POE Fix Plan

Overview

Module: POE 1.370 (Perl Object Environment - event-driven multitasking framework) Test command: ./jcpan -t POE Status: 35/53 unit+resource tests pass, ses_session.t 37/41 (up from 7/41), ses_nfa.t 39/39, k_alarms.t 37/37, k_aliases.t 20/20, k_selects.t 17/17

Dependency Tree

POE 1.370
├── POE::Test::Loops 1.360       PASS (2/2 tests)
├── IO::Pipely 0.006             OK (loads successfully after fixes)
│   └── IO::Socket (>= 1.38)    FIXED (exists &sub in require context)
│   └── Symbol (>= 1.08)        FIXED ($VERSION set in Java module)
├── Time::HiRes (>= 1.59)       STUB (POE has it commented out)
├── IO::Tty 1.08                 UNAVAILABLE (XS/native, needs C compiler)
├── IO::Poll (optional)          MISSING
├── IO::Pty (optional)           MISSING (needs IO::Tty)
├── Curses (optional)            MISSING
├── Term::Cap (optional)         MISSING
├── Term::ReadKey (optional)     MISSING ($VERSION not set)
├── Socket::GetAddrInfo (opt)    MISSING
├── POSIX                        FIXED (added uname, signals, errno consts)
├── Errno                        OK (pure Perl, complete)
├── Storable                     OK (XSLoader backend)
└── HTTP::Request/Response       PARTIAL (for Filter::HTTPD)

Bugs Fixed (Commits 743c26461 through f119640a5)

Bug 1: `exists(&Errno::EINVAL)` fails in require context - FIXED

Root cause: ConstantFoldingVisitor.java was replacing IdentifierNode("Errno::EINVAL") with NumberNode("22") inside the & sigil operator, so EmitOperatorDeleteExists.java couldn't match the pattern.

Fix: Added guard in ConstantFoldingVisitor.visit(OperatorNode) to skip folding when operator is "&".

Bug 2: Cross-file `use vars` under strict - NOT A BUG

Investigation confirmed this was a cascading failure from Bug 1. When POE::Kernel.pm crashed during loading (before reaching use vars qw($poe_kernel)), the variable was never declared. Once Bug 1 was fixed, POE loads correctly and $poe_kernel is visible across files.

Bug 3: Symbol.pm `$VERSION` not set - FIXED

Fix: Added GlobalVariable.getGlobalVariable("Symbol::VERSION").set("1.09") in Symbol.java.

Bug 4: Indirect object syntax `import $package ()` - FIXED

Root cause: Variable.java line 168 errored when it saw ( after a $var in indirect object context.

Fix: Added parsingIndirectObject flag to Parser, set it in SubroutineParser before parsing the class variable in indirect object syntax.

Bug 5: POSIX missing functions and constants - FIXED

Fix: Added to POSIX.java: signal constants (SIGHUP-SIGTSTP), errno constants (EPERM-ERANGE), uname(), sigprocmask() stub. Added to POSIX.pm: stub classes for SigSet/SigAction.

Bug 6: ConcurrentModificationException in hash each() - FIXED

Root cause: Java's HashMap iterator throws when the hash is modified during iteration. Perl's each tolerates this (common idiom: while (each %h) { delete ... }).

Fix: RuntimeHash.RuntimeHashIterator now snapshots entries at creation time via new ArrayList<>(elements.entrySet()).iterator().

Bug 7: Socket.pm missing IPPROTO constants - FIXED

Fix: Added IPPROTO_TCP, IPPROTO_UDP, IPPROTO_ICMP to both Socket.java and Socket.pm.

Bug 8: %SIG not pre-populated with signal names - FIXED (commit ba803dc49)

Root cause: %SIG was empty in PerlOnJava. Perl pre-populates it with signal names as keys (undef values). POE discovers available signals via keys %SIG.

Fix: RuntimeSigHash.java constructor now pre-populates with POSIX signals plus platform-specific signals (macOS: EMT, INFO, IOT; Linux: CLD, STKFLT, PWR, IOT).

Bug 9: DESTROY not called for blessed objects - ATTEMPTED AND REVERTED

Root cause: PerlOnJava had no DESTROY support. POE and many modules rely on DESTROY for cleanup.

Attempted fix: DestroyManager.java using java.lang.ref.Cleaner to detect GC-unreachable blessed objects and reconstruct proxy objects for DESTROY calls.

