fix: AnyEvent CPAN module — parser/warnings fixes (part 1)

dev/modules/anyevent_fixes.md

@@ -0,0 +1,260 @@
+# AnyEvent CPAN Module — Test Fix Plan
+
+This document tracks the work needed to make `./jcpan -t AnyEvent` pass
+all 83 test programs. The project policy requires all tests to pass,
+including low-priority ones.
+
+## Status
+
+| Date | Failed | Passed | Subtests running | Subtests failed |
+|------|--------|--------|------------------|-----------------|
+| 2026-04-20 initial | 82/83 | 1/83 | 24 | 12 |
+| 2026-04-20 after parser/warnings fixes | 17/83 | 66/83 | 93 | 12 |
+| 2026-04-20 after ternary `:` fix | 14/83 | 69/83 | 103 | 5 |
+| 2026-04-20 after `()` overload marker + `pipe` + delete chain | 13/83 | 70/83 | 157 | 13 |
+| 2026-04-20 after `/gc` in list ctx keeps pos() | 12/83 | 71/83 | 157 | 8 |
+| 2026-04-20 after further fixes (sysopen O_EXCL, ex_data, my(undef,%h) in eval, require package) | **see below** | | | |
+
+**Note**: `./jcpan -t AnyEvent` stops at `t/02_signals.t` because that
+test outputs `Bail out!` on failure, which aborts the entire harness
+run after only 3 files. The signal failure is downstream of the
+`weaken`/cooperative-refcount limitation documented in `AGENTS.md`
+(timer/io watchers are destroyed immediately because `weaken` too
+eagerly clears the last strong ref). This is being addressed in a
+separate branch. Running tests individually would reveal the per-file
+status, which is broadly unchanged from row 6 above aside from:
+
+- `t/11_io_perl.t`: subtest 6 (aio_open with O_EXCL on existing file)
+  was fixed via sysopen O_EXCL handling.
+
+## Already Fixed (PR fix/anyevent-cpan-tests)
+
+- [x] `Unknown warnings category 'internal'` — registered aliases for
+  `debugging`, `inplace`, `internal`, `malloc` → `severe::*` in
+  `WarningFlags.java`.
+- [x] `Bad name after Foo::Bar::::` — added `]` to the set of terminators
+  accepted after a package-name bareword in `IdentifierParser.java`.
+- [x] `EV:: => "val"` autoquoting left trailing `::` — stripped in
+  `ParseInfix.java`.
+- [x] `my $var : $fallback` misparsed as attribute inside ternary —
+  check that what follows `:` is an identifier before treating it as
+  an attribute introducer in `OperatorParser.java`.
+- [x] `&{}` overload via glob aliasing — recognize `()` (not just `((`)
+  as an overload marker in `NameNormalizer.hasOverloadMarker`, so
+  classes that hand-roll overloading (as AnyEvent::CondVar does to
+  avoid loading overload.pm) are properly detected.
+- [x] `pipe` / `socketpair` missing from the bytecode interpreter
+  (eval STRING path) — added PIPE/SOCKETPAIR opcodes and dispatch.
+- [x] `delete $ref->[I][J]` with elided arrow — allowed `[` to have
+  a scalar-expression left side in `CompileExistsDelete.visitDeleteArray`.
+- [x] `elsif` "masks earlier declaration" — verified real Perl emits
+  the same warning, no fix needed (was a red herring).
+- [x] `/gc` in list context now preserves pos() — was unconditionally
+  resetting pos after any list-context /g match. Now honours `/c`.
+  Fixed `AnyEvent::Socket::parse_hostport` IPv6 handling; t/06_socket.t
+  now 19/19 (was 14/19).
+- [x] `sysopen` now honours `O_EXCL` — failed to report `$! = "File
+  exists"` when `O_CREAT|O_EXCL` was used on an existing file.
+  Fixes AnyEvent::IO::Perl's `aio_open` tests.
+- [x] `Net::SSLeay::get_ex_new_index` / `set_ex_data` / `get_ex_data`
+  — previously undefined; AnyEvent::TLS's load-time
+  `until $REF_IDX = get_ex_new_index(...)` looped forever. This does
+  NOT make the SSL test suite pass — AnyEvent::TLS uses ~30 further
+  Net::SSLeay functions that remain unimplemented (`CTX_set_options`,
+  `set_accept_state`, etc.) — but the module now loads.
