Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

v0.1.0 - 2026-09-21

This started as a silly little diversion in February but I'm using this in actual projects now. I've even used it to shake out bugs in Affix.

I might move it out of the Acme namespace...

Anyway, the major win is that fiber hot path has been moved from Perl into C and roughly tripled context swapping throughput with no change to the public API.

Added

• Acme::Parataxis::Channel: a buffered FIFO message queue for producer/consumer patterns between fibers. Writers block when full, readers when empty; a capacity of 1 makes it a rendezvous point. Built on two semaphores.
• Acme::Parataxis::Future: a one-shot placeholder for an eventual computation result. A producer fires set_result/set_error exactly once; consumers pick it up with await/result or register an on_ready callback.
• Acme::Parataxis::Semaphore: a counting semaphore with no ownership: blocked fibers are parked (no busy-wait) and resumed FIFO as permits become available.
• Acme::Parataxis::Signal: a two-state flag with a FIFO queue of waiters. send latches the signal so a later wait consumes it immediately, while broadcast wakes every queued waiter at once (and drops if nobody is waiting).

Fixed

• Fixed crash (double-free / use-after-free) when fibers call Affix'd functions on non-threaded Perl. The bug was in Affix's SAVEVPTR/SAVEDESTRUCTOR_X arena pattern, which was not fiber-safe; now fixed upstream in Affix v1.2.5+.
• Fixed SIGSEGV on macOS and FreeBSD caused by fiber stacks being only 512KB (via posix_memalign). All POSIX platforms now use a 64MB mmap-backed stack with a PROT_NONE guard page, matching the Linux path. The SIGSEGV guard handler is also available on macOS/FreeBSD now.
• Fixed FreeBSD compilation: added MAP_ANONYMOUS w/ MAP_ANON fallback.
• Fixed macOS SIGBUS: the guard region size is now derived from sysconf(_SC_PAGESIZE) at runtime so it always covers at least one full page (16 KiB on Apple Silicon). Also fixed cleanup() to use munmap() instead of free() on non-Linux POSIX platforms.
• is_finished() now rejects fiber ids of MAX_FIBERS or greater instead of reading out of bounds of the fiber table.
• Closed a busy-spin footgun: wait, fiber await, Semaphore waits, and Signal->wait now croak instead of burning 100% CPU when called from outside the scheduler, and ->new croaks when the 1024-slot fiber table is exhausted rather than creating a fiber that can never run.
• A fiber that yields during its initial run is now re-enqueued by the scheduler instead of being dropped, which previously could hang a regex-heavy workload.
• The scheduler no longer hangs when a fiber object is created but never spawned (->new without spawn): live-fiber tracking only counts fibers that have actually started, matching Coro's ready-queue semantics.
• async/run is now re-entrant: a nested async inside another async or inside a fiber shares the one run loop (like Coro's single global scheduler) and returns the block's value, instead of clobbering the outer scheduler and deadlocking.
• A destroyed fiber's id is kept out of the free list until every job it submitted has been reclaimed, so a stale completion can never be misdelivered to a (or corrupt) fiber that later reuses the id.
• Pending-job tracking now reads the C-side outstanding job count instead of a run-local counter, so jobs left over from a stopped run are drained and handled by the next run instead of tripping FATAL: deadlock detected or sitting in the done-queue forever.
• Semaphore up/adjust skip stale (already destroyed) waiters instead of consuming a wake that should go to a live fiber.
• Channel constructors now reject a capacity below 1 instead of deadlocking on it at load time.
• The 1024-slot job queue is no longer fatal on the first try: _submit_job yields once and retries before croaking.
• Future::set_result/set_error wake awaiters exactly once instead of appending a duplicate _wake_waiters callback on every await.