Why it was reverted: The proxy reconstruction approach is fundamentally fragile:

close() inside DESTROY on a proxy hash corrupts subsequent hash access (File::Temp "Not a HASH reference" at line 205)
Overloaded classes get negative blessIds; Math.abs() on cache keys collided with normal class IDs
Proxy can't fully replicate tied/magic/overloaded behavior of original objects

Current state: DESTROY is not called for regular blessed objects. Tied variable DESTROY still works (uses scope-based cleanup via TieScalar.tiedDestroy()). See dev/design/object_lifecycle.md for future directions (scope-based ref counting recommended).

Impact on POE: POE's core event loop works without DESTROY. The only affected feature is POE::Session::AnonEvent postback cleanup — sessions using postbacks won't get automatic refcount decrement. Workaround: explicit cleanup or patching POE.

Bug 10: foreach doesn't see array modifications during iteration - FIXED (commit f79f9f6e8)

Root cause: RuntimeArrayIterator cached elements.size() at creation time. Perl's foreach sees elements pushed to the array during the loop body. This broke POE's Kernel->stop() which uses exactly this pattern to walk the session tree:

my @children = ($self);
foreach my $session (@children) {
    push @children, $self->_data_ses_get_children($session->ID);
}

Fix: Changed hasNext() to check elements.size() dynamically instead of using a cached value. This was the root cause of ses_session.t hanging after test 31 — nested child sessions were not being found during stop(), leaving orphan sessions keeping the event loop alive.

Bug 11: `require File::Spec->catfile(...)` parsed as module name - FIXED (commit 6b9fa2c30)

Root cause: POE's Resource/Clock.pm does require File::Spec->catfile(qw(Time HiRes.pm)). The parser's parseRequire method consumed File::Spec as a bareword module name without checking for -> method call after it.

Fix: In OperatorParser.java, save parser position after consuming identifier, peek for ->, and if found, restore position and fall through to expression parsing. This allows Time::HiRes to load correctly, making monotime() return float time instead of integer seconds.

Bug 12: Non-blocking I/O for pipe handles - FIXED (commit 6b9fa2c30)

Root cause: POE's _data_handle_condition calls IO::Handle::blocking($handle, 0) on the signal pipe, but PerlOnJava's pipes didn't support non-blocking mode. sysread on empty non-blocking pipe blocked forever.

Fix: Added isBlocking()/setBlocking() to IOHandle interface, implemented in InternalPipeHandle with EAGAIN (errno 11) return on empty non-blocking read. Fixed IO::Handle::blocking() Perl method to use ($fh, @args) = @_ instead of shift (which copied the glob and lost connection to the underlying handle).

Bug 13: DestroyManager crash with overloaded classes - FIXED (commit cddf4b121)

Root cause: DestroyManager.registerForDestroy used Math.abs(blessId) as cache keys, but overloaded classes get negative blessIds (-1, -2, ...). Math.abs(-1) = 1 collided with the first normal class ID, causing getBlessStr to return null and NPE in normalizeVariableName.

Fix: Used original blessId directly as cache key (fixed before DestroyManager was removed).

Bug 14: 4-arg select() marks pipes as always ready - FIXED (commit f119640a5)

Root cause: The NIO-based selectWithNIO() in IOOperator.java treated all non-socket handles (pipes, files) as unconditionally ready (nonSocketReady++). This caused select() to return immediately when monitoring POE's signal pipe, preventing the event loop from blocking for timer timeouts.

Impact: POE's ses_session.t hung at test 7 (before POE::Kernel->run()) because the event loop never slept — select() always returned immediately with the pipe "ready", POE tried to read (got nothing), and looped back.

Fix: Replaced the "always ready" assumption with proper polling:

InternalPipeHandle.hasDataAvailable() checks if data is actually in the pipe
Write ends and regular files remain always-ready
A poll loop with 10ms intervals respects the timeout parameter
Both pollable fds and NIO selector are checked each iteration

Current Test Results (2026-04-04)

Unit Tests (t/10_units/)