+- [x] `my (undef, %hash) = @_` inside eval STRING — the bytecode
+  interpreter was silently skipping `undef` placeholders on the LHS
+  of a `my` declaration, mis-pairing keys and values in the hash.
+  Added a LOAD_UNDEF_READONLY opcode that emits the shared read-only
+  `scalarUndef` so `RuntimeList.assign` recognises the placeholder.
+  Triggered by AnyEvent's signal-setup `my (undef, %arg) = @_;` inside
+  `eval q{ *signal = ... }`.
+- [x] `require FILE` / `do FILE` now compile the loaded file in the
+  caller's package. Perl 5 semantics: `sub foo { ... }` inside a
+  required .pl lands in the caller's namespace, not in `main::`. The
+  JVM backend's `package Foo;` now also updates the runtime tracker
+  (`InterpreterState.currentPackage`) so downstream tools see the
+  right package. Fixed via new `CompilerOptions.initialPackage` and
+  an INVOKESTATIC hook in `handlePackageOperator`.
+
+## Remaining Failures (13 test programs)
+
+Each of the items below is one PerlOnJava bug. Items are ordered by
+expected effort/impact.
+
+### A — Investigate: ASM NegativeArraySize compiling DESTROY
+
+**Symptom**:
+```
+java.lang.NegativeArraySizeException: -1
+    at org.objectweb.asm.Frame.merge
+    at org.perlonjava.backend.jvm.EmitterMethodCreator.getBytecode
+```
+then `ASM bytecode generation failed: -1`.
+
+**Trigger**: `AnyEvent::Loop::io::DESTROY`, which combines `delete
+$fds->[W][$fd]` (now fixed in compiler), lvalue `(vec $fds->[V], $fd, 1)
+= 0`, `weaken`, and a `pop @$q`.
+
+**Affects**: `t/07_io.t`, any AnyEvent watcher cleanup path.
+
+**Approach**:
+1. Bisect the DESTROY body to isolate which construct produces invalid
+   stack frames.
+2. Fix the root cause — probably the stack-map frame computation during
+   `emitVarAttrsIfNeeded` or lvalue vec, which both push/pop values in
+   unusual ways.
+
+### B — `lvalue vec` with complex base expression
+
+**Symptom**: Same kind of error as the delete-chain fix that was already
+landed — expected to be fixed by the same pattern, but the assignment
+path (`(vec $expr, $idx, $bits) = value`) has its own lvalue compile
+path and is separate. Likely small.
+
+**Affects**: `t/07_io.t`.
+
+### C — `t/03_child.t`: fork unsupported
+
+`fork` returns undef under jperl (per AGENTS.md). The test dies with
+`unable to fork at t/03_child.t line 35` because it doesn't check
+`$Config{d_fork}` first.
+
+Options:
+1. Implement a minimal `fork` using sub-JVM processes (major work —
+   process state copying).
+2. Make the test skip when `d_fork` is empty.
+3. Patch the extracted CPAN tree before `make test` to add a SKIP block.
+
+**Recommendation**: (2) — ship a jperl-side `$SIG{__DIE__}` convention
+or a `JPERL_FORK_FAIL=skip` environment variable that the bundled
+`jperl` wrapper sets so that `fork` returning failure prints the
+TAP SKIP plan and exits 0. Or: detect the upstream pattern `my $pid =
+fork; defined $pid or die "unable to fork"` and short-circuit. None of
+these are ideal; the cleanest path is still (1) but scope-heavy.
+
+### D — `t/80_ssltest.t` (415 subtests): Net::SSLeay not available
+
+The test has `BEGIN { eval "use Net::SSLeay; 1" or exit 0 }`. It's
+dying before the SKIP takes effect. Either our `use Net::SSLeay;` fails
+outside the `eval ""` scope, or the `BEGIN` ordering is wrong. Trace to
+find why the SKIP path isn't hit. Likely a PerlOnJava bug around
+`eval "use Missing::Module"`.
+
+### E — `t/handle/01_readline.t`, `t/handle/02_write.t` — socket-type detection
+
+AnyEvent's `Handle.pm` bails out with "only stream sockets supported"
+because it gets an unexpected SO_TYPE / getsockname. Verify that our
+socket creation returns the correct packed type and that
+`getsockopt(SOL_SOCKET, SO_TYPE)` is implemented.
