feat: futures on real OS threads (ev/thread)

Implement clojure.core futures backed by Janet's ev/thread for genuine
parallelism (CPU-bound work can use a second core, unlike cooperative go
blocks):

- future / future-call, deref + (deref f timeout-ms timeout-val), future?,
  future-done?, future-cancel, future-cancelled?; realized? on futures.
- A worker OS thread computes and marshals back a [:ok v]/[:error e] result
  over a thread-chan; a parent-side collector fiber caches it and closes a
  broadcast latch so any number of deref-ers unpark.
- Snapshot semantics: separate heaps mean the body + captured state are copied
  to the worker and only the result is copied back (mutating a captured atom
  does not propagate). Documented in README.
- future-cancel can't interrupt a Janet OS thread, so it marks the future
  cancelled/done (deref throws, predicates flip) while the worker runs out.

clojure-test-suite baseline 3915 -> 3913: implementing future unskips
realized_qmark.cljc's (when-var-exists future ...) block, which depends on
JVM Thread/sleep + real thread interruption jolt can't provide; deref then
re-raises the unresolved-Thread/sleep error. Documented at the baseline.

Spec: test/spec/futures-spec.janet (18 cases).
This commit is contained in:
Yogthos 2026-06-05 20:00:02 -04:00
parent 120d6d73fa
commit 8be7743b26
4 changed files with 132 additions and 4 deletions

View file

@ -125,7 +125,8 @@ Jolt targets Clojure semantics but runs on Janet, not the JVM. The notable diver
- **Numbers.** Janet integers and doubles. `(/ 1 3)` is `0.3333…` and large products lose precision. No ratios or `BigDecimal` (`ratio?` is always false, `bigdec` falls back to a double); `bigint`/`biginteger` use Janet's 64-bit `int/s64`, not arbitrary precision. The reader still accepts Clojure's numeric literal syntaxes — the BigInt/BigDecimal suffixes (`42N`, `1.5M`), ratios (`1/2`), radixed integers (`2r1010`, `16rFF`), and exponents (`1e3`) — but reads them as plain Janet numbers (a ratio becomes its double quotient). The auto-promoting `+'`/`-'`/`*'`/`inc'`/`dec'` are aliases for the plain ops, since Janet numbers don't overflow. `quot`/`rem`/`mod` follow Clojure's sign rules. The symbolic values `##Inf`/`##-Inf`/`##NaN` read, and `infinite?`/`NaN?` work. Janet represents an integer and an integer-valued double identically, so `1` and `1.0` are indistinguishable: `(float?/double? 1.0)` is `false` and `(int? 1.0)` is `true``float?`/`double?` are true only for values with a fractional part or `##Inf`/`##NaN`.
- **Collections.** By default Jolt uses immutable persistent data structures: vectors are 32-way branching tries (structural-sharing persistent vectors with O(log₃₂ n) `conj`/`assoc`/`nth`), lists are persistent singly-linked cons cells (O(1) `conj`/`cons` prepend with structural sharing), and maps/sets are persistent hash structures. Value equality and sequence operations are Clojure-compatible, but hash-map/hash-set iteration order is unspecified and differs from Clojure — use `sorted-map`/`sorted-set` when order matters.
- **Mutable build mode.** Jolt can be compiled to use fast Janet-native *mutable* collections instead, via a build-time flag: `JOLT_MUTABLE=1 jpm build` (default `jpm build` is immutable). In mutable mode vectors and lists share one mutable array representation (so `conj` mutates in place and appends, and `vector?`/`list?` no longer distinguish them) — a performance/looseness trade-off. The default immutable build has full Clojure value semantics.
- **Concurrency / STM.** Single-threaded. No refs, `dosync`, agents, or `send`; `locking` evaluates its body without real locking. Atoms, volatiles, and delays are supported.
- **Concurrency / STM.** No refs, `dosync`, agents, or `send`; `locking` evaluates its body without real locking. Atoms, volatiles, promises, and delays are supported.
- **Futures.** `future` runs its body on a *real* OS thread (Janet's `ev/thread`), so it can use a second core for CPU-bound work — unlike the cooperatively-scheduled `go` blocks. `deref`/`@` parks until the result is ready (with the optional `(deref f timeout-ms timeout-val)` arity); `future?`, `future-done?`, `realized?`, `future-cancel`, and `future-cancelled?` are supported. Two important divergences from the JVM: (1) **snapshot semantics** — Janet threads have separate heaps, so the body and the state it closes over are *copied* to the worker thread and only the return value is copied back; mutating a captured atom does not propagate to the parent (communicate via the return value). (2) **no thread interruption** — Janet OS threads can't be cancelled mid-run, so `future-cancel` marks the *future* cancelled (deref then throws and the predicates flip) but the underlying computation still runs to completion in the background. As on the JVM, a live future thread keeps the process alive until it finishes (the JVM's non-daemon future pool behaves the same).
- **core.async.** `clojure.core.async` runs on Janet fibers and channels (`chan`, `go`, `go-loop`, `<!`/`>!`/`<!!`/`>!!`, `close!`, `alts!`, `timeout`, `put!`/`take!`, `buffer`/`dropping-buffer`/`sliding-buffer`, and channel transducers via `(chan n xform)`). Because Janet fibers are stackful coroutines, a `go` block is just its body run in a fiber — no CPS/state-machine rewrite — so `<!`/`>!` work *anywhere*, including inside `try`, nested `fn`s, and loops (positions Clojure's `go` macro forbids). Go blocks are cooperatively scheduled on one OS thread, so parking (`<!`) and blocking (`<!!`) coincide; `thread` runs cooperatively too. Dynamic-var bindings are conveyed into `go` blocks (each go block sees the bindings in effect when it was spawned).
- **Regex.** Compiled to Janet's PEG engine (Janet has no regex). Supported: capturing groups (`[whole g1 …]`), greedy and lazy quantifiers with backtracking, `(?:…)`, lookahead `(?=…)`/`(?!…)`, alternation, anchors `^ $ \b \B`, character classes, and the `(?i)` flag. Not supported: lookbehind, backreferences (`\1`), and named groups (`(?<name>…)`).
- **Arrays.** Java-style arrays map onto Janet's native types: `byte-array` is a Janet buffer (contiguous, C-backed); `object-array`/`int-array`/`double-array`/etc. are Janet arrays. `aget`/`aset`/`alength`/`aclone` work over both.