Test File	Result	Notes
03_base/01_poe.t	PASS (4/4)
03_base/03_component.t	PASS (1/1)
03_base/04_driver.t	PASS (2/2)
03_base/05_filter.t	PASS (2/2)
03_base/06_loop.t	PASS (1/1)
03_base/07_queue.t	PASS (2/2)
03_base/08_resource.t	PASS (1/1)
03_base/09_resources.t	FAIL (1/7)	CORE::GLOBAL::require not supported
03_base/10_wheel.t	PASS (7/7)
03_base/11_assert_usage.t	PASS (76/76)
03_base/12_assert_retval.t	PASS (22/22)
03_base/13_assert_data.t	PASS (7/7)
03_base/14_kernel.t	PASS (6/6)
03_base/15_kernel_internal.t	PARTIAL (7/12)	File handle dup bug
03_base/16_nfa_usage.t	PASS (11/11)
03_base/17_detach_start.t	PASS (14/14)
04_drivers/01_sysrw.t	TIMEOUT	Hangs on I/O test
05_filters/01_block.t	PASS (42/42)
05_filters/02_grep.t	PASS (48/48)
05_filters/03_http.t	PARTIAL (79/137)	HTTP::Message bytes issue
05_filters/04_line.t	PASS (50/50)
05_filters/05_map.t	FAIL	Minor test failure
05_filters/06_recordblock.t	PASS (36/36)
05_filters/07_reference.t	FAIL	Storable not available at test time
05_filters/08_stream.t	PASS (24/24)
05_filters/50_stackable.t	PASS (29/29)
05_filters/51_reference_die.t	FAIL (0/5)	Storable not available at test time
05_filters/99_filterchange.t	FAIL	Filter::Reference compilation
06_queues/01_array.t	PASS (2047/2047)
07_exceptions/01_normal.t	PASS (7/7)
07_exceptions/02_turn_off.t	PASS (4/4)
07_exceptions/03_not_handled.t	PASS (8/8)
08_loops/01_explicit_loop.t	PASS (2/2)
08_loops/02_explicit_loop_fail.t	PASS (1/1)
08_loops/03_explicit_loop_poll.t	FAIL	IO::Poll not available
08_loops/04_explicit_loop_envvar.t	FAIL	IO::Poll not available
08_loops/05_kernel_loop.t	PASS (2/2)
08_loops/06_kernel_loop_poll.t	FAIL	IO::Poll not available
08_loops/07_kernel_loop_fail.t	PASS (1/1)
08_loops/08_kernel_loop_search_poll.t	FAIL	IO::Poll not available
08_loops/09_naive_loop_load.t	TODO	Feature not implemented yet
08_loops/10_naive_loop_load_poll.t	TODO	Feature not implemented yet
08_loops/11_double_loop.t	TODO	Feature not implemented yet

Resource Tests (t/20_resources/10_perl/)

Test File	Result	Notes
aliases.t	PASS (14/14)	Fixed ConcurrentModificationException
caller_state.t	PASS (6/6)
events.t	PASS (38/38)
extrefs.t	PASS (31/31)
extrefs_gc.t	PASS (5/5)
filehandles.t	FAIL (1/132)	Socket.getChannel() null
sessions.t	PASS (58/58)
sids.t	PASS (7/7)
signals.t	PARTIAL (45/46)	1 test failure

Summary: 35 test files fully pass, 18 fail/partial

Remaining Issues

Pre-existing PerlOnJava limitations (not POE-specific)

Issue	Impact	Category
CORE::GLOBAL::require override not supported	09_resources.t	Runtime feature
File handle dup (open FH, ">&OTHER") shares state	15_kernel_internal.t	I/O subsystem
Socket.getChannel() returns null	filehandles.t	Network I/O
IO::Poll not available	4 loop tests	Missing module

Issues worth fixing

Issue	Impact	Difficulty
Storable not found by POE test runner	3 filter tests	Low (path issue?)
HTTP::Message bytes handling	03_http.t (58 tests)	Medium
01_sysrw.t hangs	1 driver test	Medium (I/O)
signals.t 1 failure	1 test	Low

Event Loop Tests (t/30_loops/select/)