+
+### F — `t/handle/04_listen.t`, `t/08_idna.t`: OOM (exit 137)
+
+Run with `JPERL_OPTS="-Xmx2g"`. If still OOM, profile and reduce
+retention.
+
+### G — `t/06_socket.t`: 5 subtests — IPv6 packed-address returns SCALAR ref
+
+```
+not ok 18 # 'SCALAR(0x101952da),443' => ',' eq '2002:58c6:438b::10.0.0.17,443'
+```
+
+Our `pack N*` / `inet_pton` returns a SCALAR reference instead of a
+packed string in the specific path AnyEvent exercises. Narrow the
+minimal reproduction in `AnyEvent::Socket::parse_address` and fix.
+
+### H — `t/09_multi.t` and `t/02_signals.t`: signal delivery / Ctrl+C
+
+Exit 130 is SIGINT. The test's timer/signal infrastructure is probably
+reaching a deadlock and the outer harness sends SIGINT. Likely related
+to AnyEvent::Base using `pipe` + signals, which now compiles but may
+not actually wake up the select loop.
+
+### I — `t/13_weaken.t`: 3 subtests — weaken semantics
+
+```
+not ok 5    # weakened timer still fires
+not ok 6    # twin (expected/unexpected) of 5
+```
+
+Our `weaken` is cooperative-refcount based (per AGENTS.md) and doesn't
+match Perl's eager-free semantics in this specific pattern:
+
+```perl
+Scalar::Util::weaken $t2;
+print $t2 ? "not " : "", "ok 5\n";    # expects $t2 to be undef here
+```
+
+This is a known limitation. If we want 83/83, we need to make weakened
+refs go to undef when the last strong ref drops, even if the referent
+is still reachable from the refcount "wait list".
+
+### J — `t/11_io_perl.t`: 1 failing subtest (#6)
+
+Narrow after all upstream fixes land; may resolve spontaneously.
+
+### K — `t/01_basic.t`: 2 failing subtests (#5, #6 out of sequence)
+
+```
+ok 4
+not ok 5
+ok 5      <-- test 5 appears twice
+ok 6
+```
+
+`$cv->recv` after `croak` is emitting two lines where one is expected.
+Low priority, narrow after A–I land.
+
+## Progress Tracking
+
+### Current status
+
+Tier 1 fixes landed. Remaining work:
+
+- **Quick wins likely**: B (lvalue vec), G (inet_pton), J (single
+  subtest), K (single out-of-sequence).
+- **Moderate**: A (ASM stack frame), D (eval vs use), E (socket type),
+  F (OOM).
+- **Heavy / policy-dependent**: C (fork), I (weaken).
+
+### Completed in this PR (fix/anyevent-cpan-tests)
+
+- caa49bf78: parser + warnings (internal/debugging/inplace/malloc,
+  trailing-`::` terminators, `=>` autoquote)
+- 27a31d5fc: `my $var :` inside ternary
+- 20123cb85: overload detection via `()` marker
+- ce11a2a96: pipe / socketpair in bytecode interpreter
+- 30519d9d2: delete `$ref->[I][J]` with elided arrow
+
+Result: 82/83 → 13/83 test-program failures; subtests 24 → 157.
+
+### Open Questions
+
+- `fork`: implement, or reach agreement that fork-dependent tests are
+  exempt from the "all tests must pass" rule?
+- `weaken` semantics: cooperative-refcount is documented in `AGENTS.md`
+  — do we strengthen it for this test, or treat t/13_weaken #5–#6 as a
+  known deviation?
+
+## Appendix: Minimal reproductions (still open)
+
+### H. Signal delivery via the pipe
+```perl
+use AnyEvent;
+use AnyEvent::Impl::Perl;
+my $cv = AnyEvent->condvar;
+my $w = AnyEvent->signal(signal => "USR1", cb => sub { print "got\n"; $cv->send });
+kill USR1 => $$;
+$cv->recv;
+```
+Should print "got". Currently hangs / SIGINT-killed.
+
+### G. IPv6 formatting
+```perl
+use AnyEvent::Socket;
+print format_address(parse_address("2002:58c6:438b::10.0.0.17")), "\n";
+```
+Should print the roundtrip string. Currently prints a SCALAR(0x...) ref.
+
+### A. ASM failure
+(still to be bisected from `AnyEvent::Loop::io::DESTROY`)