003 test suite's legacy require 't/…' paths under modern Perl.
  • PerlOnJava JVM and interpreter: t/dclone.t fails test 4; t/06refcnt.t fails tests 6–9. Both focused probes above reproduce on both backends.
  • The focused PerlOnJava runs used launcher build b9c9a5f54d0dfa6367c85148d064d9f6d1fa8a89 on Darwin arm64, Java 24.0.2. The repository checkout was at 506371c13.
  • Duplicate search found no existing Clone-specific issue. The ownership behavior is adjacent to the active work in #2043; coordinate changes to owner tracking with that effort, while keeping this graph-cloning defect independently testable.

Acceptance criteria

  • Add project-owned system-Perl-validated regressions for the reduced graph case and cloned-object DESTROY timing.
  • Fix reference-graph caching so the cloned structure matches Perl for shared scalar referents, cycles, and repeated references.
  • Use anonymous-reference creation semantics for newly cloned aggregates and correctly account for cloned array element owner slots.
  • Pass the focused regressions on JVM and interpreter, then pass Clone-Any 1.01 unchanged on system Perl and both PerlOnJava backends.
主要语言
Perl
星标
67
派生
8
平均合并
4 小时 36 分钟
30 天内合并 PR
123

环境准备

  • 提供 Dockerfile 或 Docker Compose 文件
  • 没有 Pull Request 模板
  • 阅读贡献指南

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。

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 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。