Test File	Result	Notes
00_info.t	PASS (2/2)
all_errors.t	SKIP
k_alarms.t	PASS (37/37)	Alarm scheduling works
k_aliases.t	PASS (20/20)	Session aliases work
k_detach.t	PASS (9/9)	Session detach works
k_run_returns.t	PASS (1/1)
k_selects.t	PASS (17/17)	File handle watchers
k_sig_child.t	PARTIAL (5/15)	Child signal handling
k_signals.t	PARTIAL (2/8)	Signal delivery
k_signals_rerun.t	FAIL
sbk_signal_init.t	PASS (1/1)
ses_nfa.t	TIMEOUT	NFA session hangs
ses_session.t	PARTIAL (35/41)	Signal delivery + DESTROY timing
comp_tcp.t	FAIL (0/34)	TCP networking
wheel_accept.t	PARTIAL (1/2)	Hangs after test 1
wheel_readwrite.t	PARTIAL (16/28)	I/O events don't fire, hangs
wheel_run.t	PARTIAL (42/103)	10 pass, 32 skip (IO::Pty), blocked by TIOCSWINSZ
wheel_sf_tcp.t	PARTIAL (4/9)	Hangs after test 4
wheel_sf_udp.t	PARTIAL (4/10)	UDP datagrams never delivered
wheel_sf_unix.t	FAIL (0/12)	Socket factory Unix
wheel_tail.t	PARTIAL (4/10)	sysseek now works; hangs due to DESTROY
z_kogman_sig_order.t	PASS (7/7)
z_merijn_sigchld_system.t	PASS (4/4)
z_steinert_signal_integrity.t	PASS (2/2)
connect_errors.t	PASS (3/3)
k_signals_rerun.t	PARTIAL (1/9)	TIOCSWINSZ error in child processes

Event loop summary: 13/35 fully pass. Core event loop works (alarms, aliases, detach, signals).

Fix Plan - Remaining Phases

Phase 3: Event loop and session hardening (high impact)

Step	Target	Expected Impact	Difficulty
3.1	Fix ses_session.t (7/41)	Core session lifecycle validation	Medium
3.2	Fix k_selects.t (5/17)	File handle watcher support	Medium
3.3	Fix k_signals.t (2/8) and k_sig_child.t (5/15)	Signal delivery	Medium
3.4	Fix signals.t (45/46)	1 remaining test failure	Low
3.5	Fix Storable path for POE test runner	Unblocks 3 filter tests	Low
3.6	Fix ses_nfa.t timeout	NFA state machine tests	Medium

Phase 4: Extended features (lower priority)

Step	Target	Expected Impact	Difficulty
4.1	HTTP::Message bytes handling	03_http.t (58 more tests)	Medium
4.2	Socket/network tests (comp_tcp, wheel_sf_*)	TCP/UDP networking	Hard
4.3	IO::Poll stub	4 poll-related loop tests	Medium
4.4	File handle dup fix	15_kernel_internal.t (5 tests)	Hard
4.5	wheel_tail.t (FollowTail)	File watching	Medium

Phase 5: JVM limitations (not fixable without major work)

Feature	Reason
wheel_run.t (103 tests)	Needs fork + IO::Pty (native)
IO::Tty / IO::Pty	XS module, needs C compiler
wheel_curses.t	Needs Curses (native)
wheel_readline.t	Needs terminal

Progress Tracking

Current Status: Phase 4.3 analysis complete — DESTROY is the root cause of all wheel hangs

Completed Phases

Key Findings (Phase 3.1-3.4)

foreach-push pattern: Perl's foreach dynamically sees elements pushed during iteration. PerlOnJava's RuntimeArrayIterator was caching size at creation. This broke POE::Kernel->stop() which walks the session tree by pushing children during foreach.
DESTROY not feasible via GC: Java's Cleaner/GC-based DESTROY is unreliable across JVM implementations and cannot guarantee deterministic timing. Perl's DESTROY depends on reference counting (fires immediately when last reference drops). The JVM's tracing GC is fundamentally incompatible with this semantic. DESTROY is not implemented.
DESTROY workaround for POE: POE::Session::AnonEvent postbacks/callbacks now auto-cleanup when called (decrement session refcount on first invocation). This replaces DESTROY-based cleanup for the common one-shot postback pattern. 4 ses_session.t failures remain for tests that explicitly count DESTROY invocations.
Dual fd registry: pipe() registered handles in FileDescriptorTable but not RuntimeIO.filenoToIO. select() only consulted RuntimeIO, making pipes invisible. Fixed by registerExternalFd().
Fd counter collision: FileDescriptorTable and RuntimeIO maintained separate fd counters. socketpair/socket advanced RuntimeIO.nextFileno but not FileDescriptorTable.nextFd, causing pipe() to allocate fds that overlapped with existing socket fds. Fixed by cross-synchronizing both counters (advancePast/advanceFilenoCounterPast).
select() bitvector write-back: 4-arg select() created snapshot copies of bitvector args for tied-variable safety, but never wrote the modified bitvectors back to the original variables. Callers always saw their original bitvectors unchanged. Fixed to write back after selectWithNIO returns.
socketpair() stream init: socketpair() used SocketIO(channel, family) constructor which doesn't initialize inputStream/outputStream. sysread() failed with "No input stream available" on blocking sockets. Fixed to use SocketIO(socket) constructor, and made sysread() fall back to channel-based I/O when streams are unavailable.
Platform errno: Hard-coded Linux errno values (EAGAIN=11) caused mismatches on macOS (EAGAIN=35). Fixed to use ErrnoVariable.EAGAIN() which probes the platform.
Signal delivery: Now works end-to-end: kill() → %SIG handler → signal pipe write → select() detects pipe → POE dispatches signal event.
require expression parsing: require File::Spec->catfile(...) was parsed as require File::Spec (module) instead of require <expr>. This prevented Time::HiRes from loading, causing monotime() to return integer seconds instead of float.