Changed

• Spawned fibers run inline at spawn time.
• The fiber registry is replaced by strong references to each fiber object in C.
• Fiber completion moved from a Perl method into C: the entry point writes state directly into the object's slots with av_store, and only dispatches callbacks when callbacks were actually registered.
• To save time on FFI boundary crossings, spawn now performs the whole create run sequence in a single call and builds the fiber object in C.
• Fiber objects are incrementally-filled AVs instead of HVs.
• On x86_64 ELF, context switching uses a hand written trampoline that only saves the callee-saved registers and stack pointer, avoiding swapcontext's signal-mask syscall.
• spawn and await hot paths flattened by inlining helpers.
• Worker threads block on select() for the full await_read/await_write timeout instead of polling every 10ms, cutting idle syscalls by ~50x. On POSIX a shutdown pipe wakes any worker blocked in select() during cleanup().

v0.0.10 - 2026-02-22

This version comes with a dynamic thread pool and an improved API.

Added

• New ergonomic API using exported functions like async { ... }, fiber { ... }, and await( $target ).

Changed

• Refactored native thread pool to use cond vars (PARA_COND_*) instead of busy polling, reducing idle CPU usage to near zero.
• Switched to a global job queue for the thread pool for better load balancing across worker threads.
• Reduced default fiber stack size from 4MB to 512K.
• Worker threads are now only spawned when the first asynchronous job is submitted.
• Increased MAX_FIBERS limit to 1024.
• Expose thread pool config with set_max_threads and max_threads.

v0.0.9 - 2026-02-21

Asynchronous HTTP::Tiny is basically a semi-automatic footgun.

Fixed

• Resolved AvFILLp(av) == -1 and !AvREAL(av) assertion failures in Perl_pp_entersub on DEBUGGING builds of Perl. This was fixed by ensuring Slot 0 (the argument array) of the next pad depth is correctly initialized during fiber context switches.

Changed

• Increased fiber stack size to 4MB to provide better support for deep Perl calls and regex operations. This is a temp solution.

v0.0.8 - 2026-02-19

All the remaining failing smokers all had old versions of Affix and sure enough when I installed v1.0.6, I saw the same failure. Always the most obvious thing...

Changed

• Require Affix v1.0.7

v0.0.7 - 2026-02-18

Another dist targetting a specific CPAN smoker. I cannot replicate the failure in https://www.cpantesters.org/cpan/report/f0ca1d14-0cfa-11f1-9988-e7d94c615303, so I'm just trying different things...

Fixed?

• Arguments passed to a fiber might not be released until the fiber object was destroyed.

v0.0.6 - 2026-02-18

Fixed

• Resolved assertion failures in Perl_cx_popsub_args and Perl_pp_entersub when running on a DEBUGGING build of Perl. This was fixed by ensuring CvDEPTH and pads are correctly restored during context switches. (I hope...)

Added

• Added --debug build to GitHub Actions matrix to ensure future compatibility with Perl debugging builds.

Changed

• Refactored swap_perl_state to be more robust regarding Perl's internal stack management.

v0.0.5 - 2026-02-18

Changed

• I'm honeslty just throwing stuff at the wall. Between my local machines and GH CI workflows, I cannot replicate some of the failures I'm seeing from smokers which makes them virtually impossible to resolve.

v0.0.4 - 2026-02-17

Changed

• Attempt to only spawn max X threads in t/006_parallel.t where X is 3 or the get_thread_pool_size()? See https://www.cpantesters.org/cpan/report/ecf1410e-0c46-11f1-8628-aee76d8775ea
• Recalculate PL_curpad = AvARRAY(PL_comppad) in swap_perl_state? See https://www.cpantesters.org/cpan/report/e7244bd8-0c44-11f1-b3ab-94362698fc84

v0.0.3 - 2026-02-17

Changed

• Adding an optional timeout to await_read and await_write.
• Allow fibers to return complex data (AV*, HV*).

v0.0.2 - 2026-02-17

Fixed

• Fixed segfault in coro_yield by adding NULL checks for destroyed or missing fibers.
• Resolved stall in exception handling by introducing last_sender tracking to prevent parent_id cycles.

Changed

• Made unit tests a lot more noisy

v0.0.1 - 2026-02-16

Changes

• It exists! It shouldn't but it does.