Key Findings (Phase 4.3) — DESTROY Root Cause Analysis

The event loop I/O hang is caused by POE::Wheel DESTROY not being called, not by I/O subsystem bugs. The underlying I/O works correctly:

fileno() now works for all handle types (regular files, JARs, scalars, pipes)
select() correctly detects readability on regular file handles
sysread()/syswrite() work on tmpfiles, pipes, and sockets
sysseek() works correctly (returns position, "0 but true" at offset 0)
POE::Wheel::ReadWrite I/O events fire correctly in isolation

Verified working in isolation: A standalone POE session with a ReadWrite wheel on a tmpfile receives all input events, flushes writes, and completes. The state machine (read→pause→seek→resume→write→shutdown) works as expected.

What fails: When wheels are created in eval (test_new validation) or deleted via delete $heap->{wheel}, DESTROY never fires. This leaves orphan select() watchers and anonymous event handlers registered in the kernel, preventing sessions from stopping.

POE::Wheel DESTROY cleanup pattern (all wheels follow this):

Remove I/O watchers: $poe_kernel->select_read($handle), select_write($handle)
Cancel timers: $poe_kernel->delay($state_name) (FollowTail only)
Remove anonymous states: $poe_kernel->state($state_name)
Free wheel ID: POE::Wheel::free_wheel_id($id)

Specific test impacts:

wheel_readwrite.t test 12: ReadWrite->new(Handle=>\*DATA, LowMark=>3, HighMark=>8, LowEvent=>"low") succeeds when it should die (missing HighEvent validation in POE). The wheel registers a select watcher on *DATA, and without DESTROY, it's never removed. This keeps Part 1's session alive forever, preventing Part 2 from running.
wheel_tail.t: FollowTail file-based watching works (creates file, reads lines, detects resets), but the session hangs after delete $heap->{wheel} because the FollowTail's delay timer and select watcher aren't cleaned up.
wheel_sf_tcp.t, wheel_accept.t: TCP server accepts connections (test 1 passes), but subsequent wheel lifecycle depends on DESTROY for cleanup between phases.

DESTROY Workaround Options (Not Yet Implemented)

Option	Approach	Pros	Cons
A	Trigger DESTROY on `delete`/`set` — when overwriting a blessed reference in a hash/scalar, check if the class defines DESTROY and call it	Simple to implement; covers the `delete $heap->{wheel}` pattern	May call DESTROY too early if other references exist; not refcount-accurate
B	Scope-based cleanup via tied proxy — wrap wheel references in a tied scalar that calls DESTROY when the scalar is overwritten	More accurate lifecycle tracking	Complex; requires patching POE internals
C	Patch POE::Wheel subclasses — add explicit `_cleanup()` methods, patch POE::Kernel to call them when sessions stop	Clean, no Java changes needed	Fragile; must patch every Wheel subclass
D	Implement reference counting — track refcounts for blessed objects at the RuntimeScalar level	Correct Perl 5 semantics	Very complex; massive changes to RuntimeScalar, affects performance
E	GC-based DESTROY with Cleaner — register cleanup actions with java.lang.ref.Cleaner, run when GC collects	No refcounting needed	Unpredictable timing; previously attempted and reverted
F	POE::Kernel session GC — when a session has no pending events/timers except select watchers, force-remove all watchers	Targeted to POE's specific deadlock pattern	Doesn't generalize; requires understanding POE's invariants

Recommended approach: Option A (trigger DESTROY on delete/set) is the most pragmatic. POE's pattern is always $heap->{wheel} = Wheel->new(...) with a single reference, so calling DESTROY when the hash value is overwritten or deleted is correct 99% of the time. For safety, DESTROY could be made idempotent (track whether it's already been called).

Next Steps (Phase 4)

Current Event Loop Test Inventory (35 test files, ~596 tests total)

Fully passing (13 files, 178 tests):

00_info (2/2), k_alarms (37/37), k_aliases (20/20), k_detach (9/9), k_run_returns (1/1), k_selects (17/17), sbk_signal_init (1/1), ses_nfa (39/39), ses_session (37/41), z_kogman_sig_order (7/7), z_merijn_sigchld_system (4/4), z_steinert_signal_integrity (2/2), connect_errors (3/3)

Partially passing (10 files):

ses_session: 37/41 — 4 failures from DESTROY (JVM limitation, won't fix)
k_signals: 2/8 — remaining tests need fork()
k_sig_child: 5/15 — remaining tests need fork()
wheel_tail: 4/10 — sysseek works; hangs due to DESTROY (FollowTail cleanup)
wheel_run: 42/103 — 10 pass, 32 skip (IO::Pty), blocked by TIOCSWINSZ constant
wheel_sf_tcp: 4/9 — hangs after test 4 (event loop stalls after first TCP message)
wheel_sf_udp: 4/10 — UDP sockets created but datagrams never delivered
wheel_accept: 1/2 — hangs after test 1 (accept callback never fires)
wheel_readwrite: 16/28 — constructor tests pass, I/O events don't fire, hangs
k_signals_rerun: 1/9 — child processes fail with TIOCSWINSZ error

Blocked by missing sysseek — FIXED (commit 5b0ca1383):

sysseek now implemented for both JVM and interpreter backends
wheel_tail FollowTail file-based watching works in isolation
Remaining wheel_tail failures are DESTROY-related, not sysseek-related

Blocked by missing TIOCSWINSZ (Wheel::Run ioctl):

wheel_run additional tests beyond test 42, k_signals_rerun (8 of 9 fail)
TIOCSWINSZ is an ioctl constant from sys/ioctl.ph; needs stub or ioctl.ph generation

Event loop I/O hang pattern (root cause: DESTROY):

wheel_readwrite, wheel_sf_tcp, wheel_accept, wheel_sf_udp, wheel_tail all hang because POE::Wheel DESTROY never fires when wheels go out of scope
The I/O subsystem itself works — select(), sysread/syswrite, fileno all verified
Constructor/setup tests pass, then sessions hang because orphan watchers remain

Skipped (platform/network):

all_errors (0, skip), comp_tcp (0, skip network), comp_tcp_concurrent (0), wheel_curses (0, skip IO::Pty), wheel_readline (0), wheel_sf_unix (0, skip), wheel_sf_ipv6 (0, skip GetAddrInfo), z_rt53302_fh_watchers (0, skip network)

Fork-dependent (JVM limitation, won't fix):

wheel_run (103), k_sig_child, k_signals_rerun, z_rt39872_sigchld*, z_leolo_wheel_run, z_merijn_sigchld_system (passes via system())

Phase 4.1: Add Socket pack_sockaddr_un/unpack_sockaddr_un stubs — DONE

Impact: Unblocks POE::Wheel::SocketFactory loading
Enables: wheel_sf_tcp (9), wheel_sf_udp (10), wheel_accept (2), connect_errors (3)
Difficulty: Low (stub functions that die on actual use)
Result: connect_errors 3/3 PASS, wheel_sf_tcp 4/9 (hangs after test 4), wheel_accept 1/2 (hangs)

Phase 4.2: Add POSIX terminal/file constants — DONE

Added: 80+ constants (stat permissions, terminal I/O, baud rates, sysconf/_SC_OPEN_MAX, setsid)
Added: S_IS* file type test functions (S_ISBLK, S_ISCHR, S_ISDIR, etc.) as pure Perl
Impact: Unblocks POE::Wheel::Run and Wheel::FollowTail loading
Result:
- Wheel::FollowTail loads, 4/10 pass — blocked by missing sysseek operator
- Wheel::Run loads, 42/103 (6 pass, 36 skip for IO::Pty) — blocked by missing TIOCSWINSZ ioctl constant
New blockers found:
- sysseek: Not implemented in PerlOnJava. FollowTail uses sysseek($fh, 0, SEEK_CUR) to get current file position. Error: "Operator sysseek doesn't have a defined JVM descriptor"
- TIOCSWINSZ: Bareword ioctl constant used in Wheel::Run for terminal window size. Error: 'Bareword "TIOCSWINSZ" not allowed while "strict subs"'. This is a require inside an eval — a constant from sys/ioctl.ph that doesn't exist on JVM.

Phase 4.3: Debug event loop I/O hang — DONE (analysis complete)

Root cause: DESTROY not called for POE::Wheel objects (see analysis above)
The I/O subsystem works correctly; all hangs traced to orphan select watchers
Fixed fileno() for regular files (Bug 21) — unrelated but needed for POE
See "DESTROY Workaround Options" section for implementation plan

Phase 4.4: Implement sysseek — DONE (commit 5b0ca1383)

sysseek implemented for JVM backend (CoreOperatorResolver, EmitBinaryOperatorNode, OperatorHandler, CompileBinaryOperator) and interpreter backend
Returns new position or "0 but true", unlike seek which returns 1/0
POE::Wheel::FollowTail file-based watching now works in isolation

Phase 4.5: Implement DESTROY workaround (highest remaining impact)

Affects: wheel_readwrite (28), wheel_tail (10), wheel_sf_tcp (9), wheel_accept (2), wheel_sf_udp (10), ses_session (4), plus any module using DESTROY for cleanup
Recommended: Option A — trigger DESTROY on hash delete/set when blessed ref is overwritten
Expected impact: 20-30+ additional test passes across 5+ test files
Difficulty: Medium-Hard (requires changes to RuntimeHash.delete/RuntimeScalar.set)

Phase 4.6: Add TIOCSWINSZ stub (unblocks Wheel::Run child processes)

Affects: wheel_run (42/103), k_signals_rerun (1/9)
TIOCSWINSZ is loaded via require 'sys/ioctl.ph' inside an eval
Options: (a) create a stub sys/ioctl.ph, or (b) make the eval silently fail
Most wheel_run tests also need fork, so impact is limited
k_signals_rerun would benefit most (8 failures all from TIOCSWINSZ in child)

Phase 4.7: Windows platform support

POE's core event loop uses Java NIO (cross-platform), and POE itself has $^O eq 'MSWin32' guards. But several PerlOnJava subsystems only have macOS/Linux branches.

Step	Target	Severity	Difficulty
4.7.1	Fix EAGAIN/errno on Windows — `FFMPosixWindows.strerror()` is a 10-entry stub. `ErrnoVariable` probes strerror to discover EAGAIN/EINPROGRESS/etc., which all resolve to 0 on Windows. POE's non-blocking I/O depends on these. Fix: add Windows errno constants directly, bypassing strerror probing.	Critical	Medium
4.7.2	Add Windows errno table — `ErrnoHash.java` and `Errno.pm` have macOS/Linux tables only, fall through to Linux on Windows. Fix: add `MSWin32` branch with MSVC CRT errno values.	High	Low
4.7.3	Add Windows branch to RuntimeSigHash — `%SIG` is pre-populated with Unix signals (HUP, USR1, PIPE, etc.) that don't exist on Windows. POE iterates `keys %SIG` and tries to install handlers. Fix: Windows branch with only INT, TERM, ABRT, FPE, ILL, SEGV, BREAK.	High	Low
4.7.4	Add Windows branch to POSIX.java — IS_MAC flag gives macOS vs "everything else" (Linux). Windows gets Linux signal/termios values. Fix: add IS_WINDOWS flag, return Windows-correct signal constants, skip termios.	Medium	Low
4.7.5	Fix Socket.java constants — SOL_SOCKET=1 (Linux) vs 0xFFFF (Windows), SO_REUSEADDR=2 vs 4, etc. These matter if passed to native setsockopt. Java abstracts most of this, so impact depends on implementation path. Fix: platform-detect and use correct values.	Medium	Low
4.7.6	Fix sysconf for Windows — Runs `ulimit -n` which doesn't exist on Windows. Already has catch block defaulting to 1024, but should use a cleaner approach (e.g., `_getmaxstdio()` or just return 2048).	Low	Low

What already works on Windows (no changes needed):

select() via Java NIO Selector — cross-platform
Pipe handling via Java PipedInputStream/PipedOutputStream — cross-platform
POE::Loop::Select — has $^O eq 'MSWin32' guards
socketpair via loopback TCP — the standard Windows approach
$^O correctly set to MSWin32 on Windows

FilesExpand file tree

poe.md

Latest commit

History